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"
44 class AArch64FastISel final : public FastISel {
54 AArch64_AM::ShiftExtendType ExtType;
62 const GlobalValue *GV;
65 Address() : Kind(RegBase), ExtType(AArch64_AM::InvalidShiftExtend),
66 OffsetReg(0), Shift(0), Offset(0), GV(nullptr) { Base.Reg = 0; }
67 void setKind(BaseKind K) { Kind = K; }
68 BaseKind getKind() const { return Kind; }
69 void setExtendType(AArch64_AM::ShiftExtendType E) { ExtType = E; }
70 AArch64_AM::ShiftExtendType getExtendType() const { return ExtType; }
71 bool isRegBase() const { return Kind == RegBase; }
72 bool isFIBase() const { return Kind == FrameIndexBase; }
73 void setReg(unsigned Reg) {
74 assert(isRegBase() && "Invalid base register access!");
77 unsigned getReg() const {
78 assert(isRegBase() && "Invalid base register access!");
81 void setOffsetReg(unsigned Reg) {
84 unsigned getOffsetReg() const {
87 void setFI(unsigned FI) {
88 assert(isFIBase() && "Invalid base frame index access!");
91 unsigned getFI() const {
92 assert(isFIBase() && "Invalid base frame index access!");
95 void setOffset(int64_t O) { Offset = O; }
96 int64_t getOffset() { return Offset; }
97 void setShift(unsigned S) { Shift = S; }
98 unsigned getShift() { return Shift; }
100 void setGlobalValue(const GlobalValue *G) { GV = G; }
101 const GlobalValue *getGlobalValue() { return GV; }
104 /// Subtarget - Keep a pointer to the AArch64Subtarget around so that we can
105 /// make the right decision when generating code for different targets.
106 const AArch64Subtarget *Subtarget;
107 LLVMContext *Context;
109 bool fastLowerArguments() override;
110 bool fastLowerCall(CallLoweringInfo &CLI) override;
111 bool fastLowerIntrinsicCall(const IntrinsicInst *II) override;
114 // Selection routines.
115 bool selectAddSub(const Instruction *I);
116 bool selectLogicalOp(const Instruction *I);
117 bool selectLoad(const Instruction *I);
118 bool selectStore(const Instruction *I);
119 bool selectBranch(const Instruction *I);
120 bool selectIndirectBr(const Instruction *I);
121 bool selectCmp(const Instruction *I);
122 bool selectSelect(const Instruction *I);
123 bool selectFPExt(const Instruction *I);
124 bool selectFPTrunc(const Instruction *I);
125 bool selectFPToInt(const Instruction *I, bool Signed);
126 bool selectIntToFP(const Instruction *I, bool Signed);
127 bool selectRem(const Instruction *I, unsigned ISDOpcode);
128 bool selectRet(const Instruction *I);
129 bool selectTrunc(const Instruction *I);
130 bool selectIntExt(const Instruction *I);
131 bool selectMul(const Instruction *I);
132 bool selectShift(const Instruction *I);
133 bool selectBitCast(const Instruction *I);
134 bool selectFRem(const Instruction *I);
135 bool selectSDiv(const Instruction *I);
136 bool selectGetElementPtr(const Instruction *I);
138 // Utility helper routines.
139 bool isTypeLegal(Type *Ty, MVT &VT);
140 bool isTypeSupported(Type *Ty, MVT &VT, bool IsVectorAllowed = false);
141 bool isValueAvailable(const Value *V) const;
142 bool computeAddress(const Value *Obj, Address &Addr, Type *Ty = nullptr);
143 bool computeCallAddress(const Value *V, Address &Addr);
144 bool simplifyAddress(Address &Addr, MVT VT);
145 void addLoadStoreOperands(Address &Addr, const MachineInstrBuilder &MIB,
146 MachineMemOperand::Flags Flags,
147 unsigned ScaleFactor, MachineMemOperand *MMO);
148 bool isMemCpySmall(uint64_t Len, unsigned Alignment);
149 bool tryEmitSmallMemCpy(Address Dest, Address Src, uint64_t Len,
151 bool foldXALUIntrinsic(AArch64CC::CondCode &CC, const Instruction *I,
153 bool optimizeIntExtLoad(const Instruction *I, MVT RetVT, MVT SrcVT);
154 bool optimizeSelect(const SelectInst *SI);
155 std::pair<unsigned, bool> getRegForGEPIndex(const Value *Idx);
157 // Emit helper routines.
158 unsigned emitAddSub(bool UseAdd, MVT RetVT, const Value *LHS,
159 const Value *RHS, bool SetFlags = false,
160 bool WantResult = true, bool IsZExt = false);
161 unsigned emitAddSub_rr(bool UseAdd, MVT RetVT, unsigned LHSReg,
162 bool LHSIsKill, unsigned RHSReg, bool RHSIsKill,
163 bool SetFlags = false, bool WantResult = true);
164 unsigned emitAddSub_ri(bool UseAdd, MVT RetVT, unsigned LHSReg,
165 bool LHSIsKill, uint64_t Imm, bool SetFlags = false,
166 bool WantResult = true);
167 unsigned emitAddSub_rs(bool UseAdd, MVT RetVT, unsigned LHSReg,
168 bool LHSIsKill, unsigned RHSReg, bool RHSIsKill,
169 AArch64_AM::ShiftExtendType ShiftType,
170 uint64_t ShiftImm, bool SetFlags = false,
171 bool WantResult = true);
172 unsigned emitAddSub_rx(bool UseAdd, MVT RetVT, unsigned LHSReg,
173 bool LHSIsKill, unsigned RHSReg, bool RHSIsKill,
174 AArch64_AM::ShiftExtendType ExtType,
175 uint64_t ShiftImm, bool SetFlags = false,
176 bool WantResult = true);
179 bool emitCompareAndBranch(const BranchInst *BI);
180 bool emitCmp(const Value *LHS, const Value *RHS, bool IsZExt);
181 bool emitICmp(MVT RetVT, const Value *LHS, const Value *RHS, bool IsZExt);
182 bool emitICmp_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill, uint64_t Imm);
183 bool emitFCmp(MVT RetVT, const Value *LHS, const Value *RHS);
184 unsigned emitLoad(MVT VT, MVT ResultVT, Address Addr, bool WantZExt = true,
185 MachineMemOperand *MMO = nullptr);
186 bool emitStore(MVT VT, unsigned SrcReg, Address Addr,
187 MachineMemOperand *MMO = nullptr);
188 unsigned emitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, bool isZExt);
189 unsigned emiti1Ext(unsigned SrcReg, MVT DestVT, bool isZExt);
190 unsigned emitAdd(MVT RetVT, const Value *LHS, const Value *RHS,
191 bool SetFlags = false, bool WantResult = true,
192 bool IsZExt = false);
193 unsigned emitAdd_ri_(MVT VT, unsigned Op0, bool Op0IsKill, int64_t Imm);
194 unsigned emitSub(MVT RetVT, const Value *LHS, const Value *RHS,
195 bool SetFlags = false, bool WantResult = true,
196 bool IsZExt = false);
197 unsigned emitSubs_rr(MVT RetVT, unsigned LHSReg, bool LHSIsKill,
198 unsigned RHSReg, bool RHSIsKill, bool WantResult = true);
199 unsigned emitSubs_rs(MVT RetVT, unsigned LHSReg, bool LHSIsKill,
200 unsigned RHSReg, bool RHSIsKill,
201 AArch64_AM::ShiftExtendType ShiftType, uint64_t ShiftImm,
202 bool WantResult = true);
203 unsigned emitLogicalOp(unsigned ISDOpc, MVT RetVT, const Value *LHS,
205 unsigned emitLogicalOp_ri(unsigned ISDOpc, MVT RetVT, unsigned LHSReg,
206 bool LHSIsKill, uint64_t Imm);
207 unsigned emitLogicalOp_rs(unsigned ISDOpc, MVT RetVT, unsigned LHSReg,
208 bool LHSIsKill, unsigned RHSReg, bool RHSIsKill,
210 unsigned emitAnd_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill, uint64_t Imm);
211 unsigned emitMul_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
212 unsigned Op1, bool Op1IsKill);
213 unsigned emitSMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
214 unsigned Op1, bool Op1IsKill);
215 unsigned emitUMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
216 unsigned Op1, bool Op1IsKill);
217 unsigned emitLSL_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
218 unsigned Op1Reg, bool Op1IsKill);
219 unsigned emitLSL_ri(MVT RetVT, MVT SrcVT, unsigned Op0Reg, bool Op0IsKill,
220 uint64_t Imm, bool IsZExt = true);
221 unsigned emitLSR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
222 unsigned Op1Reg, bool Op1IsKill);
223 unsigned emitLSR_ri(MVT RetVT, MVT SrcVT, unsigned Op0Reg, bool Op0IsKill,
224 uint64_t Imm, bool IsZExt = true);
225 unsigned emitASR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
226 unsigned Op1Reg, bool Op1IsKill);
227 unsigned emitASR_ri(MVT RetVT, MVT SrcVT, unsigned Op0Reg, bool Op0IsKill,
228 uint64_t Imm, bool IsZExt = false);
230 unsigned materializeInt(const ConstantInt *CI, MVT VT);
231 unsigned materializeFP(const ConstantFP *CFP, MVT VT);
232 unsigned materializeGV(const GlobalValue *GV);
234 // Call handling routines.
236 CCAssignFn *CCAssignFnForCall(CallingConv::ID CC) const;
237 bool processCallArgs(CallLoweringInfo &CLI, SmallVectorImpl<MVT> &ArgVTs,
239 bool finishCall(CallLoweringInfo &CLI, MVT RetVT, unsigned NumBytes);
242 // Backend specific FastISel code.
243 unsigned fastMaterializeAlloca(const AllocaInst *AI) override;
244 unsigned fastMaterializeConstant(const Constant *C) override;
245 unsigned fastMaterializeFloatZero(const ConstantFP* CF) override;
247 explicit AArch64FastISel(FunctionLoweringInfo &FuncInfo,
248 const TargetLibraryInfo *LibInfo)
249 : FastISel(FuncInfo, LibInfo, /*SkipTargetIndependentISel=*/true) {
251 &static_cast<const AArch64Subtarget &>(FuncInfo.MF->getSubtarget());
252 Context = &FuncInfo.Fn->getContext();
255 bool fastSelectInstruction(const Instruction *I) override;
257 #include "AArch64GenFastISel.inc"
260 } // end anonymous namespace
262 #include "AArch64GenCallingConv.inc"
264 /// \brief Check if the sign-/zero-extend will be a noop.
265 static bool isIntExtFree(const Instruction *I) {
266 assert((isa<ZExtInst>(I) || isa<SExtInst>(I)) &&
267 "Unexpected integer extend instruction.");
268 assert(!I->getType()->isVectorTy() && I->getType()->isIntegerTy() &&
269 "Unexpected value type.");
270 bool IsZExt = isa<ZExtInst>(I);
272 if (const auto *LI = dyn_cast<LoadInst>(I->getOperand(0)))
276 if (const auto *Arg = dyn_cast<Argument>(I->getOperand(0)))
277 if ((IsZExt && Arg->hasZExtAttr()) || (!IsZExt && Arg->hasSExtAttr()))
283 /// \brief Determine the implicit scale factor that is applied by a memory
284 /// operation for a given value type.
285 static unsigned getImplicitScaleFactor(MVT VT) {
286 switch (VT.SimpleTy) {
289 case MVT::i1: // fall-through
294 case MVT::i32: // fall-through
297 case MVT::i64: // fall-through
303 CCAssignFn *AArch64FastISel::CCAssignFnForCall(CallingConv::ID CC) const {
304 if (CC == CallingConv::WebKit_JS)
305 return CC_AArch64_WebKit_JS;
306 if (CC == CallingConv::GHC)
307 return CC_AArch64_GHC;
308 return Subtarget->isTargetDarwin() ? CC_AArch64_DarwinPCS : CC_AArch64_AAPCS;
311 unsigned AArch64FastISel::fastMaterializeAlloca(const AllocaInst *AI) {
312 assert(TLI.getValueType(DL, AI->getType(), true) == MVT::i64 &&
313 "Alloca should always return a pointer.");
315 // Don't handle dynamic allocas.
316 if (!FuncInfo.StaticAllocaMap.count(AI))
319 DenseMap<const AllocaInst *, int>::iterator SI =
320 FuncInfo.StaticAllocaMap.find(AI);
322 if (SI != FuncInfo.StaticAllocaMap.end()) {
323 unsigned ResultReg = createResultReg(&AArch64::GPR64spRegClass);
324 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri),
326 .addFrameIndex(SI->second)
335 unsigned AArch64FastISel::materializeInt(const ConstantInt *CI, MVT VT) {
340 return fastEmit_i(VT, VT, ISD::Constant, CI->getZExtValue());
342 // Create a copy from the zero register to materialize a "0" value.
343 const TargetRegisterClass *RC = (VT == MVT::i64) ? &AArch64::GPR64RegClass
344 : &AArch64::GPR32RegClass;
345 unsigned ZeroReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
346 unsigned ResultReg = createResultReg(RC);
347 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TargetOpcode::COPY),
348 ResultReg).addReg(ZeroReg, getKillRegState(true));
352 unsigned AArch64FastISel::materializeFP(const ConstantFP *CFP, MVT VT) {
353 // Positive zero (+0.0) has to be materialized with a fmov from the zero
354 // register, because the immediate version of fmov cannot encode zero.
355 if (CFP->isNullValue())
356 return fastMaterializeFloatZero(CFP);
358 if (VT != MVT::f32 && VT != MVT::f64)
361 const APFloat Val = CFP->getValueAPF();
362 bool Is64Bit = (VT == MVT::f64);
363 // This checks to see if we can use FMOV instructions to materialize
364 // a constant, otherwise we have to materialize via the constant pool.
365 if (TLI.isFPImmLegal(Val, VT)) {
367 Is64Bit ? AArch64_AM::getFP64Imm(Val) : AArch64_AM::getFP32Imm(Val);
368 assert((Imm != -1) && "Cannot encode floating-point constant.");
369 unsigned Opc = Is64Bit ? AArch64::FMOVDi : AArch64::FMOVSi;
370 return fastEmitInst_i(Opc, TLI.getRegClassFor(VT), Imm);
373 // For the MachO large code model materialize the FP constant in code.
374 if (Subtarget->isTargetMachO() && TM.getCodeModel() == CodeModel::Large) {
375 unsigned Opc1 = Is64Bit ? AArch64::MOVi64imm : AArch64::MOVi32imm;
376 const TargetRegisterClass *RC = Is64Bit ?
377 &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
379 unsigned TmpReg = createResultReg(RC);
380 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc1), TmpReg)
381 .addImm(CFP->getValueAPF().bitcastToAPInt().getZExtValue());
383 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
384 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
385 TII.get(TargetOpcode::COPY), ResultReg)
386 .addReg(TmpReg, getKillRegState(true));
391 // Materialize via constant pool. MachineConstantPool wants an explicit
393 unsigned Align = DL.getPrefTypeAlignment(CFP->getType());
395 Align = DL.getTypeAllocSize(CFP->getType());
397 unsigned CPI = MCP.getConstantPoolIndex(cast<Constant>(CFP), Align);
398 unsigned ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
399 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
400 ADRPReg).addConstantPoolIndex(CPI, 0, AArch64II::MO_PAGE);
402 unsigned Opc = Is64Bit ? AArch64::LDRDui : AArch64::LDRSui;
403 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
404 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
406 .addConstantPoolIndex(CPI, 0, AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
410 unsigned AArch64FastISel::materializeGV(const GlobalValue *GV) {
411 // We can't handle thread-local variables quickly yet.
412 if (GV->isThreadLocal())
415 // MachO still uses GOT for large code-model accesses, but ELF requires
416 // movz/movk sequences, which FastISel doesn't handle yet.
417 if (TM.getCodeModel() != CodeModel::Small && !Subtarget->isTargetMachO())
420 unsigned char OpFlags = Subtarget->ClassifyGlobalReference(GV, TM);
422 EVT DestEVT = TLI.getValueType(DL, GV->getType(), true);
423 if (!DestEVT.isSimple())
426 unsigned ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
429 if (OpFlags & AArch64II::MO_GOT) {
431 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
433 .addGlobalAddress(GV, 0, AArch64II::MO_GOT | AArch64II::MO_PAGE);
435 ResultReg = createResultReg(&AArch64::GPR64RegClass);
436 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::LDRXui),
439 .addGlobalAddress(GV, 0, AArch64II::MO_GOT | AArch64II::MO_PAGEOFF |
443 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
445 .addGlobalAddress(GV, 0, AArch64II::MO_PAGE);
447 ResultReg = createResultReg(&AArch64::GPR64spRegClass);
448 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri),
451 .addGlobalAddress(GV, 0, AArch64II::MO_PAGEOFF | AArch64II::MO_NC)
457 unsigned AArch64FastISel::fastMaterializeConstant(const Constant *C) {
458 EVT CEVT = TLI.getValueType(DL, C->getType(), true);
460 // Only handle simple types.
461 if (!CEVT.isSimple())
463 MVT VT = CEVT.getSimpleVT();
465 if (const auto *CI = dyn_cast<ConstantInt>(C))
466 return materializeInt(CI, VT);
467 else if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
468 return materializeFP(CFP, VT);
469 else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
470 return materializeGV(GV);
475 unsigned AArch64FastISel::fastMaterializeFloatZero(const ConstantFP* CFP) {
476 assert(CFP->isNullValue() &&
477 "Floating-point constant is not a positive zero.");
479 if (!isTypeLegal(CFP->getType(), VT))
482 if (VT != MVT::f32 && VT != MVT::f64)
485 bool Is64Bit = (VT == MVT::f64);
486 unsigned ZReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
487 unsigned Opc = Is64Bit ? AArch64::FMOVXDr : AArch64::FMOVWSr;
488 return fastEmitInst_r(Opc, TLI.getRegClassFor(VT), ZReg, /*IsKill=*/true);
491 /// \brief Check if the multiply is by a power-of-2 constant.
492 static bool isMulPowOf2(const Value *I) {
493 if (const auto *MI = dyn_cast<MulOperator>(I)) {
494 if (const auto *C = dyn_cast<ConstantInt>(MI->getOperand(0)))
495 if (C->getValue().isPowerOf2())
497 if (const auto *C = dyn_cast<ConstantInt>(MI->getOperand(1)))
498 if (C->getValue().isPowerOf2())
504 // Computes the address to get to an object.
505 bool AArch64FastISel::computeAddress(const Value *Obj, Address &Addr, Type *Ty)
507 const User *U = nullptr;
508 unsigned Opcode = Instruction::UserOp1;
509 if (const Instruction *I = dyn_cast<Instruction>(Obj)) {
510 // Don't walk into other basic blocks unless the object is an alloca from
511 // another block, otherwise it may not have a virtual register assigned.
512 if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(Obj)) ||
513 FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
514 Opcode = I->getOpcode();
517 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Obj)) {
518 Opcode = C->getOpcode();
522 if (auto *Ty = dyn_cast<PointerType>(Obj->getType()))
523 if (Ty->getAddressSpace() > 255)
524 // Fast instruction selection doesn't support the special
531 case Instruction::BitCast: {
532 // Look through bitcasts.
533 return computeAddress(U->getOperand(0), Addr, Ty);
535 case Instruction::IntToPtr: {
536 // Look past no-op inttoptrs.
537 if (TLI.getValueType(DL, U->getOperand(0)->getType()) ==
538 TLI.getPointerTy(DL))
539 return computeAddress(U->getOperand(0), Addr, Ty);
542 case Instruction::PtrToInt: {
543 // Look past no-op ptrtoints.
544 if (TLI.getValueType(DL, U->getType()) == TLI.getPointerTy(DL))
545 return computeAddress(U->getOperand(0), Addr, Ty);
548 case Instruction::GetElementPtr: {
549 Address SavedAddr = Addr;
550 uint64_t TmpOffset = Addr.getOffset();
552 // Iterate through the GEP folding the constants into offsets where
554 for (gep_type_iterator GTI = gep_type_begin(U), E = gep_type_end(U);
556 const Value *Op = GTI.getOperand();
557 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
558 const StructLayout *SL = DL.getStructLayout(STy);
559 unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
560 TmpOffset += SL->getElementOffset(Idx);
562 uint64_t S = DL.getTypeAllocSize(GTI.getIndexedType());
564 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
565 // Constant-offset addressing.
566 TmpOffset += CI->getSExtValue() * S;
569 if (canFoldAddIntoGEP(U, Op)) {
570 // A compatible add with a constant operand. Fold the constant.
572 cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
573 TmpOffset += CI->getSExtValue() * S;
574 // Iterate on the other operand.
575 Op = cast<AddOperator>(Op)->getOperand(0);
579 goto unsupported_gep;
584 // Try to grab the base operand now.
585 Addr.setOffset(TmpOffset);
586 if (computeAddress(U->getOperand(0), Addr, Ty))
589 // We failed, restore everything and try the other options.
595 case Instruction::Alloca: {
596 const AllocaInst *AI = cast<AllocaInst>(Obj);
597 DenseMap<const AllocaInst *, int>::iterator SI =
598 FuncInfo.StaticAllocaMap.find(AI);
599 if (SI != FuncInfo.StaticAllocaMap.end()) {
600 Addr.setKind(Address::FrameIndexBase);
601 Addr.setFI(SI->second);
606 case Instruction::Add: {
607 // Adds of constants are common and easy enough.
