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"
17 #include "X86InstrBuilder.h"
18 #include "X86ISelLowering.h"
19 #include "X86TargetMachine.h"
20 #include "X86TargetObjectFile.h"
21 #include "Utils/X86ShuffleDecode.h"
22 #include "llvm/CallingConv.h"
23 #include "llvm/Constants.h"
24 #include "llvm/DerivedTypes.h"
25 #include "llvm/GlobalAlias.h"
26 #include "llvm/GlobalVariable.h"
27 #include "llvm/Function.h"
28 #include "llvm/Instructions.h"
29 #include "llvm/Intrinsics.h"
30 #include "llvm/LLVMContext.h"
31 #include "llvm/CodeGen/IntrinsicLowering.h"
32 #include "llvm/CodeGen/MachineFrameInfo.h"
33 #include "llvm/CodeGen/MachineFunction.h"
34 #include "llvm/CodeGen/MachineInstrBuilder.h"
35 #include "llvm/CodeGen/MachineJumpTableInfo.h"
36 #include "llvm/CodeGen/MachineModuleInfo.h"
37 #include "llvm/CodeGen/MachineRegisterInfo.h"
38 #include "llvm/CodeGen/PseudoSourceValue.h"
39 #include "llvm/MC/MCAsmInfo.h"
40 #include "llvm/MC/MCContext.h"
41 #include "llvm/MC/MCExpr.h"
42 #include "llvm/MC/MCSymbol.h"
43 #include "llvm/ADT/BitVector.h"
44 #include "llvm/ADT/SmallSet.h"
45 #include "llvm/ADT/Statistic.h"
46 #include "llvm/ADT/StringExtras.h"
47 #include "llvm/ADT/VectorExtras.h"
48 #include "llvm/Support/CallSite.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/Dwarf.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Support/raw_ostream.h"
55 using namespace dwarf;
57 STATISTIC(NumTailCalls, "Number of tail calls");
59 // Forward declarations.
60 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
63 static SDValue Insert128BitVector(SDValue Result,
69 static SDValue Extract128BitVector(SDValue Vec,
74 static SDValue ConcatVectors(SDValue Lower, SDValue Upper, SelectionDAG &DAG);
77 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
78 /// sets things up to match to an AVX VEXTRACTF128 instruction or a
79 /// simple subregister reference. Idx is an index in the 128 bits we
80 /// want. It need not be aligned to a 128-bit bounday. That makes
81 /// lowering EXTRACT_VECTOR_ELT operations easier.
82 static SDValue Extract128BitVector(SDValue Vec,
86 EVT VT = Vec.getValueType();
87 assert(VT.getSizeInBits() == 256 && "Unexpected vector size!");
89 EVT ElVT = VT.getVectorElementType();
91 int Factor = VT.getSizeInBits() / 128;
93 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(),
95 VT.getVectorNumElements() / Factor);
97 // Extract from UNDEF is UNDEF.
98 if (Vec.getOpcode() == ISD::UNDEF)
99 return DAG.getNode(ISD::UNDEF, dl, ResultVT);
101 if (isa<ConstantSDNode>(Idx)) {
102 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
104 // Extract the relevant 128 bits. Generate an EXTRACT_SUBVECTOR
105 // we can match to VEXTRACTF128.
106 unsigned ElemsPerChunk = 128 / ElVT.getSizeInBits();
108 // This is the index of the first element of the 128-bit chunk
110 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / 128)
113 SDValue VecIdx = DAG.getConstant(NormalizedIdxVal, MVT::i32);
115 SDValue Result = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec,
124 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
125 /// sets things up to match to an AVX VINSERTF128 instruction or a
126 /// simple superregister reference. Idx is an index in the 128 bits
127 /// we want. It need not be aligned to a 128-bit bounday. That makes
128 /// lowering INSERT_VECTOR_ELT operations easier.
129 static SDValue Insert128BitVector(SDValue Result,
134 if (isa<ConstantSDNode>(Idx)) {
135 EVT VT = Vec.getValueType();
136 assert(VT.getSizeInBits() == 128 && "Unexpected vector size!");
138 EVT ElVT = VT.getVectorElementType();
140 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
142 EVT ResultVT = Result.getValueType();
144 // Insert the relevant 128 bits.
145 unsigned ElemsPerChunk = 128 / ElVT.getSizeInBits();
147 // This is the index of the first element of the 128-bit chunk
149 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / 128)
152 SDValue VecIdx = DAG.getConstant(NormalizedIdxVal, MVT::i32);
154 Result = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec,
162 /// Given two vectors, concat them.
163 static SDValue ConcatVectors(SDValue Lower, SDValue Upper, SelectionDAG &DAG) {
164 DebugLoc dl = Lower.getDebugLoc();
166 assert(Lower.getValueType() == Upper.getValueType() && "Mismatched vectors!");
168 EVT VT = EVT::getVectorVT(*DAG.getContext(),
169 Lower.getValueType().getVectorElementType(),
170 Lower.getValueType().getVectorNumElements() * 2);
172 // TODO: Generalize to arbitrary vector length (this assumes 256-bit vectors).
173 assert(VT.getSizeInBits() == 256 && "Unsupported vector concat!");
175 // Insert the upper subvector.
176 SDValue Vec = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, VT), Upper,
178 // This is half the length of the result
179 // vector. Start inserting the upper 128
181 Lower.getValueType().getVectorNumElements(),
185 // Insert the lower subvector.
186 Vec = Insert128BitVector(Vec, Lower, DAG.getConstant(0, MVT::i32), DAG, dl);
190 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
191 const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>();
192 bool is64Bit = Subtarget->is64Bit();
194 if (Subtarget->isTargetEnvMacho()) {
196 return new X8664_MachoTargetObjectFile();
197 return new TargetLoweringObjectFileMachO();
200 if (Subtarget->isTargetELF()) {
202 return new X8664_ELFTargetObjectFile(TM);
203 return new X8632_ELFTargetObjectFile(TM);
205 if (Subtarget->isTargetCOFF() && !Subtarget->isTargetEnvMacho())
206 return new TargetLoweringObjectFileCOFF();
207 llvm_unreachable("unknown subtarget type");
210 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
211 : TargetLowering(TM, createTLOF(TM)) {
212 Subtarget = &TM.getSubtarget<X86Subtarget>();
213 X86ScalarSSEf64 = Subtarget->hasXMMInt();
214 X86ScalarSSEf32 = Subtarget->hasXMM();
215 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
217 RegInfo = TM.getRegisterInfo();
218 TD = getTargetData();
220 // Set up the TargetLowering object.
221 static MVT IntVTs[] = { MVT::i8, MVT::i16, MVT::i32, MVT::i64 };
223 // X86 is weird, it always uses i8 for shift amounts and setcc results.
224 setBooleanContents(ZeroOrOneBooleanContent);
226 // For 64-bit since we have so many registers use the ILP scheduler, for
227 // 32-bit code use the register pressure specific scheduling.
228 if (Subtarget->is64Bit())
229 setSchedulingPreference(Sched::ILP);
231 setSchedulingPreference(Sched::RegPressure);
232 setStackPointerRegisterToSaveRestore(X86StackPtr);
234 if (Subtarget->isTargetWindows() && !Subtarget->isTargetCygMing()) {
235 // Setup Windows compiler runtime calls.
236 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
237 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
238 setLibcallName(RTLIB::FPTOUINT_F64_I64, "_ftol2");
239 setLibcallName(RTLIB::FPTOUINT_F32_I64, "_ftol2");
240 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
241 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
242 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::C);
243 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::C);
246 if (Subtarget->isTargetDarwin()) {
247 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
248 setUseUnderscoreSetJmp(false);
249 setUseUnderscoreLongJmp(false);
250 } else if (Subtarget->isTargetMingw()) {
251 // MS runtime is weird: it exports _setjmp, but longjmp!
252 setUseUnderscoreSetJmp(true);
253 setUseUnderscoreLongJmp(false);
255 setUseUnderscoreSetJmp(true);
256 setUseUnderscoreLongJmp(true);
259 // Set up the register classes.
260 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
261 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
262 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
263 if (Subtarget->is64Bit())
264 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
266 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
268 // We don't accept any truncstore of integer registers.
269 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
270 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
271 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
272 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
273 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
274 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
276 // SETOEQ and SETUNE require checking two conditions.
277 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
278 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
279 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
280 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
281 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
282 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
284 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
286 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
287 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
288 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
290 if (Subtarget->is64Bit()) {
291 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
292 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
293 } else if (!UseSoftFloat) {
294 // We have an algorithm for SSE2->double, and we turn this into a
295 // 64-bit FILD followed by conditional FADD for other targets.
296 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
297 // We have an algorithm for SSE2, and we turn this into a 64-bit
298 // FILD for other targets.
299 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
302 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
304 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
305 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
308 // SSE has no i16 to fp conversion, only i32
309 if (X86ScalarSSEf32) {
310 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
311 // f32 and f64 cases are Legal, f80 case is not
312 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
314 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
315 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
318 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
319 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
322 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
323 // are Legal, f80 is custom lowered.
324 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
325 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
327 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
329 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
330 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
332 if (X86ScalarSSEf32) {
333 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
334 // f32 and f64 cases are Legal, f80 case is not
335 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
337 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
338 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
341 // Handle FP_TO_UINT by promoting the destination to a larger signed
343 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
344 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
345 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
347 if (Subtarget->is64Bit()) {
348 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
349 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
350 } else if (!UseSoftFloat) {
351 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
352 // Expand FP_TO_UINT into a select.
353 // FIXME: We would like to use a Custom expander here eventually to do
354 // the optimal thing for SSE vs. the default expansion in the legalizer.
355 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
357 // With SSE3 we can use fisttpll to convert to a signed i64; without
358 // SSE, we're stuck with a fistpll.
359 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
362 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
363 if (!X86ScalarSSEf64) {
364 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
365 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
366 if (Subtarget->is64Bit()) {
367 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
368 // Without SSE, i64->f64 goes through memory.
369 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
373 // Scalar integer divide and remainder are lowered to use operations that
374 // produce two results, to match the available instructions. This exposes
375 // the two-result form to trivial CSE, which is able to combine x/y and x%y
376 // into a single instruction.
378 // Scalar integer multiply-high is also lowered to use two-result
379 // operations, to match the available instructions. However, plain multiply
380 // (low) operations are left as Legal, as there are single-result
381 // instructions for this in x86. Using the two-result multiply instructions
382 // when both high and low results are needed must be arranged by dagcombine.
383 for (unsigned i = 0, e = 4; i != e; ++i) {
385 setOperationAction(ISD::MULHS, VT, Expand);
386 setOperationAction(ISD::MULHU, VT, Expand);
387 setOperationAction(ISD::SDIV, VT, Expand);
388 setOperationAction(ISD::UDIV, VT, Expand);
389 setOperationAction(ISD::SREM, VT, Expand);
390 setOperationAction(ISD::UREM, VT, Expand);
392 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
393 setOperationAction(ISD::ADDC, VT, Custom);
394 setOperationAction(ISD::ADDE, VT, Custom);
395 setOperationAction(ISD::SUBC, VT, Custom);
396 setOperationAction(ISD::SUBE, VT, Custom);
399 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
400 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
401 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
402 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
403 if (Subtarget->is64Bit())
404 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
405 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
406 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
407 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
408 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
409 setOperationAction(ISD::FREM , MVT::f32 , Expand);
410 setOperationAction(ISD::FREM , MVT::f64 , Expand);
411 setOperationAction(ISD::FREM , MVT::f80 , Expand);
412 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
414 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
415 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
416 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
417 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
418 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
419 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
420 if (Subtarget->is64Bit()) {
421 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
422 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
425 if (Subtarget->hasPOPCNT()) {
426 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
428 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
429 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
430 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
431 if (Subtarget->is64Bit())
432 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
435 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
436 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
438 // These should be promoted to a larger select which is supported.
439 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
440 // X86 wants to expand cmov itself.
441 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
442 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
443 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
444 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
445 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
446 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
447 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
448 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
449 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
450 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
451 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
452 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
453 if (Subtarget->is64Bit()) {
454 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
455 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
457 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
460 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
461 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
462 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
463 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
464 if (Subtarget->is64Bit())
465 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
466 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
467 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
468 if (Subtarget->is64Bit()) {
469 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
470 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
471 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
472 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
473 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
475 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
476 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
477 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
478 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
479 if (Subtarget->is64Bit()) {
480 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
481 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
482 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
485 if (Subtarget->hasXMM())
486 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
488 // We may not have a libcall for MEMBARRIER so we should lower this.
489 setOperationAction(ISD::MEMBARRIER , MVT::Other, Custom);
491 // On X86 and X86-64, atomic operations are lowered to locked instructions.
492 // Locked instructions, in turn, have implicit fence semantics (all memory
493 // operations are flushed before issuing the locked instruction, and they
494 // are not buffered), so we can fold away the common pattern of
495 // fence-atomic-fence.
496 setShouldFoldAtomicFences(true);
498 // Expand certain atomics
499 for (unsigned i = 0, e = 4; i != e; ++i) {
501 setOperationAction(ISD::ATOMIC_CMP_SWAP, VT, Custom);
502 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
505 if (!Subtarget->is64Bit()) {
506 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
507 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
508 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
509 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
510 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
511 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
512 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
515 // FIXME - use subtarget debug flags
516 if (!Subtarget->isTargetDarwin() &&
517 !Subtarget->isTargetELF() &&
518 !Subtarget->isTargetCygMing()) {
519 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
522 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
523 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
524 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
525 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
526 if (Subtarget->is64Bit()) {
527 setExceptionPointerRegister(X86::RAX);
528 setExceptionSelectorRegister(X86::RDX);
530 setExceptionPointerRegister(X86::EAX);
531 setExceptionSelectorRegister(X86::EDX);
533 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
534 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
536 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
538 setOperationAction(ISD::TRAP, MVT::Other, Legal);
540 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
541 setOperationAction(ISD::VASTART , MVT::Other, Custom);
542 setOperationAction(ISD::VAEND , MVT::Other, Expand);
543 if (Subtarget->is64Bit()) {
544 setOperationAction(ISD::VAARG , MVT::Other, Custom);
545 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
547 setOperationAction(ISD::VAARG , MVT::Other, Expand);
548 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
551 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
552 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
553 setOperationAction(ISD::DYNAMIC_STACKALLOC,
554 (Subtarget->is64Bit() ? MVT::i64 : MVT::i32),
555 (Subtarget->isTargetCOFF()
556 && !Subtarget->isTargetEnvMacho()
559 if (!UseSoftFloat && X86ScalarSSEf64) {
560 // f32 and f64 use SSE.
561 // Set up the FP register classes.
562 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
563 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
565 // Use ANDPD to simulate FABS.
566 setOperationAction(ISD::FABS , MVT::f64, Custom);
567 setOperationAction(ISD::FABS , MVT::f32, Custom);
569 // Use XORP to simulate FNEG.
570 setOperationAction(ISD::FNEG , MVT::f64, Custom);
571 setOperationAction(ISD::FNEG , MVT::f32, Custom);
573 // Use ANDPD and ORPD to simulate FCOPYSIGN.
574 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
575 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
577 // We don't support sin/cos/fmod
578 setOperationAction(ISD::FSIN , MVT::f64, Expand);
579 setOperationAction(ISD::FCOS , MVT::f64, Expand);
580 setOperationAction(ISD::FSIN , MVT::f32, Expand);
581 setOperationAction(ISD::FCOS , MVT::f32, Expand);
583 // Expand FP immediates into loads from the stack, except for the special
585 addLegalFPImmediate(APFloat(+0.0)); // xorpd
586 addLegalFPImmediate(APFloat(+0.0f)); // xorps
587 } else if (!UseSoftFloat && X86ScalarSSEf32) {
588 // Use SSE for f32, x87 for f64.
589 // Set up the FP register classes.
590 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
591 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
593 // Use ANDPS to simulate FABS.
594 setOperationAction(ISD::FABS , MVT::f32, Custom);
596 // Use XORP to simulate FNEG.
597 setOperationAction(ISD::FNEG , MVT::f32, Custom);
599 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
601 // Use ANDPS and ORPS to simulate FCOPYSIGN.
602 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
603 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
605 // We don't support sin/cos/fmod
606 setOperationAction(ISD::FSIN , MVT::f32, Expand);
607 setOperationAction(ISD::FCOS , MVT::f32, Expand);
609 // Special cases we handle for FP constants.
610 addLegalFPImmediate(APFloat(+0.0f)); // xorps
611 addLegalFPImmediate(APFloat(+0.0)); // FLD0
612 addLegalFPImmediate(APFloat(+1.0)); // FLD1
613 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
614 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
617 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
618 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
620 } else if (!UseSoftFloat) {
621 // f32 and f64 in x87.
622 // Set up the FP register classes.
623 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
624 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
626 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
627 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
628 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
629 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
632 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
633 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
635 addLegalFPImmediate(APFloat(+0.0)); // FLD0
636 addLegalFPImmediate(APFloat(+1.0)); // FLD1
637 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
638 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
639 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
640 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
641 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
642 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
645 // Long double always uses X87.
647 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
648 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
649 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
651 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
652 addLegalFPImmediate(TmpFlt); // FLD0
654 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
657 APFloat TmpFlt2(+1.0);
658 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
660 addLegalFPImmediate(TmpFlt2); // FLD1
661 TmpFlt2.changeSign();
662 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
666 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
667 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
671 // Always use a library call for pow.
672 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
673 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
674 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
676 setOperationAction(ISD::FLOG, MVT::f80, Expand);
677 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
678 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
679 setOperationAction(ISD::FEXP, MVT::f80, Expand);
680 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
682 // First set operation action for all vector types to either promote
683 // (for widening) or expand (for scalarization). Then we will selectively
684 // turn on ones that can be effectively codegen'd.
685 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
686 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
687 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
688 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
689 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
690 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
691 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
692 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
693 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
694 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
695 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
696 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
697 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
698 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
699 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
700 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
701 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
702 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
703 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
704 setOperationAction(ISD::INSERT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
705 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
706 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
707 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
708 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
709 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
710 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
711 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
712 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
713 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
714 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
715 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
716 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
717 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
718 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
719 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
720 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
721 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
722 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
723 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
724 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
725 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
726 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
727 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
728 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
729 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
730 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
731 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
732 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
733 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
734 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
735 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
736 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
737 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
738 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
739 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
740 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
741 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
742 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
743 setTruncStoreAction((MVT::SimpleValueType)VT,
744 (MVT::SimpleValueType)InnerVT, Expand);
745 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
746 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
747 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
750 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
751 // with -msoft-float, disable use of MMX as well.
752 if (!UseSoftFloat && Subtarget->hasMMX()) {
753 addRegisterClass(MVT::x86mmx, X86::VR64RegisterClass);
754 // No operations on x86mmx supported, everything uses intrinsics.
757 // MMX-sized vectors (other than x86mmx) are expected to be expanded
758 // into smaller operations.
759 setOperationAction(ISD::MULHS, MVT::v8i8, Expand);
760 setOperationAction(ISD::MULHS, MVT::v4i16, Expand);
761 setOperationAction(ISD::MULHS, MVT::v2i32, Expand);
762 setOperationAction(ISD::MULHS, MVT::v1i64, Expand);
763 setOperationAction(ISD::AND, MVT::v8i8, Expand);
764 setOperationAction(ISD::AND, MVT::v4i16, Expand);
765 setOperationAction(ISD::AND, MVT::v2i32, Expand);
766 setOperationAction(ISD::AND, MVT::v1i64, Expand);
767 setOperationAction(ISD::OR, MVT::v8i8, Expand);
768 setOperationAction(ISD::OR, MVT::v4i16, Expand);
769 setOperationAction(ISD::OR, MVT::v2i32, Expand);
770 setOperationAction(ISD::OR, MVT::v1i64, Expand);
771 setOperationAction(ISD::XOR, MVT::v8i8, Expand);
772 setOperationAction(ISD::XOR, MVT::v4i16, Expand);
773 setOperationAction(ISD::XOR, MVT::v2i32, Expand);
774 setOperationAction(ISD::XOR, MVT::v1i64, Expand);
775 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Expand);
776 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Expand);
777 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Expand);
778 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Expand);
779 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
780 setOperationAction(ISD::SELECT, MVT::v8i8, Expand);
781 setOperationAction(ISD::SELECT, MVT::v4i16, Expand);
782 setOperationAction(ISD::SELECT, MVT::v2i32, Expand);
783 setOperationAction(ISD::SELECT, MVT::v1i64, Expand);
784 setOperationAction(ISD::BITCAST, MVT::v8i8, Expand);
785 setOperationAction(ISD::BITCAST, MVT::v4i16, Expand);
786 setOperationAction(ISD::BITCAST, MVT::v2i32, Expand);
787 setOperationAction(ISD::BITCAST, MVT::v1i64, Expand);
789 if (!UseSoftFloat && Subtarget->hasXMM()) {
790 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
792 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
793 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
794 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
795 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
796 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
797 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
798 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
799 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
800 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
801 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
802 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
803 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
806 if (!UseSoftFloat && Subtarget->hasXMMInt()) {
807 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
809 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
810 // registers cannot be used even for integer operations.
811 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
812 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
813 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
814 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
816 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
817 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
818 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
819 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
820 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
821 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
822 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
823 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
824 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
825 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
826 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
827 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
828 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
829 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
830 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
831 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
833 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
834 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
835 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
836 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
838 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
839 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
840 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
841 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
842 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
844 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
845 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
846 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
847 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
848 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
850 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
851 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
852 EVT VT = (MVT::SimpleValueType)i;
853 // Do not attempt to custom lower non-power-of-2 vectors
854 if (!isPowerOf2_32(VT.getVectorNumElements()))
856 // Do not attempt to custom lower non-128-bit vectors
857 if (!VT.is128BitVector())
859 setOperationAction(ISD::BUILD_VECTOR,
860 VT.getSimpleVT().SimpleTy, Custom);
861 setOperationAction(ISD::VECTOR_SHUFFLE,
862 VT.getSimpleVT().SimpleTy, Custom);
863 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
864 VT.getSimpleVT().SimpleTy, Custom);
867 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
868 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
869 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
870 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
871 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
872 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
874 if (Subtarget->is64Bit()) {
875 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
876 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
879 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
880 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
881 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
884 // Do not attempt to promote non-128-bit vectors
885 if (!VT.is128BitVector())
888 setOperationAction(ISD::AND, SVT, Promote);
889 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
890 setOperationAction(ISD::OR, SVT, Promote);
891 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
892 setOperationAction(ISD::XOR, SVT, Promote);
893 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
894 setOperationAction(ISD::LOAD, SVT, Promote);
895 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
896 setOperationAction(ISD::SELECT, SVT, Promote);
897 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
900 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
902 // Custom lower v2i64 and v2f64 selects.
903 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
904 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
905 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
906 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
908 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
909 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
912 if (Subtarget->hasSSE41()) {
913 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
914 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
915 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
916 setOperationAction(ISD::FRINT, MVT::f32, Legal);
917 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
918 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
919 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
920 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
921 setOperationAction(ISD::FRINT, MVT::f64, Legal);
922 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
924 // FIXME: Do we need to handle scalar-to-vector here?
925 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
927 // Can turn SHL into an integer multiply.
928 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
929 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
931 // i8 and i16 vectors are custom , because the source register and source
932 // source memory operand types are not the same width. f32 vectors are
933 // custom since the immediate controlling the insert encodes additional
935 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
936 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
937 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
938 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
940 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
941 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
942 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
943 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
945 if (Subtarget->is64Bit()) {
946 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
947 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
951 if (Subtarget->hasSSE2()) {
952 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
953 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
954 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
956 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
957 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
958 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
960 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
961 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
964 if (Subtarget->hasSSE42())
965 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
967 if (!UseSoftFloat && Subtarget->hasAVX()) {
968 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
969 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
970 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
971 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
972 addRegisterClass(MVT::v32i8, X86::VR256RegisterClass);
974 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
975 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
976 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
977 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
979 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
980 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
981 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
982 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
983 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
984 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
986 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
987 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
988 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
989 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
990 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
991 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
993 // Custom lower build_vector, vector_shuffle, scalar_to_vector,
994 // insert_vector_elt extract_subvector and extract_vector_elt for
996 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
997 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE;
999 MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
1000 // Do not attempt to custom lower non-256-bit vectors
1001 if (!isPowerOf2_32(MVT(VT).getVectorNumElements())
1002 || (MVT(VT).getSizeInBits() < 256))
1004 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1005 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1006 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1007 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1008 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1010 // Custom-lower insert_subvector and extract_subvector based on
1012 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
1013 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE;
1015 MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
1016 // Do not attempt to custom lower non-256-bit vectors
1017 if (!isPowerOf2_32(MVT(VT).getVectorNumElements()))
1020 if (MVT(VT).getSizeInBits() == 128) {
1021 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1023 else if (MVT(VT).getSizeInBits() == 256) {
1024 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1028 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1029 // Don't promote loads because we need them for VPERM vector index versions.
1031 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
1032 VT != (unsigned)MVT::LAST_VECTOR_VALUETYPE;
1034 if (!isPowerOf2_32(MVT((MVT::SimpleValueType)VT).getVectorNumElements())
1035 || (MVT((MVT::SimpleValueType)VT).getSizeInBits() < 256))
1037 setOperationAction(ISD::AND, (MVT::SimpleValueType)VT, Promote);
1038 AddPromotedToType (ISD::AND, (MVT::SimpleValueType)VT, MVT::v4i64);
1039 setOperationAction(ISD::OR, (MVT::SimpleValueType)VT, Promote);
1040 AddPromotedToType (ISD::OR, (MVT::SimpleValueType)VT, MVT::v4i64);
1041 setOperationAction(ISD::XOR, (MVT::SimpleValueType)VT, Promote);
1042 AddPromotedToType (ISD::XOR, (MVT::SimpleValueType)VT, MVT::v4i64);
1043 //setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Promote);
1044 //AddPromotedToType (ISD::LOAD, (MVT::SimpleValueType)VT, MVT::v4i64);
1045 setOperationAction(ISD::SELECT, (MVT::SimpleValueType)VT, Promote);
1046 AddPromotedToType (ISD::SELECT, (MVT::SimpleValueType)VT, MVT::v4i64);
1050 // We want to custom lower some of our intrinsics.
1051 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1054 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1055 // handle type legalization for these operations here.
1057 // FIXME: We really should do custom legalization for addition and
1058 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1059 // than generic legalization for 64-bit multiplication-with-overflow, though.
1060 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1061 // Add/Sub/Mul with overflow operations are custom lowered.
1063 setOperationAction(ISD::SADDO, VT, Custom);
1064 setOperationAction(ISD::UADDO, VT, Custom);
1065 setOperationAction(ISD::SSUBO, VT, Custom);
1066 setOperationAction(ISD::USUBO, VT, Custom);
1067 setOperationAction(ISD::SMULO, VT, Custom);
1068 setOperationAction(ISD::UMULO, VT, Custom);
1071 // There are no 8-bit 3-address imul/mul instructions
1072 setOperationAction(ISD::SMULO, MVT::i8, Expand);
1073 setOperationAction(ISD::UMULO, MVT::i8, Expand);
1075 if (!Subtarget->is64Bit()) {
1076 // These libcalls are not available in 32-bit.
1077 setLibcallName(RTLIB::SHL_I128, 0);
1078 setLibcallName(RTLIB::SRL_I128, 0);
1079 setLibcallName(RTLIB::SRA_I128, 0);
1082 // We have target-specific dag combine patterns for the following nodes:
1083 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1084 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1085 setTargetDAGCombine(ISD::BUILD_VECTOR);
1086 setTargetDAGCombine(ISD::SELECT);
1087 setTargetDAGCombine(ISD::SHL);
1088 setTargetDAGCombine(ISD::SRA);
1089 setTargetDAGCombine(ISD::SRL);
1090 setTargetDAGCombine(ISD::OR);
1091 setTargetDAGCombine(ISD::AND);
1092 setTargetDAGCombine(ISD::ADD);
1093 setTargetDAGCombine(ISD::SUB);
1094 setTargetDAGCombine(ISD::STORE);
1095 setTargetDAGCombine(ISD::ZERO_EXTEND);
1096 if (Subtarget->is64Bit())
1097 setTargetDAGCombine(ISD::MUL);
1099 computeRegisterProperties();
1101 // On Darwin, -Os means optimize for size without hurting performance,
1102 // do not reduce the limit.
1103 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1104 maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
1105 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1106 maxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1107 maxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1108 maxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1109 setPrefLoopAlignment(16);
1110 benefitFromCodePlacementOpt = true;
1112 setPrefFunctionAlignment(4);
1116 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1121 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1122 /// the desired ByVal argument alignment.
1123 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1126 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1127 if (VTy->getBitWidth() == 128)
1129 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1130 unsigned EltAlign = 0;
1131 getMaxByValAlign(ATy->getElementType(), EltAlign);
1132 if (EltAlign > MaxAlign)
1133 MaxAlign = EltAlign;
1134 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1135 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1136 unsigned EltAlign = 0;
1137 getMaxByValAlign(STy->getElementType(i), EltAlign);
1138 if (EltAlign > MaxAlign)
1139 MaxAlign = EltAlign;
1147 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1148 /// function arguments in the caller parameter area. For X86, aggregates
1149 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1150 /// are at 4-byte boundaries.
1151 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1152 if (Subtarget->is64Bit()) {
1153 // Max of 8 and alignment of type.
1154 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1161 if (Subtarget->hasXMM())
1162 getMaxByValAlign(Ty, Align);
1166 /// getOptimalMemOpType - Returns the target specific optimal type for load
1167 /// and store operations as a result of memset, memcpy, and memmove
1168 /// lowering. If DstAlign is zero that means it's safe to destination
1169 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1170 /// means there isn't a need to check it against alignment requirement,
1171 /// probably because the source does not need to be loaded. If
1172 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1173 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1174 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1175 /// constant so it does not need to be loaded.
1176 /// It returns EVT::Other if the type should be determined using generic
1177 /// target-independent logic.
1179 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1180 unsigned DstAlign, unsigned SrcAlign,
1181 bool NonScalarIntSafe,
1183 MachineFunction &MF) const {
1184 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1185 // linux. This is because the stack realignment code can't handle certain
1186 // cases like PR2962. This should be removed when PR2962 is fixed.
1187 const Function *F = MF.getFunction();
1188 if (NonScalarIntSafe &&
1189 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1191 (Subtarget->isUnalignedMemAccessFast() ||
1192 ((DstAlign == 0 || DstAlign >= 16) &&
1193 (SrcAlign == 0 || SrcAlign >= 16))) &&
1194 Subtarget->getStackAlignment() >= 16) {
1195 if (Subtarget->hasSSE2())
1197 if (Subtarget->hasSSE1())
1199 } else if (!MemcpyStrSrc && Size >= 8 &&
1200 !Subtarget->is64Bit() &&
1201 Subtarget->getStackAlignment() >= 8 &&
1202 Subtarget->hasXMMInt()) {
1203 // Do not use f64 to lower memcpy if source is string constant. It's
1204 // better to use i32 to avoid the loads.
1208 if (Subtarget->is64Bit() && Size >= 8)
1213 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1214 /// current function. The returned value is a member of the
1215 /// MachineJumpTableInfo::JTEntryKind enum.
1216 unsigned X86TargetLowering::getJumpTableEncoding() const {
1217 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1219 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1220 Subtarget->isPICStyleGOT())
1221 return MachineJumpTableInfo::EK_Custom32;
1223 // Otherwise, use the normal jump table encoding heuristics.
1224 return TargetLowering::getJumpTableEncoding();
1228 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1229 const MachineBasicBlock *MBB,
1230 unsigned uid,MCContext &Ctx) const{
1231 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1232 Subtarget->isPICStyleGOT());
1233 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1235 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1236 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1239 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1241 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1242 SelectionDAG &DAG) const {
1243 if (!Subtarget->is64Bit())
1244 // This doesn't have DebugLoc associated with it, but is not really the
1245 // same as a Register.
1246 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1250 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1251 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1253 const MCExpr *X86TargetLowering::
1254 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1255 MCContext &Ctx) const {
1256 // X86-64 uses RIP relative addressing based on the jump table label.
1257 if (Subtarget->isPICStyleRIPRel())
1258 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1260 // Otherwise, the reference is relative to the PIC base.
1261 return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
1264 // FIXME: Why this routine is here? Move to RegInfo!
1265 std::pair<const TargetRegisterClass*, uint8_t>
1266 X86TargetLowering::findRepresentativeClass(EVT VT) const{
1267 const TargetRegisterClass *RRC = 0;
1269 switch (VT.getSimpleVT().SimpleTy) {
1271 return TargetLowering::findRepresentativeClass(VT);
1272 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1273 RRC = (Subtarget->is64Bit()
1274 ? X86::GR64RegisterClass : X86::GR32RegisterClass);
1277 RRC = X86::VR64RegisterClass;
1279 case MVT::f32: case MVT::f64:
1280 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1281 case MVT::v4f32: case MVT::v2f64:
1282 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1284 RRC = X86::VR128RegisterClass;
1287 return std::make_pair(RRC, Cost);
1290 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1291 unsigned &Offset) const {
1292 if (!Subtarget->isTargetLinux())
1295 if (Subtarget->is64Bit()) {
1296 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1298 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1311 //===----------------------------------------------------------------------===//
1312 // Return Value Calling Convention Implementation
1313 //===----------------------------------------------------------------------===//
1315 #include "X86GenCallingConv.inc"
1318 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1319 const SmallVectorImpl<ISD::OutputArg> &Outs,
1320 LLVMContext &Context) const {
1321 SmallVector<CCValAssign, 16> RVLocs;
1322 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1324 return CCInfo.CheckReturn(Outs, RetCC_X86);
1328 X86TargetLowering::LowerReturn(SDValue Chain,
1329 CallingConv::ID CallConv, bool isVarArg,
1330 const SmallVectorImpl<ISD::OutputArg> &Outs,
1331 const SmallVectorImpl<SDValue> &OutVals,
1332 DebugLoc dl, SelectionDAG &DAG) const {
1333 MachineFunction &MF = DAG.getMachineFunction();
1334 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1336 SmallVector<CCValAssign, 16> RVLocs;
1337 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1338 RVLocs, *DAG.getContext());
1339 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1341 // Add the regs to the liveout set for the function.
1342 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1343 for (unsigned i = 0; i != RVLocs.size(); ++i)
1344 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1345 MRI.addLiveOut(RVLocs[i].getLocReg());
1349 SmallVector<SDValue, 6> RetOps;
1350 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1351 // Operand #1 = Bytes To Pop
1352 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1355 // Copy the result values into the output registers.
1356 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1357 CCValAssign &VA = RVLocs[i];
1358 assert(VA.isRegLoc() && "Can only return in registers!");
1359 SDValue ValToCopy = OutVals[i];
1360 EVT ValVT = ValToCopy.getValueType();
1362 // If this is x86-64, and we disabled SSE, we can't return FP values,
1363 // or SSE or MMX vectors.
1364 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
1365 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
1366 (Subtarget->is64Bit() && !Subtarget->hasXMM())) {
1367 report_fatal_error("SSE register return with SSE disabled");
1369 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1370 // llvm-gcc has never done it right and no one has noticed, so this
1371 // should be OK for now.
1372 if (ValVT == MVT::f64 &&
1373 (Subtarget->is64Bit() && !Subtarget->hasXMMInt()))
1374 report_fatal_error("SSE2 register return with SSE2 disabled");
1376 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1377 // the RET instruction and handled by the FP Stackifier.
1378 if (VA.getLocReg() == X86::ST0 ||
1379 VA.getLocReg() == X86::ST1) {
1380 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1381 // change the value to the FP stack register class.
1382 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1383 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1384 RetOps.push_back(ValToCopy);
1385 // Don't emit a copytoreg.
1389 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1390 // which is returned in RAX / RDX.
1391 if (Subtarget->is64Bit()) {
1392 if (ValVT == MVT::x86mmx) {
1393 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1394 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ValToCopy);
1395 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
1397 // If we don't have SSE2 available, convert to v4f32 so the generated
1398 // register is legal.
1399 if (!Subtarget->hasSSE2())
1400 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,ValToCopy);
1405 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1406 Flag = Chain.getValue(1);
1409 // The x86-64 ABI for returning structs by value requires that we copy
1410 // the sret argument into %rax for the return. We saved the argument into
1411 // a virtual register in the entry block, so now we copy the value out
1413 if (Subtarget->is64Bit() &&
1414 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1415 MachineFunction &MF = DAG.getMachineFunction();
1416 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1417 unsigned Reg = FuncInfo->getSRetReturnReg();
1419 "SRetReturnReg should have been set in LowerFormalArguments().");
1420 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1422 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1423 Flag = Chain.getValue(1);
1425 // RAX now acts like a return value.
1426 MRI.addLiveOut(X86::RAX);
1429 RetOps[0] = Chain; // Update chain.
1431 // Add the flag if we have it.
1433 RetOps.push_back(Flag);
1435 return DAG.getNode(X86ISD::RET_FLAG, dl,
1436 MVT::Other, &RetOps[0], RetOps.size());
1439 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N) const {
1440 if (N->getNumValues() != 1)
1442 if (!N->hasNUsesOfValue(1, 0))
1445 SDNode *Copy = *N->use_begin();
1446 if (Copy->getOpcode() != ISD::CopyToReg &&
1447 Copy->getOpcode() != ISD::FP_EXTEND)
1450 bool HasRet = false;
1451 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1453 if (UI->getOpcode() != X86ISD::RET_FLAG)
1462 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
1463 ISD::NodeType ExtendKind) const {
1465 // TODO: Is this also valid on 32-bit?
1466 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
1467 ReturnMVT = MVT::i8;
1469 ReturnMVT = MVT::i32;
1471 EVT MinVT = getRegisterType(Context, ReturnMVT);
1472 return VT.bitsLT(MinVT) ? MinVT : VT;
1475 /// LowerCallResult - Lower the result values of a call into the
1476 /// appropriate copies out of appropriate physical registers.
1479 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1480 CallingConv::ID CallConv, bool isVarArg,
1481 const SmallVectorImpl<ISD::InputArg> &Ins,
1482 DebugLoc dl, SelectionDAG &DAG,
1483 SmallVectorImpl<SDValue> &InVals) const {
1485 // Assign locations to each value returned by this call.
1486 SmallVector<CCValAssign, 16> RVLocs;
1487 bool Is64Bit = Subtarget->is64Bit();
1488 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1489 RVLocs, *DAG.getContext());
1490 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1492 // Copy all of the result registers out of their specified physreg.
1493 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1494 CCValAssign &VA = RVLocs[i];
1495 EVT CopyVT = VA.getValVT();
1497 // If this is x86-64, and we disabled SSE, we can't return FP values
1498 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1499 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasXMM())) {
1500 report_fatal_error("SSE register return with SSE disabled");
1505 // If this is a call to a function that returns an fp value on the floating
1506 // point stack, we must guarantee the the value is popped from the stack, so
1507 // a CopyFromReg is not good enough - the copy instruction may be eliminated
1508 // if the return value is not used. We use the FpGET_ST0 instructions
1510 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
1511 // If we prefer to use the value in xmm registers, copy it out as f80 and
1512 // use a truncate to move it from fp stack reg to xmm reg.
1513 if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
1514 bool isST0 = VA.getLocReg() == X86::ST0;
1516 if (CopyVT == MVT::f32) Opc = isST0 ? X86::FpGET_ST0_32:X86::FpGET_ST1_32;
1517 if (CopyVT == MVT::f64) Opc = isST0 ? X86::FpGET_ST0_64:X86::FpGET_ST1_64;
1518 if (CopyVT == MVT::f80) Opc = isST0 ? X86::FpGET_ST0_80:X86::FpGET_ST1_80;
1519 SDValue Ops[] = { Chain, InFlag };
1520 Chain = SDValue(DAG.getMachineNode(Opc, dl, CopyVT, MVT::Other, MVT::Glue,
1522 Val = Chain.getValue(0);
1524 // Round the f80 to the right size, which also moves it to the appropriate
1526 if (CopyVT != VA.getValVT())
1527 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1528 // This truncation won't change the value.
1529 DAG.getIntPtrConstant(1));
1530 } else if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1531 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1532 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1533 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1534 MVT::v2i64, InFlag).getValue(1);
1535 Val = Chain.getValue(0);
1536 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1537 Val, DAG.getConstant(0, MVT::i64));
1539 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1540 MVT::i64, InFlag).getValue(1);
1541 Val = Chain.getValue(0);
1543 Val = DAG.getNode(ISD::BITCAST, dl, CopyVT, Val);
1545 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1546 CopyVT, InFlag).getValue(1);
1547 Val = Chain.getValue(0);
1549 InFlag = Chain.getValue(2);
1550 InVals.push_back(Val);
1557 //===----------------------------------------------------------------------===//
1558 // C & StdCall & Fast Calling Convention implementation
1559 //===----------------------------------------------------------------------===//
1560 // StdCall calling convention seems to be standard for many Windows' API
1561 // routines and around. It differs from C calling convention just a little:
1562 // callee should clean up the stack, not caller. Symbols should be also
1563 // decorated in some fancy way :) It doesn't support any vector arguments.
1564 // For info on fast calling convention see Fast Calling Convention (tail call)
1565 // implementation LowerX86_32FastCCCallTo.
1567 /// CallIsStructReturn - Determines whether a call uses struct return
1569 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1573 return Outs[0].Flags.isSRet();
1576 /// ArgsAreStructReturn - Determines whether a function uses struct
1577 /// return semantics.
1579 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1583 return Ins[0].Flags.isSRet();
1586 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1587 /// by "Src" to address "Dst" with size and alignment information specified by
1588 /// the specific parameter attribute. The copy will be passed as a byval
1589 /// function parameter.
1591 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1592 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1594 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1596 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1597 /*isVolatile*/false, /*AlwaysInline=*/true,
1598 MachinePointerInfo(), MachinePointerInfo());
1601 /// IsTailCallConvention - Return true if the calling convention is one that
1602 /// supports tail call optimization.
1603 static bool IsTailCallConvention(CallingConv::ID CC) {
1604 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1607 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
1608 if (!CI->isTailCall())
1612 CallingConv::ID CalleeCC = CS.getCallingConv();
1613 if (!IsTailCallConvention(CalleeCC) && CalleeCC != CallingConv::C)
1619 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1620 /// a tailcall target by changing its ABI.
1621 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1622 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1626 X86TargetLowering::LowerMemArgument(SDValue Chain,
1627 CallingConv::ID CallConv,
1628 const SmallVectorImpl<ISD::InputArg> &Ins,
1629 DebugLoc dl, SelectionDAG &DAG,
1630 const CCValAssign &VA,
1631 MachineFrameInfo *MFI,
1633 // Create the nodes corresponding to a load from this parameter slot.
1634 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1635 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1636 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1639 // If value is passed by pointer we have address passed instead of the value
1641 if (VA.getLocInfo() == CCValAssign::Indirect)
1642 ValVT = VA.getLocVT();
1644 ValVT = VA.getValVT();
1646 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1647 // changed with more analysis.
1648 // In case of tail call optimization mark all arguments mutable. Since they
1649 // could be overwritten by lowering of arguments in case of a tail call.
1650 if (Flags.isByVal()) {
1651 unsigned Bytes = Flags.getByValSize();
1652 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
1653 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
1654 return DAG.getFrameIndex(FI, getPointerTy());
1656 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1657 VA.getLocMemOffset(), isImmutable);
1658 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1659 return DAG.getLoad(ValVT, dl, Chain, FIN,
1660 MachinePointerInfo::getFixedStack(FI),
1666 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1667 CallingConv::ID CallConv,
1669 const SmallVectorImpl<ISD::InputArg> &Ins,
1672 SmallVectorImpl<SDValue> &InVals)
1674 MachineFunction &MF = DAG.getMachineFunction();
1675 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1677 const Function* Fn = MF.getFunction();
1678 if (Fn->hasExternalLinkage() &&
1679 Subtarget->isTargetCygMing() &&
1680 Fn->getName() == "main")
1681 FuncInfo->setForceFramePointer(true);
1683 MachineFrameInfo *MFI = MF.getFrameInfo();
1684 bool Is64Bit = Subtarget->is64Bit();
1685 bool IsWin64 = Subtarget->isTargetWin64();
1687 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1688 "Var args not supported with calling convention fastcc or ghc");
1690 // Assign locations to all of the incoming arguments.
