1 //===-- X86FastISel.cpp - X86 FastISel 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 X86-specific support for the FastISel class. Much
11 // of the target-specific code is generated by tablegen in the file
12 // X86GenFastISel.inc, which is #included here.
14 //===----------------------------------------------------------------------===//
17 #include "X86InstrBuilder.h"
18 #include "X86RegisterInfo.h"
19 #include "X86Subtarget.h"
20 #include "X86TargetMachine.h"
21 #include "llvm/CallingConv.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/GlobalVariable.h"
24 #include "llvm/Instructions.h"
25 #include "llvm/IntrinsicInst.h"
26 #include "llvm/Operator.h"
27 #include "llvm/CodeGen/Analysis.h"
28 #include "llvm/CodeGen/FastISel.h"
29 #include "llvm/CodeGen/FunctionLoweringInfo.h"
30 #include "llvm/CodeGen/MachineConstantPool.h"
31 #include "llvm/CodeGen/MachineFrameInfo.h"
32 #include "llvm/CodeGen/MachineRegisterInfo.h"
33 #include "llvm/Support/CallSite.h"
34 #include "llvm/Support/ErrorHandling.h"
35 #include "llvm/Support/GetElementPtrTypeIterator.h"
36 #include "llvm/Target/TargetOptions.h"
41 class X86FastISel : public FastISel {
42 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
43 /// make the right decision when generating code for different targets.
44 const X86Subtarget *Subtarget;
46 /// StackPtr - Register used as the stack pointer.
50 /// X86ScalarSSEf32, X86ScalarSSEf64 - Select between SSE or x87
51 /// floating point ops.
52 /// When SSE is available, use it for f32 operations.
53 /// When SSE2 is available, use it for f64 operations.
58 explicit X86FastISel(FunctionLoweringInfo &funcInfo) : FastISel(funcInfo) {
59 Subtarget = &TM.getSubtarget<X86Subtarget>();
60 StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
61 X86ScalarSSEf64 = Subtarget->hasSSE2();
62 X86ScalarSSEf32 = Subtarget->hasSSE1();
65 virtual bool TargetSelectInstruction(const Instruction *I);
67 /// TryToFoldLoad - The specified machine instr operand is a vreg, and that
68 /// vreg is being provided by the specified load instruction. If possible,
69 /// try to fold the load as an operand to the instruction, returning true if
71 virtual bool TryToFoldLoad(MachineInstr *MI, unsigned OpNo,
74 #include "X86GenFastISel.inc"
77 bool X86FastEmitCompare(const Value *LHS, const Value *RHS, EVT VT);
79 bool X86FastEmitLoad(EVT VT, const X86AddressMode &AM, unsigned &RR);
81 bool X86FastEmitStore(EVT VT, const Value *Val, const X86AddressMode &AM);
82 bool X86FastEmitStore(EVT VT, unsigned Val, const X86AddressMode &AM);
84 bool X86FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src, EVT SrcVT,
87 bool X86SelectAddress(const Value *V, X86AddressMode &AM);
88 bool X86SelectCallAddress(const Value *V, X86AddressMode &AM);
90 bool X86SelectLoad(const Instruction *I);
92 bool X86SelectStore(const Instruction *I);
94 bool X86SelectRet(const Instruction *I);
96 bool X86SelectCmp(const Instruction *I);
98 bool X86SelectZExt(const Instruction *I);
100 bool X86SelectBranch(const Instruction *I);
102 bool X86SelectShift(const Instruction *I);
104 bool X86SelectSelect(const Instruction *I);
106 bool X86SelectTrunc(const Instruction *I);
108 bool X86SelectFPExt(const Instruction *I);
109 bool X86SelectFPTrunc(const Instruction *I);
111 bool X86VisitIntrinsicCall(const IntrinsicInst &I);
112 bool X86SelectCall(const Instruction *I);
114 const X86InstrInfo *getInstrInfo() const {
115 return getTargetMachine()->getInstrInfo();
117 const X86TargetMachine *getTargetMachine() const {
118 return static_cast<const X86TargetMachine *>(&TM);
121 unsigned TargetMaterializeConstant(const Constant *C);
123 unsigned TargetMaterializeAlloca(const AllocaInst *C);
125 unsigned TargetMaterializeFloatZero(const ConstantFP *CF);
127 /// isScalarFPTypeInSSEReg - Return true if the specified scalar FP type is
128 /// computed in an SSE register, not on the X87 floating point stack.
129 bool isScalarFPTypeInSSEReg(EVT VT) const {
130 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
131 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1
134 bool isTypeLegal(const Type *Ty, MVT &VT, bool AllowI1 = false);
136 bool TryEmitSmallMemcpy(X86AddressMode DestAM,
137 X86AddressMode SrcAM, uint64_t Len);
140 } // end anonymous namespace.
142 bool X86FastISel::isTypeLegal(const Type *Ty, MVT &VT, bool AllowI1) {
143 EVT evt = TLI.getValueType(Ty, /*HandleUnknown=*/true);
144 if (evt == MVT::Other || !evt.isSimple())
145 // Unhandled type. Halt "fast" selection and bail.
148 VT = evt.getSimpleVT();
149 // For now, require SSE/SSE2 for performing floating-point operations,
150 // since x87 requires additional work.
151 if (VT == MVT::f64 && !X86ScalarSSEf64)
153 if (VT == MVT::f32 && !X86ScalarSSEf32)
155 // Similarly, no f80 support yet.
158 // We only handle legal types. For example, on x86-32 the instruction
159 // selector contains all of the 64-bit instructions from x86-64,
160 // under the assumption that i64 won't be used if the target doesn't
162 return (AllowI1 && VT == MVT::i1) || TLI.isTypeLegal(VT);
165 #include "X86GenCallingConv.inc"
167 /// X86FastEmitLoad - Emit a machine instruction to load a value of type VT.
168 /// The address is either pre-computed, i.e. Ptr, or a GlobalAddress, i.e. GV.
169 /// Return true and the result register by reference if it is possible.
170 bool X86FastISel::X86FastEmitLoad(EVT VT, const X86AddressMode &AM,
171 unsigned &ResultReg) {
172 // Get opcode and regclass of the output for the given load instruction.
174 const TargetRegisterClass *RC = NULL;
175 switch (VT.getSimpleVT().SimpleTy) {
176 default: return false;
180 RC = X86::GR8RegisterClass;
184 RC = X86::GR16RegisterClass;
188 RC = X86::GR32RegisterClass;
191 // Must be in x86-64 mode.
193 RC = X86::GR64RegisterClass;
196 if (Subtarget->hasSSE1()) {
198 RC = X86::FR32RegisterClass;
201 RC = X86::RFP32RegisterClass;
205 if (Subtarget->hasSSE2()) {
207 RC = X86::FR64RegisterClass;
210 RC = X86::RFP64RegisterClass;
214 // No f80 support yet.
218 ResultReg = createResultReg(RC);
219 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
220 DL, TII.get(Opc), ResultReg), AM);
224 /// X86FastEmitStore - Emit a machine instruction to store a value Val of
225 /// type VT. The address is either pre-computed, consisted of a base ptr, Ptr
226 /// and a displacement offset, or a GlobalAddress,
227 /// i.e. V. Return true if it is possible.
229 X86FastISel::X86FastEmitStore(EVT VT, unsigned Val, const X86AddressMode &AM) {
230 // Get opcode and regclass of the output for the given store instruction.
232 switch (VT.getSimpleVT().SimpleTy) {
233 case MVT::f80: // No f80 support yet.
234 default: return false;
236 // Mask out all but lowest bit.
237 unsigned AndResult = createResultReg(X86::GR8RegisterClass);
238 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
239 TII.get(X86::AND8ri), AndResult).addReg(Val).addImm(1);
242 // FALLTHROUGH, handling i1 as i8.
243 case MVT::i8: Opc = X86::MOV8mr; break;
244 case MVT::i16: Opc = X86::MOV16mr; break;
245 case MVT::i32: Opc = X86::MOV32mr; break;
246 case MVT::i64: Opc = X86::MOV64mr; break; // Must be in x86-64 mode.
248 Opc = Subtarget->hasSSE1() ? X86::MOVSSmr : X86::ST_Fp32m;
251 Opc = Subtarget->hasSSE2() ? X86::MOVSDmr : X86::ST_Fp64m;
255 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
256 DL, TII.get(Opc)), AM).addReg(Val);
260 bool X86FastISel::X86FastEmitStore(EVT VT, const Value *Val,
261 const X86AddressMode &AM) {
262 // Handle 'null' like i32/i64 0.
263 if (isa<ConstantPointerNull>(Val))
264 Val = Constant::getNullValue(TD.getIntPtrType(Val->getContext()));
266 // If this is a store of a simple constant, fold the constant into the store.