608 const Value *LHS = U->getOperand(0);
609 const Value *RHS = U->getOperand(1);
611 if (isa<ConstantInt>(LHS))
614 if (const ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
615 Addr.setOffset(Addr.getOffset() + CI->getSExtValue());
616 return computeAddress(LHS, Addr, Ty);
619 Address Backup = Addr;
620 if (computeAddress(LHS, Addr, Ty) && computeAddress(RHS, Addr, Ty))
626 case Instruction::Sub: {
627 // Subs of constants are common and easy enough.
628 const Value *LHS = U->getOperand(0);
629 const Value *RHS = U->getOperand(1);
631 if (const ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
632 Addr.setOffset(Addr.getOffset() - CI->getSExtValue());
633 return computeAddress(LHS, Addr, Ty);
637 case Instruction::Shl: {
638 if (Addr.getOffsetReg())
641 const auto *CI = dyn_cast<ConstantInt>(U->getOperand(1));
645 unsigned Val = CI->getZExtValue();
646 if (Val < 1 || Val > 3)
649 uint64_t NumBytes = 0;
650 if (Ty && Ty->isSized()) {
651 uint64_t NumBits = DL.getTypeSizeInBits(Ty);
652 NumBytes = NumBits / 8;
653 if (!isPowerOf2_64(NumBits))
657 if (NumBytes != (1ULL << Val))
661 Addr.setExtendType(AArch64_AM::LSL);
663 const Value *Src = U->getOperand(0);
664 if (const auto *I = dyn_cast<Instruction>(Src)) {
665 if (FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
666 // Fold the zext or sext when it won't become a noop.
667 if (const auto *ZE = dyn_cast<ZExtInst>(I)) {
668 if (!isIntExtFree(ZE) &&
669 ZE->getOperand(0)->getType()->isIntegerTy(32)) {
670 Addr.setExtendType(AArch64_AM::UXTW);
671 Src = ZE->getOperand(0);
673 } else if (const auto *SE = dyn_cast<SExtInst>(I)) {
674 if (!isIntExtFree(SE) &&
675 SE->getOperand(0)->getType()->isIntegerTy(32)) {
676 Addr.setExtendType(AArch64_AM::SXTW);
677 Src = SE->getOperand(0);
683 if (const auto *AI = dyn_cast<BinaryOperator>(Src))
684 if (AI->getOpcode() == Instruction::And) {
685 const Value *LHS = AI->getOperand(0);
686 const Value *RHS = AI->getOperand(1);
688 if (const auto *C = dyn_cast<ConstantInt>(LHS))
689 if (C->getValue() == 0xffffffff)
692 if (const auto *C = dyn_cast<ConstantInt>(RHS))
693 if (C->getValue() == 0xffffffff) {
694 Addr.setExtendType(AArch64_AM::UXTW);
695 unsigned Reg = getRegForValue(LHS);
698 bool RegIsKill = hasTrivialKill(LHS);
699 Reg = fastEmitInst_extractsubreg(MVT::i32, Reg, RegIsKill,
701 Addr.setOffsetReg(Reg);
706 unsigned Reg = getRegForValue(Src);
709 Addr.setOffsetReg(Reg);
712 case Instruction::Mul: {
713 if (Addr.getOffsetReg())
719 const Value *LHS = U->getOperand(0);
720 const Value *RHS = U->getOperand(1);
722 // Canonicalize power-of-2 value to the RHS.
723 if (const auto *C = dyn_cast<ConstantInt>(LHS))
724 if (C->getValue().isPowerOf2())
727 assert(isa<ConstantInt>(RHS) && "Expected an ConstantInt.");
728 const auto *C = cast<ConstantInt>(RHS);
729 unsigned Val = C->getValue().logBase2();
730 if (Val < 1 || Val > 3)
733 uint64_t NumBytes = 0;
734 if (Ty && Ty->isSized()) {
735 uint64_t NumBits = DL.getTypeSizeInBits(Ty);
736 NumBytes = NumBits / 8;
737 if (!isPowerOf2_64(NumBits))
741 if (NumBytes != (1ULL << Val))
745 Addr.setExtendType(AArch64_AM::LSL);
747 const Value *Src = LHS;
748 if (const auto *I = dyn_cast<Instruction>(Src)) {
749 if (FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
750 // Fold the zext or sext when it won't become a noop.
751 if (const auto *ZE = dyn_cast<ZExtInst>(I)) {
752 if (!isIntExtFree(ZE) &&
753 ZE->getOperand(0)->getType()->isIntegerTy(32)) {
754 Addr.setExtendType(AArch64_AM::UXTW);
755 Src = ZE->getOperand(0);
757 } else if (const auto *SE = dyn_cast<SExtInst>(I)) {
758 if (!isIntExtFree(SE) &&
759 SE->getOperand(0)->getType()->isIntegerTy(32)) {
760 Addr.setExtendType(AArch64_AM::SXTW);
761 Src = SE->getOperand(0);
767 unsigned Reg = getRegForValue(Src);
770 Addr.setOffsetReg(Reg);
773 case Instruction::And: {
774 if (Addr.getOffsetReg())
777 if (!Ty || DL.getTypeSizeInBits(Ty) != 8)
780 const Value *LHS = U->getOperand(0);
781 const Value *RHS = U->getOperand(1);
783 if (const auto *C = dyn_cast<ConstantInt>(LHS))
784 if (C->getValue() == 0xffffffff)
787 if (const auto *C = dyn_cast<ConstantInt>(RHS))
788 if (C->getValue() == 0xffffffff) {
790 Addr.setExtendType(AArch64_AM::LSL);
791 Addr.setExtendType(AArch64_AM::UXTW);
793 unsigned Reg = getRegForValue(LHS);
796 bool RegIsKill = hasTrivialKill(LHS);
797 Reg = fastEmitInst_extractsubreg(MVT::i32, Reg, RegIsKill,
799 Addr.setOffsetReg(Reg);
804 case Instruction::SExt:
805 case Instruction::ZExt: {
806 if (!Addr.getReg() || Addr.getOffsetReg())
809 const Value *Src = nullptr;
810 // Fold the zext or sext when it won't become a noop.
811 if (const auto *ZE = dyn_cast<ZExtInst>(U)) {
812 if (!isIntExtFree(ZE) && ZE->getOperand(0)->getType()->isIntegerTy(32)) {
813 Addr.setExtendType(AArch64_AM::UXTW);
814 Src = ZE->getOperand(0);
816 } else if (const auto *SE = dyn_cast<SExtInst>(U)) {
817 if (!isIntExtFree(SE) && SE->getOperand(0)->getType()->isIntegerTy(32)) {
818 Addr.setExtendType(AArch64_AM::SXTW);
819 Src = SE->getOperand(0);
827 unsigned Reg = getRegForValue(Src);
830 Addr.setOffsetReg(Reg);
835 if (Addr.isRegBase() && !Addr.getReg()) {
836 unsigned Reg = getRegForValue(Obj);
843 if (!Addr.getOffsetReg()) {
844 unsigned Reg = getRegForValue(Obj);
847 Addr.setOffsetReg(Reg);
854 bool AArch64FastISel::computeCallAddress(const Value *V, Address &Addr) {
855 const User *U = nullptr;
856 unsigned Opcode = Instruction::UserOp1;
859 if (const auto *I = dyn_cast<Instruction>(V)) {
860 Opcode = I->getOpcode();
862 InMBB = I->getParent() == FuncInfo.MBB->getBasicBlock();
863 } else if (const auto *C = dyn_cast<ConstantExpr>(V)) {
864 Opcode = C->getOpcode();
870 case Instruction::BitCast:
871 // Look past bitcasts if its operand is in the same BB.
873 return computeCallAddress(U->getOperand(0), Addr);
875 case Instruction::IntToPtr:
876 // Look past no-op inttoptrs if its operand is in the same BB.
878 TLI.getValueType(DL, U->getOperand(0)->getType()) ==
879 TLI.getPointerTy(DL))
880 return computeCallAddress(U->getOperand(0), Addr);
882 case Instruction::PtrToInt:
883 // Look past no-op ptrtoints if its operand is in the same BB.
884 if (InMBB && TLI.getValueType(DL, U->getType()) == TLI.getPointerTy(DL))
885 return computeCallAddress(U->getOperand(0), Addr);
889 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
890 Addr.setGlobalValue(GV);
894 // If all else fails, try to materialize the value in a register.
895 if (!Addr.getGlobalValue()) {
896 Addr.setReg(getRegForValue(V));
897 return Addr.getReg() != 0;
904 bool AArch64FastISel::isTypeLegal(Type *Ty, MVT &VT) {
905 EVT evt = TLI.getValueType(DL, Ty, true);
907 // Only handle simple types.
908 if (evt == MVT::Other || !evt.isSimple())
910 VT = evt.getSimpleVT();
912 // This is a legal type, but it's not something we handle in fast-isel.
916 // Handle all other legal types, i.e. a register that will directly hold this
918 return TLI.isTypeLegal(VT);
921 /// \brief Determine if the value type is supported by FastISel.
923 /// FastISel for AArch64 can handle more value types than are legal. This adds
924 /// simple value type such as i1, i8, and i16.
925 bool AArch64FastISel::isTypeSupported(Type *Ty, MVT &VT, bool IsVectorAllowed) {
926 if (Ty->isVectorTy() && !IsVectorAllowed)
929 if (isTypeLegal(Ty, VT))
932 // If this is a type than can be sign or zero-extended to a basic operation
933 // go ahead and accept it now.
934 if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16)
940 bool AArch64FastISel::isValueAvailable(const Value *V) const {
941 if (!isa<Instruction>(V))
944 const auto *I = cast<Instruction>(V);
945 return FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB;
948 bool AArch64FastISel::simplifyAddress(Address &Addr, MVT VT) {
949 unsigned ScaleFactor = getImplicitScaleFactor(VT);
953 bool ImmediateOffsetNeedsLowering = false;
954 bool RegisterOffsetNeedsLowering = false;
955 int64_t Offset = Addr.getOffset();
956 if (((Offset < 0) || (Offset & (ScaleFactor - 1))) && !isInt<9>(Offset))
957 ImmediateOffsetNeedsLowering = true;
958 else if (Offset > 0 && !(Offset & (ScaleFactor - 1)) &&
959 !isUInt<12>(Offset / ScaleFactor))
960 ImmediateOffsetNeedsLowering = true;
962 // Cannot encode an offset register and an immediate offset in the same
963 // instruction. Fold the immediate offset into the load/store instruction and
964 // emit an additional add to take care of the offset register.
965 if (!ImmediateOffsetNeedsLowering && Addr.getOffset() && Addr.getOffsetReg())
966 RegisterOffsetNeedsLowering = true;
968 // Cannot encode zero register as base.
969 if (Addr.isRegBase() && Addr.getOffsetReg() && !Addr.getReg())
970 RegisterOffsetNeedsLowering = true;
972 // If this is a stack pointer and the offset needs to be simplified then put
973 // the alloca address into a register, set the base type back to register and
974 // continue. This should almost never happen.
975 if ((ImmediateOffsetNeedsLowering || Addr.getOffsetReg()) && Addr.isFIBase())
977 unsigned ResultReg = createResultReg(&AArch64::GPR64spRegClass);
978 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri),
980 .addFrameIndex(Addr.getFI())
983 Addr.setKind(Address::RegBase);
984 Addr.setReg(ResultReg);
987 if (RegisterOffsetNeedsLowering) {
988 unsigned ResultReg = 0;
990 if (Addr.getExtendType() == AArch64_AM::SXTW ||
991 Addr.getExtendType() == AArch64_AM::UXTW )
992 ResultReg = emitAddSub_rx(/*UseAdd=*/true, MVT::i64, Addr.getReg(),
993 /*TODO:IsKill=*/false, Addr.getOffsetReg(),
994 /*TODO:IsKill=*/false, Addr.getExtendType(),
997 ResultReg = emitAddSub_rs(/*UseAdd=*/true, MVT::i64, Addr.getReg(),
998 /*TODO:IsKill=*/false, Addr.getOffsetReg(),
999 /*TODO:IsKill=*/false, AArch64_AM::LSL,
1002 if (Addr.getExtendType() == AArch64_AM::UXTW)
1003 ResultReg = emitLSL_ri(MVT::i64, MVT::i32, Addr.getOffsetReg(),
1004 /*Op0IsKill=*/false, Addr.getShift(),
1006 else if (Addr.getExtendType() == AArch64_AM::SXTW)
1007 ResultReg = emitLSL_ri(MVT::i64, MVT::i32, Addr.getOffsetReg(),
1008 /*Op0IsKill=*/false, Addr.getShift(),
1011 ResultReg = emitLSL_ri(MVT::i64, MVT::i64, Addr.getOffsetReg(),
1012 /*Op0IsKill=*/false, Addr.getShift());
1017 Addr.setReg(ResultReg);
1018 Addr.setOffsetReg(0);
1020 Addr.setExtendType(AArch64_AM::InvalidShiftExtend);
1023 // Since the offset is too large for the load/store instruction get the
1024 // reg+offset into a register.
1025 if (ImmediateOffsetNeedsLowering) {
1028 // Try to fold the immediate into the add instruction.
1029 ResultReg = emitAdd_ri_(MVT::i64, Addr.getReg(), /*IsKill=*/false, Offset);
1031 ResultReg = fastEmit_i(MVT::i64, MVT::i64, ISD::Constant, Offset);
1035 Addr.setReg(ResultReg);
1041 void AArch64FastISel::addLoadStoreOperands(Address &Addr,
1042 const MachineInstrBuilder &MIB,
1043 MachineMemOperand::Flags Flags,
1044 unsigned ScaleFactor,
1045 MachineMemOperand *MMO) {
1046 int64_t Offset = Addr.getOffset() / ScaleFactor;
1047 // Frame base works a bit differently. Handle it separately.
1048 if (Addr.isFIBase()) {
1049 int FI = Addr.getFI();
1050 // FIXME: We shouldn't be using getObjectSize/getObjectAlignment. The size
1051 // and alignment should be based on the VT.
1052 MMO = FuncInfo.MF->getMachineMemOperand(
1053 MachinePointerInfo::getFixedStack(*FuncInfo.MF, FI, Offset), Flags,
1054 MFI.getObjectSize(FI), MFI.getObjectAlignment(FI));
1055 // Now add the rest of the operands.
1056 MIB.addFrameIndex(FI).addImm(Offset);
1058 assert(Addr.isRegBase() && "Unexpected address kind.");
1059 const MCInstrDesc &II = MIB->getDesc();
1060 unsigned Idx = (Flags & MachineMemOperand::MOStore) ? 1 : 0;
1062 constrainOperandRegClass(II, Addr.getReg(), II.getNumDefs()+Idx));
1064 constrainOperandRegClass(II, Addr.getOffsetReg(), II.getNumDefs()+Idx+1));
1065 if (Addr.getOffsetReg()) {
1066 assert(Addr.getOffset() == 0 && "Unexpected offset");
1067 bool IsSigned = Addr.getExtendType() == AArch64_AM::SXTW ||
1068 Addr.getExtendType() == AArch64_AM::SXTX;
1069 MIB.addReg(Addr.getReg());
1070 MIB.addReg(Addr.getOffsetReg());
1071 MIB.addImm(IsSigned);
1072 MIB.addImm(Addr.getShift() != 0);
1074 MIB.addReg(Addr.getReg()).addImm(Offset);
1078 MIB.addMemOperand(MMO);
1081 unsigned AArch64FastISel::emitAddSub(bool UseAdd, MVT RetVT, const Value *LHS,
1082 const Value *RHS, bool SetFlags,
1083 bool WantResult, bool IsZExt) {
1084 AArch64_AM::ShiftExtendType ExtendType = AArch64_AM::InvalidShiftExtend;
1085 bool NeedExtend = false;
1086 switch (RetVT.SimpleTy) {
1094 ExtendType = IsZExt ? AArch64_AM::UXTB : AArch64_AM::SXTB;
1098 ExtendType = IsZExt ? AArch64_AM::UXTH : AArch64_AM::SXTH;
1100 case MVT::i32: // fall-through
1105 RetVT.SimpleTy = std::max(RetVT.SimpleTy, MVT::i32);
1107 // Canonicalize immediates to the RHS first.
1108 if (UseAdd && isa<Constant>(LHS) && !isa<Constant>(RHS))
1109 std::swap(LHS, RHS);
1111 // Canonicalize mul by power of 2 to the RHS.
1112 if (UseAdd && LHS->hasOneUse() && isValueAvailable(LHS))
1113 if (isMulPowOf2(LHS))
1114 std::swap(LHS, RHS);
1116 // Canonicalize shift immediate to the RHS.
1117 if (UseAdd && LHS->hasOneUse() && isValueAvailable(LHS))
1118 if (const auto *SI = dyn_cast<BinaryOperator>(LHS))
1119 if (isa<ConstantInt>(SI->getOperand(1)))
1120 if (SI->getOpcode() == Instruction::Shl ||
1121 SI->getOpcode() == Instruction::LShr ||
1122 SI->getOpcode() == Instruction::AShr )
1123 std::swap(LHS, RHS);
1125 unsigned LHSReg = getRegForValue(LHS);
1128 bool LHSIsKill = hasTrivialKill(LHS);
1131 LHSReg = emitIntExt(SrcVT, LHSReg, RetVT, IsZExt);
1133 unsigned ResultReg = 0;
1134 if (const auto *C = dyn_cast<ConstantInt>(RHS)) {
1135 uint64_t Imm = IsZExt ? C->getZExtValue() : C->getSExtValue();
1136 if (C->isNegative())
1137 ResultReg = emitAddSub_ri(!UseAdd, RetVT, LHSReg, LHSIsKill, -Imm,
1138 SetFlags, WantResult);
1140 ResultReg = emitAddSub_ri(UseAdd, RetVT, LHSReg, LHSIsKill, Imm, SetFlags,
1142 } else if (const auto *C = dyn_cast<Constant>(RHS))
1143 if (C->isNullValue())
1144 ResultReg = emitAddSub_ri(UseAdd, RetVT, LHSReg, LHSIsKill, 0, SetFlags,
1150 // Only extend the RHS within the instruction if there is a valid extend type.
1151 if (ExtendType != AArch64_AM::InvalidShiftExtend && RHS->hasOneUse() &&
1152 isValueAvailable(RHS)) {
1153 if (const auto *SI = dyn_cast<BinaryOperator>(RHS))
1154 if (const auto *C = dyn_cast<ConstantInt>(SI->getOperand(1)))
1155 if ((SI->getOpcode() == Instruction::Shl) && (C->getZExtValue() < 4)) {
1156 unsigned RHSReg = getRegForValue(SI->getOperand(0));
1159 bool RHSIsKill = hasTrivialKill(SI->getOperand(0));
1160 return emitAddSub_rx(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg,
1161 RHSIsKill, ExtendType, C->getZExtValue(),
1162 SetFlags, WantResult);
1164 unsigned RHSReg = getRegForValue(RHS);
1167 bool RHSIsKill = hasTrivialKill(RHS);
1168 return emitAddSub_rx(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg, RHSIsKill,
1169 ExtendType, 0, SetFlags, WantResult);
1172 // Check if the mul can be folded into the instruction.
1173 if (RHS->hasOneUse() && isValueAvailable(RHS)) {
1174 if (isMulPowOf2(RHS)) {
1175 const Value *MulLHS = cast<MulOperator>(RHS)->getOperand(0);
1176 const Value *MulRHS = cast<MulOperator>(RHS)->getOperand(1);
1178 if (const auto *C = dyn_cast<ConstantInt>(MulLHS))
1179 if (C->getValue().isPowerOf2())
1180 std::swap(MulLHS, MulRHS);
1182 assert(isa<ConstantInt>(MulRHS) && "Expected a ConstantInt.");
1183 uint64_t ShiftVal = cast<ConstantInt>(MulRHS)->getValue().logBase2();
1184 unsigned RHSReg = getRegForValue(MulLHS);
1187 bool RHSIsKill = hasTrivialKill(MulLHS);
1188 ResultReg = emitAddSub_rs(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg,
1189 RHSIsKill, AArch64_AM::LSL, ShiftVal, SetFlags,
1196 // Check if the shift can be folded into the instruction.
1197 if (RHS->hasOneUse() && isValueAvailable(RHS)) {
1198 if (const auto *SI = dyn_cast<BinaryOperator>(RHS)) {
1199 if (const auto *C = dyn_cast<ConstantInt>(SI->getOperand(1))) {
1200 AArch64_AM::ShiftExtendType ShiftType = AArch64_AM::InvalidShiftExtend;
1201 switch (SI->getOpcode()) {
1203 case Instruction::Shl: ShiftType = AArch64_AM::LSL; break;
1204 case Instruction::LShr: ShiftType = AArch64_AM::LSR; break;
1205 case Instruction::AShr: ShiftType = AArch64_AM::ASR; break;
1207 uint64_t ShiftVal = C->getZExtValue();
1208 if (ShiftType != AArch64_AM::InvalidShiftExtend) {
1209 unsigned RHSReg = getRegForValue(SI->getOperand(0));
1212 bool RHSIsKill = hasTrivialKill(SI->getOperand(0));
1213 ResultReg = emitAddSub_rs(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg,
1214 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 // Don't deal with undefined shifts.
