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/Support/CommandLine.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 unsigned Flags, unsigned ScaleFactor,
147 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) {
250 Subtarget = &TM.getSubtarget<AArch64Subtarget>();
251 Context = &FuncInfo.Fn->getContext();
254 bool fastSelectInstruction(const Instruction *I) override;
256 #include "AArch64GenFastISel.inc"
259 } // end anonymous namespace
261 #include "AArch64GenCallingConv.inc"
263 /// \brief Check if the sign-/zero-extend will be a noop.
264 static bool isIntExtFree(const Instruction *I) {
265 assert((isa<ZExtInst>(I) || isa<SExtInst>(I)) &&
266 "Unexpected integer extend instruction.");
267 assert(!I->getType()->isVectorTy() && I->getType()->isIntegerTy() &&
268 "Unexpected value type.");
269 bool IsZExt = isa<ZExtInst>(I);
271 if (const auto *LI = dyn_cast<LoadInst>(I->getOperand(0)))
275 if (const auto *Arg = dyn_cast<Argument>(I->getOperand(0)))
276 if ((IsZExt && Arg->hasZExtAttr()) || (!IsZExt && Arg->hasSExtAttr()))
282 /// \brief Determine the implicit scale factor that is applied by a memory
283 /// operation for a given value type.
284 static unsigned getImplicitScaleFactor(MVT VT) {
285 switch (VT.SimpleTy) {
288 case MVT::i1: // fall-through
293 case MVT::i32: // fall-through
296 case MVT::i64: // fall-through
302 CCAssignFn *AArch64FastISel::CCAssignFnForCall(CallingConv::ID CC) const {
303 if (CC == CallingConv::WebKit_JS)
304 return CC_AArch64_WebKit_JS;
305 return Subtarget->isTargetDarwin() ? CC_AArch64_DarwinPCS : CC_AArch64_AAPCS;
308 unsigned AArch64FastISel::fastMaterializeAlloca(const AllocaInst *AI) {
309 assert(TLI.getValueType(AI->getType(), true) == MVT::i64 &&
310 "Alloca should always return a pointer.");
312 // Don't handle dynamic allocas.
313 if (!FuncInfo.StaticAllocaMap.count(AI))
316 DenseMap<const AllocaInst *, int>::iterator SI =
317 FuncInfo.StaticAllocaMap.find(AI);
319 if (SI != FuncInfo.StaticAllocaMap.end()) {
320 unsigned ResultReg = createResultReg(&AArch64::GPR64spRegClass);
321 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri),
323 .addFrameIndex(SI->second)
332 unsigned AArch64FastISel::materializeInt(const ConstantInt *CI, MVT VT) {
337 return fastEmit_i(VT, VT, ISD::Constant, CI->getZExtValue());
339 // Create a copy from the zero register to materialize a "0" value.
340 const TargetRegisterClass *RC = (VT == MVT::i64) ? &AArch64::GPR64RegClass
341 : &AArch64::GPR32RegClass;
342 unsigned ZeroReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
343 unsigned ResultReg = createResultReg(RC);
344 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TargetOpcode::COPY),
345 ResultReg).addReg(ZeroReg, getKillRegState(true));
349 unsigned AArch64FastISel::materializeFP(const ConstantFP *CFP, MVT VT) {
350 // Positive zero (+0.0) has to be materialized with a fmov from the zero
351 // register, because the immediate version of fmov cannot encode zero.
352 if (CFP->isNullValue())
353 return fastMaterializeFloatZero(CFP);
355 if (VT != MVT::f32 && VT != MVT::f64)
358 const APFloat Val = CFP->getValueAPF();
359 bool Is64Bit = (VT == MVT::f64);
360 // This checks to see if we can use FMOV instructions to materialize
361 // a constant, otherwise we have to materialize via the constant pool.
362 if (TLI.isFPImmLegal(Val, VT)) {
364 Is64Bit ? AArch64_AM::getFP64Imm(Val) : AArch64_AM::getFP32Imm(Val);
365 assert((Imm != -1) && "Cannot encode floating-point constant.");
366 unsigned Opc = Is64Bit ? AArch64::FMOVDi : AArch64::FMOVSi;
367 return fastEmitInst_i(Opc, TLI.getRegClassFor(VT), Imm);
370 // Materialize via constant pool. MachineConstantPool wants an explicit
372 unsigned Align = DL.getPrefTypeAlignment(CFP->getType());
374 Align = DL.getTypeAllocSize(CFP->getType());
376 unsigned CPI = MCP.getConstantPoolIndex(cast<Constant>(CFP), Align);
377 unsigned ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
378 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
379 ADRPReg).addConstantPoolIndex(CPI, 0, AArch64II::MO_PAGE);
381 unsigned Opc = Is64Bit ? AArch64::LDRDui : AArch64::LDRSui;
382 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
383 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
385 .addConstantPoolIndex(CPI, 0, AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
389 unsigned AArch64FastISel::materializeGV(const GlobalValue *GV) {
390 // We can't handle thread-local variables quickly yet.
