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 "AArch64CallingConvention.h"
18 #include "AArch64Subtarget.h"
19 #include "AArch64TargetMachine.h"
20 #include "MCTargetDesc/AArch64AddressingModes.h"
21 #include "llvm/Analysis/BranchProbabilityInfo.h"
22 #include "llvm/CodeGen/CallingConvLower.h"
23 #include "llvm/CodeGen/FastISel.h"
24 #include "llvm/CodeGen/FunctionLoweringInfo.h"
25 #include "llvm/CodeGen/MachineConstantPool.h"
26 #include "llvm/CodeGen/MachineFrameInfo.h"
27 #include "llvm/CodeGen/MachineInstrBuilder.h"
28 #include "llvm/CodeGen/MachineRegisterInfo.h"
29 #include "llvm/IR/CallingConv.h"
30 #include "llvm/IR/DataLayout.h"
31 #include "llvm/IR/DerivedTypes.h"
32 #include "llvm/IR/Function.h"
33 #include "llvm/IR/GetElementPtrTypeIterator.h"
34 #include "llvm/IR/GlobalAlias.h"
35 #include "llvm/IR/GlobalVariable.h"
36 #include "llvm/IR/Instructions.h"
37 #include "llvm/IR/IntrinsicInst.h"
38 #include "llvm/IR/Operator.h"
39 #include "llvm/MC/MCSymbol.h"
40 #include "llvm/Support/CommandLine.h"
45 class AArch64FastISel final : public FastISel {
55 AArch64_AM::ShiftExtendType ExtType;
63 const GlobalValue *GV;
66 Address() : Kind(RegBase), ExtType(AArch64_AM::InvalidShiftExtend),
67 OffsetReg(0), Shift(0), Offset(0), GV(nullptr) { Base.Reg = 0; }
68 void setKind(BaseKind K) { Kind = K; }
69 BaseKind getKind() const { return Kind; }
70 void setExtendType(AArch64_AM::ShiftExtendType E) { ExtType = E; }
71 AArch64_AM::ShiftExtendType getExtendType() const { return ExtType; }
72 bool isRegBase() const { return Kind == RegBase; }
73 bool isFIBase() const { return Kind == FrameIndexBase; }
74 void setReg(unsigned Reg) {
75 assert(isRegBase() && "Invalid base register access!");
78 unsigned getReg() const {
79 assert(isRegBase() && "Invalid base register access!");
82 void setOffsetReg(unsigned Reg) {
85 unsigned getOffsetReg() const {
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);
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);
135 bool selectFRem(const Instruction *I);
136 bool selectSDiv(const Instruction *I);
137 bool selectGetElementPtr(const Instruction *I);
139 // Utility helper routines.
140 bool isTypeLegal(Type *Ty, MVT &VT);
141 bool isTypeSupported(Type *Ty, MVT &VT, bool IsVectorAllowed = false);
142 bool isValueAvailable(const Value *V) const;
143 bool computeAddress(const Value *Obj, Address &Addr, Type *Ty = nullptr);
144 bool computeCallAddress(const Value *V, Address &Addr);
145 bool simplifyAddress(Address &Addr, MVT VT);
146 void addLoadStoreOperands(Address &Addr, const MachineInstrBuilder &MIB,
147 unsigned Flags, unsigned ScaleFactor,
148 MachineMemOperand *MMO);
149 bool isMemCpySmall(uint64_t Len, unsigned Alignment);
150 bool tryEmitSmallMemCpy(Address Dest, Address Src, uint64_t Len,
152 bool foldXALUIntrinsic(AArch64CC::CondCode &CC, const Instruction *I,
154 bool optimizeIntExtLoad(const Instruction *I, MVT RetVT, MVT SrcVT);
155 bool optimizeSelect(const SelectInst *SI);
156 std::pair<unsigned, bool> getRegForGEPIndex(const Value *Idx);
158 // Emit helper routines.
159 unsigned emitAddSub(bool UseAdd, MVT RetVT, const Value *LHS,
160 const Value *RHS, bool SetFlags = false,
161 bool WantResult = true, bool IsZExt = false);
162 unsigned emitAddSub_rr(bool UseAdd, MVT RetVT, unsigned LHSReg,
163 bool LHSIsKill, unsigned RHSReg, bool RHSIsKill,
164 bool SetFlags = false, bool WantResult = true);
165 unsigned emitAddSub_ri(bool UseAdd, MVT RetVT, unsigned LHSReg,
166 bool LHSIsKill, uint64_t Imm, bool SetFlags = false,
167 bool WantResult = true);
168 unsigned emitAddSub_rs(bool UseAdd, MVT RetVT, unsigned LHSReg,
169 bool LHSIsKill, unsigned RHSReg, bool RHSIsKill,
170 AArch64_AM::ShiftExtendType ShiftType,
171 uint64_t ShiftImm, bool SetFlags = false,
172 bool WantResult = true);
173 unsigned emitAddSub_rx(bool UseAdd, MVT RetVT, unsigned LHSReg,
174 bool LHSIsKill, unsigned RHSReg, bool RHSIsKill,
175 AArch64_AM::ShiftExtendType ExtType,
176 uint64_t ShiftImm, bool SetFlags = false,
177 bool WantResult = true);
180 bool emitCompareAndBranch(const BranchInst *BI);
181 bool emitCmp(const Value *LHS, const Value *RHS, bool IsZExt);
182 bool emitICmp(MVT RetVT, const Value *LHS, const Value *RHS, bool IsZExt);
183 bool emitICmp_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill, uint64_t Imm);
184 bool emitFCmp(MVT RetVT, const Value *LHS, const Value *RHS);
185 unsigned emitLoad(MVT VT, MVT ResultVT, Address Addr, bool WantZExt = true,
186 MachineMemOperand *MMO = nullptr);
187 bool emitStore(MVT VT, unsigned SrcReg, Address Addr,
188 MachineMemOperand *MMO = nullptr);
189 unsigned emitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, bool isZExt);
190 unsigned emiti1Ext(unsigned SrcReg, MVT DestVT, bool isZExt);
191 unsigned emitAdd(MVT RetVT, const Value *LHS, const Value *RHS,
192 bool SetFlags = false, bool WantResult = true,
193 bool IsZExt = false);
194 unsigned emitAdd_ri_(MVT VT, unsigned Op0, bool Op0IsKill, int64_t Imm);
195 unsigned emitSub(MVT RetVT, const Value *LHS, const Value *RHS,
196 bool SetFlags = false, bool WantResult = true,
197 bool IsZExt = false);
198 unsigned emitSubs_rr(MVT RetVT, unsigned LHSReg, bool LHSIsKill,
199 unsigned RHSReg, bool RHSIsKill, bool WantResult = true);
200 unsigned emitSubs_rs(MVT RetVT, unsigned LHSReg, bool LHSIsKill,
201 unsigned RHSReg, bool RHSIsKill,
202 AArch64_AM::ShiftExtendType ShiftType, uint64_t ShiftImm,
203 bool WantResult = true);
204 unsigned emitLogicalOp(unsigned ISDOpc, MVT RetVT, const Value *LHS,
206 unsigned emitLogicalOp_ri(unsigned ISDOpc, MVT RetVT, unsigned LHSReg,
207 bool LHSIsKill, uint64_t Imm);
208 unsigned emitLogicalOp_rs(unsigned ISDOpc, MVT RetVT, unsigned LHSReg,
209 bool LHSIsKill, unsigned RHSReg, bool RHSIsKill,
211 unsigned emitAnd_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill, uint64_t Imm);
212 unsigned emitMul_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
213 unsigned Op1, bool Op1IsKill);
214 unsigned emitSMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
215 unsigned Op1, bool Op1IsKill);
216 unsigned emitUMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
217 unsigned Op1, bool Op1IsKill);
218 unsigned emitLSL_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
219 unsigned Op1Reg, bool Op1IsKill);
220 unsigned emitLSL_ri(MVT RetVT, MVT SrcVT, unsigned Op0Reg, bool Op0IsKill,
221 uint64_t Imm, bool IsZExt = true);
222 unsigned emitLSR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
223 unsigned Op1Reg, bool Op1IsKill);
224 unsigned emitLSR_ri(MVT RetVT, MVT SrcVT, unsigned Op0Reg, bool Op0IsKill,
225 uint64_t Imm, bool IsZExt = true);
226 unsigned emitASR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
227 unsigned Op1Reg, bool Op1IsKill);
228 unsigned emitASR_ri(MVT RetVT, MVT SrcVT, unsigned Op0Reg, bool Op0IsKill,
229 uint64_t Imm, bool IsZExt = false);
231 unsigned materializeInt(const ConstantInt *CI, MVT VT);
232 unsigned materializeFP(const ConstantFP *CFP, MVT VT);
233 unsigned materializeGV(const GlobalValue *GV);
235 // Call handling routines.
237 CCAssignFn *CCAssignFnForCall(CallingConv::ID CC) const;
238 bool processCallArgs(CallLoweringInfo &CLI, SmallVectorImpl<MVT> &ArgVTs,
240 bool finishCall(CallLoweringInfo &CLI, MVT RetVT, unsigned NumBytes);
243 // Backend specific FastISel code.
244 unsigned fastMaterializeAlloca(const AllocaInst *AI) override;
245 unsigned fastMaterializeConstant(const Constant *C) override;
246 unsigned fastMaterializeFloatZero(const ConstantFP* CF) override;
248 explicit AArch64FastISel(FunctionLoweringInfo &FuncInfo,
249 const TargetLibraryInfo *LibInfo)
250 : FastISel(FuncInfo, LibInfo, /*SkipTargetIndependentISel=*/true) {
252 &static_cast<const AArch64Subtarget &>(FuncInfo.MF->getSubtarget());
253 Context = &FuncInfo.Fn->getContext();
256 bool fastSelectInstruction(const Instruction *I) override;
258 #include "AArch64GenFastISel.inc"
261 } // end anonymous namespace
263 #include "AArch64GenCallingConv.inc"
265 /// \brief Check if the sign-/zero-extend will be a noop.
266 static bool isIntExtFree(const Instruction *I) {
267 assert((isa<ZExtInst>(I) || isa<SExtInst>(I)) &&
268 "Unexpected integer extend instruction.");
269 assert(!I->getType()->isVectorTy() && I->getType()->isIntegerTy() &&
270 "Unexpected value type.");
271 bool IsZExt = isa<ZExtInst>(I);
273 if (const auto *LI = dyn_cast<LoadInst>(I->getOperand(0)))
277 if (const auto *Arg = dyn_cast<Argument>(I->getOperand(0)))
278 if ((IsZExt && Arg->hasZExtAttr()) || (!IsZExt && Arg->hasSExtAttr()))
284 /// \brief Determine the implicit scale factor that is applied by a memory
285 /// operation for a given value type.
286 static unsigned getImplicitScaleFactor(MVT VT) {
287 switch (VT.SimpleTy) {
290 case MVT::i1: // fall-through
295 case MVT::i32: // fall-through
298 case MVT::i64: // fall-through
304 CCAssignFn *AArch64FastISel::CCAssignFnForCall(CallingConv::ID CC) const {
305 if (CC == CallingConv::WebKit_JS)
306 return CC_AArch64_WebKit_JS;
307 if (CC == CallingConv::GHC)
308 return CC_AArch64_GHC;
309 return Subtarget->isTargetDarwin() ? CC_AArch64_DarwinPCS : CC_AArch64_AAPCS;
312 unsigned AArch64FastISel::fastMaterializeAlloca(const AllocaInst *AI) {
313 assert(TLI.getValueType(AI->getType(), true) == MVT::i64 &&
314 "Alloca should always return a pointer.");
316 // Don't handle dynamic allocas.
317 if (!FuncInfo.StaticAllocaMap.count(AI))
320 DenseMap<const AllocaInst *, int>::iterator SI =
321 FuncInfo.StaticAllocaMap.find(AI);
323 if (SI != FuncInfo.StaticAllocaMap.end()) {
324 unsigned ResultReg = createResultReg(&AArch64::GPR64spRegClass);
325 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri),
327 .addFrameIndex(SI->second)
336 unsigned AArch64FastISel::materializeInt(const ConstantInt *CI, MVT VT) {
341 return fastEmit_i(VT, VT, ISD::Constant, CI->getZExtValue());
343 // Create a copy from the zero register to materialize a "0" value.
344 const TargetRegisterClass *RC = (VT == MVT::i64) ? &AArch64::GPR64RegClass
345 : &AArch64::GPR32RegClass;
346 unsigned ZeroReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
347 unsigned ResultReg = createResultReg(RC);
348 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TargetOpcode::COPY),
349 ResultReg).addReg(ZeroReg, getKillRegState(true));
353 unsigned AArch64FastISel::materializeFP(const ConstantFP *CFP, MVT VT) {
354 // Positive zero (+0.0) has to be materialized with a fmov from the zero
355 // register, because the immediate version of fmov cannot encode zero.
356 if (CFP->isNullValue())
357 return fastMaterializeFloatZero(CFP);
359 if (VT != MVT::f32 && VT != MVT::f64)
362 const APFloat Val = CFP->getValueAPF();
363 bool Is64Bit = (VT == MVT::f64);
364 // This checks to see if we can use FMOV instructions to materialize
365 // a constant, otherwise we have to materialize via the constant pool.
366 if (TLI.isFPImmLegal(Val, VT)) {
368 Is64Bit ? AArch64_AM::getFP64Imm(Val) : AArch64_AM::getFP32Imm(Val);
369 assert((Imm != -1) && "Cannot encode floating-point constant.");
370 unsigned Opc = Is64Bit ? AArch64::FMOVDi : AArch64::FMOVSi;
371 return fastEmitInst_i(Opc, TLI.getRegClassFor(VT), Imm);
374 // For the MachO large code model materialize the FP constant in code.
375 if (Subtarget->isTargetMachO() && TM.getCodeModel() == CodeModel::Large) {
376 unsigned Opc1 = Is64Bit ? AArch64::MOVi64imm : AArch64::MOVi32imm;
377 const TargetRegisterClass *RC = Is64Bit ?
378 &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
380 unsigned TmpReg = createResultReg(RC);
381 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc1), TmpReg)
382 .addImm(CFP->getValueAPF().bitcastToAPInt().getZExtValue());
384 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
385 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
386 TII.get(TargetOpcode::COPY), ResultReg)
387 .addReg(TmpReg, getKillRegState(true));
392 // Materialize via constant pool. MachineConstantPool wants an explicit
394 unsigned Align = DL.getPrefTypeAlignment(CFP->getType());
396 Align = DL.getTypeAllocSize(CFP->getType());
398 unsigned CPI = MCP.getConstantPoolIndex(cast<Constant>(CFP), Align);
399 unsigned ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
400 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
401 ADRPReg).addConstantPoolIndex(CPI, 0, AArch64II::MO_PAGE);
403 unsigned Opc = Is64Bit ? AArch64::LDRDui : AArch64::LDRSui;
404 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
405 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
407 .addConstantPoolIndex(CPI, 0, AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
411 unsigned AArch64FastISel::materializeGV(const GlobalValue *GV) {
412 // We can't handle thread-local variables quickly yet.
413 if (GV->isThreadLocal())
416 // MachO still uses GOT for large code-model accesses, but ELF requires
417 // movz/movk sequences, which FastISel doesn't handle yet.
418 if (TM.getCodeModel() != CodeModel::Small && !Subtarget->isTargetMachO())
421 unsigned char OpFlags = Subtarget->ClassifyGlobalReference(GV, TM);
423 EVT DestEVT = TLI.getValueType(GV->getType(), true);
424 if (!DestEVT.isSimple())
427 unsigned ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
430 if (OpFlags & AArch64II::MO_GOT) {
432 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
434 .addGlobalAddress(GV, 0, AArch64II::MO_GOT | AArch64II::MO_PAGE);
436 ResultReg = createResultReg(&AArch64::GPR64RegClass);
437 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::LDRXui),
440 .addGlobalAddress(GV, 0, AArch64II::MO_GOT | AArch64II::MO_PAGEOFF |
442 } else if (OpFlags & AArch64II::MO_CONSTPOOL) {
443 // We can't handle addresses loaded from a constant pool quickly yet.
447 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
449 .addGlobalAddress(GV, 0, AArch64II::MO_PAGE);
451 ResultReg = createResultReg(&AArch64::GPR64spRegClass);
452 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri),
455 .addGlobalAddress(GV, 0, AArch64II::MO_PAGEOFF | AArch64II::MO_NC)
461 unsigned AArch64FastISel::fastMaterializeConstant(const Constant *C) {
462 EVT CEVT = TLI.getValueType(C->getType(), true);
464 // Only handle simple types.
465 if (!CEVT.isSimple())
467 MVT VT = CEVT.getSimpleVT();
469 if (const auto *CI = dyn_cast<ConstantInt>(C))
470 return materializeInt(CI, VT);
471 else if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
472 return materializeFP(CFP, VT);
473 else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
474 return materializeGV(GV);
479 unsigned AArch64FastISel::fastMaterializeFloatZero(const ConstantFP* CFP) {
480 assert(CFP->isNullValue() &&
481 "Floating-point constant is not a positive zero.");
483 if (!isTypeLegal(CFP->getType(), VT))
486 if (VT != MVT::f32 && VT != MVT::f64)
489 bool Is64Bit = (VT == MVT::f64);
490 unsigned ZReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
491 unsigned Opc = Is64Bit ? AArch64::FMOVXDr : AArch64::FMOVWSr;
492 return fastEmitInst_r(Opc, TLI.getRegClassFor(VT), ZReg, /*IsKill=*/true);
495 /// \brief Check if the multiply is by a power-of-2 constant.
496 static bool isMulPowOf2(const Value *I) {
497 if (const auto *MI = dyn_cast<MulOperator>(I)) {
498 if (const auto *C = dyn_cast<ConstantInt>(MI->getOperand(0)))
499 if (C->getValue().isPowerOf2())
501 if (const auto *C = dyn_cast<ConstantInt>(MI->getOperand(1)))
502 if (C->getValue().isPowerOf2())
508 // Computes the address to get to an object.
509 bool AArch64FastISel::computeAddress(const Value *Obj, Address &Addr, Type *Ty)
511 const User *U = nullptr;
512 unsigned Opcode = Instruction::UserOp1;
513 if (const Instruction *I = dyn_cast<Instruction>(Obj)) {
514 // Don't walk into other basic blocks unless the object is an alloca from
515 // another block, otherwise it may not have a virtual register assigned.
516 if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(Obj)) ||
517 FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
518 Opcode = I->getOpcode();
521 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Obj)) {
522 Opcode = C->getOpcode();
526 if (const PointerType *Ty = dyn_cast<PointerType>(Obj->getType()))
527 if (Ty->getAddressSpace() > 255)
528 // Fast instruction selection doesn't support the special
535 case Instruction::BitCast: {
536 // Look through bitcasts.
537 return computeAddress(U->getOperand(0), Addr, Ty);
539 case Instruction::IntToPtr: {
540 // Look past no-op inttoptrs.
541 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
542 return computeAddress(U->getOperand(0), Addr, Ty);
545 case Instruction::PtrToInt: {
546 // Look past no-op ptrtoints.
547 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
548 return computeAddress(U->getOperand(0), Addr, Ty);
551 case Instruction::GetElementPtr: {
552 Address SavedAddr = Addr;
553 uint64_t TmpOffset = Addr.getOffset();
555 // Iterate through the GEP folding the constants into offsets where
557 gep_type_iterator GTI = gep_type_begin(U);
558 for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e;
560 const Value *Op = *i;
561 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
562 const StructLayout *SL = DL.getStructLayout(STy);
563 unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
564 TmpOffset += SL->getElementOffset(Idx);
566 uint64_t S = DL.getTypeAllocSize(GTI.getIndexedType());
568 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
569 // Constant-offset addressing.
570 TmpOffset += CI->getSExtValue() * S;
573 if (canFoldAddIntoGEP(U, Op)) {
574 // A compatible add with a constant operand. Fold the constant.
576 cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
577 TmpOffset += CI->getSExtValue() * S;
578 // Iterate on the other operand.
579 Op = cast<AddOperator>(Op)->getOperand(0);
583 goto unsupported_gep;
588 // Try to grab the base operand now.
589 Addr.setOffset(TmpOffset);
590 if (computeAddress(U->getOperand(0), Addr, Ty))
593 // We failed, restore everything and try the other options.
599 case Instruction::Alloca: {
600 const AllocaInst *AI = cast<AllocaInst>(Obj);
601 DenseMap<const AllocaInst *, int>::iterator SI =
602 FuncInfo.StaticAllocaMap.find(AI);
603 if (SI != FuncInfo.StaticAllocaMap.end()) {
604 Addr.setKind(Address::FrameIndexBase);
605 Addr.setFI(SI->second);
610 case Instruction::Add: {
611 // Adds of constants are common and easy enough.
612 const Value *LHS = U->getOperand(0);
613 const Value *RHS = U->getOperand(1);
615 if (isa<ConstantInt>(LHS))
618 if (const ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
619 Addr.setOffset(Addr.getOffset() + CI->getSExtValue());
620 return computeAddress(LHS, Addr, Ty);
623 Address Backup = Addr;
624 if (computeAddress(LHS, Addr, Ty) && computeAddress(RHS, Addr, Ty))
630 case Instruction::Sub: {
631 // Subs of constants are common and easy enough.