267 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
270 switch (VT.getSimpleVT().SimpleTy) {
272 case MVT::i1: Signed = false; // FALLTHROUGH to handle as i8.
273 case MVT::i8: Opc = X86::MOV8mi; break;
274 case MVT::i16: Opc = X86::MOV16mi; break;
275 case MVT::i32: Opc = X86::MOV32mi; break;
277 // Must be a 32-bit sign extended value.
278 if ((int)CI->getSExtValue() == CI->getSExtValue())
279 Opc = X86::MOV64mi32;
284 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
285 DL, TII.get(Opc)), AM)
286 .addImm(Signed ? (uint64_t) CI->getSExtValue() :
292 unsigned ValReg = getRegForValue(Val);
296 return X86FastEmitStore(VT, ValReg, AM);
299 /// X86FastEmitExtend - Emit a machine instruction to extend a value Src of
300 /// type SrcVT to type DstVT using the specified extension opcode Opc (e.g.
301 /// ISD::SIGN_EXTEND).
302 bool X86FastISel::X86FastEmitExtend(ISD::NodeType Opc, EVT DstVT,
303 unsigned Src, EVT SrcVT,
304 unsigned &ResultReg) {
305 unsigned RR = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), Opc,
306 Src, /*TODO: Kill=*/false);
315 /// X86SelectAddress - Attempt to fill in an address from the given value.
317 bool X86FastISel::X86SelectAddress(const Value *V, X86AddressMode &AM) {
318 const User *U = NULL;
319 unsigned Opcode = Instruction::UserOp1;
320 if (const Instruction *I = dyn_cast<Instruction>(V)) {
321 // Don't walk into other basic blocks; it's possible we haven't
322 // visited them yet, so the instructions may not yet be assigned
323 // virtual registers.
324 if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(V)) ||
325 FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
326 Opcode = I->getOpcode();
329 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(V)) {
330 Opcode = C->getOpcode();
334 if (const PointerType *Ty = dyn_cast<PointerType>(V->getType()))
335 if (Ty->getAddressSpace() > 255)
336 // Fast instruction selection doesn't support the special
342 case Instruction::BitCast:
343 // Look past bitcasts.
344 return X86SelectAddress(U->getOperand(0), AM);
346 case Instruction::IntToPtr:
347 // Look past no-op inttoptrs.
348 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
349 return X86SelectAddress(U->getOperand(0), AM);
352 case Instruction::PtrToInt:
353 // Look past no-op ptrtoints.
354 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
355 return X86SelectAddress(U->getOperand(0), AM);
358 case Instruction::Alloca: {
359 // Do static allocas.
360 const AllocaInst *A = cast<AllocaInst>(V);
361 DenseMap<const AllocaInst*, int>::iterator SI =
362 FuncInfo.StaticAllocaMap.find(A);
363 if (SI != FuncInfo.StaticAllocaMap.end()) {
364 AM.BaseType = X86AddressMode::FrameIndexBase;
365 AM.Base.FrameIndex = SI->second;
371 case Instruction::Add: {
372 // Adds of constants are common and easy enough.
373 if (const ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
374 uint64_t Disp = (int32_t)AM.Disp + (uint64_t)CI->getSExtValue();
375 // They have to fit in the 32-bit signed displacement field though.
376 if (isInt<32>(Disp)) {
377 AM.Disp = (uint32_t)Disp;
378 return X86SelectAddress(U->getOperand(0), AM);
384 case Instruction::GetElementPtr: {
385 X86AddressMode SavedAM = AM;
387 // Pattern-match simple GEPs.
388 uint64_t Disp = (int32_t)AM.Disp;
389 unsigned IndexReg = AM.IndexReg;
390 unsigned Scale = AM.Scale;
391 gep_type_iterator GTI = gep_type_begin(U);
392 // Iterate through the indices, folding what we can. Constants can be
393 // folded, and one dynamic index can be handled, if the scale is supported.
394 for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end();
395 i != e; ++i, ++GTI) {
396 const Value *Op = *i;
397 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
398 const StructLayout *SL = TD.getStructLayout(STy);
399 Disp += SL->getElementOffset(cast<ConstantInt>(Op)->getZExtValue());
403 // A array/variable index is always of the form i*S where S is the
404 // constant scale size. See if we can push the scale into immediates.
405 uint64_t S = TD.getTypeAllocSize(GTI.getIndexedType());
407 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
408 // Constant-offset addressing.
409 Disp += CI->getSExtValue() * S;
412 if (isa<AddOperator>(Op) &&
413 (!isa<Instruction>(Op) ||
414 FuncInfo.MBBMap[cast<Instruction>(Op)->getParent()]
416 isa<ConstantInt>(cast<AddOperator>(Op)->getOperand(1))) {
417 // An add (in the same block) with a constant operand. Fold the
420 cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
421 Disp += CI->getSExtValue() * S;
422 // Iterate on the other operand.
423 Op = cast<AddOperator>(Op)->getOperand(0);
427 (!AM.GV || !Subtarget->isPICStyleRIPRel()) &&
428 (S == 1 || S == 2 || S == 4 || S == 8)) {
429 // Scaled-index addressing.
431 IndexReg = getRegForGEPIndex(Op).first;
437 goto unsupported_gep;
440 // Check for displacement overflow.
441 if (!isInt<32>(Disp))
443 // Ok, the GEP indices were covered by constant-offset and scaled-index
444 // addressing. Update the address state and move on to examining the base.
445 AM.IndexReg = IndexReg;
447 AM.Disp = (uint32_t)Disp;
448 if (X86SelectAddress(U->getOperand(0), AM))
451 // If we couldn't merge the gep value into this addr mode, revert back to
452 // our address and just match the value instead of completely failing.
456 // Ok, the GEP indices weren't all covered.
461 // Handle constant address.
462 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
463 // Can't handle alternate code models or TLS yet.
464 if (TM.getCodeModel() != CodeModel::Small)
467 if (const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV))
468 if (GVar->isThreadLocal())
471 // RIP-relative addresses can't have additional register operands, so if
472 // we've already folded stuff into the addressing mode, just force the
473 // global value into its own register, which we can use as the basereg.
474 if (!Subtarget->isPICStyleRIPRel() ||
475 (AM.Base.Reg == 0 && AM.IndexReg == 0)) {
476 // Okay, we've committed to selecting this global. Set up the address.
479 // Allow the subtarget to classify the global.
480 unsigned char GVFlags = Subtarget->ClassifyGlobalReference(GV, TM);
482 // If this reference is relative to the pic base, set it now.
483 if (isGlobalRelativeToPICBase(GVFlags)) {
484 // FIXME: How do we know Base.Reg is free??
485 AM.Base.Reg = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF);
488 // Unless the ABI requires an extra load, return a direct reference to
490 if (!isGlobalStubReference(GVFlags)) {
491 if (Subtarget->isPICStyleRIPRel()) {
492 // Use rip-relative addressing if we can. Above we verified that the
493 // base and index registers are unused.
494 assert(AM.Base.Reg == 0 && AM.IndexReg == 0);
495 AM.Base.Reg = X86::RIP;
497 AM.GVOpFlags = GVFlags;
501 // Ok, we need to do a load from a stub. If we've already loaded from
502 // this stub, reuse the loaded pointer, otherwise emit the load now.
503 DenseMap<const Value*, unsigned>::iterator I = LocalValueMap.find(V);
505 if (I != LocalValueMap.end() && I->second != 0) {
508 // Issue load from stub.
510 const TargetRegisterClass *RC = NULL;
511 X86AddressMode StubAM;
512 StubAM.Base.Reg = AM.Base.Reg;
514 StubAM.GVOpFlags = GVFlags;
516 // Prepare for inserting code in the local-value area.
517 SavePoint SaveInsertPt = enterLocalValueArea();
519 if (TLI.getPointerTy() == MVT::i64) {
521 RC = X86::GR64RegisterClass;
523 if (Subtarget->isPICStyleRIPRel())
524 StubAM.Base.Reg = X86::RIP;
527 RC = X86::GR32RegisterClass;
530 LoadReg = createResultReg(RC);
531 MachineInstrBuilder LoadMI =
532 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), LoadReg);
533 addFullAddress(LoadMI, StubAM);
535 // Ok, back to normal mode.
536 leaveLocalValueArea(SaveInsertPt);
538 // Prevent loading GV stub multiple times in same MBB.
539 LocalValueMap[V] = LoadReg;
542 // Now construct the final address. Note that the Disp, Scale,
543 // and Index values may already be set here.
544 AM.Base.Reg = LoadReg;
550 // If all else fails, try to materialize the value in a register.
551 if (!AM.GV || !Subtarget->isPICStyleRIPRel()) {
552 if (AM.Base.Reg == 0) {
553 AM.Base.Reg = getRegForValue(V);
554 return AM.Base.Reg != 0;
556 if (AM.IndexReg == 0) {
557 assert(AM.Scale == 1 && "Scale with no index!");
558 AM.IndexReg = getRegForValue(V);
559 return AM.IndexReg != 0;
566 /// X86SelectCallAddress - Attempt to fill in an address from the given value.
