1 //===-- AArch6464FastISel.cpp - AArch64 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 AArch64-specific support for the FastISel class. Some
11 // of the target-specific code is generated by tablegen in the file
12 // AArch64GenFastISel.inc, which is #included here.
14 //===----------------------------------------------------------------------===//
17 #include "AArch64Subtarget.h"
18 #include "AArch64TargetMachine.h"
19 #include "MCTargetDesc/AArch64AddressingModes.h"
20 #include "llvm/Analysis/BranchProbabilityInfo.h"
21 #include "llvm/CodeGen/CallingConvLower.h"
22 #include "llvm/CodeGen/FastISel.h"
23 #include "llvm/CodeGen/FunctionLoweringInfo.h"
24 #include "llvm/CodeGen/MachineConstantPool.h"
25 #include "llvm/CodeGen/MachineFrameInfo.h"
26 #include "llvm/CodeGen/MachineInstrBuilder.h"
27 #include "llvm/CodeGen/MachineRegisterInfo.h"
28 #include "llvm/IR/CallingConv.h"
29 #include "llvm/IR/DataLayout.h"
30 #include "llvm/IR/DerivedTypes.h"
31 #include "llvm/IR/Function.h"
32 #include "llvm/IR/GetElementPtrTypeIterator.h"
33 #include "llvm/IR/GlobalAlias.h"
34 #include "llvm/IR/GlobalVariable.h"
35 #include "llvm/IR/Instructions.h"
36 #include "llvm/IR/IntrinsicInst.h"
37 #include "llvm/IR/Operator.h"
38 #include "llvm/Support/CommandLine.h"
43 class AArch64FastISel : public FastISel {
53 AArch64_AM::ShiftExtendType ExtType;
61 const GlobalValue *GV;
64 Address() : Kind(RegBase), ExtType(AArch64_AM::InvalidShiftExtend),
65 OffsetReg(0), Shift(0), Offset(0), GV(nullptr) { Base.Reg = 0; }
66 void setKind(BaseKind K) { Kind = K; }
67 BaseKind getKind() const { return Kind; }
68 void setExtendType(AArch64_AM::ShiftExtendType E) { ExtType = E; }
69 AArch64_AM::ShiftExtendType getExtendType() const { return ExtType; }
70 bool isRegBase() const { return Kind == RegBase; }
71 bool isFIBase() const { return Kind == FrameIndexBase; }
72 void setReg(unsigned Reg) {
73 assert(isRegBase() && "Invalid base register access!");
76 unsigned getReg() const {
77 assert(isRegBase() && "Invalid base register access!");
80 void setOffsetReg(unsigned Reg) {
81 assert(isRegBase() && "Invalid offset register access!");
84 unsigned getOffsetReg() const {
85 assert(isRegBase() && "Invalid offset register access!");
88 void setFI(unsigned FI) {
89 assert(isFIBase() && "Invalid base frame index access!");
92 unsigned getFI() const {
93 assert(isFIBase() && "Invalid base frame index access!");
96 void setOffset(int64_t O) { Offset = O; }
97 int64_t getOffset() { return Offset; }
98 void setShift(unsigned S) { Shift = S; }
99 unsigned getShift() { return Shift; }
101 void setGlobalValue(const GlobalValue *G) { GV = G; }
102 const GlobalValue *getGlobalValue() { return GV; }
105 /// Subtarget - Keep a pointer to the AArch64Subtarget around so that we can
106 /// make the right decision when generating code for different targets.
107 const AArch64Subtarget *Subtarget;
108 LLVMContext *Context;
110 bool FastLowerArguments() override;
111 bool FastLowerCall(CallLoweringInfo &CLI) override;
112 bool FastLowerIntrinsicCall(const IntrinsicInst *II) override;
115 // Selection routines.
116 bool SelectLoad(const Instruction *I);
117 bool SelectStore(const Instruction *I);
118 bool SelectBranch(const Instruction *I);
119 bool SelectIndirectBr(const Instruction *I);
120 bool SelectCmp(const Instruction *I);
121 bool SelectSelect(const Instruction *I);
122 bool SelectFPExt(const Instruction *I);
123 bool SelectFPTrunc(const Instruction *I);
124 bool SelectFPToInt(const Instruction *I, bool Signed);
125 bool SelectIntToFP(const Instruction *I, bool Signed);
126 bool SelectRem(const Instruction *I, unsigned ISDOpcode);
127 bool SelectRet(const Instruction *I);
128 bool SelectTrunc(const Instruction *I);
129 bool SelectIntExt(const Instruction *I);
130 bool SelectMul(const Instruction *I);
131 bool SelectShift(const Instruction *I);
132 bool SelectBitCast(const Instruction *I);
134 // Utility helper routines.
135 bool isTypeLegal(Type *Ty, MVT &VT);
136 bool isLoadStoreTypeLegal(Type *Ty, MVT &VT);
137 bool ComputeAddress(const Value *Obj, Address &Addr, Type *Ty = nullptr);
138 bool ComputeCallAddress(const Value *V, Address &Addr);
139 bool SimplifyAddress(Address &Addr, MVT VT);
140 void AddLoadStoreOperands(Address &Addr, const MachineInstrBuilder &MIB,
141 unsigned Flags, unsigned ScaleFactor,
142 MachineMemOperand *MMO);
143 bool IsMemCpySmall(uint64_t Len, unsigned Alignment);
144 bool TryEmitSmallMemCpy(Address Dest, Address Src, uint64_t Len,
146 bool foldXALUIntrinsic(AArch64CC::CondCode &CC, const Instruction *I,
149 // Emit helper routines.
150 unsigned emitAddsSubs(bool UseAdds, MVT RetVT, const Value *LHS,
151 const Value *RHS, bool IsZExt = false,
152 bool WantResult = true);
153 unsigned emitAddsSubs_rr(bool UseAdds, MVT RetVT, unsigned LHSReg,
154 bool LHSIsKill, unsigned RHSReg, bool RHSIsKill,
155 bool WantResult = true);
156 unsigned emitAddsSubs_ri(bool UseAdds, MVT RetVT, unsigned LHSReg,
157 bool LHSIsKill, uint64_t Imm,
158 bool WantResult = true);
159 unsigned emitAddSub_rs(bool UseAdd, MVT RetVT, unsigned LHSReg,
160 bool LHSIsKill, unsigned RHSReg, bool RHSIsKill,
161 AArch64_AM::ShiftExtendType ShiftType,
162 uint64_t ShiftImm, bool WantResult = true);
163 unsigned emitAddsSubs_rs(bool UseAdds, MVT RetVT, unsigned LHSReg,
164 bool LHSIsKill, unsigned RHSReg, bool RHSIsKill,
165 AArch64_AM::ShiftExtendType ShiftType,
166 uint64_t ShiftImm, bool WantResult = true);
167 unsigned emitAddSub_rx(bool UseAdd, MVT RetVT, unsigned LHSReg,
168 bool LHSIsKill, unsigned RHSReg, bool RHSIsKill,
169 AArch64_AM::ShiftExtendType ExtType,
170 uint64_t ShiftImm, bool WantResult = true);
172 unsigned emitAddsSubs_rx(bool UseAdds, MVT RetVT, unsigned LHSReg,
173 bool LHSIsKill, unsigned RHSReg, bool RHSIsKill,
174 AArch64_AM::ShiftExtendType ExtType,
175 uint64_t ShiftImm, bool WantResult = true);
178 bool emitCmp(const Value *LHS, const Value *RHS, bool IsZExt);
179 bool emitICmp(MVT RetVT, const Value *LHS, const Value *RHS, bool IsZExt);
180 bool emitICmp_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill, uint64_t Imm);
181 bool emitFCmp(MVT RetVT, const Value *LHS, const Value *RHS);
182 bool EmitLoad(MVT VT, unsigned &ResultReg, Address Addr,
183 MachineMemOperand *MMO = nullptr);
184 bool EmitStore(MVT VT, unsigned SrcReg, Address Addr,
185 MachineMemOperand *MMO = nullptr);
186 unsigned EmitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, bool isZExt);
187 unsigned Emiti1Ext(unsigned SrcReg, MVT DestVT, bool isZExt);
188 unsigned emitAdds(MVT RetVT, const Value *LHS, const Value *RHS,
189 bool IsZExt = false, bool WantResult = true);
190 unsigned emitSubs(MVT RetVT, const Value *LHS, const Value *RHS,
191 bool IsZExt = false, bool WantResult = true);
192 unsigned emitSubs_rr(MVT RetVT, unsigned LHSReg, bool LHSIsKill,
193 unsigned RHSReg, bool RHSIsKill, bool WantResult = true);
194 unsigned emitSubs_rs(MVT RetVT, unsigned LHSReg, bool LHSIsKill,
195 unsigned RHSReg, bool RHSIsKill,
196 AArch64_AM::ShiftExtendType ShiftType, uint64_t ShiftImm,
197 bool WantResult = true);
198 unsigned emitAND_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill, uint64_t Imm);
199 unsigned Emit_MUL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
200 unsigned Op1, bool Op1IsKill);
201 unsigned Emit_SMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
202 unsigned Op1, bool Op1IsKill);
203 unsigned Emit_UMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
204 unsigned Op1, bool Op1IsKill);
205 unsigned emitLSL_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
206 unsigned Op1Reg, bool Op1IsKill);
207 unsigned emitLSL_ri(MVT RetVT, MVT SrcVT, unsigned Op0Reg, bool Op0IsKill,
208 uint64_t Imm, bool IsZExt = true);
209 unsigned emitLSR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
210 unsigned Op1Reg, bool Op1IsKill);
211 unsigned emitLSR_ri(MVT RetVT, MVT SrcVT, unsigned Op0Reg, bool Op0IsKill,
212 uint64_t Imm, bool IsZExt = true);
213 unsigned emitASR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
214 unsigned Op1Reg, bool Op1IsKill);
215 unsigned emitASR_ri(MVT RetVT, MVT SrcVT, unsigned Op0Reg, bool Op0IsKill,
216 uint64_t Imm, bool IsZExt = false);
218 unsigned AArch64MaterializeInt(const ConstantInt *CI, MVT VT);
219 unsigned AArch64MaterializeFP(const ConstantFP *CFP, MVT VT);
220 unsigned AArch64MaterializeGV(const GlobalValue *GV);
222 // Call handling routines.
224 CCAssignFn *CCAssignFnForCall(CallingConv::ID CC) const;
225 bool ProcessCallArgs(CallLoweringInfo &CLI, SmallVectorImpl<MVT> &ArgVTs,
227 bool FinishCall(CallLoweringInfo &CLI, MVT RetVT, unsigned NumBytes);
230 // Backend specific FastISel code.
231 unsigned TargetMaterializeAlloca(const AllocaInst *AI) override;
232 unsigned TargetMaterializeConstant(const Constant *C) override;
233 unsigned TargetMaterializeFloatZero(const ConstantFP* CF) override;
235 explicit AArch64FastISel(FunctionLoweringInfo &funcInfo,
236 const TargetLibraryInfo *libInfo)
237 : FastISel(funcInfo, libInfo) {
238 Subtarget = &TM.getSubtarget<AArch64Subtarget>();
239 Context = &funcInfo.Fn->getContext();
242 bool TargetSelectInstruction(const Instruction *I) override;
244 #include "AArch64GenFastISel.inc"
247 } // end anonymous namespace
249 #include "AArch64GenCallingConv.inc"
251 CCAssignFn *AArch64FastISel::CCAssignFnForCall(CallingConv::ID CC) const {
252 if (CC == CallingConv::WebKit_JS)
253 return CC_AArch64_WebKit_JS;
254 return Subtarget->isTargetDarwin() ? CC_AArch64_DarwinPCS : CC_AArch64_AAPCS;
257 unsigned AArch64FastISel::TargetMaterializeAlloca(const AllocaInst *AI) {
258 assert(TLI.getValueType(AI->getType(), true) == MVT::i64 &&
259 "Alloca should always return a pointer.");
261 // Don't handle dynamic allocas.
262 if (!FuncInfo.StaticAllocaMap.count(AI))
265 DenseMap<const AllocaInst *, int>::iterator SI =
266 FuncInfo.StaticAllocaMap.find(AI);
268 if (SI != FuncInfo.StaticAllocaMap.end()) {
269 unsigned ResultReg = createResultReg(&AArch64::GPR64spRegClass);
270 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri),
272 .addFrameIndex(SI->second)
281 unsigned AArch64FastISel::AArch64MaterializeInt(const ConstantInt *CI, MVT VT) {
286 return FastEmit_i(VT, VT, ISD::Constant, CI->getZExtValue());
288 // Create a copy from the zero register to materialize a "0" value.
289 const TargetRegisterClass *RC = (VT == MVT::i64) ? &AArch64::GPR64RegClass
290 : &AArch64::GPR32RegClass;
291 unsigned ZeroReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
292 unsigned ResultReg = createResultReg(RC);
293 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TargetOpcode::COPY),
294 ResultReg).addReg(ZeroReg, getKillRegState(true));
298 unsigned AArch64FastISel::AArch64MaterializeFP(const ConstantFP *CFP, MVT VT) {
299 // Positive zero (+0.0) has to be materialized with a fmov from the zero
300 // register, because the immediate version of fmov cannot encode zero.
301 if (CFP->isNullValue())
302 return TargetMaterializeFloatZero(CFP);
304 if (VT != MVT::f32 && VT != MVT::f64)
307 const APFloat Val = CFP->getValueAPF();
308 bool Is64Bit = (VT == MVT::f64);
309 // This checks to see if we can use FMOV instructions to materialize
310 // a constant, otherwise we have to materialize via the constant pool.
311 if (TLI.isFPImmLegal(Val, VT)) {
313 Is64Bit ? AArch64_AM::getFP64Imm(Val) : AArch64_AM::getFP32Imm(Val);
314 assert((Imm != -1) && "Cannot encode floating-point constant.");
315 unsigned Opc = Is64Bit ? AArch64::FMOVDi : AArch64::FMOVSi;
316 return FastEmitInst_i(Opc, TLI.getRegClassFor(VT), Imm);
319 // Materialize via constant pool. MachineConstantPool wants an explicit
321 unsigned Align = DL.getPrefTypeAlignment(CFP->getType());
323 Align = DL.getTypeAllocSize(CFP->getType());
325 unsigned CPI = MCP.getConstantPoolIndex(cast<Constant>(CFP), Align);
326 unsigned ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
327 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
328 ADRPReg).addConstantPoolIndex(CPI, 0, AArch64II::MO_PAGE);
330 unsigned Opc = Is64Bit ? AArch64::LDRDui : AArch64::LDRSui;
331 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
332 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
334 .addConstantPoolIndex(CPI, 0, AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
338 unsigned AArch64FastISel::AArch64MaterializeGV(const GlobalValue *GV) {
339 // We can't handle thread-local variables quickly yet.
340 if (GV->isThreadLocal())
343 // MachO still uses GOT for large code-model accesses, but ELF requires
344 // movz/movk sequences, which FastISel doesn't handle yet.
