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/CommandLine.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");
60 Disable256Bit("disable-256bit", cl::Hidden,
61 cl::desc("Disable use of 256-bit vectors"));
63 // Forward declarations.
64 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
67 static SDValue Insert128BitVector(SDValue Result,
73 static SDValue Extract128BitVector(SDValue Vec,
78 static SDValue ConcatVectors(SDValue Lower, SDValue Upper, SelectionDAG &DAG);
81 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
82 /// sets things up to match to an AVX VEXTRACTF128 instruction or a
83 /// simple subregister reference. Idx is an index in the 128 bits we
84 /// want. It need not be aligned to a 128-bit bounday. That makes
85 /// lowering EXTRACT_VECTOR_ELT operations easier.
86 static SDValue Extract128BitVector(SDValue Vec,
90 EVT VT = Vec.getValueType();
91 assert(VT.getSizeInBits() == 256 && "Unexpected vector size!");
93 EVT ElVT = VT.getVectorElementType();
95 int Factor = VT.getSizeInBits() / 128;
97 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(),
99 VT.getVectorNumElements() / Factor);
101 // Extract from UNDEF is UNDEF.
102 if (Vec.getOpcode() == ISD::UNDEF)
103 return DAG.getNode(ISD::UNDEF, dl, ResultVT);
105 if (isa<ConstantSDNode>(Idx)) {
106 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
108 // Extract the relevant 128 bits. Generate an EXTRACT_SUBVECTOR
109 // we can match to VEXTRACTF128.
110 unsigned ElemsPerChunk = 128 / ElVT.getSizeInBits();
112 // This is the index of the first element of the 128-bit chunk
114 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / 128)
117 SDValue VecIdx = DAG.getConstant(NormalizedIdxVal, MVT::i32);
119 SDValue Result = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec,
128 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
129 /// sets things up to match to an AVX VINSERTF128 instruction or a
130 /// simple superregister reference. Idx is an index in the 128 bits
131 /// we want. It need not be aligned to a 128-bit bounday. That makes
132 /// lowering INSERT_VECTOR_ELT operations easier.
133 static SDValue Insert128BitVector(SDValue Result,
138 if (isa<ConstantSDNode>(Idx)) {
139 EVT VT = Vec.getValueType();
140 assert(VT.getSizeInBits() == 128 && "Unexpected vector size!");
142 EVT ElVT = VT.getVectorElementType();
144 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
146 EVT ResultVT = Result.getValueType();
148 // Insert the relevant 128 bits.
149 unsigned ElemsPerChunk = 128 / ElVT.getSizeInBits();
151 // This is the index of the first element of the 128-bit chunk
153 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / 128)
156 SDValue VecIdx = DAG.getConstant(NormalizedIdxVal, MVT::i32);
158 Result = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec,
166 /// Given two vectors, concat them.
167 static SDValue ConcatVectors(SDValue Lower, SDValue Upper, SelectionDAG &DAG) {
168 DebugLoc dl = Lower.getDebugLoc();
170 assert(Lower.getValueType() == Upper.getValueType() && "Mismatched vectors!");
172 EVT VT = EVT::getVectorVT(*DAG.getContext(),
173 Lower.getValueType().getVectorElementType(),
174 Lower.getValueType().getVectorNumElements() * 2);
176 // TODO: Generalize to arbitrary vector length (this assumes 256-bit vectors).
177 assert(VT.getSizeInBits() == 256 && "Unsupported vector concat!");
179 // Insert the upper subvector.
180 SDValue Vec = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, VT), Upper,
182 // This is half the length of the result
183 // vector. Start inserting the upper 128
185 Lower.getValueType().getVectorNumElements(),
189 // Insert the lower subvector.
190 Vec = Insert128BitVector(Vec, Lower, DAG.getConstant(0, MVT::i32), DAG, dl);
194 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
195 const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>();
196 bool is64Bit = Subtarget->is64Bit();
198 if (Subtarget->isTargetEnvMacho()) {
200 return new X8664_MachoTargetObjectFile();
201 return new TargetLoweringObjectFileMachO();
204 if (Subtarget->isTargetELF()) {
206 return new X8664_ELFTargetObjectFile(TM);
207 return new X8632_ELFTargetObjectFile(TM);
209 if (Subtarget->isTargetCOFF() && !Subtarget->isTargetEnvMacho())
210 return new TargetLoweringObjectFileCOFF();
211 llvm_unreachable("unknown subtarget type");
214 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
215 : TargetLowering(TM, createTLOF(TM)) {
216 Subtarget = &TM.getSubtarget<X86Subtarget>();
217 X86ScalarSSEf64 = Subtarget->hasXMMInt();
218 X86ScalarSSEf32 = Subtarget->hasXMM();
219 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
221 RegInfo = TM.getRegisterInfo();
222 TD = getTargetData();
224 // Set up the TargetLowering object.
225 static MVT IntVTs[] = { MVT::i8, MVT::i16, MVT::i32, MVT::i64 };
227 // X86 is weird, it always uses i8 for shift amounts and setcc results.
228 setShiftAmountType(MVT::i8);
229 setBooleanContents(ZeroOrOneBooleanContent);
230 setSchedulingPreference(Sched::RegPressure);
231 setStackPointerRegisterToSaveRestore(X86StackPtr);
233 if (Subtarget->isTargetWindows() && !Subtarget->isTargetCygMing()) {
234 // Setup Windows compiler runtime calls.
235 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
236 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
237 setLibcallName(RTLIB::FPTOUINT_F64_I64, "_ftol2");
238 setLibcallName(RTLIB::FPTOUINT_F32_I64, "_ftol2");
239 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
240 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
241 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::C);
242 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::C);
245 if (Subtarget->isTargetDarwin()) {
246 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
247 setUseUnderscoreSetJmp(false);
248 setUseUnderscoreLongJmp(false);
249 } else if (Subtarget->isTargetMingw()) {
250 // MS runtime is weird: it exports _setjmp, but longjmp!
251 setUseUnderscoreSetJmp(true);
252 setUseUnderscoreLongJmp(false);
254 setUseUnderscoreSetJmp(true);
255 setUseUnderscoreLongJmp(true);
258 // Set up the register classes.
259 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
260 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
261 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
262 if (Subtarget->is64Bit())
263 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
265 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
267 // We don't accept any truncstore of integer registers.
268 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
269 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
270 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
271 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
272 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
273 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
275 // SETOEQ and SETUNE require checking two conditions.
276 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
277 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
278 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
279 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
280 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
281 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
283 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
285 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
286 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
287 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
289 if (Subtarget->is64Bit()) {
290 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
291 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
292 } else if (!UseSoftFloat) {
293 // We have an algorithm for SSE2->double, and we turn this into a
294 // 64-bit FILD followed by conditional FADD for other targets.
295 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
296 // We have an algorithm for SSE2, and we turn this into a 64-bit
297 // FILD for other targets.
298 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
301 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
303 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
304 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
307 // SSE has no i16 to fp conversion, only i32
308 if (X86ScalarSSEf32) {
309 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
310 // f32 and f64 cases are Legal, f80 case is not
311 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
313 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
314 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
317 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
318 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
321 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
322 // are Legal, f80 is custom lowered.
323 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
324 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
326 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
328 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
329 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
331 if (X86ScalarSSEf32) {
332 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
333 // f32 and f64 cases are Legal, f80 case is not
334 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
336 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
337 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
340 // Handle FP_TO_UINT by promoting the destination to a larger signed
342 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
343 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
344 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
346 if (Subtarget->is64Bit()) {
347 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
348 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
349 } else if (!UseSoftFloat) {
350 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
351 // Expand FP_TO_UINT into a select.
352 // FIXME: We would like to use a Custom expander here eventually to do
353 // the optimal thing for SSE vs. the default expansion in the legalizer.
354 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
356 // With SSE3 we can use fisttpll to convert to a signed i64; without
357 // SSE, we're stuck with a fistpll.
358 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
361 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
362 if (!X86ScalarSSEf64) {
363 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
364 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
365 if (Subtarget->is64Bit()) {
366 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
367 // Without SSE, i64->f64 goes through memory.
368 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
372 // Scalar integer divide and remainder are lowered to use operations that
373 // produce two results, to match the available instructions. This exposes
374 // the two-result form to trivial CSE, which is able to combine x/y and x%y
375 // into a single instruction.
377 // Scalar integer multiply-high is also lowered to use two-result
378 // operations, to match the available instructions. However, plain multiply
379 // (low) operations are left as Legal, as there are single-result
380 // instructions for this in x86. Using the two-result multiply instructions
381 // when both high and low results are needed must be arranged by dagcombine.
382 for (unsigned i = 0, e = 4; i != e; ++i) {
384 setOperationAction(ISD::MULHS, VT, Expand);
385 setOperationAction(ISD::MULHU, VT, Expand);
386 setOperationAction(ISD::SDIV, VT, Expand);
387 setOperationAction(ISD::UDIV, VT, Expand);
388 setOperationAction(ISD::SREM, VT, Expand);
389 setOperationAction(ISD::UREM, VT, Expand);
391 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
392 setOperationAction(ISD::ADDC, VT, Custom);
393 setOperationAction(ISD::ADDE, VT, Custom);
394 setOperationAction(ISD::SUBC, VT, Custom);
395 setOperationAction(ISD::SUBE, VT, Custom);
398 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
399 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
400 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
401 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
402 if (Subtarget->is64Bit())
403 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
404 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
405 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
406 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
407 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
408 setOperationAction(ISD::FREM , MVT::f32 , Expand);
409 setOperationAction(ISD::FREM , MVT::f64 , Expand);
410 setOperationAction(ISD::FREM , MVT::f80 , Expand);
411 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
413 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
414 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
415 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
416 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
417 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
418 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
419 if (Subtarget->is64Bit()) {
420 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
421 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
424 if (Subtarget->hasPOPCNT()) {
425 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
427 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
428 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
429 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
430 if (Subtarget->is64Bit())
431 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
434 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
435 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
437 // These should be promoted to a larger select which is supported.
438 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
439 // X86 wants to expand cmov itself.
440 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
441 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
442 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
443 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
444 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
445 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
446 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
447 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
448 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
449 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
450 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
451 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
452 if (Subtarget->is64Bit()) {
453 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
454 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
456 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
459 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
460 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
461 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
462 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
463 if (Subtarget->is64Bit())
464 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
465 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
466 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
467 if (Subtarget->is64Bit()) {
468 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
469 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
470 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
471 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
472 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
474 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
475 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
476 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
477 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
478 if (Subtarget->is64Bit()) {
479 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
480 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
481 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
484 if (Subtarget->hasXMM())
485 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
487 // We may not have a libcall for MEMBARRIER so we should lower this.
488 setOperationAction(ISD::MEMBARRIER , MVT::Other, Custom);
490 // On X86 and X86-64, atomic operations are lowered to locked instructions.
491 // Locked instructions, in turn, have implicit fence semantics (all memory
492 // operations are flushed before issuing the locked instruction, and they
493 // are not buffered), so we can fold away the common pattern of
494 // fence-atomic-fence.
495 setShouldFoldAtomicFences(true);
497 // Expand certain atomics
498 for (unsigned i = 0, e = 4; i != e; ++i) {
500 setOperationAction(ISD::ATOMIC_CMP_SWAP, VT, Custom);
501 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
504 if (!Subtarget->is64Bit()) {
505 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
506 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
507 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
508 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
509 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
510 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
511 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
514 // FIXME - use subtarget debug flags
515 if (!Subtarget->isTargetDarwin() &&
516 !Subtarget->isTargetELF() &&
517 !Subtarget->isTargetCygMing()) {
518 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
521 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
522 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
523 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
524 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
525 if (Subtarget->is64Bit()) {
526 setExceptionPointerRegister(X86::RAX);
527 setExceptionSelectorRegister(X86::RDX);
529 setExceptionPointerRegister(X86::EAX);
530 setExceptionSelectorRegister(X86::EDX);
532 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
533 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
535 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
537 setOperationAction(ISD::TRAP, MVT::Other, Legal);
539 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
540 setOperationAction(ISD::VASTART , MVT::Other, Custom);
541 setOperationAction(ISD::VAEND , MVT::Other, Expand);
542 if (Subtarget->is64Bit()) {
543 setOperationAction(ISD::VAARG , MVT::Other, Custom);
544 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
546 setOperationAction(ISD::VAARG , MVT::Other, Expand);
547 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
550 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
551 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
552 if (Subtarget->is64Bit())
553 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
554 if (Subtarget->isTargetCygMing() || Subtarget->isTargetWindows())
555 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
557 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
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->hasSSE42())
952 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
954 if (!UseSoftFloat && Subtarget->hasAVX()) {
955 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
956 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
957 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
958 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
959 addRegisterClass(MVT::v32i8, X86::VR256RegisterClass);
961 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
962 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
963 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
964 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
966 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
967 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
968 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
969 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
970 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
971 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
973 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
974 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
975 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
976 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
977 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
978 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
980 // Custom lower build_vector, vector_shuffle, scalar_to_vector,
981 // insert_vector_elt extract_subvector and extract_vector_elt for
983 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
984 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE;
986 MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
987 // Do not attempt to custom lower non-256-bit vectors
988 if (!isPowerOf2_32(MVT(VT).getVectorNumElements())
989 || (MVT(VT).getSizeInBits() < 256))
991 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
992 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
993 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
994 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
995 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
997 // Custom-lower insert_subvector and extract_subvector based on
999 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
1000 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE;
1002 MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
1003 // Do not attempt to custom lower non-256-bit vectors
1004 if (!isPowerOf2_32(MVT(VT).getVectorNumElements()))
1007 if (MVT(VT).getSizeInBits() == 128) {
1008 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1010 else if (MVT(VT).getSizeInBits() == 256) {
1011 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1015 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1016 // Don't promote loads because we need them for VPERM vector index versions.
1018 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
1019 VT != (unsigned)MVT::LAST_VECTOR_VALUETYPE;
1021 if (!isPowerOf2_32(MVT((MVT::SimpleValueType)VT).getVectorNumElements())
1022 || (MVT((MVT::SimpleValueType)VT).getSizeInBits() < 256))
1024 setOperationAction(ISD::AND, (MVT::SimpleValueType)VT, Promote);
1025 AddPromotedToType (ISD::AND, (MVT::SimpleValueType)VT, MVT::v4i64);
1026 setOperationAction(ISD::OR, (MVT::SimpleValueType)VT, Promote);
1027 AddPromotedToType (ISD::OR, (MVT::SimpleValueType)VT, MVT::v4i64);
1028 setOperationAction(ISD::XOR, (MVT::SimpleValueType)VT, Promote);
1029 AddPromotedToType (ISD::XOR, (MVT::SimpleValueType)VT, MVT::v4i64);
1030 //setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Promote);
1031 //AddPromotedToType (ISD::LOAD, (MVT::SimpleValueType)VT, MVT::v4i64);
1032 setOperationAction(ISD::SELECT, (MVT::SimpleValueType)VT, Promote);
1033 AddPromotedToType (ISD::SELECT, (MVT::SimpleValueType)VT, MVT::v4i64);
1037 // We want to custom lower some of our intrinsics.
1038 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1041 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1042 // handle type legalization for these operations here.
1044 // FIXME: We really should do custom legalization for addition and
1045 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1046 // than generic legalization for 64-bit multiplication-with-overflow, though.
1047 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1048 // Add/Sub/Mul with overflow operations are custom lowered.
1050 setOperationAction(ISD::SADDO, VT, Custom);
1051 setOperationAction(ISD::UADDO, VT, Custom);
1052 setOperationAction(ISD::SSUBO, VT, Custom);
1053 setOperationAction(ISD::USUBO, VT, Custom);
1054 setOperationAction(ISD::SMULO, VT, Custom);
1055 setOperationAction(ISD::UMULO, VT, Custom);
1058 // There are no 8-bit 3-address imul/mul instructions
1059 setOperationAction(ISD::SMULO, MVT::i8, Expand);
1060 setOperationAction(ISD::UMULO, MVT::i8, Expand);
1062 if (!Subtarget->is64Bit()) {
1063 // These libcalls are not available in 32-bit.
1064 setLibcallName(RTLIB::SHL_I128, 0);
1065 setLibcallName(RTLIB::SRL_I128, 0);
1066 setLibcallName(RTLIB::SRA_I128, 0);
1069 // We have target-specific dag combine patterns for the following nodes:
1070 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1071 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1072 setTargetDAGCombine(ISD::BUILD_VECTOR);
1073 setTargetDAGCombine(ISD::SELECT);
1074 setTargetDAGCombine(ISD::SHL);
1075 setTargetDAGCombine(ISD::SRA);
1076 setTargetDAGCombine(ISD::SRL);
1077 setTargetDAGCombine(ISD::OR);
1078 setTargetDAGCombine(ISD::AND);
1079 setTargetDAGCombine(ISD::ADD);
1080 setTargetDAGCombine(ISD::SUB);
1081 setTargetDAGCombine(ISD::STORE);
1082 setTargetDAGCombine(ISD::ZERO_EXTEND);
1083 if (Subtarget->is64Bit())
1084 setTargetDAGCombine(ISD::MUL);
1086 computeRegisterProperties();
1088 // On Darwin, -Os means optimize for size without hurting performance,
1089 // do not reduce the limit.
1090 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1091 maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
1092 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1093 maxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1094 maxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1095 maxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1096 setPrefLoopAlignment(16);
1097 benefitFromCodePlacementOpt = true;
1101 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1106 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1107 /// the desired ByVal argument alignment.
1108 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1111 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1112 if (VTy->getBitWidth() == 128)
1114 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1115 unsigned EltAlign = 0;
1116 getMaxByValAlign(ATy->getElementType(), EltAlign);
1117 if (EltAlign > MaxAlign)
1118 MaxAlign = EltAlign;
1119 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1120 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1121 unsigned EltAlign = 0;
1122 getMaxByValAlign(STy->getElementType(i), EltAlign);
1123 if (EltAlign > MaxAlign)
1124 MaxAlign = EltAlign;
1132 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1133 /// function arguments in the caller parameter area. For X86, aggregates
1134 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1135 /// are at 4-byte boundaries.
1136 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1137 if (Subtarget->is64Bit()) {
1138 // Max of 8 and alignment of type.
1139 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1146 if (Subtarget->hasXMM())
1147 getMaxByValAlign(Ty, Align);
1151 /// getOptimalMemOpType - Returns the target specific optimal type for load
1152 /// and store operations as a result of memset, memcpy, and memmove
1153 /// lowering. If DstAlign is zero that means it's safe to destination
1154 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1155 /// means there isn't a need to check it against alignment requirement,
1156 /// probably because the source does not need to be loaded. If
1157 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1158 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1159 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1160 /// constant so it does not need to be loaded.
1161 /// It returns EVT::Other if the type should be determined using generic
1162 /// target-independent logic.
1164 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1165 unsigned DstAlign, unsigned SrcAlign,
1166 bool NonScalarIntSafe,
1168 MachineFunction &MF) const {
1169 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1170 // linux. This is because the stack realignment code can't handle certain
1171 // cases like PR2962. This should be removed when PR2962 is fixed.
1172 const Function *F = MF.getFunction();
1173 if (NonScalarIntSafe &&
1174 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1176 (Subtarget->isUnalignedMemAccessFast() ||
1177 ((DstAlign == 0 || DstAlign >= 16) &&
1178 (SrcAlign == 0 || SrcAlign >= 16))) &&
1179 Subtarget->getStackAlignment() >= 16) {
1180 if (Subtarget->hasSSE2())
1182 if (Subtarget->hasSSE1())
1184 } else if (!MemcpyStrSrc && Size >= 8 &&
1185 !Subtarget->is64Bit() &&
1186 Subtarget->getStackAlignment() >= 8 &&
1187 Subtarget->hasXMMInt()) {
1188 // Do not use f64 to lower memcpy if source is string constant. It's
1189 // better to use i32 to avoid the loads.
1193 if (Subtarget->is64Bit() && Size >= 8)
1198 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1199 /// current function. The returned value is a member of the
1200 /// MachineJumpTableInfo::JTEntryKind enum.
1201 unsigned X86TargetLowering::getJumpTableEncoding() const {
1202 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1204 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1205 Subtarget->isPICStyleGOT())
1206 return MachineJumpTableInfo::EK_Custom32;
1208 // Otherwise, use the normal jump table encoding heuristics.
1209 return TargetLowering::getJumpTableEncoding();
1213 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1214 const MachineBasicBlock *MBB,
1215 unsigned uid,MCContext &Ctx) const{
1216 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1217 Subtarget->isPICStyleGOT());
1218 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1220 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1221 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1224 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1226 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1227 SelectionDAG &DAG) const {
1228 if (!Subtarget->is64Bit())
1229 // This doesn't have DebugLoc associated with it, but is not really the
1230 // same as a Register.
1231 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1235 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1236 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1238 const MCExpr *X86TargetLowering::
1239 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1240 MCContext &Ctx) const {
1241 // X86-64 uses RIP relative addressing based on the jump table label.
1242 if (Subtarget->isPICStyleRIPRel())
1243 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1245 // Otherwise, the reference is relative to the PIC base.
1246 return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
1249 /// getFunctionAlignment - Return the Log2 alignment of this function.
1250 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1251 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1254 // FIXME: Why this routine is here? Move to RegInfo!
1255 std::pair<const TargetRegisterClass*, uint8_t>
1256 X86TargetLowering::findRepresentativeClass(EVT VT) const{
1257 const TargetRegisterClass *RRC = 0;
1259 switch (VT.getSimpleVT().SimpleTy) {
1261 return TargetLowering::findRepresentativeClass(VT);
1262 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1263 RRC = (Subtarget->is64Bit()
1264 ? X86::GR64RegisterClass : X86::GR32RegisterClass);
1267 RRC = X86::VR64RegisterClass;
1269 case MVT::f32: case MVT::f64:
1270 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1271 case MVT::v4f32: case MVT::v2f64:
1272 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1274 RRC = X86::VR128RegisterClass;
1277 return std::make_pair(RRC, Cost);
1280 // FIXME: Why this routine is here? Move to RegInfo!
1282 X86TargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
1283 MachineFunction &MF) const {
1284 const TargetFrameLowering *TFI = MF.getTarget().getFrameLowering();
1286 unsigned FPDiff = TFI->hasFP(MF) ? 1 : 0;
1287 switch (RC->getID()) {
1290 case X86::GR32RegClassID:
1292 case X86::GR64RegClassID:
1294 case X86::VR128RegClassID:
1295 return Subtarget->is64Bit() ? 10 : 4;
1296 case X86::VR64RegClassID:
1301 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1302 unsigned &Offset) const {
1303 if (!Subtarget->isTargetLinux())
1306 if (Subtarget->is64Bit()) {
1307 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1309 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1322 //===----------------------------------------------------------------------===//
1323 // Return Value Calling Convention Implementation
1324 //===----------------------------------------------------------------------===//
1326 #include "X86GenCallingConv.inc"
1329 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1330 const SmallVectorImpl<ISD::OutputArg> &Outs,
1331 LLVMContext &Context) const {
1332 SmallVector<CCValAssign, 16> RVLocs;
1333 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1335 return CCInfo.CheckReturn(Outs, RetCC_X86);
1339 X86TargetLowering::LowerReturn(SDValue Chain,
1340 CallingConv::ID CallConv, bool isVarArg,
1341 const SmallVectorImpl<ISD::OutputArg> &Outs,
1342 const SmallVectorImpl<SDValue> &OutVals,
1343 DebugLoc dl, SelectionDAG &DAG) const {
1344 MachineFunction &MF = DAG.getMachineFunction();
1345 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1347 SmallVector<CCValAssign, 16> RVLocs;
1348 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1349 RVLocs, *DAG.getContext());
1350 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1352 // Add the regs to the liveout set for the function.
1353 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1354 for (unsigned i = 0; i != RVLocs.size(); ++i)
1355 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1356 MRI.addLiveOut(RVLocs[i].getLocReg());
1360 SmallVector<SDValue, 6> RetOps;
1361 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1362 // Operand #1 = Bytes To Pop
1363 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1366 // Copy the result values into the output registers.
1367 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1368 CCValAssign &VA = RVLocs[i];
1369 assert(VA.isRegLoc() && "Can only return in registers!");
1370 SDValue ValToCopy = OutVals[i];
1371 EVT ValVT = ValToCopy.getValueType();
1373 // If this is x86-64, and we disabled SSE, we can't return FP values,
1374 // or SSE or MMX vectors.
1375 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
1376 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
1377 (Subtarget->is64Bit() && !Subtarget->hasXMM())) {
1378 report_fatal_error("SSE register return with SSE disabled");
1380 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1381 // llvm-gcc has never done it right and no one has noticed, so this
1382 // should be OK for now.
1383 if (ValVT == MVT::f64 &&
1384 (Subtarget->is64Bit() && !Subtarget->hasXMMInt()))
1385 report_fatal_error("SSE2 register return with SSE2 disabled");
1387 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1388 // the RET instruction and handled by the FP Stackifier.
1389 if (VA.getLocReg() == X86::ST0 ||
1390 VA.getLocReg() == X86::ST1) {
1391 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1392 // change the value to the FP stack register class.
1393 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1394 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1395 RetOps.push_back(ValToCopy);
1396 // Don't emit a copytoreg.
1400 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1401 // which is returned in RAX / RDX.
1402 if (Subtarget->is64Bit()) {
1403 if (ValVT == MVT::x86mmx) {
1404 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1405 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ValToCopy);
1406 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
1408 // If we don't have SSE2 available, convert to v4f32 so the generated
1409 // register is legal.
1410 if (!Subtarget->hasSSE2())
1411 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,ValToCopy);
1416 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1417 Flag = Chain.getValue(1);
1420 // The x86-64 ABI for returning structs by value requires that we copy
1421 // the sret argument into %rax for the return. We saved the argument into
1422 // a virtual register in the entry block, so now we copy the value out
1424 if (Subtarget->is64Bit() &&
1425 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1426 MachineFunction &MF = DAG.getMachineFunction();
1427 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1428 unsigned Reg = FuncInfo->getSRetReturnReg();
1430 "SRetReturnReg should have been set in LowerFormalArguments().");
1431 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1433 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1434 Flag = Chain.getValue(1);
1436 // RAX now acts like a return value.
1437 MRI.addLiveOut(X86::RAX);
1440 RetOps[0] = Chain; // Update chain.
1442 // Add the flag if we have it.
1444 RetOps.push_back(Flag);
1446 return DAG.getNode(X86ISD::RET_FLAG, dl,
1447 MVT::Other, &RetOps[0], RetOps.size());
1450 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N) const {
1451 if (N->getNumValues() != 1)
1453 if (!N->hasNUsesOfValue(1, 0))
1456 SDNode *Copy = *N->use_begin();
1457 if (Copy->getOpcode() != ISD::CopyToReg &&
1458 Copy->getOpcode() != ISD::FP_EXTEND)
1461 bool HasRet = false;
1462 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1464 if (UI->getOpcode() != X86ISD::RET_FLAG)
1472 /// LowerCallResult - Lower the result values of a call into the
1473 /// appropriate copies out of appropriate physical registers.
1476 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1477 CallingConv::ID CallConv, bool isVarArg,
1478 const SmallVectorImpl<ISD::InputArg> &Ins,
1479 DebugLoc dl, SelectionDAG &DAG,
1480 SmallVectorImpl<SDValue> &InVals) const {
1482 // Assign locations to each value returned by this call.
1483 SmallVector<CCValAssign, 16> RVLocs;
1484 bool Is64Bit = Subtarget->is64Bit();
1485 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1486 RVLocs, *DAG.getContext());
1487 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1489 // Copy all of the result registers out of their specified physreg.
1490 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1491 CCValAssign &VA = RVLocs[i];
1492 EVT CopyVT = VA.getValVT();
1494 // If this is x86-64, and we disabled SSE, we can't return FP values
1495 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1496 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasXMM())) {
1497 report_fatal_error("SSE register return with SSE disabled");
1502 // If this is a call to a function that returns an fp value on the floating
1503 // point stack, we must guarantee the the value is popped from the stack, so
1504 // a CopyFromReg is not good enough - the copy instruction may be eliminated
1505 // if the return value is not used. We use the FpGET_ST0 instructions
1507 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
1508 // If we prefer to use the value in xmm registers, copy it out as f80 and
1509 // use a truncate to move it from fp stack reg to xmm reg.
1510 if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
1511 bool isST0 = VA.getLocReg() == X86::ST0;
1513 if (CopyVT == MVT::f32) Opc = isST0 ? X86::FpGET_ST0_32:X86::FpGET_ST1_32;
1514 if (CopyVT == MVT::f64) Opc = isST0 ? X86::FpGET_ST0_64:X86::FpGET_ST1_64;
1515 if (CopyVT == MVT::f80) Opc = isST0 ? X86::FpGET_ST0_80:X86::FpGET_ST1_80;
1516 SDValue Ops[] = { Chain, InFlag };
1517 Chain = SDValue(DAG.getMachineNode(Opc, dl, CopyVT, MVT::Other, MVT::Glue,
1519 Val = Chain.getValue(0);
1521 // Round the f80 to the right size, which also moves it to the appropriate
1523 if (CopyVT != VA.getValVT())
1524 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1525 // This truncation won't change the value.
1526 DAG.getIntPtrConstant(1));
1527 } else if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1528 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1529 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1530 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1531 MVT::v2i64, InFlag).getValue(1);
1532 Val = Chain.getValue(0);
1533 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1534 Val, DAG.getConstant(0, MVT::i64));
1536 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1537 MVT::i64, InFlag).getValue(1);
1538 Val = Chain.getValue(0);
1540 Val = DAG.getNode(ISD::BITCAST, dl, CopyVT, Val);
1542 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1543 CopyVT, InFlag).getValue(1);
1544 Val = Chain.getValue(0);
1546 InFlag = Chain.getValue(2);
1547 InVals.push_back(Val);
1554 //===----------------------------------------------------------------------===//
1555 // C & StdCall & Fast Calling Convention implementation
1556 //===----------------------------------------------------------------------===//
1557 // StdCall calling convention seems to be standard for many Windows' API
1558 // routines and around. It differs from C calling convention just a little:
1559 // callee should clean up the stack, not caller. Symbols should be also
1560 // decorated in some fancy way :) It doesn't support any vector arguments.
1561 // For info on fast calling convention see Fast Calling Convention (tail call)
1562 // implementation LowerX86_32FastCCCallTo.
1564 /// CallIsStructReturn - Determines whether a call uses struct return
1566 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1570 return Outs[0].Flags.isSRet();
1573 /// ArgsAreStructReturn - Determines whether a function uses struct
1574 /// return semantics.
1576 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1580 return Ins[0].Flags.isSRet();
1583 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1584 /// by "Src" to address "Dst" with size and alignment information specified by
1585 /// the specific parameter attribute. The copy will be passed as a byval
1586 /// function parameter.
1588 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1589 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1591 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1593 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1594 /*isVolatile*/false, /*AlwaysInline=*/true,
1595 MachinePointerInfo(), MachinePointerInfo());
1598 /// IsTailCallConvention - Return true if the calling convention is one that
1599 /// supports tail call optimization.
1600 static bool IsTailCallConvention(CallingConv::ID CC) {
1601 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1604 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1605 /// a tailcall target by changing its ABI.
1606 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1607 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1611 X86TargetLowering::LowerMemArgument(SDValue Chain,
1612 CallingConv::ID CallConv,
1613 const SmallVectorImpl<ISD::InputArg> &Ins,
1614 DebugLoc dl, SelectionDAG &DAG,
1615 const CCValAssign &VA,
1616 MachineFrameInfo *MFI,
1618 // Create the nodes corresponding to a load from this parameter slot.
1619 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1620 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1621 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1624 // If value is passed by pointer we have address passed instead of the value
1626 if (VA.getLocInfo() == CCValAssign::Indirect)
1627 ValVT = VA.getLocVT();
1629 ValVT = VA.getValVT();
1631 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1632 // changed with more analysis.
1633 // In case of tail call optimization mark all arguments mutable. Since they
1634 // could be overwritten by lowering of arguments in case of a tail call.
1635 if (Flags.isByVal()) {
1636 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1637 VA.getLocMemOffset(), isImmutable);
1638 return DAG.getFrameIndex(FI, getPointerTy());
1640 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1641 VA.getLocMemOffset(), isImmutable);
1642 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1643 return DAG.getLoad(ValVT, dl, Chain, FIN,
1644 MachinePointerInfo::getFixedStack(FI),
1650 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1651 CallingConv::ID CallConv,
1653 const SmallVectorImpl<ISD::InputArg> &Ins,
1656 SmallVectorImpl<SDValue> &InVals)
1658 MachineFunction &MF = DAG.getMachineFunction();
1659 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1661 const Function* Fn = MF.getFunction();
1662 if (Fn->hasExternalLinkage() &&
1663 Subtarget->isTargetCygMing() &&
1664 Fn->getName() == "main")
1665 FuncInfo->setForceFramePointer(true);
1667 MachineFrameInfo *MFI = MF.getFrameInfo();
1668 bool Is64Bit = Subtarget->is64Bit();
1669 bool IsWin64 = Subtarget->isTargetWin64();
1671 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1672 "Var args not supported with calling convention fastcc or ghc");
1674 // Assign locations to all of the incoming arguments.
1675 SmallVector<CCValAssign, 16> ArgLocs;
1676 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1677 ArgLocs, *DAG.getContext());
1679 // Allocate shadow area for Win64
1681 CCInfo.AllocateStack(32, 8);
1684 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
1686 unsigned LastVal = ~0U;
1688 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1689 CCValAssign &VA = ArgLocs[i];
1690 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1692 assert(VA.getValNo() != LastVal &&
1693 "Don't support value assigned to multiple locs yet");
1694 LastVal = VA.getValNo();
1696 if (VA.isRegLoc()) {
1697 EVT RegVT = VA.getLocVT();
1698 TargetRegisterClass *RC = NULL;
1699 if (RegVT == MVT::i32)
1700 RC = X86::GR32RegisterClass;
1701 else if (Is64Bit && RegVT == MVT::i64)
1702 RC = X86::GR64RegisterClass;
1703 else if (RegVT == MVT::f32)
1704 RC = X86::FR32RegisterClass;
1705 else if (RegVT == MVT::f64)
1706 RC = X86::FR64RegisterClass;
1707 else if (RegVT.isVector() && RegVT.getSizeInBits() == 256)
1708 RC = X86::VR256RegisterClass;
1709 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1710 RC = X86::VR128RegisterClass;
1711 else if (RegVT == MVT::x86mmx)
1712 RC = X86::VR64RegisterClass;
1714 llvm_unreachable("Unknown argument type!");
1716 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC, dl);
1717 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1719 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1720 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1722 if (VA.getLocInfo() == CCValAssign::SExt)
1723 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1724 DAG.getValueType(VA.getValVT()));
1725 else if (VA.getLocInfo() == CCValAssign::ZExt)
1726 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1727 DAG.getValueType(VA.getValVT()));
1728 else if (VA.getLocInfo() == CCValAssign::BCvt)
1729 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
1731 if (VA.isExtInLoc()) {
1732 // Handle MMX values passed in XMM regs.
1733 if (RegVT.isVector()) {
1734 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(),
1737 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1740 assert(VA.isMemLoc());
1741 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1744 // If value is passed via pointer - do a load.
1745 if (VA.getLocInfo() == CCValAssign::Indirect)
1746 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
1747 MachinePointerInfo(), false, false, 0);
1749 InVals.push_back(ArgValue);
1752 // The x86-64 ABI for returning structs by value requires that we copy
1753 // the sret argument into %rax for the return. Save the argument into
1754 // a virtual register so that we can access it from the return points.
1755 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1756 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1757 unsigned Reg = FuncInfo->getSRetReturnReg();
1759 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1760 FuncInfo->setSRetReturnReg(Reg);
1762 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1763 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1766 unsigned StackSize = CCInfo.getNextStackOffset();
1767 // Align stack specially for tail calls.
1768 if (FuncIsMadeTailCallSafe(CallConv))
1769 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1771 // If the function takes variable number of arguments, make a frame index for
1772 // the start of the first vararg value... for expansion of llvm.va_start.
1774 if (!IsWin64 && (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1775 CallConv != CallingConv::X86_ThisCall))) {
1776 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
1779 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1781 // FIXME: We should really autogenerate these arrays
1782 static const unsigned GPR64ArgRegsWin64[] = {
1783 X86::RCX, X86::RDX, X86::R8, X86::R9
1785 static const unsigned GPR64ArgRegs64Bit[] = {
1786 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1788 static const unsigned XMMArgRegs64Bit[] = {
1789 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1790 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1792 const unsigned *GPR64ArgRegs;
1793 unsigned NumXMMRegs = 0;
1796 // The XMM registers which might contain var arg parameters are shadowed
1797 // in their paired GPR. So we only need to save the GPR to their home
1799 TotalNumIntRegs = 4;
1800 GPR64ArgRegs = GPR64ArgRegsWin64;
1802 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1803 GPR64ArgRegs = GPR64ArgRegs64Bit;
1805 NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs64Bit, TotalNumXMMRegs);
1807 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1810 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1811 assert(!(NumXMMRegs && !Subtarget->hasXMM()) &&
1812 "SSE register cannot be used when SSE is disabled!");
1813 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1814 "SSE register cannot be used when SSE is disabled!");
1815 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasXMM())
1816 // Kernel mode asks for SSE to be disabled, so don't push them
1818 TotalNumXMMRegs = 0;
1821 const TargetFrameLowering &TFI = *getTargetMachine().getFrameLowering();
1822 // Get to the caller-allocated home save location. Add 8 to account
1823 // for the return address.
1824 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
1825 FuncInfo->setRegSaveFrameIndex(
1826 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
1827 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
1829 // For X86-64, if there are vararg parameters that are passed via
1830 // registers, then we must store them to their spots on the stack so they
1831 // may be loaded by deferencing the result of va_next.
1832 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1833 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1834 FuncInfo->setRegSaveFrameIndex(
1835 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1839 // Store the integer parameter registers.
1840 SmallVector<SDValue, 8> MemOps;
1841 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1843 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1844 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1845 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1846 DAG.getIntPtrConstant(Offset));
1847 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1848 X86::GR64RegisterClass, dl);
1849 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1851 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1852 MachinePointerInfo::getFixedStack(
1853 FuncInfo->getRegSaveFrameIndex(), Offset),
1855 MemOps.push_back(Store);
1859 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1860 // Now store the XMM (fp + vector) parameter registers.
1861 SmallVector<SDValue, 11> SaveXMMOps;
1862 SaveXMMOps.push_back(Chain);
1864 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass, dl);
1865 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1866 SaveXMMOps.push_back(ALVal);
1868 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1869 FuncInfo->getRegSaveFrameIndex()));
1870 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1871 FuncInfo->getVarArgsFPOffset()));
1873 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1874 unsigned VReg = MF.addLiveIn(XMMArgRegs64Bit[NumXMMRegs],
1875 X86::VR128RegisterClass, dl);
1876 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1877 SaveXMMOps.push_back(Val);
1879 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1881 &SaveXMMOps[0], SaveXMMOps.size()));
1884 if (!MemOps.empty())
1885 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1886 &MemOps[0], MemOps.size());
1890 // Some CCs need callee pop.
1891 if (Subtarget->IsCalleePop(isVarArg, CallConv)) {
1892 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1894 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1895 // If this is an sret function, the return should pop the hidden pointer.
1896 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1897 FuncInfo->setBytesToPopOnReturn(4);
1901 // RegSaveFrameIndex is X86-64 only.
1902 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1903 if (CallConv == CallingConv::X86_FastCall ||
1904 CallConv == CallingConv::X86_ThisCall)
1905 // fastcc functions can't have varargs.
1906 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1913 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1914 SDValue StackPtr, SDValue Arg,
1915 DebugLoc dl, SelectionDAG &DAG,
1916 const CCValAssign &VA,
1917 ISD::ArgFlagsTy Flags) const {
1918 unsigned LocMemOffset = VA.getLocMemOffset();
1919 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1920 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1921 if (Flags.isByVal())
1922 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1924 return DAG.getStore(Chain, dl, Arg, PtrOff,
1925 MachinePointerInfo::getStack(LocMemOffset),
1929 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1930 /// optimization is performed and it is required.
1932 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1933 SDValue &OutRetAddr, SDValue Chain,
1934 bool IsTailCall, bool Is64Bit,
1935 int FPDiff, DebugLoc dl) const {
1936 // Adjust the Return address stack slot.
1937 EVT VT = getPointerTy();
1938 OutRetAddr = getReturnAddressFrameIndex(DAG);
1940 // Load the "old" Return address.
1941 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
1943 return SDValue(OutRetAddr.getNode(), 1);
1946 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1947 /// optimization is performed and it is required (FPDiff!=0).
1949 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1950 SDValue Chain, SDValue RetAddrFrIdx,
1951 bool Is64Bit, int FPDiff, DebugLoc dl) {
1952 // Store the return address to the appropriate stack slot.
1953 if (!FPDiff) return Chain;
1954 // Calculate the new stack slot for the return address.
1955 int SlotSize = Is64Bit ? 8 : 4;
1956 int NewReturnAddrFI =
1957 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
1958 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1959 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1960 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1961 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
1967 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1968 CallingConv::ID CallConv, bool isVarArg,
1970 const SmallVectorImpl<ISD::OutputArg> &Outs,
1971 const SmallVectorImpl<SDValue> &OutVals,
1972 const SmallVectorImpl<ISD::InputArg> &Ins,
1973 DebugLoc dl, SelectionDAG &DAG,
1974 SmallVectorImpl<SDValue> &InVals) const {
1975 MachineFunction &MF = DAG.getMachineFunction();
1976 bool Is64Bit = Subtarget->is64Bit();
1977 bool IsWin64 = Subtarget->isTargetWin64();
1978 bool IsStructRet = CallIsStructReturn(Outs);
1979 bool IsSibcall = false;
1982 // Check if it's really possible to do a tail call.