1326 if (ShiftImm >= RetVT.getSizeInBits())
1329 static const unsigned OpcTable[2][2][2] = {
1330 { { AArch64::SUBWrs, AArch64::SUBXrs },
1331 { AArch64::ADDWrs, AArch64::ADDXrs } },
1332 { { AArch64::SUBSWrs, AArch64::SUBSXrs },
1333 { AArch64::ADDSWrs, AArch64::ADDSXrs } }
1335 bool Is64Bit = RetVT == MVT::i64;
1336 unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit];
1337 const TargetRegisterClass *RC =
1338 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
1341 ResultReg = createResultReg(RC);
1343 ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
1345 const MCInstrDesc &II = TII.get(Opc);
1346 LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
1347 RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1);
1348 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
1349 .addReg(LHSReg, getKillRegState(LHSIsKill))
1350 .addReg(RHSReg, getKillRegState(RHSIsKill))
1351 .addImm(getShifterImm(ShiftType, ShiftImm));
1355 unsigned AArch64FastISel::emitAddSub_rx(bool UseAdd, MVT RetVT, unsigned LHSReg,
1356 bool LHSIsKill, unsigned RHSReg,
1358 AArch64_AM::ShiftExtendType ExtType,
1359 uint64_t ShiftImm, bool SetFlags,
1361 assert(LHSReg && RHSReg && "Invalid register number.");
1363 if (RetVT != MVT::i32 && RetVT != MVT::i64)
1369 static const unsigned OpcTable[2][2][2] = {
1370 { { AArch64::SUBWrx, AArch64::SUBXrx },
1371 { AArch64::ADDWrx, AArch64::ADDXrx } },
1372 { { AArch64::SUBSWrx, AArch64::SUBSXrx },
1373 { AArch64::ADDSWrx, AArch64::ADDSXrx } }
1375 bool Is64Bit = RetVT == MVT::i64;
1376 unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit];
1377 const TargetRegisterClass *RC = nullptr;
1379 RC = Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
1381 RC = Is64Bit ? &AArch64::GPR64spRegClass : &AArch64::GPR32spRegClass;
1384 ResultReg = createResultReg(RC);
1386 ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
1388 const MCInstrDesc &II = TII.get(Opc);
1389 LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
1390 RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1);
1391 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
1392 .addReg(LHSReg, getKillRegState(LHSIsKill))
1393 .addReg(RHSReg, getKillRegState(RHSIsKill))
1394 .addImm(getArithExtendImm(ExtType, ShiftImm));
1398 bool AArch64FastISel::emitCmp(const Value *LHS, const Value *RHS, bool IsZExt) {
1399 Type *Ty = LHS->getType();
1400 EVT EVT = TLI.getValueType(DL, Ty, true);
1401 if (!EVT.isSimple())
1403 MVT VT = EVT.getSimpleVT();
1405 switch (VT.SimpleTy) {
1413 return emitICmp(VT, LHS, RHS, IsZExt);
1416 return emitFCmp(VT, LHS, RHS);
1420 bool AArch64FastISel::emitICmp(MVT RetVT, const Value *LHS, const Value *RHS,
1422 return emitSub(RetVT, LHS, RHS, /*SetFlags=*/true, /*WantResult=*/false,
1426 bool AArch64FastISel::emitICmp_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill,
1428 return emitAddSub_ri(/*UseAdd=*/false, RetVT, LHSReg, LHSIsKill, Imm,
1429 /*SetFlags=*/true, /*WantResult=*/false) != 0;
1432 bool AArch64FastISel::emitFCmp(MVT RetVT, const Value *LHS, const Value *RHS) {
1433 if (RetVT != MVT::f32 && RetVT != MVT::f64)
1436 // Check to see if the 2nd operand is a constant that we can encode directly
1438 bool UseImm = false;
1439 if (const auto *CFP = dyn_cast<ConstantFP>(RHS))
1440 if (CFP->isZero() && !CFP->isNegative())
1443 unsigned LHSReg = getRegForValue(LHS);
1446 bool LHSIsKill = hasTrivialKill(LHS);
1449 unsigned Opc = (RetVT == MVT::f64) ? AArch64::FCMPDri : AArch64::FCMPSri;
1450 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc))
1451 .addReg(LHSReg, getKillRegState(LHSIsKill));
1455 unsigned RHSReg = getRegForValue(RHS);
1458 bool RHSIsKill = hasTrivialKill(RHS);
1460 unsigned Opc = (RetVT == MVT::f64) ? AArch64::FCMPDrr : AArch64::FCMPSrr;
1461 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc))
1462 .addReg(LHSReg, getKillRegState(LHSIsKill))
1463 .addReg(RHSReg, getKillRegState(RHSIsKill));
1467 unsigned AArch64FastISel::emitAdd(MVT RetVT, const Value *LHS, const Value *RHS,
1468 bool SetFlags, bool WantResult, bool IsZExt) {
1469 return emitAddSub(/*UseAdd=*/true, RetVT, LHS, RHS, SetFlags, WantResult,
1473 /// \brief This method is a wrapper to simplify add emission.
1475 /// First try to emit an add with an immediate operand using emitAddSub_ri. If
1476 /// that fails, then try to materialize the immediate into a register and use
1477 /// emitAddSub_rr instead.
1478 unsigned AArch64FastISel::emitAdd_ri_(MVT VT, unsigned Op0, bool Op0IsKill,
1482 ResultReg = emitAddSub_ri(false, VT, Op0, Op0IsKill, -Imm);
1484 ResultReg = emitAddSub_ri(true, VT, Op0, Op0IsKill, Imm);
1489 unsigned CReg = fastEmit_i(VT, VT, ISD::Constant, Imm);
1493 ResultReg = emitAddSub_rr(true, VT, Op0, Op0IsKill, CReg, true);
1497 unsigned AArch64FastISel::emitSub(MVT RetVT, const Value *LHS, const Value *RHS,
1498 bool SetFlags, bool WantResult, bool IsZExt) {
1499 return emitAddSub(/*UseAdd=*/false, RetVT, LHS, RHS, SetFlags, WantResult,
1503 unsigned AArch64FastISel::emitSubs_rr(MVT RetVT, unsigned LHSReg,
1504 bool LHSIsKill, unsigned RHSReg,
1505 bool RHSIsKill, bool WantResult) {
1506 return emitAddSub_rr(/*UseAdd=*/false, RetVT, LHSReg, LHSIsKill, RHSReg,
1507 RHSIsKill, /*SetFlags=*/true, WantResult);
1510 unsigned AArch64FastISel::emitSubs_rs(MVT RetVT, unsigned LHSReg,
1511 bool LHSIsKill, unsigned RHSReg,
1513 AArch64_AM::ShiftExtendType ShiftType,
1514 uint64_t ShiftImm, bool WantResult) {
1515 return emitAddSub_rs(/*UseAdd=*/false, RetVT, LHSReg, LHSIsKill, RHSReg,
1516 RHSIsKill, ShiftType, ShiftImm, /*SetFlags=*/true,
1520 unsigned AArch64FastISel::emitLogicalOp(unsigned ISDOpc, MVT RetVT,
1521 const Value *LHS, const Value *RHS) {
1522 // Canonicalize immediates to the RHS first.
1523 if (isa<ConstantInt>(LHS) && !isa<ConstantInt>(RHS))
1524 std::swap(LHS, RHS);
1526 // Canonicalize mul by power-of-2 to the RHS.
1527 if (LHS->hasOneUse() && isValueAvailable(LHS))
1528 if (isMulPowOf2(LHS))
1529 std::swap(LHS, RHS);
1531 // Canonicalize shift immediate to the RHS.
1532 if (LHS->hasOneUse() && isValueAvailable(LHS))
1533 if (const auto *SI = dyn_cast<ShlOperator>(LHS))
1534 if (isa<ConstantInt>(SI->getOperand(1)))
1535 std::swap(LHS, RHS);
1537 unsigned LHSReg = getRegForValue(LHS);
1540 bool LHSIsKill = hasTrivialKill(LHS);
1542 unsigned ResultReg = 0;
1543 if (const auto *C = dyn_cast<ConstantInt>(RHS)) {
1544 uint64_t Imm = C->getZExtValue();
1545 ResultReg = emitLogicalOp_ri(ISDOpc, RetVT, LHSReg, LHSIsKill, Imm);
1550 // Check if the mul can be folded into the instruction.
1551 if (RHS->hasOneUse() && isValueAvailable(RHS)) {
1552 if (isMulPowOf2(RHS)) {
1553 const Value *MulLHS = cast<MulOperator>(RHS)->getOperand(0);
1554 const Value *MulRHS = cast<MulOperator>(RHS)->getOperand(1);
1556 if (const auto *C = dyn_cast<ConstantInt>(MulLHS))
1557 if (C->getValue().isPowerOf2())
1558 std::swap(MulLHS, MulRHS);
1560 assert(isa<ConstantInt>(MulRHS) && "Expected a ConstantInt.");
1561 uint64_t ShiftVal = cast<ConstantInt>(MulRHS)->getValue().logBase2();
1563 unsigned RHSReg = getRegForValue(MulLHS);
1566 bool RHSIsKill = hasTrivialKill(MulLHS);
1567 ResultReg = emitLogicalOp_rs(ISDOpc, RetVT, LHSReg, LHSIsKill, RHSReg,
1568 RHSIsKill, ShiftVal);
1574 // Check if the shift can be folded into the instruction.
1575 if (RHS->hasOneUse() && isValueAvailable(RHS)) {
1576 if (const auto *SI = dyn_cast<ShlOperator>(RHS))
1577 if (const auto *C = dyn_cast<ConstantInt>(SI->getOperand(1))) {
1578 uint64_t ShiftVal = C->getZExtValue();
1579 unsigned RHSReg = getRegForValue(SI->getOperand(0));
1582 bool RHSIsKill = hasTrivialKill(SI->getOperand(0));
1583 ResultReg = emitLogicalOp_rs(ISDOpc, RetVT, LHSReg, LHSIsKill, RHSReg,
1584 RHSIsKill, ShiftVal);
1590 unsigned RHSReg = getRegForValue(RHS);
1593 bool RHSIsKill = hasTrivialKill(RHS);
1595 MVT VT = std::max(MVT::i32, RetVT.SimpleTy);
1596 ResultReg = fastEmit_rr(VT, VT, ISDOpc, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
1597 if (RetVT >= MVT::i8 && RetVT <= MVT::i16) {
1598 uint64_t Mask = (RetVT == MVT::i8) ? 0xff : 0xffff;
1599 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
1604 unsigned AArch64FastISel::emitLogicalOp_ri(unsigned ISDOpc, MVT RetVT,
1605 unsigned LHSReg, bool LHSIsKill,
1607 static_assert((ISD::AND + 1 == ISD::OR) && (ISD::AND + 2 == ISD::XOR),
1608 "ISD nodes are not consecutive!");
1609 static const unsigned OpcTable[3][2] = {
1610 { AArch64::ANDWri, AArch64::ANDXri },
1611 { AArch64::ORRWri, AArch64::ORRXri },
1612 { AArch64::EORWri, AArch64::EORXri }
1614 const TargetRegisterClass *RC;
1617 switch (RetVT.SimpleTy) {
1624 unsigned Idx = ISDOpc - ISD::AND;
1625 Opc = OpcTable[Idx][0];
1626 RC = &AArch64::GPR32spRegClass;
1631 Opc = OpcTable[ISDOpc - ISD::AND][1];
1632 RC = &AArch64::GPR64spRegClass;
1637 if (!AArch64_AM::isLogicalImmediate(Imm, RegSize))
1640 unsigned ResultReg =
1641 fastEmitInst_ri(Opc, RC, LHSReg, LHSIsKill,
1642 AArch64_AM::encodeLogicalImmediate(Imm, RegSize));
1643 if (RetVT >= MVT::i8 && RetVT <= MVT::i16 && ISDOpc != ISD::AND) {
1644 uint64_t Mask = (RetVT == MVT::i8) ? 0xff : 0xffff;
1645 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
1650 unsigned AArch64FastISel::emitLogicalOp_rs(unsigned ISDOpc, MVT RetVT,
1651 unsigned LHSReg, bool LHSIsKill,
1652 unsigned RHSReg, bool RHSIsKill,
1653 uint64_t ShiftImm) {
1654 static_assert((ISD::AND + 1 == ISD::OR) && (ISD::AND + 2 == ISD::XOR),
1655 "ISD nodes are not consecutive!");
1656 static const unsigned OpcTable[3][2] = {
1657 { AArch64::ANDWrs, AArch64::ANDXrs },
1658 { AArch64::ORRWrs, AArch64::ORRXrs },
1659 { AArch64::EORWrs, AArch64::EORXrs }
1662 // Don't deal with undefined shifts.
1663 if (ShiftImm >= RetVT.getSizeInBits())
1666 const TargetRegisterClass *RC;
1668 switch (RetVT.SimpleTy) {
1675 Opc = OpcTable[ISDOpc - ISD::AND][0];
1676 RC = &AArch64::GPR32RegClass;
1679 Opc = OpcTable[ISDOpc - ISD::AND][1];
1680 RC = &AArch64::GPR64RegClass;
1683 unsigned ResultReg =
1684 fastEmitInst_rri(Opc, RC, LHSReg, LHSIsKill, RHSReg, RHSIsKill,
1685 AArch64_AM::getShifterImm(AArch64_AM::LSL, ShiftImm));
1686 if (RetVT >= MVT::i8 && RetVT <= MVT::i16) {
1687 uint64_t Mask = (RetVT == MVT::i8) ? 0xff : 0xffff;
1688 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
1693 unsigned AArch64FastISel::emitAnd_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill,
1695 return emitLogicalOp_ri(ISD::AND, RetVT, LHSReg, LHSIsKill, Imm);
1698 unsigned AArch64FastISel::emitLoad(MVT VT, MVT RetVT, Address Addr,
1699 bool WantZExt, MachineMemOperand *MMO) {
1700 if (!TLI.allowsMisalignedMemoryAccesses(VT))
1703 // Simplify this down to something we can handle.
1704 if (!simplifyAddress(Addr, VT))
1707 unsigned ScaleFactor = getImplicitScaleFactor(VT);
1709 llvm_unreachable("Unexpected value type.");
1711 // Negative offsets require unscaled, 9-bit, signed immediate offsets.
1712 // Otherwise, we try using scaled, 12-bit, unsigned immediate offsets.
1713 bool UseScaled = true;
1714 if ((Addr.getOffset() < 0) || (Addr.getOffset() & (ScaleFactor - 1))) {
1719 static const unsigned GPOpcTable[2][8][4] = {
1721 { { AArch64::LDURSBWi, AArch64::LDURSHWi, AArch64::LDURWi,
1723 { AArch64::LDURSBXi, AArch64::LDURSHXi, AArch64::LDURSWi,
1725 { AArch64::LDRSBWui, AArch64::LDRSHWui, AArch64::LDRWui,
1727 { AArch64::LDRSBXui, AArch64::LDRSHXui, AArch64::LDRSWui,
1729 { AArch64::LDRSBWroX, AArch64::LDRSHWroX, AArch64::LDRWroX,
1731 { AArch64::LDRSBXroX, AArch64::LDRSHXroX, AArch64::LDRSWroX,
1733 { AArch64::LDRSBWroW, AArch64::LDRSHWroW, AArch64::LDRWroW,
1735 { AArch64::LDRSBXroW, AArch64::LDRSHXroW, AArch64::LDRSWroW,
1739 { { AArch64::LDURBBi, AArch64::LDURHHi, AArch64::LDURWi,
1741 { AArch64::LDURBBi, AArch64::LDURHHi, AArch64::LDURWi,
1743 { AArch64::LDRBBui, AArch64::LDRHHui, AArch64::LDRWui,
1745 { AArch64::LDRBBui, AArch64::LDRHHui, AArch64::LDRWui,
1747 { AArch64::LDRBBroX, AArch64::LDRHHroX, AArch64::LDRWroX,
1749 { AArch64::LDRBBroX, AArch64::LDRHHroX, AArch64::LDRWroX,
1751 { AArch64::LDRBBroW, AArch64::LDRHHroW, AArch64::LDRWroW,
1753 { AArch64::LDRBBroW, AArch64::LDRHHroW, AArch64::LDRWroW,
1758 static const unsigned FPOpcTable[4][2] = {
1759 { AArch64::LDURSi, AArch64::LDURDi },
1760 { AArch64::LDRSui, AArch64::LDRDui },
1761 { AArch64::LDRSroX, AArch64::LDRDroX },
1762 { AArch64::LDRSroW, AArch64::LDRDroW }
1766 const TargetRegisterClass *RC;
1767 bool UseRegOffset = Addr.isRegBase() && !Addr.getOffset() && Addr.getReg() &&
1768 Addr.getOffsetReg();
1769 unsigned Idx = UseRegOffset ? 2 : UseScaled ? 1 : 0;
1770 if (Addr.getExtendType() == AArch64_AM::UXTW ||
1771 Addr.getExtendType() == AArch64_AM::SXTW)
1774 bool IsRet64Bit = RetVT == MVT::i64;
1775 switch (VT.SimpleTy) {
1777 llvm_unreachable("Unexpected value type.");
1778 case MVT::i1: // Intentional fall-through.
1780 Opc = GPOpcTable[WantZExt][2 * Idx + IsRet64Bit][0];
1781 RC = (IsRet64Bit && !WantZExt) ?
1782 &AArch64::GPR64RegClass: &AArch64::GPR32RegClass;
1785 Opc = GPOpcTable[WantZExt][2 * Idx + IsRet64Bit][1];
1786 RC = (IsRet64Bit && !WantZExt) ?
1787 &AArch64::GPR64RegClass: &AArch64::GPR32RegClass;
1790 Opc = GPOpcTable[WantZExt][2 * Idx + IsRet64Bit][2];
1791 RC = (IsRet64Bit && !WantZExt) ?
1792 &AArch64::GPR64RegClass: &AArch64::GPR32RegClass;
1795 Opc = GPOpcTable[WantZExt][2 * Idx + IsRet64Bit][3];
1796 RC = &AArch64::GPR64RegClass;
1799 Opc = FPOpcTable[Idx][0];
1800 RC = &AArch64::FPR32RegClass;
1803 Opc = FPOpcTable[Idx][1];
1804 RC = &AArch64::FPR64RegClass;
1808 // Create the base instruction, then add the operands.
1809 unsigned ResultReg = createResultReg(RC);
1810 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1811 TII.get(Opc), ResultReg);
1812 addLoadStoreOperands(Addr, MIB, MachineMemOperand::MOLoad, ScaleFactor, MMO);
1814 // Loading an i1 requires special handling.
1815 if (VT == MVT::i1) {
1816 unsigned ANDReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, 1);
1817 assert(ANDReg && "Unexpected AND instruction emission failure.");
1821 // For zero-extending loads to 64bit we emit a 32bit load and then convert
1822 // the 32bit reg to a 64bit reg.
1823 if (WantZExt && RetVT == MVT::i64 && VT <= MVT::i32) {
1824 unsigned Reg64 = createResultReg(&AArch64::GPR64RegClass);
1825 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1826 TII.get(AArch64::SUBREG_TO_REG), Reg64)
1828 .addReg(ResultReg, getKillRegState(true))
1829 .addImm(AArch64::sub_32);
1835 bool AArch64FastISel::selectAddSub(const Instruction *I) {
1837 if (!isTypeSupported(I->getType(), VT, /*IsVectorAllowed=*/true))
1841 return selectOperator(I, I->getOpcode());
1844 switch (I->getOpcode()) {
1846 llvm_unreachable("Unexpected instruction.");
1847 case Instruction::Add:
1848 ResultReg = emitAdd(VT, I->getOperand(0), I->getOperand(1));
1850 case Instruction::Sub:
1851 ResultReg = emitSub(VT, I->getOperand(0), I->getOperand(1));
1857 updateValueMap(I, ResultReg);
1861 bool AArch64FastISel::selectLogicalOp(const Instruction *I) {
1863 if (!isTypeSupported(I->getType(), VT, /*IsVectorAllowed=*/true))
1867 return selectOperator(I, I->getOpcode());
1870 switch (I->getOpcode()) {
1872 llvm_unreachable("Unexpected instruction.");
1873 case Instruction::And:
1874 ResultReg = emitLogicalOp(ISD::AND, VT, I->getOperand(0), I->getOperand(1));
1876 case Instruction::Or:
1877 ResultReg = emitLogicalOp(ISD::OR, VT, I->getOperand(0), I->getOperand(1));
1879 case Instruction::Xor:
1880 ResultReg = emitLogicalOp(ISD::XOR, VT, I->getOperand(0), I->getOperand(1));
1886 updateValueMap(I, ResultReg);
1890 bool AArch64FastISel::selectLoad(const Instruction *I) {
1892 // Verify we have a legal type before going any further. Currently, we handle
1893 // simple types that will directly fit in a register (i32/f32/i64/f64) or
1894 // those that can be sign or zero-extended to a basic operation (i1/i8/i16).
1895 if (!isTypeSupported(I->getType(), VT, /*IsVectorAllowed=*/true) ||
1896 cast<LoadInst>(I)->isAtomic())
1899 const Value *SV = I->getOperand(0);
1900 if (TLI.supportSwiftError()) {
1901 // Swifterror values can come from either a function parameter with
1902 // swifterror attribute or an alloca with swifterror attribute.