1691 SmallVector<CCValAssign, 16> ArgLocs;
1692 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1693 ArgLocs, *DAG.getContext());
1695 // Allocate shadow area for Win64
1697 CCInfo.AllocateStack(32, 8);
1700 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
1702 unsigned LastVal = ~0U;
1704 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1705 CCValAssign &VA = ArgLocs[i];
1706 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1708 assert(VA.getValNo() != LastVal &&
1709 "Don't support value assigned to multiple locs yet");
1710 LastVal = VA.getValNo();
1712 if (VA.isRegLoc()) {
1713 EVT RegVT = VA.getLocVT();
1714 TargetRegisterClass *RC = NULL;
1715 if (RegVT == MVT::i32)
1716 RC = X86::GR32RegisterClass;
1717 else if (Is64Bit && RegVT == MVT::i64)
1718 RC = X86::GR64RegisterClass;
1719 else if (RegVT == MVT::f32)
1720 RC = X86::FR32RegisterClass;
1721 else if (RegVT == MVT::f64)
1722 RC = X86::FR64RegisterClass;
1723 else if (RegVT.isVector() && RegVT.getSizeInBits() == 256)
1724 RC = X86::VR256RegisterClass;
1725 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1726 RC = X86::VR128RegisterClass;
1727 else if (RegVT == MVT::x86mmx)
1728 RC = X86::VR64RegisterClass;
1730 llvm_unreachable("Unknown argument type!");
1732 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1733 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1735 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1736 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1738 if (VA.getLocInfo() == CCValAssign::SExt)
1739 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1740 DAG.getValueType(VA.getValVT()));
1741 else if (VA.getLocInfo() == CCValAssign::ZExt)
1742 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1743 DAG.getValueType(VA.getValVT()));
1744 else if (VA.getLocInfo() == CCValAssign::BCvt)
1745 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
1747 if (VA.isExtInLoc()) {
1748 // Handle MMX values passed in XMM regs.
1749 if (RegVT.isVector()) {
1750 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(),
1753 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1756 assert(VA.isMemLoc());
1757 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1760 // If value is passed via pointer - do a load.
1761 if (VA.getLocInfo() == CCValAssign::Indirect)
1762 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
1763 MachinePointerInfo(), false, false, 0);
1765 InVals.push_back(ArgValue);
1768 // The x86-64 ABI for returning structs by value requires that we copy
1769 // the sret argument into %rax for the return. Save the argument into
1770 // a virtual register so that we can access it from the return points.
1771 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1772 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1773 unsigned Reg = FuncInfo->getSRetReturnReg();
1775 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1776 FuncInfo->setSRetReturnReg(Reg);
1778 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1779 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1782 unsigned StackSize = CCInfo.getNextStackOffset();
1783 // Align stack specially for tail calls.
1784 if (FuncIsMadeTailCallSafe(CallConv))
1785 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1787 // If the function takes variable number of arguments, make a frame index for
1788 // the start of the first vararg value... for expansion of llvm.va_start.
1790 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1791 CallConv != CallingConv::X86_ThisCall)) {
1792 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
1795 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1797 // FIXME: We should really autogenerate these arrays
1798 static const unsigned GPR64ArgRegsWin64[] = {
1799 X86::RCX, X86::RDX, X86::R8, X86::R9
1801 static const unsigned GPR64ArgRegs64Bit[] = {
1802 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1804 static const unsigned XMMArgRegs64Bit[] = {
1805 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1806 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1808 const unsigned *GPR64ArgRegs;
1809 unsigned NumXMMRegs = 0;
1812 // The XMM registers which might contain var arg parameters are shadowed
1813 // in their paired GPR. So we only need to save the GPR to their home
1815 TotalNumIntRegs = 4;
1816 GPR64ArgRegs = GPR64ArgRegsWin64;
1818 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1819 GPR64ArgRegs = GPR64ArgRegs64Bit;
1821 NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs64Bit, TotalNumXMMRegs);
1823 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1826 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1827 assert(!(NumXMMRegs && !Subtarget->hasXMM()) &&
1828 "SSE register cannot be used when SSE is disabled!");
1829 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1830 "SSE register cannot be used when SSE is disabled!");
1831 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasXMM())
1832 // Kernel mode asks for SSE to be disabled, so don't push them
1834 TotalNumXMMRegs = 0;
1837 const TargetFrameLowering &TFI = *getTargetMachine().getFrameLowering();
1838 // Get to the caller-allocated home save location. Add 8 to account
1839 // for the return address.
1840 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
1841 FuncInfo->setRegSaveFrameIndex(
1842 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
1843 // Fixup to set vararg frame on shadow area (4 x i64).
1845 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
1847 // For X86-64, if there are vararg parameters that are passed via
1848 // registers, then we must store them to their spots on the stack so they
1849 // may be loaded by deferencing the result of va_next.
1850 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1851 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1852 FuncInfo->setRegSaveFrameIndex(
1853 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1857 // Store the integer parameter registers.
1858 SmallVector<SDValue, 8> MemOps;
1859 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1861 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1862 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1863 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1864 DAG.getIntPtrConstant(Offset));
1865 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1866 X86::GR64RegisterClass);
1867 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1869 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1870 MachinePointerInfo::getFixedStack(
1871 FuncInfo->getRegSaveFrameIndex(), Offset),
1873 MemOps.push_back(Store);
1877 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1878 // Now store the XMM (fp + vector) parameter registers.
1879 SmallVector<SDValue, 11> SaveXMMOps;
1880 SaveXMMOps.push_back(Chain);
1882 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1883 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1884 SaveXMMOps.push_back(ALVal);
1886 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1887 FuncInfo->getRegSaveFrameIndex()));
1888 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1889 FuncInfo->getVarArgsFPOffset()));
1891 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1892 unsigned VReg = MF.addLiveIn(XMMArgRegs64Bit[NumXMMRegs],
1893 X86::VR128RegisterClass);
1894 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1895 SaveXMMOps.push_back(Val);
1897 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1899 &SaveXMMOps[0], SaveXMMOps.size()));
1902 if (!MemOps.empty())
1903 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1904 &MemOps[0], MemOps.size());
1908 // Some CCs need callee pop.
1909 if (Subtarget->IsCalleePop(isVarArg, CallConv)) {
1910 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1912 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1913 // If this is an sret function, the return should pop the hidden pointer.
1914 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1915 FuncInfo->setBytesToPopOnReturn(4);
1919 // RegSaveFrameIndex is X86-64 only.
1920 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1921 if (CallConv == CallingConv::X86_FastCall ||
1922 CallConv == CallingConv::X86_ThisCall)
1923 // fastcc functions can't have varargs.
1924 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1931 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1932 SDValue StackPtr, SDValue Arg,
1933 DebugLoc dl, SelectionDAG &DAG,
1934 const CCValAssign &VA,
1935 ISD::ArgFlagsTy Flags) const {
1936 unsigned LocMemOffset = VA.getLocMemOffset();
1937 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1938 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1939 if (Flags.isByVal())
1940 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1942 return DAG.getStore(Chain, dl, Arg, PtrOff,
1943 MachinePointerInfo::getStack(LocMemOffset),
1947 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1948 /// optimization is performed and it is required.
1950 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1951 SDValue &OutRetAddr, SDValue Chain,
1952 bool IsTailCall, bool Is64Bit,
1953 int FPDiff, DebugLoc dl) const {
1954 // Adjust the Return address stack slot.
1955 EVT VT = getPointerTy();
1956 OutRetAddr = getReturnAddressFrameIndex(DAG);
1958 // Load the "old" Return address.
1959 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
1961 return SDValue(OutRetAddr.getNode(), 1);
1964 /// EmitTailCallStoreRetAddr - Emit a store of the return address if tail call
1965 /// optimization is performed and it is required (FPDiff!=0).
1967 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1968 SDValue Chain, SDValue RetAddrFrIdx,
1969 bool Is64Bit, int FPDiff, DebugLoc dl) {
1970 // Store the return address to the appropriate stack slot.
1971 if (!FPDiff) return Chain;
1972 // Calculate the new stack slot for the return address.
1973 int SlotSize = Is64Bit ? 8 : 4;
1974 int NewReturnAddrFI =
1975 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
1976 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1977 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1978 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1979 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
1985 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1986 CallingConv::ID CallConv, bool isVarArg,
1988 const SmallVectorImpl<ISD::OutputArg> &Outs,
1989 const SmallVectorImpl<SDValue> &OutVals,
1990 const SmallVectorImpl<ISD::InputArg> &Ins,
1991 DebugLoc dl, SelectionDAG &DAG,
1992 SmallVectorImpl<SDValue> &InVals) const {
1993 MachineFunction &MF = DAG.getMachineFunction();
1994 bool Is64Bit = Subtarget->is64Bit();
1995 bool IsWin64 = Subtarget->isTargetWin64();
1996 bool IsStructRet = CallIsStructReturn(Outs);
1997 bool IsSibcall = false;
2000 // Check if it's really possible to do a tail call.
2001 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
2002 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
2003 Outs, OutVals, Ins, DAG);
2005 // Sibcalls are automatically detected tailcalls which do not require
2007 if (!GuaranteedTailCallOpt && isTailCall)
2014 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2015 "Var args not supported with calling convention fastcc or ghc");
2017 // Analyze operands of the call, assigning locations to each operand.
2018 SmallVector<CCValAssign, 16> ArgLocs;
2019 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
2020 ArgLocs, *DAG.getContext());
2022 // Allocate shadow area for Win64
2024 CCInfo.AllocateStack(32, 8);
2027 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2029 // Get a count of how many bytes are to be pushed on the stack.
2030 unsigned NumBytes = CCInfo.getNextStackOffset();
2032 // This is a sibcall. The memory operands are available in caller's
2033 // own caller's stack.
2035 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
2036 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2039 if (isTailCall && !IsSibcall) {
2040 // Lower arguments at fp - stackoffset + fpdiff.
2041 unsigned NumBytesCallerPushed =
2042 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
2043 FPDiff = NumBytesCallerPushed - NumBytes;
2045 // Set the delta of movement of the returnaddr stackslot.
2046 // But only set if delta is greater than previous delta.
2047 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
2048 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
2052 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
2054 SDValue RetAddrFrIdx;
2055 // Load return address for tail calls.
2056 if (isTailCall && FPDiff)
2057 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
2058 Is64Bit, FPDiff, dl);
2060 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2061 SmallVector<SDValue, 8> MemOpChains;
2064 // Walk the register/memloc assignments, inserting copies/loads. In the case
2065 // of tail call optimization arguments are handle later.
2066 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2067 CCValAssign &VA = ArgLocs[i];
2068 EVT RegVT = VA.getLocVT();
2069 SDValue Arg = OutVals[i];
2070 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2071 bool isByVal = Flags.isByVal();
2073 // Promote the value if needed.
2074 switch (VA.getLocInfo()) {
2075 default: llvm_unreachable("Unknown loc info!");
2076 case CCValAssign::Full: break;
2077 case CCValAssign::SExt:
2078 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2080 case CCValAssign::ZExt:
2081 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
2083 case CCValAssign::AExt:
2084 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
2085 // Special case: passing MMX values in XMM registers.
2086 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
2087 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
2088 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
2090 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
2092 case CCValAssign::BCvt:
2093 Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg);
2095 case CCValAssign::Indirect: {
2096 // Store the argument.
2097 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2098 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2099 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2100 MachinePointerInfo::getFixedStack(FI),
2107 if (VA.isRegLoc()) {
2108 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2109 if (isVarArg && IsWin64) {
2110 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2111 // shadow reg if callee is a varargs function.
2112 unsigned ShadowReg = 0;
2113 switch (VA.getLocReg()) {
2114 case X86::XMM0: ShadowReg = X86::RCX; break;
2115 case X86::XMM1: ShadowReg = X86::RDX; break;
2116 case X86::XMM2: ShadowReg = X86::R8; break;
2117 case X86::XMM3: ShadowReg = X86::R9; break;
2120 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2122 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2123 assert(VA.isMemLoc());
2124 if (StackPtr.getNode() == 0)
2125 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
2126 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2127 dl, DAG, VA, Flags));
2131 if (!MemOpChains.empty())
2132 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2133 &MemOpChains[0], MemOpChains.size());
2135 // Build a sequence of copy-to-reg nodes chained together with token chain
2136 // and flag operands which copy the outgoing args into registers.
2138 // Tail call byval lowering might overwrite argument registers so in case of
2139 // tail call optimization the copies to registers are lowered later.
2141 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2142 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2143 RegsToPass[i].second, InFlag);
2144 InFlag = Chain.getValue(1);
2147 if (Subtarget->isPICStyleGOT()) {
2148 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2151 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
2152 DAG.getNode(X86ISD::GlobalBaseReg,
2153 DebugLoc(), getPointerTy()),
2155 InFlag = Chain.getValue(1);
2157 // If we are tail calling and generating PIC/GOT style code load the
2158 // address of the callee into ECX. The value in ecx is used as target of
2159 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2160 // for tail calls on PIC/GOT architectures. Normally we would just put the
2161 // address of GOT into ebx and then call target@PLT. But for tail calls
2162 // ebx would be restored (since ebx is callee saved) before jumping to the
2165 // Note: The actual moving to ECX is done further down.
2166 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2167 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2168 !G->getGlobal()->hasProtectedVisibility())
2169 Callee = LowerGlobalAddress(Callee, DAG);
2170 else if (isa<ExternalSymbolSDNode>(Callee))
2171 Callee = LowerExternalSymbol(Callee, DAG);
2175 if (Is64Bit && isVarArg && !IsWin64) {
2176 // From AMD64 ABI document:
2177 // For calls that may call functions that use varargs or stdargs
2178 // (prototype-less calls or calls to functions containing ellipsis (...) in
2179 // the declaration) %al is used as hidden argument to specify the number
2180 // of SSE registers used. The contents of %al do not need to match exactly
2181 // the number of registers, but must be an ubound on the number of SSE
2182 // registers used and is in the range 0 - 8 inclusive.
2184 // Count the number of XMM registers allocated.
2185 static const unsigned XMMArgRegs[] = {
2186 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2187 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2189 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2190 assert((Subtarget->hasXMM() || !NumXMMRegs)
2191 && "SSE registers cannot be used when SSE is disabled");
2193 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
2194 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
2195 InFlag = Chain.getValue(1);
2199 // For tail calls lower the arguments to the 'real' stack slot.
2201 // Force all the incoming stack arguments to be loaded from the stack
2202 // before any new outgoing arguments are stored to the stack, because the
2203 // outgoing stack slots may alias the incoming argument stack slots, and
2204 // the alias isn't otherwise explicit. This is slightly more conservative
2205 // than necessary, because it means that each store effectively depends
2206 // on every argument instead of just those arguments it would clobber.
2207 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2209 SmallVector<SDValue, 8> MemOpChains2;
2212 // Do not flag preceding copytoreg stuff together with the following stuff.
2214 if (GuaranteedTailCallOpt) {
2215 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2216 CCValAssign &VA = ArgLocs[i];
2219 assert(VA.isMemLoc());
2220 SDValue Arg = OutVals[i];
2221 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2222 // Create frame index.
2223 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2224 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2225 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2226 FIN = DAG.getFrameIndex(FI, getPointerTy());
2228 if (Flags.isByVal()) {
2229 // Copy relative to framepointer.
2230 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2231 if (StackPtr.getNode() == 0)
2232 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2234 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2236 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2240 // Store relative to framepointer.
2241 MemOpChains2.push_back(
2242 DAG.getStore(ArgChain, dl, Arg, FIN,
2243 MachinePointerInfo::getFixedStack(FI),
2249 if (!MemOpChains2.empty())
2250 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2251 &MemOpChains2[0], MemOpChains2.size());
2253 // Copy arguments to their registers.
2254 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2255 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2256 RegsToPass[i].second, InFlag);
2257 InFlag = Chain.getValue(1);
2261 // Store the return address to the appropriate stack slot.
2262 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2266 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2267 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2268 // In the 64-bit large code model, we have to make all calls
2269 // through a register, since the call instruction's 32-bit
2270 // pc-relative offset may not be large enough to hold the whole
2272 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2273 // If the callee is a GlobalAddress node (quite common, every direct call
2274 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2277 // We should use extra load for direct calls to dllimported functions in
2279 const GlobalValue *GV = G->getGlobal();
2280 if (!GV->hasDLLImportLinkage()) {
2281 unsigned char OpFlags = 0;
2283 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2284 // external symbols most go through the PLT in PIC mode. If the symbol
2285 // has hidden or protected visibility, or if it is static or local, then
2286 // we don't need to use the PLT - we can directly call it.
2287 if (Subtarget->isTargetELF() &&
2288 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2289 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2290 OpFlags = X86II::MO_PLT;
2291 } else if (Subtarget->isPICStyleStubAny() &&
2292 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2293 (!Subtarget->getTargetTriple().isMacOSX() ||
2294 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2295 // PC-relative references to external symbols should go through $stub,
2296 // unless we're building with the leopard linker or later, which
2297 // automatically synthesizes these stubs.
2298 OpFlags = X86II::MO_DARWIN_STUB;
2301 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2302 G->getOffset(), OpFlags);
2304 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2305 unsigned char OpFlags = 0;
2307 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
2308 // external symbols should go through the PLT.
2309 if (Subtarget->isTargetELF() &&
2310 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2311 OpFlags = X86II::MO_PLT;
2312 } else if (Subtarget->isPICStyleStubAny() &&
2313 (!Subtarget->getTargetTriple().isMacOSX() ||
2314 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2315 // PC-relative references to external symbols should go through $stub,
2316 // unless we're building with the leopard linker or later, which
2317 // automatically synthesizes these stubs.
2318 OpFlags = X86II::MO_DARWIN_STUB;
2321 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2325 // Returns a chain & a flag for retval copy to use.
2326 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2327 SmallVector<SDValue, 8> Ops;
2329 if (!IsSibcall && isTailCall) {
2330 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2331 DAG.getIntPtrConstant(0, true), InFlag);
2332 InFlag = Chain.getValue(1);
2335 Ops.push_back(Chain);
2336 Ops.push_back(Callee);
2339 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2341 // Add argument registers to the end of the list so that they are known live
2343 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2344 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2345 RegsToPass[i].second.getValueType()));
2347 // Add an implicit use GOT pointer in EBX.
2348 if (!isTailCall && Subtarget->isPICStyleGOT())
2349 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2351 // Add an implicit use of AL for non-Windows x86 64-bit vararg functions.
2352 if (Is64Bit && isVarArg && !IsWin64)
2353 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2355 if (InFlag.getNode())
2356 Ops.push_back(InFlag);
2360 //// If this is the first return lowered for this function, add the regs
2361 //// to the liveout set for the function.
2362 // This isn't right, although it's probably harmless on x86; liveouts
2363 // should be computed from returns not tail calls. Consider a void
2364 // function making a tail call to a function returning int.
2365 return DAG.getNode(X86ISD::TC_RETURN, dl,
2366 NodeTys, &Ops[0], Ops.size());
2369 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2370 InFlag = Chain.getValue(1);
2372 // Create the CALLSEQ_END node.
2373 unsigned NumBytesForCalleeToPush;
2374 if (Subtarget->IsCalleePop(isVarArg, CallConv))
2375 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2376 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2377 // If this is a call to a struct-return function, the callee
2378 // pops the hidden struct pointer, so we have to push it back.
2379 // This is common for Darwin/X86, Linux & Mingw32 targets.
2380 NumBytesForCalleeToPush = 4;
2382 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2384 // Returns a flag for retval copy to use.
2386 Chain = DAG.getCALLSEQ_END(Chain,
2387 DAG.getIntPtrConstant(NumBytes, true),
2388 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2391 InFlag = Chain.getValue(1);
2394 // Handle result values, copying them out of physregs into vregs that we
2396 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2397 Ins, dl, DAG, InVals);
2401 //===----------------------------------------------------------------------===//
2402 // Fast Calling Convention (tail call) implementation
2403 //===----------------------------------------------------------------------===//
2405 // Like std call, callee cleans arguments, convention except that ECX is
2406 // reserved for storing the tail called function address. Only 2 registers are
2407 // free for argument passing (inreg). Tail call optimization is performed
2409 // * tailcallopt is enabled
2410 // * caller/callee are fastcc
2411 // On X86_64 architecture with GOT-style position independent code only local
2412 // (within module) calls are supported at the moment.
2413 // To keep the stack aligned according to platform abi the function
2414 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2415 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2416 // If a tail called function callee has more arguments than the caller the
2417 // caller needs to make sure that there is room to move the RETADDR to. This is
2418 // achieved by reserving an area the size of the argument delta right after the
2419 // original REtADDR, but before the saved framepointer or the spilled registers
2420 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2432 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2433 /// for a 16 byte align requirement.
2435 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2436 SelectionDAG& DAG) const {
2437 MachineFunction &MF = DAG.getMachineFunction();
2438 const TargetMachine &TM = MF.getTarget();
2439 const TargetFrameLowering &TFI = *TM.getFrameLowering();
2440 unsigned StackAlignment = TFI.getStackAlignment();
2441 uint64_t AlignMask = StackAlignment - 1;
2442 int64_t Offset = StackSize;
2443 uint64_t SlotSize = TD->getPointerSize();
2444 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2445 // Number smaller than 12 so just add the difference.
2446 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2448 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2449 Offset = ((~AlignMask) & Offset) + StackAlignment +
2450 (StackAlignment-SlotSize);
2455 /// MatchingStackOffset - Return true if the given stack call argument is
2456 /// already available in the same position (relatively) of the caller's
2457 /// incoming argument stack.
2459 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2460 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2461 const X86InstrInfo *TII) {
2462 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2464 if (Arg.getOpcode() == ISD::CopyFromReg) {
2465 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2466 if (!TargetRegisterInfo::isVirtualRegister(VR))
2468 MachineInstr *Def = MRI->getVRegDef(VR);
2471 if (!Flags.isByVal()) {
2472 if (!TII->isLoadFromStackSlot(Def, FI))
2475 unsigned Opcode = Def->getOpcode();
2476 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2477 Def->getOperand(1).isFI()) {
2478 FI = Def->getOperand(1).getIndex();
2479 Bytes = Flags.getByValSize();
2483 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2484 if (Flags.isByVal())
2485 // ByVal argument is passed in as a pointer but it's now being
2486 // dereferenced. e.g.
2487 // define @foo(%struct.X* %A) {
2488 // tail call @bar(%struct.X* byval %A)
2491 SDValue Ptr = Ld->getBasePtr();
2492 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2495 FI = FINode->getIndex();
2499 assert(FI != INT_MAX);
2500 if (!MFI->isFixedObjectIndex(FI))
2502 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2505 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2506 /// for tail call optimization. Targets which want to do tail call
2507 /// optimization should implement this function.
2509 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2510 CallingConv::ID CalleeCC,
2512 bool isCalleeStructRet,
2513 bool isCallerStructRet,
2514 const SmallVectorImpl<ISD::OutputArg> &Outs,
2515 const SmallVectorImpl<SDValue> &OutVals,
2516 const SmallVectorImpl<ISD::InputArg> &Ins,
2517 SelectionDAG& DAG) const {
2518 if (!IsTailCallConvention(CalleeCC) &&
2519 CalleeCC != CallingConv::C)
2522 // If -tailcallopt is specified, make fastcc functions tail-callable.
2523 const MachineFunction &MF = DAG.getMachineFunction();
2524 const Function *CallerF = DAG.getMachineFunction().getFunction();
2525 CallingConv::ID CallerCC = CallerF->getCallingConv();
2526 bool CCMatch = CallerCC == CalleeCC;
2528 if (GuaranteedTailCallOpt) {
2529 if (IsTailCallConvention(CalleeCC) && CCMatch)
2534 // Look for obvious safe cases to perform tail call optimization that do not
2535 // require ABI changes. This is what gcc calls sibcall.
2537 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2538 // emit a special epilogue.
2539 if (RegInfo->needsStackRealignment(MF))
2542 // Do not sibcall optimize vararg calls unless the call site is not passing
2544 if (isVarArg && !Outs.empty())
2547 // Also avoid sibcall optimization if either caller or callee uses struct
2548 // return semantics.
2549 if (isCalleeStructRet || isCallerStructRet)
2552 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2553 // Therefore if it's not used by the call it is not safe to optimize this into
2555 bool Unused = false;
2556 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2563 SmallVector<CCValAssign, 16> RVLocs;
2564 CCState CCInfo(CalleeCC, false, getTargetMachine(),
2565 RVLocs, *DAG.getContext());
2566 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2567 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2568 CCValAssign &VA = RVLocs[i];
2569 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2574 // If the calling conventions do not match, then we'd better make sure the
2575 // results are returned in the same way as what the caller expects.
2577 SmallVector<CCValAssign, 16> RVLocs1;
2578 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
2579 RVLocs1, *DAG.getContext());
2580 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2582 SmallVector<CCValAssign, 16> RVLocs2;
2583 CCState CCInfo2(CallerCC, false, getTargetMachine(),
2584 RVLocs2, *DAG.getContext());
2585 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2587 if (RVLocs1.size() != RVLocs2.size())
2589 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2590 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2592 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2594 if (RVLocs1[i].isRegLoc()) {
2595 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2598 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2604 // If the callee takes no arguments then go on to check the results of the
2606 if (!Outs.empty()) {
2607 // Check if stack adjustment is needed. For now, do not do this if any
2608 // argument is passed on the stack.
2609 SmallVector<CCValAssign, 16> ArgLocs;
2610 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2611 ArgLocs, *DAG.getContext());
2613 // Allocate shadow area for Win64
2614 if (Subtarget->isTargetWin64()) {
2615 CCInfo.AllocateStack(32, 8);
2618 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2619 if (CCInfo.getNextStackOffset()) {
2620 MachineFunction &MF = DAG.getMachineFunction();
2621 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2624 // Check if the arguments are already laid out in the right way as
2625 // the caller's fixed stack objects.
2626 MachineFrameInfo *MFI = MF.getFrameInfo();
2627 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2628 const X86InstrInfo *TII =
2629 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2630 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2631 CCValAssign &VA = ArgLocs[i];
2632 SDValue Arg = OutVals[i];
2633 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2634 if (VA.getLocInfo() == CCValAssign::Indirect)
2636 if (!VA.isRegLoc()) {
2637 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2644 // If the tailcall address may be in a register, then make sure it's
2645 // possible to register allocate for it. In 32-bit, the call address can
2646 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2647 // callee-saved registers are restored. These happen to be the same
2648 // registers used to pass 'inreg' arguments so watch out for those.
2649 if (!Subtarget->is64Bit() &&
2650 !isa<GlobalAddressSDNode>(Callee) &&
2651 !isa<ExternalSymbolSDNode>(Callee)) {
2652 unsigned NumInRegs = 0;
2653 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2654 CCValAssign &VA = ArgLocs[i];
2657 unsigned Reg = VA.getLocReg();
2660 case X86::EAX: case X86::EDX: case X86::ECX:
2661 if (++NumInRegs == 3)
2669 // An stdcall caller is expected to clean up its arguments; the callee
2670 // isn't going to do that.
2671 if (!CCMatch && CallerCC==CallingConv::X86_StdCall)
2678 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
2679 return X86::createFastISel(funcInfo);
2683 //===----------------------------------------------------------------------===//
2684 // Other Lowering Hooks
2685 //===----------------------------------------------------------------------===//
2687 static bool MayFoldLoad(SDValue Op) {
2688 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
2691 static bool MayFoldIntoStore(SDValue Op) {
2692 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
2695 static bool isTargetShuffle(unsigned Opcode) {
2697 default: return false;
2698 case X86ISD::PSHUFD:
2699 case X86ISD::PSHUFHW:
2700 case X86ISD::PSHUFLW:
2701 case X86ISD::SHUFPD:
2702 case X86ISD::PALIGN:
2703 case X86ISD::SHUFPS:
2704 case X86ISD::MOVLHPS:
2705 case X86ISD::MOVLHPD:
2706 case X86ISD::MOVHLPS:
2707 case X86ISD::MOVLPS:
2708 case X86ISD::MOVLPD:
2709 case X86ISD::MOVSHDUP:
2710 case X86ISD::MOVSLDUP:
2711 case X86ISD::MOVDDUP:
2714 case X86ISD::UNPCKLPS:
2715 case X86ISD::UNPCKLPD:
2716 case X86ISD::VUNPCKLPS:
2717 case X86ISD::VUNPCKLPD:
2718 case X86ISD::VUNPCKLPSY:
2719 case X86ISD::VUNPCKLPDY:
2720 case X86ISD::PUNPCKLWD:
2721 case X86ISD::PUNPCKLBW:
2722 case X86ISD::PUNPCKLDQ:
2723 case X86ISD::PUNPCKLQDQ:
2724 case X86ISD::UNPCKHPS:
2725 case X86ISD::UNPCKHPD:
2726 case X86ISD::PUNPCKHWD:
2727 case X86ISD::PUNPCKHBW:
2728 case X86ISD::PUNPCKHDQ:
2729 case X86ISD::PUNPCKHQDQ:
2735 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2736 SDValue V1, SelectionDAG &DAG) {
2738 default: llvm_unreachable("Unknown x86 shuffle node");
2739 case X86ISD::MOVSHDUP:
2740 case X86ISD::MOVSLDUP:
2741 case X86ISD::MOVDDUP:
2742 return DAG.getNode(Opc, dl, VT, V1);
2748 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2749 SDValue V1, unsigned TargetMask, SelectionDAG &DAG) {
2751 default: llvm_unreachable("Unknown x86 shuffle node");
2752 case X86ISD::PSHUFD:
2753 case X86ISD::PSHUFHW:
2754 case X86ISD::PSHUFLW:
2755 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
2761 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2762 SDValue V1, SDValue V2, unsigned TargetMask, SelectionDAG &DAG) {
2764 default: llvm_unreachable("Unknown x86 shuffle node");
2765 case X86ISD::PALIGN:
2766 case X86ISD::SHUFPD:
2767 case X86ISD::SHUFPS:
2768 return DAG.getNode(Opc, dl, VT, V1, V2,
2769 DAG.getConstant(TargetMask, MVT::i8));
2774 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2775 SDValue V1, SDValue V2, SelectionDAG &DAG) {
2777 default: llvm_unreachable("Unknown x86 shuffle node");
2778 case X86ISD::MOVLHPS:
2779 case X86ISD::MOVLHPD:
2780 case X86ISD::MOVHLPS:
2781 case X86ISD::MOVLPS:
2782 case X86ISD::MOVLPD:
2785 case X86ISD::UNPCKLPS:
2786 case X86ISD::UNPCKLPD:
2787 case X86ISD::VUNPCKLPS:
2788 case X86ISD::VUNPCKLPD:
2789 case X86ISD::VUNPCKLPSY:
2790 case X86ISD::VUNPCKLPDY:
2791 case X86ISD::PUNPCKLWD:
2792 case X86ISD::PUNPCKLBW:
2793 case X86ISD::PUNPCKLDQ:
2794 case X86ISD::PUNPCKLQDQ:
2795 case X86ISD::UNPCKHPS:
2796 case X86ISD::UNPCKHPD:
2797 case X86ISD::PUNPCKHWD:
2798 case X86ISD::PUNPCKHBW:
2799 case X86ISD::PUNPCKHDQ:
2800 case X86ISD::PUNPCKHQDQ:
2801 return DAG.getNode(Opc, dl, VT, V1, V2);
2806 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2807 MachineFunction &MF = DAG.getMachineFunction();
2808 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2809 int ReturnAddrIndex = FuncInfo->getRAIndex();
2811 if (ReturnAddrIndex == 0) {
2812 // Set up a frame object for the return address.
2813 uint64_t SlotSize = TD->getPointerSize();
2814 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2816 FuncInfo->setRAIndex(ReturnAddrIndex);
2819 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2823 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2824 bool hasSymbolicDisplacement) {
2825 // Offset should fit into 32 bit immediate field.
2826 if (!isInt<32>(Offset))
2829 // If we don't have a symbolic displacement - we don't have any extra
2831 if (!hasSymbolicDisplacement)
2834 // FIXME: Some tweaks might be needed for medium code model.
2835 if (M != CodeModel::Small && M != CodeModel::Kernel)
2838 // For small code model we assume that latest object is 16MB before end of 31
2839 // bits boundary. We may also accept pretty large negative constants knowing
2840 // that all objects are in the positive half of address space.
2841 if (M == CodeModel::Small && Offset < 16*1024*1024)
2844 // For kernel code model we know that all object resist in the negative half
2845 // of 32bits address space. We may not accept negative offsets, since they may
2846 // be just off and we may accept pretty large positive ones.
2847 if (M == CodeModel::Kernel && Offset > 0)
2853 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2854 /// specific condition code, returning the condition code and the LHS/RHS of the
2855 /// comparison to make.
2856 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2857 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2859 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2860 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2861 // X > -1 -> X == 0, jump !sign.
2862 RHS = DAG.getConstant(0, RHS.getValueType());
2863 return X86::COND_NS;
2864 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2865 // X < 0 -> X == 0, jump on sign.
2867 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2869 RHS = DAG.getConstant(0, RHS.getValueType());
2870 return X86::COND_LE;
2874 switch (SetCCOpcode) {
2875 default: llvm_unreachable("Invalid integer condition!");
2876 case ISD::SETEQ: return X86::COND_E;
2877 case ISD::SETGT: return X86::COND_G;
2878 case ISD::SETGE: return X86::COND_GE;
2879 case ISD::SETLT: return X86::COND_L;
2880 case ISD::SETLE: return X86::COND_LE;
2881 case ISD::SETNE: return X86::COND_NE;
2882 case ISD::SETULT: return X86::COND_B;
2883 case ISD::SETUGT: return X86::COND_A;
2884 case ISD::SETULE: return X86::COND_BE;
2885 case ISD::SETUGE: return X86::COND_AE;
2889 // First determine if it is required or is profitable to flip the operands.
2891 // If LHS is a foldable load, but RHS is not, flip the condition.
2892 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
2893 !ISD::isNON_EXTLoad(RHS.getNode())) {
2894 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2895 std::swap(LHS, RHS);
2898 switch (SetCCOpcode) {
2904 std::swap(LHS, RHS);
2908 // On a floating point condition, the flags are set as follows:
2910 // 0 | 0 | 0 | X > Y
2911 // 0 | 0 | 1 | X < Y
2912 // 1 | 0 | 0 | X == Y
2913 // 1 | 1 | 1 | unordered
2914 switch (SetCCOpcode) {
2915 default: llvm_unreachable("Condcode should be pre-legalized away");
2917 case ISD::SETEQ: return X86::COND_E;
2918 case ISD::SETOLT: // flipped
2920 case ISD::SETGT: return X86::COND_A;
2921 case ISD::SETOLE: // flipped
2923 case ISD::SETGE: return X86::COND_AE;
2924 case ISD::SETUGT: // flipped
2926 case ISD::SETLT: return X86::COND_B;
2927 case ISD::SETUGE: // flipped
2929 case ISD::SETLE: return X86::COND_BE;
2931 case ISD::SETNE: return X86::COND_NE;
2932 case ISD::SETUO: return X86::COND_P;
2933 case ISD::SETO: return X86::COND_NP;
2935 case ISD::SETUNE: return X86::COND_INVALID;
2939 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2940 /// code. Current x86 isa includes the following FP cmov instructions:
2941 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2942 static bool hasFPCMov(unsigned X86CC) {
2958 /// isFPImmLegal - Returns true if the target can instruction select the
2959 /// specified FP immediate natively. If false, the legalizer will
2960 /// materialize the FP immediate as a load from a constant pool.
2961 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2962 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2963 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2969 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2970 /// the specified range (L, H].
2971 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2972 return (Val < 0) || (Val >= Low && Val < Hi);
2975 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2976 /// specified value.
2977 static bool isUndefOrEqual(int Val, int CmpVal) {
2978 if (Val < 0 || Val == CmpVal)
2983 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2984 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2985 /// the second operand.
2986 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2987 if (VT == MVT::v4f32 || VT == MVT::v4i32 )
2988 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2989 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2990 return (Mask[0] < 2 && Mask[1] < 2);
2994 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2995 SmallVector<int, 8> M;
2997 return ::isPSHUFDMask(M, N->getValueType(0));
3000 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
3001 /// is suitable for input to PSHUFHW.
3002 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3003 if (VT != MVT::v8i16)
3006 // Lower quadword copied in order or undef.
3007 for (int i = 0; i != 4; ++i)
3008 if (Mask[i] >= 0 && Mask[i] != i)
3011 // Upper quadword shuffled.
3012 for (int i = 4; i != 8; ++i)
3013 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
3019 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
3020 SmallVector<int, 8> M;
3022 return ::isPSHUFHWMask(M, N->getValueType(0));
3025 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
3026 /// is suitable for input to PSHUFLW.
3027 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3028 if (VT != MVT::v8i16)
3031 // Upper quadword copied in order.
3032 for (int i = 4; i != 8; ++i)
3033 if (Mask[i] >= 0 && Mask[i] != i)
3036 // Lower quadword shuffled.
3037 for (int i = 0; i != 4; ++i)
3044 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
3045 SmallVector<int, 8> M;
3047 return ::isPSHUFLWMask(M, N->getValueType(0));
3050 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
3051 /// is suitable for input to PALIGNR.
3052 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
3054 int i, e = VT.getVectorNumElements();
3056 // Do not handle v2i64 / v2f64 shuffles with palignr.
3057 if (e < 4 || !hasSSSE3)
3060 for (i = 0; i != e; ++i)
3064 // All undef, not a palignr.
3068 // Determine if it's ok to perform a palignr with only the LHS, since we
3069 // don't have access to the actual shuffle elements to see if RHS is undef.
3070 bool Unary = Mask[i] < (int)e;
3071 bool NeedsUnary = false;
3073 int s = Mask[i] - i;
3075 // Check the rest of the elements to see if they are consecutive.
3076 for (++i; i != e; ++i) {
3081 Unary = Unary && (m < (int)e);
3082 NeedsUnary = NeedsUnary || (m < s);
3084 if (NeedsUnary && !Unary)
3086 if (Unary && m != ((s+i) & (e-1)))
3088 if (!Unary && m != (s+i))
3094 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
3095 SmallVector<int, 8> M;
3097 return ::isPALIGNRMask(M, N->getValueType(0), true);
3100 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
3101 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
3102 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3103 int NumElems = VT.getVectorNumElements();
3104 if (NumElems != 2 && NumElems != 4)
3107 int Half = NumElems / 2;
3108 for (int i = 0; i < Half; ++i)
3109 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3111 for (int i = Half; i < NumElems; ++i)
3112 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3118 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
3119 SmallVector<int, 8> M;
3121 return ::isSHUFPMask(M, N->getValueType(0));
3124 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
3125 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
3126 /// half elements to come from vector 1 (which would equal the dest.) and
3127 /// the upper half to come from vector 2.
3128 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3129 int NumElems = VT.getVectorNumElements();
3131 if (NumElems != 2 && NumElems != 4)
3134 int Half = NumElems / 2;
3135 for (int i = 0; i < Half; ++i)
3136 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3138 for (int i = Half; i < NumElems; ++i)
3139 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3144 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
3145 SmallVector<int, 8> M;
3147 return isCommutedSHUFPMask(M, N->getValueType(0));
3150 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
3151 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
3152 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
3153 if (N->getValueType(0).getVectorNumElements() != 4)
3156 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
3157 return isUndefOrEqual(N->getMaskElt(0), 6) &&
3158 isUndefOrEqual(N->getMaskElt(1), 7) &&
3159 isUndefOrEqual(N->getMaskElt(2), 2) &&
3160 isUndefOrEqual(N->getMaskElt(3), 3);
3163 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
3164 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
3166 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
3167 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3172 return isUndefOrEqual(N->getMaskElt(0), 2) &&
3173 isUndefOrEqual(N->getMaskElt(1), 3) &&
3174 isUndefOrEqual(N->getMaskElt(2), 2) &&
3175 isUndefOrEqual(N->getMaskElt(3), 3);
3178 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
3179 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
3180 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
3181 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3183 if (NumElems != 2 && NumElems != 4)
3186 for (unsigned i = 0; i < NumElems/2; ++i)
3187 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
3190 for (unsigned i = NumElems/2; i < NumElems; ++i)
3191 if (!isUndefOrEqual(N->getMaskElt(i), i))
3197 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
3198 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
3199 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
3200 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3202 if ((NumElems != 2 && NumElems != 4)
3203 || N->getValueType(0).getSizeInBits() > 128)
3206 for (unsigned i = 0; i < NumElems/2; ++i)
3207 if (!isUndefOrEqual(N->getMaskElt(i), i))
3210 for (unsigned i = 0; i < NumElems/2; ++i)
3211 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
3217 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
3218 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
3219 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3220 bool V2IsSplat = false) {
3221 int NumElts = VT.getVectorNumElements();
3222 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3225 // Handle vector lengths > 128 bits. Define a "section" as a set of
3226 // 128 bits. AVX defines UNPCK* to operate independently on 128-bit
3228 unsigned NumSections = VT.getSizeInBits() / 128;
3229 if (NumSections == 0 ) NumSections = 1; // Handle MMX
3230 unsigned NumSectionElts = NumElts / NumSections;
3233 unsigned End = NumSectionElts;
3234 for (unsigned s = 0; s < NumSections; ++s) {
3235 for (unsigned i = Start, j = s * NumSectionElts;
3239 int BitI1 = Mask[i+1];
3240 if (!isUndefOrEqual(BitI, j))
3243 if (!isUndefOrEqual(BitI1, NumElts))
3246 if (!isUndefOrEqual(BitI1, j + NumElts))
3250 // Process the next 128 bits.
3251 Start += NumSectionElts;
3252 End += NumSectionElts;
3258 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3259 SmallVector<int, 8> M;
3261 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
3264 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
3265 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
3266 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
3267 bool V2IsSplat = false) {
3268 int NumElts = VT.getVectorNumElements();
3269 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3272 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
3274 int BitI1 = Mask[i+1];
3275 if (!isUndefOrEqual(BitI, j + NumElts/2))
3278 if (isUndefOrEqual(BitI1, NumElts))
3281 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
3288 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3289 SmallVector<int, 8> M;
3291 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
3294 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
3295 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
3297 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3298 int NumElems = VT.getVectorNumElements();
3299 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3302 // Handle vector lengths > 128 bits. Define a "section" as a set of
3303 // 128 bits. AVX defines UNPCK* to operate independently on 128-bit
3305 unsigned NumSections = VT.getSizeInBits() / 128;
3306 if (NumSections == 0 ) NumSections = 1; // Handle MMX
3307 unsigned NumSectionElts = NumElems / NumSections;
3309 for (unsigned s = 0; s < NumSections; ++s) {
3310 for (unsigned i = s * NumSectionElts, j = s * NumSectionElts;
3311 i != NumSectionElts * (s + 1);
3314 int BitI1 = Mask[i+1];
3316 if (!isUndefOrEqual(BitI, j))
3318 if (!isUndefOrEqual(BitI1, j))
3326 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
3327 SmallVector<int, 8> M;
3329 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
3332 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
3333 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
3335 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3336 int NumElems = VT.getVectorNumElements();
3337 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3340 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
3342 int BitI1 = Mask[i+1];
3343 if (!isUndefOrEqual(BitI, j))
3345 if (!isUndefOrEqual(BitI1, j))
3351 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
3352 SmallVector<int, 8> M;
3354 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3357 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3358 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3359 /// MOVSD, and MOVD, i.e. setting the lowest element.
