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 "X86InstrInfo.h"
20 #include "X86MachineFunctionInfo.h"
21 #include "X86RegisterInfo.h"
22 #include "X86Subtarget.h"
23 #include "X86TargetMachine.h"
24 #include "llvm/Analysis/BranchProbabilityInfo.h"
25 #include "llvm/CodeGen/Analysis.h"
26 #include "llvm/CodeGen/FastISel.h"
27 #include "llvm/CodeGen/FunctionLoweringInfo.h"
28 #include "llvm/CodeGen/MachineConstantPool.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineRegisterInfo.h"
31 #include "llvm/IR/CallSite.h"
32 #include "llvm/IR/CallingConv.h"
33 #include "llvm/IR/DerivedTypes.h"
34 #include "llvm/IR/GetElementPtrTypeIterator.h"
35 #include "llvm/IR/GlobalAlias.h"
36 #include "llvm/IR/GlobalVariable.h"
37 #include "llvm/IR/Instructions.h"
38 #include "llvm/IR/IntrinsicInst.h"
39 #include "llvm/IR/Operator.h"
40 #include "llvm/MC/MCAsmInfo.h"
41 #include "llvm/MC/MCSymbol.h"
42 #include "llvm/Support/ErrorHandling.h"
43 #include "llvm/Target/TargetOptions.h"
48 class X86FastISel final : public FastISel {
49 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
50 /// make the right decision when generating code for different targets.
51 const X86Subtarget *Subtarget;
53 /// X86ScalarSSEf32, X86ScalarSSEf64 - Select between SSE or x87
54 /// floating point ops.
55 /// When SSE is available, use it for f32 operations.
56 /// When SSE2 is available, use it for f64 operations.
61 explicit X86FastISel(FunctionLoweringInfo &funcInfo,
62 const TargetLibraryInfo *libInfo)
63 : FastISel(funcInfo, libInfo) {
64 Subtarget = &funcInfo.MF->getSubtarget<X86Subtarget>();
65 X86ScalarSSEf64 = Subtarget->hasSSE2();
66 X86ScalarSSEf32 = Subtarget->hasSSE1();
69 bool fastSelectInstruction(const Instruction *I) override;
71 /// \brief The specified machine instr operand is a vreg, and that
72 /// vreg is being provided by the specified load instruction. If possible,
73 /// try to fold the load as an operand to the instruction, returning true if
75 bool tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo,
76 const LoadInst *LI) override;
78 bool fastLowerArguments() override;
79 bool fastLowerCall(CallLoweringInfo &CLI) override;
80 bool fastLowerIntrinsicCall(const IntrinsicInst *II) override;
82 #include "X86GenFastISel.inc"
85 bool X86FastEmitCompare(const Value *LHS, const Value *RHS, EVT VT, DebugLoc DL);
87 bool X86FastEmitLoad(EVT VT, X86AddressMode &AM, MachineMemOperand *MMO,
88 unsigned &ResultReg, unsigned Alignment = 1);
90 bool X86FastEmitStore(EVT VT, const Value *Val, X86AddressMode &AM,
91 MachineMemOperand *MMO = nullptr, bool Aligned = false);
92 bool X86FastEmitStore(EVT VT, unsigned ValReg, bool ValIsKill,
94 MachineMemOperand *MMO = nullptr, bool Aligned = false);
96 bool X86FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src, EVT SrcVT,
99 bool X86SelectAddress(const Value *V, X86AddressMode &AM);
100 bool X86SelectCallAddress(const Value *V, X86AddressMode &AM);
102 bool X86SelectLoad(const Instruction *I);
104 bool X86SelectStore(const Instruction *I);
106 bool X86SelectRet(const Instruction *I);
108 bool X86SelectCmp(const Instruction *I);
110 bool X86SelectZExt(const Instruction *I);
112 bool X86SelectBranch(const Instruction *I);
114 bool X86SelectShift(const Instruction *I);
116 bool X86SelectDivRem(const Instruction *I);
118 bool X86FastEmitCMoveSelect(MVT RetVT, const Instruction *I);
120 bool X86FastEmitSSESelect(MVT RetVT, const Instruction *I);
122 bool X86FastEmitPseudoSelect(MVT RetVT, const Instruction *I);
124 bool X86SelectSelect(const Instruction *I);
126 bool X86SelectTrunc(const Instruction *I);
128 bool X86SelectFPExtOrFPTrunc(const Instruction *I, unsigned Opc,
129 const TargetRegisterClass *RC);
131 bool X86SelectFPExt(const Instruction *I);
132 bool X86SelectFPTrunc(const Instruction *I);
133 bool X86SelectSIToFP(const Instruction *I);
135 const X86InstrInfo *getInstrInfo() const {
136 return Subtarget->getInstrInfo();
138 const X86TargetMachine *getTargetMachine() const {
139 return static_cast<const X86TargetMachine *>(&TM);
142 bool handleConstantAddresses(const Value *V, X86AddressMode &AM);
144 unsigned X86MaterializeInt(const ConstantInt *CI, MVT VT);
145 unsigned X86MaterializeFP(const ConstantFP *CFP, MVT VT);
146 unsigned X86MaterializeGV(const GlobalValue *GV, MVT VT);
147 unsigned fastMaterializeConstant(const Constant *C) override;
149 unsigned fastMaterializeAlloca(const AllocaInst *C) override;
151 unsigned fastMaterializeFloatZero(const ConstantFP *CF) override;
153 /// isScalarFPTypeInSSEReg - Return true if the specified scalar FP type is
154 /// computed in an SSE register, not on the X87 floating point stack.
155 bool isScalarFPTypeInSSEReg(EVT VT) const {
156 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
157 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1
160 bool isTypeLegal(Type *Ty, MVT &VT, bool AllowI1 = false);
162 bool IsMemcpySmall(uint64_t Len);
164 bool TryEmitSmallMemcpy(X86AddressMode DestAM,
165 X86AddressMode SrcAM, uint64_t Len);
167 bool foldX86XALUIntrinsic(X86::CondCode &CC, const Instruction *I,
170 const MachineInstrBuilder &addFullAddress(const MachineInstrBuilder &MIB,
174 } // end anonymous namespace.
176 static std::pair<X86::CondCode, bool>
177 getX86ConditionCode(CmpInst::Predicate Predicate) {
178 X86::CondCode CC = X86::COND_INVALID;
179 bool NeedSwap = false;
182 // Floating-point Predicates
183 case CmpInst::FCMP_UEQ: CC = X86::COND_E; break;
184 case CmpInst::FCMP_OLT: NeedSwap = true; // fall-through
185 case CmpInst::FCMP_OGT: CC = X86::COND_A; break;
186 case CmpInst::FCMP_OLE: NeedSwap = true; // fall-through
187 case CmpInst::FCMP_OGE: CC = X86::COND_AE; break;
188 case CmpInst::FCMP_UGT: NeedSwap = true; // fall-through
189 case CmpInst::FCMP_ULT: CC = X86::COND_B; break;
190 case CmpInst::FCMP_UGE: NeedSwap = true; // fall-through
191 case CmpInst::FCMP_ULE: CC = X86::COND_BE; break;
192 case CmpInst::FCMP_ONE: CC = X86::COND_NE; break;
193 case CmpInst::FCMP_UNO: CC = X86::COND_P; break;
194 case CmpInst::FCMP_ORD: CC = X86::COND_NP; break;
195 case CmpInst::FCMP_OEQ: // fall-through
196 case CmpInst::FCMP_UNE: CC = X86::COND_INVALID; break;
198 // Integer Predicates
199 case CmpInst::ICMP_EQ: CC = X86::COND_E; break;
200 case CmpInst::ICMP_NE: CC = X86::COND_NE; break;
201 case CmpInst::ICMP_UGT: CC = X86::COND_A; break;
202 case CmpInst::ICMP_UGE: CC = X86::COND_AE; break;
203 case CmpInst::ICMP_ULT: CC = X86::COND_B; break;
204 case CmpInst::ICMP_ULE: CC = X86::COND_BE; break;
205 case CmpInst::ICMP_SGT: CC = X86::COND_G; break;
206 case CmpInst::ICMP_SGE: CC = X86::COND_GE; break;
207 case CmpInst::ICMP_SLT: CC = X86::COND_L; break;
208 case CmpInst::ICMP_SLE: CC = X86::COND_LE; break;
211 return std::make_pair(CC, NeedSwap);
214 static std::pair<unsigned, bool>
215 getX86SSEConditionCode(CmpInst::Predicate Predicate) {
217 bool NeedSwap = false;
219 // SSE Condition code mapping:
229 default: llvm_unreachable("Unexpected predicate");
230 case CmpInst::FCMP_OEQ: CC = 0; break;
231 case CmpInst::FCMP_OGT: NeedSwap = true; // fall-through
232 case CmpInst::FCMP_OLT: CC = 1; break;
233 case CmpInst::FCMP_OGE: NeedSwap = true; // fall-through
234 case CmpInst::FCMP_OLE: CC = 2; break;
235 case CmpInst::FCMP_UNO: CC = 3; break;
236 case CmpInst::FCMP_UNE: CC = 4; break;
237 case CmpInst::FCMP_ULE: NeedSwap = true; // fall-through
238 case CmpInst::FCMP_UGE: CC = 5; break;
239 case CmpInst::FCMP_ULT: NeedSwap = true; // fall-through
240 case CmpInst::FCMP_UGT: CC = 6; break;
241 case CmpInst::FCMP_ORD: CC = 7; break;
242 case CmpInst::FCMP_UEQ:
243 case CmpInst::FCMP_ONE: CC = 8; break;
246 return std::make_pair(CC, NeedSwap);
249 /// \brief Adds a complex addressing mode to the given machine instr builder.
250 /// Note, this will constrain the index register. If its not possible to
251 /// constrain the given index register, then a new one will be created. The
252 /// IndexReg field of the addressing mode will be updated to match in this case.
253 const MachineInstrBuilder &
254 X86FastISel::addFullAddress(const MachineInstrBuilder &MIB,
255 X86AddressMode &AM) {
256 // First constrain the index register. It needs to be a GR64_NOSP.
257 AM.IndexReg = constrainOperandRegClass(MIB->getDesc(), AM.IndexReg,
258 MIB->getNumOperands() +
260 return ::addFullAddress(MIB, AM);
263 /// \brief Check if it is possible to fold the condition from the XALU intrinsic
264 /// into the user. The condition code will only be updated on success.
265 bool X86FastISel::foldX86XALUIntrinsic(X86::CondCode &CC, const Instruction *I,
267 if (!isa<ExtractValueInst>(Cond))
270 const auto *EV = cast<ExtractValueInst>(Cond);
271 if (!isa<IntrinsicInst>(EV->getAggregateOperand()))
274 const auto *II = cast<IntrinsicInst>(EV->getAggregateOperand());
276 const Function *Callee = II->getCalledFunction();
278 cast<StructType>(Callee->getReturnType())->getTypeAtIndex(0U);
279 if (!isTypeLegal(RetTy, RetVT))
282 if (RetVT != MVT::i32 && RetVT != MVT::i64)
286 switch (II->getIntrinsicID()) {
287 default: return false;
288 case Intrinsic::sadd_with_overflow:
289 case Intrinsic::ssub_with_overflow:
290 case Intrinsic::smul_with_overflow:
291 case Intrinsic::umul_with_overflow: TmpCC = X86::COND_O; break;
292 case Intrinsic::uadd_with_overflow:
293 case Intrinsic::usub_with_overflow: TmpCC = X86::COND_B; break;
296 // Check if both instructions are in the same basic block.
297 if (II->getParent() != I->getParent())
300 // Make sure nothing is in the way
301 BasicBlock::const_iterator Start(I);
302 BasicBlock::const_iterator End(II);
303 for (auto Itr = std::prev(Start); Itr != End; --Itr) {
304 // We only expect extractvalue instructions between the intrinsic and the
305 // instruction to be selected.
306 if (!isa<ExtractValueInst>(Itr))
309 // Check that the extractvalue operand comes from the intrinsic.
310 const auto *EVI = cast<ExtractValueInst>(Itr);
311 if (EVI->getAggregateOperand() != II)
319 bool X86FastISel::isTypeLegal(Type *Ty, MVT &VT, bool AllowI1) {
320 EVT evt = TLI.getValueType(DL, Ty, /*HandleUnknown=*/true);
321 if (evt == MVT::Other || !evt.isSimple())
322 // Unhandled type. Halt "fast" selection and bail.
325 VT = evt.getSimpleVT();
326 // For now, require SSE/SSE2 for performing floating-point operations,
327 // since x87 requires additional work.
328 if (VT == MVT::f64 && !X86ScalarSSEf64)
330 if (VT == MVT::f32 && !X86ScalarSSEf32)
332 // Similarly, no f80 support yet.
335 // We only handle legal types. For example, on x86-32 the instruction
336 // selector contains all of the 64-bit instructions from x86-64,
337 // under the assumption that i64 won't be used if the target doesn't
339 return (AllowI1 && VT == MVT::i1) || TLI.isTypeLegal(VT);
342 #include "X86GenCallingConv.inc"
344 /// X86FastEmitLoad - Emit a machine instruction to load a value of type VT.
345 /// The address is either pre-computed, i.e. Ptr, or a GlobalAddress, i.e. GV.
346 /// Return true and the result register by reference if it is possible.
347 bool X86FastISel::X86FastEmitLoad(EVT VT, X86AddressMode &AM,
348 MachineMemOperand *MMO, unsigned &ResultReg,
349 unsigned Alignment) {
350 // Get opcode and regclass of the output for the given load instruction.
352 const TargetRegisterClass *RC = nullptr;
353 switch (VT.getSimpleVT().SimpleTy) {
354 default: return false;
358 RC = &X86::GR8RegClass;
362 RC = &X86::GR16RegClass;
366 RC = &X86::GR32RegClass;
369 // Must be in x86-64 mode.
371 RC = &X86::GR64RegClass;
374 if (X86ScalarSSEf32) {
375 Opc = Subtarget->hasAVX() ? X86::VMOVSSrm : X86::MOVSSrm;
376 RC = &X86::FR32RegClass;
379 RC = &X86::RFP32RegClass;
383 if (X86ScalarSSEf64) {
384 Opc = Subtarget->hasAVX() ? X86::VMOVSDrm : X86::MOVSDrm;
385 RC = &X86::FR64RegClass;
388 RC = &X86::RFP64RegClass;
392 // No f80 support yet.
396 Opc = Subtarget->hasAVX() ? X86::VMOVAPSrm : X86::MOVAPSrm;
398 Opc = Subtarget->hasAVX() ? X86::VMOVUPSrm : X86::MOVUPSrm;
399 RC = &X86::VR128RegClass;
403 Opc = Subtarget->hasAVX() ? X86::VMOVAPDrm : X86::MOVAPDrm;
405 Opc = Subtarget->hasAVX() ? X86::VMOVUPDrm : X86::MOVUPDrm;
406 RC = &X86::VR128RegClass;
413 Opc = Subtarget->hasAVX() ? X86::VMOVDQArm : X86::MOVDQArm;
415 Opc = Subtarget->hasAVX() ? X86::VMOVDQUrm : X86::MOVDQUrm;
416 RC = &X86::VR128RegClass;
420 ResultReg = createResultReg(RC);
421 MachineInstrBuilder MIB =
422 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg);
423 addFullAddress(MIB, AM);
425 MIB->addMemOperand(*FuncInfo.MF, MMO);
429 /// X86FastEmitStore - Emit a machine instruction to store a value Val of
430 /// type VT. The address is either pre-computed, consisted of a base ptr, Ptr
431 /// and a displacement offset, or a GlobalAddress,
432 /// i.e. V. Return true if it is possible.
433 bool X86FastISel::X86FastEmitStore(EVT VT, unsigned ValReg, bool ValIsKill,
435 MachineMemOperand *MMO, bool Aligned) {
436 bool HasSSE2 = Subtarget->hasSSE2();
437 bool HasSSE4A = Subtarget->hasSSE4A();
438 bool HasAVX = Subtarget->hasAVX();
439 bool IsNonTemporal = MMO && MMO->isNonTemporal();
441 // Get opcode and regclass of the output for the given store instruction.
443 switch (VT.getSimpleVT().SimpleTy) {
444 case MVT::f80: // No f80 support yet.
445 default: return false;
447 // Mask out all but lowest bit.
