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 "X86CallingConv.h"
18 #include "X86InstrBuilder.h"
19 #include "X86MachineFunctionInfo.h"
20 #include "X86RegisterInfo.h"
21 #include "X86Subtarget.h"
22 #include "X86TargetMachine.h"
23 #include "llvm/CodeGen/Analysis.h"
24 #include "llvm/CodeGen/FastISel.h"
25 #include "llvm/CodeGen/FunctionLoweringInfo.h"
26 #include "llvm/CodeGen/MachineConstantPool.h"
27 #include "llvm/CodeGen/MachineFrameInfo.h"
28 #include "llvm/CodeGen/MachineRegisterInfo.h"
29 #include "llvm/IR/CallSite.h"
30 #include "llvm/IR/CallingConv.h"
31 #include "llvm/IR/DerivedTypes.h"
32 #include "llvm/IR/GetElementPtrTypeIterator.h"
33 #include "llvm/IR/GlobalAlias.h"
34 #include "llvm/IR/GlobalVariable.h"
35 #include "llvm/IR/Instructions.h"
36 #include "llvm/IR/IntrinsicInst.h"
37 #include "llvm/IR/Operator.h"
38 #include "llvm/Support/ErrorHandling.h"
39 #include "llvm/Target/TargetOptions.h"
44 class X86FastISel final : public FastISel {
45 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
46 /// make the right decision when generating code for different targets.
47 const X86Subtarget *Subtarget;
49 /// X86ScalarSSEf32, X86ScalarSSEf64 - Select between SSE or x87
50 /// floating point ops.
51 /// When SSE is available, use it for f32 operations.
52 /// When SSE2 is available, use it for f64 operations.
57 explicit X86FastISel(FunctionLoweringInfo &funcInfo,
58 const TargetLibraryInfo *libInfo)
59 : FastISel(funcInfo, libInfo) {
60 Subtarget = &TM.getSubtarget<X86Subtarget>();
61 X86ScalarSSEf64 = Subtarget->hasSSE2();
62 X86ScalarSSEf32 = Subtarget->hasSSE1();
65 bool TargetSelectInstruction(const Instruction *I) override;
67 /// \brief 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 bool tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo,
72 const LoadInst *LI) override;
74 bool FastLowerArguments() override;
76 #include "X86GenFastISel.inc"
79 bool X86FastEmitCompare(const Value *LHS, const Value *RHS, EVT VT);
81 bool X86FastEmitLoad(EVT VT, const X86AddressMode &AM, MachineMemOperand *MMO,
84 bool X86FastEmitStore(EVT VT, const Value *Val, const X86AddressMode &AM,
85 MachineMemOperand *MMO = nullptr, bool Aligned = false);
86 bool X86FastEmitStore(EVT VT, unsigned ValReg, bool ValIsKill,
87 const X86AddressMode &AM,
88 MachineMemOperand *MMO = nullptr, bool Aligned = false);
90 bool X86FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src, EVT SrcVT,
93 bool X86SelectAddress(const Value *V, X86AddressMode &AM);
94 bool X86SelectCallAddress(const Value *V, X86AddressMode &AM);
96 bool X86SelectLoad(const Instruction *I);
98 bool X86SelectStore(const Instruction *I);
100 bool X86SelectRet(const Instruction *I);
102 bool X86SelectCmp(const Instruction *I);
104 bool X86SelectZExt(const Instruction *I);
106 bool X86SelectBranch(const Instruction *I);
108 bool X86SelectShift(const Instruction *I);
110 bool X86SelectDivRem(const Instruction *I);
112 bool X86SelectSelect(const Instruction *I);
114 bool X86SelectTrunc(const Instruction *I);
116 bool X86SelectFPExt(const Instruction *I);
117 bool X86SelectFPTrunc(const Instruction *I);
119 bool X86VisitIntrinsicCall(const IntrinsicInst &I);
120 bool X86SelectCall(const Instruction *I);
122 bool DoSelectCall(const Instruction *I, const char *MemIntName);
124 const X86InstrInfo *getInstrInfo() const {
125 return getTargetMachine()->getInstrInfo();
127 const X86TargetMachine *getTargetMachine() const {
128 return static_cast<const X86TargetMachine *>(&TM);
131 bool handleConstantAddresses(const Value *V, X86AddressMode &AM);
133 unsigned TargetMaterializeConstant(const Constant *C) override;
135 unsigned TargetMaterializeAlloca(const AllocaInst *C) override;
137 unsigned TargetMaterializeFloatZero(const ConstantFP *CF) override;
139 /// isScalarFPTypeInSSEReg - Return true if the specified scalar FP type is
140 /// computed in an SSE register, not on the X87 floating point stack.
141 bool isScalarFPTypeInSSEReg(EVT VT) const {
142 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
143 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1
146 bool isTypeLegal(Type *Ty, MVT &VT, bool AllowI1 = false);
148 bool IsMemcpySmall(uint64_t Len);
150 bool TryEmitSmallMemcpy(X86AddressMode DestAM,
151 X86AddressMode SrcAM, uint64_t Len);
154 } // end anonymous namespace.
156 bool X86FastISel::isTypeLegal(Type *Ty, MVT &VT, bool AllowI1) {
157 EVT evt = TLI.getValueType(Ty, /*HandleUnknown=*/true);
158 if (evt == MVT::Other || !evt.isSimple())
159 // Unhandled type. Halt "fast" selection and bail.
162 VT = evt.getSimpleVT();
163 // For now, require SSE/SSE2 for performing floating-point operations,
164 // since x87 requires additional work.
165 if (VT == MVT::f64 && !X86ScalarSSEf64)
167 if (VT == MVT::f32 && !X86ScalarSSEf32)
169 // Similarly, no f80 support yet.
172 // We only handle legal types. For example, on x86-32 the instruction
173 // selector contains all of the 64-bit instructions from x86-64,
174 // under the assumption that i64 won't be used if the target doesn't
176 return (AllowI1 && VT == MVT::i1) || TLI.isTypeLegal(VT);
179 #include "X86GenCallingConv.inc"
181 /// X86FastEmitLoad - Emit a machine instruction to load a value of type VT.
182 /// The address is either pre-computed, i.e. Ptr, or a GlobalAddress, i.e. GV.
183 /// Return true and the result register by reference if it is possible.
184 bool X86FastISel::X86FastEmitLoad(EVT VT, const X86AddressMode &AM,
185 MachineMemOperand *MMO, unsigned &ResultReg) {
186 // Get opcode and regclass of the output for the given load instruction.
188 const TargetRegisterClass *RC = nullptr;
189 switch (VT.getSimpleVT().SimpleTy) {
190 default: return false;
194 RC = &X86::GR8RegClass;
198 RC = &X86::GR16RegClass;
202 RC = &X86::GR32RegClass;
205 // Must be in x86-64 mode.
207 RC = &X86::GR64RegClass;
210 if (X86ScalarSSEf32) {
211 Opc = Subtarget->hasAVX() ? X86::VMOVSSrm : X86::MOVSSrm;
212 RC = &X86::FR32RegClass;
215 RC = &X86::RFP32RegClass;
219 if (X86ScalarSSEf64) {
220 Opc = Subtarget->hasAVX() ? X86::VMOVSDrm : X86::MOVSDrm;
221 RC = &X86::FR64RegClass;
224 RC = &X86::RFP64RegClass;
228 // No f80 support yet.
232 ResultReg = createResultReg(RC);
233 MachineInstrBuilder MIB =
234 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg);
235 addFullAddress(MIB, AM);
237 MIB->addMemOperand(*FuncInfo.MF, MMO);
241 /// X86FastEmitStore - Emit a machine instruction to store a value Val of
242 /// type VT. The address is either pre-computed, consisted of a base ptr, Ptr
243 /// and a displacement offset, or a GlobalAddress,
244 /// i.e. V. Return true if it is possible.
245 bool X86FastISel::X86FastEmitStore(EVT VT, unsigned ValReg, bool ValIsKill,
246 const X86AddressMode &AM,
247 MachineMemOperand *MMO, bool Aligned) {
248 // Get opcode and regclass of the output for the given store instruction.
250 switch (VT.getSimpleVT().SimpleTy) {
251 case MVT::f80: // No f80 support yet.
252 default: return false;
254 // Mask out all but lowest bit.
255 unsigned AndResult = createResultReg(&X86::GR8RegClass);
256 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
257 TII.get(X86::AND8ri), AndResult)
258 .addReg(ValReg, getKillRegState(ValIsKill)).addImm(1);
261 // FALLTHROUGH, handling i1 as i8.
262 case MVT::i8: Opc = X86::MOV8mr; break;
263 case MVT::i16: Opc = X86::MOV16mr; break;
264 case MVT::i32: Opc = X86::MOV32mr; break;
265 case MVT::i64: Opc = X86::MOV64mr; break; // Must be in x86-64 mode.
267 Opc = X86ScalarSSEf32 ?
268 (Subtarget->hasAVX() ? X86::VMOVSSmr : X86::MOVSSmr) : X86::ST_Fp32m;
271 Opc = X86ScalarSSEf64 ?
272 (Subtarget->hasAVX() ? X86::VMOVSDmr : X86::MOVSDmr) : X86::ST_Fp64m;
276 Opc = Subtarget->hasAVX() ? X86::VMOVAPSmr : X86::MOVAPSmr;
278 Opc = Subtarget->hasAVX() ? X86::VMOVUPSmr : X86::MOVUPSmr;
282 Opc = Subtarget->hasAVX() ? X86::VMOVAPDmr : X86::MOVAPDmr;
284 Opc = Subtarget->hasAVX() ? X86::VMOVUPDmr : X86::MOVUPDmr;
291 Opc = Subtarget->hasAVX() ? X86::VMOVDQAmr : X86::MOVDQAmr;
293 Opc = Subtarget->hasAVX() ? X86::VMOVDQUmr : X86::MOVDQUmr;
297 MachineInstrBuilder MIB =
298 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc));
299 addFullAddress(MIB, AM).addReg(ValReg, getKillRegState(ValIsKill));
301 MIB->addMemOperand(*FuncInfo.MF, MMO);
306 bool X86FastISel::X86FastEmitStore(EVT VT, const Value *Val,
307 const X86AddressMode &AM,
308 MachineMemOperand *MMO, bool Aligned) {
309 // Handle 'null' like i32/i64 0.
310 if (isa<ConstantPointerNull>(Val))
311 Val = Constant::getNullValue(DL.getIntPtrType(Val->getContext()));
313 // If this is a store of a simple constant, fold the constant into the store.
314 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
317 switch (VT.getSimpleVT().SimpleTy) {
319 case MVT::i1: Signed = false; // FALLTHROUGH to handle as i8.
320 case MVT::i8: Opc = X86::MOV8mi; break;
321 case MVT::i16: Opc = X86::MOV16mi; break;
322 case MVT::i32: Opc = X86::MOV32mi; break;
324 // Must be a 32-bit sign extended value.
325 if (isInt<32>(CI->getSExtValue()))
326 Opc = X86::MOV64mi32;
331 MachineInstrBuilder MIB =
332 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc));
333 addFullAddress(MIB, AM).addImm(Signed ? (uint64_t) CI->getSExtValue()
334 : CI->getZExtValue());
336 MIB->addMemOperand(*FuncInfo.MF, MMO);
341 unsigned ValReg = getRegForValue(Val);
345 bool ValKill = hasTrivialKill(Val);
346 return X86FastEmitStore(VT, ValReg, ValKill, AM, MMO, Aligned);
349 /// X86FastEmitExtend - Emit a machine instruction to extend a value Src of
350 /// type SrcVT to type DstVT using the specified extension opcode Opc (e.g.
351 /// ISD::SIGN_EXTEND).
352 bool X86FastISel::X86FastEmitExtend(ISD::NodeType Opc, EVT DstVT,
353 unsigned Src, EVT SrcVT,
354 unsigned &ResultReg) {
355 unsigned RR = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), Opc,
356 Src, /*TODO: Kill=*/false);
364 bool X86FastISel::handleConstantAddresses(const Value *V, X86AddressMode &AM) {
365 // Handle constant address.
366 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
367 // Can't handle alternate code models yet.
368 if (TM.getCodeModel() != CodeModel::Small)
371 // Can't handle TLS yet.
372 if (GV->isThreadLocal())
375 // RIP-relative addresses can't have additional register operands, so if
376 // we've already folded stuff into the addressing mode, just force the
377 // global value into its own register, which we can use as the basereg.
378 if (!Subtarget->isPICStyleRIPRel() ||
379 (AM.Base.Reg == 0 && AM.IndexReg == 0)) {
380 // Okay, we've committed to selecting this global. Set up the address.