345 if (TM.getCodeModel() != CodeModel::Small && !Subtarget->isTargetMachO())
348 unsigned char OpFlags = Subtarget->ClassifyGlobalReference(GV, TM);
350 EVT DestEVT = TLI.getValueType(GV->getType(), true);
351 if (!DestEVT.isSimple())
354 unsigned ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
357 if (OpFlags & AArch64II::MO_GOT) {
359 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
361 .addGlobalAddress(GV, 0, AArch64II::MO_GOT | AArch64II::MO_PAGE);
363 ResultReg = createResultReg(&AArch64::GPR64RegClass);
364 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::LDRXui),
367 .addGlobalAddress(GV, 0, AArch64II::MO_GOT | AArch64II::MO_PAGEOFF |
371 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
373 .addGlobalAddress(GV, 0, AArch64II::MO_PAGE);
375 ResultReg = createResultReg(&AArch64::GPR64spRegClass);
376 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri),
379 .addGlobalAddress(GV, 0, AArch64II::MO_PAGEOFF | AArch64II::MO_NC)
385 unsigned AArch64FastISel::TargetMaterializeConstant(const Constant *C) {
386 EVT CEVT = TLI.getValueType(C->getType(), true);
388 // Only handle simple types.
389 if (!CEVT.isSimple())
391 MVT VT = CEVT.getSimpleVT();
393 if (const auto *CI = dyn_cast<ConstantInt>(C))
394 return AArch64MaterializeInt(CI, VT);
395 else if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
396 return AArch64MaterializeFP(CFP, VT);
397 else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
398 return AArch64MaterializeGV(GV);
403 unsigned AArch64FastISel::TargetMaterializeFloatZero(const ConstantFP* CFP) {
404 assert(CFP->isNullValue() &&
405 "Floating-point constant is not a positive zero.");
407 if (!isTypeLegal(CFP->getType(), VT))
410 if (VT != MVT::f32 && VT != MVT::f64)
413 bool Is64Bit = (VT == MVT::f64);
414 unsigned ZReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
415 unsigned Opc = Is64Bit ? AArch64::FMOVXDr : AArch64::FMOVWSr;
416 return FastEmitInst_r(Opc, TLI.getRegClassFor(VT), ZReg, /*IsKill=*/true);
419 // Computes the address to get to an object.
420 bool AArch64FastISel::ComputeAddress(const Value *Obj, Address &Addr, Type *Ty)
422 const User *U = nullptr;
423 unsigned Opcode = Instruction::UserOp1;
424 if (const Instruction *I = dyn_cast<Instruction>(Obj)) {
425 // Don't walk into other basic blocks unless the object is an alloca from
426 // another block, otherwise it may not have a virtual register assigned.
427 if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(Obj)) ||
428 FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
429 Opcode = I->getOpcode();
432 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Obj)) {
433 Opcode = C->getOpcode();
437 if (const PointerType *Ty = dyn_cast<PointerType>(Obj->getType()))
438 if (Ty->getAddressSpace() > 255)
439 // Fast instruction selection doesn't support the special
446 case Instruction::BitCast: {
447 // Look through bitcasts.
448 return ComputeAddress(U->getOperand(0), Addr, Ty);
450 case Instruction::IntToPtr: {
451 // Look past no-op inttoptrs.
452 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
453 return ComputeAddress(U->getOperand(0), Addr, Ty);
456 case Instruction::PtrToInt: {
457 // Look past no-op ptrtoints.
458 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
459 return ComputeAddress(U->getOperand(0), Addr, Ty);
462 case Instruction::GetElementPtr: {
463 Address SavedAddr = Addr;
464 uint64_t TmpOffset = Addr.getOffset();
466 // Iterate through the GEP folding the constants into offsets where
468 gep_type_iterator GTI = gep_type_begin(U);
469 for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e;
471 const Value *Op = *i;
472 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
473 const StructLayout *SL = DL.getStructLayout(STy);
474 unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
475 TmpOffset += SL->getElementOffset(Idx);
477 uint64_t S = DL.getTypeAllocSize(GTI.getIndexedType());
479 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
480 // Constant-offset addressing.
481 TmpOffset += CI->getSExtValue() * S;
484 if (canFoldAddIntoGEP(U, Op)) {
485 // A compatible add with a constant operand. Fold the constant.
487 cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
488 TmpOffset += CI->getSExtValue() * S;
489 // Iterate on the other operand.
490 Op = cast<AddOperator>(Op)->getOperand(0);
494 goto unsupported_gep;
499 // Try to grab the base operand now.
500 Addr.setOffset(TmpOffset);
501 if (ComputeAddress(U->getOperand(0), Addr, Ty))
504 // We failed, restore everything and try the other options.
510 case Instruction::Alloca: {
511 const AllocaInst *AI = cast<AllocaInst>(Obj);
512 DenseMap<const AllocaInst *, int>::iterator SI =
513 FuncInfo.StaticAllocaMap.find(AI);
514 if (SI != FuncInfo.StaticAllocaMap.end()) {
515 Addr.setKind(Address::FrameIndexBase);
516 Addr.setFI(SI->second);
521 case Instruction::Add: {
522 // Adds of constants are common and easy enough.
523 const Value *LHS = U->getOperand(0);
524 const Value *RHS = U->getOperand(1);
526 if (isa<ConstantInt>(LHS))
529 if (const ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
530 Addr.setOffset(Addr.getOffset() + (uint64_t)CI->getSExtValue());
531 return ComputeAddress(LHS, Addr, Ty);
534 Address Backup = Addr;
535 if (ComputeAddress(LHS, Addr, Ty) && ComputeAddress(RHS, Addr, Ty))
541 case Instruction::Shl:
542 if (Addr.getOffsetReg())
545 if (const auto *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
546 unsigned Val = CI->getZExtValue();
547 if (Val < 1 || Val > 3)
550 uint64_t NumBytes = 0;
551 if (Ty && Ty->isSized()) {
552 uint64_t NumBits = DL.getTypeSizeInBits(Ty);
553 NumBytes = NumBits / 8;
554 if (!isPowerOf2_64(NumBits))
558 if (NumBytes != (1ULL << Val))
562 Addr.setExtendType(AArch64_AM::LSL);
564 if (const auto *I = dyn_cast<Instruction>(U->getOperand(0)))
565 if (FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB)
568 if (const auto *ZE = dyn_cast<ZExtInst>(U))
569 if (ZE->getOperand(0)->getType()->isIntegerTy(32))
570 Addr.setExtendType(AArch64_AM::UXTW);
572 if (const auto *SE = dyn_cast<SExtInst>(U))
573 if (SE->getOperand(0)->getType()->isIntegerTy(32))
574 Addr.setExtendType(AArch64_AM::SXTW);
576 unsigned Reg = getRegForValue(U->getOperand(0));
579 Addr.setOffsetReg(Reg);
586 if (!Addr.getOffsetReg()) {
587 unsigned Reg = getRegForValue(Obj);
590 Addr.setOffsetReg(Reg);
596 unsigned Reg = getRegForValue(Obj);
603 bool AArch64FastISel::ComputeCallAddress(const Value *V, Address &Addr) {
604 const User *U = nullptr;
605 unsigned Opcode = Instruction::UserOp1;
608 if (const auto *I = dyn_cast<Instruction>(V)) {
609 Opcode = I->getOpcode();
611 InMBB = I->getParent() == FuncInfo.MBB->getBasicBlock();
612 } else if (const auto *C = dyn_cast<ConstantExpr>(V)) {
613 Opcode = C->getOpcode();
619 case Instruction::BitCast:
620 // Look past bitcasts if its operand is in the same BB.
622 return ComputeCallAddress(U->getOperand(0), Addr);
624 case Instruction::IntToPtr:
625 // Look past no-op inttoptrs if its operand is in the same BB.
627 TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
628 return ComputeCallAddress(U->getOperand(0), Addr);
630 case Instruction::PtrToInt:
631 // Look past no-op ptrtoints if its operand is in the same BB.
633 TLI.getValueType(U->getType()) == TLI.getPointerTy())
634 return ComputeCallAddress(U->getOperand(0), Addr);
638 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
639 Addr.setGlobalValue(GV);
643 // If all else fails, try to materialize the value in a register.
644 if (!Addr.getGlobalValue()) {
645 Addr.setReg(getRegForValue(V));
646 return Addr.getReg() != 0;
653 bool AArch64FastISel::isTypeLegal(Type *Ty, MVT &VT) {
654 EVT evt = TLI.getValueType(Ty, true);
656 // Only handle simple types.
657 if (evt == MVT::Other || !evt.isSimple())
659 VT = evt.getSimpleVT();
661 // This is a legal type, but it's not something we handle in fast-isel.
665 // Handle all other legal types, i.e. a register that will directly hold this
667 return TLI.isTypeLegal(VT);
670 bool AArch64FastISel::isLoadStoreTypeLegal(Type *Ty, MVT &VT) {
671 if (isTypeLegal(Ty, VT))
674 // If this is a type than can be sign or zero-extended to a basic operation
675 // go ahead and accept it now. For stores, this reflects truncation.
676 if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16)
682 bool AArch64FastISel::SimplifyAddress(Address &Addr, MVT VT) {
683 unsigned ScaleFactor;
684 switch (VT.SimpleTy) {
685 default: return false;
686 case MVT::i1: // fall-through
687 case MVT::i8: ScaleFactor = 1; break;
688 case MVT::i16: ScaleFactor = 2; break;
689 case MVT::i32: // fall-through
690 case MVT::f32: ScaleFactor = 4; break;
691 case MVT::i64: // fall-through
692 case MVT::f64: ScaleFactor = 8; break;
695 bool ImmediateOffsetNeedsLowering = false;
696 bool RegisterOffsetNeedsLowering = false;
697 int64_t Offset = Addr.getOffset();
698 if (((Offset < 0) || (Offset & (ScaleFactor - 1))) && !isInt<9>(Offset))
699 ImmediateOffsetNeedsLowering = true;
700 else if (Offset > 0 && !(Offset & (ScaleFactor - 1)) &&
701 !isUInt<12>(Offset / ScaleFactor))
702 ImmediateOffsetNeedsLowering = true;
704 // Cannot encode an offset register and an immediate offset in the same
705 // instruction. Fold the immediate offset into the load/store instruction and
706 // emit an additonal add to take care of the offset register.
707 if (!ImmediateOffsetNeedsLowering && Addr.getOffset() && Addr.isRegBase() &&
709 RegisterOffsetNeedsLowering = true;
711 // If this is a stack pointer and the offset needs to be simplified then put
712 // the alloca address into a register, set the base type back to register and
713 // continue. This should almost never happen.
714 if (ImmediateOffsetNeedsLowering && Addr.isFIBase()) {
715 unsigned ResultReg = createResultReg(&AArch64::GPR64spRegClass);
716 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri),
718 .addFrameIndex(Addr.getFI())
721 Addr.setKind(Address::RegBase);
722 Addr.setReg(ResultReg);
725 if (RegisterOffsetNeedsLowering) {
726 unsigned ResultReg = 0;
728 if (Addr.getExtendType() == AArch64_AM::SXTW ||
729 Addr.getExtendType() == AArch64_AM::UXTW )
730 ResultReg = emitAddSub_rx(/*UseAdd=*/true, MVT::i64, Addr.getReg(),
731 /*TODO:IsKill=*/false, Addr.getOffsetReg(),
732 /*TODO:IsKill=*/false, Addr.getExtendType(),
735 ResultReg = emitAddSub_rs(/*UseAdd=*/true, MVT::i64, Addr.getReg(),
736 /*TODO:IsKill=*/false, Addr.getOffsetReg(),
737 /*TODO:IsKill=*/false, AArch64_AM::LSL,
740 if (Addr.getExtendType() == AArch64_AM::UXTW)
741 ResultReg = emitLSL_ri(MVT::i64, MVT::i32, Addr.getOffsetReg(),
742 /*Op0IsKill=*/false, Addr.getShift(),
744 else if (Addr.getExtendType() == AArch64_AM::SXTW)
745 ResultReg = emitLSL_ri(MVT::i64, MVT::i32, Addr.getOffsetReg(),
746 /*Op0IsKill=*/false, Addr.getShift(),
749 ResultReg = emitLSL_ri(MVT::i64, MVT::i64, Addr.getOffsetReg(),
750 /*Op0IsKill=*/false, Addr.getShift());
755 Addr.setReg(ResultReg);
756 Addr.setOffsetReg(0);
758 Addr.setExtendType(AArch64_AM::InvalidShiftExtend);
761 // Since the offset is too large for the load/store instruction get the
762 // reg+offset into a register.
763 if (ImmediateOffsetNeedsLowering) {
764 unsigned ResultReg = 0;
766 ResultReg = FastEmit_ri_(MVT::i64, ISD::ADD, Addr.getReg(),
767 /*IsKill=*/false, Offset, MVT::i64);
769 ResultReg = FastEmit_i(MVT::i64, MVT::i64, ISD::Constant, Offset);
773 Addr.setReg(ResultReg);
779 void AArch64FastISel::AddLoadStoreOperands(Address &Addr,
780 const MachineInstrBuilder &MIB,
782 unsigned ScaleFactor,
783 MachineMemOperand *MMO) {
784 int64_t Offset = Addr.getOffset() / ScaleFactor;
785 // Frame base works a bit differently. Handle it separately.
786 if (Addr.isFIBase()) {
787 int FI = Addr.getFI();
788 // FIXME: We shouldn't be using getObjectSize/getObjectAlignment. The size
789 // and alignment should be based on the VT.
790 MMO = FuncInfo.MF->getMachineMemOperand(
791 MachinePointerInfo::getFixedStack(FI, Offset), Flags,
792 MFI.getObjectSize(FI), MFI.getObjectAlignment(FI));
793 // Now add the rest of the operands.
794 MIB.addFrameIndex(FI).addImm(Offset);
796 assert(Addr.isRegBase() && "Unexpected address kind.");
797 const MCInstrDesc &II = MIB->getDesc();
798 unsigned Idx = (Flags & MachineMemOperand::MOStore) ? 1 : 0;
800 constrainOperandRegClass(II, Addr.getReg(), II.getNumDefs()+Idx));
802 constrainOperandRegClass(II, Addr.getOffsetReg(), II.getNumDefs()+Idx+1));
803 if (Addr.getOffsetReg()) {
804 assert(Addr.getOffset() == 0 && "Unexpected offset");
805 bool IsSigned = Addr.getExtendType() == AArch64_AM::SXTW ||
806 Addr.getExtendType() == AArch64_AM::SXTX;
807 MIB.addReg(Addr.getReg());
808 MIB.addReg(Addr.getOffsetReg());
809 MIB.addImm(IsSigned);
810 MIB.addImm(Addr.getShift() != 0);
812 MIB.addReg(Addr.getReg());
818 MIB.addMemOperand(MMO);
821 unsigned AArch64FastISel::emitAddsSubs(bool UseAdds, MVT RetVT,
822 const Value *LHS, const Value *RHS,
823 bool IsZExt, bool WantResult) {
824 AArch64_AM::ShiftExtendType ExtendType = AArch64_AM::InvalidShiftExtend;
825 bool NeedExtend = false;
826 switch (RetVT.SimpleTy) {
834 ExtendType = IsZExt ? AArch64_AM::UXTB : AArch64_AM::SXTB;
838 ExtendType = IsZExt ? AArch64_AM::UXTH : AArch64_AM::SXTH;
840 case MVT::i32: // fall-through
845 RetVT.SimpleTy = std::max(RetVT.SimpleTy, MVT::i32);
847 // Canonicalize immediates to the RHS first.
848 if (UseAdds && isa<ConstantInt>(LHS) && !isa<ConstantInt>(RHS))
851 // Canonicalize shift immediate to the RHS.