448 unsigned AndResult = createResultReg(&X86::GR8RegClass);
449 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
450 TII.get(X86::AND8ri), AndResult)
451 .addReg(ValReg, getKillRegState(ValIsKill)).addImm(1);
454 // FALLTHROUGH, handling i1 as i8.
455 case MVT::i8: Opc = X86::MOV8mr; break;
456 case MVT::i16: Opc = X86::MOV16mr; break;
458 Opc = (IsNonTemporal && HasSSE2) ? X86::MOVNTImr : X86::MOV32mr;
461 // Must be in x86-64 mode.
462 Opc = (IsNonTemporal && HasSSE2) ? X86::MOVNTI_64mr : X86::MOV64mr;
465 if (X86ScalarSSEf32) {
466 if (IsNonTemporal && HasSSE4A)
469 Opc = HasAVX ? X86::VMOVSSmr : X86::MOVSSmr;
474 if (X86ScalarSSEf32) {
475 if (IsNonTemporal && HasSSE4A)
478 Opc = HasAVX ? X86::VMOVSDmr : X86::MOVSDmr;
485 Opc = HasAVX ? X86::VMOVNTPSmr : X86::MOVNTPSmr;
487 Opc = HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr;
489 Opc = HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr;
494 Opc = HasAVX ? X86::VMOVNTPDmr : X86::MOVNTPDmr;
496 Opc = HasAVX ? X86::VMOVAPDmr : X86::MOVAPDmr;
498 Opc = HasAVX ? X86::VMOVUPDmr : X86::MOVUPDmr;
506 Opc = HasAVX ? X86::VMOVNTDQmr : X86::MOVNTDQmr;
508 Opc = HasAVX ? X86::VMOVDQAmr : X86::MOVDQAmr;
510 Opc = Subtarget->hasAVX() ? X86::VMOVDQUmr : X86::MOVDQUmr;
514 MachineInstrBuilder MIB =
515 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc));
516 addFullAddress(MIB, AM).addReg(ValReg, getKillRegState(ValIsKill));
518 MIB->addMemOperand(*FuncInfo.MF, MMO);
523 bool X86FastISel::X86FastEmitStore(EVT VT, const Value *Val,
525 MachineMemOperand *MMO, bool Aligned) {
526 // Handle 'null' like i32/i64 0.
527 if (isa<ConstantPointerNull>(Val))
528 Val = Constant::getNullValue(DL.getIntPtrType(Val->getContext()));
530 // If this is a store of a simple constant, fold the constant into the store.
531 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
534 switch (VT.getSimpleVT().SimpleTy) {
536 case MVT::i1: Signed = false; // FALLTHROUGH to handle as i8.
537 case MVT::i8: Opc = X86::MOV8mi; break;
538 case MVT::i16: Opc = X86::MOV16mi; break;
539 case MVT::i32: Opc = X86::MOV32mi; break;
541 // Must be a 32-bit sign extended value.
542 if (isInt<32>(CI->getSExtValue()))
543 Opc = X86::MOV64mi32;
548 MachineInstrBuilder MIB =
549 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc));
550 addFullAddress(MIB, AM).addImm(Signed ? (uint64_t) CI->getSExtValue()
551 : CI->getZExtValue());
553 MIB->addMemOperand(*FuncInfo.MF, MMO);
558 unsigned ValReg = getRegForValue(Val);
562 bool ValKill = hasTrivialKill(Val);
563 return X86FastEmitStore(VT, ValReg, ValKill, AM, MMO, Aligned);
566 /// X86FastEmitExtend - Emit a machine instruction to extend a value Src of
567 /// type SrcVT to type DstVT using the specified extension opcode Opc (e.g.
568 /// ISD::SIGN_EXTEND).
569 bool X86FastISel::X86FastEmitExtend(ISD::NodeType Opc, EVT DstVT,
570 unsigned Src, EVT SrcVT,
571 unsigned &ResultReg) {
572 unsigned RR = fastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), Opc,
573 Src, /*TODO: Kill=*/false);
581 bool X86FastISel::handleConstantAddresses(const Value *V, X86AddressMode &AM) {
582 // Handle constant address.
583 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
584 // Can't handle alternate code models yet.
585 if (TM.getCodeModel() != CodeModel::Small)
588 // Can't handle TLS yet.
589 if (GV->isThreadLocal())
592 // RIP-relative addresses can't have additional register operands, so if
593 // we've already folded stuff into the addressing mode, just force the
594 // global value into its own register, which we can use as the basereg.
595 if (!Subtarget->isPICStyleRIPRel() ||
596 (AM.Base.Reg == 0 && AM.IndexReg == 0)) {
597 // Okay, we've committed to selecting this global. Set up the address.
600 // Allow the subtarget to classify the global.
601 unsigned char GVFlags = Subtarget->ClassifyGlobalReference(GV, TM);
603 // If this reference is relative to the pic base, set it now.
604 if (isGlobalRelativeToPICBase(GVFlags)) {
605 // FIXME: How do we know Base.Reg is free??
606 AM.Base.Reg = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF);
609 // Unless the ABI requires an extra load, return a direct reference to
611 if (!isGlobalStubReference(GVFlags)) {
612 if (Subtarget->isPICStyleRIPRel()) {
613 // Use rip-relative addressing if we can. Above we verified that the
614 // base and index registers are unused.
615 assert(AM.Base.Reg == 0 && AM.IndexReg == 0);
616 AM.Base.Reg = X86::RIP;
618 AM.GVOpFlags = GVFlags;
622 // Ok, we need to do a load from a stub. If we've already loaded from
623 // this stub, reuse the loaded pointer, otherwise emit the load now.
624 DenseMap<const Value *, unsigned>::iterator I = LocalValueMap.find(V);
626 if (I != LocalValueMap.end() && I->second != 0) {
629 // Issue load from stub.
631 const TargetRegisterClass *RC = nullptr;
632 X86AddressMode StubAM;
633 StubAM.Base.Reg = AM.Base.Reg;
635 StubAM.GVOpFlags = GVFlags;
637 // Prepare for inserting code in the local-value area.
638 SavePoint SaveInsertPt = enterLocalValueArea();
640 if (TLI.getPointerTy(DL) == MVT::i64) {
642 RC = &X86::GR64RegClass;
644 if (Subtarget->isPICStyleRIPRel())
645 StubAM.Base.Reg = X86::RIP;
648 RC = &X86::GR32RegClass;
651 LoadReg = createResultReg(RC);
652 MachineInstrBuilder LoadMI =
653 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), LoadReg);
654 addFullAddress(LoadMI, StubAM);
656 // Ok, back to normal mode.
657 leaveLocalValueArea(SaveInsertPt);
659 // Prevent loading GV stub multiple times in same MBB.
660 LocalValueMap[V] = LoadReg;
663 // Now construct the final address. Note that the Disp, Scale,
664 // and Index values may already be set here.
665 AM.Base.Reg = LoadReg;
671 // If all else fails, try to materialize the value in a register.
672 if (!AM.GV || !Subtarget->isPICStyleRIPRel()) {
673 if (AM.Base.Reg == 0) {
674 AM.Base.Reg = getRegForValue(V);
675 return AM.Base.Reg != 0;
677 if (AM.IndexReg == 0) {
678 assert(AM.Scale == 1 && "Scale with no index!");
679 AM.IndexReg = getRegForValue(V);
680 return AM.IndexReg != 0;
687 /// X86SelectAddress - Attempt to fill in an address from the given value.
689 bool X86FastISel::X86SelectAddress(const Value *V, X86AddressMode &AM) {
690 SmallVector<const Value *, 32> GEPs;
692 const User *U = nullptr;
693 unsigned Opcode = Instruction::UserOp1;
694 if (const Instruction *I = dyn_cast<Instruction>(V)) {
695 // Don't walk into other basic blocks; it's possible we haven't
696 // visited them yet, so the instructions may not yet be assigned
697 // virtual registers.
698 if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(V)) ||
699 FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
700 Opcode = I->getOpcode();
703 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(V)) {
704 Opcode = C->getOpcode();
708 if (PointerType *Ty = dyn_cast<PointerType>(V->getType()))
709 if (Ty->getAddressSpace() > 255)
710 // Fast instruction selection doesn't support the special
716 case Instruction::BitCast:
717 // Look past bitcasts.
718 return X86SelectAddress(U->getOperand(0), AM);
720 case Instruction::IntToPtr:
721 // Look past no-op inttoptrs.
722 if (TLI.getValueType(DL, U->getOperand(0)->getType()) ==
723 TLI.getPointerTy(DL))
724 return X86SelectAddress(U->getOperand(0), AM);
727 case Instruction::PtrToInt:
728 // Look past no-op ptrtoints.
729 if (TLI.getValueType(DL, U->getType()) == TLI.getPointerTy(DL))
730 return X86SelectAddress(U->getOperand(0), AM);
733 case Instruction::Alloca: {
734 // Do static allocas.
735 const AllocaInst *A = cast<AllocaInst>(V);
736 DenseMap<const AllocaInst *, int>::iterator SI =
737 FuncInfo.StaticAllocaMap.find(A);
738 if (SI != FuncInfo.StaticAllocaMap.end()) {
739 AM.BaseType = X86AddressMode::FrameIndexBase;
740 AM.Base.FrameIndex = SI->second;
746 case Instruction::Add: {
747 // Adds of constants are common and easy enough.
748 if (const ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
749 uint64_t Disp = (int32_t)AM.Disp + (uint64_t)CI->getSExtValue();
750 // They have to fit in the 32-bit signed displacement field though.
751 if (isInt<32>(Disp)) {
752 AM.Disp = (uint32_t)Disp;
753 return X86SelectAddress(U->getOperand(0), AM);
759 case Instruction::GetElementPtr: {
760 X86AddressMode SavedAM = AM;
762 // Pattern-match simple GEPs.
763 uint64_t Disp = (int32_t)AM.Disp;
764 unsigned IndexReg = AM.IndexReg;
765 unsigned Scale = AM.Scale;
766 gep_type_iterator GTI = gep_type_begin(U);
767 // Iterate through the indices, folding what we can. Constants can be
768 // folded, and one dynamic index can be handled, if the scale is supported.
769 for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end();
770 i != e; ++i, ++GTI) {
771 const Value *Op = *i;
772 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
773 const StructLayout *SL = DL.getStructLayout(STy);
774 Disp += SL->getElementOffset(cast<ConstantInt>(Op)->getZExtValue());
778 // A array/variable index is always of the form i*S where S is the
779 // constant scale size. See if we can push the scale into immediates.
780 uint64_t S = DL.getTypeAllocSize(GTI.getIndexedType());
782 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
783 // Constant-offset addressing.
784 Disp += CI->getSExtValue() * S;
787 if (canFoldAddIntoGEP(U, Op)) {
788 // A compatible add with a constant operand. Fold the constant.
790 cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
791 Disp += CI->getSExtValue() * S;
792 // Iterate on the other operand.
793 Op = cast<AddOperator>(Op)->getOperand(0);
797 (!AM.GV || !Subtarget->isPICStyleRIPRel()) &&
798 (S == 1 || S == 2 || S == 4 || S == 8)) {
799 // Scaled-index addressing.
801 IndexReg = getRegForGEPIndex(Op).first;
807 goto unsupported_gep;
811 // Check for displacement overflow.
812 if (!isInt<32>(Disp))
815 AM.IndexReg = IndexReg;
817 AM.Disp = (uint32_t)Disp;
820 if (const GetElementPtrInst *GEP =
821 dyn_cast<GetElementPtrInst>(U->getOperand(0))) {
822 // Ok, the GEP indices were covered by constant-offset and scaled-index
823 // addressing. Update the address state and move on to examining the base.
826 } else if (X86SelectAddress(U->getOperand(0), AM)) {
830 // If we couldn't merge the gep value into this addr mode, revert back to
831 // our address and just match the value instead of completely failing.
834 for (SmallVectorImpl<const Value *>::reverse_iterator
835 I = GEPs.rbegin(), E = GEPs.rend(); I != E; ++I)
836 if (handleConstantAddresses(*I, AM))
841 // Ok, the GEP indices weren't all covered.
846 return handleConstantAddresses(V, AM);
849 /// X86SelectCallAddress - Attempt to fill in an address from the given value.
851 bool X86FastISel::X86SelectCallAddress(const Value *V, X86AddressMode &AM) {
852 const User *U = nullptr;
853 unsigned Opcode = Instruction::UserOp1;
854 const Instruction *I = dyn_cast<Instruction>(V);
855 // Record if the value is defined in the same basic block.
857 // This information is crucial to know whether or not folding an
859 // Indeed, FastISel generates or reuses a virtual register for all
860 // operands of all instructions it selects. Obviously, the definition and
861 // its uses must use the same virtual register otherwise the produced
862 // code is incorrect.
863 // Before instruction selection, FunctionLoweringInfo::set sets the virtual
864 // registers for values that are alive across basic blocks. This ensures
865 // that the values are consistently set between across basic block, even
866 // if different instruction selection mechanisms are used (e.g., a mix of
867 // SDISel and FastISel).
868 // For values local to a basic block, the instruction selection process
869 // generates these virtual registers with whatever method is appropriate
870 // for its needs. In particular, FastISel and SDISel do not share the way
871 // local virtual registers are set.
872 // Therefore, this is impossible (or at least unsafe) to share values
873 // between basic blocks unless they use the same instruction selection
874 // method, which is not guarantee for X86.
875 // Moreover, things like hasOneUse could not be used accurately, if we
876 // allow to reference values across basic blocks whereas they are not
877 // alive across basic blocks initially.
880 Opcode = I->getOpcode();
882 InMBB = I->getParent() == FuncInfo.MBB->getBasicBlock();
883 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(V)) {
884 Opcode = C->getOpcode();
890 case Instruction::BitCast:
891 // Look past bitcasts if its operand is in the same BB.
893 return X86SelectCallAddress(U->getOperand(0), AM);
896 case Instruction::IntToPtr:
897 // Look past no-op inttoptrs if its operand is in the same BB.
899 TLI.getValueType(DL, U->getOperand(0)->getType()) ==
900 TLI.getPointerTy(DL))
901 return X86SelectCallAddress(U->getOperand(0), AM);
904 case Instruction::PtrToInt:
905 // Look past no-op ptrtoints if its operand is in the same BB.
906 if (InMBB && TLI.getValueType(DL, U->getType()) == TLI.getPointerTy(DL))
907 return X86SelectCallAddress(U->getOperand(0), AM);
911 // Handle constant address.
912 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
913 // Can't handle alternate code models yet.
914 if (TM.getCodeModel() != CodeModel::Small)
917 // RIP-relative addresses can't have additional register operands.
918 if (Subtarget->isPICStyleRIPRel() &&
919 (AM.Base.Reg != 0 || AM.IndexReg != 0))
922 // Can't handle DLL Import.
923 if (GV->hasDLLImportStorageClass())
927 if (const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV))
928 if (GVar->isThreadLocal())
931 // Okay, we've committed to selecting this global. Set up the basic address.
934 // No ABI requires an extra load for anything other than DLLImport, which
935 // we rejected above. Return a direct reference to the global.
936 if (Subtarget->isPICStyleRIPRel()) {
937 // Use rip-relative addressing if we can. Above we verified that the
938 // base and index registers are unused.
939 assert(AM.Base.Reg == 0 && AM.IndexReg == 0);
940 AM.Base.Reg = X86::RIP;
941 } else if (Subtarget->isPICStyleStubPIC()) {
942 AM.GVOpFlags = X86II::MO_PIC_BASE_OFFSET;
943 } else if (Subtarget->isPICStyleGOT()) {
944 AM.GVOpFlags = X86II::MO_GOTOFF;
950 // If all else fails, try to materialize the value in a register.
951 if (!AM.GV || !Subtarget->isPICStyleRIPRel()) {
952 if (AM.Base.Reg == 0) {
953 AM.Base.Reg = getRegForValue(V);
954 return AM.Base.Reg != 0;
956 if (AM.IndexReg == 0) {
957 assert(AM.Scale == 1 && "Scale with no index!");
958 AM.IndexReg = getRegForValue(V);
959 return AM.IndexReg != 0;
967 /// X86SelectStore - Select and emit code to implement store instructions.
968 bool X86FastISel::X86SelectStore(const Instruction *I) {
969 // Atomic stores need special handling.