1983 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1984 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1985 Outs, OutVals, Ins, DAG);
1987 // Sibcalls are automatically detected tailcalls which do not require
1989 if (!GuaranteedTailCallOpt && isTailCall)
1996 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1997 "Var args not supported with calling convention fastcc or ghc");
1999 // Analyze operands of the call, assigning locations to each operand.
2000 SmallVector<CCValAssign, 16> ArgLocs;
2001 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
2002 ArgLocs, *DAG.getContext());
2004 // Allocate shadow area for Win64
2006 CCInfo.AllocateStack(32, 8);
2009 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2011 // Get a count of how many bytes are to be pushed on the stack.
2012 unsigned NumBytes = CCInfo.getNextStackOffset();
2014 // This is a sibcall. The memory operands are available in caller's
2015 // own caller's stack.
2017 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
2018 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2021 if (isTailCall && !IsSibcall) {
2022 // Lower arguments at fp - stackoffset + fpdiff.
2023 unsigned NumBytesCallerPushed =
2024 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
2025 FPDiff = NumBytesCallerPushed - NumBytes;
2027 // Set the delta of movement of the returnaddr stackslot.
2028 // But only set if delta is greater than previous delta.
2029 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
2030 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
2034 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
2036 SDValue RetAddrFrIdx;
2037 // Load return adress for tail calls.
2038 if (isTailCall && FPDiff)
2039 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
2040 Is64Bit, FPDiff, dl);
2042 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2043 SmallVector<SDValue, 8> MemOpChains;
2046 // Walk the register/memloc assignments, inserting copies/loads. In the case
2047 // of tail call optimization arguments are handle later.
2048 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2049 CCValAssign &VA = ArgLocs[i];
2050 EVT RegVT = VA.getLocVT();
2051 SDValue Arg = OutVals[i];
2052 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2053 bool isByVal = Flags.isByVal();
2055 // Promote the value if needed.
2056 switch (VA.getLocInfo()) {
2057 default: llvm_unreachable("Unknown loc info!");
2058 case CCValAssign::Full: break;
2059 case CCValAssign::SExt:
2060 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2062 case CCValAssign::ZExt:
2063 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
2065 case CCValAssign::AExt:
2066 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
2067 // Special case: passing MMX values in XMM registers.
2068 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
2069 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
2070 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
2072 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
2074 case CCValAssign::BCvt:
2075 Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg);
2077 case CCValAssign::Indirect: {
2078 // Store the argument.
2079 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2080 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2081 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2082 MachinePointerInfo::getFixedStack(FI),
2089 if (VA.isRegLoc()) {
2090 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2091 if (isVarArg && IsWin64) {
2092 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2093 // shadow reg if callee is a varargs function.
2094 unsigned ShadowReg = 0;
2095 switch (VA.getLocReg()) {
2096 case X86::XMM0: ShadowReg = X86::RCX; break;
2097 case X86::XMM1: ShadowReg = X86::RDX; break;
2098 case X86::XMM2: ShadowReg = X86::R8; break;
2099 case X86::XMM3: ShadowReg = X86::R9; break;
2102 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2104 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2105 assert(VA.isMemLoc());
2106 if (StackPtr.getNode() == 0)
2107 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
2108 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2109 dl, DAG, VA, Flags));
2113 if (!MemOpChains.empty())
2114 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2115 &MemOpChains[0], MemOpChains.size());
2117 // Build a sequence of copy-to-reg nodes chained together with token chain
2118 // and flag operands which copy the outgoing args into registers.
2120 // Tail call byval lowering might overwrite argument registers so in case of
2121 // tail call optimization the copies to registers are lowered later.
2123 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2124 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2125 RegsToPass[i].second, InFlag);
2126 InFlag = Chain.getValue(1);
2129 if (Subtarget->isPICStyleGOT()) {
2130 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2133 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
2134 DAG.getNode(X86ISD::GlobalBaseReg,
2135 DebugLoc(), getPointerTy()),
2137 InFlag = Chain.getValue(1);
2139 // If we are tail calling and generating PIC/GOT style code load the
2140 // address of the callee into ECX. The value in ecx is used as target of
2141 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2142 // for tail calls on PIC/GOT architectures. Normally we would just put the
2143 // address of GOT into ebx and then call target@PLT. But for tail calls
2144 // ebx would be restored (since ebx is callee saved) before jumping to the
2147 // Note: The actual moving to ECX is done further down.
2148 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2149 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2150 !G->getGlobal()->hasProtectedVisibility())
2151 Callee = LowerGlobalAddress(Callee, DAG);
2152 else if (isa<ExternalSymbolSDNode>(Callee))
2153 Callee = LowerExternalSymbol(Callee, DAG);
2157 if (Is64Bit && isVarArg && !IsWin64) {
2158 // From AMD64 ABI document:
2159 // For calls that may call functions that use varargs or stdargs
2160 // (prototype-less calls or calls to functions containing ellipsis (...) in
2161 // the declaration) %al is used as hidden argument to specify the number
2162 // of SSE registers used. The contents of %al do not need to match exactly
2163 // the number of registers, but must be an ubound on the number of SSE
2164 // registers used and is in the range 0 - 8 inclusive.
2166 // Count the number of XMM registers allocated.
2167 static const unsigned XMMArgRegs[] = {
2168 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2169 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2171 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2172 assert((Subtarget->hasXMM() || !NumXMMRegs)
2173 && "SSE registers cannot be used when SSE is disabled");
2175 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
2176 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
2177 InFlag = Chain.getValue(1);
2181 // For tail calls lower the arguments to the 'real' stack slot.
2183 // Force all the incoming stack arguments to be loaded from the stack
2184 // before any new outgoing arguments are stored to the stack, because the
2185 // outgoing stack slots may alias the incoming argument stack slots, and
2186 // the alias isn't otherwise explicit. This is slightly more conservative
2187 // than necessary, because it means that each store effectively depends
2188 // on every argument instead of just those arguments it would clobber.
2189 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2191 SmallVector<SDValue, 8> MemOpChains2;
2194 // Do not flag preceeding copytoreg stuff together with the following stuff.
2196 if (GuaranteedTailCallOpt) {
2197 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2198 CCValAssign &VA = ArgLocs[i];
2201 assert(VA.isMemLoc());
2202 SDValue Arg = OutVals[i];
2203 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2204 // Create frame index.
2205 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2206 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2207 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2208 FIN = DAG.getFrameIndex(FI, getPointerTy());
2210 if (Flags.isByVal()) {
2211 // Copy relative to framepointer.
2212 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2213 if (StackPtr.getNode() == 0)
2214 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2216 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2218 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2222 // Store relative to framepointer.
2223 MemOpChains2.push_back(
2224 DAG.getStore(ArgChain, dl, Arg, FIN,
2225 MachinePointerInfo::getFixedStack(FI),
2231 if (!MemOpChains2.empty())
2232 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2233 &MemOpChains2[0], MemOpChains2.size());
2235 // Copy arguments to their registers.
2236 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2237 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2238 RegsToPass[i].second, InFlag);
2239 InFlag = Chain.getValue(1);
2243 // Store the return address to the appropriate stack slot.
2244 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2248 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2249 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2250 // In the 64-bit large code model, we have to make all calls
2251 // through a register, since the call instruction's 32-bit
2252 // pc-relative offset may not be large enough to hold the whole
2254 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2255 // If the callee is a GlobalAddress node (quite common, every direct call
2256 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2259 // We should use extra load for direct calls to dllimported functions in
2261 const GlobalValue *GV = G->getGlobal();
2262 if (!GV->hasDLLImportLinkage()) {
2263 unsigned char OpFlags = 0;
2265 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2266 // external symbols most go through the PLT in PIC mode. If the symbol
2267 // has hidden or protected visibility, or if it is static or local, then
2268 // we don't need to use the PLT - we can directly call it.
2269 if (Subtarget->isTargetELF() &&
2270 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2271 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2272 OpFlags = X86II::MO_PLT;
2273 } else if (Subtarget->isPICStyleStubAny() &&
2274 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2275 Subtarget->getDarwinVers() < 9) {
2276 // PC-relative references to external symbols should go through $stub,
2277 // unless we're building with the leopard linker or later, which
2278 // automatically synthesizes these stubs.
2279 OpFlags = X86II::MO_DARWIN_STUB;
2282 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2283 G->getOffset(), OpFlags);
2285 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2286 unsigned char OpFlags = 0;
2288 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
2289 // external symbols should go through the PLT.
2290 if (Subtarget->isTargetELF() &&
2291 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2292 OpFlags = X86II::MO_PLT;
2293 } else if (Subtarget->isPICStyleStubAny() &&
2294 Subtarget->getDarwinVers() < 9) {
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.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2305 // Returns a chain & a flag for retval copy to use.
2306 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2307 SmallVector<SDValue, 8> Ops;
2309 if (!IsSibcall && isTailCall) {
2310 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2311 DAG.getIntPtrConstant(0, true), InFlag);
2312 InFlag = Chain.getValue(1);
2315 Ops.push_back(Chain);
2316 Ops.push_back(Callee);
2319 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2321 // Add argument registers to the end of the list so that they are known live
2323 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2324 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2325 RegsToPass[i].second.getValueType()));
2327 // Add an implicit use GOT pointer in EBX.
2328 if (!isTailCall && Subtarget->isPICStyleGOT())
2329 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2331 // Add an implicit use of AL for non-Windows x86 64-bit vararg functions.
2332 if (Is64Bit && isVarArg && !IsWin64)
2333 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2335 if (InFlag.getNode())
2336 Ops.push_back(InFlag);
2340 //// If this is the first return lowered for this function, add the regs
2341 //// to the liveout set for the function.
2342 // This isn't right, although it's probably harmless on x86; liveouts
2343 // should be computed from returns not tail calls. Consider a void
2344 // function making a tail call to a function returning int.
2345 return DAG.getNode(X86ISD::TC_RETURN, dl,
2346 NodeTys, &Ops[0], Ops.size());
2349 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2350 InFlag = Chain.getValue(1);
2352 // Create the CALLSEQ_END node.
2353 unsigned NumBytesForCalleeToPush;
2354 if (Subtarget->IsCalleePop(isVarArg, CallConv))
2355 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2356 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2357 // If this is a call to a struct-return function, the callee
2358 // pops the hidden struct pointer, so we have to push it back.
2359 // This is common for Darwin/X86, Linux & Mingw32 targets.
2360 NumBytesForCalleeToPush = 4;
2362 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2364 // Returns a flag for retval copy to use.
2366 Chain = DAG.getCALLSEQ_END(Chain,
2367 DAG.getIntPtrConstant(NumBytes, true),
2368 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2371 InFlag = Chain.getValue(1);
2374 // Handle result values, copying them out of physregs into vregs that we
2376 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2377 Ins, dl, DAG, InVals);
2381 //===----------------------------------------------------------------------===//
2382 // Fast Calling Convention (tail call) implementation
2383 //===----------------------------------------------------------------------===//
2385 // Like std call, callee cleans arguments, convention except that ECX is
2386 // reserved for storing the tail called function address. Only 2 registers are
2387 // free for argument passing (inreg). Tail call optimization is performed
2389 // * tailcallopt is enabled
2390 // * caller/callee are fastcc
2391 // On X86_64 architecture with GOT-style position independent code only local
2392 // (within module) calls are supported at the moment.
2393 // To keep the stack aligned according to platform abi the function
2394 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2395 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2396 // If a tail called function callee has more arguments than the caller the
2397 // caller needs to make sure that there is room to move the RETADDR to. This is
2398 // achieved by reserving an area the size of the argument delta right after the
2399 // original REtADDR, but before the saved framepointer or the spilled registers
2400 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2412 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2413 /// for a 16 byte align requirement.
2415 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2416 SelectionDAG& DAG) const {
2417 MachineFunction &MF = DAG.getMachineFunction();
2418 const TargetMachine &TM = MF.getTarget();
2419 const TargetFrameLowering &TFI = *TM.getFrameLowering();
2420 unsigned StackAlignment = TFI.getStackAlignment();
2421 uint64_t AlignMask = StackAlignment - 1;
2422 int64_t Offset = StackSize;
2423 uint64_t SlotSize = TD->getPointerSize();
2424 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2425 // Number smaller than 12 so just add the difference.
2426 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2428 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2429 Offset = ((~AlignMask) & Offset) + StackAlignment +
2430 (StackAlignment-SlotSize);
2435 /// MatchingStackOffset - Return true if the given stack call argument is
2436 /// already available in the same position (relatively) of the caller's
2437 /// incoming argument stack.
2439 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2440 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2441 const X86InstrInfo *TII) {
2442 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2444 if (Arg.getOpcode() == ISD::CopyFromReg) {
2445 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2446 if (!TargetRegisterInfo::isVirtualRegister(VR))
2448 MachineInstr *Def = MRI->getVRegDef(VR);
2451 if (!Flags.isByVal()) {
2452 if (!TII->isLoadFromStackSlot(Def, FI))
2455 unsigned Opcode = Def->getOpcode();
2456 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2457 Def->getOperand(1).isFI()) {
2458 FI = Def->getOperand(1).getIndex();
2459 Bytes = Flags.getByValSize();
2463 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2464 if (Flags.isByVal())
2465 // ByVal argument is passed in as a pointer but it's now being
2466 // dereferenced. e.g.
2467 // define @foo(%struct.X* %A) {
2468 // tail call @bar(%struct.X* byval %A)
2471 SDValue Ptr = Ld->getBasePtr();
2472 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2475 FI = FINode->getIndex();
2479 assert(FI != INT_MAX);
2480 if (!MFI->isFixedObjectIndex(FI))
2482 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2485 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2486 /// for tail call optimization. Targets which want to do tail call
2487 /// optimization should implement this function.
2489 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2490 CallingConv::ID CalleeCC,
2492 bool isCalleeStructRet,
2493 bool isCallerStructRet,
2494 const SmallVectorImpl<ISD::OutputArg> &Outs,
2495 const SmallVectorImpl<SDValue> &OutVals,
2496 const SmallVectorImpl<ISD::InputArg> &Ins,
2497 SelectionDAG& DAG) const {
2498 if (!IsTailCallConvention(CalleeCC) &&
2499 CalleeCC != CallingConv::C)
2502 // If -tailcallopt is specified, make fastcc functions tail-callable.
2503 const MachineFunction &MF = DAG.getMachineFunction();
2504 const Function *CallerF = DAG.getMachineFunction().getFunction();
2505 CallingConv::ID CallerCC = CallerF->getCallingConv();
2506 bool CCMatch = CallerCC == CalleeCC;
2508 if (GuaranteedTailCallOpt) {
2509 if (IsTailCallConvention(CalleeCC) && CCMatch)
2514 // Look for obvious safe cases to perform tail call optimization that do not
2515 // require ABI changes. This is what gcc calls sibcall.
2517 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2518 // emit a special epilogue.
2519 if (RegInfo->needsStackRealignment(MF))
2522 // Do not sibcall optimize vararg calls unless the call site is not passing
2524 if (isVarArg && !Outs.empty())
2527 // Also avoid sibcall optimization if either caller or callee uses struct
2528 // return semantics.
2529 if (isCalleeStructRet || isCallerStructRet)
2532 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2533 // Therefore if it's not used by the call it is not safe to optimize this into
2535 bool Unused = false;
2536 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2543 SmallVector<CCValAssign, 16> RVLocs;
2544 CCState CCInfo(CalleeCC, false, getTargetMachine(),
2545 RVLocs, *DAG.getContext());
2546 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2547 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2548 CCValAssign &VA = RVLocs[i];
2549 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2554 // If the calling conventions do not match, then we'd better make sure the
2555 // results are returned in the same way as what the caller expects.
2557 SmallVector<CCValAssign, 16> RVLocs1;
2558 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
2559 RVLocs1, *DAG.getContext());
2560 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2562 SmallVector<CCValAssign, 16> RVLocs2;
2563 CCState CCInfo2(CallerCC, false, getTargetMachine(),
2564 RVLocs2, *DAG.getContext());
2565 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2567 if (RVLocs1.size() != RVLocs2.size())
2569 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2570 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2572 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2574 if (RVLocs1[i].isRegLoc()) {
2575 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2578 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2584 // If the callee takes no arguments then go on to check the results of the
2586 if (!Outs.empty()) {
2587 // Check if stack adjustment is needed. For now, do not do this if any
2588 // argument is passed on the stack.
2589 SmallVector<CCValAssign, 16> ArgLocs;
2590 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2591 ArgLocs, *DAG.getContext());
2593 // Allocate shadow area for Win64
2594 if (Subtarget->isTargetWin64()) {
2595 CCInfo.AllocateStack(32, 8);
2598 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2599 if (CCInfo.getNextStackOffset()) {
2600 MachineFunction &MF = DAG.getMachineFunction();
2601 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2604 // Check if the arguments are already laid out in the right way as
2605 // the caller's fixed stack objects.
2606 MachineFrameInfo *MFI = MF.getFrameInfo();
2607 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2608 const X86InstrInfo *TII =
2609 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2610 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2611 CCValAssign &VA = ArgLocs[i];
2612 SDValue Arg = OutVals[i];
2613 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2614 if (VA.getLocInfo() == CCValAssign::Indirect)
2616 if (!VA.isRegLoc()) {
2617 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2624 // If the tailcall address may be in a register, then make sure it's
2625 // possible to register allocate for it. In 32-bit, the call address can
2626 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2627 // callee-saved registers are restored. These happen to be the same
2628 // registers used to pass 'inreg' arguments so watch out for those.
2629 if (!Subtarget->is64Bit() &&
2630 !isa<GlobalAddressSDNode>(Callee) &&
2631 !isa<ExternalSymbolSDNode>(Callee)) {
2632 unsigned NumInRegs = 0;
2633 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2634 CCValAssign &VA = ArgLocs[i];
2637 unsigned Reg = VA.getLocReg();
2640 case X86::EAX: case X86::EDX: case X86::ECX:
2641 if (++NumInRegs == 3)
2649 // An stdcall caller is expected to clean up its arguments; the callee
2650 // isn't going to do that.
2651 if (!CCMatch && CallerCC==CallingConv::X86_StdCall)
2658 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
2659 return X86::createFastISel(funcInfo);
2663 //===----------------------------------------------------------------------===//
2664 // Other Lowering Hooks
2665 //===----------------------------------------------------------------------===//
2667 static bool MayFoldLoad(SDValue Op) {
2668 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
2671 static bool MayFoldIntoStore(SDValue Op) {
2672 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
2675 static bool isTargetShuffle(unsigned Opcode) {
2677 default: return false;
2678 case X86ISD::PSHUFD:
2679 case X86ISD::PSHUFHW:
2680 case X86ISD::PSHUFLW:
2681 case X86ISD::SHUFPD:
2682 case X86ISD::PALIGN:
2683 case X86ISD::SHUFPS:
2684 case X86ISD::MOVLHPS:
2685 case X86ISD::MOVLHPD:
2686 case X86ISD::MOVHLPS:
2687 case X86ISD::MOVLPS:
2688 case X86ISD::MOVLPD:
2689 case X86ISD::MOVSHDUP:
2690 case X86ISD::MOVSLDUP:
2691 case X86ISD::MOVDDUP:
2694 case X86ISD::UNPCKLPS:
2695 case X86ISD::UNPCKLPD:
2696 case X86ISD::PUNPCKLWD:
2697 case X86ISD::PUNPCKLBW:
2698 case X86ISD::PUNPCKLDQ:
2699 case X86ISD::PUNPCKLQDQ:
2700 case X86ISD::UNPCKHPS:
2701 case X86ISD::UNPCKHPD:
2702 case X86ISD::PUNPCKHWD:
2703 case X86ISD::PUNPCKHBW:
2704 case X86ISD::PUNPCKHDQ:
2705 case X86ISD::PUNPCKHQDQ:
2711 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2712 SDValue V1, SelectionDAG &DAG) {
2714 default: llvm_unreachable("Unknown x86 shuffle node");
2715 case X86ISD::MOVSHDUP:
2716 case X86ISD::MOVSLDUP:
2717 case X86ISD::MOVDDUP:
2718 return DAG.getNode(Opc, dl, VT, V1);
2724 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2725 SDValue V1, unsigned TargetMask, SelectionDAG &DAG) {
2727 default: llvm_unreachable("Unknown x86 shuffle node");
2728 case X86ISD::PSHUFD:
2729 case X86ISD::PSHUFHW:
2730 case X86ISD::PSHUFLW:
2731 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
2737 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2738 SDValue V1, SDValue V2, unsigned TargetMask, SelectionDAG &DAG) {
2740 default: llvm_unreachable("Unknown x86 shuffle node");
2741 case X86ISD::PALIGN:
2742 case X86ISD::SHUFPD:
2743 case X86ISD::SHUFPS:
2744 return DAG.getNode(Opc, dl, VT, V1, V2,
2745 DAG.getConstant(TargetMask, MVT::i8));
2750 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2751 SDValue V1, SDValue V2, SelectionDAG &DAG) {
2753 default: llvm_unreachable("Unknown x86 shuffle node");
2754 case X86ISD::MOVLHPS:
2755 case X86ISD::MOVLHPD:
2756 case X86ISD::MOVHLPS:
2757 case X86ISD::MOVLPS:
2758 case X86ISD::MOVLPD:
2761 case X86ISD::UNPCKLPS:
2762 case X86ISD::UNPCKLPD:
2763 case X86ISD::PUNPCKLWD:
2764 case X86ISD::PUNPCKLBW:
2765 case X86ISD::PUNPCKLDQ:
2766 case X86ISD::PUNPCKLQDQ:
2767 case X86ISD::UNPCKHPS:
2768 case X86ISD::UNPCKHPD:
2769 case X86ISD::PUNPCKHWD:
2770 case X86ISD::PUNPCKHBW:
2771 case X86ISD::PUNPCKHDQ:
2772 case X86ISD::PUNPCKHQDQ:
2773 return DAG.getNode(Opc, dl, VT, V1, V2);
2778 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2779 MachineFunction &MF = DAG.getMachineFunction();
2780 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2781 int ReturnAddrIndex = FuncInfo->getRAIndex();
2783 if (ReturnAddrIndex == 0) {
2784 // Set up a frame object for the return address.
2785 uint64_t SlotSize = TD->getPointerSize();
2786 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2788 FuncInfo->setRAIndex(ReturnAddrIndex);
2791 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2795 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2796 bool hasSymbolicDisplacement) {
2797 // Offset should fit into 32 bit immediate field.
2798 if (!isInt<32>(Offset))
2801 // If we don't have a symbolic displacement - we don't have any extra
2803 if (!hasSymbolicDisplacement)
2806 // FIXME: Some tweaks might be needed for medium code model.
2807 if (M != CodeModel::Small && M != CodeModel::Kernel)
2810 // For small code model we assume that latest object is 16MB before end of 31
2811 // bits boundary. We may also accept pretty large negative constants knowing
2812 // that all objects are in the positive half of address space.
2813 if (M == CodeModel::Small && Offset < 16*1024*1024)
2816 // For kernel code model we know that all object resist in the negative half
2817 // of 32bits address space. We may not accept negative offsets, since they may
2818 // be just off and we may accept pretty large positive ones.
2819 if (M == CodeModel::Kernel && Offset > 0)
2825 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2826 /// specific condition code, returning the condition code and the LHS/RHS of the
2827 /// comparison to make.
2828 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2829 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2831 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2832 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2833 // X > -1 -> X == 0, jump !sign.
2834 RHS = DAG.getConstant(0, RHS.getValueType());
2835 return X86::COND_NS;
2836 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2837 // X < 0 -> X == 0, jump on sign.
2839 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2841 RHS = DAG.getConstant(0, RHS.getValueType());
2842 return X86::COND_LE;
2846 switch (SetCCOpcode) {
2847 default: llvm_unreachable("Invalid integer condition!");
2848 case ISD::SETEQ: return X86::COND_E;
2849 case ISD::SETGT: return X86::COND_G;
2850 case ISD::SETGE: return X86::COND_GE;
2851 case ISD::SETLT: return X86::COND_L;
2852 case ISD::SETLE: return X86::COND_LE;
2853 case ISD::SETNE: return X86::COND_NE;
2854 case ISD::SETULT: return X86::COND_B;
2855 case ISD::SETUGT: return X86::COND_A;
2856 case ISD::SETULE: return X86::COND_BE;
2857 case ISD::SETUGE: return X86::COND_AE;
2861 // First determine if it is required or is profitable to flip the operands.
2863 // If LHS is a foldable load, but RHS is not, flip the condition.
2864 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
2865 !ISD::isNON_EXTLoad(RHS.getNode())) {
2866 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2867 std::swap(LHS, RHS);
2870 switch (SetCCOpcode) {
2876 std::swap(LHS, RHS);
2880 // On a floating point condition, the flags are set as follows:
2882 // 0 | 0 | 0 | X > Y
2883 // 0 | 0 | 1 | X < Y
2884 // 1 | 0 | 0 | X == Y
2885 // 1 | 1 | 1 | unordered
2886 switch (SetCCOpcode) {
2887 default: llvm_unreachable("Condcode should be pre-legalized away");
2889 case ISD::SETEQ: return X86::COND_E;
2890 case ISD::SETOLT: // flipped
2892 case ISD::SETGT: return X86::COND_A;
2893 case ISD::SETOLE: // flipped
2895 case ISD::SETGE: return X86::COND_AE;
2896 case ISD::SETUGT: // flipped
2898 case ISD::SETLT: return X86::COND_B;
2899 case ISD::SETUGE: // flipped
2901 case ISD::SETLE: return X86::COND_BE;
2903 case ISD::SETNE: return X86::COND_NE;
2904 case ISD::SETUO: return X86::COND_P;
2905 case ISD::SETO: return X86::COND_NP;
2907 case ISD::SETUNE: return X86::COND_INVALID;
2911 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2912 /// code. Current x86 isa includes the following FP cmov instructions:
2913 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2914 static bool hasFPCMov(unsigned X86CC) {
2930 /// isFPImmLegal - Returns true if the target can instruction select the
2931 /// specified FP immediate natively. If false, the legalizer will
2932 /// materialize the FP immediate as a load from a constant pool.
2933 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2934 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2935 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2941 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2942 /// the specified range (L, H].
2943 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2944 return (Val < 0) || (Val >= Low && Val < Hi);
2947 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2948 /// specified value.
2949 static bool isUndefOrEqual(int Val, int CmpVal) {
2950 if (Val < 0 || Val == CmpVal)
2955 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2956 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2957 /// the second operand.
2958 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2959 if (VT == MVT::v4f32 || VT == MVT::v4i32 )
2960 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2961 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2962 return (Mask[0] < 2 && Mask[1] < 2);
2966 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2967 SmallVector<int, 8> M;
2969 return ::isPSHUFDMask(M, N->getValueType(0));
2972 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2973 /// is suitable for input to PSHUFHW.
2974 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2975 if (VT != MVT::v8i16)
2978 // Lower quadword copied in order or undef.
2979 for (int i = 0; i != 4; ++i)
2980 if (Mask[i] >= 0 && Mask[i] != i)
2983 // Upper quadword shuffled.
2984 for (int i = 4; i != 8; ++i)
2985 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2991 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2992 SmallVector<int, 8> M;
2994 return ::isPSHUFHWMask(M, N->getValueType(0));
2997 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2998 /// is suitable for input to PSHUFLW.
2999 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3000 if (VT != MVT::v8i16)
3003 // Upper quadword copied in order.
3004 for (int i = 4; i != 8; ++i)
3005 if (Mask[i] >= 0 && Mask[i] != i)
3008 // Lower quadword shuffled.
3009 for (int i = 0; i != 4; ++i)
3016 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
3017 SmallVector<int, 8> M;
3019 return ::isPSHUFLWMask(M, N->getValueType(0));
3022 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
3023 /// is suitable for input to PALIGNR.
3024 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
3026 int i, e = VT.getVectorNumElements();
3028 // Do not handle v2i64 / v2f64 shuffles with palignr.
3029 if (e < 4 || !hasSSSE3)
3032 for (i = 0; i != e; ++i)
3036 // All undef, not a palignr.
3040 // Determine if it's ok to perform a palignr with only the LHS, since we
3041 // don't have access to the actual shuffle elements to see if RHS is undef.
3042 bool Unary = Mask[i] < (int)e;
3043 bool NeedsUnary = false;
3045 int s = Mask[i] - i;
3047 // Check the rest of the elements to see if they are consecutive.
3048 for (++i; i != e; ++i) {
3053 Unary = Unary && (m < (int)e);
3054 NeedsUnary = NeedsUnary || (m < s);
3056 if (NeedsUnary && !Unary)
3058 if (Unary && m != ((s+i) & (e-1)))
3060 if (!Unary && m != (s+i))
3066 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
3067 SmallVector<int, 8> M;
3069 return ::isPALIGNRMask(M, N->getValueType(0), true);
3072 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
3073 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
3074 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3075 int NumElems = VT.getVectorNumElements();
3076 if (NumElems != 2 && NumElems != 4)
3079 int Half = NumElems / 2;
3080 for (int i = 0; i < Half; ++i)
3081 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3083 for (int i = Half; i < NumElems; ++i)
3084 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3090 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
3091 SmallVector<int, 8> M;
3093 return ::isSHUFPMask(M, N->getValueType(0));
3096 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
3097 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
3098 /// half elements to come from vector 1 (which would equal the dest.) and
3099 /// the upper half to come from vector 2.
3100 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3101 int NumElems = VT.getVectorNumElements();
3103 if (NumElems != 2 && NumElems != 4)
3106 int Half = NumElems / 2;
3107 for (int i = 0; i < Half; ++i)
3108 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3110 for (int i = Half; i < NumElems; ++i)
3111 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3116 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
3117 SmallVector<int, 8> M;
3119 return isCommutedSHUFPMask(M, N->getValueType(0));
3122 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
3123 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
3124 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
3125 if (N->getValueType(0).getVectorNumElements() != 4)
3128 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
3129 return isUndefOrEqual(N->getMaskElt(0), 6) &&
3130 isUndefOrEqual(N->getMaskElt(1), 7) &&
3131 isUndefOrEqual(N->getMaskElt(2), 2) &&
3132 isUndefOrEqual(N->getMaskElt(3), 3);
3135 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
3136 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
3138 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
3139 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3144 return isUndefOrEqual(N->getMaskElt(0), 2) &&
3145 isUndefOrEqual(N->getMaskElt(1), 3) &&
3146 isUndefOrEqual(N->getMaskElt(2), 2) &&
3147 isUndefOrEqual(N->getMaskElt(3), 3);
3150 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
3151 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
3152 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
3153 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3155 if (NumElems != 2 && NumElems != 4)
3158 for (unsigned i = 0; i < NumElems/2; ++i)
3159 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
3162 for (unsigned i = NumElems/2; i < NumElems; ++i)
3163 if (!isUndefOrEqual(N->getMaskElt(i), i))
3169 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
3170 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
3171 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
3172 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3174 if (NumElems != 2 && NumElems != 4)
3177 for (unsigned i = 0; i < NumElems/2; ++i)
3178 if (!isUndefOrEqual(N->getMaskElt(i), i))
3181 for (unsigned i = 0; i < NumElems/2; ++i)
3182 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
3188 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
3189 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
3190 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3191 bool V2IsSplat = false) {
3192 int NumElts = VT.getVectorNumElements();
3193 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3196 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
3198 int BitI1 = Mask[i+1];
3199 if (!isUndefOrEqual(BitI, j))
3202 if (!isUndefOrEqual(BitI1, NumElts))
3205 if (!isUndefOrEqual(BitI1, j + NumElts))
3212 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3213 SmallVector<int, 8> M;
3215 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
3218 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
3219 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
3220 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
3221 bool V2IsSplat = false) {
3222 int NumElts = VT.getVectorNumElements();
3223 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3226 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
3228 int BitI1 = Mask[i+1];
3229 if (!isUndefOrEqual(BitI, j + NumElts/2))
3232 if (isUndefOrEqual(BitI1, NumElts))
3235 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
3242 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3243 SmallVector<int, 8> M;
3245 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
3248 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
3249 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
3251 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3252 int NumElems = VT.getVectorNumElements();
3253 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3256 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
3258 int BitI1 = Mask[i+1];
3259 if (!isUndefOrEqual(BitI, j))
3261 if (!isUndefOrEqual(BitI1, j))
3267 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
3268 SmallVector<int, 8> M;
3270 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
3273 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
3274 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
3276 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3277 int NumElems = VT.getVectorNumElements();
3278 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3281 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
3283 int BitI1 = Mask[i+1];
3284 if (!isUndefOrEqual(BitI, j))
3286 if (!isUndefOrEqual(BitI1, j))
3292 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
3293 SmallVector<int, 8> M;
3295 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3298 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3299 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3300 /// MOVSD, and MOVD, i.e. setting the lowest element.
3301 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3302 if (VT.getVectorElementType().getSizeInBits() < 32)
3305 int NumElts = VT.getVectorNumElements();
3307 if (!isUndefOrEqual(Mask[0], NumElts))
3310 for (int i = 1; i < NumElts; ++i)
3311 if (!isUndefOrEqual(Mask[i], i))
3317 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3318 SmallVector<int, 8> M;
3320 return ::isMOVLMask(M, N->getValueType(0));
3323 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3324 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3325 /// element of vector 2 and the other elements to come from vector 1 in order.
3326 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3327 bool V2IsSplat = false, bool V2IsUndef = false) {
3328 int NumOps = VT.getVectorNumElements();
3329 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3332 if (!isUndefOrEqual(Mask[0], 0))
3335 for (int i = 1; i < NumOps; ++i)
3336 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3337 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3338 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3344 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3345 bool V2IsUndef = false) {
3346 SmallVector<int, 8> M;
3348 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3351 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3352 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3353 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3354 if (N->getValueType(0).getVectorNumElements() != 4)
3357 // Expect 1, 1, 3, 3
3358 for (unsigned i = 0; i < 2; ++i) {
3359 int Elt = N->getMaskElt(i);
3360 if (Elt >= 0 && Elt != 1)
3365 for (unsigned i = 2; i < 4; ++i) {
3366 int Elt = N->getMaskElt(i);
3367 if (Elt >= 0 && Elt != 3)
3372 // Don't use movshdup if it can be done with a shufps.
3373 // FIXME: verify that matching u, u, 3, 3 is what we want.
3377 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3378 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3379 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3380 if (N->getValueType(0).getVectorNumElements() != 4)
3383 // Expect 0, 0, 2, 2
3384 for (unsigned i = 0; i < 2; ++i)
3385 if (N->getMaskElt(i) > 0)
3389 for (unsigned i = 2; i < 4; ++i) {
3390 int Elt = N->getMaskElt(i);
3391 if (Elt >= 0 && Elt != 2)
3396 // Don't use movsldup if it can be done with a shufps.
3400 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3401 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3402 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3403 int e = N->getValueType(0).getVectorNumElements() / 2;
3405 for (int i = 0; i < e; ++i)
3406 if (!isUndefOrEqual(N->getMaskElt(i), i))
3408 for (int i = 0; i < e; ++i)
3409 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3414 /// isVEXTRACTF128Index - Return true if the specified
3415 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
3416 /// suitable for input to VEXTRACTF128.
3417 bool X86::isVEXTRACTF128Index(SDNode *N) {
3418 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3421 // The index should be aligned on a 128-bit boundary.
3423 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3425 unsigned VL = N->getValueType(0).getVectorNumElements();
3426 unsigned VBits = N->getValueType(0).getSizeInBits();
3427 unsigned ElSize = VBits / VL;
3428 bool Result = (Index * ElSize) % 128 == 0;
3433 /// isVINSERTF128Index - Return true if the specified INSERT_SUBVECTOR
3434 /// operand specifies a subvector insert that is suitable for input to
3436 bool X86::isVINSERTF128Index(SDNode *N) {
3437 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3440 // The index should be aligned on a 128-bit boundary.
3442 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3444 unsigned VL = N->getValueType(0).getVectorNumElements();
3445 unsigned VBits = N->getValueType(0).getSizeInBits();
3446 unsigned ElSize = VBits / VL;
3447 bool Result = (Index * ElSize) % 128 == 0;
3452 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3453 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3454 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3455 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3456 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3458 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3460 for (int i = 0; i < NumOperands; ++i) {
3461 int Val = SVOp->getMaskElt(NumOperands-i-1);
3462 if (Val < 0) Val = 0;
3463 if (Val >= NumOperands) Val -= NumOperands;
3465 if (i != NumOperands - 1)
3471 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3472 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3473 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3474 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3476 // 8 nodes, but we only care about the last 4.
3477 for (unsigned i = 7; i >= 4; --i) {
3478 int Val = SVOp->getMaskElt(i);
3487 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3488 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3489 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3490 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3492 // 8 nodes, but we only care about the first 4.
3493 for (int i = 3; i >= 0; --i) {
3494 int Val = SVOp->getMaskElt(i);
3503 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3504 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3505 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3506 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3507 EVT VVT = N->getValueType(0);
3508 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3512 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3513 Val = SVOp->getMaskElt(i);
3517 return (Val - i) * EltSize;
3520 /// getExtractVEXTRACTF128Immediate - Return the appropriate immediate
3521 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
3523 unsigned X86::getExtractVEXTRACTF128Immediate(SDNode *N) {
3524 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3525 llvm_unreachable("Illegal extract subvector for VEXTRACTF128");
3528 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3530 EVT VecVT = N->getOperand(0).getValueType();
3531 EVT ElVT = VecVT.getVectorElementType();
3533 unsigned NumElemsPerChunk = 128 / ElVT.getSizeInBits();
3535 return Index / NumElemsPerChunk;
3538 /// getInsertVINSERTF128Immediate - Return the appropriate immediate
3539 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
3541 unsigned X86::getInsertVINSERTF128Immediate(SDNode *N) {
3542 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3543 llvm_unreachable("Illegal insert subvector for VINSERTF128");
3546 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3548 EVT VecVT = N->getValueType(0);
3549 EVT ElVT = VecVT.getVectorElementType();
3551 unsigned NumElemsPerChunk = 128 / ElVT.getSizeInBits();
3553 return Index / NumElemsPerChunk;
3556 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3558 bool X86::isZeroNode(SDValue Elt) {
3559 return ((isa<ConstantSDNode>(Elt) &&
3560 cast<ConstantSDNode>(Elt)->isNullValue()) ||
3561 (isa<ConstantFPSDNode>(Elt) &&
3562 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3565 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3566 /// their permute mask.
3567 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3568 SelectionDAG &DAG) {
3569 EVT VT = SVOp->getValueType(0);
3570 unsigned NumElems = VT.getVectorNumElements();
3571 SmallVector<int, 8> MaskVec;
3573 for (unsigned i = 0; i != NumElems; ++i) {
3574 int idx = SVOp->getMaskElt(i);
3576 MaskVec.push_back(idx);
3577 else if (idx < (int)NumElems)
3578 MaskVec.push_back(idx + NumElems);
3580 MaskVec.push_back(idx - NumElems);
3582 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3583 SVOp->getOperand(0), &MaskVec[0]);
3586 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3587 /// the two vector operands have swapped position.
3588 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3589 unsigned NumElems = VT.getVectorNumElements();
3590 for (unsigned i = 0; i != NumElems; ++i) {
3594 else if (idx < (int)NumElems)
3595 Mask[i] = idx + NumElems;
3597 Mask[i] = idx - NumElems;
3601 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3602 /// match movhlps. The lower half elements should come from upper half of
3603 /// V1 (and in order), and the upper half elements should come from the upper
3604 /// half of V2 (and in order).
3605 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3606 if (Op->getValueType(0).getVectorNumElements() != 4)
3608 for (unsigned i = 0, e = 2; i != e; ++i)
3609 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3611 for (unsigned i = 2; i != 4; ++i)
3612 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3617 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3618 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3620 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3621 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3623 N = N->getOperand(0).getNode();
3624 if (!ISD::isNON_EXTLoad(N))
3627 *LD = cast<LoadSDNode>(N);
3631 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3632 /// match movlp{s|d}. The lower half elements should come from lower half of
3633 /// V1 (and in order), and the upper half elements should come from the upper
3634 /// half of V2 (and in order). And since V1 will become the source of the
3635 /// MOVLP, it must be either a vector load or a scalar load to vector.
3636 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3637 ShuffleVectorSDNode *Op) {
3638 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3640 // Is V2 is a vector load, don't do this transformation. We will try to use
3641 // load folding shufps op.
3642 if (ISD::isNON_EXTLoad(V2))
3645 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3647 if (NumElems != 2 && NumElems != 4)
3649 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3650 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3652 for (unsigned i = NumElems/2; i != NumElems; ++i)
3653 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3658 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3660 static bool isSplatVector(SDNode *N) {
3661 if (N->getOpcode() != ISD::BUILD_VECTOR)
3664 SDValue SplatValue = N->getOperand(0);
3665 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3666 if (N->getOperand(i) != SplatValue)
3671 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3672 /// to an zero vector.
3673 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3674 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3675 SDValue V1 = N->getOperand(0);
3676 SDValue V2 = N->getOperand(1);
3677 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3678 for (unsigned i = 0; i != NumElems; ++i) {
3679 int Idx = N->getMaskElt(i);
3680 if (Idx >= (int)NumElems) {
3681 unsigned Opc = V2.getOpcode();
3682 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3684 if (Opc != ISD::BUILD_VECTOR ||
3685 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3687 } else if (Idx >= 0) {
3688 unsigned Opc = V1.getOpcode();
3689 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3691 if (Opc != ISD::BUILD_VECTOR ||
3692 !X86::isZeroNode(V1.getOperand(Idx)))
3699 /// getZeroVector - Returns a vector of specified type with all zero elements.