1903 if (const Argument *Arg = dyn_cast<Argument>(SV)) {
1904 if (Arg->hasSwiftErrorAttr())
1908 if (const AllocaInst *Alloca = dyn_cast<AllocaInst>(SV)) {
1909 if (Alloca->isSwiftError())
1914 // See if we can handle this address.
1916 if (!computeAddress(I->getOperand(0), Addr, I->getType()))
1919 // Fold the following sign-/zero-extend into the load instruction.
1920 bool WantZExt = true;
1922 const Value *IntExtVal = nullptr;
1923 if (I->hasOneUse()) {
1924 if (const auto *ZE = dyn_cast<ZExtInst>(I->use_begin()->getUser())) {
1925 if (isTypeSupported(ZE->getType(), RetVT))
1929 } else if (const auto *SE = dyn_cast<SExtInst>(I->use_begin()->getUser())) {
1930 if (isTypeSupported(SE->getType(), RetVT))
1938 unsigned ResultReg =
1939 emitLoad(VT, RetVT, Addr, WantZExt, createMachineMemOperandFor(I));
1943 // There are a few different cases we have to handle, because the load or the
1944 // sign-/zero-extend might not be selected by FastISel if we fall-back to
1945 // SelectionDAG. There is also an ordering issue when both instructions are in
1946 // different basic blocks.
1947 // 1.) The load instruction is selected by FastISel, but the integer extend
1948 // not. This usually happens when the integer extend is in a different
1949 // basic block and SelectionDAG took over for that basic block.
1950 // 2.) The load instruction is selected before the integer extend. This only
1951 // happens when the integer extend is in a different basic block.
1952 // 3.) The load instruction is selected by SelectionDAG and the integer extend
1953 // by FastISel. This happens if there are instructions between the load
1954 // and the integer extend that couldn't be selected by FastISel.
1956 // The integer extend hasn't been emitted yet. FastISel or SelectionDAG
1957 // could select it. Emit a copy to subreg if necessary. FastISel will remove
1958 // it when it selects the integer extend.
1959 unsigned Reg = lookUpRegForValue(IntExtVal);
1960 auto *MI = MRI.getUniqueVRegDef(Reg);
1962 if (RetVT == MVT::i64 && VT <= MVT::i32) {
1964 // Delete the last emitted instruction from emitLoad (SUBREG_TO_REG).
1965 std::prev(FuncInfo.InsertPt)->eraseFromParent();
1966 ResultReg = std::prev(FuncInfo.InsertPt)->getOperand(0).getReg();
1968 ResultReg = fastEmitInst_extractsubreg(MVT::i32, ResultReg,
1972 updateValueMap(I, ResultReg);
1976 // The integer extend has already been emitted - delete all the instructions
1977 // that have been emitted by the integer extend lowering code and use the
1978 // result from the load instruction directly.
1981 for (auto &Opnd : MI->uses()) {
1983 Reg = Opnd.getReg();
1987 MI->eraseFromParent();
1990 MI = MRI.getUniqueVRegDef(Reg);
1992 updateValueMap(IntExtVal, ResultReg);
1996 updateValueMap(I, ResultReg);
2000 bool AArch64FastISel::emitStore(MVT VT, unsigned SrcReg, Address Addr,
2001 MachineMemOperand *MMO) {
2002 if (!TLI.allowsMisalignedMemoryAccesses(VT))
2005 // Simplify this down to something we can handle.
2006 if (!simplifyAddress(Addr, VT))
2009 unsigned ScaleFactor = getImplicitScaleFactor(VT);
2011 llvm_unreachable("Unexpected value type.");
2013 // Negative offsets require unscaled, 9-bit, signed immediate offsets.
2014 // Otherwise, we try using scaled, 12-bit, unsigned immediate offsets.
2015 bool UseScaled = true;
2016 if ((Addr.getOffset() < 0) || (Addr.getOffset() & (ScaleFactor - 1))) {
2021 static const unsigned OpcTable[4][6] = {
2022 { AArch64::STURBBi, AArch64::STURHHi, AArch64::STURWi, AArch64::STURXi,
2023 AArch64::STURSi, AArch64::STURDi },
2024 { AArch64::STRBBui, AArch64::STRHHui, AArch64::STRWui, AArch64::STRXui,
2025 AArch64::STRSui, AArch64::STRDui },
2026 { AArch64::STRBBroX, AArch64::STRHHroX, AArch64::STRWroX, AArch64::STRXroX,
2027 AArch64::STRSroX, AArch64::STRDroX },
2028 { AArch64::STRBBroW, AArch64::STRHHroW, AArch64::STRWroW, AArch64::STRXroW,
2029 AArch64::STRSroW, AArch64::STRDroW }
2033 bool VTIsi1 = false;
2034 bool UseRegOffset = Addr.isRegBase() && !Addr.getOffset() && Addr.getReg() &&
2035 Addr.getOffsetReg();
2036 unsigned Idx = UseRegOffset ? 2 : UseScaled ? 1 : 0;
2037 if (Addr.getExtendType() == AArch64_AM::UXTW ||
2038 Addr.getExtendType() == AArch64_AM::SXTW)
2041 switch (VT.SimpleTy) {
2042 default: llvm_unreachable("Unexpected value type.");
2043 case MVT::i1: VTIsi1 = true;
2044 case MVT::i8: Opc = OpcTable[Idx][0]; break;
2045 case MVT::i16: Opc = OpcTable[Idx][1]; break;
2046 case MVT::i32: Opc = OpcTable[Idx][2]; break;
2047 case MVT::i64: Opc = OpcTable[Idx][3]; break;
2048 case MVT::f32: Opc = OpcTable[Idx][4]; break;
2049 case MVT::f64: Opc = OpcTable[Idx][5]; break;
2052 // Storing an i1 requires special handling.
2053 if (VTIsi1 && SrcReg != AArch64::WZR) {
2054 unsigned ANDReg = emitAnd_ri(MVT::i32, SrcReg, /*TODO:IsKill=*/false, 1);
2055 assert(ANDReg && "Unexpected AND instruction emission failure.");
2058 // Create the base instruction, then add the operands.
2059 const MCInstrDesc &II = TII.get(Opc);
2060 SrcReg = constrainOperandRegClass(II, SrcReg, II.getNumDefs());
2061 MachineInstrBuilder MIB =
2062 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II).addReg(SrcReg);
2063 addLoadStoreOperands(Addr, MIB, MachineMemOperand::MOStore, ScaleFactor, MMO);
2068 bool AArch64FastISel::selectStore(const Instruction *I) {
2070 const Value *Op0 = I->getOperand(0);
2071 // Verify we have a legal type before going any further. Currently, we handle
2072 // simple types that will directly fit in a register (i32/f32/i64/f64) or
2073 // those that can be sign or zero-extended to a basic operation (i1/i8/i16).
2074 if (!isTypeSupported(Op0->getType(), VT, /*IsVectorAllowed=*/true) ||
2075 cast<StoreInst>(I)->isAtomic())
2078 const Value *PtrV = I->getOperand(1);
2079 if (TLI.supportSwiftError()) {
2080 // Swifterror values can come from either a function parameter with
2081 // swifterror attribute or an alloca with swifterror attribute.
2082 if (const Argument *Arg = dyn_cast<Argument>(PtrV)) {
2083 if (Arg->hasSwiftErrorAttr())
2087 if (const AllocaInst *Alloca = dyn_cast<AllocaInst>(PtrV)) {
2088 if (Alloca->isSwiftError())
2093 // Get the value to be stored into a register. Use the zero register directly
2094 // when possible to avoid an unnecessary copy and a wasted register.
2095 unsigned SrcReg = 0;
2096 if (const auto *CI = dyn_cast<ConstantInt>(Op0)) {
2098 SrcReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
2099 } else if (const auto *CF = dyn_cast<ConstantFP>(Op0)) {
2100 if (CF->isZero() && !CF->isNegative()) {
2101 VT = MVT::getIntegerVT(VT.getSizeInBits());
2102 SrcReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
2107 SrcReg = getRegForValue(Op0);
2112 // See if we can handle this address.
2114 if (!computeAddress(I->getOperand(1), Addr, I->getOperand(0)->getType()))
2117 if (!emitStore(VT, SrcReg, Addr, createMachineMemOperandFor(I)))
2122 static AArch64CC::CondCode getCompareCC(CmpInst::Predicate Pred) {
2124 case CmpInst::FCMP_ONE:
2125 case CmpInst::FCMP_UEQ:
2127 // AL is our "false" for now. The other two need more compares.
2128 return AArch64CC::AL;
2129 case CmpInst::ICMP_EQ:
2130 case CmpInst::FCMP_OEQ:
2131 return AArch64CC::EQ;
2132 case CmpInst::ICMP_SGT:
2133 case CmpInst::FCMP_OGT:
2134 return AArch64CC::GT;
2135 case CmpInst::ICMP_SGE:
2136 case CmpInst::FCMP_OGE:
2137 return AArch64CC::GE;
2138 case CmpInst::ICMP_UGT:
2139 case CmpInst::FCMP_UGT:
2140 return AArch64CC::HI;
2141 case CmpInst::FCMP_OLT:
2142 return AArch64CC::MI;
2143 case CmpInst::ICMP_ULE:
2144 case CmpInst::FCMP_OLE:
2145 return AArch64CC::LS;
2146 case CmpInst::FCMP_ORD:
2147 return AArch64CC::VC;
2148 case CmpInst::FCMP_UNO:
2149 return AArch64CC::VS;
2150 case CmpInst::FCMP_UGE:
2151 return AArch64CC::PL;
2152 case CmpInst::ICMP_SLT:
2153 case CmpInst::FCMP_ULT:
2154 return AArch64CC::LT;
2155 case CmpInst::ICMP_SLE:
2156 case CmpInst::FCMP_ULE:
2157 return AArch64CC::LE;
2158 case CmpInst::FCMP_UNE:
2159 case CmpInst::ICMP_NE:
2160 return AArch64CC::NE;
2161 case CmpInst::ICMP_UGE:
2162 return AArch64CC::HS;
2163 case CmpInst::ICMP_ULT:
2164 return AArch64CC::LO;
2168 /// \brief Try to emit a combined compare-and-branch instruction.
2169 bool AArch64FastISel::emitCompareAndBranch(const BranchInst *BI) {
2170 assert(isa<CmpInst>(BI->getCondition()) && "Expected cmp instruction");
2171 const CmpInst *CI = cast<CmpInst>(BI->getCondition());
2172 CmpInst::Predicate Predicate = optimizeCmpPredicate(CI);
2174 const Value *LHS = CI->getOperand(0);
2175 const Value *RHS = CI->getOperand(1);
2178 if (!isTypeSupported(LHS->getType(), VT))
2181 unsigned BW = VT.getSizeInBits();
2185 MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
2186 MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
2188 // Try to take advantage of fallthrough opportunities.
2189 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
2190 std::swap(TBB, FBB);
2191 Predicate = CmpInst::getInversePredicate(Predicate);
2196 switch (Predicate) {
2199 case CmpInst::ICMP_EQ:
2200 case CmpInst::ICMP_NE:
2201 if (isa<Constant>(LHS) && cast<Constant>(LHS)->isNullValue())
2202 std::swap(LHS, RHS);
2204 if (!isa<Constant>(RHS) || !cast<Constant>(RHS)->isNullValue())
2207 if (const auto *AI = dyn_cast<BinaryOperator>(LHS))
2208 if (AI->getOpcode() == Instruction::And && isValueAvailable(AI)) {
2209 const Value *AndLHS = AI->getOperand(0);
2210 const Value *AndRHS = AI->getOperand(1);
2212 if (const auto *C = dyn_cast<ConstantInt>(AndLHS))
2213 if (C->getValue().isPowerOf2())
2214 std::swap(AndLHS, AndRHS);
2216 if (const auto *C = dyn_cast<ConstantInt>(AndRHS))
2217 if (C->getValue().isPowerOf2()) {
2218 TestBit = C->getValue().logBase2();
2226 IsCmpNE = Predicate == CmpInst::ICMP_NE;
2228 case CmpInst::ICMP_SLT:
2229 case CmpInst::ICMP_SGE:
2230 if (!isa<Constant>(RHS) || !cast<Constant>(RHS)->isNullValue())
2234 IsCmpNE = Predicate == CmpInst::ICMP_SLT;
2236 case CmpInst::ICMP_SGT:
2237 case CmpInst::ICMP_SLE:
2238 if (!isa<ConstantInt>(RHS))
2241 if (cast<ConstantInt>(RHS)->getValue() != APInt(BW, -1, true))
2245 IsCmpNE = Predicate == CmpInst::ICMP_SLE;
2249 static const unsigned OpcTable[2][2][2] = {
2250 { {AArch64::CBZW, AArch64::CBZX },
2251 {AArch64::CBNZW, AArch64::CBNZX} },
2252 { {AArch64::TBZW, AArch64::TBZX },
2253 {AArch64::TBNZW, AArch64::TBNZX} }
2256 bool IsBitTest = TestBit != -1;
2257 bool Is64Bit = BW == 64;
2258 if (TestBit < 32 && TestBit >= 0)
2261 unsigned Opc = OpcTable[IsBitTest][IsCmpNE][Is64Bit];
2262 const MCInstrDesc &II = TII.get(Opc);
2264 unsigned SrcReg = getRegForValue(LHS);
2267 bool SrcIsKill = hasTrivialKill(LHS);
2269 if (BW == 64 && !Is64Bit)
2270 SrcReg = fastEmitInst_extractsubreg(MVT::i32, SrcReg, SrcIsKill,
2273 if ((BW < 32) && !IsBitTest)
2274 SrcReg = emitIntExt(VT, SrcReg, MVT::i32, /*IsZExt=*/true);
2276 // Emit the combined compare and branch instruction.
2277 SrcReg = constrainOperandRegClass(II, SrcReg, II.getNumDefs());
2278 MachineInstrBuilder MIB =
2279 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc))
2280 .addReg(SrcReg, getKillRegState(SrcIsKill));
2282 MIB.addImm(TestBit);
2285 finishCondBranch(BI->getParent(), TBB, FBB);
2289 bool AArch64FastISel::selectBranch(const Instruction *I) {
2290 const BranchInst *BI = cast<BranchInst>(I);
2291 if (BI->isUnconditional()) {
2292 MachineBasicBlock *MSucc = FuncInfo.MBBMap[BI->getSuccessor(0)];
2293 fastEmitBranch(MSucc, BI->getDebugLoc());
2297 MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
2298 MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
2300 if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
2301 if (CI->hasOneUse() && isValueAvailable(CI)) {
2302 // Try to optimize or fold the cmp.
2303 CmpInst::Predicate Predicate = optimizeCmpPredicate(CI);
2304 switch (Predicate) {
2307 case CmpInst::FCMP_FALSE:
2308 fastEmitBranch(FBB, DbgLoc);
2310 case CmpInst::FCMP_TRUE:
2311 fastEmitBranch(TBB, DbgLoc);
2315 // Try to emit a combined compare-and-branch first.
2316 if (emitCompareAndBranch(BI))
2319 // Try to take advantage of fallthrough opportunities.
2320 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
2321 std::swap(TBB, FBB);
2322 Predicate = CmpInst::getInversePredicate(Predicate);
2326 if (!emitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
2329 // FCMP_UEQ and FCMP_ONE cannot be checked with a single branch
2331 AArch64CC::CondCode CC = getCompareCC(Predicate);
2332 AArch64CC::CondCode ExtraCC = AArch64CC::AL;
2333 switch (Predicate) {
2336 case CmpInst::FCMP_UEQ:
2337 ExtraCC = AArch64CC::EQ;
2340 case CmpInst::FCMP_ONE:
2341 ExtraCC = AArch64CC::MI;
2345 assert((CC != AArch64CC::AL) && "Unexpected condition code.");
2347 // Emit the extra branch for FCMP_UEQ and FCMP_ONE.
2348 if (ExtraCC != AArch64CC::AL) {
2349 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
2355 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
2359 finishCondBranch(BI->getParent(), TBB, FBB);
2362 } else if (const auto *CI = dyn_cast<ConstantInt>(BI->getCondition())) {
2363 uint64_t Imm = CI->getZExtValue();
2364 MachineBasicBlock *Target = (Imm == 0) ? FBB : TBB;
2365 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::B))
2368 // Obtain the branch probability and add the target to the successor list.
2370 auto BranchProbability = FuncInfo.BPI->getEdgeProbability(
2371 BI->getParent(), Target->getBasicBlock());
2372 FuncInfo.MBB->addSuccessor(Target, BranchProbability);
2374 FuncInfo.MBB->addSuccessorWithoutProb(Target);
2377 AArch64CC::CondCode CC = AArch64CC::NE;
2378 if (foldXALUIntrinsic(CC, I, BI->getCondition())) {
2379 // Fake request the condition, otherwise the intrinsic might be completely
2381 unsigned CondReg = getRegForValue(BI->getCondition());
2386 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
2390 finishCondBranch(BI->getParent(), TBB, FBB);
2395 unsigned CondReg = getRegForValue(BI->getCondition());
2398 bool CondRegIsKill = hasTrivialKill(BI->getCondition());
2400 // i1 conditions come as i32 values, test the lowest bit with tb(n)z.
2401 unsigned Opcode = AArch64::TBNZW;
2402 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
2403 std::swap(TBB, FBB);
2404 Opcode = AArch64::TBZW;
2407 const MCInstrDesc &II = TII.get(Opcode);
2408 unsigned ConstrainedCondReg
2409 = constrainOperandRegClass(II, CondReg, II.getNumDefs());
2410 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
2411 .addReg(ConstrainedCondReg, getKillRegState(CondRegIsKill))
2415 finishCondBranch(BI->getParent(), TBB, FBB);
2419 bool AArch64FastISel::selectIndirectBr(const Instruction *I) {
2420 const IndirectBrInst *BI = cast<IndirectBrInst>(I);
2421 unsigned AddrReg = getRegForValue(BI->getOperand(0));
2425 // Emit the indirect branch.
2426 const MCInstrDesc &II = TII.get(AArch64::BR);
2427 AddrReg = constrainOperandRegClass(II, AddrReg, II.getNumDefs());
2428 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II).addReg(AddrReg);
2430 // Make sure the CFG is up-to-date.
2431 for (auto *Succ : BI->successors())
2432 FuncInfo.MBB->addSuccessor(FuncInfo.MBBMap[Succ]);
2437 bool AArch64FastISel::selectCmp(const Instruction *I) {
2438 const CmpInst *CI = cast<CmpInst>(I);
2440 // Vectors of i1 are weird: bail out.
2441 if (CI->getType()->isVectorTy())
2444 // Try to optimize or fold the cmp.
2445 CmpInst::Predicate Predicate = optimizeCmpPredicate(CI);
2446 unsigned ResultReg = 0;
2447 switch (Predicate) {
2450 case CmpInst::FCMP_FALSE:
2451 ResultReg = createResultReg(&AArch64::GPR32RegClass);
2452 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2453 TII.get(TargetOpcode::COPY), ResultReg)
2454 .addReg(AArch64::WZR, getKillRegState(true));
2456 case CmpInst::FCMP_TRUE:
2457 ResultReg = fastEmit_i(MVT::i32, MVT::i32, ISD::Constant, 1);
2462 updateValueMap(I, ResultReg);
2467 if (!emitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
2470 ResultReg = createResultReg(&AArch64::GPR32RegClass);
2472 // FCMP_UEQ and FCMP_ONE cannot be checked with a single instruction. These
2473 // condition codes are inverted, because they are used by CSINC.
2474 static unsigned CondCodeTable[2][2] = {
2475 { AArch64CC::NE, AArch64CC::VC },
2476 { AArch64CC::PL, AArch64CC::LE }
2478 unsigned *CondCodes = nullptr;
2479 switch (Predicate) {
2482 case CmpInst::FCMP_UEQ:
2483 CondCodes = &CondCodeTable[0][0];
2485 case CmpInst::FCMP_ONE:
2486 CondCodes = &CondCodeTable[1][0];
2491 unsigned TmpReg1 = createResultReg(&AArch64::GPR32RegClass);
2492 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::CSINCWr),
2494 .addReg(AArch64::WZR, getKillRegState(true))
2495 .addReg(AArch64::WZR, getKillRegState(true))
2496 .addImm(CondCodes[0]);
2497 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::CSINCWr),
2499 .addReg(TmpReg1, getKillRegState(true))
2500 .addReg(AArch64::WZR, getKillRegState(true))
2501 .addImm(CondCodes[1]);
2503 updateValueMap(I, ResultReg);
2507 // Now set a register based on the comparison.