632 const Value *LHS = U->getOperand(0);
633 const Value *RHS = U->getOperand(1);
635 if (const ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
636 Addr.setOffset(Addr.getOffset() - CI->getSExtValue());
637 return computeAddress(LHS, Addr, Ty);
641 case Instruction::Shl: {
642 if (Addr.getOffsetReg())
645 const auto *CI = dyn_cast<ConstantInt>(U->getOperand(1));
649 unsigned Val = CI->getZExtValue();
650 if (Val < 1 || Val > 3)
653 uint64_t NumBytes = 0;
654 if (Ty && Ty->isSized()) {
655 uint64_t NumBits = DL.getTypeSizeInBits(Ty);
656 NumBytes = NumBits / 8;
657 if (!isPowerOf2_64(NumBits))
661 if (NumBytes != (1ULL << Val))
665 Addr.setExtendType(AArch64_AM::LSL);
667 const Value *Src = U->getOperand(0);
668 if (const auto *I = dyn_cast<Instruction>(Src)) {
669 if (FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
670 // Fold the zext or sext when it won't become a noop.
671 if (const auto *ZE = dyn_cast<ZExtInst>(I)) {
672 if (!isIntExtFree(ZE) &&
673 ZE->getOperand(0)->getType()->isIntegerTy(32)) {
674 Addr.setExtendType(AArch64_AM::UXTW);
675 Src = ZE->getOperand(0);
677 } else if (const auto *SE = dyn_cast<SExtInst>(I)) {
678 if (!isIntExtFree(SE) &&
679 SE->getOperand(0)->getType()->isIntegerTy(32)) {
680 Addr.setExtendType(AArch64_AM::SXTW);
681 Src = SE->getOperand(0);
687 if (const auto *AI = dyn_cast<BinaryOperator>(Src))
688 if (AI->getOpcode() == Instruction::And) {
689 const Value *LHS = AI->getOperand(0);
690 const Value *RHS = AI->getOperand(1);
692 if (const auto *C = dyn_cast<ConstantInt>(LHS))
693 if (C->getValue() == 0xffffffff)
696 if (const auto *C = dyn_cast<ConstantInt>(RHS))
697 if (C->getValue() == 0xffffffff) {
698 Addr.setExtendType(AArch64_AM::UXTW);
699 unsigned Reg = getRegForValue(LHS);
702 bool RegIsKill = hasTrivialKill(LHS);
703 Reg = fastEmitInst_extractsubreg(MVT::i32, Reg, RegIsKill,
705 Addr.setOffsetReg(Reg);
710 unsigned Reg = getRegForValue(Src);
713 Addr.setOffsetReg(Reg);
716 case Instruction::Mul: {
717 if (Addr.getOffsetReg())
723 const Value *LHS = U->getOperand(0);
724 const Value *RHS = U->getOperand(1);
726 // Canonicalize power-of-2 value to the RHS.
727 if (const auto *C = dyn_cast<ConstantInt>(LHS))
728 if (C->getValue().isPowerOf2())
731 assert(isa<ConstantInt>(RHS) && "Expected an ConstantInt.");
732 const auto *C = cast<ConstantInt>(RHS);
733 unsigned Val = C->getValue().logBase2();
734 if (Val < 1 || Val > 3)
737 uint64_t NumBytes = 0;
738 if (Ty && Ty->isSized()) {
739 uint64_t NumBits = DL.getTypeSizeInBits(Ty);
740 NumBytes = NumBits / 8;
741 if (!isPowerOf2_64(NumBits))
745 if (NumBytes != (1ULL << Val))
749 Addr.setExtendType(AArch64_AM::LSL);
751 const Value *Src = LHS;
752 if (const auto *I = dyn_cast<Instruction>(Src)) {
753 if (FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
754 // Fold the zext or sext when it won't become a noop.
755 if (const auto *ZE = dyn_cast<ZExtInst>(I)) {
756 if (!isIntExtFree(ZE) &&
757 ZE->getOperand(0)->getType()->isIntegerTy(32)) {
758 Addr.setExtendType(AArch64_AM::UXTW);
759 Src = ZE->getOperand(0);
761 } else if (const auto *SE = dyn_cast<SExtInst>(I)) {
762 if (!isIntExtFree(SE) &&
763 SE->getOperand(0)->getType()->isIntegerTy(32)) {
764 Addr.setExtendType(AArch64_AM::SXTW);
765 Src = SE->getOperand(0);
771 unsigned Reg = getRegForValue(Src);
774 Addr.setOffsetReg(Reg);
777 case Instruction::And: {
778 if (Addr.getOffsetReg())
781 if (!Ty || DL.getTypeSizeInBits(Ty) != 8)
784 const Value *LHS = U->getOperand(0);
785 const Value *RHS = U->getOperand(1);
787 if (const auto *C = dyn_cast<ConstantInt>(LHS))
788 if (C->getValue() == 0xffffffff)
791 if (const auto *C = dyn_cast<ConstantInt>(RHS))
792 if (C->getValue() == 0xffffffff) {
794 Addr.setExtendType(AArch64_AM::LSL);
795 Addr.setExtendType(AArch64_AM::UXTW);
797 unsigned Reg = getRegForValue(LHS);
800 bool RegIsKill = hasTrivialKill(LHS);
801 Reg = fastEmitInst_extractsubreg(MVT::i32, Reg, RegIsKill,
803 Addr.setOffsetReg(Reg);
808 case Instruction::SExt:
809 case Instruction::ZExt: {
810 if (!Addr.getReg() || Addr.getOffsetReg())
813 const Value *Src = nullptr;
814 // Fold the zext or sext when it won't become a noop.
815 if (const auto *ZE = dyn_cast<ZExtInst>(U)) {
816 if (!isIntExtFree(ZE) && ZE->getOperand(0)->getType()->isIntegerTy(32)) {
817 Addr.setExtendType(AArch64_AM::UXTW);
818 Src = ZE->getOperand(0);
820 } else if (const auto *SE = dyn_cast<SExtInst>(U)) {
821 if (!isIntExtFree(SE) && SE->getOperand(0)->getType()->isIntegerTy(32)) {
822 Addr.setExtendType(AArch64_AM::SXTW);
823 Src = SE->getOperand(0);
831 unsigned Reg = getRegForValue(Src);
834 Addr.setOffsetReg(Reg);
839 if (Addr.isRegBase() && !Addr.getReg()) {
840 unsigned Reg = getRegForValue(Obj);
847 if (!Addr.getOffsetReg()) {
848 unsigned Reg = getRegForValue(Obj);
851 Addr.setOffsetReg(Reg);
858 bool AArch64FastISel::computeCallAddress(const Value *V, Address &Addr) {
859 const User *U = nullptr;
860 unsigned Opcode = Instruction::UserOp1;
863 if (const auto *I = dyn_cast<Instruction>(V)) {
864 Opcode = I->getOpcode();
866 InMBB = I->getParent() == FuncInfo.MBB->getBasicBlock();
867 } else if (const auto *C = dyn_cast<ConstantExpr>(V)) {
868 Opcode = C->getOpcode();
874 case Instruction::BitCast:
875 // Look past bitcasts if its operand is in the same BB.
877 return computeCallAddress(U->getOperand(0), Addr);
879 case Instruction::IntToPtr:
880 // Look past no-op inttoptrs if its operand is in the same BB.
882 TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
883 return computeCallAddress(U->getOperand(0), Addr);
885 case Instruction::PtrToInt:
886 // Look past no-op ptrtoints if its operand is in the same BB.
888 TLI.getValueType(U->getType()) == TLI.getPointerTy())
889 return computeCallAddress(U->getOperand(0), Addr);
893 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
894 Addr.setGlobalValue(GV);
898 // If all else fails, try to materialize the value in a register.
899 if (!Addr.getGlobalValue()) {
900 Addr.setReg(getRegForValue(V));
901 return Addr.getReg() != 0;
908 bool AArch64FastISel::isTypeLegal(Type *Ty, MVT &VT) {
909 EVT evt = TLI.getValueType(Ty, true);
911 // Only handle simple types.
912 if (evt == MVT::Other || !evt.isSimple())
914 VT = evt.getSimpleVT();
916 // This is a legal type, but it's not something we handle in fast-isel.
920 // Handle all other legal types, i.e. a register that will directly hold this
922 return TLI.isTypeLegal(VT);
925 /// \brief Determine if the value type is supported by FastISel.
927 /// FastISel for AArch64 can handle more value types than are legal. This adds
928 /// simple value type such as i1, i8, and i16.
929 bool AArch64FastISel::isTypeSupported(Type *Ty, MVT &VT, bool IsVectorAllowed) {
930 if (Ty->isVectorTy() && !IsVectorAllowed)
933 if (isTypeLegal(Ty, VT))
936 // If this is a type than can be sign or zero-extended to a basic operation
937 // go ahead and accept it now.
938 if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16)
944 bool AArch64FastISel::isValueAvailable(const Value *V) const {
945 if (!isa<Instruction>(V))
948 const auto *I = cast<Instruction>(V);
949 if (FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB)
955 bool AArch64FastISel::simplifyAddress(Address &Addr, MVT VT) {
956 unsigned ScaleFactor = getImplicitScaleFactor(VT);
960 bool ImmediateOffsetNeedsLowering = false;
961 bool RegisterOffsetNeedsLowering = false;
962 int64_t Offset = Addr.getOffset();
963 if (((Offset < 0) || (Offset & (ScaleFactor - 1))) && !isInt<9>(Offset))
964 ImmediateOffsetNeedsLowering = true;
965 else if (Offset > 0 && !(Offset & (ScaleFactor - 1)) &&
966 !isUInt<12>(Offset / ScaleFactor))
967 ImmediateOffsetNeedsLowering = true;
969 // Cannot encode an offset register and an immediate offset in the same
970 // instruction. Fold the immediate offset into the load/store instruction and
971 // emit an additonal add to take care of the offset register.
972 if (!ImmediateOffsetNeedsLowering && Addr.getOffset() && Addr.getOffsetReg())
973 RegisterOffsetNeedsLowering = true;
975 // Cannot encode zero register as base.
976 if (Addr.isRegBase() && Addr.getOffsetReg() && !Addr.getReg())
977 RegisterOffsetNeedsLowering = true;
979 // If this is a stack pointer and the offset needs to be simplified then put
980 // the alloca address into a register, set the base type back to register and
981 // continue. This should almost never happen.
982 if ((ImmediateOffsetNeedsLowering || Addr.getOffsetReg()) && Addr.isFIBase())
984 unsigned ResultReg = createResultReg(&AArch64::GPR64spRegClass);
985 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri),
987 .addFrameIndex(Addr.getFI())
990 Addr.setKind(Address::RegBase);
991 Addr.setReg(ResultReg);
994 if (RegisterOffsetNeedsLowering) {
995 unsigned ResultReg = 0;
997 if (Addr.getExtendType() == AArch64_AM::SXTW ||
998 Addr.getExtendType() == AArch64_AM::UXTW )
999 ResultReg = emitAddSub_rx(/*UseAdd=*/true, MVT::i64, Addr.getReg(),
1000 /*TODO:IsKill=*/false, Addr.getOffsetReg(),
1001 /*TODO:IsKill=*/false, Addr.getExtendType(),
1004 ResultReg = emitAddSub_rs(/*UseAdd=*/true, MVT::i64, Addr.getReg(),
1005 /*TODO:IsKill=*/false, Addr.getOffsetReg(),
1006 /*TODO:IsKill=*/false, AArch64_AM::LSL,
1009 if (Addr.getExtendType() == AArch64_AM::UXTW)
1010 ResultReg = emitLSL_ri(MVT::i64, MVT::i32, Addr.getOffsetReg(),
1011 /*Op0IsKill=*/false, Addr.getShift(),
1013 else if (Addr.getExtendType() == AArch64_AM::SXTW)
1014 ResultReg = emitLSL_ri(MVT::i64, MVT::i32, Addr.getOffsetReg(),
1015 /*Op0IsKill=*/false, Addr.getShift(),
1018 ResultReg = emitLSL_ri(MVT::i64, MVT::i64, Addr.getOffsetReg(),
1019 /*Op0IsKill=*/false, Addr.getShift());
1024 Addr.setReg(ResultReg);
1025 Addr.setOffsetReg(0);
1027 Addr.setExtendType(AArch64_AM::InvalidShiftExtend);
1030 // Since the offset is too large for the load/store instruction get the
1031 // reg+offset into a register.
1032 if (ImmediateOffsetNeedsLowering) {
1035 // Try to fold the immediate into the add instruction.
1036 ResultReg = emitAdd_ri_(MVT::i64, Addr.getReg(), /*IsKill=*/false, Offset);
1038 ResultReg = fastEmit_i(MVT::i64, MVT::i64, ISD::Constant, Offset);
1042 Addr.setReg(ResultReg);
1048 void AArch64FastISel::addLoadStoreOperands(Address &Addr,
1049 const MachineInstrBuilder &MIB,
1051 unsigned ScaleFactor,
1052 MachineMemOperand *MMO) {
1053 int64_t Offset = Addr.getOffset() / ScaleFactor;
1054 // Frame base works a bit differently. Handle it separately.
1055 if (Addr.isFIBase()) {
1056 int FI = Addr.getFI();
1057 // FIXME: We shouldn't be using getObjectSize/getObjectAlignment. The size
1058 // and alignment should be based on the VT.
1059 MMO = FuncInfo.MF->getMachineMemOperand(
1060 MachinePointerInfo::getFixedStack(FI, Offset), Flags,
1061 MFI.getObjectSize(FI), MFI.getObjectAlignment(FI));
1062 // Now add the rest of the operands.
1063 MIB.addFrameIndex(FI).addImm(Offset);
1065 assert(Addr.isRegBase() && "Unexpected address kind.");
1066 const MCInstrDesc &II = MIB->getDesc();
1067 unsigned Idx = (Flags & MachineMemOperand::MOStore) ? 1 : 0;
1069 constrainOperandRegClass(II, Addr.getReg(), II.getNumDefs()+Idx));
1071 constrainOperandRegClass(II, Addr.getOffsetReg(), II.getNumDefs()+Idx+1));
1072 if (Addr.getOffsetReg()) {
1073 assert(Addr.getOffset() == 0 && "Unexpected offset");
1074 bool IsSigned = Addr.getExtendType() == AArch64_AM::SXTW ||
1075 Addr.getExtendType() == AArch64_AM::SXTX;
1076 MIB.addReg(Addr.getReg());
1077 MIB.addReg(Addr.getOffsetReg());
1078 MIB.addImm(IsSigned);
1079 MIB.addImm(Addr.getShift() != 0);
1081 MIB.addReg(Addr.getReg()).addImm(Offset);
1085 MIB.addMemOperand(MMO);
1088 unsigned AArch64FastISel::emitAddSub(bool UseAdd, MVT RetVT, const Value *LHS,
1089 const Value *RHS, bool SetFlags,
1090 bool WantResult, bool IsZExt) {
1091 AArch64_AM::ShiftExtendType ExtendType = AArch64_AM::InvalidShiftExtend;
1092 bool NeedExtend = false;
1093 switch (RetVT.SimpleTy) {
1101 ExtendType = IsZExt ? AArch64_AM::UXTB : AArch64_AM::SXTB;
1105 ExtendType = IsZExt ? AArch64_AM::UXTH : AArch64_AM::SXTH;
1107 case MVT::i32: // fall-through
1112 RetVT.SimpleTy = std::max(RetVT.SimpleTy, MVT::i32);
1114 // Canonicalize immediates to the RHS first.
1115 if (UseAdd && isa<Constant>(LHS) && !isa<Constant>(RHS))
1116 std::swap(LHS, RHS);
1118 // Canonicalize mul by power of 2 to the RHS.
1119 if (UseAdd && LHS->hasOneUse() && isValueAvailable(LHS))
1120 if (isMulPowOf2(LHS))
1121 std::swap(LHS, RHS);
1123 // Canonicalize shift immediate to the RHS.
1124 if (UseAdd && LHS->hasOneUse() && isValueAvailable(LHS))
1125 if (const auto *SI = dyn_cast<BinaryOperator>(LHS))
1126 if (isa<ConstantInt>(SI->getOperand(1)))
1127 if (SI->getOpcode() == Instruction::Shl ||
1128 SI->getOpcode() == Instruction::LShr ||
1129 SI->getOpcode() == Instruction::AShr )
1130 std::swap(LHS, RHS);
1132 unsigned LHSReg = getRegForValue(LHS);
1135 bool LHSIsKill = hasTrivialKill(LHS);
1138 LHSReg = emitIntExt(SrcVT, LHSReg, RetVT, IsZExt);
1140 unsigned ResultReg = 0;
1141 if (const auto *C = dyn_cast<ConstantInt>(RHS)) {
1142 uint64_t Imm = IsZExt ? C->getZExtValue() : C->getSExtValue();
1143 if (C->isNegative())
1144 ResultReg = emitAddSub_ri(!UseAdd, RetVT, LHSReg, LHSIsKill, -Imm,
1145 SetFlags, WantResult);
1147 ResultReg = emitAddSub_ri(UseAdd, RetVT, LHSReg, LHSIsKill, Imm, SetFlags,
1149 } else if (const auto *C = dyn_cast<Constant>(RHS))
1150 if (C->isNullValue())
1151 ResultReg = emitAddSub_ri(UseAdd, RetVT, LHSReg, LHSIsKill, 0, SetFlags,
1157 // Only extend the RHS within the instruction if there is a valid extend type.
1158 if (ExtendType != AArch64_AM::InvalidShiftExtend && RHS->hasOneUse() &&
1159 isValueAvailable(RHS)) {
1160 if (const auto *SI = dyn_cast<BinaryOperator>(RHS))
1161 if (const auto *C = dyn_cast<ConstantInt>(SI->getOperand(1)))
1162 if ((SI->getOpcode() == Instruction::Shl) && (C->getZExtValue() < 4)) {
1163 unsigned RHSReg = getRegForValue(SI->getOperand(0));
1166 bool RHSIsKill = hasTrivialKill(SI->getOperand(0));
1167 return emitAddSub_rx(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg,
1168 RHSIsKill, ExtendType, C->getZExtValue(),
1169 SetFlags, WantResult);
1171 unsigned RHSReg = getRegForValue(RHS);
1174 bool RHSIsKill = hasTrivialKill(RHS);
1175 return emitAddSub_rx(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg, RHSIsKill,
1176 ExtendType, 0, SetFlags, WantResult);
1179 // Check if the mul can be folded into the instruction.
1180 if (RHS->hasOneUse() && isValueAvailable(RHS))
1181 if (isMulPowOf2(RHS)) {
1182 const Value *MulLHS = cast<MulOperator>(RHS)->getOperand(0);
1183 const Value *MulRHS = cast<MulOperator>(RHS)->getOperand(1);
1185 if (const auto *C = dyn_cast<ConstantInt>(MulLHS))
1186 if (C->getValue().isPowerOf2())
1187 std::swap(MulLHS, MulRHS);
1189 assert(isa<ConstantInt>(MulRHS) && "Expected a ConstantInt.");
1190 uint64_t ShiftVal = cast<ConstantInt>(MulRHS)->getValue().logBase2();
1191 unsigned RHSReg = getRegForValue(MulLHS);
1194 bool RHSIsKill = hasTrivialKill(MulLHS);
1195 return emitAddSub_rs(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg, RHSIsKill,
1196 AArch64_AM::LSL, ShiftVal, SetFlags, WantResult);
1199 // Check if the shift can be folded into the instruction.