970 const StoreInst *S = cast<StoreInst>(I);
975 const Value *Val = S->getValueOperand();
976 const Value *Ptr = S->getPointerOperand();
979 if (!isTypeLegal(Val->getType(), VT, /*AllowI1=*/true))
982 unsigned Alignment = S->getAlignment();
983 unsigned ABIAlignment = DL.getABITypeAlignment(Val->getType());
984 if (Alignment == 0) // Ensure that codegen never sees alignment 0
985 Alignment = ABIAlignment;
986 bool Aligned = Alignment >= ABIAlignment;
989 if (!X86SelectAddress(Ptr, AM))
992 return X86FastEmitStore(VT, Val, AM, createMachineMemOperandFor(I), Aligned);
995 /// X86SelectRet - Select and emit code to implement ret instructions.
996 bool X86FastISel::X86SelectRet(const Instruction *I) {
997 const ReturnInst *Ret = cast<ReturnInst>(I);
998 const Function &F = *I->getParent()->getParent();
999 const X86MachineFunctionInfo *X86MFInfo =
1000 FuncInfo.MF->getInfo<X86MachineFunctionInfo>();
1002 if (!FuncInfo.CanLowerReturn)
1005 CallingConv::ID CC = F.getCallingConv();
1006 if (CC != CallingConv::C &&
1007 CC != CallingConv::Fast &&
1008 CC != CallingConv::X86_FastCall &&
1009 CC != CallingConv::X86_64_SysV)
1012 if (Subtarget->isCallingConvWin64(CC))
1015 // Don't handle popping bytes on return for now.
1016 if (X86MFInfo->getBytesToPopOnReturn() != 0)
1019 // fastcc with -tailcallopt is intended to provide a guaranteed
1020 // tail call optimization. Fastisel doesn't know how to do that.
1021 if (CC == CallingConv::Fast && TM.Options.GuaranteedTailCallOpt)
1024 // Let SDISel handle vararg functions.
1028 // Build a list of return value registers.
1029 SmallVector<unsigned, 4> RetRegs;
1031 if (Ret->getNumOperands() > 0) {
1032 SmallVector<ISD::OutputArg, 4> Outs;
1033 GetReturnInfo(F.getReturnType(), F.getAttributes(), Outs, TLI, DL);
1035 // Analyze operands of the call, assigning locations to each operand.
1036 SmallVector<CCValAssign, 16> ValLocs;
1037 CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, ValLocs, I->getContext());
1038 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1040 const Value *RV = Ret->getOperand(0);
1041 unsigned Reg = getRegForValue(RV);
1045 // Only handle a single return value for now.
1046 if (ValLocs.size() != 1)
1049 CCValAssign &VA = ValLocs[0];
1051 // Don't bother handling odd stuff for now.
1052 if (VA.getLocInfo() != CCValAssign::Full)
1054 // Only handle register returns for now.
1058 // The calling-convention tables for x87 returns don't tell
1060 if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
1063 unsigned SrcReg = Reg + VA.getValNo();
1064 EVT SrcVT = TLI.getValueType(DL, RV->getType());
1065 EVT DstVT = VA.getValVT();
1066 // Special handling for extended integers.
1067 if (SrcVT != DstVT) {
1068 if (SrcVT != MVT::i1 && SrcVT != MVT::i8 && SrcVT != MVT::i16)
1071 if (!Outs[0].Flags.isZExt() && !Outs[0].Flags.isSExt())
1074 assert(DstVT == MVT::i32 && "X86 should always ext to i32");
1076 if (SrcVT == MVT::i1) {
1077 if (Outs[0].Flags.isSExt())
1079 SrcReg = fastEmitZExtFromI1(MVT::i8, SrcReg, /*TODO: Kill=*/false);
1082 unsigned Op = Outs[0].Flags.isZExt() ? ISD::ZERO_EXTEND :
1084 SrcReg = fastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), Op,
1085 SrcReg, /*TODO: Kill=*/false);
1089 unsigned DstReg = VA.getLocReg();
1090 const TargetRegisterClass *SrcRC = MRI.getRegClass(SrcReg);
1091 // Avoid a cross-class copy. This is very unlikely.
1092 if (!SrcRC->contains(DstReg))
1094 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1095 TII.get(TargetOpcode::COPY), DstReg).addReg(SrcReg);
1097 // Add register to return instruction.
1098 RetRegs.push_back(VA.getLocReg());
1101 // All x86 ABIs require that for returning structs by value we copy
1102 // the sret argument into %rax/%eax (depending on ABI) for the return.
1103 // We saved the argument into a virtual register in the entry block,
1104 // so now we copy the value out and into %rax/%eax.
1105 if (F.hasStructRetAttr()) {
1106 unsigned Reg = X86MFInfo->getSRetReturnReg();
1108 "SRetReturnReg should have been set in LowerFormalArguments()!");
1109 unsigned RetReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
1110 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1111 TII.get(TargetOpcode::COPY), RetReg).addReg(Reg);
1112 RetRegs.push_back(RetReg);
1115 // Now emit the RET.
1116 MachineInstrBuilder MIB =
1117 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1118 TII.get(Subtarget->is64Bit() ? X86::RETQ : X86::RETL));
1119 for (unsigned i = 0, e = RetRegs.size(); i != e; ++i)
1120 MIB.addReg(RetRegs[i], RegState::Implicit);
1124 /// X86SelectLoad - Select and emit code to implement load instructions.
1126 bool X86FastISel::X86SelectLoad(const Instruction *I) {
1127 const LoadInst *LI = cast<LoadInst>(I);
1129 // Atomic loads need special handling.
1134 if (!isTypeLegal(LI->getType(), VT, /*AllowI1=*/true))
1137 const Value *Ptr = LI->getPointerOperand();
1140 if (!X86SelectAddress(Ptr, AM))
1143 unsigned Alignment = LI->getAlignment();
1144 unsigned ABIAlignment = DL.getABITypeAlignment(LI->getType());
1145 if (Alignment == 0) // Ensure that codegen never sees alignment 0
1146 Alignment = ABIAlignment;
1148 unsigned ResultReg = 0;
1149 if (!X86FastEmitLoad(VT, AM, createMachineMemOperandFor(LI), ResultReg,
1153 updateValueMap(I, ResultReg);
1157 static unsigned X86ChooseCmpOpcode(EVT VT, const X86Subtarget *Subtarget) {
1158 bool HasAVX = Subtarget->hasAVX();
1159 bool X86ScalarSSEf32 = Subtarget->hasSSE1();
1160 bool X86ScalarSSEf64 = Subtarget->hasSSE2();
1162 switch (VT.getSimpleVT().SimpleTy) {
1164 case MVT::i8: return X86::CMP8rr;
1165 case MVT::i16: return X86::CMP16rr;
1166 case MVT::i32: return X86::CMP32rr;
1167 case MVT::i64: return X86::CMP64rr;
1169 return X86ScalarSSEf32 ? (HasAVX ? X86::VUCOMISSrr : X86::UCOMISSrr) : 0;
1171 return X86ScalarSSEf64 ? (HasAVX ? X86::VUCOMISDrr : X86::UCOMISDrr) : 0;
1175 /// If we have a comparison with RHS as the RHS of the comparison, return an
1176 /// opcode that works for the compare (e.g. CMP32ri) otherwise return 0.
1177 static unsigned X86ChooseCmpImmediateOpcode(EVT VT, const ConstantInt *RHSC) {
1178 int64_t Val = RHSC->getSExtValue();
1179 switch (VT.getSimpleVT().SimpleTy) {
1180 // Otherwise, we can't fold the immediate into this comparison.
1187 return X86::CMP16ri8;
1188 return X86::CMP16ri;
1191 return X86::CMP32ri8;
1192 return X86::CMP32ri;
1195 return X86::CMP64ri8;
1196 // 64-bit comparisons are only valid if the immediate fits in a 32-bit sext
1199 return X86::CMP64ri32;
1204 bool X86FastISel::X86FastEmitCompare(const Value *Op0, const Value *Op1,
1205 EVT VT, DebugLoc CurDbgLoc) {
1206 unsigned Op0Reg = getRegForValue(Op0);
1207 if (Op0Reg == 0) return false;
1209 // Handle 'null' like i32/i64 0.
1210 if (isa<ConstantPointerNull>(Op1))
1211 Op1 = Constant::getNullValue(DL.getIntPtrType(Op0->getContext()));
1213 // We have two options: compare with register or immediate. If the RHS of
1214 // the compare is an immediate that we can fold into this compare, use
1215 // CMPri, otherwise use CMPrr.
1216 if (const ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
1217 if (unsigned CompareImmOpc = X86ChooseCmpImmediateOpcode(VT, Op1C)) {
1218 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, CurDbgLoc, TII.get(CompareImmOpc))
1220 .addImm(Op1C->getSExtValue());
1225 unsigned CompareOpc = X86ChooseCmpOpcode(VT, Subtarget);
1226 if (CompareOpc == 0) return false;
1228 unsigned Op1Reg = getRegForValue(Op1);
1229 if (Op1Reg == 0) return false;
1230 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, CurDbgLoc, TII.get(CompareOpc))
1237 bool X86FastISel::X86SelectCmp(const Instruction *I) {
1238 const CmpInst *CI = cast<CmpInst>(I);
1241 if (!isTypeLegal(I->getOperand(0)->getType(), VT))
1244 // Try to optimize or fold the cmp.
1245 CmpInst::Predicate Predicate = optimizeCmpPredicate(CI);
1246 unsigned ResultReg = 0;
1247 switch (Predicate) {
1249 case CmpInst::FCMP_FALSE: {
1250 ResultReg = createResultReg(&X86::GR32RegClass);
1251 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::MOV32r0),
1253 ResultReg = fastEmitInst_extractsubreg(MVT::i8, ResultReg, /*Kill=*/true,
1259 case CmpInst::FCMP_TRUE: {
1260 ResultReg = createResultReg(&X86::GR8RegClass);
1261 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::MOV8ri),
1262 ResultReg).addImm(1);
1268 updateValueMap(I, ResultReg);
1272 const Value *LHS = CI->getOperand(0);
1273 const Value *RHS = CI->getOperand(1);
1275 // The optimizer might have replaced fcmp oeq %x, %x with fcmp ord %x, 0.0.
1276 // We don't have to materialize a zero constant for this case and can just use
1277 // %x again on the RHS.
1278 if (Predicate == CmpInst::FCMP_ORD || Predicate == CmpInst::FCMP_UNO) {
1279 const auto *RHSC = dyn_cast<ConstantFP>(RHS);
1280 if (RHSC && RHSC->isNullValue())
1284 // FCMP_OEQ and FCMP_UNE cannot be checked with a single instruction.
1285 static unsigned SETFOpcTable[2][3] = {
1286 { X86::SETEr, X86::SETNPr, X86::AND8rr },
1287 { X86::SETNEr, X86::SETPr, X86::OR8rr }
1289 unsigned *SETFOpc = nullptr;
1290 switch (Predicate) {
1292 case CmpInst::FCMP_OEQ: SETFOpc = &SETFOpcTable[0][0]; break;
1293 case CmpInst::FCMP_UNE: SETFOpc = &SETFOpcTable[1][0]; break;
1296 ResultReg = createResultReg(&X86::GR8RegClass);
1298 if (!X86FastEmitCompare(LHS, RHS, VT, I->getDebugLoc()))
1301 unsigned FlagReg1 = createResultReg(&X86::GR8RegClass);
1302 unsigned FlagReg2 = createResultReg(&X86::GR8RegClass);
1303 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(SETFOpc[0]),
1305 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(SETFOpc[1]),
1307 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(SETFOpc[2]),
1308 ResultReg).addReg(FlagReg1).addReg(FlagReg2);
1309 updateValueMap(I, ResultReg);
1315 std::tie(CC, SwapArgs) = getX86ConditionCode(Predicate);
1316 assert(CC <= X86::LAST_VALID_COND && "Unexpected condition code.");
1317 unsigned Opc = X86::getSETFromCond(CC);
1320 std::swap(LHS, RHS);
1322 // Emit a compare of LHS/RHS.
1323 if (!X86FastEmitCompare(LHS, RHS, VT, I->getDebugLoc()))
1326 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg);
1327 updateValueMap(I, ResultReg);
1331 bool X86FastISel::X86SelectZExt(const Instruction *I) {
1332 EVT DstVT = TLI.getValueType(DL, I->getType());
1333 if (!TLI.isTypeLegal(DstVT))
1336 unsigned ResultReg = getRegForValue(I->getOperand(0));
1340 // Handle zero-extension from i1 to i8, which is common.
1341 MVT SrcVT = TLI.getSimpleValueType(DL, I->getOperand(0)->getType());
1342 if (SrcVT.SimpleTy == MVT::i1) {
1343 // Set the high bits to zero.
1344 ResultReg = fastEmitZExtFromI1(MVT::i8, ResultReg, /*TODO: Kill=*/false);
1351 if (DstVT == MVT::i64) {
1352 // Handle extension to 64-bits via sub-register shenanigans.
1355 switch (SrcVT.SimpleTy) {
1356 case MVT::i8: MovInst = X86::MOVZX32rr8; break;
1357 case MVT::i16: MovInst = X86::MOVZX32rr16; break;
1358 case MVT::i32: MovInst = X86::MOV32rr; break;
1359 default: llvm_unreachable("Unexpected zext to i64 source type");
1362 unsigned Result32 = createResultReg(&X86::GR32RegClass);
1363 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(MovInst), Result32)
1366 ResultReg = createResultReg(&X86::GR64RegClass);
1367 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TargetOpcode::SUBREG_TO_REG),
1369 .addImm(0).addReg(Result32).addImm(X86::sub_32bit);
1370 } else if (DstVT != MVT::i8) {
1371 ResultReg = fastEmit_r(MVT::i8, DstVT.getSimpleVT(), ISD::ZERO_EXTEND,
1372 ResultReg, /*Kill=*/true);
1377 updateValueMap(I, ResultReg);
1381 bool X86FastISel::X86SelectBranch(const Instruction *I) {
1382 // Unconditional branches are selected by tablegen-generated code.
1383 // Handle a conditional branch.
1384 const BranchInst *BI = cast<BranchInst>(I);
1385 MachineBasicBlock *TrueMBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
1386 MachineBasicBlock *FalseMBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
1388 // Fold the common case of a conditional branch with a comparison
1389 // in the same block (values defined on other blocks may not have
1390 // initialized registers).
1392 if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
1393 if (CI->hasOneUse() && CI->getParent() == I->getParent()) {
1394 EVT VT = TLI.getValueType(DL, CI->getOperand(0)->getType());
1396 // Try to optimize or fold the cmp.
1397 CmpInst::Predicate Predicate = optimizeCmpPredicate(CI);
1398 switch (Predicate) {
1400 case CmpInst::FCMP_FALSE: fastEmitBranch(FalseMBB, DbgLoc); return true;
1401 case CmpInst::FCMP_TRUE: fastEmitBranch(TrueMBB, DbgLoc); return true;
1404 const Value *CmpLHS = CI->getOperand(0);
1405 const Value *CmpRHS = CI->getOperand(1);
1407 // The optimizer might have replaced fcmp oeq %x, %x with fcmp ord %x,
1409 // We don't have to materialize a zero constant for this case and can just
1410 // use %x again on the RHS.
1411 if (Predicate == CmpInst::FCMP_ORD || Predicate == CmpInst::FCMP_UNO) {
1412 const auto *CmpRHSC = dyn_cast<ConstantFP>(CmpRHS);
1413 if (CmpRHSC && CmpRHSC->isNullValue())
1417 // Try to take advantage of fallthrough opportunities.
1418 if (FuncInfo.MBB->isLayoutSuccessor(TrueMBB)) {
1419 std::swap(TrueMBB, FalseMBB);
1420 Predicate = CmpInst::getInversePredicate(Predicate);
1423 // FCMP_OEQ and FCMP_UNE cannot be expressed with a single flag/condition
1424 // code check. Instead two branch instructions are required to check all
1425 // the flags. First we change the predicate to a supported condition code,
1426 // which will be the first branch. Later one we will emit the second
1428 bool NeedExtraBranch = false;
1429 switch (Predicate) {
1431 case CmpInst::FCMP_OEQ:
1432 std::swap(TrueMBB, FalseMBB); // fall-through
1433 case CmpInst::FCMP_UNE:
1434 NeedExtraBranch = true;
1435 Predicate = CmpInst::FCMP_ONE;
1441 std::tie(CC, SwapArgs) = getX86ConditionCode(Predicate);
1442 assert(CC <= X86::LAST_VALID_COND && "Unexpected condition code.");
1444 BranchOpc = X86::GetCondBranchFromCond(CC);
1446 std::swap(CmpLHS, CmpRHS);
1448 // Emit a compare of the LHS and RHS, setting the flags.