3360 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3361 if (VT.getVectorElementType().getSizeInBits() < 32)
3364 int NumElts = VT.getVectorNumElements();
3366 if (!isUndefOrEqual(Mask[0], NumElts))
3369 for (int i = 1; i < NumElts; ++i)
3370 if (!isUndefOrEqual(Mask[i], i))
3376 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3377 SmallVector<int, 8> M;
3379 return ::isMOVLMask(M, N->getValueType(0));
3382 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3383 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3384 /// element of vector 2 and the other elements to come from vector 1 in order.
3385 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3386 bool V2IsSplat = false, bool V2IsUndef = false) {
3387 int NumOps = VT.getVectorNumElements();
3388 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3391 if (!isUndefOrEqual(Mask[0], 0))
3394 for (int i = 1; i < NumOps; ++i)
3395 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3396 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3397 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3403 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3404 bool V2IsUndef = false) {
3405 SmallVector<int, 8> M;
3407 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3410 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3411 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3412 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3413 if (N->getValueType(0).getVectorNumElements() != 4)
3416 // Expect 1, 1, 3, 3
3417 for (unsigned i = 0; i < 2; ++i) {
3418 int Elt = N->getMaskElt(i);
3419 if (Elt >= 0 && Elt != 1)
3424 for (unsigned i = 2; i < 4; ++i) {
3425 int Elt = N->getMaskElt(i);
3426 if (Elt >= 0 && Elt != 3)
3431 // Don't use movshdup if it can be done with a shufps.
3432 // FIXME: verify that matching u, u, 3, 3 is what we want.
3436 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3437 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3438 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3439 if (N->getValueType(0).getVectorNumElements() != 4)
3442 // Expect 0, 0, 2, 2
3443 for (unsigned i = 0; i < 2; ++i)
3444 if (N->getMaskElt(i) > 0)
3448 for (unsigned i = 2; i < 4; ++i) {
3449 int Elt = N->getMaskElt(i);
3450 if (Elt >= 0 && Elt != 2)
3455 // Don't use movsldup if it can be done with a shufps.
3459 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3460 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3461 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3462 int e = N->getValueType(0).getVectorNumElements() / 2;
3464 for (int i = 0; i < e; ++i)
3465 if (!isUndefOrEqual(N->getMaskElt(i), i))
3467 for (int i = 0; i < e; ++i)
3468 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3473 /// isVEXTRACTF128Index - Return true if the specified
3474 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
3475 /// suitable for input to VEXTRACTF128.
3476 bool X86::isVEXTRACTF128Index(SDNode *N) {
3477 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3480 // The index should be aligned on a 128-bit boundary.
3482 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3484 unsigned VL = N->getValueType(0).getVectorNumElements();
3485 unsigned VBits = N->getValueType(0).getSizeInBits();
3486 unsigned ElSize = VBits / VL;
3487 bool Result = (Index * ElSize) % 128 == 0;
3492 /// isVINSERTF128Index - Return true if the specified INSERT_SUBVECTOR
3493 /// operand specifies a subvector insert that is suitable for input to
3495 bool X86::isVINSERTF128Index(SDNode *N) {
3496 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3499 // The index should be aligned on a 128-bit boundary.
3501 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3503 unsigned VL = N->getValueType(0).getVectorNumElements();
3504 unsigned VBits = N->getValueType(0).getSizeInBits();
3505 unsigned ElSize = VBits / VL;
3506 bool Result = (Index * ElSize) % 128 == 0;
3511 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3512 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3513 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3514 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3515 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3517 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3519 for (int i = 0; i < NumOperands; ++i) {
3520 int Val = SVOp->getMaskElt(NumOperands-i-1);
3521 if (Val < 0) Val = 0;
3522 if (Val >= NumOperands) Val -= NumOperands;
3524 if (i != NumOperands - 1)
3530 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3531 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3532 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3533 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3535 // 8 nodes, but we only care about the last 4.
3536 for (unsigned i = 7; i >= 4; --i) {
3537 int Val = SVOp->getMaskElt(i);
3546 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3547 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3548 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3549 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3551 // 8 nodes, but we only care about the first 4.
3552 for (int i = 3; i >= 0; --i) {
3553 int Val = SVOp->getMaskElt(i);
3562 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3563 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3564 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3565 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3566 EVT VVT = N->getValueType(0);
3567 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3571 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3572 Val = SVOp->getMaskElt(i);
3576 return (Val - i) * EltSize;
3579 /// getExtractVEXTRACTF128Immediate - Return the appropriate immediate
3580 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
3582 unsigned X86::getExtractVEXTRACTF128Immediate(SDNode *N) {
3583 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3584 llvm_unreachable("Illegal extract subvector for VEXTRACTF128");
3587 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3589 EVT VecVT = N->getOperand(0).getValueType();
3590 EVT ElVT = VecVT.getVectorElementType();
3592 unsigned NumElemsPerChunk = 128 / ElVT.getSizeInBits();
3594 return Index / NumElemsPerChunk;
3597 /// getInsertVINSERTF128Immediate - Return the appropriate immediate
3598 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
3600 unsigned X86::getInsertVINSERTF128Immediate(SDNode *N) {
3601 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3602 llvm_unreachable("Illegal insert subvector for VINSERTF128");
3605 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3607 EVT VecVT = N->getValueType(0);
3608 EVT ElVT = VecVT.getVectorElementType();
3610 unsigned NumElemsPerChunk = 128 / ElVT.getSizeInBits();
3612 return Index / NumElemsPerChunk;
3615 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3617 bool X86::isZeroNode(SDValue Elt) {
3618 return ((isa<ConstantSDNode>(Elt) &&
3619 cast<ConstantSDNode>(Elt)->isNullValue()) ||
3620 (isa<ConstantFPSDNode>(Elt) &&
3621 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3624 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3625 /// their permute mask.
3626 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3627 SelectionDAG &DAG) {
3628 EVT VT = SVOp->getValueType(0);
3629 unsigned NumElems = VT.getVectorNumElements();
3630 SmallVector<int, 8> MaskVec;
3632 for (unsigned i = 0; i != NumElems; ++i) {
3633 int idx = SVOp->getMaskElt(i);
3635 MaskVec.push_back(idx);
3636 else if (idx < (int)NumElems)
3637 MaskVec.push_back(idx + NumElems);
3639 MaskVec.push_back(idx - NumElems);
3641 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3642 SVOp->getOperand(0), &MaskVec[0]);
3645 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3646 /// the two vector operands have swapped position.
3647 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3648 unsigned NumElems = VT.getVectorNumElements();
3649 for (unsigned i = 0; i != NumElems; ++i) {
3653 else if (idx < (int)NumElems)
3654 Mask[i] = idx + NumElems;
3656 Mask[i] = idx - NumElems;
3660 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3661 /// match movhlps. The lower half elements should come from upper half of
3662 /// V1 (and in order), and the upper half elements should come from the upper
3663 /// half of V2 (and in order).
3664 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3665 if (Op->getValueType(0).getVectorNumElements() != 4)
3667 for (unsigned i = 0, e = 2; i != e; ++i)
3668 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3670 for (unsigned i = 2; i != 4; ++i)
3671 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3676 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3677 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3679 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3680 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3682 N = N->getOperand(0).getNode();
3683 if (!ISD::isNON_EXTLoad(N))
3686 *LD = cast<LoadSDNode>(N);
3690 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3691 /// match movlp{s|d}. The lower half elements should come from lower half of
3692 /// V1 (and in order), and the upper half elements should come from the upper
3693 /// half of V2 (and in order). And since V1 will become the source of the
3694 /// MOVLP, it must be either a vector load or a scalar load to vector.
3695 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3696 ShuffleVectorSDNode *Op) {
3697 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3699 // Is V2 is a vector load, don't do this transformation. We will try to use
3700 // load folding shufps op.
3701 if (ISD::isNON_EXTLoad(V2))
3704 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3706 if (NumElems != 2 && NumElems != 4)
3708 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3709 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3711 for (unsigned i = NumElems/2; i != NumElems; ++i)
3712 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3717 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3719 static bool isSplatVector(SDNode *N) {
3720 if (N->getOpcode() != ISD::BUILD_VECTOR)
3723 SDValue SplatValue = N->getOperand(0);
3724 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3725 if (N->getOperand(i) != SplatValue)
3730 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3731 /// to an zero vector.
3732 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3733 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3734 SDValue V1 = N->getOperand(0);
3735 SDValue V2 = N->getOperand(1);
3736 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3737 for (unsigned i = 0; i != NumElems; ++i) {
3738 int Idx = N->getMaskElt(i);
3739 if (Idx >= (int)NumElems) {
3740 unsigned Opc = V2.getOpcode();
3741 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3743 if (Opc != ISD::BUILD_VECTOR ||
3744 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3746 } else if (Idx >= 0) {
3747 unsigned Opc = V1.getOpcode();
3748 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3750 if (Opc != ISD::BUILD_VECTOR ||
3751 !X86::isZeroNode(V1.getOperand(Idx)))
3758 /// getZeroVector - Returns a vector of specified type with all zero elements.
3760 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3762 assert(VT.isVector() && "Expected a vector type");
3764 // Always build SSE zero vectors as <4 x i32> bitcasted
3765 // to their dest type. This ensures they get CSE'd.
3767 if (VT.getSizeInBits() == 128) { // SSE
3768 if (HasSSE2) { // SSE2
3769 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3770 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3772 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3773 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3775 } else if (VT.getSizeInBits() == 256) { // AVX
3776 // 256-bit logic and arithmetic instructions in AVX are
3777 // all floating-point, no support for integer ops. Default
3778 // to emitting fp zeroed vectors then.
3779 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3780 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
3781 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops, 8);
3783 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
3786 /// getOnesVector - Returns a vector of specified type with all bits set.
3788 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3789 assert(VT.isVector() && "Expected a vector type");
3791 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3792 // type. This ensures they get CSE'd.
3793 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3795 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3796 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
3800 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3801 /// that point to V2 points to its first element.
3802 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3803 EVT VT = SVOp->getValueType(0);
3804 unsigned NumElems = VT.getVectorNumElements();
3806 bool Changed = false;
3807 SmallVector<int, 8> MaskVec;
3808 SVOp->getMask(MaskVec);
3810 for (unsigned i = 0; i != NumElems; ++i) {
3811 if (MaskVec[i] > (int)NumElems) {
3812 MaskVec[i] = NumElems;
3817 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3818 SVOp->getOperand(1), &MaskVec[0]);
3819 return SDValue(SVOp, 0);
3822 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3823 /// operation of specified width.
3824 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3826 unsigned NumElems = VT.getVectorNumElements();
3827 SmallVector<int, 8> Mask;
3828 Mask.push_back(NumElems);
3829 for (unsigned i = 1; i != NumElems; ++i)
3831 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3834 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3835 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3837 unsigned NumElems = VT.getVectorNumElements();
3838 SmallVector<int, 8> Mask;
3839 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3841 Mask.push_back(i + NumElems);
3843 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3846 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3847 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3849 unsigned NumElems = VT.getVectorNumElements();
3850 unsigned Half = NumElems/2;
3851 SmallVector<int, 8> Mask;
3852 for (unsigned i = 0; i != Half; ++i) {
3853 Mask.push_back(i + Half);
3854 Mask.push_back(i + NumElems + Half);
3856 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3859 /// PromoteSplat - Promote a splat of v4i32, v8i16 or v16i8 to v4f32.
3860 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
3861 EVT PVT = MVT::v4f32;
3862 EVT VT = SV->getValueType(0);
3863 DebugLoc dl = SV->getDebugLoc();
3864 SDValue V1 = SV->getOperand(0);
3865 int NumElems = VT.getVectorNumElements();
3866 int EltNo = SV->getSplatIndex();
3868 // unpack elements to the correct location
3869 while (NumElems > 4) {
3870 if (EltNo < NumElems/2) {
3871 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3873 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3874 EltNo -= NumElems/2;
3879 // Perform the splat.
3880 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3881 V1 = DAG.getNode(ISD::BITCAST, dl, PVT, V1);
3882 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3883 return DAG.getNode(ISD::BITCAST, dl, VT, V1);
3886 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3887 /// vector of zero or undef vector. This produces a shuffle where the low
3888 /// element of V2 is swizzled into the zero/undef vector, landing at element
3889 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3890 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3891 bool isZero, bool HasSSE2,
3892 SelectionDAG &DAG) {
3893 EVT VT = V2.getValueType();
3895 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3896 unsigned NumElems = VT.getVectorNumElements();
3897 SmallVector<int, 16> MaskVec;
3898 for (unsigned i = 0; i != NumElems; ++i)
3899 // If this is the insertion idx, put the low elt of V2 here.
3900 MaskVec.push_back(i == Idx ? NumElems : i);
3901 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3904 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
3905 /// element of the result of the vector shuffle.
3906 static SDValue getShuffleScalarElt(SDNode *N, int Index, SelectionDAG &DAG,
3909 return SDValue(); // Limit search depth.
3911 SDValue V = SDValue(N, 0);
3912 EVT VT = V.getValueType();
3913 unsigned Opcode = V.getOpcode();
3915 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
3916 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
3917 Index = SV->getMaskElt(Index);
3920 return DAG.getUNDEF(VT.getVectorElementType());
3922 int NumElems = VT.getVectorNumElements();
3923 SDValue NewV = (Index < NumElems) ? SV->getOperand(0) : SV->getOperand(1);
3924 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG, Depth+1);
3927 // Recurse into target specific vector shuffles to find scalars.
3928 if (isTargetShuffle(Opcode)) {
3929 int NumElems = VT.getVectorNumElements();
3930 SmallVector<unsigned, 16> ShuffleMask;
3934 case X86ISD::SHUFPS:
3935 case X86ISD::SHUFPD:
3936 ImmN = N->getOperand(N->getNumOperands()-1);
3937 DecodeSHUFPSMask(NumElems,
3938 cast<ConstantSDNode>(ImmN)->getZExtValue(),
3941 case X86ISD::PUNPCKHBW:
3942 case X86ISD::PUNPCKHWD:
3943 case X86ISD::PUNPCKHDQ:
3944 case X86ISD::PUNPCKHQDQ:
3945 DecodePUNPCKHMask(NumElems, ShuffleMask);
3947 case X86ISD::UNPCKHPS:
3948 case X86ISD::UNPCKHPD:
3949 DecodeUNPCKHPMask(NumElems, ShuffleMask);
3951 case X86ISD::PUNPCKLBW:
3952 case X86ISD::PUNPCKLWD:
3953 case X86ISD::PUNPCKLDQ:
3954 case X86ISD::PUNPCKLQDQ:
3955 DecodePUNPCKLMask(VT, ShuffleMask);
3957 case X86ISD::UNPCKLPS:
3958 case X86ISD::UNPCKLPD:
3959 case X86ISD::VUNPCKLPS:
3960 case X86ISD::VUNPCKLPD:
3961 case X86ISD::VUNPCKLPSY:
3962 case X86ISD::VUNPCKLPDY:
3963 DecodeUNPCKLPMask(VT, ShuffleMask);
3965 case X86ISD::MOVHLPS:
3966 DecodeMOVHLPSMask(NumElems, ShuffleMask);
3968 case X86ISD::MOVLHPS:
3969 DecodeMOVLHPSMask(NumElems, ShuffleMask);
3971 case X86ISD::PSHUFD:
3972 ImmN = N->getOperand(N->getNumOperands()-1);
3973 DecodePSHUFMask(NumElems,
3974 cast<ConstantSDNode>(ImmN)->getZExtValue(),
3977 case X86ISD::PSHUFHW:
3978 ImmN = N->getOperand(N->getNumOperands()-1);
3979 DecodePSHUFHWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
3982 case X86ISD::PSHUFLW:
3983 ImmN = N->getOperand(N->getNumOperands()-1);
3984 DecodePSHUFLWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
3988 case X86ISD::MOVSD: {
3989 // The index 0 always comes from the first element of the second source,
3990 // this is why MOVSS and MOVSD are used in the first place. The other
3991 // elements come from the other positions of the first source vector.
3992 unsigned OpNum = (Index == 0) ? 1 : 0;
3993 return getShuffleScalarElt(V.getOperand(OpNum).getNode(), Index, DAG,
3997 assert("not implemented for target shuffle node");
4001 Index = ShuffleMask[Index];
4003 return DAG.getUNDEF(VT.getVectorElementType());
4005 SDValue NewV = (Index < NumElems) ? N->getOperand(0) : N->getOperand(1);
4006 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG,
4010 // Actual nodes that may contain scalar elements
4011 if (Opcode == ISD::BITCAST) {
4012 V = V.getOperand(0);
4013 EVT SrcVT = V.getValueType();
4014 unsigned NumElems = VT.getVectorNumElements();
4016 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4020 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4021 return (Index == 0) ? V.getOperand(0)
4022 : DAG.getUNDEF(VT.getVectorElementType());
4024 if (V.getOpcode() == ISD::BUILD_VECTOR)
4025 return V.getOperand(Index);
4030 /// getNumOfConsecutiveZeros - Return the number of elements of a vector
4031 /// shuffle operation which come from a consecutively from a zero. The
4032 /// search can start in two different directions, from left or right.
4034 unsigned getNumOfConsecutiveZeros(SDNode *N, int NumElems,
4035 bool ZerosFromLeft, SelectionDAG &DAG) {
4038 while (i < NumElems) {
4039 unsigned Index = ZerosFromLeft ? i : NumElems-i-1;
4040 SDValue Elt = getShuffleScalarElt(N, Index, DAG, 0);
4041 if (!(Elt.getNode() &&
4042 (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt))))
4050 /// isShuffleMaskConsecutive - Check if the shuffle mask indicies from MaskI to
4051 /// MaskE correspond consecutively to elements from one of the vector operands,
4052 /// starting from its index OpIdx. Also tell OpNum which source vector operand.
4054 bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp, int MaskI, int MaskE,
4055 int OpIdx, int NumElems, unsigned &OpNum) {
4056 bool SeenV1 = false;
4057 bool SeenV2 = false;
4059 for (int i = MaskI; i <= MaskE; ++i, ++OpIdx) {
4060 int Idx = SVOp->getMaskElt(i);
4061 // Ignore undef indicies
4070 // Only accept consecutive elements from the same vector
4071 if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2))
4075 OpNum = SeenV1 ? 0 : 1;
4079 /// isVectorShiftRight - Returns true if the shuffle can be implemented as a
4080 /// logical left shift of a vector.
4081 static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4082 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4083 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
4084 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
4085 false /* check zeros from right */, DAG);
4091 // Considering the elements in the mask that are not consecutive zeros,
4092 // check if they consecutively come from only one of the source vectors.
4094 // V1 = {X, A, B, C} 0
4096 // vector_shuffle V1, V2 <1, 2, 3, X>
4098 if (!isShuffleMaskConsecutive(SVOp,
4099 0, // Mask Start Index
4100 NumElems-NumZeros-1, // Mask End Index
4101 NumZeros, // Where to start looking in the src vector
4102 NumElems, // Number of elements in vector
4103 OpSrc)) // Which source operand ?
4108 ShVal = SVOp->getOperand(OpSrc);
4112 /// isVectorShiftLeft - Returns true if the shuffle can be implemented as a
4113 /// logical left shift of a vector.
4114 static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4115 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4116 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
4117 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
4118 true /* check zeros from left */, DAG);
4124 // Considering the elements in the mask that are not consecutive zeros,
4125 // check if they consecutively come from only one of the source vectors.
4127 // 0 { A, B, X, X } = V2
4129 // vector_shuffle V1, V2 <X, X, 4, 5>
4131 if (!isShuffleMaskConsecutive(SVOp,
4132 NumZeros, // Mask Start Index
4133 NumElems-1, // Mask End Index
4134 0, // Where to start looking in the src vector
4135 NumElems, // Number of elements in vector
4136 OpSrc)) // Which source operand ?
4141 ShVal = SVOp->getOperand(OpSrc);
4145 /// isVectorShift - Returns true if the shuffle can be implemented as a
4146 /// logical left or right shift of a vector.
4147 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4148 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4149 if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) ||
4150 isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt))
4156 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
4158 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4159 unsigned NumNonZero, unsigned NumZero,
4161 const TargetLowering &TLI) {
4165 DebugLoc dl = Op.getDebugLoc();
4168 for (unsigned i = 0; i < 16; ++i) {
4169 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4170 if (ThisIsNonZero && First) {
4172 V = getZeroVector(MVT::v8i16, true, DAG, dl);
4174 V = DAG.getUNDEF(MVT::v8i16);
4179 SDValue ThisElt(0, 0), LastElt(0, 0);
4180 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4181 if (LastIsNonZero) {
4182 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4183 MVT::i16, Op.getOperand(i-1));
4185 if (ThisIsNonZero) {
4186 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4187 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4188 ThisElt, DAG.getConstant(8, MVT::i8));
4190 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
4194 if (ThisElt.getNode())
4195 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
4196 DAG.getIntPtrConstant(i/2));
4200 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
4203 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
4205 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
4206 unsigned NumNonZero, unsigned NumZero,
4208 const TargetLowering &TLI) {
4212 DebugLoc dl = Op.getDebugLoc();
4215 for (unsigned i = 0; i < 8; ++i) {
4216 bool isNonZero = (NonZeros & (1 << i)) != 0;
4220 V = getZeroVector(MVT::v8i16, true, DAG, dl);
4222 V = DAG.getUNDEF(MVT::v8i16);
4225 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4226 MVT::v8i16, V, Op.getOperand(i),
4227 DAG.getIntPtrConstant(i));
4234 /// getVShift - Return a vector logical shift node.
4236 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
4237 unsigned NumBits, SelectionDAG &DAG,
4238 const TargetLowering &TLI, DebugLoc dl) {
4239 EVT ShVT = MVT::v2i64;
4240 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
4241 SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
4242 return DAG.getNode(ISD::BITCAST, dl, VT,
4243 DAG.getNode(Opc, dl, ShVT, SrcOp,
4244 DAG.getConstant(NumBits,
4245 TLI.getShiftAmountTy(SrcOp.getValueType()))));
4249 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
4250 SelectionDAG &DAG) const {
4252 // Check if the scalar load can be widened into a vector load. And if
4253 // the address is "base + cst" see if the cst can be "absorbed" into
4254 // the shuffle mask.
4255 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
4256 SDValue Ptr = LD->getBasePtr();
4257 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
4259 EVT PVT = LD->getValueType(0);
4260 if (PVT != MVT::i32 && PVT != MVT::f32)
4265 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4266 FI = FINode->getIndex();
4268 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
4269 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4270 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4271 Offset = Ptr.getConstantOperandVal(1);
4272 Ptr = Ptr.getOperand(0);
4277 SDValue Chain = LD->getChain();
4278 // Make sure the stack object alignment is at least 16.
4279 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4280 if (DAG.InferPtrAlignment(Ptr) < 16) {
4281 if (MFI->isFixedObjectIndex(FI)) {
4282 // Can't change the alignment. FIXME: It's possible to compute
4283 // the exact stack offset and reference FI + adjust offset instead.
4284 // If someone *really* cares about this. That's the way to implement it.
4287 MFI->setObjectAlignment(FI, 16);
4291 // (Offset % 16) must be multiple of 4. Then address is then
4292 // Ptr + (Offset & ~15).
4295 if ((Offset % 16) & 3)
4297 int64_t StartOffset = Offset & ~15;
4299 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
4300 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
4302 int EltNo = (Offset - StartOffset) >> 2;
4303 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
4304 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
4305 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,
4306 LD->getPointerInfo().getWithOffset(StartOffset),
4308 // Canonicalize it to a v4i32 shuffle.
4309 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V1);
4310 return DAG.getNode(ISD::BITCAST, dl, VT,
4311 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
4312 DAG.getUNDEF(MVT::v4i32),&Mask[0]));
4318 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
4319 /// vector of type 'VT', see if the elements can be replaced by a single large
4320 /// load which has the same value as a build_vector whose operands are 'elts'.
4322 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
4324 /// FIXME: we'd also like to handle the case where the last elements are zero
4325 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
4326 /// There's even a handy isZeroNode for that purpose.
4327 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
4328 DebugLoc &DL, SelectionDAG &DAG) {
4329 EVT EltVT = VT.getVectorElementType();
4330 unsigned NumElems = Elts.size();
4332 LoadSDNode *LDBase = NULL;
4333 unsigned LastLoadedElt = -1U;
4335 // For each element in the initializer, see if we've found a load or an undef.
4336 // If we don't find an initial load element, or later load elements are
4337 // non-consecutive, bail out.
4338 for (unsigned i = 0; i < NumElems; ++i) {
4339 SDValue Elt = Elts[i];
4341 if (!Elt.getNode() ||
4342 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
4345 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
4347 LDBase = cast<LoadSDNode>(Elt.getNode());
4351 if (Elt.getOpcode() == ISD::UNDEF)
4354 LoadSDNode *LD = cast<LoadSDNode>(Elt);
4355 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
4360 // If we have found an entire vector of loads and undefs, then return a large
4361 // load of the entire vector width starting at the base pointer. If we found
4362 // consecutive loads for the low half, generate a vzext_load node.
4363 if (LastLoadedElt == NumElems - 1) {
4364 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
4365 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4366 LDBase->getPointerInfo(),
4367 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
4368 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4369 LDBase->getPointerInfo(),
4370 LDBase->isVolatile(), LDBase->isNonTemporal(),
4371 LDBase->getAlignment());
4372 } else if (NumElems == 4 && LastLoadedElt == 1) {
4373 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
4374 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
4375 SDValue ResNode = DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys,
4377 LDBase->getMemOperand());
4378 return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
4384 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
4385 DebugLoc dl = Op.getDebugLoc();
4387 EVT VT = Op.getValueType();
4388 EVT ExtVT = VT.getVectorElementType();
4390 unsigned NumElems = Op.getNumOperands();
4392 // For AVX-length vectors, build the individual 128-bit pieces and
4393 // use shuffles to put them in place.
4394 if (VT.getSizeInBits() > 256 &&
4395 Subtarget->hasAVX() &&
4396 !ISD::isBuildVectorAllZeros(Op.getNode())) {
4397 SmallVector<SDValue, 8> V;
4399 for (unsigned i = 0; i < NumElems; ++i) {
4400 V[i] = Op.getOperand(i);
4403 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
4405 // Build the lower subvector.
4406 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[0], NumElems/2);
4407 // Build the upper subvector.
4408 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[NumElems / 2],
4411 return ConcatVectors(Lower, Upper, DAG);
4414 // All zero's are handled with pxor in SSE2 and above, xorps in SSE1.
4415 // All one's are handled with pcmpeqd. In AVX, zero's are handled with
4416 // vpxor in 128-bit and xor{pd,ps} in 256-bit, but no 256 version of pcmpeqd
4417 // is present, so AllOnes is ignored.
4418 if (ISD::isBuildVectorAllZeros(Op.getNode()) ||
4419 (Op.getValueType().getSizeInBits() != 256 &&
4420 ISD::isBuildVectorAllOnes(Op.getNode()))) {
4421 // Canonicalize this to <4 x i32> (SSE) to
4422 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
4423 // eliminated on x86-32 hosts.
4424 if (Op.getValueType() == MVT::v4i32)
4427 if (ISD::isBuildVectorAllOnes(Op.getNode()))
4428 return getOnesVector(Op.getValueType(), DAG, dl);
4429 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
4432 unsigned EVTBits = ExtVT.getSizeInBits();
4434 unsigned NumZero = 0;
4435 unsigned NumNonZero = 0;
4436 unsigned NonZeros = 0;
4437 bool IsAllConstants = true;
4438 SmallSet<SDValue, 8> Values;
4439 for (unsigned i = 0; i < NumElems; ++i) {
4440 SDValue Elt = Op.getOperand(i);
4441 if (Elt.getOpcode() == ISD::UNDEF)
4444 if (Elt.getOpcode() != ISD::Constant &&
4445 Elt.getOpcode() != ISD::ConstantFP)
4446 IsAllConstants = false;
4447 if (X86::isZeroNode(Elt))
4450 NonZeros |= (1 << i);
4455 // All undef vector. Return an UNDEF. All zero vectors were handled above.
4456 if (NumNonZero == 0)
4457 return DAG.getUNDEF(VT);
4459 // Special case for single non-zero, non-undef, element.
4460 if (NumNonZero == 1) {
4461 unsigned Idx = CountTrailingZeros_32(NonZeros);
4462 SDValue Item = Op.getOperand(Idx);
4464 // If this is an insertion of an i64 value on x86-32, and if the top bits of
4465 // the value are obviously zero, truncate the value to i32 and do the
4466 // insertion that way. Only do this if the value is non-constant or if the
4467 // value is a constant being inserted into element 0. It is cheaper to do
4468 // a constant pool load than it is to do a movd + shuffle.
4469 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
4470 (!IsAllConstants || Idx == 0)) {
4471 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
4473 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
4474 EVT VecVT = MVT::v4i32;
4475 unsigned VecElts = 4;
4477 // Truncate the value (which may itself be a constant) to i32, and
4478 // convert it to a vector with movd (S2V+shuffle to zero extend).
4479 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
4480 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
4481 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4482 Subtarget->hasSSE2(), DAG);
4484 // Now we have our 32-bit value zero extended in the low element of
4485 // a vector. If Idx != 0, swizzle it into place.
4487 SmallVector<int, 4> Mask;
4488 Mask.push_back(Idx);
4489 for (unsigned i = 1; i != VecElts; ++i)
4491 Item = DAG.getVectorShuffle(VecVT, dl, Item,
4492 DAG.getUNDEF(Item.getValueType()),
4495 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Item);
4499 // If we have a constant or non-constant insertion into the low element of
4500 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
4501 // the rest of the elements. This will be matched as movd/movq/movss/movsd
4502 // depending on what the source datatype is.
4505 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4506 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
4507 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
4508 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4509 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
4510 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
4512 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
4513 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
4514 assert(VT.getSizeInBits() == 128 && "Expected an SSE value type!");
4515 EVT MiddleVT = MVT::v4i32;
4516 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
4517 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4518 Subtarget->hasSSE2(), DAG);
4519 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
4523 // Is it a vector logical left shift?
4524 if (NumElems == 2 && Idx == 1 &&
4525 X86::isZeroNode(Op.getOperand(0)) &&
4526 !X86::isZeroNode(Op.getOperand(1))) {
4527 unsigned NumBits = VT.getSizeInBits();
4528 return getVShift(true, VT,
4529 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4530 VT, Op.getOperand(1)),
4531 NumBits/2, DAG, *this, dl);
4534 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
4537 // Otherwise, if this is a vector with i32 or f32 elements, and the element
4538 // is a non-constant being inserted into an element other than the low one,
4539 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
4540 // movd/movss) to move this into the low element, then shuffle it into
4542 if (EVTBits == 32) {
4543 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4545 // Turn it into a shuffle of zero and zero-extended scalar to vector.
4546 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
4547 Subtarget->hasSSE2(), DAG);
4548 SmallVector<int, 8> MaskVec;
4549 for (unsigned i = 0; i < NumElems; i++)
4550 MaskVec.push_back(i == Idx ? 0 : 1);
4551 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
4555 // Splat is obviously ok. Let legalizer expand it to a shuffle.
4556 if (Values.size() == 1) {
4557 if (EVTBits == 32) {
4558 // Instead of a shuffle like this:
4559 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
4560 // Check if it's possible to issue this instead.
4561 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
4562 unsigned Idx = CountTrailingZeros_32(NonZeros);
4563 SDValue Item = Op.getOperand(Idx);
4564 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
4565 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
4570 // A vector full of immediates; various special cases are already
4571 // handled, so this is best done with a single constant-pool load.
4575 // Let legalizer expand 2-wide build_vectors.
4576 if (EVTBits == 64) {
4577 if (NumNonZero == 1) {
4578 // One half is zero or undef.
4579 unsigned Idx = CountTrailingZeros_32(NonZeros);
4580 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
4581 Op.getOperand(Idx));
4582 return getShuffleVectorZeroOrUndef(V2, Idx, true,
4583 Subtarget->hasSSE2(), DAG);
4588 // If element VT is < 32 bits, convert it to inserts into a zero vector.
4589 if (EVTBits == 8 && NumElems == 16) {
4590 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
4592 if (V.getNode()) return V;
4595 if (EVTBits == 16 && NumElems == 8) {
4596 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
4598 if (V.getNode()) return V;
4601 // If element VT is == 32 bits, turn it into a number of shuffles.
4602 SmallVector<SDValue, 8> V;
4604 if (NumElems == 4 && NumZero > 0) {
4605 for (unsigned i = 0; i < 4; ++i) {
4606 bool isZero = !(NonZeros & (1 << i));
4608 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4610 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4613 for (unsigned i = 0; i < 2; ++i) {
4614 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
4617 V[i] = V[i*2]; // Must be a zero vector.
4620 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
4623 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
4626 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
4631 SmallVector<int, 8> MaskVec;
4632 bool Reverse = (NonZeros & 0x3) == 2;
4633 for (unsigned i = 0; i < 2; ++i)
4634 MaskVec.push_back(Reverse ? 1-i : i);
4635 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
4636 for (unsigned i = 0; i < 2; ++i)
4637 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
4638 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
4641 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
4642 // Check for a build vector of consecutive loads.
4643 for (unsigned i = 0; i < NumElems; ++i)
4644 V[i] = Op.getOperand(i);
4646 // Check for elements which are consecutive loads.
4647 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
4651 // For SSE 4.1, use insertps to put the high elements into the low element.
4652 if (getSubtarget()->hasSSE41()) {
4654 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
4655 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
4657 Result = DAG.getUNDEF(VT);
4659 for (unsigned i = 1; i < NumElems; ++i) {
4660 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
4661 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
4662 Op.getOperand(i), DAG.getIntPtrConstant(i));
4667 // Otherwise, expand into a number of unpckl*, start by extending each of
4668 // our (non-undef) elements to the full vector width with the element in the
4669 // bottom slot of the vector (which generates no code for SSE).
4670 for (unsigned i = 0; i < NumElems; ++i) {
4671 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
4672 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4674 V[i] = DAG.getUNDEF(VT);
4677 // Next, we iteratively mix elements, e.g. for v4f32:
4678 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
4679 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
4680 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
4681 unsigned EltStride = NumElems >> 1;
4682 while (EltStride != 0) {
4683 for (unsigned i = 0; i < EltStride; ++i) {
4684 // If V[i+EltStride] is undef and this is the first round of mixing,
4685 // then it is safe to just drop this shuffle: V[i] is already in the
4686 // right place, the one element (since it's the first round) being
4687 // inserted as undef can be dropped. This isn't safe for successive
4688 // rounds because they will permute elements within both vectors.
4689 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
4690 EltStride == NumElems/2)
4693 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
4703 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
4704 // We support concatenate two MMX registers and place them in a MMX
4705 // register. This is better than doing a stack convert.
4706 DebugLoc dl = Op.getDebugLoc();
4707 EVT ResVT = Op.getValueType();
4708 assert(Op.getNumOperands() == 2);
4709 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
4710 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
4712 SDValue InVec = DAG.getNode(ISD::BITCAST,dl, MVT::v1i64, Op.getOperand(0));
4713 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4714 InVec = Op.getOperand(1);
4715 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4716 unsigned NumElts = ResVT.getVectorNumElements();
4717 VecOp = DAG.getNode(ISD::BITCAST, dl, ResVT, VecOp);
4718 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4719 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4721 InVec = DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, InVec);
4722 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4723 Mask[0] = 0; Mask[1] = 2;
4724 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4726 return DAG.getNode(ISD::BITCAST, dl, ResVT, VecOp);
4729 // v8i16 shuffles - Prefer shuffles in the following order:
4730 // 1. [all] pshuflw, pshufhw, optional move
4731 // 2. [ssse3] 1 x pshufb
4732 // 3. [ssse3] 2 x pshufb + 1 x por
4733 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4735 X86TargetLowering::LowerVECTOR_SHUFFLEv8i16(SDValue Op,
4736 SelectionDAG &DAG) const {
4737 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4738 SDValue V1 = SVOp->getOperand(0);
4739 SDValue V2 = SVOp->getOperand(1);
4740 DebugLoc dl = SVOp->getDebugLoc();
4741 SmallVector<int, 8> MaskVals;
4743 // Determine if more than 1 of the words in each of the low and high quadwords
4744 // of the result come from the same quadword of one of the two inputs. Undef
4745 // mask values count as coming from any quadword, for better codegen.
4746 SmallVector<unsigned, 4> LoQuad(4);
4747 SmallVector<unsigned, 4> HiQuad(4);
4748 BitVector InputQuads(4);
4749 for (unsigned i = 0; i < 8; ++i) {
4750 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4751 int EltIdx = SVOp->getMaskElt(i);
4752 MaskVals.push_back(EltIdx);
4761 InputQuads.set(EltIdx / 4);
4764 int BestLoQuad = -1;
4765 unsigned MaxQuad = 1;
4766 for (unsigned i = 0; i < 4; ++i) {
4767 if (LoQuad[i] > MaxQuad) {
4769 MaxQuad = LoQuad[i];
4773 int BestHiQuad = -1;
4775 for (unsigned i = 0; i < 4; ++i) {
4776 if (HiQuad[i] > MaxQuad) {
4778 MaxQuad = HiQuad[i];
4782 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4783 // of the two input vectors, shuffle them into one input vector so only a
4784 // single pshufb instruction is necessary. If There are more than 2 input
4785 // quads, disable the next transformation since it does not help SSSE3.
4786 bool V1Used = InputQuads[0] || InputQuads[1];
4787 bool V2Used = InputQuads[2] || InputQuads[3];
4788 if (Subtarget->hasSSSE3()) {
4789 if (InputQuads.count() == 2 && V1Used && V2Used) {
4790 BestLoQuad = InputQuads.find_first();
4791 BestHiQuad = InputQuads.find_next(BestLoQuad);
4793 if (InputQuads.count() > 2) {
4799 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4800 // the shuffle mask. If a quad is scored as -1, that means that it contains
4801 // words from all 4 input quadwords.
4803 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4804 SmallVector<int, 8> MaskV;
4805 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4806 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4807 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4808 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1),
4809 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V2), &MaskV[0]);
4810 NewV = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, NewV);
4812 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4813 // source words for the shuffle, to aid later transformations.
4814 bool AllWordsInNewV = true;
4815 bool InOrder[2] = { true, true };
4816 for (unsigned i = 0; i != 8; ++i) {
4817 int idx = MaskVals[i];
4819 InOrder[i/4] = false;
4820 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4822 AllWordsInNewV = false;
4826 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4827 if (AllWordsInNewV) {
4828 for (int i = 0; i != 8; ++i) {
4829 int idx = MaskVals[i];
4832 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4833 if ((idx != i) && idx < 4)
4835 if ((idx != i) && idx > 3)
4844 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4845 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4846 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4847 unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
4848 unsigned TargetMask = 0;
4849 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4850 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4851 TargetMask = pshufhw ? X86::getShufflePSHUFHWImmediate(NewV.getNode()):
4852 X86::getShufflePSHUFLWImmediate(NewV.getNode());
4853 V1 = NewV.getOperand(0);
4854 return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
4858 // If we have SSSE3, and all words of the result are from 1 input vector,
4859 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4860 // is present, fall back to case 4.
4861 if (Subtarget->hasSSSE3()) {
4862 SmallVector<SDValue,16> pshufbMask;
4864 // If we have elements from both input vectors, set the high bit of the
4865 // shuffle mask element to zero out elements that come from V2 in the V1
4866 // mask, and elements that come from V1 in the V2 mask, so that the two
4867 // results can be OR'd together.
4868 bool TwoInputs = V1Used && V2Used;
4869 for (unsigned i = 0; i != 8; ++i) {
4870 int EltIdx = MaskVals[i] * 2;
4871 if (TwoInputs && (EltIdx >= 16)) {
4872 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4873 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4876 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4877 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4879 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V1);
4880 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4881 DAG.getNode(ISD::BUILD_VECTOR, dl,
4882 MVT::v16i8, &pshufbMask[0], 16));
4884 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
4886 // Calculate the shuffle mask for the second input, shuffle it, and
4887 // OR it with the first shuffled input.
4889 for (unsigned i = 0; i != 8; ++i) {
4890 int EltIdx = MaskVals[i] * 2;
4892 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4893 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4896 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4897 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4899 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V2);
4900 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4901 DAG.getNode(ISD::BUILD_VECTOR, dl,
4902 MVT::v16i8, &pshufbMask[0], 16));
4903 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4904 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
4907 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4908 // and update MaskVals with new element order.
4909 BitVector InOrder(8);
4910 if (BestLoQuad >= 0) {
4911 SmallVector<int, 8> MaskV;
4912 for (int i = 0; i != 4; ++i) {
4913 int idx = MaskVals[i];
4915 MaskV.push_back(-1);
4917 } else if ((idx / 4) == BestLoQuad) {
4918 MaskV.push_back(idx & 3);
4921 MaskV.push_back(-1);
4924 for (unsigned i = 4; i != 8; ++i)
4926 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4929 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
4930 NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
4932 X86::getShufflePSHUFLWImmediate(NewV.getNode()),
4936 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4937 // and update MaskVals with the new element order.
4938 if (BestHiQuad >= 0) {
4939 SmallVector<int, 8> MaskV;
4940 for (unsigned i = 0; i != 4; ++i)
4942 for (unsigned i = 4; i != 8; ++i) {
4943 int idx = MaskVals[i];
4945 MaskV.push_back(-1);
4947 } else if ((idx / 4) == BestHiQuad) {
4948 MaskV.push_back((idx & 3) + 4);
4951 MaskV.push_back(-1);
4954 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4957 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
4958 NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
4960 X86::getShufflePSHUFHWImmediate(NewV.getNode()),
4964 // In case BestHi & BestLo were both -1, which means each quadword has a word
4965 // from each of the four input quadwords, calculate the InOrder bitvector now
4966 // before falling through to the insert/extract cleanup.
4967 if (BestLoQuad == -1 && BestHiQuad == -1) {
4969 for (int i = 0; i != 8; ++i)
4970 if (MaskVals[i] < 0 || MaskVals[i] == i)
4974 // The other elements are put in the right place using pextrw and pinsrw.
4975 for (unsigned i = 0; i != 8; ++i) {
4978 int EltIdx = MaskVals[i];
4981 SDValue ExtOp = (EltIdx < 8)
4982 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4983 DAG.getIntPtrConstant(EltIdx))
4984 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4985 DAG.getIntPtrConstant(EltIdx - 8));
4986 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4987 DAG.getIntPtrConstant(i));
4992 // v16i8 shuffles - Prefer shuffles in the following order:
4993 // 1. [ssse3] 1 x pshufb
4994 // 2. [ssse3] 2 x pshufb + 1 x por
4995 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4997 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4999 const X86TargetLowering &TLI) {
5000 SDValue V1 = SVOp->getOperand(0);
5001 SDValue V2 = SVOp->getOperand(1);
5002 DebugLoc dl = SVOp->getDebugLoc();
5003 SmallVector<int, 16> MaskVals;
5004 SVOp->getMask(MaskVals);
5006 // If we have SSSE3, case 1 is generated when all result bytes come from
5007 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
5008 // present, fall back to case 3.
5009 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
5012 for (unsigned i = 0; i < 16; ++i) {
5013 int EltIdx = MaskVals[i];
5022 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
5023 if (TLI.getSubtarget()->hasSSSE3()) {
5024 SmallVector<SDValue,16> pshufbMask;
5026 // If all result elements are from one input vector, then only translate
5027 // undef mask values to 0x80 (zero out result) in the pshufb mask.
5029 // Otherwise, we have elements from both input vectors, and must zero out
5030 // elements that come from V2 in the first mask, and V1 in the second mask
5031 // so that we can OR them together.
5032 bool TwoInputs = !(V1Only || V2Only);
5033 for (unsigned i = 0; i != 16; ++i) {
5034 int EltIdx = MaskVals[i];
5035 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
5036 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5039 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
5041 // If all the elements are from V2, assign it to V1 and return after
5042 // building the first pshufb.
5045 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
5046 DAG.getNode(ISD::BUILD_VECTOR, dl,
5047 MVT::v16i8, &pshufbMask[0], 16));
5051 // Calculate the shuffle mask for the second input, shuffle it, and
5052 // OR it with the first shuffled input.