3701 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3703 assert(VT.isVector() && "Expected a vector type");
3705 // Always build SSE zero vectors as <4 x i32> bitcasted
3706 // to their dest type. This ensures they get CSE'd.
3708 if (VT.getSizeInBits() == 128) { // SSE
3709 if (HasSSE2) { // SSE2
3710 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3711 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3713 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3714 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3716 } else if (VT.getSizeInBits() == 256) { // AVX
3717 // 256-bit logic and arithmetic instructions in AVX are
3718 // all floating-point, no support for integer ops. Default
3719 // to emitting fp zeroed vectors then.
3720 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3721 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
3722 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops, 8);
3724 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
3727 /// getOnesVector - Returns a vector of specified type with all bits set.
3729 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3730 assert(VT.isVector() && "Expected a vector type");
3732 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3733 // type. This ensures they get CSE'd.
3734 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3736 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3737 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
3741 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3742 /// that point to V2 points to its first element.
3743 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3744 EVT VT = SVOp->getValueType(0);
3745 unsigned NumElems = VT.getVectorNumElements();
3747 bool Changed = false;
3748 SmallVector<int, 8> MaskVec;
3749 SVOp->getMask(MaskVec);
3751 for (unsigned i = 0; i != NumElems; ++i) {
3752 if (MaskVec[i] > (int)NumElems) {
3753 MaskVec[i] = NumElems;
3758 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3759 SVOp->getOperand(1), &MaskVec[0]);
3760 return SDValue(SVOp, 0);
3763 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3764 /// operation of specified width.
3765 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3767 unsigned NumElems = VT.getVectorNumElements();
3768 SmallVector<int, 8> Mask;
3769 Mask.push_back(NumElems);
3770 for (unsigned i = 1; i != NumElems; ++i)
3772 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3775 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3776 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3778 unsigned NumElems = VT.getVectorNumElements();
3779 SmallVector<int, 8> Mask;
3780 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3782 Mask.push_back(i + NumElems);
3784 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3787 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3788 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3790 unsigned NumElems = VT.getVectorNumElements();
3791 unsigned Half = NumElems/2;
3792 SmallVector<int, 8> Mask;
3793 for (unsigned i = 0; i != Half; ++i) {
3794 Mask.push_back(i + Half);
3795 Mask.push_back(i + NumElems + Half);
3797 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3800 /// PromoteSplat - Promote a splat of v4i32, v8i16 or v16i8 to v4f32.
3801 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
3802 EVT PVT = MVT::v4f32;
3803 EVT VT = SV->getValueType(0);
3804 DebugLoc dl = SV->getDebugLoc();
3805 SDValue V1 = SV->getOperand(0);
3806 int NumElems = VT.getVectorNumElements();
3807 int EltNo = SV->getSplatIndex();
3809 // unpack elements to the correct location
3810 while (NumElems > 4) {
3811 if (EltNo < NumElems/2) {
3812 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3814 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3815 EltNo -= NumElems/2;
3820 // Perform the splat.
3821 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3822 V1 = DAG.getNode(ISD::BITCAST, dl, PVT, V1);
3823 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3824 return DAG.getNode(ISD::BITCAST, dl, VT, V1);
3827 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3828 /// vector of zero or undef vector. This produces a shuffle where the low
3829 /// element of V2 is swizzled into the zero/undef vector, landing at element
3830 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3831 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3832 bool isZero, bool HasSSE2,
3833 SelectionDAG &DAG) {
3834 EVT VT = V2.getValueType();
3836 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3837 unsigned NumElems = VT.getVectorNumElements();
3838 SmallVector<int, 16> MaskVec;
3839 for (unsigned i = 0; i != NumElems; ++i)
3840 // If this is the insertion idx, put the low elt of V2 here.
3841 MaskVec.push_back(i == Idx ? NumElems : i);
3842 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3845 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
3846 /// element of the result of the vector shuffle.
3847 SDValue getShuffleScalarElt(SDNode *N, int Index, SelectionDAG &DAG,
3850 return SDValue(); // Limit search depth.
3852 SDValue V = SDValue(N, 0);
3853 EVT VT = V.getValueType();
3854 unsigned Opcode = V.getOpcode();
3856 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
3857 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
3858 Index = SV->getMaskElt(Index);
3861 return DAG.getUNDEF(VT.getVectorElementType());
3863 int NumElems = VT.getVectorNumElements();
3864 SDValue NewV = (Index < NumElems) ? SV->getOperand(0) : SV->getOperand(1);
3865 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG, Depth+1);
3868 // Recurse into target specific vector shuffles to find scalars.
3869 if (isTargetShuffle(Opcode)) {
3870 int NumElems = VT.getVectorNumElements();
3871 SmallVector<unsigned, 16> ShuffleMask;
3875 case X86ISD::SHUFPS:
3876 case X86ISD::SHUFPD:
3877 ImmN = N->getOperand(N->getNumOperands()-1);
3878 DecodeSHUFPSMask(NumElems,
3879 cast<ConstantSDNode>(ImmN)->getZExtValue(),
3882 case X86ISD::PUNPCKHBW:
3883 case X86ISD::PUNPCKHWD:
3884 case X86ISD::PUNPCKHDQ:
3885 case X86ISD::PUNPCKHQDQ:
3886 DecodePUNPCKHMask(NumElems, ShuffleMask);
3888 case X86ISD::UNPCKHPS:
3889 case X86ISD::UNPCKHPD:
3890 DecodeUNPCKHPMask(NumElems, ShuffleMask);
3892 case X86ISD::PUNPCKLBW:
3893 case X86ISD::PUNPCKLWD:
3894 case X86ISD::PUNPCKLDQ:
3895 case X86ISD::PUNPCKLQDQ:
3896 DecodePUNPCKLMask(NumElems, ShuffleMask);
3898 case X86ISD::UNPCKLPS:
3899 case X86ISD::UNPCKLPD:
3900 DecodeUNPCKLPMask(NumElems, ShuffleMask);
3902 case X86ISD::MOVHLPS:
3903 DecodeMOVHLPSMask(NumElems, ShuffleMask);
3905 case X86ISD::MOVLHPS:
3906 DecodeMOVLHPSMask(NumElems, ShuffleMask);
3908 case X86ISD::PSHUFD:
3909 ImmN = N->getOperand(N->getNumOperands()-1);
3910 DecodePSHUFMask(NumElems,
3911 cast<ConstantSDNode>(ImmN)->getZExtValue(),
3914 case X86ISD::PSHUFHW:
3915 ImmN = N->getOperand(N->getNumOperands()-1);
3916 DecodePSHUFHWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
3919 case X86ISD::PSHUFLW:
3920 ImmN = N->getOperand(N->getNumOperands()-1);
3921 DecodePSHUFLWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
3925 case X86ISD::MOVSD: {
3926 // The index 0 always comes from the first element of the second source,
3927 // this is why MOVSS and MOVSD are used in the first place. The other
3928 // elements come from the other positions of the first source vector.
3929 unsigned OpNum = (Index == 0) ? 1 : 0;
3930 return getShuffleScalarElt(V.getOperand(OpNum).getNode(), Index, DAG,
3934 assert("not implemented for target shuffle node");
3938 Index = ShuffleMask[Index];
3940 return DAG.getUNDEF(VT.getVectorElementType());
3942 SDValue NewV = (Index < NumElems) ? N->getOperand(0) : N->getOperand(1);
3943 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG,
3947 // Actual nodes that may contain scalar elements
3948 if (Opcode == ISD::BITCAST) {
3949 V = V.getOperand(0);
3950 EVT SrcVT = V.getValueType();
3951 unsigned NumElems = VT.getVectorNumElements();
3953 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
3957 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
3958 return (Index == 0) ? V.getOperand(0)
3959 : DAG.getUNDEF(VT.getVectorElementType());
3961 if (V.getOpcode() == ISD::BUILD_VECTOR)
3962 return V.getOperand(Index);
3967 /// getNumOfConsecutiveZeros - Return the number of elements of a vector
3968 /// shuffle operation which come from a consecutively from a zero. The
3969 /// search can start in two diferent directions, from left or right.
3971 unsigned getNumOfConsecutiveZeros(SDNode *N, int NumElems,
3972 bool ZerosFromLeft, SelectionDAG &DAG) {
3975 while (i < NumElems) {
3976 unsigned Index = ZerosFromLeft ? i : NumElems-i-1;
3977 SDValue Elt = getShuffleScalarElt(N, Index, DAG, 0);
3978 if (!(Elt.getNode() &&
3979 (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt))))
3987 /// isShuffleMaskConsecutive - Check if the shuffle mask indicies from MaskI to
3988 /// MaskE correspond consecutively to elements from one of the vector operands,
3989 /// starting from its index OpIdx. Also tell OpNum which source vector operand.
3991 bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp, int MaskI, int MaskE,
3992 int OpIdx, int NumElems, unsigned &OpNum) {
3993 bool SeenV1 = false;
3994 bool SeenV2 = false;
3996 for (int i = MaskI; i <= MaskE; ++i, ++OpIdx) {
3997 int Idx = SVOp->getMaskElt(i);
3998 // Ignore undef indicies
4007 // Only accept consecutive elements from the same vector
4008 if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2))
4012 OpNum = SeenV1 ? 0 : 1;
4016 /// isVectorShiftRight - Returns true if the shuffle can be implemented as a
4017 /// logical left shift of a vector.
4018 static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4019 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4020 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
4021 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
4022 false /* check zeros from right */, DAG);
4028 // Considering the elements in the mask that are not consecutive zeros,
4029 // check if they consecutively come from only one of the source vectors.
4031 // V1 = {X, A, B, C} 0
4033 // vector_shuffle V1, V2 <1, 2, 3, X>
4035 if (!isShuffleMaskConsecutive(SVOp,
4036 0, // Mask Start Index
4037 NumElems-NumZeros-1, // Mask End Index
4038 NumZeros, // Where to start looking in the src vector
4039 NumElems, // Number of elements in vector
4040 OpSrc)) // Which source operand ?
4045 ShVal = SVOp->getOperand(OpSrc);
4049 /// isVectorShiftLeft - Returns true if the shuffle can be implemented as a
4050 /// logical left shift of a vector.
4051 static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4052 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4053 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
4054 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
4055 true /* check zeros from left */, DAG);
4061 // Considering the elements in the mask that are not consecutive zeros,
4062 // check if they consecutively come from only one of the source vectors.
4064 // 0 { A, B, X, X } = V2
4066 // vector_shuffle V1, V2 <X, X, 4, 5>
4068 if (!isShuffleMaskConsecutive(SVOp,
4069 NumZeros, // Mask Start Index
4070 NumElems-1, // Mask End Index
4071 0, // Where to start looking in the src vector
4072 NumElems, // Number of elements in vector
4073 OpSrc)) // Which source operand ?
4078 ShVal = SVOp->getOperand(OpSrc);
4082 /// isVectorShift - Returns true if the shuffle can be implemented as a
4083 /// logical left or right shift of a vector.
4084 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4085 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4086 if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) ||
4087 isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt))
4093 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
4095 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4096 unsigned NumNonZero, unsigned NumZero,
4098 const TargetLowering &TLI) {
4102 DebugLoc dl = Op.getDebugLoc();
4105 for (unsigned i = 0; i < 16; ++i) {
4106 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4107 if (ThisIsNonZero && First) {
4109 V = getZeroVector(MVT::v8i16, true, DAG, dl);
4111 V = DAG.getUNDEF(MVT::v8i16);
4116 SDValue ThisElt(0, 0), LastElt(0, 0);
4117 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4118 if (LastIsNonZero) {
4119 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4120 MVT::i16, Op.getOperand(i-1));
4122 if (ThisIsNonZero) {
4123 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4124 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4125 ThisElt, DAG.getConstant(8, MVT::i8));
4127 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
4131 if (ThisElt.getNode())
4132 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
4133 DAG.getIntPtrConstant(i/2));
4137 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
4140 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
4142 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
4143 unsigned NumNonZero, unsigned NumZero,
4145 const TargetLowering &TLI) {
4149 DebugLoc dl = Op.getDebugLoc();
4152 for (unsigned i = 0; i < 8; ++i) {
4153 bool isNonZero = (NonZeros & (1 << i)) != 0;
4157 V = getZeroVector(MVT::v8i16, true, DAG, dl);
4159 V = DAG.getUNDEF(MVT::v8i16);
4162 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4163 MVT::v8i16, V, Op.getOperand(i),
4164 DAG.getIntPtrConstant(i));
4171 /// getVShift - Return a vector logical shift node.
4173 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
4174 unsigned NumBits, SelectionDAG &DAG,
4175 const TargetLowering &TLI, DebugLoc dl) {
4176 EVT ShVT = MVT::v2i64;
4177 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
4178 SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
4179 return DAG.getNode(ISD::BITCAST, dl, VT,
4180 DAG.getNode(Opc, dl, ShVT, SrcOp,
4181 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
4185 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
4186 SelectionDAG &DAG) const {
4188 // Check if the scalar load can be widened into a vector load. And if
4189 // the address is "base + cst" see if the cst can be "absorbed" into
4190 // the shuffle mask.
4191 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
4192 SDValue Ptr = LD->getBasePtr();
4193 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
4195 EVT PVT = LD->getValueType(0);
4196 if (PVT != MVT::i32 && PVT != MVT::f32)
4201 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4202 FI = FINode->getIndex();
4204 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
4205 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4206 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4207 Offset = Ptr.getConstantOperandVal(1);
4208 Ptr = Ptr.getOperand(0);
4213 SDValue Chain = LD->getChain();
4214 // Make sure the stack object alignment is at least 16.
4215 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4216 if (DAG.InferPtrAlignment(Ptr) < 16) {
4217 if (MFI->isFixedObjectIndex(FI)) {
4218 // Can't change the alignment. FIXME: It's possible to compute
4219 // the exact stack offset and reference FI + adjust offset instead.
4220 // If someone *really* cares about this. That's the way to implement it.
4223 MFI->setObjectAlignment(FI, 16);
4227 // (Offset % 16) must be multiple of 4. Then address is then
4228 // Ptr + (Offset & ~15).
4231 if ((Offset % 16) & 3)
4233 int64_t StartOffset = Offset & ~15;
4235 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
4236 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
4238 int EltNo = (Offset - StartOffset) >> 2;
4239 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
4240 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
4241 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,
4242 LD->getPointerInfo().getWithOffset(StartOffset),
4244 // Canonicalize it to a v4i32 shuffle.
4245 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V1);
4246 return DAG.getNode(ISD::BITCAST, dl, VT,
4247 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
4248 DAG.getUNDEF(MVT::v4i32),&Mask[0]));
4254 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
4255 /// vector of type 'VT', see if the elements can be replaced by a single large
4256 /// load which has the same value as a build_vector whose operands are 'elts'.
4258 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
4260 /// FIXME: we'd also like to handle the case where the last elements are zero
4261 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
4262 /// There's even a handy isZeroNode for that purpose.
4263 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
4264 DebugLoc &DL, SelectionDAG &DAG) {
4265 EVT EltVT = VT.getVectorElementType();
4266 unsigned NumElems = Elts.size();
4268 LoadSDNode *LDBase = NULL;
4269 unsigned LastLoadedElt = -1U;
4271 // For each element in the initializer, see if we've found a load or an undef.
4272 // If we don't find an initial load element, or later load elements are
4273 // non-consecutive, bail out.
4274 for (unsigned i = 0; i < NumElems; ++i) {
4275 SDValue Elt = Elts[i];
4277 if (!Elt.getNode() ||
4278 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
4281 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
4283 LDBase = cast<LoadSDNode>(Elt.getNode());
4287 if (Elt.getOpcode() == ISD::UNDEF)
4290 LoadSDNode *LD = cast<LoadSDNode>(Elt);
4291 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
4296 // If we have found an entire vector of loads and undefs, then return a large
4297 // load of the entire vector width starting at the base pointer. If we found
4298 // consecutive loads for the low half, generate a vzext_load node.
4299 if (LastLoadedElt == NumElems - 1) {
4300 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
4301 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4302 LDBase->getPointerInfo(),
4303 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
4304 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4305 LDBase->getPointerInfo(),
4306 LDBase->isVolatile(), LDBase->isNonTemporal(),
4307 LDBase->getAlignment());
4308 } else if (NumElems == 4 && LastLoadedElt == 1) {
4309 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
4310 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
4311 SDValue ResNode = DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys,
4313 LDBase->getMemOperand());
4314 return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
4320 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
4321 DebugLoc dl = Op.getDebugLoc();
4323 EVT VT = Op.getValueType();
4324 EVT ExtVT = VT.getVectorElementType();
4326 unsigned NumElems = Op.getNumOperands();
4328 // For AVX-length vectors, build the individual 128-bit pieces and
4329 // use shuffles to put them in place.
4330 if (VT.getSizeInBits() > 256 &&
4331 Subtarget->hasAVX() &&
4333 !ISD::isBuildVectorAllZeros(Op.getNode())) {
4334 SmallVector<SDValue, 8> V;
4336 for (unsigned i = 0; i < NumElems; ++i) {
4337 V[i] = Op.getOperand(i);
4340 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
4342 // Build the lower subvector.
4343 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[0], NumElems/2);
4344 // Build the upper subvector.
4345 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[NumElems / 2],
4348 return ConcatVectors(Lower, Upper, DAG);
4351 // All zero's are handled with pxor in SSE2 and above, xorps in SSE1.
4352 // All one's are handled with pcmpeqd. In AVX, zero's are handled with
4353 // vpxor in 128-bit and xor{pd,ps} in 256-bit, but no 256 version of pcmpeqd
4354 // is present, so AllOnes is ignored.
4355 if (ISD::isBuildVectorAllZeros(Op.getNode()) ||
4356 (Op.getValueType().getSizeInBits() != 256 &&
4357 ISD::isBuildVectorAllOnes(Op.getNode()))) {
4358 // Canonicalize this to <4 x i32> (SSE) to
4359 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
4360 // eliminated on x86-32 hosts.
4361 if (Op.getValueType() == MVT::v4i32)
4364 if (ISD::isBuildVectorAllOnes(Op.getNode()))
4365 return getOnesVector(Op.getValueType(), DAG, dl);
4366 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
4369 unsigned EVTBits = ExtVT.getSizeInBits();
4371 unsigned NumZero = 0;
4372 unsigned NumNonZero = 0;
4373 unsigned NonZeros = 0;
4374 bool IsAllConstants = true;
4375 SmallSet<SDValue, 8> Values;
4376 for (unsigned i = 0; i < NumElems; ++i) {
4377 SDValue Elt = Op.getOperand(i);
4378 if (Elt.getOpcode() == ISD::UNDEF)
4381 if (Elt.getOpcode() != ISD::Constant &&
4382 Elt.getOpcode() != ISD::ConstantFP)
4383 IsAllConstants = false;
4384 if (X86::isZeroNode(Elt))
4387 NonZeros |= (1 << i);
4392 // All undef vector. Return an UNDEF. All zero vectors were handled above.
4393 if (NumNonZero == 0)
4394 return DAG.getUNDEF(VT);
4396 // Special case for single non-zero, non-undef, element.
4397 if (NumNonZero == 1) {
4398 unsigned Idx = CountTrailingZeros_32(NonZeros);
4399 SDValue Item = Op.getOperand(Idx);
4401 // If this is an insertion of an i64 value on x86-32, and if the top bits of
4402 // the value are obviously zero, truncate the value to i32 and do the
4403 // insertion that way. Only do this if the value is non-constant or if the
4404 // value is a constant being inserted into element 0. It is cheaper to do
4405 // a constant pool load than it is to do a movd + shuffle.
4406 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
4407 (!IsAllConstants || Idx == 0)) {
4408 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
4410 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
4411 EVT VecVT = MVT::v4i32;
4412 unsigned VecElts = 4;
4414 // Truncate the value (which may itself be a constant) to i32, and
4415 // convert it to a vector with movd (S2V+shuffle to zero extend).
4416 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
4417 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
4418 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4419 Subtarget->hasSSE2(), DAG);
4421 // Now we have our 32-bit value zero extended in the low element of
4422 // a vector. If Idx != 0, swizzle it into place.
4424 SmallVector<int, 4> Mask;
4425 Mask.push_back(Idx);
4426 for (unsigned i = 1; i != VecElts; ++i)
4428 Item = DAG.getVectorShuffle(VecVT, dl, Item,
4429 DAG.getUNDEF(Item.getValueType()),
4432 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Item);
4436 // If we have a constant or non-constant insertion into the low element of
4437 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
4438 // the rest of the elements. This will be matched as movd/movq/movss/movsd
4439 // depending on what the source datatype is.
4442 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4443 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
4444 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
4445 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4446 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
4447 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
4449 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
4450 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
4451 assert(VT.getSizeInBits() == 128 && "Expected an SSE value type!");
4452 EVT MiddleVT = MVT::v4i32;
4453 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
4454 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4455 Subtarget->hasSSE2(), DAG);
4456 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
4460 // Is it a vector logical left shift?
4461 if (NumElems == 2 && Idx == 1 &&
4462 X86::isZeroNode(Op.getOperand(0)) &&
4463 !X86::isZeroNode(Op.getOperand(1))) {
4464 unsigned NumBits = VT.getSizeInBits();
4465 return getVShift(true, VT,
4466 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4467 VT, Op.getOperand(1)),
4468 NumBits/2, DAG, *this, dl);
4471 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
4474 // Otherwise, if this is a vector with i32 or f32 elements, and the element
4475 // is a non-constant being inserted into an element other than the low one,
4476 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
4477 // movd/movss) to move this into the low element, then shuffle it into
4479 if (EVTBits == 32) {
4480 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4482 // Turn it into a shuffle of zero and zero-extended scalar to vector.
4483 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
4484 Subtarget->hasSSE2(), DAG);
4485 SmallVector<int, 8> MaskVec;
4486 for (unsigned i = 0; i < NumElems; i++)
4487 MaskVec.push_back(i == Idx ? 0 : 1);
4488 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
4492 // Splat is obviously ok. Let legalizer expand it to a shuffle.
4493 if (Values.size() == 1) {
4494 if (EVTBits == 32) {
4495 // Instead of a shuffle like this:
4496 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
4497 // Check if it's possible to issue this instead.
4498 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
4499 unsigned Idx = CountTrailingZeros_32(NonZeros);
4500 SDValue Item = Op.getOperand(Idx);
4501 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
4502 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
4507 // A vector full of immediates; various special cases are already
4508 // handled, so this is best done with a single constant-pool load.
4512 // Let legalizer expand 2-wide build_vectors.
4513 if (EVTBits == 64) {
4514 if (NumNonZero == 1) {
4515 // One half is zero or undef.
4516 unsigned Idx = CountTrailingZeros_32(NonZeros);
4517 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
4518 Op.getOperand(Idx));
4519 return getShuffleVectorZeroOrUndef(V2, Idx, true,
4520 Subtarget->hasSSE2(), DAG);
4525 // If element VT is < 32 bits, convert it to inserts into a zero vector.
4526 if (EVTBits == 8 && NumElems == 16) {
4527 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
4529 if (V.getNode()) return V;
4532 if (EVTBits == 16 && NumElems == 8) {
4533 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
4535 if (V.getNode()) return V;
4538 // If element VT is == 32 bits, turn it into a number of shuffles.
4539 SmallVector<SDValue, 8> V;
4541 if (NumElems == 4 && NumZero > 0) {
4542 for (unsigned i = 0; i < 4; ++i) {
4543 bool isZero = !(NonZeros & (1 << i));
4545 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4547 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4550 for (unsigned i = 0; i < 2; ++i) {
4551 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
4554 V[i] = V[i*2]; // Must be a zero vector.
4557 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
4560 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
4563 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
4568 SmallVector<int, 8> MaskVec;
4569 bool Reverse = (NonZeros & 0x3) == 2;
4570 for (unsigned i = 0; i < 2; ++i)
4571 MaskVec.push_back(Reverse ? 1-i : i);
4572 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
4573 for (unsigned i = 0; i < 2; ++i)
4574 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
4575 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
4578 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
4579 // Check for a build vector of consecutive loads.
4580 for (unsigned i = 0; i < NumElems; ++i)
4581 V[i] = Op.getOperand(i);
4583 // Check for elements which are consecutive loads.
4584 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
4588 // For SSE 4.1, use insertps to put the high elements into the low element.
4589 if (getSubtarget()->hasSSE41()) {
4591 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
4592 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
4594 Result = DAG.getUNDEF(VT);
4596 for (unsigned i = 1; i < NumElems; ++i) {
4597 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
4598 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
4599 Op.getOperand(i), DAG.getIntPtrConstant(i));
4604 // Otherwise, expand into a number of unpckl*, start by extending each of
4605 // our (non-undef) elements to the full vector width with the element in the
4606 // bottom slot of the vector (which generates no code for SSE).
4607 for (unsigned i = 0; i < NumElems; ++i) {
4608 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
4609 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4611 V[i] = DAG.getUNDEF(VT);
4614 // Next, we iteratively mix elements, e.g. for v4f32:
4615 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
4616 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
4617 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
4618 unsigned EltStride = NumElems >> 1;
4619 while (EltStride != 0) {
4620 for (unsigned i = 0; i < EltStride; ++i) {
4621 // If V[i+EltStride] is undef and this is the first round of mixing,
4622 // then it is safe to just drop this shuffle: V[i] is already in the
4623 // right place, the one element (since it's the first round) being
4624 // inserted as undef can be dropped. This isn't safe for successive
4625 // rounds because they will permute elements within both vectors.
4626 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
4627 EltStride == NumElems/2)
4630 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
4640 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
4641 // We support concatenate two MMX registers and place them in a MMX
4642 // register. This is better than doing a stack convert.
4643 DebugLoc dl = Op.getDebugLoc();
4644 EVT ResVT = Op.getValueType();
4645 assert(Op.getNumOperands() == 2);
4646 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
4647 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
4649 SDValue InVec = DAG.getNode(ISD::BITCAST,dl, MVT::v1i64, Op.getOperand(0));
4650 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4651 InVec = Op.getOperand(1);
4652 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4653 unsigned NumElts = ResVT.getVectorNumElements();
4654 VecOp = DAG.getNode(ISD::BITCAST, dl, ResVT, VecOp);
4655 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4656 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4658 InVec = DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, InVec);
4659 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4660 Mask[0] = 0; Mask[1] = 2;
4661 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4663 return DAG.getNode(ISD::BITCAST, dl, ResVT, VecOp);
4666 // v8i16 shuffles - Prefer shuffles in the following order:
4667 // 1. [all] pshuflw, pshufhw, optional move
4668 // 2. [ssse3] 1 x pshufb
4669 // 3. [ssse3] 2 x pshufb + 1 x por
4670 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4672 X86TargetLowering::LowerVECTOR_SHUFFLEv8i16(SDValue Op,
4673 SelectionDAG &DAG) const {
4674 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4675 SDValue V1 = SVOp->getOperand(0);
4676 SDValue V2 = SVOp->getOperand(1);
4677 DebugLoc dl = SVOp->getDebugLoc();
4678 SmallVector<int, 8> MaskVals;
4680 // Determine if more than 1 of the words in each of the low and high quadwords
4681 // of the result come from the same quadword of one of the two inputs. Undef
4682 // mask values count as coming from any quadword, for better codegen.
4683 SmallVector<unsigned, 4> LoQuad(4);
4684 SmallVector<unsigned, 4> HiQuad(4);
4685 BitVector InputQuads(4);
4686 for (unsigned i = 0; i < 8; ++i) {
4687 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4688 int EltIdx = SVOp->getMaskElt(i);
4689 MaskVals.push_back(EltIdx);
4698 InputQuads.set(EltIdx / 4);
4701 int BestLoQuad = -1;
4702 unsigned MaxQuad = 1;
4703 for (unsigned i = 0; i < 4; ++i) {
4704 if (LoQuad[i] > MaxQuad) {
4706 MaxQuad = LoQuad[i];
4710 int BestHiQuad = -1;
4712 for (unsigned i = 0; i < 4; ++i) {
4713 if (HiQuad[i] > MaxQuad) {
4715 MaxQuad = HiQuad[i];
4719 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4720 // of the two input vectors, shuffle them into one input vector so only a
4721 // single pshufb instruction is necessary. If There are more than 2 input
4722 // quads, disable the next transformation since it does not help SSSE3.
4723 bool V1Used = InputQuads[0] || InputQuads[1];
4724 bool V2Used = InputQuads[2] || InputQuads[3];
4725 if (Subtarget->hasSSSE3()) {
4726 if (InputQuads.count() == 2 && V1Used && V2Used) {
4727 BestLoQuad = InputQuads.find_first();
4728 BestHiQuad = InputQuads.find_next(BestLoQuad);
4730 if (InputQuads.count() > 2) {
4736 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4737 // the shuffle mask. If a quad is scored as -1, that means that it contains
4738 // words from all 4 input quadwords.
4740 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4741 SmallVector<int, 8> MaskV;
4742 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4743 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4744 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4745 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1),
4746 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V2), &MaskV[0]);
4747 NewV = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, NewV);
4749 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4750 // source words for the shuffle, to aid later transformations.
4751 bool AllWordsInNewV = true;
4752 bool InOrder[2] = { true, true };
4753 for (unsigned i = 0; i != 8; ++i) {
4754 int idx = MaskVals[i];
4756 InOrder[i/4] = false;
4757 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4759 AllWordsInNewV = false;
4763 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4764 if (AllWordsInNewV) {
4765 for (int i = 0; i != 8; ++i) {
4766 int idx = MaskVals[i];
4769 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4770 if ((idx != i) && idx < 4)
4772 if ((idx != i) && idx > 3)
4781 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4782 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4783 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4784 unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
4785 unsigned TargetMask = 0;
4786 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4787 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4788 TargetMask = pshufhw ? X86::getShufflePSHUFHWImmediate(NewV.getNode()):
4789 X86::getShufflePSHUFLWImmediate(NewV.getNode());
4790 V1 = NewV.getOperand(0);
4791 return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
4795 // If we have SSSE3, and all words of the result are from 1 input vector,
4796 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4797 // is present, fall back to case 4.
4798 if (Subtarget->hasSSSE3()) {
4799 SmallVector<SDValue,16> pshufbMask;
4801 // If we have elements from both input vectors, set the high bit of the
4802 // shuffle mask element to zero out elements that come from V2 in the V1
4803 // mask, and elements that come from V1 in the V2 mask, so that the two
4804 // results can be OR'd together.
4805 bool TwoInputs = V1Used && V2Used;
4806 for (unsigned i = 0; i != 8; ++i) {
4807 int EltIdx = MaskVals[i] * 2;
4808 if (TwoInputs && (EltIdx >= 16)) {
4809 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4810 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4813 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4814 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4816 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V1);
4817 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4818 DAG.getNode(ISD::BUILD_VECTOR, dl,
4819 MVT::v16i8, &pshufbMask[0], 16));
4821 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
4823 // Calculate the shuffle mask for the second input, shuffle it, and
4824 // OR it with the first shuffled input.
4826 for (unsigned i = 0; i != 8; ++i) {
4827 int EltIdx = MaskVals[i] * 2;
4829 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4830 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4833 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4834 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4836 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V2);
4837 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4838 DAG.getNode(ISD::BUILD_VECTOR, dl,
4839 MVT::v16i8, &pshufbMask[0], 16));
4840 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4841 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
4844 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4845 // and update MaskVals with new element order.
4846 BitVector InOrder(8);
4847 if (BestLoQuad >= 0) {
4848 SmallVector<int, 8> MaskV;
4849 for (int i = 0; i != 4; ++i) {
4850 int idx = MaskVals[i];
4852 MaskV.push_back(-1);
4854 } else if ((idx / 4) == BestLoQuad) {
4855 MaskV.push_back(idx & 3);
4858 MaskV.push_back(-1);
4861 for (unsigned i = 4; i != 8; ++i)
4863 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4866 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
4867 NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
4869 X86::getShufflePSHUFLWImmediate(NewV.getNode()),
4873 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4874 // and update MaskVals with the new element order.
4875 if (BestHiQuad >= 0) {
4876 SmallVector<int, 8> MaskV;
4877 for (unsigned i = 0; i != 4; ++i)
4879 for (unsigned i = 4; i != 8; ++i) {
4880 int idx = MaskVals[i];
4882 MaskV.push_back(-1);
4884 } else if ((idx / 4) == BestHiQuad) {
4885 MaskV.push_back((idx & 3) + 4);
4888 MaskV.push_back(-1);
4891 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4894 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
4895 NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
4897 X86::getShufflePSHUFHWImmediate(NewV.getNode()),
4901 // In case BestHi & BestLo were both -1, which means each quadword has a word
4902 // from each of the four input quadwords, calculate the InOrder bitvector now
4903 // before falling through to the insert/extract cleanup.
4904 if (BestLoQuad == -1 && BestHiQuad == -1) {
4906 for (int i = 0; i != 8; ++i)
4907 if (MaskVals[i] < 0 || MaskVals[i] == i)
4911 // The other elements are put in the right place using pextrw and pinsrw.
4912 for (unsigned i = 0; i != 8; ++i) {
4915 int EltIdx = MaskVals[i];
4918 SDValue ExtOp = (EltIdx < 8)
4919 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4920 DAG.getIntPtrConstant(EltIdx))
4921 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4922 DAG.getIntPtrConstant(EltIdx - 8));
4923 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4924 DAG.getIntPtrConstant(i));
4929 // v16i8 shuffles - Prefer shuffles in the following order:
4930 // 1. [ssse3] 1 x pshufb
4931 // 2. [ssse3] 2 x pshufb + 1 x por
4932 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4934 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4936 const X86TargetLowering &TLI) {
4937 SDValue V1 = SVOp->getOperand(0);
4938 SDValue V2 = SVOp->getOperand(1);
4939 DebugLoc dl = SVOp->getDebugLoc();
4940 SmallVector<int, 16> MaskVals;
4941 SVOp->getMask(MaskVals);
4943 // If we have SSSE3, case 1 is generated when all result bytes come from
4944 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4945 // present, fall back to case 3.
4946 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4949 for (unsigned i = 0; i < 16; ++i) {
4950 int EltIdx = MaskVals[i];
4959 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4960 if (TLI.getSubtarget()->hasSSSE3()) {
4961 SmallVector<SDValue,16> pshufbMask;
4963 // If all result elements are from one input vector, then only translate
4964 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4966 // Otherwise, we have elements from both input vectors, and must zero out
4967 // elements that come from V2 in the first mask, and V1 in the second mask
4968 // so that we can OR them together.
4969 bool TwoInputs = !(V1Only || V2Only);
4970 for (unsigned i = 0; i != 16; ++i) {
4971 int EltIdx = MaskVals[i];
4972 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4973 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4976 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4978 // If all the elements are from V2, assign it to V1 and return after
4979 // building the first pshufb.
4982 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4983 DAG.getNode(ISD::BUILD_VECTOR, dl,
4984 MVT::v16i8, &pshufbMask[0], 16));
4988 // Calculate the shuffle mask for the second input, shuffle it, and
4989 // OR it with the first shuffled input.
4991 for (unsigned i = 0; i != 16; ++i) {
4992 int EltIdx = MaskVals[i];
4994 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4997 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4999 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
5000 DAG.getNode(ISD::BUILD_VECTOR, dl,
5001 MVT::v16i8, &pshufbMask[0], 16));
5002 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
5005 // No SSSE3 - Calculate in place words and then fix all out of place words
5006 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
5007 // the 16 different words that comprise the two doublequadword input vectors.
5008 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5009 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V2);
5010 SDValue NewV = V2Only ? V2 : V1;
5011 for (int i = 0; i != 8; ++i) {
5012 int Elt0 = MaskVals[i*2];
5013 int Elt1 = MaskVals[i*2+1];
5015 // This word of the result is all undef, skip it.
5016 if (Elt0 < 0 && Elt1 < 0)
5019 // This word of the result is already in the correct place, skip it.
5020 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
5022 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
5025 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
5026 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
5029 // If Elt0 and Elt1 are defined, are consecutive, and can be load
5030 // using a single extract together, load it and store it.
5031 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
5032 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
5033 DAG.getIntPtrConstant(Elt1 / 2));
5034 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
5035 DAG.getIntPtrConstant(i));
5039 // If Elt1 is defined, extract it from the appropriate source. If the
5040 // source byte is not also odd, shift the extracted word left 8 bits
5041 // otherwise clear the bottom 8 bits if we need to do an or.
5043 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
5044 DAG.getIntPtrConstant(Elt1 / 2));
5045 if ((Elt1 & 1) == 0)
5046 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
5047 DAG.getConstant(8, TLI.getShiftAmountTy()));
5049 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
5050 DAG.getConstant(0xFF00, MVT::i16));
5052 // If Elt0 is defined, extract it from the appropriate source. If the
5053 // source byte is not also even, shift the extracted word right 8 bits. If
5054 // Elt1 was also defined, OR the extracted values together before
5055 // inserting them in the result.
5057 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
5058 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
5059 if ((Elt0 & 1) != 0)
5060 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
5061 DAG.getConstant(8, TLI.getShiftAmountTy()));
5063 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
5064 DAG.getConstant(0x00FF, MVT::i16));
5065 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
5068 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
5069 DAG.getIntPtrConstant(i));
5071 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, NewV);
5074 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
5075 /// ones, or rewriting v4i32 / v4f32 as 2 wide ones if possible. This can be
5076 /// done when every pair / quad of shuffle mask elements point to elements in
5077 /// the right sequence. e.g.
5078 /// vector_shuffle X, Y, <2, 3, | 10, 11, | 0, 1, | 14, 15>
5080 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
5081 SelectionDAG &DAG, DebugLoc dl) {
5082 EVT VT = SVOp->getValueType(0);
5083 SDValue V1 = SVOp->getOperand(0);
5084 SDValue V2 = SVOp->getOperand(1);
5085 unsigned NumElems = VT.getVectorNumElements();
5086 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
5088 switch (VT.getSimpleVT().SimpleTy) {
5089 default: assert(false && "Unexpected!");
5090 case MVT::v4f32: NewVT = MVT::v2f64; break;
5091 case MVT::v4i32: NewVT = MVT::v2i64; break;
5092 case MVT::v8i16: NewVT = MVT::v4i32; break;
5093 case MVT::v16i8: NewVT = MVT::v4i32; break;
5096 int Scale = NumElems / NewWidth;
5097 SmallVector<int, 8> MaskVec;
5098 for (unsigned i = 0; i < NumElems; i += Scale) {
5100 for (int j = 0; j < Scale; ++j) {
5101 int EltIdx = SVOp->getMaskElt(i+j);
5105 StartIdx = EltIdx - (EltIdx % Scale);
5106 if (EltIdx != StartIdx + j)
5110 MaskVec.push_back(-1);
5112 MaskVec.push_back(StartIdx / Scale);
5115 V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1);
5116 V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2);
5117 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
5120 /// getVZextMovL - Return a zero-extending vector move low node.
5122 static SDValue getVZextMovL(EVT VT, EVT OpVT,
5123 SDValue SrcOp, SelectionDAG &DAG,
5124 const X86Subtarget *Subtarget, DebugLoc dl) {
5125 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
5126 LoadSDNode *LD = NULL;
5127 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
5128 LD = dyn_cast<LoadSDNode>(SrcOp);
5130 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
5132 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
5133 if ((ExtVT != MVT::i64 || Subtarget->is64Bit()) &&
5134 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
5135 SrcOp.getOperand(0).getOpcode() == ISD::BITCAST &&
5136 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
5138 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
5139 return DAG.getNode(ISD::BITCAST, dl, VT,
5140 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
5141 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5149 return DAG.getNode(ISD::BITCAST, dl, VT,
5150 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
5151 DAG.getNode(ISD::BITCAST, dl,
5155 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
5158 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
5159 SDValue V1 = SVOp->getOperand(0);
5160 SDValue V2 = SVOp->getOperand(1);
5161 DebugLoc dl = SVOp->getDebugLoc();
5162 EVT VT = SVOp->getValueType(0);
5164 SmallVector<std::pair<int, int>, 8> Locs;
5166 SmallVector<int, 8> Mask1(4U, -1);
5167 SmallVector<int, 8> PermMask;
5168 SVOp->getMask(PermMask);
5172 for (unsigned i = 0; i != 4; ++i) {
5173 int Idx = PermMask[i];
5175 Locs[i] = std::make_pair(-1, -1);
5177 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
5179 Locs[i] = std::make_pair(0, NumLo);
5183 Locs[i] = std::make_pair(1, NumHi);
5185 Mask1[2+NumHi] = Idx;
5191 if (NumLo <= 2 && NumHi <= 2) {
5192 // If no more than two elements come from either vector. This can be
5193 // implemented with two shuffles. First shuffle gather the elements.
5194 // The second shuffle, which takes the first shuffle as both of its
5195 // vector operands, put the elements into the right order.
5196 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5198 SmallVector<int, 8> Mask2(4U, -1);
5200 for (unsigned i = 0; i != 4; ++i) {
5201 if (Locs[i].first == -1)
5204 unsigned Idx = (i < 2) ? 0 : 4;
5205 Idx += Locs[i].first * 2 + Locs[i].second;
5210 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
5211 } else if (NumLo == 3 || NumHi == 3) {
5212 // Otherwise, we must have three elements from one vector, call it X, and
5213 // one element from the other, call it Y. First, use a shufps to build an
5214 // intermediate vector with the one element from Y and the element from X
5215 // that will be in the same half in the final destination (the indexes don't
5216 // matter). Then, use a shufps to build the final vector, taking the half
5217 // containing the element from Y from the intermediate, and the other half
5220 // Normalize it so the 3 elements come from V1.