383 // Allow the subtarget to classify the global.
384 unsigned char GVFlags = Subtarget->ClassifyGlobalReference(GV, TM);
386 // If this reference is relative to the pic base, set it now.
387 if (isGlobalRelativeToPICBase(GVFlags)) {
388 // FIXME: How do we know Base.Reg is free??
389 AM.Base.Reg = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF);
392 // Unless the ABI requires an extra load, return a direct reference to
394 if (!isGlobalStubReference(GVFlags)) {
395 if (Subtarget->isPICStyleRIPRel()) {
396 // Use rip-relative addressing if we can. Above we verified that the
397 // base and index registers are unused.
398 assert(AM.Base.Reg == 0 && AM.IndexReg == 0);
399 AM.Base.Reg = X86::RIP;
401 AM.GVOpFlags = GVFlags;
405 // Ok, we need to do a load from a stub. If we've already loaded from
406 // this stub, reuse the loaded pointer, otherwise emit the load now.
407 DenseMap<const Value*, unsigned>::iterator I = LocalValueMap.find(V);
409 if (I != LocalValueMap.end() && I->second != 0) {
412 // Issue load from stub.
414 const TargetRegisterClass *RC = nullptr;
415 X86AddressMode StubAM;
416 StubAM.Base.Reg = AM.Base.Reg;
418 StubAM.GVOpFlags = GVFlags;
420 // Prepare for inserting code in the local-value area.
421 SavePoint SaveInsertPt = enterLocalValueArea();
423 if (TLI.getPointerTy() == MVT::i64) {
425 RC = &X86::GR64RegClass;
427 if (Subtarget->isPICStyleRIPRel())
428 StubAM.Base.Reg = X86::RIP;
431 RC = &X86::GR32RegClass;
434 LoadReg = createResultReg(RC);
435 MachineInstrBuilder LoadMI =
436 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), LoadReg);
437 addFullAddress(LoadMI, StubAM);
439 // Ok, back to normal mode.
440 leaveLocalValueArea(SaveInsertPt);
442 // Prevent loading GV stub multiple times in same MBB.
443 LocalValueMap[V] = LoadReg;
446 // Now construct the final address. Note that the Disp, Scale,
447 // and Index values may already be set here.
448 AM.Base.Reg = LoadReg;
454 // If all else fails, try to materialize the value in a register.
455 if (!AM.GV || !Subtarget->isPICStyleRIPRel()) {
456 if (AM.Base.Reg == 0) {
457 AM.Base.Reg = getRegForValue(V);
458 return AM.Base.Reg != 0;
460 if (AM.IndexReg == 0) {
461 assert(AM.Scale == 1 && "Scale with no index!");
462 AM.IndexReg = getRegForValue(V);
463 return AM.IndexReg != 0;
470 /// X86SelectAddress - Attempt to fill in an address from the given value.
472 bool X86FastISel::X86SelectAddress(const Value *V, X86AddressMode &AM) {
473 SmallVector<const Value *, 32> GEPs;
475 const User *U = nullptr;
476 unsigned Opcode = Instruction::UserOp1;
477 if (const Instruction *I = dyn_cast<Instruction>(V)) {
478 // Don't walk into other basic blocks; it's possible we haven't
479 // visited them yet, so the instructions may not yet be assigned
480 // virtual registers.
481 if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(V)) ||
482 FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
483 Opcode = I->getOpcode();
486 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(V)) {
487 Opcode = C->getOpcode();
491 if (PointerType *Ty = dyn_cast<PointerType>(V->getType()))
492 if (Ty->getAddressSpace() > 255)
493 // Fast instruction selection doesn't support the special
499 case Instruction::BitCast:
500 // Look past bitcasts.
501 return X86SelectAddress(U->getOperand(0), AM);
503 case Instruction::IntToPtr:
504 // Look past no-op inttoptrs.
505 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
506 return X86SelectAddress(U->getOperand(0), AM);
509 case Instruction::PtrToInt:
510 // Look past no-op ptrtoints.
511 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
512 return X86SelectAddress(U->getOperand(0), AM);
515 case Instruction::Alloca: {
516 // Do static allocas.
517 const AllocaInst *A = cast<AllocaInst>(V);
518 DenseMap<const AllocaInst*, int>::iterator SI =
519 FuncInfo.StaticAllocaMap.find(A);
520 if (SI != FuncInfo.StaticAllocaMap.end()) {
521 AM.BaseType = X86AddressMode::FrameIndexBase;
522 AM.Base.FrameIndex = SI->second;
528 case Instruction::Add: {
529 // Adds of constants are common and easy enough.
530 if (const ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
531 uint64_t Disp = (int32_t)AM.Disp + (uint64_t)CI->getSExtValue();
532 // They have to fit in the 32-bit signed displacement field though.
533 if (isInt<32>(Disp)) {
534 AM.Disp = (uint32_t)Disp;
535 return X86SelectAddress(U->getOperand(0), AM);
541 case Instruction::GetElementPtr: {
542 X86AddressMode SavedAM = AM;
544 // Pattern-match simple GEPs.
545 uint64_t Disp = (int32_t)AM.Disp;
546 unsigned IndexReg = AM.IndexReg;
547 unsigned Scale = AM.Scale;
548 gep_type_iterator GTI = gep_type_begin(U);
549 // Iterate through the indices, folding what we can. Constants can be
550 // folded, and one dynamic index can be handled, if the scale is supported.
551 for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end();
552 i != e; ++i, ++GTI) {
553 const Value *Op = *i;
554 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
555 const StructLayout *SL = DL.getStructLayout(STy);
556 Disp += SL->getElementOffset(cast<ConstantInt>(Op)->getZExtValue());
560 // A array/variable index is always of the form i*S where S is the
561 // constant scale size. See if we can push the scale into immediates.
562 uint64_t S = DL.getTypeAllocSize(GTI.getIndexedType());
564 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
565 // Constant-offset addressing.
566 Disp += CI->getSExtValue() * S;
569 if (canFoldAddIntoGEP(U, Op)) {
570 // A compatible add with a constant operand. Fold the constant.
572 cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
573 Disp += CI->getSExtValue() * S;
574 // Iterate on the other operand.
575 Op = cast<AddOperator>(Op)->getOperand(0);
579 (!AM.GV || !Subtarget->isPICStyleRIPRel()) &&
580 (S == 1 || S == 2 || S == 4 || S == 8)) {
581 // Scaled-index addressing.
583 IndexReg = getRegForGEPIndex(Op).first;
589 goto unsupported_gep;
593 // Check for displacement overflow.
594 if (!isInt<32>(Disp))
597 AM.IndexReg = IndexReg;
599 AM.Disp = (uint32_t)Disp;
602 if (const GetElementPtrInst *GEP =
603 dyn_cast<GetElementPtrInst>(U->getOperand(0))) {
604 // Ok, the GEP indices were covered by constant-offset and scaled-index
605 // addressing. Update the address state and move on to examining the base.
608 } else if (X86SelectAddress(U->getOperand(0), AM)) {
612 // If we couldn't merge the gep value into this addr mode, revert back to
613 // our address and just match the value instead of completely failing.
616 for (SmallVectorImpl<const Value *>::reverse_iterator
617 I = GEPs.rbegin(), E = GEPs.rend(); I != E; ++I)
618 if (handleConstantAddresses(*I, AM))
623 // Ok, the GEP indices weren't all covered.
628 return handleConstantAddresses(V, AM);
631 /// X86SelectCallAddress - Attempt to fill in an address from the given value.
633 bool X86FastISel::X86SelectCallAddress(const Value *V, X86AddressMode &AM) {
634 const User *U = nullptr;
635 unsigned Opcode = Instruction::UserOp1;
636 const Instruction *I = dyn_cast<Instruction>(V);
637 // Record if the value is defined in the same basic block.
639 // This information is crucial to know whether or not folding an
641 // Indeed, FastISel generates or reuses a virtual register for all
642 // operands of all instructions it selects. Obviously, the definition and
643 // its uses must use the same virtual register otherwise the produced
644 // code is incorrect.
645 // Before instruction selection, FunctionLoweringInfo::set sets the virtual
646 // registers for values that are alive across basic blocks. This ensures
647 // that the values are consistently set between across basic block, even
648 // if different instruction selection mechanisms are used (e.g., a mix of
649 // SDISel and FastISel).
650 // For values local to a basic block, the instruction selection process
651 // generates these virtual registers with whatever method is appropriate
652 // for its needs. In particular, FastISel and SDISel do not share the way
653 // local virtual registers are set.
654 // Therefore, this is impossible (or at least unsafe) to share values
655 // between basic blocks unless they use the same instruction selection
656 // method, which is not guarantee for X86.
657 // Moreover, things like hasOneUse could not be used accurately, if we
658 // allow to reference values across basic blocks whereas they are not
659 // alive across basic blocks initially.
662 Opcode = I->getOpcode();
664 InMBB = I->getParent() == FuncInfo.MBB->getBasicBlock();
665 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(V)) {
666 Opcode = C->getOpcode();
672 case Instruction::BitCast:
673 // Look past bitcasts if its operand is in the same BB.
675 return X86SelectCallAddress(U->getOperand(0), AM);
678 case Instruction::IntToPtr:
679 // Look past no-op inttoptrs if its operand is in the same BB.
681 TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
682 return X86SelectCallAddress(U->getOperand(0), AM);
685 case Instruction::PtrToInt:
686 // Look past no-op ptrtoints if its operand is in the same BB.
688 TLI.getValueType(U->getType()) == TLI.getPointerTy())
689 return X86SelectCallAddress(U->getOperand(0), AM);
693 // Handle constant address.
694 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
695 // Can't handle alternate code models yet.
696 if (TM.getCodeModel() != CodeModel::Small)
699 // RIP-relative addresses can't have additional register operands.
700 if (Subtarget->isPICStyleRIPRel() &&
701 (AM.Base.Reg != 0 || AM.IndexReg != 0))
704 // Can't handle DbgLocLImport.
705 if (GV->hasDLLImportStorageClass())
709 if (const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV))
710 if (GVar->isThreadLocal())
713 // Okay, we've committed to selecting this global. Set up the basic address.
716 // No ABI requires an extra load for anything other than DLLImport, which
717 // we rejected above. Return a direct reference to the global.
718 if (Subtarget->isPICStyleRIPRel()) {
719 // Use rip-relative addressing if we can. Above we verified that the
720 // base and index registers are unused.
721 assert(AM.Base.Reg == 0 && AM.IndexReg == 0);
722 AM.Base.Reg = X86::RIP;
723 } else if (Subtarget->isPICStyleStubPIC()) {
724 AM.GVOpFlags = X86II::MO_PIC_BASE_OFFSET;
725 } else if (Subtarget->isPICStyleGOT()) {
726 AM.GVOpFlags = X86II::MO_GOTOFF;
732 // If all else fails, try to materialize the value in a register.
733 if (!AM.GV || !Subtarget->isPICStyleRIPRel()) {
734 if (AM.Base.Reg == 0) {
735 AM.Base.Reg = getRegForValue(V);
736 return AM.Base.Reg != 0;
738 if (AM.IndexReg == 0) {
739 assert(AM.Scale == 1 && "Scale with no index!");
740 AM.IndexReg = getRegForValue(V);
741 return AM.IndexReg != 0;
749 /// X86SelectStore - Select and emit code to implement store instructions.
750 bool X86FastISel::X86SelectStore(const Instruction *I) {
751 // Atomic stores need special handling.
752 const StoreInst *S = cast<StoreInst>(I);
757 const Value *Val = S->getValueOperand();
758 const Value *Ptr = S->getPointerOperand();
761 if (!isTypeLegal(Val->getType(), VT, /*AllowI1=*/true))
764 unsigned Alignment = S->getAlignment();
765 unsigned ABIAlignment = DL.getABITypeAlignment(Val->getType());
766 if (Alignment == 0) // Ensure that codegen never sees alignment 0
767 Alignment = ABIAlignment;
768 bool Aligned = Alignment >= ABIAlignment;
771 if (!X86SelectAddress(Ptr, AM))
774 return X86FastEmitStore(VT, Val, AM, createMachineMemOperandFor(I), Aligned);
777 /// X86SelectRet - Select and emit code to implement ret instructions.
778 bool X86FastISel::X86SelectRet(const Instruction *I) {
779 const ReturnInst *Ret = cast<ReturnInst>(I);
780 const Function &F = *I->getParent()->getParent();
781 const X86MachineFunctionInfo *X86MFInfo =
782 FuncInfo.MF->getInfo<X86MachineFunctionInfo>();
784 if (!FuncInfo.CanLowerReturn)
787 CallingConv::ID CC = F.getCallingConv();
788 if (CC != CallingConv::C &&
789 CC != CallingConv::Fast &&
790 CC != CallingConv::X86_FastCall &&
791 CC != CallingConv::X86_64_SysV)
794 if (Subtarget->isCallingConvWin64(CC))
797 // Don't handle popping bytes on return for now.