2508 AArch64CC::CondCode CC = getCompareCC(Predicate);
2509 assert((CC != AArch64CC::AL) && "Unexpected condition code.");
2510 AArch64CC::CondCode invertedCC = getInvertedCondCode(CC);
2511 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::CSINCWr),
2513 .addReg(AArch64::WZR, getKillRegState(true))
2514 .addReg(AArch64::WZR, getKillRegState(true))
2515 .addImm(invertedCC);
2517 updateValueMap(I, ResultReg);
2521 /// \brief Optimize selects of i1 if one of the operands has a 'true' or 'false'
2523 bool AArch64FastISel::optimizeSelect(const SelectInst *SI) {
2524 if (!SI->getType()->isIntegerTy(1))
2527 const Value *Src1Val, *Src2Val;
2529 bool NeedExtraOp = false;
2530 if (auto *CI = dyn_cast<ConstantInt>(SI->getTrueValue())) {
2532 Src1Val = SI->getCondition();
2533 Src2Val = SI->getFalseValue();
2534 Opc = AArch64::ORRWrr;
2536 assert(CI->isZero());
2537 Src1Val = SI->getFalseValue();
2538 Src2Val = SI->getCondition();
2539 Opc = AArch64::BICWrr;
2541 } else if (auto *CI = dyn_cast<ConstantInt>(SI->getFalseValue())) {
2543 Src1Val = SI->getCondition();
2544 Src2Val = SI->getTrueValue();
2545 Opc = AArch64::ORRWrr;
2548 assert(CI->isZero());
2549 Src1Val = SI->getCondition();
2550 Src2Val = SI->getTrueValue();
2551 Opc = AArch64::ANDWrr;
2558 unsigned Src1Reg = getRegForValue(Src1Val);
2561 bool Src1IsKill = hasTrivialKill(Src1Val);
2563 unsigned Src2Reg = getRegForValue(Src2Val);
2566 bool Src2IsKill = hasTrivialKill(Src2Val);
2569 Src1Reg = emitLogicalOp_ri(ISD::XOR, MVT::i32, Src1Reg, Src1IsKill, 1);
2572 unsigned ResultReg = fastEmitInst_rr(Opc, &AArch64::GPR32RegClass, Src1Reg,
2573 Src1IsKill, Src2Reg, Src2IsKill);
2574 updateValueMap(SI, ResultReg);
2578 bool AArch64FastISel::selectSelect(const Instruction *I) {
2579 assert(isa<SelectInst>(I) && "Expected a select instruction.");
2581 if (!isTypeSupported(I->getType(), VT))
2585 const TargetRegisterClass *RC;
2586 switch (VT.SimpleTy) {
2593 Opc = AArch64::CSELWr;
2594 RC = &AArch64::GPR32RegClass;
2597 Opc = AArch64::CSELXr;
2598 RC = &AArch64::GPR64RegClass;
2601 Opc = AArch64::FCSELSrrr;
2602 RC = &AArch64::FPR32RegClass;
2605 Opc = AArch64::FCSELDrrr;
2606 RC = &AArch64::FPR64RegClass;
2610 const SelectInst *SI = cast<SelectInst>(I);
2611 const Value *Cond = SI->getCondition();
2612 AArch64CC::CondCode CC = AArch64CC::NE;
2613 AArch64CC::CondCode ExtraCC = AArch64CC::AL;
2615 if (optimizeSelect(SI))
2618 // Try to pickup the flags, so we don't have to emit another compare.
2619 if (foldXALUIntrinsic(CC, I, Cond)) {
2620 // Fake request the condition to force emission of the XALU intrinsic.
2621 unsigned CondReg = getRegForValue(Cond);
2624 } else if (isa<CmpInst>(Cond) && cast<CmpInst>(Cond)->hasOneUse() &&
2625 isValueAvailable(Cond)) {
2626 const auto *Cmp = cast<CmpInst>(Cond);
2627 // Try to optimize or fold the cmp.
2628 CmpInst::Predicate Predicate = optimizeCmpPredicate(Cmp);
2629 const Value *FoldSelect = nullptr;
2630 switch (Predicate) {
2633 case CmpInst::FCMP_FALSE:
2634 FoldSelect = SI->getFalseValue();
2636 case CmpInst::FCMP_TRUE:
2637 FoldSelect = SI->getTrueValue();
2642 unsigned SrcReg = getRegForValue(FoldSelect);
2645 unsigned UseReg = lookUpRegForValue(SI);
2647 MRI.clearKillFlags(UseReg);
2649 updateValueMap(I, SrcReg);
2654 if (!emitCmp(Cmp->getOperand(0), Cmp->getOperand(1), Cmp->isUnsigned()))
2657 // FCMP_UEQ and FCMP_ONE cannot be checked with a single select instruction.
2658 CC = getCompareCC(Predicate);
2659 switch (Predicate) {
2662 case CmpInst::FCMP_UEQ:
2663 ExtraCC = AArch64CC::EQ;
2666 case CmpInst::FCMP_ONE:
2667 ExtraCC = AArch64CC::MI;
2671 assert((CC != AArch64CC::AL) && "Unexpected condition code.");
2673 unsigned CondReg = getRegForValue(Cond);
2676 bool CondIsKill = hasTrivialKill(Cond);
2678 const MCInstrDesc &II = TII.get(AArch64::ANDSWri);
2679 CondReg = constrainOperandRegClass(II, CondReg, 1);
2681 // Emit a TST instruction (ANDS wzr, reg, #imm).
2682 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II,
2684 .addReg(CondReg, getKillRegState(CondIsKill))
2685 .addImm(AArch64_AM::encodeLogicalImmediate(1, 32));
2688 unsigned Src1Reg = getRegForValue(SI->getTrueValue());
2689 bool Src1IsKill = hasTrivialKill(SI->getTrueValue());
2691 unsigned Src2Reg = getRegForValue(SI->getFalseValue());
2692 bool Src2IsKill = hasTrivialKill(SI->getFalseValue());
2694 if (!Src1Reg || !Src2Reg)
2697 if (ExtraCC != AArch64CC::AL) {
2698 Src2Reg = fastEmitInst_rri(Opc, RC, Src1Reg, Src1IsKill, Src2Reg,
2699 Src2IsKill, ExtraCC);
2702 unsigned ResultReg = fastEmitInst_rri(Opc, RC, Src1Reg, Src1IsKill, Src2Reg,
2704 updateValueMap(I, ResultReg);
2708 bool AArch64FastISel::selectFPExt(const Instruction *I) {
2709 Value *V = I->getOperand(0);
2710 if (!I->getType()->isDoubleTy() || !V->getType()->isFloatTy())
2713 unsigned Op = getRegForValue(V);
2717 unsigned ResultReg = createResultReg(&AArch64::FPR64RegClass);
2718 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::FCVTDSr),
2719 ResultReg).addReg(Op);
2720 updateValueMap(I, ResultReg);
2724 bool AArch64FastISel::selectFPTrunc(const Instruction *I) {
2725 Value *V = I->getOperand(0);
2726 if (!I->getType()->isFloatTy() || !V->getType()->isDoubleTy())
2729 unsigned Op = getRegForValue(V);
2733 unsigned ResultReg = createResultReg(&AArch64::FPR32RegClass);
2734 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::FCVTSDr),
2735 ResultReg).addReg(Op);
2736 updateValueMap(I, ResultReg);
2740 // FPToUI and FPToSI
2741 bool AArch64FastISel::selectFPToInt(const Instruction *I, bool Signed) {
2743 if (!isTypeLegal(I->getType(), DestVT) || DestVT.isVector())
2746 unsigned SrcReg = getRegForValue(I->getOperand(0));
2750 EVT SrcVT = TLI.getValueType(DL, I->getOperand(0)->getType(), true);
2751 if (SrcVT == MVT::f128)
2755 if (SrcVT == MVT::f64) {
2757 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZSUWDr : AArch64::FCVTZSUXDr;
2759 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZUUWDr : AArch64::FCVTZUUXDr;
2762 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZSUWSr : AArch64::FCVTZSUXSr;
2764 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZUUWSr : AArch64::FCVTZUUXSr;
2766 unsigned ResultReg = createResultReg(
2767 DestVT == MVT::i32 ? &AArch64::GPR32RegClass : &AArch64::GPR64RegClass);
2768 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
2770 updateValueMap(I, ResultReg);
2774 bool AArch64FastISel::selectIntToFP(const Instruction *I, bool Signed) {
2776 if (!isTypeLegal(I->getType(), DestVT) || DestVT.isVector())
2778 assert ((DestVT == MVT::f32 || DestVT == MVT::f64) &&
2779 "Unexpected value type.");
2781 unsigned SrcReg = getRegForValue(I->getOperand(0));
2784 bool SrcIsKill = hasTrivialKill(I->getOperand(0));
2786 EVT SrcVT = TLI.getValueType(DL, I->getOperand(0)->getType(), true);
2788 // Handle sign-extension.
2789 if (SrcVT == MVT::i16 || SrcVT == MVT::i8 || SrcVT == MVT::i1) {
2791 emitIntExt(SrcVT.getSimpleVT(), SrcReg, MVT::i32, /*isZExt*/ !Signed);
2798 if (SrcVT == MVT::i64) {
2800 Opc = (DestVT == MVT::f32) ? AArch64::SCVTFUXSri : AArch64::SCVTFUXDri;
2802 Opc = (DestVT == MVT::f32) ? AArch64::UCVTFUXSri : AArch64::UCVTFUXDri;
2805 Opc = (DestVT == MVT::f32) ? AArch64::SCVTFUWSri : AArch64::SCVTFUWDri;
2807 Opc = (DestVT == MVT::f32) ? AArch64::UCVTFUWSri : AArch64::UCVTFUWDri;
2810 unsigned ResultReg = fastEmitInst_r(Opc, TLI.getRegClassFor(DestVT), SrcReg,
2812 updateValueMap(I, ResultReg);
2816 bool AArch64FastISel::fastLowerArguments() {
2817 if (!FuncInfo.CanLowerReturn)
2820 const Function *F = FuncInfo.Fn;
2824 CallingConv::ID CC = F->getCallingConv();
2825 if (CC != CallingConv::C)
2828 // Only handle simple cases of up to 8 GPR and FPR each.
2829 unsigned GPRCnt = 0;
2830 unsigned FPRCnt = 0;
2832 for (auto const &Arg : F->args()) {
2833 // The first argument is at index 1.
2835 if (F->getAttributes().hasAttribute(Idx, Attribute::ByVal) ||
2836 F->getAttributes().hasAttribute(Idx, Attribute::InReg) ||
2837 F->getAttributes().hasAttribute(Idx, Attribute::StructRet) ||
2838 F->getAttributes().hasAttribute(Idx, Attribute::SwiftSelf) ||
2839 F->getAttributes().hasAttribute(Idx, Attribute::SwiftError) ||
2840 F->getAttributes().hasAttribute(Idx, Attribute::Nest))
2843 Type *ArgTy = Arg.getType();
2844 if (ArgTy->isStructTy() || ArgTy->isArrayTy())
2847 EVT ArgVT = TLI.getValueType(DL, ArgTy);
2848 if (!ArgVT.isSimple())
2851 MVT VT = ArgVT.getSimpleVT().SimpleTy;
2852 if (VT.isFloatingPoint() && !Subtarget->hasFPARMv8())
2855 if (VT.isVector() &&
2856 (!Subtarget->hasNEON() || !Subtarget->isLittleEndian()))
2859 if (VT >= MVT::i1 && VT <= MVT::i64)
2861 else if ((VT >= MVT::f16 && VT <= MVT::f64) || VT.is64BitVector() ||
2862 VT.is128BitVector())
2867 if (GPRCnt > 8 || FPRCnt > 8)
2871 static const MCPhysReg Registers[6][8] = {
2872 { AArch64::W0, AArch64::W1, AArch64::W2, AArch64::W3, AArch64::W4,
2873 AArch64::W5, AArch64::W6, AArch64::W7 },
2874 { AArch64::X0, AArch64::X1, AArch64::X2, AArch64::X3, AArch64::X4,
2875 AArch64::X5, AArch64::X6, AArch64::X7 },
2876 { AArch64::H0, AArch64::H1, AArch64::H2, AArch64::H3, AArch64::H4,
2877 AArch64::H5, AArch64::H6, AArch64::H7 },
2878 { AArch64::S0, AArch64::S1, AArch64::S2, AArch64::S3, AArch64::S4,
2879 AArch64::S5, AArch64::S6, AArch64::S7 },
2880 { AArch64::D0, AArch64::D1, AArch64::D2, AArch64::D3, AArch64::D4,
2881 AArch64::D5, AArch64::D6, AArch64::D7 },
2882 { AArch64::Q0, AArch64::Q1, AArch64::Q2, AArch64::Q3, AArch64::Q4,
2883 AArch64::Q5, AArch64::Q6, AArch64::Q7 }
2886 unsigned GPRIdx = 0;
2887 unsigned FPRIdx = 0;
2888 for (auto const &Arg : F->args()) {
2889 MVT VT = TLI.getSimpleValueType(DL, Arg.getType());
2891 const TargetRegisterClass *RC;
2892 if (VT >= MVT::i1 && VT <= MVT::i32) {
2893 SrcReg = Registers[0][GPRIdx++];
2894 RC = &AArch64::GPR32RegClass;
2896 } else if (VT == MVT::i64) {
2897 SrcReg = Registers[1][GPRIdx++];
2898 RC = &AArch64::GPR64RegClass;
2899 } else if (VT == MVT::f16) {
2900 SrcReg = Registers[2][FPRIdx++];
2901 RC = &AArch64::FPR16RegClass;
2902 } else if (VT == MVT::f32) {
2903 SrcReg = Registers[3][FPRIdx++];
2904 RC = &AArch64::FPR32RegClass;
2905 } else if ((VT == MVT::f64) || VT.is64BitVector()) {
2906 SrcReg = Registers[4][FPRIdx++];
2907 RC = &AArch64::FPR64RegClass;
2908 } else if (VT.is128BitVector()) {
2909 SrcReg = Registers[5][FPRIdx++];
2910 RC = &AArch64::FPR128RegClass;
2912 llvm_unreachable("Unexpected value type.");
2914 unsigned DstReg = FuncInfo.MF->addLiveIn(SrcReg, RC);
2915 // FIXME: Unfortunately it's necessary to emit a copy from the livein copy.
2916 // Without this, EmitLiveInCopies may eliminate the livein if its only
2917 // use is a bitcast (which isn't turned into an instruction).
2918 unsigned ResultReg = createResultReg(RC);
2919 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2920 TII.get(TargetOpcode::COPY), ResultReg)
2921 .addReg(DstReg, getKillRegState(true));
2922 updateValueMap(&Arg, ResultReg);
2927 bool AArch64FastISel::processCallArgs(CallLoweringInfo &CLI,
2928 SmallVectorImpl<MVT> &OutVTs,
2929 unsigned &NumBytes) {
2930 CallingConv::ID CC = CLI.CallConv;
2931 SmallVector<CCValAssign, 16> ArgLocs;
2932 CCState CCInfo(CC, false, *FuncInfo.MF, ArgLocs, *Context);
2933 CCInfo.AnalyzeCallOperands(OutVTs, CLI.OutFlags, CCAssignFnForCall(CC));
2935 // Get a count of how many bytes are to be pushed on the stack.
2936 NumBytes = CCInfo.getNextStackOffset();
2938 // Issue CALLSEQ_START
2939 unsigned AdjStackDown = TII.getCallFrameSetupOpcode();
2940 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackDown))
2943 // Process the args.
2944 for (CCValAssign &VA : ArgLocs) {
2945 const Value *ArgVal = CLI.OutVals[VA.getValNo()];
2946 MVT ArgVT = OutVTs[VA.getValNo()];
2948 unsigned ArgReg = getRegForValue(ArgVal);
2952 // Handle arg promotion: SExt, ZExt, AExt.
2953 switch (VA.getLocInfo()) {
2954 case CCValAssign::Full:
2956 case CCValAssign::SExt: {
2957 MVT DestVT = VA.getLocVT();
2959 ArgReg = emitIntExt(SrcVT, ArgReg, DestVT, /*isZExt=*/false);
2964 case CCValAssign::AExt:
2965 // Intentional fall-through.
2966 case CCValAssign::ZExt: {
2967 MVT DestVT = VA.getLocVT();
2969 ArgReg = emitIntExt(SrcVT, ArgReg, DestVT, /*isZExt=*/true);
2975 llvm_unreachable("Unknown arg promotion!");
2978 // Now copy/store arg to correct locations.
2979 if (VA.isRegLoc() && !VA.needsCustom()) {
2980 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2981 TII.get(TargetOpcode::COPY), VA.getLocReg()).addReg(ArgReg);
2982 CLI.OutRegs.push_back(VA.getLocReg());
2983 } else if (VA.needsCustom()) {
2984 // FIXME: Handle custom args.
2987 assert(VA.isMemLoc() && "Assuming store on stack.");
2989 // Don't emit stores for undef values.
2990 if (isa<UndefValue>(ArgVal))
2993 // Need to store on the stack.
2994 unsigned ArgSize = (ArgVT.getSizeInBits() + 7) / 8;
2996 unsigned BEAlign = 0;
2997 if (ArgSize < 8 && !Subtarget->isLittleEndian())
2998 BEAlign = 8 - ArgSize;
3001 Addr.setKind(Address::RegBase);
3002 Addr.setReg(AArch64::SP);
3003 Addr.setOffset(VA.getLocMemOffset() + BEAlign);
3005 unsigned Alignment = DL.getABITypeAlignment(ArgVal->getType());
3006 MachineMemOperand *MMO = FuncInfo.MF->getMachineMemOperand(
3007 MachinePointerInfo::getStack(*FuncInfo.MF, Addr.getOffset()),
3008 MachineMemOperand::MOStore, ArgVT.getStoreSize(), Alignment);
3010 if (!emitStore(ArgVT, ArgReg, Addr, MMO))
3017 bool AArch64FastISel::finishCall(CallLoweringInfo &CLI, MVT RetVT,
3018 unsigned NumBytes) {
3019 CallingConv::ID CC = CLI.CallConv;
3021 // Issue CALLSEQ_END
3022 unsigned AdjStackUp = TII.getCallFrameDestroyOpcode();
3023 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackUp))
3024 .addImm(NumBytes).addImm(0);
3026 // Now the return value.
3027 if (RetVT != MVT::isVoid) {
3028 SmallVector<CCValAssign, 16> RVLocs;
3029 CCState CCInfo(CC, false, *FuncInfo.MF, RVLocs, *Context);
3030 CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC));
3032 // Only handle a single return value.
3033 if (RVLocs.size() != 1)
3036 // Copy all of the result registers out of their specified physreg.
3037 MVT CopyVT = RVLocs[0].getValVT();
3039 // TODO: Handle big-endian results
3040 if (CopyVT.isVector() && !Subtarget->isLittleEndian())
3043 unsigned ResultReg = createResultReg(TLI.getRegClassFor(CopyVT));
3044 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3045 TII.get(TargetOpcode::COPY), ResultReg)
3046 .addReg(RVLocs[0].getLocReg());
3047 CLI.InRegs.push_back(RVLocs[0].getLocReg());
3049 CLI.ResultReg = ResultReg;
3050 CLI.NumResultRegs = 1;
3056 bool AArch64FastISel::fastLowerCall(CallLoweringInfo &CLI) {
3057 CallingConv::ID CC = CLI.CallConv;
3058 bool IsTailCall = CLI.IsTailCall;
3059 bool IsVarArg = CLI.IsVarArg;
3060 const Value *Callee = CLI.Callee;
3061 MCSymbol *Symbol = CLI.Symbol;
3063 if (!Callee && !Symbol)
3066 // Allow SelectionDAG isel to handle tail calls.
3070 CodeModel::Model CM = TM.getCodeModel();
3071 // Only support the small and large code model.
3072 if (CM != CodeModel::Small && CM != CodeModel::Large)
3075 // FIXME: Add large code model support for ELF.
3076 if (CM == CodeModel::Large && !Subtarget->isTargetMachO())
3079 // Let SDISel handle vararg functions.
3083 // FIXME: Only handle *simple* calls for now.
3085 if (CLI.RetTy->isVoidTy())
3086 RetVT = MVT::isVoid;
3087 else if (!isTypeLegal(CLI.RetTy, RetVT))
3090 for (auto Flag : CLI.OutFlags)
3091 if (Flag.isInReg() || Flag.isSRet() || Flag.isNest() || Flag.isByVal() ||
3092 Flag.isSwiftSelf() || Flag.isSwiftError())
3095 // Set up the argument vectors.
3096 SmallVector<MVT, 16> OutVTs;
3097 OutVTs.reserve(CLI.OutVals.size());
3099 for (auto *Val : CLI.OutVals) {
3101 if (!isTypeLegal(Val->getType(), VT) &&
3102 !(VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16))
3105 // We don't handle vector parameters yet.
3106 if (VT.isVector() || VT.getSizeInBits() > 64)
3109 OutVTs.push_back(VT);
3113 if (Callee && !computeCallAddress(Callee, Addr))
3116 // Handle the arguments now that we've gotten them.
3118 if (!processCallArgs(CLI, OutVTs, NumBytes))
3122 MachineInstrBuilder MIB;
3123 if (CM == CodeModel::Small) {
3124 const MCInstrDesc &II = TII.get(Addr.getReg() ? AArch64::BLR : AArch64::BL);
3125 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II);
3127 MIB.addSym(Symbol, 0);
3128 else if (Addr.getGlobalValue())
3129 MIB.addGlobalAddress(Addr.getGlobalValue(), 0, 0);
3130 else if (Addr.getReg()) {
3131 unsigned Reg = constrainOperandRegClass(II, Addr.getReg(), 0);
3136 unsigned CallReg = 0;
3138 unsigned ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
3139 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
3141 .addSym(Symbol, AArch64II::MO_GOT | AArch64II::MO_PAGE);
3143 CallReg = createResultReg(&AArch64::GPR64RegClass);
3144 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3145 TII.get(AArch64::LDRXui), CallReg)
3148 AArch64II::MO_GOT | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
3149 } else if (Addr.getGlobalValue())
3150 CallReg = materializeGV(Addr.getGlobalValue());
3151 else if (Addr.getReg())
3152 CallReg = Addr.getReg();
3157 const MCInstrDesc &II = TII.get(AArch64::BLR);
3158 CallReg = constrainOperandRegClass(II, CallReg, 0);
3159 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II).addReg(CallReg);
3162 // Add implicit physical register uses to the call.