1200 if (RHS->hasOneUse() && isValueAvailable(RHS))
1201 if (const auto *SI = dyn_cast<BinaryOperator>(RHS)) {
1202 if (const auto *C = dyn_cast<ConstantInt>(SI->getOperand(1))) {
1203 AArch64_AM::ShiftExtendType ShiftType = AArch64_AM::InvalidShiftExtend;
1204 switch (SI->getOpcode()) {
1206 case Instruction::Shl: ShiftType = AArch64_AM::LSL; break;
1207 case Instruction::LShr: ShiftType = AArch64_AM::LSR; break;
1208 case Instruction::AShr: ShiftType = AArch64_AM::ASR; break;
1210 uint64_t ShiftVal = C->getZExtValue();
1211 if (ShiftType != AArch64_AM::InvalidShiftExtend) {
1212 unsigned RHSReg = getRegForValue(SI->getOperand(0));
1215 bool RHSIsKill = hasTrivialKill(SI->getOperand(0));
1216 return emitAddSub_rs(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg,
1217 RHSIsKill, ShiftType, ShiftVal, SetFlags,
1223 unsigned RHSReg = getRegForValue(RHS);
1226 bool RHSIsKill = hasTrivialKill(RHS);
1229 RHSReg = emitIntExt(SrcVT, RHSReg, RetVT, IsZExt);
1231 return emitAddSub_rr(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg, RHSIsKill,
1232 SetFlags, WantResult);
1235 unsigned AArch64FastISel::emitAddSub_rr(bool UseAdd, MVT RetVT, unsigned LHSReg,
1236 bool LHSIsKill, unsigned RHSReg,
1237 bool RHSIsKill, bool SetFlags,
1239 assert(LHSReg && RHSReg && "Invalid register number.");
1241 if (RetVT != MVT::i32 && RetVT != MVT::i64)
1244 static const unsigned OpcTable[2][2][2] = {
1245 { { AArch64::SUBWrr, AArch64::SUBXrr },
1246 { AArch64::ADDWrr, AArch64::ADDXrr } },
1247 { { AArch64::SUBSWrr, AArch64::SUBSXrr },
1248 { AArch64::ADDSWrr, AArch64::ADDSXrr } }
1250 bool Is64Bit = RetVT == MVT::i64;
1251 unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit];
1252 const TargetRegisterClass *RC =
1253 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
1256 ResultReg = createResultReg(RC);
1258 ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
1260 const MCInstrDesc &II = TII.get(Opc);
1261 LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
1262 RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1);
1263 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
1264 .addReg(LHSReg, getKillRegState(LHSIsKill))
1265 .addReg(RHSReg, getKillRegState(RHSIsKill));
1269 unsigned AArch64FastISel::emitAddSub_ri(bool UseAdd, MVT RetVT, unsigned LHSReg,
1270 bool LHSIsKill, uint64_t Imm,
1271 bool SetFlags, bool WantResult) {
1272 assert(LHSReg && "Invalid register number.");
1274 if (RetVT != MVT::i32 && RetVT != MVT::i64)
1278 if (isUInt<12>(Imm))
1280 else if ((Imm & 0xfff000) == Imm) {
1286 static const unsigned OpcTable[2][2][2] = {
1287 { { AArch64::SUBWri, AArch64::SUBXri },
1288 { AArch64::ADDWri, AArch64::ADDXri } },
1289 { { AArch64::SUBSWri, AArch64::SUBSXri },
1290 { AArch64::ADDSWri, AArch64::ADDSXri } }
1292 bool Is64Bit = RetVT == MVT::i64;
1293 unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit];
1294 const TargetRegisterClass *RC;
1296 RC = Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
1298 RC = Is64Bit ? &AArch64::GPR64spRegClass : &AArch64::GPR32spRegClass;
1301 ResultReg = createResultReg(RC);
1303 ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
1305 const MCInstrDesc &II = TII.get(Opc);
1306 LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
1307 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
1308 .addReg(LHSReg, getKillRegState(LHSIsKill))
1310 .addImm(getShifterImm(AArch64_AM::LSL, ShiftImm));
1314 unsigned AArch64FastISel::emitAddSub_rs(bool UseAdd, MVT RetVT, unsigned LHSReg,
1315 bool LHSIsKill, unsigned RHSReg,
1317 AArch64_AM::ShiftExtendType ShiftType,
1318 uint64_t ShiftImm, bool SetFlags,
1320 assert(LHSReg && RHSReg && "Invalid register number.");
1322 if (RetVT != MVT::i32 && RetVT != MVT::i64)
1325 static const unsigned OpcTable[2][2][2] = {
1326 { { AArch64::SUBWrs, AArch64::SUBXrs },
1327 { AArch64::ADDWrs, AArch64::ADDXrs } },
1328 { { AArch64::SUBSWrs, AArch64::SUBSXrs },
1329 { AArch64::ADDSWrs, AArch64::ADDSXrs } }
1331 bool Is64Bit = RetVT == MVT::i64;
1332 unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit];
1333 const TargetRegisterClass *RC =
1334 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
1337 ResultReg = createResultReg(RC);
1339 ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
1341 const MCInstrDesc &II = TII.get(Opc);
1342 LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
1343 RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1);
1344 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
1345 .addReg(LHSReg, getKillRegState(LHSIsKill))
1346 .addReg(RHSReg, getKillRegState(RHSIsKill))
1347 .addImm(getShifterImm(ShiftType, ShiftImm));
1351 unsigned AArch64FastISel::emitAddSub_rx(bool UseAdd, MVT RetVT, unsigned LHSReg,
1352 bool LHSIsKill, unsigned RHSReg,
1354 AArch64_AM::ShiftExtendType ExtType,
1355 uint64_t ShiftImm, bool SetFlags,
1357 assert(LHSReg && RHSReg && "Invalid register number.");
1359 if (RetVT != MVT::i32 && RetVT != MVT::i64)
1362 static const unsigned OpcTable[2][2][2] = {
1363 { { AArch64::SUBWrx, AArch64::SUBXrx },
1364 { AArch64::ADDWrx, AArch64::ADDXrx } },
1365 { { AArch64::SUBSWrx, AArch64::SUBSXrx },
1366 { AArch64::ADDSWrx, AArch64::ADDSXrx } }
1368 bool Is64Bit = RetVT == MVT::i64;
1369 unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit];
1370 const TargetRegisterClass *RC = nullptr;
1372 RC = Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
1374 RC = Is64Bit ? &AArch64::GPR64spRegClass : &AArch64::GPR32spRegClass;
1377 ResultReg = createResultReg(RC);
1379 ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
1381 const MCInstrDesc &II = TII.get(Opc);
1382 LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
1383 RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1);
1384 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
1385 .addReg(LHSReg, getKillRegState(LHSIsKill))
1386 .addReg(RHSReg, getKillRegState(RHSIsKill))
1387 .addImm(getArithExtendImm(ExtType, ShiftImm));
1391 bool AArch64FastISel::emitCmp(const Value *LHS, const Value *RHS, bool IsZExt) {
1392 Type *Ty = LHS->getType();
1393 EVT EVT = TLI.getValueType(Ty, true);
1394 if (!EVT.isSimple())
1396 MVT VT = EVT.getSimpleVT();
1398 switch (VT.SimpleTy) {
1406 return emitICmp(VT, LHS, RHS, IsZExt);
1409 return emitFCmp(VT, LHS, RHS);
1413 bool AArch64FastISel::emitICmp(MVT RetVT, const Value *LHS, const Value *RHS,
1415 return emitSub(RetVT, LHS, RHS, /*SetFlags=*/true, /*WantResult=*/false,
1419 bool AArch64FastISel::emitICmp_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill,
1421 return emitAddSub_ri(/*UseAdd=*/false, RetVT, LHSReg, LHSIsKill, Imm,
1422 /*SetFlags=*/true, /*WantResult=*/false) != 0;
1425 bool AArch64FastISel::emitFCmp(MVT RetVT, const Value *LHS, const Value *RHS) {
1426 if (RetVT != MVT::f32 && RetVT != MVT::f64)
1429 // Check to see if the 2nd operand is a constant that we can encode directly
1431 bool UseImm = false;
1432 if (const auto *CFP = dyn_cast<ConstantFP>(RHS))
1433 if (CFP->isZero() && !CFP->isNegative())
1436 unsigned LHSReg = getRegForValue(LHS);
1439 bool LHSIsKill = hasTrivialKill(LHS);
1442 unsigned Opc = (RetVT == MVT::f64) ? AArch64::FCMPDri : AArch64::FCMPSri;
1443 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc))
1444 .addReg(LHSReg, getKillRegState(LHSIsKill));
1448 unsigned RHSReg = getRegForValue(RHS);
1451 bool RHSIsKill = hasTrivialKill(RHS);
1453 unsigned Opc = (RetVT == MVT::f64) ? AArch64::FCMPDrr : AArch64::FCMPSrr;
1454 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc))
1455 .addReg(LHSReg, getKillRegState(LHSIsKill))
1456 .addReg(RHSReg, getKillRegState(RHSIsKill));
1460 unsigned AArch64FastISel::emitAdd(MVT RetVT, const Value *LHS, const Value *RHS,
1461 bool SetFlags, bool WantResult, bool IsZExt) {
1462 return emitAddSub(/*UseAdd=*/true, RetVT, LHS, RHS, SetFlags, WantResult,
1466 /// \brief This method is a wrapper to simplify add emission.
1468 /// First try to emit an add with an immediate operand using emitAddSub_ri. If
1469 /// that fails, then try to materialize the immediate into a register and use
1470 /// emitAddSub_rr instead.
1471 unsigned AArch64FastISel::emitAdd_ri_(MVT VT, unsigned Op0, bool Op0IsKill,
1475 ResultReg = emitAddSub_ri(false, VT, Op0, Op0IsKill, -Imm);
1477 ResultReg = emitAddSub_ri(true, VT, Op0, Op0IsKill, Imm);
1482 unsigned CReg = fastEmit_i(VT, VT, ISD::Constant, Imm);
1486 ResultReg = emitAddSub_rr(true, VT, Op0, Op0IsKill, CReg, true);
1490 unsigned AArch64FastISel::emitSub(MVT RetVT, const Value *LHS, const Value *RHS,
1491 bool SetFlags, bool WantResult, bool IsZExt) {
1492 return emitAddSub(/*UseAdd=*/false, RetVT, LHS, RHS, SetFlags, WantResult,
1496 unsigned AArch64FastISel::emitSubs_rr(MVT RetVT, unsigned LHSReg,
1497 bool LHSIsKill, unsigned RHSReg,
1498 bool RHSIsKill, bool WantResult) {
1499 return emitAddSub_rr(/*UseAdd=*/false, RetVT, LHSReg, LHSIsKill, RHSReg,
1500 RHSIsKill, /*SetFlags=*/true, WantResult);
1503 unsigned AArch64FastISel::emitSubs_rs(MVT RetVT, unsigned LHSReg,
1504 bool LHSIsKill, unsigned RHSReg,
1506 AArch64_AM::ShiftExtendType ShiftType,
1507 uint64_t ShiftImm, bool WantResult) {
1508 return emitAddSub_rs(/*UseAdd=*/false, RetVT, LHSReg, LHSIsKill, RHSReg,
1509 RHSIsKill, ShiftType, ShiftImm, /*SetFlags=*/true,
1513 unsigned AArch64FastISel::emitLogicalOp(unsigned ISDOpc, MVT RetVT,
1514 const Value *LHS, const Value *RHS) {
1515 // Canonicalize immediates to the RHS first.
1516 if (isa<ConstantInt>(LHS) && !isa<ConstantInt>(RHS))
1517 std::swap(LHS, RHS);
1519 // Canonicalize mul by power-of-2 to the RHS.
1520 if (LHS->hasOneUse() && isValueAvailable(LHS))
1521 if (isMulPowOf2(LHS))
1522 std::swap(LHS, RHS);
1524 // Canonicalize shift immediate to the RHS.
1525 if (LHS->hasOneUse() && isValueAvailable(LHS))
1526 if (const auto *SI = dyn_cast<ShlOperator>(LHS))
1527 if (isa<ConstantInt>(SI->getOperand(1)))
1528 std::swap(LHS, RHS);
1530 unsigned LHSReg = getRegForValue(LHS);
1533 bool LHSIsKill = hasTrivialKill(LHS);
1535 unsigned ResultReg = 0;
1536 if (const auto *C = dyn_cast<ConstantInt>(RHS)) {
1537 uint64_t Imm = C->getZExtValue();
1538 ResultReg = emitLogicalOp_ri(ISDOpc, RetVT, LHSReg, LHSIsKill, Imm);
1543 // Check if the mul can be folded into the instruction.
1544 if (RHS->hasOneUse() && isValueAvailable(RHS))
1545 if (isMulPowOf2(RHS)) {
1546 const Value *MulLHS = cast<MulOperator>(RHS)->getOperand(0);
1547 const Value *MulRHS = cast<MulOperator>(RHS)->getOperand(1);
1549 if (const auto *C = dyn_cast<ConstantInt>(MulLHS))
1550 if (C->getValue().isPowerOf2())
1551 std::swap(MulLHS, MulRHS);
1553 assert(isa<ConstantInt>(MulRHS) && "Expected a ConstantInt.");
1554 uint64_t ShiftVal = cast<ConstantInt>(MulRHS)->getValue().logBase2();
1556 unsigned RHSReg = getRegForValue(MulLHS);
1559 bool RHSIsKill = hasTrivialKill(MulLHS);
1560 return emitLogicalOp_rs(ISDOpc, RetVT, LHSReg, LHSIsKill, RHSReg,
1561 RHSIsKill, ShiftVal);
1564 // Check if the shift can be folded into the instruction.
1565 if (RHS->hasOneUse() && isValueAvailable(RHS))
1566 if (const auto *SI = dyn_cast<ShlOperator>(RHS))
1567 if (const auto *C = dyn_cast<ConstantInt>(SI->getOperand(1))) {
1568 uint64_t ShiftVal = C->getZExtValue();
1569 unsigned RHSReg = getRegForValue(SI->getOperand(0));
1572 bool RHSIsKill = hasTrivialKill(SI->getOperand(0));
1573 return emitLogicalOp_rs(ISDOpc, RetVT, LHSReg, LHSIsKill, RHSReg,
1574 RHSIsKill, ShiftVal);
1577 unsigned RHSReg = getRegForValue(RHS);
1580 bool RHSIsKill = hasTrivialKill(RHS);
1582 MVT VT = std::max(MVT::i32, RetVT.SimpleTy);
1583 ResultReg = fastEmit_rr(VT, VT, ISDOpc, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
1584 if (RetVT >= MVT::i8 && RetVT <= MVT::i16) {
1585 uint64_t Mask = (RetVT == MVT::i8) ? 0xff : 0xffff;
1586 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
1591 unsigned AArch64FastISel::emitLogicalOp_ri(unsigned ISDOpc, MVT RetVT,
1592 unsigned LHSReg, bool LHSIsKill,
1594 assert((ISD::AND + 1 == ISD::OR) && (ISD::AND + 2 == ISD::XOR) &&
1595 "ISD nodes are not consecutive!");
1596 static const unsigned OpcTable[3][2] = {
1597 { AArch64::ANDWri, AArch64::ANDXri },
1598 { AArch64::ORRWri, AArch64::ORRXri },
1599 { AArch64::EORWri, AArch64::EORXri }
1601 const TargetRegisterClass *RC;
1604 switch (RetVT.SimpleTy) {
1611 unsigned Idx = ISDOpc - ISD::AND;
1612 Opc = OpcTable[Idx][0];
1613 RC = &AArch64::GPR32spRegClass;
1618 Opc = OpcTable[ISDOpc - ISD::AND][1];
1619 RC = &AArch64::GPR64spRegClass;
1624 if (!AArch64_AM::isLogicalImmediate(Imm, RegSize))
1627 unsigned ResultReg =
1628 fastEmitInst_ri(Opc, RC, LHSReg, LHSIsKill,
1629 AArch64_AM::encodeLogicalImmediate(Imm, RegSize));
1630 if (RetVT >= MVT::i8 && RetVT <= MVT::i16 && ISDOpc != ISD::AND) {
1631 uint64_t Mask = (RetVT == MVT::i8) ? 0xff : 0xffff;
1632 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
1637 unsigned AArch64FastISel::emitLogicalOp_rs(unsigned ISDOpc, MVT RetVT,
1638 unsigned LHSReg, bool LHSIsKill,
1639 unsigned RHSReg, bool RHSIsKill,
1640 uint64_t ShiftImm) {
1641 assert((ISD::AND + 1 == ISD::OR) && (ISD::AND + 2 == ISD::XOR) &&
1642 "ISD nodes are not consecutive!");
1643 static const unsigned OpcTable[3][2] = {
1644 { AArch64::ANDWrs, AArch64::ANDXrs },
1645 { AArch64::ORRWrs, AArch64::ORRXrs },
1646 { AArch64::EORWrs, AArch64::EORXrs }
1648 const TargetRegisterClass *RC;
1650 switch (RetVT.SimpleTy) {
1657 Opc = OpcTable[ISDOpc - ISD::AND][0];
1658 RC = &AArch64::GPR32RegClass;
1661 Opc = OpcTable[ISDOpc - ISD::AND][1];
1662 RC = &AArch64::GPR64RegClass;
1665 unsigned ResultReg =
1666 fastEmitInst_rri(Opc, RC, LHSReg, LHSIsKill, RHSReg, RHSIsKill,
1667 AArch64_AM::getShifterImm(AArch64_AM::LSL, ShiftImm));
1668 if (RetVT >= MVT::i8 && RetVT <= MVT::i16) {
1669 uint64_t Mask = (RetVT == MVT::i8) ? 0xff : 0xffff;
1670 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
1675 unsigned AArch64FastISel::emitAnd_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill,
1677 return emitLogicalOp_ri(ISD::AND, RetVT, LHSReg, LHSIsKill, Imm);
1680 unsigned AArch64FastISel::emitLoad(MVT VT, MVT RetVT, Address Addr,
1681 bool WantZExt, MachineMemOperand *MMO) {
1682 if(!TLI.allowsMisalignedMemoryAccesses(VT))
1685 // Simplify this down to something we can handle.
1686 if (!simplifyAddress(Addr, VT))
1689 unsigned ScaleFactor = getImplicitScaleFactor(VT);
1691 llvm_unreachable("Unexpected value type.");
1693 // Negative offsets require unscaled, 9-bit, signed immediate offsets.
1694 // Otherwise, we try using scaled, 12-bit, unsigned immediate offsets.
1695 bool UseScaled = true;
1696 if ((Addr.getOffset() < 0) || (Addr.getOffset() & (ScaleFactor - 1))) {
1701 static const unsigned GPOpcTable[2][8][4] = {
1703 { { AArch64::LDURSBWi, AArch64::LDURSHWi, AArch64::LDURWi,
1705 { AArch64::LDURSBXi, AArch64::LDURSHXi, AArch64::LDURSWi,
1707 { AArch64::LDRSBWui, AArch64::LDRSHWui, AArch64::LDRWui,
1709 { AArch64::LDRSBXui, AArch64::LDRSHXui, AArch64::LDRSWui,
1711 { AArch64::LDRSBWroX, AArch64::LDRSHWroX, AArch64::LDRWroX,
1713 { AArch64::LDRSBXroX, AArch64::LDRSHXroX, AArch64::LDRSWroX,
1715 { AArch64::LDRSBWroW, AArch64::LDRSHWroW, AArch64::LDRWroW,
1717 { AArch64::LDRSBXroW, AArch64::LDRSHXroW, AArch64::LDRSWroW,
1721 { { AArch64::LDURBBi, AArch64::LDURHHi, AArch64::LDURWi,
1723 { AArch64::LDURBBi, AArch64::LDURHHi, AArch64::LDURWi,
1725 { AArch64::LDRBBui, AArch64::LDRHHui, AArch64::LDRWui,
1727 { AArch64::LDRBBui, AArch64::LDRHHui, AArch64::LDRWui,
1729 { AArch64::LDRBBroX, AArch64::LDRHHroX, AArch64::LDRWroX,
1731 { AArch64::LDRBBroX, AArch64::LDRHHroX, AArch64::LDRWroX,
1733 { AArch64::LDRBBroW, AArch64::LDRHHroW, AArch64::LDRWroW,
1735 { AArch64::LDRBBroW, AArch64::LDRHHroW, AArch64::LDRWroW,
1740 static const unsigned FPOpcTable[4][2] = {
1741 { AArch64::LDURSi, AArch64::LDURDi },
1742 { AArch64::LDRSui, AArch64::LDRDui },
1743 { AArch64::LDRSroX, AArch64::LDRDroX },
1744 { AArch64::LDRSroW, AArch64::LDRDroW }
1748 const TargetRegisterClass *RC;
1749 bool UseRegOffset = Addr.isRegBase() && !Addr.getOffset() && Addr.getReg() &&
1750 Addr.getOffsetReg();
1751 unsigned Idx = UseRegOffset ? 2 : UseScaled ? 1 : 0;
1752 if (Addr.getExtendType() == AArch64_AM::UXTW ||
1753 Addr.getExtendType() == AArch64_AM::SXTW)
1756 bool IsRet64Bit = RetVT == MVT::i64;
1757 switch (VT.SimpleTy) {
1759 llvm_unreachable("Unexpected value type.");
1760 case MVT::i1: // Intentional fall-through.
1762 Opc = GPOpcTable[WantZExt][2 * Idx + IsRet64Bit][0];
1763 RC = (IsRet64Bit && !WantZExt) ?
1764 &AArch64::GPR64RegClass: &AArch64::GPR32RegClass;
1767 Opc = GPOpcTable[WantZExt][2 * Idx + IsRet64Bit][1];
1768 RC = (IsRet64Bit && !WantZExt) ?
1769 &AArch64::GPR64RegClass: &AArch64::GPR32RegClass;
1772 Opc = GPOpcTable[WantZExt][2 * Idx + IsRet64Bit][2];
1773 RC = (IsRet64Bit && !WantZExt) ?
1774 &AArch64::GPR64RegClass: &AArch64::GPR32RegClass;
1777 Opc = GPOpcTable[WantZExt][2 * Idx + IsRet64Bit][3];
1778 RC = &AArch64::GPR64RegClass;
1781 Opc = FPOpcTable[Idx][0];
1782 RC = &AArch64::FPR32RegClass;
1785 Opc = FPOpcTable[Idx][1];
1786 RC = &AArch64::FPR64RegClass;
1790 // Create the base instruction, then add the operands.
1791 unsigned ResultReg = createResultReg(RC);
1792 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1793 TII.get(Opc), ResultReg);
1794 addLoadStoreOperands(Addr, MIB, MachineMemOperand::MOLoad, ScaleFactor, MMO);
1796 // Loading an i1 requires special handling.
1797 if (VT == MVT::i1) {
1798 unsigned ANDReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, 1);
1799 assert(ANDReg && "Unexpected AND instruction emission failure.");
1803 // For zero-extending loads to 64bit we emit a 32bit load and then convert
1804 // the 32bit reg to a 64bit reg.
1805 if (WantZExt && RetVT == MVT::i64 && VT <= MVT::i32) {
1806 unsigned Reg64 = createResultReg(&AArch64::GPR64RegClass);
1807 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1808 TII.get(AArch64::SUBREG_TO_REG), Reg64)
1810 .addReg(ResultReg, getKillRegState(true))
1811 .addImm(AArch64::sub_32);
1817 bool AArch64FastISel::selectAddSub(const Instruction *I) {
1819 if (!isTypeSupported(I->getType(), VT, /*IsVectorAllowed=*/true))
1823 return selectOperator(I, I->getOpcode());
1826 switch (I->getOpcode()) {
1828 llvm_unreachable("Unexpected instruction.");
1829 case Instruction::Add:
1830 ResultReg = emitAdd(VT, I->getOperand(0), I->getOperand(1));
1832 case Instruction::Sub:
1833 ResultReg = emitSub(VT, I->getOperand(0), I->getOperand(1));
1839 updateValueMap(I, ResultReg);
1843 bool AArch64FastISel::selectLogicalOp(const Instruction *I) {
1845 if (!isTypeSupported(I->getType(), VT, /*IsVectorAllowed=*/true))
1849 return selectOperator(I, I->getOpcode());
1852 switch (I->getOpcode()) {
1854 llvm_unreachable("Unexpected instruction.");
1855 case Instruction::And:
1856 ResultReg = emitLogicalOp(ISD::AND, VT, I->getOperand(0), I->getOperand(1));
1858 case Instruction::Or:
1859 ResultReg = emitLogicalOp(ISD::OR, VT, I->getOperand(0), I->getOperand(1));
1861 case Instruction::Xor:
1862 ResultReg = emitLogicalOp(ISD::XOR, VT, I->getOperand(0), I->getOperand(1));
1868 updateValueMap(I, ResultReg);
1872 bool AArch64FastISel::selectLoad(const Instruction *I) {
1874 // Verify we have a legal type before going any further. Currently, we handle
1875 // simple types that will directly fit in a register (i32/f32/i64/f64) or
1876 // those that can be sign or zero-extended to a basic operation (i1/i8/i16).