798 if (X86MFInfo->getBytesToPopOnReturn() != 0)
801 // fastcc with -tailcallopt is intended to provide a guaranteed
802 // tail call optimization. Fastisel doesn't know how to do that.
803 if (CC == CallingConv::Fast && TM.Options.GuaranteedTailCallOpt)
806 // Let SDISel handle vararg functions.
810 // Build a list of return value registers.
811 SmallVector<unsigned, 4> RetRegs;
813 if (Ret->getNumOperands() > 0) {
814 SmallVector<ISD::OutputArg, 4> Outs;
815 GetReturnInfo(F.getReturnType(), F.getAttributes(), Outs, TLI);
817 // Analyze operands of the call, assigning locations to each operand.
818 SmallVector<CCValAssign, 16> ValLocs;
819 CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, TM, ValLocs,
821 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
823 const Value *RV = Ret->getOperand(0);
824 unsigned Reg = getRegForValue(RV);
828 // Only handle a single return value for now.
829 if (ValLocs.size() != 1)
832 CCValAssign &VA = ValLocs[0];
834 // Don't bother handling odd stuff for now.
835 if (VA.getLocInfo() != CCValAssign::Full)
837 // Only handle register returns for now.
841 // The calling-convention tables for x87 returns don't tell
843 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
846 unsigned SrcReg = Reg + VA.getValNo();
847 EVT SrcVT = TLI.getValueType(RV->getType());
848 EVT DstVT = VA.getValVT();
849 // Special handling for extended integers.
850 if (SrcVT != DstVT) {
851 if (SrcVT != MVT::i1 && SrcVT != MVT::i8 && SrcVT != MVT::i16)
854 if (!Outs[0].Flags.isZExt() && !Outs[0].Flags.isSExt())
857 assert(DstVT == MVT::i32 && "X86 should always ext to i32");
859 if (SrcVT == MVT::i1) {
860 if (Outs[0].Flags.isSExt())
862 SrcReg = FastEmitZExtFromI1(MVT::i8, SrcReg, /*TODO: Kill=*/false);
865 unsigned Op = Outs[0].Flags.isZExt() ? ISD::ZERO_EXTEND :
867 SrcReg = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), Op,
868 SrcReg, /*TODO: Kill=*/false);
872 unsigned DstReg = VA.getLocReg();
873 const TargetRegisterClass* SrcRC = MRI.getRegClass(SrcReg);
874 // Avoid a cross-class copy. This is very unlikely.
875 if (!SrcRC->contains(DstReg))
877 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TargetOpcode::COPY),
878 DstReg).addReg(SrcReg);
880 // Add register to return instruction.
881 RetRegs.push_back(VA.getLocReg());
884 // The x86-64 ABI for returning structs by value requires that we copy
885 // the sret argument into %rax for the return. We saved the argument into
886 // a virtual register in the entry block, so now we copy the value out
887 // and into %rax. We also do the same with %eax for Win32.
888 if (F.hasStructRetAttr() &&
889 (Subtarget->is64Bit() || Subtarget->isTargetKnownWindowsMSVC())) {
890 unsigned Reg = X86MFInfo->getSRetReturnReg();
892 "SRetReturnReg should have been set in LowerFormalArguments()!");
893 unsigned RetReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
894 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TargetOpcode::COPY),
896 RetRegs.push_back(RetReg);
900 MachineInstrBuilder MIB =
901 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Subtarget->is64Bit() ? X86::RETQ : X86::RETL));
902 for (unsigned i = 0, e = RetRegs.size(); i != e; ++i)
903 MIB.addReg(RetRegs[i], RegState::Implicit);
907 /// X86SelectLoad - Select and emit code to implement load instructions.
909 bool X86FastISel::X86SelectLoad(const Instruction *I) {
910 const LoadInst *LI = cast<LoadInst>(I);
912 // Atomic loads need special handling.
917 if (!isTypeLegal(LI->getType(), VT, /*AllowI1=*/true))
920 const Value *Ptr = LI->getPointerOperand();
923 if (!X86SelectAddress(Ptr, AM))
926 unsigned ResultReg = 0;
927 if (!X86FastEmitLoad(VT, AM, createMachineMemOperandFor(LI), ResultReg))
930 UpdateValueMap(I, ResultReg);
934 static unsigned X86ChooseCmpOpcode(EVT VT, const X86Subtarget *Subtarget) {
935 bool HasAVX = Subtarget->hasAVX();
936 bool X86ScalarSSEf32 = Subtarget->hasSSE1();
937 bool X86ScalarSSEf64 = Subtarget->hasSSE2();
939 switch (VT.getSimpleVT().SimpleTy) {
941 case MVT::i8: return X86::CMP8rr;
942 case MVT::i16: return X86::CMP16rr;
943 case MVT::i32: return X86::CMP32rr;
944 case MVT::i64: return X86::CMP64rr;
946 return X86ScalarSSEf32 ? (HasAVX ? X86::VUCOMISSrr : X86::UCOMISSrr) : 0;
948 return X86ScalarSSEf64 ? (HasAVX ? X86::VUCOMISDrr : X86::UCOMISDrr) : 0;
952 /// X86ChooseCmpImmediateOpcode - If we have a comparison with RHS as the RHS
953 /// of the comparison, return an opcode that works for the compare (e.g.
954 /// CMP32ri) otherwise return 0.
955 static unsigned X86ChooseCmpImmediateOpcode(EVT VT, const ConstantInt *RHSC) {
956 switch (VT.getSimpleVT().SimpleTy) {
957 // Otherwise, we can't fold the immediate into this comparison.
959 case MVT::i8: return X86::CMP8ri;
960 case MVT::i16: return X86::CMP16ri;
961 case MVT::i32: return X86::CMP32ri;
963 // 64-bit comparisons are only valid if the immediate fits in a 32-bit sext
965 if ((int)RHSC->getSExtValue() == RHSC->getSExtValue())
966 return X86::CMP64ri32;
971 bool X86FastISel::X86FastEmitCompare(const Value *Op0, const Value *Op1,
973 unsigned Op0Reg = getRegForValue(Op0);
974 if (Op0Reg == 0) return false;
976 // Handle 'null' like i32/i64 0.
977 if (isa<ConstantPointerNull>(Op1))
978 Op1 = Constant::getNullValue(DL.getIntPtrType(Op0->getContext()));
980 // We have two options: compare with register or immediate. If the RHS of
981 // the compare is an immediate that we can fold into this compare, use
982 // CMPri, otherwise use CMPrr.
983 if (const ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
984 if (unsigned CompareImmOpc = X86ChooseCmpImmediateOpcode(VT, Op1C)) {
985 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CompareImmOpc))
987 .addImm(Op1C->getSExtValue());
992 unsigned CompareOpc = X86ChooseCmpOpcode(VT, Subtarget);
993 if (CompareOpc == 0) return false;
995 unsigned Op1Reg = getRegForValue(Op1);
996 if (Op1Reg == 0) return false;
997 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CompareOpc))
1004 bool X86FastISel::X86SelectCmp(const Instruction *I) {
1005 const CmpInst *CI = cast<CmpInst>(I);
1008 if (!isTypeLegal(I->getOperand(0)->getType(), VT))
1011 unsigned ResultReg = createResultReg(&X86::GR8RegClass);
1013 bool SwapArgs; // false -> compare Op0, Op1. true -> compare Op1, Op0.
1014 switch (CI->getPredicate()) {
1015 case CmpInst::FCMP_OEQ: {
1016 if (!X86FastEmitCompare(CI->getOperand(0), CI->getOperand(1), VT))
1019 unsigned EReg = createResultReg(&X86::GR8RegClass);
1020 unsigned NPReg = createResultReg(&X86::GR8RegClass);
1021 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::SETEr), EReg);
1022 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1023 TII.get(X86::SETNPr), NPReg);
1024 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1025 TII.get(X86::AND8rr), ResultReg).addReg(NPReg).addReg(EReg);
1026 UpdateValueMap(I, ResultReg);
1029 case CmpInst::FCMP_UNE: {
1030 if (!X86FastEmitCompare(CI->getOperand(0), CI->getOperand(1), VT))
1033 unsigned NEReg = createResultReg(&X86::GR8RegClass);
1034 unsigned PReg = createResultReg(&X86::GR8RegClass);
1035 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::SETNEr), NEReg);
1036 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::SETPr), PReg);
1037 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::OR8rr),ResultReg)
1038 .addReg(PReg).addReg(NEReg);
1039 UpdateValueMap(I, ResultReg);
1042 case CmpInst::FCMP_OGT: SwapArgs = false; SetCCOpc = X86::SETAr; break;
1043 case CmpInst::FCMP_OGE: SwapArgs = false; SetCCOpc = X86::SETAEr; break;
1044 case CmpInst::FCMP_OLT: SwapArgs = true; SetCCOpc = X86::SETAr; break;
1045 case CmpInst::FCMP_OLE: SwapArgs = true; SetCCOpc = X86::SETAEr; break;
1046 case CmpInst::FCMP_ONE: SwapArgs = false; SetCCOpc = X86::SETNEr; break;
1047 case CmpInst::FCMP_ORD: SwapArgs = false; SetCCOpc = X86::SETNPr; break;
1048 case CmpInst::FCMP_UNO: SwapArgs = false; SetCCOpc = X86::SETPr; break;
1049 case CmpInst::FCMP_UEQ: SwapArgs = false; SetCCOpc = X86::SETEr; break;
1050 case CmpInst::FCMP_UGT: SwapArgs = true; SetCCOpc = X86::SETBr; break;
1051 case CmpInst::FCMP_UGE: SwapArgs = true; SetCCOpc = X86::SETBEr; break;
1052 case CmpInst::FCMP_ULT: SwapArgs = false; SetCCOpc = X86::SETBr; break;
1053 case CmpInst::FCMP_ULE: SwapArgs = false; SetCCOpc = X86::SETBEr; break;
1055 case CmpInst::ICMP_EQ: SwapArgs = false; SetCCOpc = X86::SETEr; break;
1056 case CmpInst::ICMP_NE: SwapArgs = false; SetCCOpc = X86::SETNEr; break;
1057 case CmpInst::ICMP_UGT: SwapArgs = false; SetCCOpc = X86::SETAr; break;
1058 case CmpInst::ICMP_UGE: SwapArgs = false; SetCCOpc = X86::SETAEr; break;
1059 case CmpInst::ICMP_ULT: SwapArgs = false; SetCCOpc = X86::SETBr; break;
1060 case CmpInst::ICMP_ULE: SwapArgs = false; SetCCOpc = X86::SETBEr; break;
1061 case CmpInst::ICMP_SGT: SwapArgs = false; SetCCOpc = X86::SETGr; break;
1062 case CmpInst::ICMP_SGE: SwapArgs = false; SetCCOpc = X86::SETGEr; break;
1063 case CmpInst::ICMP_SLT: SwapArgs = false; SetCCOpc = X86::SETLr; break;
1064 case CmpInst::ICMP_SLE: SwapArgs = false; SetCCOpc = X86::SETLEr; break;
1069 const Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
1071 std::swap(Op0, Op1);
1073 // Emit a compare of Op0/Op1.
1074 if (!X86FastEmitCompare(Op0, Op1, VT))
1077 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(SetCCOpc), ResultReg);
1078 UpdateValueMap(I, ResultReg);
1082 bool X86FastISel::X86SelectZExt(const Instruction *I) {
1083 EVT DstVT = TLI.getValueType(I->getType());
1084 if (!TLI.isTypeLegal(DstVT))
1087 unsigned ResultReg = getRegForValue(I->getOperand(0));
1091 // Handle zero-extension from i1 to i8, which is common.
1092 MVT SrcVT = TLI.getSimpleValueType(I->getOperand(0)->getType());
1093 if (SrcVT.SimpleTy == MVT::i1) {
1094 // Set the high bits to zero.
1095 ResultReg = FastEmitZExtFromI1(MVT::i8, ResultReg, /*TODO: Kill=*/false);
1102 if (DstVT == MVT::i64) {
1103 // Handle extension to 64-bits via sub-register shenanigans.