568 bool X86FastISel::X86SelectCallAddress(const Value *V, X86AddressMode &AM) {
569 const User *U = NULL;
570 unsigned Opcode = Instruction::UserOp1;
571 if (const Instruction *I = dyn_cast<Instruction>(V)) {
572 Opcode = I->getOpcode();
574 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(V)) {
575 Opcode = C->getOpcode();
581 case Instruction::BitCast:
582 // Look past bitcasts.
583 return X86SelectCallAddress(U->getOperand(0), AM);
585 case Instruction::IntToPtr:
586 // Look past no-op inttoptrs.
587 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
588 return X86SelectCallAddress(U->getOperand(0), AM);
591 case Instruction::PtrToInt:
592 // Look past no-op ptrtoints.
593 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
594 return X86SelectCallAddress(U->getOperand(0), AM);
598 // Handle constant address.
599 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
600 // Can't handle alternate code models yet.
601 if (TM.getCodeModel() != CodeModel::Small)
604 // RIP-relative addresses can't have additional register operands.
605 if (Subtarget->isPICStyleRIPRel() &&
606 (AM.Base.Reg != 0 || AM.IndexReg != 0))
609 // Can't handle DLLImport.
610 if (GV->hasDLLImportLinkage())
614 if (const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV))
615 if (GVar->isThreadLocal())
618 // Okay, we've committed to selecting this global. Set up the basic address.
621 // No ABI requires an extra load for anything other than DLLImport, which
622 // we rejected above. Return a direct reference to the global.
623 if (Subtarget->isPICStyleRIPRel()) {
624 // Use rip-relative addressing if we can. Above we verified that the
625 // base and index registers are unused.
626 assert(AM.Base.Reg == 0 && AM.IndexReg == 0);
627 AM.Base.Reg = X86::RIP;
628 } else if (Subtarget->isPICStyleStubPIC()) {
629 AM.GVOpFlags = X86II::MO_PIC_BASE_OFFSET;
630 } else if (Subtarget->isPICStyleGOT()) {
631 AM.GVOpFlags = X86II::MO_GOTOFF;
637 // If all else fails, try to materialize the value in a register.
638 if (!AM.GV || !Subtarget->isPICStyleRIPRel()) {
639 if (AM.Base.Reg == 0) {
640 AM.Base.Reg = getRegForValue(V);
641 return AM.Base.Reg != 0;
643 if (AM.IndexReg == 0) {
644 assert(AM.Scale == 1 && "Scale with no index!");
645 AM.IndexReg = getRegForValue(V);
646 return AM.IndexReg != 0;
654 /// X86SelectStore - Select and emit code to implement store instructions.
655 bool X86FastISel::X86SelectStore(const Instruction *I) {
657 if (!isTypeLegal(I->getOperand(0)->getType(), VT, /*AllowI1=*/true))
661 if (!X86SelectAddress(I->getOperand(1), AM))
664 return X86FastEmitStore(VT, I->getOperand(0), AM);
667 /// X86SelectRet - Select and emit code to implement ret instructions.
668 bool X86FastISel::X86SelectRet(const Instruction *I) {
669 const ReturnInst *Ret = cast<ReturnInst>(I);
670 const Function &F = *I->getParent()->getParent();
672 if (!FuncInfo.CanLowerReturn)
675 CallingConv::ID CC = F.getCallingConv();
676 if (CC != CallingConv::C &&
677 CC != CallingConv::Fast &&
678 CC != CallingConv::X86_FastCall)
681 if (Subtarget->isTargetWin64())
684 // Don't handle popping bytes on return for now.
685 if (FuncInfo.MF->getInfo<X86MachineFunctionInfo>()
686 ->getBytesToPopOnReturn() != 0)
689 // fastcc with -tailcallopt is intended to provide a guaranteed
690 // tail call optimization. Fastisel doesn't know how to do that.
691 if (CC == CallingConv::Fast && GuaranteedTailCallOpt)
694 // Let SDISel handle vararg functions.
698 if (Ret->getNumOperands() > 0) {
699 SmallVector<ISD::OutputArg, 4> Outs;
700 GetReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(),
703 // Analyze operands of the call, assigning locations to each operand.
704 SmallVector<CCValAssign, 16> ValLocs;
705 CCState CCInfo(CC, F.isVarArg(), TM, ValLocs, I->getContext());
706 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
708 const Value *RV = Ret->getOperand(0);
709 unsigned Reg = getRegForValue(RV);
713 // Only handle a single return value for now.
714 if (ValLocs.size() != 1)
717 CCValAssign &VA = ValLocs[0];
719 // Don't bother handling odd stuff for now.
720 if (VA.getLocInfo() != CCValAssign::Full)
722 // Only handle register returns for now.
725 // TODO: For now, don't try to handle cases where getLocInfo()
726 // says Full but the types don't match.
727 if (TLI.getValueType(RV->getType()) != VA.getValVT())
730 // The calling-convention tables for x87 returns don't tell
732 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
736 unsigned SrcReg = Reg + VA.getValNo();
737 unsigned DstReg = VA.getLocReg();
738 const TargetRegisterClass* SrcRC = MRI.getRegClass(SrcReg);
739 // Avoid a cross-class copy. This is very unlikely.
740 if (!SrcRC->contains(DstReg))
742 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
743 DstReg).addReg(SrcReg);
745 // Mark the register as live out of the function.
746 MRI.addLiveOut(VA.getLocReg());
750 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::RET));
754 /// X86SelectLoad - Select and emit code to implement load instructions.
756 bool X86FastISel::X86SelectLoad(const Instruction *I) {
758 if (!isTypeLegal(I->getType(), VT, /*AllowI1=*/true))
762 if (!X86SelectAddress(I->getOperand(0), AM))
765 unsigned ResultReg = 0;
766 if (X86FastEmitLoad(VT, AM, ResultReg)) {
767 UpdateValueMap(I, ResultReg);
773 static unsigned X86ChooseCmpOpcode(EVT VT, const X86Subtarget *Subtarget) {
774 switch (VT.getSimpleVT().SimpleTy) {
776 case MVT::i8: return X86::CMP8rr;
777 case MVT::i16: return X86::CMP16rr;
778 case MVT::i32: return X86::CMP32rr;
779 case MVT::i64: return X86::CMP64rr;
780 case MVT::f32: return Subtarget->hasSSE1() ? X86::UCOMISSrr : 0;
781 case MVT::f64: return Subtarget->hasSSE2() ? X86::UCOMISDrr : 0;
785 /// X86ChooseCmpImmediateOpcode - If we have a comparison with RHS as the RHS
786 /// of the comparison, return an opcode that works for the compare (e.g.
787 /// CMP32ri) otherwise return 0.
788 static unsigned X86ChooseCmpImmediateOpcode(EVT VT, const ConstantInt *RHSC) {
789 switch (VT.getSimpleVT().SimpleTy) {
790 // Otherwise, we can't fold the immediate into this comparison.
792 case MVT::i8: return X86::CMP8ri;
793 case MVT::i16: return X86::CMP16ri;
794 case MVT::i32: return X86::CMP32ri;
796 // 64-bit comparisons are only valid if the immediate fits in a 32-bit sext
798 if ((int)RHSC->getSExtValue() == RHSC->getSExtValue())
799 return X86::CMP64ri32;
804 bool X86FastISel::X86FastEmitCompare(const Value *Op0, const Value *Op1,
806 unsigned Op0Reg = getRegForValue(Op0);
807 if (Op0Reg == 0) return false;
809 // Handle 'null' like i32/i64 0.
810 if (isa<ConstantPointerNull>(Op1))
811 Op1 = Constant::getNullValue(TD.getIntPtrType(Op0->getContext()));
813 // We have two options: compare with register or immediate. If the RHS of
814 // the compare is an immediate that we can fold into this compare, use
815 // CMPri, otherwise use CMPrr.
816 if (const ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
817 if (unsigned CompareImmOpc = X86ChooseCmpImmediateOpcode(VT, Op1C)) {
818 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CompareImmOpc))
820 .addImm(Op1C->getSExtValue());
825 unsigned CompareOpc = X86ChooseCmpOpcode(VT, Subtarget);
826 if (CompareOpc == 0) return false;
828 unsigned Op1Reg = getRegForValue(Op1);
829 if (Op1Reg == 0) return false;
830 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CompareOpc))
837 bool X86FastISel::X86SelectCmp(const Instruction *I) {
838 const CmpInst *CI = cast<CmpInst>(I);
841 if (!isTypeLegal(I->getOperand(0)->getType(), VT))
844 unsigned ResultReg = createResultReg(&X86::GR8RegClass);
846 bool SwapArgs; // false -> compare Op0, Op1. true -> compare Op1, Op0.