853 if (const auto *SI = dyn_cast<BinaryOperator>(LHS))
854 if (isa<ConstantInt>(SI->getOperand(1)))
855 if (SI->getOpcode() == Instruction::Shl ||
856 SI->getOpcode() == Instruction::LShr ||
857 SI->getOpcode() == Instruction::AShr )
860 unsigned LHSReg = getRegForValue(LHS);
863 bool LHSIsKill = hasTrivialKill(LHS);
866 LHSReg = EmitIntExt(SrcVT, LHSReg, RetVT, IsZExt);
868 unsigned ResultReg = 0;
869 if (const auto *C = dyn_cast<ConstantInt>(RHS)) {
870 uint64_t Imm = IsZExt ? C->getZExtValue() : C->getSExtValue();
873 emitAddsSubs_ri(!UseAdds, RetVT, LHSReg, LHSIsKill, -Imm, WantResult);
876 emitAddsSubs_ri(UseAdds, RetVT, LHSReg, LHSIsKill, Imm, WantResult);
881 // Only extend the RHS within the instruction if there is a valid extend type.
882 if (ExtendType != AArch64_AM::InvalidShiftExtend) {
883 if (const auto *SI = dyn_cast<BinaryOperator>(RHS))
884 if (const auto *C = dyn_cast<ConstantInt>(SI->getOperand(1)))
885 if ((SI->getOpcode() == Instruction::Shl) && (C->getZExtValue() < 4)) {
886 unsigned RHSReg = getRegForValue(SI->getOperand(0));
889 bool RHSIsKill = hasTrivialKill(SI->getOperand(0));
890 return emitAddsSubs_rx(UseAdds, RetVT, LHSReg, LHSIsKill, RHSReg,
891 RHSIsKill, ExtendType, C->getZExtValue(),
894 unsigned RHSReg = getRegForValue(RHS);
897 bool RHSIsKill = hasTrivialKill(RHS);
898 return emitAddsSubs_rx(UseAdds, RetVT, LHSReg, LHSIsKill, RHSReg, RHSIsKill,
899 ExtendType, 0, WantResult);
902 // Check if the shift can be folded into the instruction.
903 if (const auto *SI = dyn_cast<BinaryOperator>(RHS)) {
904 if (const auto *C = dyn_cast<ConstantInt>(SI->getOperand(1))) {
905 AArch64_AM::ShiftExtendType ShiftType = AArch64_AM::InvalidShiftExtend;
906 switch (SI->getOpcode()) {
908 case Instruction::Shl: ShiftType = AArch64_AM::LSL; break;
909 case Instruction::LShr: ShiftType = AArch64_AM::LSR; break;
910 case Instruction::AShr: ShiftType = AArch64_AM::ASR; break;
912 uint64_t ShiftVal = C->getZExtValue();
913 if (ShiftType != AArch64_AM::InvalidShiftExtend) {
914 unsigned RHSReg = getRegForValue(SI->getOperand(0));
917 bool RHSIsKill = hasTrivialKill(SI->getOperand(0));
918 return emitAddsSubs_rs(UseAdds, RetVT, LHSReg, LHSIsKill, RHSReg,
919 RHSIsKill, ShiftType, ShiftVal, WantResult);
924 unsigned RHSReg = getRegForValue(RHS);
927 bool RHSIsKill = hasTrivialKill(RHS);
930 RHSReg = EmitIntExt(SrcVT, RHSReg, RetVT, IsZExt);
932 return emitAddsSubs_rr(UseAdds, RetVT, LHSReg, LHSIsKill, RHSReg, RHSIsKill,
936 unsigned AArch64FastISel::emitAddsSubs_rr(bool UseAdds, MVT RetVT,
937 unsigned LHSReg, bool LHSIsKill,
938 unsigned RHSReg, bool RHSIsKill,
940 assert(LHSReg && RHSReg && "Invalid register number.");
942 if (RetVT != MVT::i32 && RetVT != MVT::i64)
945 static const unsigned OpcTable[2][2] = {
946 { AArch64::ADDSWrr, AArch64::ADDSXrr },
947 { AArch64::SUBSWrr, AArch64::SUBSXrr }
949 unsigned Opc = OpcTable[!UseAdds][(RetVT == MVT::i64)];
952 const TargetRegisterClass *RC =
953 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
954 ResultReg = createResultReg(RC);
956 ResultReg = (RetVT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
958 const MCInstrDesc &II = TII.get(Opc);
959 LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
960 RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1);
961 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
962 .addReg(LHSReg, getKillRegState(LHSIsKill))
963 .addReg(RHSReg, getKillRegState(RHSIsKill));
968 unsigned AArch64FastISel::emitAddsSubs_ri(bool UseAdds, MVT RetVT,
969 unsigned LHSReg, bool LHSIsKill,
970 uint64_t Imm, bool WantResult) {
971 assert(LHSReg && "Invalid register number.");
973 if (RetVT != MVT::i32 && RetVT != MVT::i64)
979 else if ((Imm & 0xfff000) == Imm) {
985 static const unsigned OpcTable[2][2] = {
986 { AArch64::ADDSWri, AArch64::ADDSXri },
987 { AArch64::SUBSWri, AArch64::SUBSXri }
989 unsigned Opc = OpcTable[!UseAdds][(RetVT == MVT::i64)];
992 const TargetRegisterClass *RC =
993 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
994 ResultReg = createResultReg(RC);
996 ResultReg = (RetVT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
998 const MCInstrDesc &II = TII.get(Opc);
999 LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
1000 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
1001 .addReg(LHSReg, getKillRegState(LHSIsKill))
1003 .addImm(getShifterImm(AArch64_AM::LSL, ShiftImm));
1008 unsigned AArch64FastISel::emitAddSub_rs(bool UseAdd, MVT RetVT,
1009 unsigned LHSReg, bool LHSIsKill,
1010 unsigned RHSReg, bool RHSIsKill,
1011 AArch64_AM::ShiftExtendType ShiftType,
1012 uint64_t ShiftImm, bool WantResult) {
1013 assert(LHSReg && RHSReg && "Invalid register number.");
1015 if (RetVT != MVT::i32 && RetVT != MVT::i64)
1018 static const unsigned OpcTable[2][2] = {
1019 { AArch64::ADDWrs, AArch64::ADDXrs },
1020 { AArch64::SUBWrs, AArch64::SUBXrs }
1022 unsigned Opc = OpcTable[!UseAdd][(RetVT == MVT::i64)];
1025 const TargetRegisterClass *RC =
1026 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
1027 ResultReg = createResultReg(RC);
1029 ResultReg = (RetVT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
1031 const MCInstrDesc &II = TII.get(Opc);
1032 LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
1033 RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1);
1034 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
1035 .addReg(LHSReg, getKillRegState(LHSIsKill))
1036 .addReg(RHSReg, getKillRegState(RHSIsKill))
1037 .addImm(getShifterImm(ShiftType, ShiftImm));
1042 unsigned AArch64FastISel::emitAddsSubs_rs(bool UseAdds, MVT RetVT,
1043 unsigned LHSReg, bool LHSIsKill,
1044 unsigned RHSReg, bool RHSIsKill,
1045 AArch64_AM::ShiftExtendType ShiftType,
1046 uint64_t ShiftImm, bool WantResult) {
1047 assert(LHSReg && RHSReg && "Invalid register number.");
1049 if (RetVT != MVT::i32 && RetVT != MVT::i64)
1052 static const unsigned OpcTable[2][2] = {
1053 { AArch64::ADDSWrs, AArch64::ADDSXrs },
1054 { AArch64::SUBSWrs, AArch64::SUBSXrs }
1056 unsigned Opc = OpcTable[!UseAdds][(RetVT == MVT::i64)];
1059 const TargetRegisterClass *RC =
1060 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
1061 ResultReg = createResultReg(RC);
1063 ResultReg = (RetVT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
1065 const MCInstrDesc &II = TII.get(Opc);
1066 LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
1067 RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1);
1068 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
1069 .addReg(LHSReg, getKillRegState(LHSIsKill))
1070 .addReg(RHSReg, getKillRegState(RHSIsKill))
1071 .addImm(getShifterImm(ShiftType, ShiftImm));
1076 unsigned AArch64FastISel::emitAddSub_rx(bool UseAdd, MVT RetVT,
1077 unsigned LHSReg, bool LHSIsKill,
1078 unsigned RHSReg, bool RHSIsKill,
1079 AArch64_AM::ShiftExtendType ExtType,
1080 uint64_t ShiftImm, bool WantResult) {
1081 assert(LHSReg && RHSReg && "Invalid register number.");
1083 if (RetVT != MVT::i32 && RetVT != MVT::i64)
1086 static const unsigned OpcTable[2][2] = {
1087 { AArch64::ADDWrx, AArch64::ADDXrx },
1088 { AArch64::SUBWrx, AArch64::SUBXrx }
1090 unsigned Opc = OpcTable[!UseAdd][(RetVT == MVT::i64)];
1093 const TargetRegisterClass *RC =
1094 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
1095 ResultReg = createResultReg(RC);
1097 ResultReg = (RetVT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
1099 const MCInstrDesc &II = TII.get(Opc);
1100 LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
1101 RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1);
1102 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
1103 .addReg(LHSReg, getKillRegState(LHSIsKill))
1104 .addReg(RHSReg, getKillRegState(RHSIsKill))
1105 .addImm(getArithExtendImm(ExtType, ShiftImm));
1110 unsigned AArch64FastISel::emitAddsSubs_rx(bool UseAdds, MVT RetVT,
1111 unsigned LHSReg, bool LHSIsKill,
1112 unsigned RHSReg, bool RHSIsKill,
1113 AArch64_AM::ShiftExtendType ExtType,
1114 uint64_t ShiftImm, bool WantResult) {
1115 assert(LHSReg && RHSReg && "Invalid register number.");
1117 if (RetVT != MVT::i32 && RetVT != MVT::i64)
1120 static const unsigned OpcTable[2][2] = {
1121 { AArch64::ADDSWrx, AArch64::ADDSXrx },
1122 { AArch64::SUBSWrx, AArch64::SUBSXrx }
1124 unsigned Opc = OpcTable[!UseAdds][(RetVT == MVT::i64)];
1127 const TargetRegisterClass *RC =
1128 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
1129 ResultReg = createResultReg(RC);
1131 ResultReg = (RetVT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
1133 const MCInstrDesc &II = TII.get(Opc);
1134 LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
1135 RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1);
1136 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
1137 .addReg(LHSReg, getKillRegState(LHSIsKill))
1138 .addReg(RHSReg, getKillRegState(RHSIsKill))
1139 .addImm(getArithExtendImm(ExtType, ShiftImm));
1144 bool AArch64FastISel::emitCmp(const Value *LHS, const Value *RHS, bool IsZExt) {
1145 Type *Ty = LHS->getType();
1146 EVT EVT = TLI.getValueType(Ty, true);
1147 if (!EVT.isSimple())
1149 MVT VT = EVT.getSimpleVT();
1151 switch (VT.SimpleTy) {
1159 return emitICmp(VT, LHS, RHS, IsZExt);
1162 return emitFCmp(VT, LHS, RHS);
1166 bool AArch64FastISel::emitICmp(MVT RetVT, const Value *LHS, const Value *RHS,
1168 return emitSubs(RetVT, LHS, RHS, IsZExt, /*WantResult=*/false) != 0;
1171 bool AArch64FastISel::emitICmp_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill,
1173 return emitAddsSubs_ri(false, RetVT, LHSReg, LHSIsKill, Imm,
1174 /*WantResult=*/false) != 0;
1177 bool AArch64FastISel::emitFCmp(MVT RetVT, const Value *LHS, const Value *RHS) {
1178 if (RetVT != MVT::f32 && RetVT != MVT::f64)
1181 // Check to see if the 2nd operand is a constant that we can encode directly
1183 bool UseImm = false;
1184 if (const auto *CFP = dyn_cast<ConstantFP>(RHS))
1185 if (CFP->isZero() && !CFP->isNegative())
1188 unsigned LHSReg = getRegForValue(LHS);
1191 bool LHSIsKill = hasTrivialKill(LHS);
1194 unsigned Opc = (RetVT == MVT::f64) ? AArch64::FCMPDri : AArch64::FCMPSri;
1195 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc))
1196 .addReg(LHSReg, getKillRegState(LHSIsKill));
1200 unsigned RHSReg = getRegForValue(RHS);
1203 bool RHSIsKill = hasTrivialKill(RHS);
1205 unsigned Opc = (RetVT == MVT::f64) ? AArch64::FCMPDrr : AArch64::FCMPSrr;
1206 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc))
1207 .addReg(LHSReg, getKillRegState(LHSIsKill))
1208 .addReg(RHSReg, getKillRegState(RHSIsKill));
1212 unsigned AArch64FastISel::emitAdds(MVT RetVT, const Value *LHS,
1213 const Value *RHS, bool IsZExt,
1215 return emitAddsSubs(true, RetVT, LHS, RHS, IsZExt, WantResult);
1218 unsigned AArch64FastISel::emitSubs(MVT RetVT, const Value *LHS,
1219 const Value *RHS, bool IsZExt,
1221 return emitAddsSubs(false, RetVT, LHS, RHS, IsZExt, WantResult);
1224 unsigned AArch64FastISel::emitSubs_rr(MVT RetVT, unsigned LHSReg,
1225 bool LHSIsKill, unsigned RHSReg,
1226 bool RHSIsKill, bool WantResult) {
1227 return emitAddsSubs_rr(false, RetVT, LHSReg, LHSIsKill, RHSReg, RHSIsKill,
1231 unsigned AArch64FastISel::emitSubs_rs(MVT RetVT, unsigned LHSReg,
1232 bool LHSIsKill, unsigned RHSReg,
1234 AArch64_AM::ShiftExtendType ShiftType,
1235 uint64_t ShiftImm, bool WantResult) {
1236 return emitAddsSubs_rs(false, RetVT, LHSReg, LHSIsKill, RHSReg, RHSIsKill,
1237 ShiftType, ShiftImm, WantResult);
1240 // FIXME: This should be eventually generated automatically by tblgen.
1241 unsigned AArch64FastISel::emitAND_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill,
1243 const TargetRegisterClass *RC = nullptr;
1245 unsigned RegSize = 0;
1246 switch (RetVT.SimpleTy) {
1250 Opc = AArch64::ANDWri;
1251 RC = &AArch64::GPR32spRegClass;
1255 Opc = AArch64::ANDXri;
1256 RC = &AArch64::GPR64spRegClass;
1261 if (!AArch64_AM::isLogicalImmediate(Imm, RegSize))
1264 return FastEmitInst_ri(Opc, RC, LHSReg, LHSIsKill,
1265 AArch64_AM::encodeLogicalImmediate(Imm, RegSize));
1268 bool AArch64FastISel::EmitLoad(MVT VT, unsigned &ResultReg, Address Addr,
1269 MachineMemOperand *MMO) {
1270 // Simplify this down to something we can handle.
1271 if (!SimplifyAddress(Addr, VT))
1274 unsigned ScaleFactor;
1275 switch (VT.SimpleTy) {
1276 default: llvm_unreachable("Unexpected value type.");
1277 case MVT::i1: // fall-through
1278 case MVT::i8: ScaleFactor = 1; break;
1279 case MVT::i16: ScaleFactor = 2; break;
1280 case MVT::i32: // fall-through
1281 case MVT::f32: ScaleFactor = 4; break;
1282 case MVT::i64: // fall-through
1283 case MVT::f64: ScaleFactor = 8; break;
1286 // Negative offsets require unscaled, 9-bit, signed immediate offsets.
1287 // Otherwise, we try using scaled, 12-bit, unsigned immediate offsets.
