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 selectAddSub(const Instruction *I);
117 bool selectLogicalOp(const Instruction *I, unsigned ISDOpcode);
118 bool SelectLoad(const Instruction *I);
119 bool SelectStore(const Instruction *I);
120 bool SelectBranch(const Instruction *I);
121 bool SelectIndirectBr(const Instruction *I);
122 bool SelectCmp(const Instruction *I);
123 bool SelectSelect(const Instruction *I);
124 bool SelectFPExt(const Instruction *I);
125 bool SelectFPTrunc(const Instruction *I);
126 bool SelectFPToInt(const Instruction *I, bool Signed);
127 bool SelectIntToFP(const Instruction *I, bool Signed);
128 bool SelectRem(const Instruction *I, unsigned ISDOpcode);
129 bool SelectRet(const Instruction *I);
130 bool SelectTrunc(const Instruction *I);
131 bool SelectIntExt(const Instruction *I);
132 bool SelectMul(const Instruction *I);
133 bool SelectShift(const Instruction *I);
134 bool SelectBitCast(const Instruction *I);
136 // Utility helper routines.
137 bool isTypeLegal(Type *Ty, MVT &VT);
138 bool isLoadStoreTypeLegal(Type *Ty, MVT &VT);
139 bool isTypeSupported(Type *Ty, MVT &VT);
140 bool isValueAvailable(const Value *V) const;
141 bool ComputeAddress(const Value *Obj, Address &Addr, Type *Ty = nullptr);
142 bool ComputeCallAddress(const Value *V, Address &Addr);
143 bool SimplifyAddress(Address &Addr, MVT VT);
144 void AddLoadStoreOperands(Address &Addr, const MachineInstrBuilder &MIB,
145 unsigned Flags, unsigned ScaleFactor,
146 MachineMemOperand *MMO);
147 bool IsMemCpySmall(uint64_t Len, unsigned Alignment);
148 bool TryEmitSmallMemCpy(Address Dest, Address Src, uint64_t Len,
150 bool foldXALUIntrinsic(AArch64CC::CondCode &CC, const Instruction *I,
153 // Emit helper routines.
154 unsigned emitAddSub(bool UseAdd, MVT RetVT, const Value *LHS,
155 const Value *RHS, bool SetFlags = false,
156 bool WantResult = true, bool IsZExt = false);
157 unsigned emitAddSub_rr(bool UseAdd, MVT RetVT, unsigned LHSReg,
158 bool LHSIsKill, unsigned RHSReg, bool RHSIsKill,
159 bool SetFlags = false, bool WantResult = true);
160 unsigned emitAddSub_ri(bool UseAdd, MVT RetVT, unsigned LHSReg,
161 bool LHSIsKill, uint64_t Imm, bool SetFlags = false,
162 bool WantResult = true);
163 unsigned emitAddSub_rs(bool UseAdd, MVT RetVT, unsigned LHSReg,
164 bool LHSIsKill, unsigned RHSReg, bool RHSIsKill,
165 AArch64_AM::ShiftExtendType ShiftType,
166 uint64_t ShiftImm, bool SetFlags = false,
167 bool WantResult = true);
168 unsigned emitAddSub_rx(bool UseAdd, MVT RetVT, unsigned LHSReg,
169 bool LHSIsKill, unsigned RHSReg, bool RHSIsKill,
170 AArch64_AM::ShiftExtendType ExtType,
171 uint64_t ShiftImm, bool SetFlags = false,
172 bool WantResult = true);
175 bool emitCmp(const Value *LHS, const Value *RHS, bool IsZExt);
176 bool emitICmp(MVT RetVT, const Value *LHS, const Value *RHS, bool IsZExt);
177 bool emitICmp_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill, uint64_t Imm);
178 bool emitFCmp(MVT RetVT, const Value *LHS, const Value *RHS);
179 bool EmitLoad(MVT VT, unsigned &ResultReg, Address Addr,
180 MachineMemOperand *MMO = nullptr);
181 bool EmitStore(MVT VT, unsigned SrcReg, Address Addr,
182 MachineMemOperand *MMO = nullptr);
183 unsigned EmitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, bool isZExt);
184 unsigned Emiti1Ext(unsigned SrcReg, MVT DestVT, bool isZExt);
185 unsigned emitAdd(MVT RetVT, const Value *LHS, const Value *RHS,
186 bool SetFlags = false, bool WantResult = true,
187 bool IsZExt = false);
188 unsigned emitSub(MVT RetVT, const Value *LHS, const Value *RHS,
189 bool SetFlags = false, bool WantResult = true,
190 bool IsZExt = false);
191 unsigned emitSubs_rr(MVT RetVT, unsigned LHSReg, bool LHSIsKill,
192 unsigned RHSReg, bool RHSIsKill, bool WantResult = true);
193 unsigned emitSubs_rs(MVT RetVT, unsigned LHSReg, bool LHSIsKill,
194 unsigned RHSReg, bool RHSIsKill,
195 AArch64_AM::ShiftExtendType ShiftType, uint64_t ShiftImm,
196 bool WantResult = true);
197 unsigned emitLogicalOp(unsigned ISDOpc, MVT RetVT, const Value *LHS,
199 unsigned emitLogicalOp_ri(unsigned ISDOpc, MVT RetVT, unsigned LHSReg,
200 bool LHSIsKill, uint64_t Imm);
201 unsigned emitLogicalOp_rs(unsigned ISDOpc, MVT RetVT, unsigned LHSReg,
202 bool LHSIsKill, unsigned RHSReg, bool RHSIsKill,
204 unsigned emitAnd_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill, uint64_t Imm);
205 unsigned Emit_MUL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
206 unsigned Op1, bool Op1IsKill);
207 unsigned Emit_SMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
208 unsigned Op1, bool Op1IsKill);
209 unsigned Emit_UMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
210 unsigned Op1, bool Op1IsKill);
211 unsigned emitLSL_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
212 unsigned Op1Reg, bool Op1IsKill);
213 unsigned emitLSL_ri(MVT RetVT, MVT SrcVT, unsigned Op0Reg, bool Op0IsKill,
214 uint64_t Imm, bool IsZExt = true);
215 unsigned emitLSR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
216 unsigned Op1Reg, bool Op1IsKill);
217 unsigned emitLSR_ri(MVT RetVT, MVT SrcVT, unsigned Op0Reg, bool Op0IsKill,
218 uint64_t Imm, bool IsZExt = true);
219 unsigned emitASR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
220 unsigned Op1Reg, bool Op1IsKill);
221 unsigned emitASR_ri(MVT RetVT, MVT SrcVT, unsigned Op0Reg, bool Op0IsKill,
222 uint64_t Imm, bool IsZExt = false);
224 unsigned AArch64MaterializeInt(const ConstantInt *CI, MVT VT);
225 unsigned AArch64MaterializeFP(const ConstantFP *CFP, MVT VT);
226 unsigned AArch64MaterializeGV(const GlobalValue *GV);
228 // Call handling routines.
230 CCAssignFn *CCAssignFnForCall(CallingConv::ID CC) const;
231 bool ProcessCallArgs(CallLoweringInfo &CLI, SmallVectorImpl<MVT> &ArgVTs,
233 bool FinishCall(CallLoweringInfo &CLI, MVT RetVT, unsigned NumBytes);
236 // Backend specific FastISel code.
237 unsigned fastMaterializeAlloca(const AllocaInst *AI) override;
238 unsigned fastMaterializeConstant(const Constant *C) override;
239 unsigned fastMaterializeFloatZero(const ConstantFP* CF) override;
241 explicit AArch64FastISel(FunctionLoweringInfo &FuncInfo,
242 const TargetLibraryInfo *LibInfo)
243 : FastISel(FuncInfo, LibInfo, /*SkipTargetIndependentISel=*/true) {
244 Subtarget = &TM.getSubtarget<AArch64Subtarget>();
245 Context = &FuncInfo.Fn->getContext();
248 bool fastSelectInstruction(const Instruction *I) override;
250 #include "AArch64GenFastISel.inc"
253 } // end anonymous namespace
255 #include "AArch64GenCallingConv.inc"
257 CCAssignFn *AArch64FastISel::CCAssignFnForCall(CallingConv::ID CC) const {
258 if (CC == CallingConv::WebKit_JS)
259 return CC_AArch64_WebKit_JS;
260 return Subtarget->isTargetDarwin() ? CC_AArch64_DarwinPCS : CC_AArch64_AAPCS;
263 unsigned AArch64FastISel::fastMaterializeAlloca(const AllocaInst *AI) {
264 assert(TLI.getValueType(AI->getType(), true) == MVT::i64 &&
265 "Alloca should always return a pointer.");
267 // Don't handle dynamic allocas.
268 if (!FuncInfo.StaticAllocaMap.count(AI))
271 DenseMap<const AllocaInst *, int>::iterator SI =
272 FuncInfo.StaticAllocaMap.find(AI);
274 if (SI != FuncInfo.StaticAllocaMap.end()) {
275 unsigned ResultReg = createResultReg(&AArch64::GPR64spRegClass);
276 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri),
278 .addFrameIndex(SI->second)
287 unsigned AArch64FastISel::AArch64MaterializeInt(const ConstantInt *CI, MVT VT) {
292 return fastEmit_i(VT, VT, ISD::Constant, CI->getZExtValue());
294 // Create a copy from the zero register to materialize a "0" value.
295 const TargetRegisterClass *RC = (VT == MVT::i64) ? &AArch64::GPR64RegClass
296 : &AArch64::GPR32RegClass;
297 unsigned ZeroReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
298 unsigned ResultReg = createResultReg(RC);
299 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TargetOpcode::COPY),
300 ResultReg).addReg(ZeroReg, getKillRegState(true));
304 unsigned AArch64FastISel::AArch64MaterializeFP(const ConstantFP *CFP, MVT VT) {
305 // Positive zero (+0.0) has to be materialized with a fmov from the zero
306 // register, because the immediate version of fmov cannot encode zero.
307 if (CFP->isNullValue())
308 return fastMaterializeFloatZero(CFP);
310 if (VT != MVT::f32 && VT != MVT::f64)
313 const APFloat Val = CFP->getValueAPF();
314 bool Is64Bit = (VT == MVT::f64);
315 // This checks to see if we can use FMOV instructions to materialize
316 // a constant, otherwise we have to materialize via the constant pool.
317 if (TLI.isFPImmLegal(Val, VT)) {
319 Is64Bit ? AArch64_AM::getFP64Imm(Val) : AArch64_AM::getFP32Imm(Val);
320 assert((Imm != -1) && "Cannot encode floating-point constant.");
321 unsigned Opc = Is64Bit ? AArch64::FMOVDi : AArch64::FMOVSi;
322 return fastEmitInst_i(Opc, TLI.getRegClassFor(VT), Imm);
325 // Materialize via constant pool. MachineConstantPool wants an explicit
327 unsigned Align = DL.getPrefTypeAlignment(CFP->getType());
329 Align = DL.getTypeAllocSize(CFP->getType());
331 unsigned CPI = MCP.getConstantPoolIndex(cast<Constant>(CFP), Align);
332 unsigned ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
333 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
334 ADRPReg).addConstantPoolIndex(CPI, 0, AArch64II::MO_PAGE);
336 unsigned Opc = Is64Bit ? AArch64::LDRDui : AArch64::LDRSui;
337 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
338 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
340 .addConstantPoolIndex(CPI, 0, AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
344 unsigned AArch64FastISel::AArch64MaterializeGV(const GlobalValue *GV) {
345 // We can't handle thread-local variables quickly yet.
346 if (GV->isThreadLocal())
349 // MachO still uses GOT for large code-model accesses, but ELF requires
350 // movz/movk sequences, which FastISel doesn't handle yet.
351 if (TM.getCodeModel() != CodeModel::Small && !Subtarget->isTargetMachO())
354 unsigned char OpFlags = Subtarget->ClassifyGlobalReference(GV, TM);
356 EVT DestEVT = TLI.getValueType(GV->getType(), true);
357 if (!DestEVT.isSimple())
360 unsigned ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
363 if (OpFlags & AArch64II::MO_GOT) {
365 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
367 .addGlobalAddress(GV, 0, AArch64II::MO_GOT | AArch64II::MO_PAGE);
369 ResultReg = createResultReg(&AArch64::GPR64RegClass);
370 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::LDRXui),
373 .addGlobalAddress(GV, 0, AArch64II::MO_GOT | AArch64II::MO_PAGEOFF |
375 } else if (OpFlags & AArch64II::MO_CONSTPOOL) {
376 // We can't handle addresses loaded from a constant pool quickly yet.
380 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
382 .addGlobalAddress(GV, 0, AArch64II::MO_PAGE);
384 ResultReg = createResultReg(&AArch64::GPR64spRegClass);
385 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri),
388 .addGlobalAddress(GV, 0, AArch64II::MO_PAGEOFF | AArch64II::MO_NC)
394 unsigned AArch64FastISel::fastMaterializeConstant(const Constant *C) {
395 EVT CEVT = TLI.getValueType(C->getType(), true);
397 // Only handle simple types.
398 if (!CEVT.isSimple())
400 MVT VT = CEVT.getSimpleVT();
402 if (const auto *CI = dyn_cast<ConstantInt>(C))
403 return AArch64MaterializeInt(CI, VT);
404 else if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
405 return AArch64MaterializeFP(CFP, VT);
406 else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
407 return AArch64MaterializeGV(GV);
412 unsigned AArch64FastISel::fastMaterializeFloatZero(const ConstantFP* CFP) {
413 assert(CFP->isNullValue() &&
414 "Floating-point constant is not a positive zero.");
416 if (!isTypeLegal(CFP->getType(), VT))
419 if (VT != MVT::f32 && VT != MVT::f64)
422 bool Is64Bit = (VT == MVT::f64);
423 unsigned ZReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
424 unsigned Opc = Is64Bit ? AArch64::FMOVXDr : AArch64::FMOVWSr;
425 return fastEmitInst_r(Opc, TLI.getRegClassFor(VT), ZReg, /*IsKill=*/true);
428 // Computes the address to get to an object.
429 bool AArch64FastISel::ComputeAddress(const Value *Obj, Address &Addr, Type *Ty)
431 const User *U = nullptr;
432 unsigned Opcode = Instruction::UserOp1;
433 if (const Instruction *I = dyn_cast<Instruction>(Obj)) {
434 // Don't walk into other basic blocks unless the object is an alloca from
435 // another block, otherwise it may not have a virtual register assigned.
436 if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(Obj)) ||
437 FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
438 Opcode = I->getOpcode();
441 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Obj)) {
442 Opcode = C->getOpcode();
446 if (const PointerType *Ty = dyn_cast<PointerType>(Obj->getType()))
447 if (Ty->getAddressSpace() > 255)
448 // Fast instruction selection doesn't support the special
455 case Instruction::BitCast: {
456 // Look through bitcasts.
457 return ComputeAddress(U->getOperand(0), Addr, Ty);
459 case Instruction::IntToPtr: {
460 // Look past no-op inttoptrs.
461 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
462 return ComputeAddress(U->getOperand(0), Addr, Ty);
465 case Instruction::PtrToInt: {
466 // Look past no-op ptrtoints.
467 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
468 return ComputeAddress(U->getOperand(0), Addr, Ty);
471 case Instruction::GetElementPtr: {
472 Address SavedAddr = Addr;
473 uint64_t TmpOffset = Addr.getOffset();
475 // Iterate through the GEP folding the constants into offsets where
477 gep_type_iterator GTI = gep_type_begin(U);
478 for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e;
480 const Value *Op = *i;
481 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
482 const StructLayout *SL = DL.getStructLayout(STy);
483 unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
484 TmpOffset += SL->getElementOffset(Idx);
486 uint64_t S = DL.getTypeAllocSize(GTI.getIndexedType());
488 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
489 // Constant-offset addressing.
490 TmpOffset += CI->getSExtValue() * S;
493 if (canFoldAddIntoGEP(U, Op)) {
494 // A compatible add with a constant operand. Fold the constant.
496 cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
497 TmpOffset += CI->getSExtValue() * S;
498 // Iterate on the other operand.