1106 switch (SrcVT.SimpleTy) {
1107 case MVT::i8: MovInst = X86::MOVZX32rr8; break;
1108 case MVT::i16: MovInst = X86::MOVZX32rr16; break;
1109 case MVT::i32: MovInst = X86::MOV32rr; break;
1110 default: llvm_unreachable("Unexpected zext to i64 source type");
1113 unsigned Result32 = createResultReg(&X86::GR32RegClass);
1114 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(MovInst), Result32)
1117 ResultReg = createResultReg(&X86::GR64RegClass);
1118 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TargetOpcode::SUBREG_TO_REG),
1120 .addImm(0).addReg(Result32).addImm(X86::sub_32bit);
1121 } else if (DstVT != MVT::i8) {
1122 ResultReg = FastEmit_r(MVT::i8, DstVT.getSimpleVT(), ISD::ZERO_EXTEND,
1123 ResultReg, /*Kill=*/true);
1128 UpdateValueMap(I, ResultReg);
1133 bool X86FastISel::X86SelectBranch(const Instruction *I) {
1134 // Unconditional branches are selected by tablegen-generated code.
1135 // Handle a conditional branch.
1136 const BranchInst *BI = cast<BranchInst>(I);
1137 MachineBasicBlock *TrueMBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
1138 MachineBasicBlock *FalseMBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
1140 // Fold the common case of a conditional branch with a comparison
1141 // in the same block (values defined on other blocks may not have
1142 // initialized registers).
1143 if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
1144 if (CI->hasOneUse() && CI->getParent() == I->getParent()) {
1145 EVT VT = TLI.getValueType(CI->getOperand(0)->getType());
1147 // Try to take advantage of fallthrough opportunities.
1148 CmpInst::Predicate Predicate = CI->getPredicate();
1149 if (FuncInfo.MBB->isLayoutSuccessor(TrueMBB)) {
1150 std::swap(TrueMBB, FalseMBB);
1151 Predicate = CmpInst::getInversePredicate(Predicate);
1154 bool SwapArgs; // false -> compare Op0, Op1. true -> compare Op1, Op0.
1155 unsigned BranchOpc; // Opcode to jump on, e.g. "X86::JA"
1157 switch (Predicate) {
1158 case CmpInst::FCMP_OEQ:
1159 std::swap(TrueMBB, FalseMBB);
1160 Predicate = CmpInst::FCMP_UNE;
1162 case CmpInst::FCMP_UNE: SwapArgs = false; BranchOpc = X86::JNE_4; break;
1163 case CmpInst::FCMP_OGT: SwapArgs = false; BranchOpc = X86::JA_4; break;
1164 case CmpInst::FCMP_OGE: SwapArgs = false; BranchOpc = X86::JAE_4; break;
1165 case CmpInst::FCMP_OLT: SwapArgs = true; BranchOpc = X86::JA_4; break;
1166 case CmpInst::FCMP_OLE: SwapArgs = true; BranchOpc = X86::JAE_4; break;
1167 case CmpInst::FCMP_ONE: SwapArgs = false; BranchOpc = X86::JNE_4; break;
1168 case CmpInst::FCMP_ORD: SwapArgs = false; BranchOpc = X86::JNP_4; break;
1169 case CmpInst::FCMP_UNO: SwapArgs = false; BranchOpc = X86::JP_4; break;
1170 case CmpInst::FCMP_UEQ: SwapArgs = false; BranchOpc = X86::JE_4; break;
1171 case CmpInst::FCMP_UGT: SwapArgs = true; BranchOpc = X86::JB_4; break;
1172 case CmpInst::FCMP_UGE: SwapArgs = true; BranchOpc = X86::JBE_4; break;
1173 case CmpInst::FCMP_ULT: SwapArgs = false; BranchOpc = X86::JB_4; break;
1174 case CmpInst::FCMP_ULE: SwapArgs = false; BranchOpc = X86::JBE_4; break;
1176 case CmpInst::ICMP_EQ: SwapArgs = false; BranchOpc = X86::JE_4; break;
1177 case CmpInst::ICMP_NE: SwapArgs = false; BranchOpc = X86::JNE_4; break;
1178 case CmpInst::ICMP_UGT: SwapArgs = false; BranchOpc = X86::JA_4; break;
1179 case CmpInst::ICMP_UGE: SwapArgs = false; BranchOpc = X86::JAE_4; break;
1180 case CmpInst::ICMP_ULT: SwapArgs = false; BranchOpc = X86::JB_4; break;
1181 case CmpInst::ICMP_ULE: SwapArgs = false; BranchOpc = X86::JBE_4; break;
1182 case CmpInst::ICMP_SGT: SwapArgs = false; BranchOpc = X86::JG_4; break;
1183 case CmpInst::ICMP_SGE: SwapArgs = false; BranchOpc = X86::JGE_4; break;
1184 case CmpInst::ICMP_SLT: SwapArgs = false; BranchOpc = X86::JL_4; break;
1185 case CmpInst::ICMP_SLE: SwapArgs = false; BranchOpc = X86::JLE_4; break;
1190 const Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
1192 std::swap(Op0, Op1);
1194 // Emit a compare of the LHS and RHS, setting the flags.
1195 if (!X86FastEmitCompare(Op0, Op1, VT))
1198 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(BranchOpc))
1201 if (Predicate == CmpInst::FCMP_UNE) {
1202 // X86 requires a second branch to handle UNE (and OEQ,
1203 // which is mapped to UNE above).
1204 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::JP_4))
1208 FastEmitBranch(FalseMBB, DbgLoc);
1209 FuncInfo.MBB->addSuccessor(TrueMBB);
1212 } else if (TruncInst *TI = dyn_cast<TruncInst>(BI->getCondition())) {
1213 // Handle things like "%cond = trunc i32 %X to i1 / br i1 %cond", which
1214 // typically happen for _Bool and C++ bools.
1216 if (TI->hasOneUse() && TI->getParent() == I->getParent() &&
1217 isTypeLegal(TI->getOperand(0)->getType(), SourceVT)) {
1218 unsigned TestOpc = 0;
1219 switch (SourceVT.SimpleTy) {
1221 case MVT::i8: TestOpc = X86::TEST8ri; break;
1222 case MVT::i16: TestOpc = X86::TEST16ri; break;
1223 case MVT::i32: TestOpc = X86::TEST32ri; break;
1224 case MVT::i64: TestOpc = X86::TEST64ri32; break;
1227 unsigned OpReg = getRegForValue(TI->getOperand(0));
1228 if (OpReg == 0) return false;
1229 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TestOpc))
1230 .addReg(OpReg).addImm(1);
1232 unsigned JmpOpc = X86::JNE_4;
1233 if (FuncInfo.MBB->isLayoutSuccessor(TrueMBB)) {
1234 std::swap(TrueMBB, FalseMBB);
1238 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(JmpOpc))
1240 FastEmitBranch(FalseMBB, DbgLoc);
1241 FuncInfo.MBB->addSuccessor(TrueMBB);
1247 // Otherwise do a clumsy setcc and re-test it.
1248 // Note that i1 essentially gets ANY_EXTEND'ed to i8 where it isn't used
1249 // in an explicit cast, so make sure to handle that correctly.
1250 unsigned OpReg = getRegForValue(BI->getCondition());
1251 if (OpReg == 0) return false;
1253 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::TEST8ri))
1254 .addReg(OpReg).addImm(1);
1255 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::JNE_4))
1257 FastEmitBranch(FalseMBB, DbgLoc);
1258 FuncInfo.MBB->addSuccessor(TrueMBB);
1262 bool X86FastISel::X86SelectShift(const Instruction *I) {
1263 unsigned CReg = 0, OpReg = 0;
1264 const TargetRegisterClass *RC = nullptr;
1265 if (I->getType()->isIntegerTy(8)) {
1267 RC = &X86::GR8RegClass;
1268 switch (I->getOpcode()) {
1269 case Instruction::LShr: OpReg = X86::SHR8rCL; break;
1270 case Instruction::AShr: OpReg = X86::SAR8rCL; break;
1271 case Instruction::Shl: OpReg = X86::SHL8rCL; break;
1272 default: return false;
1274 } else if (I->getType()->isIntegerTy(16)) {
1276 RC = &X86::GR16RegClass;
1277 switch (I->getOpcode()) {
1278 case Instruction::LShr: OpReg = X86::SHR16rCL; break;
1279 case Instruction::AShr: OpReg = X86::SAR16rCL; break;
1280 case Instruction::Shl: OpReg = X86::SHL16rCL; break;
1281 default: return false;
1283 } else if (I->getType()->isIntegerTy(32)) {
1285 RC = &X86::GR32RegClass;
1286 switch (I->getOpcode()) {
1287 case Instruction::LShr: OpReg = X86::SHR32rCL; break;
1288 case Instruction::AShr: OpReg = X86::SAR32rCL; break;
1289 case Instruction::Shl: OpReg = X86::SHL32rCL; break;
1290 default: return false;
1292 } else if (I->getType()->isIntegerTy(64)) {
1294 RC = &X86::GR64RegClass;
1295 switch (I->getOpcode()) {
1296 case Instruction::LShr: OpReg = X86::SHR64rCL; break;
1297 case Instruction::AShr: OpReg = X86::SAR64rCL; break;
1298 case Instruction::Shl: OpReg = X86::SHL64rCL; break;
1299 default: return false;
1306 if (!isTypeLegal(I->getType(), VT))
1309 unsigned Op0Reg = getRegForValue(I->getOperand(0));
1310 if (Op0Reg == 0) return false;
1312 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1313 if (Op1Reg == 0) return false;
1314 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TargetOpcode::COPY),
1315 CReg).addReg(Op1Reg);
1317 // The shift instruction uses X86::CL. If we defined a super-register
1318 // of X86::CL, emit a subreg KILL to precisely describe what we're doing here.
1319 if (CReg != X86::CL)
1320 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1321 TII.get(TargetOpcode::KILL), X86::CL)
1322 .addReg(CReg, RegState::Kill);
1324 unsigned ResultReg = createResultReg(RC);
1325 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(OpReg), ResultReg)
1327 UpdateValueMap(I, ResultReg);
1331 bool X86FastISel::X86SelectDivRem(const Instruction *I) {
1332 const static unsigned NumTypes = 4; // i8, i16, i32, i64
1333 const static unsigned NumOps = 4; // SDiv, SRem, UDiv, URem
1334 const static bool S = true; // IsSigned
1335 const static bool U = false; // !IsSigned
1336 const static unsigned Copy = TargetOpcode::COPY;
1337 // For the X86 DIV/IDIV instruction, in most cases the dividend
1338 // (numerator) must be in a specific register pair highreg:lowreg,
1339 // producing the quotient in lowreg and the remainder in highreg.
1340 // For most data types, to set up the instruction, the dividend is
1341 // copied into lowreg, and lowreg is sign-extended or zero-extended
1342 // into highreg. The exception is i8, where the dividend is defined
1343 // as a single register rather than a register pair, and we
1344 // therefore directly sign-extend or zero-extend the dividend into
1345 // lowreg, instead of copying, and ignore the highreg.
1346 const static struct DivRemEntry {
1347 // The following portion depends only on the data type.
1348 const TargetRegisterClass *RC;
1349 unsigned LowInReg; // low part of the register pair
1350 unsigned HighInReg; // high part of the register pair
1351 // The following portion depends on both the data type and the operation.
1352 struct DivRemResult {
1353 unsigned OpDivRem; // The specific DIV/IDIV opcode to use.
1354 unsigned OpSignExtend; // Opcode for sign-extending lowreg into
1355 // highreg, or copying a zero into highreg.
1356 unsigned OpCopy; // Opcode for copying dividend into lowreg, or
1357 // zero/sign-extending into lowreg for i8.
1358 unsigned DivRemResultReg; // Register containing the desired result.
1359 bool IsOpSigned; // Whether to use signed or unsigned form.