1288 bool UseScaled = true;
1289 if ((Addr.getOffset() < 0) || (Addr.getOffset() & (ScaleFactor - 1))) {
1294 static const unsigned OpcTable[4][6] = {
1295 { AArch64::LDURBBi, AArch64::LDURHHi, AArch64::LDURWi, AArch64::LDURXi,
1296 AArch64::LDURSi, AArch64::LDURDi },
1297 { AArch64::LDRBBui, AArch64::LDRHHui, AArch64::LDRWui, AArch64::LDRXui,
1298 AArch64::LDRSui, AArch64::LDRDui },
1299 { AArch64::LDRBBroX, AArch64::LDRHHroX, AArch64::LDRWroX, AArch64::LDRXroX,
1300 AArch64::LDRSroX, AArch64::LDRDroX },
1301 { AArch64::LDRBBroW, AArch64::LDRHHroW, AArch64::LDRWroW, AArch64::LDRXroW,
1302 AArch64::LDRSroW, AArch64::LDRDroW }
1306 const TargetRegisterClass *RC;
1307 bool VTIsi1 = false;
1308 bool UseRegOffset = Addr.isRegBase() && !Addr.getOffset() && Addr.getReg() &&
1309 Addr.getOffsetReg();
1310 unsigned Idx = UseRegOffset ? 2 : UseScaled ? 1 : 0;
1311 if (Addr.getExtendType() == AArch64_AM::UXTW ||
1312 Addr.getExtendType() == AArch64_AM::SXTW)
1315 switch (VT.SimpleTy) {
1316 default: llvm_unreachable("Unexpected value type.");
1317 case MVT::i1: VTIsi1 = true; // Intentional fall-through.
1318 case MVT::i8: Opc = OpcTable[Idx][0]; RC = &AArch64::GPR32RegClass; break;
1319 case MVT::i16: Opc = OpcTable[Idx][1]; RC = &AArch64::GPR32RegClass; break;
1320 case MVT::i32: Opc = OpcTable[Idx][2]; RC = &AArch64::GPR32RegClass; break;
1321 case MVT::i64: Opc = OpcTable[Idx][3]; RC = &AArch64::GPR64RegClass; break;
1322 case MVT::f32: Opc = OpcTable[Idx][4]; RC = &AArch64::FPR32RegClass; break;
1323 case MVT::f64: Opc = OpcTable[Idx][5]; RC = &AArch64::FPR64RegClass; break;
1326 // Create the base instruction, then add the operands.
1327 ResultReg = createResultReg(RC);
1328 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1329 TII.get(Opc), ResultReg);
1330 AddLoadStoreOperands(Addr, MIB, MachineMemOperand::MOLoad, ScaleFactor, MMO);
1332 // Loading an i1 requires special handling.
1334 unsigned ANDReg = emitAND_ri(MVT::i32, ResultReg, /*IsKill=*/true, 1);
1335 assert(ANDReg && "Unexpected AND instruction emission failure.");
1341 bool AArch64FastISel::SelectLoad(const Instruction *I) {
1343 // Verify we have a legal type before going any further. Currently, we handle
1344 // simple types that will directly fit in a register (i32/f32/i64/f64) or
1345 // those that can be sign or zero-extended to a basic operation (i1/i8/i16).
1346 if (!isLoadStoreTypeLegal(I->getType(), VT) || cast<LoadInst>(I)->isAtomic())
1349 // See if we can handle this address.
1351 if (!ComputeAddress(I->getOperand(0), Addr, I->getType()))
1355 if (!EmitLoad(VT, ResultReg, Addr, createMachineMemOperandFor(I)))
1358 UpdateValueMap(I, ResultReg);
1362 bool AArch64FastISel::EmitStore(MVT VT, unsigned SrcReg, Address Addr,
1363 MachineMemOperand *MMO) {
1364 // Simplify this down to something we can handle.
1365 if (!SimplifyAddress(Addr, VT))
1368 unsigned ScaleFactor;
1369 switch (VT.SimpleTy) {
1370 default: llvm_unreachable("Unexpected value type.");
1371 case MVT::i1: // fall-through
1372 case MVT::i8: ScaleFactor = 1; break;
1373 case MVT::i16: ScaleFactor = 2; break;
1374 case MVT::i32: // fall-through
1375 case MVT::f32: ScaleFactor = 4; break;
1376 case MVT::i64: // fall-through
1377 case MVT::f64: ScaleFactor = 8; break;
1380 // Negative offsets require unscaled, 9-bit, signed immediate offsets.
1381 // Otherwise, we try using scaled, 12-bit, unsigned immediate offsets.
1382 bool UseScaled = true;
1383 if ((Addr.getOffset() < 0) || (Addr.getOffset() & (ScaleFactor - 1))) {
1389 static const unsigned OpcTable[4][6] = {
1390 { AArch64::STURBBi, AArch64::STURHHi, AArch64::STURWi, AArch64::STURXi,
1391 AArch64::STURSi, AArch64::STURDi },
1392 { AArch64::STRBBui, AArch64::STRHHui, AArch64::STRWui, AArch64::STRXui,
1393 AArch64::STRSui, AArch64::STRDui },
1394 { AArch64::STRBBroX, AArch64::STRHHroX, AArch64::STRWroX, AArch64::STRXroX,
1395 AArch64::STRSroX, AArch64::STRDroX },
1396 { AArch64::STRBBroW, AArch64::STRHHroW, AArch64::STRWroW, AArch64::STRXroW,
1397 AArch64::STRSroW, AArch64::STRDroW }
1402 bool VTIsi1 = false;
1403 bool UseRegOffset = Addr.isRegBase() && !Addr.getOffset() && Addr.getReg() &&
1404 Addr.getOffsetReg();
1405 unsigned Idx = UseRegOffset ? 2 : UseScaled ? 1 : 0;
1406 if (Addr.getExtendType() == AArch64_AM::UXTW ||
1407 Addr.getExtendType() == AArch64_AM::SXTW)
1410 switch (VT.SimpleTy) {
1411 default: llvm_unreachable("Unexpected value type.");
1412 case MVT::i1: VTIsi1 = true;
1413 case MVT::i8: Opc = OpcTable[Idx][0]; break;
1414 case MVT::i16: Opc = OpcTable[Idx][1]; break;
1415 case MVT::i32: Opc = OpcTable[Idx][2]; break;
1416 case MVT::i64: Opc = OpcTable[Idx][3]; break;
1417 case MVT::f32: Opc = OpcTable[Idx][4]; break;
1418 case MVT::f64: Opc = OpcTable[Idx][5]; break;
1421 // Storing an i1 requires special handling.
1423 unsigned ANDReg = emitAND_ri(MVT::i32, SrcReg, /*TODO:IsKill=*/false, 1);
1424 assert(ANDReg && "Unexpected AND instruction emission failure.");
1427 // Create the base instruction, then add the operands.
1428 const MCInstrDesc &II = TII.get(Opc);
1429 SrcReg = constrainOperandRegClass(II, SrcReg, II.getNumDefs());
1430 MachineInstrBuilder MIB =
1431 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II).addReg(SrcReg);
1432 AddLoadStoreOperands(Addr, MIB, MachineMemOperand::MOStore, ScaleFactor, MMO);
1437 bool AArch64FastISel::SelectStore(const Instruction *I) {
1439 Value *Op0 = I->getOperand(0);
1440 // Verify we have a legal type before going any further. Currently, we handle
1441 // simple types that will directly fit in a register (i32/f32/i64/f64) or
1442 // those that can be sign or zero-extended to a basic operation (i1/i8/i16).
1443 if (!isLoadStoreTypeLegal(Op0->getType(), VT) ||
1444 cast<StoreInst>(I)->isAtomic())
1447 // Get the value to be stored into a register.
1448 unsigned SrcReg = getRegForValue(Op0);
1452 // See if we can handle this address.
1454 if (!ComputeAddress(I->getOperand(1), Addr, I->getOperand(0)->getType()))
1457 if (!EmitStore(VT, SrcReg, Addr, createMachineMemOperandFor(I)))
1462 static AArch64CC::CondCode getCompareCC(CmpInst::Predicate Pred) {
1464 case CmpInst::FCMP_ONE:
1465 case CmpInst::FCMP_UEQ:
1467 // AL is our "false" for now. The other two need more compares.
1468 return AArch64CC::AL;
1469 case CmpInst::ICMP_EQ:
1470 case CmpInst::FCMP_OEQ:
1471 return AArch64CC::EQ;
1472 case CmpInst::ICMP_SGT:
1473 case CmpInst::FCMP_OGT:
1474 return AArch64CC::GT;
1475 case CmpInst::ICMP_SGE:
1476 case CmpInst::FCMP_OGE:
1477 return AArch64CC::GE;
1478 case CmpInst::ICMP_UGT:
1479 case CmpInst::FCMP_UGT:
1480 return AArch64CC::HI;
1481 case CmpInst::FCMP_OLT:
1482 return AArch64CC::MI;
1483 case CmpInst::ICMP_ULE:
1484 case CmpInst::FCMP_OLE:
1485 return AArch64CC::LS;
1486 case CmpInst::FCMP_ORD:
1487 return AArch64CC::VC;
1488 case CmpInst::FCMP_UNO:
1489 return AArch64CC::VS;
1490 case CmpInst::FCMP_UGE:
1491 return AArch64CC::PL;
1492 case CmpInst::ICMP_SLT:
1493 case CmpInst::FCMP_ULT:
1494 return AArch64CC::LT;
1495 case CmpInst::ICMP_SLE:
1496 case CmpInst::FCMP_ULE:
1497 return AArch64CC::LE;
1498 case CmpInst::FCMP_UNE:
1499 case CmpInst::ICMP_NE:
1500 return AArch64CC::NE;
1501 case CmpInst::ICMP_UGE:
1502 return AArch64CC::HS;
1503 case CmpInst::ICMP_ULT:
1504 return AArch64CC::LO;
1508 bool AArch64FastISel::SelectBranch(const Instruction *I) {
1509 const BranchInst *BI = cast<BranchInst>(I);
1510 MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
1511 MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
1513 AArch64CC::CondCode CC = AArch64CC::NE;
1514 if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
1515 if (CI->hasOneUse() && (CI->getParent() == I->getParent())) {
1516 // We may not handle every CC for now.
1517 CC = getCompareCC(CI->getPredicate());
1518 if (CC == AArch64CC::AL)
1522 if (!emitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
1526 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
1530 // Obtain the branch weight and add the TrueBB to the successor list.
1531 uint32_t BranchWeight = 0;
1533 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
1534 TBB->getBasicBlock());
1535 FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
1537 FastEmitBranch(FBB, DbgLoc);
1540 } else if (TruncInst *TI = dyn_cast<TruncInst>(BI->getCondition())) {
1542 if (TI->hasOneUse() && TI->getParent() == I->getParent() &&
1543 (isLoadStoreTypeLegal(TI->getOperand(0)->getType(), SrcVT))) {
1544 unsigned CondReg = getRegForValue(TI->getOperand(0));
1547 bool CondIsKill = hasTrivialKill(TI->getOperand(0));
1549 // Issue an extract_subreg to get the lower 32-bits.
1550 if (SrcVT == MVT::i64) {
1551 CondReg = FastEmitInst_extractsubreg(MVT::i32, CondReg, CondIsKill,
1556 unsigned ANDReg = emitAND_ri(MVT::i32, CondReg, CondIsKill, 1);
1557 assert(ANDReg && "Unexpected AND instruction emission failure.");
1558 emitICmp_ri(MVT::i32, ANDReg, /*IsKill=*/true, 0);
1560 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
1561 std::swap(TBB, FBB);
1564 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
1568 // Obtain the branch weight and add the TrueBB to the successor list.
1569 uint32_t BranchWeight = 0;
1571 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
1572 TBB->getBasicBlock());
1573 FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
1575 FastEmitBranch(FBB, DbgLoc);
1578 } else if (const ConstantInt *CI =
1579 dyn_cast<ConstantInt>(BI->getCondition())) {
1580 uint64_t Imm = CI->getZExtValue();
1581 MachineBasicBlock *Target = (Imm == 0) ? FBB : TBB;
1582 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::B))
1585 // Obtain the branch weight and add the target to the successor list.
1586 uint32_t BranchWeight = 0;
1588 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
1589 Target->getBasicBlock());
1590 FuncInfo.MBB->addSuccessor(Target, BranchWeight);
1592 } else if (foldXALUIntrinsic(CC, I, BI->getCondition())) {
1593 // Fake request the condition, otherwise the intrinsic might be completely
1595 unsigned CondReg = getRegForValue(BI->getCondition());
1600 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
1604 // Obtain the branch weight and add the TrueBB to the successor list.
1605 uint32_t BranchWeight = 0;
1607 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
1608 TBB->getBasicBlock());
1609 FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
1611 FastEmitBranch(FBB, DbgLoc);
1615 unsigned CondReg = getRegForValue(BI->getCondition());
1618 bool CondRegIsKill = hasTrivialKill(BI->getCondition());
1620 // We've been divorced from our compare! Our block was split, and
1621 // now our compare lives in a predecessor block. We musn't
1622 // re-compare here, as the children of the compare aren't guaranteed
1623 // live across the block boundary (we *could* check for this).
1624 // Regardless, the compare has been done in the predecessor block,
1625 // and it left a value for us in a virtual register. Ergo, we test
1626 // the one-bit value left in the virtual register.
1627 emitICmp_ri(MVT::i32, CondReg, CondRegIsKill, 0);
1629 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
1630 std::swap(TBB, FBB);
1634 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
1638 // Obtain the branch weight and add the TrueBB to the successor list.
1639 uint32_t BranchWeight = 0;
1641 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
1642 TBB->getBasicBlock());
1643 FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
1645 FastEmitBranch(FBB, DbgLoc);
1649 bool AArch64FastISel::SelectIndirectBr(const Instruction *I) {
1650 const IndirectBrInst *BI = cast<IndirectBrInst>(I);
1651 unsigned AddrReg = getRegForValue(BI->getOperand(0));
1655 // Emit the indirect branch.
1656 const MCInstrDesc &II = TII.get(AArch64::BR);
1657 AddrReg = constrainOperandRegClass(II, AddrReg, II.getNumDefs());
1658 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II).addReg(AddrReg);
1660 // Make sure the CFG is up-to-date.
1661 for (unsigned i = 0, e = BI->getNumSuccessors(); i != e; ++i)
1662 FuncInfo.MBB->addSuccessor(FuncInfo.MBBMap[BI->getSuccessor(i)]);
1667 bool AArch64FastISel::SelectCmp(const Instruction *I) {
1668 const CmpInst *CI = cast<CmpInst>(I);
1670 // We may not handle every CC for now.
1671 AArch64CC::CondCode CC = getCompareCC(CI->getPredicate());
1672 if (CC == AArch64CC::AL)
1676 if (!emitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
1679 // Now set a register based on the comparison.