5054 for (unsigned i = 0; i != 16; ++i) {
5055 int EltIdx = MaskVals[i];
5057 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5060 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
5062 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
5063 DAG.getNode(ISD::BUILD_VECTOR, dl,
5064 MVT::v16i8, &pshufbMask[0], 16));
5065 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
5068 // No SSSE3 - Calculate in place words and then fix all out of place words
5069 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
5070 // the 16 different words that comprise the two doublequadword input vectors.
5071 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5072 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V2);
5073 SDValue NewV = V2Only ? V2 : V1;
5074 for (int i = 0; i != 8; ++i) {
5075 int Elt0 = MaskVals[i*2];
5076 int Elt1 = MaskVals[i*2+1];
5078 // This word of the result is all undef, skip it.
5079 if (Elt0 < 0 && Elt1 < 0)
5082 // This word of the result is already in the correct place, skip it.
5083 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
5085 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
5088 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
5089 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
5092 // If Elt0 and Elt1 are defined, are consecutive, and can be load
5093 // using a single extract together, load it and store it.
5094 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
5095 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
5096 DAG.getIntPtrConstant(Elt1 / 2));
5097 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
5098 DAG.getIntPtrConstant(i));
5102 // If Elt1 is defined, extract it from the appropriate source. If the
5103 // source byte is not also odd, shift the extracted word left 8 bits
5104 // otherwise clear the bottom 8 bits if we need to do an or.
5106 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
5107 DAG.getIntPtrConstant(Elt1 / 2));
5108 if ((Elt1 & 1) == 0)
5109 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
5111 TLI.getShiftAmountTy(InsElt.getValueType())));
5113 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
5114 DAG.getConstant(0xFF00, MVT::i16));
5116 // If Elt0 is defined, extract it from the appropriate source. If the
5117 // source byte is not also even, shift the extracted word right 8 bits. If
5118 // Elt1 was also defined, OR the extracted values together before
5119 // inserting them in the result.
5121 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
5122 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
5123 if ((Elt0 & 1) != 0)
5124 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
5126 TLI.getShiftAmountTy(InsElt0.getValueType())));
5128 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
5129 DAG.getConstant(0x00FF, MVT::i16));
5130 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
5133 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
5134 DAG.getIntPtrConstant(i));
5136 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, NewV);
5139 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
5140 /// ones, or rewriting v4i32 / v4f32 as 2 wide ones if possible. This can be
5141 /// done when every pair / quad of shuffle mask elements point to elements in
5142 /// the right sequence. e.g.
5143 /// vector_shuffle X, Y, <2, 3, | 10, 11, | 0, 1, | 14, 15>
5145 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
5146 SelectionDAG &DAG, DebugLoc dl) {
5147 EVT VT = SVOp->getValueType(0);
5148 SDValue V1 = SVOp->getOperand(0);
5149 SDValue V2 = SVOp->getOperand(1);
5150 unsigned NumElems = VT.getVectorNumElements();
5151 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
5153 switch (VT.getSimpleVT().SimpleTy) {
5154 default: assert(false && "Unexpected!");
5155 case MVT::v4f32: NewVT = MVT::v2f64; break;
5156 case MVT::v4i32: NewVT = MVT::v2i64; break;
5157 case MVT::v8i16: NewVT = MVT::v4i32; break;
5158 case MVT::v16i8: NewVT = MVT::v4i32; break;
5161 int Scale = NumElems / NewWidth;
5162 SmallVector<int, 8> MaskVec;
5163 for (unsigned i = 0; i < NumElems; i += Scale) {
5165 for (int j = 0; j < Scale; ++j) {
5166 int EltIdx = SVOp->getMaskElt(i+j);
5170 StartIdx = EltIdx - (EltIdx % Scale);
5171 if (EltIdx != StartIdx + j)
5175 MaskVec.push_back(-1);
5177 MaskVec.push_back(StartIdx / Scale);
5180 V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1);
5181 V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2);
5182 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
5185 /// getVZextMovL - Return a zero-extending vector move low node.
5187 static SDValue getVZextMovL(EVT VT, EVT OpVT,
5188 SDValue SrcOp, SelectionDAG &DAG,
5189 const X86Subtarget *Subtarget, DebugLoc dl) {
5190 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
5191 LoadSDNode *LD = NULL;
5192 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
5193 LD = dyn_cast<LoadSDNode>(SrcOp);
5195 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
5197 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
5198 if ((ExtVT != MVT::i64 || Subtarget->is64Bit()) &&
5199 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
5200 SrcOp.getOperand(0).getOpcode() == ISD::BITCAST &&
5201 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
5203 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
5204 return DAG.getNode(ISD::BITCAST, dl, VT,
5205 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
5206 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5214 return DAG.getNode(ISD::BITCAST, dl, VT,
5215 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
5216 DAG.getNode(ISD::BITCAST, dl,
5220 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
5223 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
5224 SDValue V1 = SVOp->getOperand(0);
5225 SDValue V2 = SVOp->getOperand(1);
5226 DebugLoc dl = SVOp->getDebugLoc();
5227 EVT VT = SVOp->getValueType(0);
5229 SmallVector<std::pair<int, int>, 8> Locs;
5231 SmallVector<int, 8> Mask1(4U, -1);
5232 SmallVector<int, 8> PermMask;
5233 SVOp->getMask(PermMask);
5237 for (unsigned i = 0; i != 4; ++i) {
5238 int Idx = PermMask[i];
5240 Locs[i] = std::make_pair(-1, -1);
5242 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
5244 Locs[i] = std::make_pair(0, NumLo);
5248 Locs[i] = std::make_pair(1, NumHi);
5250 Mask1[2+NumHi] = Idx;
5256 if (NumLo <= 2 && NumHi <= 2) {
5257 // If no more than two elements come from either vector. This can be
5258 // implemented with two shuffles. First shuffle gather the elements.
5259 // The second shuffle, which takes the first shuffle as both of its
5260 // vector operands, put the elements into the right order.
5261 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5263 SmallVector<int, 8> Mask2(4U, -1);
5265 for (unsigned i = 0; i != 4; ++i) {
5266 if (Locs[i].first == -1)
5269 unsigned Idx = (i < 2) ? 0 : 4;
5270 Idx += Locs[i].first * 2 + Locs[i].second;
5275 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
5276 } else if (NumLo == 3 || NumHi == 3) {
5277 // Otherwise, we must have three elements from one vector, call it X, and
5278 // one element from the other, call it Y. First, use a shufps to build an
5279 // intermediate vector with the one element from Y and the element from X
5280 // that will be in the same half in the final destination (the indexes don't
5281 // matter). Then, use a shufps to build the final vector, taking the half
5282 // containing the element from Y from the intermediate, and the other half
5285 // Normalize it so the 3 elements come from V1.
5286 CommuteVectorShuffleMask(PermMask, VT);
5290 // Find the element from V2.
5292 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
5293 int Val = PermMask[HiIndex];
5300 Mask1[0] = PermMask[HiIndex];
5302 Mask1[2] = PermMask[HiIndex^1];
5304 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5307 Mask1[0] = PermMask[0];
5308 Mask1[1] = PermMask[1];
5309 Mask1[2] = HiIndex & 1 ? 6 : 4;
5310 Mask1[3] = HiIndex & 1 ? 4 : 6;
5311 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5313 Mask1[0] = HiIndex & 1 ? 2 : 0;
5314 Mask1[1] = HiIndex & 1 ? 0 : 2;
5315 Mask1[2] = PermMask[2];
5316 Mask1[3] = PermMask[3];
5321 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
5325 // Break it into (shuffle shuffle_hi, shuffle_lo).
5328 SmallVector<int,8> LoMask(4U, -1);
5329 SmallVector<int,8> HiMask(4U, -1);
5331 SmallVector<int,8> *MaskPtr = &LoMask;
5332 unsigned MaskIdx = 0;
5335 for (unsigned i = 0; i != 4; ++i) {
5342 int Idx = PermMask[i];
5344 Locs[i] = std::make_pair(-1, -1);
5345 } else if (Idx < 4) {
5346 Locs[i] = std::make_pair(MaskIdx, LoIdx);
5347 (*MaskPtr)[LoIdx] = Idx;
5350 Locs[i] = std::make_pair(MaskIdx, HiIdx);
5351 (*MaskPtr)[HiIdx] = Idx;
5356 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
5357 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
5358 SmallVector<int, 8> MaskOps;
5359 for (unsigned i = 0; i != 4; ++i) {
5360 if (Locs[i].first == -1) {
5361 MaskOps.push_back(-1);
5363 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
5364 MaskOps.push_back(Idx);
5367 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
5370 static bool MayFoldVectorLoad(SDValue V) {
5371 if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
5372 V = V.getOperand(0);
5373 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5374 V = V.getOperand(0);
5380 // FIXME: the version above should always be used. Since there's
5381 // a bug where several vector shuffles can't be folded because the
5382 // DAG is not updated during lowering and a node claims to have two
5383 // uses while it only has one, use this version, and let isel match
5384 // another instruction if the load really happens to have more than
5385 // one use. Remove this version after this bug get fixed.
5386 // rdar://8434668, PR8156
5387 static bool RelaxedMayFoldVectorLoad(SDValue V) {
5388 if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
5389 V = V.getOperand(0);
5390 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5391 V = V.getOperand(0);
5392 if (ISD::isNormalLoad(V.getNode()))
5397 /// CanFoldShuffleIntoVExtract - Check if the current shuffle is used by
5398 /// a vector extract, and if both can be later optimized into a single load.
5399 /// This is done in visitEXTRACT_VECTOR_ELT and the conditions are checked
5400 /// here because otherwise a target specific shuffle node is going to be
5401 /// emitted for this shuffle, and the optimization not done.
5402 /// FIXME: This is probably not the best approach, but fix the problem
5403 /// until the right path is decided.
5405 bool CanXFormVExtractWithShuffleIntoLoad(SDValue V, SelectionDAG &DAG,
5406 const TargetLowering &TLI) {
5407 EVT VT = V.getValueType();
5408 ShuffleVectorSDNode *SVOp = dyn_cast<ShuffleVectorSDNode>(V);
5410 // Be sure that the vector shuffle is present in a pattern like this:
5411 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), c) -> (f32 load $addr)
5415 SDNode *N = *V.getNode()->use_begin();
5416 if (N->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
5419 SDValue EltNo = N->getOperand(1);
5420 if (!isa<ConstantSDNode>(EltNo))
5423 // If the bit convert changed the number of elements, it is unsafe
5424 // to examine the mask.
5425 bool HasShuffleIntoBitcast = false;
5426 if (V.getOpcode() == ISD::BITCAST) {
5427 EVT SrcVT = V.getOperand(0).getValueType();
5428 if (SrcVT.getVectorNumElements() != VT.getVectorNumElements())
5430 V = V.getOperand(0);
5431 HasShuffleIntoBitcast = true;
5434 // Select the input vector, guarding against out of range extract vector.
5435 unsigned NumElems = VT.getVectorNumElements();
5436 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
5437 int Idx = (Elt > NumElems) ? -1 : SVOp->getMaskElt(Elt);
5438 V = (Idx < (int)NumElems) ? V.getOperand(0) : V.getOperand(1);
5440 // Skip one more bit_convert if necessary
5441 if (V.getOpcode() == ISD::BITCAST)
5442 V = V.getOperand(0);
5444 if (ISD::isNormalLoad(V.getNode())) {
5445 // Is the original load suitable?
5446 LoadSDNode *LN0 = cast<LoadSDNode>(V);
5448 // FIXME: avoid the multi-use bug that is preventing lots of
5449 // of foldings to be detected, this is still wrong of course, but
5450 // give the temporary desired behavior, and if it happens that
5451 // the load has real more uses, during isel it will not fold, and
5452 // will generate poor code.
5453 if (!LN0 || LN0->isVolatile()) // || !LN0->hasOneUse()
5456 if (!HasShuffleIntoBitcast)
5459 // If there's a bitcast before the shuffle, check if the load type and
5460 // alignment is valid.
5461 unsigned Align = LN0->getAlignment();
5463 TLI.getTargetData()->getABITypeAlignment(
5464 VT.getTypeForEVT(*DAG.getContext()));
5466 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VT))
5474 SDValue getMOVDDup(SDValue &Op, DebugLoc &dl, SDValue V1, SelectionDAG &DAG) {
5475 EVT VT = Op.getValueType();
5477 // Canonizalize to v2f64.
5478 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, V1);
5479 return DAG.getNode(ISD::BITCAST, dl, VT,
5480 getTargetShuffleNode(X86ISD::MOVDDUP, dl, MVT::v2f64,
5485 SDValue getMOVLowToHigh(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG,
5487 SDValue V1 = Op.getOperand(0);
5488 SDValue V2 = Op.getOperand(1);
5489 EVT VT = Op.getValueType();
5491 assert(VT != MVT::v2i64 && "unsupported shuffle type");
5493 if (HasSSE2 && VT == MVT::v2f64)
5494 return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG);
5497 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V2, DAG);
5501 SDValue getMOVHighToLow(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG) {
5502 SDValue V1 = Op.getOperand(0);
5503 SDValue V2 = Op.getOperand(1);
5504 EVT VT = Op.getValueType();
5506 assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&
5507 "unsupported shuffle type");
5509 if (V2.getOpcode() == ISD::UNDEF)
5513 return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG);
5517 SDValue getMOVLP(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
5518 SDValue V1 = Op.getOperand(0);
5519 SDValue V2 = Op.getOperand(1);
5520 EVT VT = Op.getValueType();
5521 unsigned NumElems = VT.getVectorNumElements();
5523 // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second
5524 // operand of these instructions is only memory, so check if there's a
5525 // potencial load folding here, otherwise use SHUFPS or MOVSD to match the
5527 bool CanFoldLoad = false;
5529 // Trivial case, when V2 comes from a load.
5530 if (MayFoldVectorLoad(V2))
5533 // When V1 is a load, it can be folded later into a store in isel, example:
5534 // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1)
5536 // (MOVLPSmr addr:$src1, VR128:$src2)
5537 // So, recognize this potential and also use MOVLPS or MOVLPD
5538 if (MayFoldVectorLoad(V1) && MayFoldIntoStore(Op))
5541 // Both of them can't be memory operations though.
5542 if (MayFoldVectorLoad(V1) && MayFoldVectorLoad(V2))
5543 CanFoldLoad = false;
5546 if (HasSSE2 && NumElems == 2)
5547 return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
5550 return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
5553 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5554 // movl and movlp will both match v2i64, but v2i64 is never matched by
5555 // movl earlier because we make it strict to avoid messing with the movlp load
5556 // folding logic (see the code above getMOVLP call). Match it here then,
5557 // this is horrible, but will stay like this until we move all shuffle
5558 // matching to x86 specific nodes. Note that for the 1st condition all
5559 // types are matched with movsd.
5560 if ((HasSSE2 && NumElems == 2) || !X86::isMOVLMask(SVOp))
5561 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5563 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5566 assert(VT != MVT::v4i32 && "unsupported shuffle type");
5568 // Invert the operand order and use SHUFPS to match it.
5569 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V2, V1,
5570 X86::getShuffleSHUFImmediate(SVOp), DAG);
5573 static inline unsigned getUNPCKLOpcode(EVT VT, const X86Subtarget *Subtarget) {
5574 switch(VT.getSimpleVT().SimpleTy) {
5575 case MVT::v4i32: return X86ISD::PUNPCKLDQ;
5576 case MVT::v2i64: return X86ISD::PUNPCKLQDQ;
5578 return Subtarget->hasAVX() ? X86ISD::VUNPCKLPS : X86ISD::UNPCKLPS;
5580 return Subtarget->hasAVX() ? X86ISD::VUNPCKLPD : X86ISD::UNPCKLPD;
5581 case MVT::v8f32: return X86ISD::VUNPCKLPSY;
5582 case MVT::v4f64: return X86ISD::VUNPCKLPDY;
5583 case MVT::v16i8: return X86ISD::PUNPCKLBW;
5584 case MVT::v8i16: return X86ISD::PUNPCKLWD;
5586 llvm_unreachable("Unknown type for unpckl");
5591 static inline unsigned getUNPCKHOpcode(EVT VT) {
5592 switch(VT.getSimpleVT().SimpleTy) {
5593 case MVT::v4i32: return X86ISD::PUNPCKHDQ;
5594 case MVT::v2i64: return X86ISD::PUNPCKHQDQ;
5595 case MVT::v4f32: return X86ISD::UNPCKHPS;
5596 case MVT::v2f64: return X86ISD::UNPCKHPD;
5597 case MVT::v16i8: return X86ISD::PUNPCKHBW;
5598 case MVT::v8i16: return X86ISD::PUNPCKHWD;
5600 llvm_unreachable("Unknown type for unpckh");
5606 SDValue NormalizeVectorShuffle(SDValue Op, SelectionDAG &DAG,
5607 const TargetLowering &TLI,
5608 const X86Subtarget *Subtarget) {
5609 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5610 EVT VT = Op.getValueType();
5611 DebugLoc dl = Op.getDebugLoc();
5612 SDValue V1 = Op.getOperand(0);
5613 SDValue V2 = Op.getOperand(1);
5615 if (isZeroShuffle(SVOp))
5616 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
5618 // Handle splat operations
5619 if (SVOp->isSplat()) {
5620 // Special case, this is the only place now where it's
5621 // allowed to return a vector_shuffle operation without
5622 // using a target specific node, because *hopefully* it
5623 // will be optimized away by the dag combiner.
5624 if (VT.getVectorNumElements() <= 4 &&
5625 CanXFormVExtractWithShuffleIntoLoad(Op, DAG, TLI))
5628 // Handle splats by matching through known masks
5629 if (VT.getVectorNumElements() <= 4)
5632 // Canonicalize all of the remaining to v4f32.
5633 return PromoteSplat(SVOp, DAG);
5636 // If the shuffle can be profitably rewritten as a narrower shuffle, then
5638 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
5639 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
5640 if (NewOp.getNode())
5641 return DAG.getNode(ISD::BITCAST, dl, VT, NewOp);
5642 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
5643 // FIXME: Figure out a cleaner way to do this.
5644 // Try to make use of movq to zero out the top part.
5645 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
5646 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
5647 if (NewOp.getNode()) {
5648 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
5649 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
5650 DAG, Subtarget, dl);
5652 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
5653 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
5654 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
5655 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
5656 DAG, Subtarget, dl);
5663 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
5664 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5665 SDValue V1 = Op.getOperand(0);
5666 SDValue V2 = Op.getOperand(1);
5667 EVT VT = Op.getValueType();
5668 DebugLoc dl = Op.getDebugLoc();
5669 unsigned NumElems = VT.getVectorNumElements();
5670 bool isMMX = VT.getSizeInBits() == 64;
5671 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
5672 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
5673 bool V1IsSplat = false;
5674 bool V2IsSplat = false;
5675 bool HasSSE2 = Subtarget->hasSSE2() || Subtarget->hasAVX();
5676 bool HasSSE3 = Subtarget->hasSSE3() || Subtarget->hasAVX();
5677 bool HasSSSE3 = Subtarget->hasSSSE3() || Subtarget->hasAVX();
5678 MachineFunction &MF = DAG.getMachineFunction();
5679 bool OptForSize = MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize);
5681 // Shuffle operations on MMX not supported.
5685 // Vector shuffle lowering takes 3 steps:
5687 // 1) Normalize the input vectors. Here splats, zeroed vectors, profitable
5688 // narrowing and commutation of operands should be handled.
5689 // 2) Matching of shuffles with known shuffle masks to x86 target specific
5691 // 3) Rewriting of unmatched masks into new generic shuffle operations,
5692 // so the shuffle can be broken into other shuffles and the legalizer can
5693 // try the lowering again.
5695 // The general ideia is that no vector_shuffle operation should be left to
5696 // be matched during isel, all of them must be converted to a target specific
5699 // Normalize the input vectors. Here splats, zeroed vectors, profitable
5700 // narrowing and commutation of operands should be handled. The actual code
5701 // doesn't include all of those, work in progress...
5702 SDValue NewOp = NormalizeVectorShuffle(Op, DAG, *this, Subtarget);
5703 if (NewOp.getNode())
5706 // NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
5707 // unpckh_undef). Only use pshufd if speed is more important than size.
5708 if (OptForSize && X86::isUNPCKL_v_undef_Mask(SVOp))
5709 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5710 return getTargetShuffleNode(getUNPCKLOpcode(VT, getSubtarget()), dl, VT, V1, V1, DAG);
5711 if (OptForSize && X86::isUNPCKH_v_undef_Mask(SVOp))
5712 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5713 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5715 if (X86::isMOVDDUPMask(SVOp) && HasSSE3 && V2IsUndef &&
5716 RelaxedMayFoldVectorLoad(V1))
5717 return getMOVDDup(Op, dl, V1, DAG);
5719 if (X86::isMOVHLPS_v_undef_Mask(SVOp))
5720 return getMOVHighToLow(Op, dl, DAG);
5722 // Use to match splats
5723 if (HasSSE2 && X86::isUNPCKHMask(SVOp) && V2IsUndef &&
5724 (VT == MVT::v2f64 || VT == MVT::v2i64))
5725 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5727 if (X86::isPSHUFDMask(SVOp)) {
5728 // The actual implementation will match the mask in the if above and then
5729 // during isel it can match several different instructions, not only pshufd
5730 // as its name says, sad but true, emulate the behavior for now...
5731 if (X86::isMOVDDUPMask(SVOp) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
5732 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
5734 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
5736 if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
5737 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
5739 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
5740 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V1,
5743 if (VT == MVT::v4f32)
5744 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V1,
5748 // Check if this can be converted into a logical shift.
5749 bool isLeft = false;
5752 bool isShift = getSubtarget()->hasSSE2() &&
5753 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
5754 if (isShift && ShVal.hasOneUse()) {
5755 // If the shifted value has multiple uses, it may be cheaper to use
5756 // v_set0 + movlhps or movhlps, etc.
5757 EVT EltVT = VT.getVectorElementType();
5758 ShAmt *= EltVT.getSizeInBits();
5759 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
5762 if (X86::isMOVLMask(SVOp)) {
5765 if (ISD::isBuildVectorAllZeros(V1.getNode()))
5766 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
5767 if (!X86::isMOVLPMask(SVOp)) {
5768 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
5769 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5771 if (VT == MVT::v4i32 || VT == MVT::v4f32)
5772 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5776 // FIXME: fold these into legal mask.
5777 if (X86::isMOVLHPSMask(SVOp) && !X86::isUNPCKLMask(SVOp))
5778 return getMOVLowToHigh(Op, dl, DAG, HasSSE2);
5780 if (X86::isMOVHLPSMask(SVOp))
5781 return getMOVHighToLow(Op, dl, DAG);
5783 if (X86::isMOVSHDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
5784 return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
5786 if (X86::isMOVSLDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
5787 return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
5789 if (X86::isMOVLPMask(SVOp))
5790 return getMOVLP(Op, dl, DAG, HasSSE2);
5792 if (ShouldXformToMOVHLPS(SVOp) ||
5793 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
5794 return CommuteVectorShuffle(SVOp, DAG);
5797 // No better options. Use a vshl / vsrl.
5798 EVT EltVT = VT.getVectorElementType();
5799 ShAmt *= EltVT.getSizeInBits();
5800 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
5803 bool Commuted = false;
5804 // FIXME: This should also accept a bitcast of a splat? Be careful, not
5805 // 1,1,1,1 -> v8i16 though.
5806 V1IsSplat = isSplatVector(V1.getNode());
5807 V2IsSplat = isSplatVector(V2.getNode());
5809 // Canonicalize the splat or undef, if present, to be on the RHS.
5810 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
5811 Op = CommuteVectorShuffle(SVOp, DAG);
5812 SVOp = cast<ShuffleVectorSDNode>(Op);
5813 V1 = SVOp->getOperand(0);
5814 V2 = SVOp->getOperand(1);
5815 std::swap(V1IsSplat, V2IsSplat);
5816 std::swap(V1IsUndef, V2IsUndef);
5820 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
5821 // Shuffling low element of v1 into undef, just return v1.
5824 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
5825 // the instruction selector will not match, so get a canonical MOVL with
5826 // swapped operands to undo the commute.
5827 return getMOVL(DAG, dl, VT, V2, V1);
5830 if (X86::isUNPCKLMask(SVOp))
5831 return getTargetShuffleNode(getUNPCKLOpcode(VT, getSubtarget()),
5832 dl, VT, V1, V2, DAG);
5834 if (X86::isUNPCKHMask(SVOp))
5835 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V2, DAG);
5838 // Normalize mask so all entries that point to V2 points to its first
5839 // element then try to match unpck{h|l} again. If match, return a
5840 // new vector_shuffle with the corrected mask.
5841 SDValue NewMask = NormalizeMask(SVOp, DAG);
5842 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
5843 if (NSVOp != SVOp) {
5844 if (X86::isUNPCKLMask(NSVOp, true)) {
5846 } else if (X86::isUNPCKHMask(NSVOp, true)) {
5853 // Commute is back and try unpck* again.
5854 // FIXME: this seems wrong.
5855 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
5856 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
5858 if (X86::isUNPCKLMask(NewSVOp))
5859 return getTargetShuffleNode(getUNPCKLOpcode(VT, getSubtarget()),
5860 dl, VT, V2, V1, DAG);
5862 if (X86::isUNPCKHMask(NewSVOp))
5863 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V2, V1, DAG);
5866 // Normalize the node to match x86 shuffle ops if needed
5867 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
5868 return CommuteVectorShuffle(SVOp, DAG);
5870 // The checks below are all present in isShuffleMaskLegal, but they are
5871 // inlined here right now to enable us to directly emit target specific
5872 // nodes, and remove one by one until they don't return Op anymore.
5873 SmallVector<int, 16> M;
5876 if (isPALIGNRMask(M, VT, HasSSSE3))
5877 return getTargetShuffleNode(X86ISD::PALIGN, dl, VT, V1, V2,
5878 X86::getShufflePALIGNRImmediate(SVOp),
5881 if (ShuffleVectorSDNode::isSplatMask(&M[0], VT) &&
5882 SVOp->getSplatIndex() == 0 && V2IsUndef) {
5883 if (VT == MVT::v2f64) {
5884 X86ISD::NodeType Opcode =
5885 getSubtarget()->hasAVX() ? X86ISD::VUNPCKLPD : X86ISD::UNPCKLPD;
5886 return getTargetShuffleNode(Opcode, dl, VT, V1, V1, DAG);
5888 if (VT == MVT::v2i64)
5889 return getTargetShuffleNode(X86ISD::PUNPCKLQDQ, dl, VT, V1, V1, DAG);
5892 if (isPSHUFHWMask(M, VT))
5893 return getTargetShuffleNode(X86ISD::PSHUFHW, dl, VT, V1,
5894 X86::getShufflePSHUFHWImmediate(SVOp),
5897 if (isPSHUFLWMask(M, VT))
5898 return getTargetShuffleNode(X86ISD::PSHUFLW, dl, VT, V1,
5899 X86::getShufflePSHUFLWImmediate(SVOp),
5902 if (isSHUFPMask(M, VT)) {
5903 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
5904 if (VT == MVT::v4f32 || VT == MVT::v4i32)
5905 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V2,
5907 if (VT == MVT::v2f64 || VT == MVT::v2i64)
5908 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V2,
5912 if (X86::isUNPCKL_v_undef_Mask(SVOp))
5913 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5914 return getTargetShuffleNode(getUNPCKLOpcode(VT, getSubtarget()),
5915 dl, VT, V1, V1, DAG);
5916 if (X86::isUNPCKH_v_undef_Mask(SVOp))
5917 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5918 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5920 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
5921 if (VT == MVT::v8i16) {
5922 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, DAG);
5923 if (NewOp.getNode())
5927 if (VT == MVT::v16i8) {
5928 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
5929 if (NewOp.getNode())
5933 // Handle all 4 wide cases with a number of shuffles.
5935 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
5941 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
5942 SelectionDAG &DAG) const {
5943 EVT VT = Op.getValueType();
5944 DebugLoc dl = Op.getDebugLoc();
5945 if (VT.getSizeInBits() == 8) {
5946 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
5947 Op.getOperand(0), Op.getOperand(1));
5948 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
5949 DAG.getValueType(VT));
5950 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5951 } else if (VT.getSizeInBits() == 16) {
5952 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5953 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
5955 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
5956 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5957 DAG.getNode(ISD::BITCAST, dl,
5961 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
5962 Op.getOperand(0), Op.getOperand(1));
5963 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
5964 DAG.getValueType(VT));
5965 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5966 } else if (VT == MVT::f32) {
5967 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
5968 // the result back to FR32 register. It's only worth matching if the
5969 // result has a single use which is a store or a bitcast to i32. And in
5970 // the case of a store, it's not worth it if the index is a constant 0,
5971 // because a MOVSSmr can be used instead, which is smaller and faster.
5972 if (!Op.hasOneUse())
5974 SDNode *User = *Op.getNode()->use_begin();
5975 if ((User->getOpcode() != ISD::STORE ||
5976 (isa<ConstantSDNode>(Op.getOperand(1)) &&
5977 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
5978 (User->getOpcode() != ISD::BITCAST ||
5979 User->getValueType(0) != MVT::i32))
5981 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5982 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32,
5985 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract);
5986 } else if (VT == MVT::i32) {
5987 // ExtractPS works with constant index.
5988 if (isa<ConstantSDNode>(Op.getOperand(1)))
5996 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
5997 SelectionDAG &DAG) const {
5998 if (!isa<ConstantSDNode>(Op.getOperand(1)))
6001 SDValue Vec = Op.getOperand(0);
6002 EVT VecVT = Vec.getValueType();
6004 // If this is a 256-bit vector result, first extract the 128-bit
6005 // vector and then extract from the 128-bit vector.
6006 if (VecVT.getSizeInBits() > 128) {
6007 DebugLoc dl = Op.getNode()->getDebugLoc();
6008 unsigned NumElems = VecVT.getVectorNumElements();
6009 SDValue Idx = Op.getOperand(1);
6011 if (!isa<ConstantSDNode>(Idx))
6014 unsigned ExtractNumElems = NumElems / (VecVT.getSizeInBits() / 128);
6015 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
6017 // Get the 128-bit vector.
6018 bool Upper = IdxVal >= ExtractNumElems;
6019 Vec = Extract128BitVector(Vec, Idx, DAG, dl);
6022 SDValue ScaledIdx = Idx;
6024 ScaledIdx = DAG.getNode(ISD::SUB, dl, Idx.getValueType(), Idx,
6025 DAG.getConstant(ExtractNumElems,
6026 Idx.getValueType()));
6027 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
6031 assert(Vec.getValueSizeInBits() <= 128 && "Unexpected vector length");
6033 if (Subtarget->hasSSE41()) {
6034 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
6039 EVT VT = Op.getValueType();
6040 DebugLoc dl = Op.getDebugLoc();
6041 // TODO: handle v16i8.
6042 if (VT.getSizeInBits() == 16) {
6043 SDValue Vec = Op.getOperand(0);
6044 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6046 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
6047 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
6048 DAG.getNode(ISD::BITCAST, dl,
6051 // Transform it so it match pextrw which produces a 32-bit result.
6052 EVT EltVT = MVT::i32;
6053 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
6054 Op.getOperand(0), Op.getOperand(1));
6055 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
6056 DAG.getValueType(VT));
6057 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
6058 } else if (VT.getSizeInBits() == 32) {
6059 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6063 // SHUFPS the element to the lowest double word, then movss.
6064 int Mask[4] = { Idx, -1, -1, -1 };
6065 EVT VVT = Op.getOperand(0).getValueType();
6066 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
6067 DAG.getUNDEF(VVT), Mask);
6068 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
6069 DAG.getIntPtrConstant(0));
6070 } else if (VT.getSizeInBits() == 64) {
6071 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
6072 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
6073 // to match extract_elt for f64.
6074 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6078 // UNPCKHPD the element to the lowest double word, then movsd.
6079 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
6080 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
6081 int Mask[2] = { 1, -1 };
6082 EVT VVT = Op.getOperand(0).getValueType();
6083 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
6084 DAG.getUNDEF(VVT), Mask);
6085 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
6086 DAG.getIntPtrConstant(0));
6093 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
6094 SelectionDAG &DAG) const {
6095 EVT VT = Op.getValueType();
6096 EVT EltVT = VT.getVectorElementType();
6097 DebugLoc dl = Op.getDebugLoc();
6099 SDValue N0 = Op.getOperand(0);
6100 SDValue N1 = Op.getOperand(1);
6101 SDValue N2 = Op.getOperand(2);
6103 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
6104 isa<ConstantSDNode>(N2)) {
6106 if (VT == MVT::v8i16)
6107 Opc = X86ISD::PINSRW;
6108 else if (VT == MVT::v16i8)
6109 Opc = X86ISD::PINSRB;
6111 Opc = X86ISD::PINSRB;
6113 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
6115 if (N1.getValueType() != MVT::i32)
6116 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
6117 if (N2.getValueType() != MVT::i32)
6118 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
6119 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
6120 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
6121 // Bits [7:6] of the constant are the source select. This will always be
6122 // zero here. The DAG Combiner may combine an extract_elt index into these
6123 // bits. For example (insert (extract, 3), 2) could be matched by putting
6124 // the '3' into bits [7:6] of X86ISD::INSERTPS.
6125 // Bits [5:4] of the constant are the destination select. This is the
6126 // value of the incoming immediate.
6127 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
6128 // combine either bitwise AND or insert of float 0.0 to set these bits.
6129 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
6130 // Create this as a scalar to vector..
6131 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
6132 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
6133 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
6134 // PINSR* works with constant index.
6141 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
6142 EVT VT = Op.getValueType();
6143 EVT EltVT = VT.getVectorElementType();
6145 DebugLoc dl = Op.getDebugLoc();
6146 SDValue N0 = Op.getOperand(0);
6147 SDValue N1 = Op.getOperand(1);
6148 SDValue N2 = Op.getOperand(2);
6150 // If this is a 256-bit vector result, first insert into a 128-bit
6151 // vector and then insert into the 256-bit vector.
6152 if (VT.getSizeInBits() > 128) {
6153 if (!isa<ConstantSDNode>(N2))
6156 // Get the 128-bit vector.
6157 unsigned NumElems = VT.getVectorNumElements();
6158 unsigned IdxVal = cast<ConstantSDNode>(N2)->getZExtValue();
6159 bool Upper = IdxVal >= NumElems / 2;
6161 SDValue SubN0 = Extract128BitVector(N0, N2, DAG, dl);
6164 SDValue ScaledN2 = N2;
6166 ScaledN2 = DAG.getNode(ISD::SUB, dl, N2.getValueType(), N2,
6167 DAG.getConstant(NumElems /
6168 (VT.getSizeInBits() / 128),
6169 N2.getValueType()));
6170 Op = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, SubN0.getValueType(), SubN0,
6173 // Insert the 128-bit vector
6174 // FIXME: Why UNDEF?
6175 return Insert128BitVector(N0, Op, N2, DAG, dl);
6178 if (Subtarget->hasSSE41())
6179 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
6181 if (EltVT == MVT::i8)
6184 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
6185 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
6186 // as its second argument.
6187 if (N1.getValueType() != MVT::i32)
6188 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
6189 if (N2.getValueType() != MVT::i32)
6190 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
6191 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
6197 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
6198 LLVMContext *Context = DAG.getContext();
6199 DebugLoc dl = Op.getDebugLoc();
6200 EVT OpVT = Op.getValueType();
6202 // If this is a 256-bit vector result, first insert into a 128-bit
6203 // vector and then insert into the 256-bit vector.
6204 if (OpVT.getSizeInBits() > 128) {
6205 // Insert into a 128-bit vector.
6206 EVT VT128 = EVT::getVectorVT(*Context,
6207 OpVT.getVectorElementType(),
6208 OpVT.getVectorNumElements() / 2);
6210 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
6212 // Insert the 128-bit vector.
6213 return Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, OpVT), Op,
6214 DAG.getConstant(0, MVT::i32),
6218 if (Op.getValueType() == MVT::v1i64 &&
6219 Op.getOperand(0).getValueType() == MVT::i64)
6220 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
6222 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
6223 assert(Op.getValueType().getSimpleVT().getSizeInBits() == 128 &&
6224 "Expected an SSE type!");
6225 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(),
6226 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
6229 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
6230 // a simple subregister reference or explicit instructions to grab
6231 // upper bits of a vector.
6233 X86TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const {
6234 if (Subtarget->hasAVX()) {
6235 DebugLoc dl = Op.getNode()->getDebugLoc();
6236 SDValue Vec = Op.getNode()->getOperand(0);
6237 SDValue Idx = Op.getNode()->getOperand(1);
6239 if (Op.getNode()->getValueType(0).getSizeInBits() == 128
6240 && Vec.getNode()->getValueType(0).getSizeInBits() == 256) {
6241 return Extract128BitVector(Vec, Idx, DAG, dl);
6247 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
6248 // simple superregister reference or explicit instructions to insert
6249 // the upper bits of a vector.
6251 X86TargetLowering::LowerINSERT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const {
6252 if (Subtarget->hasAVX()) {
6253 DebugLoc dl = Op.getNode()->getDebugLoc();
6254 SDValue Vec = Op.getNode()->getOperand(0);
6255 SDValue SubVec = Op.getNode()->getOperand(1);
6256 SDValue Idx = Op.getNode()->getOperand(2);
6258 if (Op.getNode()->getValueType(0).getSizeInBits() == 256
6259 && SubVec.getNode()->getValueType(0).getSizeInBits() == 128) {
6260 return Insert128BitVector(Vec, SubVec, Idx, DAG, dl);
6266 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
6267 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
6268 // one of the above mentioned nodes. It has to be wrapped because otherwise
6269 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
6270 // be used to form addressing mode. These wrapped nodes will be selected
6273 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
6274 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
6276 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6278 unsigned char OpFlag = 0;
6279 unsigned WrapperKind = X86ISD::Wrapper;
6280 CodeModel::Model M = getTargetMachine().getCodeModel();
6282 if (Subtarget->isPICStyleRIPRel() &&
6283 (M == CodeModel::Small || M == CodeModel::Kernel))
6284 WrapperKind = X86ISD::WrapperRIP;
6285 else if (Subtarget->isPICStyleGOT())
6286 OpFlag = X86II::MO_GOTOFF;
6287 else if (Subtarget->isPICStyleStubPIC())
6288 OpFlag = X86II::MO_PIC_BASE_OFFSET;
6290 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
6292 CP->getOffset(), OpFlag);
6293 DebugLoc DL = CP->getDebugLoc();
6294 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6295 // With PIC, the address is actually $g + Offset.
6297 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6298 DAG.getNode(X86ISD::GlobalBaseReg,
6299 DebugLoc(), getPointerTy()),
6306 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
6307 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
6309 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6311 unsigned char OpFlag = 0;
6312 unsigned WrapperKind = X86ISD::Wrapper;
6313 CodeModel::Model M = getTargetMachine().getCodeModel();
6315 if (Subtarget->isPICStyleRIPRel() &&
6316 (M == CodeModel::Small || M == CodeModel::Kernel))
6317 WrapperKind = X86ISD::WrapperRIP;
6318 else if (Subtarget->isPICStyleGOT())
6319 OpFlag = X86II::MO_GOTOFF;
6320 else if (Subtarget->isPICStyleStubPIC())
6321 OpFlag = X86II::MO_PIC_BASE_OFFSET;
6323 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
6325 DebugLoc DL = JT->getDebugLoc();
6326 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6328 // With PIC, the address is actually $g + Offset.
6330 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6331 DAG.getNode(X86ISD::GlobalBaseReg,
6332 DebugLoc(), getPointerTy()),
6339 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
6340 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
6342 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6344 unsigned char OpFlag = 0;
6345 unsigned WrapperKind = X86ISD::Wrapper;
6346 CodeModel::Model M = getTargetMachine().getCodeModel();
6348 if (Subtarget->isPICStyleRIPRel() &&
6349 (M == CodeModel::Small || M == CodeModel::Kernel))
6350 WrapperKind = X86ISD::WrapperRIP;
6351 else if (Subtarget->isPICStyleGOT())
6352 OpFlag = X86II::MO_GOTOFF;
6353 else if (Subtarget->isPICStyleStubPIC())
6354 OpFlag = X86II::MO_PIC_BASE_OFFSET;
6356 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
6358 DebugLoc DL = Op.getDebugLoc();
6359 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6362 // With PIC, the address is actually $g + Offset.
6363 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
6364 !Subtarget->is64Bit()) {
6365 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6366 DAG.getNode(X86ISD::GlobalBaseReg,
6367 DebugLoc(), getPointerTy()),
6375 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
6376 // Create the TargetBlockAddressAddress node.
6377 unsigned char OpFlags =
6378 Subtarget->ClassifyBlockAddressReference();
6379 CodeModel::Model M = getTargetMachine().getCodeModel();
6380 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
6381 DebugLoc dl = Op.getDebugLoc();
6382 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
6383 /*isTarget=*/true, OpFlags);
6385 if (Subtarget->isPICStyleRIPRel() &&
6386 (M == CodeModel::Small || M == CodeModel::Kernel))
6387 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
6389 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
6391 // With PIC, the address is actually $g + Offset.
6392 if (isGlobalRelativeToPICBase(OpFlags)) {
6393 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6394 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
6402 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
6404 SelectionDAG &DAG) const {
6405 // Create the TargetGlobalAddress node, folding in the constant
6406 // offset if it is legal.
6407 unsigned char OpFlags =
6408 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
6409 CodeModel::Model M = getTargetMachine().getCodeModel();
6411 if (OpFlags == X86II::MO_NO_FLAG &&
6412 X86::isOffsetSuitableForCodeModel(Offset, M)) {
6413 // A direct static reference to a global.
6414 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
6417 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
6420 if (Subtarget->isPICStyleRIPRel() &&
6421 (M == CodeModel::Small || M == CodeModel::Kernel))
6422 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
6424 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
6426 // With PIC, the address is actually $g + Offset.
6427 if (isGlobalRelativeToPICBase(OpFlags)) {
6428 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6429 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
6433 // For globals that require a load from a stub to get the address, emit the
6435 if (isGlobalStubReference(OpFlags))
6436 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
6437 MachinePointerInfo::getGOT(), false, false, 0);
6439 // If there was a non-zero offset that we didn't fold, create an explicit
6442 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
6443 DAG.getConstant(Offset, getPointerTy()));
6449 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
6450 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
6451 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
6452 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
6456 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
6457 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
6458 unsigned char OperandFlags) {
6459 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6460 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
6461 DebugLoc dl = GA->getDebugLoc();
6462 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
6463 GA->getValueType(0),
6467 SDValue Ops[] = { Chain, TGA, *InFlag };
6468 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
6470 SDValue Ops[] = { Chain, TGA };
6471 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
6474 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
6475 MFI->setAdjustsStack(true);
6477 SDValue Flag = Chain.getValue(1);
6478 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
6481 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
6483 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6486 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
6487 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
6488 DAG.getNode(X86ISD::GlobalBaseReg,
6489 DebugLoc(), PtrVT), InFlag);
6490 InFlag = Chain.getValue(1);
6492 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
6495 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
6497 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6499 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
6500 X86::RAX, X86II::MO_TLSGD);
6503 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
6504 // "local exec" model.
6505 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6506 const EVT PtrVT, TLSModel::Model model,
6508 DebugLoc dl = GA->getDebugLoc();
6510 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
6511 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
6512 is64Bit ? 257 : 256));
6514 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
6515 DAG.getIntPtrConstant(0),
6516 MachinePointerInfo(Ptr), false, false, 0);
6518 unsigned char OperandFlags = 0;
6519 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
6521 unsigned WrapperKind = X86ISD::Wrapper;
6522 if (model == TLSModel::LocalExec) {
6523 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
6524 } else if (is64Bit) {
6525 assert(model == TLSModel::InitialExec);
6526 OperandFlags = X86II::MO_GOTTPOFF;
6527 WrapperKind = X86ISD::WrapperRIP;
6529 assert(model == TLSModel::InitialExec);
6530 OperandFlags = X86II::MO_INDNTPOFF;
6533 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
6535 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
6536 GA->getValueType(0),
6537 GA->getOffset(), OperandFlags);
6538 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
6540 if (model == TLSModel::InitialExec)
6541 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
6542 MachinePointerInfo::getGOT(), false, false, 0);
6544 // The address of the thread local variable is the add of the thread
6545 // pointer with the offset of the variable.