1360 } ResultTable[NumOps];
1361 } OpTable[NumTypes] = {
1362 { &X86::GR8RegClass, X86::AX, 0, {
1363 { X86::IDIV8r, 0, X86::MOVSX16rr8, X86::AL, S }, // SDiv
1364 { X86::IDIV8r, 0, X86::MOVSX16rr8, X86::AH, S }, // SRem
1365 { X86::DIV8r, 0, X86::MOVZX16rr8, X86::AL, U }, // UDiv
1366 { X86::DIV8r, 0, X86::MOVZX16rr8, X86::AH, U }, // URem
1369 { &X86::GR16RegClass, X86::AX, X86::DX, {
1370 { X86::IDIV16r, X86::CWD, Copy, X86::AX, S }, // SDiv
1371 { X86::IDIV16r, X86::CWD, Copy, X86::DX, S }, // SRem
1372 { X86::DIV16r, X86::MOV32r0, Copy, X86::AX, U }, // UDiv
1373 { X86::DIV16r, X86::MOV32r0, Copy, X86::DX, U }, // URem
1376 { &X86::GR32RegClass, X86::EAX, X86::EDX, {
1377 { X86::IDIV32r, X86::CDQ, Copy, X86::EAX, S }, // SDiv
1378 { X86::IDIV32r, X86::CDQ, Copy, X86::EDX, S }, // SRem
1379 { X86::DIV32r, X86::MOV32r0, Copy, X86::EAX, U }, // UDiv
1380 { X86::DIV32r, X86::MOV32r0, Copy, X86::EDX, U }, // URem
1383 { &X86::GR64RegClass, X86::RAX, X86::RDX, {
1384 { X86::IDIV64r, X86::CQO, Copy, X86::RAX, S }, // SDiv
1385 { X86::IDIV64r, X86::CQO, Copy, X86::RDX, S }, // SRem
1386 { X86::DIV64r, X86::MOV32r0, Copy, X86::RAX, U }, // UDiv
1387 { X86::DIV64r, X86::MOV32r0, Copy, X86::RDX, U }, // URem
1393 if (!isTypeLegal(I->getType(), VT))
1396 unsigned TypeIndex, OpIndex;
1397 switch (VT.SimpleTy) {
1398 default: return false;
1399 case MVT::i8: TypeIndex = 0; break;
1400 case MVT::i16: TypeIndex = 1; break;
1401 case MVT::i32: TypeIndex = 2; break;
1402 case MVT::i64: TypeIndex = 3;
1403 if (!Subtarget->is64Bit())
1408 switch (I->getOpcode()) {
1409 default: llvm_unreachable("Unexpected div/rem opcode");
1410 case Instruction::SDiv: OpIndex = 0; break;
1411 case Instruction::SRem: OpIndex = 1; break;
1412 case Instruction::UDiv: OpIndex = 2; break;
1413 case Instruction::URem: OpIndex = 3; break;
1416 const DivRemEntry &TypeEntry = OpTable[TypeIndex];
1417 const DivRemEntry::DivRemResult &OpEntry = TypeEntry.ResultTable[OpIndex];
1418 unsigned Op0Reg = getRegForValue(I->getOperand(0));
1421 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1425 // Move op0 into low-order input register.
1426 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1427 TII.get(OpEntry.OpCopy), TypeEntry.LowInReg).addReg(Op0Reg);
1428 // Zero-extend or sign-extend into high-order input register.
1429 if (OpEntry.OpSignExtend) {
1430 if (OpEntry.IsOpSigned)
1431 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1432 TII.get(OpEntry.OpSignExtend));
1434 unsigned Zero32 = createResultReg(&X86::GR32RegClass);
1435 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1436 TII.get(X86::MOV32r0), Zero32);
1438 // Copy the zero into the appropriate sub/super/identical physical
1439 // register. Unfortunately the operations needed are not uniform enough to
1440 // fit neatly into the table above.
1441 if (VT.SimpleTy == MVT::i16) {
1442 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1443 TII.get(Copy), TypeEntry.HighInReg)
1444 .addReg(Zero32, 0, X86::sub_16bit);
1445 } else if (VT.SimpleTy == MVT::i32) {
1446 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1447 TII.get(Copy), TypeEntry.HighInReg)
1449 } else if (VT.SimpleTy == MVT::i64) {
1450 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1451 TII.get(TargetOpcode::SUBREG_TO_REG), TypeEntry.HighInReg)
1452 .addImm(0).addReg(Zero32).addImm(X86::sub_32bit);
1456 // Generate the DIV/IDIV instruction.
1457 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1458 TII.get(OpEntry.OpDivRem)).addReg(Op1Reg);
1459 // For i8 remainder, we can't reference AH directly, as we'll end
1460 // up with bogus copies like %R9B = COPY %AH. Reference AX
1461 // instead to prevent AH references in a REX instruction.
1463 // The current assumption of the fast register allocator is that isel
1464 // won't generate explicit references to the GPR8_NOREX registers. If
1465 // the allocator and/or the backend get enhanced to be more robust in
1466 // that regard, this can be, and should be, removed.
1467 unsigned ResultReg = 0;
1468 if ((I->getOpcode() == Instruction::SRem ||
1469 I->getOpcode() == Instruction::URem) &&
1470 OpEntry.DivRemResultReg == X86::AH && Subtarget->is64Bit()) {
1471 unsigned SourceSuperReg = createResultReg(&X86::GR16RegClass);
1472 unsigned ResultSuperReg = createResultReg(&X86::GR16RegClass);
1473 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1474 TII.get(Copy), SourceSuperReg).addReg(X86::AX);
1476 // Shift AX right by 8 bits instead of using AH.
1477 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::SHR16ri),
1478 ResultSuperReg).addReg(SourceSuperReg).addImm(8);
1480 // Now reference the 8-bit subreg of the result.
1481 ResultReg = FastEmitInst_extractsubreg(MVT::i8, ResultSuperReg,
1482 /*Kill=*/true, X86::sub_8bit);
1484 // Copy the result out of the physreg if we haven't already.
1486 ResultReg = createResultReg(TypeEntry.RC);
1487 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Copy), ResultReg)
1488 .addReg(OpEntry.DivRemResultReg);
1490 UpdateValueMap(I, ResultReg);
1495 bool X86FastISel::X86SelectSelect(const Instruction *I) {
1497 if (!isTypeLegal(I->getType(), VT))
1500 // We only use cmov here, if we don't have a cmov instruction bail.
1501 if (!Subtarget->hasCMov()) return false;
1504 const TargetRegisterClass *RC = nullptr;
1505 if (VT == MVT::i16) {
1506 Opc = X86::CMOVE16rr;
1507 RC = &X86::GR16RegClass;
1508 } else if (VT == MVT::i32) {
1509 Opc = X86::CMOVE32rr;
1510 RC = &X86::GR32RegClass;
1511 } else if (VT == MVT::i64) {
1512 Opc = X86::CMOVE64rr;
1513 RC = &X86::GR64RegClass;
1518 unsigned Op0Reg = getRegForValue(I->getOperand(0));
1519 if (Op0Reg == 0) return false;
1520 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1521 if (Op1Reg == 0) return false;
1522 unsigned Op2Reg = getRegForValue(I->getOperand(2));
1523 if (Op2Reg == 0) return false;
1525 // Selects operate on i1, however, Op0Reg is 8 bits width and may contain
1526 // garbage. Indeed, only the less significant bit is supposed to be accurate.
1527 // If we read more than the lsb, we may see non-zero values whereas lsb
1528 // is zero. Therefore, we have to truncate Op0Reg to i1 for the select.
1529 // This is achieved by performing TEST against 1.
1530 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::TEST8ri))
1531 .addReg(Op0Reg).addImm(1);
1532 unsigned ResultReg = createResultReg(RC);
1533 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
1534 .addReg(Op1Reg).addReg(Op2Reg);
1535 UpdateValueMap(I, ResultReg);
1539 bool X86FastISel::X86SelectFPExt(const Instruction *I) {
1540 // fpext from float to double.
1541 if (X86ScalarSSEf64 &&
1542 I->getType()->isDoubleTy()) {
1543 const Value *V = I->getOperand(0);
1544 if (V->getType()->isFloatTy()) {
1545 unsigned OpReg = getRegForValue(V);
1546 if (OpReg == 0) return false;
1547 unsigned ResultReg = createResultReg(&X86::FR64RegClass);
1548 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1549 TII.get(X86::CVTSS2SDrr), ResultReg)
1551 UpdateValueMap(I, ResultReg);
1559 bool X86FastISel::X86SelectFPTrunc(const Instruction *I) {
1560 if (X86ScalarSSEf64) {
1561 if (I->getType()->isFloatTy()) {
1562 const Value *V = I->getOperand(0);
1563 if (V->getType()->isDoubleTy()) {
1564 unsigned OpReg = getRegForValue(V);
1565 if (OpReg == 0) return false;
1566 unsigned ResultReg = createResultReg(&X86::FR32RegClass);
1567 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1568 TII.get(X86::CVTSD2SSrr), ResultReg)
1570 UpdateValueMap(I, ResultReg);
1579 bool X86FastISel::X86SelectTrunc(const Instruction *I) {
1580 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
1581 EVT DstVT = TLI.getValueType(I->getType());
1583 // This code only handles truncation to byte.
1584 if (DstVT != MVT::i8 && DstVT != MVT::i1)
1586 if (!TLI.isTypeLegal(SrcVT))
1589 unsigned InputReg = getRegForValue(I->getOperand(0));
1591 // Unhandled operand. Halt "fast" selection and bail.
1594 if (SrcVT == MVT::i8) {
1595 // Truncate from i8 to i1; no code needed.
1596 UpdateValueMap(I, InputReg);
1600 if (!Subtarget->is64Bit()) {
1601 // If we're on x86-32; we can't extract an i8 from a general register.
1602 // First issue a copy to GR16_ABCD or GR32_ABCD.
1603 const TargetRegisterClass *CopyRC = (SrcVT == MVT::i16) ?
1604 (const TargetRegisterClass*)&X86::GR16_ABCDRegClass :
1605 (const TargetRegisterClass*)&X86::GR32_ABCDRegClass;
1606 unsigned CopyReg = createResultReg(CopyRC);
1607 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TargetOpcode::COPY),
1608 CopyReg).addReg(InputReg);
1612 // Issue an extract_subreg.
1613 unsigned ResultReg = FastEmitInst_extractsubreg(MVT::i8,
1614 InputReg, /*Kill=*/true,
1619 UpdateValueMap(I, ResultReg);
1623 bool X86FastISel::IsMemcpySmall(uint64_t Len) {
1624 return Len <= (Subtarget->is64Bit() ? 32 : 16);
1627 bool X86FastISel::TryEmitSmallMemcpy(X86AddressMode DestAM,
1628 X86AddressMode SrcAM, uint64_t Len) {
1630 // Make sure we don't bloat code by inlining very large memcpy's.
1631 if (!IsMemcpySmall(Len))
1634 bool i64Legal = Subtarget->is64Bit();
1636 // We don't care about alignment here since we just emit integer accesses.
1639 if (Len >= 8 && i64Legal)
1650 bool RV = X86FastEmitLoad(VT, SrcAM, nullptr, Reg);
1651 RV &= X86FastEmitStore(VT, Reg, /*Kill=*/true, DestAM);
1652 assert(RV && "Failed to emit load or store??");
1654 unsigned Size = VT.getSizeInBits()/8;
1656 DestAM.Disp += Size;
1663 static bool isCommutativeIntrinsic(IntrinsicInst const &I) {
1664 switch (I.getIntrinsicID()) {
1665 case Intrinsic::sadd_with_overflow:
1666 case Intrinsic::uadd_with_overflow:
1667 case Intrinsic::smul_with_overflow:
1668 case Intrinsic::umul_with_overflow:
1675 bool X86FastISel::X86VisitIntrinsicCall(const IntrinsicInst &I) {
1676 // FIXME: Handle more intrinsics.
1677 switch (I.getIntrinsicID()) {
1678 default: return false;
1679 case Intrinsic::frameaddress: {
1680 Type *RetTy = I.getCalledFunction()->getReturnType();
1683 if (!isTypeLegal(RetTy, VT))
1687 const TargetRegisterClass *RC = nullptr;
1689 switch (VT.SimpleTy) {
1690 default: llvm_unreachable("Invalid result type for frameaddress.");
1691 case MVT::i32: Opc = X86::MOV32rm; RC = &X86::GR32RegClass; break;
1692 case MVT::i64: Opc = X86::MOV64rm; RC = &X86::GR64RegClass; break;
1695 // This needs to be set before we call getFrameRegister, otherwise we get
1696 // the wrong frame register.
1697 MachineFrameInfo *MFI = FuncInfo.MF->getFrameInfo();
1698 MFI->setFrameAddressIsTaken(true);
1700 const X86RegisterInfo *RegInfo =
1701 static_cast<const X86RegisterInfo*>(TM.getRegisterInfo());
1702 unsigned FrameReg = RegInfo->getFrameRegister(*(FuncInfo.MF));
1703 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
1704 (FrameReg == X86::EBP && VT == MVT::i32)) &&
1705 "Invalid Frame Register!");
1707 // Always make a copy of the frame register to to a vreg first, so that we
1708 // never directly reference the frame register (the TwoAddressInstruction-
1709 // Pass doesn't like that).
1710 unsigned SrcReg = createResultReg(RC);
1711 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1712 TII.get(TargetOpcode::COPY), SrcReg).addReg(FrameReg);
1714 // Now recursively load from the frame address.