1877 if (!isTypeSupported(I->getType(), VT, /*IsVectorAllowed=*/true) ||
1878 cast<LoadInst>(I)->isAtomic())
1881 // See if we can handle this address.
1883 if (!computeAddress(I->getOperand(0), Addr, I->getType()))
1886 // Fold the following sign-/zero-extend into the load instruction.
1887 bool WantZExt = true;
1889 const Value *IntExtVal = nullptr;
1890 if (I->hasOneUse()) {
1891 if (const auto *ZE = dyn_cast<ZExtInst>(I->use_begin()->getUser())) {
1892 if (isTypeSupported(ZE->getType(), RetVT))
1896 } else if (const auto *SE = dyn_cast<SExtInst>(I->use_begin()->getUser())) {
1897 if (isTypeSupported(SE->getType(), RetVT))
1905 unsigned ResultReg =
1906 emitLoad(VT, RetVT, Addr, WantZExt, createMachineMemOperandFor(I));
1910 // There are a few different cases we have to handle, because the load or the
1911 // sign-/zero-extend might not be selected by FastISel if we fall-back to
1912 // SelectionDAG. There is also an ordering issue when both instructions are in
1913 // different basic blocks.
1914 // 1.) The load instruction is selected by FastISel, but the integer extend
1915 // not. This usually happens when the integer extend is in a different
1916 // basic block and SelectionDAG took over for that basic block.
1917 // 2.) The load instruction is selected before the integer extend. This only
1918 // happens when the integer extend is in a different basic block.
1919 // 3.) The load instruction is selected by SelectionDAG and the integer extend
1920 // by FastISel. This happens if there are instructions between the load
1921 // and the integer extend that couldn't be selected by FastISel.
1923 // The integer extend hasn't been emitted yet. FastISel or SelectionDAG
1924 // could select it. Emit a copy to subreg if necessary. FastISel will remove
1925 // it when it selects the integer extend.
1926 unsigned Reg = lookUpRegForValue(IntExtVal);
1927 auto *MI = MRI.getUniqueVRegDef(Reg);
1929 if (RetVT == MVT::i64 && VT <= MVT::i32) {
1931 // Delete the last emitted instruction from emitLoad (SUBREG_TO_REG).
1932 std::prev(FuncInfo.InsertPt)->eraseFromParent();
1933 ResultReg = std::prev(FuncInfo.InsertPt)->getOperand(0).getReg();
1935 ResultReg = fastEmitInst_extractsubreg(MVT::i32, ResultReg,
1939 updateValueMap(I, ResultReg);
1943 // The integer extend has already been emitted - delete all the instructions
1944 // that have been emitted by the integer extend lowering code and use the
1945 // result from the load instruction directly.
1948 for (auto &Opnd : MI->uses()) {
1950 Reg = Opnd.getReg();
1954 MI->eraseFromParent();
1957 MI = MRI.getUniqueVRegDef(Reg);
1959 updateValueMap(IntExtVal, ResultReg);
1963 updateValueMap(I, ResultReg);
1967 bool AArch64FastISel::emitStore(MVT VT, unsigned SrcReg, Address Addr,
1968 MachineMemOperand *MMO) {
1969 if(!TLI.allowsMisalignedMemoryAccesses(VT))
1972 // Simplify this down to something we can handle.
1973 if (!simplifyAddress(Addr, VT))
1976 unsigned ScaleFactor = getImplicitScaleFactor(VT);
1978 llvm_unreachable("Unexpected value type.");
1980 // Negative offsets require unscaled, 9-bit, signed immediate offsets.
1981 // Otherwise, we try using scaled, 12-bit, unsigned immediate offsets.
1982 bool UseScaled = true;
1983 if ((Addr.getOffset() < 0) || (Addr.getOffset() & (ScaleFactor - 1))) {
1988 static const unsigned OpcTable[4][6] = {
1989 { AArch64::STURBBi, AArch64::STURHHi, AArch64::STURWi, AArch64::STURXi,
1990 AArch64::STURSi, AArch64::STURDi },
1991 { AArch64::STRBBui, AArch64::STRHHui, AArch64::STRWui, AArch64::STRXui,
1992 AArch64::STRSui, AArch64::STRDui },
1993 { AArch64::STRBBroX, AArch64::STRHHroX, AArch64::STRWroX, AArch64::STRXroX,
1994 AArch64::STRSroX, AArch64::STRDroX },
1995 { AArch64::STRBBroW, AArch64::STRHHroW, AArch64::STRWroW, AArch64::STRXroW,
1996 AArch64::STRSroW, AArch64::STRDroW }
2000 bool VTIsi1 = false;
2001 bool UseRegOffset = Addr.isRegBase() && !Addr.getOffset() && Addr.getReg() &&
2002 Addr.getOffsetReg();
2003 unsigned Idx = UseRegOffset ? 2 : UseScaled ? 1 : 0;
2004 if (Addr.getExtendType() == AArch64_AM::UXTW ||
2005 Addr.getExtendType() == AArch64_AM::SXTW)
2008 switch (VT.SimpleTy) {
2009 default: llvm_unreachable("Unexpected value type.");
2010 case MVT::i1: VTIsi1 = true;
2011 case MVT::i8: Opc = OpcTable[Idx][0]; break;
2012 case MVT::i16: Opc = OpcTable[Idx][1]; break;
2013 case MVT::i32: Opc = OpcTable[Idx][2]; break;
2014 case MVT::i64: Opc = OpcTable[Idx][3]; break;
2015 case MVT::f32: Opc = OpcTable[Idx][4]; break;
2016 case MVT::f64: Opc = OpcTable[Idx][5]; break;
2019 // Storing an i1 requires special handling.
2020 if (VTIsi1 && SrcReg != AArch64::WZR) {
2021 unsigned ANDReg = emitAnd_ri(MVT::i32, SrcReg, /*TODO:IsKill=*/false, 1);
2022 assert(ANDReg && "Unexpected AND instruction emission failure.");
2025 // Create the base instruction, then add the operands.
2026 const MCInstrDesc &II = TII.get(Opc);
2027 SrcReg = constrainOperandRegClass(II, SrcReg, II.getNumDefs());
2028 MachineInstrBuilder MIB =
2029 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II).addReg(SrcReg);
2030 addLoadStoreOperands(Addr, MIB, MachineMemOperand::MOStore, ScaleFactor, MMO);
2035 bool AArch64FastISel::selectStore(const Instruction *I) {
2037 const Value *Op0 = I->getOperand(0);
2038 // Verify we have a legal type before going any further. Currently, we handle
2039 // simple types that will directly fit in a register (i32/f32/i64/f64) or
2040 // those that can be sign or zero-extended to a basic operation (i1/i8/i16).
2041 if (!isTypeSupported(Op0->getType(), VT, /*IsVectorAllowed=*/true) ||
2042 cast<StoreInst>(I)->isAtomic())
2045 // Get the value to be stored into a register. Use the zero register directly
2046 // when possible to avoid an unnecessary copy and a wasted register.
2047 unsigned SrcReg = 0;
2048 if (const auto *CI = dyn_cast<ConstantInt>(Op0)) {
2050 SrcReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
2051 } else if (const auto *CF = dyn_cast<ConstantFP>(Op0)) {
2052 if (CF->isZero() && !CF->isNegative()) {
2053 VT = MVT::getIntegerVT(VT.getSizeInBits());
2054 SrcReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
2059 SrcReg = getRegForValue(Op0);
2064 // See if we can handle this address.
2066 if (!computeAddress(I->getOperand(1), Addr, I->getOperand(0)->getType()))
2069 if (!emitStore(VT, SrcReg, Addr, createMachineMemOperandFor(I)))
2074 static AArch64CC::CondCode getCompareCC(CmpInst::Predicate Pred) {
2076 case CmpInst::FCMP_ONE:
2077 case CmpInst::FCMP_UEQ:
2079 // AL is our "false" for now. The other two need more compares.
2080 return AArch64CC::AL;
2081 case CmpInst::ICMP_EQ:
2082 case CmpInst::FCMP_OEQ:
2083 return AArch64CC::EQ;
2084 case CmpInst::ICMP_SGT:
2085 case CmpInst::FCMP_OGT:
2086 return AArch64CC::GT;
2087 case CmpInst::ICMP_SGE:
2088 case CmpInst::FCMP_OGE:
2089 return AArch64CC::GE;
2090 case CmpInst::ICMP_UGT:
2091 case CmpInst::FCMP_UGT:
2092 return AArch64CC::HI;
2093 case CmpInst::FCMP_OLT:
2094 return AArch64CC::MI;
2095 case CmpInst::ICMP_ULE:
2096 case CmpInst::FCMP_OLE:
2097 return AArch64CC::LS;
2098 case CmpInst::FCMP_ORD:
2099 return AArch64CC::VC;
2100 case CmpInst::FCMP_UNO:
2101 return AArch64CC::VS;
2102 case CmpInst::FCMP_UGE:
2103 return AArch64CC::PL;
2104 case CmpInst::ICMP_SLT:
2105 case CmpInst::FCMP_ULT:
2106 return AArch64CC::LT;
2107 case CmpInst::ICMP_SLE:
2108 case CmpInst::FCMP_ULE:
2109 return AArch64CC::LE;
2110 case CmpInst::FCMP_UNE:
2111 case CmpInst::ICMP_NE:
2112 return AArch64CC::NE;
2113 case CmpInst::ICMP_UGE:
2114 return AArch64CC::HS;
2115 case CmpInst::ICMP_ULT:
2116 return AArch64CC::LO;
2120 /// \brief Try to emit a combined compare-and-branch instruction.
2121 bool AArch64FastISel::emitCompareAndBranch(const BranchInst *BI) {
2122 assert(isa<CmpInst>(BI->getCondition()) && "Expected cmp instruction");
2123 const CmpInst *CI = cast<CmpInst>(BI->getCondition());
2124 CmpInst::Predicate Predicate = optimizeCmpPredicate(CI);
2126 const Value *LHS = CI->getOperand(0);
2127 const Value *RHS = CI->getOperand(1);
2130 if (!isTypeSupported(LHS->getType(), VT))
2133 unsigned BW = VT.getSizeInBits();
2137 MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
2138 MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
2140 // Try to take advantage of fallthrough opportunities.
2141 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
2142 std::swap(TBB, FBB);
2143 Predicate = CmpInst::getInversePredicate(Predicate);
2148 switch (Predicate) {
2151 case CmpInst::ICMP_EQ:
2152 case CmpInst::ICMP_NE:
2153 if (isa<Constant>(LHS) && cast<Constant>(LHS)->isNullValue())
2154 std::swap(LHS, RHS);
2156 if (!isa<Constant>(RHS) || !cast<Constant>(RHS)->isNullValue())
2159 if (const auto *AI = dyn_cast<BinaryOperator>(LHS))
2160 if (AI->getOpcode() == Instruction::And && isValueAvailable(AI)) {
2161 const Value *AndLHS = AI->getOperand(0);
2162 const Value *AndRHS = AI->getOperand(1);
2164 if (const auto *C = dyn_cast<ConstantInt>(AndLHS))
2165 if (C->getValue().isPowerOf2())
2166 std::swap(AndLHS, AndRHS);
2168 if (const auto *C = dyn_cast<ConstantInt>(AndRHS))
2169 if (C->getValue().isPowerOf2()) {
2170 TestBit = C->getValue().logBase2();
2178 IsCmpNE = Predicate == CmpInst::ICMP_NE;
2180 case CmpInst::ICMP_SLT:
2181 case CmpInst::ICMP_SGE:
2182 if (!isa<Constant>(RHS) || !cast<Constant>(RHS)->isNullValue())
2186 IsCmpNE = Predicate == CmpInst::ICMP_SLT;
2188 case CmpInst::ICMP_SGT:
2189 case CmpInst::ICMP_SLE:
2190 if (!isa<ConstantInt>(RHS))
2193 if (cast<ConstantInt>(RHS)->getValue() != APInt(BW, -1, true))
2197 IsCmpNE = Predicate == CmpInst::ICMP_SLE;
2201 static const unsigned OpcTable[2][2][2] = {
2202 { {AArch64::CBZW, AArch64::CBZX },
2203 {AArch64::CBNZW, AArch64::CBNZX} },
2204 { {AArch64::TBZW, AArch64::TBZX },
2205 {AArch64::TBNZW, AArch64::TBNZX} }
2208 bool IsBitTest = TestBit != -1;
2209 bool Is64Bit = BW == 64;
2210 if (TestBit < 32 && TestBit >= 0)
2213 unsigned Opc = OpcTable[IsBitTest][IsCmpNE][Is64Bit];
2214 const MCInstrDesc &II = TII.get(Opc);
2216 unsigned SrcReg = getRegForValue(LHS);
2219 bool SrcIsKill = hasTrivialKill(LHS);
2221 if (BW == 64 && !Is64Bit)
2222 SrcReg = fastEmitInst_extractsubreg(MVT::i32, SrcReg, SrcIsKill,
2225 if ((BW < 32) && !IsBitTest)
2226 SrcReg = emitIntExt(VT, SrcReg, MVT::i32, /*IsZExt=*/true);
2228 // Emit the combined compare and branch instruction.
2229 SrcReg = constrainOperandRegClass(II, SrcReg, II.getNumDefs());
2230 MachineInstrBuilder MIB =
2231 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc))
2232 .addReg(SrcReg, getKillRegState(SrcIsKill));
2234 MIB.addImm(TestBit);
2237 // Obtain the branch weight and add the TrueBB to the successor list.
2238 uint32_t BranchWeight = 0;
2240 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
2241 TBB->getBasicBlock());
2242 FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
2243 fastEmitBranch(FBB, DbgLoc);
2248 bool AArch64FastISel::selectBranch(const Instruction *I) {
2249 const BranchInst *BI = cast<BranchInst>(I);
2250 if (BI->isUnconditional()) {
2251 MachineBasicBlock *MSucc = FuncInfo.MBBMap[BI->getSuccessor(0)];
2252 fastEmitBranch(MSucc, BI->getDebugLoc());
2256 MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
2257 MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
2259 AArch64CC::CondCode CC = AArch64CC::NE;
2260 if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
2261 if (CI->hasOneUse() && isValueAvailable(CI)) {
2262 // Try to optimize or fold the cmp.
2263 CmpInst::Predicate Predicate = optimizeCmpPredicate(CI);
2264 switch (Predicate) {
2267 case CmpInst::FCMP_FALSE:
2268 fastEmitBranch(FBB, DbgLoc);
2270 case CmpInst::FCMP_TRUE:
2271 fastEmitBranch(TBB, DbgLoc);
2275 // Try to emit a combined compare-and-branch first.
2276 if (emitCompareAndBranch(BI))
2279 // Try to take advantage of fallthrough opportunities.
2280 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
2281 std::swap(TBB, FBB);
2282 Predicate = CmpInst::getInversePredicate(Predicate);
2286 if (!emitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
2289 // FCMP_UEQ and FCMP_ONE cannot be checked with a single branch
2291 CC = getCompareCC(Predicate);
2292 AArch64CC::CondCode ExtraCC = AArch64CC::AL;
2293 switch (Predicate) {
2296 case CmpInst::FCMP_UEQ:
2297 ExtraCC = AArch64CC::EQ;
2300 case CmpInst::FCMP_ONE:
2301 ExtraCC = AArch64CC::MI;
2305 assert((CC != AArch64CC::AL) && "Unexpected condition code.");
2307 // Emit the extra branch for FCMP_UEQ and FCMP_ONE.
2308 if (ExtraCC != AArch64CC::AL) {
2309 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
2315 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
2319 // Obtain the branch weight and add the TrueBB to the successor list.
2320 uint32_t BranchWeight = 0;
2322 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
2323 TBB->getBasicBlock());
2324 FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
2326 fastEmitBranch(FBB, DbgLoc);
2329 } else if (TruncInst *TI = dyn_cast<TruncInst>(BI->getCondition())) {
2331 if (TI->hasOneUse() && isValueAvailable(TI) &&
2332 isTypeSupported(TI->getOperand(0)->getType(), SrcVT)) {
2333 unsigned CondReg = getRegForValue(TI->getOperand(0));
2336 bool CondIsKill = hasTrivialKill(TI->getOperand(0));
2338 // Issue an extract_subreg to get the lower 32-bits.
2339 if (SrcVT == MVT::i64) {
2340 CondReg = fastEmitInst_extractsubreg(MVT::i32, CondReg, CondIsKill,
2345 unsigned ANDReg = emitAnd_ri(MVT::i32, CondReg, CondIsKill, 1);
2346 assert(ANDReg && "Unexpected AND instruction emission failure.");
2347 emitICmp_ri(MVT::i32, ANDReg, /*IsKill=*/true, 0);
2349 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
2350 std::swap(TBB, FBB);
2353 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
2357 // Obtain the branch weight and add the TrueBB to the successor list.
2358 uint32_t BranchWeight = 0;
2360 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
2361 TBB->getBasicBlock());
2362 FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
2364 fastEmitBranch(FBB, DbgLoc);
2367 } else if (const auto *CI = dyn_cast<ConstantInt>(BI->getCondition())) {
2368 uint64_t Imm = CI->getZExtValue();
2369 MachineBasicBlock *Target = (Imm == 0) ? FBB : TBB;
2370 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::B))
2373 // Obtain the branch weight and add the target to the successor list.
2374 uint32_t BranchWeight = 0;
2376 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
2377 Target->getBasicBlock());
2378 FuncInfo.MBB->addSuccessor(Target, BranchWeight);
2380 } else if (foldXALUIntrinsic(CC, I, BI->getCondition())) {
2381 // Fake request the condition, otherwise the intrinsic might be completely
2383 unsigned CondReg = getRegForValue(BI->getCondition());
2388 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
2392 // Obtain the branch weight and add the TrueBB to the successor list.
2393 uint32_t BranchWeight = 0;
2395 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
2396 TBB->getBasicBlock());
2397 FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
2399 fastEmitBranch(FBB, DbgLoc);
2403 unsigned CondReg = getRegForValue(BI->getCondition());
2406 bool CondRegIsKill = hasTrivialKill(BI->getCondition());
2408 // We've been divorced from our compare! Our block was split, and
2409 // now our compare lives in a predecessor block. We musn't
2410 // re-compare here, as the children of the compare aren't guaranteed
2411 // live across the block boundary (we *could* check for this).
2412 // Regardless, the compare has been done in the predecessor block,
2413 // and it left a value for us in a virtual register. Ergo, we test
2414 // the one-bit value left in the virtual register.
2415 emitICmp_ri(MVT::i32, CondReg, CondRegIsKill, 0);
2417 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
2418 std::swap(TBB, FBB);
2422 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
2426 // Obtain the branch weight and add the TrueBB to the successor list.
2427 uint32_t BranchWeight = 0;
2429 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
2430 TBB->getBasicBlock());
2431 FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
2433 fastEmitBranch(FBB, DbgLoc);
2437 bool AArch64FastISel::selectIndirectBr(const Instruction *I) {
2438 const IndirectBrInst *BI = cast<IndirectBrInst>(I);
2439 unsigned AddrReg = getRegForValue(BI->getOperand(0));
2443 // Emit the indirect branch.
2444 const MCInstrDesc &II = TII.get(AArch64::BR);
2445 AddrReg = constrainOperandRegClass(II, AddrReg, II.getNumDefs());
2446 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II).addReg(AddrReg);
2448 // Make sure the CFG is up-to-date.
2449 for (unsigned i = 0, e = BI->getNumSuccessors(); i != e; ++i)
2450 FuncInfo.MBB->addSuccessor(FuncInfo.MBBMap[BI->getSuccessor(i)]);
2455 bool AArch64FastISel::selectCmp(const Instruction *I) {
2456 const CmpInst *CI = cast<CmpInst>(I);
2458 // Try to optimize or fold the cmp.
2459 CmpInst::Predicate Predicate = optimizeCmpPredicate(CI);
2460 unsigned ResultReg = 0;
2461 switch (Predicate) {
2464 case CmpInst::FCMP_FALSE:
2465 ResultReg = createResultReg(&AArch64::GPR32RegClass);
2466 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2467 TII.get(TargetOpcode::COPY), ResultReg)
2468 .addReg(AArch64::WZR, getKillRegState(true));
2470 case CmpInst::FCMP_TRUE:
2471 ResultReg = fastEmit_i(MVT::i32, MVT::i32, ISD::Constant, 1);
2476 updateValueMap(I, ResultReg);
2481 if (!emitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
2484 ResultReg = createResultReg(&AArch64::GPR32RegClass);
2486 // FCMP_UEQ and FCMP_ONE cannot be checked with a single instruction. These
2487 // condition codes are inverted, because they are used by CSINC.