499 Op = cast<AddOperator>(Op)->getOperand(0);
503 goto unsupported_gep;
508 // Try to grab the base operand now.
509 Addr.setOffset(TmpOffset);
510 if (ComputeAddress(U->getOperand(0), Addr, Ty))
513 // We failed, restore everything and try the other options.
519 case Instruction::Alloca: {
520 const AllocaInst *AI = cast<AllocaInst>(Obj);
521 DenseMap<const AllocaInst *, int>::iterator SI =
522 FuncInfo.StaticAllocaMap.find(AI);
523 if (SI != FuncInfo.StaticAllocaMap.end()) {
524 Addr.setKind(Address::FrameIndexBase);
525 Addr.setFI(SI->second);
530 case Instruction::Add: {
531 // Adds of constants are common and easy enough.
532 const Value *LHS = U->getOperand(0);
533 const Value *RHS = U->getOperand(1);
535 if (isa<ConstantInt>(LHS))
538 if (const ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
539 Addr.setOffset(Addr.getOffset() + (uint64_t)CI->getSExtValue());
540 return ComputeAddress(LHS, Addr, Ty);
543 Address Backup = Addr;
544 if (ComputeAddress(LHS, Addr, Ty) && ComputeAddress(RHS, Addr, Ty))
550 case Instruction::Shl:
551 if (Addr.getOffsetReg())
554 if (const auto *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
555 unsigned Val = CI->getZExtValue();
556 if (Val < 1 || Val > 3)
559 uint64_t NumBytes = 0;
560 if (Ty && Ty->isSized()) {
561 uint64_t NumBits = DL.getTypeSizeInBits(Ty);
562 NumBytes = NumBits / 8;
563 if (!isPowerOf2_64(NumBits))
567 if (NumBytes != (1ULL << Val))
571 Addr.setExtendType(AArch64_AM::LSL);
573 if (const auto *I = dyn_cast<Instruction>(U->getOperand(0)))
574 if (FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB)
577 if (const auto *ZE = dyn_cast<ZExtInst>(U))
578 if (ZE->getOperand(0)->getType()->isIntegerTy(32))
579 Addr.setExtendType(AArch64_AM::UXTW);
581 if (const auto *SE = dyn_cast<SExtInst>(U))
582 if (SE->getOperand(0)->getType()->isIntegerTy(32))
583 Addr.setExtendType(AArch64_AM::SXTW);
585 unsigned Reg = getRegForValue(U->getOperand(0));
588 Addr.setOffsetReg(Reg);
595 if (!Addr.getOffsetReg()) {
596 unsigned Reg = getRegForValue(Obj);
599 Addr.setOffsetReg(Reg);
605 unsigned Reg = getRegForValue(Obj);
612 bool AArch64FastISel::ComputeCallAddress(const Value *V, Address &Addr) {
613 const User *U = nullptr;
614 unsigned Opcode = Instruction::UserOp1;
617 if (const auto *I = dyn_cast<Instruction>(V)) {
618 Opcode = I->getOpcode();
620 InMBB = I->getParent() == FuncInfo.MBB->getBasicBlock();
621 } else if (const auto *C = dyn_cast<ConstantExpr>(V)) {
622 Opcode = C->getOpcode();
628 case Instruction::BitCast:
629 // Look past bitcasts if its operand is in the same BB.
631 return ComputeCallAddress(U->getOperand(0), Addr);
633 case Instruction::IntToPtr:
634 // Look past no-op inttoptrs if its operand is in the same BB.
636 TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
637 return ComputeCallAddress(U->getOperand(0), Addr);
639 case Instruction::PtrToInt:
640 // Look past no-op ptrtoints if its operand is in the same BB.
642 TLI.getValueType(U->getType()) == TLI.getPointerTy())
643 return ComputeCallAddress(U->getOperand(0), Addr);
647 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
648 Addr.setGlobalValue(GV);
652 // If all else fails, try to materialize the value in a register.
653 if (!Addr.getGlobalValue()) {
654 Addr.setReg(getRegForValue(V));
655 return Addr.getReg() != 0;
662 bool AArch64FastISel::isTypeLegal(Type *Ty, MVT &VT) {
663 EVT evt = TLI.getValueType(Ty, true);
665 // Only handle simple types.
666 if (evt == MVT::Other || !evt.isSimple())
668 VT = evt.getSimpleVT();
670 // This is a legal type, but it's not something we handle in fast-isel.
674 // Handle all other legal types, i.e. a register that will directly hold this
676 return TLI.isTypeLegal(VT);
679 bool AArch64FastISel::isLoadStoreTypeLegal(Type *Ty, MVT &VT) {
680 if (isTypeLegal(Ty, VT))
683 // If this is a type than can be sign or zero-extended to a basic operation
684 // go ahead and accept it now. For stores, this reflects truncation.
685 if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16)
691 /// \brief Determine if the value type is supported by FastISel.
693 /// FastISel for AArch64 can handle more value types than are legal. This adds
694 /// simple value type such as i1, i8, and i16.
695 /// Vectors on the other side are not supported yet.
696 bool AArch64FastISel::isTypeSupported(Type *Ty, MVT &VT) {
697 if (Ty->isVectorTy())
700 if (isTypeLegal(Ty, VT))
703 // If this is a type than can be sign or zero-extended to a basic operation
704 // go ahead and accept it now.
705 if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16)
711 bool AArch64FastISel::isValueAvailable(const Value *V) const {
712 if (!isa<Instruction>(V))
715 const auto *I = cast<Instruction>(V);
716 if (FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB)
722 bool AArch64FastISel::SimplifyAddress(Address &Addr, MVT VT) {
723 unsigned ScaleFactor;
724 switch (VT.SimpleTy) {
725 default: return false;
726 case MVT::i1: // fall-through
727 case MVT::i8: ScaleFactor = 1; break;
728 case MVT::i16: ScaleFactor = 2; break;
729 case MVT::i32: // fall-through
730 case MVT::f32: ScaleFactor = 4; break;
731 case MVT::i64: // fall-through
732 case MVT::f64: ScaleFactor = 8; break;
735 bool ImmediateOffsetNeedsLowering = false;
736 bool RegisterOffsetNeedsLowering = false;
737 int64_t Offset = Addr.getOffset();
738 if (((Offset < 0) || (Offset & (ScaleFactor - 1))) && !isInt<9>(Offset))
739 ImmediateOffsetNeedsLowering = true;
740 else if (Offset > 0 && !(Offset & (ScaleFactor - 1)) &&
741 !isUInt<12>(Offset / ScaleFactor))
742 ImmediateOffsetNeedsLowering = true;
744 // Cannot encode an offset register and an immediate offset in the same
745 // instruction. Fold the immediate offset into the load/store instruction and
746 // emit an additonal add to take care of the offset register.
747 if (!ImmediateOffsetNeedsLowering && Addr.getOffset() && Addr.isRegBase() &&
749 RegisterOffsetNeedsLowering = true;
751 // Cannot encode zero register as base.
752 if (Addr.isRegBase() && Addr.getOffsetReg() && !Addr.getReg())
753 RegisterOffsetNeedsLowering = true;
755 // If this is a stack pointer and the offset needs to be simplified then put
756 // the alloca address into a register, set the base type back to register and
757 // continue. This should almost never happen.
758 if (ImmediateOffsetNeedsLowering && Addr.isFIBase()) {
759 unsigned ResultReg = createResultReg(&AArch64::GPR64spRegClass);
760 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri),
762 .addFrameIndex(Addr.getFI())
765 Addr.setKind(Address::RegBase);
766 Addr.setReg(ResultReg);
769 if (RegisterOffsetNeedsLowering) {
770 unsigned ResultReg = 0;
772 if (Addr.getExtendType() == AArch64_AM::SXTW ||
773 Addr.getExtendType() == AArch64_AM::UXTW )
774 ResultReg = emitAddSub_rx(/*UseAdd=*/true, MVT::i64, Addr.getReg(),
775 /*TODO:IsKill=*/false, Addr.getOffsetReg(),
776 /*TODO:IsKill=*/false, Addr.getExtendType(),
779 ResultReg = emitAddSub_rs(/*UseAdd=*/true, MVT::i64, Addr.getReg(),
780 /*TODO:IsKill=*/false, Addr.getOffsetReg(),
781 /*TODO:IsKill=*/false, AArch64_AM::LSL,
784 if (Addr.getExtendType() == AArch64_AM::UXTW)
785 ResultReg = emitLSL_ri(MVT::i64, MVT::i32, Addr.getOffsetReg(),
786 /*Op0IsKill=*/false, Addr.getShift(),
788 else if (Addr.getExtendType() == AArch64_AM::SXTW)
789 ResultReg = emitLSL_ri(MVT::i64, MVT::i32, Addr.getOffsetReg(),
790 /*Op0IsKill=*/false, Addr.getShift(),
793 ResultReg = emitLSL_ri(MVT::i64, MVT::i64, Addr.getOffsetReg(),
794 /*Op0IsKill=*/false, Addr.getShift());
799 Addr.setReg(ResultReg);
800 Addr.setOffsetReg(0);
802 Addr.setExtendType(AArch64_AM::InvalidShiftExtend);
805 // Since the offset is too large for the load/store instruction get the
806 // reg+offset into a register.
807 if (ImmediateOffsetNeedsLowering) {
808 unsigned ResultReg = 0;
810 ResultReg = fastEmit_ri_(MVT::i64, ISD::ADD, Addr.getReg(),
811 /*IsKill=*/false, Offset, MVT::i64);
813 ResultReg = fastEmit_i(MVT::i64, MVT::i64, ISD::Constant, Offset);
817 Addr.setReg(ResultReg);
823 void AArch64FastISel::AddLoadStoreOperands(Address &Addr,
824 const MachineInstrBuilder &MIB,
826 unsigned ScaleFactor,
827 MachineMemOperand *MMO) {
828 int64_t Offset = Addr.getOffset() / ScaleFactor;
829 // Frame base works a bit differently. Handle it separately.
830 if (Addr.isFIBase()) {
831 int FI = Addr.getFI();
832 // FIXME: We shouldn't be using getObjectSize/getObjectAlignment. The size
833 // and alignment should be based on the VT.
834 MMO = FuncInfo.MF->getMachineMemOperand(
835 MachinePointerInfo::getFixedStack(FI, Offset), Flags,
836 MFI.getObjectSize(FI), MFI.getObjectAlignment(FI));
837 // Now add the rest of the operands.
838 MIB.addFrameIndex(FI).addImm(Offset);
840 assert(Addr.isRegBase() && "Unexpected address kind.");
841 const MCInstrDesc &II = MIB->getDesc();
842 unsigned Idx = (Flags & MachineMemOperand::MOStore) ? 1 : 0;
844 constrainOperandRegClass(II, Addr.getReg(), II.getNumDefs()+Idx));
846 constrainOperandRegClass(II, Addr.getOffsetReg(), II.getNumDefs()+Idx+1));
847 if (Addr.getOffsetReg()) {
848 assert(Addr.getOffset() == 0 && "Unexpected offset");
849 bool IsSigned = Addr.getExtendType() == AArch64_AM::SXTW ||
850 Addr.getExtendType() == AArch64_AM::SXTX;
851 MIB.addReg(Addr.getReg());
852 MIB.addReg(Addr.getOffsetReg());
853 MIB.addImm(IsSigned);
854 MIB.addImm(Addr.getShift() != 0);
856 MIB.addReg(Addr.getReg());
862 MIB.addMemOperand(MMO);
865 unsigned AArch64FastISel::emitAddSub(bool UseAdd, MVT RetVT, const Value *LHS,
866 const Value *RHS, bool SetFlags,
867 bool WantResult, bool IsZExt) {
868 AArch64_AM::ShiftExtendType ExtendType = AArch64_AM::InvalidShiftExtend;
869 bool NeedExtend = false;
870 switch (RetVT.SimpleTy) {
878 ExtendType = IsZExt ? AArch64_AM::UXTB : AArch64_AM::SXTB;
882 ExtendType = IsZExt ? AArch64_AM::UXTH : AArch64_AM::SXTH;
884 case MVT::i32: // fall-through
889 RetVT.SimpleTy = std::max(RetVT.SimpleTy, MVT::i32);
891 // Canonicalize immediates to the RHS first.
892 if (UseAdd && isa<ConstantInt>(LHS) && !isa<ConstantInt>(RHS))
895 // Canonicalize shift immediate to the RHS.
896 if (UseAdd && isValueAvailable(LHS))
897 if (const auto *SI = dyn_cast<BinaryOperator>(LHS))
898 if (isa<ConstantInt>(SI->getOperand(1)))
899 if (SI->getOpcode() == Instruction::Shl ||
900 SI->getOpcode() == Instruction::LShr ||
901 SI->getOpcode() == Instruction::AShr )
904 unsigned LHSReg = getRegForValue(LHS);
907 bool LHSIsKill = hasTrivialKill(LHS);
910 LHSReg = EmitIntExt(SrcVT, LHSReg, RetVT, IsZExt);
912 unsigned ResultReg = 0;
913 if (const auto *C = dyn_cast<ConstantInt>(RHS)) {
914 uint64_t Imm = IsZExt ? C->getZExtValue() : C->getSExtValue();
916 ResultReg = emitAddSub_ri(!UseAdd, RetVT, LHSReg, LHSIsKill, -Imm,
917 SetFlags, WantResult);
919 ResultReg = emitAddSub_ri(UseAdd, RetVT, LHSReg, LHSIsKill, Imm, SetFlags,
925 // Only extend the RHS within the instruction if there is a valid extend type.
926 if (ExtendType != AArch64_AM::InvalidShiftExtend && isValueAvailable(RHS)) {
927 if (const auto *SI = dyn_cast<BinaryOperator>(RHS))
928 if (const auto *C = dyn_cast<ConstantInt>(SI->getOperand(1)))
929 if ((SI->getOpcode() == Instruction::Shl) && (C->getZExtValue() < 4)) {
930 unsigned RHSReg = getRegForValue(SI->getOperand(0));
933 bool RHSIsKill = hasTrivialKill(SI->getOperand(0));
934 return emitAddSub_rx(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg,
935 RHSIsKill, ExtendType, C->getZExtValue(),
936 SetFlags, WantResult);
938 unsigned RHSReg = getRegForValue(RHS);
941 bool RHSIsKill = hasTrivialKill(RHS);
942 return emitAddSub_rx(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg, RHSIsKill,
943 ExtendType, 0, SetFlags, WantResult);
946 // Check if the shift can be folded into the instruction.