3163 for (auto Reg : CLI.OutRegs)
3164 MIB.addReg(Reg, RegState::Implicit);
3166 // Add a register mask with the call-preserved registers.
3167 // Proper defs for return values will be added by setPhysRegsDeadExcept().
3168 MIB.addRegMask(TRI.getCallPreservedMask(*FuncInfo.MF, CC));
3172 // Finish off the call including any return values.
3173 return finishCall(CLI, RetVT, NumBytes);
3176 bool AArch64FastISel::isMemCpySmall(uint64_t Len, unsigned Alignment) {
3178 return Len / Alignment <= 4;
3183 bool AArch64FastISel::tryEmitSmallMemCpy(Address Dest, Address Src,
3184 uint64_t Len, unsigned Alignment) {
3185 // Make sure we don't bloat code by inlining very large memcpy's.
3186 if (!isMemCpySmall(Len, Alignment))
3189 int64_t UnscaledOffset = 0;
3190 Address OrigDest = Dest;
3191 Address OrigSrc = Src;
3195 if (!Alignment || Alignment >= 8) {
3206 // Bound based on alignment.
3207 if (Len >= 4 && Alignment == 4)
3209 else if (Len >= 2 && Alignment == 2)
3216 unsigned ResultReg = emitLoad(VT, VT, Src);
3220 if (!emitStore(VT, ResultReg, Dest))
3223 int64_t Size = VT.getSizeInBits() / 8;
3225 UnscaledOffset += Size;
3227 // We need to recompute the unscaled offset for each iteration.
3228 Dest.setOffset(OrigDest.getOffset() + UnscaledOffset);
3229 Src.setOffset(OrigSrc.getOffset() + UnscaledOffset);
3235 /// \brief Check if it is possible to fold the condition from the XALU intrinsic
3236 /// into the user. The condition code will only be updated on success.
3237 bool AArch64FastISel::foldXALUIntrinsic(AArch64CC::CondCode &CC,
3238 const Instruction *I,
3239 const Value *Cond) {
3240 if (!isa<ExtractValueInst>(Cond))
3243 const auto *EV = cast<ExtractValueInst>(Cond);
3244 if (!isa<IntrinsicInst>(EV->getAggregateOperand()))
3247 const auto *II = cast<IntrinsicInst>(EV->getAggregateOperand());
3249 const Function *Callee = II->getCalledFunction();
3251 cast<StructType>(Callee->getReturnType())->getTypeAtIndex(0U);
3252 if (!isTypeLegal(RetTy, RetVT))
3255 if (RetVT != MVT::i32 && RetVT != MVT::i64)
3258 const Value *LHS = II->getArgOperand(0);
3259 const Value *RHS = II->getArgOperand(1);
3261 // Canonicalize immediate to the RHS.
3262 if (isa<ConstantInt>(LHS) && !isa<ConstantInt>(RHS) &&
3263 isCommutativeIntrinsic(II))
3264 std::swap(LHS, RHS);
3266 // Simplify multiplies.
3267 Intrinsic::ID IID = II->getIntrinsicID();
3271 case Intrinsic::smul_with_overflow:
3272 if (const auto *C = dyn_cast<ConstantInt>(RHS))
3273 if (C->getValue() == 2)
3274 IID = Intrinsic::sadd_with_overflow;
3276 case Intrinsic::umul_with_overflow:
3277 if (const auto *C = dyn_cast<ConstantInt>(RHS))
3278 if (C->getValue() == 2)
3279 IID = Intrinsic::uadd_with_overflow;
3283 AArch64CC::CondCode TmpCC;
3287 case Intrinsic::sadd_with_overflow:
3288 case Intrinsic::ssub_with_overflow:
3289 TmpCC = AArch64CC::VS;
3291 case Intrinsic::uadd_with_overflow:
3292 TmpCC = AArch64CC::HS;
3294 case Intrinsic::usub_with_overflow:
3295 TmpCC = AArch64CC::LO;
3297 case Intrinsic::smul_with_overflow:
3298 case Intrinsic::umul_with_overflow:
3299 TmpCC = AArch64CC::NE;
3303 // Check if both instructions are in the same basic block.
3304 if (!isValueAvailable(II))
3307 // Make sure nothing is in the way
3308 BasicBlock::const_iterator Start(I);
3309 BasicBlock::const_iterator End(II);
3310 for (auto Itr = std::prev(Start); Itr != End; --Itr) {
3311 // We only expect extractvalue instructions between the intrinsic and the
3312 // instruction to be selected.
3313 if (!isa<ExtractValueInst>(Itr))
3316 // Check that the extractvalue operand comes from the intrinsic.
3317 const auto *EVI = cast<ExtractValueInst>(Itr);
3318 if (EVI->getAggregateOperand() != II)
3326 bool AArch64FastISel::fastLowerIntrinsicCall(const IntrinsicInst *II) {
3327 // FIXME: Handle more intrinsics.
3328 switch (II->getIntrinsicID()) {
3329 default: return false;
3330 case Intrinsic::frameaddress: {
3331 MachineFrameInfo *MFI = FuncInfo.MF->getFrameInfo();
3332 MFI->setFrameAddressIsTaken(true);
3334 const AArch64RegisterInfo *RegInfo =
3335 static_cast<const AArch64RegisterInfo *>(Subtarget->getRegisterInfo());
3336 unsigned FramePtr = RegInfo->getFrameRegister(*(FuncInfo.MF));
3337 unsigned SrcReg = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
3338 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3339 TII.get(TargetOpcode::COPY), SrcReg).addReg(FramePtr);
3340 // Recursively load frame address
3346 unsigned Depth = cast<ConstantInt>(II->getOperand(0))->getZExtValue();
3348 DestReg = fastEmitInst_ri(AArch64::LDRXui, &AArch64::GPR64RegClass,
3349 SrcReg, /*IsKill=*/true, 0);
3350 assert(DestReg && "Unexpected LDR instruction emission failure.");
3354 updateValueMap(II, SrcReg);
3357 case Intrinsic::memcpy:
3358 case Intrinsic::memmove: {
3359 const auto *MTI = cast<MemTransferInst>(II);
3360 // Don't handle volatile.
3361 if (MTI->isVolatile())
3364 // Disable inlining for memmove before calls to ComputeAddress. Otherwise,
3365 // we would emit dead code because we don't currently handle memmoves.
3366 bool IsMemCpy = (II->getIntrinsicID() == Intrinsic::memcpy);
3367 if (isa<ConstantInt>(MTI->getLength()) && IsMemCpy) {
3368 // Small memcpy's are common enough that we want to do them without a call
3370 uint64_t Len = cast<ConstantInt>(MTI->getLength())->getZExtValue();
3371 unsigned Alignment = MTI->getAlignment();
3372 if (isMemCpySmall(Len, Alignment)) {
3374 if (!computeAddress(MTI->getRawDest(), Dest) ||
3375 !computeAddress(MTI->getRawSource(), Src))
3377 if (tryEmitSmallMemCpy(Dest, Src, Len, Alignment))
3382 if (!MTI->getLength()->getType()->isIntegerTy(64))
3385 if (MTI->getSourceAddressSpace() > 255 || MTI->getDestAddressSpace() > 255)
3386 // Fast instruction selection doesn't support the special
3390 const char *IntrMemName = isa<MemCpyInst>(II) ? "memcpy" : "memmove";
3391 return lowerCallTo(II, IntrMemName, II->getNumArgOperands() - 2);
3393 case Intrinsic::memset: {
3394 const MemSetInst *MSI = cast<MemSetInst>(II);
3395 // Don't handle volatile.
3396 if (MSI->isVolatile())
3399 if (!MSI->getLength()->getType()->isIntegerTy(64))
3402 if (MSI->getDestAddressSpace() > 255)
3403 // Fast instruction selection doesn't support the special
3407 return lowerCallTo(II, "memset", II->getNumArgOperands() - 2);
3409 case Intrinsic::sin:
3410 case Intrinsic::cos:
3411 case Intrinsic::pow: {
3413 if (!isTypeLegal(II->getType(), RetVT))
3416 if (RetVT != MVT::f32 && RetVT != MVT::f64)
3419 static const RTLIB::Libcall LibCallTable[3][2] = {
3420 { RTLIB::SIN_F32, RTLIB::SIN_F64 },
3421 { RTLIB::COS_F32, RTLIB::COS_F64 },
3422 { RTLIB::POW_F32, RTLIB::POW_F64 }
3425 bool Is64Bit = RetVT == MVT::f64;
3426 switch (II->getIntrinsicID()) {
3428 llvm_unreachable("Unexpected intrinsic.");
3429 case Intrinsic::sin:
3430 LC = LibCallTable[0][Is64Bit];
3432 case Intrinsic::cos:
3433 LC = LibCallTable[1][Is64Bit];
3435 case Intrinsic::pow:
3436 LC = LibCallTable[2][Is64Bit];
3441 Args.reserve(II->getNumArgOperands());
3443 // Populate the argument list.
3444 for (auto &Arg : II->arg_operands()) {
3447 Entry.Ty = Arg->getType();
3448 Args.push_back(Entry);
3451 CallLoweringInfo CLI;
3452 MCContext &Ctx = MF->getContext();
3453 CLI.setCallee(DL, Ctx, TLI.getLibcallCallingConv(LC), II->getType(),
3454 TLI.getLibcallName(LC), std::move(Args));
3455 if (!lowerCallTo(CLI))
3457 updateValueMap(II, CLI.ResultReg);
3460 case Intrinsic::fabs: {
3462 if (!isTypeLegal(II->getType(), VT))
3466 switch (VT.SimpleTy) {
3470 Opc = AArch64::FABSSr;
3473 Opc = AArch64::FABSDr;
3476 unsigned SrcReg = getRegForValue(II->getOperand(0));
3479 bool SrcRegIsKill = hasTrivialKill(II->getOperand(0));
3480 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
3481 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
3482 .addReg(SrcReg, getKillRegState(SrcRegIsKill));
3483 updateValueMap(II, ResultReg);
3486 case Intrinsic::trap: {
3487 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::BRK))
3491 case Intrinsic::sqrt: {
3492 Type *RetTy = II->getCalledFunction()->getReturnType();
3495 if (!isTypeLegal(RetTy, VT))
3498 unsigned Op0Reg = getRegForValue(II->getOperand(0));
3501 bool Op0IsKill = hasTrivialKill(II->getOperand(0));
3503 unsigned ResultReg = fastEmit_r(VT, VT, ISD::FSQRT, Op0Reg, Op0IsKill);
3507 updateValueMap(II, ResultReg);
3510 case Intrinsic::sadd_with_overflow:
3511 case Intrinsic::uadd_with_overflow:
3512 case Intrinsic::ssub_with_overflow:
3513 case Intrinsic::usub_with_overflow:
3514 case Intrinsic::smul_with_overflow:
3515 case Intrinsic::umul_with_overflow: {
3516 // This implements the basic lowering of the xalu with overflow intrinsics.
3517 const Function *Callee = II->getCalledFunction();
3518 auto *Ty = cast<StructType>(Callee->getReturnType());
3519 Type *RetTy = Ty->getTypeAtIndex(0U);
3522 if (!isTypeLegal(RetTy, VT))
3525 if (VT != MVT::i32 && VT != MVT::i64)
3528 const Value *LHS = II->getArgOperand(0);
3529 const Value *RHS = II->getArgOperand(1);
3530 // Canonicalize immediate to the RHS.
3531 if (isa<ConstantInt>(LHS) && !isa<ConstantInt>(RHS) &&
3532 isCommutativeIntrinsic(II))
3533 std::swap(LHS, RHS);
3535 // Simplify multiplies.
3536 Intrinsic::ID IID = II->getIntrinsicID();
3540 case Intrinsic::smul_with_overflow:
3541 if (const auto *C = dyn_cast<ConstantInt>(RHS))
3542 if (C->getValue() == 2) {
3543 IID = Intrinsic::sadd_with_overflow;
3547 case Intrinsic::umul_with_overflow:
3548 if (const auto *C = dyn_cast<ConstantInt>(RHS))
3549 if (C->getValue() == 2) {
3550 IID = Intrinsic::uadd_with_overflow;
3556 unsigned ResultReg1 = 0, ResultReg2 = 0, MulReg = 0;
3557 AArch64CC::CondCode CC = AArch64CC::Invalid;
3559 default: llvm_unreachable("Unexpected intrinsic!");
3560 case Intrinsic::sadd_with_overflow:
3561 ResultReg1 = emitAdd(VT, LHS, RHS, /*SetFlags=*/true);
3564 case Intrinsic::uadd_with_overflow:
3565 ResultReg1 = emitAdd(VT, LHS, RHS, /*SetFlags=*/true);
3568 case Intrinsic::ssub_with_overflow:
3569 ResultReg1 = emitSub(VT, LHS, RHS, /*SetFlags=*/true);
3572 case Intrinsic::usub_with_overflow:
3573 ResultReg1 = emitSub(VT, LHS, RHS, /*SetFlags=*/true);
3576 case Intrinsic::smul_with_overflow: {
3578 unsigned LHSReg = getRegForValue(LHS);
3581 bool LHSIsKill = hasTrivialKill(LHS);
3583 unsigned RHSReg = getRegForValue(RHS);
3586 bool RHSIsKill = hasTrivialKill(RHS);
3588 if (VT == MVT::i32) {
3589 MulReg = emitSMULL_rr(MVT::i64, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
3590 unsigned ShiftReg = emitLSR_ri(MVT::i64, MVT::i64, MulReg,
3591 /*IsKill=*/false, 32);
3592 MulReg = fastEmitInst_extractsubreg(VT, MulReg, /*IsKill=*/true,
3594 ShiftReg = fastEmitInst_extractsubreg(VT, ShiftReg, /*IsKill=*/true,
3596 emitSubs_rs(VT, ShiftReg, /*IsKill=*/true, MulReg, /*IsKill=*/false,
3597 AArch64_AM::ASR, 31, /*WantResult=*/false);
3599 assert(VT == MVT::i64 && "Unexpected value type.");
3600 // LHSReg and RHSReg cannot be killed by this Mul, since they are
3601 // reused in the next instruction.
3602 MulReg = emitMul_rr(VT, LHSReg, /*IsKill=*/false, RHSReg,
3604 unsigned SMULHReg = fastEmit_rr(VT, VT, ISD::MULHS, LHSReg, LHSIsKill,
3606 emitSubs_rs(VT, SMULHReg, /*IsKill=*/true, MulReg, /*IsKill=*/false,
3607 AArch64_AM::ASR, 63, /*WantResult=*/false);
3611 case Intrinsic::umul_with_overflow: {
3613 unsigned LHSReg = getRegForValue(LHS);
3616 bool LHSIsKill = hasTrivialKill(LHS);
3618 unsigned RHSReg = getRegForValue(RHS);
3621 bool RHSIsKill = hasTrivialKill(RHS);
3623 if (VT == MVT::i32) {
3624 MulReg = emitUMULL_rr(MVT::i64, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
3625 emitSubs_rs(MVT::i64, AArch64::XZR, /*IsKill=*/true, MulReg,
3626 /*IsKill=*/false, AArch64_AM::LSR, 32,
3627 /*WantResult=*/false);
3628 MulReg = fastEmitInst_extractsubreg(VT, MulReg, /*IsKill=*/true,
3631 assert(VT == MVT::i64 && "Unexpected value type.");
3632 // LHSReg and RHSReg cannot be killed by this Mul, since they are
3633 // reused in the next instruction.
3634 MulReg = emitMul_rr(VT, LHSReg, /*IsKill=*/false, RHSReg,
3636 unsigned UMULHReg = fastEmit_rr(VT, VT, ISD::MULHU, LHSReg, LHSIsKill,
3638 emitSubs_rr(VT, AArch64::XZR, /*IsKill=*/true, UMULHReg,
3639 /*IsKill=*/false, /*WantResult=*/false);
3646 ResultReg1 = createResultReg(TLI.getRegClassFor(VT));
3647 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3648 TII.get(TargetOpcode::COPY), ResultReg1).addReg(MulReg);
3651 ResultReg2 = fastEmitInst_rri(AArch64::CSINCWr, &AArch64::GPR32RegClass,
3652 AArch64::WZR, /*IsKill=*/true, AArch64::WZR,
3653 /*IsKill=*/true, getInvertedCondCode(CC));
3655 assert((ResultReg1 + 1) == ResultReg2 &&
3656 "Nonconsecutive result registers.");
3657 updateValueMap(II, ResultReg1, 2);
3664 bool AArch64FastISel::selectRet(const Instruction *I) {
3665 const ReturnInst *Ret = cast<ReturnInst>(I);
3666 const Function &F = *I->getParent()->getParent();
3668 if (!FuncInfo.CanLowerReturn)
3674 if (TLI.supportSwiftError() &&
3675 F.getAttributes().hasAttrSomewhere(Attribute::SwiftError))
3678 if (TLI.supportSplitCSR(FuncInfo.MF))
3681 // Build a list of return value registers.
3682 SmallVector<unsigned, 4> RetRegs;
3684 if (Ret->getNumOperands() > 0) {
3685 CallingConv::ID CC = F.getCallingConv();
3686 SmallVector<ISD::OutputArg, 4> Outs;
3687 GetReturnInfo(F.getReturnType(), F.getAttributes(), Outs, TLI, DL);
3689 // Analyze operands of the call, assigning locations to each operand.
3690 SmallVector<CCValAssign, 16> ValLocs;
3691 CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, ValLocs, I->getContext());
3692 CCAssignFn *RetCC = CC == CallingConv::WebKit_JS ? RetCC_AArch64_WebKit_JS
3693 : RetCC_AArch64_AAPCS;
3694 CCInfo.AnalyzeReturn(Outs, RetCC);
3696 // Only handle a single return value for now.
3697 if (ValLocs.size() != 1)
3700 CCValAssign &VA = ValLocs[0];
3701 const Value *RV = Ret->getOperand(0);
3703 // Don't bother handling odd stuff for now.
3704 if ((VA.getLocInfo() != CCValAssign::Full) &&
3705 (VA.getLocInfo() != CCValAssign::BCvt))
3708 // Only handle register returns for now.
3712 unsigned Reg = getRegForValue(RV);
3716 unsigned SrcReg = Reg + VA.getValNo();
3717 unsigned DestReg = VA.getLocReg();
3718 // Avoid a cross-class copy. This is very unlikely.
3719 if (!MRI.getRegClass(SrcReg)->contains(DestReg))
3722 EVT RVEVT = TLI.getValueType(DL, RV->getType());
3723 if (!RVEVT.isSimple())
3726 // Vectors (of > 1 lane) in big endian need tricky handling.
3727 if (RVEVT.isVector() && RVEVT.getVectorNumElements() > 1 &&
3728 !Subtarget->isLittleEndian())
3731 MVT RVVT = RVEVT.getSimpleVT();
3732 if (RVVT == MVT::f128)
3735 MVT DestVT = VA.getValVT();
3736 // Special handling for extended integers.
3737 if (RVVT != DestVT) {
3738 if (RVVT != MVT::i1 && RVVT != MVT::i8 && RVVT != MVT::i16)
3741 if (!Outs[0].Flags.isZExt() && !Outs[0].Flags.isSExt())
3744 bool IsZExt = Outs[0].Flags.isZExt();
3745 SrcReg = emitIntExt(RVVT, SrcReg, DestVT, IsZExt);
3751 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3752 TII.get(TargetOpcode::COPY), DestReg).addReg(SrcReg);
3754 // Add register to return instruction.
3755 RetRegs.push_back(VA.getLocReg());
3758 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3759 TII.get(AArch64::RET_ReallyLR));
3760 for (unsigned RetReg : RetRegs)
3761 MIB.addReg(RetReg, RegState::Implicit);
3765 bool AArch64FastISel::selectTrunc(const Instruction *I) {
3766 Type *DestTy = I->getType();
3767 Value *Op = I->getOperand(0);
3768 Type *SrcTy = Op->getType();
3770 EVT SrcEVT = TLI.getValueType(DL, SrcTy, true);
3771 EVT DestEVT = TLI.getValueType(DL, DestTy, true);
3772 if (!SrcEVT.isSimple())
3774 if (!DestEVT.isSimple())
3777 MVT SrcVT = SrcEVT.getSimpleVT();
3778 MVT DestVT = DestEVT.getSimpleVT();
3780 if (SrcVT != MVT::i64 && SrcVT != MVT::i32 && SrcVT != MVT::i16 &&
3783 if (DestVT != MVT::i32 && DestVT != MVT::i16 && DestVT != MVT::i8 &&
3787 unsigned SrcReg = getRegForValue(Op);
3790 bool SrcIsKill = hasTrivialKill(Op);
3792 // If we're truncating from i64 to a smaller non-legal type then generate an
3793 // AND. Otherwise, we know the high bits are undefined and a truncate only
3794 // generate a COPY. We cannot mark the source register also as result
3795 // register, because this can incorrectly transfer the kill flag onto the
3798 if (SrcVT == MVT::i64) {
3800 switch (DestVT.SimpleTy) {
3802 // Trunc i64 to i32 is handled by the target-independent fast-isel.
3814 // Issue an extract_subreg to get the lower 32-bits.