5221 CommuteVectorShuffleMask(PermMask, VT);
5225 // Find the element from V2.
5227 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
5228 int Val = PermMask[HiIndex];
5235 Mask1[0] = PermMask[HiIndex];
5237 Mask1[2] = PermMask[HiIndex^1];
5239 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5242 Mask1[0] = PermMask[0];
5243 Mask1[1] = PermMask[1];
5244 Mask1[2] = HiIndex & 1 ? 6 : 4;
5245 Mask1[3] = HiIndex & 1 ? 4 : 6;
5246 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5248 Mask1[0] = HiIndex & 1 ? 2 : 0;
5249 Mask1[1] = HiIndex & 1 ? 0 : 2;
5250 Mask1[2] = PermMask[2];
5251 Mask1[3] = PermMask[3];
5256 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
5260 // Break it into (shuffle shuffle_hi, shuffle_lo).
5262 SmallVector<int,8> LoMask(4U, -1);
5263 SmallVector<int,8> HiMask(4U, -1);
5265 SmallVector<int,8> *MaskPtr = &LoMask;
5266 unsigned MaskIdx = 0;
5269 for (unsigned i = 0; i != 4; ++i) {
5276 int Idx = PermMask[i];
5278 Locs[i] = std::make_pair(-1, -1);
5279 } else if (Idx < 4) {
5280 Locs[i] = std::make_pair(MaskIdx, LoIdx);
5281 (*MaskPtr)[LoIdx] = Idx;
5284 Locs[i] = std::make_pair(MaskIdx, HiIdx);
5285 (*MaskPtr)[HiIdx] = Idx;
5290 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
5291 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
5292 SmallVector<int, 8> MaskOps;
5293 for (unsigned i = 0; i != 4; ++i) {
5294 if (Locs[i].first == -1) {
5295 MaskOps.push_back(-1);
5297 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
5298 MaskOps.push_back(Idx);
5301 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
5304 static bool MayFoldVectorLoad(SDValue V) {
5305 if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
5306 V = V.getOperand(0);
5307 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5308 V = V.getOperand(0);
5314 // FIXME: the version above should always be used. Since there's
5315 // a bug where several vector shuffles can't be folded because the
5316 // DAG is not updated during lowering and a node claims to have two
5317 // uses while it only has one, use this version, and let isel match
5318 // another instruction if the load really happens to have more than
5319 // one use. Remove this version after this bug get fixed.
5320 // rdar://8434668, PR8156
5321 static bool RelaxedMayFoldVectorLoad(SDValue V) {
5322 if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
5323 V = V.getOperand(0);
5324 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5325 V = V.getOperand(0);
5326 if (ISD::isNormalLoad(V.getNode()))
5331 /// CanFoldShuffleIntoVExtract - Check if the current shuffle is used by
5332 /// a vector extract, and if both can be later optimized into a single load.
5333 /// This is done in visitEXTRACT_VECTOR_ELT and the conditions are checked
5334 /// here because otherwise a target specific shuffle node is going to be
5335 /// emitted for this shuffle, and the optimization not done.
5336 /// FIXME: This is probably not the best approach, but fix the problem
5337 /// until the right path is decided.
5339 bool CanXFormVExtractWithShuffleIntoLoad(SDValue V, SelectionDAG &DAG,
5340 const TargetLowering &TLI) {
5341 EVT VT = V.getValueType();
5342 ShuffleVectorSDNode *SVOp = dyn_cast<ShuffleVectorSDNode>(V);
5344 // Be sure that the vector shuffle is present in a pattern like this:
5345 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), c) -> (f32 load $addr)
5349 SDNode *N = *V.getNode()->use_begin();
5350 if (N->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
5353 SDValue EltNo = N->getOperand(1);
5354 if (!isa<ConstantSDNode>(EltNo))
5357 // If the bit convert changed the number of elements, it is unsafe
5358 // to examine the mask.
5359 bool HasShuffleIntoBitcast = false;
5360 if (V.getOpcode() == ISD::BITCAST) {
5361 EVT SrcVT = V.getOperand(0).getValueType();
5362 if (SrcVT.getVectorNumElements() != VT.getVectorNumElements())
5364 V = V.getOperand(0);
5365 HasShuffleIntoBitcast = true;
5368 // Select the input vector, guarding against out of range extract vector.
5369 unsigned NumElems = VT.getVectorNumElements();
5370 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
5371 int Idx = (Elt > NumElems) ? -1 : SVOp->getMaskElt(Elt);
5372 V = (Idx < (int)NumElems) ? V.getOperand(0) : V.getOperand(1);
5374 // Skip one more bit_convert if necessary
5375 if (V.getOpcode() == ISD::BITCAST)
5376 V = V.getOperand(0);
5378 if (ISD::isNormalLoad(V.getNode())) {
5379 // Is the original load suitable?
5380 LoadSDNode *LN0 = cast<LoadSDNode>(V);
5382 // FIXME: avoid the multi-use bug that is preventing lots of
5383 // of foldings to be detected, this is still wrong of course, but
5384 // give the temporary desired behavior, and if it happens that
5385 // the load has real more uses, during isel it will not fold, and
5386 // will generate poor code.
5387 if (!LN0 || LN0->isVolatile()) // || !LN0->hasOneUse()
5390 if (!HasShuffleIntoBitcast)
5393 // If there's a bitcast before the shuffle, check if the load type and
5394 // alignment is valid.
5395 unsigned Align = LN0->getAlignment();
5397 TLI.getTargetData()->getABITypeAlignment(
5398 VT.getTypeForEVT(*DAG.getContext()));
5400 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VT))
5408 SDValue getMOVDDup(SDValue &Op, DebugLoc &dl, SDValue V1, SelectionDAG &DAG) {
5409 EVT VT = Op.getValueType();
5411 // Canonizalize to v2f64.
5412 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, V1);
5413 return DAG.getNode(ISD::BITCAST, dl, VT,
5414 getTargetShuffleNode(X86ISD::MOVDDUP, dl, MVT::v2f64,
5419 SDValue getMOVLowToHigh(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG,
5421 SDValue V1 = Op.getOperand(0);
5422 SDValue V2 = Op.getOperand(1);
5423 EVT VT = Op.getValueType();
5425 assert(VT != MVT::v2i64 && "unsupported shuffle type");
5427 if (HasSSE2 && VT == MVT::v2f64)
5428 return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG);
5431 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V2, DAG);
5435 SDValue getMOVHighToLow(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG) {
5436 SDValue V1 = Op.getOperand(0);
5437 SDValue V2 = Op.getOperand(1);
5438 EVT VT = Op.getValueType();
5440 assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&
5441 "unsupported shuffle type");
5443 if (V2.getOpcode() == ISD::UNDEF)
5447 return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG);
5451 SDValue getMOVLP(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
5452 SDValue V1 = Op.getOperand(0);
5453 SDValue V2 = Op.getOperand(1);
5454 EVT VT = Op.getValueType();
5455 unsigned NumElems = VT.getVectorNumElements();
5457 // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second
5458 // operand of these instructions is only memory, so check if there's a
5459 // potencial load folding here, otherwise use SHUFPS or MOVSD to match the
5461 bool CanFoldLoad = false;
5463 // Trivial case, when V2 comes from a load.
5464 if (MayFoldVectorLoad(V2))
5467 // When V1 is a load, it can be folded later into a store in isel, example:
5468 // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1)
5470 // (MOVLPSmr addr:$src1, VR128:$src2)
5471 // So, recognize this potential and also use MOVLPS or MOVLPD
5472 if (MayFoldVectorLoad(V1) && MayFoldIntoStore(Op))
5476 if (HasSSE2 && NumElems == 2)
5477 return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
5480 return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
5483 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5484 // movl and movlp will both match v2i64, but v2i64 is never matched by
5485 // movl earlier because we make it strict to avoid messing with the movlp load
5486 // folding logic (see the code above getMOVLP call). Match it here then,
5487 // this is horrible, but will stay like this until we move all shuffle
5488 // matching to x86 specific nodes. Note that for the 1st condition all
5489 // types are matched with movsd.
5490 if ((HasSSE2 && NumElems == 2) || !X86::isMOVLMask(SVOp))
5491 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5493 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5496 assert(VT != MVT::v4i32 && "unsupported shuffle type");
5498 // Invert the operand order and use SHUFPS to match it.
5499 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V2, V1,
5500 X86::getShuffleSHUFImmediate(SVOp), DAG);
5503 static inline unsigned getUNPCKLOpcode(EVT VT) {
5504 switch(VT.getSimpleVT().SimpleTy) {
5505 case MVT::v4i32: return X86ISD::PUNPCKLDQ;
5506 case MVT::v2i64: return X86ISD::PUNPCKLQDQ;
5507 case MVT::v4f32: return X86ISD::UNPCKLPS;
5508 case MVT::v2f64: return X86ISD::UNPCKLPD;
5509 case MVT::v16i8: return X86ISD::PUNPCKLBW;
5510 case MVT::v8i16: return X86ISD::PUNPCKLWD;
5512 llvm_unreachable("Unknown type for unpckl");
5517 static inline unsigned getUNPCKHOpcode(EVT VT) {
5518 switch(VT.getSimpleVT().SimpleTy) {
5519 case MVT::v4i32: return X86ISD::PUNPCKHDQ;
5520 case MVT::v2i64: return X86ISD::PUNPCKHQDQ;
5521 case MVT::v4f32: return X86ISD::UNPCKHPS;
5522 case MVT::v2f64: return X86ISD::UNPCKHPD;
5523 case MVT::v16i8: return X86ISD::PUNPCKHBW;
5524 case MVT::v8i16: return X86ISD::PUNPCKHWD;
5526 llvm_unreachable("Unknown type for unpckh");
5532 SDValue NormalizeVectorShuffle(SDValue Op, SelectionDAG &DAG,
5533 const TargetLowering &TLI,
5534 const X86Subtarget *Subtarget) {
5535 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5536 EVT VT = Op.getValueType();
5537 DebugLoc dl = Op.getDebugLoc();
5538 SDValue V1 = Op.getOperand(0);
5539 SDValue V2 = Op.getOperand(1);
5541 if (isZeroShuffle(SVOp))
5542 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
5544 // Handle splat operations
5545 if (SVOp->isSplat()) {
5546 // Special case, this is the only place now where it's
5547 // allowed to return a vector_shuffle operation without
5548 // using a target specific node, because *hopefully* it
5549 // will be optimized away by the dag combiner.
5550 if (VT.getVectorNumElements() <= 4 &&
5551 CanXFormVExtractWithShuffleIntoLoad(Op, DAG, TLI))
5554 // Handle splats by matching through known masks
5555 if (VT.getVectorNumElements() <= 4)
5558 // Canonicalize all of the remaining to v4f32.
5559 return PromoteSplat(SVOp, DAG);
5562 // If the shuffle can be profitably rewritten as a narrower shuffle, then
5564 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
5565 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
5566 if (NewOp.getNode())
5567 return DAG.getNode(ISD::BITCAST, dl, VT, NewOp);
5568 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
5569 // FIXME: Figure out a cleaner way to do this.
5570 // Try to make use of movq to zero out the top part.
5571 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
5572 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
5573 if (NewOp.getNode()) {
5574 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
5575 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
5576 DAG, Subtarget, dl);
5578 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
5579 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
5580 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
5581 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
5582 DAG, Subtarget, dl);
5589 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
5590 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5591 SDValue V1 = Op.getOperand(0);
5592 SDValue V2 = Op.getOperand(1);
5593 EVT VT = Op.getValueType();
5594 DebugLoc dl = Op.getDebugLoc();
5595 unsigned NumElems = VT.getVectorNumElements();
5596 bool isMMX = VT.getSizeInBits() == 64;
5597 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
5598 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
5599 bool V1IsSplat = false;
5600 bool V2IsSplat = false;
5601 bool HasSSE2 = Subtarget->hasSSE2() || Subtarget->hasAVX();
5602 bool HasSSE3 = Subtarget->hasSSE3() || Subtarget->hasAVX();
5603 bool HasSSSE3 = Subtarget->hasSSSE3() || Subtarget->hasAVX();
5604 MachineFunction &MF = DAG.getMachineFunction();
5605 bool OptForSize = MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize);
5607 // Shuffle operations on MMX not supported.
5611 // Vector shuffle lowering takes 3 steps:
5613 // 1) Normalize the input vectors. Here splats, zeroed vectors, profitable
5614 // narrowing and commutation of operands should be handled.
5615 // 2) Matching of shuffles with known shuffle masks to x86 target specific
5617 // 3) Rewriting of unmatched masks into new generic shuffle operations,
5618 // so the shuffle can be broken into other shuffles and the legalizer can
5619 // try the lowering again.
5621 // The general ideia is that no vector_shuffle operation should be left to
5622 // be matched during isel, all of them must be converted to a target specific
5625 // Normalize the input vectors. Here splats, zeroed vectors, profitable
5626 // narrowing and commutation of operands should be handled. The actual code
5627 // doesn't include all of those, work in progress...
5628 SDValue NewOp = NormalizeVectorShuffle(Op, DAG, *this, Subtarget);
5629 if (NewOp.getNode())
5632 // NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
5633 // unpckh_undef). Only use pshufd if speed is more important than size.
5634 if (OptForSize && X86::isUNPCKL_v_undef_Mask(SVOp))
5635 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5636 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V1, DAG);
5637 if (OptForSize && X86::isUNPCKH_v_undef_Mask(SVOp))
5638 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5639 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5641 if (X86::isMOVDDUPMask(SVOp) && HasSSE3 && V2IsUndef &&
5642 RelaxedMayFoldVectorLoad(V1))
5643 return getMOVDDup(Op, dl, V1, DAG);
5645 if (X86::isMOVHLPS_v_undef_Mask(SVOp))
5646 return getMOVHighToLow(Op, dl, DAG);
5648 // Use to match splats
5649 if (HasSSE2 && X86::isUNPCKHMask(SVOp) && V2IsUndef &&
5650 (VT == MVT::v2f64 || VT == MVT::v2i64))
5651 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5653 if (X86::isPSHUFDMask(SVOp)) {
5654 // The actual implementation will match the mask in the if above and then
5655 // during isel it can match several different instructions, not only pshufd
5656 // as its name says, sad but true, emulate the behavior for now...
5657 if (X86::isMOVDDUPMask(SVOp) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
5658 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
5660 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
5662 if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
5663 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
5665 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
5666 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V1,
5669 if (VT == MVT::v4f32)
5670 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V1,
5674 // Check if this can be converted into a logical shift.
5675 bool isLeft = false;
5678 bool isShift = getSubtarget()->hasSSE2() &&
5679 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
5680 if (isShift && ShVal.hasOneUse()) {
5681 // If the shifted value has multiple uses, it may be cheaper to use
5682 // v_set0 + movlhps or movhlps, etc.
5683 EVT EltVT = VT.getVectorElementType();
5684 ShAmt *= EltVT.getSizeInBits();
5685 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
5688 if (X86::isMOVLMask(SVOp)) {
5691 if (ISD::isBuildVectorAllZeros(V1.getNode()))
5692 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
5693 if (!X86::isMOVLPMask(SVOp)) {
5694 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
5695 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5697 if (VT == MVT::v4i32 || VT == MVT::v4f32)
5698 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5702 // FIXME: fold these into legal mask.
5703 if (X86::isMOVLHPSMask(SVOp) && !X86::isUNPCKLMask(SVOp))
5704 return getMOVLowToHigh(Op, dl, DAG, HasSSE2);
5706 if (X86::isMOVHLPSMask(SVOp))
5707 return getMOVHighToLow(Op, dl, DAG);
5709 if (X86::isMOVSHDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
5710 return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
5712 if (X86::isMOVSLDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
5713 return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
5715 if (X86::isMOVLPMask(SVOp))
5716 return getMOVLP(Op, dl, DAG, HasSSE2);
5718 if (ShouldXformToMOVHLPS(SVOp) ||
5719 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
5720 return CommuteVectorShuffle(SVOp, DAG);
5723 // No better options. Use a vshl / vsrl.
5724 EVT EltVT = VT.getVectorElementType();
5725 ShAmt *= EltVT.getSizeInBits();
5726 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
5729 bool Commuted = false;
5730 // FIXME: This should also accept a bitcast of a splat? Be careful, not
5731 // 1,1,1,1 -> v8i16 though.
5732 V1IsSplat = isSplatVector(V1.getNode());
5733 V2IsSplat = isSplatVector(V2.getNode());
5735 // Canonicalize the splat or undef, if present, to be on the RHS.
5736 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
5737 Op = CommuteVectorShuffle(SVOp, DAG);
5738 SVOp = cast<ShuffleVectorSDNode>(Op);
5739 V1 = SVOp->getOperand(0);
5740 V2 = SVOp->getOperand(1);
5741 std::swap(V1IsSplat, V2IsSplat);
5742 std::swap(V1IsUndef, V2IsUndef);
5746 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
5747 // Shuffling low element of v1 into undef, just return v1.
5750 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
5751 // the instruction selector will not match, so get a canonical MOVL with
5752 // swapped operands to undo the commute.
5753 return getMOVL(DAG, dl, VT, V2, V1);
5756 if (X86::isUNPCKLMask(SVOp))
5757 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V2, DAG);
5759 if (X86::isUNPCKHMask(SVOp))
5760 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V2, DAG);
5763 // Normalize mask so all entries that point to V2 points to its first
5764 // element then try to match unpck{h|l} again. If match, return a
5765 // new vector_shuffle with the corrected mask.
5766 SDValue NewMask = NormalizeMask(SVOp, DAG);
5767 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
5768 if (NSVOp != SVOp) {
5769 if (X86::isUNPCKLMask(NSVOp, true)) {
5771 } else if (X86::isUNPCKHMask(NSVOp, true)) {
5778 // Commute is back and try unpck* again.
5779 // FIXME: this seems wrong.
5780 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
5781 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
5783 if (X86::isUNPCKLMask(NewSVOp))
5784 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V2, V1, DAG);
5786 if (X86::isUNPCKHMask(NewSVOp))
5787 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V2, V1, DAG);
5790 // Normalize the node to match x86 shuffle ops if needed
5791 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
5792 return CommuteVectorShuffle(SVOp, DAG);
5794 // The checks below are all present in isShuffleMaskLegal, but they are
5795 // inlined here right now to enable us to directly emit target specific
5796 // nodes, and remove one by one until they don't return Op anymore.
5797 SmallVector<int, 16> M;
5800 if (isPALIGNRMask(M, VT, HasSSSE3))
5801 return getTargetShuffleNode(X86ISD::PALIGN, dl, VT, V1, V2,
5802 X86::getShufflePALIGNRImmediate(SVOp),
5805 if (ShuffleVectorSDNode::isSplatMask(&M[0], VT) &&
5806 SVOp->getSplatIndex() == 0 && V2IsUndef) {
5807 if (VT == MVT::v2f64)
5808 return getTargetShuffleNode(X86ISD::UNPCKLPD, dl, VT, V1, V1, DAG);
5809 if (VT == MVT::v2i64)
5810 return getTargetShuffleNode(X86ISD::PUNPCKLQDQ, dl, VT, V1, V1, DAG);
5813 if (isPSHUFHWMask(M, VT))
5814 return getTargetShuffleNode(X86ISD::PSHUFHW, dl, VT, V1,
5815 X86::getShufflePSHUFHWImmediate(SVOp),
5818 if (isPSHUFLWMask(M, VT))
5819 return getTargetShuffleNode(X86ISD::PSHUFLW, dl, VT, V1,
5820 X86::getShufflePSHUFLWImmediate(SVOp),
5823 if (isSHUFPMask(M, VT)) {
5824 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
5825 if (VT == MVT::v4f32 || VT == MVT::v4i32)
5826 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V2,
5828 if (VT == MVT::v2f64 || VT == MVT::v2i64)
5829 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V2,
5833 if (X86::isUNPCKL_v_undef_Mask(SVOp))
5834 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5835 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V1, DAG);
5836 if (X86::isUNPCKH_v_undef_Mask(SVOp))
5837 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5838 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5840 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
5841 if (VT == MVT::v8i16) {
5842 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, DAG);
5843 if (NewOp.getNode())
5847 if (VT == MVT::v16i8) {
5848 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
5849 if (NewOp.getNode())
5853 // Handle all 4 wide cases with a number of shuffles.
5855 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
5861 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
5862 SelectionDAG &DAG) const {
5863 EVT VT = Op.getValueType();
5864 DebugLoc dl = Op.getDebugLoc();
5865 if (VT.getSizeInBits() == 8) {
5866 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
5867 Op.getOperand(0), Op.getOperand(1));
5868 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
5869 DAG.getValueType(VT));
5870 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5871 } else if (VT.getSizeInBits() == 16) {
5872 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5873 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
5875 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
5876 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5877 DAG.getNode(ISD::BITCAST, dl,
5881 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
5882 Op.getOperand(0), Op.getOperand(1));
5883 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
5884 DAG.getValueType(VT));
5885 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5886 } else if (VT == MVT::f32) {
5887 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
5888 // the result back to FR32 register. It's only worth matching if the
5889 // result has a single use which is a store or a bitcast to i32. And in
5890 // the case of a store, it's not worth it if the index is a constant 0,
5891 // because a MOVSSmr can be used instead, which is smaller and faster.
5892 if (!Op.hasOneUse())
5894 SDNode *User = *Op.getNode()->use_begin();
5895 if ((User->getOpcode() != ISD::STORE ||
5896 (isa<ConstantSDNode>(Op.getOperand(1)) &&
5897 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
5898 (User->getOpcode() != ISD::BITCAST ||
5899 User->getValueType(0) != MVT::i32))
5901 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5902 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32,
5905 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract);
5906 } else if (VT == MVT::i32) {
5907 // ExtractPS works with constant index.
5908 if (isa<ConstantSDNode>(Op.getOperand(1)))
5916 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
5917 SelectionDAG &DAG) const {
5918 if (!isa<ConstantSDNode>(Op.getOperand(1)))
5921 SDValue Vec = Op.getOperand(0);
5922 EVT VecVT = Vec.getValueType();
5924 // If this is a 256-bit vector result, first extract the 128-bit
5925 // vector and then extract from the 128-bit vector.
5926 if (VecVT.getSizeInBits() > 128) {
5927 DebugLoc dl = Op.getNode()->getDebugLoc();
5928 unsigned NumElems = VecVT.getVectorNumElements();
5929 SDValue Idx = Op.getOperand(1);
5931 if (!isa<ConstantSDNode>(Idx))
5934 unsigned ExtractNumElems = NumElems / (VecVT.getSizeInBits() / 128);
5935 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
5937 // Get the 128-bit vector.
5938 bool Upper = IdxVal >= ExtractNumElems;
5939 Vec = Extract128BitVector(Vec, Idx, DAG, dl);
5942 SDValue ScaledIdx = Idx;
5944 ScaledIdx = DAG.getNode(ISD::SUB, dl, Idx.getValueType(), Idx,
5945 DAG.getConstant(ExtractNumElems,
5946 Idx.getValueType()));
5947 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
5951 assert(Vec.getValueSizeInBits() <= 128 && "Unexpected vector length");
5953 if (Subtarget->hasSSE41()) {
5954 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
5959 EVT VT = Op.getValueType();
5960 DebugLoc dl = Op.getDebugLoc();
5961 // TODO: handle v16i8.
5962 if (VT.getSizeInBits() == 16) {
5963 SDValue Vec = Op.getOperand(0);
5964 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5966 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
5967 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5968 DAG.getNode(ISD::BITCAST, dl,
5971 // Transform it so it match pextrw which produces a 32-bit result.
5972 EVT EltVT = MVT::i32;
5973 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
5974 Op.getOperand(0), Op.getOperand(1));
5975 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
5976 DAG.getValueType(VT));
5977 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5978 } else if (VT.getSizeInBits() == 32) {
5979 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5983 // SHUFPS the element to the lowest double word, then movss.
5984 int Mask[4] = { Idx, -1, -1, -1 };
5985 EVT VVT = Op.getOperand(0).getValueType();
5986 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
5987 DAG.getUNDEF(VVT), Mask);
5988 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
5989 DAG.getIntPtrConstant(0));
5990 } else if (VT.getSizeInBits() == 64) {
5991 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
5992 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
5993 // to match extract_elt for f64.
5994 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5998 // UNPCKHPD the element to the lowest double word, then movsd.
5999 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
6000 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
6001 int Mask[2] = { 1, -1 };
6002 EVT VVT = Op.getOperand(0).getValueType();
6003 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
6004 DAG.getUNDEF(VVT), Mask);
6005 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
6006 DAG.getIntPtrConstant(0));
6013 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
6014 SelectionDAG &DAG) const {
6015 EVT VT = Op.getValueType();
6016 EVT EltVT = VT.getVectorElementType();
6017 DebugLoc dl = Op.getDebugLoc();
6019 SDValue N0 = Op.getOperand(0);
6020 SDValue N1 = Op.getOperand(1);
6021 SDValue N2 = Op.getOperand(2);
6023 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
6024 isa<ConstantSDNode>(N2)) {
6026 if (VT == MVT::v8i16)
6027 Opc = X86ISD::PINSRW;
6028 else if (VT == MVT::v16i8)
6029 Opc = X86ISD::PINSRB;
6031 Opc = X86ISD::PINSRB;
6033 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
6035 if (N1.getValueType() != MVT::i32)
6036 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
6037 if (N2.getValueType() != MVT::i32)
6038 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
6039 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
6040 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
6041 // Bits [7:6] of the constant are the source select. This will always be
6042 // zero here. The DAG Combiner may combine an extract_elt index into these
6043 // bits. For example (insert (extract, 3), 2) could be matched by putting
6044 // the '3' into bits [7:6] of X86ISD::INSERTPS.
6045 // Bits [5:4] of the constant are the destination select. This is the
6046 // value of the incoming immediate.
6047 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
6048 // combine either bitwise AND or insert of float 0.0 to set these bits.
6049 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
6050 // Create this as a scalar to vector..
6051 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
6052 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
6053 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
6054 // PINSR* works with constant index.
6061 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
6062 EVT VT = Op.getValueType();
6063 EVT EltVT = VT.getVectorElementType();
6065 DebugLoc dl = Op.getDebugLoc();
6066 SDValue N0 = Op.getOperand(0);
6067 SDValue N1 = Op.getOperand(1);
6068 SDValue N2 = Op.getOperand(2);
6070 // If this is a 256-bit vector result, first insert into a 128-bit
6071 // vector and then insert into the 256-bit vector.
6072 if (VT.getSizeInBits() > 128) {
6073 if (!isa<ConstantSDNode>(N2))
6076 // Get the 128-bit vector.
6077 unsigned NumElems = VT.getVectorNumElements();
6078 unsigned IdxVal = cast<ConstantSDNode>(N2)->getZExtValue();
6079 bool Upper = IdxVal >= NumElems / 2;
6081 SDValue SubN0 = Extract128BitVector(N0, N2, DAG, dl);
6084 SDValue ScaledN2 = N2;
6086 ScaledN2 = DAG.getNode(ISD::SUB, dl, N2.getValueType(), N2,
6087 DAG.getConstant(NumElems /
6088 (VT.getSizeInBits() / 128),
6089 N2.getValueType()));
6090 Op = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, SubN0.getValueType(), SubN0,
6093 // Insert the 128-bit vector
6094 // FIXME: Why UNDEF?
6095 return Insert128BitVector(N0, Op, N2, DAG, dl);
6098 if (Subtarget->hasSSE41())
6099 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
6101 if (EltVT == MVT::i8)
6104 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
6105 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
6106 // as its second argument.
6107 if (N1.getValueType() != MVT::i32)
6108 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
6109 if (N2.getValueType() != MVT::i32)
6110 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
6111 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
6117 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
6118 LLVMContext *Context = DAG.getContext();
6119 DebugLoc dl = Op.getDebugLoc();
6120 EVT OpVT = Op.getValueType();
6122 // If this is a 256-bit vector result, first insert into a 128-bit
6123 // vector and then insert into the 256-bit vector.
6124 if (OpVT.getSizeInBits() > 128) {
6125 // Insert into a 128-bit vector.
6126 EVT VT128 = EVT::getVectorVT(*Context,
6127 OpVT.getVectorElementType(),
6128 OpVT.getVectorNumElements() / 2);
6130 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
6132 // Insert the 128-bit vector.
6133 return Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, OpVT), Op,
6134 DAG.getConstant(0, MVT::i32),
6138 if (Op.getValueType() == MVT::v1i64 &&
6139 Op.getOperand(0).getValueType() == MVT::i64)
6140 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
6142 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
6143 assert(Op.getValueType().getSimpleVT().getSizeInBits() == 128 &&
6144 "Expected an SSE type!");
6145 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(),
6146 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
6149 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
6150 // a simple subregister reference or explicit instructions to grab
6151 // upper bits of a vector.
6153 X86TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const {
6154 if (Subtarget->hasAVX()) {
6155 DebugLoc dl = Op.getNode()->getDebugLoc();
6156 SDValue Vec = Op.getNode()->getOperand(0);
6157 SDValue Idx = Op.getNode()->getOperand(1);
6159 if (Op.getNode()->getValueType(0).getSizeInBits() == 128
6160 && Vec.getNode()->getValueType(0).getSizeInBits() == 256) {
6161 return Extract128BitVector(Vec, Idx, DAG, dl);
6167 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
6168 // simple superregister reference or explicit instructions to insert
6169 // the upper bits of a vector.
6171 X86TargetLowering::LowerINSERT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const {
6172 if (Subtarget->hasAVX()) {
6173 DebugLoc dl = Op.getNode()->getDebugLoc();
6174 SDValue Vec = Op.getNode()->getOperand(0);
6175 SDValue SubVec = Op.getNode()->getOperand(1);
6176 SDValue Idx = Op.getNode()->getOperand(2);
6178 if (Op.getNode()->getValueType(0).getSizeInBits() == 256
6179 && SubVec.getNode()->getValueType(0).getSizeInBits() == 128) {
6180 return Insert128BitVector(Vec, SubVec, Idx, DAG, dl);
6186 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
6187 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
6188 // one of the above mentioned nodes. It has to be wrapped because otherwise
6189 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
6190 // be used to form addressing mode. These wrapped nodes will be selected
6193 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
6194 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
6196 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6198 unsigned char OpFlag = 0;
6199 unsigned WrapperKind = X86ISD::Wrapper;
6200 CodeModel::Model M = getTargetMachine().getCodeModel();
6202 if (Subtarget->isPICStyleRIPRel() &&
6203 (M == CodeModel::Small || M == CodeModel::Kernel))
6204 WrapperKind = X86ISD::WrapperRIP;
6205 else if (Subtarget->isPICStyleGOT())
6206 OpFlag = X86II::MO_GOTOFF;
6207 else if (Subtarget->isPICStyleStubPIC())
6208 OpFlag = X86II::MO_PIC_BASE_OFFSET;
6210 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
6212 CP->getOffset(), OpFlag);
6213 DebugLoc DL = CP->getDebugLoc();
6214 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6215 // With PIC, the address is actually $g + Offset.
6217 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6218 DAG.getNode(X86ISD::GlobalBaseReg,
6219 DebugLoc(), getPointerTy()),
6226 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
6227 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
6229 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6231 unsigned char OpFlag = 0;
6232 unsigned WrapperKind = X86ISD::Wrapper;
6233 CodeModel::Model M = getTargetMachine().getCodeModel();
6235 if (Subtarget->isPICStyleRIPRel() &&
6236 (M == CodeModel::Small || M == CodeModel::Kernel))
6237 WrapperKind = X86ISD::WrapperRIP;
6238 else if (Subtarget->isPICStyleGOT())
6239 OpFlag = X86II::MO_GOTOFF;
6240 else if (Subtarget->isPICStyleStubPIC())
6241 OpFlag = X86II::MO_PIC_BASE_OFFSET;
6243 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
6245 DebugLoc DL = JT->getDebugLoc();
6246 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6248 // With PIC, the address is actually $g + Offset.
6250 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6251 DAG.getNode(X86ISD::GlobalBaseReg,
6252 DebugLoc(), getPointerTy()),
6259 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
6260 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
6262 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6264 unsigned char OpFlag = 0;
6265 unsigned WrapperKind = X86ISD::Wrapper;
6266 CodeModel::Model M = getTargetMachine().getCodeModel();
6268 if (Subtarget->isPICStyleRIPRel() &&
6269 (M == CodeModel::Small || M == CodeModel::Kernel))
6270 WrapperKind = X86ISD::WrapperRIP;
6271 else if (Subtarget->isPICStyleGOT())
6272 OpFlag = X86II::MO_GOTOFF;
6273 else if (Subtarget->isPICStyleStubPIC())
6274 OpFlag = X86II::MO_PIC_BASE_OFFSET;
6276 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
6278 DebugLoc DL = Op.getDebugLoc();
6279 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6282 // With PIC, the address is actually $g + Offset.
6283 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
6284 !Subtarget->is64Bit()) {
6285 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6286 DAG.getNode(X86ISD::GlobalBaseReg,
6287 DebugLoc(), getPointerTy()),
6295 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
6296 // Create the TargetBlockAddressAddress node.
6297 unsigned char OpFlags =
6298 Subtarget->ClassifyBlockAddressReference();
6299 CodeModel::Model M = getTargetMachine().getCodeModel();
6300 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
6301 DebugLoc dl = Op.getDebugLoc();
6302 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
6303 /*isTarget=*/true, OpFlags);
6305 if (Subtarget->isPICStyleRIPRel() &&
6306 (M == CodeModel::Small || M == CodeModel::Kernel))
6307 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
6309 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
6311 // With PIC, the address is actually $g + Offset.
6312 if (isGlobalRelativeToPICBase(OpFlags)) {
6313 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6314 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
6322 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
6324 SelectionDAG &DAG) const {
6325 // Create the TargetGlobalAddress node, folding in the constant
6326 // offset if it is legal.
6327 unsigned char OpFlags =
6328 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
6329 CodeModel::Model M = getTargetMachine().getCodeModel();
6331 if (OpFlags == X86II::MO_NO_FLAG &&
6332 X86::isOffsetSuitableForCodeModel(Offset, M)) {
6333 // A direct static reference to a global.
6334 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
6337 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
6340 if (Subtarget->isPICStyleRIPRel() &&
6341 (M == CodeModel::Small || M == CodeModel::Kernel))
6342 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
6344 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
6346 // With PIC, the address is actually $g + Offset.
6347 if (isGlobalRelativeToPICBase(OpFlags)) {
6348 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6349 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
6353 // For globals that require a load from a stub to get the address, emit the
6355 if (isGlobalStubReference(OpFlags))
6356 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
6357 MachinePointerInfo::getGOT(), false, false, 0);
6359 // If there was a non-zero offset that we didn't fold, create an explicit
6362 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
6363 DAG.getConstant(Offset, getPointerTy()));
6369 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
6370 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
6371 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
6372 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
6376 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
6377 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
6378 unsigned char OperandFlags) {
6379 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6380 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
6381 DebugLoc dl = GA->getDebugLoc();
6382 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
6383 GA->getValueType(0),
6387 SDValue Ops[] = { Chain, TGA, *InFlag };
6388 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
6390 SDValue Ops[] = { Chain, TGA };
6391 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
6394 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
6395 MFI->setAdjustsStack(true);
6397 SDValue Flag = Chain.getValue(1);
6398 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
6401 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
6403 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6406 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
6407 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
6408 DAG.getNode(X86ISD::GlobalBaseReg,
6409 DebugLoc(), PtrVT), InFlag);
6410 InFlag = Chain.getValue(1);
6412 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
6415 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
6417 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6419 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
6420 X86::RAX, X86II::MO_TLSGD);
6423 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
6424 // "local exec" model.
6425 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6426 const EVT PtrVT, TLSModel::Model model,
6428 DebugLoc dl = GA->getDebugLoc();
6430 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
6431 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
6432 is64Bit ? 257 : 256));
6434 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
6435 DAG.getIntPtrConstant(0),
6436 MachinePointerInfo(Ptr), false, false, 0);
6438 unsigned char OperandFlags = 0;
6439 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
6441 unsigned WrapperKind = X86ISD::Wrapper;
6442 if (model == TLSModel::LocalExec) {
6443 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
6444 } else if (is64Bit) {
6445 assert(model == TLSModel::InitialExec);
6446 OperandFlags = X86II::MO_GOTTPOFF;
6447 WrapperKind = X86ISD::WrapperRIP;
6449 assert(model == TLSModel::InitialExec);
6450 OperandFlags = X86II::MO_INDNTPOFF;
6453 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
6455 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
6456 GA->getValueType(0),
6457 GA->getOffset(), OperandFlags);
6458 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
6460 if (model == TLSModel::InitialExec)
6461 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
6462 MachinePointerInfo::getGOT(), false, false, 0);
6464 // The address of the thread local variable is the add of the thread
6465 // pointer with the offset of the variable.
6466 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
6470 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
6472 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
6473 const GlobalValue *GV = GA->getGlobal();
6475 if (Subtarget->isTargetELF()) {
6476 // TODO: implement the "local dynamic" model
6477 // TODO: implement the "initial exec"model for pic executables
6479 // If GV is an alias then use the aliasee for determining
6480 // thread-localness.
6481 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
6482 GV = GA->resolveAliasedGlobal(false);
6484 TLSModel::Model model
6485 = getTLSModel(GV, getTargetMachine().getRelocationModel());
6488 case TLSModel::GeneralDynamic:
6489 case TLSModel::LocalDynamic: // not implemented
6490 if (Subtarget->is64Bit())
6491 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
6492 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
6494 case TLSModel::InitialExec:
6495 case TLSModel::LocalExec:
6496 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
6497 Subtarget->is64Bit());
6499 } else if (Subtarget->isTargetDarwin()) {
6500 // Darwin only has one model of TLS. Lower to that.
6501 unsigned char OpFlag = 0;
6502 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
6503 X86ISD::WrapperRIP : X86ISD::Wrapper;
6505 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6507 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
6508 !Subtarget->is64Bit();
6510 OpFlag = X86II::MO_TLVP_PIC_BASE;
6512 OpFlag = X86II::MO_TLVP;
6513 DebugLoc DL = Op.getDebugLoc();
6514 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
6515 GA->getValueType(0),
6516 GA->getOffset(), OpFlag);
6517 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6519 // With PIC32, the address is actually $g + Offset.
6521 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6522 DAG.getNode(X86ISD::GlobalBaseReg,
6523 DebugLoc(), getPointerTy()),
6526 // Lowering the machine isd will make sure everything is in the right
6528 SDValue Chain = DAG.getEntryNode();
6529 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
6530 SDValue Args[] = { Chain, Offset };
6531 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args, 2);
6533 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
6534 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6535 MFI->setAdjustsStack(true);
6537 // And our return value (tls address) is in the standard call return value
6539 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
6540 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
6544 "TLS not implemented for this target.");
6546 llvm_unreachable("Unreachable");
6551 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
6552 /// take a 2 x i32 value to shift plus a shift amount.
6553 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
6554 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
6555 EVT VT = Op.getValueType();
6556 unsigned VTBits = VT.getSizeInBits();
6557 DebugLoc dl = Op.getDebugLoc();
6558 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
6559 SDValue ShOpLo = Op.getOperand(0);
6560 SDValue ShOpHi = Op.getOperand(1);
6561 SDValue ShAmt = Op.getOperand(2);
6562 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
6563 DAG.getConstant(VTBits - 1, MVT::i8))
6564 : DAG.getConstant(0, VT);
6567 if (Op.getOpcode() == ISD::SHL_PARTS) {
6568 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
6569 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
6571 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
6572 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
6575 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
6576 DAG.getConstant(VTBits, MVT::i8));
6577 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6578 AndNode, DAG.getConstant(0, MVT::i8));
6581 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6582 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
6583 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
6585 if (Op.getOpcode() == ISD::SHL_PARTS) {
6586 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
6587 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
6589 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
6590 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
6593 SDValue Ops[2] = { Lo, Hi };
6594 return DAG.getMergeValues(Ops, 2, dl);
6597 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
6598 SelectionDAG &DAG) const {
6599 EVT SrcVT = Op.getOperand(0).getValueType();
6601 if (SrcVT.isVector())
6604 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
6605 "Unknown SINT_TO_FP to lower!");
6607 // These are really Legal; return the operand so the caller accepts it as
6609 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
6611 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
6612 Subtarget->is64Bit()) {
6616 DebugLoc dl = Op.getDebugLoc();
6617 unsigned Size = SrcVT.getSizeInBits()/8;
6618 MachineFunction &MF = DAG.getMachineFunction();
6619 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
6620 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6621 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6623 MachinePointerInfo::getFixedStack(SSFI),
6625 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
6628 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
6630 SelectionDAG &DAG) const {
6632 DebugLoc DL = Op.getDebugLoc();
6634 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
6636 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
6638 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
6640 unsigned ByteSize = SrcVT.getSizeInBits()/8;
6642 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
6643 MachineMemOperand *MMO =
6644 DAG.getMachineFunction()
6645 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6646 MachineMemOperand::MOLoad, ByteSize, ByteSize);
6648 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
6649 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
6651 Tys, Ops, array_lengthof(Ops),
6655 Chain = Result.getValue(1);
6656 SDValue InFlag = Result.getValue(2);
6658 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
6659 // shouldn't be necessary except that RFP cannot be live across
6660 // multiple blocks. When stackifier is fixed, they can be uncoupled.