1680 AArch64CC::CondCode invertedCC = getInvertedCondCode(CC);
1681 unsigned ResultReg = createResultReg(&AArch64::GPR32RegClass);
1682 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::CSINCWr),
1684 .addReg(AArch64::WZR)
1685 .addReg(AArch64::WZR)
1686 .addImm(invertedCC);
1688 UpdateValueMap(I, ResultReg);
1692 bool AArch64FastISel::SelectSelect(const Instruction *I) {
1693 const SelectInst *SI = cast<SelectInst>(I);
1695 EVT DestEVT = TLI.getValueType(SI->getType(), true);
1696 if (!DestEVT.isSimple())
1699 MVT DestVT = DestEVT.getSimpleVT();
1700 if (DestVT != MVT::i32 && DestVT != MVT::i64 && DestVT != MVT::f32 &&
1705 const TargetRegisterClass *RC = nullptr;
1706 switch (DestVT.SimpleTy) {
1707 default: return false;
1709 SelectOpc = AArch64::CSELWr; RC = &AArch64::GPR32RegClass; break;
1711 SelectOpc = AArch64::CSELXr; RC = &AArch64::GPR64RegClass; break;
1713 SelectOpc = AArch64::FCSELSrrr; RC = &AArch64::FPR32RegClass; break;
1715 SelectOpc = AArch64::FCSELDrrr; RC = &AArch64::FPR64RegClass; break;
1718 const Value *Cond = SI->getCondition();
1719 bool NeedTest = true;
1720 AArch64CC::CondCode CC = AArch64CC::NE;
1721 if (foldXALUIntrinsic(CC, I, Cond))
1724 unsigned CondReg = getRegForValue(Cond);
1727 bool CondIsKill = hasTrivialKill(Cond);
1730 unsigned ANDReg = emitAND_ri(MVT::i32, CondReg, CondIsKill, 1);
1731 assert(ANDReg && "Unexpected AND instruction emission failure.");
1732 emitICmp_ri(MVT::i32, ANDReg, /*IsKill=*/true, 0);
1735 unsigned TrueReg = getRegForValue(SI->getTrueValue());
1736 bool TrueIsKill = hasTrivialKill(SI->getTrueValue());
1738 unsigned FalseReg = getRegForValue(SI->getFalseValue());
1739 bool FalseIsKill = hasTrivialKill(SI->getFalseValue());
1741 if (!TrueReg || !FalseReg)
1744 unsigned ResultReg = FastEmitInst_rri(SelectOpc, RC, TrueReg, TrueIsKill,
1745 FalseReg, FalseIsKill, CC);
1746 UpdateValueMap(I, ResultReg);
1750 bool AArch64FastISel::SelectFPExt(const Instruction *I) {
1751 Value *V = I->getOperand(0);
1752 if (!I->getType()->isDoubleTy() || !V->getType()->isFloatTy())
1755 unsigned Op = getRegForValue(V);
1759 unsigned ResultReg = createResultReg(&AArch64::FPR64RegClass);
1760 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::FCVTDSr),
1761 ResultReg).addReg(Op);
1762 UpdateValueMap(I, ResultReg);
1766 bool AArch64FastISel::SelectFPTrunc(const Instruction *I) {
1767 Value *V = I->getOperand(0);
1768 if (!I->getType()->isFloatTy() || !V->getType()->isDoubleTy())
1771 unsigned Op = getRegForValue(V);
1775 unsigned ResultReg = createResultReg(&AArch64::FPR32RegClass);
1776 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::FCVTSDr),
1777 ResultReg).addReg(Op);
1778 UpdateValueMap(I, ResultReg);
1782 // FPToUI and FPToSI
1783 bool AArch64FastISel::SelectFPToInt(const Instruction *I, bool Signed) {
1785 if (!isTypeLegal(I->getType(), DestVT) || DestVT.isVector())
1788 unsigned SrcReg = getRegForValue(I->getOperand(0));
1792 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType(), true);
1793 if (SrcVT == MVT::f128)
1797 if (SrcVT == MVT::f64) {
1799 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZSUWDr : AArch64::FCVTZSUXDr;
1801 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZUUWDr : AArch64::FCVTZUUXDr;
1804 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZSUWSr : AArch64::FCVTZSUXSr;
1806 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZUUWSr : AArch64::FCVTZUUXSr;
1808 unsigned ResultReg = createResultReg(
1809 DestVT == MVT::i32 ? &AArch64::GPR32RegClass : &AArch64::GPR64RegClass);
1810 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
1812 UpdateValueMap(I, ResultReg);
1816 bool AArch64FastISel::SelectIntToFP(const Instruction *I, bool Signed) {
1818 if (!isTypeLegal(I->getType(), DestVT) || DestVT.isVector())
1820 assert ((DestVT == MVT::f32 || DestVT == MVT::f64) &&
1821 "Unexpected value type.");
1823 unsigned SrcReg = getRegForValue(I->getOperand(0));
1826 bool SrcIsKill = hasTrivialKill(I->getOperand(0));
1828 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType(), true);
1830 // Handle sign-extension.
1831 if (SrcVT == MVT::i16 || SrcVT == MVT::i8 || SrcVT == MVT::i1) {
1833 EmitIntExt(SrcVT.getSimpleVT(), SrcReg, MVT::i32, /*isZExt*/ !Signed);
1840 if (SrcVT == MVT::i64) {
1842 Opc = (DestVT == MVT::f32) ? AArch64::SCVTFUXSri : AArch64::SCVTFUXDri;
1844 Opc = (DestVT == MVT::f32) ? AArch64::UCVTFUXSri : AArch64::UCVTFUXDri;
1847 Opc = (DestVT == MVT::f32) ? AArch64::SCVTFUWSri : AArch64::SCVTFUWDri;
1849 Opc = (DestVT == MVT::f32) ? AArch64::UCVTFUWSri : AArch64::UCVTFUWDri;
1852 unsigned ResultReg = FastEmitInst_r(Opc, TLI.getRegClassFor(DestVT), SrcReg,
1854 UpdateValueMap(I, ResultReg);
1858 bool AArch64FastISel::FastLowerArguments() {
1859 if (!FuncInfo.CanLowerReturn)
1862 const Function *F = FuncInfo.Fn;
1866 CallingConv::ID CC = F->getCallingConv();
1867 if (CC != CallingConv::C)
1870 // Only handle simple cases like i1/i8/i16/i32/i64/f32/f64 of up to 8 GPR and
1872 unsigned GPRCnt = 0;
1873 unsigned FPRCnt = 0;
1875 for (auto const &Arg : F->args()) {
1876 // The first argument is at index 1.
1878 if (F->getAttributes().hasAttribute(Idx, Attribute::ByVal) ||
1879 F->getAttributes().hasAttribute(Idx, Attribute::InReg) ||
1880 F->getAttributes().hasAttribute(Idx, Attribute::StructRet) ||
1881 F->getAttributes().hasAttribute(Idx, Attribute::Nest))
1884 Type *ArgTy = Arg.getType();
1885 if (ArgTy->isStructTy() || ArgTy->isArrayTy() || ArgTy->isVectorTy())
1888 EVT ArgVT = TLI.getValueType(ArgTy);
1889 if (!ArgVT.isSimple()) return false;
1890 switch (ArgVT.getSimpleVT().SimpleTy) {
1891 default: return false;
1906 if (GPRCnt > 8 || FPRCnt > 8)
1910 static const MCPhysReg Registers[5][8] = {
1911 { AArch64::W0, AArch64::W1, AArch64::W2, AArch64::W3, AArch64::W4,
1912 AArch64::W5, AArch64::W6, AArch64::W7 },
1913 { AArch64::X0, AArch64::X1, AArch64::X2, AArch64::X3, AArch64::X4,
1914 AArch64::X5, AArch64::X6, AArch64::X7 },
1915 { AArch64::H0, AArch64::H1, AArch64::H2, AArch64::H3, AArch64::H4,
1916 AArch64::H5, AArch64::H6, AArch64::H7 },
1917 { AArch64::S0, AArch64::S1, AArch64::S2, AArch64::S3, AArch64::S4,
1918 AArch64::S5, AArch64::S6, AArch64::S7 },
1919 { AArch64::D0, AArch64::D1, AArch64::D2, AArch64::D3, AArch64::D4,
1920 AArch64::D5, AArch64::D6, AArch64::D7 }
1923 unsigned GPRIdx = 0;
1924 unsigned FPRIdx = 0;
1925 for (auto const &Arg : F->args()) {
1926 MVT VT = TLI.getSimpleValueType(Arg.getType());
1928 const TargetRegisterClass *RC = nullptr;
1929 switch (VT.SimpleTy) {
1930 default: llvm_unreachable("Unexpected value type.");
1933 case MVT::i16: VT = MVT::i32; // fall-through
1935 SrcReg = Registers[0][GPRIdx++]; RC = &AArch64::GPR32RegClass; break;
1937 SrcReg = Registers[1][GPRIdx++]; RC = &AArch64::GPR64RegClass; break;
1939 SrcReg = Registers[2][FPRIdx++]; RC = &AArch64::FPR16RegClass; break;
1941 SrcReg = Registers[3][FPRIdx++]; RC = &AArch64::FPR32RegClass; break;
1943 SrcReg = Registers[4][FPRIdx++]; RC = &AArch64::FPR64RegClass; break;
1946 // Skip unused arguments.
1947 if (Arg.use_empty()) {
1948 UpdateValueMap(&Arg, 0);
1952 unsigned DstReg = FuncInfo.MF->addLiveIn(SrcReg, RC);
1953 // FIXME: Unfortunately it's necessary to emit a copy from the livein copy.
1954 // Without this, EmitLiveInCopies may eliminate the livein if its only
1955 // use is a bitcast (which isn't turned into an instruction).
1956 unsigned ResultReg = createResultReg(RC);
1957 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1958 TII.get(TargetOpcode::COPY), ResultReg)
1959 .addReg(DstReg, getKillRegState(true));
1960 UpdateValueMap(&Arg, ResultReg);
1965 bool AArch64FastISel::ProcessCallArgs(CallLoweringInfo &CLI,
1966 SmallVectorImpl<MVT> &OutVTs,
1967 unsigned &NumBytes) {
1968 CallingConv::ID CC = CLI.CallConv;
1969 SmallVector<CCValAssign, 16> ArgLocs;
1970 CCState CCInfo(CC, false, *FuncInfo.MF, ArgLocs, *Context);
1971 CCInfo.AnalyzeCallOperands(OutVTs, CLI.OutFlags, CCAssignFnForCall(CC));
1973 // Get a count of how many bytes are to be pushed on the stack.
1974 NumBytes = CCInfo.getNextStackOffset();
1976 // Issue CALLSEQ_START
1977 unsigned AdjStackDown = TII.getCallFrameSetupOpcode();
1978 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackDown))
1981 // Process the args.
1982 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1983 CCValAssign &VA = ArgLocs[i];
1984 const Value *ArgVal = CLI.OutVals[VA.getValNo()];
1985 MVT ArgVT = OutVTs[VA.getValNo()];
1987 unsigned ArgReg = getRegForValue(ArgVal);
1991 // Handle arg promotion: SExt, ZExt, AExt.
1992 switch (VA.getLocInfo()) {
1993 case CCValAssign::Full:
1995 case CCValAssign::SExt: {
1996 MVT DestVT = VA.getLocVT();
1998 ArgReg = EmitIntExt(SrcVT, ArgReg, DestVT, /*isZExt=*/false);
2003 case CCValAssign::AExt:
2004 // Intentional fall-through.
2005 case CCValAssign::ZExt: {
2006 MVT DestVT = VA.getLocVT();
2008 ArgReg = EmitIntExt(SrcVT, ArgReg, DestVT, /*isZExt=*/true);
2014 llvm_unreachable("Unknown arg promotion!");
2017 // Now copy/store arg to correct locations.
2018 if (VA.isRegLoc() && !VA.needsCustom()) {
2019 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2020 TII.get(TargetOpcode::COPY), VA.getLocReg()).addReg(ArgReg);
2021 CLI.OutRegs.push_back(VA.getLocReg());
2022 } else if (VA.needsCustom()) {
2023 // FIXME: Handle custom args.
2026 assert(VA.isMemLoc() && "Assuming store on stack.");
2028 // Don't emit stores for undef values.
2029 if (isa<UndefValue>(ArgVal))
2032 // Need to store on the stack.
2033 unsigned ArgSize = (ArgVT.getSizeInBits() + 7) / 8;
2035 unsigned BEAlign = 0;
2036 if (ArgSize < 8 && !Subtarget->isLittleEndian())
2037 BEAlign = 8 - ArgSize;
2040 Addr.setKind(Address::RegBase);
2041 Addr.setReg(AArch64::SP);
2042 Addr.setOffset(VA.getLocMemOffset() + BEAlign);
2044 unsigned Alignment = DL.getABITypeAlignment(ArgVal->getType());
2045 MachineMemOperand *MMO = FuncInfo.MF->getMachineMemOperand(
2046 MachinePointerInfo::getStack(Addr.getOffset()),
2047 MachineMemOperand::MOStore, ArgVT.getStoreSize(), Alignment);
2049 if (!EmitStore(ArgVT, ArgReg, Addr, MMO))
2056 bool AArch64FastISel::FinishCall(CallLoweringInfo &CLI, MVT RetVT,
2057 unsigned NumBytes) {
2058 CallingConv::ID CC = CLI.CallConv;
2060 // Issue CALLSEQ_END
2061 unsigned AdjStackUp = TII.getCallFrameDestroyOpcode();
2062 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackUp))
2063 .addImm(NumBytes).addImm(0);
2065 // Now the return value.
2066 if (RetVT != MVT::isVoid) {
2067 SmallVector<CCValAssign, 16> RVLocs;
2068 CCState CCInfo(CC, false, *FuncInfo.MF, RVLocs, *Context);
2069 CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC));
2071 // Only handle a single return value.
2072 if (RVLocs.size() != 1)
2075 // Copy all of the result registers out of their specified physreg.
2076 MVT CopyVT = RVLocs[0].getValVT();
2077 unsigned ResultReg = createResultReg(TLI.getRegClassFor(CopyVT));
2078 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2079 TII.get(TargetOpcode::COPY), ResultReg)
2080 .addReg(RVLocs[0].getLocReg());
2081 CLI.InRegs.push_back(RVLocs[0].getLocReg());
2083 CLI.ResultReg = ResultReg;
2084 CLI.NumResultRegs = 1;
2090 bool AArch64FastISel::FastLowerCall(CallLoweringInfo &CLI) {
2091 CallingConv::ID CC = CLI.CallConv;
2092 bool IsTailCall = CLI.IsTailCall;
2093 bool IsVarArg = CLI.IsVarArg;
2094 const Value *Callee = CLI.Callee;
2095 const char *SymName = CLI.SymName;
2097 // Allow SelectionDAG isel to handle tail calls.
2101 CodeModel::Model CM = TM.getCodeModel();
2102 // Only support the small and large code model.
2103 if (CM != CodeModel::Small && CM != CodeModel::Large)
2106 // FIXME: Add large code model support for ELF.
2107 if (CM == CodeModel::Large && !Subtarget->isTargetMachO())
2110 // Let SDISel handle vararg functions.
2114 // FIXME: Only handle *simple* calls for now.
2116 if (CLI.RetTy->isVoidTy())
2117 RetVT = MVT::isVoid;
2118 else if (!isTypeLegal(CLI.RetTy, RetVT))
2121 for (auto Flag : CLI.OutFlags)
2122 if (Flag.isInReg() || Flag.isSRet() || Flag.isNest() || Flag.isByVal())
2125 // Set up the argument vectors.
2126 SmallVector<MVT, 16> OutVTs;
2127 OutVTs.reserve(CLI.OutVals.size());
2129 for (auto *Val : CLI.OutVals) {
2131 if (!isTypeLegal(Val->getType(), VT) &&
2132 !(VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16))
2135 // We don't handle vector parameters yet.
2136 if (VT.isVector() || VT.getSizeInBits() > 64)
2139 OutVTs.push_back(VT);
2143 if (!ComputeCallAddress(Callee, Addr))
2146 // Handle the arguments now that we've gotten them.