2488 static unsigned CondCodeTable[2][2] = {
2489 { AArch64CC::NE, AArch64CC::VC },
2490 { AArch64CC::PL, AArch64CC::LE }
2492 unsigned *CondCodes = nullptr;
2493 switch (Predicate) {
2496 case CmpInst::FCMP_UEQ:
2497 CondCodes = &CondCodeTable[0][0];
2499 case CmpInst::FCMP_ONE:
2500 CondCodes = &CondCodeTable[1][0];
2505 unsigned TmpReg1 = createResultReg(&AArch64::GPR32RegClass);
2506 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::CSINCWr),
2508 .addReg(AArch64::WZR, getKillRegState(true))
2509 .addReg(AArch64::WZR, getKillRegState(true))
2510 .addImm(CondCodes[0]);
2511 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::CSINCWr),
2513 .addReg(TmpReg1, getKillRegState(true))
2514 .addReg(AArch64::WZR, getKillRegState(true))
2515 .addImm(CondCodes[1]);
2517 updateValueMap(I, ResultReg);
2521 // Now set a register based on the comparison.
2522 AArch64CC::CondCode CC = getCompareCC(Predicate);
2523 assert((CC != AArch64CC::AL) && "Unexpected condition code.");
2524 AArch64CC::CondCode invertedCC = getInvertedCondCode(CC);
2525 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::CSINCWr),
2527 .addReg(AArch64::WZR, getKillRegState(true))
2528 .addReg(AArch64::WZR, getKillRegState(true))
2529 .addImm(invertedCC);
2531 updateValueMap(I, ResultReg);
2535 /// \brief Optimize selects of i1 if one of the operands has a 'true' or 'false'
2537 bool AArch64FastISel::optimizeSelect(const SelectInst *SI) {
2538 if (!SI->getType()->isIntegerTy(1))
2541 const Value *Src1Val, *Src2Val;
2543 bool NeedExtraOp = false;
2544 if (auto *CI = dyn_cast<ConstantInt>(SI->getTrueValue())) {
2546 Src1Val = SI->getCondition();
2547 Src2Val = SI->getFalseValue();
2548 Opc = AArch64::ORRWrr;
2550 assert(CI->isZero());
2551 Src1Val = SI->getFalseValue();
2552 Src2Val = SI->getCondition();
2553 Opc = AArch64::BICWrr;
2555 } else if (auto *CI = dyn_cast<ConstantInt>(SI->getFalseValue())) {
2557 Src1Val = SI->getCondition();
2558 Src2Val = SI->getTrueValue();
2559 Opc = AArch64::ORRWrr;
2562 assert(CI->isZero());
2563 Src1Val = SI->getCondition();
2564 Src2Val = SI->getTrueValue();
2565 Opc = AArch64::ANDWrr;
2572 unsigned Src1Reg = getRegForValue(Src1Val);
2575 bool Src1IsKill = hasTrivialKill(Src1Val);
2577 unsigned Src2Reg = getRegForValue(Src2Val);
2580 bool Src2IsKill = hasTrivialKill(Src2Val);
2583 Src1Reg = emitLogicalOp_ri(ISD::XOR, MVT::i32, Src1Reg, Src1IsKill, 1);
2586 unsigned ResultReg = fastEmitInst_rr(Opc, &AArch64::GPR32RegClass, Src1Reg,
2587 Src1IsKill, Src2Reg, Src2IsKill);
2588 updateValueMap(SI, ResultReg);
2592 bool AArch64FastISel::selectSelect(const Instruction *I) {
2593 assert(isa<SelectInst>(I) && "Expected a select instruction.");
2595 if (!isTypeSupported(I->getType(), VT))
2599 const TargetRegisterClass *RC;
2600 switch (VT.SimpleTy) {
2607 Opc = AArch64::CSELWr;
2608 RC = &AArch64::GPR32RegClass;
2611 Opc = AArch64::CSELXr;
2612 RC = &AArch64::GPR64RegClass;
2615 Opc = AArch64::FCSELSrrr;
2616 RC = &AArch64::FPR32RegClass;
2619 Opc = AArch64::FCSELDrrr;
2620 RC = &AArch64::FPR64RegClass;
2624 const SelectInst *SI = cast<SelectInst>(I);
2625 const Value *Cond = SI->getCondition();
2626 AArch64CC::CondCode CC = AArch64CC::NE;
2627 AArch64CC::CondCode ExtraCC = AArch64CC::AL;
2629 if (optimizeSelect(SI))
2632 // Try to pickup the flags, so we don't have to emit another compare.
2633 if (foldXALUIntrinsic(CC, I, Cond)) {
2634 // Fake request the condition to force emission of the XALU intrinsic.
2635 unsigned CondReg = getRegForValue(Cond);
2638 } else if (isa<CmpInst>(Cond) && cast<CmpInst>(Cond)->hasOneUse() &&
2639 isValueAvailable(Cond)) {
2640 const auto *Cmp = cast<CmpInst>(Cond);
2641 // Try to optimize or fold the cmp.
2642 CmpInst::Predicate Predicate = optimizeCmpPredicate(Cmp);
2643 const Value *FoldSelect = nullptr;
2644 switch (Predicate) {
2647 case CmpInst::FCMP_FALSE:
2648 FoldSelect = SI->getFalseValue();
2650 case CmpInst::FCMP_TRUE:
2651 FoldSelect = SI->getTrueValue();
2656 unsigned SrcReg = getRegForValue(FoldSelect);
2659 unsigned UseReg = lookUpRegForValue(SI);
2661 MRI.clearKillFlags(UseReg);
2663 updateValueMap(I, SrcReg);
2668 if (!emitCmp(Cmp->getOperand(0), Cmp->getOperand(1), Cmp->isUnsigned()))
2671 // FCMP_UEQ and FCMP_ONE cannot be checked with a single select instruction.
2672 CC = getCompareCC(Predicate);
2673 switch (Predicate) {
2676 case CmpInst::FCMP_UEQ:
2677 ExtraCC = AArch64CC::EQ;
2680 case CmpInst::FCMP_ONE:
2681 ExtraCC = AArch64CC::MI;
2685 assert((CC != AArch64CC::AL) && "Unexpected condition code.");
2687 unsigned CondReg = getRegForValue(Cond);
2690 bool CondIsKill = hasTrivialKill(Cond);
2692 const MCInstrDesc &II = TII.get(AArch64::ANDSWri);
2693 CondReg = constrainOperandRegClass(II, CondReg, 1);
2695 // Emit a TST instruction (ANDS wzr, reg, #imm).
2696 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II,
2698 .addReg(CondReg, getKillRegState(CondIsKill))
2699 .addImm(AArch64_AM::encodeLogicalImmediate(1, 32));
2702 unsigned Src1Reg = getRegForValue(SI->getTrueValue());
2703 bool Src1IsKill = hasTrivialKill(SI->getTrueValue());
2705 unsigned Src2Reg = getRegForValue(SI->getFalseValue());
2706 bool Src2IsKill = hasTrivialKill(SI->getFalseValue());
2708 if (!Src1Reg || !Src2Reg)
2711 if (ExtraCC != AArch64CC::AL) {
2712 Src2Reg = fastEmitInst_rri(Opc, RC, Src1Reg, Src1IsKill, Src2Reg,
2713 Src2IsKill, ExtraCC);
2716 unsigned ResultReg = fastEmitInst_rri(Opc, RC, Src1Reg, Src1IsKill, Src2Reg,
2718 updateValueMap(I, ResultReg);
2722 bool AArch64FastISel::selectFPExt(const Instruction *I) {
2723 Value *V = I->getOperand(0);
2724 if (!I->getType()->isDoubleTy() || !V->getType()->isFloatTy())
2727 unsigned Op = getRegForValue(V);
2731 unsigned ResultReg = createResultReg(&AArch64::FPR64RegClass);
2732 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::FCVTDSr),
2733 ResultReg).addReg(Op);
2734 updateValueMap(I, ResultReg);
2738 bool AArch64FastISel::selectFPTrunc(const Instruction *I) {
2739 Value *V = I->getOperand(0);
2740 if (!I->getType()->isFloatTy() || !V->getType()->isDoubleTy())
2743 unsigned Op = getRegForValue(V);
2747 unsigned ResultReg = createResultReg(&AArch64::FPR32RegClass);
2748 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::FCVTSDr),
2749 ResultReg).addReg(Op);
2750 updateValueMap(I, ResultReg);
2754 // FPToUI and FPToSI
2755 bool AArch64FastISel::selectFPToInt(const Instruction *I, bool Signed) {
2757 if (!isTypeLegal(I->getType(), DestVT) || DestVT.isVector())
2760 unsigned SrcReg = getRegForValue(I->getOperand(0));
2764 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType(), true);
2765 if (SrcVT == MVT::f128)
2769 if (SrcVT == MVT::f64) {
2771 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZSUWDr : AArch64::FCVTZSUXDr;
2773 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZUUWDr : AArch64::FCVTZUUXDr;
2776 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZSUWSr : AArch64::FCVTZSUXSr;
2778 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZUUWSr : AArch64::FCVTZUUXSr;
2780 unsigned ResultReg = createResultReg(
2781 DestVT == MVT::i32 ? &AArch64::GPR32RegClass : &AArch64::GPR64RegClass);
2782 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
2784 updateValueMap(I, ResultReg);
2788 bool AArch64FastISel::selectIntToFP(const Instruction *I, bool Signed) {
2790 if (!isTypeLegal(I->getType(), DestVT) || DestVT.isVector())
2792 assert ((DestVT == MVT::f32 || DestVT == MVT::f64) &&
2793 "Unexpected value type.");
2795 unsigned SrcReg = getRegForValue(I->getOperand(0));
2798 bool SrcIsKill = hasTrivialKill(I->getOperand(0));
2800 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType(), true);
2802 // Handle sign-extension.
2803 if (SrcVT == MVT::i16 || SrcVT == MVT::i8 || SrcVT == MVT::i1) {
2805 emitIntExt(SrcVT.getSimpleVT(), SrcReg, MVT::i32, /*isZExt*/ !Signed);
2812 if (SrcVT == MVT::i64) {
2814 Opc = (DestVT == MVT::f32) ? AArch64::SCVTFUXSri : AArch64::SCVTFUXDri;
2816 Opc = (DestVT == MVT::f32) ? AArch64::UCVTFUXSri : AArch64::UCVTFUXDri;
2819 Opc = (DestVT == MVT::f32) ? AArch64::SCVTFUWSri : AArch64::SCVTFUWDri;
2821 Opc = (DestVT == MVT::f32) ? AArch64::UCVTFUWSri : AArch64::UCVTFUWDri;
2824 unsigned ResultReg = fastEmitInst_r(Opc, TLI.getRegClassFor(DestVT), SrcReg,
2826 updateValueMap(I, ResultReg);
2830 bool AArch64FastISel::fastLowerArguments() {
2831 if (!FuncInfo.CanLowerReturn)
2834 const Function *F = FuncInfo.Fn;
2838 CallingConv::ID CC = F->getCallingConv();
2839 if (CC != CallingConv::C)
2842 // Only handle simple cases of up to 8 GPR and FPR each.
2843 unsigned GPRCnt = 0;
2844 unsigned FPRCnt = 0;
2846 for (auto const &Arg : F->args()) {
2847 // The first argument is at index 1.
2849 if (F->getAttributes().hasAttribute(Idx, Attribute::ByVal) ||
2850 F->getAttributes().hasAttribute(Idx, Attribute::InReg) ||
2851 F->getAttributes().hasAttribute(Idx, Attribute::StructRet) ||
2852 F->getAttributes().hasAttribute(Idx, Attribute::Nest))
2855 Type *ArgTy = Arg.getType();
2856 if (ArgTy->isStructTy() || ArgTy->isArrayTy())
2859 EVT ArgVT = TLI.getValueType(ArgTy);
2860 if (!ArgVT.isSimple())
2863 MVT VT = ArgVT.getSimpleVT().SimpleTy;
2864 if (VT.isFloatingPoint() && !Subtarget->hasFPARMv8())
2867 if (VT.isVector() &&
2868 (!Subtarget->hasNEON() || !Subtarget->isLittleEndian()))
2871 if (VT >= MVT::i1 && VT <= MVT::i64)
2873 else if ((VT >= MVT::f16 && VT <= MVT::f64) || VT.is64BitVector() ||
2874 VT.is128BitVector())
2879 if (GPRCnt > 8 || FPRCnt > 8)
2883 static const MCPhysReg Registers[6][8] = {
2884 { AArch64::W0, AArch64::W1, AArch64::W2, AArch64::W3, AArch64::W4,
2885 AArch64::W5, AArch64::W6, AArch64::W7 },
2886 { AArch64::X0, AArch64::X1, AArch64::X2, AArch64::X3, AArch64::X4,
2887 AArch64::X5, AArch64::X6, AArch64::X7 },
2888 { AArch64::H0, AArch64::H1, AArch64::H2, AArch64::H3, AArch64::H4,
2889 AArch64::H5, AArch64::H6, AArch64::H7 },
2890 { AArch64::S0, AArch64::S1, AArch64::S2, AArch64::S3, AArch64::S4,
2891 AArch64::S5, AArch64::S6, AArch64::S7 },
2892 { AArch64::D0, AArch64::D1, AArch64::D2, AArch64::D3, AArch64::D4,
2893 AArch64::D5, AArch64::D6, AArch64::D7 },
2894 { AArch64::Q0, AArch64::Q1, AArch64::Q2, AArch64::Q3, AArch64::Q4,
2895 AArch64::Q5, AArch64::Q6, AArch64::Q7 }
2898 unsigned GPRIdx = 0;
2899 unsigned FPRIdx = 0;
2900 for (auto const &Arg : F->args()) {
2901 MVT VT = TLI.getSimpleValueType(Arg.getType());
2903 const TargetRegisterClass *RC;
2904 if (VT >= MVT::i1 && VT <= MVT::i32) {
2905 SrcReg = Registers[0][GPRIdx++];
2906 RC = &AArch64::GPR32RegClass;
2908 } else if (VT == MVT::i64) {
2909 SrcReg = Registers[1][GPRIdx++];
2910 RC = &AArch64::GPR64RegClass;
2911 } else if (VT == MVT::f16) {
2912 SrcReg = Registers[2][FPRIdx++];
2913 RC = &AArch64::FPR16RegClass;
2914 } else if (VT == MVT::f32) {
2915 SrcReg = Registers[3][FPRIdx++];
2916 RC = &AArch64::FPR32RegClass;
2917 } else if ((VT == MVT::f64) || VT.is64BitVector()) {
2918 SrcReg = Registers[4][FPRIdx++];
2919 RC = &AArch64::FPR64RegClass;
2920 } else if (VT.is128BitVector()) {
2921 SrcReg = Registers[5][FPRIdx++];
2922 RC = &AArch64::FPR128RegClass;
2924 llvm_unreachable("Unexpected value type.");
2926 unsigned DstReg = FuncInfo.MF->addLiveIn(SrcReg, RC);
2927 // FIXME: Unfortunately it's necessary to emit a copy from the livein copy.
2928 // Without this, EmitLiveInCopies may eliminate the livein if its only
2929 // use is a bitcast (which isn't turned into an instruction).
2930 unsigned ResultReg = createResultReg(RC);
2931 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2932 TII.get(TargetOpcode::COPY), ResultReg)
2933 .addReg(DstReg, getKillRegState(true));
2934 updateValueMap(&Arg, ResultReg);
2939 bool AArch64FastISel::processCallArgs(CallLoweringInfo &CLI,
2940 SmallVectorImpl<MVT> &OutVTs,
2941 unsigned &NumBytes) {
2942 CallingConv::ID CC = CLI.CallConv;
2943 SmallVector<CCValAssign, 16> ArgLocs;
2944 CCState CCInfo(CC, false, *FuncInfo.MF, ArgLocs, *Context);
2945 CCInfo.AnalyzeCallOperands(OutVTs, CLI.OutFlags, CCAssignFnForCall(CC));
2947 // Get a count of how many bytes are to be pushed on the stack.
2948 NumBytes = CCInfo.getNextStackOffset();
2950 // Issue CALLSEQ_START
2951 unsigned AdjStackDown = TII.getCallFrameSetupOpcode();
2952 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackDown))
2955 // Process the args.
2956 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2957 CCValAssign &VA = ArgLocs[i];
2958 const Value *ArgVal = CLI.OutVals[VA.getValNo()];
2959 MVT ArgVT = OutVTs[VA.getValNo()];
2961 unsigned ArgReg = getRegForValue(ArgVal);
2965 // Handle arg promotion: SExt, ZExt, AExt.
2966 switch (VA.getLocInfo()) {
2967 case CCValAssign::Full:
2969 case CCValAssign::SExt: {
2970 MVT DestVT = VA.getLocVT();
2972 ArgReg = emitIntExt(SrcVT, ArgReg, DestVT, /*isZExt=*/false);
2977 case CCValAssign::AExt:
2978 // Intentional fall-through.
2979 case CCValAssign::ZExt: {
2980 MVT DestVT = VA.getLocVT();
2982 ArgReg = emitIntExt(SrcVT, ArgReg, DestVT, /*isZExt=*/true);
2988 llvm_unreachable("Unknown arg promotion!");
2991 // Now copy/store arg to correct locations.
2992 if (VA.isRegLoc() && !VA.needsCustom()) {
2993 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2994 TII.get(TargetOpcode::COPY), VA.getLocReg()).addReg(ArgReg);
2995 CLI.OutRegs.push_back(VA.getLocReg());
2996 } else if (VA.needsCustom()) {
2997 // FIXME: Handle custom args.
3000 assert(VA.isMemLoc() && "Assuming store on stack.");
3002 // Don't emit stores for undef values.
3003 if (isa<UndefValue>(ArgVal))
3006 // Need to store on the stack.
3007 unsigned ArgSize = (ArgVT.getSizeInBits() + 7) / 8;
3009 unsigned BEAlign = 0;
3010 if (ArgSize < 8 && !Subtarget->isLittleEndian())
3011 BEAlign = 8 - ArgSize;
3014 Addr.setKind(Address::RegBase);
3015 Addr.setReg(AArch64::SP);
3016 Addr.setOffset(VA.getLocMemOffset() + BEAlign);
3018 unsigned Alignment = DL.getABITypeAlignment(ArgVal->getType());
3019 MachineMemOperand *MMO = FuncInfo.MF->getMachineMemOperand(
3020 MachinePointerInfo::getStack(Addr.getOffset()),
3021 MachineMemOperand::MOStore, ArgVT.getStoreSize(), Alignment);
3023 if (!emitStore(ArgVT, ArgReg, Addr, MMO))
3030 bool AArch64FastISel::finishCall(CallLoweringInfo &CLI, MVT RetVT,
3031 unsigned NumBytes) {
3032 CallingConv::ID CC = CLI.CallConv;
3034 // Issue CALLSEQ_END
3035 unsigned AdjStackUp = TII.getCallFrameDestroyOpcode();
3036 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackUp))
3037 .addImm(NumBytes).addImm(0);
3039 // Now the return value.
3040 if (RetVT != MVT::isVoid) {
3041 SmallVector<CCValAssign, 16> RVLocs;
3042 CCState CCInfo(CC, false, *FuncInfo.MF, RVLocs, *Context);
3043 CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC));
3045 // Only handle a single return value.
3046 if (RVLocs.size() != 1)
3049 // Copy all of the result registers out of their specified physreg.
3050 MVT CopyVT = RVLocs[0].getValVT();
3052 // TODO: Handle big-endian results
3053 if (CopyVT.isVector() && !Subtarget->isLittleEndian())
3056 unsigned ResultReg = createResultReg(TLI.getRegClassFor(CopyVT));
3057 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3058 TII.get(TargetOpcode::COPY), ResultReg)
3059 .addReg(RVLocs[0].getLocReg());
3060 CLI.InRegs.push_back(RVLocs[0].getLocReg());
3062 CLI.ResultReg = ResultReg;
3063 CLI.NumResultRegs = 1;
3069 bool AArch64FastISel::fastLowerCall(CallLoweringInfo &CLI) {
3070 CallingConv::ID CC = CLI.CallConv;
3071 bool IsTailCall = CLI.IsTailCall;
3072 bool IsVarArg = CLI.IsVarArg;
3073 const Value *Callee = CLI.Callee;
3074 MCSymbol *Symbol = CLI.Symbol;
3076 if (!Callee && !Symbol)
3079 // Allow SelectionDAG isel to handle tail calls.
3083 CodeModel::Model CM = TM.getCodeModel();
3084 // Only support the small and large code model.
3085 if (CM != CodeModel::Small && CM != CodeModel::Large)
3088 // FIXME: Add large code model support for ELF.
3089 if (CM == CodeModel::Large && !Subtarget->isTargetMachO())
3092 // Let SDISel handle vararg functions.
3096 // FIXME: Only handle *simple* calls for now.
3098 if (CLI.RetTy->isVoidTy())
3099 RetVT = MVT::isVoid;
3100 else if (!isTypeLegal(CLI.RetTy, RetVT))
3103 for (auto Flag : CLI.OutFlags)
3104 if (Flag.isInReg() || Flag.isSRet() || Flag.isNest() || Flag.isByVal())
3107 // Set up the argument vectors.
3108 SmallVector<MVT, 16> OutVTs;
3109 OutVTs.reserve(CLI.OutVals.size());
3111 for (auto *Val : CLI.OutVals) {
3113 if (!isTypeLegal(Val->getType(), VT) &&
3114 !(VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16))
3117 // We don't handle vector parameters yet.
3118 if (VT.isVector() || VT.getSizeInBits() > 64)
3121 OutVTs.push_back(VT);
3125 if (Callee && !computeCallAddress(Callee, Addr))
3128 // Handle the arguments now that we've gotten them.