6661 MachineFunction &MF = DAG.getMachineFunction();
6662 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
6663 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
6664 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6665 Tys = DAG.getVTList(MVT::Other);
6667 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
6669 MachineMemOperand *MMO =
6670 DAG.getMachineFunction()
6671 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6672 MachineMemOperand::MOStore, SSFISize, SSFISize);
6674 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
6675 Ops, array_lengthof(Ops),
6676 Op.getValueType(), MMO);
6677 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
6678 MachinePointerInfo::getFixedStack(SSFI),
6685 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
6686 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
6687 SelectionDAG &DAG) const {
6688 // This algorithm is not obvious. Here it is in C code, more or less:
6690 double uint64_to_double( uint32_t hi, uint32_t lo ) {
6691 static const __m128i exp = { 0x4330000045300000ULL, 0 };
6692 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
6694 // Copy ints to xmm registers.
6695 __m128i xh = _mm_cvtsi32_si128( hi );
6696 __m128i xl = _mm_cvtsi32_si128( lo );
6698 // Combine into low half of a single xmm register.
6699 __m128i x = _mm_unpacklo_epi32( xh, xl );
6703 // Merge in appropriate exponents to give the integer bits the right
6705 x = _mm_unpacklo_epi32( x, exp );
6707 // Subtract away the biases to deal with the IEEE-754 double precision
6709 d = _mm_sub_pd( (__m128d) x, bias );
6711 // All conversions up to here are exact. The correctly rounded result is
6712 // calculated using the current rounding mode using the following
6714 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
6715 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
6716 // store doesn't really need to be here (except
6717 // maybe to zero the other double)
6722 DebugLoc dl = Op.getDebugLoc();
6723 LLVMContext *Context = DAG.getContext();
6725 // Build some magic constants.
6726 std::vector<Constant*> CV0;
6727 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
6728 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
6729 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
6730 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
6731 Constant *C0 = ConstantVector::get(CV0);
6732 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
6734 std::vector<Constant*> CV1;
6736 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
6738 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
6739 Constant *C1 = ConstantVector::get(CV1);
6740 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
6742 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6743 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6745 DAG.getIntPtrConstant(1)));
6746 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6747 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6749 DAG.getIntPtrConstant(0)));
6750 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
6751 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
6752 MachinePointerInfo::getConstantPool(),
6754 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
6755 SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck2);
6756 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
6757 MachinePointerInfo::getConstantPool(),
6759 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
6761 // Add the halves; easiest way is to swap them into another reg first.
6762 int ShufMask[2] = { 1, -1 };
6763 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
6764 DAG.getUNDEF(MVT::v2f64), ShufMask);
6765 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
6766 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
6767 DAG.getIntPtrConstant(0));
6770 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
6771 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
6772 SelectionDAG &DAG) const {
6773 DebugLoc dl = Op.getDebugLoc();
6774 // FP constant to bias correct the final result.
6775 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
6778 // Load the 32-bit value into an XMM register.
6779 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6780 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6782 DAG.getIntPtrConstant(0)));
6784 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
6785 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load),
6786 DAG.getIntPtrConstant(0));
6788 // Or the load with the bias.
6789 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
6790 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
6791 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6793 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
6794 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6795 MVT::v2f64, Bias)));
6796 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
6797 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or),
6798 DAG.getIntPtrConstant(0));
6800 // Subtract the bias.
6801 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
6803 // Handle final rounding.
6804 EVT DestVT = Op.getValueType();
6806 if (DestVT.bitsLT(MVT::f64)) {
6807 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
6808 DAG.getIntPtrConstant(0));
6809 } else if (DestVT.bitsGT(MVT::f64)) {
6810 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
6813 // Handle final rounding.
6817 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
6818 SelectionDAG &DAG) const {
6819 SDValue N0 = Op.getOperand(0);
6820 DebugLoc dl = Op.getDebugLoc();
6822 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
6823 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
6824 // the optimization here.
6825 if (DAG.SignBitIsZero(N0))
6826 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
6828 EVT SrcVT = N0.getValueType();
6829 EVT DstVT = Op.getValueType();
6830 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
6831 return LowerUINT_TO_FP_i64(Op, DAG);
6832 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
6833 return LowerUINT_TO_FP_i32(Op, DAG);
6835 // Make a 64-bit buffer, and use it to build an FILD.
6836 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
6837 if (SrcVT == MVT::i32) {
6838 SDValue WordOff = DAG.getConstant(4, getPointerTy());
6839 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
6840 getPointerTy(), StackSlot, WordOff);
6841 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6842 StackSlot, MachinePointerInfo(),
6844 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
6845 OffsetSlot, MachinePointerInfo(),
6847 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
6851 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
6852 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6853 StackSlot, MachinePointerInfo(),
6855 // For i64 source, we need to add the appropriate power of 2 if the input
6856 // was negative. This is the same as the optimization in
6857 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
6858 // we must be careful to do the computation in x87 extended precision, not
6859 // in SSE. (The generic code can't know it's OK to do this, or how to.)
6860 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
6861 MachineMemOperand *MMO =
6862 DAG.getMachineFunction()
6863 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6864 MachineMemOperand::MOLoad, 8, 8);
6866 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
6867 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
6868 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops, 3,
6871 APInt FF(32, 0x5F800000ULL);
6873 // Check whether the sign bit is set.
6874 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
6875 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
6878 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
6879 SDValue FudgePtr = DAG.getConstantPool(
6880 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
6883 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
6884 SDValue Zero = DAG.getIntPtrConstant(0);
6885 SDValue Four = DAG.getIntPtrConstant(4);
6886 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
6888 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
6890 // Load the value out, extending it from f32 to f80.
6891 // FIXME: Avoid the extend by constructing the right constant pool?
6892 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
6893 FudgePtr, MachinePointerInfo::getConstantPool(),
6894 MVT::f32, false, false, 4);
6895 // Extend everything to 80 bits to force it to be done on x87.
6896 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
6897 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
6900 std::pair<SDValue,SDValue> X86TargetLowering::
6901 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
6902 DebugLoc DL = Op.getDebugLoc();
6904 EVT DstTy = Op.getValueType();
6907 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
6911 assert(DstTy.getSimpleVT() <= MVT::i64 &&
6912 DstTy.getSimpleVT() >= MVT::i16 &&
6913 "Unknown FP_TO_SINT to lower!");
6915 // These are really Legal.
6916 if (DstTy == MVT::i32 &&
6917 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
6918 return std::make_pair(SDValue(), SDValue());
6919 if (Subtarget->is64Bit() &&
6920 DstTy == MVT::i64 &&
6921 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
6922 return std::make_pair(SDValue(), SDValue());
6924 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
6926 MachineFunction &MF = DAG.getMachineFunction();
6927 unsigned MemSize = DstTy.getSizeInBits()/8;
6928 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
6929 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6934 switch (DstTy.getSimpleVT().SimpleTy) {
6935 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
6936 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
6937 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
6938 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
6941 SDValue Chain = DAG.getEntryNode();
6942 SDValue Value = Op.getOperand(0);
6943 EVT TheVT = Op.getOperand(0).getValueType();
6944 if (isScalarFPTypeInSSEReg(TheVT)) {
6945 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
6946 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
6947 MachinePointerInfo::getFixedStack(SSFI),
6949 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
6951 Chain, StackSlot, DAG.getValueType(TheVT)
6954 MachineMemOperand *MMO =
6955 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6956 MachineMemOperand::MOLoad, MemSize, MemSize);
6957 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, 3,
6959 Chain = Value.getValue(1);
6960 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
6961 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6964 MachineMemOperand *MMO =
6965 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6966 MachineMemOperand::MOStore, MemSize, MemSize);
6968 // Build the FP_TO_INT*_IN_MEM
6969 SDValue Ops[] = { Chain, Value, StackSlot };
6970 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
6971 Ops, 3, DstTy, MMO);
6973 return std::make_pair(FIST, StackSlot);
6976 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
6977 SelectionDAG &DAG) const {
6978 if (Op.getValueType().isVector())
6981 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
6982 SDValue FIST = Vals.first, StackSlot = Vals.second;
6983 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
6984 if (FIST.getNode() == 0) return Op;
6987 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
6988 FIST, StackSlot, MachinePointerInfo(), false, false, 0);
6991 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
6992 SelectionDAG &DAG) const {
6993 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
6994 SDValue FIST = Vals.first, StackSlot = Vals.second;
6995 assert(FIST.getNode() && "Unexpected failure");
6998 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
6999 FIST, StackSlot, MachinePointerInfo(), false, false, 0);
7002 SDValue X86TargetLowering::LowerFABS(SDValue Op,
7003 SelectionDAG &DAG) const {
7004 LLVMContext *Context = DAG.getContext();
7005 DebugLoc dl = Op.getDebugLoc();
7006 EVT VT = Op.getValueType();
7009 EltVT = VT.getVectorElementType();
7010 std::vector<Constant*> CV;
7011 if (EltVT == MVT::f64) {
7012 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
7016 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
7022 Constant *C = ConstantVector::get(CV);
7023 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7024 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7025 MachinePointerInfo::getConstantPool(),
7027 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
7030 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
7031 LLVMContext *Context = DAG.getContext();
7032 DebugLoc dl = Op.getDebugLoc();
7033 EVT VT = Op.getValueType();
7036 EltVT = VT.getVectorElementType();
7037 std::vector<Constant*> CV;
7038 if (EltVT == MVT::f64) {
7039 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
7043 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
7049 Constant *C = ConstantVector::get(CV);
7050 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7051 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7052 MachinePointerInfo::getConstantPool(),
7054 if (VT.isVector()) {
7055 return DAG.getNode(ISD::BITCAST, dl, VT,
7056 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
7057 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
7059 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, Mask)));
7061 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
7065 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
7066 LLVMContext *Context = DAG.getContext();
7067 SDValue Op0 = Op.getOperand(0);
7068 SDValue Op1 = Op.getOperand(1);
7069 DebugLoc dl = Op.getDebugLoc();
7070 EVT VT = Op.getValueType();
7071 EVT SrcVT = Op1.getValueType();
7073 // If second operand is smaller, extend it first.
7074 if (SrcVT.bitsLT(VT)) {
7075 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
7078 // And if it is bigger, shrink it first.
7079 if (SrcVT.bitsGT(VT)) {
7080 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
7084 // At this point the operands and the result should have the same
7085 // type, and that won't be f80 since that is not custom lowered.
7087 // First get the sign bit of second operand.
7088 std::vector<Constant*> CV;
7089 if (SrcVT == MVT::f64) {
7090 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
7091 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
7093 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
7094 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7095 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7096 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7098 Constant *C = ConstantVector::get(CV);
7099 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7100 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
7101 MachinePointerInfo::getConstantPool(),
7103 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
7105 // Shift sign bit right or left if the two operands have different types.
7106 if (SrcVT.bitsGT(VT)) {
7107 // Op0 is MVT::f32, Op1 is MVT::f64.
7108 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
7109 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
7110 DAG.getConstant(32, MVT::i32));
7111 SignBit = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, SignBit);
7112 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
7113 DAG.getIntPtrConstant(0));
7116 // Clear first operand sign bit.
7118 if (VT == MVT::f64) {
7119 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
7120 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
7122 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
7123 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7124 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7125 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7127 C = ConstantVector::get(CV);
7128 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7129 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7130 MachinePointerInfo::getConstantPool(),
7132 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
7134 // Or the value with the sign bit.
7135 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
7138 /// Emit nodes that will be selected as "test Op0,Op0", or something
7140 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
7141 SelectionDAG &DAG) const {
7142 DebugLoc dl = Op.getDebugLoc();
7144 // CF and OF aren't always set the way we want. Determine which
7145 // of these we need.
7146 bool NeedCF = false;
7147 bool NeedOF = false;
7150 case X86::COND_A: case X86::COND_AE:
7151 case X86::COND_B: case X86::COND_BE:
7154 case X86::COND_G: case X86::COND_GE:
7155 case X86::COND_L: case X86::COND_LE:
7156 case X86::COND_O: case X86::COND_NO:
7161 // See if we can use the EFLAGS value from the operand instead of
7162 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
7163 // we prove that the arithmetic won't overflow, we can't use OF or CF.
7164 if (Op.getResNo() != 0 || NeedOF || NeedCF)
7165 // Emit a CMP with 0, which is the TEST pattern.
7166 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
7167 DAG.getConstant(0, Op.getValueType()));
7169 unsigned Opcode = 0;
7170 unsigned NumOperands = 0;
7171 switch (Op.getNode()->getOpcode()) {
7173 // Due to an isel shortcoming, be conservative if this add is likely to be
7174 // selected as part of a load-modify-store instruction. When the root node
7175 // in a match is a store, isel doesn't know how to remap non-chain non-flag
7176 // uses of other nodes in the match, such as the ADD in this case. This
7177 // leads to the ADD being left around and reselected, with the result being
7178 // two adds in the output. Alas, even if none our users are stores, that
7179 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
7180 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
7181 // climbing the DAG back to the root, and it doesn't seem to be worth the
7183 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
7184 UE = Op.getNode()->use_end(); UI != UE; ++UI)
7185 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
7188 if (ConstantSDNode *C =
7189 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
7190 // An add of one will be selected as an INC.
7191 if (C->getAPIntValue() == 1) {
7192 Opcode = X86ISD::INC;
7197 // An add of negative one (subtract of one) will be selected as a DEC.
7198 if (C->getAPIntValue().isAllOnesValue()) {
7199 Opcode = X86ISD::DEC;
7205 // Otherwise use a regular EFLAGS-setting add.
7206 Opcode = X86ISD::ADD;
7210 // If the primary and result isn't used, don't bother using X86ISD::AND,
7211 // because a TEST instruction will be better.
7212 bool NonFlagUse = false;
7213 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
7214 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
7216 unsigned UOpNo = UI.getOperandNo();
7217 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
7218 // Look pass truncate.
7219 UOpNo = User->use_begin().getOperandNo();
7220 User = *User->use_begin();
7223 if (User->getOpcode() != ISD::BRCOND &&
7224 User->getOpcode() != ISD::SETCC &&
7225 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
7238 // Due to the ISEL shortcoming noted above, be conservative if this op is
7239 // likely to be selected as part of a load-modify-store instruction.
7240 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
7241 UE = Op.getNode()->use_end(); UI != UE; ++UI)
7242 if (UI->getOpcode() == ISD::STORE)
7245 // Otherwise use a regular EFLAGS-setting instruction.
7246 switch (Op.getNode()->getOpcode()) {
7247 default: llvm_unreachable("unexpected operator!");
7248 case ISD::SUB: Opcode = X86ISD::SUB; break;
7249 case ISD::OR: Opcode = X86ISD::OR; break;
7250 case ISD::XOR: Opcode = X86ISD::XOR; break;
7251 case ISD::AND: Opcode = X86ISD::AND; break;
7263 return SDValue(Op.getNode(), 1);
7270 // Emit a CMP with 0, which is the TEST pattern.
7271 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
7272 DAG.getConstant(0, Op.getValueType()));
7274 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
7275 SmallVector<SDValue, 4> Ops;
7276 for (unsigned i = 0; i != NumOperands; ++i)
7277 Ops.push_back(Op.getOperand(i));
7279 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
7280 DAG.ReplaceAllUsesWith(Op, New);
7281 return SDValue(New.getNode(), 1);
7284 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
7286 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
7287 SelectionDAG &DAG) const {
7288 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
7289 if (C->getAPIntValue() == 0)
7290 return EmitTest(Op0, X86CC, DAG);
7292 DebugLoc dl = Op0.getDebugLoc();
7293 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
7296 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
7297 /// if it's possible.
7298 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
7299 DebugLoc dl, SelectionDAG &DAG) const {
7300 SDValue Op0 = And.getOperand(0);
7301 SDValue Op1 = And.getOperand(1);
7302 if (Op0.getOpcode() == ISD::TRUNCATE)
7303 Op0 = Op0.getOperand(0);
7304 if (Op1.getOpcode() == ISD::TRUNCATE)
7305 Op1 = Op1.getOperand(0);
7308 if (Op1.getOpcode() == ISD::SHL)
7309 std::swap(Op0, Op1);
7310 if (Op0.getOpcode() == ISD::SHL) {
7311 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
7312 if (And00C->getZExtValue() == 1) {
7313 // If we looked past a truncate, check that it's only truncating away
7315 unsigned BitWidth = Op0.getValueSizeInBits();
7316 unsigned AndBitWidth = And.getValueSizeInBits();
7317 if (BitWidth > AndBitWidth) {
7318 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
7319 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
7320 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
7324 RHS = Op0.getOperand(1);
7326 } else if (Op1.getOpcode() == ISD::Constant) {
7327 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
7328 SDValue AndLHS = Op0;
7329 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
7330 LHS = AndLHS.getOperand(0);
7331 RHS = AndLHS.getOperand(1);
7335 if (LHS.getNode()) {
7336 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
7337 // instruction. Since the shift amount is in-range-or-undefined, we know
7338 // that doing a bittest on the i32 value is ok. We extend to i32 because
7339 // the encoding for the i16 version is larger than the i32 version.
7340 // Also promote i16 to i32 for performance / code size reason.
7341 if (LHS.getValueType() == MVT::i8 ||
7342 LHS.getValueType() == MVT::i16)
7343 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
7345 // If the operand types disagree, extend the shift amount to match. Since
7346 // BT ignores high bits (like shifts) we can use anyextend.
7347 if (LHS.getValueType() != RHS.getValueType())
7348 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
7350 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
7351 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
7352 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7353 DAG.getConstant(Cond, MVT::i8), BT);
7359 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
7360 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
7361 SDValue Op0 = Op.getOperand(0);
7362 SDValue Op1 = Op.getOperand(1);
7363 DebugLoc dl = Op.getDebugLoc();
7364 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
7366 // Optimize to BT if possible.
7367 // Lower (X & (1 << N)) == 0 to BT(X, N).
7368 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
7369 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
7370 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
7371 Op1.getOpcode() == ISD::Constant &&
7372 cast<ConstantSDNode>(Op1)->isNullValue() &&
7373 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
7374 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
7375 if (NewSetCC.getNode())
7379 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
7381 if (Op1.getOpcode() == ISD::Constant &&
7382 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
7383 cast<ConstantSDNode>(Op1)->isNullValue()) &&
7384 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
7386 // If the input is a setcc, then reuse the input setcc or use a new one with
7387 // the inverted condition.
7388 if (Op0.getOpcode() == X86ISD::SETCC) {
7389 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
7390 bool Invert = (CC == ISD::SETNE) ^
7391 cast<ConstantSDNode>(Op1)->isNullValue();
7392 if (!Invert) return Op0;
7394 CCode = X86::GetOppositeBranchCondition(CCode);
7395 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7396 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
7400 bool isFP = Op1.getValueType().isFloatingPoint();
7401 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
7402 if (X86CC == X86::COND_INVALID)
7405 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, DAG);
7406 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7407 DAG.getConstant(X86CC, MVT::i8), EFLAGS);
7410 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
7412 SDValue Op0 = Op.getOperand(0);
7413 SDValue Op1 = Op.getOperand(1);
7414 SDValue CC = Op.getOperand(2);
7415 EVT VT = Op.getValueType();
7416 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
7417 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
7418 DebugLoc dl = Op.getDebugLoc();
7422 EVT VT0 = Op0.getValueType();
7423 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
7424 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
7427 switch (SetCCOpcode) {
7430 case ISD::SETEQ: SSECC = 0; break;
7432 case ISD::SETGT: Swap = true; // Fallthrough
7434 case ISD::SETOLT: SSECC = 1; break;
7436 case ISD::SETGE: Swap = true; // Fallthrough
7438 case ISD::SETOLE: SSECC = 2; break;
7439 case ISD::SETUO: SSECC = 3; break;
7441 case ISD::SETNE: SSECC = 4; break;
7442 case ISD::SETULE: Swap = true;
7443 case ISD::SETUGE: SSECC = 5; break;
7444 case ISD::SETULT: Swap = true;
7445 case ISD::SETUGT: SSECC = 6; break;
7446 case ISD::SETO: SSECC = 7; break;
7449 std::swap(Op0, Op1);
7451 // In the two special cases we can't handle, emit two comparisons.
7453 if (SetCCOpcode == ISD::SETUEQ) {
7455 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
7456 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
7457 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
7459 else if (SetCCOpcode == ISD::SETONE) {
7461 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
7462 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
7463 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
7465 llvm_unreachable("Illegal FP comparison");
7467 // Handle all other FP comparisons here.
7468 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
7471 // We are handling one of the integer comparisons here. Since SSE only has
7472 // GT and EQ comparisons for integer, swapping operands and multiple
7473 // operations may be required for some comparisons.
7474 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
7475 bool Swap = false, Invert = false, FlipSigns = false;
7477 switch (VT.getSimpleVT().SimpleTy) {
7479 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
7480 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
7481 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
7482 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
7485 switch (SetCCOpcode) {
7487 case ISD::SETNE: Invert = true;
7488 case ISD::SETEQ: Opc = EQOpc; break;
7489 case ISD::SETLT: Swap = true;
7490 case ISD::SETGT: Opc = GTOpc; break;
7491 case ISD::SETGE: Swap = true;
7492 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
7493 case ISD::SETULT: Swap = true;
7494 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
7495 case ISD::SETUGE: Swap = true;
7496 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
7499 std::swap(Op0, Op1);
7501 // Since SSE has no unsigned integer comparisons, we need to flip the sign
7502 // bits of the inputs before performing those operations.
7504 EVT EltVT = VT.getVectorElementType();
7505 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
7507 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
7508 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
7510 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
7511 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
7514 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
7516 // If the logical-not of the result is required, perform that now.
7518 Result = DAG.getNOT(dl, Result, VT);
7523 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
7524 static bool isX86LogicalCmp(SDValue Op) {
7525 unsigned Opc = Op.getNode()->getOpcode();
7526 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
7528 if (Op.getResNo() == 1 &&
7529 (Opc == X86ISD::ADD ||
7530 Opc == X86ISD::SUB ||
7531 Opc == X86ISD::ADC ||
7532 Opc == X86ISD::SBB ||
7533 Opc == X86ISD::SMUL ||
7534 Opc == X86ISD::UMUL ||
7535 Opc == X86ISD::INC ||
7536 Opc == X86ISD::DEC ||
7537 Opc == X86ISD::OR ||
7538 Opc == X86ISD::XOR ||
7539 Opc == X86ISD::AND))
7542 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
7548 static bool isZero(SDValue V) {
7549 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
7550 return C && C->isNullValue();
7553 static bool isAllOnes(SDValue V) {
7554 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
7555 return C && C->isAllOnesValue();
7558 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
7559 bool addTest = true;
7560 SDValue Cond = Op.getOperand(0);
7561 SDValue Op1 = Op.getOperand(1);
7562 SDValue Op2 = Op.getOperand(2);
7563 DebugLoc DL = Op.getDebugLoc();
7566 if (Cond.getOpcode() == ISD::SETCC) {
7567 SDValue NewCond = LowerSETCC(Cond, DAG);
7568 if (NewCond.getNode())
7572 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
7573 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
7574 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
7575 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
7576 if (Cond.getOpcode() == X86ISD::SETCC &&
7577 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
7578 isZero(Cond.getOperand(1).getOperand(1))) {
7579 SDValue Cmp = Cond.getOperand(1);
7581 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
7583 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
7584 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
7585 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
7587 SDValue CmpOp0 = Cmp.getOperand(0);
7588 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
7589 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
7591 SDValue Res = // Res = 0 or -1.
7592 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
7593 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
7595 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
7596 Res = DAG.getNOT(DL, Res, Res.getValueType());
7598 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
7599 if (N2C == 0 || !N2C->isNullValue())
7600 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
7605 // Look past (and (setcc_carry (cmp ...)), 1).
7606 if (Cond.getOpcode() == ISD::AND &&
7607 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
7608 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
7609 if (C && C->getAPIntValue() == 1)
7610 Cond = Cond.getOperand(0);
7613 // If condition flag is set by a X86ISD::CMP, then use it as the condition
7614 // setting operand in place of the X86ISD::SETCC.
7615 if (Cond.getOpcode() == X86ISD::SETCC ||
7616 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
7617 CC = Cond.getOperand(0);
7619 SDValue Cmp = Cond.getOperand(1);
7620 unsigned Opc = Cmp.getOpcode();
7621 EVT VT = Op.getValueType();
7623 bool IllegalFPCMov = false;
7624 if (VT.isFloatingPoint() && !VT.isVector() &&
7625 !isScalarFPTypeInSSEReg(VT)) // FPStack?
7626 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
7628 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
7629 Opc == X86ISD::BT) { // FIXME
7636 // Look pass the truncate.
7637 if (Cond.getOpcode() == ISD::TRUNCATE)
7638 Cond = Cond.getOperand(0);
7640 // We know the result of AND is compared against zero. Try to match
7642 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
7643 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
7644 if (NewSetCC.getNode()) {
7645 CC = NewSetCC.getOperand(0);
7646 Cond = NewSetCC.getOperand(1);
7653 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7654 Cond = EmitTest(Cond, X86::COND_NE, DAG);
7657 // a < b ? -1 : 0 -> RES = ~setcc_carry
7658 // a < b ? 0 : -1 -> RES = setcc_carry
7659 // a >= b ? -1 : 0 -> RES = setcc_carry
7660 // a >= b ? 0 : -1 -> RES = ~setcc_carry
7661 if (Cond.getOpcode() == X86ISD::CMP) {
7662 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
7664 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
7665 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
7666 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
7667 DAG.getConstant(X86::COND_B, MVT::i8), Cond);
7668 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
7669 return DAG.getNOT(DL, Res, Res.getValueType());
7674 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
7675 // condition is true.
7676 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
7677 SDValue Ops[] = { Op2, Op1, CC, Cond };
7678 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops, array_lengthof(Ops));
7681 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
7682 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
7683 // from the AND / OR.
7684 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
7685 Opc = Op.getOpcode();
7686 if (Opc != ISD::OR && Opc != ISD::AND)
7688 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
7689 Op.getOperand(0).hasOneUse() &&
7690 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
7691 Op.getOperand(1).hasOneUse());
7694 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
7695 // 1 and that the SETCC node has a single use.
7696 static bool isXor1OfSetCC(SDValue Op) {
7697 if (Op.getOpcode() != ISD::XOR)
7699 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7700 if (N1C && N1C->getAPIntValue() == 1) {
7701 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
7702 Op.getOperand(0).hasOneUse();
7707 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
7708 bool addTest = true;
7709 SDValue Chain = Op.getOperand(0);
7710 SDValue Cond = Op.getOperand(1);
7711 SDValue Dest = Op.getOperand(2);
7712 DebugLoc dl = Op.getDebugLoc();
7715 if (Cond.getOpcode() == ISD::SETCC) {
7716 SDValue NewCond = LowerSETCC(Cond, DAG);
7717 if (NewCond.getNode())
7721 // FIXME: LowerXALUO doesn't handle these!!
7722 else if (Cond.getOpcode() == X86ISD::ADD ||
7723 Cond.getOpcode() == X86ISD::SUB ||
7724 Cond.getOpcode() == X86ISD::SMUL ||
7725 Cond.getOpcode() == X86ISD::UMUL)
7726 Cond = LowerXALUO(Cond, DAG);
7729 // Look pass (and (setcc_carry (cmp ...)), 1).
7730 if (Cond.getOpcode() == ISD::AND &&
7731 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
7732 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
7733 if (C && C->getAPIntValue() == 1)
7734 Cond = Cond.getOperand(0);
7737 // If condition flag is set by a X86ISD::CMP, then use it as the condition
7738 // setting operand in place of the X86ISD::SETCC.
7739 if (Cond.getOpcode() == X86ISD::SETCC ||
7740 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
7741 CC = Cond.getOperand(0);
7743 SDValue Cmp = Cond.getOperand(1);
7744 unsigned Opc = Cmp.getOpcode();
7745 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
7746 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
7750 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
7754 // These can only come from an arithmetic instruction with overflow,
7755 // e.g. SADDO, UADDO.
7756 Cond = Cond.getNode()->getOperand(1);
7763 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
7764 SDValue Cmp = Cond.getOperand(0).getOperand(1);
7765 if (CondOpc == ISD::OR) {
7766 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
7767 // two branches instead of an explicit OR instruction with a
7769 if (Cmp == Cond.getOperand(1).getOperand(1) &&
7770 isX86LogicalCmp(Cmp)) {
7771 CC = Cond.getOperand(0).getOperand(0);
7772 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7773 Chain, Dest, CC, Cmp);
7774 CC = Cond.getOperand(1).getOperand(0);
7778 } else { // ISD::AND
7779 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
7780 // two branches instead of an explicit AND instruction with a
7781 // separate test. However, we only do this if this block doesn't
7782 // have a fall-through edge, because this requires an explicit
7783 // jmp when the condition is false.
7784 if (Cmp == Cond.getOperand(1).getOperand(1) &&
7785 isX86LogicalCmp(Cmp) &&
7786 Op.getNode()->hasOneUse()) {
7787 X86::CondCode CCode =
7788 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
7789 CCode = X86::GetOppositeBranchCondition(CCode);
7790 CC = DAG.getConstant(CCode, MVT::i8);
7791 SDNode *User = *Op.getNode()->use_begin();
7792 // Look for an unconditional branch following this conditional branch.
7793 // We need this because we need to reverse the successors in order
7794 // to implement FCMP_OEQ.
7795 if (User->getOpcode() == ISD::BR) {
7796 SDValue FalseBB = User->getOperand(1);
7798 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
7799 assert(NewBR == User);
7803 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7804 Chain, Dest, CC, Cmp);
7805 X86::CondCode CCode =
7806 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
7807 CCode = X86::GetOppositeBranchCondition(CCode);
7808 CC = DAG.getConstant(CCode, MVT::i8);
7814 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
7815 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
7816 // It should be transformed during dag combiner except when the condition
7817 // is set by a arithmetics with overflow node.
7818 X86::CondCode CCode =
7819 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
7820 CCode = X86::GetOppositeBranchCondition(CCode);
7821 CC = DAG.getConstant(CCode, MVT::i8);
7822 Cond = Cond.getOperand(0).getOperand(1);
7828 // Look pass the truncate.
7829 if (Cond.getOpcode() == ISD::TRUNCATE)
7830 Cond = Cond.getOperand(0);
7832 // We know the result of AND is compared against zero. Try to match
7834 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
7835 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
7836 if (NewSetCC.getNode()) {
7837 CC = NewSetCC.getOperand(0);
7838 Cond = NewSetCC.getOperand(1);
7845 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7846 Cond = EmitTest(Cond, X86::COND_NE, DAG);
7848 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7849 Chain, Dest, CC, Cond);
7853 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
7854 // Calls to _alloca is needed to probe the stack when allocating more than 4k
7855 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
7856 // that the guard pages used by the OS virtual memory manager are allocated in
7857 // correct sequence.
7859 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
7860 SelectionDAG &DAG) const {
7861 assert((Subtarget->isTargetCygMing() || Subtarget->isTargetWindows()) &&
7862 "This should be used only on Windows targets");
7863 DebugLoc dl = Op.getDebugLoc();
7866 SDValue Chain = Op.getOperand(0);
7867 SDValue Size = Op.getOperand(1);
7868 // FIXME: Ensure alignment here
7872 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
7874 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
7875 Flag = Chain.getValue(1);
7877 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
7879 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
7880 Flag = Chain.getValue(1);
7882 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
7884 SDValue Ops1[2] = { Chain.getValue(0), Chain };
7885 return DAG.getMergeValues(Ops1, 2, dl);
7888 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
7889 MachineFunction &MF = DAG.getMachineFunction();
7890 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
7892 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
7893 DebugLoc DL = Op.getDebugLoc();
7895 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
7896 // vastart just stores the address of the VarArgsFrameIndex slot into the
7897 // memory location argument.
7898 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
7900 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
7901 MachinePointerInfo(SV), false, false, 0);
7905 // gp_offset (0 - 6 * 8)
7906 // fp_offset (48 - 48 + 8 * 16)
7907 // overflow_arg_area (point to parameters coming in memory).
7909 SmallVector<SDValue, 8> MemOps;
7910 SDValue FIN = Op.getOperand(1);
7912 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
7913 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
7915 FIN, MachinePointerInfo(SV), false, false, 0);
7916 MemOps.push_back(Store);
7919 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7920 FIN, DAG.getIntPtrConstant(4));
7921 Store = DAG.getStore(Op.getOperand(0), DL,
7922 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
7924 FIN, MachinePointerInfo(SV, 4), false, false, 0);
7925 MemOps.push_back(Store);
7927 // Store ptr to overflow_arg_area
7928 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7929 FIN, DAG.getIntPtrConstant(4));
7930 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
7932 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
7933 MachinePointerInfo(SV, 8),
7935 MemOps.push_back(Store);
7937 // Store ptr to reg_save_area.
7938 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7939 FIN, DAG.getIntPtrConstant(8));
7940 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
7942 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
7943 MachinePointerInfo(SV, 16), false, false, 0);
7944 MemOps.push_back(Store);
7945 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
7946 &MemOps[0], MemOps.size());
7949 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
7950 assert(Subtarget->is64Bit() &&
7951 "LowerVAARG only handles 64-bit va_arg!");
7952 assert((Subtarget->isTargetLinux() ||
7953 Subtarget->isTargetDarwin()) &&
7954 "Unhandled target in LowerVAARG");
7955 assert(Op.getNode()->getNumOperands() == 4);
7956 SDValue Chain = Op.getOperand(0);
7957 SDValue SrcPtr = Op.getOperand(1);
7958 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
7959 unsigned Align = Op.getConstantOperandVal(3);
7960 DebugLoc dl = Op.getDebugLoc();
7962 EVT ArgVT = Op.getNode()->getValueType(0);
7963 const Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
7964 uint32_t ArgSize = getTargetData()->getTypeAllocSize(ArgTy);
7967 // Decide which area this value should be read from.
7968 // TODO: Implement the AMD64 ABI in its entirety. This simple
7969 // selection mechanism works only for the basic types.
7970 if (ArgVT == MVT::f80) {
7971 llvm_unreachable("va_arg for f80 not yet implemented");
7972 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
7973 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
7974 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
7975 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
7977 llvm_unreachable("Unhandled argument type in LowerVAARG");
7981 // Sanity Check: Make sure using fp_offset makes sense.
7982 assert(!UseSoftFloat &&
7983 !(DAG.getMachineFunction()
7984 .getFunction()->hasFnAttr(Attribute::NoImplicitFloat)) &&
7985 Subtarget->hasXMM());
7988 // Insert VAARG_64 node into the DAG
7989 // VAARG_64 returns two values: Variable Argument Address, Chain
7990 SmallVector<SDValue, 11> InstOps;
7991 InstOps.push_back(Chain);
7992 InstOps.push_back(SrcPtr);
7993 InstOps.push_back(DAG.getConstant(ArgSize, MVT::i32));
7994 InstOps.push_back(DAG.getConstant(ArgMode, MVT::i8));
7995 InstOps.push_back(DAG.getConstant(Align, MVT::i32));
7996 SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
7997 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
7998 VTs, &InstOps[0], InstOps.size(),
8000 MachinePointerInfo(SV),
8005 Chain = VAARG.getValue(1);
8007 // Load the next argument and return it
8008 return DAG.getLoad(ArgVT, dl,
8011 MachinePointerInfo(),
8015 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
8016 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
8017 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
8018 SDValue Chain = Op.getOperand(0);
8019 SDValue DstPtr = Op.getOperand(1);
8020 SDValue SrcPtr = Op.getOperand(2);
8021 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
8022 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
8023 DebugLoc DL = Op.getDebugLoc();
8025 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
8026 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
8028 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
8032 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
8033 DebugLoc dl = Op.getDebugLoc();
8034 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8036 default: return SDValue(); // Don't custom lower most intrinsics.
8037 // Comparison intrinsics.
8038 case Intrinsic::x86_sse_comieq_ss:
8039 case Intrinsic::x86_sse_comilt_ss:
8040 case Intrinsic::x86_sse_comile_ss:
8041 case Intrinsic::x86_sse_comigt_ss:
8042 case Intrinsic::x86_sse_comige_ss:
8043 case Intrinsic::x86_sse_comineq_ss:
8044 case Intrinsic::x86_sse_ucomieq_ss:
8045 case Intrinsic::x86_sse_ucomilt_ss:
8046 case Intrinsic::x86_sse_ucomile_ss:
8047 case Intrinsic::x86_sse_ucomigt_ss:
8048 case Intrinsic::x86_sse_ucomige_ss:
8049 case Intrinsic::x86_sse_ucomineq_ss:
8050 case Intrinsic::x86_sse2_comieq_sd:
8051 case Intrinsic::x86_sse2_comilt_sd:
8052 case Intrinsic::x86_sse2_comile_sd:
8053 case Intrinsic::x86_sse2_comigt_sd:
8054 case Intrinsic::x86_sse2_comige_sd:
8055 case Intrinsic::x86_sse2_comineq_sd:
8056 case Intrinsic::x86_sse2_ucomieq_sd:
8057 case Intrinsic::x86_sse2_ucomilt_sd:
8058 case Intrinsic::x86_sse2_ucomile_sd:
8059 case Intrinsic::x86_sse2_ucomigt_sd:
8060 case Intrinsic::x86_sse2_ucomige_sd:
8061 case Intrinsic::x86_sse2_ucomineq_sd: {
8063 ISD::CondCode CC = ISD::SETCC_INVALID;
8066 case Intrinsic::x86_sse_comieq_ss:
8067 case Intrinsic::x86_sse2_comieq_sd:
8071 case Intrinsic::x86_sse_comilt_ss:
8072 case Intrinsic::x86_sse2_comilt_sd:
8076 case Intrinsic::x86_sse_comile_ss:
8077 case Intrinsic::x86_sse2_comile_sd:
8081 case Intrinsic::x86_sse_comigt_ss:
8082 case Intrinsic::x86_sse2_comigt_sd:
8086 case Intrinsic::x86_sse_comige_ss:
8087 case Intrinsic::x86_sse2_comige_sd:
8091 case Intrinsic::x86_sse_comineq_ss:
8092 case Intrinsic::x86_sse2_comineq_sd:
8096 case Intrinsic::x86_sse_ucomieq_ss:
8097 case Intrinsic::x86_sse2_ucomieq_sd:
8098 Opc = X86ISD::UCOMI;
8101 case Intrinsic::x86_sse_ucomilt_ss:
8102 case Intrinsic::x86_sse2_ucomilt_sd:
8103 Opc = X86ISD::UCOMI;
8106 case Intrinsic::x86_sse_ucomile_ss:
8107 case Intrinsic::x86_sse2_ucomile_sd:
8108 Opc = X86ISD::UCOMI;
8111 case Intrinsic::x86_sse_ucomigt_ss:
8112 case Intrinsic::x86_sse2_ucomigt_sd:
8113 Opc = X86ISD::UCOMI;
8116 case Intrinsic::x86_sse_ucomige_ss:
8117 case Intrinsic::x86_sse2_ucomige_sd:
8118 Opc = X86ISD::UCOMI;
8121 case Intrinsic::x86_sse_ucomineq_ss:
8122 case Intrinsic::x86_sse2_ucomineq_sd:
8123 Opc = X86ISD::UCOMI;
8128 SDValue LHS = Op.getOperand(1);
8129 SDValue RHS = Op.getOperand(2);
8130 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
8131 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
8132 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
8133 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
8134 DAG.getConstant(X86CC, MVT::i8), Cond);
8135 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
8137 // ptest and testp intrinsics. The intrinsic these come from are designed to
8138 // return an integer value, not just an instruction so lower it to the ptest
8139 // or testp pattern and a setcc for the result.