1715 // movq (%rbp), %rax
1716 // movq (%rax), %rax
1717 // movq (%rax), %rax
1720 unsigned Depth = cast<ConstantInt>(I.getOperand(0))->getZExtValue();
1722 DestReg = createResultReg(RC);
1723 addDirectMem(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1724 TII.get(Opc), DestReg), SrcReg);
1728 UpdateValueMap(&I, SrcReg);
1731 case Intrinsic::memcpy: {
1732 const MemCpyInst &MCI = cast<MemCpyInst>(I);
1733 // Don't handle volatile or variable length memcpys.
1734 if (MCI.isVolatile())
1737 if (isa<ConstantInt>(MCI.getLength())) {
1738 // Small memcpy's are common enough that we want to do them
1739 // without a call if possible.
1740 uint64_t Len = cast<ConstantInt>(MCI.getLength())->getZExtValue();
1741 if (IsMemcpySmall(Len)) {
1742 X86AddressMode DestAM, SrcAM;
1743 if (!X86SelectAddress(MCI.getRawDest(), DestAM) ||
1744 !X86SelectAddress(MCI.getRawSource(), SrcAM))
1746 TryEmitSmallMemcpy(DestAM, SrcAM, Len);
1751 unsigned SizeWidth = Subtarget->is64Bit() ? 64 : 32;
1752 if (!MCI.getLength()->getType()->isIntegerTy(SizeWidth))
1755 if (MCI.getSourceAddressSpace() > 255 || MCI.getDestAddressSpace() > 255)
1758 return DoSelectCall(&I, "memcpy");
1760 case Intrinsic::memset: {
1761 const MemSetInst &MSI = cast<MemSetInst>(I);
1763 if (MSI.isVolatile())
1766 unsigned SizeWidth = Subtarget->is64Bit() ? 64 : 32;
1767 if (!MSI.getLength()->getType()->isIntegerTy(SizeWidth))
1770 if (MSI.getDestAddressSpace() > 255)
1773 return DoSelectCall(&I, "memset");
1775 case Intrinsic::stackprotector: {
1776 // Emit code to store the stack guard onto the stack.
1777 EVT PtrTy = TLI.getPointerTy();
1779 const Value *Op1 = I.getArgOperand(0); // The guard's value.
1780 const AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1));
1782 MFI.setStackProtectorIndex(FuncInfo.StaticAllocaMap[Slot]);
1784 // Grab the frame index.
1786 if (!X86SelectAddress(Slot, AM)) return false;
1787 if (!X86FastEmitStore(PtrTy, Op1, AM)) return false;
1790 case Intrinsic::dbg_declare: {
1791 const DbgDeclareInst *DI = cast<DbgDeclareInst>(&I);
1793 assert(DI->getAddress() && "Null address should be checked earlier!");
1794 if (!X86SelectAddress(DI->getAddress(), AM))
1796 const MCInstrDesc &II = TII.get(TargetOpcode::DBG_VALUE);
1797 // FIXME may need to add RegState::Debug to any registers produced,
1798 // although ESP/EBP should be the only ones at the moment.
1799 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II), AM).
1800 addImm(0).addMetadata(DI->getVariable());
1803 case Intrinsic::trap: {
1804 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::TRAP));
1807 case Intrinsic::sqrt: {
1808 if (!Subtarget->hasSSE1())
1811 Type *RetTy = I.getCalledFunction()->getReturnType();
1814 if (!isTypeLegal(RetTy, VT))
1817 // Unfortunatelly we can't use FastEmit_r, because the AVX version of FSQRT
1818 // is not generated by FastISel yet.
1819 // FIXME: Update this code once tablegen can handle it.
1820 static const unsigned SqrtOpc[2][2] = {
1821 {X86::SQRTSSr, X86::VSQRTSSr},
1822 {X86::SQRTSDr, X86::VSQRTSDr}
1824 bool HasAVX = Subtarget->hasAVX();
1826 const TargetRegisterClass *RC;
1827 switch (VT.SimpleTy) {
1828 default: return false;
1829 case MVT::f32: Opc = SqrtOpc[0][HasAVX]; RC = &X86::FR32RegClass; break;
1830 case MVT::f64: Opc = SqrtOpc[1][HasAVX]; RC = &X86::FR64RegClass; break;
1833 const Value *SrcVal = I.getArgOperand(0);
1834 unsigned SrcReg = getRegForValue(SrcVal);
1839 unsigned ImplicitDefReg = 0;
1841 ImplicitDefReg = createResultReg(RC);
1842 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1843 TII.get(TargetOpcode::IMPLICIT_DEF), ImplicitDefReg);
1846 unsigned ResultReg = createResultReg(RC);
1847 MachineInstrBuilder MIB;
1848 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc),
1852 MIB.addReg(ImplicitDefReg);
1856 UpdateValueMap(&I, ResultReg);
1859 case Intrinsic::sadd_with_overflow:
1860 case Intrinsic::uadd_with_overflow:
1861 case Intrinsic::ssub_with_overflow:
1862 case Intrinsic::usub_with_overflow:
1863 case Intrinsic::smul_with_overflow:
1864 case Intrinsic::umul_with_overflow: {
1865 // This implements the basic lowering of the xalu with overflow intrinsics
1866 // into add/sub/mul folowed by either seto or setb.
1867 const Function *Callee = I.getCalledFunction();
1868 auto *Ty = cast<StructType>(Callee->getReturnType());
1869 Type *RetTy = Ty->getTypeAtIndex(0U);
1870 Type *CondTy = Ty->getTypeAtIndex(1);
1873 if (!isTypeLegal(RetTy, VT))
1876 if (VT < MVT::i8 || VT > MVT::i64)
1879 const Value *LHS = I.getArgOperand(0);
1880 const Value *RHS = I.getArgOperand(1);
1882 // Canonicalize immediates to the RHS.
1883 if (isa<ConstantInt>(LHS) && !isa<ConstantInt>(RHS) &&
1884 isCommutativeIntrinsic(I))
1885 std::swap(LHS, RHS);
1887 unsigned BaseOpc, CondOpc;
1888 switch (I.getIntrinsicID()) {
1889 default: llvm_unreachable("Unexpected intrinsic!");
1890 case Intrinsic::sadd_with_overflow:
1891 BaseOpc = ISD::ADD; CondOpc = X86::SETOr; break;
1892 case Intrinsic::uadd_with_overflow:
1893 BaseOpc = ISD::ADD; CondOpc = X86::SETBr; break;
1894 case Intrinsic::ssub_with_overflow:
1895 BaseOpc = ISD::SUB; CondOpc = X86::SETOr; break;
1896 case Intrinsic::usub_with_overflow:
1897 BaseOpc = ISD::SUB; CondOpc = X86::SETBr; break;
1898 case Intrinsic::smul_with_overflow:
1899 BaseOpc = ISD::MUL; CondOpc = X86::SETOr; break;
1900 case Intrinsic::umul_with_overflow:
1901 BaseOpc = X86ISD::UMUL; CondOpc = X86::SETOr; break;
1904 unsigned LHSReg = getRegForValue(LHS);
1907 bool LHSIsKill = hasTrivialKill(LHS);
1909 unsigned ResultReg = 0;
1910 // Check if we have an immediate version.
1911 if (auto const *C = dyn_cast<ConstantInt>(RHS)) {
1912 ResultReg = FastEmit_ri(VT, VT, BaseOpc, LHSReg, LHSIsKill,
1919 RHSReg = getRegForValue(RHS);
1922 RHSIsKill = hasTrivialKill(RHS);
1923 ResultReg = FastEmit_rr(VT, VT, BaseOpc, LHSReg, LHSIsKill, RHSReg,
1927 // FastISel doesn't have a pattern for X86::MUL*r. Emit it manually.
1928 if (BaseOpc == X86ISD::UMUL && !ResultReg) {
1929 static const unsigned MULOpc[] =
1930 { X86::MUL8r, X86::MUL16r, X86::MUL32r, X86::MUL64r };
1931 static const unsigned Reg[] = { X86::AL, X86::AX, X86::EAX, X86::RAX };
1932 // First copy the first operand into RAX, which is an implicit input to
1933 // the X86::MUL*r instruction.
1934 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1935 TII.get(TargetOpcode::COPY), Reg[VT.SimpleTy-MVT::i8])
1936 .addReg(LHSReg, getKillRegState(LHSIsKill));
1937 ResultReg = FastEmitInst_r(MULOpc[VT.SimpleTy-MVT::i8],
1938 TLI.getRegClassFor(VT), RHSReg, RHSIsKill);
1944 unsigned ResultReg2 = FuncInfo.CreateRegs(CondTy);
1945 assert((ResultReg+1) == ResultReg2 && "Nonconsecutive result registers.");
1946 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CondOpc),
1949 UpdateValueMap(&I, ResultReg, 2);
1955 bool X86FastISel::FastLowerArguments() {
1956 if (!FuncInfo.CanLowerReturn)
1959 const Function *F = FuncInfo.Fn;
1963 CallingConv::ID CC = F->getCallingConv();
1964 if (CC != CallingConv::C)
1967 if (Subtarget->isCallingConvWin64(CC))
1970 if (!Subtarget->is64Bit())
1973 // Only handle simple cases. i.e. Up to 6 i32/i64 scalar arguments.
1974 unsigned GPRCnt = 0;
1975 unsigned FPRCnt = 0;
1977 for (auto const &Arg : F->args()) {
1978 // The first argument is at index 1.
1980 if (F->getAttributes().hasAttribute(Idx, Attribute::ByVal) ||
1981 F->getAttributes().hasAttribute(Idx, Attribute::InReg) ||
1982 F->getAttributes().hasAttribute(Idx, Attribute::StructRet) ||
1983 F->getAttributes().hasAttribute(Idx, Attribute::Nest))
1986 Type *ArgTy = Arg.getType();
1987 if (ArgTy->isStructTy() || ArgTy->isArrayTy() || ArgTy->isVectorTy())
1990 EVT ArgVT = TLI.getValueType(ArgTy);
1991 if (!ArgVT.isSimple()) return false;
1992 switch (ArgVT.getSimpleVT().SimpleTy) {
1993 default: return false;
2000 if (!Subtarget->hasSSE1())
2013 static const MCPhysReg GPR32ArgRegs[] = {
2014 X86::EDI, X86::ESI, X86::EDX, X86::ECX, X86::R8D, X86::R9D
2016 static const MCPhysReg GPR64ArgRegs[] = {
2017 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8 , X86::R9
2019 static const MCPhysReg XMMArgRegs[] = {
2020 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2021 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2024 unsigned GPRIdx = 0;
2025 unsigned FPRIdx = 0;
2026 for (auto const &Arg : F->args()) {
2027 MVT VT = TLI.getSimpleValueType(Arg.getType());
2028 const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
2030 switch (VT.SimpleTy) {
2031 default: llvm_unreachable("Unexpected value type.");
2032 case MVT::i32: SrcReg = GPR32ArgRegs[GPRIdx++]; break;
2033 case MVT::i64: SrcReg = GPR64ArgRegs[GPRIdx++]; break;
2034 case MVT::f32: // fall-through
2035 case MVT::f64: SrcReg = XMMArgRegs[FPRIdx++]; break;
2037 unsigned DstReg = FuncInfo.MF->addLiveIn(SrcReg, RC);
2038 // FIXME: Unfortunately it's necessary to emit a copy from the livein copy.
2039 // Without this, EmitLiveInCopies may eliminate the livein if its only
2040 // use is a bitcast (which isn't turned into an instruction).
2041 unsigned ResultReg = createResultReg(RC);
2042 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2043 TII.get(TargetOpcode::COPY), ResultReg)
2044 .addReg(DstReg, getKillRegState(true));
2045 UpdateValueMap(&Arg, ResultReg);
2050 bool X86FastISel::X86SelectCall(const Instruction *I) {
2051 const CallInst *CI = cast<CallInst>(I);
2052 const Value *Callee = CI->getCalledValue();
2054 // Can't handle inline asm yet.
2055 if (isa<InlineAsm>(Callee))
2058 // Handle intrinsic calls.
2059 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI))
2060 return X86VisitIntrinsicCall(*II);
2062 // Allow SelectionDAG isel to handle tail calls.
2063 if (cast<CallInst>(I)->isTailCall())
2066 return DoSelectCall(I, nullptr);
2069 static unsigned computeBytesPoppedByCallee(const X86Subtarget &Subtarget,
2070 const ImmutableCallSite &CS) {
2071 if (Subtarget.is64Bit())
2073 if (Subtarget.getTargetTriple().isOSMSVCRT())
2075 CallingConv::ID CC = CS.getCallingConv();
2076 if (CC == CallingConv::Fast || CC == CallingConv::GHC)
2078 if (!CS.paramHasAttr(1, Attribute::StructRet))
2080 if (CS.paramHasAttr(1, Attribute::InReg))
2085 // Select either a call, or an llvm.memcpy/memmove/memset intrinsic
2086 bool X86FastISel::DoSelectCall(const Instruction *I, const char *MemIntName) {
2087 const CallInst *CI = cast<CallInst>(I);
2088 const Value *Callee = CI->getCalledValue();
2090 // Handle only C and fastcc calling conventions for now.