1449 if (!X86FastEmitCompare(CmpLHS, CmpRHS, VT, CI->getDebugLoc()))
1452 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(BranchOpc))
1455 // X86 requires a second branch to handle UNE (and OEQ, which is mapped
1457 if (NeedExtraBranch) {
1458 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::JP_1))
1462 finishCondBranch(BI->getParent(), TrueMBB, FalseMBB);
1465 } else if (TruncInst *TI = dyn_cast<TruncInst>(BI->getCondition())) {
1466 // Handle things like "%cond = trunc i32 %X to i1 / br i1 %cond", which
1467 // typically happen for _Bool and C++ bools.
1469 if (TI->hasOneUse() && TI->getParent() == I->getParent() &&
1470 isTypeLegal(TI->getOperand(0)->getType(), SourceVT)) {
1471 unsigned TestOpc = 0;
1472 switch (SourceVT.SimpleTy) {
1474 case MVT::i8: TestOpc = X86::TEST8ri; break;
1475 case MVT::i16: TestOpc = X86::TEST16ri; break;
1476 case MVT::i32: TestOpc = X86::TEST32ri; break;
1477 case MVT::i64: TestOpc = X86::TEST64ri32; break;
1480 unsigned OpReg = getRegForValue(TI->getOperand(0));
1481 if (OpReg == 0) return false;
1482 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TestOpc))
1483 .addReg(OpReg).addImm(1);
1485 unsigned JmpOpc = X86::JNE_1;
1486 if (FuncInfo.MBB->isLayoutSuccessor(TrueMBB)) {
1487 std::swap(TrueMBB, FalseMBB);
1491 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(JmpOpc))
1494 finishCondBranch(BI->getParent(), TrueMBB, FalseMBB);
1498 } else if (foldX86XALUIntrinsic(CC, BI, BI->getCondition())) {
1499 // Fake request the condition, otherwise the intrinsic might be completely
1501 unsigned TmpReg = getRegForValue(BI->getCondition());
1505 unsigned BranchOpc = X86::GetCondBranchFromCond(CC);
1507 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(BranchOpc))
1509 finishCondBranch(BI->getParent(), TrueMBB, FalseMBB);
1513 // Otherwise do a clumsy setcc and re-test it.
1514 // Note that i1 essentially gets ANY_EXTEND'ed to i8 where it isn't used
1515 // in an explicit cast, so make sure to handle that correctly.
1516 unsigned OpReg = getRegForValue(BI->getCondition());
1517 if (OpReg == 0) return false;
1519 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::TEST8ri))
1520 .addReg(OpReg).addImm(1);
1521 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::JNE_1))
1523 finishCondBranch(BI->getParent(), TrueMBB, FalseMBB);
1527 bool X86FastISel::X86SelectShift(const Instruction *I) {
1528 unsigned CReg = 0, OpReg = 0;
1529 const TargetRegisterClass *RC = nullptr;
1530 if (I->getType()->isIntegerTy(8)) {
1532 RC = &X86::GR8RegClass;
1533 switch (I->getOpcode()) {
1534 case Instruction::LShr: OpReg = X86::SHR8rCL; break;
1535 case Instruction::AShr: OpReg = X86::SAR8rCL; break;
1536 case Instruction::Shl: OpReg = X86::SHL8rCL; break;
1537 default: return false;
1539 } else if (I->getType()->isIntegerTy(16)) {
1541 RC = &X86::GR16RegClass;
1542 switch (I->getOpcode()) {
1543 case Instruction::LShr: OpReg = X86::SHR16rCL; break;
1544 case Instruction::AShr: OpReg = X86::SAR16rCL; break;
1545 case Instruction::Shl: OpReg = X86::SHL16rCL; break;
1546 default: return false;
1548 } else if (I->getType()->isIntegerTy(32)) {
1550 RC = &X86::GR32RegClass;
1551 switch (I->getOpcode()) {
1552 case Instruction::LShr: OpReg = X86::SHR32rCL; break;
1553 case Instruction::AShr: OpReg = X86::SAR32rCL; break;
1554 case Instruction::Shl: OpReg = X86::SHL32rCL; break;
1555 default: return false;
1557 } else if (I->getType()->isIntegerTy(64)) {
1559 RC = &X86::GR64RegClass;
1560 switch (I->getOpcode()) {
1561 case Instruction::LShr: OpReg = X86::SHR64rCL; break;
1562 case Instruction::AShr: OpReg = X86::SAR64rCL; break;
1563 case Instruction::Shl: OpReg = X86::SHL64rCL; break;
1564 default: return false;
1571 if (!isTypeLegal(I->getType(), VT))
1574 unsigned Op0Reg = getRegForValue(I->getOperand(0));
1575 if (Op0Reg == 0) return false;
1577 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1578 if (Op1Reg == 0) return false;
1579 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TargetOpcode::COPY),
1580 CReg).addReg(Op1Reg);
1582 // The shift instruction uses X86::CL. If we defined a super-register
1583 // of X86::CL, emit a subreg KILL to precisely describe what we're doing here.
1584 if (CReg != X86::CL)
1585 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1586 TII.get(TargetOpcode::KILL), X86::CL)
1587 .addReg(CReg, RegState::Kill);
1589 unsigned ResultReg = createResultReg(RC);
1590 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(OpReg), ResultReg)
1592 updateValueMap(I, ResultReg);
1596 bool X86FastISel::X86SelectDivRem(const Instruction *I) {
1597 const static unsigned NumTypes = 4; // i8, i16, i32, i64
1598 const static unsigned NumOps = 4; // SDiv, SRem, UDiv, URem
1599 const static bool S = true; // IsSigned
1600 const static bool U = false; // !IsSigned
1601 const static unsigned Copy = TargetOpcode::COPY;
1602 // For the X86 DIV/IDIV instruction, in most cases the dividend
1603 // (numerator) must be in a specific register pair highreg:lowreg,
1604 // producing the quotient in lowreg and the remainder in highreg.
1605 // For most data types, to set up the instruction, the dividend is
1606 // copied into lowreg, and lowreg is sign-extended or zero-extended
1607 // into highreg. The exception is i8, where the dividend is defined
1608 // as a single register rather than a register pair, and we
1609 // therefore directly sign-extend or zero-extend the dividend into
1610 // lowreg, instead of copying, and ignore the highreg.
1611 const static struct DivRemEntry {
1612 // The following portion depends only on the data type.
1613 const TargetRegisterClass *RC;
1614 unsigned LowInReg; // low part of the register pair
1615 unsigned HighInReg; // high part of the register pair
1616 // The following portion depends on both the data type and the operation.
1617 struct DivRemResult {
1618 unsigned OpDivRem; // The specific DIV/IDIV opcode to use.
1619 unsigned OpSignExtend; // Opcode for sign-extending lowreg into
1620 // highreg, or copying a zero into highreg.
1621 unsigned OpCopy; // Opcode for copying dividend into lowreg, or
1622 // zero/sign-extending into lowreg for i8.
1623 unsigned DivRemResultReg; // Register containing the desired result.
1624 bool IsOpSigned; // Whether to use signed or unsigned form.
1625 } ResultTable[NumOps];
1626 } OpTable[NumTypes] = {
1627 { &X86::GR8RegClass, X86::AX, 0, {
1628 { X86::IDIV8r, 0, X86::MOVSX16rr8, X86::AL, S }, // SDiv
1629 { X86::IDIV8r, 0, X86::MOVSX16rr8, X86::AH, S }, // SRem
1630 { X86::DIV8r, 0, X86::MOVZX16rr8, X86::AL, U }, // UDiv
1631 { X86::DIV8r, 0, X86::MOVZX16rr8, X86::AH, U }, // URem
1634 { &X86::GR16RegClass, X86::AX, X86::DX, {
1635 { X86::IDIV16r, X86::CWD, Copy, X86::AX, S }, // SDiv
1636 { X86::IDIV16r, X86::CWD, Copy, X86::DX, S }, // SRem
1637 { X86::DIV16r, X86::MOV32r0, Copy, X86::AX, U }, // UDiv
1638 { X86::DIV16r, X86::MOV32r0, Copy, X86::DX, U }, // URem
1641 { &X86::GR32RegClass, X86::EAX, X86::EDX, {
1642 { X86::IDIV32r, X86::CDQ, Copy, X86::EAX, S }, // SDiv
1643 { X86::IDIV32r, X86::CDQ, Copy, X86::EDX, S }, // SRem
1644 { X86::DIV32r, X86::MOV32r0, Copy, X86::EAX, U }, // UDiv
1645 { X86::DIV32r, X86::MOV32r0, Copy, X86::EDX, U }, // URem
1648 { &X86::GR64RegClass, X86::RAX, X86::RDX, {
1649 { X86::IDIV64r, X86::CQO, Copy, X86::RAX, S }, // SDiv
1650 { X86::IDIV64r, X86::CQO, Copy, X86::RDX, S }, // SRem
1651 { X86::DIV64r, X86::MOV32r0, Copy, X86::RAX, U }, // UDiv
1652 { X86::DIV64r, X86::MOV32r0, Copy, X86::RDX, U }, // URem
1658 if (!isTypeLegal(I->getType(), VT))
1661 unsigned TypeIndex, OpIndex;
1662 switch (VT.SimpleTy) {
1663 default: return false;
1664 case MVT::i8: TypeIndex = 0; break;
1665 case MVT::i16: TypeIndex = 1; break;
1666 case MVT::i32: TypeIndex = 2; break;
1667 case MVT::i64: TypeIndex = 3;
1668 if (!Subtarget->is64Bit())
1673 switch (I->getOpcode()) {
1674 default: llvm_unreachable("Unexpected div/rem opcode");
1675 case Instruction::SDiv: OpIndex = 0; break;
1676 case Instruction::SRem: OpIndex = 1; break;
1677 case Instruction::UDiv: OpIndex = 2; break;
1678 case Instruction::URem: OpIndex = 3; break;
1681 const DivRemEntry &TypeEntry = OpTable[TypeIndex];
1682 const DivRemEntry::DivRemResult &OpEntry = TypeEntry.ResultTable[OpIndex];
1683 unsigned Op0Reg = getRegForValue(I->getOperand(0));
1686 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1690 // Move op0 into low-order input register.
1691 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1692 TII.get(OpEntry.OpCopy), TypeEntry.LowInReg).addReg(Op0Reg);
1693 // Zero-extend or sign-extend into high-order input register.
1694 if (OpEntry.OpSignExtend) {
1695 if (OpEntry.IsOpSigned)
1696 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1697 TII.get(OpEntry.OpSignExtend));
1699 unsigned Zero32 = createResultReg(&X86::GR32RegClass);
1700 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1701 TII.get(X86::MOV32r0), Zero32);
1703 // Copy the zero into the appropriate sub/super/identical physical
1704 // register. Unfortunately the operations needed are not uniform enough
1705 // to fit neatly into the table above.
1706 if (VT.SimpleTy == MVT::i16) {
1707 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1708 TII.get(Copy), TypeEntry.HighInReg)
1709 .addReg(Zero32, 0, X86::sub_16bit);
1710 } else if (VT.SimpleTy == MVT::i32) {
1711 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1712 TII.get(Copy), TypeEntry.HighInReg)
1714 } else if (VT.SimpleTy == MVT::i64) {
1715 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1716 TII.get(TargetOpcode::SUBREG_TO_REG), TypeEntry.HighInReg)
1717 .addImm(0).addReg(Zero32).addImm(X86::sub_32bit);
1721 // Generate the DIV/IDIV instruction.
1722 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1723 TII.get(OpEntry.OpDivRem)).addReg(Op1Reg);
1724 // For i8 remainder, we can't reference AH directly, as we'll end
1725 // up with bogus copies like %R9B = COPY %AH. Reference AX
1726 // instead to prevent AH references in a REX instruction.
1728 // The current assumption of the fast register allocator is that isel
1729 // won't generate explicit references to the GPR8_NOREX registers. If
1730 // the allocator and/or the backend get enhanced to be more robust in
1731 // that regard, this can be, and should be, removed.
1732 unsigned ResultReg = 0;
1733 if ((I->getOpcode() == Instruction::SRem ||
1734 I->getOpcode() == Instruction::URem) &&
1735 OpEntry.DivRemResultReg == X86::AH && Subtarget->is64Bit()) {
1736 unsigned SourceSuperReg = createResultReg(&X86::GR16RegClass);
1737 unsigned ResultSuperReg = createResultReg(&X86::GR16RegClass);
1738 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1739 TII.get(Copy), SourceSuperReg).addReg(X86::AX);
1741 // Shift AX right by 8 bits instead of using AH.
1742 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::SHR16ri),
1743 ResultSuperReg).addReg(SourceSuperReg).addImm(8);
1745 // Now reference the 8-bit subreg of the result.
1746 ResultReg = fastEmitInst_extractsubreg(MVT::i8, ResultSuperReg,
1747 /*Kill=*/true, X86::sub_8bit);
1749 // Copy the result out of the physreg if we haven't already.
1751 ResultReg = createResultReg(TypeEntry.RC);
1752 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Copy), ResultReg)
1753 .addReg(OpEntry.DivRemResultReg);
1755 updateValueMap(I, ResultReg);
1760 /// \brief Emit a conditional move instruction (if the are supported) to lower
1762 bool X86FastISel::X86FastEmitCMoveSelect(MVT RetVT, const Instruction *I) {
1763 // Check if the subtarget supports these instructions.
1764 if (!Subtarget->hasCMov())
1767 // FIXME: Add support for i8.
1768 if (RetVT < MVT::i16 || RetVT > MVT::i64)
1771 const Value *Cond = I->getOperand(0);
1772 const TargetRegisterClass *RC = TLI.getRegClassFor(RetVT);
1773 bool NeedTest = true;
1774 X86::CondCode CC = X86::COND_NE;
1776 // Optimize conditions coming from a compare if both instructions are in the
1777 // same basic block (values defined in other basic blocks may not have
1778 // initialized registers).
1779 const auto *CI = dyn_cast<CmpInst>(Cond);
1780 if (CI && (CI->getParent() == I->getParent())) {
1781 CmpInst::Predicate Predicate = optimizeCmpPredicate(CI);
1783 // FCMP_OEQ and FCMP_UNE cannot be checked with a single instruction.
1784 static unsigned SETFOpcTable[2][3] = {
1785 { X86::SETNPr, X86::SETEr , X86::TEST8rr },
1786 { X86::SETPr, X86::SETNEr, X86::OR8rr }
1788 unsigned *SETFOpc = nullptr;
1789 switch (Predicate) {
1791 case CmpInst::FCMP_OEQ:
1792 SETFOpc = &SETFOpcTable[0][0];
1793 Predicate = CmpInst::ICMP_NE;
1795 case CmpInst::FCMP_UNE:
1796 SETFOpc = &SETFOpcTable[1][0];
1797 Predicate = CmpInst::ICMP_NE;
1802 std::tie(CC, NeedSwap) = getX86ConditionCode(Predicate);
1803 assert(CC <= X86::LAST_VALID_COND && "Unexpected condition code.");
1805 const Value *CmpLHS = CI->getOperand(0);
1806 const Value *CmpRHS = CI->getOperand(1);
1808 std::swap(CmpLHS, CmpRHS);
1810 EVT CmpVT = TLI.getValueType(DL, CmpLHS->getType());
1811 // Emit a compare of the LHS and RHS, setting the flags.
1812 if (!X86FastEmitCompare(CmpLHS, CmpRHS, CmpVT, CI->getDebugLoc()))
1816 unsigned FlagReg1 = createResultReg(&X86::GR8RegClass);
1817 unsigned FlagReg2 = createResultReg(&X86::GR8RegClass);
1818 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(SETFOpc[0]),
1820 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(SETFOpc[1]),
1822 auto const &II = TII.get(SETFOpc[2]);
1823 if (II.getNumDefs()) {
1824 unsigned TmpReg = createResultReg(&X86::GR8RegClass);
1825 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, TmpReg)
1826 .addReg(FlagReg2).addReg(FlagReg1);
1828 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
1829 .addReg(FlagReg2).addReg(FlagReg1);
1833 } else if (foldX86XALUIntrinsic(CC, I, Cond)) {
1834 // Fake request the condition, otherwise the intrinsic might be completely
1836 unsigned TmpReg = getRegForValue(Cond);
1844 // Selects operate on i1, however, CondReg is 8 bits width and may contain
1845 // garbage. Indeed, only the less significant bit is supposed to be
1846 // accurate. If we read more than the lsb, we may see non-zero values
1847 // whereas lsb is zero. Therefore, we have to truncate Op0Reg to i1 for
1848 // the select. This is achieved by performing TEST against 1.