2148 if (!ProcessCallArgs(CLI, OutVTs, NumBytes))
2152 MachineInstrBuilder MIB;
2153 if (CM == CodeModel::Small) {
2154 unsigned CallOpc = Addr.getReg() ? AArch64::BLR : AArch64::BL;
2155 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CallOpc));
2157 MIB.addExternalSymbol(SymName, 0);
2158 else if (Addr.getGlobalValue())
2159 MIB.addGlobalAddress(Addr.getGlobalValue(), 0, 0);
2160 else if (Addr.getReg())
2161 MIB.addReg(Addr.getReg());
2165 unsigned CallReg = 0;
2167 unsigned ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
2168 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
2170 .addExternalSymbol(SymName, AArch64II::MO_GOT | AArch64II::MO_PAGE);
2172 CallReg = createResultReg(&AArch64::GPR64RegClass);
2173 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::LDRXui),
2176 .addExternalSymbol(SymName, AArch64II::MO_GOT | AArch64II::MO_PAGEOFF |
2178 } else if (Addr.getGlobalValue()) {
2179 CallReg = AArch64MaterializeGV(Addr.getGlobalValue());
2180 } else if (Addr.getReg())
2181 CallReg = Addr.getReg();
2186 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2187 TII.get(AArch64::BLR)).addReg(CallReg);
2190 // Add implicit physical register uses to the call.
2191 for (auto Reg : CLI.OutRegs)
2192 MIB.addReg(Reg, RegState::Implicit);
2194 // Add a register mask with the call-preserved registers.
2195 // Proper defs for return values will be added by setPhysRegsDeadExcept().
2196 MIB.addRegMask(TRI.getCallPreservedMask(CC));
2200 // Finish off the call including any return values.
2201 return FinishCall(CLI, RetVT, NumBytes);
2204 bool AArch64FastISel::IsMemCpySmall(uint64_t Len, unsigned Alignment) {
2206 return Len / Alignment <= 4;
2211 bool AArch64FastISel::TryEmitSmallMemCpy(Address Dest, Address Src,
2212 uint64_t Len, unsigned Alignment) {
2213 // Make sure we don't bloat code by inlining very large memcpy's.
2214 if (!IsMemCpySmall(Len, Alignment))
2217 int64_t UnscaledOffset = 0;
2218 Address OrigDest = Dest;
2219 Address OrigSrc = Src;
2223 if (!Alignment || Alignment >= 8) {
2234 // Bound based on alignment.
2235 if (Len >= 4 && Alignment == 4)
2237 else if (Len >= 2 && Alignment == 2)
2246 RV = EmitLoad(VT, ResultReg, Src);
2250 RV = EmitStore(VT, ResultReg, Dest);
2254 int64_t Size = VT.getSizeInBits() / 8;
2256 UnscaledOffset += Size;
2258 // We need to recompute the unscaled offset for each iteration.
2259 Dest.setOffset(OrigDest.getOffset() + UnscaledOffset);
2260 Src.setOffset(OrigSrc.getOffset() + UnscaledOffset);
2266 /// \brief Check if it is possible to fold the condition from the XALU intrinsic
2267 /// into the user. The condition code will only be updated on success.
2268 bool AArch64FastISel::foldXALUIntrinsic(AArch64CC::CondCode &CC,
2269 const Instruction *I,
2270 const Value *Cond) {
2271 if (!isa<ExtractValueInst>(Cond))
2274 const auto *EV = cast<ExtractValueInst>(Cond);
2275 if (!isa<IntrinsicInst>(EV->getAggregateOperand()))
2278 const auto *II = cast<IntrinsicInst>(EV->getAggregateOperand());
2280 const Function *Callee = II->getCalledFunction();
2282 cast<StructType>(Callee->getReturnType())->getTypeAtIndex(0U);
2283 if (!isTypeLegal(RetTy, RetVT))
2286 if (RetVT != MVT::i32 && RetVT != MVT::i64)
2289 AArch64CC::CondCode TmpCC;
2290 switch (II->getIntrinsicID()) {
2291 default: return false;
2292 case Intrinsic::sadd_with_overflow:
2293 case Intrinsic::ssub_with_overflow: TmpCC = AArch64CC::VS; break;
2294 case Intrinsic::uadd_with_overflow: TmpCC = AArch64CC::HS; break;
2295 case Intrinsic::usub_with_overflow: TmpCC = AArch64CC::LO; break;
2296 case Intrinsic::smul_with_overflow:
2297 case Intrinsic::umul_with_overflow: TmpCC = AArch64CC::NE; break;
2300 // Check if both instructions are in the same basic block.
2301 if (II->getParent() != I->getParent())
2304 // Make sure nothing is in the way
2305 BasicBlock::const_iterator Start = I;
2306 BasicBlock::const_iterator End = II;
2307 for (auto Itr = std::prev(Start); Itr != End; --Itr) {
2308 // We only expect extractvalue instructions between the intrinsic and the
2309 // instruction to be selected.
2310 if (!isa<ExtractValueInst>(Itr))
2313 // Check that the extractvalue operand comes from the intrinsic.
2314 const auto *EVI = cast<ExtractValueInst>(Itr);
2315 if (EVI->getAggregateOperand() != II)
2323 bool AArch64FastISel::FastLowerIntrinsicCall(const IntrinsicInst *II) {
2324 // FIXME: Handle more intrinsics.
2325 switch (II->getIntrinsicID()) {
2326 default: return false;
2327 case Intrinsic::frameaddress: {
2328 MachineFrameInfo *MFI = FuncInfo.MF->getFrameInfo();
2329 MFI->setFrameAddressIsTaken(true);
2331 const AArch64RegisterInfo *RegInfo =
2332 static_cast<const AArch64RegisterInfo *>(
2333 TM.getSubtargetImpl()->getRegisterInfo());
2334 unsigned FramePtr = RegInfo->getFrameRegister(*(FuncInfo.MF));
2335 unsigned SrcReg = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
2336 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2337 TII.get(TargetOpcode::COPY), SrcReg).addReg(FramePtr);
2338 // Recursively load frame address
2344 unsigned Depth = cast<ConstantInt>(II->getOperand(0))->getZExtValue();
2346 DestReg = FastEmitInst_ri(AArch64::LDRXui, &AArch64::GPR64RegClass,
2347 SrcReg, /*IsKill=*/true, 0);
2348 assert(DestReg && "Unexpected LDR instruction emission failure.");
2352 UpdateValueMap(II, SrcReg);
2355 case Intrinsic::memcpy:
2356 case Intrinsic::memmove: {
2357 const auto *MTI = cast<MemTransferInst>(II);
2358 // Don't handle volatile.
2359 if (MTI->isVolatile())
2362 // Disable inlining for memmove before calls to ComputeAddress. Otherwise,
2363 // we would emit dead code because we don't currently handle memmoves.
2364 bool IsMemCpy = (II->getIntrinsicID() == Intrinsic::memcpy);
2365 if (isa<ConstantInt>(MTI->getLength()) && IsMemCpy) {
2366 // Small memcpy's are common enough that we want to do them without a call
2368 uint64_t Len = cast<ConstantInt>(MTI->getLength())->getZExtValue();
2369 unsigned Alignment = MTI->getAlignment();
2370 if (IsMemCpySmall(Len, Alignment)) {
2372 if (!ComputeAddress(MTI->getRawDest(), Dest) ||
2373 !ComputeAddress(MTI->getRawSource(), Src))
2375 if (TryEmitSmallMemCpy(Dest, Src, Len, Alignment))
2380 if (!MTI->getLength()->getType()->isIntegerTy(64))
2383 if (MTI->getSourceAddressSpace() > 255 || MTI->getDestAddressSpace() > 255)
2384 // Fast instruction selection doesn't support the special
2388 const char *IntrMemName = isa<MemCpyInst>(II) ? "memcpy" : "memmove";
2389 return LowerCallTo(II, IntrMemName, II->getNumArgOperands() - 2);
2391 case Intrinsic::memset: {
2392 const MemSetInst *MSI = cast<MemSetInst>(II);
2393 // Don't handle volatile.
2394 if (MSI->isVolatile())
2397 if (!MSI->getLength()->getType()->isIntegerTy(64))
2400 if (MSI->getDestAddressSpace() > 255)
2401 // Fast instruction selection doesn't support the special
2405 return LowerCallTo(II, "memset", II->getNumArgOperands() - 2);
2407 case Intrinsic::trap: {
2408 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::BRK))
2412 case Intrinsic::sqrt: {
2413 Type *RetTy = II->getCalledFunction()->getReturnType();
2416 if (!isTypeLegal(RetTy, VT))
2419 unsigned Op0Reg = getRegForValue(II->getOperand(0));
2422 bool Op0IsKill = hasTrivialKill(II->getOperand(0));
2424 unsigned ResultReg = FastEmit_r(VT, VT, ISD::FSQRT, Op0Reg, Op0IsKill);
2428 UpdateValueMap(II, ResultReg);
2431 case Intrinsic::sadd_with_overflow:
2432 case Intrinsic::uadd_with_overflow:
2433 case Intrinsic::ssub_with_overflow:
2434 case Intrinsic::usub_with_overflow:
2435 case Intrinsic::smul_with_overflow:
2436 case Intrinsic::umul_with_overflow: {
2437 // This implements the basic lowering of the xalu with overflow intrinsics.
2438 const Function *Callee = II->getCalledFunction();
2439 auto *Ty = cast<StructType>(Callee->getReturnType());
2440 Type *RetTy = Ty->getTypeAtIndex(0U);
2443 if (!isTypeLegal(RetTy, VT))
2446 if (VT != MVT::i32 && VT != MVT::i64)
2449 const Value *LHS = II->getArgOperand(0);
2450 const Value *RHS = II->getArgOperand(1);
2451 // Canonicalize immediate to the RHS.
2452 if (isa<ConstantInt>(LHS) && !isa<ConstantInt>(RHS) &&
2453 isCommutativeIntrinsic(II))
2454 std::swap(LHS, RHS);
2456 unsigned ResultReg1 = 0, ResultReg2 = 0, MulReg = 0;
2457 AArch64CC::CondCode CC = AArch64CC::Invalid;
2458 switch (II->getIntrinsicID()) {
2459 default: llvm_unreachable("Unexpected intrinsic!");
2460 case Intrinsic::sadd_with_overflow:
2461 ResultReg1 = emitAdds(VT, LHS, RHS); CC = AArch64CC::VS; break;
2462 case Intrinsic::uadd_with_overflow:
2463 ResultReg1 = emitAdds(VT, LHS, RHS); CC = AArch64CC::HS; break;
2464 case Intrinsic::ssub_with_overflow:
2465 ResultReg1 = emitSubs(VT, LHS, RHS); CC = AArch64CC::VS; break;
2466 case Intrinsic::usub_with_overflow:
2467 ResultReg1 = emitSubs(VT, LHS, RHS); CC = AArch64CC::LO; break;
2468 case Intrinsic::smul_with_overflow: {
2470 unsigned LHSReg = getRegForValue(LHS);
2473 bool LHSIsKill = hasTrivialKill(LHS);
2475 unsigned RHSReg = getRegForValue(RHS);
2478 bool RHSIsKill = hasTrivialKill(RHS);
2480 if (VT == MVT::i32) {
2481 MulReg = Emit_SMULL_rr(MVT::i64, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
2482 unsigned ShiftReg = emitLSR_ri(MVT::i64, MVT::i64, MulReg,
2483 /*IsKill=*/false, 32);
2484 MulReg = FastEmitInst_extractsubreg(VT, MulReg, /*IsKill=*/true,
2486 ShiftReg = FastEmitInst_extractsubreg(VT, ShiftReg, /*IsKill=*/true,
2488 emitSubs_rs(VT, ShiftReg, /*IsKill=*/true, MulReg, /*IsKill=*/false,
2489 AArch64_AM::ASR, 31, /*WantResult=*/false);
2491 assert(VT == MVT::i64 && "Unexpected value type.");
2492 MulReg = Emit_MUL_rr(VT, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
2493 unsigned SMULHReg = FastEmit_rr(VT, VT, ISD::MULHS, LHSReg, LHSIsKill,
2495 emitSubs_rs(VT, SMULHReg, /*IsKill=*/true, MulReg, /*IsKill=*/false,
2496 AArch64_AM::ASR, 63, /*WantResult=*/false);
2500 case Intrinsic::umul_with_overflow: {
2502 unsigned LHSReg = getRegForValue(LHS);
2505 bool LHSIsKill = hasTrivialKill(LHS);
2507 unsigned RHSReg = getRegForValue(RHS);
2510 bool RHSIsKill = hasTrivialKill(RHS);
2512 if (VT == MVT::i32) {
2513 MulReg = Emit_UMULL_rr(MVT::i64, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
2514 emitSubs_rs(MVT::i64, AArch64::XZR, /*IsKill=*/true, MulReg,
2515 /*IsKill=*/false, AArch64_AM::LSR, 32,
2516 /*WantResult=*/false);
2517 MulReg = FastEmitInst_extractsubreg(VT, MulReg, /*IsKill=*/true,
2520 assert(VT == MVT::i64 && "Unexpected value type.");
2521 MulReg = Emit_MUL_rr(VT, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
2522 unsigned UMULHReg = FastEmit_rr(VT, VT, ISD::MULHU, LHSReg, LHSIsKill,
2524 emitSubs_rr(VT, AArch64::XZR, /*IsKill=*/true, UMULHReg,
2525 /*IsKill=*/false, /*WantResult=*/false);
2532 ResultReg1 = createResultReg(TLI.getRegClassFor(VT));
2533 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2534 TII.get(TargetOpcode::COPY), ResultReg1).addReg(MulReg);
2537 ResultReg2 = FastEmitInst_rri(AArch64::CSINCWr, &AArch64::GPR32RegClass,
2538 AArch64::WZR, /*IsKill=*/true, AArch64::WZR,
2539 /*IsKill=*/true, getInvertedCondCode(CC));
2540 assert((ResultReg1 + 1) == ResultReg2 &&
2541 "Nonconsecutive result registers.");
2542 UpdateValueMap(II, ResultReg1, 2);
2549 bool AArch64FastISel::SelectRet(const Instruction *I) {
2550 const ReturnInst *Ret = cast<ReturnInst>(I);
2551 const Function &F = *I->getParent()->getParent();
2553 if (!FuncInfo.CanLowerReturn)
2559 // Build a list of return value registers.
2560 SmallVector<unsigned, 4> RetRegs;
2562 if (Ret->getNumOperands() > 0) {
2563 CallingConv::ID CC = F.getCallingConv();
2564 SmallVector<ISD::OutputArg, 4> Outs;
2565 GetReturnInfo(F.getReturnType(), F.getAttributes(), Outs, TLI);
2567 // Analyze operands of the call, assigning locations to each operand.
2568 SmallVector<CCValAssign, 16> ValLocs;
2569 CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, ValLocs, I->getContext());
2570 CCAssignFn *RetCC = CC == CallingConv::WebKit_JS ? RetCC_AArch64_WebKit_JS
2571 : RetCC_AArch64_AAPCS;
2572 CCInfo.AnalyzeReturn(Outs, RetCC);
2574 // Only handle a single return value for now.
2575 if (ValLocs.size() != 1)
2578 CCValAssign &VA = ValLocs[0];
2579 const Value *RV = Ret->getOperand(0);
2581 // Don't bother handling odd stuff for now.
2582 if (VA.getLocInfo() != CCValAssign::Full)
2584 // Only handle register returns for now.
2587 unsigned Reg = getRegForValue(RV);
2591 unsigned SrcReg = Reg + VA.getValNo();
2592 unsigned DestReg = VA.getLocReg();
2593 // Avoid a cross-class copy. This is very unlikely.
2594 if (!MRI.getRegClass(SrcReg)->contains(DestReg))
2597 EVT RVEVT = TLI.getValueType(RV->getType());
2598 if (!RVEVT.isSimple())
2601 // Vectors (of > 1 lane) in big endian need tricky handling.