847 switch (CI->getPredicate()) {
848 case CmpInst::FCMP_OEQ: {
849 if (!X86FastEmitCompare(CI->getOperand(0), CI->getOperand(1), VT))
852 unsigned EReg = createResultReg(&X86::GR8RegClass);
853 unsigned NPReg = createResultReg(&X86::GR8RegClass);
854 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::SETEr), EReg);
855 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
856 TII.get(X86::SETNPr), NPReg);
857 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
858 TII.get(X86::AND8rr), ResultReg).addReg(NPReg).addReg(EReg);
859 UpdateValueMap(I, ResultReg);
862 case CmpInst::FCMP_UNE: {
863 if (!X86FastEmitCompare(CI->getOperand(0), CI->getOperand(1), VT))
866 unsigned NEReg = createResultReg(&X86::GR8RegClass);
867 unsigned PReg = createResultReg(&X86::GR8RegClass);
868 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::SETNEr), NEReg);
869 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::SETPr), PReg);
870 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::OR8rr),ResultReg)
871 .addReg(PReg).addReg(NEReg);
872 UpdateValueMap(I, ResultReg);
875 case CmpInst::FCMP_OGT: SwapArgs = false; SetCCOpc = X86::SETAr; break;
876 case CmpInst::FCMP_OGE: SwapArgs = false; SetCCOpc = X86::SETAEr; break;
877 case CmpInst::FCMP_OLT: SwapArgs = true; SetCCOpc = X86::SETAr; break;
878 case CmpInst::FCMP_OLE: SwapArgs = true; SetCCOpc = X86::SETAEr; break;
879 case CmpInst::FCMP_ONE: SwapArgs = false; SetCCOpc = X86::SETNEr; break;
880 case CmpInst::FCMP_ORD: SwapArgs = false; SetCCOpc = X86::SETNPr; break;
881 case CmpInst::FCMP_UNO: SwapArgs = false; SetCCOpc = X86::SETPr; break;
882 case CmpInst::FCMP_UEQ: SwapArgs = false; SetCCOpc = X86::SETEr; break;
883 case CmpInst::FCMP_UGT: SwapArgs = true; SetCCOpc = X86::SETBr; break;
884 case CmpInst::FCMP_UGE: SwapArgs = true; SetCCOpc = X86::SETBEr; break;
885 case CmpInst::FCMP_ULT: SwapArgs = false; SetCCOpc = X86::SETBr; break;
886 case CmpInst::FCMP_ULE: SwapArgs = false; SetCCOpc = X86::SETBEr; break;
888 case CmpInst::ICMP_EQ: SwapArgs = false; SetCCOpc = X86::SETEr; break;
889 case CmpInst::ICMP_NE: SwapArgs = false; SetCCOpc = X86::SETNEr; break;
890 case CmpInst::ICMP_UGT: SwapArgs = false; SetCCOpc = X86::SETAr; break;
891 case CmpInst::ICMP_UGE: SwapArgs = false; SetCCOpc = X86::SETAEr; break;
892 case CmpInst::ICMP_ULT: SwapArgs = false; SetCCOpc = X86::SETBr; break;
893 case CmpInst::ICMP_ULE: SwapArgs = false; SetCCOpc = X86::SETBEr; break;
894 case CmpInst::ICMP_SGT: SwapArgs = false; SetCCOpc = X86::SETGr; break;
895 case CmpInst::ICMP_SGE: SwapArgs = false; SetCCOpc = X86::SETGEr; break;
896 case CmpInst::ICMP_SLT: SwapArgs = false; SetCCOpc = X86::SETLr; break;
897 case CmpInst::ICMP_SLE: SwapArgs = false; SetCCOpc = X86::SETLEr; break;
902 const Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
906 // Emit a compare of Op0/Op1.
907 if (!X86FastEmitCompare(Op0, Op1, VT))
910 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(SetCCOpc), ResultReg);
911 UpdateValueMap(I, ResultReg);
915 bool X86FastISel::X86SelectZExt(const Instruction *I) {
916 // Handle zero-extension from i1 to i8, which is common.
917 if (I->getType()->isIntegerTy(8) &&
918 I->getOperand(0)->getType()->isIntegerTy(1)) {
919 unsigned ResultReg = getRegForValue(I->getOperand(0));
920 if (ResultReg == 0) return false;
921 // Set the high bits to zero.
922 ResultReg = FastEmitZExtFromI1(MVT::i8, ResultReg, /*TODO: Kill=*/false);
923 if (ResultReg == 0) return false;
924 UpdateValueMap(I, ResultReg);
932 bool X86FastISel::X86SelectBranch(const Instruction *I) {
933 // Unconditional branches are selected by tablegen-generated code.
934 // Handle a conditional branch.
935 const BranchInst *BI = cast<BranchInst>(I);
936 MachineBasicBlock *TrueMBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
937 MachineBasicBlock *FalseMBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
939 // Fold the common case of a conditional branch with a comparison
940 // in the same block (values defined on other blocks may not have
941 // initialized registers).
942 if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
943 if (CI->hasOneUse() && CI->getParent() == I->getParent()) {
944 EVT VT = TLI.getValueType(CI->getOperand(0)->getType());
946 // Try to take advantage of fallthrough opportunities.
947 CmpInst::Predicate Predicate = CI->getPredicate();
948 if (FuncInfo.MBB->isLayoutSuccessor(TrueMBB)) {
949 std::swap(TrueMBB, FalseMBB);
950 Predicate = CmpInst::getInversePredicate(Predicate);
953 bool SwapArgs; // false -> compare Op0, Op1. true -> compare Op1, Op0.
954 unsigned BranchOpc; // Opcode to jump on, e.g. "X86::JA"
957 case CmpInst::FCMP_OEQ:
958 std::swap(TrueMBB, FalseMBB);
959 Predicate = CmpInst::FCMP_UNE;
961 case CmpInst::FCMP_UNE: SwapArgs = false; BranchOpc = X86::JNE_4; break;
962 case CmpInst::FCMP_OGT: SwapArgs = false; BranchOpc = X86::JA_4; break;
963 case CmpInst::FCMP_OGE: SwapArgs = false; BranchOpc = X86::JAE_4; break;
964 case CmpInst::FCMP_OLT: SwapArgs = true; BranchOpc = X86::JA_4; break;
965 case CmpInst::FCMP_OLE: SwapArgs = true; BranchOpc = X86::JAE_4; break;
966 case CmpInst::FCMP_ONE: SwapArgs = false; BranchOpc = X86::JNE_4; break;
967 case CmpInst::FCMP_ORD: SwapArgs = false; BranchOpc = X86::JNP_4; break;
968 case CmpInst::FCMP_UNO: SwapArgs = false; BranchOpc = X86::JP_4; break;
969 case CmpInst::FCMP_UEQ: SwapArgs = false; BranchOpc = X86::JE_4; break;
970 case CmpInst::FCMP_UGT: SwapArgs = true; BranchOpc = X86::JB_4; break;
971 case CmpInst::FCMP_UGE: SwapArgs = true; BranchOpc = X86::JBE_4; break;
972 case CmpInst::FCMP_ULT: SwapArgs = false; BranchOpc = X86::JB_4; break;
973 case CmpInst::FCMP_ULE: SwapArgs = false; BranchOpc = X86::JBE_4; break;
975 case CmpInst::ICMP_EQ: SwapArgs = false; BranchOpc = X86::JE_4; break;
976 case CmpInst::ICMP_NE: SwapArgs = false; BranchOpc = X86::JNE_4; break;
977 case CmpInst::ICMP_UGT: SwapArgs = false; BranchOpc = X86::JA_4; break;
978 case CmpInst::ICMP_UGE: SwapArgs = false; BranchOpc = X86::JAE_4; break;
979 case CmpInst::ICMP_ULT: SwapArgs = false; BranchOpc = X86::JB_4; break;
980 case CmpInst::ICMP_ULE: SwapArgs = false; BranchOpc = X86::JBE_4; break;
981 case CmpInst::ICMP_SGT: SwapArgs = false; BranchOpc = X86::JG_4; break;
982 case CmpInst::ICMP_SGE: SwapArgs = false; BranchOpc = X86::JGE_4; break;
983 case CmpInst::ICMP_SLT: SwapArgs = false; BranchOpc = X86::JL_4; break;
984 case CmpInst::ICMP_SLE: SwapArgs = false; BranchOpc = X86::JLE_4; break;
989 const Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
993 // Emit a compare of the LHS and RHS, setting the flags.
994 if (!X86FastEmitCompare(Op0, Op1, VT))
997 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BranchOpc))
1000 if (Predicate == CmpInst::FCMP_UNE) {
1001 // X86 requires a second branch to handle UNE (and OEQ,
1002 // which is mapped to UNE above).