1849 unsigned CondReg = getRegForValue(Cond);
1852 bool CondIsKill = hasTrivialKill(Cond);
1854 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::TEST8ri))
1855 .addReg(CondReg, getKillRegState(CondIsKill)).addImm(1);
1858 const Value *LHS = I->getOperand(1);
1859 const Value *RHS = I->getOperand(2);
1861 unsigned RHSReg = getRegForValue(RHS);
1862 bool RHSIsKill = hasTrivialKill(RHS);
1864 unsigned LHSReg = getRegForValue(LHS);
1865 bool LHSIsKill = hasTrivialKill(LHS);
1867 if (!LHSReg || !RHSReg)
1870 unsigned Opc = X86::getCMovFromCond(CC, RC->getSize());
1871 unsigned ResultReg = fastEmitInst_rr(Opc, RC, RHSReg, RHSIsKill,
1873 updateValueMap(I, ResultReg);
1877 /// \brief Emit SSE or AVX instructions to lower the select.
1879 /// Try to use SSE1/SSE2 instructions to simulate a select without branches.
1880 /// This lowers fp selects into a CMP/AND/ANDN/OR sequence when the necessary
1881 /// SSE instructions are available. If AVX is available, try to use a VBLENDV.
1882 bool X86FastISel::X86FastEmitSSESelect(MVT RetVT, const Instruction *I) {
1883 // Optimize conditions coming from a compare if both instructions are in the
1884 // same basic block (values defined in other basic blocks may not have
1885 // initialized registers).
1886 const auto *CI = dyn_cast<FCmpInst>(I->getOperand(0));
1887 if (!CI || (CI->getParent() != I->getParent()))
1890 if (I->getType() != CI->getOperand(0)->getType() ||
1891 !((Subtarget->hasSSE1() && RetVT == MVT::f32) ||
1892 (Subtarget->hasSSE2() && RetVT == MVT::f64)))
1895 const Value *CmpLHS = CI->getOperand(0);
1896 const Value *CmpRHS = CI->getOperand(1);
1897 CmpInst::Predicate Predicate = optimizeCmpPredicate(CI);
1899 // The optimizer might have replaced fcmp oeq %x, %x with fcmp ord %x, 0.0.
1900 // We don't have to materialize a zero constant for this case and can just use
1901 // %x again on the RHS.
1902 if (Predicate == CmpInst::FCMP_ORD || Predicate == CmpInst::FCMP_UNO) {
1903 const auto *CmpRHSC = dyn_cast<ConstantFP>(CmpRHS);
1904 if (CmpRHSC && CmpRHSC->isNullValue())
1910 std::tie(CC, NeedSwap) = getX86SSEConditionCode(Predicate);
1915 std::swap(CmpLHS, CmpRHS);
1917 // Choose the SSE instruction sequence based on data type (float or double).
1918 static unsigned OpcTable[2][4] = {
1919 { X86::CMPSSrr, X86::FsANDPSrr, X86::FsANDNPSrr, X86::FsORPSrr },
1920 { X86::CMPSDrr, X86::FsANDPDrr, X86::FsANDNPDrr, X86::FsORPDrr }
1923 unsigned *Opc = nullptr;
1924 switch (RetVT.SimpleTy) {
1925 default: return false;
1926 case MVT::f32: Opc = &OpcTable[0][0]; break;
1927 case MVT::f64: Opc = &OpcTable[1][0]; break;
1930 const Value *LHS = I->getOperand(1);
1931 const Value *RHS = I->getOperand(2);
1933 unsigned LHSReg = getRegForValue(LHS);
1934 bool LHSIsKill = hasTrivialKill(LHS);
1936 unsigned RHSReg = getRegForValue(RHS);
1937 bool RHSIsKill = hasTrivialKill(RHS);
1939 unsigned CmpLHSReg = getRegForValue(CmpLHS);
1940 bool CmpLHSIsKill = hasTrivialKill(CmpLHS);
1942 unsigned CmpRHSReg = getRegForValue(CmpRHS);
1943 bool CmpRHSIsKill = hasTrivialKill(CmpRHS);
1945 if (!LHSReg || !RHSReg || !CmpLHS || !CmpRHS)
1948 const TargetRegisterClass *RC = TLI.getRegClassFor(RetVT);
1951 if (Subtarget->hasAVX()) {
1952 const TargetRegisterClass *FR32 = &X86::FR32RegClass;
1953 const TargetRegisterClass *VR128 = &X86::VR128RegClass;
1955 // If we have AVX, create 1 blendv instead of 3 logic instructions.
1956 // Blendv was introduced with SSE 4.1, but the 2 register form implicitly
1957 // uses XMM0 as the selection register. That may need just as many
1958 // instructions as the AND/ANDN/OR sequence due to register moves, so
1960 unsigned CmpOpcode =
1961 (RetVT.SimpleTy == MVT::f32) ? X86::VCMPSSrr : X86::VCMPSDrr;
1962 unsigned BlendOpcode =
1963 (RetVT.SimpleTy == MVT::f32) ? X86::VBLENDVPSrr : X86::VBLENDVPDrr;
1965 unsigned CmpReg = fastEmitInst_rri(CmpOpcode, FR32, CmpLHSReg, CmpLHSIsKill,
1966 CmpRHSReg, CmpRHSIsKill, CC);
1967 unsigned VBlendReg = fastEmitInst_rrr(BlendOpcode, VR128, RHSReg, RHSIsKill,
1968 LHSReg, LHSIsKill, CmpReg, true);
1969 ResultReg = createResultReg(RC);
1970 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1971 TII.get(TargetOpcode::COPY), ResultReg).addReg(VBlendReg);
1973 unsigned CmpReg = fastEmitInst_rri(Opc[0], RC, CmpLHSReg, CmpLHSIsKill,
1974 CmpRHSReg, CmpRHSIsKill, CC);
1975 unsigned AndReg = fastEmitInst_rr(Opc[1], RC, CmpReg, /*IsKill=*/false,
1977 unsigned AndNReg = fastEmitInst_rr(Opc[2], RC, CmpReg, /*IsKill=*/true,
1979 ResultReg = fastEmitInst_rr(Opc[3], RC, AndNReg, /*IsKill=*/true,
1980 AndReg, /*IsKill=*/true);
1982 updateValueMap(I, ResultReg);
1986 bool X86FastISel::X86FastEmitPseudoSelect(MVT RetVT, const Instruction *I) {
1987 // These are pseudo CMOV instructions and will be later expanded into control-
1990 switch (RetVT.SimpleTy) {
1991 default: return false;
1992 case MVT::i8: Opc = X86::CMOV_GR8; break;
1993 case MVT::i16: Opc = X86::CMOV_GR16; break;
1994 case MVT::i32: Opc = X86::CMOV_GR32; break;
1995 case MVT::f32: Opc = X86::CMOV_FR32; break;
1996 case MVT::f64: Opc = X86::CMOV_FR64; break;
1999 const Value *Cond = I->getOperand(0);
2000 X86::CondCode CC = X86::COND_NE;
2002 // Optimize conditions coming from a compare if both instructions are in the
2003 // same basic block (values defined in other basic blocks may not have
2004 // initialized registers).
2005 const auto *CI = dyn_cast<CmpInst>(Cond);
2006 if (CI && (CI->getParent() == I->getParent())) {
2008 std::tie(CC, NeedSwap) = getX86ConditionCode(CI->getPredicate());
2009 if (CC > X86::LAST_VALID_COND)
2012 const Value *CmpLHS = CI->getOperand(0);
2013 const Value *CmpRHS = CI->getOperand(1);
2016 std::swap(CmpLHS, CmpRHS);
2018 EVT CmpVT = TLI.getValueType(DL, CmpLHS->getType());
2019 if (!X86FastEmitCompare(CmpLHS, CmpRHS, CmpVT, CI->getDebugLoc()))
2022 unsigned CondReg = getRegForValue(Cond);
2025 bool CondIsKill = hasTrivialKill(Cond);
2026 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::TEST8ri))
2027 .addReg(CondReg, getKillRegState(CondIsKill)).addImm(1);
2030 const Value *LHS = I->getOperand(1);
2031 const Value *RHS = I->getOperand(2);
2033 unsigned LHSReg = getRegForValue(LHS);
2034 bool LHSIsKill = hasTrivialKill(LHS);
2036 unsigned RHSReg = getRegForValue(RHS);
2037 bool RHSIsKill = hasTrivialKill(RHS);
2039 if (!LHSReg || !RHSReg)
2042 const TargetRegisterClass *RC = TLI.getRegClassFor(RetVT);
2044 unsigned ResultReg =
2045 fastEmitInst_rri(Opc, RC, RHSReg, RHSIsKill, LHSReg, LHSIsKill, CC);
2046 updateValueMap(I, ResultReg);
2050 bool X86FastISel::X86SelectSelect(const Instruction *I) {
2052 if (!isTypeLegal(I->getType(), RetVT))
2055 // Check if we can fold the select.
2056 if (const auto *CI = dyn_cast<CmpInst>(I->getOperand(0))) {
2057 CmpInst::Predicate Predicate = optimizeCmpPredicate(CI);
2058 const Value *Opnd = nullptr;
2059 switch (Predicate) {
2061 case CmpInst::FCMP_FALSE: Opnd = I->getOperand(2); break;
2062 case CmpInst::FCMP_TRUE: Opnd = I->getOperand(1); break;
2064 // No need for a select anymore - this is an unconditional move.
2066 unsigned OpReg = getRegForValue(Opnd);
2069 bool OpIsKill = hasTrivialKill(Opnd);
2070 const TargetRegisterClass *RC = TLI.getRegClassFor(RetVT);
2071 unsigned ResultReg = createResultReg(RC);
2072 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2073 TII.get(TargetOpcode::COPY), ResultReg)
2074 .addReg(OpReg, getKillRegState(OpIsKill));
2075 updateValueMap(I, ResultReg);
2080 // First try to use real conditional move instructions.
2081 if (X86FastEmitCMoveSelect(RetVT, I))
2084 // Try to use a sequence of SSE instructions to simulate a conditional move.
2085 if (X86FastEmitSSESelect(RetVT, I))
2088 // Fall-back to pseudo conditional move instructions, which will be later
2089 // converted to control-flow.
2090 if (X86FastEmitPseudoSelect(RetVT, I))
2096 bool X86FastISel::X86SelectSIToFP(const Instruction *I) {
2097 // The target-independent selection algorithm in FastISel already knows how
2098 // to select a SINT_TO_FP if the target is SSE but not AVX.
2099 // Early exit if the subtarget doesn't have AVX.
2100 if (!Subtarget->hasAVX())
2103 if (!I->getOperand(0)->getType()->isIntegerTy(32))
2106 // Select integer to float/double conversion.
2107 unsigned OpReg = getRegForValue(I->getOperand(0));
2111 const TargetRegisterClass *RC = nullptr;
2114 if (I->getType()->isDoubleTy()) {
2115 // sitofp int -> double
2116 Opcode = X86::VCVTSI2SDrr;
2117 RC = &X86::FR64RegClass;
2118 } else if (I->getType()->isFloatTy()) {
2119 // sitofp int -> float
2120 Opcode = X86::VCVTSI2SSrr;
2121 RC = &X86::FR32RegClass;
2125 unsigned ImplicitDefReg = createResultReg(RC);
2126 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2127 TII.get(TargetOpcode::IMPLICIT_DEF), ImplicitDefReg);
2128 unsigned ResultReg =
2129 fastEmitInst_rr(Opcode, RC, ImplicitDefReg, true, OpReg, false);
2130 updateValueMap(I, ResultReg);
2134 // Helper method used by X86SelectFPExt and X86SelectFPTrunc.
2135 bool X86FastISel::X86SelectFPExtOrFPTrunc(const Instruction *I,
2137 const TargetRegisterClass *RC) {
2138 assert((I->getOpcode() == Instruction::FPExt ||
2139 I->getOpcode() == Instruction::FPTrunc) &&
2140 "Instruction must be an FPExt or FPTrunc!");
2142 unsigned OpReg = getRegForValue(I->getOperand(0));
2146 unsigned ResultReg = createResultReg(RC);
2147 MachineInstrBuilder MIB;
2148 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TargetOpc),
2150 if (Subtarget->hasAVX())
2153 updateValueMap(I, ResultReg);
2157 bool X86FastISel::X86SelectFPExt(const Instruction *I) {
2158 if (X86ScalarSSEf64 && I->getType()->isDoubleTy() &&
2159 I->getOperand(0)->getType()->isFloatTy()) {
2160 // fpext from float to double.
2161 unsigned Opc = Subtarget->hasAVX() ? X86::VCVTSS2SDrr : X86::CVTSS2SDrr;
2162 return X86SelectFPExtOrFPTrunc(I, Opc, &X86::FR64RegClass);
2168 bool X86FastISel::X86SelectFPTrunc(const Instruction *I) {
2169 if (X86ScalarSSEf64 && I->getType()->isFloatTy() &&
2170 I->getOperand(0)->getType()->isDoubleTy()) {
2171 // fptrunc from double to float.
2172 unsigned Opc = Subtarget->hasAVX() ? X86::VCVTSD2SSrr : X86::CVTSD2SSrr;
2173 return X86SelectFPExtOrFPTrunc(I, Opc, &X86::FR32RegClass);
2179 bool X86FastISel::X86SelectTrunc(const Instruction *I) {
2180 EVT SrcVT = TLI.getValueType(DL, I->getOperand(0)->getType());
2181 EVT DstVT = TLI.getValueType(DL, I->getType());
2183 // This code only handles truncation to byte.
2184 if (DstVT != MVT::i8 && DstVT != MVT::i1)
2186 if (!TLI.isTypeLegal(SrcVT))
2189 unsigned InputReg = getRegForValue(I->getOperand(0));
2191 // Unhandled operand. Halt "fast" selection and bail.
2194 if (SrcVT == MVT::i8) {
2195 // Truncate from i8 to i1; no code needed.
2196 updateValueMap(I, InputReg);
2200 bool KillInputReg = false;
2201 if (!Subtarget->is64Bit()) {
2202 // If we're on x86-32; we can't extract an i8 from a general register.
2203 // First issue a copy to GR16_ABCD or GR32_ABCD.
2204 const TargetRegisterClass *CopyRC =
2205 (SrcVT == MVT::i16) ? &X86::GR16_ABCDRegClass : &X86::GR32_ABCDRegClass;
2206 unsigned CopyReg = createResultReg(CopyRC);
2207 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2208 TII.get(TargetOpcode::COPY), CopyReg).addReg(InputReg);
2210 KillInputReg = true;
2213 // Issue an extract_subreg.
2214 unsigned ResultReg = fastEmitInst_extractsubreg(MVT::i8,
2215 InputReg, KillInputReg,
2220 updateValueMap(I, ResultReg);
2224 bool X86FastISel::IsMemcpySmall(uint64_t Len) {
2225 return Len <= (Subtarget->is64Bit() ? 32 : 16);
2228 bool X86FastISel::TryEmitSmallMemcpy(X86AddressMode DestAM,
2229 X86AddressMode SrcAM, uint64_t Len) {
2231 // Make sure we don't bloat code by inlining very large memcpy's.
2232 if (!IsMemcpySmall(Len))
2235 bool i64Legal = Subtarget->is64Bit();
2237 // We don't care about alignment here since we just emit integer accesses.
2240 if (Len >= 8 && i64Legal)
2250 bool RV = X86FastEmitLoad(VT, SrcAM, nullptr, Reg);
2251 RV &= X86FastEmitStore(VT, Reg, /*Kill=*/true, DestAM);
2252 assert(RV && "Failed to emit load or store??");
2254 unsigned Size = VT.getSizeInBits()/8;
2256 DestAM.Disp += Size;
2263 bool X86FastISel::fastLowerIntrinsicCall(const IntrinsicInst *II) {
2264 // FIXME: Handle more intrinsics.
2265 switch (II->getIntrinsicID()) {
2266 default: return false;
2267 case Intrinsic::convert_from_fp16:
2268 case Intrinsic::convert_to_fp16: {
2269 if (Subtarget->useSoftFloat() || !Subtarget->hasF16C())
2272 const Value *Op = II->getArgOperand(0);
2273 unsigned InputReg = getRegForValue(Op);
2277 // F16C only allows converting from float to half and from half to float.