8140 case Intrinsic::x86_sse41_ptestz:
8141 case Intrinsic::x86_sse41_ptestc:
8142 case Intrinsic::x86_sse41_ptestnzc:
8143 case Intrinsic::x86_avx_ptestz_256:
8144 case Intrinsic::x86_avx_ptestc_256:
8145 case Intrinsic::x86_avx_ptestnzc_256:
8146 case Intrinsic::x86_avx_vtestz_ps:
8147 case Intrinsic::x86_avx_vtestc_ps:
8148 case Intrinsic::x86_avx_vtestnzc_ps:
8149 case Intrinsic::x86_avx_vtestz_pd:
8150 case Intrinsic::x86_avx_vtestc_pd:
8151 case Intrinsic::x86_avx_vtestnzc_pd:
8152 case Intrinsic::x86_avx_vtestz_ps_256:
8153 case Intrinsic::x86_avx_vtestc_ps_256:
8154 case Intrinsic::x86_avx_vtestnzc_ps_256:
8155 case Intrinsic::x86_avx_vtestz_pd_256:
8156 case Intrinsic::x86_avx_vtestc_pd_256:
8157 case Intrinsic::x86_avx_vtestnzc_pd_256: {
8158 bool IsTestPacked = false;
8161 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
8162 case Intrinsic::x86_avx_vtestz_ps:
8163 case Intrinsic::x86_avx_vtestz_pd:
8164 case Intrinsic::x86_avx_vtestz_ps_256:
8165 case Intrinsic::x86_avx_vtestz_pd_256:
8166 IsTestPacked = true; // Fallthrough
8167 case Intrinsic::x86_sse41_ptestz:
8168 case Intrinsic::x86_avx_ptestz_256:
8170 X86CC = X86::COND_E;
8172 case Intrinsic::x86_avx_vtestc_ps:
8173 case Intrinsic::x86_avx_vtestc_pd:
8174 case Intrinsic::x86_avx_vtestc_ps_256:
8175 case Intrinsic::x86_avx_vtestc_pd_256:
8176 IsTestPacked = true; // Fallthrough
8177 case Intrinsic::x86_sse41_ptestc:
8178 case Intrinsic::x86_avx_ptestc_256:
8180 X86CC = X86::COND_B;
8182 case Intrinsic::x86_avx_vtestnzc_ps:
8183 case Intrinsic::x86_avx_vtestnzc_pd:
8184 case Intrinsic::x86_avx_vtestnzc_ps_256:
8185 case Intrinsic::x86_avx_vtestnzc_pd_256:
8186 IsTestPacked = true; // Fallthrough
8187 case Intrinsic::x86_sse41_ptestnzc:
8188 case Intrinsic::x86_avx_ptestnzc_256:
8190 X86CC = X86::COND_A;
8194 SDValue LHS = Op.getOperand(1);
8195 SDValue RHS = Op.getOperand(2);
8196 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
8197 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
8198 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
8199 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
8200 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
8203 // Fix vector shift instructions where the last operand is a non-immediate
8205 case Intrinsic::x86_sse2_pslli_w:
8206 case Intrinsic::x86_sse2_pslli_d:
8207 case Intrinsic::x86_sse2_pslli_q:
8208 case Intrinsic::x86_sse2_psrli_w:
8209 case Intrinsic::x86_sse2_psrli_d:
8210 case Intrinsic::x86_sse2_psrli_q:
8211 case Intrinsic::x86_sse2_psrai_w:
8212 case Intrinsic::x86_sse2_psrai_d:
8213 case Intrinsic::x86_mmx_pslli_w:
8214 case Intrinsic::x86_mmx_pslli_d:
8215 case Intrinsic::x86_mmx_pslli_q:
8216 case Intrinsic::x86_mmx_psrli_w:
8217 case Intrinsic::x86_mmx_psrli_d:
8218 case Intrinsic::x86_mmx_psrli_q:
8219 case Intrinsic::x86_mmx_psrai_w:
8220 case Intrinsic::x86_mmx_psrai_d: {
8221 SDValue ShAmt = Op.getOperand(2);
8222 if (isa<ConstantSDNode>(ShAmt))
8225 unsigned NewIntNo = 0;
8226 EVT ShAmtVT = MVT::v4i32;
8228 case Intrinsic::x86_sse2_pslli_w:
8229 NewIntNo = Intrinsic::x86_sse2_psll_w;
8231 case Intrinsic::x86_sse2_pslli_d:
8232 NewIntNo = Intrinsic::x86_sse2_psll_d;
8234 case Intrinsic::x86_sse2_pslli_q:
8235 NewIntNo = Intrinsic::x86_sse2_psll_q;
8237 case Intrinsic::x86_sse2_psrli_w:
8238 NewIntNo = Intrinsic::x86_sse2_psrl_w;
8240 case Intrinsic::x86_sse2_psrli_d:
8241 NewIntNo = Intrinsic::x86_sse2_psrl_d;
8243 case Intrinsic::x86_sse2_psrli_q:
8244 NewIntNo = Intrinsic::x86_sse2_psrl_q;
8246 case Intrinsic::x86_sse2_psrai_w:
8247 NewIntNo = Intrinsic::x86_sse2_psra_w;
8249 case Intrinsic::x86_sse2_psrai_d:
8250 NewIntNo = Intrinsic::x86_sse2_psra_d;
8253 ShAmtVT = MVT::v2i32;
8255 case Intrinsic::x86_mmx_pslli_w:
8256 NewIntNo = Intrinsic::x86_mmx_psll_w;
8258 case Intrinsic::x86_mmx_pslli_d:
8259 NewIntNo = Intrinsic::x86_mmx_psll_d;
8261 case Intrinsic::x86_mmx_pslli_q:
8262 NewIntNo = Intrinsic::x86_mmx_psll_q;
8264 case Intrinsic::x86_mmx_psrli_w:
8265 NewIntNo = Intrinsic::x86_mmx_psrl_w;
8267 case Intrinsic::x86_mmx_psrli_d:
8268 NewIntNo = Intrinsic::x86_mmx_psrl_d;
8270 case Intrinsic::x86_mmx_psrli_q:
8271 NewIntNo = Intrinsic::x86_mmx_psrl_q;
8273 case Intrinsic::x86_mmx_psrai_w:
8274 NewIntNo = Intrinsic::x86_mmx_psra_w;
8276 case Intrinsic::x86_mmx_psrai_d:
8277 NewIntNo = Intrinsic::x86_mmx_psra_d;
8279 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
8285 // The vector shift intrinsics with scalars uses 32b shift amounts but
8286 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
8290 ShOps[1] = DAG.getConstant(0, MVT::i32);
8291 if (ShAmtVT == MVT::v4i32) {
8292 ShOps[2] = DAG.getUNDEF(MVT::i32);
8293 ShOps[3] = DAG.getUNDEF(MVT::i32);
8294 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
8296 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
8297 // FIXME this must be lowered to get rid of the invalid type.
8300 EVT VT = Op.getValueType();
8301 ShAmt = DAG.getNode(ISD::BITCAST, dl, VT, ShAmt);
8302 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8303 DAG.getConstant(NewIntNo, MVT::i32),
8304 Op.getOperand(1), ShAmt);
8309 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
8310 SelectionDAG &DAG) const {
8311 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8312 MFI->setReturnAddressIsTaken(true);
8314 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8315 DebugLoc dl = Op.getDebugLoc();
8318 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
8320 DAG.getConstant(TD->getPointerSize(),
8321 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
8322 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
8323 DAG.getNode(ISD::ADD, dl, getPointerTy(),
8325 MachinePointerInfo(), false, false, 0);
8328 // Just load the return address.
8329 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
8330 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
8331 RetAddrFI, MachinePointerInfo(), false, false, 0);
8334 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
8335 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8336 MFI->setFrameAddressIsTaken(true);
8338 EVT VT = Op.getValueType();
8339 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
8340 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8341 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
8342 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
8344 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
8345 MachinePointerInfo(),
8350 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
8351 SelectionDAG &DAG) const {
8352 return DAG.getIntPtrConstant(2*TD->getPointerSize());
8355 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
8356 MachineFunction &MF = DAG.getMachineFunction();
8357 SDValue Chain = Op.getOperand(0);
8358 SDValue Offset = Op.getOperand(1);
8359 SDValue Handler = Op.getOperand(2);
8360 DebugLoc dl = Op.getDebugLoc();
8362 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
8363 Subtarget->is64Bit() ? X86::RBP : X86::EBP,
8365 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
8367 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Frame,
8368 DAG.getIntPtrConstant(TD->getPointerSize()));
8369 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
8370 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
8372 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
8373 MF.getRegInfo().addLiveOut(StoreAddrReg);
8375 return DAG.getNode(X86ISD::EH_RETURN, dl,
8377 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
8380 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
8381 SelectionDAG &DAG) const {
8382 SDValue Root = Op.getOperand(0);
8383 SDValue Trmp = Op.getOperand(1); // trampoline
8384 SDValue FPtr = Op.getOperand(2); // nested function
8385 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
8386 DebugLoc dl = Op.getDebugLoc();
8388 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
8390 if (Subtarget->is64Bit()) {
8391 SDValue OutChains[6];
8393 // Large code-model.
8394 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
8395 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
8397 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
8398 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
8400 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
8402 // Load the pointer to the nested function into R11.
8403 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
8404 SDValue Addr = Trmp;
8405 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8406 Addr, MachinePointerInfo(TrmpAddr),
8409 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8410 DAG.getConstant(2, MVT::i64));
8411 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
8412 MachinePointerInfo(TrmpAddr, 2),
8415 // Load the 'nest' parameter value into R10.
8416 // R10 is specified in X86CallingConv.td
8417 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
8418 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8419 DAG.getConstant(10, MVT::i64));
8420 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8421 Addr, MachinePointerInfo(TrmpAddr, 10),
8424 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8425 DAG.getConstant(12, MVT::i64));
8426 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
8427 MachinePointerInfo(TrmpAddr, 12),
8430 // Jump to the nested function.
8431 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
8432 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8433 DAG.getConstant(20, MVT::i64));
8434 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8435 Addr, MachinePointerInfo(TrmpAddr, 20),
8438 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
8439 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8440 DAG.getConstant(22, MVT::i64));
8441 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
8442 MachinePointerInfo(TrmpAddr, 22),
8446 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
8447 return DAG.getMergeValues(Ops, 2, dl);
8449 const Function *Func =
8450 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
8451 CallingConv::ID CC = Func->getCallingConv();
8456 llvm_unreachable("Unsupported calling convention");
8457 case CallingConv::C:
8458 case CallingConv::X86_StdCall: {
8459 // Pass 'nest' parameter in ECX.
8460 // Must be kept in sync with X86CallingConv.td
8463 // Check that ECX wasn't needed by an 'inreg' parameter.
8464 const FunctionType *FTy = Func->getFunctionType();
8465 const AttrListPtr &Attrs = Func->getAttributes();
8467 if (!Attrs.isEmpty() && !Func->isVarArg()) {
8468 unsigned InRegCount = 0;
8471 for (FunctionType::param_iterator I = FTy->param_begin(),
8472 E = FTy->param_end(); I != E; ++I, ++Idx)
8473 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
8474 // FIXME: should only count parameters that are lowered to integers.
8475 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
8477 if (InRegCount > 2) {
8478 report_fatal_error("Nest register in use - reduce number of inreg"
8484 case CallingConv::X86_FastCall:
8485 case CallingConv::X86_ThisCall:
8486 case CallingConv::Fast:
8487 // Pass 'nest' parameter in EAX.
8488 // Must be kept in sync with X86CallingConv.td
8493 SDValue OutChains[4];
8496 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8497 DAG.getConstant(10, MVT::i32));
8498 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
8500 // This is storing the opcode for MOV32ri.
8501 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
8502 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
8503 OutChains[0] = DAG.getStore(Root, dl,
8504 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
8505 Trmp, MachinePointerInfo(TrmpAddr),
8508 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8509 DAG.getConstant(1, MVT::i32));
8510 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
8511 MachinePointerInfo(TrmpAddr, 1),
8514 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
8515 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8516 DAG.getConstant(5, MVT::i32));
8517 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
8518 MachinePointerInfo(TrmpAddr, 5),
8521 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8522 DAG.getConstant(6, MVT::i32));
8523 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
8524 MachinePointerInfo(TrmpAddr, 6),
8528 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
8529 return DAG.getMergeValues(Ops, 2, dl);
8533 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
8534 SelectionDAG &DAG) const {
8536 The rounding mode is in bits 11:10 of FPSR, and has the following
8543 FLT_ROUNDS, on the other hand, expects the following:
8550 To perform the conversion, we do:
8551 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
8554 MachineFunction &MF = DAG.getMachineFunction();
8555 const TargetMachine &TM = MF.getTarget();
8556 const TargetFrameLowering &TFI = *TM.getFrameLowering();
8557 unsigned StackAlignment = TFI.getStackAlignment();
8558 EVT VT = Op.getValueType();
8559 DebugLoc DL = Op.getDebugLoc();
8561 // Save FP Control Word to stack slot
8562 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
8563 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
8566 MachineMemOperand *MMO =
8567 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
8568 MachineMemOperand::MOStore, 2, 2);
8570 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
8571 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
8572 DAG.getVTList(MVT::Other),
8573 Ops, 2, MVT::i16, MMO);
8575 // Load FP Control Word from stack slot
8576 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
8577 MachinePointerInfo(), false, false, 0);
8579 // Transform as necessary
8581 DAG.getNode(ISD::SRL, DL, MVT::i16,
8582 DAG.getNode(ISD::AND, DL, MVT::i16,
8583 CWD, DAG.getConstant(0x800, MVT::i16)),
8584 DAG.getConstant(11, MVT::i8));
8586 DAG.getNode(ISD::SRL, DL, MVT::i16,
8587 DAG.getNode(ISD::AND, DL, MVT::i16,
8588 CWD, DAG.getConstant(0x400, MVT::i16)),
8589 DAG.getConstant(9, MVT::i8));
8592 DAG.getNode(ISD::AND, DL, MVT::i16,
8593 DAG.getNode(ISD::ADD, DL, MVT::i16,
8594 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
8595 DAG.getConstant(1, MVT::i16)),
8596 DAG.getConstant(3, MVT::i16));
8599 return DAG.getNode((VT.getSizeInBits() < 16 ?
8600 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
8603 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
8604 EVT VT = Op.getValueType();
8606 unsigned NumBits = VT.getSizeInBits();
8607 DebugLoc dl = Op.getDebugLoc();
8609 Op = Op.getOperand(0);
8610 if (VT == MVT::i8) {
8611 // Zero extend to i32 since there is not an i8 bsr.
8613 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
8616 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
8617 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
8618 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
8620 // If src is zero (i.e. bsr sets ZF), returns NumBits.
8623 DAG.getConstant(NumBits+NumBits-1, OpVT),
8624 DAG.getConstant(X86::COND_E, MVT::i8),
8627 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
8629 // Finally xor with NumBits-1.
8630 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
8633 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
8637 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
8638 EVT VT = Op.getValueType();
8640 unsigned NumBits = VT.getSizeInBits();
8641 DebugLoc dl = Op.getDebugLoc();
8643 Op = Op.getOperand(0);
8644 if (VT == MVT::i8) {
8646 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
8649 // Issue a bsf (scan bits forward) which also sets EFLAGS.
8650 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
8651 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
8653 // If src is zero (i.e. bsf sets ZF), returns NumBits.
8656 DAG.getConstant(NumBits, OpVT),
8657 DAG.getConstant(X86::COND_E, MVT::i8),
8660 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
8663 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
8667 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
8668 EVT VT = Op.getValueType();
8669 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
8670 DebugLoc dl = Op.getDebugLoc();
8672 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
8673 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
8674 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
8675 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
8676 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
8678 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
8679 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
8680 // return AloBlo + AloBhi + AhiBlo;
8682 SDValue A = Op.getOperand(0);
8683 SDValue B = Op.getOperand(1);
8685 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8686 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8687 A, DAG.getConstant(32, MVT::i32));
8688 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8689 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8690 B, DAG.getConstant(32, MVT::i32));
8691 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8692 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8694 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8695 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8697 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8698 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8700 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8701 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8702 AloBhi, DAG.getConstant(32, MVT::i32));
8703 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8704 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8705 AhiBlo, DAG.getConstant(32, MVT::i32));
8706 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
8707 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
8711 SDValue X86TargetLowering::LowerSHL(SDValue Op, SelectionDAG &DAG) const {
8712 EVT VT = Op.getValueType();
8713 DebugLoc dl = Op.getDebugLoc();
8714 SDValue R = Op.getOperand(0);
8716 LLVMContext *Context = DAG.getContext();
8718 assert(Subtarget->hasSSE41() && "Cannot lower SHL without SSE4.1 or later");
8720 if (VT == MVT::v4i32) {
8721 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8722 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8723 Op.getOperand(1), DAG.getConstant(23, MVT::i32));
8725 ConstantInt *CI = ConstantInt::get(*Context, APInt(32, 0x3f800000U));
8727 std::vector<Constant*> CV(4, CI);
8728 Constant *C = ConstantVector::get(CV);
8729 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8730 SDValue Addend = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8731 MachinePointerInfo::getConstantPool(),
8734 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Addend);
8735 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
8736 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
8737 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
8739 if (VT == MVT::v16i8) {
8741 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8742 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8743 Op.getOperand(1), DAG.getConstant(5, MVT::i32));
8745 ConstantInt *CM1 = ConstantInt::get(*Context, APInt(8, 15));
8746 ConstantInt *CM2 = ConstantInt::get(*Context, APInt(8, 63));
8748 std::vector<Constant*> CVM1(16, CM1);
8749 std::vector<Constant*> CVM2(16, CM2);
8750 Constant *C = ConstantVector::get(CVM1);
8751 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8752 SDValue M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8753 MachinePointerInfo::getConstantPool(),
8756 // r = pblendv(r, psllw(r & (char16)15, 4), a);
8757 M = DAG.getNode(ISD::AND, dl, VT, R, M);
8758 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8759 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
8760 DAG.getConstant(4, MVT::i32));
8761 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
8763 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
8765 C = ConstantVector::get(CVM2);
8766 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8767 M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8768 MachinePointerInfo::getConstantPool(),
8771 // r = pblendv(r, psllw(r & (char16)63, 2), a);
8772 M = DAG.getNode(ISD::AND, dl, VT, R, M);
8773 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8774 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
8775 DAG.getConstant(2, MVT::i32));
8776 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
8778 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
8780 // return pblendv(r, r+r, a);
8781 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT,
8782 R, DAG.getNode(ISD::ADD, dl, VT, R, R), Op);
8788 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
8789 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
8790 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
8791 // looks for this combo and may remove the "setcc" instruction if the "setcc"
8792 // has only one use.
8793 SDNode *N = Op.getNode();
8794 SDValue LHS = N->getOperand(0);
8795 SDValue RHS = N->getOperand(1);
8796 unsigned BaseOp = 0;
8798 DebugLoc DL = Op.getDebugLoc();
8799 switch (Op.getOpcode()) {
8800 default: llvm_unreachable("Unknown ovf instruction!");
8802 // A subtract of one will be selected as a INC. Note that INC doesn't
8803 // set CF, so we can't do this for UADDO.
8804 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
8805 if (C->getAPIntValue() == 1) {
8806 BaseOp = X86ISD::INC;
8810 BaseOp = X86ISD::ADD;
8814 BaseOp = X86ISD::ADD;
8818 // A subtract of one will be selected as a DEC. Note that DEC doesn't
8819 // set CF, so we can't do this for USUBO.
8820 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
8821 if (C->getAPIntValue() == 1) {
8822 BaseOp = X86ISD::DEC;
8826 BaseOp = X86ISD::SUB;
8830 BaseOp = X86ISD::SUB;
8834 BaseOp = X86ISD::SMUL;
8837 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
8838 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
8840 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
8843 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8844 DAG.getConstant(X86::COND_O, MVT::i32),
8845 SDValue(Sum.getNode(), 2));
8847 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
8852 // Also sets EFLAGS.
8853 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
8854 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
8857 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
8858 DAG.getConstant(Cond, MVT::i32),
8859 SDValue(Sum.getNode(), 1));
8861 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
8865 SDValue X86TargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const{
8866 DebugLoc dl = Op.getDebugLoc();
8868 if (!Subtarget->hasSSE2()) {
8869 SDValue Chain = Op.getOperand(0);
8870 SDValue Zero = DAG.getConstant(0,
8871 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
8873 DAG.getRegister(X86::ESP, MVT::i32), // Base
8874 DAG.getTargetConstant(1, MVT::i8), // Scale
8875 DAG.getRegister(0, MVT::i32), // Index
8876 DAG.getTargetConstant(0, MVT::i32), // Disp
8877 DAG.getRegister(0, MVT::i32), // Segment.
8882 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
8883 array_lengthof(Ops));
8884 return SDValue(Res, 0);
8887 unsigned isDev = cast<ConstantSDNode>(Op.getOperand(5))->getZExtValue();
8889 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
8891 unsigned Op1 = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
8892 unsigned Op2 = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
8893 unsigned Op3 = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
8894 unsigned Op4 = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
8896 // def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
8897 if (!Op1 && !Op2 && !Op3 && Op4)
8898 return DAG.getNode(X86ISD::SFENCE, dl, MVT::Other, Op.getOperand(0));
8900 // def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
8901 if (Op1 && !Op2 && !Op3 && !Op4)
8902 return DAG.getNode(X86ISD::LFENCE, dl, MVT::Other, Op.getOperand(0));
8904 // def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 1)),
8906 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
8909 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
8910 EVT T = Op.getValueType();
8911 DebugLoc DL = Op.getDebugLoc();
8914 switch(T.getSimpleVT().SimpleTy) {
8916 assert(false && "Invalid value type!");
8917 case MVT::i8: Reg = X86::AL; size = 1; break;
8918 case MVT::i16: Reg = X86::AX; size = 2; break;
8919 case MVT::i32: Reg = X86::EAX; size = 4; break;
8921 assert(Subtarget->is64Bit() && "Node not type legal!");
8922 Reg = X86::RAX; size = 8;
8925 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
8926 Op.getOperand(2), SDValue());
8927 SDValue Ops[] = { cpIn.getValue(0),
8930 DAG.getTargetConstant(size, MVT::i8),
8932 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
8933 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
8934 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
8937 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
8941 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
8942 SelectionDAG &DAG) const {
8943 assert(Subtarget->is64Bit() && "Result not type legalized?");
8944 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
8945 SDValue TheChain = Op.getOperand(0);
8946 DebugLoc dl = Op.getDebugLoc();
8947 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
8948 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
8949 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
8951 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
8952 DAG.getConstant(32, MVT::i8));
8954 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
8957 return DAG.getMergeValues(Ops, 2, dl);
8960 SDValue X86TargetLowering::LowerBITCAST(SDValue Op,
8961 SelectionDAG &DAG) const {
8962 EVT SrcVT = Op.getOperand(0).getValueType();
8963 EVT DstVT = Op.getValueType();
8964 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
8965 Subtarget->hasMMX() && "Unexpected custom BITCAST");
8966 assert((DstVT == MVT::i64 ||
8967 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
8968 "Unexpected custom BITCAST");
8969 // i64 <=> MMX conversions are Legal.
8970 if (SrcVT==MVT::i64 && DstVT.isVector())
8972 if (DstVT==MVT::i64 && SrcVT.isVector())
8974 // MMX <=> MMX conversions are Legal.
8975 if (SrcVT.isVector() && DstVT.isVector())
8977 // All other conversions need to be expanded.
8981 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
8982 SDNode *Node = Op.getNode();
8983 DebugLoc dl = Node->getDebugLoc();
8984 EVT T = Node->getValueType(0);
8985 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
8986 DAG.getConstant(0, T), Node->getOperand(2));
8987 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
8988 cast<AtomicSDNode>(Node)->getMemoryVT(),
8989 Node->getOperand(0),
8990 Node->getOperand(1), negOp,
8991 cast<AtomicSDNode>(Node)->getSrcValue(),
8992 cast<AtomicSDNode>(Node)->getAlignment());
8995 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
8996 EVT VT = Op.getNode()->getValueType(0);
8998 // Let legalize expand this if it isn't a legal type yet.
8999 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
9002 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
9005 bool ExtraOp = false;
9006 switch (Op.getOpcode()) {
9007 default: assert(0 && "Invalid code");
9008 case ISD::ADDC: Opc = X86ISD::ADD; break;
9009 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
9010 case ISD::SUBC: Opc = X86ISD::SUB; break;
9011 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
9015 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
9017 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
9018 Op.getOperand(1), Op.getOperand(2));
9021 /// LowerOperation - Provide custom lowering hooks for some operations.
9023 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
9024 switch (Op.getOpcode()) {
9025 default: llvm_unreachable("Should not custom lower this!");
9026 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op,DAG);
9027 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
9028 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
9029 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
9030 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
9031 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
9032 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
9033 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
9034 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
9035 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, DAG);
9036 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
9037 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
9038 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
9039 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
9040 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
9041 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
9042 case ISD::SHL_PARTS:
9043 case ISD::SRA_PARTS:
9044 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
9045 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
9046 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
9047 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
9048 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
9049 case ISD::FABS: return LowerFABS(Op, DAG);
9050 case ISD::FNEG: return LowerFNEG(Op, DAG);
9051 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
9052 case ISD::SETCC: return LowerSETCC(Op, DAG);
9053 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
9054 case ISD::SELECT: return LowerSELECT(Op, DAG);
9055 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
9056 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
9057 case ISD::VASTART: return LowerVASTART(Op, DAG);
9058 case ISD::VAARG: return LowerVAARG(Op, DAG);
9059 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
9060 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
9061 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
9062 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
9063 case ISD::FRAME_TO_ARGS_OFFSET:
9064 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
9065 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
9066 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
9067 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
9068 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
9069 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
9070 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
9071 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
9072 case ISD::SHL: return LowerSHL(Op, DAG);
9078 case ISD::UMULO: return LowerXALUO(Op, DAG);
9079 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
9080 case ISD::BITCAST: return LowerBITCAST(Op, DAG);
9084 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
9088 void X86TargetLowering::
9089 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
9090 SelectionDAG &DAG, unsigned NewOp) const {
9091 EVT T = Node->getValueType(0);
9092 DebugLoc dl = Node->getDebugLoc();
9093 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
9095 SDValue Chain = Node->getOperand(0);
9096 SDValue In1 = Node->getOperand(1);
9097 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
9098 Node->getOperand(2), DAG.getIntPtrConstant(0));
9099 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
9100 Node->getOperand(2), DAG.getIntPtrConstant(1));
9101 SDValue Ops[] = { Chain, In1, In2L, In2H };
9102 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
9104 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
9105 cast<MemSDNode>(Node)->getMemOperand());
9106 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
9107 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
9108 Results.push_back(Result.getValue(2));
9111 /// ReplaceNodeResults - Replace a node with an illegal result type
9112 /// with a new node built out of custom code.
9113 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
9114 SmallVectorImpl<SDValue>&Results,
9115 SelectionDAG &DAG) const {
9116 DebugLoc dl = N->getDebugLoc();
9117 switch (N->getOpcode()) {
9119 assert(false && "Do not know how to custom type legalize this operation!");
9125 // We don't want to expand or promote these.
9127 case ISD::FP_TO_SINT: {
9128 std::pair<SDValue,SDValue> Vals =
9129 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
9130 SDValue FIST = Vals.first, StackSlot = Vals.second;
9131 if (FIST.getNode() != 0) {
9132 EVT VT = N->getValueType(0);
9133 // Return a load from the stack slot.
9134 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
9135 MachinePointerInfo(), false, false, 0));
9139 case ISD::READCYCLECOUNTER: {
9140 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9141 SDValue TheChain = N->getOperand(0);
9142 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
9143 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
9145 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
9147 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
9148 SDValue Ops[] = { eax, edx };
9149 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
9150 Results.push_back(edx.getValue(1));
9153 case ISD::ATOMIC_CMP_SWAP: {
9154 EVT T = N->getValueType(0);
9155 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
9156 SDValue cpInL, cpInH;
9157 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
9158 DAG.getConstant(0, MVT::i32));
9159 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
9160 DAG.getConstant(1, MVT::i32));
9161 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
9162 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
9164 SDValue swapInL, swapInH;
9165 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
9166 DAG.getConstant(0, MVT::i32));
9167 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
9168 DAG.getConstant(1, MVT::i32));
9169 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
9171 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
9172 swapInL.getValue(1));
9173 SDValue Ops[] = { swapInH.getValue(0),
9175 swapInH.getValue(1) };
9176 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9177 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
9178 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG8_DAG, dl, Tys,
9180 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
9181 MVT::i32, Result.getValue(1));
9182 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
9183 MVT::i32, cpOutL.getValue(2));
9184 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
9185 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
9186 Results.push_back(cpOutH.getValue(1));
9189 case ISD::ATOMIC_LOAD_ADD:
9190 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
9192 case ISD::ATOMIC_LOAD_AND:
9193 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
9195 case ISD::ATOMIC_LOAD_NAND:
9196 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
9198 case ISD::ATOMIC_LOAD_OR:
9199 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
9201 case ISD::ATOMIC_LOAD_SUB:
9202 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
9204 case ISD::ATOMIC_LOAD_XOR:
9205 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
9207 case ISD::ATOMIC_SWAP:
9208 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
9213 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
9215 default: return NULL;
9216 case X86ISD::BSF: return "X86ISD::BSF";
9217 case X86ISD::BSR: return "X86ISD::BSR";
9218 case X86ISD::SHLD: return "X86ISD::SHLD";
9219 case X86ISD::SHRD: return "X86ISD::SHRD";
9220 case X86ISD::FAND: return "X86ISD::FAND";
9221 case X86ISD::FOR: return "X86ISD::FOR";
9222 case X86ISD::FXOR: return "X86ISD::FXOR";
9223 case X86ISD::FSRL: return "X86ISD::FSRL";
9224 case X86ISD::FILD: return "X86ISD::FILD";
9225 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
9226 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
9227 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
9228 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
9229 case X86ISD::FLD: return "X86ISD::FLD";
9230 case X86ISD::FST: return "X86ISD::FST";
9231 case X86ISD::CALL: return "X86ISD::CALL";
9232 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
9233 case X86ISD::BT: return "X86ISD::BT";
9234 case X86ISD::CMP: return "X86ISD::CMP";
9235 case X86ISD::COMI: return "X86ISD::COMI";
9236 case X86ISD::UCOMI: return "X86ISD::UCOMI";
9237 case X86ISD::SETCC: return "X86ISD::SETCC";
9238 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
9239 case X86ISD::CMOV: return "X86ISD::CMOV";
9240 case X86ISD::BRCOND: return "X86ISD::BRCOND";
9241 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
9242 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
9243 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
9244 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
9245 case X86ISD::Wrapper: return "X86ISD::Wrapper";
9246 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
9247 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
9248 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
9249 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
9250 case X86ISD::PINSRB: return "X86ISD::PINSRB";
9251 case X86ISD::PINSRW: return "X86ISD::PINSRW";
9252 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
9253 case X86ISD::PANDN: return "X86ISD::PANDN";
9254 case X86ISD::PSIGNB: return "X86ISD::PSIGNB";
9255 case X86ISD::PSIGNW: return "X86ISD::PSIGNW";
9256 case X86ISD::PSIGND: return "X86ISD::PSIGND";
9257 case X86ISD::PBLENDVB: return "X86ISD::PBLENDVB";
9258 case X86ISD::FMAX: return "X86ISD::FMAX";
9259 case X86ISD::FMIN: return "X86ISD::FMIN";
9260 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
9261 case X86ISD::FRCP: return "X86ISD::FRCP";
9262 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
9263 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
9264 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
9265 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
9266 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
9267 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
9268 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
9269 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
9270 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
9271 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
9272 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
9273 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
9274 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
9275 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
9276 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
9277 case X86ISD::VSHL: return "X86ISD::VSHL";
9278 case X86ISD::VSRL: return "X86ISD::VSRL";
9279 case X86ISD::CMPPD: return "X86ISD::CMPPD";
9280 case X86ISD::CMPPS: return "X86ISD::CMPPS";
9281 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
9282 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
9283 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
9284 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
9285 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
9286 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
9287 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
9288 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
9289 case X86ISD::ADD: return "X86ISD::ADD";
9290 case X86ISD::SUB: return "X86ISD::SUB";
9291 case X86ISD::ADC: return "X86ISD::ADC";
9292 case X86ISD::SBB: return "X86ISD::SBB";
9293 case X86ISD::SMUL: return "X86ISD::SMUL";
9294 case X86ISD::UMUL: return "X86ISD::UMUL";
9295 case X86ISD::INC: return "X86ISD::INC";
9296 case X86ISD::DEC: return "X86ISD::DEC";
9297 case X86ISD::OR: return "X86ISD::OR";
9298 case X86ISD::XOR: return "X86ISD::XOR";
9299 case X86ISD::AND: return "X86ISD::AND";
9300 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
9301 case X86ISD::PTEST: return "X86ISD::PTEST";
9302 case X86ISD::TESTP: return "X86ISD::TESTP";
9303 case X86ISD::PALIGN: return "X86ISD::PALIGN";
9304 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
9305 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
9306 case X86ISD::PSHUFHW_LD: return "X86ISD::PSHUFHW_LD";
9307 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
9308 case X86ISD::PSHUFLW_LD: return "X86ISD::PSHUFLW_LD";
9309 case X86ISD::SHUFPS: return "X86ISD::SHUFPS";
9310 case X86ISD::SHUFPD: return "X86ISD::SHUFPD";
9311 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
9312 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
9313 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
9314 case X86ISD::MOVHLPD: return "X86ISD::MOVHLPD";
9315 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
9316 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
9317 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
9318 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
9319 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
9320 case X86ISD::MOVSHDUP_LD: return "X86ISD::MOVSHDUP_LD";
9321 case X86ISD::MOVSLDUP_LD: return "X86ISD::MOVSLDUP_LD";
9322 case X86ISD::MOVSD: return "X86ISD::MOVSD";
9323 case X86ISD::MOVSS: return "X86ISD::MOVSS";
9324 case X86ISD::UNPCKLPS: return "X86ISD::UNPCKLPS";
9325 case X86ISD::UNPCKLPD: return "X86ISD::UNPCKLPD";
9326 case X86ISD::UNPCKHPS: return "X86ISD::UNPCKHPS";
9327 case X86ISD::UNPCKHPD: return "X86ISD::UNPCKHPD";
9328 case X86ISD::PUNPCKLBW: return "X86ISD::PUNPCKLBW";
9329 case X86ISD::PUNPCKLWD: return "X86ISD::PUNPCKLWD";
9330 case X86ISD::PUNPCKLDQ: return "X86ISD::PUNPCKLDQ";
9331 case X86ISD::PUNPCKLQDQ: return "X86ISD::PUNPCKLQDQ";
9332 case X86ISD::PUNPCKHBW: return "X86ISD::PUNPCKHBW";
9333 case X86ISD::PUNPCKHWD: return "X86ISD::PUNPCKHWD";
9334 case X86ISD::PUNPCKHDQ: return "X86ISD::PUNPCKHDQ";
9335 case X86ISD::PUNPCKHQDQ: return "X86ISD::PUNPCKHQDQ";
9336 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
9337 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
9338 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
9342 // isLegalAddressingMode - Return true if the addressing mode represented
9343 // by AM is legal for this target, for a load/store of the specified type.
9344 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
9345 const Type *Ty) const {
9346 // X86 supports extremely general addressing modes.
9347 CodeModel::Model M = getTargetMachine().getCodeModel();
9348 Reloc::Model R = getTargetMachine().getRelocationModel();
9350 // X86 allows a sign-extended 32-bit immediate field as a displacement.
9351 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
9356 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
9358 // If a reference to this global requires an extra load, we can't fold it.
9359 if (isGlobalStubReference(GVFlags))
9362 // If BaseGV requires a register for the PIC base, we cannot also have a
9363 // BaseReg specified.
9364 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
9367 // If lower 4G is not available, then we must use rip-relative addressing.
9368 if ((M != CodeModel::Small || R != Reloc::Static) &&
9369 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
9379 // These scales always work.
9384 // These scales are formed with basereg+scalereg. Only accept if there is
9389 default: // Other stuff never works.
9397 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
9398 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
9400 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
9401 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
9402 if (NumBits1 <= NumBits2)
9407 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
9408 if (!VT1.isInteger() || !VT2.isInteger())
9410 unsigned NumBits1 = VT1.getSizeInBits();
9411 unsigned NumBits2 = VT2.getSizeInBits();
9412 if (NumBits1 <= NumBits2)
9417 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
9418 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
9419 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
9422 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
9423 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
9424 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
9427 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
9428 // i16 instructions are longer (0x66 prefix) and potentially slower.
9429 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
9432 /// isShuffleMaskLegal - Targets can use this to indicate that they only
9433 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
9434 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
9435 /// are assumed to be legal.
9437 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
9439 // Very little shuffling can be done for 64-bit vectors right now.
9440 if (VT.getSizeInBits() == 64)
9441 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
9443 // FIXME: pshufb, blends, shifts.
9444 return (VT.getVectorNumElements() == 2 ||
9445 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
9446 isMOVLMask(M, VT) ||
9447 isSHUFPMask(M, VT) ||
9448 isPSHUFDMask(M, VT) ||
9449 isPSHUFHWMask(M, VT) ||
9450 isPSHUFLWMask(M, VT) ||
9451 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
9452 isUNPCKLMask(M, VT) ||
9453 isUNPCKHMask(M, VT) ||
9454 isUNPCKL_v_undef_Mask(M, VT) ||
9455 isUNPCKH_v_undef_Mask(M, VT));
9459 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
9461 unsigned NumElts = VT.getVectorNumElements();
9462 // FIXME: This collection of masks seems suspect.
9465 if (NumElts == 4 && VT.getSizeInBits() == 128) {
9466 return (isMOVLMask(Mask, VT) ||
9467 isCommutedMOVLMask(Mask, VT, true) ||
9468 isSHUFPMask(Mask, VT) ||
9469 isCommutedSHUFPMask(Mask, VT));
9474 //===----------------------------------------------------------------------===//
9475 // X86 Scheduler Hooks
9476 //===----------------------------------------------------------------------===//
9478 // private utility function
9480 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
9481 MachineBasicBlock *MBB,
9488 TargetRegisterClass *RC,
9489 bool invSrc) const {
9490 // For the atomic bitwise operator, we generate
9493 // ld t1 = [bitinstr.addr]
9494 // op t2 = t1, [bitinstr.val]
9496 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
9498 // fallthrough -->nextMBB
9499 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9500 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9501 MachineFunction::iterator MBBIter = MBB;
9504 /// First build the CFG
9505 MachineFunction *F = MBB->getParent();
9506 MachineBasicBlock *thisMBB = MBB;
9507 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
9508 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
9509 F->insert(MBBIter, newMBB);
9510 F->insert(MBBIter, nextMBB);
9512 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
9513 nextMBB->splice(nextMBB->begin(), thisMBB,
9514 llvm::next(MachineBasicBlock::iterator(bInstr)),
9516 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
9518 // Update thisMBB to fall through to newMBB
9519 thisMBB->addSuccessor(newMBB);
9521 // newMBB jumps to itself and fall through to nextMBB
9522 newMBB->addSuccessor(nextMBB);
9523 newMBB->addSuccessor(newMBB);
9525 // Insert instructions into newMBB based on incoming instruction
9526 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
9527 "unexpected number of operands");
9528 DebugLoc dl = bInstr->getDebugLoc();
9529 MachineOperand& destOper = bInstr->getOperand(0);
9530 MachineOperand* argOpers[2 + X86::AddrNumOperands];
9531 int numArgs = bInstr->getNumOperands() - 1;
9532 for (int i=0; i < numArgs; ++i)
9533 argOpers[i] = &bInstr->getOperand(i+1);
9535 // x86 address has 4 operands: base, index, scale, and displacement
9536 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
9537 int valArgIndx = lastAddrIndx + 1;
9539 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
9540 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
9541 for (int i=0; i <= lastAddrIndx; ++i)
9542 (*MIB).addOperand(*argOpers[i]);
9544 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
9546 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
9551 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
9552 assert((argOpers[valArgIndx]->isReg() ||
9553 argOpers[valArgIndx]->isImm()) &&
9555 if (argOpers[valArgIndx]->isReg())
9556 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
9558 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
9560 (*MIB).addOperand(*argOpers[valArgIndx]);
9562 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), EAXreg);
9565 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
9566 for (int i=0; i <= lastAddrIndx; ++i)
9567 (*MIB).addOperand(*argOpers[i]);
9569 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
9570 (*MIB).setMemRefs(bInstr->memoperands_begin(),
9571 bInstr->memoperands_end());
9573 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
9577 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
9579 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
9583 // private utility function: 64 bit atomics on 32 bit host.
9585 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
9586 MachineBasicBlock *MBB,
9591 bool invSrc) const {
9592 // For the atomic bitwise operator, we generate
9593 // thisMBB (instructions are in pairs, except cmpxchg8b)
9594 // ld t1,t2 = [bitinstr.addr]
9596 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
9597 // op t5, t6 <- out1, out2, [bitinstr.val]
9598 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
9599 // mov ECX, EBX <- t5, t6
9600 // mov EAX, EDX <- t1, t2
9601 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
9602 // mov t3, t4 <- EAX, EDX
9604 // result in out1, out2
9605 // fallthrough -->nextMBB
9607 const TargetRegisterClass *RC = X86::GR32RegisterClass;
9608 const unsigned LoadOpc = X86::MOV32rm;
9609 const unsigned NotOpc = X86::NOT32r;
9610 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9611 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9612 MachineFunction::iterator MBBIter = MBB;
9615 /// First build the CFG
9616 MachineFunction *F = MBB->getParent();
9617 MachineBasicBlock *thisMBB = MBB;
9618 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
9619 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
9620 F->insert(MBBIter, newMBB);
9621 F->insert(MBBIter, nextMBB);
9623 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
9624 nextMBB->splice(nextMBB->begin(), thisMBB,
9625 llvm::next(MachineBasicBlock::iterator(bInstr)),
9627 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
9629 // Update thisMBB to fall through to newMBB
9630 thisMBB->addSuccessor(newMBB);
9632 // newMBB jumps to itself and fall through to nextMBB
9633 newMBB->addSuccessor(nextMBB);
9634 newMBB->addSuccessor(newMBB);
9636 DebugLoc dl = bInstr->getDebugLoc();
9637 // Insert instructions into newMBB based on incoming instruction
9638 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
9639 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 14 &&
9640 "unexpected number of operands");
9641 MachineOperand& dest1Oper = bInstr->getOperand(0);
9642 MachineOperand& dest2Oper = bInstr->getOperand(1);
9643 MachineOperand* argOpers[2 + X86::AddrNumOperands];
9644 for (int i=0; i < 2 + X86::AddrNumOperands; ++i) {
9645 argOpers[i] = &bInstr->getOperand(i+2);
9647 // We use some of the operands multiple times, so conservatively just
9648 // clear any kill flags that might be present.
9649 if (argOpers[i]->isReg() && argOpers[i]->isUse())
9650 argOpers[i]->setIsKill(false);
9653 // x86 address has 5 operands: base, index, scale, displacement, and segment.
9654 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
9656 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
9657 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
9658 for (int i=0; i <= lastAddrIndx; ++i)
9659 (*MIB).addOperand(*argOpers[i]);
9660 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
9661 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
9662 // add 4 to displacement.
9663 for (int i=0; i <= lastAddrIndx-2; ++i)
9664 (*MIB).addOperand(*argOpers[i]);
9665 MachineOperand newOp3 = *(argOpers[3]);
9667 newOp3.setImm(newOp3.getImm()+4);
9669 newOp3.setOffset(newOp3.getOffset()+4);
9670 (*MIB).addOperand(newOp3);
9671 (*MIB).addOperand(*argOpers[lastAddrIndx]);
9673 // t3/4 are defined later, at the bottom of the loop
9674 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
9675 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
9676 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
9677 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
9678 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
9679 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
9681 // The subsequent operations should be using the destination registers of
9682 //the PHI instructions.