3815 unsigned Reg32 = fastEmitInst_extractsubreg(MVT::i32, SrcReg, SrcIsKill,
3817 // Create the AND instruction which performs the actual truncation.
3818 ResultReg = emitAnd_ri(MVT::i32, Reg32, /*IsKill=*/true, Mask);
3819 assert(ResultReg && "Unexpected AND instruction emission failure.");
3821 ResultReg = createResultReg(&AArch64::GPR32RegClass);
3822 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3823 TII.get(TargetOpcode::COPY), ResultReg)
3824 .addReg(SrcReg, getKillRegState(SrcIsKill));
3827 updateValueMap(I, ResultReg);
3831 unsigned AArch64FastISel::emiti1Ext(unsigned SrcReg, MVT DestVT, bool IsZExt) {
3832 assert((DestVT == MVT::i8 || DestVT == MVT::i16 || DestVT == MVT::i32 ||
3833 DestVT == MVT::i64) &&
3834 "Unexpected value type.");
3835 // Handle i8 and i16 as i32.
3836 if (DestVT == MVT::i8 || DestVT == MVT::i16)
3840 unsigned ResultReg = emitAnd_ri(MVT::i32, SrcReg, /*TODO:IsKill=*/false, 1);
3841 assert(ResultReg && "Unexpected AND instruction emission failure.");
3842 if (DestVT == MVT::i64) {
3843 // We're ZExt i1 to i64. The ANDWri Wd, Ws, #1 implicitly clears the
3844 // upper 32 bits. Emit a SUBREG_TO_REG to extend from Wd to Xd.
3845 unsigned Reg64 = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
3846 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3847 TII.get(AArch64::SUBREG_TO_REG), Reg64)
3850 .addImm(AArch64::sub_32);
3855 if (DestVT == MVT::i64) {
3856 // FIXME: We're SExt i1 to i64.
3859 return fastEmitInst_rii(AArch64::SBFMWri, &AArch64::GPR32RegClass, SrcReg,
3860 /*TODO:IsKill=*/false, 0, 0);
3864 unsigned AArch64FastISel::emitMul_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
3865 unsigned Op1, bool Op1IsKill) {
3867 switch (RetVT.SimpleTy) {
3873 Opc = AArch64::MADDWrrr; ZReg = AArch64::WZR; break;
3875 Opc = AArch64::MADDXrrr; ZReg = AArch64::XZR; break;
3878 const TargetRegisterClass *RC =
3879 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3880 return fastEmitInst_rrr(Opc, RC, Op0, Op0IsKill, Op1, Op1IsKill,
3881 /*IsKill=*/ZReg, true);
3884 unsigned AArch64FastISel::emitSMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
3885 unsigned Op1, bool Op1IsKill) {
3886 if (RetVT != MVT::i64)
3889 return fastEmitInst_rrr(AArch64::SMADDLrrr, &AArch64::GPR64RegClass,
3890 Op0, Op0IsKill, Op1, Op1IsKill,
3891 AArch64::XZR, /*IsKill=*/true);
3894 unsigned AArch64FastISel::emitUMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
3895 unsigned Op1, bool Op1IsKill) {
3896 if (RetVT != MVT::i64)
3899 return fastEmitInst_rrr(AArch64::UMADDLrrr, &AArch64::GPR64RegClass,
3900 Op0, Op0IsKill, Op1, Op1IsKill,
3901 AArch64::XZR, /*IsKill=*/true);
3904 unsigned AArch64FastISel::emitLSL_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
3905 unsigned Op1Reg, bool Op1IsKill) {
3907 bool NeedTrunc = false;
3909 switch (RetVT.SimpleTy) {
3911 case MVT::i8: Opc = AArch64::LSLVWr; NeedTrunc = true; Mask = 0xff; break;
3912 case MVT::i16: Opc = AArch64::LSLVWr; NeedTrunc = true; Mask = 0xffff; break;
3913 case MVT::i32: Opc = AArch64::LSLVWr; break;
3914 case MVT::i64: Opc = AArch64::LSLVXr; break;
3917 const TargetRegisterClass *RC =
3918 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3920 Op1Reg = emitAnd_ri(MVT::i32, Op1Reg, Op1IsKill, Mask);
3923 unsigned ResultReg = fastEmitInst_rr(Opc, RC, Op0Reg, Op0IsKill, Op1Reg,
3926 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
3930 unsigned AArch64FastISel::emitLSL_ri(MVT RetVT, MVT SrcVT, unsigned Op0,
3931 bool Op0IsKill, uint64_t Shift,
3933 assert(RetVT.SimpleTy >= SrcVT.SimpleTy &&
3934 "Unexpected source/return type pair.");
3935 assert((SrcVT == MVT::i1 || SrcVT == MVT::i8 || SrcVT == MVT::i16 ||
3936 SrcVT == MVT::i32 || SrcVT == MVT::i64) &&
3937 "Unexpected source value type.");
3938 assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 ||
3939 RetVT == MVT::i64) && "Unexpected return value type.");
3941 bool Is64Bit = (RetVT == MVT::i64);
3942 unsigned RegSize = Is64Bit ? 64 : 32;
3943 unsigned DstBits = RetVT.getSizeInBits();
3944 unsigned SrcBits = SrcVT.getSizeInBits();
3945 const TargetRegisterClass *RC =
3946 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3948 // Just emit a copy for "zero" shifts.
3950 if (RetVT == SrcVT) {
3951 unsigned ResultReg = createResultReg(RC);
3952 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3953 TII.get(TargetOpcode::COPY), ResultReg)
3954 .addReg(Op0, getKillRegState(Op0IsKill));
3957 return emitIntExt(SrcVT, Op0, RetVT, IsZExt);
3960 // Don't deal with undefined shifts.
3961 if (Shift >= DstBits)
3964 // For immediate shifts we can fold the zero-/sign-extension into the shift.
3965 // {S|U}BFM Wd, Wn, #r, #s
3966 // Wd<32+s-r,32-r> = Wn<s:0> when r > s
3968 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3969 // %2 = shl i16 %1, 4
3970 // Wd<32+7-28,32-28> = Wn<7:0> <- clamp s to 7
3971 // 0b1111_1111_1111_1111__1111_1010_1010_0000 sext
3972 // 0b0000_0000_0000_0000__0000_0101_0101_0000 sext | zext
3973 // 0b0000_0000_0000_0000__0000_1010_1010_0000 zext
3975 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3976 // %2 = shl i16 %1, 8
3977 // Wd<32+7-24,32-24> = Wn<7:0>
3978 // 0b1111_1111_1111_1111__1010_1010_0000_0000 sext
3979 // 0b0000_0000_0000_0000__0101_0101_0000_0000 sext | zext
3980 // 0b0000_0000_0000_0000__1010_1010_0000_0000 zext
3982 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3983 // %2 = shl i16 %1, 12
3984 // Wd<32+3-20,32-20> = Wn<3:0>
3985 // 0b1111_1111_1111_1111__1010_0000_0000_0000 sext
3986 // 0b0000_0000_0000_0000__0101_0000_0000_0000 sext | zext
3987 // 0b0000_0000_0000_0000__1010_0000_0000_0000 zext
3989 unsigned ImmR = RegSize - Shift;
3990 // Limit the width to the length of the source type.
3991 unsigned ImmS = std::min<unsigned>(SrcBits - 1, DstBits - 1 - Shift);
3992 static const unsigned OpcTable[2][2] = {
3993 {AArch64::SBFMWri, AArch64::SBFMXri},
3994 {AArch64::UBFMWri, AArch64::UBFMXri}
3996 unsigned Opc = OpcTable[IsZExt][Is64Bit];
3997 if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) {
3998 unsigned TmpReg = MRI.createVirtualRegister(RC);
3999 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
4000 TII.get(AArch64::SUBREG_TO_REG), TmpReg)
4002 .addReg(Op0, getKillRegState(Op0IsKill))
4003 .addImm(AArch64::sub_32);
4007 return fastEmitInst_rii(Opc, RC, Op0, Op0IsKill, ImmR, ImmS);
4010 unsigned AArch64FastISel::emitLSR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
4011 unsigned Op1Reg, bool Op1IsKill) {
4013 bool NeedTrunc = false;
4015 switch (RetVT.SimpleTy) {
4017 case MVT::i8: Opc = AArch64::LSRVWr; NeedTrunc = true; Mask = 0xff; break;
4018 case MVT::i16: Opc = AArch64::LSRVWr; NeedTrunc = true; Mask = 0xffff; break;
4019 case MVT::i32: Opc = AArch64::LSRVWr; break;
4020 case MVT::i64: Opc = AArch64::LSRVXr; break;
4023 const TargetRegisterClass *RC =
4024 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
4026 Op0Reg = emitAnd_ri(MVT::i32, Op0Reg, Op0IsKill, Mask);
4027 Op1Reg = emitAnd_ri(MVT::i32, Op1Reg, Op1IsKill, Mask);
4028 Op0IsKill = Op1IsKill = true;
4030 unsigned ResultReg = fastEmitInst_rr(Opc, RC, Op0Reg, Op0IsKill, Op1Reg,
4033 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
4037 unsigned AArch64FastISel::emitLSR_ri(MVT RetVT, MVT SrcVT, unsigned Op0,
4038 bool Op0IsKill, uint64_t Shift,
4040 assert(RetVT.SimpleTy >= SrcVT.SimpleTy &&
4041 "Unexpected source/return type pair.");
4042 assert((SrcVT == MVT::i1 || SrcVT == MVT::i8 || SrcVT == MVT::i16 ||
4043 SrcVT == MVT::i32 || SrcVT == MVT::i64) &&
4044 "Unexpected source value type.");
4045 assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 ||
4046 RetVT == MVT::i64) && "Unexpected return value type.");
4048 bool Is64Bit = (RetVT == MVT::i64);
4049 unsigned RegSize = Is64Bit ? 64 : 32;
4050 unsigned DstBits = RetVT.getSizeInBits();
4051 unsigned SrcBits = SrcVT.getSizeInBits();
4052 const TargetRegisterClass *RC =
4053 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
4055 // Just emit a copy for "zero" shifts.
4057 if (RetVT == SrcVT) {
4058 unsigned ResultReg = createResultReg(RC);
4059 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
4060 TII.get(TargetOpcode::COPY), ResultReg)
4061 .addReg(Op0, getKillRegState(Op0IsKill));
4064 return emitIntExt(SrcVT, Op0, RetVT, IsZExt);
4067 // Don't deal with undefined shifts.
4068 if (Shift >= DstBits)
4071 // For immediate shifts we can fold the zero-/sign-extension into the shift.
4072 // {S|U}BFM Wd, Wn, #r, #s
4073 // Wd<s-r:0> = Wn<s:r> when r <= s
4075 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
4076 // %2 = lshr i16 %1, 4
4077 // Wd<7-4:0> = Wn<7:4>
4078 // 0b0000_0000_0000_0000__0000_1111_1111_1010 sext
4079 // 0b0000_0000_0000_0000__0000_0000_0000_0101 sext | zext
4080 // 0b0000_0000_0000_0000__0000_0000_0000_1010 zext
4082 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
4083 // %2 = lshr i16 %1, 8
4084 // Wd<7-7,0> = Wn<7:7>
4085 // 0b0000_0000_0000_0000__0000_0000_1111_1111 sext
4086 // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
4087 // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
4089 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
4090 // %2 = lshr i16 %1, 12
4091 // Wd<7-7,0> = Wn<7:7> <- clamp r to 7
4092 // 0b0000_0000_0000_0000__0000_0000_0000_1111 sext
4093 // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
4094 // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
4096 if (Shift >= SrcBits && IsZExt)
4097 return materializeInt(ConstantInt::get(*Context, APInt(RegSize, 0)), RetVT);
4099 // It is not possible to fold a sign-extend into the LShr instruction. In this
4100 // case emit a sign-extend.
4102 Op0 = emitIntExt(SrcVT, Op0, RetVT, IsZExt);
4107 SrcBits = SrcVT.getSizeInBits();
4111 unsigned ImmR = std::min<unsigned>(SrcBits - 1, Shift);
4112 unsigned ImmS = SrcBits - 1;
4113 static const unsigned OpcTable[2][2] = {
4114 {AArch64::SBFMWri, AArch64::SBFMXri},
4115 {AArch64::UBFMWri, AArch64::UBFMXri}
4117 unsigned Opc = OpcTable[IsZExt][Is64Bit];
4118 if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) {
4119 unsigned TmpReg = MRI.createVirtualRegister(RC);
4120 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
4121 TII.get(AArch64::SUBREG_TO_REG), TmpReg)
4123 .addReg(Op0, getKillRegState(Op0IsKill))
4124 .addImm(AArch64::sub_32);
4128 return fastEmitInst_rii(Opc, RC, Op0, Op0IsKill, ImmR, ImmS);
4131 unsigned AArch64FastISel::emitASR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
4132 unsigned Op1Reg, bool Op1IsKill) {
4134 bool NeedTrunc = false;
4136 switch (RetVT.SimpleTy) {
4138 case MVT::i8: Opc = AArch64::ASRVWr; NeedTrunc = true; Mask = 0xff; break;
4139 case MVT::i16: Opc = AArch64::ASRVWr; NeedTrunc = true; Mask = 0xffff; break;
4140 case MVT::i32: Opc = AArch64::ASRVWr; break;
4141 case MVT::i64: Opc = AArch64::ASRVXr; break;
4144 const TargetRegisterClass *RC =
4145 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
4147 Op0Reg = emitIntExt(RetVT, Op0Reg, MVT::i32, /*IsZExt=*/false);
4148 Op1Reg = emitAnd_ri(MVT::i32, Op1Reg, Op1IsKill, Mask);
4149 Op0IsKill = Op1IsKill = true;
4151 unsigned ResultReg = fastEmitInst_rr(Opc, RC, Op0Reg, Op0IsKill, Op1Reg,
4154 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
4158 unsigned AArch64FastISel::emitASR_ri(MVT RetVT, MVT SrcVT, unsigned Op0,
4159 bool Op0IsKill, uint64_t Shift,
4161 assert(RetVT.SimpleTy >= SrcVT.SimpleTy &&
4162 "Unexpected source/return type pair.");
4163 assert((SrcVT == MVT::i1 || SrcVT == MVT::i8 || SrcVT == MVT::i16 ||
4164 SrcVT == MVT::i32 || SrcVT == MVT::i64) &&
4165 "Unexpected source value type.");
4166 assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 ||
4167 RetVT == MVT::i64) && "Unexpected return value type.");
4169 bool Is64Bit = (RetVT == MVT::i64);
4170 unsigned RegSize = Is64Bit ? 64 : 32;
4171 unsigned DstBits = RetVT.getSizeInBits();
4172 unsigned SrcBits = SrcVT.getSizeInBits();
4173 const TargetRegisterClass *RC =
4174 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
4176 // Just emit a copy for "zero" shifts.
4178 if (RetVT == SrcVT) {
4179 unsigned ResultReg = createResultReg(RC);
4180 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
4181 TII.get(TargetOpcode::COPY), ResultReg)
4182 .addReg(Op0, getKillRegState(Op0IsKill));
4185 return emitIntExt(SrcVT, Op0, RetVT, IsZExt);
4188 // Don't deal with undefined shifts.
4189 if (Shift >= DstBits)
4192 // For immediate shifts we can fold the zero-/sign-extension into the shift.
4193 // {S|U}BFM Wd, Wn, #r, #s
4194 // Wd<s-r:0> = Wn<s:r> when r <= s
4196 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
4197 // %2 = ashr i16 %1, 4
4198 // Wd<7-4:0> = Wn<7:4>
4199 // 0b1111_1111_1111_1111__1111_1111_1111_1010 sext
4200 // 0b0000_0000_0000_0000__0000_0000_0000_0101 sext | zext
4201 // 0b0000_0000_0000_0000__0000_0000_0000_1010 zext
4203 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
4204 // %2 = ashr i16 %1, 8
4205 // Wd<7-7,0> = Wn<7:7>
4206 // 0b1111_1111_1111_1111__1111_1111_1111_1111 sext
4207 // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
4208 // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
4210 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
4211 // %2 = ashr i16 %1, 12
4212 // Wd<7-7,0> = Wn<7:7> <- clamp r to 7
4213 // 0b1111_1111_1111_1111__1111_1111_1111_1111 sext
4214 // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
4215 // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
4217 if (Shift >= SrcBits && IsZExt)
4218 return materializeInt(ConstantInt::get(*Context, APInt(RegSize, 0)), RetVT);
4220 unsigned ImmR = std::min<unsigned>(SrcBits - 1, Shift);
4221 unsigned ImmS = SrcBits - 1;
4222 static const unsigned OpcTable[2][2] = {
4223 {AArch64::SBFMWri, AArch64::SBFMXri},
4224 {AArch64::UBFMWri, AArch64::UBFMXri}
4226 unsigned Opc = OpcTable[IsZExt][Is64Bit];
4227 if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) {
4228 unsigned TmpReg = MRI.createVirtualRegister(RC);
4229 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
4230 TII.get(AArch64::SUBREG_TO_REG), TmpReg)
4232 .addReg(Op0, getKillRegState(Op0IsKill))
4233 .addImm(AArch64::sub_32);
4237 return fastEmitInst_rii(Opc, RC, Op0, Op0IsKill, ImmR, ImmS);
4240 unsigned AArch64FastISel::emitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT,
4242 assert(DestVT != MVT::i1 && "ZeroExt/SignExt an i1?");
4244 // FastISel does not have plumbing to deal with extensions where the SrcVT or
4245 // DestVT are odd things, so test to make sure that they are both types we can
4246 // handle (i1/i8/i16/i32 for SrcVT and i8/i16/i32/i64 for DestVT), otherwise
4247 // bail out to SelectionDAG.
4248 if (((DestVT != MVT::i8) && (DestVT != MVT::i16) &&
4249 (DestVT != MVT::i32) && (DestVT != MVT::i64)) ||
4250 ((SrcVT != MVT::i1) && (SrcVT != MVT::i8) &&
4251 (SrcVT != MVT::i16) && (SrcVT != MVT::i32)))
4257 switch (SrcVT.SimpleTy) {
4261 return emiti1Ext(SrcReg, DestVT, IsZExt);
4263 if (DestVT == MVT::i64)
4264 Opc = IsZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
4266 Opc = IsZExt ? AArch64::UBFMWri : AArch64::SBFMWri;
4270 if (DestVT == MVT::i64)
4271 Opc = IsZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
4273 Opc = IsZExt ? AArch64::UBFMWri : AArch64::SBFMWri;
4277 assert(DestVT == MVT::i64 && "IntExt i32 to i32?!?");
4278 Opc = IsZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
4283 // Handle i8 and i16 as i32.
4284 if (DestVT == MVT::i8 || DestVT == MVT::i16)
4286 else if (DestVT == MVT::i64) {
4287 unsigned Src64 = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
4288 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
4289 TII.get(AArch64::SUBREG_TO_REG), Src64)
4292 .addImm(AArch64::sub_32);
4296 const TargetRegisterClass *RC =
4297 (DestVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
4298 return fastEmitInst_rii(Opc, RC, SrcReg, /*TODO:IsKill=*/false, 0, Imm);
4301 static bool isZExtLoad(const MachineInstr *LI) {
4302 switch (LI->getOpcode()) {
4305 case AArch64::LDURBBi:
4306 case AArch64::LDURHHi:
4307 case AArch64::LDURWi:
4308 case AArch64::LDRBBui:
4309 case AArch64::LDRHHui:
4310 case AArch64::LDRWui:
4311 case AArch64::LDRBBroX:
4312 case AArch64::LDRHHroX:
4313 case AArch64::LDRWroX:
4314 case AArch64::LDRBBroW:
4315 case AArch64::LDRHHroW:
4316 case AArch64::LDRWroW:
4321 static bool isSExtLoad(const MachineInstr *LI) {
4322 switch (LI->getOpcode()) {
4325 case AArch64::LDURSBWi:
4326 case AArch64::LDURSHWi:
4327 case AArch64::LDURSBXi:
4328 case AArch64::LDURSHXi:
4329 case AArch64::LDURSWi:
4330 case AArch64::LDRSBWui:
4331 case AArch64::LDRSHWui:
4332 case AArch64::LDRSBXui:
4333 case AArch64::LDRSHXui:
4334 case AArch64::LDRSWui:
4335 case AArch64::LDRSBWroX:
4336 case AArch64::LDRSHWroX:
4337 case AArch64::LDRSBXroX:
4338 case AArch64::LDRSHXroX:
4339 case AArch64::LDRSWroX:
4340 case AArch64::LDRSBWroW:
4341 case AArch64::LDRSHWroW:
4342 case AArch64::LDRSBXroW:
4343 case AArch64::LDRSHXroW:
4344 case AArch64::LDRSWroW:
4349 bool AArch64FastISel::optimizeIntExtLoad(const Instruction *I, MVT RetVT,
4351 const auto *LI = dyn_cast<LoadInst>(I->getOperand(0));
4352 if (!LI || !LI->hasOneUse())
4355 // Check if the load instruction has already been selected.
4356 unsigned Reg = lookUpRegForValue(LI);
4360 MachineInstr *MI = MRI.getUniqueVRegDef(Reg);
4364 // Check if the correct load instruction has been emitted - SelectionDAG might
4365 // have emitted a zero-extending load, but we need a sign-extending load.