2091 ImmutableCallSite CS(CI);
2092 CallingConv::ID CC = CS.getCallingConv();
2093 bool isWin64 = Subtarget->isCallingConvWin64(CC);
2094 if (CC != CallingConv::C && CC != CallingConv::Fast &&
2095 CC != CallingConv::X86_FastCall && CC != CallingConv::X86_64_Win64 &&
2096 CC != CallingConv::X86_64_SysV)
2099 // fastcc with -tailcallopt is intended to provide a guaranteed
2100 // tail call optimization. Fastisel doesn't know how to do that.
2101 if (CC == CallingConv::Fast && TM.Options.GuaranteedTailCallOpt)
2104 PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
2105 FunctionType *FTy = cast<FunctionType>(PT->getElementType());
2106 bool isVarArg = FTy->isVarArg();
2108 // Don't know how to handle Win64 varargs yet. Nothing special needed for
2109 // x86-32. Special handling for x86-64 is implemented.
2110 if (isVarArg && isWin64)
2113 // Don't know about inalloca yet.
2114 if (CS.hasInAllocaArgument())
2117 // Fast-isel doesn't know about callee-pop yet.
2118 if (X86::isCalleePop(CC, Subtarget->is64Bit(), isVarArg,
2119 TM.Options.GuaranteedTailCallOpt))
2122 // Check whether the function can return without sret-demotion.
2123 SmallVector<ISD::OutputArg, 4> Outs;
2124 GetReturnInfo(I->getType(), CS.getAttributes(), Outs, TLI);
2125 bool CanLowerReturn = TLI.CanLowerReturn(CS.getCallingConv(),
2126 *FuncInfo.MF, FTy->isVarArg(),
2127 Outs, FTy->getContext());
2128 if (!CanLowerReturn)
2131 // Materialize callee address in a register. FIXME: GV address can be
2132 // handled with a CALLpcrel32 instead.
2133 X86AddressMode CalleeAM;
2134 if (!X86SelectCallAddress(Callee, CalleeAM))
2136 unsigned CalleeOp = 0;
2137 const GlobalValue *GV = nullptr;
2138 if (CalleeAM.GV != nullptr) {
2140 } else if (CalleeAM.Base.Reg != 0) {
2141 CalleeOp = CalleeAM.Base.Reg;
2145 // Deal with call operands first.
2146 SmallVector<const Value *, 8> ArgVals;
2147 SmallVector<unsigned, 8> Args;
2148 SmallVector<MVT, 8> ArgVTs;
2149 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
2150 unsigned arg_size = CS.arg_size();
2151 Args.reserve(arg_size);
2152 ArgVals.reserve(arg_size);
2153 ArgVTs.reserve(arg_size);
2154 ArgFlags.reserve(arg_size);
2155 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
2157 // If we're lowering a mem intrinsic instead of a regular call, skip the
2158 // last two arguments, which should not passed to the underlying functions.
2159 if (MemIntName && e-i <= 2)
2162 ISD::ArgFlagsTy Flags;
2163 unsigned AttrInd = i - CS.arg_begin() + 1;
2164 if (CS.paramHasAttr(AttrInd, Attribute::SExt))
2166 if (CS.paramHasAttr(AttrInd, Attribute::ZExt))
2169 if (CS.paramHasAttr(AttrInd, Attribute::ByVal)) {
2170 PointerType *Ty = cast<PointerType>(ArgVal->getType());
2171 Type *ElementTy = Ty->getElementType();
2172 unsigned FrameSize = DL.getTypeAllocSize(ElementTy);
2173 unsigned FrameAlign = CS.getParamAlignment(AttrInd);
2175 FrameAlign = TLI.getByValTypeAlignment(ElementTy);
2177 Flags.setByValSize(FrameSize);
2178 Flags.setByValAlign(FrameAlign);
2179 if (!IsMemcpySmall(FrameSize))
2183 if (CS.paramHasAttr(AttrInd, Attribute::InReg))
2185 if (CS.paramHasAttr(AttrInd, Attribute::Nest))
2188 // If this is an i1/i8/i16 argument, promote to i32 to avoid an extra
2189 // instruction. This is safe because it is common to all fastisel supported
2190 // calling conventions on x86.
2191 if (ConstantInt *CI = dyn_cast<ConstantInt>(ArgVal)) {
2192 if (CI->getBitWidth() == 1 || CI->getBitWidth() == 8 ||
2193 CI->getBitWidth() == 16) {
2195 ArgVal = ConstantExpr::getSExt(CI,Type::getInt32Ty(CI->getContext()));
2197 ArgVal = ConstantExpr::getZExt(CI,Type::getInt32Ty(CI->getContext()));
2203 // Passing bools around ends up doing a trunc to i1 and passing it.
2204 // Codegen this as an argument + "and 1".
2205 if (ArgVal->getType()->isIntegerTy(1) && isa<TruncInst>(ArgVal) &&
2206 cast<TruncInst>(ArgVal)->getParent() == I->getParent() &&
2207 ArgVal->hasOneUse()) {
2208 ArgVal = cast<TruncInst>(ArgVal)->getOperand(0);
2209 ArgReg = getRegForValue(ArgVal);
2210 if (ArgReg == 0) return false;
2213 if (!isTypeLegal(ArgVal->getType(), ArgVT)) return false;
2215 ArgReg = FastEmit_ri(ArgVT, ArgVT, ISD::AND, ArgReg,
2216 ArgVal->hasOneUse(), 1);
2218 ArgReg = getRegForValue(ArgVal);
2221 if (ArgReg == 0) return false;
2223 Type *ArgTy = ArgVal->getType();
2225 if (!isTypeLegal(ArgTy, ArgVT))
2227 if (ArgVT == MVT::x86mmx)
2229 unsigned OriginalAlignment = DL.getABITypeAlignment(ArgTy);
2230 Flags.setOrigAlign(OriginalAlignment);
2232 Args.push_back(ArgReg);
2233 ArgVals.push_back(ArgVal);
2234 ArgVTs.push_back(ArgVT);
2235 ArgFlags.push_back(Flags);
2238 // Analyze operands of the call, assigning locations to each operand.
2239 SmallVector<CCValAssign, 16> ArgLocs;
2240 CCState CCInfo(CC, isVarArg, *FuncInfo.MF, TM, ArgLocs,
2241 I->getParent()->getContext());
2243 // Allocate shadow area for Win64
2245 CCInfo.AllocateStack(32, 8);
2247 CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CC_X86);
2249 // Get a count of how many bytes are to be pushed on the stack.
2250 unsigned NumBytes = CCInfo.getNextStackOffset();
2252 // Issue CALLSEQ_START
2253 unsigned AdjStackDown = TII.getCallFrameSetupOpcode();
2254 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackDown))
2257 // Process argument: walk the register/memloc assignments, inserting
2259 SmallVector<unsigned, 4> RegArgs;
2260 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2261 CCValAssign &VA = ArgLocs[i];
2262 unsigned Arg = Args[VA.getValNo()];
2263 EVT ArgVT = ArgVTs[VA.getValNo()];
2265 // Promote the value if needed.
2266 switch (VA.getLocInfo()) {
2267 case CCValAssign::Full: break;
2268 case CCValAssign::SExt: {
2269 assert(VA.getLocVT().isInteger() && !VA.getLocVT().isVector() &&
2270 "Unexpected extend");
2271 bool Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
2273 assert(Emitted && "Failed to emit a sext!"); (void)Emitted;
2274 ArgVT = VA.getLocVT();
2277 case CCValAssign::ZExt: {
2278 assert(VA.getLocVT().isInteger() && !VA.getLocVT().isVector() &&
2279 "Unexpected extend");
2280 bool Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
2282 assert(Emitted && "Failed to emit a zext!"); (void)Emitted;
2283 ArgVT = VA.getLocVT();
2286 case CCValAssign::AExt: {
2287 assert(VA.getLocVT().isInteger() && !VA.getLocVT().isVector() &&
2288 "Unexpected extend");
2289 bool Emitted = X86FastEmitExtend(ISD::ANY_EXTEND, VA.getLocVT(),
2292 Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
2295 Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
2298 assert(Emitted && "Failed to emit a aext!"); (void)Emitted;
2299 ArgVT = VA.getLocVT();
2302 case CCValAssign::BCvt: {
2303 unsigned BC = FastEmit_r(ArgVT.getSimpleVT(), VA.getLocVT(),
2304 ISD::BITCAST, Arg, /*TODO: Kill=*/false);
2305 assert(BC != 0 && "Failed to emit a bitcast!");
2307 ArgVT = VA.getLocVT();
2310 case CCValAssign::VExt:
2311 // VExt has not been implemented, so this should be impossible to reach
2312 // for now. However, fallback to Selection DAG isel once implemented.
2314 case CCValAssign::Indirect:
2315 // FIXME: Indirect doesn't need extending, but fast-isel doesn't fully
2318 case CCValAssign::FPExt:
2319 llvm_unreachable("Unexpected loc info!");
2322 if (VA.isRegLoc()) {
2323 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2324 TII.get(TargetOpcode::COPY), VA.getLocReg()).addReg(Arg);
2325 RegArgs.push_back(VA.getLocReg());
2327 unsigned LocMemOffset = VA.getLocMemOffset();
2329 const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo*>(
2330 getTargetMachine()->getRegisterInfo());
2331 AM.Base.Reg = RegInfo->getStackRegister();
2332 AM.Disp = LocMemOffset;
2333 const Value *ArgVal = ArgVals[VA.getValNo()];
2334 ISD::ArgFlagsTy Flags = ArgFlags[VA.getValNo()];
2336 if (Flags.isByVal()) {
2337 X86AddressMode SrcAM;
2338 SrcAM.Base.Reg = Arg;
2339 bool Res = TryEmitSmallMemcpy(AM, SrcAM, Flags.getByValSize());
2340 assert(Res && "memcpy length already checked!"); (void)Res;
2341 } else if (isa<ConstantInt>(ArgVal) || isa<ConstantPointerNull>(ArgVal)) {
2342 // If this is a really simple value, emit this with the Value* version
2343 // of X86FastEmitStore. If it isn't simple, we don't want to do this,
2344 // as it can cause us to reevaluate the argument.
2345 if (!X86FastEmitStore(ArgVT, ArgVal, AM))
2348 if (!X86FastEmitStore(ArgVT, Arg, /*ValIsKill=*/false, AM))
2354 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2356 if (Subtarget->isPICStyleGOT()) {
2357 unsigned Base = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF);
2358 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2359 TII.get(TargetOpcode::COPY), X86::EBX).addReg(Base);
2362 if (Subtarget->is64Bit() && isVarArg && !isWin64) {
2363 // Count the number of XMM registers allocated.
2364 static const MCPhysReg XMMArgRegs[] = {
2365 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2366 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2368 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2369 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::MOV8ri),
2370 X86::AL).addImm(NumXMMRegs);
2374 MachineInstrBuilder MIB;
2376 // Register-indirect call.
2378 if (Subtarget->is64Bit())
2379 CallOpc = X86::CALL64r;
2381 CallOpc = X86::CALL32r;
2382 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CallOpc))
2387 assert(GV && "Not a direct call");
2389 if (Subtarget->is64Bit())
2390 CallOpc = X86::CALL64pcrel32;
2392 CallOpc = X86::CALLpcrel32;
2394 // See if we need any target-specific flags on the GV operand.
2395 unsigned char OpFlags = 0;
2397 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2398 // external symbols most go through the PLT in PIC mode. If the symbol
2399 // has hidden or protected visibility, or if it is static or local, then
2400 // we don't need to use the PLT - we can directly call it.
2401 if (Subtarget->isTargetELF() &&
2402 TM.getRelocationModel() == Reloc::PIC_ &&
2403 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2404 OpFlags = X86II::MO_PLT;
2405 } else if (Subtarget->isPICStyleStubAny() &&
2406 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2407 (!Subtarget->getTargetTriple().isMacOSX() ||
2408 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2409 // PC-relative references to external symbols should go through $stub,
2410 // unless we're building with the leopard linker or later, which
2411 // automatically synthesizes these stubs.
2412 OpFlags = X86II::MO_DARWIN_STUB;
2416 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CallOpc));
2418 MIB.addExternalSymbol(MemIntName, OpFlags);
2420 MIB.addGlobalAddress(GV, 0, OpFlags);
2423 // Add a register mask with the call-preserved registers.