2278 bool IsFloatToHalf = II->getIntrinsicID() == Intrinsic::convert_to_fp16;
2279 if (IsFloatToHalf) {
2280 if (!Op->getType()->isFloatTy())
2283 if (!II->getType()->isFloatTy())
2287 unsigned ResultReg = 0;
2288 const TargetRegisterClass *RC = TLI.getRegClassFor(MVT::v8i16);
2289 if (IsFloatToHalf) {
2290 // 'InputReg' is implicitly promoted from register class FR32 to
2291 // register class VR128 by method 'constrainOperandRegClass' which is
2292 // directly called by 'fastEmitInst_ri'.
2293 // Instruction VCVTPS2PHrr takes an extra immediate operand which is
2294 // used to provide rounding control.
2295 InputReg = fastEmitInst_ri(X86::VCVTPS2PHrr, RC, InputReg, false, 0);
2297 // Move the lower 32-bits of ResultReg to another register of class GR32.
2298 ResultReg = createResultReg(&X86::GR32RegClass);
2299 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2300 TII.get(X86::VMOVPDI2DIrr), ResultReg)
2301 .addReg(InputReg, RegState::Kill);
2303 // The result value is in the lower 16-bits of ResultReg.
2304 unsigned RegIdx = X86::sub_16bit;
2305 ResultReg = fastEmitInst_extractsubreg(MVT::i16, ResultReg, true, RegIdx);
2307 assert(Op->getType()->isIntegerTy(16) && "Expected a 16-bit integer!");
2308 // Explicitly sign-extend the input to 32-bit.
2309 InputReg = fastEmit_r(MVT::i16, MVT::i32, ISD::SIGN_EXTEND, InputReg,
2312 // The following SCALAR_TO_VECTOR will be expanded into a VMOVDI2PDIrr.
2313 InputReg = fastEmit_r(MVT::i32, MVT::v4i32, ISD::SCALAR_TO_VECTOR,
2314 InputReg, /*Kill=*/true);
2316 InputReg = fastEmitInst_r(X86::VCVTPH2PSrr, RC, InputReg, /*Kill=*/true);
2318 // The result value is in the lower 32-bits of ResultReg.
2319 // Emit an explicit copy from register class VR128 to register class FR32.
2320 ResultReg = createResultReg(&X86::FR32RegClass);
2321 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2322 TII.get(TargetOpcode::COPY), ResultReg)
2323 .addReg(InputReg, RegState::Kill);
2326 updateValueMap(II, ResultReg);
2329 case Intrinsic::frameaddress: {
2330 MachineFunction *MF = FuncInfo.MF;
2331 if (MF->getTarget().getMCAsmInfo()->usesWindowsCFI())
2334 Type *RetTy = II->getCalledFunction()->getReturnType();
2337 if (!isTypeLegal(RetTy, VT))
2341 const TargetRegisterClass *RC = nullptr;
2343 switch (VT.SimpleTy) {
2344 default: llvm_unreachable("Invalid result type for frameaddress.");
2345 case MVT::i32: Opc = X86::MOV32rm; RC = &X86::GR32RegClass; break;
2346 case MVT::i64: Opc = X86::MOV64rm; RC = &X86::GR64RegClass; break;
2349 // This needs to be set before we call getPtrSizedFrameRegister, otherwise
2350 // we get the wrong frame register.
2351 MachineFrameInfo *MFI = MF->getFrameInfo();
2352 MFI->setFrameAddressIsTaken(true);
2354 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
2355 unsigned FrameReg = RegInfo->getPtrSizedFrameRegister(*MF);
2356 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
2357 (FrameReg == X86::EBP && VT == MVT::i32)) &&
2358 "Invalid Frame Register!");
2360 // Always make a copy of the frame register to to a vreg first, so that we
2361 // never directly reference the frame register (the TwoAddressInstruction-
2362 // Pass doesn't like that).
2363 unsigned SrcReg = createResultReg(RC);
2364 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2365 TII.get(TargetOpcode::COPY), SrcReg).addReg(FrameReg);
2367 // Now recursively load from the frame address.
2368 // movq (%rbp), %rax
2369 // movq (%rax), %rax
2370 // movq (%rax), %rax
2373 unsigned Depth = cast<ConstantInt>(II->getOperand(0))->getZExtValue();
2375 DestReg = createResultReg(RC);
2376 addDirectMem(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2377 TII.get(Opc), DestReg), SrcReg);
2381 updateValueMap(II, SrcReg);
2384 case Intrinsic::memcpy: {
2385 const MemCpyInst *MCI = cast<MemCpyInst>(II);
2386 // Don't handle volatile or variable length memcpys.
2387 if (MCI->isVolatile())
2390 if (isa<ConstantInt>(MCI->getLength())) {
2391 // Small memcpy's are common enough that we want to do them
2392 // without a call if possible.
2393 uint64_t Len = cast<ConstantInt>(MCI->getLength())->getZExtValue();
2394 if (IsMemcpySmall(Len)) {
2395 X86AddressMode DestAM, SrcAM;
2396 if (!X86SelectAddress(MCI->getRawDest(), DestAM) ||
2397 !X86SelectAddress(MCI->getRawSource(), SrcAM))
2399 TryEmitSmallMemcpy(DestAM, SrcAM, Len);
2404 unsigned SizeWidth = Subtarget->is64Bit() ? 64 : 32;
2405 if (!MCI->getLength()->getType()->isIntegerTy(SizeWidth))
2408 if (MCI->getSourceAddressSpace() > 255 || MCI->getDestAddressSpace() > 255)
2411 return lowerCallTo(II, "memcpy", II->getNumArgOperands() - 2);
2413 case Intrinsic::memset: {
2414 const MemSetInst *MSI = cast<MemSetInst>(II);
2416 if (MSI->isVolatile())
2419 unsigned SizeWidth = Subtarget->is64Bit() ? 64 : 32;
2420 if (!MSI->getLength()->getType()->isIntegerTy(SizeWidth))
2423 if (MSI->getDestAddressSpace() > 255)
2426 return lowerCallTo(II, "memset", II->getNumArgOperands() - 2);
2428 case Intrinsic::stackprotector: {
2429 // Emit code to store the stack guard onto the stack.
2430 EVT PtrTy = TLI.getPointerTy(DL);
2432 const Value *Op1 = II->getArgOperand(0); // The guard's value.
2433 const AllocaInst *Slot = cast<AllocaInst>(II->getArgOperand(1));
2435 MFI.setStackProtectorIndex(FuncInfo.StaticAllocaMap[Slot]);
2437 // Grab the frame index.
2439 if (!X86SelectAddress(Slot, AM)) return false;
2440 if (!X86FastEmitStore(PtrTy, Op1, AM)) return false;
2443 case Intrinsic::dbg_declare: {
2444 const DbgDeclareInst *DI = cast<DbgDeclareInst>(II);
2446 assert(DI->getAddress() && "Null address should be checked earlier!");
2447 if (!X86SelectAddress(DI->getAddress(), AM))
2449 const MCInstrDesc &II = TII.get(TargetOpcode::DBG_VALUE);
2450 // FIXME may need to add RegState::Debug to any registers produced,
2451 // although ESP/EBP should be the only ones at the moment.
2452 assert(DI->getVariable()->isValidLocationForIntrinsic(DbgLoc) &&
2453 "Expected inlined-at fields to agree");
2454 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II), AM)
2456 .addMetadata(DI->getVariable())
2457 .addMetadata(DI->getExpression());
2460 case Intrinsic::trap: {
2461 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::TRAP));
2464 case Intrinsic::sqrt: {
2465 if (!Subtarget->hasSSE1())
2468 Type *RetTy = II->getCalledFunction()->getReturnType();
2471 if (!isTypeLegal(RetTy, VT))
2474 // Unfortunately we can't use fastEmit_r, because the AVX version of FSQRT
2475 // is not generated by FastISel yet.
2476 // FIXME: Update this code once tablegen can handle it.
2477 static const unsigned SqrtOpc[2][2] = {
2478 {X86::SQRTSSr, X86::VSQRTSSr},
2479 {X86::SQRTSDr, X86::VSQRTSDr}
2481 bool HasAVX = Subtarget->hasAVX();
2483 const TargetRegisterClass *RC;
2484 switch (VT.SimpleTy) {
2485 default: return false;
2486 case MVT::f32: Opc = SqrtOpc[0][HasAVX]; RC = &X86::FR32RegClass; break;
2487 case MVT::f64: Opc = SqrtOpc[1][HasAVX]; RC = &X86::FR64RegClass; break;
2490 const Value *SrcVal = II->getArgOperand(0);
2491 unsigned SrcReg = getRegForValue(SrcVal);
2496 unsigned ImplicitDefReg = 0;
2498 ImplicitDefReg = createResultReg(RC);
2499 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2500 TII.get(TargetOpcode::IMPLICIT_DEF), ImplicitDefReg);
2503 unsigned ResultReg = createResultReg(RC);
2504 MachineInstrBuilder MIB;
2505 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc),
2509 MIB.addReg(ImplicitDefReg);
2513 updateValueMap(II, ResultReg);
2516 case Intrinsic::sadd_with_overflow:
2517 case Intrinsic::uadd_with_overflow:
2518 case Intrinsic::ssub_with_overflow:
2519 case Intrinsic::usub_with_overflow:
2520 case Intrinsic::smul_with_overflow:
2521 case Intrinsic::umul_with_overflow: {
2522 // This implements the basic lowering of the xalu with overflow intrinsics
2523 // into add/sub/mul followed by either seto or setb.
2524 const Function *Callee = II->getCalledFunction();
2525 auto *Ty = cast<StructType>(Callee->getReturnType());
2526 Type *RetTy = Ty->getTypeAtIndex(0U);
2527 Type *CondTy = Ty->getTypeAtIndex(1);
2530 if (!isTypeLegal(RetTy, VT))
2533 if (VT < MVT::i8 || VT > MVT::i64)
2536 const Value *LHS = II->getArgOperand(0);
2537 const Value *RHS = II->getArgOperand(1);
2539 // Canonicalize immediate to the RHS.
2540 if (isa<ConstantInt>(LHS) && !isa<ConstantInt>(RHS) &&
2541 isCommutativeIntrinsic(II))
2542 std::swap(LHS, RHS);
2544 bool UseIncDec = false;
2545 if (isa<ConstantInt>(RHS) && cast<ConstantInt>(RHS)->isOne())
2548 unsigned BaseOpc, CondOpc;
2549 switch (II->getIntrinsicID()) {
2550 default: llvm_unreachable("Unexpected intrinsic!");
2551 case Intrinsic::sadd_with_overflow:
2552 BaseOpc = UseIncDec ? unsigned(X86ISD::INC) : unsigned(ISD::ADD);
2553 CondOpc = X86::SETOr;
2555 case Intrinsic::uadd_with_overflow:
2556 BaseOpc = ISD::ADD; CondOpc = X86::SETBr; break;
2557 case Intrinsic::ssub_with_overflow:
2558 BaseOpc = UseIncDec ? unsigned(X86ISD::DEC) : unsigned(ISD::SUB);
2559 CondOpc = X86::SETOr;
2561 case Intrinsic::usub_with_overflow:
2562 BaseOpc = ISD::SUB; CondOpc = X86::SETBr; break;
2563 case Intrinsic::smul_with_overflow:
2564 BaseOpc = X86ISD::SMUL; CondOpc = X86::SETOr; break;
2565 case Intrinsic::umul_with_overflow:
2566 BaseOpc = X86ISD::UMUL; CondOpc = X86::SETOr; break;
2569 unsigned LHSReg = getRegForValue(LHS);
2572 bool LHSIsKill = hasTrivialKill(LHS);
2574 unsigned ResultReg = 0;
2575 // Check if we have an immediate version.
2576 if (const auto *CI = dyn_cast<ConstantInt>(RHS)) {
2577 static const unsigned Opc[2][4] = {
2578 { X86::INC8r, X86::INC16r, X86::INC32r, X86::INC64r },
2579 { X86::DEC8r, X86::DEC16r, X86::DEC32r, X86::DEC64r }
2582 if (BaseOpc == X86ISD::INC || BaseOpc == X86ISD::DEC) {
2583 ResultReg = createResultReg(TLI.getRegClassFor(VT));
2584 bool IsDec = BaseOpc == X86ISD::DEC;
2585 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2586 TII.get(Opc[IsDec][VT.SimpleTy-MVT::i8]), ResultReg)
2587 .addReg(LHSReg, getKillRegState(LHSIsKill));
2589 ResultReg = fastEmit_ri(VT, VT, BaseOpc, LHSReg, LHSIsKill,
2590 CI->getZExtValue());
2596 RHSReg = getRegForValue(RHS);
2599 RHSIsKill = hasTrivialKill(RHS);
2600 ResultReg = fastEmit_rr(VT, VT, BaseOpc, LHSReg, LHSIsKill, RHSReg,
2604 // FastISel doesn't have a pattern for all X86::MUL*r and X86::IMUL*r. Emit
2606 if (BaseOpc == X86ISD::UMUL && !ResultReg) {
2607 static const unsigned MULOpc[] =
2608 { X86::MUL8r, X86::MUL16r, X86::MUL32r, X86::MUL64r };
2609 static const unsigned Reg[] = { X86::AL, X86::AX, X86::EAX, X86::RAX };
2610 // First copy the first operand into RAX, which is an implicit input to
2611 // the X86::MUL*r instruction.
2612 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2613 TII.get(TargetOpcode::COPY), Reg[VT.SimpleTy-MVT::i8])
2614 .addReg(LHSReg, getKillRegState(LHSIsKill));
2615 ResultReg = fastEmitInst_r(MULOpc[VT.SimpleTy-MVT::i8],
2616 TLI.getRegClassFor(VT), RHSReg, RHSIsKill);
2617 } else if (BaseOpc == X86ISD::SMUL && !ResultReg) {
2618 static const unsigned MULOpc[] =
2619 { X86::IMUL8r, X86::IMUL16rr, X86::IMUL32rr, X86::IMUL64rr };
2620 if (VT == MVT::i8) {
2621 // Copy the first operand into AL, which is an implicit input to the
2622 // X86::IMUL8r instruction.
2623 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2624 TII.get(TargetOpcode::COPY), X86::AL)
2625 .addReg(LHSReg, getKillRegState(LHSIsKill));
2626 ResultReg = fastEmitInst_r(MULOpc[0], TLI.getRegClassFor(VT), RHSReg,
2629 ResultReg = fastEmitInst_rr(MULOpc[VT.SimpleTy-MVT::i8],
2630 TLI.getRegClassFor(VT), LHSReg, LHSIsKill,
2637 unsigned ResultReg2 = FuncInfo.CreateRegs(CondTy);
2638 assert((ResultReg+1) == ResultReg2 && "Nonconsecutive result registers.");
2639 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CondOpc),
2642 updateValueMap(II, ResultReg, 2);
2645 case Intrinsic::x86_sse_cvttss2si:
2646 case Intrinsic::x86_sse_cvttss2si64:
2647 case Intrinsic::x86_sse2_cvttsd2si:
2648 case Intrinsic::x86_sse2_cvttsd2si64: {
2650 switch (II->getIntrinsicID()) {
2651 default: llvm_unreachable("Unexpected intrinsic.");
2652 case Intrinsic::x86_sse_cvttss2si:
2653 case Intrinsic::x86_sse_cvttss2si64:
2654 if (!Subtarget->hasSSE1())
2656 IsInputDouble = false;
2658 case Intrinsic::x86_sse2_cvttsd2si:
2659 case Intrinsic::x86_sse2_cvttsd2si64:
2660 if (!Subtarget->hasSSE2())
2662 IsInputDouble = true;
2666 Type *RetTy = II->getCalledFunction()->getReturnType();
2668 if (!isTypeLegal(RetTy, VT))
2671 static const unsigned CvtOpc[2][2][2] = {
2672 { { X86::CVTTSS2SIrr, X86::VCVTTSS2SIrr },
2673 { X86::CVTTSS2SI64rr, X86::VCVTTSS2SI64rr } },
2674 { { X86::CVTTSD2SIrr, X86::VCVTTSD2SIrr },
2675 { X86::CVTTSD2SI64rr, X86::VCVTTSD2SI64rr } }
2677 bool HasAVX = Subtarget->hasAVX();
2679 switch (VT.SimpleTy) {
2680 default: llvm_unreachable("Unexpected result type.");
2681 case MVT::i32: Opc = CvtOpc[IsInputDouble][0][HasAVX]; break;
2682 case MVT::i64: Opc = CvtOpc[IsInputDouble][1][HasAVX]; break;
2685 // Check if we can fold insertelement instructions into the convert.