2602 if (RVEVT.isVector() && RVEVT.getVectorNumElements() > 1)
2605 MVT RVVT = RVEVT.getSimpleVT();
2606 if (RVVT == MVT::f128)
2608 MVT DestVT = VA.getValVT();
2609 // Special handling for extended integers.
2610 if (RVVT != DestVT) {
2611 if (RVVT != MVT::i1 && RVVT != MVT::i8 && RVVT != MVT::i16)
2614 if (!Outs[0].Flags.isZExt() && !Outs[0].Flags.isSExt())
2617 bool isZExt = Outs[0].Flags.isZExt();
2618 SrcReg = EmitIntExt(RVVT, SrcReg, DestVT, isZExt);
2624 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2625 TII.get(TargetOpcode::COPY), DestReg).addReg(SrcReg);
2627 // Add register to return instruction.
2628 RetRegs.push_back(VA.getLocReg());
2631 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2632 TII.get(AArch64::RET_ReallyLR));
2633 for (unsigned i = 0, e = RetRegs.size(); i != e; ++i)
2634 MIB.addReg(RetRegs[i], RegState::Implicit);
2638 bool AArch64FastISel::SelectTrunc(const Instruction *I) {
2639 Type *DestTy = I->getType();
2640 Value *Op = I->getOperand(0);
2641 Type *SrcTy = Op->getType();
2643 EVT SrcEVT = TLI.getValueType(SrcTy, true);
2644 EVT DestEVT = TLI.getValueType(DestTy, true);
2645 if (!SrcEVT.isSimple())
2647 if (!DestEVT.isSimple())
2650 MVT SrcVT = SrcEVT.getSimpleVT();
2651 MVT DestVT = DestEVT.getSimpleVT();
2653 if (SrcVT != MVT::i64 && SrcVT != MVT::i32 && SrcVT != MVT::i16 &&
2656 if (DestVT != MVT::i32 && DestVT != MVT::i16 && DestVT != MVT::i8 &&
2660 unsigned SrcReg = getRegForValue(Op);
2663 bool SrcIsKill = hasTrivialKill(Op);
2665 // If we're truncating from i64 to a smaller non-legal type then generate an
2666 // AND. Otherwise, we know the high bits are undefined and a truncate doesn't
2667 // generate any code.
2668 if (SrcVT == MVT::i64) {
2670 switch (DestVT.SimpleTy) {
2672 // Trunc i64 to i32 is handled by the target-independent fast-isel.
2684 // Issue an extract_subreg to get the lower 32-bits.
2685 unsigned Reg32 = FastEmitInst_extractsubreg(MVT::i32, SrcReg, SrcIsKill,
2687 // Create the AND instruction which performs the actual truncation.
2688 unsigned ANDReg = emitAND_ri(MVT::i32, Reg32, /*IsKill=*/true, Mask);
2689 assert(ANDReg && "Unexpected AND instruction emission failure.");
2693 UpdateValueMap(I, SrcReg);
2697 unsigned AArch64FastISel::Emiti1Ext(unsigned SrcReg, MVT DestVT, bool isZExt) {
2698 assert((DestVT == MVT::i8 || DestVT == MVT::i16 || DestVT == MVT::i32 ||
2699 DestVT == MVT::i64) &&
2700 "Unexpected value type.");
2701 // Handle i8 and i16 as i32.
2702 if (DestVT == MVT::i8 || DestVT == MVT::i16)
2706 unsigned ResultReg = emitAND_ri(MVT::i32, SrcReg, /*TODO:IsKill=*/false, 1);
2707 assert(ResultReg && "Unexpected AND instruction emission failure.");
2708 if (DestVT == MVT::i64) {
2709 // We're ZExt i1 to i64. The ANDWri Wd, Ws, #1 implicitly clears the
2710 // upper 32 bits. Emit a SUBREG_TO_REG to extend from Wd to Xd.
2711 unsigned Reg64 = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
2712 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2713 TII.get(AArch64::SUBREG_TO_REG), Reg64)
2716 .addImm(AArch64::sub_32);
2721 if (DestVT == MVT::i64) {
2722 // FIXME: We're SExt i1 to i64.
2725 return FastEmitInst_rii(AArch64::SBFMWri, &AArch64::GPR32RegClass, SrcReg,
2726 /*TODO:IsKill=*/false, 0, 0);
2730 unsigned AArch64FastISel::Emit_MUL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
2731 unsigned Op1, bool Op1IsKill) {
2733 switch (RetVT.SimpleTy) {
2739 Opc = AArch64::MADDWrrr; ZReg = AArch64::WZR; break;
2741 Opc = AArch64::MADDXrrr; ZReg = AArch64::XZR; break;
2744 const TargetRegisterClass *RC =
2745 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
2746 return FastEmitInst_rrr(Opc, RC, Op0, Op0IsKill, Op1, Op1IsKill,
2747 /*IsKill=*/ZReg, true);
2750 unsigned AArch64FastISel::Emit_SMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
2751 unsigned Op1, bool Op1IsKill) {
2752 if (RetVT != MVT::i64)
2755 return FastEmitInst_rrr(AArch64::SMADDLrrr, &AArch64::GPR64RegClass,
2756 Op0, Op0IsKill, Op1, Op1IsKill,
2757 AArch64::XZR, /*IsKill=*/true);
2760 unsigned AArch64FastISel::Emit_UMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
2761 unsigned Op1, bool Op1IsKill) {
2762 if (RetVT != MVT::i64)
2765 return FastEmitInst_rrr(AArch64::UMADDLrrr, &AArch64::GPR64RegClass,
2766 Op0, Op0IsKill, Op1, Op1IsKill,
2767 AArch64::XZR, /*IsKill=*/true);
2770 unsigned AArch64FastISel::emitLSL_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
2771 unsigned Op1Reg, bool Op1IsKill) {
2773 bool NeedTrunc = false;
2775 switch (RetVT.SimpleTy) {
2777 case MVT::i8: Opc = AArch64::LSLVWr; NeedTrunc = true; Mask = 0xff; break;
2778 case MVT::i16: Opc = AArch64::LSLVWr; NeedTrunc = true; Mask = 0xffff; break;
2779 case MVT::i32: Opc = AArch64::LSLVWr; break;
2780 case MVT::i64: Opc = AArch64::LSLVXr; break;
2783 const TargetRegisterClass *RC =
2784 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
2786 Op1Reg = emitAND_ri(MVT::i32, Op1Reg, Op1IsKill, Mask);
2789 unsigned ResultReg = FastEmitInst_rr(Opc, RC, Op0Reg, Op0IsKill, Op1Reg,
2792 ResultReg = emitAND_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
2796 unsigned AArch64FastISel::emitLSL_ri(MVT RetVT, MVT SrcVT, unsigned Op0,
2797 bool Op0IsKill, uint64_t Shift,
2799 assert(RetVT.SimpleTy >= SrcVT.SimpleTy &&
2800 "Unexpected source/return type pair.");
2801 assert((SrcVT == MVT::i8 || SrcVT == MVT::i16 || SrcVT == MVT::i32 ||
2802 SrcVT == MVT::i64) && "Unexpected source value type.");
2803 assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 ||
2804 RetVT == MVT::i64) && "Unexpected return value type.");
2806 bool Is64Bit = (RetVT == MVT::i64);
2807 unsigned RegSize = Is64Bit ? 64 : 32;
2808 unsigned DstBits = RetVT.getSizeInBits();
2809 unsigned SrcBits = SrcVT.getSizeInBits();
2811 // Don't deal with undefined shifts.
2812 if (Shift >= DstBits)
2815 // For immediate shifts we can fold the zero-/sign-extension into the shift.
2816 // {S|U}BFM Wd, Wn, #r, #s
2817 // Wd<32+s-r,32-r> = Wn<s:0> when r > s
2819 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
2820 // %2 = shl i16 %1, 4
2821 // Wd<32+7-28,32-28> = Wn<7:0> <- clamp s to 7
2822 // 0b1111_1111_1111_1111__1111_1010_1010_0000 sext
2823 // 0b0000_0000_0000_0000__0000_0101_0101_0000 sext | zext
2824 // 0b0000_0000_0000_0000__0000_1010_1010_0000 zext
2826 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
2827 // %2 = shl i16 %1, 8
2828 // Wd<32+7-24,32-24> = Wn<7:0>
2829 // 0b1111_1111_1111_1111__1010_1010_0000_0000 sext
2830 // 0b0000_0000_0000_0000__0101_0101_0000_0000 sext | zext
2831 // 0b0000_0000_0000_0000__1010_1010_0000_0000 zext
2833 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
2834 // %2 = shl i16 %1, 12
2835 // Wd<32+3-20,32-20> = Wn<3:0>
2836 // 0b1111_1111_1111_1111__1010_0000_0000_0000 sext
2837 // 0b0000_0000_0000_0000__0101_0000_0000_0000 sext | zext
2838 // 0b0000_0000_0000_0000__1010_0000_0000_0000 zext
2840 unsigned ImmR = RegSize - Shift;
2841 // Limit the width to the length of the source type.
2842 unsigned ImmS = std::min<unsigned>(SrcBits - 1, DstBits - 1 - Shift);
2843 static const unsigned OpcTable[2][2] = {
2844 {AArch64::SBFMWri, AArch64::SBFMXri},
2845 {AArch64::UBFMWri, AArch64::UBFMXri}
2847 unsigned Opc = OpcTable[IsZext][Is64Bit];
2848 const TargetRegisterClass *RC =
2849 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
2850 if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) {
2851 unsigned TmpReg = MRI.createVirtualRegister(RC);
2852 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2853 TII.get(AArch64::SUBREG_TO_REG), TmpReg)
2855 .addReg(Op0, getKillRegState(Op0IsKill))
2856 .addImm(AArch64::sub_32);
2860 return FastEmitInst_rii(Opc, RC, Op0, Op0IsKill, ImmR, ImmS);
2863 unsigned AArch64FastISel::emitLSR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
2864 unsigned Op1Reg, bool Op1IsKill) {
2866 bool NeedTrunc = false;
2868 switch (RetVT.SimpleTy) {
2870 case MVT::i8: Opc = AArch64::LSRVWr; NeedTrunc = true; Mask = 0xff; break;
2871 case MVT::i16: Opc = AArch64::LSRVWr; NeedTrunc = true; Mask = 0xffff; break;
2872 case MVT::i32: Opc = AArch64::LSRVWr; break;
2873 case MVT::i64: Opc = AArch64::LSRVXr; break;
2876 const TargetRegisterClass *RC =
2877 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
2879 Op0Reg = emitAND_ri(MVT::i32, Op0Reg, Op0IsKill, Mask);
2880 Op1Reg = emitAND_ri(MVT::i32, Op1Reg, Op1IsKill, Mask);
2881 Op0IsKill = Op1IsKill = true;
2883 unsigned ResultReg = FastEmitInst_rr(Opc, RC, Op0Reg, Op0IsKill, Op1Reg,
2886 ResultReg = emitAND_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
2890 unsigned AArch64FastISel::emitLSR_ri(MVT RetVT, MVT SrcVT, unsigned Op0,
2891 bool Op0IsKill, uint64_t Shift,
2893 assert(RetVT.SimpleTy >= SrcVT.SimpleTy &&
2894 "Unexpected source/return type pair.");
2895 assert((SrcVT == MVT::i8 || SrcVT == MVT::i16 || SrcVT == MVT::i32 ||
2896 SrcVT == MVT::i64) && "Unexpected source value type.");
2897 assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 ||
2898 RetVT == MVT::i64) && "Unexpected return value type.");
2900 bool Is64Bit = (RetVT == MVT::i64);
2901 unsigned RegSize = Is64Bit ? 64 : 32;
2902 unsigned DstBits = RetVT.getSizeInBits();
2903 unsigned SrcBits = SrcVT.getSizeInBits();
2905 // Don't deal with undefined shifts.
2906 if (Shift >= DstBits)
2909 // For immediate shifts we can fold the zero-/sign-extension into the shift.
2910 // {S|U}BFM Wd, Wn, #r, #s
2911 // Wd<s-r:0> = Wn<s:r> when r <= s
2913 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
2914 // %2 = lshr i16 %1, 4
2915 // Wd<7-4:0> = Wn<7:4>
2916 // 0b0000_0000_0000_0000__0000_1111_1111_1010 sext
2917 // 0b0000_0000_0000_0000__0000_0000_0000_0101 sext | zext
2918 // 0b0000_0000_0000_0000__0000_0000_0000_1010 zext
2920 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
2921 // %2 = lshr i16 %1, 8
2922 // Wd<7-7,0> = Wn<7:7>
2923 // 0b0000_0000_0000_0000__0000_0000_1111_1111 sext
2924 // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
2925 // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
2927 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
2928 // %2 = lshr i16 %1, 12
2929 // Wd<7-7,0> = Wn<7:7> <- clamp r to 7
2930 // 0b0000_0000_0000_0000__0000_0000_0000_1111 sext
2931 // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
2932 // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
2934 if (Shift >= SrcBits && IsZExt)
2935 return AArch64MaterializeInt(ConstantInt::get(*Context, APInt(RegSize, 0)),
2938 // It is not possible to fold a sign-extend into the LShr instruction. In this
2939 // case emit a sign-extend.
2941 Op0 = EmitIntExt(SrcVT, Op0, RetVT, IsZExt);
2946 SrcBits = SrcVT.getSizeInBits();
2950 unsigned ImmR = std::min<unsigned>(SrcBits - 1, Shift);
2951 unsigned ImmS = SrcBits - 1;
2952 static const unsigned OpcTable[2][2] = {
2953 {AArch64::SBFMWri, AArch64::SBFMXri},
2954 {AArch64::UBFMWri, AArch64::UBFMXri}
2956 unsigned Opc = OpcTable[IsZExt][Is64Bit];
2957 const TargetRegisterClass *RC =
2958 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
2959 if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) {
2960 unsigned TmpReg = MRI.createVirtualRegister(RC);
2961 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2962 TII.get(AArch64::SUBREG_TO_REG), TmpReg)
2964 .addReg(Op0, getKillRegState(Op0IsKill))
2965 .addImm(AArch64::sub_32);
2969 return FastEmitInst_rii(Opc, RC, Op0, Op0IsKill, ImmR, ImmS);
2972 unsigned AArch64FastISel::emitASR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
2973 unsigned Op1Reg, bool Op1IsKill) {
2975 bool NeedTrunc = false;
2977 switch (RetVT.SimpleTy) {
2979 case MVT::i8: Opc = AArch64::ASRVWr; NeedTrunc = true; Mask = 0xff; break;
2980 case MVT::i16: Opc = AArch64::ASRVWr; NeedTrunc = true; Mask = 0xffff; break;
2981 case MVT::i32: Opc = AArch64::ASRVWr; break;
2982 case MVT::i64: Opc = AArch64::ASRVXr; break;
2985 const TargetRegisterClass *RC =
2986 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
2988 Op0Reg = EmitIntExt(RetVT, Op0Reg, MVT::i32, /*IsZExt=*/false);
2989 Op1Reg = emitAND_ri(MVT::i32, Op1Reg, Op1IsKill, Mask);
2990 Op0IsKill = Op1IsKill = true;
2992 unsigned ResultReg = FastEmitInst_rr(Opc, RC, Op0Reg, Op0IsKill, Op1Reg,
2995 ResultReg = emitAND_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
2999 unsigned AArch64FastISel::emitASR_ri(MVT RetVT, MVT SrcVT, unsigned Op0,
3000 bool Op0IsKill, uint64_t Shift,
3002 assert(RetVT.SimpleTy >= SrcVT.SimpleTy &&
3003 "Unexpected source/return type pair.");
3004 assert((SrcVT == MVT::i8 || SrcVT == MVT::i16 || SrcVT == MVT::i32 ||
3005 SrcVT == MVT::i64) && "Unexpected source value type.");
3006 assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 ||
3007 RetVT == MVT::i64) && "Unexpected return value type.");
3009 bool Is64Bit = (RetVT == MVT::i64);
3010 unsigned RegSize = Is64Bit ? 64 : 32;
3011 unsigned DstBits = RetVT.getSizeInBits();
3012 unsigned SrcBits = SrcVT.getSizeInBits();
3014 // Don't deal with undefined shifts.