947 if (isValueAvailable(RHS))
948 if (const auto *SI = dyn_cast<BinaryOperator>(RHS)) {
949 if (const auto *C = dyn_cast<ConstantInt>(SI->getOperand(1))) {
950 AArch64_AM::ShiftExtendType ShiftType = AArch64_AM::InvalidShiftExtend;
951 switch (SI->getOpcode()) {
953 case Instruction::Shl: ShiftType = AArch64_AM::LSL; break;
954 case Instruction::LShr: ShiftType = AArch64_AM::LSR; break;
955 case Instruction::AShr: ShiftType = AArch64_AM::ASR; break;
957 uint64_t ShiftVal = C->getZExtValue();
958 if (ShiftType != AArch64_AM::InvalidShiftExtend) {
959 unsigned RHSReg = getRegForValue(SI->getOperand(0));
962 bool RHSIsKill = hasTrivialKill(SI->getOperand(0));
963 return emitAddSub_rs(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg,
964 RHSIsKill, ShiftType, ShiftVal, SetFlags,
970 unsigned RHSReg = getRegForValue(RHS);
973 bool RHSIsKill = hasTrivialKill(RHS);
976 RHSReg = EmitIntExt(SrcVT, RHSReg, RetVT, IsZExt);
978 return emitAddSub_rr(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg, RHSIsKill,
979 SetFlags, WantResult);
982 unsigned AArch64FastISel::emitAddSub_rr(bool UseAdd, MVT RetVT, unsigned LHSReg,
983 bool LHSIsKill, unsigned RHSReg,
984 bool RHSIsKill, bool SetFlags,
986 assert(LHSReg && RHSReg && "Invalid register number.");
988 if (RetVT != MVT::i32 && RetVT != MVT::i64)
991 static const unsigned OpcTable[2][2][2] = {
992 { { AArch64::SUBWrr, AArch64::SUBXrr },
993 { AArch64::ADDWrr, AArch64::ADDXrr } },
994 { { AArch64::SUBSWrr, AArch64::SUBSXrr },
995 { AArch64::ADDSWrr, AArch64::ADDSXrr } }
997 bool Is64Bit = RetVT == MVT::i64;
998 unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit];
999 const TargetRegisterClass *RC =
1000 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
1003 ResultReg = createResultReg(RC);
1005 ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
1007 const MCInstrDesc &II = TII.get(Opc);
1008 LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
1009 RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1);
1010 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
1011 .addReg(LHSReg, getKillRegState(LHSIsKill))
1012 .addReg(RHSReg, getKillRegState(RHSIsKill));
1016 unsigned AArch64FastISel::emitAddSub_ri(bool UseAdd, MVT RetVT, unsigned LHSReg,
1017 bool LHSIsKill, uint64_t Imm,
1018 bool SetFlags, bool WantResult) {
1019 assert(LHSReg && "Invalid register number.");
1021 if (RetVT != MVT::i32 && RetVT != MVT::i64)
1025 if (isUInt<12>(Imm))
1027 else if ((Imm & 0xfff000) == Imm) {
1033 static const unsigned OpcTable[2][2][2] = {
1034 { { AArch64::SUBWri, AArch64::SUBXri },
1035 { AArch64::ADDWri, AArch64::ADDXri } },
1036 { { AArch64::SUBSWri, AArch64::SUBSXri },
1037 { AArch64::ADDSWri, AArch64::ADDSXri } }
1039 bool Is64Bit = RetVT == MVT::i64;
1040 unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit];
1041 const TargetRegisterClass *RC;
1043 RC = Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
1045 RC = Is64Bit ? &AArch64::GPR64spRegClass : &AArch64::GPR32spRegClass;
1048 ResultReg = createResultReg(RC);
1050 ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
1052 const MCInstrDesc &II = TII.get(Opc);
1053 LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
1054 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
1055 .addReg(LHSReg, getKillRegState(LHSIsKill))
1057 .addImm(getShifterImm(AArch64_AM::LSL, ShiftImm));
1061 unsigned AArch64FastISel::emitAddSub_rs(bool UseAdd, MVT RetVT, unsigned LHSReg,
1062 bool LHSIsKill, unsigned RHSReg,
1064 AArch64_AM::ShiftExtendType ShiftType,
1065 uint64_t ShiftImm, bool SetFlags,
1067 assert(LHSReg && RHSReg && "Invalid register number.");
1069 if (RetVT != MVT::i32 && RetVT != MVT::i64)
1072 static const unsigned OpcTable[2][2][2] = {
1073 { { AArch64::SUBWrs, AArch64::SUBXrs },
1074 { AArch64::ADDWrs, AArch64::ADDXrs } },
1075 { { AArch64::SUBSWrs, AArch64::SUBSXrs },
1076 { AArch64::ADDSWrs, AArch64::ADDSXrs } }
1078 bool Is64Bit = RetVT == MVT::i64;
1079 unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit];
1080 const TargetRegisterClass *RC =
1081 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
1084 ResultReg = createResultReg(RC);
1086 ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
1088 const MCInstrDesc &II = TII.get(Opc);
1089 LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
1090 RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1);
1091 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
1092 .addReg(LHSReg, getKillRegState(LHSIsKill))
1093 .addReg(RHSReg, getKillRegState(RHSIsKill))
1094 .addImm(getShifterImm(ShiftType, ShiftImm));
1098 unsigned AArch64FastISel::emitAddSub_rx(bool UseAdd, MVT RetVT, unsigned LHSReg,
1099 bool LHSIsKill, unsigned RHSReg,
1101 AArch64_AM::ShiftExtendType ExtType,
1102 uint64_t ShiftImm, bool SetFlags,
1104 assert(LHSReg && RHSReg && "Invalid register number.");
1106 if (RetVT != MVT::i32 && RetVT != MVT::i64)
1109 static const unsigned OpcTable[2][2][2] = {
1110 { { AArch64::SUBWrx, AArch64::SUBXrx },
1111 { AArch64::ADDWrx, AArch64::ADDXrx } },
1112 { { AArch64::SUBSWrx, AArch64::SUBSXrx },
1113 { AArch64::ADDSWrx, AArch64::ADDSXrx } }
1115 bool Is64Bit = RetVT == MVT::i64;
1116 unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit];
1117 const TargetRegisterClass *RC = nullptr;
1119 RC = Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
1121 RC = Is64Bit ? &AArch64::GPR64spRegClass : &AArch64::GPR32spRegClass;
1124 ResultReg = createResultReg(RC);
1126 ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
1128 const MCInstrDesc &II = TII.get(Opc);
1129 LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
1130 RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1);
1131 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
1132 .addReg(LHSReg, getKillRegState(LHSIsKill))
1133 .addReg(RHSReg, getKillRegState(RHSIsKill))
1134 .addImm(getArithExtendImm(ExtType, ShiftImm));
1138 bool AArch64FastISel::emitCmp(const Value *LHS, const Value *RHS, bool IsZExt) {
1139 Type *Ty = LHS->getType();
1140 EVT EVT = TLI.getValueType(Ty, true);
1141 if (!EVT.isSimple())
1143 MVT VT = EVT.getSimpleVT();
1145 switch (VT.SimpleTy) {
1153 return emitICmp(VT, LHS, RHS, IsZExt);
1156 return emitFCmp(VT, LHS, RHS);
1160 bool AArch64FastISel::emitICmp(MVT RetVT, const Value *LHS, const Value *RHS,
1162 return emitSub(RetVT, LHS, RHS, /*SetFlags=*/true, /*WantResult=*/false,
1166 bool AArch64FastISel::emitICmp_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill,
1168 return emitAddSub_ri(/*UseAdd=*/false, RetVT, LHSReg, LHSIsKill, Imm,
1169 /*SetFlags=*/true, /*WantResult=*/false) != 0;
1172 bool AArch64FastISel::emitFCmp(MVT RetVT, const Value *LHS, const Value *RHS) {
1173 if (RetVT != MVT::f32 && RetVT != MVT::f64)
1176 // Check to see if the 2nd operand is a constant that we can encode directly
1178 bool UseImm = false;
1179 if (const auto *CFP = dyn_cast<ConstantFP>(RHS))
1180 if (CFP->isZero() && !CFP->isNegative())
1183 unsigned LHSReg = getRegForValue(LHS);
1186 bool LHSIsKill = hasTrivialKill(LHS);
1189 unsigned Opc = (RetVT == MVT::f64) ? AArch64::FCMPDri : AArch64::FCMPSri;
1190 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc))
1191 .addReg(LHSReg, getKillRegState(LHSIsKill));
1195 unsigned RHSReg = getRegForValue(RHS);
1198 bool RHSIsKill = hasTrivialKill(RHS);
1200 unsigned Opc = (RetVT == MVT::f64) ? AArch64::FCMPDrr : AArch64::FCMPSrr;
1201 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc))
1202 .addReg(LHSReg, getKillRegState(LHSIsKill))
1203 .addReg(RHSReg, getKillRegState(RHSIsKill));
1207 unsigned AArch64FastISel::emitAdd(MVT RetVT, const Value *LHS, const Value *RHS,
1208 bool SetFlags, bool WantResult, bool IsZExt) {
1209 return emitAddSub(/*UseAdd=*/true, RetVT, LHS, RHS, SetFlags, WantResult,
1213 unsigned AArch64FastISel::emitSub(MVT RetVT, const Value *LHS, const Value *RHS,
1214 bool SetFlags, bool WantResult, bool IsZExt) {
1215 return emitAddSub(/*UseAdd=*/false, RetVT, LHS, RHS, SetFlags, WantResult,
1219 unsigned AArch64FastISel::emitSubs_rr(MVT RetVT, unsigned LHSReg,
1220 bool LHSIsKill, unsigned RHSReg,
1221 bool RHSIsKill, bool WantResult) {
1222 return emitAddSub_rr(/*UseAdd=*/false, RetVT, LHSReg, LHSIsKill, RHSReg,
1223 RHSIsKill, /*SetFlags=*/true, WantResult);
1226 unsigned AArch64FastISel::emitSubs_rs(MVT RetVT, unsigned LHSReg,
1227 bool LHSIsKill, unsigned RHSReg,
1229 AArch64_AM::ShiftExtendType ShiftType,
1230 uint64_t ShiftImm, bool WantResult) {
1231 return emitAddSub_rs(/*UseAdd=*/false, RetVT, LHSReg, LHSIsKill, RHSReg,
1232 RHSIsKill, ShiftType, ShiftImm, /*SetFlags=*/true,
1236 unsigned AArch64FastISel::emitLogicalOp(unsigned ISDOpc, MVT RetVT,
1237 const Value *LHS, const Value *RHS) {
1238 // Canonicalize immediates to the RHS first.
1239 if (isa<ConstantInt>(LHS) && !isa<ConstantInt>(RHS))
1240 std::swap(LHS, RHS);
1242 // Canonicalize shift immediate to the RHS.
1243 if (isValueAvailable(LHS))
1244 if (const auto *SI = dyn_cast<BinaryOperator>(LHS))
1245 if (isa<ConstantInt>(SI->getOperand(1)))
1246 if (SI->getOpcode() == Instruction::Shl)
1247 std::swap(LHS, RHS);
1249 unsigned LHSReg = getRegForValue(LHS);
1252 bool LHSIsKill = hasTrivialKill(LHS);
1254 unsigned ResultReg = 0;
1255 if (const auto *C = dyn_cast<ConstantInt>(RHS)) {
1256 uint64_t Imm = C->getZExtValue();
1257 ResultReg = emitLogicalOp_ri(ISDOpc, RetVT, LHSReg, LHSIsKill, Imm);
1262 // Check if the shift can be folded into the instruction.
1263 if (isValueAvailable(RHS))
1264 if (const auto *SI = dyn_cast<BinaryOperator>(RHS))
1265 if (const auto *C = dyn_cast<ConstantInt>(SI->getOperand(1)))
1266 if (SI->getOpcode() == Instruction::Shl) {
1267 uint64_t ShiftVal = C->getZExtValue();
1268 unsigned RHSReg = getRegForValue(SI->getOperand(0));
1271 bool RHSIsKill = hasTrivialKill(SI->getOperand(0));
1272 return emitLogicalOp_rs(ISDOpc, RetVT, LHSReg, LHSIsKill, RHSReg,
1273 RHSIsKill, ShiftVal);
1276 unsigned RHSReg = getRegForValue(RHS);
1279 bool RHSIsKill = hasTrivialKill(RHS);
1281 MVT VT = std::max(MVT::i32, RetVT.SimpleTy);
1282 ResultReg = fastEmit_rr(VT, VT, ISDOpc, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
1283 if (RetVT >= MVT::i8 && RetVT <= MVT::i16) {
1284 uint64_t Mask = (RetVT == MVT::i8) ? 0xff : 0xffff;
1285 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
1290 unsigned AArch64FastISel::emitLogicalOp_ri(unsigned ISDOpc, MVT RetVT,
1291 unsigned LHSReg, bool LHSIsKill,
1293 assert((ISD::AND + 1 == ISD::OR) && (ISD::AND + 2 == ISD::XOR) &&
1294 "ISD nodes are not consecutive!");
1295 static const unsigned OpcTable[3][2] = {
1296 { AArch64::ANDWri, AArch64::ANDXri },
1297 { AArch64::ORRWri, AArch64::ORRXri },
1298 { AArch64::EORWri, AArch64::EORXri }
1300 const TargetRegisterClass *RC;
1303 switch (RetVT.SimpleTy) {
1310 unsigned Idx = ISDOpc - ISD::AND;
1311 Opc = OpcTable[Idx][0];
1312 RC = &AArch64::GPR32spRegClass;
1317 Opc = OpcTable[ISDOpc - ISD::AND][1];
1318 RC = &AArch64::GPR64spRegClass;
1323 if (!AArch64_AM::isLogicalImmediate(Imm, RegSize))
1326 unsigned ResultReg =
1327 fastEmitInst_ri(Opc, RC, LHSReg, LHSIsKill,
1328 AArch64_AM::encodeLogicalImmediate(Imm, RegSize));
1329 if (RetVT >= MVT::i8 && RetVT <= MVT::i16 && ISDOpc != ISD::AND) {
1330 uint64_t Mask = (RetVT == MVT::i8) ? 0xff : 0xffff;
1331 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
1336 unsigned AArch64FastISel::emitLogicalOp_rs(unsigned ISDOpc, MVT RetVT,
1337 unsigned LHSReg, bool LHSIsKill,
1338 unsigned RHSReg, bool RHSIsKill,
1339 uint64_t ShiftImm) {
1340 assert((ISD::AND + 1 == ISD::OR) && (ISD::AND + 2 == ISD::XOR) &&
1341 "ISD nodes are not consecutive!");
1342 static const unsigned OpcTable[3][2] = {
1343 { AArch64::ANDWrs, AArch64::ANDXrs },
1344 { AArch64::ORRWrs, AArch64::ORRXrs },
1345 { AArch64::EORWrs, AArch64::EORXrs }
1347 const TargetRegisterClass *RC;
1349 switch (RetVT.SimpleTy) {
1356 Opc = OpcTable[ISDOpc - ISD::AND][0];
1357 RC = &AArch64::GPR32RegClass;
1360 Opc = OpcTable[ISDOpc - ISD::AND][1];
1361 RC = &AArch64::GPR64RegClass;
1364 unsigned ResultReg =
1365 fastEmitInst_rri(Opc, RC, LHSReg, LHSIsKill, RHSReg, RHSIsKill,
1366 AArch64_AM::getShifterImm(AArch64_AM::LSL, ShiftImm));
1367 if (RetVT >= MVT::i8 && RetVT <= MVT::i16) {
1368 uint64_t Mask = (RetVT == MVT::i8) ? 0xff : 0xffff;
1369 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
1374 unsigned AArch64FastISel::emitAnd_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill,
1376 return emitLogicalOp_ri(ISD::AND, RetVT, LHSReg, LHSIsKill, Imm);
1379 bool AArch64FastISel::EmitLoad(MVT VT, unsigned &ResultReg, Address Addr,
1380 MachineMemOperand *MMO) {
1381 // Simplify this down to something we can handle.
1382 if (!SimplifyAddress(Addr, VT))
1385 unsigned ScaleFactor;
1386 switch (VT.SimpleTy) {
1387 default: llvm_unreachable("Unexpected value type.");
1388 case MVT::i1: // fall-through
1389 case MVT::i8: ScaleFactor = 1; break;
1390 case MVT::i16: ScaleFactor = 2; break;
1391 case MVT::i32: // fall-through
1392 case MVT::f32: ScaleFactor = 4; break;
1393 case MVT::i64: // fall-through
1394 case MVT::f64: ScaleFactor = 8; break;
1397 // Negative offsets require unscaled, 9-bit, signed immediate offsets.
1398 // Otherwise, we try using scaled, 12-bit, unsigned immediate offsets.