1003 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::JP_4))
1007 FastEmitBranch(FalseMBB, DL);
1008 FuncInfo.MBB->addSuccessor(TrueMBB);
1011 } else if (TruncInst *TI = dyn_cast<TruncInst>(BI->getCondition())) {
1012 // Handle things like "%cond = trunc i32 %X to i1 / br i1 %cond", which
1013 // typically happen for _Bool and C++ bools.
1015 if (TI->hasOneUse() && TI->getParent() == I->getParent() &&
1016 isTypeLegal(TI->getOperand(0)->getType(), SourceVT)) {
1017 unsigned TestOpc = 0;
1018 switch (SourceVT.SimpleTy) {
1020 case MVT::i8: TestOpc = X86::TEST8ri; break;
1021 case MVT::i16: TestOpc = X86::TEST16ri; break;
1022 case MVT::i32: TestOpc = X86::TEST32ri; break;
1023 case MVT::i64: TestOpc = X86::TEST64ri32; break;
1026 unsigned OpReg = getRegForValue(TI->getOperand(0));
1027 if (OpReg == 0) return false;
1028 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TestOpc))
1029 .addReg(OpReg).addImm(1);
1031 unsigned JmpOpc = X86::JNE_4;
1032 if (FuncInfo.MBB->isLayoutSuccessor(TrueMBB)) {
1033 std::swap(TrueMBB, FalseMBB);
1037 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(JmpOpc))
1039 FastEmitBranch(FalseMBB, DL);
1040 FuncInfo.MBB->addSuccessor(TrueMBB);
1046 // Otherwise do a clumsy setcc and re-test it.
1047 // Note that i1 essentially gets ANY_EXTEND'ed to i8 where it isn't used
1048 // in an explicit cast, so make sure to handle that correctly.
1049 unsigned OpReg = getRegForValue(BI->getCondition());
1050 if (OpReg == 0) return false;
1052 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::TEST8ri))
1053 .addReg(OpReg).addImm(1);
1054 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::JNE_4))
1056 FastEmitBranch(FalseMBB, DL);
1057 FuncInfo.MBB->addSuccessor(TrueMBB);
1061 bool X86FastISel::X86SelectShift(const Instruction *I) {
1062 unsigned CReg = 0, OpReg = 0;
1063 const TargetRegisterClass *RC = NULL;
1064 if (I->getType()->isIntegerTy(8)) {
1066 RC = &X86::GR8RegClass;
1067 switch (I->getOpcode()) {
1068 case Instruction::LShr: OpReg = X86::SHR8rCL; break;
1069 case Instruction::AShr: OpReg = X86::SAR8rCL; break;
1070 case Instruction::Shl: OpReg = X86::SHL8rCL; break;
1071 default: return false;
1073 } else if (I->getType()->isIntegerTy(16)) {
1075 RC = &X86::GR16RegClass;
1076 switch (I->getOpcode()) {
1077 case Instruction::LShr: OpReg = X86::SHR16rCL; break;
1078 case Instruction::AShr: OpReg = X86::SAR16rCL; break;
1079 case Instruction::Shl: OpReg = X86::SHL16rCL; break;
1080 default: return false;
1082 } else if (I->getType()->isIntegerTy(32)) {
1084 RC = &X86::GR32RegClass;
1085 switch (I->getOpcode()) {
1086 case Instruction::LShr: OpReg = X86::SHR32rCL; break;
1087 case Instruction::AShr: OpReg = X86::SAR32rCL; break;
1088 case Instruction::Shl: OpReg = X86::SHL32rCL; break;
1089 default: return false;
1091 } else if (I->getType()->isIntegerTy(64)) {
1093 RC = &X86::GR64RegClass;
1094 switch (I->getOpcode()) {
1095 case Instruction::LShr: OpReg = X86::SHR64rCL; break;
1096 case Instruction::AShr: OpReg = X86::SAR64rCL; break;
1097 case Instruction::Shl: OpReg = X86::SHL64rCL; break;
1098 default: return false;
1105 if (!isTypeLegal(I->getType(), VT))
1108 unsigned Op0Reg = getRegForValue(I->getOperand(0));
1109 if (Op0Reg == 0) return false;
1111 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1112 if (Op1Reg == 0) return false;
1113 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1114 CReg).addReg(Op1Reg);
1116 // The shift instruction uses X86::CL. If we defined a super-register
1117 // of X86::CL, emit a subreg KILL to precisely describe what we're doing here.
1118 if (CReg != X86::CL)
1119 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1120 TII.get(TargetOpcode::KILL), X86::CL)
1121 .addReg(CReg, RegState::Kill);
1123 unsigned ResultReg = createResultReg(RC);
1124 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(OpReg), ResultReg)
1126 UpdateValueMap(I, ResultReg);
1130 bool X86FastISel::X86SelectSelect(const Instruction *I) {
1132 if (!isTypeLegal(I->getType(), VT))
1135 // We only use cmov here, if we don't have a cmov instruction bail.
1136 if (!Subtarget->hasCMov()) return false;
1139 const TargetRegisterClass *RC = NULL;
1140 if (VT == MVT::i16) {
1141 Opc = X86::CMOVE16rr;
1142 RC = &X86::GR16RegClass;
1143 } else if (VT == MVT::i32) {
1144 Opc = X86::CMOVE32rr;
1145 RC = &X86::GR32RegClass;
1146 } else if (VT == MVT::i64) {
1147 Opc = X86::CMOVE64rr;
1148 RC = &X86::GR64RegClass;
1153 unsigned Op0Reg = getRegForValue(I->getOperand(0));
1154 if (Op0Reg == 0) return false;
1155 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1156 if (Op1Reg == 0) return false;
1157 unsigned Op2Reg = getRegForValue(I->getOperand(2));
1158 if (Op2Reg == 0) return false;
1160 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::TEST8rr))
1161 .addReg(Op0Reg).addReg(Op0Reg);
1162 unsigned ResultReg = createResultReg(RC);
1163 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), ResultReg)
1164 .addReg(Op1Reg).addReg(Op2Reg);
1165 UpdateValueMap(I, ResultReg);
1169 bool X86FastISel::X86SelectFPExt(const Instruction *I) {
1170 // fpext from float to double.
1171 if (Subtarget->hasSSE2() &&
1172 I->getType()->isDoubleTy()) {
1173 const Value *V = I->getOperand(0);
1174 if (V->getType()->isFloatTy()) {
1175 unsigned OpReg = getRegForValue(V);
1176 if (OpReg == 0) return false;
1177 unsigned ResultReg = createResultReg(X86::FR64RegisterClass);
1178 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1179 TII.get(X86::CVTSS2SDrr), ResultReg)
1181 UpdateValueMap(I, ResultReg);
1189 bool X86FastISel::X86SelectFPTrunc(const Instruction *I) {
1190 if (Subtarget->hasSSE2()) {
1191 if (I->getType()->isFloatTy()) {
1192 const Value *V = I->getOperand(0);
1193 if (V->getType()->isDoubleTy()) {
1194 unsigned OpReg = getRegForValue(V);
1195 if (OpReg == 0) return false;
1196 unsigned ResultReg = createResultReg(X86::FR32RegisterClass);
1197 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1198 TII.get(X86::CVTSD2SSrr), ResultReg)
1200 UpdateValueMap(I, ResultReg);
1209 bool X86FastISel::X86SelectTrunc(const Instruction *I) {
1210 if (Subtarget->is64Bit())
1211 // All other cases should be handled by the tblgen generated code.
1213 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
1214 EVT DstVT = TLI.getValueType(I->getType());
1216 // This code only handles truncation to byte right now.
1217 if (DstVT != MVT::i8 && DstVT != MVT::i1)
1218 // All other cases should be handled by the tblgen generated code.
1220 if (SrcVT != MVT::i16 && SrcVT != MVT::i32)
1221 // All other cases should be handled by the tblgen generated code.
1224 unsigned InputReg = getRegForValue(I->getOperand(0));
1226 // Unhandled operand. Halt "fast" selection and bail.
1229 // First issue a copy to GR16_ABCD or GR32_ABCD.
1230 const TargetRegisterClass *CopyRC = (SrcVT == MVT::i16)
1231 ? X86::GR16_ABCDRegisterClass : X86::GR32_ABCDRegisterClass;
1232 unsigned CopyReg = createResultReg(CopyRC);
1233 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1234 CopyReg).addReg(InputReg);
1236 // Then issue an extract_subreg.
1237 unsigned ResultReg = FastEmitInst_extractsubreg(MVT::i8,
1238 CopyReg, /*Kill=*/true,
1243 UpdateValueMap(I, ResultReg);
1247 bool X86FastISel::TryEmitSmallMemcpy(X86AddressMode DestAM,
1248 X86AddressMode SrcAM, uint64_t Len) {
1249 // Make sure we don't bloat code by inlining very large memcpy's.