2686 const Value *Op = II->getArgOperand(0);
2687 while (auto *IE = dyn_cast<InsertElementInst>(Op)) {
2688 const Value *Index = IE->getOperand(2);
2689 if (!isa<ConstantInt>(Index))
2691 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
2694 Op = IE->getOperand(1);
2697 Op = IE->getOperand(0);
2700 unsigned Reg = getRegForValue(Op);
2704 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
2705 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
2708 updateValueMap(II, ResultReg);
2714 bool X86FastISel::fastLowerArguments() {
2715 if (!FuncInfo.CanLowerReturn)
2718 const Function *F = FuncInfo.Fn;
2722 CallingConv::ID CC = F->getCallingConv();
2723 if (CC != CallingConv::C)
2726 if (Subtarget->isCallingConvWin64(CC))
2729 if (!Subtarget->is64Bit())
2732 // Only handle simple cases. i.e. Up to 6 i32/i64 scalar arguments.
2733 unsigned GPRCnt = 0;
2734 unsigned FPRCnt = 0;
2736 for (auto const &Arg : F->args()) {
2737 // The first argument is at index 1.
2739 if (F->getAttributes().hasAttribute(Idx, Attribute::ByVal) ||
2740 F->getAttributes().hasAttribute(Idx, Attribute::InReg) ||
2741 F->getAttributes().hasAttribute(Idx, Attribute::StructRet) ||
2742 F->getAttributes().hasAttribute(Idx, Attribute::Nest))
2745 Type *ArgTy = Arg.getType();
2746 if (ArgTy->isStructTy() || ArgTy->isArrayTy() || ArgTy->isVectorTy())
2749 EVT ArgVT = TLI.getValueType(DL, ArgTy);
2750 if (!ArgVT.isSimple()) return false;
2751 switch (ArgVT.getSimpleVT().SimpleTy) {
2752 default: return false;
2759 if (!Subtarget->hasSSE1())
2772 static const MCPhysReg GPR32ArgRegs[] = {
2773 X86::EDI, X86::ESI, X86::EDX, X86::ECX, X86::R8D, X86::R9D
2775 static const MCPhysReg GPR64ArgRegs[] = {
2776 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8 , X86::R9
2778 static const MCPhysReg XMMArgRegs[] = {
2779 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2780 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2783 unsigned GPRIdx = 0;
2784 unsigned FPRIdx = 0;
2785 for (auto const &Arg : F->args()) {
2786 MVT VT = TLI.getSimpleValueType(DL, Arg.getType());
2787 const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
2789 switch (VT.SimpleTy) {
2790 default: llvm_unreachable("Unexpected value type.");
2791 case MVT::i32: SrcReg = GPR32ArgRegs[GPRIdx++]; break;
2792 case MVT::i64: SrcReg = GPR64ArgRegs[GPRIdx++]; break;
2793 case MVT::f32: // fall-through
2794 case MVT::f64: SrcReg = XMMArgRegs[FPRIdx++]; break;
2796 unsigned DstReg = FuncInfo.MF->addLiveIn(SrcReg, RC);
2797 // FIXME: Unfortunately it's necessary to emit a copy from the livein copy.
2798 // Without this, EmitLiveInCopies may eliminate the livein if its only
2799 // use is a bitcast (which isn't turned into an instruction).
2800 unsigned ResultReg = createResultReg(RC);
2801 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2802 TII.get(TargetOpcode::COPY), ResultReg)
2803 .addReg(DstReg, getKillRegState(true));
2804 updateValueMap(&Arg, ResultReg);
2809 static unsigned computeBytesPoppedByCallee(const X86Subtarget *Subtarget,
2811 ImmutableCallSite *CS) {
2812 if (Subtarget->is64Bit())
2814 if (Subtarget->getTargetTriple().isOSMSVCRT())
2816 if (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2817 CC == CallingConv::HiPE)
2821 if (CS->arg_empty() || !CS->paramHasAttr(1, Attribute::StructRet) ||
2822 CS->paramHasAttr(1, Attribute::InReg) || Subtarget->isTargetMCU())
2828 bool X86FastISel::fastLowerCall(CallLoweringInfo &CLI) {
2829 auto &OutVals = CLI.OutVals;
2830 auto &OutFlags = CLI.OutFlags;
2831 auto &OutRegs = CLI.OutRegs;
2832 auto &Ins = CLI.Ins;
2833 auto &InRegs = CLI.InRegs;
2834 CallingConv::ID CC = CLI.CallConv;
2835 bool &IsTailCall = CLI.IsTailCall;
2836 bool IsVarArg = CLI.IsVarArg;
2837 const Value *Callee = CLI.Callee;
2838 MCSymbol *Symbol = CLI.Symbol;
2840 bool Is64Bit = Subtarget->is64Bit();
2841 bool IsWin64 = Subtarget->isCallingConvWin64(CC);
2843 // Handle only C, fastcc, and webkit_js calling conventions for now.
2845 default: return false;
2846 case CallingConv::C:
2847 case CallingConv::Fast:
2848 case CallingConv::WebKit_JS:
2849 case CallingConv::X86_FastCall:
2850 case CallingConv::X86_64_Win64:
2851 case CallingConv::X86_64_SysV:
2855 // Allow SelectionDAG isel to handle tail calls.
2859 // fastcc with -tailcallopt is intended to provide a guaranteed
2860 // tail call optimization. Fastisel doesn't know how to do that.
2861 if (CC == CallingConv::Fast && TM.Options.GuaranteedTailCallOpt)
2864 // Don't know how to handle Win64 varargs yet. Nothing special needed for
2865 // x86-32. Special handling for x86-64 is implemented.
2866 if (IsVarArg && IsWin64)
2869 // Don't know about inalloca yet.
2870 if (CLI.CS && CLI.CS->hasInAllocaArgument())
2873 // Fast-isel doesn't know about callee-pop yet.
2874 if (X86::isCalleePop(CC, Subtarget->is64Bit(), IsVarArg,
2875 TM.Options.GuaranteedTailCallOpt))
2878 SmallVector<MVT, 16> OutVTs;
2879 SmallVector<unsigned, 16> ArgRegs;
2881 // If this is a constant i1/i8/i16 argument, promote to i32 to avoid an extra
2882 // instruction. This is safe because it is common to all FastISel supported
2883 // calling conventions on x86.
2884 for (int i = 0, e = OutVals.size(); i != e; ++i) {
2885 Value *&Val = OutVals[i];
2886 ISD::ArgFlagsTy Flags = OutFlags[i];
2887 if (auto *CI = dyn_cast<ConstantInt>(Val)) {
2888 if (CI->getBitWidth() < 32) {
2890 Val = ConstantExpr::getSExt(CI, Type::getInt32Ty(CI->getContext()));
2892 Val = ConstantExpr::getZExt(CI, Type::getInt32Ty(CI->getContext()));
2896 // Passing bools around ends up doing a trunc to i1 and passing it.
2897 // Codegen this as an argument + "and 1".
2899 auto *TI = dyn_cast<TruncInst>(Val);
2901 if (TI && TI->getType()->isIntegerTy(1) && CLI.CS &&
2902 (TI->getParent() == CLI.CS->getInstruction()->getParent()) &&
2904 Value *PrevVal = TI->getOperand(0);
2905 ResultReg = getRegForValue(PrevVal);
2910 if (!isTypeLegal(PrevVal->getType(), VT))
2914 fastEmit_ri(VT, VT, ISD::AND, ResultReg, hasTrivialKill(PrevVal), 1);
2916 if (!isTypeLegal(Val->getType(), VT))
2918 ResultReg = getRegForValue(Val);
2924 ArgRegs.push_back(ResultReg);
2925 OutVTs.push_back(VT);
2928 // Analyze operands of the call, assigning locations to each operand.
2929 SmallVector<CCValAssign, 16> ArgLocs;
2930 CCState CCInfo(CC, IsVarArg, *FuncInfo.MF, ArgLocs, CLI.RetTy->getContext());
2932 // Allocate shadow area for Win64
2934 CCInfo.AllocateStack(32, 8);
2936 CCInfo.AnalyzeCallOperands(OutVTs, OutFlags, CC_X86);
2938 // Get a count of how many bytes are to be pushed on the stack.
2939 unsigned NumBytes = CCInfo.getAlignedCallFrameSize();
2941 // Issue CALLSEQ_START
2942 unsigned AdjStackDown = TII.getCallFrameSetupOpcode();
2943 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackDown))
2944 .addImm(NumBytes).addImm(0);
2946 // Walk the register/memloc assignments, inserting copies/loads.
2947 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
2948 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2949 CCValAssign const &VA = ArgLocs[i];
2950 const Value *ArgVal = OutVals[VA.getValNo()];
2951 MVT ArgVT = OutVTs[VA.getValNo()];
2953 if (ArgVT == MVT::x86mmx)
2956 unsigned ArgReg = ArgRegs[VA.getValNo()];
2958 // Promote the value if needed.
2959 switch (VA.getLocInfo()) {
2960 case CCValAssign::Full: break;
2961 case CCValAssign::SExt: {
2962 assert(VA.getLocVT().isInteger() && !VA.getLocVT().isVector() &&
2963 "Unexpected extend");
2964 bool Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(), ArgReg,
2966 assert(Emitted && "Failed to emit a sext!"); (void)Emitted;
2967 ArgVT = VA.getLocVT();
2970 case CCValAssign::ZExt: {
2971 assert(VA.getLocVT().isInteger() && !VA.getLocVT().isVector() &&
2972 "Unexpected extend");
2973 bool Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(), ArgReg,
2975 assert(Emitted && "Failed to emit a zext!"); (void)Emitted;
2976 ArgVT = VA.getLocVT();
2979 case CCValAssign::AExt: {
2980 assert(VA.getLocVT().isInteger() && !VA.getLocVT().isVector() &&
2981 "Unexpected extend");
2982 bool Emitted = X86FastEmitExtend(ISD::ANY_EXTEND, VA.getLocVT(), ArgReg,
2985 Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(), ArgReg,
2988 Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(), ArgReg,
2991 assert(Emitted && "Failed to emit a aext!"); (void)Emitted;
2992 ArgVT = VA.getLocVT();
2995 case CCValAssign::BCvt: {
2996 ArgReg = fastEmit_r(ArgVT, VA.getLocVT(), ISD::BITCAST, ArgReg,
2997 /*TODO: Kill=*/false);
2998 assert(ArgReg && "Failed to emit a bitcast!");
2999 ArgVT = VA.getLocVT();
3002 case CCValAssign::VExt:
3003 // VExt has not been implemented, so this should be impossible to reach
3004 // for now. However, fallback to Selection DAG isel once implemented.
3006 case CCValAssign::AExtUpper:
3007 case CCValAssign::SExtUpper:
3008 case CCValAssign::ZExtUpper:
3009 case CCValAssign::FPExt:
3010 llvm_unreachable("Unexpected loc info!");
3011 case CCValAssign::Indirect:
3012 // FIXME: Indirect doesn't need extending, but fast-isel doesn't fully
3017 if (VA.isRegLoc()) {
3018 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3019 TII.get(TargetOpcode::COPY), VA.getLocReg()).addReg(ArgReg);
3020 OutRegs.push_back(VA.getLocReg());
3022 assert(VA.isMemLoc());
3024 // Don't emit stores for undef values.
3025 if (isa<UndefValue>(ArgVal))
3028 unsigned LocMemOffset = VA.getLocMemOffset();
3030 AM.Base.Reg = RegInfo->getStackRegister();
3031 AM.Disp = LocMemOffset;
3032 ISD::ArgFlagsTy Flags = OutFlags[VA.getValNo()];
3033 unsigned Alignment = DL.getABITypeAlignment(ArgVal->getType());
3034 MachineMemOperand *MMO = FuncInfo.MF->getMachineMemOperand(
3035 MachinePointerInfo::getStack(*FuncInfo.MF, LocMemOffset),
3036 MachineMemOperand::MOStore, ArgVT.getStoreSize(), Alignment);
3037 if (Flags.isByVal()) {
3038 X86AddressMode SrcAM;
3039 SrcAM.Base.Reg = ArgReg;
3040 if (!TryEmitSmallMemcpy(AM, SrcAM, Flags.getByValSize()))
3042 } else if (isa<ConstantInt>(ArgVal) || isa<ConstantPointerNull>(ArgVal)) {
3043 // If this is a really simple value, emit this with the Value* version
3044 // of X86FastEmitStore. If it isn't simple, we don't want to do this,
3045 // as it can cause us to reevaluate the argument.
3046 if (!X86FastEmitStore(ArgVT, ArgVal, AM, MMO))
3049 bool ValIsKill = hasTrivialKill(ArgVal);
3050 if (!X86FastEmitStore(ArgVT, ArgReg, ValIsKill, AM, MMO))
3056 // ELF / PIC requires GOT in the EBX register before function calls via PLT
3058 if (Subtarget->isPICStyleGOT()) {
3059 unsigned Base = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF);
3060 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3061 TII.get(TargetOpcode::COPY), X86::EBX).addReg(Base);
3064 if (Is64Bit && IsVarArg && !IsWin64) {
3065 // From AMD64 ABI document:
3066 // For calls that may call functions that use varargs or stdargs
3067 // (prototype-less calls or calls to functions containing ellipsis (...) in
3068 // the declaration) %al is used as hidden argument to specify the number
3069 // of SSE registers used. The contents of %al do not need to match exactly
3070 // the number of registers, but must be an ubound on the number of SSE
3071 // registers used and is in the range 0 - 8 inclusive.
3073 // Count the number of XMM registers allocated.
3074 static const MCPhysReg XMMArgRegs[] = {
3075 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
3076 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
3078 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs);
3079 assert((Subtarget->hasSSE1() || !NumXMMRegs)
3080 && "SSE registers cannot be used when SSE is disabled");
3081 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::MOV8ri),
3082 X86::AL).addImm(NumXMMRegs);
3085 // Materialize callee address in a register. FIXME: GV address can be
3086 // handled with a CALLpcrel32 instead.
3087 X86AddressMode CalleeAM;
3088 if (!X86SelectCallAddress(Callee, CalleeAM))
3091 unsigned CalleeOp = 0;
3092 const GlobalValue *GV = nullptr;
3093 if (CalleeAM.GV != nullptr) {
3095 } else if (CalleeAM.Base.Reg != 0) {
3096 CalleeOp = CalleeAM.Base.Reg;
3101 MachineInstrBuilder MIB;
3103 // Register-indirect call.
3104 unsigned CallOpc = Is64Bit ? X86::CALL64r : X86::CALL32r;
3105 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CallOpc))
3109 assert(GV && "Not a direct call");
3110 unsigned CallOpc = Is64Bit ? X86::CALL64pcrel32 : X86::CALLpcrel32;
3112 // See if we need any target-specific flags on the GV operand.
3113 unsigned char OpFlags = 0;
3115 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
3116 // external symbols most go through the PLT in PIC mode. If the symbol
3117 // has hidden or protected visibility, or if it is static or local, then
3118 // we don't need to use the PLT - we can directly call it.
3119 if (Subtarget->isTargetELF() &&
3120 TM.getRelocationModel() == Reloc::PIC_ &&
3121 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
3122 OpFlags = X86II::MO_PLT;
3123 } else if (Subtarget->isPICStyleStubAny() &&
3124 !GV->isStrongDefinitionForLinker() &&
3125 (!Subtarget->getTargetTriple().isMacOSX() ||
3126 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3127 // PC-relative references to external symbols should go through $stub,
3128 // unless we're building with the leopard linker or later, which
3129 // automatically synthesizes these stubs.
3130 OpFlags = X86II::MO_DARWIN_STUB;
3133 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CallOpc));
3135 MIB.addSym(Symbol, OpFlags);
3137 MIB.addGlobalAddress(GV, 0, OpFlags);
3140 // Add a register mask operand representing the call-preserved registers.
3141 // Proper defs for return values will be added by setPhysRegsDeadExcept().
3142 MIB.addRegMask(TRI.getCallPreservedMask(*FuncInfo.MF, CC));
3144 // Add an implicit use GOT pointer in EBX.