391 if (GV->isThreadLocal())
394 // MachO still uses GOT for large code-model accesses, but ELF requires
395 // movz/movk sequences, which FastISel doesn't handle yet.
396 if (TM.getCodeModel() != CodeModel::Small && !Subtarget->isTargetMachO())
399 unsigned char OpFlags = Subtarget->ClassifyGlobalReference(GV, TM);
401 EVT DestEVT = TLI.getValueType(GV->getType(), true);
402 if (!DestEVT.isSimple())
405 unsigned ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
408 if (OpFlags & AArch64II::MO_GOT) {
410 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
412 .addGlobalAddress(GV, 0, AArch64II::MO_GOT | AArch64II::MO_PAGE);
414 ResultReg = createResultReg(&AArch64::GPR64RegClass);
415 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::LDRXui),
418 .addGlobalAddress(GV, 0, AArch64II::MO_GOT | AArch64II::MO_PAGEOFF |
420 } else if (OpFlags & AArch64II::MO_CONSTPOOL) {
421 // We can't handle addresses loaded from a constant pool quickly yet.
425 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
427 .addGlobalAddress(GV, 0, AArch64II::MO_PAGE);
429 ResultReg = createResultReg(&AArch64::GPR64spRegClass);
430 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri),
433 .addGlobalAddress(GV, 0, AArch64II::MO_PAGEOFF | AArch64II::MO_NC)
439 unsigned AArch64FastISel::fastMaterializeConstant(const Constant *C) {
440 EVT CEVT = TLI.getValueType(C->getType(), true);
442 // Only handle simple types.
443 if (!CEVT.isSimple())
445 MVT VT = CEVT.getSimpleVT();
447 if (const auto *CI = dyn_cast<ConstantInt>(C))
448 return materializeInt(CI, VT);
449 else if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
450 return materializeFP(CFP, VT);
451 else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
452 return materializeGV(GV);
457 unsigned AArch64FastISel::fastMaterializeFloatZero(const ConstantFP* CFP) {
458 assert(CFP->isNullValue() &&
459 "Floating-point constant is not a positive zero.");
461 if (!isTypeLegal(CFP->getType(), VT))
464 if (VT != MVT::f32 && VT != MVT::f64)
467 bool Is64Bit = (VT == MVT::f64);
468 unsigned ZReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
469 unsigned Opc = Is64Bit ? AArch64::FMOVXDr : AArch64::FMOVWSr;
470 return fastEmitInst_r(Opc, TLI.getRegClassFor(VT), ZReg, /*IsKill=*/true);
473 /// \brief Check if the multiply is by a power-of-2 constant.
474 static bool isMulPowOf2(const Value *I) {
475 if (const auto *MI = dyn_cast<MulOperator>(I)) {
476 if (const auto *C = dyn_cast<ConstantInt>(MI->getOperand(0)))
477 if (C->getValue().isPowerOf2())
479 if (const auto *C = dyn_cast<ConstantInt>(MI->getOperand(1)))
480 if (C->getValue().isPowerOf2())
486 // Computes the address to get to an object.
487 bool AArch64FastISel::computeAddress(const Value *Obj, Address &Addr, Type *Ty)
489 const User *U = nullptr;
490 unsigned Opcode = Instruction::UserOp1;
491 if (const Instruction *I = dyn_cast<Instruction>(Obj)) {
492 // Don't walk into other basic blocks unless the object is an alloca from
493 // another block, otherwise it may not have a virtual register assigned.
494 if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(Obj)) ||
495 FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
496 Opcode = I->getOpcode();
499 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Obj)) {
500 Opcode = C->getOpcode();
504 if (const PointerType *Ty = dyn_cast<PointerType>(Obj->getType()))
505 if (Ty->getAddressSpace() > 255)
506 // Fast instruction selection doesn't support the special
513 case Instruction::BitCast: {
514 // Look through bitcasts.
515 return computeAddress(U->getOperand(0), Addr, Ty);
517 case Instruction::IntToPtr: {
518 // Look past no-op inttoptrs.
519 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
520 return computeAddress(U->getOperand(0), Addr, Ty);
523 case Instruction::PtrToInt: {
524 // Look past no-op ptrtoints.