3130 if (!processCallArgs(CLI, OutVTs, NumBytes))
3134 MachineInstrBuilder MIB;
3135 if (CM == CodeModel::Small) {
3136 const MCInstrDesc &II = TII.get(Addr.getReg() ? AArch64::BLR : AArch64::BL);
3137 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II);
3139 MIB.addSym(Symbol, 0);
3140 else if (Addr.getGlobalValue())
3141 MIB.addGlobalAddress(Addr.getGlobalValue(), 0, 0);
3142 else if (Addr.getReg()) {
3143 unsigned Reg = constrainOperandRegClass(II, Addr.getReg(), 0);
3148 unsigned CallReg = 0;
3150 unsigned ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
3151 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
3153 .addSym(Symbol, AArch64II::MO_GOT | AArch64II::MO_PAGE);
3155 CallReg = createResultReg(&AArch64::GPR64RegClass);
3156 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3157 TII.get(AArch64::LDRXui), CallReg)
3160 AArch64II::MO_GOT | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3161 } else if (Addr.getGlobalValue())
3162 CallReg = materializeGV(Addr.getGlobalValue());
3163 else if (Addr.getReg())
3164 CallReg = Addr.getReg();
3169 const MCInstrDesc &II = TII.get(AArch64::BLR);
3170 CallReg = constrainOperandRegClass(II, CallReg, 0);
3171 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II).addReg(CallReg);
3174 // Add implicit physical register uses to the call.
3175 for (auto Reg : CLI.OutRegs)
3176 MIB.addReg(Reg, RegState::Implicit);
3178 // Add a register mask with the call-preserved registers.
3179 // Proper defs for return values will be added by setPhysRegsDeadExcept().
3180 MIB.addRegMask(TRI.getCallPreservedMask(*FuncInfo.MF, CC));
3184 // Finish off the call including any return values.
3185 return finishCall(CLI, RetVT, NumBytes);
3188 bool AArch64FastISel::isMemCpySmall(uint64_t Len, unsigned Alignment) {
3190 return Len / Alignment <= 4;
3195 bool AArch64FastISel::tryEmitSmallMemCpy(Address Dest, Address Src,
3196 uint64_t Len, unsigned Alignment) {
3197 // Make sure we don't bloat code by inlining very large memcpy's.
3198 if (!isMemCpySmall(Len, Alignment))
3201 int64_t UnscaledOffset = 0;
3202 Address OrigDest = Dest;
3203 Address OrigSrc = Src;
3207 if (!Alignment || Alignment >= 8) {
3218 // Bound based on alignment.
3219 if (Len >= 4 && Alignment == 4)
3221 else if (Len >= 2 && Alignment == 2)
3228 unsigned ResultReg = emitLoad(VT, VT, Src);
3232 if (!emitStore(VT, ResultReg, Dest))
3235 int64_t Size = VT.getSizeInBits() / 8;
3237 UnscaledOffset += Size;
3239 // We need to recompute the unscaled offset for each iteration.
3240 Dest.setOffset(OrigDest.getOffset() + UnscaledOffset);
3241 Src.setOffset(OrigSrc.getOffset() + UnscaledOffset);
3247 /// \brief Check if it is possible to fold the condition from the XALU intrinsic
3248 /// into the user. The condition code will only be updated on success.
3249 bool AArch64FastISel::foldXALUIntrinsic(AArch64CC::CondCode &CC,
3250 const Instruction *I,
3251 const Value *Cond) {
3252 if (!isa<ExtractValueInst>(Cond))
3255 const auto *EV = cast<ExtractValueInst>(Cond);
3256 if (!isa<IntrinsicInst>(EV->getAggregateOperand()))
3259 const auto *II = cast<IntrinsicInst>(EV->getAggregateOperand());
3261 const Function *Callee = II->getCalledFunction();
3263 cast<StructType>(Callee->getReturnType())->getTypeAtIndex(0U);
3264 if (!isTypeLegal(RetTy, RetVT))
3267 if (RetVT != MVT::i32 && RetVT != MVT::i64)
3270 const Value *LHS = II->getArgOperand(0);
3271 const Value *RHS = II->getArgOperand(1);
3273 // Canonicalize immediate to the RHS.
3274 if (isa<ConstantInt>(LHS) && !isa<ConstantInt>(RHS) &&
3275 isCommutativeIntrinsic(II))
3276 std::swap(LHS, RHS);
3278 // Simplify multiplies.
3279 Intrinsic::ID IID = II->getIntrinsicID();
3283 case Intrinsic::smul_with_overflow:
3284 if (const auto *C = dyn_cast<ConstantInt>(RHS))
3285 if (C->getValue() == 2)
3286 IID = Intrinsic::sadd_with_overflow;
3288 case Intrinsic::umul_with_overflow:
3289 if (const auto *C = dyn_cast<ConstantInt>(RHS))
3290 if (C->getValue() == 2)
3291 IID = Intrinsic::uadd_with_overflow;
3295 AArch64CC::CondCode TmpCC;
3299 case Intrinsic::sadd_with_overflow:
3300 case Intrinsic::ssub_with_overflow:
3301 TmpCC = AArch64CC::VS;
3303 case Intrinsic::uadd_with_overflow:
3304 TmpCC = AArch64CC::HS;
3306 case Intrinsic::usub_with_overflow:
3307 TmpCC = AArch64CC::LO;
3309 case Intrinsic::smul_with_overflow:
3310 case Intrinsic::umul_with_overflow:
3311 TmpCC = AArch64CC::NE;
3315 // Check if both instructions are in the same basic block.
3316 if (!isValueAvailable(II))
3319 // Make sure nothing is in the way
3320 BasicBlock::const_iterator Start = I;
3321 BasicBlock::const_iterator End = II;
3322 for (auto Itr = std::prev(Start); Itr != End; --Itr) {
3323 // We only expect extractvalue instructions between the intrinsic and the
3324 // instruction to be selected.
3325 if (!isa<ExtractValueInst>(Itr))
3328 // Check that the extractvalue operand comes from the intrinsic.
3329 const auto *EVI = cast<ExtractValueInst>(Itr);
3330 if (EVI->getAggregateOperand() != II)
3338 bool AArch64FastISel::fastLowerIntrinsicCall(const IntrinsicInst *II) {
3339 // FIXME: Handle more intrinsics.
3340 switch (II->getIntrinsicID()) {
3341 default: return false;
3342 case Intrinsic::frameaddress: {
3343 MachineFrameInfo *MFI = FuncInfo.MF->getFrameInfo();
3344 MFI->setFrameAddressIsTaken(true);
3346 const AArch64RegisterInfo *RegInfo =
3347 static_cast<const AArch64RegisterInfo *>(Subtarget->getRegisterInfo());
3348 unsigned FramePtr = RegInfo->getFrameRegister(*(FuncInfo.MF));
3349 unsigned SrcReg = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
3350 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3351 TII.get(TargetOpcode::COPY), SrcReg).addReg(FramePtr);
3352 // Recursively load frame address
3358 unsigned Depth = cast<ConstantInt>(II->getOperand(0))->getZExtValue();
3360 DestReg = fastEmitInst_ri(AArch64::LDRXui, &AArch64::GPR64RegClass,
3361 SrcReg, /*IsKill=*/true, 0);
3362 assert(DestReg && "Unexpected LDR instruction emission failure.");
3366 updateValueMap(II, SrcReg);
3369 case Intrinsic::memcpy:
3370 case Intrinsic::memmove: {
3371 const auto *MTI = cast<MemTransferInst>(II);
3372 // Don't handle volatile.
3373 if (MTI->isVolatile())
3376 // Disable inlining for memmove before calls to ComputeAddress. Otherwise,
3377 // we would emit dead code because we don't currently handle memmoves.
3378 bool IsMemCpy = (II->getIntrinsicID() == Intrinsic::memcpy);
3379 if (isa<ConstantInt>(MTI->getLength()) && IsMemCpy) {
3380 // Small memcpy's are common enough that we want to do them without a call
3382 uint64_t Len = cast<ConstantInt>(MTI->getLength())->getZExtValue();
3383 unsigned Alignment = MTI->getAlignment();
3384 if (isMemCpySmall(Len, Alignment)) {
3386 if (!computeAddress(MTI->getRawDest(), Dest) ||
3387 !computeAddress(MTI->getRawSource(), Src))
3389 if (tryEmitSmallMemCpy(Dest, Src, Len, Alignment))
3394 if (!MTI->getLength()->getType()->isIntegerTy(64))
3397 if (MTI->getSourceAddressSpace() > 255 || MTI->getDestAddressSpace() > 255)
3398 // Fast instruction selection doesn't support the special
3402 const char *IntrMemName = isa<MemCpyInst>(II) ? "memcpy" : "memmove";
3403 return lowerCallTo(II, IntrMemName, II->getNumArgOperands() - 2);
3405 case Intrinsic::memset: {
3406 const MemSetInst *MSI = cast<MemSetInst>(II);
3407 // Don't handle volatile.
3408 if (MSI->isVolatile())
3411 if (!MSI->getLength()->getType()->isIntegerTy(64))
3414 if (MSI->getDestAddressSpace() > 255)
3415 // Fast instruction selection doesn't support the special
3419 return lowerCallTo(II, "memset", II->getNumArgOperands() - 2);
3421 case Intrinsic::sin:
3422 case Intrinsic::cos:
3423 case Intrinsic::pow: {
3425 if (!isTypeLegal(II->getType(), RetVT))
3428 if (RetVT != MVT::f32 && RetVT != MVT::f64)
3431 static const RTLIB::Libcall LibCallTable[3][2] = {
3432 { RTLIB::SIN_F32, RTLIB::SIN_F64 },
3433 { RTLIB::COS_F32, RTLIB::COS_F64 },
3434 { RTLIB::POW_F32, RTLIB::POW_F64 }
3437 bool Is64Bit = RetVT == MVT::f64;
3438 switch (II->getIntrinsicID()) {
3440 llvm_unreachable("Unexpected intrinsic.");
3441 case Intrinsic::sin:
3442 LC = LibCallTable[0][Is64Bit];
3444 case Intrinsic::cos:
3445 LC = LibCallTable[1][Is64Bit];
3447 case Intrinsic::pow:
3448 LC = LibCallTable[2][Is64Bit];
3453 Args.reserve(II->getNumArgOperands());
3455 // Populate the argument list.
3456 for (auto &Arg : II->arg_operands()) {
3459 Entry.Ty = Arg->getType();
3460 Args.push_back(Entry);
3463 CallLoweringInfo CLI;
3464 MCContext &Ctx = MF->getContext();
3465 CLI.setCallee(DL, Ctx, TLI.getLibcallCallingConv(LC), II->getType(),
3466 TLI.getLibcallName(LC), std::move(Args));
3467 if (!lowerCallTo(CLI))
3469 updateValueMap(II, CLI.ResultReg);
3472 case Intrinsic::fabs: {
3474 if (!isTypeLegal(II->getType(), VT))
3478 switch (VT.SimpleTy) {
3482 Opc = AArch64::FABSSr;
3485 Opc = AArch64::FABSDr;
3488 unsigned SrcReg = getRegForValue(II->getOperand(0));
3491 bool SrcRegIsKill = hasTrivialKill(II->getOperand(0));
3492 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
3493 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
3494 .addReg(SrcReg, getKillRegState(SrcRegIsKill));
3495 updateValueMap(II, ResultReg);
3498 case Intrinsic::trap: {
3499 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::BRK))
3503 case Intrinsic::sqrt: {
3504 Type *RetTy = II->getCalledFunction()->getReturnType();
3507 if (!isTypeLegal(RetTy, VT))
3510 unsigned Op0Reg = getRegForValue(II->getOperand(0));
3513 bool Op0IsKill = hasTrivialKill(II->getOperand(0));
3515 unsigned ResultReg = fastEmit_r(VT, VT, ISD::FSQRT, Op0Reg, Op0IsKill);
3519 updateValueMap(II, ResultReg);
3522 case Intrinsic::sadd_with_overflow:
3523 case Intrinsic::uadd_with_overflow:
3524 case Intrinsic::ssub_with_overflow:
3525 case Intrinsic::usub_with_overflow:
3526 case Intrinsic::smul_with_overflow:
3527 case Intrinsic::umul_with_overflow: {
3528 // This implements the basic lowering of the xalu with overflow intrinsics.
3529 const Function *Callee = II->getCalledFunction();
3530 auto *Ty = cast<StructType>(Callee->getReturnType());
3531 Type *RetTy = Ty->getTypeAtIndex(0U);
3534 if (!isTypeLegal(RetTy, VT))
3537 if (VT != MVT::i32 && VT != MVT::i64)
3540 const Value *LHS = II->getArgOperand(0);
3541 const Value *RHS = II->getArgOperand(1);
3542 // Canonicalize immediate to the RHS.
3543 if (isa<ConstantInt>(LHS) && !isa<ConstantInt>(RHS) &&
3544 isCommutativeIntrinsic(II))
3545 std::swap(LHS, RHS);
3547 // Simplify multiplies.
3548 Intrinsic::ID IID = II->getIntrinsicID();
3552 case Intrinsic::smul_with_overflow:
3553 if (const auto *C = dyn_cast<ConstantInt>(RHS))
3554 if (C->getValue() == 2) {
3555 IID = Intrinsic::sadd_with_overflow;
3559 case Intrinsic::umul_with_overflow:
3560 if (const auto *C = dyn_cast<ConstantInt>(RHS))
3561 if (C->getValue() == 2) {
3562 IID = Intrinsic::uadd_with_overflow;
3568 unsigned ResultReg1 = 0, ResultReg2 = 0, MulReg = 0;
3569 AArch64CC::CondCode CC = AArch64CC::Invalid;
3571 default: llvm_unreachable("Unexpected intrinsic!");
3572 case Intrinsic::sadd_with_overflow:
3573 ResultReg1 = emitAdd(VT, LHS, RHS, /*SetFlags=*/true);
3576 case Intrinsic::uadd_with_overflow:
3577 ResultReg1 = emitAdd(VT, LHS, RHS, /*SetFlags=*/true);
3580 case Intrinsic::ssub_with_overflow:
3581 ResultReg1 = emitSub(VT, LHS, RHS, /*SetFlags=*/true);
3584 case Intrinsic::usub_with_overflow:
3585 ResultReg1 = emitSub(VT, LHS, RHS, /*SetFlags=*/true);
3588 case Intrinsic::smul_with_overflow: {
3590 unsigned LHSReg = getRegForValue(LHS);
3593 bool LHSIsKill = hasTrivialKill(LHS);
3595 unsigned RHSReg = getRegForValue(RHS);
3598 bool RHSIsKill = hasTrivialKill(RHS);
3600 if (VT == MVT::i32) {
3601 MulReg = emitSMULL_rr(MVT::i64, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
3602 unsigned ShiftReg = emitLSR_ri(MVT::i64, MVT::i64, MulReg,
3603 /*IsKill=*/false, 32);
3604 MulReg = fastEmitInst_extractsubreg(VT, MulReg, /*IsKill=*/true,
3606 ShiftReg = fastEmitInst_extractsubreg(VT, ShiftReg, /*IsKill=*/true,
3608 emitSubs_rs(VT, ShiftReg, /*IsKill=*/true, MulReg, /*IsKill=*/false,
3609 AArch64_AM::ASR, 31, /*WantResult=*/false);
3611 assert(VT == MVT::i64 && "Unexpected value type.");
3612 // LHSReg and RHSReg cannot be killed by this Mul, since they are
3613 // reused in the next instruction.
3614 MulReg = emitMul_rr(VT, LHSReg, /*IsKill=*/false, RHSReg,
3616 unsigned SMULHReg = fastEmit_rr(VT, VT, ISD::MULHS, LHSReg, LHSIsKill,
3618 emitSubs_rs(VT, SMULHReg, /*IsKill=*/true, MulReg, /*IsKill=*/false,
3619 AArch64_AM::ASR, 63, /*WantResult=*/false);
3623 case Intrinsic::umul_with_overflow: {
3625 unsigned LHSReg = getRegForValue(LHS);
3628 bool LHSIsKill = hasTrivialKill(LHS);
3630 unsigned RHSReg = getRegForValue(RHS);
3633 bool RHSIsKill = hasTrivialKill(RHS);
3635 if (VT == MVT::i32) {
3636 MulReg = emitUMULL_rr(MVT::i64, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
3637 emitSubs_rs(MVT::i64, AArch64::XZR, /*IsKill=*/true, MulReg,
3638 /*IsKill=*/false, AArch64_AM::LSR, 32,
3639 /*WantResult=*/false);
3640 MulReg = fastEmitInst_extractsubreg(VT, MulReg, /*IsKill=*/true,
3643 assert(VT == MVT::i64 && "Unexpected value type.");
3644 // LHSReg and RHSReg cannot be killed by this Mul, since they are
3645 // reused in the next instruction.
3646 MulReg = emitMul_rr(VT, LHSReg, /*IsKill=*/false, RHSReg,
3648 unsigned UMULHReg = fastEmit_rr(VT, VT, ISD::MULHU, LHSReg, LHSIsKill,
3650 emitSubs_rr(VT, AArch64::XZR, /*IsKill=*/true, UMULHReg,
3651 /*IsKill=*/false, /*WantResult=*/false);
3658 ResultReg1 = createResultReg(TLI.getRegClassFor(VT));
3659 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3660 TII.get(TargetOpcode::COPY), ResultReg1).addReg(MulReg);
3663 ResultReg2 = fastEmitInst_rri(AArch64::CSINCWr, &AArch64::GPR32RegClass,
3664 AArch64::WZR, /*IsKill=*/true, AArch64::WZR,
3665 /*IsKill=*/true, getInvertedCondCode(CC));
3667 assert((ResultReg1 + 1) == ResultReg2 &&
3668 "Nonconsecutive result registers.");
3669 updateValueMap(II, ResultReg1, 2);
3676 bool AArch64FastISel::selectRet(const Instruction *I) {
3677 const ReturnInst *Ret = cast<ReturnInst>(I);
3678 const Function &F = *I->getParent()->getParent();
3680 if (!FuncInfo.CanLowerReturn)
3686 // Build a list of return value registers.
3687 SmallVector<unsigned, 4> RetRegs;
3689 if (Ret->getNumOperands() > 0) {
3690 CallingConv::ID CC = F.getCallingConv();
3691 SmallVector<ISD::OutputArg, 4> Outs;
3692 GetReturnInfo(F.getReturnType(), F.getAttributes(), Outs, TLI);
3694 // Analyze operands of the call, assigning locations to each operand.
3695 SmallVector<CCValAssign, 16> ValLocs;
3696 CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, ValLocs, I->getContext());
3697 CCAssignFn *RetCC = CC == CallingConv::WebKit_JS ? RetCC_AArch64_WebKit_JS
3698 : RetCC_AArch64_AAPCS;
3699 CCInfo.AnalyzeReturn(Outs, RetCC);
3701 // Only handle a single return value for now.
3702 if (ValLocs.size() != 1)
3705 CCValAssign &VA = ValLocs[0];
3706 const Value *RV = Ret->getOperand(0);
3708 // Don't bother handling odd stuff for now.
3709 if ((VA.getLocInfo() != CCValAssign::Full) &&
3710 (VA.getLocInfo() != CCValAssign::BCvt))
3713 // Only handle register returns for now.
3717 unsigned Reg = getRegForValue(RV);
3721 unsigned SrcReg = Reg + VA.getValNo();
3722 unsigned DestReg = VA.getLocReg();
3723 // Avoid a cross-class copy. This is very unlikely.
3724 if (!MRI.getRegClass(SrcReg)->contains(DestReg))
3727 EVT RVEVT = TLI.getValueType(RV->getType());
3728 if (!RVEVT.isSimple())
3731 // Vectors (of > 1 lane) in big endian need tricky handling.
3732 if (RVEVT.isVector() && RVEVT.getVectorNumElements() > 1 &&
3733 !Subtarget->isLittleEndian())
3736 MVT RVVT = RVEVT.getSimpleVT();
3737 if (RVVT == MVT::f128)
3740 MVT DestVT = VA.getValVT();
3741 // Special handling for extended integers.
3742 if (RVVT != DestVT) {
3743 if (RVVT != MVT::i1 && RVVT != MVT::i8 && RVVT != MVT::i16)
3746 if (!Outs[0].Flags.isZExt() && !Outs[0].Flags.isSExt())
3749 bool IsZExt = Outs[0].Flags.isZExt();
3750 SrcReg = emitIntExt(RVVT, SrcReg, DestVT, IsZExt);
3756 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3757 TII.get(TargetOpcode::COPY), DestReg).addReg(SrcReg);
3759 // Add register to return instruction.
3760 RetRegs.push_back(VA.getLocReg());
3763 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3764 TII.get(AArch64::RET_ReallyLR));
3765 for (unsigned i = 0, e = RetRegs.size(); i != e; ++i)
3766 MIB.addReg(RetRegs[i], RegState::Implicit);
3770 bool AArch64FastISel::selectTrunc(const Instruction *I) {
3771 Type *DestTy = I->getType();
3772 Value *Op = I->getOperand(0);
3773 Type *SrcTy = Op->getType();
3775 EVT SrcEVT = TLI.getValueType(SrcTy, true);
3776 EVT DestEVT = TLI.getValueType(DestTy, true);
3777 if (!SrcEVT.isSimple())
3779 if (!DestEVT.isSimple())
3782 MVT SrcVT = SrcEVT.getSimpleVT();
3783 MVT DestVT = DestEVT.getSimpleVT();
3785 if (SrcVT != MVT::i64 && SrcVT != MVT::i32 && SrcVT != MVT::i16 &&
3788 if (DestVT != MVT::i32 && DestVT != MVT::i16 && DestVT != MVT::i8 &&
3792 unsigned SrcReg = getRegForValue(Op);
3795 bool SrcIsKill = hasTrivialKill(Op);
3797 // If we're truncating from i64 to a smaller non-legal type then generate an
3798 // AND. Otherwise, we know the high bits are undefined and a truncate only
3799 // generate a COPY. We cannot mark the source register also as result
3800 // register, because this can incorrectly transfer the kill flag onto the
3803 if (SrcVT == MVT::i64) {
3805 switch (DestVT.SimpleTy) {
3807 // Trunc i64 to i32 is handled by the target-independent fast-isel.