4366 bool IsZExt = isa<ZExtInst>(I);
4367 const auto *LoadMI = MI;
4368 if (LoadMI->getOpcode() == TargetOpcode::COPY &&
4369 LoadMI->getOperand(1).getSubReg() == AArch64::sub_32) {
4370 unsigned LoadReg = MI->getOperand(1).getReg();
4371 LoadMI = MRI.getUniqueVRegDef(LoadReg);
4372 assert(LoadMI && "Expected valid instruction");
4374 if (!(IsZExt && isZExtLoad(LoadMI)) && !(!IsZExt && isSExtLoad(LoadMI)))
4377 // Nothing to be done.
4378 if (RetVT != MVT::i64 || SrcVT > MVT::i32) {
4379 updateValueMap(I, Reg);
4384 unsigned Reg64 = createResultReg(&AArch64::GPR64RegClass);
4385 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
4386 TII.get(AArch64::SUBREG_TO_REG), Reg64)
4388 .addReg(Reg, getKillRegState(true))
4389 .addImm(AArch64::sub_32);
4392 assert((MI->getOpcode() == TargetOpcode::COPY &&
4393 MI->getOperand(1).getSubReg() == AArch64::sub_32) &&
4394 "Expected copy instruction");
4395 Reg = MI->getOperand(1).getReg();
4396 MI->eraseFromParent();
4398 updateValueMap(I, Reg);
4402 bool AArch64FastISel::selectIntExt(const Instruction *I) {
4403 assert((isa<ZExtInst>(I) || isa<SExtInst>(I)) &&
4404 "Unexpected integer extend instruction.");
4407 if (!isTypeSupported(I->getType(), RetVT))
4410 if (!isTypeSupported(I->getOperand(0)->getType(), SrcVT))
4413 // Try to optimize already sign-/zero-extended values from load instructions.
4414 if (optimizeIntExtLoad(I, RetVT, SrcVT))
4417 unsigned SrcReg = getRegForValue(I->getOperand(0));
4420 bool SrcIsKill = hasTrivialKill(I->getOperand(0));
4422 // Try to optimize already sign-/zero-extended values from function arguments.
4423 bool IsZExt = isa<ZExtInst>(I);
4424 if (const auto *Arg = dyn_cast<Argument>(I->getOperand(0))) {
4425 if ((IsZExt && Arg->hasZExtAttr()) || (!IsZExt && Arg->hasSExtAttr())) {
4426 if (RetVT == MVT::i64 && SrcVT != MVT::i64) {
4427 unsigned ResultReg = createResultReg(&AArch64::GPR64RegClass);
4428 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
4429 TII.get(AArch64::SUBREG_TO_REG), ResultReg)
4431 .addReg(SrcReg, getKillRegState(SrcIsKill))
4432 .addImm(AArch64::sub_32);
4435 // Conservatively clear all kill flags from all uses, because we are
4436 // replacing a sign-/zero-extend instruction at IR level with a nop at MI
4437 // level. The result of the instruction at IR level might have been
4438 // trivially dead, which is now not longer true.
4439 unsigned UseReg = lookUpRegForValue(I);
4441 MRI.clearKillFlags(UseReg);
4443 updateValueMap(I, SrcReg);
4448 unsigned ResultReg = emitIntExt(SrcVT, SrcReg, RetVT, IsZExt);
4452 updateValueMap(I, ResultReg);
4456 bool AArch64FastISel::selectRem(const Instruction *I, unsigned ISDOpcode) {
4457 EVT DestEVT = TLI.getValueType(DL, I->getType(), true);
4458 if (!DestEVT.isSimple())
4461 MVT DestVT = DestEVT.getSimpleVT();
4462 if (DestVT != MVT::i64 && DestVT != MVT::i32)
4466 bool Is64bit = (DestVT == MVT::i64);
4467 switch (ISDOpcode) {
4471 DivOpc = Is64bit ? AArch64::SDIVXr : AArch64::SDIVWr;
4474 DivOpc = Is64bit ? AArch64::UDIVXr : AArch64::UDIVWr;
4477 unsigned MSubOpc = Is64bit ? AArch64::MSUBXrrr : AArch64::MSUBWrrr;
4478 unsigned Src0Reg = getRegForValue(I->getOperand(0));
4481 bool Src0IsKill = hasTrivialKill(I->getOperand(0));
4483 unsigned Src1Reg = getRegForValue(I->getOperand(1));
4486 bool Src1IsKill = hasTrivialKill(I->getOperand(1));
4488 const TargetRegisterClass *RC =
4489 (DestVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
4490 unsigned QuotReg = fastEmitInst_rr(DivOpc, RC, Src0Reg, /*IsKill=*/false,
4491 Src1Reg, /*IsKill=*/false);
4492 assert(QuotReg && "Unexpected DIV instruction emission failure.");
4493 // The remainder is computed as numerator - (quotient * denominator) using the
4494 // MSUB instruction.
4495 unsigned ResultReg = fastEmitInst_rrr(MSubOpc, RC, QuotReg, /*IsKill=*/true,
4496 Src1Reg, Src1IsKill, Src0Reg,
4498 updateValueMap(I, ResultReg);
4502 bool AArch64FastISel::selectMul(const Instruction *I) {
4504 if (!isTypeSupported(I->getType(), VT, /*IsVectorAllowed=*/true))
4508 return selectBinaryOp(I, ISD::MUL);
4510 const Value *Src0 = I->getOperand(0);
4511 const Value *Src1 = I->getOperand(1);
4512 if (const auto *C = dyn_cast<ConstantInt>(Src0))
4513 if (C->getValue().isPowerOf2())
4514 std::swap(Src0, Src1);
4516 // Try to simplify to a shift instruction.
4517 if (const auto *C = dyn_cast<ConstantInt>(Src1))
4518 if (C->getValue().isPowerOf2()) {
4519 uint64_t ShiftVal = C->getValue().logBase2();
4522 if (const auto *ZExt = dyn_cast<ZExtInst>(Src0)) {
4523 if (!isIntExtFree(ZExt)) {
4525 if (isValueAvailable(ZExt) && isTypeSupported(ZExt->getSrcTy(), VT)) {
4528 Src0 = ZExt->getOperand(0);
4531 } else if (const auto *SExt = dyn_cast<SExtInst>(Src0)) {
4532 if (!isIntExtFree(SExt)) {
4534 if (isValueAvailable(SExt) && isTypeSupported(SExt->getSrcTy(), VT)) {
4537 Src0 = SExt->getOperand(0);
4542 unsigned Src0Reg = getRegForValue(Src0);
4545 bool Src0IsKill = hasTrivialKill(Src0);
4547 unsigned ResultReg =
4548 emitLSL_ri(VT, SrcVT, Src0Reg, Src0IsKill, ShiftVal, IsZExt);
4551 updateValueMap(I, ResultReg);
4556 unsigned Src0Reg = getRegForValue(I->getOperand(0));
4559 bool Src0IsKill = hasTrivialKill(I->getOperand(0));
4561 unsigned Src1Reg = getRegForValue(I->getOperand(1));
4564 bool Src1IsKill = hasTrivialKill(I->getOperand(1));
4566 unsigned ResultReg = emitMul_rr(VT, Src0Reg, Src0IsKill, Src1Reg, Src1IsKill);
4571 updateValueMap(I, ResultReg);
4575 bool AArch64FastISel::selectShift(const Instruction *I) {
4577 if (!isTypeSupported(I->getType(), RetVT, /*IsVectorAllowed=*/true))
4580 if (RetVT.isVector())
4581 return selectOperator(I, I->getOpcode());
4583 if (const auto *C = dyn_cast<ConstantInt>(I->getOperand(1))) {
4584 unsigned ResultReg = 0;
4585 uint64_t ShiftVal = C->getZExtValue();
4587 bool IsZExt = I->getOpcode() != Instruction::AShr;
4588 const Value *Op0 = I->getOperand(0);
4589 if (const auto *ZExt = dyn_cast<ZExtInst>(Op0)) {
4590 if (!isIntExtFree(ZExt)) {
4592 if (isValueAvailable(ZExt) && isTypeSupported(ZExt->getSrcTy(), TmpVT)) {
4595 Op0 = ZExt->getOperand(0);
4598 } else if (const auto *SExt = dyn_cast<SExtInst>(Op0)) {
4599 if (!isIntExtFree(SExt)) {
4601 if (isValueAvailable(SExt) && isTypeSupported(SExt->getSrcTy(), TmpVT)) {
4604 Op0 = SExt->getOperand(0);
4609 unsigned Op0Reg = getRegForValue(Op0);
4612 bool Op0IsKill = hasTrivialKill(Op0);
4614 switch (I->getOpcode()) {
4615 default: llvm_unreachable("Unexpected instruction.");
4616 case Instruction::Shl:
4617 ResultReg = emitLSL_ri(RetVT, SrcVT, Op0Reg, Op0IsKill, ShiftVal, IsZExt);
4619 case Instruction::AShr:
4620 ResultReg = emitASR_ri(RetVT, SrcVT, Op0Reg, Op0IsKill, ShiftVal, IsZExt);
4622 case Instruction::LShr:
4623 ResultReg = emitLSR_ri(RetVT, SrcVT, Op0Reg, Op0IsKill, ShiftVal, IsZExt);
4629 updateValueMap(I, ResultReg);
4633 unsigned Op0Reg = getRegForValue(I->getOperand(0));
4636 bool Op0IsKill = hasTrivialKill(I->getOperand(0));
4638 unsigned Op1Reg = getRegForValue(I->getOperand(1));
4641 bool Op1IsKill = hasTrivialKill(I->getOperand(1));
4643 unsigned ResultReg = 0;
4644 switch (I->getOpcode()) {
4645 default: llvm_unreachable("Unexpected instruction.");
4646 case Instruction::Shl:
4647 ResultReg = emitLSL_rr(RetVT, Op0Reg, Op0IsKill, Op1Reg, Op1IsKill);
4649 case Instruction::AShr:
4650 ResultReg = emitASR_rr(RetVT, Op0Reg, Op0IsKill, Op1Reg, Op1IsKill);
4652 case Instruction::LShr:
4653 ResultReg = emitLSR_rr(RetVT, Op0Reg, Op0IsKill, Op1Reg, Op1IsKill);
4660 updateValueMap(I, ResultReg);
4664 bool AArch64FastISel::selectBitCast(const Instruction *I) {
4667 if (!isTypeLegal(I->getOperand(0)->getType(), SrcVT))
4669 if (!isTypeLegal(I->getType(), RetVT))
4673 if (RetVT == MVT::f32 && SrcVT == MVT::i32)
4674 Opc = AArch64::FMOVWSr;
4675 else if (RetVT == MVT::f64 && SrcVT == MVT::i64)
4676 Opc = AArch64::FMOVXDr;
4677 else if (RetVT == MVT::i32 && SrcVT == MVT::f32)
4678 Opc = AArch64::FMOVSWr;
4679 else if (RetVT == MVT::i64 && SrcVT == MVT::f64)
4680 Opc = AArch64::FMOVDXr;
4684 const TargetRegisterClass *RC = nullptr;
4685 switch (RetVT.SimpleTy) {
4686 default: llvm_unreachable("Unexpected value type.");
4687 case MVT::i32: RC = &AArch64::GPR32RegClass; break;
4688 case MVT::i64: RC = &AArch64::GPR64RegClass; break;
4689 case MVT::f32: RC = &AArch64::FPR32RegClass; break;
4690 case MVT::f64: RC = &AArch64::FPR64RegClass; break;
4692 unsigned Op0Reg = getRegForValue(I->getOperand(0));
4695 bool Op0IsKill = hasTrivialKill(I->getOperand(0));
4696 unsigned ResultReg = fastEmitInst_r(Opc, RC, Op0Reg, Op0IsKill);
4701 updateValueMap(I, ResultReg);
4705 bool AArch64FastISel::selectFRem(const Instruction *I) {
4707 if (!isTypeLegal(I->getType(), RetVT))
4711 switch (RetVT.SimpleTy) {
4715 LC = RTLIB::REM_F32;
4718 LC = RTLIB::REM_F64;
4723 Args.reserve(I->getNumOperands());
4725 // Populate the argument list.
4726 for (auto &Arg : I->operands()) {
4729 Entry.Ty = Arg->getType();
4730 Args.push_back(Entry);
4733 CallLoweringInfo CLI;
4734 MCContext &Ctx = MF->getContext();
4735 CLI.setCallee(DL, Ctx, TLI.getLibcallCallingConv(LC), I->getType(),
4736 TLI.getLibcallName(LC), std::move(Args));
4737 if (!lowerCallTo(CLI))
4739 updateValueMap(I, CLI.ResultReg);
4743 bool AArch64FastISel::selectSDiv(const Instruction *I) {
4745 if (!isTypeLegal(I->getType(), VT))
4748 if (!isa<ConstantInt>(I->getOperand(1)))
4749 return selectBinaryOp(I, ISD::SDIV);
4751 const APInt &C = cast<ConstantInt>(I->getOperand(1))->getValue();
4752 if ((VT != MVT::i32 && VT != MVT::i64) || !C ||
4753 !(C.isPowerOf2() || (-C).isPowerOf2()))
4754 return selectBinaryOp(I, ISD::SDIV);
4756 unsigned Lg2 = C.countTrailingZeros();
4757 unsigned Src0Reg = getRegForValue(I->getOperand(0));
4760 bool Src0IsKill = hasTrivialKill(I->getOperand(0));
4762 if (cast<BinaryOperator>(I)->isExact()) {
4763 unsigned ResultReg = emitASR_ri(VT, VT, Src0Reg, Src0IsKill, Lg2);
4766 updateValueMap(I, ResultReg);
4770 int64_t Pow2MinusOne = (1ULL << Lg2) - 1;
4771 unsigned AddReg = emitAdd_ri_(VT, Src0Reg, /*IsKill=*/false, Pow2MinusOne);
4775 // (Src0 < 0) ? Pow2 - 1 : 0;
4776 if (!emitICmp_ri(VT, Src0Reg, /*IsKill=*/false, 0))
4780 const TargetRegisterClass *RC;
4781 if (VT == MVT::i64) {
4782 SelectOpc = AArch64::CSELXr;
4783 RC = &AArch64::GPR64RegClass;
4785 SelectOpc = AArch64::CSELWr;
4786 RC = &AArch64::GPR32RegClass;
4788 unsigned SelectReg =
4789 fastEmitInst_rri(SelectOpc, RC, AddReg, /*IsKill=*/true, Src0Reg,
4790 Src0IsKill, AArch64CC::LT);
4794 // Divide by Pow2 --> ashr. If we're dividing by a negative value we must also
4795 // negate the result.
4796 unsigned ZeroReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
4799 ResultReg = emitAddSub_rs(/*UseAdd=*/false, VT, ZeroReg, /*IsKill=*/true,
4800 SelectReg, /*IsKill=*/true, AArch64_AM::ASR, Lg2);
4802 ResultReg = emitASR_ri(VT, VT, SelectReg, /*IsKill=*/true, Lg2);
4807 updateValueMap(I, ResultReg);
4811 /// This is mostly a copy of the existing FastISel getRegForGEPIndex code. We
4812 /// have to duplicate it for AArch64, because otherwise we would fail during the
4813 /// sign-extend emission.
4814 std::pair<unsigned, bool> AArch64FastISel::getRegForGEPIndex(const Value *Idx) {
4815 unsigned IdxN = getRegForValue(Idx);
4817 // Unhandled operand. Halt "fast" selection and bail.
4818 return std::pair<unsigned, bool>(0, false);
4820 bool IdxNIsKill = hasTrivialKill(Idx);
4822 // If the index is smaller or larger than intptr_t, truncate or extend it.
4823 MVT PtrVT = TLI.getPointerTy(DL);
4824 EVT IdxVT = EVT::getEVT(Idx->getType(), /*HandleUnknown=*/false);
4825 if (IdxVT.bitsLT(PtrVT)) {
4826 IdxN = emitIntExt(IdxVT.getSimpleVT(), IdxN, PtrVT, /*IsZExt=*/false);
4828 } else if (IdxVT.bitsGT(PtrVT))
4829 llvm_unreachable("AArch64 FastISel doesn't support types larger than i64");
4830 return std::pair<unsigned, bool>(IdxN, IdxNIsKill);
4833 /// This is mostly a copy of the existing FastISel GEP code, but we have to
4834 /// duplicate it for AArch64, because otherwise we would bail out even for
4835 /// simple cases. This is because the standard fastEmit functions don't cover
4836 /// MUL at all and ADD is lowered very inefficientily.
4837 bool AArch64FastISel::selectGetElementPtr(const Instruction *I) {
4838 unsigned N = getRegForValue(I->getOperand(0));
4841 bool NIsKill = hasTrivialKill(I->getOperand(0));
4843 // Keep a running tab of the total offset to coalesce multiple N = N + Offset
4844 // into a single N = N + TotalOffset.
4845 uint64_t TotalOffs = 0;
4846 MVT VT = TLI.getPointerTy(DL);
4847 for (gep_type_iterator GTI = gep_type_begin(I), E = gep_type_end(I);
4849 const Value *Idx = GTI.getOperand();
4850 if (auto *StTy = dyn_cast<StructType>(*GTI)) {
4851 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
4854 TotalOffs += DL.getStructLayout(StTy)->getElementOffset(Field);
4856 Type *Ty = GTI.getIndexedType();
4858 // If this is a constant subscript, handle it quickly.
4859 if (const auto *CI = dyn_cast<ConstantInt>(Idx)) {
4864 DL.getTypeAllocSize(Ty) * cast<ConstantInt>(CI)->getSExtValue();
4868 N = emitAdd_ri_(VT, N, NIsKill, TotalOffs);
4875 // N = N + Idx * ElementSize;
4876 uint64_t ElementSize = DL.getTypeAllocSize(Ty);
4877 std::pair<unsigned, bool> Pair = getRegForGEPIndex(Idx);
4878 unsigned IdxN = Pair.first;
4879 bool IdxNIsKill = Pair.second;
4883 if (ElementSize != 1) {
4884 unsigned C = fastEmit_i(VT, VT, ISD::Constant, ElementSize);
4887 IdxN = emitMul_rr(VT, IdxN, IdxNIsKill, C, true);
4892 N = fastEmit_rr(VT, VT, ISD::ADD, N, NIsKill, IdxN, IdxNIsKill);
4898 N = emitAdd_ri_(VT, N, NIsKill, TotalOffs);
4902 updateValueMap(I, N);
4906 bool AArch64FastISel::fastSelectInstruction(const Instruction *I) {
4907 switch (I->getOpcode()) {
4910 case Instruction::Add:
4911 case Instruction::Sub:
4912 return selectAddSub(I);
4913 case Instruction::Mul:
4914 return selectMul(I);
4915 case Instruction::SDiv:
4916 return selectSDiv(I);
4917 case Instruction::SRem:
4918 if (!selectBinaryOp(I, ISD::SREM))
4919 return selectRem(I, ISD::SREM);
4921 case Instruction::URem:
4922 if (!selectBinaryOp(I, ISD::UREM))
4923 return selectRem(I, ISD::UREM);
4925 case Instruction::Shl:
4926 case Instruction::LShr:
4927 case Instruction::AShr:
4928 return selectShift(I);
4929 case Instruction::And:
4930 case Instruction::Or:
4931 case Instruction::Xor:
4932 return selectLogicalOp(I);
4933 case Instruction::Br:
4934 return selectBranch(I);
4935 case Instruction::IndirectBr:
4936 return selectIndirectBr(I);
4937 case Instruction::BitCast:
4938 if (!FastISel::selectBitCast(I))
4939 return selectBitCast(I);
4941 case Instruction::FPToSI:
4942 if (!selectCast(I, ISD::FP_TO_SINT))
4943 return selectFPToInt(I, /*Signed=*/true);
4945 case Instruction::FPToUI:
4946 return selectFPToInt(I, /*Signed=*/false);
4947 case Instruction::ZExt:
4948 case Instruction::SExt:
4949 return selectIntExt(I);
4950 case Instruction::Trunc:
4951 if (!selectCast(I, ISD::TRUNCATE))
4952 return selectTrunc(I);
4954 case Instruction::FPExt:
4955 return selectFPExt(I);
4956 case Instruction::FPTrunc:
4957 return selectFPTrunc(I);
4958 case Instruction::SIToFP:
4959 if (!selectCast(I, ISD::SINT_TO_FP))
4960 return selectIntToFP(I, /*Signed=*/true);
4962 case Instruction::UIToFP:
4963 return selectIntToFP(I, /*Signed=*/false);
4964 case Instruction::Load:
4965 return selectLoad(I);
4966 case Instruction::Store:
4967 return selectStore(I);
4968 case Instruction::FCmp:
4969 case Instruction::ICmp:
4970 return selectCmp(I);
4971 case Instruction::Select:
4972 return selectSelect(I);
4973 case Instruction::Ret:
4974 return selectRet(I);
4975 case Instruction::FRem:
4976 return selectFRem(I);
4977 case Instruction::GetElementPtr:
4978 return selectGetElementPtr(I);
4981 // fall-back to target-independent instruction selection.
4982 return selectOperator(I, I->getOpcode());
4983 // Silence warnings.
4984 (void)&CC_AArch64_DarwinPCS_VarArg;
4988 llvm::FastISel *AArch64::createFastISel(FunctionLoweringInfo &FuncInfo,
4989 const TargetLibraryInfo *LibInfo) {
4990 return new AArch64FastISel(FuncInfo, LibInfo);