2424 // Proper defs for return values will be added by setPhysRegsDeadExcept().
2425 MIB.addRegMask(TRI.getCallPreservedMask(CS.getCallingConv()));
2427 // Add an implicit use GOT pointer in EBX.
2428 if (Subtarget->isPICStyleGOT())
2429 MIB.addReg(X86::EBX, RegState::Implicit);
2431 if (Subtarget->is64Bit() && isVarArg && !isWin64)
2432 MIB.addReg(X86::AL, RegState::Implicit);
2434 // Add implicit physical register uses to the call.
2435 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
2436 MIB.addReg(RegArgs[i], RegState::Implicit);
2438 // Issue CALLSEQ_END
2439 unsigned AdjStackUp = TII.getCallFrameDestroyOpcode();
2440 const unsigned NumBytesCallee = computeBytesPoppedByCallee(*Subtarget, CS);
2441 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackUp))
2442 .addImm(NumBytes).addImm(NumBytesCallee);
2444 // Build info for return calling conv lowering code.
2445 // FIXME: This is practically a copy-paste from TargetLowering::LowerCallTo.
2446 SmallVector<ISD::InputArg, 32> Ins;
2447 SmallVector<EVT, 4> RetTys;
2448 ComputeValueVTs(TLI, I->getType(), RetTys);
2449 for (unsigned i = 0, e = RetTys.size(); i != e; ++i) {
2451 MVT RegisterVT = TLI.getRegisterType(I->getParent()->getContext(), VT);
2452 unsigned NumRegs = TLI.getNumRegisters(I->getParent()->getContext(), VT);
2453 for (unsigned j = 0; j != NumRegs; ++j) {
2454 ISD::InputArg MyFlags;
2455 MyFlags.VT = RegisterVT;
2456 MyFlags.Used = !CS.getInstruction()->use_empty();
2457 if (CS.paramHasAttr(0, Attribute::SExt))
2458 MyFlags.Flags.setSExt();
2459 if (CS.paramHasAttr(0, Attribute::ZExt))
2460 MyFlags.Flags.setZExt();
2461 if (CS.paramHasAttr(0, Attribute::InReg))
2462 MyFlags.Flags.setInReg();
2463 Ins.push_back(MyFlags);
2467 // Now handle call return values.
2468 SmallVector<unsigned, 4> UsedRegs;
2469 SmallVector<CCValAssign, 16> RVLocs;
2470 CCState CCRetInfo(CC, false, *FuncInfo.MF, TM, RVLocs,
2471 I->getParent()->getContext());
2472 unsigned ResultReg = FuncInfo.CreateRegs(I->getType());
2473 CCRetInfo.AnalyzeCallResult(Ins, RetCC_X86);
2474 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2475 EVT CopyVT = RVLocs[i].getValVT();
2476 unsigned CopyReg = ResultReg + i;
2478 // If this is a call to a function that returns an fp value on the x87 fp
2479 // stack, but where we prefer to use the value in xmm registers, copy it
2480 // out as F80 and use a truncate to move it from fp stack reg to xmm reg.
2481 if ((RVLocs[i].getLocReg() == X86::ST0 ||
2482 RVLocs[i].getLocReg() == X86::ST1)) {
2483 if (isScalarFPTypeInSSEReg(RVLocs[i].getValVT())) {
2485 CopyReg = createResultReg(&X86::RFP80RegClass);
2487 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2488 TII.get(X86::FpPOP_RETVAL), CopyReg);
2490 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2491 TII.get(TargetOpcode::COPY),
2492 CopyReg).addReg(RVLocs[i].getLocReg());
2493 UsedRegs.push_back(RVLocs[i].getLocReg());
2496 if (CopyVT != RVLocs[i].getValVT()) {
2497 // Round the F80 the right size, which also moves to the appropriate xmm
2498 // register. This is accomplished by storing the F80 value in memory and
2499 // then loading it back. Ewww...
2500 EVT ResVT = RVLocs[i].getValVT();
2501 unsigned Opc = ResVT == MVT::f32 ? X86::ST_Fp80m32 : X86::ST_Fp80m64;
2502 unsigned MemSize = ResVT.getSizeInBits()/8;
2503 int FI = MFI.CreateStackObject(MemSize, MemSize, false);
2504 addFrameReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2507 Opc = ResVT == MVT::f32 ? X86::MOVSSrm : X86::MOVSDrm;
2508 addFrameReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2509 TII.get(Opc), ResultReg + i), FI);
2514 UpdateValueMap(I, ResultReg, RVLocs.size());
2516 // Set all unused physreg defs as dead.
2517 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
2524 X86FastISel::TargetSelectInstruction(const Instruction *I) {
2525 switch (I->getOpcode()) {
2527 case Instruction::Load:
2528 return X86SelectLoad(I);
2529 case Instruction::Store:
2530 return X86SelectStore(I);
2531 case Instruction::Ret:
2532 return X86SelectRet(I);
2533 case Instruction::ICmp:
2534 case Instruction::FCmp:
2535 return X86SelectCmp(I);
2536 case Instruction::ZExt:
2537 return X86SelectZExt(I);
2538 case Instruction::Br:
2539 return X86SelectBranch(I);
2540 case Instruction::Call:
2541 return X86SelectCall(I);
2542 case Instruction::LShr:
2543 case Instruction::AShr:
2544 case Instruction::Shl:
2545 return X86SelectShift(I);
2546 case Instruction::SDiv:
2547 case Instruction::UDiv:
2548 case Instruction::SRem:
2549 case Instruction::URem:
2550 return X86SelectDivRem(I);
2551 case Instruction::Select:
2552 return X86SelectSelect(I);
2553 case Instruction::Trunc:
2554 return X86SelectTrunc(I);
2555 case Instruction::FPExt:
2556 return X86SelectFPExt(I);
2557 case Instruction::FPTrunc:
2558 return X86SelectFPTrunc(I);
2559 case Instruction::IntToPtr: // Deliberate fall-through.
2560 case Instruction::PtrToInt: {
2561 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
2562 EVT DstVT = TLI.getValueType(I->getType());
2563 if (DstVT.bitsGT(SrcVT))
2564 return X86SelectZExt(I);
2565 if (DstVT.bitsLT(SrcVT))
2566 return X86SelectTrunc(I);
2567 unsigned Reg = getRegForValue(I->getOperand(0));
2568 if (Reg == 0) return false;
2569 UpdateValueMap(I, Reg);
2577 unsigned X86FastISel::TargetMaterializeConstant(const Constant *C) {
2579 if (!isTypeLegal(C->getType(), VT))
2582 // Can't handle alternate code models yet.
2583 if (TM.getCodeModel() != CodeModel::Small)
2586 // Get opcode and regclass of the output for the given load instruction.
2588 const TargetRegisterClass *RC = nullptr;
2589 switch (VT.SimpleTy) {
2593 RC = &X86::GR8RegClass;
2597 RC = &X86::GR16RegClass;
2601 RC = &X86::GR32RegClass;
2604 // Must be in x86-64 mode.
2606 RC = &X86::GR64RegClass;
2609 if (X86ScalarSSEf32) {
2610 Opc = Subtarget->hasAVX() ? X86::VMOVSSrm : X86::MOVSSrm;
2611 RC = &X86::FR32RegClass;
2613 Opc = X86::LD_Fp32m;
2614 RC = &X86::RFP32RegClass;
2618 if (X86ScalarSSEf64) {
2619 Opc = Subtarget->hasAVX() ? X86::VMOVSDrm : X86::MOVSDrm;
2620 RC = &X86::FR64RegClass;
2622 Opc = X86::LD_Fp64m;
2623 RC = &X86::RFP64RegClass;
2627 // No f80 support yet.
2631 // Materialize addresses with LEA instructions.
2632 if (isa<GlobalValue>(C)) {
2634 if (X86SelectAddress(C, AM)) {
2635 // If the expression is just a basereg, then we're done, otherwise we need
2637 if (AM.BaseType == X86AddressMode::RegBase &&
2638 AM.IndexReg == 0 && AM.Disp == 0 && AM.GV == nullptr)
2641 Opc = TLI.getPointerTy() == MVT::i32 ? X86::LEA32r : X86::LEA64r;
2642 unsigned ResultReg = createResultReg(RC);
2643 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2644 TII.get(Opc), ResultReg), AM);
2650 // MachineConstantPool wants an explicit alignment.
2651 unsigned Align = DL.getPrefTypeAlignment(C->getType());
2653 // Alignment of vector types. FIXME!
2654 Align = DL.getTypeAllocSize(C->getType());
2657 // x86-32 PIC requires a PIC base register for constant pools.
2658 unsigned PICBase = 0;
2659 unsigned char OpFlag = 0;
2660 if (Subtarget->isPICStyleStubPIC()) { // Not dynamic-no-pic
2661 OpFlag = X86II::MO_PIC_BASE_OFFSET;
2662 PICBase = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF);
2663 } else if (Subtarget->isPICStyleGOT()) {
2664 OpFlag = X86II::MO_GOTOFF;
2665 PICBase = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF);
2666 } else if (Subtarget->isPICStyleRIPRel() &&
2667 TM.getCodeModel() == CodeModel::Small) {
2671 // Create the load from the constant pool.
2672 unsigned MCPOffset = MCP.getConstantPoolIndex(C, Align);
2673 unsigned ResultReg = createResultReg(RC);
2674 addConstantPoolReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2675 TII.get(Opc), ResultReg),
2676 MCPOffset, PICBase, OpFlag);
2681 unsigned X86FastISel::TargetMaterializeAlloca(const AllocaInst *C) {
2682 // Fail on dynamic allocas. At this point, getRegForValue has already
2683 // checked its CSE maps, so if we're here trying to handle a dynamic
2684 // alloca, we're not going to succeed. X86SelectAddress has a
2685 // check for dynamic allocas, because it's called directly from
2686 // various places, but TargetMaterializeAlloca also needs a check
2687 // in order to avoid recursion between getRegForValue,
2688 // X86SelectAddrss, and TargetMaterializeAlloca.
2689 if (!FuncInfo.StaticAllocaMap.count(C))
2691 assert(C->isStaticAlloca() && "dynamic alloca in the static alloca map?");
2694 if (!X86SelectAddress(C, AM))
2696 unsigned Opc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
2697 const TargetRegisterClass* RC = TLI.getRegClassFor(TLI.getPointerTy());
2698 unsigned ResultReg = createResultReg(RC);
2699 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2700 TII.get(Opc), ResultReg), AM);
2704 unsigned X86FastISel::TargetMaterializeFloatZero(const ConstantFP *CF) {
2706 if (!isTypeLegal(CF->getType(), VT))
2709 // Get opcode and regclass for the given zero.
2711 const TargetRegisterClass *RC = nullptr;
2712 switch (VT.SimpleTy) {
2715 if (X86ScalarSSEf32) {
2716 Opc = X86::FsFLD0SS;
2717 RC = &X86::FR32RegClass;
2719 Opc = X86::LD_Fp032;
2720 RC = &X86::RFP32RegClass;
2724 if (X86ScalarSSEf64) {
2725 Opc = X86::FsFLD0SD;
2726 RC = &X86::FR64RegClass;
2728 Opc = X86::LD_Fp064;
2729 RC = &X86::RFP64RegClass;
2733 // No f80 support yet.
2737 unsigned ResultReg = createResultReg(RC);
2738 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg);
2743 bool X86FastISel::tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo,
2744 const LoadInst *LI) {
2745 const Value *Ptr = LI->getPointerOperand();
2747 if (!X86SelectAddress(Ptr, AM))
2750 const X86InstrInfo &XII = (const X86InstrInfo&)TII;
2752 unsigned Size = DL.getTypeAllocSize(LI->getType());
2753 unsigned Alignment = LI->getAlignment();
2755 if (Alignment == 0) // Ensure that codegen never sees alignment 0
2756 Alignment = DL.getABITypeAlignment(LI->getType());
2758 SmallVector<MachineOperand, 8> AddrOps;
2759 AM.getFullAddress(AddrOps);
2761 MachineInstr *Result =
2762 XII.foldMemoryOperandImpl(*FuncInfo.MF, MI, OpNo, AddrOps, Size, Alignment);
2766 Result->addMemOperand(*FuncInfo.MF, createMachineMemOperandFor(LI));
2767 FuncInfo.MBB->insert(FuncInfo.InsertPt, Result);
2768 MI->eraseFromParent();
2774 FastISel *X86::createFastISel(FunctionLoweringInfo &funcInfo,
2775 const TargetLibraryInfo *libInfo) {
2776 return new X86FastISel(funcInfo, libInfo);