1399 bool UseScaled = true;
1400 if ((Addr.getOffset() < 0) || (Addr.getOffset() & (ScaleFactor - 1))) {
1405 static const unsigned OpcTable[4][6] = {
1406 { AArch64::LDURBBi, AArch64::LDURHHi, AArch64::LDURWi, AArch64::LDURXi,
1407 AArch64::LDURSi, AArch64::LDURDi },
1408 { AArch64::LDRBBui, AArch64::LDRHHui, AArch64::LDRWui, AArch64::LDRXui,
1409 AArch64::LDRSui, AArch64::LDRDui },
1410 { AArch64::LDRBBroX, AArch64::LDRHHroX, AArch64::LDRWroX, AArch64::LDRXroX,
1411 AArch64::LDRSroX, AArch64::LDRDroX },
1412 { AArch64::LDRBBroW, AArch64::LDRHHroW, AArch64::LDRWroW, AArch64::LDRXroW,
1413 AArch64::LDRSroW, AArch64::LDRDroW }
1417 const TargetRegisterClass *RC;
1418 bool VTIsi1 = false;
1419 bool UseRegOffset = Addr.isRegBase() && !Addr.getOffset() && Addr.getReg() &&
1420 Addr.getOffsetReg();
1421 unsigned Idx = UseRegOffset ? 2 : UseScaled ? 1 : 0;
1422 if (Addr.getExtendType() == AArch64_AM::UXTW ||
1423 Addr.getExtendType() == AArch64_AM::SXTW)
1426 switch (VT.SimpleTy) {
1427 default: llvm_unreachable("Unexpected value type.");
1428 case MVT::i1: VTIsi1 = true; // Intentional fall-through.
1429 case MVT::i8: Opc = OpcTable[Idx][0]; RC = &AArch64::GPR32RegClass; break;
1430 case MVT::i16: Opc = OpcTable[Idx][1]; RC = &AArch64::GPR32RegClass; break;
1431 case MVT::i32: Opc = OpcTable[Idx][2]; RC = &AArch64::GPR32RegClass; break;
1432 case MVT::i64: Opc = OpcTable[Idx][3]; RC = &AArch64::GPR64RegClass; break;
1433 case MVT::f32: Opc = OpcTable[Idx][4]; RC = &AArch64::FPR32RegClass; break;
1434 case MVT::f64: Opc = OpcTable[Idx][5]; RC = &AArch64::FPR64RegClass; break;
1437 // Create the base instruction, then add the operands.
1438 ResultReg = createResultReg(RC);
1439 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1440 TII.get(Opc), ResultReg);
1441 AddLoadStoreOperands(Addr, MIB, MachineMemOperand::MOLoad, ScaleFactor, MMO);
1443 // Loading an i1 requires special handling.
1445 unsigned ANDReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, 1);
1446 assert(ANDReg && "Unexpected AND instruction emission failure.");
1452 bool AArch64FastISel::selectAddSub(const Instruction *I) {
1454 if (!isTypeSupported(I->getType(), VT))
1458 if (I->getOpcode() == Instruction::Add)
1459 ResultReg = emitAdd(VT, I->getOperand(0), I->getOperand(1));
1460 else if (I->getOpcode() == Instruction::Sub)
1461 ResultReg = emitSub(VT, I->getOperand(0), I->getOperand(1));
1463 llvm_unreachable("Unexpected instruction.");
1465 assert(ResultReg && "Couldn't select Add/Sub instruction.");
1466 updateValueMap(I, ResultReg);
1470 bool AArch64FastISel::selectLogicalOp(const Instruction *I, unsigned ISDOpc) {
1472 if (!isTypeSupported(I->getType(), VT))
1475 unsigned ResultReg =
1476 emitLogicalOp(ISDOpc, VT, I->getOperand(0), I->getOperand(1));
1480 updateValueMap(I, ResultReg);
1484 bool AArch64FastISel::SelectLoad(const Instruction *I) {
1486 // Verify we have a legal type before going any further. Currently, we handle
1487 // simple types that will directly fit in a register (i32/f32/i64/f64) or
1488 // those that can be sign or zero-extended to a basic operation (i1/i8/i16).
1489 if (!isLoadStoreTypeLegal(I->getType(), VT) || cast<LoadInst>(I)->isAtomic())
1492 // See if we can handle this address.
1494 if (!ComputeAddress(I->getOperand(0), Addr, I->getType()))
1498 if (!EmitLoad(VT, ResultReg, Addr, createMachineMemOperandFor(I)))
1501 updateValueMap(I, ResultReg);
1505 bool AArch64FastISel::EmitStore(MVT VT, unsigned SrcReg, Address Addr,
1506 MachineMemOperand *MMO) {
1507 // Simplify this down to something we can handle.
1508 if (!SimplifyAddress(Addr, VT))
1511 unsigned ScaleFactor;
1512 switch (VT.SimpleTy) {
1513 default: llvm_unreachable("Unexpected value type.");
1514 case MVT::i1: // fall-through
1515 case MVT::i8: ScaleFactor = 1; break;
1516 case MVT::i16: ScaleFactor = 2; break;
1517 case MVT::i32: // fall-through
1518 case MVT::f32: ScaleFactor = 4; break;
1519 case MVT::i64: // fall-through
1520 case MVT::f64: ScaleFactor = 8; break;
1523 // Negative offsets require unscaled, 9-bit, signed immediate offsets.
1524 // Otherwise, we try using scaled, 12-bit, unsigned immediate offsets.
1525 bool UseScaled = true;
1526 if ((Addr.getOffset() < 0) || (Addr.getOffset() & (ScaleFactor - 1))) {
1532 static const unsigned OpcTable[4][6] = {
1533 { AArch64::STURBBi, AArch64::STURHHi, AArch64::STURWi, AArch64::STURXi,
1534 AArch64::STURSi, AArch64::STURDi },
1535 { AArch64::STRBBui, AArch64::STRHHui, AArch64::STRWui, AArch64::STRXui,
1536 AArch64::STRSui, AArch64::STRDui },
1537 { AArch64::STRBBroX, AArch64::STRHHroX, AArch64::STRWroX, AArch64::STRXroX,
1538 AArch64::STRSroX, AArch64::STRDroX },
1539 { AArch64::STRBBroW, AArch64::STRHHroW, AArch64::STRWroW, AArch64::STRXroW,
1540 AArch64::STRSroW, AArch64::STRDroW }
1545 bool VTIsi1 = false;
1546 bool UseRegOffset = Addr.isRegBase() && !Addr.getOffset() && Addr.getReg() &&
1547 Addr.getOffsetReg();
1548 unsigned Idx = UseRegOffset ? 2 : UseScaled ? 1 : 0;
1549 if (Addr.getExtendType() == AArch64_AM::UXTW ||
1550 Addr.getExtendType() == AArch64_AM::SXTW)
1553 switch (VT.SimpleTy) {
1554 default: llvm_unreachable("Unexpected value type.");
1555 case MVT::i1: VTIsi1 = true;
1556 case MVT::i8: Opc = OpcTable[Idx][0]; break;
1557 case MVT::i16: Opc = OpcTable[Idx][1]; break;
1558 case MVT::i32: Opc = OpcTable[Idx][2]; break;
1559 case MVT::i64: Opc = OpcTable[Idx][3]; break;
1560 case MVT::f32: Opc = OpcTable[Idx][4]; break;
1561 case MVT::f64: Opc = OpcTable[Idx][5]; break;
1564 // Storing an i1 requires special handling.
1565 if (VTIsi1 && SrcReg != AArch64::WZR) {
1566 unsigned ANDReg = emitAnd_ri(MVT::i32, SrcReg, /*TODO:IsKill=*/false, 1);
1567 assert(ANDReg && "Unexpected AND instruction emission failure.");
1570 // Create the base instruction, then add the operands.
1571 const MCInstrDesc &II = TII.get(Opc);
1572 SrcReg = constrainOperandRegClass(II, SrcReg, II.getNumDefs());
1573 MachineInstrBuilder MIB =
1574 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II).addReg(SrcReg);
1575 AddLoadStoreOperands(Addr, MIB, MachineMemOperand::MOStore, ScaleFactor, MMO);
1580 bool AArch64FastISel::SelectStore(const Instruction *I) {
1582 const Value *Op0 = I->getOperand(0);
1583 // Verify we have a legal type before going any further. Currently, we handle
1584 // simple types that will directly fit in a register (i32/f32/i64/f64) or
1585 // those that can be sign or zero-extended to a basic operation (i1/i8/i16).
1586 if (!isLoadStoreTypeLegal(Op0->getType(), VT) ||
1587 cast<StoreInst>(I)->isAtomic())
1590 // Get the value to be stored into a register. Use the zero register directly
1591 // when possible to avoid an unnecessary copy and a wasted register.
1592 unsigned SrcReg = 0;
1593 if (const auto *CI = dyn_cast<ConstantInt>(Op0)) {
1595 SrcReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
1596 } else if (const auto *CF = dyn_cast<ConstantFP>(Op0)) {
1597 if (CF->isZero() && !CF->isNegative()) {
1598 VT = MVT::getIntegerVT(VT.getSizeInBits());
1599 SrcReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
1604 SrcReg = getRegForValue(Op0);
1609 // See if we can handle this address.
1611 if (!ComputeAddress(I->getOperand(1), Addr, I->getOperand(0)->getType()))
1614 if (!EmitStore(VT, SrcReg, Addr, createMachineMemOperandFor(I)))
1619 static AArch64CC::CondCode getCompareCC(CmpInst::Predicate Pred) {
1621 case CmpInst::FCMP_ONE:
1622 case CmpInst::FCMP_UEQ:
1624 // AL is our "false" for now. The other two need more compares.
1625 return AArch64CC::AL;
1626 case CmpInst::ICMP_EQ:
1627 case CmpInst::FCMP_OEQ:
1628 return AArch64CC::EQ;
1629 case CmpInst::ICMP_SGT:
1630 case CmpInst::FCMP_OGT:
1631 return AArch64CC::GT;
1632 case CmpInst::ICMP_SGE:
1633 case CmpInst::FCMP_OGE:
1634 return AArch64CC::GE;
1635 case CmpInst::ICMP_UGT:
1636 case CmpInst::FCMP_UGT:
1637 return AArch64CC::HI;
1638 case CmpInst::FCMP_OLT:
1639 return AArch64CC::MI;
1640 case CmpInst::ICMP_ULE:
1641 case CmpInst::FCMP_OLE:
1642 return AArch64CC::LS;
1643 case CmpInst::FCMP_ORD:
1644 return AArch64CC::VC;
1645 case CmpInst::FCMP_UNO:
1646 return AArch64CC::VS;
1647 case CmpInst::FCMP_UGE:
1648 return AArch64CC::PL;
1649 case CmpInst::ICMP_SLT:
1650 case CmpInst::FCMP_ULT:
1651 return AArch64CC::LT;
1652 case CmpInst::ICMP_SLE:
1653 case CmpInst::FCMP_ULE:
1654 return AArch64CC::LE;
1655 case CmpInst::FCMP_UNE:
1656 case CmpInst::ICMP_NE:
1657 return AArch64CC::NE;
1658 case CmpInst::ICMP_UGE:
1659 return AArch64CC::HS;
1660 case CmpInst::ICMP_ULT:
1661 return AArch64CC::LO;
1665 bool AArch64FastISel::SelectBranch(const Instruction *I) {
1666 const BranchInst *BI = cast<BranchInst>(I);
1667 if (BI->isUnconditional()) {
1668 MachineBasicBlock *MSucc = FuncInfo.MBBMap[BI->getSuccessor(0)];
1669 fastEmitBranch(MSucc, BI->getDebugLoc());
1673 MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
1674 MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
1676 AArch64CC::CondCode CC = AArch64CC::NE;
1677 if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
1678 if (CI->hasOneUse() && (CI->getParent() == I->getParent())) {
1679 // We may not handle every CC for now.
1680 CC = getCompareCC(CI->getPredicate());
1681 if (CC == AArch64CC::AL)
1685 if (!emitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
1689 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
1693 // Obtain the branch weight and add the TrueBB to the successor list.
1694 uint32_t BranchWeight = 0;
1696 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
1697 TBB->getBasicBlock());
1698 FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
1700 fastEmitBranch(FBB, DbgLoc);
1703 } else if (TruncInst *TI = dyn_cast<TruncInst>(BI->getCondition())) {
1705 if (TI->hasOneUse() && TI->getParent() == I->getParent() &&
1706 (isTypeSupported(TI->getOperand(0)->getType(), SrcVT))) {
1707 unsigned CondReg = getRegForValue(TI->getOperand(0));
1710 bool CondIsKill = hasTrivialKill(TI->getOperand(0));
1712 // Issue an extract_subreg to get the lower 32-bits.
1713 if (SrcVT == MVT::i64) {
1714 CondReg = fastEmitInst_extractsubreg(MVT::i32, CondReg, CondIsKill,
1719 unsigned ANDReg = emitAnd_ri(MVT::i32, CondReg, CondIsKill, 1);
1720 assert(ANDReg && "Unexpected AND instruction emission failure.");
1721 emitICmp_ri(MVT::i32, ANDReg, /*IsKill=*/true, 0);
1723 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
1724 std::swap(TBB, FBB);
1727 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
1731 // Obtain the branch weight and add the TrueBB to the successor list.
1732 uint32_t BranchWeight = 0;
1734 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
1735 TBB->getBasicBlock());
1736 FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
1738 fastEmitBranch(FBB, DbgLoc);
1741 } else if (const ConstantInt *CI =
1742 dyn_cast<ConstantInt>(BI->getCondition())) {
1743 uint64_t Imm = CI->getZExtValue();
1744 MachineBasicBlock *Target = (Imm == 0) ? FBB : TBB;
1745 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::B))
1748 // Obtain the branch weight and add the target to the successor list.
1749 uint32_t BranchWeight = 0;
1751 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
1752 Target->getBasicBlock());
1753 FuncInfo.MBB->addSuccessor(Target, BranchWeight);
1755 } else if (foldXALUIntrinsic(CC, I, BI->getCondition())) {
1756 // Fake request the condition, otherwise the intrinsic might be completely
1758 unsigned CondReg = getRegForValue(BI->getCondition());
1763 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
1767 // Obtain the branch weight and add the TrueBB to the successor list.
1768 uint32_t BranchWeight = 0;
1770 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
1771 TBB->getBasicBlock());
1772 FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
1774 fastEmitBranch(FBB, DbgLoc);
1778 unsigned CondReg = getRegForValue(BI->getCondition());
1781 bool CondRegIsKill = hasTrivialKill(BI->getCondition());
1783 // We've been divorced from our compare! Our block was split, and
1784 // now our compare lives in a predecessor block. We musn't
1785 // re-compare here, as the children of the compare aren't guaranteed
1786 // live across the block boundary (we *could* check for this).
1787 // Regardless, the compare has been done in the predecessor block,
1788 // and it left a value for us in a virtual register. Ergo, we test
1789 // the one-bit value left in the virtual register.
1790 emitICmp_ri(MVT::i32, CondReg, CondRegIsKill, 0);
1792 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
1793 std::swap(TBB, FBB);
1797 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
1801 // Obtain the branch weight and add the TrueBB to the successor list.
1802 uint32_t BranchWeight = 0;
1804 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
1805 TBB->getBasicBlock());
1806 FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
1808 fastEmitBranch(FBB, DbgLoc);
1812 bool AArch64FastISel::SelectIndirectBr(const Instruction *I) {
1813 const IndirectBrInst *BI = cast<IndirectBrInst>(I);
1814 unsigned AddrReg = getRegForValue(BI->getOperand(0));
1818 // Emit the indirect branch.
1819 const MCInstrDesc &II = TII.get(AArch64::BR);
1820 AddrReg = constrainOperandRegClass(II, AddrReg, II.getNumDefs());
1821 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II).addReg(AddrReg);
1823 // Make sure the CFG is up-to-date.
1824 for (unsigned i = 0, e = BI->getNumSuccessors(); i != e; ++i)
1825 FuncInfo.MBB->addSuccessor(FuncInfo.MBBMap[BI->getSuccessor(i)]);
1830 bool AArch64FastISel::SelectCmp(const Instruction *I) {
1831 const CmpInst *CI = cast<CmpInst>(I);
1833 // We may not handle every CC for now.
1834 AArch64CC::CondCode CC = getCompareCC(CI->getPredicate());
1835 if (CC == AArch64CC::AL)
1839 if (!emitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
1842 // Now set a register based on the comparison.