6546 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
6550 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
6552 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
6553 const GlobalValue *GV = GA->getGlobal();
6555 if (Subtarget->isTargetELF()) {
6556 // TODO: implement the "local dynamic" model
6557 // TODO: implement the "initial exec"model for pic executables
6559 // If GV is an alias then use the aliasee for determining
6560 // thread-localness.
6561 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
6562 GV = GA->resolveAliasedGlobal(false);
6564 TLSModel::Model model
6565 = getTLSModel(GV, getTargetMachine().getRelocationModel());
6568 case TLSModel::GeneralDynamic:
6569 case TLSModel::LocalDynamic: // not implemented
6570 if (Subtarget->is64Bit())
6571 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
6572 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
6574 case TLSModel::InitialExec:
6575 case TLSModel::LocalExec:
6576 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
6577 Subtarget->is64Bit());
6579 } else if (Subtarget->isTargetDarwin()) {
6580 // Darwin only has one model of TLS. Lower to that.
6581 unsigned char OpFlag = 0;
6582 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
6583 X86ISD::WrapperRIP : X86ISD::Wrapper;
6585 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6587 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
6588 !Subtarget->is64Bit();
6590 OpFlag = X86II::MO_TLVP_PIC_BASE;
6592 OpFlag = X86II::MO_TLVP;
6593 DebugLoc DL = Op.getDebugLoc();
6594 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
6595 GA->getValueType(0),
6596 GA->getOffset(), OpFlag);
6597 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6599 // With PIC32, the address is actually $g + Offset.
6601 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6602 DAG.getNode(X86ISD::GlobalBaseReg,
6603 DebugLoc(), getPointerTy()),
6606 // Lowering the machine isd will make sure everything is in the right
6608 SDValue Chain = DAG.getEntryNode();
6609 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
6610 SDValue Args[] = { Chain, Offset };
6611 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args, 2);
6613 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
6614 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6615 MFI->setAdjustsStack(true);
6617 // And our return value (tls address) is in the standard call return value
6619 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
6620 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
6624 "TLS not implemented for this target.");
6626 llvm_unreachable("Unreachable");
6631 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values and
6632 /// take a 2 x i32 value to shift plus a shift amount.
6633 SDValue X86TargetLowering::LowerShiftParts(SDValue Op, SelectionDAG &DAG) const {
6634 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
6635 EVT VT = Op.getValueType();
6636 unsigned VTBits = VT.getSizeInBits();
6637 DebugLoc dl = Op.getDebugLoc();
6638 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
6639 SDValue ShOpLo = Op.getOperand(0);
6640 SDValue ShOpHi = Op.getOperand(1);
6641 SDValue ShAmt = Op.getOperand(2);
6642 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
6643 DAG.getConstant(VTBits - 1, MVT::i8))
6644 : DAG.getConstant(0, VT);
6647 if (Op.getOpcode() == ISD::SHL_PARTS) {
6648 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
6649 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
6651 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
6652 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
6655 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
6656 DAG.getConstant(VTBits, MVT::i8));
6657 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6658 AndNode, DAG.getConstant(0, MVT::i8));
6661 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6662 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
6663 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
6665 if (Op.getOpcode() == ISD::SHL_PARTS) {
6666 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
6667 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
6669 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
6670 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
6673 SDValue Ops[2] = { Lo, Hi };
6674 return DAG.getMergeValues(Ops, 2, dl);
6677 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
6678 SelectionDAG &DAG) const {
6679 EVT SrcVT = Op.getOperand(0).getValueType();
6681 if (SrcVT.isVector())
6684 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
6685 "Unknown SINT_TO_FP to lower!");
6687 // These are really Legal; return the operand so the caller accepts it as
6689 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
6691 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
6692 Subtarget->is64Bit()) {
6696 DebugLoc dl = Op.getDebugLoc();
6697 unsigned Size = SrcVT.getSizeInBits()/8;
6698 MachineFunction &MF = DAG.getMachineFunction();
6699 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
6700 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6701 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6703 MachinePointerInfo::getFixedStack(SSFI),
6705 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
6708 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
6710 SelectionDAG &DAG) const {
6712 DebugLoc DL = Op.getDebugLoc();
6714 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
6716 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
6718 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
6720 unsigned ByteSize = SrcVT.getSizeInBits()/8;
6722 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
6723 MachineMemOperand *MMO =
6724 DAG.getMachineFunction()
6725 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6726 MachineMemOperand::MOLoad, ByteSize, ByteSize);
6728 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
6729 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
6731 Tys, Ops, array_lengthof(Ops),
6735 Chain = Result.getValue(1);
6736 SDValue InFlag = Result.getValue(2);
6738 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
6739 // shouldn't be necessary except that RFP cannot be live across
6740 // multiple blocks. When stackifier is fixed, they can be uncoupled.
6741 MachineFunction &MF = DAG.getMachineFunction();
6742 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
6743 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
6744 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6745 Tys = DAG.getVTList(MVT::Other);
6747 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
6749 MachineMemOperand *MMO =
6750 DAG.getMachineFunction()
6751 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6752 MachineMemOperand::MOStore, SSFISize, SSFISize);
6754 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
6755 Ops, array_lengthof(Ops),
6756 Op.getValueType(), MMO);
6757 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
6758 MachinePointerInfo::getFixedStack(SSFI),
6765 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
6766 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
6767 SelectionDAG &DAG) const {
6768 // This algorithm is not obvious. Here it is in C code, more or less:
6770 double uint64_to_double( uint32_t hi, uint32_t lo ) {
6771 static const __m128i exp = { 0x4330000045300000ULL, 0 };
6772 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
6774 // Copy ints to xmm registers.
6775 __m128i xh = _mm_cvtsi32_si128( hi );
6776 __m128i xl = _mm_cvtsi32_si128( lo );
6778 // Combine into low half of a single xmm register.
6779 __m128i x = _mm_unpacklo_epi32( xh, xl );
6783 // Merge in appropriate exponents to give the integer bits the right
6785 x = _mm_unpacklo_epi32( x, exp );
6787 // Subtract away the biases to deal with the IEEE-754 double precision
6789 d = _mm_sub_pd( (__m128d) x, bias );
6791 // All conversions up to here are exact. The correctly rounded result is
6792 // calculated using the current rounding mode using the following
6794 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
6795 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
6796 // store doesn't really need to be here (except
6797 // maybe to zero the other double)
6802 DebugLoc dl = Op.getDebugLoc();
6803 LLVMContext *Context = DAG.getContext();
6805 // Build some magic constants.
6806 std::vector<Constant*> CV0;
6807 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
6808 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
6809 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
6810 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
6811 Constant *C0 = ConstantVector::get(CV0);
6812 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
6814 std::vector<Constant*> CV1;
6816 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
6818 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
6819 Constant *C1 = ConstantVector::get(CV1);
6820 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
6822 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6823 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6825 DAG.getIntPtrConstant(1)));
6826 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6827 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6829 DAG.getIntPtrConstant(0)));
6830 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
6831 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
6832 MachinePointerInfo::getConstantPool(),
6834 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
6835 SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck2);
6836 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
6837 MachinePointerInfo::getConstantPool(),
6839 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
6841 // Add the halves; easiest way is to swap them into another reg first.
6842 int ShufMask[2] = { 1, -1 };
6843 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
6844 DAG.getUNDEF(MVT::v2f64), ShufMask);
6845 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
6846 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
6847 DAG.getIntPtrConstant(0));
6850 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
6851 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
6852 SelectionDAG &DAG) const {
6853 DebugLoc dl = Op.getDebugLoc();
6854 // FP constant to bias correct the final result.
6855 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
6858 // Load the 32-bit value into an XMM register.
6859 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6860 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6862 DAG.getIntPtrConstant(0)));
6864 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
6865 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load),
6866 DAG.getIntPtrConstant(0));
6868 // Or the load with the bias.
6869 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
6870 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
6871 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6873 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
6874 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6875 MVT::v2f64, Bias)));
6876 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
6877 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or),
6878 DAG.getIntPtrConstant(0));
6880 // Subtract the bias.
6881 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
6883 // Handle final rounding.
6884 EVT DestVT = Op.getValueType();
6886 if (DestVT.bitsLT(MVT::f64)) {
6887 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
6888 DAG.getIntPtrConstant(0));
6889 } else if (DestVT.bitsGT(MVT::f64)) {
6890 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
6893 // Handle final rounding.
6897 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
6898 SelectionDAG &DAG) const {
6899 SDValue N0 = Op.getOperand(0);
6900 DebugLoc dl = Op.getDebugLoc();
6902 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
6903 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
6904 // the optimization here.
6905 if (DAG.SignBitIsZero(N0))
6906 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
6908 EVT SrcVT = N0.getValueType();
6909 EVT DstVT = Op.getValueType();
6910 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
6911 return LowerUINT_TO_FP_i64(Op, DAG);
6912 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
6913 return LowerUINT_TO_FP_i32(Op, DAG);
6915 // Make a 64-bit buffer, and use it to build an FILD.
6916 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
6917 if (SrcVT == MVT::i32) {
6918 SDValue WordOff = DAG.getConstant(4, getPointerTy());
6919 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
6920 getPointerTy(), StackSlot, WordOff);
6921 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6922 StackSlot, MachinePointerInfo(),
6924 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
6925 OffsetSlot, MachinePointerInfo(),
6927 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
6931 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
6932 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6933 StackSlot, MachinePointerInfo(),
6935 // For i64 source, we need to add the appropriate power of 2 if the input
6936 // was negative. This is the same as the optimization in
6937 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
6938 // we must be careful to do the computation in x87 extended precision, not
6939 // in SSE. (The generic code can't know it's OK to do this, or how to.)
6940 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
6941 MachineMemOperand *MMO =
6942 DAG.getMachineFunction()
6943 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6944 MachineMemOperand::MOLoad, 8, 8);
6946 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
6947 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
6948 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops, 3,
6951 APInt FF(32, 0x5F800000ULL);
6953 // Check whether the sign bit is set.
6954 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
6955 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
6958 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
6959 SDValue FudgePtr = DAG.getConstantPool(
6960 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
6963 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
6964 SDValue Zero = DAG.getIntPtrConstant(0);
6965 SDValue Four = DAG.getIntPtrConstant(4);
6966 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
6968 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
6970 // Load the value out, extending it from f32 to f80.
6971 // FIXME: Avoid the extend by constructing the right constant pool?
6972 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
6973 FudgePtr, MachinePointerInfo::getConstantPool(),
6974 MVT::f32, false, false, 4);
6975 // Extend everything to 80 bits to force it to be done on x87.
6976 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
6977 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
6980 std::pair<SDValue,SDValue> X86TargetLowering::
6981 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
6982 DebugLoc DL = Op.getDebugLoc();
6984 EVT DstTy = Op.getValueType();
6987 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
6991 assert(DstTy.getSimpleVT() <= MVT::i64 &&
6992 DstTy.getSimpleVT() >= MVT::i16 &&
6993 "Unknown FP_TO_SINT to lower!");
6995 // These are really Legal.
6996 if (DstTy == MVT::i32 &&
6997 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
6998 return std::make_pair(SDValue(), SDValue());
6999 if (Subtarget->is64Bit() &&
7000 DstTy == MVT::i64 &&
7001 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
7002 return std::make_pair(SDValue(), SDValue());
7004 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
7006 MachineFunction &MF = DAG.getMachineFunction();
7007 unsigned MemSize = DstTy.getSizeInBits()/8;
7008 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
7009 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7014 switch (DstTy.getSimpleVT().SimpleTy) {
7015 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
7016 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
7017 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
7018 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
7021 SDValue Chain = DAG.getEntryNode();
7022 SDValue Value = Op.getOperand(0);
7023 EVT TheVT = Op.getOperand(0).getValueType();
7024 if (isScalarFPTypeInSSEReg(TheVT)) {
7025 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
7026 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
7027 MachinePointerInfo::getFixedStack(SSFI),
7029 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
7031 Chain, StackSlot, DAG.getValueType(TheVT)
7034 MachineMemOperand *MMO =
7035 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7036 MachineMemOperand::MOLoad, MemSize, MemSize);
7037 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, 3,
7039 Chain = Value.getValue(1);
7040 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
7041 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7044 MachineMemOperand *MMO =
7045 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7046 MachineMemOperand::MOStore, MemSize, MemSize);
7048 // Build the FP_TO_INT*_IN_MEM
7049 SDValue Ops[] = { Chain, Value, StackSlot };
7050 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
7051 Ops, 3, DstTy, MMO);
7053 return std::make_pair(FIST, StackSlot);
7056 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
7057 SelectionDAG &DAG) const {
7058 if (Op.getValueType().isVector())
7061 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
7062 SDValue FIST = Vals.first, StackSlot = Vals.second;
7063 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
7064 if (FIST.getNode() == 0) return Op;
7067 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
7068 FIST, StackSlot, MachinePointerInfo(), false, false, 0);
7071 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
7072 SelectionDAG &DAG) const {
7073 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
7074 SDValue FIST = Vals.first, StackSlot = Vals.second;
7075 assert(FIST.getNode() && "Unexpected failure");
7078 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
7079 FIST, StackSlot, MachinePointerInfo(), false, false, 0);
7082 SDValue X86TargetLowering::LowerFABS(SDValue Op,
7083 SelectionDAG &DAG) const {
7084 LLVMContext *Context = DAG.getContext();
7085 DebugLoc dl = Op.getDebugLoc();
7086 EVT VT = Op.getValueType();
7089 EltVT = VT.getVectorElementType();
7090 std::vector<Constant*> CV;
7091 if (EltVT == MVT::f64) {
7092 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
7096 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
7102 Constant *C = ConstantVector::get(CV);
7103 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7104 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7105 MachinePointerInfo::getConstantPool(),
7107 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
7110 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
7111 LLVMContext *Context = DAG.getContext();
7112 DebugLoc dl = Op.getDebugLoc();
7113 EVT VT = Op.getValueType();
7116 EltVT = VT.getVectorElementType();
7117 std::vector<Constant*> CV;
7118 if (EltVT == MVT::f64) {
7119 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
7123 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
7129 Constant *C = ConstantVector::get(CV);
7130 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7131 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7132 MachinePointerInfo::getConstantPool(),
7134 if (VT.isVector()) {
7135 return DAG.getNode(ISD::BITCAST, dl, VT,
7136 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
7137 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
7139 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, Mask)));
7141 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
7145 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
7146 LLVMContext *Context = DAG.getContext();
7147 SDValue Op0 = Op.getOperand(0);
7148 SDValue Op1 = Op.getOperand(1);
7149 DebugLoc dl = Op.getDebugLoc();
7150 EVT VT = Op.getValueType();
7151 EVT SrcVT = Op1.getValueType();
7153 // If second operand is smaller, extend it first.
7154 if (SrcVT.bitsLT(VT)) {
7155 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
7158 // And if it is bigger, shrink it first.
7159 if (SrcVT.bitsGT(VT)) {
7160 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
7164 // At this point the operands and the result should have the same
7165 // type, and that won't be f80 since that is not custom lowered.
7167 // First get the sign bit of second operand.
7168 std::vector<Constant*> CV;
7169 if (SrcVT == MVT::f64) {
7170 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
7171 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
7173 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
7174 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7175 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7176 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7178 Constant *C = ConstantVector::get(CV);
7179 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7180 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
7181 MachinePointerInfo::getConstantPool(),
7183 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
7185 // Shift sign bit right or left if the two operands have different types.
7186 if (SrcVT.bitsGT(VT)) {
7187 // Op0 is MVT::f32, Op1 is MVT::f64.
7188 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
7189 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
7190 DAG.getConstant(32, MVT::i32));
7191 SignBit = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, SignBit);
7192 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
7193 DAG.getIntPtrConstant(0));
7196 // Clear first operand sign bit.
7198 if (VT == MVT::f64) {
7199 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
7200 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
7202 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
7203 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7204 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7205 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7207 C = ConstantVector::get(CV);
7208 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7209 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7210 MachinePointerInfo::getConstantPool(),
7212 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
7214 // Or the value with the sign bit.
7215 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
7218 /// Emit nodes that will be selected as "test Op0,Op0", or something
7220 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
7221 SelectionDAG &DAG) const {
7222 DebugLoc dl = Op.getDebugLoc();
7224 // CF and OF aren't always set the way we want. Determine which
7225 // of these we need.
7226 bool NeedCF = false;
7227 bool NeedOF = false;
7230 case X86::COND_A: case X86::COND_AE:
7231 case X86::COND_B: case X86::COND_BE:
7234 case X86::COND_G: case X86::COND_GE:
7235 case X86::COND_L: case X86::COND_LE:
7236 case X86::COND_O: case X86::COND_NO:
7241 // See if we can use the EFLAGS value from the operand instead of
7242 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
7243 // we prove that the arithmetic won't overflow, we can't use OF or CF.
7244 if (Op.getResNo() != 0 || NeedOF || NeedCF)
7245 // Emit a CMP with 0, which is the TEST pattern.
7246 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
7247 DAG.getConstant(0, Op.getValueType()));
7249 unsigned Opcode = 0;
7250 unsigned NumOperands = 0;
7251 switch (Op.getNode()->getOpcode()) {
7253 // Due to an isel shortcoming, be conservative if this add is likely to be
7254 // selected as part of a load-modify-store instruction. When the root node
7255 // in a match is a store, isel doesn't know how to remap non-chain non-flag
7256 // uses of other nodes in the match, such as the ADD in this case. This
7257 // leads to the ADD being left around and reselected, with the result being
7258 // two adds in the output. Alas, even if none our users are stores, that
7259 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
7260 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
7261 // climbing the DAG back to the root, and it doesn't seem to be worth the
7263 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
7264 UE = Op.getNode()->use_end(); UI != UE; ++UI)
7265 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
7268 if (ConstantSDNode *C =
7269 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
7270 // An add of one will be selected as an INC.
7271 if (C->getAPIntValue() == 1) {
7272 Opcode = X86ISD::INC;
7277 // An add of negative one (subtract of one) will be selected as a DEC.
7278 if (C->getAPIntValue().isAllOnesValue()) {
7279 Opcode = X86ISD::DEC;
7285 // Otherwise use a regular EFLAGS-setting add.
7286 Opcode = X86ISD::ADD;
7290 // If the primary and result isn't used, don't bother using X86ISD::AND,
7291 // because a TEST instruction will be better.
7292 bool NonFlagUse = false;
7293 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
7294 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
7296 unsigned UOpNo = UI.getOperandNo();
7297 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
7298 // Look pass truncate.
7299 UOpNo = User->use_begin().getOperandNo();
7300 User = *User->use_begin();
7303 if (User->getOpcode() != ISD::BRCOND &&
7304 User->getOpcode() != ISD::SETCC &&
7305 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
7318 // Due to the ISEL shortcoming noted above, be conservative if this op is
7319 // likely to be selected as part of a load-modify-store instruction.
7320 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
7321 UE = Op.getNode()->use_end(); UI != UE; ++UI)
7322 if (UI->getOpcode() == ISD::STORE)
7325 // Otherwise use a regular EFLAGS-setting instruction.
7326 switch (Op.getNode()->getOpcode()) {
7327 default: llvm_unreachable("unexpected operator!");
7328 case ISD::SUB: Opcode = X86ISD::SUB; break;
7329 case ISD::OR: Opcode = X86ISD::OR; break;
7330 case ISD::XOR: Opcode = X86ISD::XOR; break;
7331 case ISD::AND: Opcode = X86ISD::AND; break;
7343 return SDValue(Op.getNode(), 1);
7350 // Emit a CMP with 0, which is the TEST pattern.
7351 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
7352 DAG.getConstant(0, Op.getValueType()));
7354 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
7355 SmallVector<SDValue, 4> Ops;
7356 for (unsigned i = 0; i != NumOperands; ++i)
7357 Ops.push_back(Op.getOperand(i));
7359 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
7360 DAG.ReplaceAllUsesWith(Op, New);
7361 return SDValue(New.getNode(), 1);
7364 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
7366 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
7367 SelectionDAG &DAG) const {
7368 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
7369 if (C->getAPIntValue() == 0)
7370 return EmitTest(Op0, X86CC, DAG);
7372 DebugLoc dl = Op0.getDebugLoc();
7373 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
7376 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
7377 /// if it's possible.
7378 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
7379 DebugLoc dl, SelectionDAG &DAG) const {
7380 SDValue Op0 = And.getOperand(0);
7381 SDValue Op1 = And.getOperand(1);
7382 if (Op0.getOpcode() == ISD::TRUNCATE)
7383 Op0 = Op0.getOperand(0);
7384 if (Op1.getOpcode() == ISD::TRUNCATE)
7385 Op1 = Op1.getOperand(0);
7388 if (Op1.getOpcode() == ISD::SHL)
7389 std::swap(Op0, Op1);
7390 if (Op0.getOpcode() == ISD::SHL) {
7391 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
7392 if (And00C->getZExtValue() == 1) {
7393 // If we looked past a truncate, check that it's only truncating away
7395 unsigned BitWidth = Op0.getValueSizeInBits();
7396 unsigned AndBitWidth = And.getValueSizeInBits();
7397 if (BitWidth > AndBitWidth) {
7398 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
7399 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
7400 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
7404 RHS = Op0.getOperand(1);
7406 } else if (Op1.getOpcode() == ISD::Constant) {
7407 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
7408 SDValue AndLHS = Op0;
7409 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
7410 LHS = AndLHS.getOperand(0);
7411 RHS = AndLHS.getOperand(1);
7415 if (LHS.getNode()) {
7416 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
7417 // instruction. Since the shift amount is in-range-or-undefined, we know
7418 // that doing a bittest on the i32 value is ok. We extend to i32 because
7419 // the encoding for the i16 version is larger than the i32 version.
7420 // Also promote i16 to i32 for performance / code size reason.
7421 if (LHS.getValueType() == MVT::i8 ||
7422 LHS.getValueType() == MVT::i16)
7423 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
7425 // If the operand types disagree, extend the shift amount to match. Since
7426 // BT ignores high bits (like shifts) we can use anyextend.
7427 if (LHS.getValueType() != RHS.getValueType())
7428 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
7430 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
7431 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
7432 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7433 DAG.getConstant(Cond, MVT::i8), BT);
7439 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
7440 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
7441 SDValue Op0 = Op.getOperand(0);
7442 SDValue Op1 = Op.getOperand(1);
7443 DebugLoc dl = Op.getDebugLoc();
7444 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
7446 // Optimize to BT if possible.
7447 // Lower (X & (1 << N)) == 0 to BT(X, N).
7448 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
7449 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
7450 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
7451 Op1.getOpcode() == ISD::Constant &&
7452 cast<ConstantSDNode>(Op1)->isNullValue() &&
7453 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
7454 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
7455 if (NewSetCC.getNode())
7459 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
7461 if (Op1.getOpcode() == ISD::Constant &&
7462 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
7463 cast<ConstantSDNode>(Op1)->isNullValue()) &&
7464 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
7466 // If the input is a setcc, then reuse the input setcc or use a new one with
7467 // the inverted condition.
7468 if (Op0.getOpcode() == X86ISD::SETCC) {
7469 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
7470 bool Invert = (CC == ISD::SETNE) ^
7471 cast<ConstantSDNode>(Op1)->isNullValue();
7472 if (!Invert) return Op0;
7474 CCode = X86::GetOppositeBranchCondition(CCode);
7475 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7476 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
7480 bool isFP = Op1.getValueType().isFloatingPoint();
7481 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
7482 if (X86CC == X86::COND_INVALID)
7485 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, DAG);
7486 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7487 DAG.getConstant(X86CC, MVT::i8), EFLAGS);
7490 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
7492 SDValue Op0 = Op.getOperand(0);
7493 SDValue Op1 = Op.getOperand(1);
7494 SDValue CC = Op.getOperand(2);
7495 EVT VT = Op.getValueType();
7496 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
7497 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
7498 DebugLoc dl = Op.getDebugLoc();
7502 EVT VT0 = Op0.getValueType();
7503 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
7504 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
7507 switch (SetCCOpcode) {
7510 case ISD::SETEQ: SSECC = 0; break;
7512 case ISD::SETGT: Swap = true; // Fallthrough
7514 case ISD::SETOLT: SSECC = 1; break;
7516 case ISD::SETGE: Swap = true; // Fallthrough
7518 case ISD::SETOLE: SSECC = 2; break;
7519 case ISD::SETUO: SSECC = 3; break;
7521 case ISD::SETNE: SSECC = 4; break;
7522 case ISD::SETULE: Swap = true;
7523 case ISD::SETUGE: SSECC = 5; break;
7524 case ISD::SETULT: Swap = true;
7525 case ISD::SETUGT: SSECC = 6; break;
7526 case ISD::SETO: SSECC = 7; break;
7529 std::swap(Op0, Op1);
7531 // In the two special cases we can't handle, emit two comparisons.
7533 if (SetCCOpcode == ISD::SETUEQ) {
7535 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
7536 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
7537 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
7539 else if (SetCCOpcode == ISD::SETONE) {
7541 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
7542 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
7543 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
7545 llvm_unreachable("Illegal FP comparison");
7547 // Handle all other FP comparisons here.
7548 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
7551 // We are handling one of the integer comparisons here. Since SSE only has
7552 // GT and EQ comparisons for integer, swapping operands and multiple
7553 // operations may be required for some comparisons.
7554 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
7555 bool Swap = false, Invert = false, FlipSigns = false;
7557 switch (VT.getSimpleVT().SimpleTy) {
7559 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
7560 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
7561 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
7562 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
7565 switch (SetCCOpcode) {
7567 case ISD::SETNE: Invert = true;
7568 case ISD::SETEQ: Opc = EQOpc; break;
7569 case ISD::SETLT: Swap = true;
7570 case ISD::SETGT: Opc = GTOpc; break;
7571 case ISD::SETGE: Swap = true;
7572 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
7573 case ISD::SETULT: Swap = true;
7574 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
7575 case ISD::SETUGE: Swap = true;
7576 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
7579 std::swap(Op0, Op1);
7581 // Since SSE has no unsigned integer comparisons, we need to flip the sign
7582 // bits of the inputs before performing those operations.
7584 EVT EltVT = VT.getVectorElementType();
7585 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
7587 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
7588 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
7590 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
7591 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
7594 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
7596 // If the logical-not of the result is required, perform that now.
7598 Result = DAG.getNOT(dl, Result, VT);
7603 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
7604 static bool isX86LogicalCmp(SDValue Op) {
7605 unsigned Opc = Op.getNode()->getOpcode();
7606 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
7608 if (Op.getResNo() == 1 &&
7609 (Opc == X86ISD::ADD ||
7610 Opc == X86ISD::SUB ||
7611 Opc == X86ISD::ADC ||
7612 Opc == X86ISD::SBB ||
7613 Opc == X86ISD::SMUL ||
7614 Opc == X86ISD::UMUL ||
7615 Opc == X86ISD::INC ||
7616 Opc == X86ISD::DEC ||
7617 Opc == X86ISD::OR ||
7618 Opc == X86ISD::XOR ||
7619 Opc == X86ISD::AND))
7622 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
7628 static bool isZero(SDValue V) {
7629 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
7630 return C && C->isNullValue();
7633 static bool isAllOnes(SDValue V) {
7634 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
7635 return C && C->isAllOnesValue();
7638 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
7639 bool addTest = true;
7640 SDValue Cond = Op.getOperand(0);
7641 SDValue Op1 = Op.getOperand(1);
7642 SDValue Op2 = Op.getOperand(2);
7643 DebugLoc DL = Op.getDebugLoc();
7646 if (Cond.getOpcode() == ISD::SETCC) {
7647 SDValue NewCond = LowerSETCC(Cond, DAG);
7648 if (NewCond.getNode())
7652 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
7653 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
7654 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
7655 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
7656 if (Cond.getOpcode() == X86ISD::SETCC &&
7657 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
7658 isZero(Cond.getOperand(1).getOperand(1))) {
7659 SDValue Cmp = Cond.getOperand(1);
7661 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
7663 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
7664 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
7665 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
7667 SDValue CmpOp0 = Cmp.getOperand(0);
7668 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
7669 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
7671 SDValue Res = // Res = 0 or -1.
7672 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
7673 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
7675 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
7676 Res = DAG.getNOT(DL, Res, Res.getValueType());
7678 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
7679 if (N2C == 0 || !N2C->isNullValue())
7680 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
7685 // Look past (and (setcc_carry (cmp ...)), 1).
7686 if (Cond.getOpcode() == ISD::AND &&
7687 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
7688 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
7689 if (C && C->getAPIntValue() == 1)
7690 Cond = Cond.getOperand(0);
7693 // If condition flag is set by a X86ISD::CMP, then use it as the condition
7694 // setting operand in place of the X86ISD::SETCC.
7695 if (Cond.getOpcode() == X86ISD::SETCC ||
7696 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
7697 CC = Cond.getOperand(0);
7699 SDValue Cmp = Cond.getOperand(1);
7700 unsigned Opc = Cmp.getOpcode();
7701 EVT VT = Op.getValueType();
7703 bool IllegalFPCMov = false;
7704 if (VT.isFloatingPoint() && !VT.isVector() &&
7705 !isScalarFPTypeInSSEReg(VT)) // FPStack?
7706 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
7708 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
7709 Opc == X86ISD::BT) { // FIXME
7716 // Look pass the truncate.
7717 if (Cond.getOpcode() == ISD::TRUNCATE)
7718 Cond = Cond.getOperand(0);
7720 // We know the result of AND is compared against zero. Try to match
7722 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
7723 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
7724 if (NewSetCC.getNode()) {
7725 CC = NewSetCC.getOperand(0);
7726 Cond = NewSetCC.getOperand(1);
7733 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7734 Cond = EmitTest(Cond, X86::COND_NE, DAG);
7737 // a < b ? -1 : 0 -> RES = ~setcc_carry
7738 // a < b ? 0 : -1 -> RES = setcc_carry
7739 // a >= b ? -1 : 0 -> RES = setcc_carry
7740 // a >= b ? 0 : -1 -> RES = ~setcc_carry
7741 if (Cond.getOpcode() == X86ISD::CMP) {
7742 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
7744 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
7745 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
7746 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
7747 DAG.getConstant(X86::COND_B, MVT::i8), Cond);
7748 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
7749 return DAG.getNOT(DL, Res, Res.getValueType());
7754 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
7755 // condition is true.
7756 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
7757 SDValue Ops[] = { Op2, Op1, CC, Cond };
7758 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops, array_lengthof(Ops));
7761 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
7762 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
7763 // from the AND / OR.
7764 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
7765 Opc = Op.getOpcode();
7766 if (Opc != ISD::OR && Opc != ISD::AND)
7768 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
7769 Op.getOperand(0).hasOneUse() &&
7770 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
7771 Op.getOperand(1).hasOneUse());
7774 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
7775 // 1 and that the SETCC node has a single use.
7776 static bool isXor1OfSetCC(SDValue Op) {
7777 if (Op.getOpcode() != ISD::XOR)
7779 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7780 if (N1C && N1C->getAPIntValue() == 1) {
7781 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
7782 Op.getOperand(0).hasOneUse();
7787 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
7788 bool addTest = true;
7789 SDValue Chain = Op.getOperand(0);
7790 SDValue Cond = Op.getOperand(1);
7791 SDValue Dest = Op.getOperand(2);
7792 DebugLoc dl = Op.getDebugLoc();
7795 if (Cond.getOpcode() == ISD::SETCC) {
7796 SDValue NewCond = LowerSETCC(Cond, DAG);
7797 if (NewCond.getNode())
7801 // FIXME: LowerXALUO doesn't handle these!!
7802 else if (Cond.getOpcode() == X86ISD::ADD ||
7803 Cond.getOpcode() == X86ISD::SUB ||
7804 Cond.getOpcode() == X86ISD::SMUL ||
7805 Cond.getOpcode() == X86ISD::UMUL)
7806 Cond = LowerXALUO(Cond, DAG);
7809 // Look pass (and (setcc_carry (cmp ...)), 1).
7810 if (Cond.getOpcode() == ISD::AND &&
7811 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
7812 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
7813 if (C && C->getAPIntValue() == 1)
7814 Cond = Cond.getOperand(0);
7817 // If condition flag is set by a X86ISD::CMP, then use it as the condition
7818 // setting operand in place of the X86ISD::SETCC.
7819 if (Cond.getOpcode() == X86ISD::SETCC ||
7820 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
7821 CC = Cond.getOperand(0);
7823 SDValue Cmp = Cond.getOperand(1);
7824 unsigned Opc = Cmp.getOpcode();
7825 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
7826 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
7830 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
7834 // These can only come from an arithmetic instruction with overflow,
7835 // e.g. SADDO, UADDO.
7836 Cond = Cond.getNode()->getOperand(1);
7843 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
7844 SDValue Cmp = Cond.getOperand(0).getOperand(1);
7845 if (CondOpc == ISD::OR) {
7846 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
7847 // two branches instead of an explicit OR instruction with a
7849 if (Cmp == Cond.getOperand(1).getOperand(1) &&
7850 isX86LogicalCmp(Cmp)) {
7851 CC = Cond.getOperand(0).getOperand(0);
7852 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7853 Chain, Dest, CC, Cmp);
7854 CC = Cond.getOperand(1).getOperand(0);
7858 } else { // ISD::AND
7859 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
7860 // two branches instead of an explicit AND instruction with a
7861 // separate test. However, we only do this if this block doesn't
7862 // have a fall-through edge, because this requires an explicit
7863 // jmp when the condition is false.
7864 if (Cmp == Cond.getOperand(1).getOperand(1) &&
7865 isX86LogicalCmp(Cmp) &&
7866 Op.getNode()->hasOneUse()) {
7867 X86::CondCode CCode =
7868 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
7869 CCode = X86::GetOppositeBranchCondition(CCode);
7870 CC = DAG.getConstant(CCode, MVT::i8);
7871 SDNode *User = *Op.getNode()->use_begin();
7872 // Look for an unconditional branch following this conditional branch.
7873 // We need this because we need to reverse the successors in order
7874 // to implement FCMP_OEQ.
7875 if (User->getOpcode() == ISD::BR) {
7876 SDValue FalseBB = User->getOperand(1);
7878 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
7879 assert(NewBR == User);
7883 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7884 Chain, Dest, CC, Cmp);
7885 X86::CondCode CCode =
7886 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
7887 CCode = X86::GetOppositeBranchCondition(CCode);
7888 CC = DAG.getConstant(CCode, MVT::i8);
7894 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
7895 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
7896 // It should be transformed during dag combiner except when the condition
7897 // is set by a arithmetics with overflow node.
7898 X86::CondCode CCode =
7899 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
7900 CCode = X86::GetOppositeBranchCondition(CCode);
7901 CC = DAG.getConstant(CCode, MVT::i8);
7902 Cond = Cond.getOperand(0).getOperand(1);
7908 // Look pass the truncate.
7909 if (Cond.getOpcode() == ISD::TRUNCATE)
7910 Cond = Cond.getOperand(0);
7912 // We know the result of AND is compared against zero. Try to match
7914 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
7915 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
7916 if (NewSetCC.getNode()) {
7917 CC = NewSetCC.getOperand(0);
7918 Cond = NewSetCC.getOperand(1);
7925 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7926 Cond = EmitTest(Cond, X86::COND_NE, DAG);
7928 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7929 Chain, Dest, CC, Cond);
7933 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
7934 // Calls to _alloca is needed to probe the stack when allocating more than 4k
7935 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
7936 // that the guard pages used by the OS virtual memory manager are allocated in
7937 // correct sequence.
7939 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
7940 SelectionDAG &DAG) const {
7941 assert((Subtarget->isTargetCygMing() || Subtarget->isTargetWindows()) &&
7942 "This should be used only on Windows targets");
7943 assert(!Subtarget->isTargetEnvMacho());
7944 DebugLoc dl = Op.getDebugLoc();
7947 SDValue Chain = Op.getOperand(0);
7948 SDValue Size = Op.getOperand(1);
7949 // FIXME: Ensure alignment here
7953 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
7954 unsigned Reg = (Subtarget->is64Bit() ? X86::RAX : X86::EAX);
7956 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
7957 Flag = Chain.getValue(1);
7959 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
7961 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
7962 Flag = Chain.getValue(1);
7964 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
7966 SDValue Ops1[2] = { Chain.getValue(0), Chain };
7967 return DAG.getMergeValues(Ops1, 2, dl);
7970 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
7971 MachineFunction &MF = DAG.getMachineFunction();
7972 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
7974 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
7975 DebugLoc DL = Op.getDebugLoc();
7977 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
7978 // vastart just stores the address of the VarArgsFrameIndex slot into the
7979 // memory location argument.
7980 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
7982 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
7983 MachinePointerInfo(SV), false, false, 0);
7987 // gp_offset (0 - 6 * 8)
7988 // fp_offset (48 - 48 + 8 * 16)
7989 // overflow_arg_area (point to parameters coming in memory).
7991 SmallVector<SDValue, 8> MemOps;
7992 SDValue FIN = Op.getOperand(1);
7994 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
7995 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
7997 FIN, MachinePointerInfo(SV), false, false, 0);
7998 MemOps.push_back(Store);
8001 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8002 FIN, DAG.getIntPtrConstant(4));
8003 Store = DAG.getStore(Op.getOperand(0), DL,
8004 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
8006 FIN, MachinePointerInfo(SV, 4), false, false, 0);
8007 MemOps.push_back(Store);
8009 // Store ptr to overflow_arg_area
8010 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8011 FIN, DAG.getIntPtrConstant(4));
8012 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
8014 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
8015 MachinePointerInfo(SV, 8),
8017 MemOps.push_back(Store);
8019 // Store ptr to reg_save_area.
8020 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8021 FIN, DAG.getIntPtrConstant(8));
8022 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
8024 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
8025 MachinePointerInfo(SV, 16), false, false, 0);
8026 MemOps.push_back(Store);
8027 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
8028 &MemOps[0], MemOps.size());
8031 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
8032 assert(Subtarget->is64Bit() &&
8033 "LowerVAARG only handles 64-bit va_arg!");
8034 assert((Subtarget->isTargetLinux() ||
8035 Subtarget->isTargetDarwin()) &&
8036 "Unhandled target in LowerVAARG");
8037 assert(Op.getNode()->getNumOperands() == 4);
8038 SDValue Chain = Op.getOperand(0);
8039 SDValue SrcPtr = Op.getOperand(1);
8040 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
8041 unsigned Align = Op.getConstantOperandVal(3);
8042 DebugLoc dl = Op.getDebugLoc();
8044 EVT ArgVT = Op.getNode()->getValueType(0);
8045 const Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
8046 uint32_t ArgSize = getTargetData()->getTypeAllocSize(ArgTy);
8049 // Decide which area this value should be read from.
8050 // TODO: Implement the AMD64 ABI in its entirety. This simple
8051 // selection mechanism works only for the basic types.
8052 if (ArgVT == MVT::f80) {
8053 llvm_unreachable("va_arg for f80 not yet implemented");
8054 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
8055 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
8056 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
8057 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
8059 llvm_unreachable("Unhandled argument type in LowerVAARG");
8063 // Sanity Check: Make sure using fp_offset makes sense.
8064 assert(!UseSoftFloat &&
8065 !(DAG.getMachineFunction()
8066 .getFunction()->hasFnAttr(Attribute::NoImplicitFloat)) &&
8067 Subtarget->hasXMM());
8070 // Insert VAARG_64 node into the DAG
8071 // VAARG_64 returns two values: Variable Argument Address, Chain
8072 SmallVector<SDValue, 11> InstOps;
8073 InstOps.push_back(Chain);
8074 InstOps.push_back(SrcPtr);
8075 InstOps.push_back(DAG.getConstant(ArgSize, MVT::i32));
8076 InstOps.push_back(DAG.getConstant(ArgMode, MVT::i8));
8077 InstOps.push_back(DAG.getConstant(Align, MVT::i32));
8078 SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
8079 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
8080 VTs, &InstOps[0], InstOps.size(),
8082 MachinePointerInfo(SV),
8087 Chain = VAARG.getValue(1);
8089 // Load the next argument and return it
8090 return DAG.getLoad(ArgVT, dl,
8093 MachinePointerInfo(),
8097 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
8098 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
8099 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
8100 SDValue Chain = Op.getOperand(0);
8101 SDValue DstPtr = Op.getOperand(1);
8102 SDValue SrcPtr = Op.getOperand(2);
8103 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
8104 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
8105 DebugLoc DL = Op.getDebugLoc();
8107 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
8108 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
8110 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
8114 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
8115 DebugLoc dl = Op.getDebugLoc();
8116 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8118 default: return SDValue(); // Don't custom lower most intrinsics.
8119 // Comparison intrinsics.
8120 case Intrinsic::x86_sse_comieq_ss:
8121 case Intrinsic::x86_sse_comilt_ss:
8122 case Intrinsic::x86_sse_comile_ss:
8123 case Intrinsic::x86_sse_comigt_ss:
8124 case Intrinsic::x86_sse_comige_ss:
8125 case Intrinsic::x86_sse_comineq_ss:
8126 case Intrinsic::x86_sse_ucomieq_ss:
8127 case Intrinsic::x86_sse_ucomilt_ss:
8128 case Intrinsic::x86_sse_ucomile_ss:
8129 case Intrinsic::x86_sse_ucomigt_ss:
8130 case Intrinsic::x86_sse_ucomige_ss:
8131 case Intrinsic::x86_sse_ucomineq_ss:
8132 case Intrinsic::x86_sse2_comieq_sd:
8133 case Intrinsic::x86_sse2_comilt_sd:
8134 case Intrinsic::x86_sse2_comile_sd:
8135 case Intrinsic::x86_sse2_comigt_sd:
8136 case Intrinsic::x86_sse2_comige_sd:
8137 case Intrinsic::x86_sse2_comineq_sd:
8138 case Intrinsic::x86_sse2_ucomieq_sd:
8139 case Intrinsic::x86_sse2_ucomilt_sd:
8140 case Intrinsic::x86_sse2_ucomile_sd:
8141 case Intrinsic::x86_sse2_ucomigt_sd:
8142 case Intrinsic::x86_sse2_ucomige_sd:
8143 case Intrinsic::x86_sse2_ucomineq_sd: {
8145 ISD::CondCode CC = ISD::SETCC_INVALID;
8148 case Intrinsic::x86_sse_comieq_ss:
8149 case Intrinsic::x86_sse2_comieq_sd:
8153 case Intrinsic::x86_sse_comilt_ss:
8154 case Intrinsic::x86_sse2_comilt_sd:
8158 case Intrinsic::x86_sse_comile_ss:
8159 case Intrinsic::x86_sse2_comile_sd:
8163 case Intrinsic::x86_sse_comigt_ss:
8164 case Intrinsic::x86_sse2_comigt_sd:
8168 case Intrinsic::x86_sse_comige_ss:
8169 case Intrinsic::x86_sse2_comige_sd:
8173 case Intrinsic::x86_sse_comineq_ss:
8174 case Intrinsic::x86_sse2_comineq_sd:
8178 case Intrinsic::x86_sse_ucomieq_ss:
8179 case Intrinsic::x86_sse2_ucomieq_sd:
8180 Opc = X86ISD::UCOMI;
8183 case Intrinsic::x86_sse_ucomilt_ss:
8184 case Intrinsic::x86_sse2_ucomilt_sd:
8185 Opc = X86ISD::UCOMI;
8188 case Intrinsic::x86_sse_ucomile_ss:
8189 case Intrinsic::x86_sse2_ucomile_sd:
8190 Opc = X86ISD::UCOMI;
8193 case Intrinsic::x86_sse_ucomigt_ss:
8194 case Intrinsic::x86_sse2_ucomigt_sd:
8195 Opc = X86ISD::UCOMI;
8198 case Intrinsic::x86_sse_ucomige_ss:
8199 case Intrinsic::x86_sse2_ucomige_sd:
8200 Opc = X86ISD::UCOMI;
8203 case Intrinsic::x86_sse_ucomineq_ss:
8204 case Intrinsic::x86_sse2_ucomineq_sd:
8205 Opc = X86ISD::UCOMI;
8210 SDValue LHS = Op.getOperand(1);
8211 SDValue RHS = Op.getOperand(2);
8212 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
8213 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
8214 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
8215 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
8216 DAG.getConstant(X86CC, MVT::i8), Cond);
8217 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
8219 // ptest and testp intrinsics. The intrinsic these come from are designed to
8220 // return an integer value, not just an instruction so lower it to the ptest
8221 // or testp pattern and a setcc for the result.