3015 if (Shift >= DstBits)
3018 // For immediate shifts we can fold the zero-/sign-extension into the shift.
3019 // {S|U}BFM Wd, Wn, #r, #s
3020 // Wd<s-r:0> = Wn<s:r> when r <= s
3022 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3023 // %2 = ashr i16 %1, 4
3024 // Wd<7-4:0> = Wn<7:4>
3025 // 0b1111_1111_1111_1111__1111_1111_1111_1010 sext
3026 // 0b0000_0000_0000_0000__0000_0000_0000_0101 sext | zext
3027 // 0b0000_0000_0000_0000__0000_0000_0000_1010 zext
3029 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3030 // %2 = ashr i16 %1, 8
3031 // Wd<7-7,0> = Wn<7:7>
3032 // 0b1111_1111_1111_1111__1111_1111_1111_1111 sext
3033 // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
3034 // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
3036 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3037 // %2 = ashr i16 %1, 12
3038 // Wd<7-7,0> = Wn<7:7> <- clamp r to 7
3039 // 0b1111_1111_1111_1111__1111_1111_1111_1111 sext
3040 // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
3041 // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
3043 if (Shift >= SrcBits && IsZExt)
3044 return AArch64MaterializeInt(ConstantInt::get(*Context, APInt(RegSize, 0)),
3047 unsigned ImmR = std::min<unsigned>(SrcBits - 1, Shift);
3048 unsigned ImmS = SrcBits - 1;
3049 static const unsigned OpcTable[2][2] = {
3050 {AArch64::SBFMWri, AArch64::SBFMXri},
3051 {AArch64::UBFMWri, AArch64::UBFMXri}
3053 unsigned Opc = OpcTable[IsZExt][Is64Bit];
3054 const TargetRegisterClass *RC =
3055 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3056 if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) {
3057 unsigned TmpReg = MRI.createVirtualRegister(RC);
3058 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3059 TII.get(AArch64::SUBREG_TO_REG), TmpReg)
3061 .addReg(Op0, getKillRegState(Op0IsKill))
3062 .addImm(AArch64::sub_32);
3066 return FastEmitInst_rii(Opc, RC, Op0, Op0IsKill, ImmR, ImmS);
3069 unsigned AArch64FastISel::EmitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT,
3071 assert(DestVT != MVT::i1 && "ZeroExt/SignExt an i1?");
3073 // FastISel does not have plumbing to deal with extensions where the SrcVT or
3074 // DestVT are odd things, so test to make sure that they are both types we can
3075 // handle (i1/i8/i16/i32 for SrcVT and i8/i16/i32/i64 for DestVT), otherwise
3076 // bail out to SelectionDAG.
3077 if (((DestVT != MVT::i8) && (DestVT != MVT::i16) &&
3078 (DestVT != MVT::i32) && (DestVT != MVT::i64)) ||
3079 ((SrcVT != MVT::i1) && (SrcVT != MVT::i8) &&
3080 (SrcVT != MVT::i16) && (SrcVT != MVT::i32)))
3086 switch (SrcVT.SimpleTy) {
3090 return Emiti1Ext(SrcReg, DestVT, isZExt);
3092 if (DestVT == MVT::i64)
3093 Opc = isZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
3095 Opc = isZExt ? AArch64::UBFMWri : AArch64::SBFMWri;
3099 if (DestVT == MVT::i64)
3100 Opc = isZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
3102 Opc = isZExt ? AArch64::UBFMWri : AArch64::SBFMWri;
3106 assert(DestVT == MVT::i64 && "IntExt i32 to i32?!?");
3107 Opc = isZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
3112 // Handle i8 and i16 as i32.
3113 if (DestVT == MVT::i8 || DestVT == MVT::i16)
3115 else if (DestVT == MVT::i64) {
3116 unsigned Src64 = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
3117 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3118 TII.get(AArch64::SUBREG_TO_REG), Src64)
3121 .addImm(AArch64::sub_32);
3125 const TargetRegisterClass *RC =
3126 (DestVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3127 return FastEmitInst_rii(Opc, RC, SrcReg, /*TODO:IsKill=*/false, 0, Imm);
3130 bool AArch64FastISel::SelectIntExt(const Instruction *I) {
3131 // On ARM, in general, integer casts don't involve legal types; this code
3132 // handles promotable integers. The high bits for a type smaller than
3133 // the register size are assumed to be undefined.
3134 Type *DestTy = I->getType();
3135 Value *Src = I->getOperand(0);
3136 Type *SrcTy = Src->getType();
3138 bool isZExt = isa<ZExtInst>(I);
3139 unsigned SrcReg = getRegForValue(Src);
3143 EVT SrcEVT = TLI.getValueType(SrcTy, true);
3144 EVT DestEVT = TLI.getValueType(DestTy, true);
3145 if (!SrcEVT.isSimple())
3147 if (!DestEVT.isSimple())
3150 MVT SrcVT = SrcEVT.getSimpleVT();
3151 MVT DestVT = DestEVT.getSimpleVT();
3152 unsigned ResultReg = 0;
3154 // Check if it is an argument and if it is already zero/sign-extended.
3155 if (const auto *Arg = dyn_cast<Argument>(Src)) {
3156 if ((isZExt && Arg->hasZExtAttr()) || (!isZExt && Arg->hasSExtAttr())) {
3157 if (DestVT == MVT::i64) {
3158 ResultReg = createResultReg(TLI.getRegClassFor(DestVT));
3159 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3160 TII.get(AArch64::SUBREG_TO_REG), ResultReg)
3163 .addImm(AArch64::sub_32);
3170 ResultReg = EmitIntExt(SrcVT, SrcReg, DestVT, isZExt);
3175 UpdateValueMap(I, ResultReg);
3179 bool AArch64FastISel::SelectRem(const Instruction *I, unsigned ISDOpcode) {
3180 EVT DestEVT = TLI.getValueType(I->getType(), true);
3181 if (!DestEVT.isSimple())
3184 MVT DestVT = DestEVT.getSimpleVT();
3185 if (DestVT != MVT::i64 && DestVT != MVT::i32)
3189 bool is64bit = (DestVT == MVT::i64);
3190 switch (ISDOpcode) {
3194 DivOpc = is64bit ? AArch64::SDIVXr : AArch64::SDIVWr;
3197 DivOpc = is64bit ? AArch64::UDIVXr : AArch64::UDIVWr;
3200 unsigned MSubOpc = is64bit ? AArch64::MSUBXrrr : AArch64::MSUBWrrr;
3201 unsigned Src0Reg = getRegForValue(I->getOperand(0));
3204 bool Src0IsKill = hasTrivialKill(I->getOperand(0));
3206 unsigned Src1Reg = getRegForValue(I->getOperand(1));
3209 bool Src1IsKill = hasTrivialKill(I->getOperand(1));
3211 const TargetRegisterClass *RC =
3212 (DestVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3213 unsigned QuotReg = FastEmitInst_rr(DivOpc, RC, Src0Reg, /*IsKill=*/false,
3214 Src1Reg, /*IsKill=*/false);
3215 assert(QuotReg && "Unexpected DIV instruction emission failure.");
3216 // The remainder is computed as numerator - (quotient * denominator) using the
3217 // MSUB instruction.
3218 unsigned ResultReg = FastEmitInst_rrr(MSubOpc, RC, QuotReg, /*IsKill=*/true,
3219 Src1Reg, Src1IsKill, Src0Reg,
3221 UpdateValueMap(I, ResultReg);
3225 bool AArch64FastISel::SelectMul(const Instruction *I) {
3226 EVT SrcEVT = TLI.getValueType(I->getOperand(0)->getType(), true);
3227 if (!SrcEVT.isSimple())
3229 MVT SrcVT = SrcEVT.getSimpleVT();
3231 // Must be simple value type. Don't handle vectors.
3232 if (SrcVT != MVT::i64 && SrcVT != MVT::i32 && SrcVT != MVT::i16 &&
3236 unsigned Src0Reg = getRegForValue(I->getOperand(0));
3239 bool Src0IsKill = hasTrivialKill(I->getOperand(0));
3241 unsigned Src1Reg = getRegForValue(I->getOperand(1));
3244 bool Src1IsKill = hasTrivialKill(I->getOperand(1));
3246 unsigned ResultReg =
3247 Emit_MUL_rr(SrcVT, Src0Reg, Src0IsKill, Src1Reg, Src1IsKill);
3252 UpdateValueMap(I, ResultReg);
3256 bool AArch64FastISel::SelectShift(const Instruction *I) {
3258 if (!isLoadStoreTypeLegal(I->getType(), RetVT))
3261 if (const auto *C = dyn_cast<ConstantInt>(I->getOperand(1))) {
3262 unsigned ResultReg = 0;
3263 uint64_t ShiftVal = C->getZExtValue();
3265 bool IsZExt = (I->getOpcode() == Instruction::AShr) ? false : true;
3266 const Value * Op0 = I->getOperand(0);
3267 if (const auto *ZExt = dyn_cast<ZExtInst>(Op0)) {
3269 if (isLoadStoreTypeLegal(ZExt->getSrcTy(), TmpVT)) {
3272 Op0 = ZExt->getOperand(0);
3274 } else if (const auto *SExt = dyn_cast<SExtInst>(Op0)) {
3276 if (isLoadStoreTypeLegal(SExt->getSrcTy(), TmpVT)) {
3279 Op0 = SExt->getOperand(0);
3283 unsigned Op0Reg = getRegForValue(Op0);
3286 bool Op0IsKill = hasTrivialKill(Op0);
3288 switch (I->getOpcode()) {
3289 default: llvm_unreachable("Unexpected instruction.");
3290 case Instruction::Shl:
3291 ResultReg = emitLSL_ri(RetVT, SrcVT, Op0Reg, Op0IsKill, ShiftVal, IsZExt);
3293 case Instruction::AShr:
3294 ResultReg = emitASR_ri(RetVT, SrcVT, Op0Reg, Op0IsKill, ShiftVal, IsZExt);
3296 case Instruction::LShr:
3297 ResultReg = emitLSR_ri(RetVT, SrcVT, Op0Reg, Op0IsKill, ShiftVal, IsZExt);
3303 UpdateValueMap(I, ResultReg);
3307 unsigned Op0Reg = getRegForValue(I->getOperand(0));
3310 bool Op0IsKill = hasTrivialKill(I->getOperand(0));
3312 unsigned Op1Reg = getRegForValue(I->getOperand(1));
3315 bool Op1IsKill = hasTrivialKill(I->getOperand(1));
3317 unsigned ResultReg = 0;
3318 switch (I->getOpcode()) {
3319 default: llvm_unreachable("Unexpected instruction.");
3320 case Instruction::Shl:
3321 ResultReg = emitLSL_rr(RetVT, Op0Reg, Op0IsKill, Op1Reg, Op1IsKill);
3323 case Instruction::AShr:
3324 ResultReg = emitASR_rr(RetVT, Op0Reg, Op0IsKill, Op1Reg, Op1IsKill);
3326 case Instruction::LShr:
3327 ResultReg = emitLSR_rr(RetVT, Op0Reg, Op0IsKill, Op1Reg, Op1IsKill);
3334 UpdateValueMap(I, ResultReg);
3338 bool AArch64FastISel::SelectBitCast(const Instruction *I) {
3341 if (!isTypeLegal(I->getOperand(0)->getType(), SrcVT))
3343 if (!isTypeLegal(I->getType(), RetVT))
3347 if (RetVT == MVT::f32 && SrcVT == MVT::i32)
3348 Opc = AArch64::FMOVWSr;
3349 else if (RetVT == MVT::f64 && SrcVT == MVT::i64)
3350 Opc = AArch64::FMOVXDr;
3351 else if (RetVT == MVT::i32 && SrcVT == MVT::f32)
3352 Opc = AArch64::FMOVSWr;
3353 else if (RetVT == MVT::i64 && SrcVT == MVT::f64)
3354 Opc = AArch64::FMOVDXr;
3358 const TargetRegisterClass *RC = nullptr;
3359 switch (RetVT.SimpleTy) {
3360 default: llvm_unreachable("Unexpected value type.");
3361 case MVT::i32: RC = &AArch64::GPR32RegClass; break;
3362 case MVT::i64: RC = &AArch64::GPR64RegClass; break;
3363 case MVT::f32: RC = &AArch64::FPR32RegClass; break;
3364 case MVT::f64: RC = &AArch64::FPR64RegClass; break;
3366 unsigned Op0Reg = getRegForValue(I->getOperand(0));
3369 bool Op0IsKill = hasTrivialKill(I->getOperand(0));
3370 unsigned ResultReg = FastEmitInst_r(Opc, RC, Op0Reg, Op0IsKill);
3375 UpdateValueMap(I, ResultReg);
3379 bool AArch64FastISel::TargetSelectInstruction(const Instruction *I) {
3380 switch (I->getOpcode()) {
3383 case Instruction::Load:
3384 return SelectLoad(I);
3385 case Instruction::Store:
3386 return SelectStore(I);
3387 case Instruction::Br:
3388 return SelectBranch(I);
3389 case Instruction::IndirectBr:
3390 return SelectIndirectBr(I);
3391 case Instruction::FCmp:
3392 case Instruction::ICmp:
3393 return SelectCmp(I);
3394 case Instruction::Select:
3395 return SelectSelect(I);
3396 case Instruction::FPExt:
3397 return SelectFPExt(I);
3398 case Instruction::FPTrunc:
3399 return SelectFPTrunc(I);
3400 case Instruction::FPToSI:
3401 return SelectFPToInt(I, /*Signed=*/true);
3402 case Instruction::FPToUI:
3403 return SelectFPToInt(I, /*Signed=*/false);
3404 case Instruction::SIToFP:
3405 return SelectIntToFP(I, /*Signed=*/true);
3406 case Instruction::UIToFP:
3407 return SelectIntToFP(I, /*Signed=*/false);
3408 case Instruction::SRem:
3409 return SelectRem(I, ISD::SREM);
3410 case Instruction::URem:
3411 return SelectRem(I, ISD::UREM);
3412 case Instruction::Ret:
3413 return SelectRet(I);
3414 case Instruction::Trunc:
3415 return SelectTrunc(I);
3416 case Instruction::ZExt:
3417 case Instruction::SExt:
3418 return SelectIntExt(I);
3420 // FIXME: All of these should really be handled by the target-independent
3421 // selector -> improve FastISel tblgen.
3422 case Instruction::Mul:
3423 return SelectMul(I);
3424 case Instruction::Shl: // fall-through
3425 case Instruction::LShr: // fall-through
3426 case Instruction::AShr:
3427 return SelectShift(I);
3428 case Instruction::BitCast:
3429 return SelectBitCast(I);
3432 // Silence warnings.
3433 (void)&CC_AArch64_DarwinPCS_VarArg;
3437 llvm::FastISel *AArch64::createFastISel(FunctionLoweringInfo &funcInfo,
3438 const TargetLibraryInfo *libInfo) {
3439 return new AArch64FastISel(funcInfo, libInfo);