1843 AArch64CC::CondCode invertedCC = getInvertedCondCode(CC);
1844 unsigned ResultReg = createResultReg(&AArch64::GPR32RegClass);
1845 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::CSINCWr),
1847 .addReg(AArch64::WZR)
1848 .addReg(AArch64::WZR)
1849 .addImm(invertedCC);
1851 updateValueMap(I, ResultReg);
1855 bool AArch64FastISel::SelectSelect(const Instruction *I) {
1856 const SelectInst *SI = cast<SelectInst>(I);
1858 EVT DestEVT = TLI.getValueType(SI->getType(), true);
1859 if (!DestEVT.isSimple())
1862 MVT DestVT = DestEVT.getSimpleVT();
1863 if (DestVT != MVT::i32 && DestVT != MVT::i64 && DestVT != MVT::f32 &&
1868 const TargetRegisterClass *RC = nullptr;
1869 switch (DestVT.SimpleTy) {
1870 default: return false;
1872 SelectOpc = AArch64::CSELWr; RC = &AArch64::GPR32RegClass; break;
1874 SelectOpc = AArch64::CSELXr; RC = &AArch64::GPR64RegClass; break;
1876 SelectOpc = AArch64::FCSELSrrr; RC = &AArch64::FPR32RegClass; break;
1878 SelectOpc = AArch64::FCSELDrrr; RC = &AArch64::FPR64RegClass; break;
1881 const Value *Cond = SI->getCondition();
1882 bool NeedTest = true;
1883 AArch64CC::CondCode CC = AArch64CC::NE;
1884 if (foldXALUIntrinsic(CC, I, Cond))
1887 unsigned CondReg = getRegForValue(Cond);
1890 bool CondIsKill = hasTrivialKill(Cond);
1893 unsigned ANDReg = emitAnd_ri(MVT::i32, CondReg, CondIsKill, 1);
1894 assert(ANDReg && "Unexpected AND instruction emission failure.");
1895 emitICmp_ri(MVT::i32, ANDReg, /*IsKill=*/true, 0);
1898 unsigned TrueReg = getRegForValue(SI->getTrueValue());
1899 bool TrueIsKill = hasTrivialKill(SI->getTrueValue());
1901 unsigned FalseReg = getRegForValue(SI->getFalseValue());
1902 bool FalseIsKill = hasTrivialKill(SI->getFalseValue());
1904 if (!TrueReg || !FalseReg)
1907 unsigned ResultReg = fastEmitInst_rri(SelectOpc, RC, TrueReg, TrueIsKill,
1908 FalseReg, FalseIsKill, CC);
1909 updateValueMap(I, ResultReg);
1913 bool AArch64FastISel::SelectFPExt(const Instruction *I) {
1914 Value *V = I->getOperand(0);
1915 if (!I->getType()->isDoubleTy() || !V->getType()->isFloatTy())
1918 unsigned Op = getRegForValue(V);
1922 unsigned ResultReg = createResultReg(&AArch64::FPR64RegClass);
1923 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::FCVTDSr),
1924 ResultReg).addReg(Op);
1925 updateValueMap(I, ResultReg);
1929 bool AArch64FastISel::SelectFPTrunc(const Instruction *I) {
1930 Value *V = I->getOperand(0);
1931 if (!I->getType()->isFloatTy() || !V->getType()->isDoubleTy())
1934 unsigned Op = getRegForValue(V);
1938 unsigned ResultReg = createResultReg(&AArch64::FPR32RegClass);
1939 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::FCVTSDr),
1940 ResultReg).addReg(Op);
1941 updateValueMap(I, ResultReg);
1945 // FPToUI and FPToSI
1946 bool AArch64FastISel::SelectFPToInt(const Instruction *I, bool Signed) {
1948 if (!isTypeLegal(I->getType(), DestVT) || DestVT.isVector())
1951 unsigned SrcReg = getRegForValue(I->getOperand(0));
1955 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType(), true);
1956 if (SrcVT == MVT::f128)
1960 if (SrcVT == MVT::f64) {
1962 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZSUWDr : AArch64::FCVTZSUXDr;
1964 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZUUWDr : AArch64::FCVTZUUXDr;
1967 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZSUWSr : AArch64::FCVTZSUXSr;
1969 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZUUWSr : AArch64::FCVTZUUXSr;
1971 unsigned ResultReg = createResultReg(
1972 DestVT == MVT::i32 ? &AArch64::GPR32RegClass : &AArch64::GPR64RegClass);
1973 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
1975 updateValueMap(I, ResultReg);
1979 bool AArch64FastISel::SelectIntToFP(const Instruction *I, bool Signed) {
1981 if (!isTypeLegal(I->getType(), DestVT) || DestVT.isVector())
1983 assert ((DestVT == MVT::f32 || DestVT == MVT::f64) &&
1984 "Unexpected value type.");
1986 unsigned SrcReg = getRegForValue(I->getOperand(0));
1989 bool SrcIsKill = hasTrivialKill(I->getOperand(0));
1991 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType(), true);
1993 // Handle sign-extension.
1994 if (SrcVT == MVT::i16 || SrcVT == MVT::i8 || SrcVT == MVT::i1) {
1996 EmitIntExt(SrcVT.getSimpleVT(), SrcReg, MVT::i32, /*isZExt*/ !Signed);
2003 if (SrcVT == MVT::i64) {
2005 Opc = (DestVT == MVT::f32) ? AArch64::SCVTFUXSri : AArch64::SCVTFUXDri;
2007 Opc = (DestVT == MVT::f32) ? AArch64::UCVTFUXSri : AArch64::UCVTFUXDri;
2010 Opc = (DestVT == MVT::f32) ? AArch64::SCVTFUWSri : AArch64::SCVTFUWDri;
2012 Opc = (DestVT == MVT::f32) ? AArch64::UCVTFUWSri : AArch64::UCVTFUWDri;
2015 unsigned ResultReg = fastEmitInst_r(Opc, TLI.getRegClassFor(DestVT), SrcReg,
2017 updateValueMap(I, ResultReg);
2021 bool AArch64FastISel::fastLowerArguments() {
2022 if (!FuncInfo.CanLowerReturn)
2025 const Function *F = FuncInfo.Fn;
2029 CallingConv::ID CC = F->getCallingConv();
2030 if (CC != CallingConv::C)
2033 // Only handle simple cases like i1/i8/i16/i32/i64/f32/f64 of up to 8 GPR and
2035 unsigned GPRCnt = 0;
2036 unsigned FPRCnt = 0;
2038 for (auto const &Arg : F->args()) {
2039 // The first argument is at index 1.
2041 if (F->getAttributes().hasAttribute(Idx, Attribute::ByVal) ||
2042 F->getAttributes().hasAttribute(Idx, Attribute::InReg) ||
2043 F->getAttributes().hasAttribute(Idx, Attribute::StructRet) ||
2044 F->getAttributes().hasAttribute(Idx, Attribute::Nest))
2047 Type *ArgTy = Arg.getType();
2048 if (ArgTy->isStructTy() || ArgTy->isArrayTy() || ArgTy->isVectorTy())
2051 EVT ArgVT = TLI.getValueType(ArgTy);
2052 if (!ArgVT.isSimple()) return false;
2053 switch (ArgVT.getSimpleVT().SimpleTy) {
2054 default: return false;
2069 if (GPRCnt > 8 || FPRCnt > 8)
2073 static const MCPhysReg Registers[5][8] = {
2074 { AArch64::W0, AArch64::W1, AArch64::W2, AArch64::W3, AArch64::W4,
2075 AArch64::W5, AArch64::W6, AArch64::W7 },
2076 { AArch64::X0, AArch64::X1, AArch64::X2, AArch64::X3, AArch64::X4,
2077 AArch64::X5, AArch64::X6, AArch64::X7 },
2078 { AArch64::H0, AArch64::H1, AArch64::H2, AArch64::H3, AArch64::H4,
2079 AArch64::H5, AArch64::H6, AArch64::H7 },
2080 { AArch64::S0, AArch64::S1, AArch64::S2, AArch64::S3, AArch64::S4,
2081 AArch64::S5, AArch64::S6, AArch64::S7 },
2082 { AArch64::D0, AArch64::D1, AArch64::D2, AArch64::D3, AArch64::D4,
2083 AArch64::D5, AArch64::D6, AArch64::D7 }
2086 unsigned GPRIdx = 0;
2087 unsigned FPRIdx = 0;
2088 for (auto const &Arg : F->args()) {
2089 MVT VT = TLI.getSimpleValueType(Arg.getType());
2091 const TargetRegisterClass *RC = nullptr;
2092 switch (VT.SimpleTy) {
2093 default: llvm_unreachable("Unexpected value type.");
2096 case MVT::i16: VT = MVT::i32; // fall-through
2098 SrcReg = Registers[0][GPRIdx++]; RC = &AArch64::GPR32RegClass; break;
2100 SrcReg = Registers[1][GPRIdx++]; RC = &AArch64::GPR64RegClass; break;
2102 SrcReg = Registers[2][FPRIdx++]; RC = &AArch64::FPR16RegClass; break;
2104 SrcReg = Registers[3][FPRIdx++]; RC = &AArch64::FPR32RegClass; break;
2106 SrcReg = Registers[4][FPRIdx++]; RC = &AArch64::FPR64RegClass; break;
2109 // Skip unused arguments.
2110 if (Arg.use_empty()) {
2111 updateValueMap(&Arg, 0);
2115 unsigned DstReg = FuncInfo.MF->addLiveIn(SrcReg, RC);
2116 // FIXME: Unfortunately it's necessary to emit a copy from the livein copy.
2117 // Without this, EmitLiveInCopies may eliminate the livein if its only
2118 // use is a bitcast (which isn't turned into an instruction).
2119 unsigned ResultReg = createResultReg(RC);
2120 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2121 TII.get(TargetOpcode::COPY), ResultReg)
2122 .addReg(DstReg, getKillRegState(true));
2123 updateValueMap(&Arg, ResultReg);
2128 bool AArch64FastISel::ProcessCallArgs(CallLoweringInfo &CLI,
2129 SmallVectorImpl<MVT> &OutVTs,
2130 unsigned &NumBytes) {
2131 CallingConv::ID CC = CLI.CallConv;
2132 SmallVector<CCValAssign, 16> ArgLocs;
2133 CCState CCInfo(CC, false, *FuncInfo.MF, ArgLocs, *Context);
2134 CCInfo.AnalyzeCallOperands(OutVTs, CLI.OutFlags, CCAssignFnForCall(CC));
2136 // Get a count of how many bytes are to be pushed on the stack.
2137 NumBytes = CCInfo.getNextStackOffset();
2139 // Issue CALLSEQ_START
2140 unsigned AdjStackDown = TII.getCallFrameSetupOpcode();
2141 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackDown))
2144 // Process the args.
2145 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2146 CCValAssign &VA = ArgLocs[i];
2147 const Value *ArgVal = CLI.OutVals[VA.getValNo()];
2148 MVT ArgVT = OutVTs[VA.getValNo()];
2150 unsigned ArgReg = getRegForValue(ArgVal);
2154 // Handle arg promotion: SExt, ZExt, AExt.
2155 switch (VA.getLocInfo()) {
2156 case CCValAssign::Full:
2158 case CCValAssign::SExt: {
2159 MVT DestVT = VA.getLocVT();
2161 ArgReg = EmitIntExt(SrcVT, ArgReg, DestVT, /*isZExt=*/false);
2166 case CCValAssign::AExt:
2167 // Intentional fall-through.
2168 case CCValAssign::ZExt: {
2169 MVT DestVT = VA.getLocVT();
2171 ArgReg = EmitIntExt(SrcVT, ArgReg, DestVT, /*isZExt=*/true);
2177 llvm_unreachable("Unknown arg promotion!");
2180 // Now copy/store arg to correct locations.
2181 if (VA.isRegLoc() && !VA.needsCustom()) {
2182 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2183 TII.get(TargetOpcode::COPY), VA.getLocReg()).addReg(ArgReg);
2184 CLI.OutRegs.push_back(VA.getLocReg());
2185 } else if (VA.needsCustom()) {
2186 // FIXME: Handle custom args.
2189 assert(VA.isMemLoc() && "Assuming store on stack.");
2191 // Don't emit stores for undef values.
2192 if (isa<UndefValue>(ArgVal))
2195 // Need to store on the stack.
2196 unsigned ArgSize = (ArgVT.getSizeInBits() + 7) / 8;
2198 unsigned BEAlign = 0;
2199 if (ArgSize < 8 && !Subtarget->isLittleEndian())
2200 BEAlign = 8 - ArgSize;
2203 Addr.setKind(Address::RegBase);
2204 Addr.setReg(AArch64::SP);
2205 Addr.setOffset(VA.getLocMemOffset() + BEAlign);
2207 unsigned Alignment = DL.getABITypeAlignment(ArgVal->getType());
2208 MachineMemOperand *MMO = FuncInfo.MF->getMachineMemOperand(
2209 MachinePointerInfo::getStack(Addr.getOffset()),
2210 MachineMemOperand::MOStore, ArgVT.getStoreSize(), Alignment);
2212 if (!EmitStore(ArgVT, ArgReg, Addr, MMO))
2219 bool AArch64FastISel::FinishCall(CallLoweringInfo &CLI, MVT RetVT,
2220 unsigned NumBytes) {
2221 CallingConv::ID CC = CLI.CallConv;
2223 // Issue CALLSEQ_END
2224 unsigned AdjStackUp = TII.getCallFrameDestroyOpcode();
2225 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackUp))
2226 .addImm(NumBytes).addImm(0);
2228 // Now the return value.
2229 if (RetVT != MVT::isVoid) {
2230 SmallVector<CCValAssign, 16> RVLocs;
2231 CCState CCInfo(CC, false, *FuncInfo.MF, RVLocs, *Context);
2232 CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC));
2234 // Only handle a single return value.
2235 if (RVLocs.size() != 1)
2238 // Copy all of the result registers out of their specified physreg.
2239 MVT CopyVT = RVLocs[0].getValVT();
2240 unsigned ResultReg = createResultReg(TLI.getRegClassFor(CopyVT));
2241 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2242 TII.get(TargetOpcode::COPY), ResultReg)
2243 .addReg(RVLocs[0].getLocReg());
2244 CLI.InRegs.push_back(RVLocs[0].getLocReg());
2246 CLI.ResultReg = ResultReg;
2247 CLI.NumResultRegs = 1;
2253 bool AArch64FastISel::fastLowerCall(CallLoweringInfo &CLI) {
2254 CallingConv::ID CC = CLI.CallConv;
2255 bool IsTailCall = CLI.IsTailCall;
2256 bool IsVarArg = CLI.IsVarArg;
2257 const Value *Callee = CLI.Callee;
2258 const char *SymName = CLI.SymName;
2260 // Allow SelectionDAG isel to handle tail calls.
2264 CodeModel::Model CM = TM.getCodeModel();
2265 // Only support the small and large code model.
2266 if (CM != CodeModel::Small && CM != CodeModel::Large)
2269 // FIXME: Add large code model support for ELF.
2270 if (CM == CodeModel::Large && !Subtarget->isTargetMachO())
2273 // Let SDISel handle vararg functions.
2277 // FIXME: Only handle *simple* calls for now.
2279 if (CLI.RetTy->isVoidTy())
2280 RetVT = MVT::isVoid;
2281 else if (!isTypeLegal(CLI.RetTy, RetVT))
2284 for (auto Flag : CLI.OutFlags)
2285 if (Flag.isInReg() || Flag.isSRet() || Flag.isNest() || Flag.isByVal())
2288 // Set up the argument vectors.
2289 SmallVector<MVT, 16> OutVTs;
2290 OutVTs.reserve(CLI.OutVals.size());
2292 for (auto *Val : CLI.OutVals) {
2294 if (!isTypeLegal(Val->getType(), VT) &&
2295 !(VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16))
2298 // We don't handle vector parameters yet.
2299 if (VT.isVector() || VT.getSizeInBits() > 64)
2302 OutVTs.push_back(VT);
2306 if (!ComputeCallAddress(Callee, Addr))
2309 // Handle the arguments now that we've gotten them.