1250 bool i64Legal = TLI.isTypeLegal(MVT::i64);
1251 if (Len > (i64Legal ? 32 : 16)) return false;
1253 // We don't care about alignment here since we just emit integer accesses.
1256 if (Len >= 8 && i64Legal)
1268 bool RV = X86FastEmitLoad(VT, SrcAM, Reg);
1269 RV &= X86FastEmitStore(VT, Reg, DestAM);
1270 assert(RV && "Failed to emit load or store??");
1272 unsigned Size = VT.getSizeInBits()/8;
1274 DestAM.Disp += Size;
1281 bool X86FastISel::X86VisitIntrinsicCall(const IntrinsicInst &I) {
1282 // FIXME: Handle more intrinsics.
1283 switch (I.getIntrinsicID()) {
1284 default: return false;
1285 case Intrinsic::memcpy: {
1286 const MemCpyInst &MCI = cast<MemCpyInst>(I);
1287 // Don't handle volatile or variable length memcpys.
1288 if (MCI.isVolatile() || !isa<ConstantInt>(MCI.getLength()))
1291 uint64_t Len = cast<ConstantInt>(MCI.getLength())->getZExtValue();
1293 // Get the address of the dest and source addresses.
1294 X86AddressMode DestAM, SrcAM;
1295 if (!X86SelectAddress(MCI.getRawDest(), DestAM) ||
1296 !X86SelectAddress(MCI.getRawSource(), SrcAM))
1299 return TryEmitSmallMemcpy(DestAM, SrcAM, Len);
1302 case Intrinsic::stackprotector: {
1303 // Emit code inline code to store the stack guard onto the stack.
1304 EVT PtrTy = TLI.getPointerTy();
1306 const Value *Op1 = I.getArgOperand(0); // The guard's value.
1307 const AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1));
1309 // Grab the frame index.
1311 if (!X86SelectAddress(Slot, AM)) return false;
1312 if (!X86FastEmitStore(PtrTy, Op1, AM)) return false;
1315 case Intrinsic::dbg_declare: {
1316 const DbgDeclareInst *DI = cast<DbgDeclareInst>(&I);
1318 assert(DI->getAddress() && "Null address should be checked earlier!");
1319 if (!X86SelectAddress(DI->getAddress(), AM))
1321 const TargetInstrDesc &II = TII.get(TargetOpcode::DBG_VALUE);
1322 // FIXME may need to add RegState::Debug to any registers produced,
1323 // although ESP/EBP should be the only ones at the moment.
1324 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II), AM).
1325 addImm(0).addMetadata(DI->getVariable());
1328 case Intrinsic::trap: {
1329 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::TRAP));
1332 case Intrinsic::sadd_with_overflow:
1333 case Intrinsic::uadd_with_overflow: {
1334 // FIXME: Should fold immediates.
1336 // Replace "add with overflow" intrinsics with an "add" instruction followed
1337 // by a seto/setc instruction.
1338 const Function *Callee = I.getCalledFunction();
1340 cast<StructType>(Callee->getReturnType())->getTypeAtIndex(unsigned(0));
1343 if (!isTypeLegal(RetTy, VT))
1346 const Value *Op1 = I.getArgOperand(0);
1347 const Value *Op2 = I.getArgOperand(1);
1348 unsigned Reg1 = getRegForValue(Op1);
1349 unsigned Reg2 = getRegForValue(Op2);
1351 if (Reg1 == 0 || Reg2 == 0)
1352 // FIXME: Handle values *not* in registers.
1358 else if (VT == MVT::i64)
1363 // The call to CreateRegs builds two sequential registers, to store the
1364 // both the the returned values.
1365 unsigned ResultReg = FuncInfo.CreateRegs(I.getType());
1366 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(OpC), ResultReg)
1367 .addReg(Reg1).addReg(Reg2);
1369 unsigned Opc = X86::SETBr;
1370 if (I.getIntrinsicID() == Intrinsic::sadd_with_overflow)
1372 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), ResultReg+1);
1374 UpdateValueMap(&I, ResultReg, 2);
1380 bool X86FastISel::X86SelectCall(const Instruction *I) {
1381 const CallInst *CI = cast<CallInst>(I);
1382 const Value *Callee = CI->getCalledValue();
1384 // Can't handle inline asm yet.
1385 if (isa<InlineAsm>(Callee))
1388 // Handle intrinsic calls.
1389 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI))
1390 return X86VisitIntrinsicCall(*II);
1392 // Handle only C and fastcc calling conventions for now.
1393 ImmutableCallSite CS(CI);
1394 CallingConv::ID CC = CS.getCallingConv();
1395 if (CC != CallingConv::C && CC != CallingConv::Fast &&
1396 CC != CallingConv::X86_FastCall)
1399 // fastcc with -tailcallopt is intended to provide a guaranteed
1400 // tail call optimization. Fastisel doesn't know how to do that.
1401 if (CC == CallingConv::Fast && GuaranteedTailCallOpt)
1404 const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
1405 const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
1406 bool isVarArg = FTy->isVarArg();
1408 // Don't know how to handle Win64 varargs yet. Nothing special needed for
1409 // x86-32. Special handling for x86-64 is implemented.
1410 if (isVarArg && Subtarget->isTargetWin64())
1413 // Fast-isel doesn't know about callee-pop yet.
1414 if (Subtarget->IsCalleePop(isVarArg, CC))
1417 // Check whether the function can return without sret-demotion.
1418 SmallVector<ISD::OutputArg, 4> Outs;
1419 SmallVector<uint64_t, 4> Offsets;
1420 GetReturnInfo(I->getType(), CS.getAttributes().getRetAttributes(),
1421 Outs, TLI, &Offsets);
1422 bool CanLowerReturn = TLI.CanLowerReturn(CS.getCallingConv(),
1423 FTy->isVarArg(), Outs, FTy->getContext());
1424 if (!CanLowerReturn)
1427 // Materialize callee address in a register. FIXME: GV address can be
1428 // handled with a CALLpcrel32 instead.
1429 X86AddressMode CalleeAM;
1430 if (!X86SelectCallAddress(Callee, CalleeAM))
1432 unsigned CalleeOp = 0;
1433 const GlobalValue *GV = 0;
1434 if (CalleeAM.GV != 0) {
1436 } else if (CalleeAM.Base.Reg != 0) {
1437 CalleeOp = CalleeAM.Base.Reg;
1441 // Deal with call operands first.
1442 SmallVector<const Value *, 8> ArgVals;
1443 SmallVector<unsigned, 8> Args;
1444 SmallVector<MVT, 8> ArgVTs;
1445 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
1446 Args.reserve(CS.arg_size());
1447 ArgVals.reserve(CS.arg_size());
1448 ArgVTs.reserve(CS.arg_size());
1449 ArgFlags.reserve(CS.arg_size());
1450 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
1453 ISD::ArgFlagsTy Flags;
1454 unsigned AttrInd = i - CS.arg_begin() + 1;
1455 if (CS.paramHasAttr(AttrInd, Attribute::SExt))
1457 if (CS.paramHasAttr(AttrInd, Attribute::ZExt))
1460 // If this is an i1/i8/i16 argument, promote to i32 to avoid an extra
1461 // instruction. This is safe because it is common to all fastisel supported
1462 // calling conventions on x86.
1463 if (ConstantInt *CI = dyn_cast<ConstantInt>(ArgVal)) {
1464 if (CI->getBitWidth() == 1 || CI->getBitWidth() == 8 ||
1465 CI->getBitWidth() == 16) {
1467 ArgVal = ConstantExpr::getSExt(CI,Type::getInt32Ty(CI->getContext()));
1469 ArgVal = ConstantExpr::getZExt(CI,Type::getInt32Ty(CI->getContext()));
1475 // Passing bools around ends up doing a trunc to i1 and passing it.
1476 // Codegen this as an argument + "and 1".
1477 if (ArgVal->getType()->isIntegerTy(1) && isa<TruncInst>(ArgVal) &&
1478 cast<TruncInst>(ArgVal)->getParent() == I->getParent() &&
1479 ArgVal->hasOneUse()) {
1480 ArgVal = cast<TruncInst>(ArgVal)->getOperand(0);
1481 ArgReg = getRegForValue(ArgVal);
1482 if (ArgReg == 0) return false;
1485 if (!isTypeLegal(ArgVal->getType(), ArgVT)) return false;
1487 ArgReg = FastEmit_ri(ArgVT, ArgVT, ISD::AND, ArgReg,
1488 ArgVal->hasOneUse(), 1);
1490 ArgReg = getRegForValue(ArgVal);
1493 if (ArgReg == 0) return false;
1495 // FIXME: Only handle *easy* calls for now.