9684 t1 = F->getRegInfo().createVirtualRegister(RC);
9685 t2 = F->getRegInfo().createVirtualRegister(RC);
9686 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
9687 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
9689 t1 = dest1Oper.getReg();
9690 t2 = dest2Oper.getReg();
9693 int valArgIndx = lastAddrIndx + 1;
9694 assert((argOpers[valArgIndx]->isReg() ||
9695 argOpers[valArgIndx]->isImm()) &&
9697 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
9698 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
9699 if (argOpers[valArgIndx]->isReg())
9700 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
9702 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
9703 if (regOpcL != X86::MOV32rr)
9705 (*MIB).addOperand(*argOpers[valArgIndx]);
9706 assert(argOpers[valArgIndx + 1]->isReg() ==
9707 argOpers[valArgIndx]->isReg());
9708 assert(argOpers[valArgIndx + 1]->isImm() ==
9709 argOpers[valArgIndx]->isImm());
9710 if (argOpers[valArgIndx + 1]->isReg())
9711 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
9713 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
9714 if (regOpcH != X86::MOV32rr)
9716 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
9718 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
9720 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EDX);
9723 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EBX);
9725 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::ECX);
9728 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
9729 for (int i=0; i <= lastAddrIndx; ++i)
9730 (*MIB).addOperand(*argOpers[i]);
9732 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
9733 (*MIB).setMemRefs(bInstr->memoperands_begin(),
9734 bInstr->memoperands_end());
9736 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t3);
9737 MIB.addReg(X86::EAX);
9738 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t4);
9739 MIB.addReg(X86::EDX);
9742 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
9744 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
9748 // private utility function
9750 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
9751 MachineBasicBlock *MBB,
9752 unsigned cmovOpc) const {
9753 // For the atomic min/max operator, we generate
9756 // ld t1 = [min/max.addr]
9757 // mov t2 = [min/max.val]
9759 // cmov[cond] t2 = t1
9761 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
9763 // fallthrough -->nextMBB
9765 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9766 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9767 MachineFunction::iterator MBBIter = MBB;
9770 /// First build the CFG
9771 MachineFunction *F = MBB->getParent();
9772 MachineBasicBlock *thisMBB = MBB;
9773 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
9774 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
9775 F->insert(MBBIter, newMBB);
9776 F->insert(MBBIter, nextMBB);
9778 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
9779 nextMBB->splice(nextMBB->begin(), thisMBB,
9780 llvm::next(MachineBasicBlock::iterator(mInstr)),
9782 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
9784 // Update thisMBB to fall through to newMBB
9785 thisMBB->addSuccessor(newMBB);
9787 // newMBB jumps to newMBB and fall through to nextMBB
9788 newMBB->addSuccessor(nextMBB);
9789 newMBB->addSuccessor(newMBB);
9791 DebugLoc dl = mInstr->getDebugLoc();
9792 // Insert instructions into newMBB based on incoming instruction
9793 assert(mInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
9794 "unexpected number of operands");
9795 MachineOperand& destOper = mInstr->getOperand(0);
9796 MachineOperand* argOpers[2 + X86::AddrNumOperands];
9797 int numArgs = mInstr->getNumOperands() - 1;
9798 for (int i=0; i < numArgs; ++i)
9799 argOpers[i] = &mInstr->getOperand(i+1);
9801 // x86 address has 4 operands: base, index, scale, and displacement
9802 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
9803 int valArgIndx = lastAddrIndx + 1;
9805 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9806 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
9807 for (int i=0; i <= lastAddrIndx; ++i)
9808 (*MIB).addOperand(*argOpers[i]);
9810 // We only support register and immediate values
9811 assert((argOpers[valArgIndx]->isReg() ||
9812 argOpers[valArgIndx]->isImm()) &&
9815 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9816 if (argOpers[valArgIndx]->isReg())
9817 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t2);
9819 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
9820 (*MIB).addOperand(*argOpers[valArgIndx]);
9822 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
9825 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
9830 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9831 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
9835 // Cmp and exchange if none has modified the memory location
9836 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
9837 for (int i=0; i <= lastAddrIndx; ++i)
9838 (*MIB).addOperand(*argOpers[i]);
9840 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
9841 (*MIB).setMemRefs(mInstr->memoperands_begin(),
9842 mInstr->memoperands_end());
9844 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
9845 MIB.addReg(X86::EAX);
9848 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
9850 mInstr->eraseFromParent(); // The pseudo instruction is gone now.
9854 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
9855 // or XMM0_V32I8 in AVX all of this code can be replaced with that
9858 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
9859 unsigned numArgs, bool memArg) const {
9860 assert((Subtarget->hasSSE42() || Subtarget->hasAVX()) &&
9861 "Target must have SSE4.2 or AVX features enabled");
9863 DebugLoc dl = MI->getDebugLoc();
9864 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9866 if (!Subtarget->hasAVX()) {
9868 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
9870 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
9873 Opc = numArgs == 3 ? X86::VPCMPISTRM128rm : X86::VPCMPESTRM128rm;
9875 Opc = numArgs == 3 ? X86::VPCMPISTRM128rr : X86::VPCMPESTRM128rr;
9878 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
9879 for (unsigned i = 0; i < numArgs; ++i) {
9880 MachineOperand &Op = MI->getOperand(i+1);
9881 if (!(Op.isReg() && Op.isImplicit()))
9884 BuildMI(*BB, MI, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
9887 MI->eraseFromParent();
9892 X86TargetLowering::EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB) const {
9893 DebugLoc dl = MI->getDebugLoc();
9894 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9896 // Address into RAX/EAX, other two args into ECX, EDX.
9897 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
9898 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
9899 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
9900 for (int i = 0; i < X86::AddrNumOperands; ++i)
9901 MIB.addOperand(MI->getOperand(i));
9903 unsigned ValOps = X86::AddrNumOperands;
9904 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
9905 .addReg(MI->getOperand(ValOps).getReg());
9906 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
9907 .addReg(MI->getOperand(ValOps+1).getReg());
9909 // The instruction doesn't actually take any operands though.
9910 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
9912 MI->eraseFromParent(); // The pseudo is gone now.
9917 X86TargetLowering::EmitMwait(MachineInstr *MI, MachineBasicBlock *BB) const {
9918 DebugLoc dl = MI->getDebugLoc();
9919 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9921 // First arg in ECX, the second in EAX.
9922 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
9923 .addReg(MI->getOperand(0).getReg());
9924 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EAX)
9925 .addReg(MI->getOperand(1).getReg());
9927 // The instruction doesn't actually take any operands though.
9928 BuildMI(*BB, MI, dl, TII->get(X86::MWAITrr));
9930 MI->eraseFromParent(); // The pseudo is gone now.
9935 X86TargetLowering::EmitVAARG64WithCustomInserter(
9937 MachineBasicBlock *MBB) const {
9938 // Emit va_arg instruction on X86-64.
9940 // Operands to this pseudo-instruction:
9941 // 0 ) Output : destination address (reg)
9942 // 1-5) Input : va_list address (addr, i64mem)
9943 // 6 ) ArgSize : Size (in bytes) of vararg type
9944 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
9945 // 8 ) Align : Alignment of type
9946 // 9 ) EFLAGS (implicit-def)
9948 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
9949 assert(X86::AddrNumOperands == 5 && "VAARG_64 assumes 5 address operands");
9951 unsigned DestReg = MI->getOperand(0).getReg();
9952 MachineOperand &Base = MI->getOperand(1);
9953 MachineOperand &Scale = MI->getOperand(2);
9954 MachineOperand &Index = MI->getOperand(3);
9955 MachineOperand &Disp = MI->getOperand(4);
9956 MachineOperand &Segment = MI->getOperand(5);
9957 unsigned ArgSize = MI->getOperand(6).getImm();
9958 unsigned ArgMode = MI->getOperand(7).getImm();
9959 unsigned Align = MI->getOperand(8).getImm();
9962 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
9963 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
9964 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
9966 // Machine Information
9967 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9968 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
9969 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
9970 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
9971 DebugLoc DL = MI->getDebugLoc();
9976 // i64 overflow_area (address)
9977 // i64 reg_save_area (address)
9979 // sizeof(va_list) = 24
9980 // alignment(va_list) = 8
9982 unsigned TotalNumIntRegs = 6;
9983 unsigned TotalNumXMMRegs = 8;
9984 bool UseGPOffset = (ArgMode == 1);
9985 bool UseFPOffset = (ArgMode == 2);
9986 unsigned MaxOffset = TotalNumIntRegs * 8 +
9987 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
9989 /* Align ArgSize to a multiple of 8 */
9990 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
9991 bool NeedsAlign = (Align > 8);
9993 MachineBasicBlock *thisMBB = MBB;
9994 MachineBasicBlock *overflowMBB;
9995 MachineBasicBlock *offsetMBB;
9996 MachineBasicBlock *endMBB;
9998 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
9999 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
10000 unsigned OffsetReg = 0;
10002 if (!UseGPOffset && !UseFPOffset) {
10003 // If we only pull from the overflow region, we don't create a branch.
10004 // We don't need to alter control flow.
10005 OffsetDestReg = 0; // unused
10006 OverflowDestReg = DestReg;
10009 overflowMBB = thisMBB;
10012 // First emit code to check if gp_offset (or fp_offset) is below the bound.
10013 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
10014 // If not, pull from overflow_area. (branch to overflowMBB)
10019 // offsetMBB overflowMBB
10024 // Registers for the PHI in endMBB
10025 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
10026 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
10028 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10029 MachineFunction *MF = MBB->getParent();
10030 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
10031 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
10032 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
10034 MachineFunction::iterator MBBIter = MBB;
10037 // Insert the new basic blocks
10038 MF->insert(MBBIter, offsetMBB);
10039 MF->insert(MBBIter, overflowMBB);
10040 MF->insert(MBBIter, endMBB);
10042 // Transfer the remainder of MBB and its successor edges to endMBB.
10043 endMBB->splice(endMBB->begin(), thisMBB,
10044 llvm::next(MachineBasicBlock::iterator(MI)),
10046 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
10048 // Make offsetMBB and overflowMBB successors of thisMBB
10049 thisMBB->addSuccessor(offsetMBB);
10050 thisMBB->addSuccessor(overflowMBB);
10052 // endMBB is a successor of both offsetMBB and overflowMBB
10053 offsetMBB->addSuccessor(endMBB);
10054 overflowMBB->addSuccessor(endMBB);
10056 // Load the offset value into a register
10057 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
10058 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
10062 .addDisp(Disp, UseFPOffset ? 4 : 0)
10063 .addOperand(Segment)
10064 .setMemRefs(MMOBegin, MMOEnd);
10066 // Check if there is enough room left to pull this argument.
10067 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
10069 .addImm(MaxOffset + 8 - ArgSizeA8);
10071 // Branch to "overflowMBB" if offset >= max
10072 // Fall through to "offsetMBB" otherwise
10073 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
10074 .addMBB(overflowMBB);
10077 // In offsetMBB, emit code to use the reg_save_area.
10079 assert(OffsetReg != 0);
10081 // Read the reg_save_area address.
10082 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
10083 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
10088 .addOperand(Segment)
10089 .setMemRefs(MMOBegin, MMOEnd);
10091 // Zero-extend the offset
10092 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
10093 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
10096 .addImm(X86::sub_32bit);
10098 // Add the offset to the reg_save_area to get the final address.
10099 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
10100 .addReg(OffsetReg64)
10101 .addReg(RegSaveReg);
10103 // Compute the offset for the next argument
10104 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
10105 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
10107 .addImm(UseFPOffset ? 16 : 8);
10109 // Store it back into the va_list.
10110 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
10114 .addDisp(Disp, UseFPOffset ? 4 : 0)
10115 .addOperand(Segment)
10116 .addReg(NextOffsetReg)
10117 .setMemRefs(MMOBegin, MMOEnd);
10120 BuildMI(offsetMBB, DL, TII->get(X86::JMP_4))
10125 // Emit code to use overflow area
10128 // Load the overflow_area address into a register.
10129 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
10130 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
10135 .addOperand(Segment)
10136 .setMemRefs(MMOBegin, MMOEnd);
10138 // If we need to align it, do so. Otherwise, just copy the address
10139 // to OverflowDestReg.
10141 // Align the overflow address
10142 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
10143 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
10145 // aligned_addr = (addr + (align-1)) & ~(align-1)
10146 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
10147 .addReg(OverflowAddrReg)
10150 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
10152 .addImm(~(uint64_t)(Align-1));
10154 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
10155 .addReg(OverflowAddrReg);
10158 // Compute the next overflow address after this argument.
10159 // (the overflow address should be kept 8-byte aligned)
10160 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
10161 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
10162 .addReg(OverflowDestReg)
10163 .addImm(ArgSizeA8);
10165 // Store the new overflow address.
10166 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
10171 .addOperand(Segment)
10172 .addReg(NextAddrReg)
10173 .setMemRefs(MMOBegin, MMOEnd);
10175 // If we branched, emit the PHI to the front of endMBB.
10177 BuildMI(*endMBB, endMBB->begin(), DL,
10178 TII->get(X86::PHI), DestReg)
10179 .addReg(OffsetDestReg).addMBB(offsetMBB)
10180 .addReg(OverflowDestReg).addMBB(overflowMBB);
10183 // Erase the pseudo instruction
10184 MI->eraseFromParent();
10189 MachineBasicBlock *
10190 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
10192 MachineBasicBlock *MBB) const {
10193 // Emit code to save XMM registers to the stack. The ABI says that the
10194 // number of registers to save is given in %al, so it's theoretically
10195 // possible to do an indirect jump trick to avoid saving all of them,
10196 // however this code takes a simpler approach and just executes all
10197 // of the stores if %al is non-zero. It's less code, and it's probably
10198 // easier on the hardware branch predictor, and stores aren't all that
10199 // expensive anyway.
10201 // Create the new basic blocks. One block contains all the XMM stores,
10202 // and one block is the final destination regardless of whether any
10203 // stores were performed.
10204 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10205 MachineFunction *F = MBB->getParent();
10206 MachineFunction::iterator MBBIter = MBB;
10208 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
10209 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
10210 F->insert(MBBIter, XMMSaveMBB);
10211 F->insert(MBBIter, EndMBB);
10213 // Transfer the remainder of MBB and its successor edges to EndMBB.
10214 EndMBB->splice(EndMBB->begin(), MBB,
10215 llvm::next(MachineBasicBlock::iterator(MI)),
10217 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
10219 // The original block will now fall through to the XMM save block.
10220 MBB->addSuccessor(XMMSaveMBB);
10221 // The XMMSaveMBB will fall through to the end block.
10222 XMMSaveMBB->addSuccessor(EndMBB);
10224 // Now add the instructions.
10225 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10226 DebugLoc DL = MI->getDebugLoc();
10228 unsigned CountReg = MI->getOperand(0).getReg();
10229 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
10230 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
10232 if (!Subtarget->isTargetWin64()) {
10233 // If %al is 0, branch around the XMM save block.
10234 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
10235 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
10236 MBB->addSuccessor(EndMBB);
10239 // In the XMM save block, save all the XMM argument registers.
10240 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
10241 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
10242 MachineMemOperand *MMO =
10243 F->getMachineMemOperand(
10244 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
10245 MachineMemOperand::MOStore,
10246 /*Size=*/16, /*Align=*/16);
10247 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
10248 .addFrameIndex(RegSaveFrameIndex)
10249 .addImm(/*Scale=*/1)
10250 .addReg(/*IndexReg=*/0)
10251 .addImm(/*Disp=*/Offset)
10252 .addReg(/*Segment=*/0)
10253 .addReg(MI->getOperand(i).getReg())
10254 .addMemOperand(MMO);
10257 MI->eraseFromParent(); // The pseudo instruction is gone now.
10262 MachineBasicBlock *
10263 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
10264 MachineBasicBlock *BB) const {
10265 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10266 DebugLoc DL = MI->getDebugLoc();
10268 // To "insert" a SELECT_CC instruction, we actually have to insert the
10269 // diamond control-flow pattern. The incoming instruction knows the
10270 // destination vreg to set, the condition code register to branch on, the
10271 // true/false values to select between, and a branch opcode to use.
10272 const BasicBlock *LLVM_BB = BB->getBasicBlock();
10273 MachineFunction::iterator It = BB;
10279 // cmpTY ccX, r1, r2
10281 // fallthrough --> copy0MBB
10282 MachineBasicBlock *thisMBB = BB;
10283 MachineFunction *F = BB->getParent();
10284 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
10285 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
10286 F->insert(It, copy0MBB);
10287 F->insert(It, sinkMBB);
10289 // If the EFLAGS register isn't dead in the terminator, then claim that it's
10290 // live into the sink and copy blocks.
10291 const MachineFunction *MF = BB->getParent();
10292 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
10293 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
10295 for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
10296 const MachineOperand &MO = MI->getOperand(I);
10297 if (!MO.isReg() || !MO.isUse() || MO.isKill()) continue;
10298 unsigned Reg = MO.getReg();
10299 if (Reg != X86::EFLAGS) continue;
10300 copy0MBB->addLiveIn(Reg);
10301 sinkMBB->addLiveIn(Reg);
10304 // Transfer the remainder of BB and its successor edges to sinkMBB.
10305 sinkMBB->splice(sinkMBB->begin(), BB,
10306 llvm::next(MachineBasicBlock::iterator(MI)),
10308 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
10310 // Add the true and fallthrough blocks as its successors.
10311 BB->addSuccessor(copy0MBB);
10312 BB->addSuccessor(sinkMBB);
10314 // Create the conditional branch instruction.
10316 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
10317 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
10320 // %FalseValue = ...
10321 // # fallthrough to sinkMBB
10322 copy0MBB->addSuccessor(sinkMBB);
10325 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
10327 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
10328 TII->get(X86::PHI), MI->getOperand(0).getReg())
10329 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
10330 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
10332 MI->eraseFromParent(); // The pseudo instruction is gone now.
10336 MachineBasicBlock *
10337 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
10338 MachineBasicBlock *BB) const {
10339 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10340 DebugLoc DL = MI->getDebugLoc();
10342 // The lowering is pretty easy: we're just emitting the call to _alloca. The
10343 // non-trivial part is impdef of ESP.
10344 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
10347 const char *StackProbeSymbol =
10348 Subtarget->isTargetWindows() ? "_chkstk" : "_alloca";
10350 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
10351 .addExternalSymbol(StackProbeSymbol)
10352 .addReg(X86::EAX, RegState::Implicit)
10353 .addReg(X86::ESP, RegState::Implicit)
10354 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
10355 .addReg(X86::ESP, RegState::Define | RegState::Implicit)
10356 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
10358 MI->eraseFromParent(); // The pseudo instruction is gone now.
10362 MachineBasicBlock *
10363 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
10364 MachineBasicBlock *BB) const {
10365 // This is pretty easy. We're taking the value that we received from
10366 // our load from the relocation, sticking it in either RDI (x86-64)
10367 // or EAX and doing an indirect call. The return value will then
10368 // be in the normal return register.
10369 const X86InstrInfo *TII
10370 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
10371 DebugLoc DL = MI->getDebugLoc();
10372 MachineFunction *F = BB->getParent();
10374 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
10375 assert(MI->getOperand(3).isGlobal() && "This should be a global");
10377 if (Subtarget->is64Bit()) {
10378 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
10379 TII->get(X86::MOV64rm), X86::RDI)
10381 .addImm(0).addReg(0)
10382 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
10383 MI->getOperand(3).getTargetFlags())
10385 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
10386 addDirectMem(MIB, X86::RDI);
10387 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
10388 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
10389 TII->get(X86::MOV32rm), X86::EAX)
10391 .addImm(0).addReg(0)
10392 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
10393 MI->getOperand(3).getTargetFlags())
10395 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
10396 addDirectMem(MIB, X86::EAX);
10398 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
10399 TII->get(X86::MOV32rm), X86::EAX)
10400 .addReg(TII->getGlobalBaseReg(F))
10401 .addImm(0).addReg(0)
10402 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
10403 MI->getOperand(3).getTargetFlags())
10405 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
10406 addDirectMem(MIB, X86::EAX);
10409 MI->eraseFromParent(); // The pseudo instruction is gone now.
10413 MachineBasicBlock *
10414 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
10415 MachineBasicBlock *BB) const {
10416 switch (MI->getOpcode()) {
10417 default: assert(false && "Unexpected instr type to insert");
10418 case X86::TAILJMPd64:
10419 case X86::TAILJMPr64:
10420 case X86::TAILJMPm64:
10421 assert(!"TAILJMP64 would not be touched here.");
10422 case X86::TCRETURNdi64:
10423 case X86::TCRETURNri64:
10424 case X86::TCRETURNmi64:
10425 // Defs of TCRETURNxx64 has Win64's callee-saved registers, as subset.
10426 // On AMD64, additional defs should be added before register allocation.
10427 if (!Subtarget->isTargetWin64()) {
10428 MI->addRegisterDefined(X86::RSI);
10429 MI->addRegisterDefined(X86::RDI);
10430 MI->addRegisterDefined(X86::XMM6);
10431 MI->addRegisterDefined(X86::XMM7);
10432 MI->addRegisterDefined(X86::XMM8);
10433 MI->addRegisterDefined(X86::XMM9);
10434 MI->addRegisterDefined(X86::XMM10);
10435 MI->addRegisterDefined(X86::XMM11);
10436 MI->addRegisterDefined(X86::XMM12);
10437 MI->addRegisterDefined(X86::XMM13);
10438 MI->addRegisterDefined(X86::XMM14);
10439 MI->addRegisterDefined(X86::XMM15);
10442 case X86::WIN_ALLOCA:
10443 return EmitLoweredWinAlloca(MI, BB);
10444 case X86::TLSCall_32:
10445 case X86::TLSCall_64:
10446 return EmitLoweredTLSCall(MI, BB);
10447 case X86::CMOV_GR8:
10448 case X86::CMOV_FR32:
10449 case X86::CMOV_FR64:
10450 case X86::CMOV_V4F32:
10451 case X86::CMOV_V2F64:
10452 case X86::CMOV_V2I64:
10453 case X86::CMOV_GR16:
10454 case X86::CMOV_GR32:
10455 case X86::CMOV_RFP32:
10456 case X86::CMOV_RFP64:
10457 case X86::CMOV_RFP80:
10458 return EmitLoweredSelect(MI, BB);
10460 case X86::FP32_TO_INT16_IN_MEM:
10461 case X86::FP32_TO_INT32_IN_MEM:
10462 case X86::FP32_TO_INT64_IN_MEM:
10463 case X86::FP64_TO_INT16_IN_MEM:
10464 case X86::FP64_TO_INT32_IN_MEM:
10465 case X86::FP64_TO_INT64_IN_MEM:
10466 case X86::FP80_TO_INT16_IN_MEM:
10467 case X86::FP80_TO_INT32_IN_MEM:
10468 case X86::FP80_TO_INT64_IN_MEM: {
10469 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10470 DebugLoc DL = MI->getDebugLoc();
10472 // Change the floating point control register to use "round towards zero"
10473 // mode when truncating to an integer value.
10474 MachineFunction *F = BB->getParent();
10475 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
10476 addFrameReference(BuildMI(*BB, MI, DL,
10477 TII->get(X86::FNSTCW16m)), CWFrameIdx);
10479 // Load the old value of the high byte of the control word...
10481 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
10482 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
10485 // Set the high part to be round to zero...
10486 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
10489 // Reload the modified control word now...
10490 addFrameReference(BuildMI(*BB, MI, DL,
10491 TII->get(X86::FLDCW16m)), CWFrameIdx);
10493 // Restore the memory image of control word to original value
10494 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
10497 // Get the X86 opcode to use.
10499 switch (MI->getOpcode()) {
10500 default: llvm_unreachable("illegal opcode!");
10501 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
10502 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
10503 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
10504 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
10505 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
10506 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
10507 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
10508 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
10509 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
10513 MachineOperand &Op = MI->getOperand(0);
10515 AM.BaseType = X86AddressMode::RegBase;
10516 AM.Base.Reg = Op.getReg();
10518 AM.BaseType = X86AddressMode::FrameIndexBase;
10519 AM.Base.FrameIndex = Op.getIndex();
10521 Op = MI->getOperand(1);
10523 AM.Scale = Op.getImm();
10524 Op = MI->getOperand(2);
10526 AM.IndexReg = Op.getImm();
10527 Op = MI->getOperand(3);
10528 if (Op.isGlobal()) {
10529 AM.GV = Op.getGlobal();
10531 AM.Disp = Op.getImm();
10533 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
10534 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
10536 // Reload the original control word now.
10537 addFrameReference(BuildMI(*BB, MI, DL,
10538 TII->get(X86::FLDCW16m)), CWFrameIdx);
10540 MI->eraseFromParent(); // The pseudo instruction is gone now.
10543 // String/text processing lowering.
10544 case X86::PCMPISTRM128REG:
10545 case X86::VPCMPISTRM128REG:
10546 return EmitPCMP(MI, BB, 3, false /* in-mem */);
10547 case X86::PCMPISTRM128MEM:
10548 case X86::VPCMPISTRM128MEM:
10549 return EmitPCMP(MI, BB, 3, true /* in-mem */);
10550 case X86::PCMPESTRM128REG:
10551 case X86::VPCMPESTRM128REG:
10552 return EmitPCMP(MI, BB, 5, false /* in mem */);
10553 case X86::PCMPESTRM128MEM:
10554 case X86::VPCMPESTRM128MEM:
10555 return EmitPCMP(MI, BB, 5, true /* in mem */);
10557 // Thread synchronization.
10559 return EmitMonitor(MI, BB);
10561 return EmitMwait(MI, BB);
10563 // Atomic Lowering.
10564 case X86::ATOMAND32:
10565 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
10566 X86::AND32ri, X86::MOV32rm,
10568 X86::NOT32r, X86::EAX,
10569 X86::GR32RegisterClass);
10570 case X86::ATOMOR32:
10571 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
10572 X86::OR32ri, X86::MOV32rm,
10574 X86::NOT32r, X86::EAX,
10575 X86::GR32RegisterClass);
10576 case X86::ATOMXOR32:
10577 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
10578 X86::XOR32ri, X86::MOV32rm,
10580 X86::NOT32r, X86::EAX,
10581 X86::GR32RegisterClass);
10582 case X86::ATOMNAND32:
10583 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
10584 X86::AND32ri, X86::MOV32rm,
10586 X86::NOT32r, X86::EAX,
10587 X86::GR32RegisterClass, true);
10588 case X86::ATOMMIN32:
10589 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
10590 case X86::ATOMMAX32:
10591 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
10592 case X86::ATOMUMIN32:
10593 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
10594 case X86::ATOMUMAX32:
10595 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
10597 case X86::ATOMAND16:
10598 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
10599 X86::AND16ri, X86::MOV16rm,
10601 X86::NOT16r, X86::AX,
10602 X86::GR16RegisterClass);
10603 case X86::ATOMOR16:
10604 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
10605 X86::OR16ri, X86::MOV16rm,
10607 X86::NOT16r, X86::AX,
10608 X86::GR16RegisterClass);
10609 case X86::ATOMXOR16:
10610 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
10611 X86::XOR16ri, X86::MOV16rm,
10613 X86::NOT16r, X86::AX,
10614 X86::GR16RegisterClass);
10615 case X86::ATOMNAND16:
10616 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
10617 X86::AND16ri, X86::MOV16rm,
10619 X86::NOT16r, X86::AX,
10620 X86::GR16RegisterClass, true);
10621 case X86::ATOMMIN16:
10622 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
10623 case X86::ATOMMAX16:
10624 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
10625 case X86::ATOMUMIN16:
10626 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
10627 case X86::ATOMUMAX16:
10628 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
10630 case X86::ATOMAND8:
10631 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
10632 X86::AND8ri, X86::MOV8rm,
10634 X86::NOT8r, X86::AL,
10635 X86::GR8RegisterClass);
10637 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
10638 X86::OR8ri, X86::MOV8rm,
10640 X86::NOT8r, X86::AL,
10641 X86::GR8RegisterClass);
10642 case X86::ATOMXOR8:
10643 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
10644 X86::XOR8ri, X86::MOV8rm,
10646 X86::NOT8r, X86::AL,
10647 X86::GR8RegisterClass);
10648 case X86::ATOMNAND8:
10649 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
10650 X86::AND8ri, X86::MOV8rm,
10652 X86::NOT8r, X86::AL,
10653 X86::GR8RegisterClass, true);
10654 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
10655 // This group is for 64-bit host.
10656 case X86::ATOMAND64:
10657 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
10658 X86::AND64ri32, X86::MOV64rm,
10660 X86::NOT64r, X86::RAX,
10661 X86::GR64RegisterClass);
10662 case X86::ATOMOR64:
10663 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
10664 X86::OR64ri32, X86::MOV64rm,
10666 X86::NOT64r, X86::RAX,
10667 X86::GR64RegisterClass);
10668 case X86::ATOMXOR64:
10669 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
10670 X86::XOR64ri32, X86::MOV64rm,
10672 X86::NOT64r, X86::RAX,
10673 X86::GR64RegisterClass);
10674 case X86::ATOMNAND64:
10675 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
10676 X86::AND64ri32, X86::MOV64rm,
10678 X86::NOT64r, X86::RAX,
10679 X86::GR64RegisterClass, true);
10680 case X86::ATOMMIN64:
10681 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
10682 case X86::ATOMMAX64:
10683 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
10684 case X86::ATOMUMIN64:
10685 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
10686 case X86::ATOMUMAX64:
10687 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
10689 // This group does 64-bit operations on a 32-bit host.
10690 case X86::ATOMAND6432:
10691 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10692 X86::AND32rr, X86::AND32rr,
10693 X86::AND32ri, X86::AND32ri,
10695 case X86::ATOMOR6432:
10696 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10697 X86::OR32rr, X86::OR32rr,
10698 X86::OR32ri, X86::OR32ri,
10700 case X86::ATOMXOR6432:
10701 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10702 X86::XOR32rr, X86::XOR32rr,
10703 X86::XOR32ri, X86::XOR32ri,
10705 case X86::ATOMNAND6432:
10706 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10707 X86::AND32rr, X86::AND32rr,
10708 X86::AND32ri, X86::AND32ri,
10710 case X86::ATOMADD6432:
10711 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10712 X86::ADD32rr, X86::ADC32rr,
10713 X86::ADD32ri, X86::ADC32ri,
10715 case X86::ATOMSUB6432:
10716 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10717 X86::SUB32rr, X86::SBB32rr,
10718 X86::SUB32ri, X86::SBB32ri,
10720 case X86::ATOMSWAP6432:
10721 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10722 X86::MOV32rr, X86::MOV32rr,
10723 X86::MOV32ri, X86::MOV32ri,
10725 case X86::VASTART_SAVE_XMM_REGS:
10726 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
10728 case X86::VAARG_64:
10729 return EmitVAARG64WithCustomInserter(MI, BB);
10733 //===----------------------------------------------------------------------===//
10734 // X86 Optimization Hooks
10735 //===----------------------------------------------------------------------===//
10737 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
10741 const SelectionDAG &DAG,
10742 unsigned Depth) const {
10743 unsigned Opc = Op.getOpcode();
10744 assert((Opc >= ISD::BUILTIN_OP_END ||
10745 Opc == ISD::INTRINSIC_WO_CHAIN ||
10746 Opc == ISD::INTRINSIC_W_CHAIN ||
10747 Opc == ISD::INTRINSIC_VOID) &&
10748 "Should use MaskedValueIsZero if you don't know whether Op"
10749 " is a target node!");
10751 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
10765 // These nodes' second result is a boolean.
10766 if (Op.getResNo() == 0)
10769 case X86ISD::SETCC:
10770 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
10771 Mask.getBitWidth() - 1);
10776 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
10777 unsigned Depth) const {
10778 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
10779 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
10780 return Op.getValueType().getScalarType().getSizeInBits();
10786 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
10787 /// node is a GlobalAddress + offset.
10788 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
10789 const GlobalValue* &GA,
10790 int64_t &Offset) const {
10791 if (N->getOpcode() == X86ISD::Wrapper) {
10792 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
10793 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
10794 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
10798 return TargetLowering::isGAPlusOffset(N, GA, Offset);
10801 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
10802 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
10803 /// if the load addresses are consecutive, non-overlapping, and in the right
10805 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
10806 TargetLowering::DAGCombinerInfo &DCI) {
10807 DebugLoc dl = N->getDebugLoc();
10808 EVT VT = N->getValueType(0);
10810 if (VT.getSizeInBits() != 128)
10813 // Don't create instructions with illegal types after legalize types has run.
10814 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
10815 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
10818 SmallVector<SDValue, 16> Elts;
10819 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
10820 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
10822 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
10825 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
10826 /// generation and convert it from being a bunch of shuffles and extracts
10827 /// to a simple store and scalar loads to extract the elements.
10828 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
10829 const TargetLowering &TLI) {
10830 SDValue InputVector = N->getOperand(0);
10832 // Only operate on vectors of 4 elements, where the alternative shuffling
10833 // gets to be more expensive.
10834 if (InputVector.getValueType() != MVT::v4i32)
10837 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
10838 // single use which is a sign-extend or zero-extend, and all elements are
10840 SmallVector<SDNode *, 4> Uses;
10841 unsigned ExtractedElements = 0;
10842 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
10843 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
10844 if (UI.getUse().getResNo() != InputVector.getResNo())
10847 SDNode *Extract = *UI;
10848 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
10851 if (Extract->getValueType(0) != MVT::i32)
10853 if (!Extract->hasOneUse())
10855 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
10856 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
10858 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
10861 // Record which element was extracted.
10862 ExtractedElements |=
10863 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
10865 Uses.push_back(Extract);
10868 // If not all the elements were used, this may not be worthwhile.
10869 if (ExtractedElements != 15)
10872 // Ok, we've now decided to do the transformation.
10873 DebugLoc dl = InputVector.getDebugLoc();
10875 // Store the value to a temporary stack slot.
10876 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
10877 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
10878 MachinePointerInfo(), false, false, 0);
10880 // Replace each use (extract) with a load of the appropriate element.
10881 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
10882 UE = Uses.end(); UI != UE; ++UI) {
10883 SDNode *Extract = *UI;
10885 // Compute the element's address.
10886 SDValue Idx = Extract->getOperand(1);
10888 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
10889 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
10890 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
10892 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(),
10893 StackPtr, OffsetVal);
10895 // Load the scalar.
10896 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
10897 ScalarAddr, MachinePointerInfo(),
10900 // Replace the exact with the load.
10901 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
10904 // The replacement was made in place; don't return anything.
10908 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
10909 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
10910 const X86Subtarget *Subtarget) {
10911 DebugLoc DL = N->getDebugLoc();
10912 SDValue Cond = N->getOperand(0);
10913 // Get the LHS/RHS of the select.
10914 SDValue LHS = N->getOperand(1);
10915 SDValue RHS = N->getOperand(2);
10917 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
10918 // instructions match the semantics of the common C idiom x<y?x:y but not
10919 // x<=y?x:y, because of how they handle negative zero (which can be
10920 // ignored in unsafe-math mode).
10921 if (Subtarget->hasSSE2() &&
10922 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
10923 Cond.getOpcode() == ISD::SETCC) {
10924 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
10926 unsigned Opcode = 0;
10927 // Check for x CC y ? x : y.
10928 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
10929 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
10933 // Converting this to a min would handle NaNs incorrectly, and swapping
10934 // the operands would cause it to handle comparisons between positive
10935 // and negative zero incorrectly.
10936 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
10937 if (!UnsafeFPMath &&
10938 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
10940 std::swap(LHS, RHS);
10942 Opcode = X86ISD::FMIN;
10945 // Converting this to a min would handle comparisons between positive
10946 // and negative zero incorrectly.
10947 if (!UnsafeFPMath &&
10948 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
10950 Opcode = X86ISD::FMIN;
10953 // Converting this to a min would handle both negative zeros and NaNs
10954 // incorrectly, but we can swap the operands to fix both.
10955 std::swap(LHS, RHS);
10959 Opcode = X86ISD::FMIN;
10963 // Converting this to a max would handle comparisons between positive
10964 // and negative zero incorrectly.
10965 if (!UnsafeFPMath &&
10966 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
10968 Opcode = X86ISD::FMAX;
10971 // Converting this to a max would handle NaNs incorrectly, and swapping
10972 // the operands would cause it to handle comparisons between positive
10973 // and negative zero incorrectly.
10974 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
10975 if (!UnsafeFPMath &&
10976 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
10978 std::swap(LHS, RHS);
10980 Opcode = X86ISD::FMAX;
10983 // Converting this to a max would handle both negative zeros and NaNs
10984 // incorrectly, but we can swap the operands to fix both.
10985 std::swap(LHS, RHS);
10989 Opcode = X86ISD::FMAX;
10992 // Check for x CC y ? y : x -- a min/max with reversed arms.
10993 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
10994 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
10998 // Converting this to a min would handle comparisons between positive
10999 // and negative zero incorrectly, and swapping the operands would
11000 // cause it to handle NaNs incorrectly.
11001 if (!UnsafeFPMath &&
11002 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
11003 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
11005 std::swap(LHS, RHS);
11007 Opcode = X86ISD::FMIN;
11010 // Converting this to a min would handle NaNs incorrectly.
11011 if (!UnsafeFPMath &&
11012 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
11014 Opcode = X86ISD::FMIN;
11017 // Converting this to a min would handle both negative zeros and NaNs
11018 // incorrectly, but we can swap the operands to fix both.
11019 std::swap(LHS, RHS);
11023 Opcode = X86ISD::FMIN;
11027 // Converting this to a max would handle NaNs incorrectly.
11028 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
11030 Opcode = X86ISD::FMAX;
11033 // Converting this to a max would handle comparisons between positive
11034 // and negative zero incorrectly, and swapping the operands would
11035 // cause it to handle NaNs incorrectly.
11036 if (!UnsafeFPMath &&
11037 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
11038 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
11040 std::swap(LHS, RHS);
11042 Opcode = X86ISD::FMAX;
11045 // Converting this to a max would handle both negative zeros and NaNs
11046 // incorrectly, but we can swap the operands to fix both.
11047 std::swap(LHS, RHS);
11051 Opcode = X86ISD::FMAX;
11057 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
11060 // If this is a select between two integer constants, try to do some
11062 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
11063 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
11064 // Don't do this for crazy integer types.
11065 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
11066 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
11067 // so that TrueC (the true value) is larger than FalseC.
11068 bool NeedsCondInvert = false;
11070 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
11071 // Efficiently invertible.
11072 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
11073 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
11074 isa<ConstantSDNode>(Cond.getOperand(1))))) {
11075 NeedsCondInvert = true;
11076 std::swap(TrueC, FalseC);
11079 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
11080 if (FalseC->getAPIntValue() == 0 &&
11081 TrueC->getAPIntValue().isPowerOf2()) {
11082 if (NeedsCondInvert) // Invert the condition if needed.
11083 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
11084 DAG.getConstant(1, Cond.getValueType()));
11086 // Zero extend the condition if needed.
11087 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
11089 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
11090 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
11091 DAG.getConstant(ShAmt, MVT::i8));
11094 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
11095 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
11096 if (NeedsCondInvert) // Invert the condition if needed.
11097 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
11098 DAG.getConstant(1, Cond.getValueType()));
11100 // Zero extend the condition if needed.
11101 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
11102 FalseC->getValueType(0), Cond);
11103 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11104 SDValue(FalseC, 0));
11107 // Optimize cases that will turn into an LEA instruction. This requires
11108 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
11109 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
11110 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
11111 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
11113 bool isFastMultiplier = false;
11115 switch ((unsigned char)Diff) {
11117 case 1: // result = add base, cond
11118 case 2: // result = lea base( , cond*2)
11119 case 3: // result = lea base(cond, cond*2)
11120 case 4: // result = lea base( , cond*4)
11121 case 5: // result = lea base(cond, cond*4)
11122 case 8: // result = lea base( , cond*8)
11123 case 9: // result = lea base(cond, cond*8)
11124 isFastMultiplier = true;
11129 if (isFastMultiplier) {
11130 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
11131 if (NeedsCondInvert) // Invert the condition if needed.
11132 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
11133 DAG.getConstant(1, Cond.getValueType()));
11135 // Zero extend the condition if needed.
11136 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
11138 // Scale the condition by the difference.
11140 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
11141 DAG.getConstant(Diff, Cond.getValueType()));
11143 // Add the base if non-zero.
11144 if (FalseC->getAPIntValue() != 0)
11145 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11146 SDValue(FalseC, 0));
11156 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
11157 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
11158 TargetLowering::DAGCombinerInfo &DCI) {
11159 DebugLoc DL = N->getDebugLoc();
11161 // If the flag operand isn't dead, don't touch this CMOV.
11162 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
11165 // If this is a select between two integer constants, try to do some
11166 // optimizations. Note that the operands are ordered the opposite of SELECT
11168 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
11169 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
11170 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
11171 // larger than FalseC (the false value).
11172 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
11174 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
11175 CC = X86::GetOppositeBranchCondition(CC);
11176 std::swap(TrueC, FalseC);
11179 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
11180 // This is efficient for any integer data type (including i8/i16) and
11182 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
11183 SDValue Cond = N->getOperand(3);
11184 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
11185 DAG.getConstant(CC, MVT::i8), Cond);
11187 // Zero extend the condition if needed.
11188 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
11190 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
11191 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
11192 DAG.getConstant(ShAmt, MVT::i8));
11193 if (N->getNumValues() == 2) // Dead flag value?
11194 return DCI.CombineTo(N, Cond, SDValue());
11198 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
11199 // for any integer data type, including i8/i16.
11200 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
11201 SDValue Cond = N->getOperand(3);
11202 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
11203 DAG.getConstant(CC, MVT::i8), Cond);
11205 // Zero extend the condition if needed.