8222 case Intrinsic::x86_sse41_ptestz:
8223 case Intrinsic::x86_sse41_ptestc:
8224 case Intrinsic::x86_sse41_ptestnzc:
8225 case Intrinsic::x86_avx_ptestz_256:
8226 case Intrinsic::x86_avx_ptestc_256:
8227 case Intrinsic::x86_avx_ptestnzc_256:
8228 case Intrinsic::x86_avx_vtestz_ps:
8229 case Intrinsic::x86_avx_vtestc_ps:
8230 case Intrinsic::x86_avx_vtestnzc_ps:
8231 case Intrinsic::x86_avx_vtestz_pd:
8232 case Intrinsic::x86_avx_vtestc_pd:
8233 case Intrinsic::x86_avx_vtestnzc_pd:
8234 case Intrinsic::x86_avx_vtestz_ps_256:
8235 case Intrinsic::x86_avx_vtestc_ps_256:
8236 case Intrinsic::x86_avx_vtestnzc_ps_256:
8237 case Intrinsic::x86_avx_vtestz_pd_256:
8238 case Intrinsic::x86_avx_vtestc_pd_256:
8239 case Intrinsic::x86_avx_vtestnzc_pd_256: {
8240 bool IsTestPacked = false;
8243 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
8244 case Intrinsic::x86_avx_vtestz_ps:
8245 case Intrinsic::x86_avx_vtestz_pd:
8246 case Intrinsic::x86_avx_vtestz_ps_256:
8247 case Intrinsic::x86_avx_vtestz_pd_256:
8248 IsTestPacked = true; // Fallthrough
8249 case Intrinsic::x86_sse41_ptestz:
8250 case Intrinsic::x86_avx_ptestz_256:
8252 X86CC = X86::COND_E;
8254 case Intrinsic::x86_avx_vtestc_ps:
8255 case Intrinsic::x86_avx_vtestc_pd:
8256 case Intrinsic::x86_avx_vtestc_ps_256:
8257 case Intrinsic::x86_avx_vtestc_pd_256:
8258 IsTestPacked = true; // Fallthrough
8259 case Intrinsic::x86_sse41_ptestc:
8260 case Intrinsic::x86_avx_ptestc_256:
8262 X86CC = X86::COND_B;
8264 case Intrinsic::x86_avx_vtestnzc_ps:
8265 case Intrinsic::x86_avx_vtestnzc_pd:
8266 case Intrinsic::x86_avx_vtestnzc_ps_256:
8267 case Intrinsic::x86_avx_vtestnzc_pd_256:
8268 IsTestPacked = true; // Fallthrough
8269 case Intrinsic::x86_sse41_ptestnzc:
8270 case Intrinsic::x86_avx_ptestnzc_256:
8272 X86CC = X86::COND_A;
8276 SDValue LHS = Op.getOperand(1);
8277 SDValue RHS = Op.getOperand(2);
8278 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
8279 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
8280 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
8281 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
8282 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
8285 // Fix vector shift instructions where the last operand is a non-immediate
8287 case Intrinsic::x86_sse2_pslli_w:
8288 case Intrinsic::x86_sse2_pslli_d:
8289 case Intrinsic::x86_sse2_pslli_q:
8290 case Intrinsic::x86_sse2_psrli_w:
8291 case Intrinsic::x86_sse2_psrli_d:
8292 case Intrinsic::x86_sse2_psrli_q:
8293 case Intrinsic::x86_sse2_psrai_w:
8294 case Intrinsic::x86_sse2_psrai_d:
8295 case Intrinsic::x86_mmx_pslli_w:
8296 case Intrinsic::x86_mmx_pslli_d:
8297 case Intrinsic::x86_mmx_pslli_q:
8298 case Intrinsic::x86_mmx_psrli_w:
8299 case Intrinsic::x86_mmx_psrli_d:
8300 case Intrinsic::x86_mmx_psrli_q:
8301 case Intrinsic::x86_mmx_psrai_w:
8302 case Intrinsic::x86_mmx_psrai_d: {
8303 SDValue ShAmt = Op.getOperand(2);
8304 if (isa<ConstantSDNode>(ShAmt))
8307 unsigned NewIntNo = 0;
8308 EVT ShAmtVT = MVT::v4i32;
8310 case Intrinsic::x86_sse2_pslli_w:
8311 NewIntNo = Intrinsic::x86_sse2_psll_w;
8313 case Intrinsic::x86_sse2_pslli_d:
8314 NewIntNo = Intrinsic::x86_sse2_psll_d;
8316 case Intrinsic::x86_sse2_pslli_q:
8317 NewIntNo = Intrinsic::x86_sse2_psll_q;
8319 case Intrinsic::x86_sse2_psrli_w:
8320 NewIntNo = Intrinsic::x86_sse2_psrl_w;
8322 case Intrinsic::x86_sse2_psrli_d:
8323 NewIntNo = Intrinsic::x86_sse2_psrl_d;
8325 case Intrinsic::x86_sse2_psrli_q:
8326 NewIntNo = Intrinsic::x86_sse2_psrl_q;
8328 case Intrinsic::x86_sse2_psrai_w:
8329 NewIntNo = Intrinsic::x86_sse2_psra_w;
8331 case Intrinsic::x86_sse2_psrai_d:
8332 NewIntNo = Intrinsic::x86_sse2_psra_d;
8335 ShAmtVT = MVT::v2i32;
8337 case Intrinsic::x86_mmx_pslli_w:
8338 NewIntNo = Intrinsic::x86_mmx_psll_w;
8340 case Intrinsic::x86_mmx_pslli_d:
8341 NewIntNo = Intrinsic::x86_mmx_psll_d;
8343 case Intrinsic::x86_mmx_pslli_q:
8344 NewIntNo = Intrinsic::x86_mmx_psll_q;
8346 case Intrinsic::x86_mmx_psrli_w:
8347 NewIntNo = Intrinsic::x86_mmx_psrl_w;
8349 case Intrinsic::x86_mmx_psrli_d:
8350 NewIntNo = Intrinsic::x86_mmx_psrl_d;
8352 case Intrinsic::x86_mmx_psrli_q:
8353 NewIntNo = Intrinsic::x86_mmx_psrl_q;
8355 case Intrinsic::x86_mmx_psrai_w:
8356 NewIntNo = Intrinsic::x86_mmx_psra_w;
8358 case Intrinsic::x86_mmx_psrai_d:
8359 NewIntNo = Intrinsic::x86_mmx_psra_d;
8361 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
8367 // The vector shift intrinsics with scalars uses 32b shift amounts but
8368 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
8372 ShOps[1] = DAG.getConstant(0, MVT::i32);
8373 if (ShAmtVT == MVT::v4i32) {
8374 ShOps[2] = DAG.getUNDEF(MVT::i32);
8375 ShOps[3] = DAG.getUNDEF(MVT::i32);
8376 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
8378 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
8379 // FIXME this must be lowered to get rid of the invalid type.
8382 EVT VT = Op.getValueType();
8383 ShAmt = DAG.getNode(ISD::BITCAST, dl, VT, ShAmt);
8384 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8385 DAG.getConstant(NewIntNo, MVT::i32),
8386 Op.getOperand(1), ShAmt);
8391 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
8392 SelectionDAG &DAG) const {
8393 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8394 MFI->setReturnAddressIsTaken(true);
8396 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8397 DebugLoc dl = Op.getDebugLoc();
8400 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
8402 DAG.getConstant(TD->getPointerSize(),
8403 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
8404 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
8405 DAG.getNode(ISD::ADD, dl, getPointerTy(),
8407 MachinePointerInfo(), false, false, 0);
8410 // Just load the return address.
8411 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
8412 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
8413 RetAddrFI, MachinePointerInfo(), false, false, 0);
8416 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
8417 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8418 MFI->setFrameAddressIsTaken(true);
8420 EVT VT = Op.getValueType();
8421 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
8422 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8423 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
8424 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
8426 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
8427 MachinePointerInfo(),
8432 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
8433 SelectionDAG &DAG) const {
8434 return DAG.getIntPtrConstant(2*TD->getPointerSize());
8437 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
8438 MachineFunction &MF = DAG.getMachineFunction();
8439 SDValue Chain = Op.getOperand(0);
8440 SDValue Offset = Op.getOperand(1);
8441 SDValue Handler = Op.getOperand(2);
8442 DebugLoc dl = Op.getDebugLoc();
8444 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
8445 Subtarget->is64Bit() ? X86::RBP : X86::EBP,
8447 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
8449 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Frame,
8450 DAG.getIntPtrConstant(TD->getPointerSize()));
8451 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
8452 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
8454 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
8455 MF.getRegInfo().addLiveOut(StoreAddrReg);
8457 return DAG.getNode(X86ISD::EH_RETURN, dl,
8459 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
8462 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
8463 SelectionDAG &DAG) const {
8464 SDValue Root = Op.getOperand(0);
8465 SDValue Trmp = Op.getOperand(1); // trampoline
8466 SDValue FPtr = Op.getOperand(2); // nested function
8467 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
8468 DebugLoc dl = Op.getDebugLoc();
8470 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
8472 if (Subtarget->is64Bit()) {
8473 SDValue OutChains[6];
8475 // Large code-model.
8476 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
8477 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
8479 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
8480 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
8482 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
8484 // Load the pointer to the nested function into R11.
8485 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
8486 SDValue Addr = Trmp;
8487 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8488 Addr, MachinePointerInfo(TrmpAddr),
8491 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8492 DAG.getConstant(2, MVT::i64));
8493 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
8494 MachinePointerInfo(TrmpAddr, 2),
8497 // Load the 'nest' parameter value into R10.
8498 // R10 is specified in X86CallingConv.td
8499 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
8500 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8501 DAG.getConstant(10, MVT::i64));
8502 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8503 Addr, MachinePointerInfo(TrmpAddr, 10),
8506 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8507 DAG.getConstant(12, MVT::i64));
8508 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
8509 MachinePointerInfo(TrmpAddr, 12),
8512 // Jump to the nested function.
8513 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
8514 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8515 DAG.getConstant(20, MVT::i64));
8516 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8517 Addr, MachinePointerInfo(TrmpAddr, 20),
8520 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
8521 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8522 DAG.getConstant(22, MVT::i64));
8523 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
8524 MachinePointerInfo(TrmpAddr, 22),
8528 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
8529 return DAG.getMergeValues(Ops, 2, dl);
8531 const Function *Func =
8532 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
8533 CallingConv::ID CC = Func->getCallingConv();
8538 llvm_unreachable("Unsupported calling convention");
8539 case CallingConv::C:
8540 case CallingConv::X86_StdCall: {
8541 // Pass 'nest' parameter in ECX.
8542 // Must be kept in sync with X86CallingConv.td
8545 // Check that ECX wasn't needed by an 'inreg' parameter.
8546 const FunctionType *FTy = Func->getFunctionType();
8547 const AttrListPtr &Attrs = Func->getAttributes();
8549 if (!Attrs.isEmpty() && !Func->isVarArg()) {
8550 unsigned InRegCount = 0;
8553 for (FunctionType::param_iterator I = FTy->param_begin(),
8554 E = FTy->param_end(); I != E; ++I, ++Idx)
8555 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
8556 // FIXME: should only count parameters that are lowered to integers.
8557 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
8559 if (InRegCount > 2) {
8560 report_fatal_error("Nest register in use - reduce number of inreg"
8566 case CallingConv::X86_FastCall:
8567 case CallingConv::X86_ThisCall:
8568 case CallingConv::Fast:
8569 // Pass 'nest' parameter in EAX.
8570 // Must be kept in sync with X86CallingConv.td
8575 SDValue OutChains[4];
8578 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8579 DAG.getConstant(10, MVT::i32));
8580 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
8582 // This is storing the opcode for MOV32ri.
8583 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
8584 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
8585 OutChains[0] = DAG.getStore(Root, dl,
8586 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
8587 Trmp, MachinePointerInfo(TrmpAddr),
8590 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8591 DAG.getConstant(1, MVT::i32));
8592 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
8593 MachinePointerInfo(TrmpAddr, 1),
8596 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
8597 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8598 DAG.getConstant(5, MVT::i32));
8599 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
8600 MachinePointerInfo(TrmpAddr, 5),
8603 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8604 DAG.getConstant(6, MVT::i32));
8605 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
8606 MachinePointerInfo(TrmpAddr, 6),
8610 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
8611 return DAG.getMergeValues(Ops, 2, dl);
8615 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
8616 SelectionDAG &DAG) const {
8618 The rounding mode is in bits 11:10 of FPSR, and has the following
8625 FLT_ROUNDS, on the other hand, expects the following:
8632 To perform the conversion, we do:
8633 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
8636 MachineFunction &MF = DAG.getMachineFunction();
8637 const TargetMachine &TM = MF.getTarget();
8638 const TargetFrameLowering &TFI = *TM.getFrameLowering();
8639 unsigned StackAlignment = TFI.getStackAlignment();
8640 EVT VT = Op.getValueType();
8641 DebugLoc DL = Op.getDebugLoc();
8643 // Save FP Control Word to stack slot
8644 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
8645 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
8648 MachineMemOperand *MMO =
8649 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
8650 MachineMemOperand::MOStore, 2, 2);
8652 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
8653 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
8654 DAG.getVTList(MVT::Other),
8655 Ops, 2, MVT::i16, MMO);
8657 // Load FP Control Word from stack slot
8658 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
8659 MachinePointerInfo(), false, false, 0);
8661 // Transform as necessary
8663 DAG.getNode(ISD::SRL, DL, MVT::i16,
8664 DAG.getNode(ISD::AND, DL, MVT::i16,
8665 CWD, DAG.getConstant(0x800, MVT::i16)),
8666 DAG.getConstant(11, MVT::i8));
8668 DAG.getNode(ISD::SRL, DL, MVT::i16,
8669 DAG.getNode(ISD::AND, DL, MVT::i16,
8670 CWD, DAG.getConstant(0x400, MVT::i16)),
8671 DAG.getConstant(9, MVT::i8));
8674 DAG.getNode(ISD::AND, DL, MVT::i16,
8675 DAG.getNode(ISD::ADD, DL, MVT::i16,
8676 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
8677 DAG.getConstant(1, MVT::i16)),
8678 DAG.getConstant(3, MVT::i16));
8681 return DAG.getNode((VT.getSizeInBits() < 16 ?
8682 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
8685 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
8686 EVT VT = Op.getValueType();
8688 unsigned NumBits = VT.getSizeInBits();
8689 DebugLoc dl = Op.getDebugLoc();
8691 Op = Op.getOperand(0);
8692 if (VT == MVT::i8) {
8693 // Zero extend to i32 since there is not an i8 bsr.
8695 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
8698 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
8699 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
8700 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
8702 // If src is zero (i.e. bsr sets ZF), returns NumBits.
8705 DAG.getConstant(NumBits+NumBits-1, OpVT),
8706 DAG.getConstant(X86::COND_E, MVT::i8),
8709 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
8711 // Finally xor with NumBits-1.
8712 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
8715 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
8719 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
8720 EVT VT = Op.getValueType();
8722 unsigned NumBits = VT.getSizeInBits();
8723 DebugLoc dl = Op.getDebugLoc();
8725 Op = Op.getOperand(0);
8726 if (VT == MVT::i8) {
8728 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
8731 // Issue a bsf (scan bits forward) which also sets EFLAGS.
8732 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
8733 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
8735 // If src is zero (i.e. bsf sets ZF), returns NumBits.
8738 DAG.getConstant(NumBits, OpVT),
8739 DAG.getConstant(X86::COND_E, MVT::i8),
8742 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
8745 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
8749 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
8750 EVT VT = Op.getValueType();
8751 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
8752 DebugLoc dl = Op.getDebugLoc();
8754 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
8755 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
8756 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
8757 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
8758 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
8760 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
8761 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
8762 // return AloBlo + AloBhi + AhiBlo;
8764 SDValue A = Op.getOperand(0);
8765 SDValue B = Op.getOperand(1);
8767 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8768 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8769 A, DAG.getConstant(32, MVT::i32));
8770 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8771 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8772 B, DAG.getConstant(32, MVT::i32));
8773 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8774 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8776 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8777 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8779 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8780 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8782 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8783 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8784 AloBhi, DAG.getConstant(32, MVT::i32));
8785 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8786 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8787 AhiBlo, DAG.getConstant(32, MVT::i32));
8788 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
8789 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
8793 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
8795 EVT VT = Op.getValueType();
8796 DebugLoc dl = Op.getDebugLoc();
8797 SDValue R = Op.getOperand(0);
8798 SDValue Amt = Op.getOperand(1);
8800 LLVMContext *Context = DAG.getContext();
8803 if (!Subtarget->hasSSE2()) return SDValue();
8805 // Optimize shl/srl/sra with constant shift amount.
8806 if (isSplatVector(Amt.getNode())) {
8807 SDValue SclrAmt = Amt->getOperand(0);
8808 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(SclrAmt)) {
8809 uint64_t ShiftAmt = C->getZExtValue();
8811 if (VT == MVT::v2i64 && Op.getOpcode() == ISD::SHL)
8812 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8813 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8814 R, DAG.getConstant(ShiftAmt, MVT::i32));
8816 if (VT == MVT::v4i32 && Op.getOpcode() == ISD::SHL)
8817 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8818 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8819 R, DAG.getConstant(ShiftAmt, MVT::i32));
8821 if (VT == MVT::v8i16 && Op.getOpcode() == ISD::SHL)
8822 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8823 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8824 R, DAG.getConstant(ShiftAmt, MVT::i32));
8826 if (VT == MVT::v2i64 && Op.getOpcode() == ISD::SRL)
8827 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8828 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8829 R, DAG.getConstant(ShiftAmt, MVT::i32));
8831 if (VT == MVT::v4i32 && Op.getOpcode() == ISD::SRL)
8832 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8833 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
8834 R, DAG.getConstant(ShiftAmt, MVT::i32));
8836 if (VT == MVT::v8i16 && Op.getOpcode() == ISD::SRL)
8837 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8838 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
8839 R, DAG.getConstant(ShiftAmt, MVT::i32));
8841 if (VT == MVT::v4i32 && Op.getOpcode() == ISD::SRA)
8842 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8843 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
8844 R, DAG.getConstant(ShiftAmt, MVT::i32));
8846 if (VT == MVT::v8i16 && Op.getOpcode() == ISD::SRA)
8847 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8848 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
8849 R, DAG.getConstant(ShiftAmt, MVT::i32));
8853 // Lower SHL with variable shift amount.
8854 // Cannot lower SHL without SSE4.1 or later.
8855 if (!Subtarget->hasSSE41()) return SDValue();
8857 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
8858 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8859 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8860 Op.getOperand(1), DAG.getConstant(23, MVT::i32));
8862 ConstantInt *CI = ConstantInt::get(*Context, APInt(32, 0x3f800000U));
8864 std::vector<Constant*> CV(4, CI);
8865 Constant *C = ConstantVector::get(CV);
8866 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8867 SDValue Addend = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8868 MachinePointerInfo::getConstantPool(),
8871 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Addend);
8872 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
8873 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
8874 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
8876 if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) {
8878 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8879 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8880 Op.getOperand(1), DAG.getConstant(5, MVT::i32));
8882 ConstantInt *CM1 = ConstantInt::get(*Context, APInt(8, 15));
8883 ConstantInt *CM2 = ConstantInt::get(*Context, APInt(8, 63));
8885 std::vector<Constant*> CVM1(16, CM1);
8886 std::vector<Constant*> CVM2(16, CM2);
8887 Constant *C = ConstantVector::get(CVM1);
8888 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8889 SDValue M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8890 MachinePointerInfo::getConstantPool(),
8893 // r = pblendv(r, psllw(r & (char16)15, 4), a);
8894 M = DAG.getNode(ISD::AND, dl, VT, R, M);
8895 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8896 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
8897 DAG.getConstant(4, MVT::i32));
8898 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
8900 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
8902 C = ConstantVector::get(CVM2);
8903 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8904 M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8905 MachinePointerInfo::getConstantPool(),
8908 // r = pblendv(r, psllw(r & (char16)63, 2), a);
8909 M = DAG.getNode(ISD::AND, dl, VT, R, M);
8910 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8911 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
8912 DAG.getConstant(2, MVT::i32));
8913 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
8915 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
8917 // return pblendv(r, r+r, a);
8918 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT,
8919 R, DAG.getNode(ISD::ADD, dl, VT, R, R), Op);
8925 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
8926 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
8927 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
8928 // looks for this combo and may remove the "setcc" instruction if the "setcc"
8929 // has only one use.
8930 SDNode *N = Op.getNode();
8931 SDValue LHS = N->getOperand(0);
8932 SDValue RHS = N->getOperand(1);
8933 unsigned BaseOp = 0;
8935 DebugLoc DL = Op.getDebugLoc();
8936 switch (Op.getOpcode()) {
8937 default: llvm_unreachable("Unknown ovf instruction!");
8939 // A subtract of one will be selected as a INC. Note that INC doesn't
8940 // set CF, so we can't do this for UADDO.
8941 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
8943 BaseOp = X86ISD::INC;
8947 BaseOp = X86ISD::ADD;
8951 BaseOp = X86ISD::ADD;
8955 // A subtract of one will be selected as a DEC. Note that DEC doesn't
8956 // set CF, so we can't do this for USUBO.
8957 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
8959 BaseOp = X86ISD::DEC;
8963 BaseOp = X86ISD::SUB;
8967 BaseOp = X86ISD::SUB;
8971 BaseOp = X86ISD::SMUL;
8974 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
8975 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
8977 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
8980 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8981 DAG.getConstant(X86::COND_O, MVT::i32),
8982 SDValue(Sum.getNode(), 2));
8984 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
8989 // Also sets EFLAGS.
8990 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
8991 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
8994 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
8995 DAG.getConstant(Cond, MVT::i32),
8996 SDValue(Sum.getNode(), 1));
8998 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
9002 SDValue X86TargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const{
9003 DebugLoc dl = Op.getDebugLoc();
9005 if (!Subtarget->hasSSE2()) {
9006 SDValue Chain = Op.getOperand(0);
9007 SDValue Zero = DAG.getConstant(0,
9008 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
9010 DAG.getRegister(X86::ESP, MVT::i32), // Base
9011 DAG.getTargetConstant(1, MVT::i8), // Scale
9012 DAG.getRegister(0, MVT::i32), // Index
9013 DAG.getTargetConstant(0, MVT::i32), // Disp
9014 DAG.getRegister(0, MVT::i32), // Segment.
9019 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
9020 array_lengthof(Ops));
9021 return SDValue(Res, 0);
9024 unsigned isDev = cast<ConstantSDNode>(Op.getOperand(5))->getZExtValue();
9026 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
9028 unsigned Op1 = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
9029 unsigned Op2 = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
9030 unsigned Op3 = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
9031 unsigned Op4 = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
9033 // def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
9034 if (!Op1 && !Op2 && !Op3 && Op4)
9035 return DAG.getNode(X86ISD::SFENCE, dl, MVT::Other, Op.getOperand(0));
9037 // def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
9038 if (Op1 && !Op2 && !Op3 && !Op4)
9039 return DAG.getNode(X86ISD::LFENCE, dl, MVT::Other, Op.getOperand(0));
9041 // def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 1)),
9043 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
9046 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
9047 EVT T = Op.getValueType();
9048 DebugLoc DL = Op.getDebugLoc();
9051 switch(T.getSimpleVT().SimpleTy) {
9053 assert(false && "Invalid value type!");
9054 case MVT::i8: Reg = X86::AL; size = 1; break;
9055 case MVT::i16: Reg = X86::AX; size = 2; break;
9056 case MVT::i32: Reg = X86::EAX; size = 4; break;
9058 assert(Subtarget->is64Bit() && "Node not type legal!");
9059 Reg = X86::RAX; size = 8;
9062 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
9063 Op.getOperand(2), SDValue());
9064 SDValue Ops[] = { cpIn.getValue(0),
9067 DAG.getTargetConstant(size, MVT::i8),
9069 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9070 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
9071 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
9074 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
9078 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
9079 SelectionDAG &DAG) const {
9080 assert(Subtarget->is64Bit() && "Result not type legalized?");
9081 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9082 SDValue TheChain = Op.getOperand(0);
9083 DebugLoc dl = Op.getDebugLoc();
9084 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
9085 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
9086 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
9088 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
9089 DAG.getConstant(32, MVT::i8));
9091 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
9094 return DAG.getMergeValues(Ops, 2, dl);
9097 SDValue X86TargetLowering::LowerBITCAST(SDValue Op,
9098 SelectionDAG &DAG) const {
9099 EVT SrcVT = Op.getOperand(0).getValueType();
9100 EVT DstVT = Op.getValueType();
9101 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
9102 Subtarget->hasMMX() && "Unexpected custom BITCAST");
9103 assert((DstVT == MVT::i64 ||
9104 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
9105 "Unexpected custom BITCAST");
9106 // i64 <=> MMX conversions are Legal.
9107 if (SrcVT==MVT::i64 && DstVT.isVector())
9109 if (DstVT==MVT::i64 && SrcVT.isVector())
9111 // MMX <=> MMX conversions are Legal.
9112 if (SrcVT.isVector() && DstVT.isVector())
9114 // All other conversions need to be expanded.
9118 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
9119 SDNode *Node = Op.getNode();
9120 DebugLoc dl = Node->getDebugLoc();
9121 EVT T = Node->getValueType(0);
9122 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
9123 DAG.getConstant(0, T), Node->getOperand(2));
9124 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
9125 cast<AtomicSDNode>(Node)->getMemoryVT(),
9126 Node->getOperand(0),
9127 Node->getOperand(1), negOp,
9128 cast<AtomicSDNode>(Node)->getSrcValue(),
9129 cast<AtomicSDNode>(Node)->getAlignment());
9132 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
9133 EVT VT = Op.getNode()->getValueType(0);
9135 // Let legalize expand this if it isn't a legal type yet.
9136 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
9139 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
9142 bool ExtraOp = false;
9143 switch (Op.getOpcode()) {
9144 default: assert(0 && "Invalid code");
9145 case ISD::ADDC: Opc = X86ISD::ADD; break;
9146 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
9147 case ISD::SUBC: Opc = X86ISD::SUB; break;
9148 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
9152 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
9154 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
9155 Op.getOperand(1), Op.getOperand(2));
9158 /// LowerOperation - Provide custom lowering hooks for some operations.
9160 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
9161 switch (Op.getOpcode()) {
9162 default: llvm_unreachable("Should not custom lower this!");
9163 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op,DAG);
9164 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
9165 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
9166 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
9167 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
9168 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
9169 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
9170 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
9171 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
9172 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, DAG);
9173 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
9174 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
9175 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
9176 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
9177 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
9178 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
9179 case ISD::SHL_PARTS:
9180 case ISD::SRA_PARTS:
9181 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
9182 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
9183 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
9184 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
9185 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
9186 case ISD::FABS: return LowerFABS(Op, DAG);
9187 case ISD::FNEG: return LowerFNEG(Op, DAG);
9188 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
9189 case ISD::SETCC: return LowerSETCC(Op, DAG);
9190 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
9191 case ISD::SELECT: return LowerSELECT(Op, DAG);
9192 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
9193 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
9194 case ISD::VASTART: return LowerVASTART(Op, DAG);
9195 case ISD::VAARG: return LowerVAARG(Op, DAG);
9196 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
9197 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
9198 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
9199 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
9200 case ISD::FRAME_TO_ARGS_OFFSET:
9201 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
9202 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
9203 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
9204 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
9205 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
9206 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
9207 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
9208 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
9211 case ISD::SHL: return LowerShift(Op, DAG);
9217 case ISD::UMULO: return LowerXALUO(Op, DAG);
9218 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
9219 case ISD::BITCAST: return LowerBITCAST(Op, DAG);
9223 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
9227 void X86TargetLowering::
9228 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
9229 SelectionDAG &DAG, unsigned NewOp) const {
9230 EVT T = Node->getValueType(0);
9231 DebugLoc dl = Node->getDebugLoc();
9232 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
9234 SDValue Chain = Node->getOperand(0);
9235 SDValue In1 = Node->getOperand(1);
9236 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
9237 Node->getOperand(2), DAG.getIntPtrConstant(0));
9238 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
9239 Node->getOperand(2), DAG.getIntPtrConstant(1));
9240 SDValue Ops[] = { Chain, In1, In2L, In2H };
9241 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
9243 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
9244 cast<MemSDNode>(Node)->getMemOperand());
9245 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
9246 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
9247 Results.push_back(Result.getValue(2));
9250 /// ReplaceNodeResults - Replace a node with an illegal result type
9251 /// with a new node built out of custom code.
9252 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
9253 SmallVectorImpl<SDValue>&Results,
9254 SelectionDAG &DAG) const {
9255 DebugLoc dl = N->getDebugLoc();
9256 switch (N->getOpcode()) {
9258 assert(false && "Do not know how to custom type legalize this operation!");
9264 // We don't want to expand or promote these.
9266 case ISD::FP_TO_SINT: {
9267 std::pair<SDValue,SDValue> Vals =
9268 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
9269 SDValue FIST = Vals.first, StackSlot = Vals.second;
9270 if (FIST.getNode() != 0) {
9271 EVT VT = N->getValueType(0);
9272 // Return a load from the stack slot.
9273 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
9274 MachinePointerInfo(), false, false, 0));
9278 case ISD::READCYCLECOUNTER: {
9279 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9280 SDValue TheChain = N->getOperand(0);
9281 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
9282 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
9284 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
9286 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
9287 SDValue Ops[] = { eax, edx };
9288 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
9289 Results.push_back(edx.getValue(1));
9292 case ISD::ATOMIC_CMP_SWAP: {
9293 EVT T = N->getValueType(0);
9294 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
9295 SDValue cpInL, cpInH;
9296 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
9297 DAG.getConstant(0, MVT::i32));
9298 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
9299 DAG.getConstant(1, MVT::i32));
9300 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
9301 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
9303 SDValue swapInL, swapInH;
9304 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
9305 DAG.getConstant(0, MVT::i32));
9306 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
9307 DAG.getConstant(1, MVT::i32));
9308 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
9310 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
9311 swapInL.getValue(1));
9312 SDValue Ops[] = { swapInH.getValue(0),
9314 swapInH.getValue(1) };
9315 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9316 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
9317 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG8_DAG, dl, Tys,
9319 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
9320 MVT::i32, Result.getValue(1));
9321 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
9322 MVT::i32, cpOutL.getValue(2));
9323 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
9324 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
9325 Results.push_back(cpOutH.getValue(1));
9328 case ISD::ATOMIC_LOAD_ADD:
9329 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
9331 case ISD::ATOMIC_LOAD_AND:
9332 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
9334 case ISD::ATOMIC_LOAD_NAND:
9335 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
9337 case ISD::ATOMIC_LOAD_OR:
9338 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
9340 case ISD::ATOMIC_LOAD_SUB:
9341 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
9343 case ISD::ATOMIC_LOAD_XOR:
9344 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
9346 case ISD::ATOMIC_SWAP:
9347 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
9352 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
9354 default: return NULL;
9355 case X86ISD::BSF: return "X86ISD::BSF";
9356 case X86ISD::BSR: return "X86ISD::BSR";
9357 case X86ISD::SHLD: return "X86ISD::SHLD";
9358 case X86ISD::SHRD: return "X86ISD::SHRD";
9359 case X86ISD::FAND: return "X86ISD::FAND";
9360 case X86ISD::FOR: return "X86ISD::FOR";
9361 case X86ISD::FXOR: return "X86ISD::FXOR";
9362 case X86ISD::FSRL: return "X86ISD::FSRL";
9363 case X86ISD::FILD: return "X86ISD::FILD";
9364 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
9365 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
9366 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
9367 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
9368 case X86ISD::FLD: return "X86ISD::FLD";
9369 case X86ISD::FST: return "X86ISD::FST";
9370 case X86ISD::CALL: return "X86ISD::CALL";
9371 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
9372 case X86ISD::BT: return "X86ISD::BT";
9373 case X86ISD::CMP: return "X86ISD::CMP";
9374 case X86ISD::COMI: return "X86ISD::COMI";
9375 case X86ISD::UCOMI: return "X86ISD::UCOMI";
9376 case X86ISD::SETCC: return "X86ISD::SETCC";
9377 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
9378 case X86ISD::CMOV: return "X86ISD::CMOV";
9379 case X86ISD::BRCOND: return "X86ISD::BRCOND";
9380 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
9381 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
9382 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
9383 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
9384 case X86ISD::Wrapper: return "X86ISD::Wrapper";
9385 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
9386 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
9387 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
9388 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
9389 case X86ISD::PINSRB: return "X86ISD::PINSRB";
9390 case X86ISD::PINSRW: return "X86ISD::PINSRW";
9391 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
9392 case X86ISD::PANDN: return "X86ISD::PANDN";
9393 case X86ISD::PSIGNB: return "X86ISD::PSIGNB";
9394 case X86ISD::PSIGNW: return "X86ISD::PSIGNW";
9395 case X86ISD::PSIGND: return "X86ISD::PSIGND";
9396 case X86ISD::PBLENDVB: return "X86ISD::PBLENDVB";
9397 case X86ISD::FMAX: return "X86ISD::FMAX";
9398 case X86ISD::FMIN: return "X86ISD::FMIN";
9399 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
9400 case X86ISD::FRCP: return "X86ISD::FRCP";
9401 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
9402 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
9403 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
9404 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
9405 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
9406 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
9407 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
9408 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
9409 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
9410 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
9411 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
9412 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
9413 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
9414 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
9415 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
9416 case X86ISD::VSHL: return "X86ISD::VSHL";
9417 case X86ISD::VSRL: return "X86ISD::VSRL";
9418 case X86ISD::CMPPD: return "X86ISD::CMPPD";
9419 case X86ISD::CMPPS: return "X86ISD::CMPPS";
9420 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
9421 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
9422 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
9423 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
9424 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
9425 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
9426 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
9427 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
9428 case X86ISD::ADD: return "X86ISD::ADD";
9429 case X86ISD::SUB: return "X86ISD::SUB";
9430 case X86ISD::ADC: return "X86ISD::ADC";
9431 case X86ISD::SBB: return "X86ISD::SBB";
9432 case X86ISD::SMUL: return "X86ISD::SMUL";
9433 case X86ISD::UMUL: return "X86ISD::UMUL";
9434 case X86ISD::INC: return "X86ISD::INC";
9435 case X86ISD::DEC: return "X86ISD::DEC";
9436 case X86ISD::OR: return "X86ISD::OR";
9437 case X86ISD::XOR: return "X86ISD::XOR";
9438 case X86ISD::AND: return "X86ISD::AND";
9439 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
9440 case X86ISD::PTEST: return "X86ISD::PTEST";
9441 case X86ISD::TESTP: return "X86ISD::TESTP";
9442 case X86ISD::PALIGN: return "X86ISD::PALIGN";
9443 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
9444 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
9445 case X86ISD::PSHUFHW_LD: return "X86ISD::PSHUFHW_LD";
9446 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
9447 case X86ISD::PSHUFLW_LD: return "X86ISD::PSHUFLW_LD";
9448 case X86ISD::SHUFPS: return "X86ISD::SHUFPS";
9449 case X86ISD::SHUFPD: return "X86ISD::SHUFPD";
9450 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
9451 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
9452 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
9453 case X86ISD::MOVHLPD: return "X86ISD::MOVHLPD";
9454 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
9455 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
9456 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
9457 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
9458 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
9459 case X86ISD::MOVSHDUP_LD: return "X86ISD::MOVSHDUP_LD";
9460 case X86ISD::MOVSLDUP_LD: return "X86ISD::MOVSLDUP_LD";
9461 case X86ISD::MOVSD: return "X86ISD::MOVSD";
9462 case X86ISD::MOVSS: return "X86ISD::MOVSS";
9463 case X86ISD::UNPCKLPS: return "X86ISD::UNPCKLPS";
9464 case X86ISD::UNPCKLPD: return "X86ISD::UNPCKLPD";
9465 case X86ISD::VUNPCKLPS: return "X86ISD::VUNPCKLPS";
9466 case X86ISD::VUNPCKLPD: return "X86ISD::VUNPCKLPD";
9467 case X86ISD::VUNPCKLPSY: return "X86ISD::VUNPCKLPSY";
9468 case X86ISD::VUNPCKLPDY: return "X86ISD::VUNPCKLPDY";
9469 case X86ISD::UNPCKHPS: return "X86ISD::UNPCKHPS";
9470 case X86ISD::UNPCKHPD: return "X86ISD::UNPCKHPD";
9471 case X86ISD::PUNPCKLBW: return "X86ISD::PUNPCKLBW";
9472 case X86ISD::PUNPCKLWD: return "X86ISD::PUNPCKLWD";
9473 case X86ISD::PUNPCKLDQ: return "X86ISD::PUNPCKLDQ";
9474 case X86ISD::PUNPCKLQDQ: return "X86ISD::PUNPCKLQDQ";
9475 case X86ISD::PUNPCKHBW: return "X86ISD::PUNPCKHBW";
9476 case X86ISD::PUNPCKHWD: return "X86ISD::PUNPCKHWD";
9477 case X86ISD::PUNPCKHDQ: return "X86ISD::PUNPCKHDQ";
9478 case X86ISD::PUNPCKHQDQ: return "X86ISD::PUNPCKHQDQ";
9479 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
9480 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
9481 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
9485 // isLegalAddressingMode - Return true if the addressing mode represented
9486 // by AM is legal for this target, for a load/store of the specified type.
9487 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
9488 const Type *Ty) const {
9489 // X86 supports extremely general addressing modes.
9490 CodeModel::Model M = getTargetMachine().getCodeModel();
9491 Reloc::Model R = getTargetMachine().getRelocationModel();
9493 // X86 allows a sign-extended 32-bit immediate field as a displacement.
9494 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
9499 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
9501 // If a reference to this global requires an extra load, we can't fold it.
9502 if (isGlobalStubReference(GVFlags))
9505 // If BaseGV requires a register for the PIC base, we cannot also have a
9506 // BaseReg specified.
9507 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
9510 // If lower 4G is not available, then we must use rip-relative addressing.
9511 if ((M != CodeModel::Small || R != Reloc::Static) &&
9512 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
9522 // These scales always work.
9527 // These scales are formed with basereg+scalereg. Only accept if there is
9532 default: // Other stuff never works.
9540 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
9541 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
9543 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
9544 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
9545 if (NumBits1 <= NumBits2)
9550 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
9551 if (!VT1.isInteger() || !VT2.isInteger())
9553 unsigned NumBits1 = VT1.getSizeInBits();
9554 unsigned NumBits2 = VT2.getSizeInBits();
9555 if (NumBits1 <= NumBits2)
9560 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
9561 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
9562 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
9565 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
9566 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
9567 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
9570 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
9571 // i16 instructions are longer (0x66 prefix) and potentially slower.
9572 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
9575 /// isShuffleMaskLegal - Targets can use this to indicate that they only
9576 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
9577 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
9578 /// are assumed to be legal.
9580 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
9582 // Very little shuffling can be done for 64-bit vectors right now.
9583 if (VT.getSizeInBits() == 64)
9584 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
9586 // FIXME: pshufb, blends, shifts.
9587 return (VT.getVectorNumElements() == 2 ||
9588 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
9589 isMOVLMask(M, VT) ||
9590 isSHUFPMask(M, VT) ||
9591 isPSHUFDMask(M, VT) ||
9592 isPSHUFHWMask(M, VT) ||
9593 isPSHUFLWMask(M, VT) ||
9594 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
9595 isUNPCKLMask(M, VT) ||
9596 isUNPCKHMask(M, VT) ||
9597 isUNPCKL_v_undef_Mask(M, VT) ||
9598 isUNPCKH_v_undef_Mask(M, VT));
9602 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
9604 unsigned NumElts = VT.getVectorNumElements();
9605 // FIXME: This collection of masks seems suspect.
9608 if (NumElts == 4 && VT.getSizeInBits() == 128) {
9609 return (isMOVLMask(Mask, VT) ||
9610 isCommutedMOVLMask(Mask, VT, true) ||
9611 isSHUFPMask(Mask, VT) ||
9612 isCommutedSHUFPMask(Mask, VT));
9617 //===----------------------------------------------------------------------===//
9618 // X86 Scheduler Hooks
9619 //===----------------------------------------------------------------------===//
9621 // private utility function
9623 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
9624 MachineBasicBlock *MBB,
9631 TargetRegisterClass *RC,
9632 bool invSrc) const {
9633 // For the atomic bitwise operator, we generate
9636 // ld t1 = [bitinstr.addr]
9637 // op t2 = t1, [bitinstr.val]
9639 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
9641 // fallthrough -->nextMBB
9642 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9643 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9644 MachineFunction::iterator MBBIter = MBB;
9647 /// First build the CFG
9648 MachineFunction *F = MBB->getParent();
9649 MachineBasicBlock *thisMBB = MBB;
9650 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
9651 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
9652 F->insert(MBBIter, newMBB);
9653 F->insert(MBBIter, nextMBB);
9655 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
9656 nextMBB->splice(nextMBB->begin(), thisMBB,
9657 llvm::next(MachineBasicBlock::iterator(bInstr)),
9659 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
9661 // Update thisMBB to fall through to newMBB
9662 thisMBB->addSuccessor(newMBB);
9664 // newMBB jumps to itself and fall through to nextMBB
9665 newMBB->addSuccessor(nextMBB);
9666 newMBB->addSuccessor(newMBB);
9668 // Insert instructions into newMBB based on incoming instruction
9669 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
9670 "unexpected number of operands");
9671 DebugLoc dl = bInstr->getDebugLoc();
9672 MachineOperand& destOper = bInstr->getOperand(0);
9673 MachineOperand* argOpers[2 + X86::AddrNumOperands];
9674 int numArgs = bInstr->getNumOperands() - 1;
9675 for (int i=0; i < numArgs; ++i)
9676 argOpers[i] = &bInstr->getOperand(i+1);
9678 // x86 address has 4 operands: base, index, scale, and displacement
9679 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
9680 int valArgIndx = lastAddrIndx + 1;
9682 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
9683 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
9684 for (int i=0; i <= lastAddrIndx; ++i)
9685 (*MIB).addOperand(*argOpers[i]);
9687 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
9689 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
9694 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
9695 assert((argOpers[valArgIndx]->isReg() ||
9696 argOpers[valArgIndx]->isImm()) &&
9698 if (argOpers[valArgIndx]->isReg())
9699 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
9701 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
9703 (*MIB).addOperand(*argOpers[valArgIndx]);
9705 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), EAXreg);
9708 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
9709 for (int i=0; i <= lastAddrIndx; ++i)
9710 (*MIB).addOperand(*argOpers[i]);
9712 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
9713 (*MIB).setMemRefs(bInstr->memoperands_begin(),
9714 bInstr->memoperands_end());
9716 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
9720 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
9722 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
9726 // private utility function: 64 bit atomics on 32 bit host.