3819 // Issue an extract_subreg to get the lower 32-bits.
3820 unsigned Reg32 = fastEmitInst_extractsubreg(MVT::i32, SrcReg, SrcIsKill,
3822 // Create the AND instruction which performs the actual truncation.
3823 ResultReg = emitAnd_ri(MVT::i32, Reg32, /*IsKill=*/true, Mask);
3824 assert(ResultReg && "Unexpected AND instruction emission failure.");
3826 ResultReg = createResultReg(&AArch64::GPR32RegClass);
3827 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3828 TII.get(TargetOpcode::COPY), ResultReg)
3829 .addReg(SrcReg, getKillRegState(SrcIsKill));
3832 updateValueMap(I, ResultReg);
3836 unsigned AArch64FastISel::emiti1Ext(unsigned SrcReg, MVT DestVT, bool IsZExt) {
3837 assert((DestVT == MVT::i8 || DestVT == MVT::i16 || DestVT == MVT::i32 ||
3838 DestVT == MVT::i64) &&
3839 "Unexpected value type.");
3840 // Handle i8 and i16 as i32.
3841 if (DestVT == MVT::i8 || DestVT == MVT::i16)
3845 unsigned ResultReg = emitAnd_ri(MVT::i32, SrcReg, /*TODO:IsKill=*/false, 1);
3846 assert(ResultReg && "Unexpected AND instruction emission failure.");
3847 if (DestVT == MVT::i64) {
3848 // We're ZExt i1 to i64. The ANDWri Wd, Ws, #1 implicitly clears the
3849 // upper 32 bits. Emit a SUBREG_TO_REG to extend from Wd to Xd.
3850 unsigned Reg64 = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
3851 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3852 TII.get(AArch64::SUBREG_TO_REG), Reg64)
3855 .addImm(AArch64::sub_32);
3860 if (DestVT == MVT::i64) {
3861 // FIXME: We're SExt i1 to i64.
3864 return fastEmitInst_rii(AArch64::SBFMWri, &AArch64::GPR32RegClass, SrcReg,
3865 /*TODO:IsKill=*/false, 0, 0);
3869 unsigned AArch64FastISel::emitMul_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
3870 unsigned Op1, bool Op1IsKill) {
3872 switch (RetVT.SimpleTy) {
3878 Opc = AArch64::MADDWrrr; ZReg = AArch64::WZR; break;
3880 Opc = AArch64::MADDXrrr; ZReg = AArch64::XZR; break;
3883 const TargetRegisterClass *RC =
3884 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3885 return fastEmitInst_rrr(Opc, RC, Op0, Op0IsKill, Op1, Op1IsKill,
3886 /*IsKill=*/ZReg, true);
3889 unsigned AArch64FastISel::emitSMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
3890 unsigned Op1, bool Op1IsKill) {
3891 if (RetVT != MVT::i64)
3894 return fastEmitInst_rrr(AArch64::SMADDLrrr, &AArch64::GPR64RegClass,
3895 Op0, Op0IsKill, Op1, Op1IsKill,
3896 AArch64::XZR, /*IsKill=*/true);
3899 unsigned AArch64FastISel::emitUMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
3900 unsigned Op1, bool Op1IsKill) {
3901 if (RetVT != MVT::i64)
3904 return fastEmitInst_rrr(AArch64::UMADDLrrr, &AArch64::GPR64RegClass,
3905 Op0, Op0IsKill, Op1, Op1IsKill,
3906 AArch64::XZR, /*IsKill=*/true);
3909 unsigned AArch64FastISel::emitLSL_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
3910 unsigned Op1Reg, bool Op1IsKill) {
3912 bool NeedTrunc = false;
3914 switch (RetVT.SimpleTy) {
3916 case MVT::i8: Opc = AArch64::LSLVWr; NeedTrunc = true; Mask = 0xff; break;
3917 case MVT::i16: Opc = AArch64::LSLVWr; NeedTrunc = true; Mask = 0xffff; break;
3918 case MVT::i32: Opc = AArch64::LSLVWr; break;
3919 case MVT::i64: Opc = AArch64::LSLVXr; break;
3922 const TargetRegisterClass *RC =
3923 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3925 Op1Reg = emitAnd_ri(MVT::i32, Op1Reg, Op1IsKill, Mask);
3928 unsigned ResultReg = fastEmitInst_rr(Opc, RC, Op0Reg, Op0IsKill, Op1Reg,
3931 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
3935 unsigned AArch64FastISel::emitLSL_ri(MVT RetVT, MVT SrcVT, unsigned Op0,
3936 bool Op0IsKill, uint64_t Shift,
3938 assert(RetVT.SimpleTy >= SrcVT.SimpleTy &&
3939 "Unexpected source/return type pair.");
3940 assert((SrcVT == MVT::i1 || SrcVT == MVT::i8 || SrcVT == MVT::i16 ||
3941 SrcVT == MVT::i32 || SrcVT == MVT::i64) &&
3942 "Unexpected source value type.");
3943 assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 ||
3944 RetVT == MVT::i64) && "Unexpected return value type.");
3946 bool Is64Bit = (RetVT == MVT::i64);
3947 unsigned RegSize = Is64Bit ? 64 : 32;
3948 unsigned DstBits = RetVT.getSizeInBits();
3949 unsigned SrcBits = SrcVT.getSizeInBits();
3950 const TargetRegisterClass *RC =
3951 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3953 // Just emit a copy for "zero" shifts.
3955 if (RetVT == SrcVT) {
3956 unsigned ResultReg = createResultReg(RC);
3957 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3958 TII.get(TargetOpcode::COPY), ResultReg)
3959 .addReg(Op0, getKillRegState(Op0IsKill));
3962 return emitIntExt(SrcVT, Op0, RetVT, IsZExt);
3965 // Don't deal with undefined shifts.
3966 if (Shift >= DstBits)
3969 // For immediate shifts we can fold the zero-/sign-extension into the shift.
3970 // {S|U}BFM Wd, Wn, #r, #s
3971 // Wd<32+s-r,32-r> = Wn<s:0> when r > s
3973 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3974 // %2 = shl i16 %1, 4
3975 // Wd<32+7-28,32-28> = Wn<7:0> <- clamp s to 7
3976 // 0b1111_1111_1111_1111__1111_1010_1010_0000 sext
3977 // 0b0000_0000_0000_0000__0000_0101_0101_0000 sext | zext
3978 // 0b0000_0000_0000_0000__0000_1010_1010_0000 zext
3980 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3981 // %2 = shl i16 %1, 8
3982 // Wd<32+7-24,32-24> = Wn<7:0>
3983 // 0b1111_1111_1111_1111__1010_1010_0000_0000 sext
3984 // 0b0000_0000_0000_0000__0101_0101_0000_0000 sext | zext
3985 // 0b0000_0000_0000_0000__1010_1010_0000_0000 zext
3987 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3988 // %2 = shl i16 %1, 12
3989 // Wd<32+3-20,32-20> = Wn<3:0>
3990 // 0b1111_1111_1111_1111__1010_0000_0000_0000 sext
3991 // 0b0000_0000_0000_0000__0101_0000_0000_0000 sext | zext
3992 // 0b0000_0000_0000_0000__1010_0000_0000_0000 zext
3994 unsigned ImmR = RegSize - Shift;
3995 // Limit the width to the length of the source type.
3996 unsigned ImmS = std::min<unsigned>(SrcBits - 1, DstBits - 1 - Shift);
3997 static const unsigned OpcTable[2][2] = {
3998 {AArch64::SBFMWri, AArch64::SBFMXri},
3999 {AArch64::UBFMWri, AArch64::UBFMXri}
4001 unsigned Opc = OpcTable[IsZExt][Is64Bit];
4002 if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) {
4003 unsigned TmpReg = MRI.createVirtualRegister(RC);
4004 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
4005 TII.get(AArch64::SUBREG_TO_REG), TmpReg)
4007 .addReg(Op0, getKillRegState(Op0IsKill))
4008 .addImm(AArch64::sub_32);
4012 return fastEmitInst_rii(Opc, RC, Op0, Op0IsKill, ImmR, ImmS);
4015 unsigned AArch64FastISel::emitLSR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
4016 unsigned Op1Reg, bool Op1IsKill) {
4018 bool NeedTrunc = false;
4020 switch (RetVT.SimpleTy) {
4022 case MVT::i8: Opc = AArch64::LSRVWr; NeedTrunc = true; Mask = 0xff; break;
4023 case MVT::i16: Opc = AArch64::LSRVWr; NeedTrunc = true; Mask = 0xffff; break;
4024 case MVT::i32: Opc = AArch64::LSRVWr; break;
4025 case MVT::i64: Opc = AArch64::LSRVXr; break;
4028 const TargetRegisterClass *RC =
4029 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
4031 Op0Reg = emitAnd_ri(MVT::i32, Op0Reg, Op0IsKill, Mask);
4032 Op1Reg = emitAnd_ri(MVT::i32, Op1Reg, Op1IsKill, Mask);
4033 Op0IsKill = Op1IsKill = true;
4035 unsigned ResultReg = fastEmitInst_rr(Opc, RC, Op0Reg, Op0IsKill, Op1Reg,
4038 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
4042 unsigned AArch64FastISel::emitLSR_ri(MVT RetVT, MVT SrcVT, unsigned Op0,
4043 bool Op0IsKill, uint64_t Shift,
4045 assert(RetVT.SimpleTy >= SrcVT.SimpleTy &&
4046 "Unexpected source/return type pair.");
4047 assert((SrcVT == MVT::i1 || SrcVT == MVT::i8 || SrcVT == MVT::i16 ||
4048 SrcVT == MVT::i32 || SrcVT == MVT::i64) &&
4049 "Unexpected source value type.");
4050 assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 ||
4051 RetVT == MVT::i64) && "Unexpected return value type.");
4053 bool Is64Bit = (RetVT == MVT::i64);
4054 unsigned RegSize = Is64Bit ? 64 : 32;
4055 unsigned DstBits = RetVT.getSizeInBits();
4056 unsigned SrcBits = SrcVT.getSizeInBits();
4057 const TargetRegisterClass *RC =
4058 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
4060 // Just emit a copy for "zero" shifts.
4062 if (RetVT == SrcVT) {
4063 unsigned ResultReg = createResultReg(RC);
4064 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
4065 TII.get(TargetOpcode::COPY), ResultReg)
4066 .addReg(Op0, getKillRegState(Op0IsKill));
4069 return emitIntExt(SrcVT, Op0, RetVT, IsZExt);
4072 // Don't deal with undefined shifts.
4073 if (Shift >= DstBits)
4076 // For immediate shifts we can fold the zero-/sign-extension into the shift.
4077 // {S|U}BFM Wd, Wn, #r, #s
4078 // Wd<s-r:0> = Wn<s:r> when r <= s
4080 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
4081 // %2 = lshr i16 %1, 4
4082 // Wd<7-4:0> = Wn<7:4>
4083 // 0b0000_0000_0000_0000__0000_1111_1111_1010 sext
4084 // 0b0000_0000_0000_0000__0000_0000_0000_0101 sext | zext
4085 // 0b0000_0000_0000_0000__0000_0000_0000_1010 zext
4087 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
4088 // %2 = lshr i16 %1, 8
4089 // Wd<7-7,0> = Wn<7:7>
4090 // 0b0000_0000_0000_0000__0000_0000_1111_1111 sext
4091 // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
4092 // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
4094 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
4095 // %2 = lshr i16 %1, 12
4096 // Wd<7-7,0> = Wn<7:7> <- clamp r to 7
4097 // 0b0000_0000_0000_0000__0000_0000_0000_1111 sext
4098 // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
4099 // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
4101 if (Shift >= SrcBits && IsZExt)
4102 return materializeInt(ConstantInt::get(*Context, APInt(RegSize, 0)), RetVT);
4104 // It is not possible to fold a sign-extend into the LShr instruction. In this
4105 // case emit a sign-extend.
4107 Op0 = emitIntExt(SrcVT, Op0, RetVT, IsZExt);
4112 SrcBits = SrcVT.getSizeInBits();
4116 unsigned ImmR = std::min<unsigned>(SrcBits - 1, Shift);
4117 unsigned ImmS = SrcBits - 1;
4118 static const unsigned OpcTable[2][2] = {
4119 {AArch64::SBFMWri, AArch64::SBFMXri},
4120 {AArch64::UBFMWri, AArch64::UBFMXri}
4122 unsigned Opc = OpcTable[IsZExt][Is64Bit];
4123 if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) {
4124 unsigned TmpReg = MRI.createVirtualRegister(RC);
4125 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
4126 TII.get(AArch64::SUBREG_TO_REG), TmpReg)
4128 .addReg(Op0, getKillRegState(Op0IsKill))
4129 .addImm(AArch64::sub_32);
4133 return fastEmitInst_rii(Opc, RC, Op0, Op0IsKill, ImmR, ImmS);
4136 unsigned AArch64FastISel::emitASR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
4137 unsigned Op1Reg, bool Op1IsKill) {
4139 bool NeedTrunc = false;
4141 switch (RetVT.SimpleTy) {
4143 case MVT::i8: Opc = AArch64::ASRVWr; NeedTrunc = true; Mask = 0xff; break;
4144 case MVT::i16: Opc = AArch64::ASRVWr; NeedTrunc = true; Mask = 0xffff; break;
4145 case MVT::i32: Opc = AArch64::ASRVWr; break;
4146 case MVT::i64: Opc = AArch64::ASRVXr; break;
4149 const TargetRegisterClass *RC =
4150 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
4152 Op0Reg = emitIntExt(RetVT, Op0Reg, MVT::i32, /*IsZExt=*/false);
4153 Op1Reg = emitAnd_ri(MVT::i32, Op1Reg, Op1IsKill, Mask);
4154 Op0IsKill = Op1IsKill = true;
4156 unsigned ResultReg = fastEmitInst_rr(Opc, RC, Op0Reg, Op0IsKill, Op1Reg,
4159 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
4163 unsigned AArch64FastISel::emitASR_ri(MVT RetVT, MVT SrcVT, unsigned Op0,
4164 bool Op0IsKill, uint64_t Shift,
4166 assert(RetVT.SimpleTy >= SrcVT.SimpleTy &&
4167 "Unexpected source/return type pair.");
4168 assert((SrcVT == MVT::i1 || SrcVT == MVT::i8 || SrcVT == MVT::i16 ||
4169 SrcVT == MVT::i32 || SrcVT == MVT::i64) &&
4170 "Unexpected source value type.");
4171 assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 ||
4172 RetVT == MVT::i64) && "Unexpected return value type.");
4174 bool Is64Bit = (RetVT == MVT::i64);
4175 unsigned RegSize = Is64Bit ? 64 : 32;
4176 unsigned DstBits = RetVT.getSizeInBits();
4177 unsigned SrcBits = SrcVT.getSizeInBits();
4178 const TargetRegisterClass *RC =
4179 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
4181 // Just emit a copy for "zero" shifts.
4183 if (RetVT == SrcVT) {
4184 unsigned ResultReg = createResultReg(RC);
4185 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
4186 TII.get(TargetOpcode::COPY), ResultReg)
4187 .addReg(Op0, getKillRegState(Op0IsKill));
4190 return emitIntExt(SrcVT, Op0, RetVT, IsZExt);
4193 // Don't deal with undefined shifts.
4194 if (Shift >= DstBits)
4197 // For immediate shifts we can fold the zero-/sign-extension into the shift.
4198 // {S|U}BFM Wd, Wn, #r, #s
4199 // Wd<s-r:0> = Wn<s:r> when r <= s
4201 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
4202 // %2 = ashr i16 %1, 4
4203 // Wd<7-4:0> = Wn<7:4>
4204 // 0b1111_1111_1111_1111__1111_1111_1111_1010 sext
4205 // 0b0000_0000_0000_0000__0000_0000_0000_0101 sext | zext
4206 // 0b0000_0000_0000_0000__0000_0000_0000_1010 zext
4208 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
4209 // %2 = ashr i16 %1, 8
4210 // Wd<7-7,0> = Wn<7:7>
4211 // 0b1111_1111_1111_1111__1111_1111_1111_1111 sext
4212 // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
4213 // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
4215 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
4216 // %2 = ashr i16 %1, 12
4217 // Wd<7-7,0> = Wn<7:7> <- clamp r to 7
4218 // 0b1111_1111_1111_1111__1111_1111_1111_1111 sext
4219 // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
4220 // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
4222 if (Shift >= SrcBits && IsZExt)
4223 return materializeInt(ConstantInt::get(*Context, APInt(RegSize, 0)), RetVT);
4225 unsigned ImmR = std::min<unsigned>(SrcBits - 1, Shift);
4226 unsigned ImmS = SrcBits - 1;
4227 static const unsigned OpcTable[2][2] = {
4228 {AArch64::SBFMWri, AArch64::SBFMXri},
4229 {AArch64::UBFMWri, AArch64::UBFMXri}
4231 unsigned Opc = OpcTable[IsZExt][Is64Bit];
4232 if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) {
4233 unsigned TmpReg = MRI.createVirtualRegister(RC);
4234 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
4235 TII.get(AArch64::SUBREG_TO_REG), TmpReg)
4237 .addReg(Op0, getKillRegState(Op0IsKill))
4238 .addImm(AArch64::sub_32);
4242 return fastEmitInst_rii(Opc, RC, Op0, Op0IsKill, ImmR, ImmS);
4245 unsigned AArch64FastISel::emitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT,
4247 assert(DestVT != MVT::i1 && "ZeroExt/SignExt an i1?");
4249 // FastISel does not have plumbing to deal with extensions where the SrcVT or
4250 // DestVT are odd things, so test to make sure that they are both types we can
4251 // handle (i1/i8/i16/i32 for SrcVT and i8/i16/i32/i64 for DestVT), otherwise
4252 // bail out to SelectionDAG.
4253 if (((DestVT != MVT::i8) && (DestVT != MVT::i16) &&
4254 (DestVT != MVT::i32) && (DestVT != MVT::i64)) ||
4255 ((SrcVT != MVT::i1) && (SrcVT != MVT::i8) &&
4256 (SrcVT != MVT::i16) && (SrcVT != MVT::i32)))
4262 switch (SrcVT.SimpleTy) {
4266 return emiti1Ext(SrcReg, DestVT, IsZExt);
4268 if (DestVT == MVT::i64)
4269 Opc = IsZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
4271 Opc = IsZExt ? AArch64::UBFMWri : AArch64::SBFMWri;
4275 if (DestVT == MVT::i64)
4276 Opc = IsZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
4278 Opc = IsZExt ? AArch64::UBFMWri : AArch64::SBFMWri;
4282 assert(DestVT == MVT::i64 && "IntExt i32 to i32?!?");
4283 Opc = IsZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
4288 // Handle i8 and i16 as i32.
4289 if (DestVT == MVT::i8 || DestVT == MVT::i16)
4291 else if (DestVT == MVT::i64) {
4292 unsigned Src64 = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
4293 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
4294 TII.get(AArch64::SUBREG_TO_REG), Src64)
4297 .addImm(AArch64::sub_32);
4301 const TargetRegisterClass *RC =
4302 (DestVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
4303 return fastEmitInst_rii(Opc, RC, SrcReg, /*TODO:IsKill=*/false, 0, Imm);
4306 static bool isZExtLoad(const MachineInstr *LI) {
4307 switch (LI->getOpcode()) {
4310 case AArch64::LDURBBi:
4311 case AArch64::LDURHHi:
4312 case AArch64::LDURWi:
4313 case AArch64::LDRBBui:
4314 case AArch64::LDRHHui:
4315 case AArch64::LDRWui:
4316 case AArch64::LDRBBroX:
4317 case AArch64::LDRHHroX:
4318 case AArch64::LDRWroX:
4319 case AArch64::LDRBBroW:
4320 case AArch64::LDRHHroW:
4321 case AArch64::LDRWroW:
4326 static bool isSExtLoad(const MachineInstr *LI) {
4327 switch (LI->getOpcode()) {
4330 case AArch64::LDURSBWi:
4331 case AArch64::LDURSHWi:
4332 case AArch64::LDURSBXi:
4333 case AArch64::LDURSHXi:
4334 case AArch64::LDURSWi:
4335 case AArch64::LDRSBWui:
4336 case AArch64::LDRSHWui:
4337 case AArch64::LDRSBXui:
4338 case AArch64::LDRSHXui:
4339 case AArch64::LDRSWui:
4340 case AArch64::LDRSBWroX:
4341 case AArch64::LDRSHWroX:
4342 case AArch64::LDRSBXroX:
4343 case AArch64::LDRSHXroX:
4344 case AArch64::LDRSWroX:
4345 case AArch64::LDRSBWroW:
4346 case AArch64::LDRSHWroW:
4347 case AArch64::LDRSBXroW:
4348 case AArch64::LDRSHXroW:
4349 case AArch64::LDRSWroW:
4354 bool AArch64FastISel::optimizeIntExtLoad(const Instruction *I, MVT RetVT,
4356 const auto *LI = dyn_cast<LoadInst>(I->getOperand(0));
4357 if (!LI || !LI->hasOneUse())
4360 // Check if the load instruction has already been selected.
4361 unsigned Reg = lookUpRegForValue(LI);
4365 MachineInstr *MI = MRI.getUniqueVRegDef(Reg);
4369 // Check if the correct load instruction has been emitted - SelectionDAG might
4370 // have emitted a zero-extending load, but we need a sign-extending load.