2311 if (!ProcessCallArgs(CLI, OutVTs, NumBytes))
2315 MachineInstrBuilder MIB;
2316 if (CM == CodeModel::Small) {
2317 const MCInstrDesc &II = TII.get(Addr.getReg() ? AArch64::BLR : AArch64::BL);
2318 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II);
2320 MIB.addExternalSymbol(SymName, 0);
2321 else if (Addr.getGlobalValue())
2322 MIB.addGlobalAddress(Addr.getGlobalValue(), 0, 0);
2323 else if (Addr.getReg()) {
2324 unsigned Reg = constrainOperandRegClass(II, Addr.getReg(), 0);
2329 unsigned CallReg = 0;
2331 unsigned ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
2332 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
2334 .addExternalSymbol(SymName, AArch64II::MO_GOT | AArch64II::MO_PAGE);
2336 CallReg = createResultReg(&AArch64::GPR64RegClass);
2337 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::LDRXui),
2340 .addExternalSymbol(SymName, AArch64II::MO_GOT | AArch64II::MO_PAGEOFF |
2342 } else if (Addr.getGlobalValue()) {
2343 CallReg = AArch64MaterializeGV(Addr.getGlobalValue());
2344 } else if (Addr.getReg())
2345 CallReg = Addr.getReg();
2350 const MCInstrDesc &II = TII.get(AArch64::BLR);
2351 CallReg = constrainOperandRegClass(II, CallReg, 0);
2352 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II).addReg(CallReg);
2355 // Add implicit physical register uses to the call.
2356 for (auto Reg : CLI.OutRegs)
2357 MIB.addReg(Reg, RegState::Implicit);
2359 // Add a register mask with the call-preserved registers.
2360 // Proper defs for return values will be added by setPhysRegsDeadExcept().
2361 MIB.addRegMask(TRI.getCallPreservedMask(CC));
2365 // Finish off the call including any return values.
2366 return FinishCall(CLI, RetVT, NumBytes);
2369 bool AArch64FastISel::IsMemCpySmall(uint64_t Len, unsigned Alignment) {
2371 return Len / Alignment <= 4;
2376 bool AArch64FastISel::TryEmitSmallMemCpy(Address Dest, Address Src,
2377 uint64_t Len, unsigned Alignment) {
2378 // Make sure we don't bloat code by inlining very large memcpy's.
2379 if (!IsMemCpySmall(Len, Alignment))
2382 int64_t UnscaledOffset = 0;
2383 Address OrigDest = Dest;
2384 Address OrigSrc = Src;
2388 if (!Alignment || Alignment >= 8) {
2399 // Bound based on alignment.
2400 if (Len >= 4 && Alignment == 4)
2402 else if (Len >= 2 && Alignment == 2)
2411 RV = EmitLoad(VT, ResultReg, Src);
2415 RV = EmitStore(VT, ResultReg, Dest);
2419 int64_t Size = VT.getSizeInBits() / 8;
2421 UnscaledOffset += Size;
2423 // We need to recompute the unscaled offset for each iteration.
2424 Dest.setOffset(OrigDest.getOffset() + UnscaledOffset);
2425 Src.setOffset(OrigSrc.getOffset() + UnscaledOffset);
2431 /// \brief Check if it is possible to fold the condition from the XALU intrinsic
2432 /// into the user. The condition code will only be updated on success.
2433 bool AArch64FastISel::foldXALUIntrinsic(AArch64CC::CondCode &CC,
2434 const Instruction *I,
2435 const Value *Cond) {
2436 if (!isa<ExtractValueInst>(Cond))
2439 const auto *EV = cast<ExtractValueInst>(Cond);
2440 if (!isa<IntrinsicInst>(EV->getAggregateOperand()))
2443 const auto *II = cast<IntrinsicInst>(EV->getAggregateOperand());
2445 const Function *Callee = II->getCalledFunction();
2447 cast<StructType>(Callee->getReturnType())->getTypeAtIndex(0U);
2448 if (!isTypeLegal(RetTy, RetVT))
2451 if (RetVT != MVT::i32 && RetVT != MVT::i64)
2454 AArch64CC::CondCode TmpCC;
2455 switch (II->getIntrinsicID()) {
2456 default: return false;
2457 case Intrinsic::sadd_with_overflow:
2458 case Intrinsic::ssub_with_overflow: TmpCC = AArch64CC::VS; break;
2459 case Intrinsic::uadd_with_overflow: TmpCC = AArch64CC::HS; break;
2460 case Intrinsic::usub_with_overflow: TmpCC = AArch64CC::LO; break;
2461 case Intrinsic::smul_with_overflow:
2462 case Intrinsic::umul_with_overflow: TmpCC = AArch64CC::NE; break;
2465 // Check if both instructions are in the same basic block.
2466 if (II->getParent() != I->getParent())
2469 // Make sure nothing is in the way
2470 BasicBlock::const_iterator Start = I;
2471 BasicBlock::const_iterator End = II;
2472 for (auto Itr = std::prev(Start); Itr != End; --Itr) {
2473 // We only expect extractvalue instructions between the intrinsic and the
2474 // instruction to be selected.
2475 if (!isa<ExtractValueInst>(Itr))
2478 // Check that the extractvalue operand comes from the intrinsic.
2479 const auto *EVI = cast<ExtractValueInst>(Itr);
2480 if (EVI->getAggregateOperand() != II)
2488 bool AArch64FastISel::fastLowerIntrinsicCall(const IntrinsicInst *II) {
2489 // FIXME: Handle more intrinsics.
2490 switch (II->getIntrinsicID()) {
2491 default: return false;
2492 case Intrinsic::frameaddress: {
2493 MachineFrameInfo *MFI = FuncInfo.MF->getFrameInfo();
2494 MFI->setFrameAddressIsTaken(true);
2496 const AArch64RegisterInfo *RegInfo =
2497 static_cast<const AArch64RegisterInfo *>(
2498 TM.getSubtargetImpl()->getRegisterInfo());
2499 unsigned FramePtr = RegInfo->getFrameRegister(*(FuncInfo.MF));
2500 unsigned SrcReg = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
2501 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2502 TII.get(TargetOpcode::COPY), SrcReg).addReg(FramePtr);
2503 // Recursively load frame address
2509 unsigned Depth = cast<ConstantInt>(II->getOperand(0))->getZExtValue();
2511 DestReg = fastEmitInst_ri(AArch64::LDRXui, &AArch64::GPR64RegClass,
2512 SrcReg, /*IsKill=*/true, 0);
2513 assert(DestReg && "Unexpected LDR instruction emission failure.");
2517 updateValueMap(II, SrcReg);
2520 case Intrinsic::memcpy:
2521 case Intrinsic::memmove: {
2522 const auto *MTI = cast<MemTransferInst>(II);
2523 // Don't handle volatile.
2524 if (MTI->isVolatile())
2527 // Disable inlining for memmove before calls to ComputeAddress. Otherwise,
2528 // we would emit dead code because we don't currently handle memmoves.
2529 bool IsMemCpy = (II->getIntrinsicID() == Intrinsic::memcpy);
2530 if (isa<ConstantInt>(MTI->getLength()) && IsMemCpy) {
2531 // Small memcpy's are common enough that we want to do them without a call
2533 uint64_t Len = cast<ConstantInt>(MTI->getLength())->getZExtValue();
2534 unsigned Alignment = MTI->getAlignment();
2535 if (IsMemCpySmall(Len, Alignment)) {
2537 if (!ComputeAddress(MTI->getRawDest(), Dest) ||
2538 !ComputeAddress(MTI->getRawSource(), Src))
2540 if (TryEmitSmallMemCpy(Dest, Src, Len, Alignment))
2545 if (!MTI->getLength()->getType()->isIntegerTy(64))
2548 if (MTI->getSourceAddressSpace() > 255 || MTI->getDestAddressSpace() > 255)
2549 // Fast instruction selection doesn't support the special
2553 const char *IntrMemName = isa<MemCpyInst>(II) ? "memcpy" : "memmove";
2554 return lowerCallTo(II, IntrMemName, II->getNumArgOperands() - 2);
2556 case Intrinsic::memset: {
2557 const MemSetInst *MSI = cast<MemSetInst>(II);
2558 // Don't handle volatile.
2559 if (MSI->isVolatile())
2562 if (!MSI->getLength()->getType()->isIntegerTy(64))
2565 if (MSI->getDestAddressSpace() > 255)
2566 // Fast instruction selection doesn't support the special
2570 return lowerCallTo(II, "memset", II->getNumArgOperands() - 2);
2572 case Intrinsic::trap: {
2573 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::BRK))
2577 case Intrinsic::sqrt: {
2578 Type *RetTy = II->getCalledFunction()->getReturnType();
2581 if (!isTypeLegal(RetTy, VT))
2584 unsigned Op0Reg = getRegForValue(II->getOperand(0));
2587 bool Op0IsKill = hasTrivialKill(II->getOperand(0));
2589 unsigned ResultReg = fastEmit_r(VT, VT, ISD::FSQRT, Op0Reg, Op0IsKill);
2593 updateValueMap(II, ResultReg);
2596 case Intrinsic::sadd_with_overflow:
2597 case Intrinsic::uadd_with_overflow:
2598 case Intrinsic::ssub_with_overflow:
2599 case Intrinsic::usub_with_overflow:
2600 case Intrinsic::smul_with_overflow:
2601 case Intrinsic::umul_with_overflow: {
2602 // This implements the basic lowering of the xalu with overflow intrinsics.
2603 const Function *Callee = II->getCalledFunction();
2604 auto *Ty = cast<StructType>(Callee->getReturnType());
2605 Type *RetTy = Ty->getTypeAtIndex(0U);
2608 if (!isTypeLegal(RetTy, VT))
2611 if (VT != MVT::i32 && VT != MVT::i64)
2614 const Value *LHS = II->getArgOperand(0);
2615 const Value *RHS = II->getArgOperand(1);
2616 // Canonicalize immediate to the RHS.
2617 if (isa<ConstantInt>(LHS) && !isa<ConstantInt>(RHS) &&
2618 isCommutativeIntrinsic(II))
2619 std::swap(LHS, RHS);
2621 unsigned ResultReg1 = 0, ResultReg2 = 0, MulReg = 0;
2622 AArch64CC::CondCode CC = AArch64CC::Invalid;
2623 switch (II->getIntrinsicID()) {
2624 default: llvm_unreachable("Unexpected intrinsic!");
2625 case Intrinsic::sadd_with_overflow:
2626 ResultReg1 = emitAdd(VT, LHS, RHS, /*SetFlags=*/true);
2629 case Intrinsic::uadd_with_overflow:
2630 ResultReg1 = emitAdd(VT, LHS, RHS, /*SetFlags=*/true);
2633 case Intrinsic::ssub_with_overflow:
2634 ResultReg1 = emitSub(VT, LHS, RHS, /*SetFlags=*/true);
2637 case Intrinsic::usub_with_overflow:
2638 ResultReg1 = emitSub(VT, LHS, RHS, /*SetFlags=*/true);
2641 case Intrinsic::smul_with_overflow: {
2643 unsigned LHSReg = getRegForValue(LHS);
2646 bool LHSIsKill = hasTrivialKill(LHS);
2648 unsigned RHSReg = getRegForValue(RHS);
2651 bool RHSIsKill = hasTrivialKill(RHS);
2653 if (VT == MVT::i32) {
2654 MulReg = Emit_SMULL_rr(MVT::i64, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
2655 unsigned ShiftReg = emitLSR_ri(MVT::i64, MVT::i64, MulReg,
2656 /*IsKill=*/false, 32);
2657 MulReg = fastEmitInst_extractsubreg(VT, MulReg, /*IsKill=*/true,
2659 ShiftReg = fastEmitInst_extractsubreg(VT, ShiftReg, /*IsKill=*/true,
2661 emitSubs_rs(VT, ShiftReg, /*IsKill=*/true, MulReg, /*IsKill=*/false,
2662 AArch64_AM::ASR, 31, /*WantResult=*/false);
2664 assert(VT == MVT::i64 && "Unexpected value type.");
2665 MulReg = Emit_MUL_rr(VT, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
2666 unsigned SMULHReg = fastEmit_rr(VT, VT, ISD::MULHS, LHSReg, LHSIsKill,
2668 emitSubs_rs(VT, SMULHReg, /*IsKill=*/true, MulReg, /*IsKill=*/false,
2669 AArch64_AM::ASR, 63, /*WantResult=*/false);
2673 case Intrinsic::umul_with_overflow: {
2675 unsigned LHSReg = getRegForValue(LHS);
2678 bool LHSIsKill = hasTrivialKill(LHS);
2680 unsigned RHSReg = getRegForValue(RHS);
2683 bool RHSIsKill = hasTrivialKill(RHS);
2685 if (VT == MVT::i32) {
2686 MulReg = Emit_UMULL_rr(MVT::i64, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
2687 emitSubs_rs(MVT::i64, AArch64::XZR, /*IsKill=*/true, MulReg,
2688 /*IsKill=*/false, AArch64_AM::LSR, 32,
2689 /*WantResult=*/false);
2690 MulReg = fastEmitInst_extractsubreg(VT, MulReg, /*IsKill=*/true,
2693 assert(VT == MVT::i64 && "Unexpected value type.");
2694 MulReg = Emit_MUL_rr(VT, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
2695 unsigned UMULHReg = fastEmit_rr(VT, VT, ISD::MULHU, LHSReg, LHSIsKill,
2697 emitSubs_rr(VT, AArch64::XZR, /*IsKill=*/true, UMULHReg,
2698 /*IsKill=*/false, /*WantResult=*/false);
2705 ResultReg1 = createResultReg(TLI.getRegClassFor(VT));
2706 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2707 TII.get(TargetOpcode::COPY), ResultReg1).addReg(MulReg);
2710 ResultReg2 = fastEmitInst_rri(AArch64::CSINCWr, &AArch64::GPR32RegClass,
2711 AArch64::WZR, /*IsKill=*/true, AArch64::WZR,
2712 /*IsKill=*/true, getInvertedCondCode(CC));
2713 assert((ResultReg1 + 1) == ResultReg2 &&
2714 "Nonconsecutive result registers.");
2715 updateValueMap(II, ResultReg1, 2);
2722 bool AArch64FastISel::SelectRet(const Instruction *I) {
2723 const ReturnInst *Ret = cast<ReturnInst>(I);
2724 const Function &F = *I->getParent()->getParent();
2726 if (!FuncInfo.CanLowerReturn)
2732 // Build a list of return value registers.
2733 SmallVector<unsigned, 4> RetRegs;
2735 if (Ret->getNumOperands() > 0) {
2736 CallingConv::ID CC = F.getCallingConv();
2737 SmallVector<ISD::OutputArg, 4> Outs;
2738 GetReturnInfo(F.getReturnType(), F.getAttributes(), Outs, TLI);
2740 // Analyze operands of the call, assigning locations to each operand.
2741 SmallVector<CCValAssign, 16> ValLocs;
2742 CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, ValLocs, I->getContext());
2743 CCAssignFn *RetCC = CC == CallingConv::WebKit_JS ? RetCC_AArch64_WebKit_JS
2744 : RetCC_AArch64_AAPCS;
2745 CCInfo.AnalyzeReturn(Outs, RetCC);
2747 // Only handle a single return value for now.
2748 if (ValLocs.size() != 1)
2751 CCValAssign &VA = ValLocs[0];
2752 const Value *RV = Ret->getOperand(0);
2754 // Don't bother handling odd stuff for now.
2755 if (VA.getLocInfo() != CCValAssign::Full)
2757 // Only handle register returns for now.
2760 unsigned Reg = getRegForValue(RV);
2764 unsigned SrcReg = Reg + VA.getValNo();
2765 unsigned DestReg = VA.getLocReg();
2766 // Avoid a cross-class copy. This is very unlikely.
2767 if (!MRI.getRegClass(SrcReg)->contains(DestReg))
2770 EVT RVEVT = TLI.getValueType(RV->getType());
2771 if (!RVEVT.isSimple())
2774 // Vectors (of > 1 lane) in big endian need tricky handling.
2775 if (RVEVT.isVector() && RVEVT.getVectorNumElements() > 1)
2778 MVT RVVT = RVEVT.getSimpleVT();
2779 if (RVVT == MVT::f128)
2781 MVT DestVT = VA.getValVT();
2782 // Special handling for extended integers.