9728 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
9729 MachineBasicBlock *MBB,
9734 bool invSrc) const {
9735 // For the atomic bitwise operator, we generate
9736 // thisMBB (instructions are in pairs, except cmpxchg8b)
9737 // ld t1,t2 = [bitinstr.addr]
9739 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
9740 // op t5, t6 <- out1, out2, [bitinstr.val]
9741 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
9742 // mov ECX, EBX <- t5, t6
9743 // mov EAX, EDX <- t1, t2
9744 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
9745 // mov t3, t4 <- EAX, EDX
9747 // result in out1, out2
9748 // fallthrough -->nextMBB
9750 const TargetRegisterClass *RC = X86::GR32RegisterClass;
9751 const unsigned LoadOpc = X86::MOV32rm;
9752 const unsigned NotOpc = X86::NOT32r;
9753 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9754 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9755 MachineFunction::iterator MBBIter = MBB;
9758 /// First build the CFG
9759 MachineFunction *F = MBB->getParent();
9760 MachineBasicBlock *thisMBB = MBB;
9761 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
9762 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
9763 F->insert(MBBIter, newMBB);
9764 F->insert(MBBIter, nextMBB);
9766 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
9767 nextMBB->splice(nextMBB->begin(), thisMBB,
9768 llvm::next(MachineBasicBlock::iterator(bInstr)),
9770 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
9772 // Update thisMBB to fall through to newMBB
9773 thisMBB->addSuccessor(newMBB);
9775 // newMBB jumps to itself and fall through to nextMBB
9776 newMBB->addSuccessor(nextMBB);
9777 newMBB->addSuccessor(newMBB);
9779 DebugLoc dl = bInstr->getDebugLoc();
9780 // Insert instructions into newMBB based on incoming instruction
9781 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
9782 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 14 &&
9783 "unexpected number of operands");
9784 MachineOperand& dest1Oper = bInstr->getOperand(0);
9785 MachineOperand& dest2Oper = bInstr->getOperand(1);
9786 MachineOperand* argOpers[2 + X86::AddrNumOperands];
9787 for (int i=0; i < 2 + X86::AddrNumOperands; ++i) {
9788 argOpers[i] = &bInstr->getOperand(i+2);
9790 // We use some of the operands multiple times, so conservatively just
9791 // clear any kill flags that might be present.
9792 if (argOpers[i]->isReg() && argOpers[i]->isUse())
9793 argOpers[i]->setIsKill(false);
9796 // x86 address has 5 operands: base, index, scale, displacement, and segment.
9797 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
9799 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
9800 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
9801 for (int i=0; i <= lastAddrIndx; ++i)
9802 (*MIB).addOperand(*argOpers[i]);
9803 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
9804 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
9805 // add 4 to displacement.
9806 for (int i=0; i <= lastAddrIndx-2; ++i)
9807 (*MIB).addOperand(*argOpers[i]);
9808 MachineOperand newOp3 = *(argOpers[3]);
9810 newOp3.setImm(newOp3.getImm()+4);
9812 newOp3.setOffset(newOp3.getOffset()+4);
9813 (*MIB).addOperand(newOp3);
9814 (*MIB).addOperand(*argOpers[lastAddrIndx]);
9816 // t3/4 are defined later, at the bottom of the loop
9817 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
9818 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
9819 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
9820 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
9821 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
9822 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
9824 // The subsequent operations should be using the destination registers of
9825 //the PHI instructions.
9827 t1 = F->getRegInfo().createVirtualRegister(RC);
9828 t2 = F->getRegInfo().createVirtualRegister(RC);
9829 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
9830 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
9832 t1 = dest1Oper.getReg();
9833 t2 = dest2Oper.getReg();
9836 int valArgIndx = lastAddrIndx + 1;
9837 assert((argOpers[valArgIndx]->isReg() ||
9838 argOpers[valArgIndx]->isImm()) &&
9840 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
9841 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
9842 if (argOpers[valArgIndx]->isReg())
9843 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
9845 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
9846 if (regOpcL != X86::MOV32rr)
9848 (*MIB).addOperand(*argOpers[valArgIndx]);
9849 assert(argOpers[valArgIndx + 1]->isReg() ==
9850 argOpers[valArgIndx]->isReg());
9851 assert(argOpers[valArgIndx + 1]->isImm() ==
9852 argOpers[valArgIndx]->isImm());
9853 if (argOpers[valArgIndx + 1]->isReg())
9854 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
9856 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
9857 if (regOpcH != X86::MOV32rr)
9859 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
9861 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
9863 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EDX);
9866 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EBX);
9868 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::ECX);
9871 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
9872 for (int i=0; i <= lastAddrIndx; ++i)
9873 (*MIB).addOperand(*argOpers[i]);
9875 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
9876 (*MIB).setMemRefs(bInstr->memoperands_begin(),
9877 bInstr->memoperands_end());
9879 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t3);
9880 MIB.addReg(X86::EAX);
9881 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t4);
9882 MIB.addReg(X86::EDX);
9885 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
9887 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
9891 // private utility function
9893 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
9894 MachineBasicBlock *MBB,
9895 unsigned cmovOpc) const {
9896 // For the atomic min/max operator, we generate
9899 // ld t1 = [min/max.addr]
9900 // mov t2 = [min/max.val]
9902 // cmov[cond] t2 = t1
9904 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
9906 // fallthrough -->nextMBB
9908 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9909 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9910 MachineFunction::iterator MBBIter = MBB;
9913 /// First build the CFG
9914 MachineFunction *F = MBB->getParent();
9915 MachineBasicBlock *thisMBB = MBB;
9916 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
9917 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
9918 F->insert(MBBIter, newMBB);
9919 F->insert(MBBIter, nextMBB);
9921 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
9922 nextMBB->splice(nextMBB->begin(), thisMBB,
9923 llvm::next(MachineBasicBlock::iterator(mInstr)),
9925 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
9927 // Update thisMBB to fall through to newMBB
9928 thisMBB->addSuccessor(newMBB);
9930 // newMBB jumps to newMBB and fall through to nextMBB
9931 newMBB->addSuccessor(nextMBB);
9932 newMBB->addSuccessor(newMBB);
9934 DebugLoc dl = mInstr->getDebugLoc();
9935 // Insert instructions into newMBB based on incoming instruction
9936 assert(mInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
9937 "unexpected number of operands");
9938 MachineOperand& destOper = mInstr->getOperand(0);
9939 MachineOperand* argOpers[2 + X86::AddrNumOperands];
9940 int numArgs = mInstr->getNumOperands() - 1;
9941 for (int i=0; i < numArgs; ++i)
9942 argOpers[i] = &mInstr->getOperand(i+1);
9944 // x86 address has 4 operands: base, index, scale, and displacement
9945 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
9946 int valArgIndx = lastAddrIndx + 1;
9948 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9949 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
9950 for (int i=0; i <= lastAddrIndx; ++i)
9951 (*MIB).addOperand(*argOpers[i]);
9953 // We only support register and immediate values
9954 assert((argOpers[valArgIndx]->isReg() ||
9955 argOpers[valArgIndx]->isImm()) &&
9958 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9959 if (argOpers[valArgIndx]->isReg())
9960 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t2);
9962 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
9963 (*MIB).addOperand(*argOpers[valArgIndx]);
9965 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
9968 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
9973 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9974 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
9978 // Cmp and exchange if none has modified the memory location
9979 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
9980 for (int i=0; i <= lastAddrIndx; ++i)
9981 (*MIB).addOperand(*argOpers[i]);
9983 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
9984 (*MIB).setMemRefs(mInstr->memoperands_begin(),
9985 mInstr->memoperands_end());
9987 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
9988 MIB.addReg(X86::EAX);
9991 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
9993 mInstr->eraseFromParent(); // The pseudo instruction is gone now.
9997 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
9998 // or XMM0_V32I8 in AVX all of this code can be replaced with that
10000 MachineBasicBlock *
10001 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
10002 unsigned numArgs, bool memArg) const {
10003 assert((Subtarget->hasSSE42() || Subtarget->hasAVX()) &&
10004 "Target must have SSE4.2 or AVX features enabled");
10006 DebugLoc dl = MI->getDebugLoc();
10007 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10009 if (!Subtarget->hasAVX()) {
10011 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
10013 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
10016 Opc = numArgs == 3 ? X86::VPCMPISTRM128rm : X86::VPCMPESTRM128rm;
10018 Opc = numArgs == 3 ? X86::VPCMPISTRM128rr : X86::VPCMPESTRM128rr;
10021 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
10022 for (unsigned i = 0; i < numArgs; ++i) {
10023 MachineOperand &Op = MI->getOperand(i+1);
10024 if (!(Op.isReg() && Op.isImplicit()))
10025 MIB.addOperand(Op);
10027 BuildMI(*BB, MI, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
10028 .addReg(X86::XMM0);
10030 MI->eraseFromParent();
10034 MachineBasicBlock *
10035 X86TargetLowering::EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB) const {
10036 DebugLoc dl = MI->getDebugLoc();
10037 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10039 // Address into RAX/EAX, other two args into ECX, EDX.
10040 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
10041 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
10042 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
10043 for (int i = 0; i < X86::AddrNumOperands; ++i)
10044 MIB.addOperand(MI->getOperand(i));
10046 unsigned ValOps = X86::AddrNumOperands;
10047 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
10048 .addReg(MI->getOperand(ValOps).getReg());
10049 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
10050 .addReg(MI->getOperand(ValOps+1).getReg());
10052 // The instruction doesn't actually take any operands though.
10053 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
10055 MI->eraseFromParent(); // The pseudo is gone now.
10059 MachineBasicBlock *
10060 X86TargetLowering::EmitMwait(MachineInstr *MI, MachineBasicBlock *BB) const {
10061 DebugLoc dl = MI->getDebugLoc();
10062 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10064 // First arg in ECX, the second in EAX.
10065 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
10066 .addReg(MI->getOperand(0).getReg());
10067 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EAX)
10068 .addReg(MI->getOperand(1).getReg());
10070 // The instruction doesn't actually take any operands though.
10071 BuildMI(*BB, MI, dl, TII->get(X86::MWAITrr));
10073 MI->eraseFromParent(); // The pseudo is gone now.
10077 MachineBasicBlock *
10078 X86TargetLowering::EmitVAARG64WithCustomInserter(
10080 MachineBasicBlock *MBB) const {
10081 // Emit va_arg instruction on X86-64.
10083 // Operands to this pseudo-instruction:
10084 // 0 ) Output : destination address (reg)
10085 // 1-5) Input : va_list address (addr, i64mem)
10086 // 6 ) ArgSize : Size (in bytes) of vararg type
10087 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
10088 // 8 ) Align : Alignment of type
10089 // 9 ) EFLAGS (implicit-def)
10091 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
10092 assert(X86::AddrNumOperands == 5 && "VAARG_64 assumes 5 address operands");
10094 unsigned DestReg = MI->getOperand(0).getReg();
10095 MachineOperand &Base = MI->getOperand(1);
10096 MachineOperand &Scale = MI->getOperand(2);
10097 MachineOperand &Index = MI->getOperand(3);
10098 MachineOperand &Disp = MI->getOperand(4);
10099 MachineOperand &Segment = MI->getOperand(5);
10100 unsigned ArgSize = MI->getOperand(6).getImm();
10101 unsigned ArgMode = MI->getOperand(7).getImm();
10102 unsigned Align = MI->getOperand(8).getImm();
10104 // Memory Reference
10105 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
10106 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
10107 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
10109 // Machine Information
10110 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10111 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
10112 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
10113 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
10114 DebugLoc DL = MI->getDebugLoc();
10116 // struct va_list {
10119 // i64 overflow_area (address)
10120 // i64 reg_save_area (address)
10122 // sizeof(va_list) = 24
10123 // alignment(va_list) = 8
10125 unsigned TotalNumIntRegs = 6;
10126 unsigned TotalNumXMMRegs = 8;
10127 bool UseGPOffset = (ArgMode == 1);
10128 bool UseFPOffset = (ArgMode == 2);
10129 unsigned MaxOffset = TotalNumIntRegs * 8 +
10130 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
10132 /* Align ArgSize to a multiple of 8 */
10133 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
10134 bool NeedsAlign = (Align > 8);
10136 MachineBasicBlock *thisMBB = MBB;
10137 MachineBasicBlock *overflowMBB;
10138 MachineBasicBlock *offsetMBB;
10139 MachineBasicBlock *endMBB;
10141 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
10142 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
10143 unsigned OffsetReg = 0;
10145 if (!UseGPOffset && !UseFPOffset) {
10146 // If we only pull from the overflow region, we don't create a branch.
10147 // We don't need to alter control flow.
10148 OffsetDestReg = 0; // unused
10149 OverflowDestReg = DestReg;
10152 overflowMBB = thisMBB;
10155 // First emit code to check if gp_offset (or fp_offset) is below the bound.
10156 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
10157 // If not, pull from overflow_area. (branch to overflowMBB)
10162 // offsetMBB overflowMBB
10167 // Registers for the PHI in endMBB
10168 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
10169 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
10171 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10172 MachineFunction *MF = MBB->getParent();
10173 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
10174 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
10175 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
10177 MachineFunction::iterator MBBIter = MBB;
10180 // Insert the new basic blocks
10181 MF->insert(MBBIter, offsetMBB);
10182 MF->insert(MBBIter, overflowMBB);
10183 MF->insert(MBBIter, endMBB);
10185 // Transfer the remainder of MBB and its successor edges to endMBB.
10186 endMBB->splice(endMBB->begin(), thisMBB,
10187 llvm::next(MachineBasicBlock::iterator(MI)),
10189 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
10191 // Make offsetMBB and overflowMBB successors of thisMBB
10192 thisMBB->addSuccessor(offsetMBB);
10193 thisMBB->addSuccessor(overflowMBB);
10195 // endMBB is a successor of both offsetMBB and overflowMBB
10196 offsetMBB->addSuccessor(endMBB);
10197 overflowMBB->addSuccessor(endMBB);
10199 // Load the offset value into a register
10200 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
10201 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
10205 .addDisp(Disp, UseFPOffset ? 4 : 0)
10206 .addOperand(Segment)
10207 .setMemRefs(MMOBegin, MMOEnd);
10209 // Check if there is enough room left to pull this argument.
10210 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
10212 .addImm(MaxOffset + 8 - ArgSizeA8);
10214 // Branch to "overflowMBB" if offset >= max
10215 // Fall through to "offsetMBB" otherwise
10216 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
10217 .addMBB(overflowMBB);
10220 // In offsetMBB, emit code to use the reg_save_area.
10222 assert(OffsetReg != 0);
10224 // Read the reg_save_area address.
10225 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
10226 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
10231 .addOperand(Segment)
10232 .setMemRefs(MMOBegin, MMOEnd);
10234 // Zero-extend the offset
10235 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
10236 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
10239 .addImm(X86::sub_32bit);
10241 // Add the offset to the reg_save_area to get the final address.
10242 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
10243 .addReg(OffsetReg64)
10244 .addReg(RegSaveReg);
10246 // Compute the offset for the next argument
10247 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
10248 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
10250 .addImm(UseFPOffset ? 16 : 8);
10252 // Store it back into the va_list.
10253 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
10257 .addDisp(Disp, UseFPOffset ? 4 : 0)
10258 .addOperand(Segment)
10259 .addReg(NextOffsetReg)
10260 .setMemRefs(MMOBegin, MMOEnd);
10263 BuildMI(offsetMBB, DL, TII->get(X86::JMP_4))
10268 // Emit code to use overflow area
10271 // Load the overflow_area address into a register.
10272 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
10273 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
10278 .addOperand(Segment)
10279 .setMemRefs(MMOBegin, MMOEnd);
10281 // If we need to align it, do so. Otherwise, just copy the address
10282 // to OverflowDestReg.
10284 // Align the overflow address
10285 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
10286 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
10288 // aligned_addr = (addr + (align-1)) & ~(align-1)
10289 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
10290 .addReg(OverflowAddrReg)
10293 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
10295 .addImm(~(uint64_t)(Align-1));
10297 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
10298 .addReg(OverflowAddrReg);
10301 // Compute the next overflow address after this argument.
10302 // (the overflow address should be kept 8-byte aligned)
10303 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
10304 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
10305 .addReg(OverflowDestReg)
10306 .addImm(ArgSizeA8);
10308 // Store the new overflow address.
10309 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
10314 .addOperand(Segment)
10315 .addReg(NextAddrReg)
10316 .setMemRefs(MMOBegin, MMOEnd);
10318 // If we branched, emit the PHI to the front of endMBB.
10320 BuildMI(*endMBB, endMBB->begin(), DL,
10321 TII->get(X86::PHI), DestReg)
10322 .addReg(OffsetDestReg).addMBB(offsetMBB)
10323 .addReg(OverflowDestReg).addMBB(overflowMBB);
10326 // Erase the pseudo instruction
10327 MI->eraseFromParent();
10332 MachineBasicBlock *
10333 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
10335 MachineBasicBlock *MBB) const {
10336 // Emit code to save XMM registers to the stack. The ABI says that the
10337 // number of registers to save is given in %al, so it's theoretically
10338 // possible to do an indirect jump trick to avoid saving all of them,
10339 // however this code takes a simpler approach and just executes all
10340 // of the stores if %al is non-zero. It's less code, and it's probably
10341 // easier on the hardware branch predictor, and stores aren't all that
10342 // expensive anyway.
10344 // Create the new basic blocks. One block contains all the XMM stores,
10345 // and one block is the final destination regardless of whether any
10346 // stores were performed.
10347 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10348 MachineFunction *F = MBB->getParent();
10349 MachineFunction::iterator MBBIter = MBB;
10351 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
10352 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
10353 F->insert(MBBIter, XMMSaveMBB);
10354 F->insert(MBBIter, EndMBB);
10356 // Transfer the remainder of MBB and its successor edges to EndMBB.
10357 EndMBB->splice(EndMBB->begin(), MBB,
10358 llvm::next(MachineBasicBlock::iterator(MI)),
10360 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
10362 // The original block will now fall through to the XMM save block.
10363 MBB->addSuccessor(XMMSaveMBB);
10364 // The XMMSaveMBB will fall through to the end block.
10365 XMMSaveMBB->addSuccessor(EndMBB);
10367 // Now add the instructions.
10368 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10369 DebugLoc DL = MI->getDebugLoc();
10371 unsigned CountReg = MI->getOperand(0).getReg();
10372 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
10373 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
10375 if (!Subtarget->isTargetWin64()) {
10376 // If %al is 0, branch around the XMM save block.
10377 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
10378 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
10379 MBB->addSuccessor(EndMBB);
10382 // In the XMM save block, save all the XMM argument registers.
10383 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
10384 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
10385 MachineMemOperand *MMO =
10386 F->getMachineMemOperand(
10387 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
10388 MachineMemOperand::MOStore,
10389 /*Size=*/16, /*Align=*/16);
10390 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
10391 .addFrameIndex(RegSaveFrameIndex)
10392 .addImm(/*Scale=*/1)
10393 .addReg(/*IndexReg=*/0)
10394 .addImm(/*Disp=*/Offset)
10395 .addReg(/*Segment=*/0)
10396 .addReg(MI->getOperand(i).getReg())
10397 .addMemOperand(MMO);
10400 MI->eraseFromParent(); // The pseudo instruction is gone now.
10405 MachineBasicBlock *
10406 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
10407 MachineBasicBlock *BB) const {
10408 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10409 DebugLoc DL = MI->getDebugLoc();
10411 // To "insert" a SELECT_CC instruction, we actually have to insert the
10412 // diamond control-flow pattern. The incoming instruction knows the
10413 // destination vreg to set, the condition code register to branch on, the
10414 // true/false values to select between, and a branch opcode to use.
10415 const BasicBlock *LLVM_BB = BB->getBasicBlock();
10416 MachineFunction::iterator It = BB;
10422 // cmpTY ccX, r1, r2
10424 // fallthrough --> copy0MBB
10425 MachineBasicBlock *thisMBB = BB;
10426 MachineFunction *F = BB->getParent();
10427 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
10428 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
10429 F->insert(It, copy0MBB);
10430 F->insert(It, sinkMBB);
10432 // If the EFLAGS register isn't dead in the terminator, then claim that it's
10433 // live into the sink and copy blocks.
10434 const MachineFunction *MF = BB->getParent();
10435 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
10436 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
10438 for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
10439 const MachineOperand &MO = MI->getOperand(I);
10440 if (!MO.isReg() || !MO.isUse() || MO.isKill()) continue;
10441 unsigned Reg = MO.getReg();
10442 if (Reg != X86::EFLAGS) continue;
10443 copy0MBB->addLiveIn(Reg);
10444 sinkMBB->addLiveIn(Reg);
10447 // Transfer the remainder of BB and its successor edges to sinkMBB.
10448 sinkMBB->splice(sinkMBB->begin(), BB,
10449 llvm::next(MachineBasicBlock::iterator(MI)),
10451 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
10453 // Add the true and fallthrough blocks as its successors.
10454 BB->addSuccessor(copy0MBB);
10455 BB->addSuccessor(sinkMBB);
10457 // Create the conditional branch instruction.
10459 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
10460 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
10463 // %FalseValue = ...
10464 // # fallthrough to sinkMBB
10465 copy0MBB->addSuccessor(sinkMBB);
10468 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
10470 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
10471 TII->get(X86::PHI), MI->getOperand(0).getReg())
10472 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
10473 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
10475 MI->eraseFromParent(); // The pseudo instruction is gone now.
10479 MachineBasicBlock *
10480 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
10481 MachineBasicBlock *BB) const {
10482 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10483 DebugLoc DL = MI->getDebugLoc();
10485 assert(!Subtarget->isTargetEnvMacho());
10487 // The lowering is pretty easy: we're just emitting the call to _alloca. The
10488 // non-trivial part is impdef of ESP.
10490 if (Subtarget->isTargetWin64()) {
10491 if (Subtarget->isTargetCygMing()) {
10492 // ___chkstk(Mingw64):
10493 // Clobbers R10, R11, RAX and EFLAGS.
10495 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
10496 .addExternalSymbol("___chkstk")
10497 .addReg(X86::RAX, RegState::Implicit)
10498 .addReg(X86::RSP, RegState::Implicit)
10499 .addReg(X86::RAX, RegState::Define | RegState::Implicit)
10500 .addReg(X86::RSP, RegState::Define | RegState::Implicit)
10501 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
10503 // __chkstk(MSVCRT): does not update stack pointer.
10504 // Clobbers R10, R11 and EFLAGS.
10505 // FIXME: RAX(allocated size) might be reused and not killed.
10506 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
10507 .addExternalSymbol("__chkstk")
10508 .addReg(X86::RAX, RegState::Implicit)
10509 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
10510 // RAX has the offset to subtracted from RSP.
10511 BuildMI(*BB, MI, DL, TII->get(X86::SUB64rr), X86::RSP)
10516 const char *StackProbeSymbol =
10517 Subtarget->isTargetWindows() ? "_chkstk" : "_alloca";
10519 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
10520 .addExternalSymbol(StackProbeSymbol)
10521 .addReg(X86::EAX, RegState::Implicit)
10522 .addReg(X86::ESP, RegState::Implicit)
10523 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
10524 .addReg(X86::ESP, RegState::Define | RegState::Implicit)
10525 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
10528 MI->eraseFromParent(); // The pseudo instruction is gone now.
10532 MachineBasicBlock *
10533 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
10534 MachineBasicBlock *BB) const {
10535 // This is pretty easy. We're taking the value that we received from
10536 // our load from the relocation, sticking it in either RDI (x86-64)
10537 // or EAX and doing an indirect call. The return value will then
10538 // be in the normal return register.
10539 const X86InstrInfo *TII
10540 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
10541 DebugLoc DL = MI->getDebugLoc();
10542 MachineFunction *F = BB->getParent();
10544 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
10545 assert(MI->getOperand(3).isGlobal() && "This should be a global");
10547 if (Subtarget->is64Bit()) {
10548 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
10549 TII->get(X86::MOV64rm), X86::RDI)
10551 .addImm(0).addReg(0)
10552 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
10553 MI->getOperand(3).getTargetFlags())
10555 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
10556 addDirectMem(MIB, X86::RDI);
10557 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
10558 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
10559 TII->get(X86::MOV32rm), X86::EAX)
10561 .addImm(0).addReg(0)
10562 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
10563 MI->getOperand(3).getTargetFlags())
10565 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
10566 addDirectMem(MIB, X86::EAX);
10568 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
10569 TII->get(X86::MOV32rm), X86::EAX)
10570 .addReg(TII->getGlobalBaseReg(F))
10571 .addImm(0).addReg(0)
10572 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
10573 MI->getOperand(3).getTargetFlags())
10575 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
10576 addDirectMem(MIB, X86::EAX);
10579 MI->eraseFromParent(); // The pseudo instruction is gone now.
10583 MachineBasicBlock *
10584 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
10585 MachineBasicBlock *BB) const {
10586 switch (MI->getOpcode()) {
10587 default: assert(false && "Unexpected instr type to insert");
10588 case X86::TAILJMPd64:
10589 case X86::TAILJMPr64:
10590 case X86::TAILJMPm64:
10591 assert(!"TAILJMP64 would not be touched here.");
10592 case X86::TCRETURNdi64:
10593 case X86::TCRETURNri64:
10594 case X86::TCRETURNmi64:
10595 // Defs of TCRETURNxx64 has Win64's callee-saved registers, as subset.
10596 // On AMD64, additional defs should be added before register allocation.
10597 if (!Subtarget->isTargetWin64()) {
10598 MI->addRegisterDefined(X86::RSI);
10599 MI->addRegisterDefined(X86::RDI);
10600 MI->addRegisterDefined(X86::XMM6);
10601 MI->addRegisterDefined(X86::XMM7);
10602 MI->addRegisterDefined(X86::XMM8);
10603 MI->addRegisterDefined(X86::XMM9);
10604 MI->addRegisterDefined(X86::XMM10);
10605 MI->addRegisterDefined(X86::XMM11);
10606 MI->addRegisterDefined(X86::XMM12);
10607 MI->addRegisterDefined(X86::XMM13);
10608 MI->addRegisterDefined(X86::XMM14);
10609 MI->addRegisterDefined(X86::XMM15);
10612 case X86::WIN_ALLOCA:
10613 return EmitLoweredWinAlloca(MI, BB);
10614 case X86::TLSCall_32:
10615 case X86::TLSCall_64:
10616 return EmitLoweredTLSCall(MI, BB);
10617 case X86::CMOV_GR8:
10618 case X86::CMOV_FR32:
10619 case X86::CMOV_FR64:
10620 case X86::CMOV_V4F32:
10621 case X86::CMOV_V2F64:
10622 case X86::CMOV_V2I64:
10623 case X86::CMOV_GR16:
10624 case X86::CMOV_GR32:
10625 case X86::CMOV_RFP32:
10626 case X86::CMOV_RFP64:
10627 case X86::CMOV_RFP80:
10628 return EmitLoweredSelect(MI, BB);
10630 case X86::FP32_TO_INT16_IN_MEM:
10631 case X86::FP32_TO_INT32_IN_MEM:
10632 case X86::FP32_TO_INT64_IN_MEM:
10633 case X86::FP64_TO_INT16_IN_MEM:
10634 case X86::FP64_TO_INT32_IN_MEM:
10635 case X86::FP64_TO_INT64_IN_MEM:
10636 case X86::FP80_TO_INT16_IN_MEM:
10637 case X86::FP80_TO_INT32_IN_MEM:
10638 case X86::FP80_TO_INT64_IN_MEM: {
10639 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10640 DebugLoc DL = MI->getDebugLoc();
10642 // Change the floating point control register to use "round towards zero"
10643 // mode when truncating to an integer value.
10644 MachineFunction *F = BB->getParent();
10645 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
10646 addFrameReference(BuildMI(*BB, MI, DL,
10647 TII->get(X86::FNSTCW16m)), CWFrameIdx);
10649 // Load the old value of the high byte of the control word...
10651 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
10652 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
10655 // Set the high part to be round to zero...
10656 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
10659 // Reload the modified control word now...
10660 addFrameReference(BuildMI(*BB, MI, DL,
10661 TII->get(X86::FLDCW16m)), CWFrameIdx);
10663 // Restore the memory image of control word to original value
10664 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
10667 // Get the X86 opcode to use.
10669 switch (MI->getOpcode()) {
10670 default: llvm_unreachable("illegal opcode!");
10671 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
10672 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
10673 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
10674 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
10675 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
10676 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
10677 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
10678 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
10679 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
10683 MachineOperand &Op = MI->getOperand(0);
10685 AM.BaseType = X86AddressMode::RegBase;
10686 AM.Base.Reg = Op.getReg();
10688 AM.BaseType = X86AddressMode::FrameIndexBase;
10689 AM.Base.FrameIndex = Op.getIndex();
10691 Op = MI->getOperand(1);
10693 AM.Scale = Op.getImm();
10694 Op = MI->getOperand(2);
10696 AM.IndexReg = Op.getImm();
10697 Op = MI->getOperand(3);
10698 if (Op.isGlobal()) {
10699 AM.GV = Op.getGlobal();
10701 AM.Disp = Op.getImm();
10703 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
10704 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
10706 // Reload the original control word now.
10707 addFrameReference(BuildMI(*BB, MI, DL,
10708 TII->get(X86::FLDCW16m)), CWFrameIdx);
10710 MI->eraseFromParent(); // The pseudo instruction is gone now.
10713 // String/text processing lowering.
10714 case X86::PCMPISTRM128REG:
10715 case X86::VPCMPISTRM128REG:
10716 return EmitPCMP(MI, BB, 3, false /* in-mem */);
10717 case X86::PCMPISTRM128MEM:
10718 case X86::VPCMPISTRM128MEM:
10719 return EmitPCMP(MI, BB, 3, true /* in-mem */);
10720 case X86::PCMPESTRM128REG:
10721 case X86::VPCMPESTRM128REG:
10722 return EmitPCMP(MI, BB, 5, false /* in mem */);
10723 case X86::PCMPESTRM128MEM:
10724 case X86::VPCMPESTRM128MEM:
10725 return EmitPCMP(MI, BB, 5, true /* in mem */);
10727 // Thread synchronization.
10729 return EmitMonitor(MI, BB);
10731 return EmitMwait(MI, BB);
10733 // Atomic Lowering.
10734 case X86::ATOMAND32:
10735 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
10736 X86::AND32ri, X86::MOV32rm,
10738 X86::NOT32r, X86::EAX,
10739 X86::GR32RegisterClass);
10740 case X86::ATOMOR32:
10741 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
10742 X86::OR32ri, X86::MOV32rm,
10744 X86::NOT32r, X86::EAX,
10745 X86::GR32RegisterClass);
10746 case X86::ATOMXOR32:
10747 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
10748 X86::XOR32ri, X86::MOV32rm,
10750 X86::NOT32r, X86::EAX,
10751 X86::GR32RegisterClass);
10752 case X86::ATOMNAND32:
10753 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
10754 X86::AND32ri, X86::MOV32rm,
10756 X86::NOT32r, X86::EAX,
10757 X86::GR32RegisterClass, true);
10758 case X86::ATOMMIN32:
10759 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
10760 case X86::ATOMMAX32:
10761 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
10762 case X86::ATOMUMIN32:
10763 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
10764 case X86::ATOMUMAX32:
10765 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
10767 case X86::ATOMAND16:
10768 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
10769 X86::AND16ri, X86::MOV16rm,
10771 X86::NOT16r, X86::AX,
10772 X86::GR16RegisterClass);
10773 case X86::ATOMOR16:
10774 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
10775 X86::OR16ri, X86::MOV16rm,
10777 X86::NOT16r, X86::AX,
10778 X86::GR16RegisterClass);
10779 case X86::ATOMXOR16:
10780 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
10781 X86::XOR16ri, X86::MOV16rm,
10783 X86::NOT16r, X86::AX,
10784 X86::GR16RegisterClass);
10785 case X86::ATOMNAND16:
10786 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
10787 X86::AND16ri, X86::MOV16rm,
10789 X86::NOT16r, X86::AX,
10790 X86::GR16RegisterClass, true);
10791 case X86::ATOMMIN16:
10792 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
10793 case X86::ATOMMAX16:
10794 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
10795 case X86::ATOMUMIN16:
10796 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
10797 case X86::ATOMUMAX16:
10798 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
10800 case X86::ATOMAND8:
10801 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
10802 X86::AND8ri, X86::MOV8rm,
10804 X86::NOT8r, X86::AL,
10805 X86::GR8RegisterClass);
10807 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
10808 X86::OR8ri, X86::MOV8rm,
10810 X86::NOT8r, X86::AL,
10811 X86::GR8RegisterClass);
10812 case X86::ATOMXOR8:
10813 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
10814 X86::XOR8ri, X86::MOV8rm,
10816 X86::NOT8r, X86::AL,
10817 X86::GR8RegisterClass);
10818 case X86::ATOMNAND8:
10819 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
10820 X86::AND8ri, X86::MOV8rm,
10822 X86::NOT8r, X86::AL,
10823 X86::GR8RegisterClass, true);
10824 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
10825 // This group is for 64-bit host.
10826 case X86::ATOMAND64:
10827 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
10828 X86::AND64ri32, X86::MOV64rm,
10830 X86::NOT64r, X86::RAX,
10831 X86::GR64RegisterClass);
10832 case X86::ATOMOR64:
10833 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
10834 X86::OR64ri32, X86::MOV64rm,
10836 X86::NOT64r, X86::RAX,
10837 X86::GR64RegisterClass);
10838 case X86::ATOMXOR64:
10839 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
10840 X86::XOR64ri32, X86::MOV64rm,
10842 X86::NOT64r, X86::RAX,
10843 X86::GR64RegisterClass);
10844 case X86::ATOMNAND64:
10845 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
10846 X86::AND64ri32, X86::MOV64rm,
10848 X86::NOT64r, X86::RAX,
10849 X86::GR64RegisterClass, true);
10850 case X86::ATOMMIN64:
10851 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
10852 case X86::ATOMMAX64:
10853 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
10854 case X86::ATOMUMIN64:
10855 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
10856 case X86::ATOMUMAX64:
10857 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
10859 // This group does 64-bit operations on a 32-bit host.
10860 case X86::ATOMAND6432:
10861 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10862 X86::AND32rr, X86::AND32rr,
10863 X86::AND32ri, X86::AND32ri,
10865 case X86::ATOMOR6432:
10866 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10867 X86::OR32rr, X86::OR32rr,
10868 X86::OR32ri, X86::OR32ri,
10870 case X86::ATOMXOR6432:
10871 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10872 X86::XOR32rr, X86::XOR32rr,
10873 X86::XOR32ri, X86::XOR32ri,
10875 case X86::ATOMNAND6432:
10876 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10877 X86::AND32rr, X86::AND32rr,
10878 X86::AND32ri, X86::AND32ri,
10880 case X86::ATOMADD6432:
10881 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10882 X86::ADD32rr, X86::ADC32rr,
10883 X86::ADD32ri, X86::ADC32ri,
10885 case X86::ATOMSUB6432:
10886 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10887 X86::SUB32rr, X86::SBB32rr,
10888 X86::SUB32ri, X86::SBB32ri,
10890 case X86::ATOMSWAP6432:
10891 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10892 X86::MOV32rr, X86::MOV32rr,
10893 X86::MOV32ri, X86::MOV32ri,
10895 case X86::VASTART_SAVE_XMM_REGS:
10896 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
10898 case X86::VAARG_64:
10899 return EmitVAARG64WithCustomInserter(MI, BB);
10903 //===----------------------------------------------------------------------===//
10904 // X86 Optimization Hooks
10905 //===----------------------------------------------------------------------===//
10907 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
10911 const SelectionDAG &DAG,
10912 unsigned Depth) const {
10913 unsigned Opc = Op.getOpcode();
10914 assert((Opc >= ISD::BUILTIN_OP_END ||
10915 Opc == ISD::INTRINSIC_WO_CHAIN ||
10916 Opc == ISD::INTRINSIC_W_CHAIN ||
10917 Opc == ISD::INTRINSIC_VOID) &&
10918 "Should use MaskedValueIsZero if you don't know whether Op"
10919 " is a target node!");
10921 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
10935 // These nodes' second result is a boolean.
10936 if (Op.getResNo() == 0)
10939 case X86ISD::SETCC:
10940 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
10941 Mask.getBitWidth() - 1);
10946 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
10947 unsigned Depth) const {
10948 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
10949 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
10950 return Op.getValueType().getScalarType().getSizeInBits();
10956 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
10957 /// node is a GlobalAddress + offset.
10958 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
10959 const GlobalValue* &GA,
10960 int64_t &Offset) const {
10961 if (N->getOpcode() == X86ISD::Wrapper) {
10962 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
10963 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
10964 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
10968 return TargetLowering::isGAPlusOffset(N, GA, Offset);
10971 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
10972 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
10973 /// if the load addresses are consecutive, non-overlapping, and in the right
10975 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
10976 TargetLowering::DAGCombinerInfo &DCI) {
10977 DebugLoc dl = N->getDebugLoc();
10978 EVT VT = N->getValueType(0);
10980 if (VT.getSizeInBits() != 128)
10983 // Don't create instructions with illegal types after legalize types has run.
10984 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
10985 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
10988 SmallVector<SDValue, 16> Elts;
10989 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
10990 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
10992 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
10995 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
10996 /// generation and convert it from being a bunch of shuffles and extracts
10997 /// to a simple store and scalar loads to extract the elements.
10998 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
10999 const TargetLowering &TLI) {
11000 SDValue InputVector = N->getOperand(0);
11002 // Only operate on vectors of 4 elements, where the alternative shuffling
11003 // gets to be more expensive.
11004 if (InputVector.getValueType() != MVT::v4i32)
11007 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
11008 // single use which is a sign-extend or zero-extend, and all elements are
11010 SmallVector<SDNode *, 4> Uses;
11011 unsigned ExtractedElements = 0;
11012 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
11013 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
11014 if (UI.getUse().getResNo() != InputVector.getResNo())
11017 SDNode *Extract = *UI;
11018 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
11021 if (Extract->getValueType(0) != MVT::i32)
11023 if (!Extract->hasOneUse())
11025 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
11026 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
11028 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
11031 // Record which element was extracted.
11032 ExtractedElements |=
11033 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
11035 Uses.push_back(Extract);
11038 // If not all the elements were used, this may not be worthwhile.
11039 if (ExtractedElements != 15)
11042 // Ok, we've now decided to do the transformation.
11043 DebugLoc dl = InputVector.getDebugLoc();
11045 // Store the value to a temporary stack slot.
11046 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
11047 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
11048 MachinePointerInfo(), false, false, 0);
11050 // Replace each use (extract) with a load of the appropriate element.
11051 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
11052 UE = Uses.end(); UI != UE; ++UI) {
11053 SDNode *Extract = *UI;
11055 // Compute the element's address.
11056 SDValue Idx = Extract->getOperand(1);
11058 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
11059 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
11060 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
11062 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(),
11063 StackPtr, OffsetVal);
11065 // Load the scalar.
11066 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
11067 ScalarAddr, MachinePointerInfo(),
11070 // Replace the exact with the load.
11071 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
11074 // The replacement was made in place; don't return anything.
11078 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
11079 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
11080 const X86Subtarget *Subtarget) {
11081 DebugLoc DL = N->getDebugLoc();
11082 SDValue Cond = N->getOperand(0);
11083 // Get the LHS/RHS of the select.
11084 SDValue LHS = N->getOperand(1);
11085 SDValue RHS = N->getOperand(2);
11087 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
11088 // instructions match the semantics of the common C idiom x<y?x:y but not
11089 // x<=y?x:y, because of how they handle negative zero (which can be
11090 // ignored in unsafe-math mode).
11091 if (Subtarget->hasSSE2() &&
11092 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
11093 Cond.getOpcode() == ISD::SETCC) {
11094 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
11096 unsigned Opcode = 0;
11097 // Check for x CC y ? x : y.
11098 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
11099 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
11103 // Converting this to a min would handle NaNs incorrectly, and swapping
11104 // the operands would cause it to handle comparisons between positive
11105 // and negative zero incorrectly.
11106 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
11107 if (!UnsafeFPMath &&
11108 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
11110 std::swap(LHS, RHS);
11112 Opcode = X86ISD::FMIN;
11115 // Converting this to a min would handle comparisons between positive
11116 // and negative zero incorrectly.
11117 if (!UnsafeFPMath &&
11118 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
11120 Opcode = X86ISD::FMIN;
11123 // Converting this to a min would handle both negative zeros and NaNs
11124 // incorrectly, but we can swap the operands to fix both.
11125 std::swap(LHS, RHS);
11129 Opcode = X86ISD::FMIN;
11133 // Converting this to a max would handle comparisons between positive
11134 // and negative zero incorrectly.
11135 if (!UnsafeFPMath &&
11136 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
11138 Opcode = X86ISD::FMAX;
11141 // Converting this to a max would handle NaNs incorrectly, and swapping
11142 // the operands would cause it to handle comparisons between positive
11143 // and negative zero incorrectly.
11144 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
11145 if (!UnsafeFPMath &&
11146 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
11148 std::swap(LHS, RHS);
11150 Opcode = X86ISD::FMAX;
11153 // Converting this to a max would handle both negative zeros and NaNs
11154 // incorrectly, but we can swap the operands to fix both.
11155 std::swap(LHS, RHS);
11159 Opcode = X86ISD::FMAX;
11162 // Check for x CC y ? y : x -- a min/max with reversed arms.
11163 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
11164 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
11168 // Converting this to a min would handle comparisons between positive
11169 // and negative zero incorrectly, and swapping the operands would
11170 // cause it to handle NaNs incorrectly.
11171 if (!UnsafeFPMath &&
11172 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
11173 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
11175 std::swap(LHS, RHS);
11177 Opcode = X86ISD::FMIN;
11180 // Converting this to a min would handle NaNs incorrectly.
11181 if (!UnsafeFPMath &&
11182 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
11184 Opcode = X86ISD::FMIN;
11187 // Converting this to a min would handle both negative zeros and NaNs
11188 // incorrectly, but we can swap the operands to fix both.
11189 std::swap(LHS, RHS);
11193 Opcode = X86ISD::FMIN;
11197 // Converting this to a max would handle NaNs incorrectly.
11198 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
11200 Opcode = X86ISD::FMAX;
11203 // Converting this to a max would handle comparisons between positive
11204 // and negative zero incorrectly, and swapping the operands would
11205 // cause it to handle NaNs incorrectly.
11206 if (!UnsafeFPMath &&
11207 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
11208 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
11210 std::swap(LHS, RHS);
11212 Opcode = X86ISD::FMAX;
11215 // Converting this to a max would handle both negative zeros and NaNs
11216 // incorrectly, but we can swap the operands to fix both.
11217 std::swap(LHS, RHS);
11221 Opcode = X86ISD::FMAX;
11227 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
11230 // If this is a select between two integer constants, try to do some
11232 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
11233 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
11234 // Don't do this for crazy integer types.
11235 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
11236 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
11237 // so that TrueC (the true value) is larger than FalseC.
11238 bool NeedsCondInvert = false;
11240 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
11241 // Efficiently invertible.
11242 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
11243 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
11244 isa<ConstantSDNode>(Cond.getOperand(1))))) {
11245 NeedsCondInvert = true;
11246 std::swap(TrueC, FalseC);
11249 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
11250 if (FalseC->getAPIntValue() == 0 &&
11251 TrueC->getAPIntValue().isPowerOf2()) {
11252 if (NeedsCondInvert) // Invert the condition if needed.
11253 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
11254 DAG.getConstant(1, Cond.getValueType()));
11256 // Zero extend the condition if needed.
11257 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
11259 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
11260 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
11261 DAG.getConstant(ShAmt, MVT::i8));
11264 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
11265 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
11266 if (NeedsCondInvert) // Invert the condition if needed.
11267 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
11268 DAG.getConstant(1, Cond.getValueType()));
11270 // Zero extend the condition if needed.