3145 if (Subtarget->isPICStyleGOT())
3146 MIB.addReg(X86::EBX, RegState::Implicit);
3148 if (Is64Bit && IsVarArg && !IsWin64)
3149 MIB.addReg(X86::AL, RegState::Implicit);
3151 // Add implicit physical register uses to the call.
3152 for (auto Reg : OutRegs)
3153 MIB.addReg(Reg, RegState::Implicit);
3155 // Issue CALLSEQ_END
3156 unsigned NumBytesForCalleeToPop =
3157 computeBytesPoppedByCallee(Subtarget, CC, CLI.CS);
3158 unsigned AdjStackUp = TII.getCallFrameDestroyOpcode();
3159 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackUp))
3160 .addImm(NumBytes).addImm(NumBytesForCalleeToPop);
3162 // Now handle call return values.
3163 SmallVector<CCValAssign, 16> RVLocs;
3164 CCState CCRetInfo(CC, IsVarArg, *FuncInfo.MF, RVLocs,
3165 CLI.RetTy->getContext());
3166 CCRetInfo.AnalyzeCallResult(Ins, RetCC_X86);
3168 // Copy all of the result registers out of their specified physreg.
3169 unsigned ResultReg = FuncInfo.CreateRegs(CLI.RetTy);
3170 for (unsigned i = 0; i != RVLocs.size(); ++i) {
3171 CCValAssign &VA = RVLocs[i];
3172 EVT CopyVT = VA.getValVT();
3173 unsigned CopyReg = ResultReg + i;
3175 // If this is x86-64, and we disabled SSE, we can't return FP values
3176 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
3177 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
3178 report_fatal_error("SSE register return with SSE disabled");
3181 // If we prefer to use the value in xmm registers, copy it out as f80 and
3182 // use a truncate to move it from fp stack reg to xmm reg.
3183 if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
3184 isScalarFPTypeInSSEReg(VA.getValVT())) {
3186 CopyReg = createResultReg(&X86::RFP80RegClass);
3189 // Copy out the result.
3190 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3191 TII.get(TargetOpcode::COPY), CopyReg).addReg(VA.getLocReg());
3192 InRegs.push_back(VA.getLocReg());
3194 // Round the f80 to the right size, which also moves it to the appropriate
3195 // xmm register. This is accomplished by storing the f80 value in memory
3196 // and then loading it back.
3197 if (CopyVT != VA.getValVT()) {
3198 EVT ResVT = VA.getValVT();
3199 unsigned Opc = ResVT == MVT::f32 ? X86::ST_Fp80m32 : X86::ST_Fp80m64;
3200 unsigned MemSize = ResVT.getSizeInBits()/8;
3201 int FI = MFI.CreateStackObject(MemSize, MemSize, false);
3202 addFrameReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3205 Opc = ResVT == MVT::f32 ? X86::MOVSSrm : X86::MOVSDrm;
3206 addFrameReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3207 TII.get(Opc), ResultReg + i), FI);
3211 CLI.ResultReg = ResultReg;
3212 CLI.NumResultRegs = RVLocs.size();
3219 X86FastISel::fastSelectInstruction(const Instruction *I) {
3220 switch (I->getOpcode()) {
3222 case Instruction::Load:
3223 return X86SelectLoad(I);
3224 case Instruction::Store:
3225 return X86SelectStore(I);
3226 case Instruction::Ret:
3227 return X86SelectRet(I);
3228 case Instruction::ICmp:
3229 case Instruction::FCmp:
3230 return X86SelectCmp(I);
3231 case Instruction::ZExt:
3232 return X86SelectZExt(I);
3233 case Instruction::Br:
3234 return X86SelectBranch(I);
3235 case Instruction::LShr:
3236 case Instruction::AShr:
3237 case Instruction::Shl:
3238 return X86SelectShift(I);
3239 case Instruction::SDiv:
3240 case Instruction::UDiv:
3241 case Instruction::SRem:
3242 case Instruction::URem:
3243 return X86SelectDivRem(I);
3244 case Instruction::Select:
3245 return X86SelectSelect(I);
3246 case Instruction::Trunc:
3247 return X86SelectTrunc(I);
3248 case Instruction::FPExt:
3249 return X86SelectFPExt(I);
3250 case Instruction::FPTrunc:
3251 return X86SelectFPTrunc(I);
3252 case Instruction::SIToFP:
3253 return X86SelectSIToFP(I);
3254 case Instruction::IntToPtr: // Deliberate fall-through.
3255 case Instruction::PtrToInt: {
3256 EVT SrcVT = TLI.getValueType(DL, I->getOperand(0)->getType());
3257 EVT DstVT = TLI.getValueType(DL, I->getType());
3258 if (DstVT.bitsGT(SrcVT))
3259 return X86SelectZExt(I);
3260 if (DstVT.bitsLT(SrcVT))
3261 return X86SelectTrunc(I);
3262 unsigned Reg = getRegForValue(I->getOperand(0));
3263 if (Reg == 0) return false;
3264 updateValueMap(I, Reg);
3267 case Instruction::BitCast: {
3268 // Select SSE2/AVX bitcasts between 128/256 bit vector types.
3269 if (!Subtarget->hasSSE2())
3272 EVT SrcVT = TLI.getValueType(DL, I->getOperand(0)->getType());
3273 EVT DstVT = TLI.getValueType(DL, I->getType());
3275 if (!SrcVT.isSimple() || !DstVT.isSimple())
3278 if (!SrcVT.is128BitVector() &&
3279 !(Subtarget->hasAVX() && SrcVT.is256BitVector()))
3282 unsigned Reg = getRegForValue(I->getOperand(0));
3286 // No instruction is needed for conversion. Reuse the register used by
3287 // the fist operand.
3288 updateValueMap(I, Reg);
3296 unsigned X86FastISel::X86MaterializeInt(const ConstantInt *CI, MVT VT) {
3300 uint64_t Imm = CI->getZExtValue();
3302 unsigned SrcReg = fastEmitInst_(X86::MOV32r0, &X86::GR32RegClass);
3303 switch (VT.SimpleTy) {
3304 default: llvm_unreachable("Unexpected value type");
3307 return fastEmitInst_extractsubreg(MVT::i8, SrcReg, /*Kill=*/true,
3310 return fastEmitInst_extractsubreg(MVT::i16, SrcReg, /*Kill=*/true,
3315 unsigned ResultReg = createResultReg(&X86::GR64RegClass);
3316 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3317 TII.get(TargetOpcode::SUBREG_TO_REG), ResultReg)
3318 .addImm(0).addReg(SrcReg).addImm(X86::sub_32bit);
3325 switch (VT.SimpleTy) {
3326 default: llvm_unreachable("Unexpected value type");
3327 case MVT::i1: VT = MVT::i8; // fall-through
3328 case MVT::i8: Opc = X86::MOV8ri; break;
3329 case MVT::i16: Opc = X86::MOV16ri; break;
3330 case MVT::i32: Opc = X86::MOV32ri; break;
3332 if (isUInt<32>(Imm))
3334 else if (isInt<32>(Imm))
3335 Opc = X86::MOV64ri32;
3341 if (VT == MVT::i64 && Opc == X86::MOV32ri) {
3342 unsigned SrcReg = fastEmitInst_i(Opc, &X86::GR32RegClass, Imm);
3343 unsigned ResultReg = createResultReg(&X86::GR64RegClass);
3344 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3345 TII.get(TargetOpcode::SUBREG_TO_REG), ResultReg)
3346 .addImm(0).addReg(SrcReg).addImm(X86::sub_32bit);
3349 return fastEmitInst_i(Opc, TLI.getRegClassFor(VT), Imm);
3352 unsigned X86FastISel::X86MaterializeFP(const ConstantFP *CFP, MVT VT) {
3353 if (CFP->isNullValue())
3354 return fastMaterializeFloatZero(CFP);
3356 // Can't handle alternate code models yet.
3357 CodeModel::Model CM = TM.getCodeModel();
3358 if (CM != CodeModel::Small && CM != CodeModel::Large)
3361 // Get opcode and regclass of the output for the given load instruction.
3363 const TargetRegisterClass *RC = nullptr;
3364 switch (VT.SimpleTy) {
3367 if (X86ScalarSSEf32) {
3368 Opc = Subtarget->hasAVX() ? X86::VMOVSSrm : X86::MOVSSrm;
3369 RC = &X86::FR32RegClass;
3371 Opc = X86::LD_Fp32m;
3372 RC = &X86::RFP32RegClass;
3376 if (X86ScalarSSEf64) {
3377 Opc = Subtarget->hasAVX() ? X86::VMOVSDrm : X86::MOVSDrm;
3378 RC = &X86::FR64RegClass;
3380 Opc = X86::LD_Fp64m;
3381 RC = &X86::RFP64RegClass;
3385 // No f80 support yet.
3389 // MachineConstantPool wants an explicit alignment.
3390 unsigned Align = DL.getPrefTypeAlignment(CFP->getType());
3392 // Alignment of vector types. FIXME!
3393 Align = DL.getTypeAllocSize(CFP->getType());
3396 // x86-32 PIC requires a PIC base register for constant pools.
3397 unsigned PICBase = 0;
3398 unsigned char OpFlag = 0;
3399 if (Subtarget->isPICStyleStubPIC()) { // Not dynamic-no-pic
3400 OpFlag = X86II::MO_PIC_BASE_OFFSET;
3401 PICBase = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF);
3402 } else if (Subtarget->isPICStyleGOT()) {
3403 OpFlag = X86II::MO_GOTOFF;
3404 PICBase = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF);
3405 } else if (Subtarget->isPICStyleRIPRel() &&
3406 TM.getCodeModel() == CodeModel::Small) {
3410 // Create the load from the constant pool.
3411 unsigned CPI = MCP.getConstantPoolIndex(CFP, Align);
3412 unsigned ResultReg = createResultReg(RC);
3414 if (CM == CodeModel::Large) {
3415 unsigned AddrReg = createResultReg(&X86::GR64RegClass);
3416 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::MOV64ri),
3418 .addConstantPoolIndex(CPI, 0, OpFlag);
3419 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3420 TII.get(Opc), ResultReg);
3421 addDirectMem(MIB, AddrReg);
3422 MachineMemOperand *MMO = FuncInfo.MF->getMachineMemOperand(
3423 MachinePointerInfo::getConstantPool(*FuncInfo.MF),
3424 MachineMemOperand::MOLoad, DL.getPointerSize(), Align);
3425 MIB->addMemOperand(*FuncInfo.MF, MMO);
3429 addConstantPoolReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3430 TII.get(Opc), ResultReg),
3431 CPI, PICBase, OpFlag);
3435 unsigned X86FastISel::X86MaterializeGV(const GlobalValue *GV, MVT VT) {
3436 // Can't handle alternate code models yet.
3437 if (TM.getCodeModel() != CodeModel::Small)
3440 // Materialize addresses with LEA/MOV instructions.
3442 if (X86SelectAddress(GV, AM)) {
3443 // If the expression is just a basereg, then we're done, otherwise we need
3445 if (AM.BaseType == X86AddressMode::RegBase &&
3446 AM.IndexReg == 0 && AM.Disp == 0 && AM.GV == nullptr)
3449 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
3450 if (TM.getRelocationModel() == Reloc::Static &&
3451 TLI.getPointerTy(DL) == MVT::i64) {
3452 // The displacement code could be more than 32 bits away so we need to use
3453 // an instruction with a 64 bit immediate
3454 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::MOV64ri),
3456 .addGlobalAddress(GV);
3459 TLI.getPointerTy(DL) == MVT::i32
3460 ? (Subtarget->isTarget64BitILP32() ? X86::LEA64_32r : X86::LEA32r)
3462 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3463 TII.get(Opc), ResultReg), AM);
3470 unsigned X86FastISel::fastMaterializeConstant(const Constant *C) {
3471 EVT CEVT = TLI.getValueType(DL, C->getType(), true);
3473 // Only handle simple types.
3474 if (!CEVT.isSimple())
3476 MVT VT = CEVT.getSimpleVT();
3478 if (const auto *CI = dyn_cast<ConstantInt>(C))
3479 return X86MaterializeInt(CI, VT);
3480 else if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
3481 return X86MaterializeFP(CFP, VT);
3482 else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
3483 return X86MaterializeGV(GV, VT);
3488 unsigned X86FastISel::fastMaterializeAlloca(const AllocaInst *C) {
3489 // Fail on dynamic allocas. At this point, getRegForValue has already
3490 // checked its CSE maps, so if we're here trying to handle a dynamic
3491 // alloca, we're not going to succeed. X86SelectAddress has a
3492 // check for dynamic allocas, because it's called directly from
3493 // various places, but targetMaterializeAlloca also needs a check
3494 // in order to avoid recursion between getRegForValue,
3495 // X86SelectAddrss, and targetMaterializeAlloca.
3496 if (!FuncInfo.StaticAllocaMap.count(C))
3498 assert(C->isStaticAlloca() && "dynamic alloca in the static alloca map?");
3501 if (!X86SelectAddress(C, AM))
3504 TLI.getPointerTy(DL) == MVT::i32
3505 ? (Subtarget->isTarget64BitILP32() ? X86::LEA64_32r : X86::LEA32r)
3507 const TargetRegisterClass *RC = TLI.getRegClassFor(TLI.getPointerTy(DL));
3508 unsigned ResultReg = createResultReg(RC);
3509 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3510 TII.get(Opc), ResultReg), AM);
3514 unsigned X86FastISel::fastMaterializeFloatZero(const ConstantFP *CF) {
3516 if (!isTypeLegal(CF->getType(), VT))
3519 // Get opcode and regclass for the given zero.
3521 const TargetRegisterClass *RC = nullptr;
3522 switch (VT.SimpleTy) {
3525 if (X86ScalarSSEf32) {
3526 Opc = X86::FsFLD0SS;
3527 RC = &X86::FR32RegClass;
3529 Opc = X86::LD_Fp032;
3530 RC = &X86::RFP32RegClass;
3534 if (X86ScalarSSEf64) {
3535 Opc = X86::FsFLD0SD;
3536 RC = &X86::FR64RegClass;
3538 Opc = X86::LD_Fp064;
3539 RC = &X86::RFP64RegClass;
3543 // No f80 support yet.
3547 unsigned ResultReg = createResultReg(RC);
3548 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg);
3553 bool X86FastISel::tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo,
3554 const LoadInst *LI) {
3555 const Value *Ptr = LI->getPointerOperand();
3557 if (!X86SelectAddress(Ptr, AM))
3560 const X86InstrInfo &XII = (const X86InstrInfo &)TII;
3562 unsigned Size = DL.getTypeAllocSize(LI->getType());
3563 unsigned Alignment = LI->getAlignment();
3565 if (Alignment == 0) // Ensure that codegen never sees alignment 0
3566 Alignment = DL.getABITypeAlignment(LI->getType());
3568 SmallVector<MachineOperand, 8> AddrOps;
3569 AM.getFullAddress(AddrOps);
3571 MachineInstr *Result = XII.foldMemoryOperandImpl(
3572 *FuncInfo.MF, MI, OpNo, AddrOps, FuncInfo.InsertPt, Size, Alignment,
3573 /*AllowCommute=*/true);
3577 // The index register could be in the wrong register class. Unfortunately,
3578 // foldMemoryOperandImpl could have commuted the instruction so its not enough
3579 // to just look at OpNo + the offset to the index reg. We actually need to
3580 // scan the instruction to find the index reg and see if its the correct reg
3582 unsigned OperandNo = 0;
3583 for (MachineInstr::mop_iterator I = Result->operands_begin(),
3584 E = Result->operands_end(); I != E; ++I, ++OperandNo) {
3585 MachineOperand &MO = *I;
3586 if (!MO.isReg() || MO.isDef() || MO.getReg() != AM.IndexReg)
3588 // Found the index reg, now try to rewrite it.
3589 unsigned IndexReg = constrainOperandRegClass(Result->getDesc(),
3590 MO.getReg(), OperandNo);
3591 if (IndexReg == MO.getReg())
3593 MO.setReg(IndexReg);
3596 Result->addMemOperand(*FuncInfo.MF, createMachineMemOperandFor(LI));
3597 MI->eraseFromParent();
3603 FastISel *X86::createFastISel(FunctionLoweringInfo &funcInfo,
3604 const TargetLibraryInfo *libInfo) {
3605 return new X86FastISel(funcInfo, libInfo);