11206 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
11207 FalseC->getValueType(0), Cond);
11208 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11209 SDValue(FalseC, 0));
11211 if (N->getNumValues() == 2) // Dead flag value?
11212 return DCI.CombineTo(N, Cond, SDValue());
11216 // Optimize cases that will turn into an LEA instruction. This requires
11217 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
11218 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
11219 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
11220 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
11222 bool isFastMultiplier = false;
11224 switch ((unsigned char)Diff) {
11226 case 1: // result = add base, cond
11227 case 2: // result = lea base( , cond*2)
11228 case 3: // result = lea base(cond, cond*2)
11229 case 4: // result = lea base( , cond*4)
11230 case 5: // result = lea base(cond, cond*4)
11231 case 8: // result = lea base( , cond*8)
11232 case 9: // result = lea base(cond, cond*8)
11233 isFastMultiplier = true;
11238 if (isFastMultiplier) {
11239 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
11240 SDValue Cond = N->getOperand(3);
11241 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
11242 DAG.getConstant(CC, MVT::i8), Cond);
11243 // Zero extend the condition if needed.
11244 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
11246 // Scale the condition by the difference.
11248 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
11249 DAG.getConstant(Diff, Cond.getValueType()));
11251 // Add the base if non-zero.
11252 if (FalseC->getAPIntValue() != 0)
11253 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11254 SDValue(FalseC, 0));
11255 if (N->getNumValues() == 2) // Dead flag value?
11256 return DCI.CombineTo(N, Cond, SDValue());
11266 /// PerformMulCombine - Optimize a single multiply with constant into two
11267 /// in order to implement it with two cheaper instructions, e.g.
11268 /// LEA + SHL, LEA + LEA.
11269 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
11270 TargetLowering::DAGCombinerInfo &DCI) {
11271 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
11274 EVT VT = N->getValueType(0);
11275 if (VT != MVT::i64)
11278 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
11281 uint64_t MulAmt = C->getZExtValue();
11282 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
11285 uint64_t MulAmt1 = 0;
11286 uint64_t MulAmt2 = 0;
11287 if ((MulAmt % 9) == 0) {
11289 MulAmt2 = MulAmt / 9;
11290 } else if ((MulAmt % 5) == 0) {
11292 MulAmt2 = MulAmt / 5;
11293 } else if ((MulAmt % 3) == 0) {
11295 MulAmt2 = MulAmt / 3;
11298 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
11299 DebugLoc DL = N->getDebugLoc();
11301 if (isPowerOf2_64(MulAmt2) &&
11302 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
11303 // If second multiplifer is pow2, issue it first. We want the multiply by
11304 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
11306 std::swap(MulAmt1, MulAmt2);
11309 if (isPowerOf2_64(MulAmt1))
11310 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
11311 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
11313 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
11314 DAG.getConstant(MulAmt1, VT));
11316 if (isPowerOf2_64(MulAmt2))
11317 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
11318 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
11320 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
11321 DAG.getConstant(MulAmt2, VT));
11323 // Do not add new nodes to DAG combiner worklist.
11324 DCI.CombineTo(N, NewMul, false);
11329 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
11330 SDValue N0 = N->getOperand(0);
11331 SDValue N1 = N->getOperand(1);
11332 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
11333 EVT VT = N0.getValueType();
11335 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
11336 // since the result of setcc_c is all zero's or all ones.
11337 if (N1C && N0.getOpcode() == ISD::AND &&
11338 N0.getOperand(1).getOpcode() == ISD::Constant) {
11339 SDValue N00 = N0.getOperand(0);
11340 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
11341 ((N00.getOpcode() == ISD::ANY_EXTEND ||
11342 N00.getOpcode() == ISD::ZERO_EXTEND) &&
11343 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
11344 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
11345 APInt ShAmt = N1C->getAPIntValue();
11346 Mask = Mask.shl(ShAmt);
11348 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
11349 N00, DAG.getConstant(Mask, VT));
11356 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
11358 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
11359 const X86Subtarget *Subtarget) {
11360 EVT VT = N->getValueType(0);
11361 if (!VT.isVector() && VT.isInteger() &&
11362 N->getOpcode() == ISD::SHL)
11363 return PerformSHLCombine(N, DAG);
11365 // On X86 with SSE2 support, we can transform this to a vector shift if
11366 // all elements are shifted by the same amount. We can't do this in legalize
11367 // because the a constant vector is typically transformed to a constant pool
11368 // so we have no knowledge of the shift amount.
11369 if (!Subtarget->hasSSE2())
11372 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
11375 SDValue ShAmtOp = N->getOperand(1);
11376 EVT EltVT = VT.getVectorElementType();
11377 DebugLoc DL = N->getDebugLoc();
11378 SDValue BaseShAmt = SDValue();
11379 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
11380 unsigned NumElts = VT.getVectorNumElements();
11382 for (; i != NumElts; ++i) {
11383 SDValue Arg = ShAmtOp.getOperand(i);
11384 if (Arg.getOpcode() == ISD::UNDEF) continue;
11388 for (; i != NumElts; ++i) {
11389 SDValue Arg = ShAmtOp.getOperand(i);
11390 if (Arg.getOpcode() == ISD::UNDEF) continue;
11391 if (Arg != BaseShAmt) {
11395 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
11396 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
11397 SDValue InVec = ShAmtOp.getOperand(0);
11398 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
11399 unsigned NumElts = InVec.getValueType().getVectorNumElements();
11401 for (; i != NumElts; ++i) {
11402 SDValue Arg = InVec.getOperand(i);
11403 if (Arg.getOpcode() == ISD::UNDEF) continue;
11407 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
11408 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
11409 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
11410 if (C->getZExtValue() == SplatIdx)
11411 BaseShAmt = InVec.getOperand(1);
11414 if (BaseShAmt.getNode() == 0)
11415 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
11416 DAG.getIntPtrConstant(0));
11420 // The shift amount is an i32.
11421 if (EltVT.bitsGT(MVT::i32))
11422 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
11423 else if (EltVT.bitsLT(MVT::i32))
11424 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
11426 // The shift amount is identical so we can do a vector shift.
11427 SDValue ValOp = N->getOperand(0);
11428 switch (N->getOpcode()) {
11430 llvm_unreachable("Unknown shift opcode!");
11433 if (VT == MVT::v2i64)
11434 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11435 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
11437 if (VT == MVT::v4i32)
11438 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11439 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
11441 if (VT == MVT::v8i16)
11442 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11443 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
11447 if (VT == MVT::v4i32)
11448 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11449 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
11451 if (VT == MVT::v8i16)
11452 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11453 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
11457 if (VT == MVT::v2i64)
11458 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11459 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
11461 if (VT == MVT::v4i32)
11462 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11463 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
11465 if (VT == MVT::v8i16)
11466 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11467 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
11475 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
11476 TargetLowering::DAGCombinerInfo &DCI,
11477 const X86Subtarget *Subtarget) {
11478 if (DCI.isBeforeLegalizeOps())
11481 // Want to form PANDN nodes, in the hopes of then easily combining them with
11482 // OR and AND nodes to form PBLEND/PSIGN.
11483 EVT VT = N->getValueType(0);
11484 if (VT != MVT::v2i64)
11487 SDValue N0 = N->getOperand(0);
11488 SDValue N1 = N->getOperand(1);
11489 DebugLoc DL = N->getDebugLoc();
11491 // Check LHS for vnot
11492 if (N0.getOpcode() == ISD::XOR &&
11493 ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
11494 return DAG.getNode(X86ISD::PANDN, DL, VT, N0.getOperand(0), N1);
11496 // Check RHS for vnot
11497 if (N1.getOpcode() == ISD::XOR &&
11498 ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
11499 return DAG.getNode(X86ISD::PANDN, DL, VT, N1.getOperand(0), N0);
11504 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
11505 TargetLowering::DAGCombinerInfo &DCI,
11506 const X86Subtarget *Subtarget) {
11507 if (DCI.isBeforeLegalizeOps())
11510 EVT VT = N->getValueType(0);
11511 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64 && VT != MVT::v2i64)
11514 SDValue N0 = N->getOperand(0);
11515 SDValue N1 = N->getOperand(1);
11517 // look for psign/blend
11518 if (Subtarget->hasSSSE3()) {
11519 if (VT == MVT::v2i64) {
11520 // Canonicalize pandn to RHS
11521 if (N0.getOpcode() == X86ISD::PANDN)
11523 // or (and (m, x), (pandn m, y))
11524 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::PANDN) {
11525 SDValue Mask = N1.getOperand(0);
11526 SDValue X = N1.getOperand(1);
11528 if (N0.getOperand(0) == Mask)
11529 Y = N0.getOperand(1);
11530 if (N0.getOperand(1) == Mask)
11531 Y = N0.getOperand(0);
11533 // Check to see if the mask appeared in both the AND and PANDN and
11537 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
11538 if (Mask.getOpcode() != ISD::BITCAST ||
11539 X.getOpcode() != ISD::BITCAST ||
11540 Y.getOpcode() != ISD::BITCAST)
11543 // Look through mask bitcast.
11544 Mask = Mask.getOperand(0);
11545 EVT MaskVT = Mask.getValueType();
11547 // Validate that the Mask operand is a vector sra node. The sra node
11548 // will be an intrinsic.
11549 if (Mask.getOpcode() != ISD::INTRINSIC_WO_CHAIN)
11552 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
11553 // there is no psrai.b
11554 switch (cast<ConstantSDNode>(Mask.getOperand(0))->getZExtValue()) {
11555 case Intrinsic::x86_sse2_psrai_w:
11556 case Intrinsic::x86_sse2_psrai_d:
11558 default: return SDValue();
11561 // Check that the SRA is all signbits.
11562 SDValue SraC = Mask.getOperand(2);
11563 unsigned SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
11564 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
11565 if ((SraAmt + 1) != EltBits)
11568 DebugLoc DL = N->getDebugLoc();
11570 // Now we know we at least have a plendvb with the mask val. See if
11571 // we can form a psignb/w/d.
11572 // psign = x.type == y.type == mask.type && y = sub(0, x);
11573 X = X.getOperand(0);
11574 Y = Y.getOperand(0);
11575 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
11576 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
11577 X.getValueType() == MaskVT && X.getValueType() == Y.getValueType()){
11580 case 8: Opc = X86ISD::PSIGNB; break;
11581 case 16: Opc = X86ISD::PSIGNW; break;
11582 case 32: Opc = X86ISD::PSIGND; break;
11586 SDValue Sign = DAG.getNode(Opc, DL, MaskVT, X, Mask.getOperand(1));
11587 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Sign);
11590 // PBLENDVB only available on SSE 4.1
11591 if (!Subtarget->hasSSE41())
11594 X = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, X);
11595 Y = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Y);
11596 Mask = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Mask);
11597 Mask = DAG.getNode(X86ISD::PBLENDVB, DL, MVT::v16i8, X, Y, Mask);
11598 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Mask);
11603 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
11604 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
11606 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
11608 if (!N0.hasOneUse() || !N1.hasOneUse())
11611 SDValue ShAmt0 = N0.getOperand(1);
11612 if (ShAmt0.getValueType() != MVT::i8)
11614 SDValue ShAmt1 = N1.getOperand(1);
11615 if (ShAmt1.getValueType() != MVT::i8)
11617 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
11618 ShAmt0 = ShAmt0.getOperand(0);
11619 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
11620 ShAmt1 = ShAmt1.getOperand(0);
11622 DebugLoc DL = N->getDebugLoc();
11623 unsigned Opc = X86ISD::SHLD;
11624 SDValue Op0 = N0.getOperand(0);
11625 SDValue Op1 = N1.getOperand(0);
11626 if (ShAmt0.getOpcode() == ISD::SUB) {
11627 Opc = X86ISD::SHRD;
11628 std::swap(Op0, Op1);
11629 std::swap(ShAmt0, ShAmt1);
11632 unsigned Bits = VT.getSizeInBits();
11633 if (ShAmt1.getOpcode() == ISD::SUB) {
11634 SDValue Sum = ShAmt1.getOperand(0);
11635 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
11636 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
11637 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
11638 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
11639 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
11640 return DAG.getNode(Opc, DL, VT,
11642 DAG.getNode(ISD::TRUNCATE, DL,
11645 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
11646 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
11648 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
11649 return DAG.getNode(Opc, DL, VT,
11650 N0.getOperand(0), N1.getOperand(0),
11651 DAG.getNode(ISD::TRUNCATE, DL,
11658 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
11659 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
11660 const X86Subtarget *Subtarget) {
11661 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
11662 // the FP state in cases where an emms may be missing.
11663 // A preferable solution to the general problem is to figure out the right
11664 // places to insert EMMS. This qualifies as a quick hack.
11666 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
11667 StoreSDNode *St = cast<StoreSDNode>(N);
11668 EVT VT = St->getValue().getValueType();
11669 if (VT.getSizeInBits() != 64)
11672 const Function *F = DAG.getMachineFunction().getFunction();
11673 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
11674 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
11675 && Subtarget->hasSSE2();
11676 if ((VT.isVector() ||
11677 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
11678 isa<LoadSDNode>(St->getValue()) &&
11679 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
11680 St->getChain().hasOneUse() && !St->isVolatile()) {
11681 SDNode* LdVal = St->getValue().getNode();
11682 LoadSDNode *Ld = 0;
11683 int TokenFactorIndex = -1;
11684 SmallVector<SDValue, 8> Ops;
11685 SDNode* ChainVal = St->getChain().getNode();
11686 // Must be a store of a load. We currently handle two cases: the load
11687 // is a direct child, and it's under an intervening TokenFactor. It is
11688 // possible to dig deeper under nested TokenFactors.
11689 if (ChainVal == LdVal)
11690 Ld = cast<LoadSDNode>(St->getChain());
11691 else if (St->getValue().hasOneUse() &&
11692 ChainVal->getOpcode() == ISD::TokenFactor) {
11693 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
11694 if (ChainVal->getOperand(i).getNode() == LdVal) {
11695 TokenFactorIndex = i;
11696 Ld = cast<LoadSDNode>(St->getValue());
11698 Ops.push_back(ChainVal->getOperand(i));
11702 if (!Ld || !ISD::isNormalLoad(Ld))
11705 // If this is not the MMX case, i.e. we are just turning i64 load/store
11706 // into f64 load/store, avoid the transformation if there are multiple
11707 // uses of the loaded value.
11708 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
11711 DebugLoc LdDL = Ld->getDebugLoc();
11712 DebugLoc StDL = N->getDebugLoc();
11713 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
11714 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
11716 if (Subtarget->is64Bit() || F64IsLegal) {
11717 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
11718 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
11719 Ld->getPointerInfo(), Ld->isVolatile(),
11720 Ld->isNonTemporal(), Ld->getAlignment());
11721 SDValue NewChain = NewLd.getValue(1);
11722 if (TokenFactorIndex != -1) {
11723 Ops.push_back(NewChain);
11724 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
11727 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
11728 St->getPointerInfo(),
11729 St->isVolatile(), St->isNonTemporal(),
11730 St->getAlignment());
11733 // Otherwise, lower to two pairs of 32-bit loads / stores.
11734 SDValue LoAddr = Ld->getBasePtr();
11735 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
11736 DAG.getConstant(4, MVT::i32));
11738 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
11739 Ld->getPointerInfo(),
11740 Ld->isVolatile(), Ld->isNonTemporal(),
11741 Ld->getAlignment());
11742 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
11743 Ld->getPointerInfo().getWithOffset(4),
11744 Ld->isVolatile(), Ld->isNonTemporal(),
11745 MinAlign(Ld->getAlignment(), 4));
11747 SDValue NewChain = LoLd.getValue(1);
11748 if (TokenFactorIndex != -1) {
11749 Ops.push_back(LoLd);
11750 Ops.push_back(HiLd);
11751 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
11755 LoAddr = St->getBasePtr();
11756 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
11757 DAG.getConstant(4, MVT::i32));
11759 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
11760 St->getPointerInfo(),
11761 St->isVolatile(), St->isNonTemporal(),
11762 St->getAlignment());
11763 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
11764 St->getPointerInfo().getWithOffset(4),
11766 St->isNonTemporal(),
11767 MinAlign(St->getAlignment(), 4));
11768 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
11773 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
11774 /// X86ISD::FXOR nodes.
11775 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
11776 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
11777 // F[X]OR(0.0, x) -> x
11778 // F[X]OR(x, 0.0) -> x
11779 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
11780 if (C->getValueAPF().isPosZero())
11781 return N->getOperand(1);
11782 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
11783 if (C->getValueAPF().isPosZero())
11784 return N->getOperand(0);
11788 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
11789 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
11790 // FAND(0.0, x) -> 0.0
11791 // FAND(x, 0.0) -> 0.0
11792 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
11793 if (C->getValueAPF().isPosZero())
11794 return N->getOperand(0);
11795 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
11796 if (C->getValueAPF().isPosZero())
11797 return N->getOperand(1);
11801 static SDValue PerformBTCombine(SDNode *N,
11803 TargetLowering::DAGCombinerInfo &DCI) {
11804 // BT ignores high bits in the bit index operand.
11805 SDValue Op1 = N->getOperand(1);
11806 if (Op1.hasOneUse()) {
11807 unsigned BitWidth = Op1.getValueSizeInBits();
11808 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
11809 APInt KnownZero, KnownOne;
11810 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
11811 !DCI.isBeforeLegalizeOps());
11812 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
11813 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
11814 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
11815 DCI.CommitTargetLoweringOpt(TLO);
11820 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
11821 SDValue Op = N->getOperand(0);
11822 if (Op.getOpcode() == ISD::BITCAST)
11823 Op = Op.getOperand(0);
11824 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
11825 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
11826 VT.getVectorElementType().getSizeInBits() ==
11827 OpVT.getVectorElementType().getSizeInBits()) {
11828 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
11833 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
11834 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
11835 // (and (i32 x86isd::setcc_carry), 1)
11836 // This eliminates the zext. This transformation is necessary because
11837 // ISD::SETCC is always legalized to i8.
11838 DebugLoc dl = N->getDebugLoc();
11839 SDValue N0 = N->getOperand(0);
11840 EVT VT = N->getValueType(0);
11841 if (N0.getOpcode() == ISD::AND &&
11843 N0.getOperand(0).hasOneUse()) {
11844 SDValue N00 = N0.getOperand(0);
11845 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
11847 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
11848 if (!C || C->getZExtValue() != 1)
11850 return DAG.getNode(ISD::AND, dl, VT,
11851 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
11852 N00.getOperand(0), N00.getOperand(1)),
11853 DAG.getConstant(1, VT));
11859 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
11860 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG) {
11861 unsigned X86CC = N->getConstantOperandVal(0);
11862 SDValue EFLAG = N->getOperand(1);
11863 DebugLoc DL = N->getDebugLoc();
11865 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
11866 // a zext and produces an all-ones bit which is more useful than 0/1 in some
11868 if (X86CC == X86::COND_B)
11869 return DAG.getNode(ISD::AND, DL, MVT::i8,
11870 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
11871 DAG.getConstant(X86CC, MVT::i8), EFLAG),
11872 DAG.getConstant(1, MVT::i8));
11877 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
11878 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
11879 X86TargetLowering::DAGCombinerInfo &DCI) {
11880 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
11881 // the result is either zero or one (depending on the input carry bit).
11882 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
11883 if (X86::isZeroNode(N->getOperand(0)) &&
11884 X86::isZeroNode(N->getOperand(1)) &&
11885 // We don't have a good way to replace an EFLAGS use, so only do this when
11887 SDValue(N, 1).use_empty()) {
11888 DebugLoc DL = N->getDebugLoc();
11889 EVT VT = N->getValueType(0);
11890 SDValue CarryOut = DAG.getConstant(0, N->getValueType(1));
11891 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
11892 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
11893 DAG.getConstant(X86::COND_B,MVT::i8),
11895 DAG.getConstant(1, VT));
11896 return DCI.CombineTo(N, Res1, CarryOut);
11902 // fold (add Y, (sete X, 0)) -> adc 0, Y
11903 // (add Y, (setne X, 0)) -> sbb -1, Y
11904 // (sub (sete X, 0), Y) -> sbb 0, Y
11905 // (sub (setne X, 0), Y) -> adc -1, Y
11906 static SDValue OptimizeConditonalInDecrement(SDNode *N, SelectionDAG &DAG) {
11907 DebugLoc DL = N->getDebugLoc();
11909 // Look through ZExts.
11910 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
11911 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
11914 SDValue SetCC = Ext.getOperand(0);
11915 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
11918 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
11919 if (CC != X86::COND_E && CC != X86::COND_NE)
11922 SDValue Cmp = SetCC.getOperand(1);
11923 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
11924 !X86::isZeroNode(Cmp.getOperand(1)) ||
11925 !Cmp.getOperand(0).getValueType().isInteger())
11928 SDValue CmpOp0 = Cmp.getOperand(0);
11929 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
11930 DAG.getConstant(1, CmpOp0.getValueType()));
11932 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
11933 if (CC == X86::COND_NE)
11934 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
11935 DL, OtherVal.getValueType(), OtherVal,
11936 DAG.getConstant(-1ULL, OtherVal.getValueType()), NewCmp);
11937 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
11938 DL, OtherVal.getValueType(), OtherVal,
11939 DAG.getConstant(0, OtherVal.getValueType()), NewCmp);
11942 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
11943 DAGCombinerInfo &DCI) const {
11944 SelectionDAG &DAG = DCI.DAG;
11945 switch (N->getOpcode()) {
11947 case ISD::EXTRACT_VECTOR_ELT:
11948 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
11949 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
11950 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
11952 case ISD::SUB: return OptimizeConditonalInDecrement(N, DAG);
11953 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
11954 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
11957 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
11958 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
11959 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
11960 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
11962 case X86ISD::FOR: return PerformFORCombine(N, DAG);
11963 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
11964 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
11965 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
11966 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
11967 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG);
11968 case X86ISD::SHUFPS: // Handle all target specific shuffles
11969 case X86ISD::SHUFPD:
11970 case X86ISD::PALIGN:
11971 case X86ISD::PUNPCKHBW:
11972 case X86ISD::PUNPCKHWD:
11973 case X86ISD::PUNPCKHDQ:
11974 case X86ISD::PUNPCKHQDQ:
11975 case X86ISD::UNPCKHPS:
11976 case X86ISD::UNPCKHPD:
11977 case X86ISD::PUNPCKLBW:
11978 case X86ISD::PUNPCKLWD:
11979 case X86ISD::PUNPCKLDQ:
11980 case X86ISD::PUNPCKLQDQ:
11981 case X86ISD::UNPCKLPS:
11982 case X86ISD::UNPCKLPD:
11983 case X86ISD::MOVHLPS:
11984 case X86ISD::MOVLHPS:
11985 case X86ISD::PSHUFD:
11986 case X86ISD::PSHUFHW:
11987 case X86ISD::PSHUFLW:
11988 case X86ISD::MOVSS:
11989 case X86ISD::MOVSD:
11990 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI);
11996 /// isTypeDesirableForOp - Return true if the target has native support for
11997 /// the specified value type and it is 'desirable' to use the type for the
11998 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
11999 /// instruction encodings are longer and some i16 instructions are slow.
12000 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
12001 if (!isTypeLegal(VT))
12003 if (VT != MVT::i16)
12010 case ISD::SIGN_EXTEND:
12011 case ISD::ZERO_EXTEND:
12012 case ISD::ANY_EXTEND:
12025 /// IsDesirableToPromoteOp - This method query the target whether it is
12026 /// beneficial for dag combiner to promote the specified node. If true, it
12027 /// should return the desired promotion type by reference.
12028 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
12029 EVT VT = Op.getValueType();
12030 if (VT != MVT::i16)
12033 bool Promote = false;
12034 bool Commute = false;
12035 switch (Op.getOpcode()) {
12038 LoadSDNode *LD = cast<LoadSDNode>(Op);
12039 // If the non-extending load has a single use and it's not live out, then it
12040 // might be folded.
12041 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
12042 Op.hasOneUse()*/) {
12043 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
12044 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
12045 // The only case where we'd want to promote LOAD (rather then it being
12046 // promoted as an operand is when it's only use is liveout.
12047 if (UI->getOpcode() != ISD::CopyToReg)
12054 case ISD::SIGN_EXTEND:
12055 case ISD::ZERO_EXTEND:
12056 case ISD::ANY_EXTEND:
12061 SDValue N0 = Op.getOperand(0);
12062 // Look out for (store (shl (load), x)).
12063 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
12076 SDValue N0 = Op.getOperand(0);
12077 SDValue N1 = Op.getOperand(1);
12078 if (!Commute && MayFoldLoad(N1))
12080 // Avoid disabling potential load folding opportunities.
12081 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
12083 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
12093 //===----------------------------------------------------------------------===//
12094 // X86 Inline Assembly Support
12095 //===----------------------------------------------------------------------===//
12097 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
12098 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
12100 std::string AsmStr = IA->getAsmString();
12102 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
12103 SmallVector<StringRef, 4> AsmPieces;
12104 SplitString(AsmStr, AsmPieces, ";\n");
12106 switch (AsmPieces.size()) {
12107 default: return false;
12109 AsmStr = AsmPieces[0];
12111 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
12113 // FIXME: this should verify that we are targetting a 486 or better. If not,
12114 // we will turn this bswap into something that will be lowered to logical ops
12115 // instead of emitting the bswap asm. For now, we don't support 486 or lower
12116 // so don't worry about this.
12118 if (AsmPieces.size() == 2 &&
12119 (AsmPieces[0] == "bswap" ||
12120 AsmPieces[0] == "bswapq" ||
12121 AsmPieces[0] == "bswapl") &&
12122 (AsmPieces[1] == "$0" ||
12123 AsmPieces[1] == "${0:q}")) {
12124 // No need to check constraints, nothing other than the equivalent of
12125 // "=r,0" would be valid here.
12126 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
12127 if (!Ty || Ty->getBitWidth() % 16 != 0)
12129 return IntrinsicLowering::LowerToByteSwap(CI);
12131 // rorw $$8, ${0:w} --> llvm.bswap.i16
12132 if (CI->getType()->isIntegerTy(16) &&
12133 AsmPieces.size() == 3 &&
12134 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
12135 AsmPieces[1] == "$$8," &&
12136 AsmPieces[2] == "${0:w}" &&
12137 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
12139 const std::string &ConstraintsStr = IA->getConstraintString();
12140 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
12141 std::sort(AsmPieces.begin(), AsmPieces.end());
12142 if (AsmPieces.size() == 4 &&
12143 AsmPieces[0] == "~{cc}" &&
12144 AsmPieces[1] == "~{dirflag}" &&
12145 AsmPieces[2] == "~{flags}" &&
12146 AsmPieces[3] == "~{fpsr}") {
12147 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
12148 if (!Ty || Ty->getBitWidth() % 16 != 0)
12150 return IntrinsicLowering::LowerToByteSwap(CI);
12155 if (CI->getType()->isIntegerTy(32) &&
12156 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
12157 SmallVector<StringRef, 4> Words;
12158 SplitString(AsmPieces[0], Words, " \t,");
12159 if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
12160 Words[2] == "${0:w}") {
12162 SplitString(AsmPieces[1], Words, " \t,");
12163 if (Words.size() == 3 && Words[0] == "rorl" && Words[1] == "$$16" &&
12164 Words[2] == "$0") {
12166 SplitString(AsmPieces[2], Words, " \t,");
12167 if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
12168 Words[2] == "${0:w}") {
12170 const std::string &ConstraintsStr = IA->getConstraintString();
12171 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
12172 std::sort(AsmPieces.begin(), AsmPieces.end());
12173 if (AsmPieces.size() == 4 &&
12174 AsmPieces[0] == "~{cc}" &&
12175 AsmPieces[1] == "~{dirflag}" &&
12176 AsmPieces[2] == "~{flags}" &&
12177 AsmPieces[3] == "~{fpsr}") {
12178 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
12179 if (!Ty || Ty->getBitWidth() % 16 != 0)
12181 return IntrinsicLowering::LowerToByteSwap(CI);
12188 if (CI->getType()->isIntegerTy(64)) {
12189 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
12190 if (Constraints.size() >= 2 &&
12191 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
12192 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
12193 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
12194 SmallVector<StringRef, 4> Words;
12195 SplitString(AsmPieces[0], Words, " \t");
12196 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
12198 SplitString(AsmPieces[1], Words, " \t");
12199 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
12201 SplitString(AsmPieces[2], Words, " \t,");
12202 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
12203 Words[2] == "%edx") {
12204 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
12205 if (!Ty || Ty->getBitWidth() % 16 != 0)
12207 return IntrinsicLowering::LowerToByteSwap(CI);
12220 /// getConstraintType - Given a constraint letter, return the type of
12221 /// constraint it is for this target.
12222 X86TargetLowering::ConstraintType
12223 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
12224 if (Constraint.size() == 1) {
12225 switch (Constraint[0]) {
12235 return C_RegisterClass;
12259 return TargetLowering::getConstraintType(Constraint);
12262 /// Examine constraint type and operand type and determine a weight value.
12263 /// This object must already have been set up with the operand type
12264 /// and the current alternative constraint selected.
12265 TargetLowering::ConstraintWeight
12266 X86TargetLowering::getSingleConstraintMatchWeight(
12267 AsmOperandInfo &info, const char *constraint) const {
12268 ConstraintWeight weight = CW_Invalid;
12269 Value *CallOperandVal = info.CallOperandVal;
12270 // If we don't have a value, we can't do a match,
12271 // but allow it at the lowest weight.
12272 if (CallOperandVal == NULL)
12274 const Type *type = CallOperandVal->getType();
12275 // Look at the constraint type.
12276 switch (*constraint) {
12278 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
12289 if (CallOperandVal->getType()->isIntegerTy())
12290 weight = CW_SpecificReg;
12295 if (type->isFloatingPointTy())
12296 weight = CW_SpecificReg;
12299 if (type->isX86_MMXTy() && Subtarget->hasMMX())
12300 weight = CW_SpecificReg;
12304 if ((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasXMM())
12305 weight = CW_Register;
12308 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
12309 if (C->getZExtValue() <= 31)
12310 weight = CW_Constant;
12314 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12315 if (C->getZExtValue() <= 63)
12316 weight = CW_Constant;
12320 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12321 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
12322 weight = CW_Constant;
12326 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12327 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
12328 weight = CW_Constant;
12332 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12333 if (C->getZExtValue() <= 3)
12334 weight = CW_Constant;
12338 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12339 if (C->getZExtValue() <= 0xff)
12340 weight = CW_Constant;
12345 if (dyn_cast<ConstantFP>(CallOperandVal)) {
12346 weight = CW_Constant;
12350 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12351 if ((C->getSExtValue() >= -0x80000000LL) &&
12352 (C->getSExtValue() <= 0x7fffffffLL))
12353 weight = CW_Constant;
12357 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12358 if (C->getZExtValue() <= 0xffffffff)
12359 weight = CW_Constant;
12366 /// LowerXConstraint - try to replace an X constraint, which matches anything,
12367 /// with another that has more specific requirements based on the type of the
12368 /// corresponding operand.
12369 const char *X86TargetLowering::
12370 LowerXConstraint(EVT ConstraintVT) const {
12371 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
12372 // 'f' like normal targets.
12373 if (ConstraintVT.isFloatingPoint()) {
12374 if (Subtarget->hasXMMInt())
12376 if (Subtarget->hasXMM())
12380 return TargetLowering::LowerXConstraint(ConstraintVT);
12383 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
12384 /// vector. If it is invalid, don't add anything to Ops.
12385 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
12387 std::vector<SDValue>&Ops,
12388 SelectionDAG &DAG) const {
12389 SDValue Result(0, 0);
12391 switch (Constraint) {
12394 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12395 if (C->getZExtValue() <= 31) {
12396 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
12402 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12403 if (C->getZExtValue() <= 63) {
12404 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
12410 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12411 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
12412 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
12418 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12419 if (C->getZExtValue() <= 255) {
12420 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
12426 // 32-bit signed value
12427 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12428 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
12429 C->getSExtValue())) {
12430 // Widen to 64 bits here to get it sign extended.
12431 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
12434 // FIXME gcc accepts some relocatable values here too, but only in certain
12435 // memory models; it's complicated.
12440 // 32-bit unsigned value
12441 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12442 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
12443 C->getZExtValue())) {
12444 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
12448 // FIXME gcc accepts some relocatable values here too, but only in certain
12449 // memory models; it's complicated.
12453 // Literal immediates are always ok.
12454 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
12455 // Widen to 64 bits here to get it sign extended.
12456 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
12460 // In any sort of PIC mode addresses need to be computed at runtime by
12461 // adding in a register or some sort of table lookup. These can't
12462 // be used as immediates.
12463 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
12466 // If we are in non-pic codegen mode, we allow the address of a global (with
12467 // an optional displacement) to be used with 'i'.
12468 GlobalAddressSDNode *GA = 0;
12469 int64_t Offset = 0;
12471 // Match either (GA), (GA+C), (GA+C1+C2), etc.
12473 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
12474 Offset += GA->getOffset();
12476 } else if (Op.getOpcode() == ISD::ADD) {
12477 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
12478 Offset += C->getZExtValue();
12479 Op = Op.getOperand(0);
12482 } else if (Op.getOpcode() == ISD::SUB) {
12483 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
12484 Offset += -C->getZExtValue();
12485 Op = Op.getOperand(0);
12490 // Otherwise, this isn't something we can handle, reject it.
12494 const GlobalValue *GV = GA->getGlobal();
12495 // If we require an extra load to get this address, as in PIC mode, we
12496 // can't accept it.
12497 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
12498 getTargetMachine())))
12501 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
12502 GA->getValueType(0), Offset);
12507 if (Result.getNode()) {
12508 Ops.push_back(Result);
12511 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
12514 std::vector<unsigned> X86TargetLowering::
12515 getRegClassForInlineAsmConstraint(const std::string &Constraint,
12517 if (Constraint.size() == 1) {
12518 // FIXME: not handling fp-stack yet!
12519 switch (Constraint[0]) { // GCC X86 Constraint Letters
12520 default: break; // Unknown constraint letter
12521 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
12522 if (Subtarget->is64Bit()) {
12523 if (VT == MVT::i32)
12524 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
12525 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
12526 X86::R10D,X86::R11D,X86::R12D,
12527 X86::R13D,X86::R14D,X86::R15D,
12528 X86::EBP, X86::ESP, 0);
12529 else if (VT == MVT::i16)
12530 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
12531 X86::SI, X86::DI, X86::R8W,X86::R9W,
12532 X86::R10W,X86::R11W,X86::R12W,
12533 X86::R13W,X86::R14W,X86::R15W,
12534 X86::BP, X86::SP, 0);
12535 else if (VT == MVT::i8)
12536 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
12537 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
12538 X86::R10B,X86::R11B,X86::R12B,
12539 X86::R13B,X86::R14B,X86::R15B,
12540 X86::BPL, X86::SPL, 0);
12542 else if (VT == MVT::i64)
12543 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
12544 X86::RSI, X86::RDI, X86::R8, X86::R9,
12545 X86::R10, X86::R11, X86::R12,
12546 X86::R13, X86::R14, X86::R15,
12547 X86::RBP, X86::RSP, 0);
12551 // 32-bit fallthrough
12552 case 'Q': // Q_REGS
12553 if (VT == MVT::i32)
12554 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
12555 else if (VT == MVT::i16)
12556 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
12557 else if (VT == MVT::i8)
12558 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
12559 else if (VT == MVT::i64)
12560 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
12565 return std::vector<unsigned>();
12568 std::pair<unsigned, const TargetRegisterClass*>
12569 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
12571 // First, see if this is a constraint that directly corresponds to an LLVM
12573 if (Constraint.size() == 1) {
12574 // GCC Constraint Letters
12575 switch (Constraint[0]) {
12577 case 'r': // GENERAL_REGS
12578 case 'l': // INDEX_REGS
12580 return std::make_pair(0U, X86::GR8RegisterClass);
12581 if (VT == MVT::i16)
12582 return std::make_pair(0U, X86::GR16RegisterClass);
12583 if (VT == MVT::i32 || !Subtarget->is64Bit())
12584 return std::make_pair(0U, X86::GR32RegisterClass);
12585 return std::make_pair(0U, X86::GR64RegisterClass);
12586 case 'R': // LEGACY_REGS
12588 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
12589 if (VT == MVT::i16)
12590 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
12591 if (VT == MVT::i32 || !Subtarget->is64Bit())
12592 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
12593 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
12594 case 'f': // FP Stack registers.
12595 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
12596 // value to the correct fpstack register class.
12597 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
12598 return std::make_pair(0U, X86::RFP32RegisterClass);
12599 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
12600 return std::make_pair(0U, X86::RFP64RegisterClass);
12601 return std::make_pair(0U, X86::RFP80RegisterClass);
12602 case 'y': // MMX_REGS if MMX allowed.
12603 if (!Subtarget->hasMMX()) break;
12604 return std::make_pair(0U, X86::VR64RegisterClass);
12605 case 'Y': // SSE_REGS if SSE2 allowed
12606 if (!Subtarget->hasXMMInt()) break;
12608 case 'x': // SSE_REGS if SSE1 allowed
12609 if (!Subtarget->hasXMM()) break;
12611 switch (VT.getSimpleVT().SimpleTy) {
12613 // Scalar SSE types.
12616 return std::make_pair(0U, X86::FR32RegisterClass);
12619 return std::make_pair(0U, X86::FR64RegisterClass);
12627 return std::make_pair(0U, X86::VR128RegisterClass);
12633 // Use the default implementation in TargetLowering to convert the register
12634 // constraint into a member of a register class.
12635 std::pair<unsigned, const TargetRegisterClass*> Res;
12636 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
12638 // Not found as a standard register?
12639 if (Res.second == 0) {
12640 // Map st(0) -> st(7) -> ST0
12641 if (Constraint.size() == 7 && Constraint[0] == '{' &&
12642 tolower(Constraint[1]) == 's' &&
12643 tolower(Constraint[2]) == 't' &&
12644 Constraint[3] == '(' &&
12645 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
12646 Constraint[5] == ')' &&
12647 Constraint[6] == '}') {
12649 Res.first = X86::ST0+Constraint[4]-'0';
12650 Res.second = X86::RFP80RegisterClass;
12654 // GCC allows "st(0)" to be called just plain "st".
12655 if (StringRef("{st}").equals_lower(Constraint)) {
12656 Res.first = X86::ST0;
12657 Res.second = X86::RFP80RegisterClass;
12662 if (StringRef("{flags}").equals_lower(Constraint)) {
12663 Res.first = X86::EFLAGS;
12664 Res.second = X86::CCRRegisterClass;
12668 // 'A' means EAX + EDX.
12669 if (Constraint == "A") {
12670 Res.first = X86::EAX;
12671 Res.second = X86::GR32_ADRegisterClass;
12677 // Otherwise, check to see if this is a register class of the wrong value
12678 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
12679 // turn into {ax},{dx}.
12680 if (Res.second->hasType(VT))
12681 return Res; // Correct type already, nothing to do.
12683 // All of the single-register GCC register classes map their values onto
12684 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
12685 // really want an 8-bit or 32-bit register, map to the appropriate register
12686 // class and return the appropriate register.
12687 if (Res.second == X86::GR16RegisterClass) {
12688 if (VT == MVT::i8) {
12689 unsigned DestReg = 0;
12690 switch (Res.first) {
12692 case X86::AX: DestReg = X86::AL; break;
12693 case X86::DX: DestReg = X86::DL; break;
12694 case X86::CX: DestReg = X86::CL; break;
12695 case X86::BX: DestReg = X86::BL; break;
12698 Res.first = DestReg;
12699 Res.second = X86::GR8RegisterClass;
12701 } else if (VT == MVT::i32) {
12702 unsigned DestReg = 0;
12703 switch (Res.first) {
12705 case X86::AX: DestReg = X86::EAX; break;
12706 case X86::DX: DestReg = X86::EDX; break;
12707 case X86::CX: DestReg = X86::ECX; break;
12708 case X86::BX: DestReg = X86::EBX; break;
12709 case X86::SI: DestReg = X86::ESI; break;
12710 case X86::DI: DestReg = X86::EDI; break;
12711 case X86::BP: DestReg = X86::EBP; break;
12712 case X86::SP: DestReg = X86::ESP; break;
12715 Res.first = DestReg;
12716 Res.second = X86::GR32RegisterClass;
12718 } else if (VT == MVT::i64) {
12719 unsigned DestReg = 0;
12720 switch (Res.first) {
12722 case X86::AX: DestReg = X86::RAX; break;
12723 case X86::DX: DestReg = X86::RDX; break;
12724 case X86::CX: DestReg = X86::RCX; break;
12725 case X86::BX: DestReg = X86::RBX; break;
12726 case X86::SI: DestReg = X86::RSI; break;
12727 case X86::DI: DestReg = X86::RDI; break;
12728 case X86::BP: DestReg = X86::RBP; break;
12729 case X86::SP: DestReg = X86::RSP; break;
12732 Res.first = DestReg;
12733 Res.second = X86::GR64RegisterClass;
12736 } else if (Res.second == X86::FR32RegisterClass ||
12737 Res.second == X86::FR64RegisterClass ||
12738 Res.second == X86::VR128RegisterClass) {
12739 // Handle references to XMM physical registers that got mapped into the
12740 // wrong class. This can happen with constraints like {xmm0} where the
12741 // target independent register mapper will just pick the first match it can
12742 // find, ignoring the required type.
12743 if (VT == MVT::f32)
12744 Res.second = X86::FR32RegisterClass;
12745 else if (VT == MVT::f64)
12746 Res.second = X86::FR64RegisterClass;
12747 else if (X86::VR128RegisterClass->hasType(VT))
12748 Res.second = X86::VR128RegisterClass;