1496 if (CS.paramHasAttr(AttrInd, Attribute::InReg) ||
1497 CS.paramHasAttr(AttrInd, Attribute::Nest) ||
1498 CS.paramHasAttr(AttrInd, Attribute::ByVal))
1501 const Type *ArgTy = ArgVal->getType();
1503 if (!isTypeLegal(ArgTy, ArgVT))
1505 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
1506 Flags.setOrigAlign(OriginalAlignment);
1508 Args.push_back(ArgReg);
1509 ArgVals.push_back(ArgVal);
1510 ArgVTs.push_back(ArgVT);
1511 ArgFlags.push_back(Flags);
1514 // Analyze operands of the call, assigning locations to each operand.
1515 SmallVector<CCValAssign, 16> ArgLocs;
1516 CCState CCInfo(CC, isVarArg, TM, ArgLocs, I->getParent()->getContext());
1518 // Allocate shadow area for Win64
1519 if (Subtarget->isTargetWin64())
1520 CCInfo.AllocateStack(32, 8);
1522 CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CC_X86);
1524 // Get a count of how many bytes are to be pushed on the stack.
1525 unsigned NumBytes = CCInfo.getNextStackOffset();
1527 // Issue CALLSEQ_START
1528 unsigned AdjStackDown = TM.getRegisterInfo()->getCallFrameSetupOpcode();
1529 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(AdjStackDown))
1532 // Process argument: walk the register/memloc assignments, inserting
1534 SmallVector<unsigned, 4> RegArgs;
1535 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1536 CCValAssign &VA = ArgLocs[i];
1537 unsigned Arg = Args[VA.getValNo()];
1538 EVT ArgVT = ArgVTs[VA.getValNo()];
1540 // Promote the value if needed.
1541 switch (VA.getLocInfo()) {
1542 default: llvm_unreachable("Unknown loc info!");
1543 case CCValAssign::Full: break;
1544 case CCValAssign::SExt: {
1545 bool Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
1547 assert(Emitted && "Failed to emit a sext!"); (void)Emitted;
1548 ArgVT = VA.getLocVT();
1551 case CCValAssign::ZExt: {
1552 bool Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
1554 assert(Emitted && "Failed to emit a zext!"); (void)Emitted;
1555 ArgVT = VA.getLocVT();
1558 case CCValAssign::AExt: {
1559 // We don't handle MMX parameters yet.
1560 if (VA.getLocVT().isVector() && VA.getLocVT().getSizeInBits() == 128)
1562 bool Emitted = X86FastEmitExtend(ISD::ANY_EXTEND, VA.getLocVT(),
1565 Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
1568 Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
1571 assert(Emitted && "Failed to emit a aext!"); (void)Emitted;
1572 ArgVT = VA.getLocVT();
1575 case CCValAssign::BCvt: {
1576 unsigned BC = FastEmit_r(ArgVT.getSimpleVT(), VA.getLocVT(),
1577 ISD::BITCAST, Arg, /*TODO: Kill=*/false);
1578 assert(BC != 0 && "Failed to emit a bitcast!");
1580 ArgVT = VA.getLocVT();
1585 if (VA.isRegLoc()) {
1586 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1587 VA.getLocReg()).addReg(Arg);
1588 RegArgs.push_back(VA.getLocReg());
1590 unsigned LocMemOffset = VA.getLocMemOffset();
1592 AM.Base.Reg = StackPtr;
1593 AM.Disp = LocMemOffset;
1594 const Value *ArgVal = ArgVals[VA.getValNo()];
1596 // If this is a really simple value, emit this with the Value* version of
1597 // X86FastEmitStore. If it isn't simple, we don't want to do this, as it
1598 // can cause us to reevaluate the argument.
1599 if (isa<ConstantInt>(ArgVal) || isa<ConstantPointerNull>(ArgVal))
1600 X86FastEmitStore(ArgVT, ArgVal, AM);
1602 X86FastEmitStore(ArgVT, Arg, AM);
1606 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1608 if (Subtarget->isPICStyleGOT()) {
1609 unsigned Base = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF);
1610 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1611 X86::EBX).addReg(Base);
1614 if (Subtarget->is64Bit() && isVarArg && !Subtarget->isTargetWin64()) {
1615 // Count the number of XMM registers allocated.
1616 static const unsigned XMMArgRegs[] = {
1617 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1618 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1620 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1621 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::MOV8ri),
1622 X86::AL).addImm(NumXMMRegs);
1626 MachineInstrBuilder MIB;
1628 // Register-indirect call.
1630 if (Subtarget->isTargetWin64())
1631 CallOpc = X86::WINCALL64r;
1632 else if (Subtarget->is64Bit())
1633 CallOpc = X86::CALL64r;
1635 CallOpc = X86::CALL32r;
1636 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CallOpc))
1641 assert(GV && "Not a direct call");
1643 if (Subtarget->isTargetWin64())
1644 CallOpc = X86::WINCALL64pcrel32;
1645 else if (Subtarget->is64Bit())
1646 CallOpc = X86::CALL64pcrel32;
1648 CallOpc = X86::CALLpcrel32;
1650 // See if we need any target-specific flags on the GV operand.
1651 unsigned char OpFlags = 0;
1653 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
1654 // external symbols most go through the PLT in PIC mode. If the symbol
1655 // has hidden or protected visibility, or if it is static or local, then
1656 // we don't need to use the PLT - we can directly call it.
1657 if (Subtarget->isTargetELF() &&
1658 TM.getRelocationModel() == Reloc::PIC_ &&
1659 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
1660 OpFlags = X86II::MO_PLT;
1661 } else if (Subtarget->isPICStyleStubAny() &&
1662 (GV->isDeclaration() || GV->isWeakForLinker()) &&
1663 (!Subtarget->getTargetTriple().isMacOSX() ||
1664 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
1665 // PC-relative references to external symbols should go through $stub,
1666 // unless we're building with the leopard linker or later, which
1667 // automatically synthesizes these stubs.
1668 OpFlags = X86II::MO_DARWIN_STUB;
1672 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CallOpc))
1673 .addGlobalAddress(GV, 0, OpFlags);
1676 // Add an implicit use GOT pointer in EBX.
1677 if (Subtarget->isPICStyleGOT())
1678 MIB.addReg(X86::EBX);
1680 if (Subtarget->is64Bit() && isVarArg && !Subtarget->isTargetWin64())
1681 MIB.addReg(X86::AL);
1683 // Add implicit physical register uses to the call.
1684 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
1685 MIB.addReg(RegArgs[i]);
1687 // Issue CALLSEQ_END
1688 unsigned AdjStackUp = TM.getRegisterInfo()->getCallFrameDestroyOpcode();
1689 unsigned NumBytesCallee = 0;
1690 if (!Subtarget->is64Bit() && CS.paramHasAttr(1, Attribute::StructRet))
1692 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(AdjStackUp))
1693 .addImm(NumBytes).addImm(NumBytesCallee);
1695 // Build info for return calling conv lowering code.
1696 // FIXME: This is practically a copy-paste from TargetLowering::LowerCallTo.
1697 SmallVector<ISD::InputArg, 32> Ins;
1698 SmallVector<EVT, 4> RetTys;
1699 ComputeValueVTs(TLI, I->getType(), RetTys);
1700 for (unsigned i = 0, e = RetTys.size(); i != e; ++i) {
1702 EVT RegisterVT = TLI.getRegisterType(I->getParent()->getContext(), VT);
1703 unsigned NumRegs = TLI.getNumRegisters(I->getParent()->getContext(), VT);
1704 for (unsigned j = 0; j != NumRegs; ++j) {
1705 ISD::InputArg MyFlags;
1706 MyFlags.VT = RegisterVT.getSimpleVT();
1707 MyFlags.Used = !CS.getInstruction()->use_empty();
1708 if (CS.paramHasAttr(0, Attribute::SExt))
1709 MyFlags.Flags.setSExt();
1710 if (CS.paramHasAttr(0, Attribute::ZExt))
1711 MyFlags.Flags.setZExt();
1712 if (CS.paramHasAttr(0, Attribute::InReg))
1713 MyFlags.Flags.setInReg();
1714 Ins.push_back(MyFlags);
1718 // Now handle call return values.
1719 SmallVector<unsigned, 4> UsedRegs;
1720 SmallVector<CCValAssign, 16> RVLocs;
1721 CCState CCRetInfo(CC, false, TM, RVLocs, I->getParent()->getContext());
1722 unsigned ResultReg = FuncInfo.CreateRegs(I->getType());
1723 CCRetInfo.AnalyzeCallResult(Ins, RetCC_X86);
1724 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1725 EVT CopyVT = RVLocs[i].getValVT();
1726 unsigned CopyReg = ResultReg + i;
1728 // If this is a call to a function that returns an fp value on the x87 fp
1729 // stack, but where we prefer to use the value in xmm registers, copy it
1730 // out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1731 if ((RVLocs[i].getLocReg() == X86::ST0 ||
1732 RVLocs[i].getLocReg() == X86::ST1) &&
1733 isScalarFPTypeInSSEReg(RVLocs[0].getValVT())) {
1735 CopyReg = createResultReg(X86::RFP80RegisterClass);
1738 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1739 CopyReg).addReg(RVLocs[i].getLocReg());
1740 UsedRegs.push_back(RVLocs[i].getLocReg());
1742 if (CopyVT != RVLocs[i].getValVT()) {
1743 // Round the F80 the right size, which also moves to the appropriate xmm
1744 // register. This is accomplished by storing the F80 value in memory and
1745 // then loading it back. Ewww...