11271 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
11272 FalseC->getValueType(0), Cond);
11273 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11274 SDValue(FalseC, 0));
11277 // Optimize cases that will turn into an LEA instruction. This requires
11278 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
11279 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
11280 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
11281 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
11283 bool isFastMultiplier = false;
11285 switch ((unsigned char)Diff) {
11287 case 1: // result = add base, cond
11288 case 2: // result = lea base( , cond*2)
11289 case 3: // result = lea base(cond, cond*2)
11290 case 4: // result = lea base( , cond*4)
11291 case 5: // result = lea base(cond, cond*4)
11292 case 8: // result = lea base( , cond*8)
11293 case 9: // result = lea base(cond, cond*8)
11294 isFastMultiplier = true;
11299 if (isFastMultiplier) {
11300 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
11301 if (NeedsCondInvert) // Invert the condition if needed.
11302 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
11303 DAG.getConstant(1, Cond.getValueType()));
11305 // Zero extend the condition if needed.
11306 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
11308 // Scale the condition by the difference.
11310 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
11311 DAG.getConstant(Diff, Cond.getValueType()));
11313 // Add the base if non-zero.
11314 if (FalseC->getAPIntValue() != 0)
11315 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11316 SDValue(FalseC, 0));
11326 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
11327 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
11328 TargetLowering::DAGCombinerInfo &DCI) {
11329 DebugLoc DL = N->getDebugLoc();
11331 // If the flag operand isn't dead, don't touch this CMOV.
11332 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
11335 // If this is a select between two integer constants, try to do some
11336 // optimizations. Note that the operands are ordered the opposite of SELECT
11338 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
11339 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
11340 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
11341 // larger than FalseC (the false value).
11342 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
11344 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
11345 CC = X86::GetOppositeBranchCondition(CC);
11346 std::swap(TrueC, FalseC);
11349 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
11350 // This is efficient for any integer data type (including i8/i16) and
11352 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
11353 SDValue Cond = N->getOperand(3);
11354 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
11355 DAG.getConstant(CC, MVT::i8), Cond);
11357 // Zero extend the condition if needed.
11358 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
11360 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
11361 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
11362 DAG.getConstant(ShAmt, MVT::i8));
11363 if (N->getNumValues() == 2) // Dead flag value?
11364 return DCI.CombineTo(N, Cond, SDValue());
11368 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
11369 // for any integer data type, including i8/i16.
11370 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
11371 SDValue Cond = N->getOperand(3);
11372 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
11373 DAG.getConstant(CC, MVT::i8), Cond);
11375 // Zero extend the condition if needed.
11376 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
11377 FalseC->getValueType(0), Cond);
11378 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11379 SDValue(FalseC, 0));
11381 if (N->getNumValues() == 2) // Dead flag value?
11382 return DCI.CombineTo(N, Cond, SDValue());
11386 // Optimize cases that will turn into an LEA instruction. This requires
11387 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
11388 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
11389 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
11390 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
11392 bool isFastMultiplier = false;
11394 switch ((unsigned char)Diff) {
11396 case 1: // result = add base, cond
11397 case 2: // result = lea base( , cond*2)
11398 case 3: // result = lea base(cond, cond*2)
11399 case 4: // result = lea base( , cond*4)
11400 case 5: // result = lea base(cond, cond*4)
11401 case 8: // result = lea base( , cond*8)
11402 case 9: // result = lea base(cond, cond*8)
11403 isFastMultiplier = true;
11408 if (isFastMultiplier) {
11409 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
11410 SDValue Cond = N->getOperand(3);
11411 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
11412 DAG.getConstant(CC, MVT::i8), Cond);
11413 // Zero extend the condition if needed.
11414 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
11416 // Scale the condition by the difference.
11418 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
11419 DAG.getConstant(Diff, Cond.getValueType()));
11421 // Add the base if non-zero.
11422 if (FalseC->getAPIntValue() != 0)
11423 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11424 SDValue(FalseC, 0));
11425 if (N->getNumValues() == 2) // Dead flag value?
11426 return DCI.CombineTo(N, Cond, SDValue());
11436 /// PerformMulCombine - Optimize a single multiply with constant into two
11437 /// in order to implement it with two cheaper instructions, e.g.
11438 /// LEA + SHL, LEA + LEA.
11439 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
11440 TargetLowering::DAGCombinerInfo &DCI) {
11441 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
11444 EVT VT = N->getValueType(0);
11445 if (VT != MVT::i64)
11448 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
11451 uint64_t MulAmt = C->getZExtValue();
11452 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
11455 uint64_t MulAmt1 = 0;
11456 uint64_t MulAmt2 = 0;
11457 if ((MulAmt % 9) == 0) {
11459 MulAmt2 = MulAmt / 9;
11460 } else if ((MulAmt % 5) == 0) {
11462 MulAmt2 = MulAmt / 5;
11463 } else if ((MulAmt % 3) == 0) {
11465 MulAmt2 = MulAmt / 3;
11468 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
11469 DebugLoc DL = N->getDebugLoc();
11471 if (isPowerOf2_64(MulAmt2) &&
11472 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
11473 // If second multiplifer is pow2, issue it first. We want the multiply by
11474 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
11476 std::swap(MulAmt1, MulAmt2);
11479 if (isPowerOf2_64(MulAmt1))
11480 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
11481 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
11483 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
11484 DAG.getConstant(MulAmt1, VT));
11486 if (isPowerOf2_64(MulAmt2))
11487 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
11488 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
11490 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
11491 DAG.getConstant(MulAmt2, VT));
11493 // Do not add new nodes to DAG combiner worklist.
11494 DCI.CombineTo(N, NewMul, false);
11499 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
11500 SDValue N0 = N->getOperand(0);
11501 SDValue N1 = N->getOperand(1);
11502 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
11503 EVT VT = N0.getValueType();
11505 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
11506 // since the result of setcc_c is all zero's or all ones.
11507 if (N1C && N0.getOpcode() == ISD::AND &&
11508 N0.getOperand(1).getOpcode() == ISD::Constant) {
11509 SDValue N00 = N0.getOperand(0);
11510 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
11511 ((N00.getOpcode() == ISD::ANY_EXTEND ||
11512 N00.getOpcode() == ISD::ZERO_EXTEND) &&
11513 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
11514 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
11515 APInt ShAmt = N1C->getAPIntValue();
11516 Mask = Mask.shl(ShAmt);
11518 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
11519 N00, DAG.getConstant(Mask, VT));
11526 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
11528 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
11529 const X86Subtarget *Subtarget) {
11530 EVT VT = N->getValueType(0);
11531 if (!VT.isVector() && VT.isInteger() &&
11532 N->getOpcode() == ISD::SHL)
11533 return PerformSHLCombine(N, DAG);
11535 // On X86 with SSE2 support, we can transform this to a vector shift if
11536 // all elements are shifted by the same amount. We can't do this in legalize
11537 // because the a constant vector is typically transformed to a constant pool
11538 // so we have no knowledge of the shift amount.
11539 if (!Subtarget->hasSSE2())
11542 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
11545 SDValue ShAmtOp = N->getOperand(1);
11546 EVT EltVT = VT.getVectorElementType();
11547 DebugLoc DL = N->getDebugLoc();
11548 SDValue BaseShAmt = SDValue();
11549 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
11550 unsigned NumElts = VT.getVectorNumElements();
11552 for (; i != NumElts; ++i) {
11553 SDValue Arg = ShAmtOp.getOperand(i);
11554 if (Arg.getOpcode() == ISD::UNDEF) continue;
11558 for (; i != NumElts; ++i) {
11559 SDValue Arg = ShAmtOp.getOperand(i);
11560 if (Arg.getOpcode() == ISD::UNDEF) continue;
11561 if (Arg != BaseShAmt) {
11565 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
11566 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
11567 SDValue InVec = ShAmtOp.getOperand(0);
11568 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
11569 unsigned NumElts = InVec.getValueType().getVectorNumElements();
11571 for (; i != NumElts; ++i) {
11572 SDValue Arg = InVec.getOperand(i);
11573 if (Arg.getOpcode() == ISD::UNDEF) continue;
11577 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
11578 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
11579 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
11580 if (C->getZExtValue() == SplatIdx)
11581 BaseShAmt = InVec.getOperand(1);
11584 if (BaseShAmt.getNode() == 0)
11585 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
11586 DAG.getIntPtrConstant(0));
11590 // The shift amount is an i32.
11591 if (EltVT.bitsGT(MVT::i32))
11592 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
11593 else if (EltVT.bitsLT(MVT::i32))
11594 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
11596 // The shift amount is identical so we can do a vector shift.
11597 SDValue ValOp = N->getOperand(0);
11598 switch (N->getOpcode()) {
11600 llvm_unreachable("Unknown shift opcode!");
11603 if (VT == MVT::v2i64)
11604 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11605 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
11607 if (VT == MVT::v4i32)
11608 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11609 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
11611 if (VT == MVT::v8i16)
11612 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11613 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
11617 if (VT == MVT::v4i32)
11618 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11619 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
11621 if (VT == MVT::v8i16)
11622 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11623 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
11627 if (VT == MVT::v2i64)
11628 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11629 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
11631 if (VT == MVT::v4i32)
11632 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11633 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
11635 if (VT == MVT::v8i16)
11636 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11637 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
11645 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
11646 TargetLowering::DAGCombinerInfo &DCI,
11647 const X86Subtarget *Subtarget) {
11648 if (DCI.isBeforeLegalizeOps())
11651 // Want to form PANDN nodes, in the hopes of then easily combining them with
11652 // OR and AND nodes to form PBLEND/PSIGN.
11653 EVT VT = N->getValueType(0);
11654 if (VT != MVT::v2i64)
11657 SDValue N0 = N->getOperand(0);
11658 SDValue N1 = N->getOperand(1);
11659 DebugLoc DL = N->getDebugLoc();
11661 // Check LHS for vnot
11662 if (N0.getOpcode() == ISD::XOR &&
11663 ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
11664 return DAG.getNode(X86ISD::PANDN, DL, VT, N0.getOperand(0), N1);
11666 // Check RHS for vnot
11667 if (N1.getOpcode() == ISD::XOR &&
11668 ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
11669 return DAG.getNode(X86ISD::PANDN, DL, VT, N1.getOperand(0), N0);
11674 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
11675 TargetLowering::DAGCombinerInfo &DCI,
11676 const X86Subtarget *Subtarget) {
11677 if (DCI.isBeforeLegalizeOps())
11680 EVT VT = N->getValueType(0);
11681 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64 && VT != MVT::v2i64)
11684 SDValue N0 = N->getOperand(0);
11685 SDValue N1 = N->getOperand(1);
11687 // look for psign/blend
11688 if (Subtarget->hasSSSE3()) {
11689 if (VT == MVT::v2i64) {
11690 // Canonicalize pandn to RHS
11691 if (N0.getOpcode() == X86ISD::PANDN)
11693 // or (and (m, x), (pandn m, y))
11694 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::PANDN) {
11695 SDValue Mask = N1.getOperand(0);
11696 SDValue X = N1.getOperand(1);
11698 if (N0.getOperand(0) == Mask)
11699 Y = N0.getOperand(1);
11700 if (N0.getOperand(1) == Mask)
11701 Y = N0.getOperand(0);
11703 // Check to see if the mask appeared in both the AND and PANDN and
11707 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
11708 if (Mask.getOpcode() != ISD::BITCAST ||
11709 X.getOpcode() != ISD::BITCAST ||
11710 Y.getOpcode() != ISD::BITCAST)
11713 // Look through mask bitcast.
11714 Mask = Mask.getOperand(0);
11715 EVT MaskVT = Mask.getValueType();
11717 // Validate that the Mask operand is a vector sra node. The sra node
11718 // will be an intrinsic.
11719 if (Mask.getOpcode() != ISD::INTRINSIC_WO_CHAIN)
11722 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
11723 // there is no psrai.b
11724 switch (cast<ConstantSDNode>(Mask.getOperand(0))->getZExtValue()) {
11725 case Intrinsic::x86_sse2_psrai_w:
11726 case Intrinsic::x86_sse2_psrai_d:
11728 default: return SDValue();
11731 // Check that the SRA is all signbits.
11732 SDValue SraC = Mask.getOperand(2);
11733 unsigned SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
11734 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
11735 if ((SraAmt + 1) != EltBits)
11738 DebugLoc DL = N->getDebugLoc();
11740 // Now we know we at least have a plendvb with the mask val. See if
11741 // we can form a psignb/w/d.
11742 // psign = x.type == y.type == mask.type && y = sub(0, x);
11743 X = X.getOperand(0);
11744 Y = Y.getOperand(0);
11745 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
11746 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
11747 X.getValueType() == MaskVT && X.getValueType() == Y.getValueType()){
11750 case 8: Opc = X86ISD::PSIGNB; break;
11751 case 16: Opc = X86ISD::PSIGNW; break;
11752 case 32: Opc = X86ISD::PSIGND; break;
11756 SDValue Sign = DAG.getNode(Opc, DL, MaskVT, X, Mask.getOperand(1));
11757 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Sign);
11760 // PBLENDVB only available on SSE 4.1
11761 if (!Subtarget->hasSSE41())
11764 X = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, X);
11765 Y = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Y);
11766 Mask = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Mask);
11767 Mask = DAG.getNode(X86ISD::PBLENDVB, DL, MVT::v16i8, X, Y, Mask);
11768 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Mask);
11773 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
11774 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
11776 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
11778 if (!N0.hasOneUse() || !N1.hasOneUse())
11781 SDValue ShAmt0 = N0.getOperand(1);
11782 if (ShAmt0.getValueType() != MVT::i8)
11784 SDValue ShAmt1 = N1.getOperand(1);
11785 if (ShAmt1.getValueType() != MVT::i8)
11787 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
11788 ShAmt0 = ShAmt0.getOperand(0);
11789 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
11790 ShAmt1 = ShAmt1.getOperand(0);
11792 DebugLoc DL = N->getDebugLoc();
11793 unsigned Opc = X86ISD::SHLD;
11794 SDValue Op0 = N0.getOperand(0);
11795 SDValue Op1 = N1.getOperand(0);
11796 if (ShAmt0.getOpcode() == ISD::SUB) {
11797 Opc = X86ISD::SHRD;
11798 std::swap(Op0, Op1);
11799 std::swap(ShAmt0, ShAmt1);
11802 unsigned Bits = VT.getSizeInBits();
11803 if (ShAmt1.getOpcode() == ISD::SUB) {
11804 SDValue Sum = ShAmt1.getOperand(0);
11805 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
11806 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
11807 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
11808 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
11809 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
11810 return DAG.getNode(Opc, DL, VT,
11812 DAG.getNode(ISD::TRUNCATE, DL,
11815 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
11816 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
11818 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
11819 return DAG.getNode(Opc, DL, VT,
11820 N0.getOperand(0), N1.getOperand(0),
11821 DAG.getNode(ISD::TRUNCATE, DL,
11828 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
11829 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
11830 const X86Subtarget *Subtarget) {
11831 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
11832 // the FP state in cases where an emms may be missing.
11833 // A preferable solution to the general problem is to figure out the right
11834 // places to insert EMMS. This qualifies as a quick hack.
11836 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
11837 StoreSDNode *St = cast<StoreSDNode>(N);
11838 EVT VT = St->getValue().getValueType();
11839 if (VT.getSizeInBits() != 64)
11842 const Function *F = DAG.getMachineFunction().getFunction();
11843 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
11844 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
11845 && Subtarget->hasSSE2();
11846 if ((VT.isVector() ||
11847 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
11848 isa<LoadSDNode>(St->getValue()) &&
11849 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
11850 St->getChain().hasOneUse() && !St->isVolatile()) {
11851 SDNode* LdVal = St->getValue().getNode();
11852 LoadSDNode *Ld = 0;
11853 int TokenFactorIndex = -1;
11854 SmallVector<SDValue, 8> Ops;
11855 SDNode* ChainVal = St->getChain().getNode();
11856 // Must be a store of a load. We currently handle two cases: the load
11857 // is a direct child, and it's under an intervening TokenFactor. It is
11858 // possible to dig deeper under nested TokenFactors.
11859 if (ChainVal == LdVal)
11860 Ld = cast<LoadSDNode>(St->getChain());
11861 else if (St->getValue().hasOneUse() &&
11862 ChainVal->getOpcode() == ISD::TokenFactor) {
11863 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
11864 if (ChainVal->getOperand(i).getNode() == LdVal) {
11865 TokenFactorIndex = i;
11866 Ld = cast<LoadSDNode>(St->getValue());
11868 Ops.push_back(ChainVal->getOperand(i));
11872 if (!Ld || !ISD::isNormalLoad(Ld))
11875 // If this is not the MMX case, i.e. we are just turning i64 load/store
11876 // into f64 load/store, avoid the transformation if there are multiple
11877 // uses of the loaded value.
11878 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
11881 DebugLoc LdDL = Ld->getDebugLoc();
11882 DebugLoc StDL = N->getDebugLoc();
11883 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
11884 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
11886 if (Subtarget->is64Bit() || F64IsLegal) {
11887 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
11888 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
11889 Ld->getPointerInfo(), Ld->isVolatile(),
11890 Ld->isNonTemporal(), Ld->getAlignment());
11891 SDValue NewChain = NewLd.getValue(1);
11892 if (TokenFactorIndex != -1) {
11893 Ops.push_back(NewChain);
11894 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
11897 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
11898 St->getPointerInfo(),
11899 St->isVolatile(), St->isNonTemporal(),
11900 St->getAlignment());
11903 // Otherwise, lower to two pairs of 32-bit loads / stores.
11904 SDValue LoAddr = Ld->getBasePtr();
11905 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
11906 DAG.getConstant(4, MVT::i32));
11908 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
11909 Ld->getPointerInfo(),
11910 Ld->isVolatile(), Ld->isNonTemporal(),
11911 Ld->getAlignment());
11912 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
11913 Ld->getPointerInfo().getWithOffset(4),
11914 Ld->isVolatile(), Ld->isNonTemporal(),
11915 MinAlign(Ld->getAlignment(), 4));
11917 SDValue NewChain = LoLd.getValue(1);
11918 if (TokenFactorIndex != -1) {
11919 Ops.push_back(LoLd);
11920 Ops.push_back(HiLd);
11921 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
11925 LoAddr = St->getBasePtr();
11926 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
11927 DAG.getConstant(4, MVT::i32));
11929 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
11930 St->getPointerInfo(),
11931 St->isVolatile(), St->isNonTemporal(),
11932 St->getAlignment());
11933 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
11934 St->getPointerInfo().getWithOffset(4),
11936 St->isNonTemporal(),
11937 MinAlign(St->getAlignment(), 4));
11938 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
11943 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
11944 /// X86ISD::FXOR nodes.
11945 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
11946 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
11947 // F[X]OR(0.0, x) -> x
11948 // F[X]OR(x, 0.0) -> x
11949 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
11950 if (C->getValueAPF().isPosZero())
11951 return N->getOperand(1);
11952 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
11953 if (C->getValueAPF().isPosZero())
11954 return N->getOperand(0);
11958 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
11959 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
11960 // FAND(0.0, x) -> 0.0
11961 // FAND(x, 0.0) -> 0.0
11962 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
11963 if (C->getValueAPF().isPosZero())
11964 return N->getOperand(0);
11965 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
11966 if (C->getValueAPF().isPosZero())
11967 return N->getOperand(1);
11971 static SDValue PerformBTCombine(SDNode *N,
11973 TargetLowering::DAGCombinerInfo &DCI) {
11974 // BT ignores high bits in the bit index operand.
11975 SDValue Op1 = N->getOperand(1);
11976 if (Op1.hasOneUse()) {
11977 unsigned BitWidth = Op1.getValueSizeInBits();
11978 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
11979 APInt KnownZero, KnownOne;
11980 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
11981 !DCI.isBeforeLegalizeOps());
11982 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
11983 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
11984 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
11985 DCI.CommitTargetLoweringOpt(TLO);
11990 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
11991 SDValue Op = N->getOperand(0);
11992 if (Op.getOpcode() == ISD::BITCAST)
11993 Op = Op.getOperand(0);
11994 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
11995 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
11996 VT.getVectorElementType().getSizeInBits() ==
11997 OpVT.getVectorElementType().getSizeInBits()) {
11998 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
12003 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
12004 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
12005 // (and (i32 x86isd::setcc_carry), 1)
12006 // This eliminates the zext. This transformation is necessary because
12007 // ISD::SETCC is always legalized to i8.
12008 DebugLoc dl = N->getDebugLoc();
12009 SDValue N0 = N->getOperand(0);
12010 EVT VT = N->getValueType(0);
12011 if (N0.getOpcode() == ISD::AND &&
12013 N0.getOperand(0).hasOneUse()) {
12014 SDValue N00 = N0.getOperand(0);
12015 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
12017 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
12018 if (!C || C->getZExtValue() != 1)
12020 return DAG.getNode(ISD::AND, dl, VT,
12021 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
12022 N00.getOperand(0), N00.getOperand(1)),
12023 DAG.getConstant(1, VT));
12029 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
12030 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG) {
12031 unsigned X86CC = N->getConstantOperandVal(0);
12032 SDValue EFLAG = N->getOperand(1);
12033 DebugLoc DL = N->getDebugLoc();
12035 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
12036 // a zext and produces an all-ones bit which is more useful than 0/1 in some
12038 if (X86CC == X86::COND_B)
12039 return DAG.getNode(ISD::AND, DL, MVT::i8,
12040 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
12041 DAG.getConstant(X86CC, MVT::i8), EFLAG),
12042 DAG.getConstant(1, MVT::i8));
12047 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
12048 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
12049 X86TargetLowering::DAGCombinerInfo &DCI) {
12050 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
12051 // the result is either zero or one (depending on the input carry bit).
12052 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
12053 if (X86::isZeroNode(N->getOperand(0)) &&
12054 X86::isZeroNode(N->getOperand(1)) &&
12055 // We don't have a good way to replace an EFLAGS use, so only do this when
12057 SDValue(N, 1).use_empty()) {
12058 DebugLoc DL = N->getDebugLoc();
12059 EVT VT = N->getValueType(0);
12060 SDValue CarryOut = DAG.getConstant(0, N->getValueType(1));
12061 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
12062 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
12063 DAG.getConstant(X86::COND_B,MVT::i8),
12065 DAG.getConstant(1, VT));
12066 return DCI.CombineTo(N, Res1, CarryOut);
12072 // fold (add Y, (sete X, 0)) -> adc 0, Y
12073 // (add Y, (setne X, 0)) -> sbb -1, Y
12074 // (sub (sete X, 0), Y) -> sbb 0, Y
12075 // (sub (setne X, 0), Y) -> adc -1, Y
12076 static SDValue OptimizeConditonalInDecrement(SDNode *N, SelectionDAG &DAG) {
12077 DebugLoc DL = N->getDebugLoc();
12079 // Look through ZExts.
12080 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
12081 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
12084 SDValue SetCC = Ext.getOperand(0);
12085 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
12088 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
12089 if (CC != X86::COND_E && CC != X86::COND_NE)
12092 SDValue Cmp = SetCC.getOperand(1);
12093 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
12094 !X86::isZeroNode(Cmp.getOperand(1)) ||
12095 !Cmp.getOperand(0).getValueType().isInteger())
12098 SDValue CmpOp0 = Cmp.getOperand(0);
12099 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
12100 DAG.getConstant(1, CmpOp0.getValueType()));
12102 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
12103 if (CC == X86::COND_NE)
12104 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
12105 DL, OtherVal.getValueType(), OtherVal,
12106 DAG.getConstant(-1ULL, OtherVal.getValueType()), NewCmp);
12107 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
12108 DL, OtherVal.getValueType(), OtherVal,
12109 DAG.getConstant(0, OtherVal.getValueType()), NewCmp);
12112 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
12113 DAGCombinerInfo &DCI) const {
12114 SelectionDAG &DAG = DCI.DAG;
12115 switch (N->getOpcode()) {
12117 case ISD::EXTRACT_VECTOR_ELT:
12118 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
12119 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
12120 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
12122 case ISD::SUB: return OptimizeConditonalInDecrement(N, DAG);
12123 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
12124 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
12127 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
12128 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
12129 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
12130 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
12132 case X86ISD::FOR: return PerformFORCombine(N, DAG);
12133 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
12134 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
12135 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
12136 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
12137 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG);
12138 case X86ISD::SHUFPS: // Handle all target specific shuffles
12139 case X86ISD::SHUFPD:
12140 case X86ISD::PALIGN:
12141 case X86ISD::PUNPCKHBW:
12142 case X86ISD::PUNPCKHWD:
12143 case X86ISD::PUNPCKHDQ:
12144 case X86ISD::PUNPCKHQDQ:
12145 case X86ISD::UNPCKHPS:
12146 case X86ISD::UNPCKHPD:
12147 case X86ISD::PUNPCKLBW:
12148 case X86ISD::PUNPCKLWD:
12149 case X86ISD::PUNPCKLDQ:
12150 case X86ISD::PUNPCKLQDQ:
12151 case X86ISD::UNPCKLPS:
12152 case X86ISD::UNPCKLPD:
12153 case X86ISD::VUNPCKLPS:
12154 case X86ISD::VUNPCKLPD:
12155 case X86ISD::VUNPCKLPSY:
12156 case X86ISD::VUNPCKLPDY:
12157 case X86ISD::MOVHLPS:
12158 case X86ISD::MOVLHPS:
12159 case X86ISD::PSHUFD:
12160 case X86ISD::PSHUFHW:
12161 case X86ISD::PSHUFLW:
12162 case X86ISD::MOVSS:
12163 case X86ISD::MOVSD:
12164 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI);
12170 /// isTypeDesirableForOp - Return true if the target has native support for
12171 /// the specified value type and it is 'desirable' to use the type for the
12172 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
12173 /// instruction encodings are longer and some i16 instructions are slow.
12174 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
12175 if (!isTypeLegal(VT))
12177 if (VT != MVT::i16)
12184 case ISD::SIGN_EXTEND:
12185 case ISD::ZERO_EXTEND:
12186 case ISD::ANY_EXTEND:
12199 /// IsDesirableToPromoteOp - This method query the target whether it is
12200 /// beneficial for dag combiner to promote the specified node. If true, it
12201 /// should return the desired promotion type by reference.
12202 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
12203 EVT VT = Op.getValueType();
12204 if (VT != MVT::i16)
12207 bool Promote = false;
12208 bool Commute = false;
12209 switch (Op.getOpcode()) {
12212 LoadSDNode *LD = cast<LoadSDNode>(Op);
12213 // If the non-extending load has a single use and it's not live out, then it
12214 // might be folded.
12215 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
12216 Op.hasOneUse()*/) {
12217 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
12218 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
12219 // The only case where we'd want to promote LOAD (rather then it being
12220 // promoted as an operand is when it's only use is liveout.
12221 if (UI->getOpcode() != ISD::CopyToReg)
12228 case ISD::SIGN_EXTEND:
12229 case ISD::ZERO_EXTEND:
12230 case ISD::ANY_EXTEND:
12235 SDValue N0 = Op.getOperand(0);
12236 // Look out for (store (shl (load), x)).
12237 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
12250 SDValue N0 = Op.getOperand(0);
12251 SDValue N1 = Op.getOperand(1);
12252 if (!Commute && MayFoldLoad(N1))
12254 // Avoid disabling potential load folding opportunities.
12255 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
12257 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
12267 //===----------------------------------------------------------------------===//
12268 // X86 Inline Assembly Support
12269 //===----------------------------------------------------------------------===//
12271 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
12272 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
12274 std::string AsmStr = IA->getAsmString();
12276 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
12277 SmallVector<StringRef, 4> AsmPieces;
12278 SplitString(AsmStr, AsmPieces, ";\n");
12280 switch (AsmPieces.size()) {
12281 default: return false;
12283 AsmStr = AsmPieces[0];
12285 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
12287 // FIXME: this should verify that we are targeting a 486 or better. If not,
12288 // we will turn this bswap into something that will be lowered to logical ops
12289 // instead of emitting the bswap asm. For now, we don't support 486 or lower
12290 // so don't worry about this.
12292 if (AsmPieces.size() == 2 &&
12293 (AsmPieces[0] == "bswap" ||
12294 AsmPieces[0] == "bswapq" ||
12295 AsmPieces[0] == "bswapl") &&
12296 (AsmPieces[1] == "$0" ||
12297 AsmPieces[1] == "${0:q}")) {
12298 // No need to check constraints, nothing other than the equivalent of
12299 // "=r,0" would be valid here.
12300 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
12301 if (!Ty || Ty->getBitWidth() % 16 != 0)
12303 return IntrinsicLowering::LowerToByteSwap(CI);
12305 // rorw $$8, ${0:w} --> llvm.bswap.i16
12306 if (CI->getType()->isIntegerTy(16) &&
12307 AsmPieces.size() == 3 &&
12308 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
12309 AsmPieces[1] == "$$8," &&
12310 AsmPieces[2] == "${0:w}" &&
12311 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
12313 const std::string &ConstraintsStr = IA->getConstraintString();
12314 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
12315 std::sort(AsmPieces.begin(), AsmPieces.end());
12316 if (AsmPieces.size() == 4 &&
12317 AsmPieces[0] == "~{cc}" &&
12318 AsmPieces[1] == "~{dirflag}" &&
12319 AsmPieces[2] == "~{flags}" &&
12320 AsmPieces[3] == "~{fpsr}") {
12321 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
12322 if (!Ty || Ty->getBitWidth() % 16 != 0)
12324 return IntrinsicLowering::LowerToByteSwap(CI);
12329 if (CI->getType()->isIntegerTy(32) &&
12330 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
12331 SmallVector<StringRef, 4> Words;
12332 SplitString(AsmPieces[0], Words, " \t,");
12333 if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
12334 Words[2] == "${0:w}") {
12336 SplitString(AsmPieces[1], Words, " \t,");
12337 if (Words.size() == 3 && Words[0] == "rorl" && Words[1] == "$$16" &&
12338 Words[2] == "$0") {
12340 SplitString(AsmPieces[2], Words, " \t,");
12341 if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
12342 Words[2] == "${0:w}") {
12344 const std::string &ConstraintsStr = IA->getConstraintString();
12345 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
12346 std::sort(AsmPieces.begin(), AsmPieces.end());
12347 if (AsmPieces.size() == 4 &&
12348 AsmPieces[0] == "~{cc}" &&
12349 AsmPieces[1] == "~{dirflag}" &&
12350 AsmPieces[2] == "~{flags}" &&
12351 AsmPieces[3] == "~{fpsr}") {
12352 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
12353 if (!Ty || Ty->getBitWidth() % 16 != 0)
12355 return IntrinsicLowering::LowerToByteSwap(CI);
12362 if (CI->getType()->isIntegerTy(64)) {
12363 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
12364 if (Constraints.size() >= 2 &&
12365 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
12366 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
12367 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
12368 SmallVector<StringRef, 4> Words;
12369 SplitString(AsmPieces[0], Words, " \t");
12370 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
12372 SplitString(AsmPieces[1], Words, " \t");
12373 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
12375 SplitString(AsmPieces[2], Words, " \t,");
12376 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
12377 Words[2] == "%edx") {
12378 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
12379 if (!Ty || Ty->getBitWidth() % 16 != 0)
12381 return IntrinsicLowering::LowerToByteSwap(CI);
12394 /// getConstraintType - Given a constraint letter, return the type of
12395 /// constraint it is for this target.
12396 X86TargetLowering::ConstraintType
12397 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
12398 if (Constraint.size() == 1) {
12399 switch (Constraint[0]) {
12409 return C_RegisterClass;
12433 return TargetLowering::getConstraintType(Constraint);
12436 /// Examine constraint type and operand type and determine a weight value.
12437 /// This object must already have been set up with the operand type
12438 /// and the current alternative constraint selected.
12439 TargetLowering::ConstraintWeight
12440 X86TargetLowering::getSingleConstraintMatchWeight(
12441 AsmOperandInfo &info, const char *constraint) const {
12442 ConstraintWeight weight = CW_Invalid;
12443 Value *CallOperandVal = info.CallOperandVal;
12444 // If we don't have a value, we can't do a match,
12445 // but allow it at the lowest weight.
12446 if (CallOperandVal == NULL)
12448 const Type *type = CallOperandVal->getType();
12449 // Look at the constraint type.
12450 switch (*constraint) {
12452 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
12463 if (CallOperandVal->getType()->isIntegerTy())
12464 weight = CW_SpecificReg;
12469 if (type->isFloatingPointTy())
12470 weight = CW_SpecificReg;
12473 if (type->isX86_MMXTy() && Subtarget->hasMMX())
12474 weight = CW_SpecificReg;
12478 if ((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasXMM())
12479 weight = CW_Register;
12482 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
12483 if (C->getZExtValue() <= 31)
12484 weight = CW_Constant;
12488 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12489 if (C->getZExtValue() <= 63)
12490 weight = CW_Constant;
12494 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12495 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
12496 weight = CW_Constant;
12500 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12501 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
12502 weight = CW_Constant;
12506 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12507 if (C->getZExtValue() <= 3)
12508 weight = CW_Constant;
12512 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12513 if (C->getZExtValue() <= 0xff)
12514 weight = CW_Constant;
12519 if (dyn_cast<ConstantFP>(CallOperandVal)) {
12520 weight = CW_Constant;
12524 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12525 if ((C->getSExtValue() >= -0x80000000LL) &&
12526 (C->getSExtValue() <= 0x7fffffffLL))
12527 weight = CW_Constant;
12531 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12532 if (C->getZExtValue() <= 0xffffffff)
12533 weight = CW_Constant;
12540 /// LowerXConstraint - try to replace an X constraint, which matches anything,
12541 /// with another that has more specific requirements based on the type of the
12542 /// corresponding operand.
12543 const char *X86TargetLowering::
12544 LowerXConstraint(EVT ConstraintVT) const {
12545 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
12546 // 'f' like normal targets.
12547 if (ConstraintVT.isFloatingPoint()) {
12548 if (Subtarget->hasXMMInt())
12550 if (Subtarget->hasXMM())
12554 return TargetLowering::LowerXConstraint(ConstraintVT);
12557 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
12558 /// vector. If it is invalid, don't add anything to Ops.
12559 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
12561 std::vector<SDValue>&Ops,
12562 SelectionDAG &DAG) const {
12563 SDValue Result(0, 0);
12565 switch (Constraint) {
12568 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12569 if (C->getZExtValue() <= 31) {
12570 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
12576 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12577 if (C->getZExtValue() <= 63) {
12578 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
12584 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12585 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
12586 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
12592 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12593 if (C->getZExtValue() <= 255) {
12594 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
12600 // 32-bit signed value
12601 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12602 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
12603 C->getSExtValue())) {
12604 // Widen to 64 bits here to get it sign extended.
12605 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
12608 // FIXME gcc accepts some relocatable values here too, but only in certain
12609 // memory models; it's complicated.
12614 // 32-bit unsigned value
12615 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12616 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
12617 C->getZExtValue())) {
12618 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
12622 // FIXME gcc accepts some relocatable values here too, but only in certain
12623 // memory models; it's complicated.
12627 // Literal immediates are always ok.
12628 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
12629 // Widen to 64 bits here to get it sign extended.
12630 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
12634 // In any sort of PIC mode addresses need to be computed at runtime by
12635 // adding in a register or some sort of table lookup. These can't
12636 // be used as immediates.
12637 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
12640 // If we are in non-pic codegen mode, we allow the address of a global (with
12641 // an optional displacement) to be used with 'i'.
12642 GlobalAddressSDNode *GA = 0;
12643 int64_t Offset = 0;
12645 // Match either (GA), (GA+C), (GA+C1+C2), etc.
12647 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
12648 Offset += GA->getOffset();
12650 } else if (Op.getOpcode() == ISD::ADD) {
12651 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
12652 Offset += C->getZExtValue();
12653 Op = Op.getOperand(0);
12656 } else if (Op.getOpcode() == ISD::SUB) {
12657 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
12658 Offset += -C->getZExtValue();
12659 Op = Op.getOperand(0);
12664 // Otherwise, this isn't something we can handle, reject it.
12668 const GlobalValue *GV = GA->getGlobal();
12669 // If we require an extra load to get this address, as in PIC mode, we
12670 // can't accept it.
12671 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
12672 getTargetMachine())))
12675 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
12676 GA->getValueType(0), Offset);
12681 if (Result.getNode()) {
12682 Ops.push_back(Result);
12685 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
12688 std::vector<unsigned> X86TargetLowering::
12689 getRegClassForInlineAsmConstraint(const std::string &Constraint,
12691 if (Constraint.size() == 1) {
12692 // FIXME: not handling fp-stack yet!
12693 switch (Constraint[0]) { // GCC X86 Constraint Letters
12694 default: break; // Unknown constraint letter
12695 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
12696 if (Subtarget->is64Bit()) {
12697 if (VT == MVT::i32)
12698 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
12699 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
12700 X86::R10D,X86::R11D,X86::R12D,
12701 X86::R13D,X86::R14D,X86::R15D,
12702 X86::EBP, X86::ESP, 0);
12703 else if (VT == MVT::i16)
12704 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
12705 X86::SI, X86::DI, X86::R8W,X86::R9W,
12706 X86::R10W,X86::R11W,X86::R12W,
12707 X86::R13W,X86::R14W,X86::R15W,
12708 X86::BP, X86::SP, 0);
12709 else if (VT == MVT::i8)
12710 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
12711 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
12712 X86::R10B,X86::R11B,X86::R12B,
12713 X86::R13B,X86::R14B,X86::R15B,
12714 X86::BPL, X86::SPL, 0);
12716 else if (VT == MVT::i64)
12717 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
12718 X86::RSI, X86::RDI, X86::R8, X86::R9,
12719 X86::R10, X86::R11, X86::R12,
12720 X86::R13, X86::R14, X86::R15,
12721 X86::RBP, X86::RSP, 0);
12725 // 32-bit fallthrough
12726 case 'Q': // Q_REGS
12727 if (VT == MVT::i32)
12728 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
12729 else if (VT == MVT::i16)
12730 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
12731 else if (VT == MVT::i8)
12732 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
12733 else if (VT == MVT::i64)
12734 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
12739 return std::vector<unsigned>();
12742 std::pair<unsigned, const TargetRegisterClass*>
12743 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
12745 // First, see if this is a constraint that directly corresponds to an LLVM
12747 if (Constraint.size() == 1) {
12748 // GCC Constraint Letters
12749 switch (Constraint[0]) {
12751 case 'r': // GENERAL_REGS
12752 case 'l': // INDEX_REGS
12754 return std::make_pair(0U, X86::GR8RegisterClass);
12755 if (VT == MVT::i16)
12756 return std::make_pair(0U, X86::GR16RegisterClass);
12757 if (VT == MVT::i32 || !Subtarget->is64Bit())
12758 return std::make_pair(0U, X86::GR32RegisterClass);
12759 return std::make_pair(0U, X86::GR64RegisterClass);
12760 case 'R': // LEGACY_REGS
12762 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
12763 if (VT == MVT::i16)
12764 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
12765 if (VT == MVT::i32 || !Subtarget->is64Bit())
12766 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
12767 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
12768 case 'f': // FP Stack registers.
12769 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
12770 // value to the correct fpstack register class.
12771 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
12772 return std::make_pair(0U, X86::RFP32RegisterClass);
12773 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
12774 return std::make_pair(0U, X86::RFP64RegisterClass);
12775 return std::make_pair(0U, X86::RFP80RegisterClass);
12776 case 'y': // MMX_REGS if MMX allowed.
12777 if (!Subtarget->hasMMX()) break;
12778 return std::make_pair(0U, X86::VR64RegisterClass);
12779 case 'Y': // SSE_REGS if SSE2 allowed
12780 if (!Subtarget->hasXMMInt()) break;
12782 case 'x': // SSE_REGS if SSE1 allowed
12783 if (!Subtarget->hasXMM()) break;
12785 switch (VT.getSimpleVT().SimpleTy) {
12787 // Scalar SSE types.
12790 return std::make_pair(0U, X86::FR32RegisterClass);
12793 return std::make_pair(0U, X86::FR64RegisterClass);
12801 return std::make_pair(0U, X86::VR128RegisterClass);
12807 // Use the default implementation in TargetLowering to convert the register
12808 // constraint into a member of a register class.
12809 std::pair<unsigned, const TargetRegisterClass*> Res;
12810 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
12812 // Not found as a standard register?
12813 if (Res.second == 0) {
12814 // Map st(0) -> st(7) -> ST0
12815 if (Constraint.size() == 7 && Constraint[0] == '{' &&
12816 tolower(Constraint[1]) == 's' &&
12817 tolower(Constraint[2]) == 't' &&
12818 Constraint[3] == '(' &&
12819 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
12820 Constraint[5] == ')' &&
12821 Constraint[6] == '}') {
12823 Res.first = X86::ST0+Constraint[4]-'0';
12824 Res.second = X86::RFP80RegisterClass;
12828 // GCC allows "st(0)" to be called just plain "st".
12829 if (StringRef("{st}").equals_lower(Constraint)) {
12830 Res.first = X86::ST0;
12831 Res.second = X86::RFP80RegisterClass;
12836 if (StringRef("{flags}").equals_lower(Constraint)) {
12837 Res.first = X86::EFLAGS;
12838 Res.second = X86::CCRRegisterClass;
12842 // 'A' means EAX + EDX.
12843 if (Constraint == "A") {
12844 Res.first = X86::EAX;
12845 Res.second = X86::GR32_ADRegisterClass;
12851 // Otherwise, check to see if this is a register class of the wrong value
12852 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
12853 // turn into {ax},{dx}.
12854 if (Res.second->hasType(VT))
12855 return Res; // Correct type already, nothing to do.
12857 // All of the single-register GCC register classes map their values onto
12858 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
12859 // really want an 8-bit or 32-bit register, map to the appropriate register
12860 // class and return the appropriate register.
12861 if (Res.second == X86::GR16RegisterClass) {
12862 if (VT == MVT::i8) {
12863 unsigned DestReg = 0;
12864 switch (Res.first) {
12866 case X86::AX: DestReg = X86::AL; break;
12867 case X86::DX: DestReg = X86::DL; break;
12868 case X86::CX: DestReg = X86::CL; break;
12869 case X86::BX: DestReg = X86::BL; break;
12872 Res.first = DestReg;
12873 Res.second = X86::GR8RegisterClass;
12875 } else if (VT == MVT::i32) {
12876 unsigned DestReg = 0;
12877 switch (Res.first) {
12879 case X86::AX: DestReg = X86::EAX; break;
12880 case X86::DX: DestReg = X86::EDX; break;
12881 case X86::CX: DestReg = X86::ECX; break;
12882 case X86::BX: DestReg = X86::EBX; break;
12883 case X86::SI: DestReg = X86::ESI; break;
12884 case X86::DI: DestReg = X86::EDI; break;
12885 case X86::BP: DestReg = X86::EBP; break;
12886 case X86::SP: DestReg = X86::ESP; break;
12889 Res.first = DestReg;
12890 Res.second = X86::GR32RegisterClass;
12892 } else if (VT == MVT::i64) {
12893 unsigned DestReg = 0;
12894 switch (Res.first) {
12896 case X86::AX: DestReg = X86::RAX; break;
12897 case X86::DX: DestReg = X86::RDX; break;
12898 case X86::CX: DestReg = X86::RCX; break;
12899 case X86::BX: DestReg = X86::RBX; break;
12900 case X86::SI: DestReg = X86::RSI; break;
12901 case X86::DI: DestReg = X86::RDI; break;
12902 case X86::BP: DestReg = X86::RBP; break;
12903 case X86::SP: DestReg = X86::RSP; break;
12906 Res.first = DestReg;
12907 Res.second = X86::GR64RegisterClass;
12910 } else if (Res.second == X86::FR32RegisterClass ||
12911 Res.second == X86::FR64RegisterClass ||
12912 Res.second == X86::VR128RegisterClass) {
12913 // Handle references to XMM physical registers that got mapped into the
12914 // wrong class. This can happen with constraints like {xmm0} where the
12915 // target independent register mapper will just pick the first match it can
12916 // find, ignoring the required type.
12917 if (VT == MVT::f32)
12918 Res.second = X86::FR32RegisterClass;
12919 else if (VT == MVT::f64)
12920 Res.second = X86::FR64RegisterClass;
12921 else if (X86::VR128RegisterClass->hasType(VT))
12922 Res.second = X86::VR128RegisterClass;