4371 bool IsZExt = isa<ZExtInst>(I);
4372 const auto *LoadMI = MI;
4373 if (LoadMI->getOpcode() == TargetOpcode::COPY &&
4374 LoadMI->getOperand(1).getSubReg() == AArch64::sub_32) {
4375 unsigned LoadReg = MI->getOperand(1).getReg();
4376 LoadMI = MRI.getUniqueVRegDef(LoadReg);
4377 assert(LoadMI && "Expected valid instruction");
4379 if (!(IsZExt && isZExtLoad(LoadMI)) && !(!IsZExt && isSExtLoad(LoadMI)))
4382 // Nothing to be done.
4383 if (RetVT != MVT::i64 || SrcVT > MVT::i32) {
4384 updateValueMap(I, Reg);
4389 unsigned Reg64 = createResultReg(&AArch64::GPR64RegClass);
4390 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
4391 TII.get(AArch64::SUBREG_TO_REG), Reg64)
4393 .addReg(Reg, getKillRegState(true))
4394 .addImm(AArch64::sub_32);
4397 assert((MI->getOpcode() == TargetOpcode::COPY &&
4398 MI->getOperand(1).getSubReg() == AArch64::sub_32) &&
4399 "Expected copy instruction");
4400 Reg = MI->getOperand(1).getReg();
4401 MI->eraseFromParent();
4403 updateValueMap(I, Reg);
4407 bool AArch64FastISel::selectIntExt(const Instruction *I) {
4408 assert((isa<ZExtInst>(I) || isa<SExtInst>(I)) &&
4409 "Unexpected integer extend instruction.");
4412 if (!isTypeSupported(I->getType(), RetVT))
4415 if (!isTypeSupported(I->getOperand(0)->getType(), SrcVT))
4418 // Try to optimize already sign-/zero-extended values from load instructions.
4419 if (optimizeIntExtLoad(I, RetVT, SrcVT))
4422 unsigned SrcReg = getRegForValue(I->getOperand(0));
4425 bool SrcIsKill = hasTrivialKill(I->getOperand(0));
4427 // Try to optimize already sign-/zero-extended values from function arguments.
4428 bool IsZExt = isa<ZExtInst>(I);
4429 if (const auto *Arg = dyn_cast<Argument>(I->getOperand(0))) {
4430 if ((IsZExt && Arg->hasZExtAttr()) || (!IsZExt && Arg->hasSExtAttr())) {
4431 if (RetVT == MVT::i64 && SrcVT != MVT::i64) {
4432 unsigned ResultReg = createResultReg(&AArch64::GPR64RegClass);
4433 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
4434 TII.get(AArch64::SUBREG_TO_REG), ResultReg)
4436 .addReg(SrcReg, getKillRegState(SrcIsKill))
4437 .addImm(AArch64::sub_32);
4440 // Conservatively clear all kill flags from all uses, because we are
4441 // replacing a sign-/zero-extend instruction at IR level with a nop at MI
4442 // level. The result of the instruction at IR level might have been
4443 // trivially dead, which is now not longer true.
4444 unsigned UseReg = lookUpRegForValue(I);
4446 MRI.clearKillFlags(UseReg);
4448 updateValueMap(I, SrcReg);
4453 unsigned ResultReg = emitIntExt(SrcVT, SrcReg, RetVT, IsZExt);
4457 updateValueMap(I, ResultReg);
4461 bool AArch64FastISel::selectRem(const Instruction *I, unsigned ISDOpcode) {
4462 EVT DestEVT = TLI.getValueType(I->getType(), true);
4463 if (!DestEVT.isSimple())
4466 MVT DestVT = DestEVT.getSimpleVT();
4467 if (DestVT != MVT::i64 && DestVT != MVT::i32)
4471 bool Is64bit = (DestVT == MVT::i64);
4472 switch (ISDOpcode) {
4476 DivOpc = Is64bit ? AArch64::SDIVXr : AArch64::SDIVWr;
4479 DivOpc = Is64bit ? AArch64::UDIVXr : AArch64::UDIVWr;
4482 unsigned MSubOpc = Is64bit ? AArch64::MSUBXrrr : AArch64::MSUBWrrr;
4483 unsigned Src0Reg = getRegForValue(I->getOperand(0));
4486 bool Src0IsKill = hasTrivialKill(I->getOperand(0));
4488 unsigned Src1Reg = getRegForValue(I->getOperand(1));
4491 bool Src1IsKill = hasTrivialKill(I->getOperand(1));
4493 const TargetRegisterClass *RC =
4494 (DestVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
4495 unsigned QuotReg = fastEmitInst_rr(DivOpc, RC, Src0Reg, /*IsKill=*/false,
4496 Src1Reg, /*IsKill=*/false);
4497 assert(QuotReg && "Unexpected DIV instruction emission failure.");
4498 // The remainder is computed as numerator - (quotient * denominator) using the
4499 // MSUB instruction.
4500 unsigned ResultReg = fastEmitInst_rrr(MSubOpc, RC, QuotReg, /*IsKill=*/true,
4501 Src1Reg, Src1IsKill, Src0Reg,
4503 updateValueMap(I, ResultReg);
4507 bool AArch64FastISel::selectMul(const Instruction *I) {
4509 if (!isTypeSupported(I->getType(), VT, /*IsVectorAllowed=*/true))
4513 return selectBinaryOp(I, ISD::MUL);
4515 const Value *Src0 = I->getOperand(0);
4516 const Value *Src1 = I->getOperand(1);
4517 if (const auto *C = dyn_cast<ConstantInt>(Src0))
4518 if (C->getValue().isPowerOf2())
4519 std::swap(Src0, Src1);
4521 // Try to simplify to a shift instruction.
4522 if (const auto *C = dyn_cast<ConstantInt>(Src1))
4523 if (C->getValue().isPowerOf2()) {
4524 uint64_t ShiftVal = C->getValue().logBase2();
4527 if (const auto *ZExt = dyn_cast<ZExtInst>(Src0)) {
4528 if (!isIntExtFree(ZExt)) {
4530 if (isValueAvailable(ZExt) && isTypeSupported(ZExt->getSrcTy(), VT)) {
4533 Src0 = ZExt->getOperand(0);
4536 } else if (const auto *SExt = dyn_cast<SExtInst>(Src0)) {
4537 if (!isIntExtFree(SExt)) {
4539 if (isValueAvailable(SExt) && isTypeSupported(SExt->getSrcTy(), VT)) {
4542 Src0 = SExt->getOperand(0);
4547 unsigned Src0Reg = getRegForValue(Src0);
4550 bool Src0IsKill = hasTrivialKill(Src0);
4552 unsigned ResultReg =
4553 emitLSL_ri(VT, SrcVT, Src0Reg, Src0IsKill, ShiftVal, IsZExt);
4556 updateValueMap(I, ResultReg);
4561 unsigned Src0Reg = getRegForValue(I->getOperand(0));
4564 bool Src0IsKill = hasTrivialKill(I->getOperand(0));
4566 unsigned Src1Reg = getRegForValue(I->getOperand(1));
4569 bool Src1IsKill = hasTrivialKill(I->getOperand(1));
4571 unsigned ResultReg = emitMul_rr(VT, Src0Reg, Src0IsKill, Src1Reg, Src1IsKill);
4576 updateValueMap(I, ResultReg);
4580 bool AArch64FastISel::selectShift(const Instruction *I) {
4582 if (!isTypeSupported(I->getType(), RetVT, /*IsVectorAllowed=*/true))
4585 if (RetVT.isVector())
4586 return selectOperator(I, I->getOpcode());
4588 if (const auto *C = dyn_cast<ConstantInt>(I->getOperand(1))) {
4589 unsigned ResultReg = 0;
4590 uint64_t ShiftVal = C->getZExtValue();
4592 bool IsZExt = I->getOpcode() != Instruction::AShr;
4593 const Value *Op0 = I->getOperand(0);
4594 if (const auto *ZExt = dyn_cast<ZExtInst>(Op0)) {
4595 if (!isIntExtFree(ZExt)) {
4597 if (isValueAvailable(ZExt) && isTypeSupported(ZExt->getSrcTy(), TmpVT)) {
4600 Op0 = ZExt->getOperand(0);
4603 } else if (const auto *SExt = dyn_cast<SExtInst>(Op0)) {
4604 if (!isIntExtFree(SExt)) {
4606 if (isValueAvailable(SExt) && isTypeSupported(SExt->getSrcTy(), TmpVT)) {
4609 Op0 = SExt->getOperand(0);
4614 unsigned Op0Reg = getRegForValue(Op0);
4617 bool Op0IsKill = hasTrivialKill(Op0);
4619 switch (I->getOpcode()) {
4620 default: llvm_unreachable("Unexpected instruction.");
4621 case Instruction::Shl:
4622 ResultReg = emitLSL_ri(RetVT, SrcVT, Op0Reg, Op0IsKill, ShiftVal, IsZExt);
4624 case Instruction::AShr:
4625 ResultReg = emitASR_ri(RetVT, SrcVT, Op0Reg, Op0IsKill, ShiftVal, IsZExt);
4627 case Instruction::LShr:
4628 ResultReg = emitLSR_ri(RetVT, SrcVT, Op0Reg, Op0IsKill, ShiftVal, IsZExt);
4634 updateValueMap(I, ResultReg);
4638 unsigned Op0Reg = getRegForValue(I->getOperand(0));
4641 bool Op0IsKill = hasTrivialKill(I->getOperand(0));
4643 unsigned Op1Reg = getRegForValue(I->getOperand(1));
4646 bool Op1IsKill = hasTrivialKill(I->getOperand(1));
4648 unsigned ResultReg = 0;
4649 switch (I->getOpcode()) {
4650 default: llvm_unreachable("Unexpected instruction.");
4651 case Instruction::Shl:
4652 ResultReg = emitLSL_rr(RetVT, Op0Reg, Op0IsKill, Op1Reg, Op1IsKill);
4654 case Instruction::AShr:
4655 ResultReg = emitASR_rr(RetVT, Op0Reg, Op0IsKill, Op1Reg, Op1IsKill);
4657 case Instruction::LShr:
4658 ResultReg = emitLSR_rr(RetVT, Op0Reg, Op0IsKill, Op1Reg, Op1IsKill);
4665 updateValueMap(I, ResultReg);
4669 bool AArch64FastISel::selectBitCast(const Instruction *I) {
4672 if (!isTypeLegal(I->getOperand(0)->getType(), SrcVT))
4674 if (!isTypeLegal(I->getType(), RetVT))
4678 if (RetVT == MVT::f32 && SrcVT == MVT::i32)
4679 Opc = AArch64::FMOVWSr;
4680 else if (RetVT == MVT::f64 && SrcVT == MVT::i64)
4681 Opc = AArch64::FMOVXDr;
4682 else if (RetVT == MVT::i32 && SrcVT == MVT::f32)
4683 Opc = AArch64::FMOVSWr;
4684 else if (RetVT == MVT::i64 && SrcVT == MVT::f64)
4685 Opc = AArch64::FMOVDXr;
4689 const TargetRegisterClass *RC = nullptr;
4690 switch (RetVT.SimpleTy) {
4691 default: llvm_unreachable("Unexpected value type.");
4692 case MVT::i32: RC = &AArch64::GPR32RegClass; break;
4693 case MVT::i64: RC = &AArch64::GPR64RegClass; break;
4694 case MVT::f32: RC = &AArch64::FPR32RegClass; break;
4695 case MVT::f64: RC = &AArch64::FPR64RegClass; break;
4697 unsigned Op0Reg = getRegForValue(I->getOperand(0));
4700 bool Op0IsKill = hasTrivialKill(I->getOperand(0));
4701 unsigned ResultReg = fastEmitInst_r(Opc, RC, Op0Reg, Op0IsKill);
4706 updateValueMap(I, ResultReg);
4710 bool AArch64FastISel::selectFRem(const Instruction *I) {
4712 if (!isTypeLegal(I->getType(), RetVT))
4716 switch (RetVT.SimpleTy) {
4720 LC = RTLIB::REM_F32;
4723 LC = RTLIB::REM_F64;
4728 Args.reserve(I->getNumOperands());
4730 // Populate the argument list.
4731 for (auto &Arg : I->operands()) {
4734 Entry.Ty = Arg->getType();
4735 Args.push_back(Entry);
4738 CallLoweringInfo CLI;
4739 MCContext &Ctx = MF->getContext();
4740 CLI.setCallee(DL, Ctx, TLI.getLibcallCallingConv(LC), I->getType(),
4741 TLI.getLibcallName(LC), std::move(Args));
4742 if (!lowerCallTo(CLI))
4744 updateValueMap(I, CLI.ResultReg);
4748 bool AArch64FastISel::selectSDiv(const Instruction *I) {
4750 if (!isTypeLegal(I->getType(), VT))
4753 if (!isa<ConstantInt>(I->getOperand(1)))
4754 return selectBinaryOp(I, ISD::SDIV);
4756 const APInt &C = cast<ConstantInt>(I->getOperand(1))->getValue();
4757 if ((VT != MVT::i32 && VT != MVT::i64) || !C ||
4758 !(C.isPowerOf2() || (-C).isPowerOf2()))
4759 return selectBinaryOp(I, ISD::SDIV);
4761 unsigned Lg2 = C.countTrailingZeros();
4762 unsigned Src0Reg = getRegForValue(I->getOperand(0));
4765 bool Src0IsKill = hasTrivialKill(I->getOperand(0));
4767 if (cast<BinaryOperator>(I)->isExact()) {
4768 unsigned ResultReg = emitASR_ri(VT, VT, Src0Reg, Src0IsKill, Lg2);
4771 updateValueMap(I, ResultReg);
4775 int64_t Pow2MinusOne = (1ULL << Lg2) - 1;
4776 unsigned AddReg = emitAdd_ri_(VT, Src0Reg, /*IsKill=*/false, Pow2MinusOne);
4780 // (Src0 < 0) ? Pow2 - 1 : 0;
4781 if (!emitICmp_ri(VT, Src0Reg, /*IsKill=*/false, 0))
4785 const TargetRegisterClass *RC;
4786 if (VT == MVT::i64) {
4787 SelectOpc = AArch64::CSELXr;
4788 RC = &AArch64::GPR64RegClass;
4790 SelectOpc = AArch64::CSELWr;
4791 RC = &AArch64::GPR32RegClass;
4793 unsigned SelectReg =
4794 fastEmitInst_rri(SelectOpc, RC, AddReg, /*IsKill=*/true, Src0Reg,
4795 Src0IsKill, AArch64CC::LT);
4799 // Divide by Pow2 --> ashr. If we're dividing by a negative value we must also
4800 // negate the result.
4801 unsigned ZeroReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
4804 ResultReg = emitAddSub_rs(/*UseAdd=*/false, VT, ZeroReg, /*IsKill=*/true,
4805 SelectReg, /*IsKill=*/true, AArch64_AM::ASR, Lg2);
4807 ResultReg = emitASR_ri(VT, VT, SelectReg, /*IsKill=*/true, Lg2);
4812 updateValueMap(I, ResultReg);
4816 /// This is mostly a copy of the existing FastISel getRegForGEPIndex code. We
4817 /// have to duplicate it for AArch64, because otherwise we would fail during the
4818 /// sign-extend emission.
4819 std::pair<unsigned, bool> AArch64FastISel::getRegForGEPIndex(const Value *Idx) {
4820 unsigned IdxN = getRegForValue(Idx);
4822 // Unhandled operand. Halt "fast" selection and bail.
4823 return std::pair<unsigned, bool>(0, false);
4825 bool IdxNIsKill = hasTrivialKill(Idx);
4827 // If the index is smaller or larger than intptr_t, truncate or extend it.
4828 MVT PtrVT = TLI.getPointerTy();
4829 EVT IdxVT = EVT::getEVT(Idx->getType(), /*HandleUnknown=*/false);
4830 if (IdxVT.bitsLT(PtrVT)) {
4831 IdxN = emitIntExt(IdxVT.getSimpleVT(), IdxN, PtrVT, /*IsZExt=*/false);
4833 } else if (IdxVT.bitsGT(PtrVT))
4834 llvm_unreachable("AArch64 FastISel doesn't support types larger than i64");
4835 return std::pair<unsigned, bool>(IdxN, IdxNIsKill);
4838 /// This is mostly a copy of the existing FastISel GEP code, but we have to
4839 /// duplicate it for AArch64, because otherwise we would bail out even for
4840 /// simple cases. This is because the standard fastEmit functions don't cover
4841 /// MUL at all and ADD is lowered very inefficientily.
4842 bool AArch64FastISel::selectGetElementPtr(const Instruction *I) {
4843 unsigned N = getRegForValue(I->getOperand(0));
4846 bool NIsKill = hasTrivialKill(I->getOperand(0));
4848 // Keep a running tab of the total offset to coalesce multiple N = N + Offset
4849 // into a single N = N + TotalOffset.
4850 uint64_t TotalOffs = 0;
4851 Type *Ty = I->getOperand(0)->getType();
4852 MVT VT = TLI.getPointerTy();
4853 for (auto OI = std::next(I->op_begin()), E = I->op_end(); OI != E; ++OI) {
4854 const Value *Idx = *OI;
4855 if (auto *StTy = dyn_cast<StructType>(Ty)) {
4856 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
4859 TotalOffs += DL.getStructLayout(StTy)->getElementOffset(Field);
4860 Ty = StTy->getElementType(Field);
4862 Ty = cast<SequentialType>(Ty)->getElementType();
4863 // If this is a constant subscript, handle it quickly.
4864 if (const auto *CI = dyn_cast<ConstantInt>(Idx)) {
4869 DL.getTypeAllocSize(Ty) * cast<ConstantInt>(CI)->getSExtValue();
4873 N = emitAdd_ri_(VT, N, NIsKill, TotalOffs);
4880 // N = N + Idx * ElementSize;
4881 uint64_t ElementSize = DL.getTypeAllocSize(Ty);
4882 std::pair<unsigned, bool> Pair = getRegForGEPIndex(Idx);
4883 unsigned IdxN = Pair.first;
4884 bool IdxNIsKill = Pair.second;
4888 if (ElementSize != 1) {
4889 unsigned C = fastEmit_i(VT, VT, ISD::Constant, ElementSize);
4892 IdxN = emitMul_rr(VT, IdxN, IdxNIsKill, C, true);
4897 N = fastEmit_rr(VT, VT, ISD::ADD, N, NIsKill, IdxN, IdxNIsKill);
4903 N = emitAdd_ri_(VT, N, NIsKill, TotalOffs);
4907 updateValueMap(I, N);
4911 bool AArch64FastISel::fastSelectInstruction(const Instruction *I) {
4912 switch (I->getOpcode()) {
4915 case Instruction::Add:
4916 case Instruction::Sub:
4917 return selectAddSub(I);
4918 case Instruction::Mul:
4919 return selectMul(I);
4920 case Instruction::SDiv:
4921 return selectSDiv(I);
4922 case Instruction::SRem:
4923 if (!selectBinaryOp(I, ISD::SREM))
4924 return selectRem(I, ISD::SREM);
4926 case Instruction::URem:
4927 if (!selectBinaryOp(I, ISD::UREM))
4928 return selectRem(I, ISD::UREM);
4930 case Instruction::Shl:
4931 case Instruction::LShr:
4932 case Instruction::AShr:
4933 return selectShift(I);
4934 case Instruction::And:
4935 case Instruction::Or:
4936 case Instruction::Xor:
4937 return selectLogicalOp(I);
4938 case Instruction::Br:
4939 return selectBranch(I);
4940 case Instruction::IndirectBr:
4941 return selectIndirectBr(I);
4942 case Instruction::BitCast:
4943 if (!FastISel::selectBitCast(I))
4944 return selectBitCast(I);
4946 case Instruction::FPToSI:
4947 if (!selectCast(I, ISD::FP_TO_SINT))
4948 return selectFPToInt(I, /*Signed=*/true);
4950 case Instruction::FPToUI:
4951 return selectFPToInt(I, /*Signed=*/false);
4952 case Instruction::ZExt:
4953 case Instruction::SExt:
4954 return selectIntExt(I);
4955 case Instruction::Trunc:
4956 if (!selectCast(I, ISD::TRUNCATE))
4957 return selectTrunc(I);
4959 case Instruction::FPExt:
4960 return selectFPExt(I);
4961 case Instruction::FPTrunc:
4962 return selectFPTrunc(I);
4963 case Instruction::SIToFP:
4964 if (!selectCast(I, ISD::SINT_TO_FP))
4965 return selectIntToFP(I, /*Signed=*/true);
4967 case Instruction::UIToFP:
4968 return selectIntToFP(I, /*Signed=*/false);
4969 case Instruction::Load:
4970 return selectLoad(I);
4971 case Instruction::Store:
4972 return selectStore(I);
4973 case Instruction::FCmp:
4974 case Instruction::ICmp:
4975 return selectCmp(I);
4976 case Instruction::Select:
4977 return selectSelect(I);
4978 case Instruction::Ret:
4979 return selectRet(I);
4980 case Instruction::FRem:
4981 return selectFRem(I);
4982 case Instruction::GetElementPtr:
4983 return selectGetElementPtr(I);
4986 // fall-back to target-independent instruction selection.
4987 return selectOperator(I, I->getOpcode());
4988 // Silence warnings.
4989 (void)&CC_AArch64_DarwinPCS_VarArg;
4993 llvm::FastISel *AArch64::createFastISel(FunctionLoweringInfo &FuncInfo,
4994 const TargetLibraryInfo *LibInfo) {
4995 return new AArch64FastISel(FuncInfo, LibInfo);