1746 EVT ResVT = RVLocs[i].getValVT();
1747 unsigned Opc = ResVT == MVT::f32 ? X86::ST_Fp80m32 : X86::ST_Fp80m64;
1748 unsigned MemSize = ResVT.getSizeInBits()/8;
1749 int FI = MFI.CreateStackObject(MemSize, MemSize, false);
1750 addFrameReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1753 Opc = ResVT == MVT::f32 ? X86::MOVSSrm : X86::MOVSDrm;
1754 addFrameReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1755 TII.get(Opc), ResultReg + i), FI);
1760 UpdateValueMap(I, ResultReg, RVLocs.size());
1762 // Set all unused physreg defs as dead.
1763 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
1770 X86FastISel::TargetSelectInstruction(const Instruction *I) {
1771 switch (I->getOpcode()) {
1773 case Instruction::Load:
1774 return X86SelectLoad(I);
1775 case Instruction::Store:
1776 return X86SelectStore(I);
1777 case Instruction::Ret:
1778 return X86SelectRet(I);
1779 case Instruction::ICmp:
1780 case Instruction::FCmp:
1781 return X86SelectCmp(I);
1782 case Instruction::ZExt:
1783 return X86SelectZExt(I);
1784 case Instruction::Br:
1785 return X86SelectBranch(I);
1786 case Instruction::Call:
1787 return X86SelectCall(I);
1788 case Instruction::LShr:
1789 case Instruction::AShr:
1790 case Instruction::Shl:
1791 return X86SelectShift(I);
1792 case Instruction::Select:
1793 return X86SelectSelect(I);
1794 case Instruction::Trunc:
1795 return X86SelectTrunc(I);
1796 case Instruction::FPExt:
1797 return X86SelectFPExt(I);
1798 case Instruction::FPTrunc:
1799 return X86SelectFPTrunc(I);
1800 case Instruction::IntToPtr: // Deliberate fall-through.
1801 case Instruction::PtrToInt: {
1802 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
1803 EVT DstVT = TLI.getValueType(I->getType());
1804 if (DstVT.bitsGT(SrcVT))
1805 return X86SelectZExt(I);
1806 if (DstVT.bitsLT(SrcVT))
1807 return X86SelectTrunc(I);
1808 unsigned Reg = getRegForValue(I->getOperand(0));
1809 if (Reg == 0) return false;
1810 UpdateValueMap(I, Reg);
1818 unsigned X86FastISel::TargetMaterializeConstant(const Constant *C) {
1820 if (!isTypeLegal(C->getType(), VT))
1823 // Get opcode and regclass of the output for the given load instruction.
1825 const TargetRegisterClass *RC = NULL;
1826 switch (VT.SimpleTy) {
1827 default: return false;
1830 RC = X86::GR8RegisterClass;
1834 RC = X86::GR16RegisterClass;
1838 RC = X86::GR32RegisterClass;
1841 // Must be in x86-64 mode.
1843 RC = X86::GR64RegisterClass;
1846 if (Subtarget->hasSSE1()) {
1848 RC = X86::FR32RegisterClass;
1850 Opc = X86::LD_Fp32m;
1851 RC = X86::RFP32RegisterClass;
1855 if (Subtarget->hasSSE2()) {
1857 RC = X86::FR64RegisterClass;
1859 Opc = X86::LD_Fp64m;
1860 RC = X86::RFP64RegisterClass;
1864 // No f80 support yet.
1868 // Materialize addresses with LEA instructions.
1869 if (isa<GlobalValue>(C)) {
1871 if (X86SelectAddress(C, AM)) {
1872 // If the expression is just a basereg, then we're done, otherwise we need
1874 if (AM.BaseType == X86AddressMode::RegBase &&
1875 AM.IndexReg == 0 && AM.Disp == 0 && AM.GV == 0)
1878 Opc = TLI.getPointerTy() == MVT::i32 ? X86::LEA32r : X86::LEA64r;
1879 unsigned ResultReg = createResultReg(RC);
1880 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1881 TII.get(Opc), ResultReg), AM);
1887 // MachineConstantPool wants an explicit alignment.
1888 unsigned Align = TD.getPrefTypeAlignment(C->getType());
1890 // Alignment of vector types. FIXME!
1891 Align = TD.getTypeAllocSize(C->getType());
1894 // x86-32 PIC requires a PIC base register for constant pools.
1895 unsigned PICBase = 0;
1896 unsigned char OpFlag = 0;
1897 if (Subtarget->isPICStyleStubPIC()) { // Not dynamic-no-pic
1898 OpFlag = X86II::MO_PIC_BASE_OFFSET;
1899 PICBase = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF);
1900 } else if (Subtarget->isPICStyleGOT()) {
1901 OpFlag = X86II::MO_GOTOFF;
1902 PICBase = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF);
1903 } else if (Subtarget->isPICStyleRIPRel() &&
1904 TM.getCodeModel() == CodeModel::Small) {
1908 // Create the load from the constant pool.
1909 unsigned MCPOffset = MCP.getConstantPoolIndex(C, Align);
1910 unsigned ResultReg = createResultReg(RC);
1911 addConstantPoolReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1912 TII.get(Opc), ResultReg),
1913 MCPOffset, PICBase, OpFlag);
1918 unsigned X86FastISel::TargetMaterializeAlloca(const AllocaInst *C) {
1919 // Fail on dynamic allocas. At this point, getRegForValue has already
1920 // checked its CSE maps, so if we're here trying to handle a dynamic
1921 // alloca, we're not going to succeed. X86SelectAddress has a
1922 // check for dynamic allocas, because it's called directly from
1923 // various places, but TargetMaterializeAlloca also needs a check
1924 // in order to avoid recursion between getRegForValue,
1925 // X86SelectAddrss, and TargetMaterializeAlloca.
1926 if (!FuncInfo.StaticAllocaMap.count(C))
1930 if (!X86SelectAddress(C, AM))
1932 unsigned Opc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
1933 TargetRegisterClass* RC = TLI.getRegClassFor(TLI.getPointerTy());
1934 unsigned ResultReg = createResultReg(RC);
1935 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1936 TII.get(Opc), ResultReg), AM);
1940 unsigned X86FastISel::TargetMaterializeFloatZero(const ConstantFP *CF) {
1942 if (!isTypeLegal(CF->getType(), VT))
1945 // Get opcode and regclass for the given zero.
1947 const TargetRegisterClass *RC = NULL;
1948 switch (VT.SimpleTy) {
1949 default: return false;
1951 if (Subtarget->hasSSE1()) {
1952 Opc = X86::FsFLD0SS;
1953 RC = X86::FR32RegisterClass;
1955 Opc = X86::LD_Fp032;
1956 RC = X86::RFP32RegisterClass;
1960 if (Subtarget->hasSSE2()) {
1961 Opc = X86::FsFLD0SD;
1962 RC = X86::FR64RegisterClass;
1964 Opc = X86::LD_Fp064;
1965 RC = X86::RFP64RegisterClass;
1969 // No f80 support yet.
1973 unsigned ResultReg = createResultReg(RC);
1974 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), ResultReg);
1979 /// TryToFoldLoad - The specified machine instr operand is a vreg, and that
1980 /// vreg is being provided by the specified load instruction. If possible,
1981 /// try to fold the load as an operand to the instruction, returning true if
1983 bool X86FastISel::TryToFoldLoad(MachineInstr *MI, unsigned OpNo,
1984 const LoadInst *LI) {
1986 if (!X86SelectAddress(LI->getOperand(0), AM))
1989 X86InstrInfo &XII = (X86InstrInfo&)TII;
1991 unsigned Size = TD.getTypeAllocSize(LI->getType());
1992 unsigned Alignment = LI->getAlignment();
1994 SmallVector<MachineOperand, 8> AddrOps;
1995 AM.getFullAddress(AddrOps);
1997 MachineInstr *Result =
1998 XII.foldMemoryOperandImpl(*FuncInfo.MF, MI, OpNo, AddrOps, Size, Alignment);
1999 if (Result == 0) return false;
2001 FuncInfo.MBB->insert(FuncInfo.InsertPt, Result);
2002 MI->eraseFromParent();
2008 llvm::FastISel *X86::createFastISel(FunctionLoweringInfo &funcInfo) {
2009 return new X86FastISel(funcInfo);