2783 if (RVVT != DestVT) {
2784 if (RVVT != MVT::i1 && RVVT != MVT::i8 && RVVT != MVT::i16)
2787 if (!Outs[0].Flags.isZExt() && !Outs[0].Flags.isSExt())
2790 bool isZExt = Outs[0].Flags.isZExt();
2791 SrcReg = EmitIntExt(RVVT, SrcReg, DestVT, isZExt);
2797 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2798 TII.get(TargetOpcode::COPY), DestReg).addReg(SrcReg);
2800 // Add register to return instruction.
2801 RetRegs.push_back(VA.getLocReg());
2804 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2805 TII.get(AArch64::RET_ReallyLR));
2806 for (unsigned i = 0, e = RetRegs.size(); i != e; ++i)
2807 MIB.addReg(RetRegs[i], RegState::Implicit);
2811 bool AArch64FastISel::SelectTrunc(const Instruction *I) {
2812 Type *DestTy = I->getType();
2813 Value *Op = I->getOperand(0);
2814 Type *SrcTy = Op->getType();
2816 EVT SrcEVT = TLI.getValueType(SrcTy, true);
2817 EVT DestEVT = TLI.getValueType(DestTy, true);
2818 if (!SrcEVT.isSimple())
2820 if (!DestEVT.isSimple())
2823 MVT SrcVT = SrcEVT.getSimpleVT();
2824 MVT DestVT = DestEVT.getSimpleVT();
2826 if (SrcVT != MVT::i64 && SrcVT != MVT::i32 && SrcVT != MVT::i16 &&
2829 if (DestVT != MVT::i32 && DestVT != MVT::i16 && DestVT != MVT::i8 &&
2833 unsigned SrcReg = getRegForValue(Op);
2836 bool SrcIsKill = hasTrivialKill(Op);
2838 // If we're truncating from i64 to a smaller non-legal type then generate an
2839 // AND. Otherwise, we know the high bits are undefined and a truncate only
2840 // generate a COPY. We cannot mark the source register also as result
2841 // register, because this can incorrectly transfer the kill flag onto the
2844 if (SrcVT == MVT::i64) {
2846 switch (DestVT.SimpleTy) {
2848 // Trunc i64 to i32 is handled by the target-independent fast-isel.
2860 // Issue an extract_subreg to get the lower 32-bits.
2861 unsigned Reg32 = fastEmitInst_extractsubreg(MVT::i32, SrcReg, SrcIsKill,
2863 // Create the AND instruction which performs the actual truncation.
2864 ResultReg = emitAnd_ri(MVT::i32, Reg32, /*IsKill=*/true, Mask);
2865 assert(ResultReg && "Unexpected AND instruction emission failure.");
2867 ResultReg = createResultReg(&AArch64::GPR32RegClass);
2868 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2869 TII.get(TargetOpcode::COPY), ResultReg)
2870 .addReg(SrcReg, getKillRegState(SrcIsKill));
2873 updateValueMap(I, ResultReg);
2877 unsigned AArch64FastISel::Emiti1Ext(unsigned SrcReg, MVT DestVT, bool isZExt) {
2878 assert((DestVT == MVT::i8 || DestVT == MVT::i16 || DestVT == MVT::i32 ||
2879 DestVT == MVT::i64) &&
2880 "Unexpected value type.");
2881 // Handle i8 and i16 as i32.
2882 if (DestVT == MVT::i8 || DestVT == MVT::i16)
2886 unsigned ResultReg = emitAnd_ri(MVT::i32, SrcReg, /*TODO:IsKill=*/false, 1);
2887 assert(ResultReg && "Unexpected AND instruction emission failure.");
2888 if (DestVT == MVT::i64) {
2889 // We're ZExt i1 to i64. The ANDWri Wd, Ws, #1 implicitly clears the
2890 // upper 32 bits. Emit a SUBREG_TO_REG to extend from Wd to Xd.
2891 unsigned Reg64 = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
2892 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2893 TII.get(AArch64::SUBREG_TO_REG), Reg64)
2896 .addImm(AArch64::sub_32);
2901 if (DestVT == MVT::i64) {
2902 // FIXME: We're SExt i1 to i64.
2905 return fastEmitInst_rii(AArch64::SBFMWri, &AArch64::GPR32RegClass, SrcReg,
2906 /*TODO:IsKill=*/false, 0, 0);
2910 unsigned AArch64FastISel::Emit_MUL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
2911 unsigned Op1, bool Op1IsKill) {
2913 switch (RetVT.SimpleTy) {
2919 Opc = AArch64::MADDWrrr; ZReg = AArch64::WZR; break;
2921 Opc = AArch64::MADDXrrr; ZReg = AArch64::XZR; break;
2924 const TargetRegisterClass *RC =
2925 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
2926 return fastEmitInst_rrr(Opc, RC, Op0, Op0IsKill, Op1, Op1IsKill,
2927 /*IsKill=*/ZReg, true);
2930 unsigned AArch64FastISel::Emit_SMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
2931 unsigned Op1, bool Op1IsKill) {
2932 if (RetVT != MVT::i64)
2935 return fastEmitInst_rrr(AArch64::SMADDLrrr, &AArch64::GPR64RegClass,
2936 Op0, Op0IsKill, Op1, Op1IsKill,
2937 AArch64::XZR, /*IsKill=*/true);
2940 unsigned AArch64FastISel::Emit_UMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
2941 unsigned Op1, bool Op1IsKill) {
2942 if (RetVT != MVT::i64)
2945 return fastEmitInst_rrr(AArch64::UMADDLrrr, &AArch64::GPR64RegClass,
2946 Op0, Op0IsKill, Op1, Op1IsKill,
2947 AArch64::XZR, /*IsKill=*/true);
2950 unsigned AArch64FastISel::emitLSL_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
2951 unsigned Op1Reg, bool Op1IsKill) {
2953 bool NeedTrunc = false;
2955 switch (RetVT.SimpleTy) {
2957 case MVT::i8: Opc = AArch64::LSLVWr; NeedTrunc = true; Mask = 0xff; break;
2958 case MVT::i16: Opc = AArch64::LSLVWr; NeedTrunc = true; Mask = 0xffff; break;
2959 case MVT::i32: Opc = AArch64::LSLVWr; break;
2960 case MVT::i64: Opc = AArch64::LSLVXr; break;
2963 const TargetRegisterClass *RC =
2964 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
2966 Op1Reg = emitAnd_ri(MVT::i32, Op1Reg, Op1IsKill, Mask);
2969 unsigned ResultReg = fastEmitInst_rr(Opc, RC, Op0Reg, Op0IsKill, Op1Reg,
2972 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
2976 unsigned AArch64FastISel::emitLSL_ri(MVT RetVT, MVT SrcVT, unsigned Op0,
2977 bool Op0IsKill, uint64_t Shift,
2979 assert(RetVT.SimpleTy >= SrcVT.SimpleTy &&
2980 "Unexpected source/return type pair.");
2981 assert((SrcVT == MVT::i8 || SrcVT == MVT::i16 || SrcVT == MVT::i32 ||
2982 SrcVT == MVT::i64) && "Unexpected source value type.");
2983 assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 ||
2984 RetVT == MVT::i64) && "Unexpected return value type.");
2986 bool Is64Bit = (RetVT == MVT::i64);
2987 unsigned RegSize = Is64Bit ? 64 : 32;
2988 unsigned DstBits = RetVT.getSizeInBits();
2989 unsigned SrcBits = SrcVT.getSizeInBits();
2991 // Don't deal with undefined shifts.
2992 if (Shift >= DstBits)
2995 // For immediate shifts we can fold the zero-/sign-extension into the shift.
2996 // {S|U}BFM Wd, Wn, #r, #s
2997 // Wd<32+s-r,32-r> = Wn<s:0> when r > s
2999 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3000 // %2 = shl i16 %1, 4
3001 // Wd<32+7-28,32-28> = Wn<7:0> <- clamp s to 7
3002 // 0b1111_1111_1111_1111__1111_1010_1010_0000 sext
3003 // 0b0000_0000_0000_0000__0000_0101_0101_0000 sext | zext
3004 // 0b0000_0000_0000_0000__0000_1010_1010_0000 zext
3006 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3007 // %2 = shl i16 %1, 8
3008 // Wd<32+7-24,32-24> = Wn<7:0>
3009 // 0b1111_1111_1111_1111__1010_1010_0000_0000 sext
3010 // 0b0000_0000_0000_0000__0101_0101_0000_0000 sext | zext
3011 // 0b0000_0000_0000_0000__1010_1010_0000_0000 zext
3013 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3014 // %2 = shl i16 %1, 12
3015 // Wd<32+3-20,32-20> = Wn<3:0>
3016 // 0b1111_1111_1111_1111__1010_0000_0000_0000 sext
3017 // 0b0000_0000_0000_0000__0101_0000_0000_0000 sext | zext
3018 // 0b0000_0000_0000_0000__1010_0000_0000_0000 zext
3020 unsigned ImmR = RegSize - Shift;
3021 // Limit the width to the length of the source type.
3022 unsigned ImmS = std::min<unsigned>(SrcBits - 1, DstBits - 1 - Shift);
3023 static const unsigned OpcTable[2][2] = {
3024 {AArch64::SBFMWri, AArch64::SBFMXri},
3025 {AArch64::UBFMWri, AArch64::UBFMXri}
3027 unsigned Opc = OpcTable[IsZext][Is64Bit];
3028 const TargetRegisterClass *RC =
3029 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3030 if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) {
3031 unsigned TmpReg = MRI.createVirtualRegister(RC);
3032 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3033 TII.get(AArch64::SUBREG_TO_REG), TmpReg)
3035 .addReg(Op0, getKillRegState(Op0IsKill))
3036 .addImm(AArch64::sub_32);
3040 return fastEmitInst_rii(Opc, RC, Op0, Op0IsKill, ImmR, ImmS);
3043 unsigned AArch64FastISel::emitLSR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
3044 unsigned Op1Reg, bool Op1IsKill) {
3046 bool NeedTrunc = false;
3048 switch (RetVT.SimpleTy) {
3050 case MVT::i8: Opc = AArch64::LSRVWr; NeedTrunc = true; Mask = 0xff; break;
3051 case MVT::i16: Opc = AArch64::LSRVWr; NeedTrunc = true; Mask = 0xffff; break;
3052 case MVT::i32: Opc = AArch64::LSRVWr; break;
3053 case MVT::i64: Opc = AArch64::LSRVXr; break;
3056 const TargetRegisterClass *RC =
3057 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3059 Op0Reg = emitAnd_ri(MVT::i32, Op0Reg, Op0IsKill, Mask);
3060 Op1Reg = emitAnd_ri(MVT::i32, Op1Reg, Op1IsKill, Mask);
3061 Op0IsKill = Op1IsKill = true;
3063 unsigned ResultReg = fastEmitInst_rr(Opc, RC, Op0Reg, Op0IsKill, Op1Reg,
3066 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
3070 unsigned AArch64FastISel::emitLSR_ri(MVT RetVT, MVT SrcVT, unsigned Op0,
3071 bool Op0IsKill, uint64_t Shift,
3073 assert(RetVT.SimpleTy >= SrcVT.SimpleTy &&
3074 "Unexpected source/return type pair.");
3075 assert((SrcVT == MVT::i8 || SrcVT == MVT::i16 || SrcVT == MVT::i32 ||
3076 SrcVT == MVT::i64) && "Unexpected source value type.");
3077 assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 ||
3078 RetVT == MVT::i64) && "Unexpected return value type.");
3080 bool Is64Bit = (RetVT == MVT::i64);
3081 unsigned RegSize = Is64Bit ? 64 : 32;
3082 unsigned DstBits = RetVT.getSizeInBits();
3083 unsigned SrcBits = SrcVT.getSizeInBits();
3085 // Don't deal with undefined shifts.
3086 if (Shift >= DstBits)
3089 // For immediate shifts we can fold the zero-/sign-extension into the shift.
3090 // {S|U}BFM Wd, Wn, #r, #s
3091 // Wd<s-r:0> = Wn<s:r> when r <= s
3093 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3094 // %2 = lshr i16 %1, 4
3095 // Wd<7-4:0> = Wn<7:4>
3096 // 0b0000_0000_0000_0000__0000_1111_1111_1010 sext
3097 // 0b0000_0000_0000_0000__0000_0000_0000_0101 sext | zext
3098 // 0b0000_0000_0000_0000__0000_0000_0000_1010 zext
3100 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3101 // %2 = lshr i16 %1, 8
3102 // Wd<7-7,0> = Wn<7:7>
3103 // 0b0000_0000_0000_0000__0000_0000_1111_1111 sext
3104 // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
3105 // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
3107 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3108 // %2 = lshr i16 %1, 12
3109 // Wd<7-7,0> = Wn<7:7> <- clamp r to 7
3110 // 0b0000_0000_0000_0000__0000_0000_0000_1111 sext
3111 // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
3112 // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
3114 if (Shift >= SrcBits && IsZExt)
3115 return AArch64MaterializeInt(ConstantInt::get(*Context, APInt(RegSize, 0)),
3118 // It is not possible to fold a sign-extend into the LShr instruction. In this
3119 // case emit a sign-extend.
3121 Op0 = EmitIntExt(SrcVT, Op0, RetVT, IsZExt);
3126 SrcBits = SrcVT.getSizeInBits();
3130 unsigned ImmR = std::min<unsigned>(SrcBits - 1, Shift);
3131 unsigned ImmS = SrcBits - 1;
3132 static const unsigned OpcTable[2][2] = {
3133 {AArch64::SBFMWri, AArch64::SBFMXri},
3134 {AArch64::UBFMWri, AArch64::UBFMXri}
3136 unsigned Opc = OpcTable[IsZExt][Is64Bit];
3137 const TargetRegisterClass *RC =
3138 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3139 if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) {
3140 unsigned TmpReg = MRI.createVirtualRegister(RC);
3141 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3142 TII.get(AArch64::SUBREG_TO_REG), TmpReg)
3144 .addReg(Op0, getKillRegState(Op0IsKill))
3145 .addImm(AArch64::sub_32);
3149 return fastEmitInst_rii(Opc, RC, Op0, Op0IsKill, ImmR, ImmS);
3152 unsigned AArch64FastISel::emitASR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
3153 unsigned Op1Reg, bool Op1IsKill) {
3155 bool NeedTrunc = false;
3157 switch (RetVT.SimpleTy) {
3159 case MVT::i8: Opc = AArch64::ASRVWr; NeedTrunc = true; Mask = 0xff; break;
3160 case MVT::i16: Opc = AArch64::ASRVWr; NeedTrunc = true; Mask = 0xffff; break;
3161 case MVT::i32: Opc = AArch64::ASRVWr; break;
3162 case MVT::i64: Opc = AArch64::ASRVXr; break;
3165 const TargetRegisterClass *RC =
3166 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3168 Op0Reg = EmitIntExt(RetVT, Op0Reg, MVT::i32, /*IsZExt=*/false);
3169 Op1Reg = emitAnd_ri(MVT::i32, Op1Reg, Op1IsKill, Mask);
3170 Op0IsKill = Op1IsKill = true;
3172 unsigned ResultReg = fastEmitInst_rr(Opc, RC, Op0Reg, Op0IsKill, Op1Reg,
3175 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
3179 unsigned AArch64FastISel::emitASR_ri(MVT RetVT, MVT SrcVT, unsigned Op0,
3180 bool Op0IsKill, uint64_t Shift,
3182 assert(RetVT.SimpleTy >= SrcVT.SimpleTy &&
3183 "Unexpected source/return type pair.");
3184 assert((SrcVT == MVT::i8 || SrcVT == MVT::i16 || SrcVT == MVT::i32 ||
3185 SrcVT == MVT::i64) && "Unexpected source value type.");
3186 assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 ||
3187 RetVT == MVT::i64) && "Unexpected return value type.");
3189 bool Is64Bit = (RetVT == MVT::i64);
3190 unsigned RegSize = Is64Bit ? 64 : 32;
3191 unsigned DstBits = RetVT.getSizeInBits();
3192 unsigned SrcBits = SrcVT.getSizeInBits();
3194 // Don't deal with undefined shifts.
3195 if (Shift >= DstBits)
3198 // For immediate shifts we can fold the zero-/sign-extension into the shift.
3199 // {S|U}BFM Wd, Wn, #r, #s
3200 // Wd<s-r:0> = Wn<s:r> when r <= s
3202 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3203 // %2 = ashr i16 %1, 4
3204 // Wd<7-4:0> = Wn<7:4>
3205 // 0b1111_1111_1111_1111__1111_1111_1111_1010 sext
3206 // 0b0000_0000_0000_0000__0000_0000_0000_0101 sext | zext
3207 // 0b0000_0000_0000_0000__0000_0000_0000_1010 zext
3209 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3210 // %2 = ashr i16 %1, 8
3211 // Wd<7-7,0> = Wn<7:7>
3212 // 0b1111_1111_1111_1111__1111_1111_1111_1111 sext
3213 // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
3214 // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
3216 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3217 // %2 = ashr i16 %1, 12
3218 // Wd<7-7,0> = Wn<7:7> <- clamp r to 7
3219 // 0b1111_1111_1111_1111__1111_1111_1111_1111 sext
3220 // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
3221 // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
3223 if (Shift >= SrcBits && IsZExt)
3224 return AArch64MaterializeInt(ConstantInt::get(*Context, APInt(RegSize, 0)),
3227 unsigned ImmR = std::min<unsigned>(SrcBits - 1, Shift);
3228 unsigned ImmS = SrcBits - 1;
3229 static const unsigned OpcTable[2][2] = {
3230 {AArch64::SBFMWri, AArch64::SBFMXri},
3231 {AArch64::UBFMWri, AArch64::UBFMXri}
3233 unsigned Opc = OpcTable[IsZExt][Is64Bit];
3234 const TargetRegisterClass *RC =
3235 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3236 if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) {
3237 unsigned TmpReg = MRI.createVirtualRegister(RC);
3238 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3239 TII.get(AArch64::SUBREG_TO_REG), TmpReg)
3241 .addReg(Op0, getKillRegState(Op0IsKill))
3242 .addImm(AArch64::sub_32);
3246 return fastEmitInst_rii(Opc, RC, Op0, Op0IsKill, ImmR, ImmS);
3249 unsigned AArch64FastISel::EmitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT,
3251 assert(DestVT != MVT::i1 && "ZeroExt/SignExt an i1?");
3253 // FastISel does not have plumbing to deal with extensions where the SrcVT or
3254 // DestVT are odd things, so test to make sure that they are both types we can
3255 // handle (i1/i8/i16/i32 for SrcVT and i8/i16/i32/i64 for DestVT), otherwise
3256 // bail out to SelectionDAG.
3257 if (((DestVT != MVT::i8) && (DestVT != MVT::i16) &&
3258 (DestVT != MVT::i32) && (DestVT != MVT::i64)) ||
3259 ((SrcVT != MVT::i1) && (SrcVT != MVT::i8) &&
3260 (SrcVT != MVT::i16) && (SrcVT != MVT::i32)))
3266 switch (SrcVT.SimpleTy) {
3270 return Emiti1Ext(SrcReg, DestVT, isZExt);
3272 if (DestVT == MVT::i64)
3273 Opc = isZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
3275 Opc = isZExt ? AArch64::UBFMWri : AArch64::SBFMWri;
3279 if (DestVT == MVT::i64)
3280 Opc = isZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
3282 Opc = isZExt ? AArch64::UBFMWri : AArch64::SBFMWri;
3286 assert(DestVT == MVT::i64 && "IntExt i32 to i32?!?");
3287 Opc = isZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
3292 // Handle i8 and i16 as i32.
3293 if (DestVT == MVT::i8 || DestVT == MVT::i16)
3295 else if (DestVT == MVT::i64) {
3296 unsigned Src64 = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
3297 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3298 TII.get(AArch64::SUBREG_TO_REG), Src64)
3301 .addImm(AArch64::sub_32);
3305 const TargetRegisterClass *RC =
3306 (DestVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3307 return fastEmitInst_rii(Opc, RC, SrcReg, /*TODO:IsKill=*/false, 0, Imm);
3310 bool AArch64FastISel::SelectIntExt(const Instruction *I) {
3311 // On ARM, in general, integer casts don't involve legal types; this code
3312 // handles promotable integers. The high bits for a type smaller than
3313 // the register size are assumed to be undefined.
3314 Type *DestTy = I->getType();
3315 Value *Src = I->getOperand(0);
3316 Type *SrcTy = Src->getType();
3318 bool isZExt = isa<ZExtInst>(I);
3319 unsigned SrcReg = getRegForValue(Src);
3323 EVT SrcEVT = TLI.getValueType(SrcTy, true);
3324 EVT DestEVT = TLI.getValueType(DestTy, true);
3325 if (!SrcEVT.isSimple())
3327 if (!DestEVT.isSimple())
3330 MVT SrcVT = SrcEVT.getSimpleVT();
3331 MVT DestVT = DestEVT.getSimpleVT();
3332 unsigned ResultReg = 0;
3334 // Check if it is an argument and if it is already zero/sign-extended.
3335 if (const auto *Arg = dyn_cast<Argument>(Src)) {
3336 if ((isZExt && Arg->hasZExtAttr()) || (!isZExt && Arg->hasSExtAttr())) {
3337 if (DestVT == MVT::i64) {
3338 ResultReg = createResultReg(TLI.getRegClassFor(DestVT));
3339 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3340 TII.get(AArch64::SUBREG_TO_REG), ResultReg)
3343 .addImm(AArch64::sub_32);
3350 ResultReg = EmitIntExt(SrcVT, SrcReg, DestVT, isZExt);
3355 updateValueMap(I, ResultReg);
3359 bool AArch64FastISel::SelectRem(const Instruction *I, unsigned ISDOpcode) {
3360 EVT DestEVT = TLI.getValueType(I->getType(), true);
3361 if (!DestEVT.isSimple())
3364 MVT DestVT = DestEVT.getSimpleVT();
3365 if (DestVT != MVT::i64 && DestVT != MVT::i32)
3369 bool is64bit = (DestVT == MVT::i64);
3370 switch (ISDOpcode) {
3374 DivOpc = is64bit ? AArch64::SDIVXr : AArch64::SDIVWr;
3377 DivOpc = is64bit ? AArch64::UDIVXr : AArch64::UDIVWr;
3380 unsigned MSubOpc = is64bit ? AArch64::MSUBXrrr : AArch64::MSUBWrrr;
3381 unsigned Src0Reg = getRegForValue(I->getOperand(0));
3384 bool Src0IsKill = hasTrivialKill(I->getOperand(0));
3386 unsigned Src1Reg = getRegForValue(I->getOperand(1));
3389 bool Src1IsKill = hasTrivialKill(I->getOperand(1));
3391 const TargetRegisterClass *RC =
3392 (DestVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3393 unsigned QuotReg = fastEmitInst_rr(DivOpc, RC, Src0Reg, /*IsKill=*/false,
3394 Src1Reg, /*IsKill=*/false);
3395 assert(QuotReg && "Unexpected DIV instruction emission failure.");
3396 // The remainder is computed as numerator - (quotient * denominator) using the
3397 // MSUB instruction.
3398 unsigned ResultReg = fastEmitInst_rrr(MSubOpc, RC, QuotReg, /*IsKill=*/true,
3399 Src1Reg, Src1IsKill, Src0Reg,
3401 updateValueMap(I, ResultReg);
3405 bool AArch64FastISel::SelectMul(const Instruction *I) {
3406 EVT SrcEVT = TLI.getValueType(I->getOperand(0)->getType(), true);
3407 if (!SrcEVT.isSimple())
3409 MVT SrcVT = SrcEVT.getSimpleVT();
3411 // Must be simple value type. Don't handle vectors.
3412 if (SrcVT != MVT::i64 && SrcVT != MVT::i32 && SrcVT != MVT::i16 &&
3416 unsigned Src0Reg = getRegForValue(I->getOperand(0));
3419 bool Src0IsKill = hasTrivialKill(I->getOperand(0));
3421 unsigned Src1Reg = getRegForValue(I->getOperand(1));
3424 bool Src1IsKill = hasTrivialKill(I->getOperand(1));
3426 unsigned ResultReg =
3427 Emit_MUL_rr(SrcVT, Src0Reg, Src0IsKill, Src1Reg, Src1IsKill);
3432 updateValueMap(I, ResultReg);
3436 bool AArch64FastISel::SelectShift(const Instruction *I) {
3438 if (!isTypeSupported(I->getType(), RetVT))
3441 if (const auto *C = dyn_cast<ConstantInt>(I->getOperand(1))) {
3442 unsigned ResultReg = 0;
3443 uint64_t ShiftVal = C->getZExtValue();
3445 bool IsZExt = (I->getOpcode() == Instruction::AShr) ? false : true;
3446 const Value *Op0 = I->getOperand(0);
3447 if (const auto *ZExt = dyn_cast<ZExtInst>(Op0)) {
3449 if (isValueAvailable(ZExt) && isTypeSupported(ZExt->getSrcTy(), TmpVT)) {
3452 Op0 = ZExt->getOperand(0);
3454 } else if (const auto *SExt = dyn_cast<SExtInst>(Op0)) {
3456 if (isValueAvailable(SExt) && isTypeSupported(SExt->getSrcTy(), TmpVT)) {
3459 Op0 = SExt->getOperand(0);
3463 unsigned Op0Reg = getRegForValue(Op0);
3466 bool Op0IsKill = hasTrivialKill(Op0);
3468 switch (I->getOpcode()) {
3469 default: llvm_unreachable("Unexpected instruction.");
3470 case Instruction::Shl:
3471 ResultReg = emitLSL_ri(RetVT, SrcVT, Op0Reg, Op0IsKill, ShiftVal, IsZExt);
3473 case Instruction::AShr:
3474 ResultReg = emitASR_ri(RetVT, SrcVT, Op0Reg, Op0IsKill, ShiftVal, IsZExt);
3476 case Instruction::LShr:
3477 ResultReg = emitLSR_ri(RetVT, SrcVT, Op0Reg, Op0IsKill, ShiftVal, IsZExt);
3483 updateValueMap(I, ResultReg);
3487 unsigned Op0Reg = getRegForValue(I->getOperand(0));
3490 bool Op0IsKill = hasTrivialKill(I->getOperand(0));
3492 unsigned Op1Reg = getRegForValue(I->getOperand(1));
3495 bool Op1IsKill = hasTrivialKill(I->getOperand(1));
3497 unsigned ResultReg = 0;
3498 switch (I->getOpcode()) {
3499 default: llvm_unreachable("Unexpected instruction.");
3500 case Instruction::Shl:
3501 ResultReg = emitLSL_rr(RetVT, Op0Reg, Op0IsKill, Op1Reg, Op1IsKill);
3503 case Instruction::AShr:
3504 ResultReg = emitASR_rr(RetVT, Op0Reg, Op0IsKill, Op1Reg, Op1IsKill);
3506 case Instruction::LShr:
3507 ResultReg = emitLSR_rr(RetVT, Op0Reg, Op0IsKill, Op1Reg, Op1IsKill);
3514 updateValueMap(I, ResultReg);
3518 bool AArch64FastISel::SelectBitCast(const Instruction *I) {
3521 if (!isTypeLegal(I->getOperand(0)->getType(), SrcVT))
3523 if (!isTypeLegal(I->getType(), RetVT))
3527 if (RetVT == MVT::f32 && SrcVT == MVT::i32)
3528 Opc = AArch64::FMOVWSr;
3529 else if (RetVT == MVT::f64 && SrcVT == MVT::i64)
3530 Opc = AArch64::FMOVXDr;
3531 else if (RetVT == MVT::i32 && SrcVT == MVT::f32)
3532 Opc = AArch64::FMOVSWr;
3533 else if (RetVT == MVT::i64 && SrcVT == MVT::f64)
3534 Opc = AArch64::FMOVDXr;
3538 const TargetRegisterClass *RC = nullptr;
3539 switch (RetVT.SimpleTy) {
3540 default: llvm_unreachable("Unexpected value type.");
3541 case MVT::i32: RC = &AArch64::GPR32RegClass; break;
3542 case MVT::i64: RC = &AArch64::GPR64RegClass; break;
3543 case MVT::f32: RC = &AArch64::FPR32RegClass; break;
3544 case MVT::f64: RC = &AArch64::FPR64RegClass; break;
3546 unsigned Op0Reg = getRegForValue(I->getOperand(0));
3549 bool Op0IsKill = hasTrivialKill(I->getOperand(0));
3550 unsigned ResultReg = fastEmitInst_r(Opc, RC, Op0Reg, Op0IsKill);
3555 updateValueMap(I, ResultReg);
3559 bool AArch64FastISel::fastSelectInstruction(const Instruction *I) {
3560 switch (I->getOpcode()) {
3563 case Instruction::Add:
3564 case Instruction::Sub:
3565 if (selectAddSub(I))
3568 case Instruction::Mul:
3569 if (!selectBinaryOp(I, ISD::MUL))
3570 return SelectMul(I);
3572 case Instruction::SRem:
3573 if (!selectBinaryOp(I, ISD::SREM))
3574 return SelectRem(I, ISD::SREM);
3576 case Instruction::URem:
3577 if (!selectBinaryOp(I, ISD::UREM))
3578 return SelectRem(I, ISD::UREM);
3580 case Instruction::Shl:
3581 case Instruction::LShr:
3582 case Instruction::AShr:
3586 case Instruction::And:
3587 if (selectLogicalOp(I, ISD::AND))
3590 case Instruction::Or:
3591 if (selectLogicalOp(I, ISD::OR))
3594 case Instruction::Xor:
3595 if (selectLogicalOp(I, ISD::XOR))
3598 case Instruction::Br:
3599 return SelectBranch(I);
3600 case Instruction::IndirectBr:
3601 return SelectIndirectBr(I);
3602 case Instruction::BitCast:
3603 if (!FastISel::selectBitCast(I))
3604 return SelectBitCast(I);
3606 case Instruction::FPToSI:
3607 if (!selectCast(I, ISD::FP_TO_SINT))
3608 return SelectFPToInt(I, /*Signed=*/true);
3610 case Instruction::FPToUI:
3611 return SelectFPToInt(I, /*Signed=*/false);
3612 case Instruction::ZExt:
3613 if (!selectCast(I, ISD::ZERO_EXTEND))
3614 return SelectIntExt(I);
3616 case Instruction::SExt:
3617 if (!selectCast(I, ISD::SIGN_EXTEND))
3618 return SelectIntExt(I);
3620 case Instruction::Trunc:
3621 if (!selectCast(I, ISD::TRUNCATE))
3622 return SelectTrunc(I);
3624 case Instruction::FPExt:
3625 return SelectFPExt(I);
3626 case Instruction::FPTrunc:
3627 return SelectFPTrunc(I);
3628 case Instruction::SIToFP:
3629 if (!selectCast(I, ISD::SINT_TO_FP))
3630 return SelectIntToFP(I, /*Signed=*/true);
3632 case Instruction::UIToFP:
3633 return SelectIntToFP(I, /*Signed=*/false);
3634 case Instruction::Load:
3635 return SelectLoad(I);
3636 case Instruction::Store:
3637 return SelectStore(I);
3638 case Instruction::FCmp:
3639 case Instruction::ICmp:
3640 return SelectCmp(I);
3641 case Instruction::Select:
3642 return SelectSelect(I);
3643 case Instruction::Ret:
3644 return SelectRet(I);
3647 // fall-back to target-independent instruction selection.
3648 return selectOperator(I, I->getOpcode());
3649 // Silence warnings.
3650 (void)&CC_AArch64_DarwinPCS_VarArg;
3654 llvm::FastISel *AArch64::createFastISel(FunctionLoweringInfo &funcInfo,
3655 const TargetLibraryInfo *libInfo) {
3656 return new AArch64FastISel(funcInfo, libInfo);