525 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
526 return computeAddress(U->getOperand(0), Addr, Ty);
529 case Instruction::GetElementPtr: {
530 Address SavedAddr = Addr;
531 uint64_t TmpOffset = Addr.getOffset();
533 // Iterate through the GEP folding the constants into offsets where
535 gep_type_iterator GTI = gep_type_begin(U);
536 for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e;
538 const Value *Op = *i;
539 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
540 const StructLayout *SL = DL.getStructLayout(STy);
541 unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
542 TmpOffset += SL->getElementOffset(Idx);
544 uint64_t S = DL.getTypeAllocSize(GTI.getIndexedType());
546 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
547 // Constant-offset addressing.
548 TmpOffset += CI->getSExtValue() * S;
551 if (canFoldAddIntoGEP(U, Op)) {
552 // A compatible add with a constant operand. Fold the constant.
554 cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
555 TmpOffset += CI->getSExtValue() * S;
556 // Iterate on the other operand.
557 Op = cast<AddOperator>(Op)->getOperand(0);
561 goto unsupported_gep;
566 // Try to grab the base operand now.
567 Addr.setOffset(TmpOffset);
568 if (computeAddress(U->getOperand(0), Addr, Ty))
571 // We failed, restore everything and try the other options.
577 case Instruction::Alloca: {
578 const AllocaInst *AI = cast<AllocaInst>(Obj);
579 DenseMap<const AllocaInst *, int>::iterator SI =
580 FuncInfo.StaticAllocaMap.find(AI);
581 if (SI != FuncInfo.StaticAllocaMap.end()) {
582 Addr.setKind(Address::FrameIndexBase);
583 Addr.setFI(SI->second);
588 case Instruction::Add: {
589 // Adds of constants are common and easy enough.
590 const Value *LHS = U->getOperand(0);
591 const Value *RHS = U->getOperand(1);
593 if (isa<ConstantInt>(LHS))
596 if (const ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
597 Addr.setOffset(Addr.getOffset() + CI->getSExtValue());
598 return computeAddress(LHS, Addr, Ty);
601 Address Backup = Addr;
602 if (computeAddress(LHS, Addr, Ty) && computeAddress(RHS, Addr, Ty))
608 case Instruction::Sub: {
609 // Subs of constants are common and easy enough.
610 const Value *LHS = U->getOperand(0);
611 const Value *RHS = U->getOperand(1);
613 if (const ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
614 Addr.setOffset(Addr.getOffset() - CI->getSExtValue());
615 return computeAddress(LHS, Addr, Ty);
619 case Instruction::Shl: {
620 if (Addr.getOffsetReg())
623 const auto *CI = dyn_cast<ConstantInt>(U->getOperand(1));
627 unsigned Val = CI->getZExtValue();
628 if (Val < 1 || Val > 3)
631 uint64_t NumBytes = 0;
632 if (Ty && Ty->isSized()) {
633 uint64_t NumBits = DL.getTypeSizeInBits(Ty);
634 NumBytes = NumBits / 8;
635 if (!isPowerOf2_64(NumBits))
639 if (NumBytes != (1ULL << Val))
643 Addr.setExtendType(AArch64_AM::LSL);
645 const Value *Src = U->getOperand(0);
646 if (const auto *I = dyn_cast<Instruction>(Src))
647 if (FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB)
650 // Fold the zext or sext when it won't become a noop.
651 if (const auto *ZE = dyn_cast<ZExtInst>(Src)) {
652 if (!isIntExtFree(ZE) && ZE->getOperand(0)->getType()->isIntegerTy(32)) {
653 Addr.setExtendType(AArch64_AM::UXTW);
654 Src = ZE->getOperand(0);
656 } else if (const auto *SE = dyn_cast<SExtInst>(Src)) {
657 if (!isIntExtFree(SE) && SE->getOperand(0)->getType()->isIntegerTy(32)) {
658 Addr.setExtendType(AArch64_AM::SXTW);
659 Src = SE->getOperand(0);
663 if (const auto *AI = dyn_cast<BinaryOperator>(Src))
664 if (AI->getOpcode() == Instruction::And) {
665 const Value *LHS = AI->getOperand(0);
666 const Value *RHS = AI->getOperand(1);
668 if (const auto *C = dyn_cast<ConstantInt>(LHS))
669 if (C->getValue() == 0xffffffff)
672 if (const auto *C = dyn_cast<ConstantInt>(RHS))
673 if (C->getValue() == 0xffffffff) {
674 Addr.setExtendType(AArch64_AM::UXTW);
675 unsigned Reg = getRegForValue(LHS);
678 bool RegIsKill = hasTrivialKill(LHS);
679 Reg = fastEmitInst_extractsubreg(MVT::i32, Reg, RegIsKill,
681 Addr.setOffsetReg(Reg);
686 unsigned Reg = getRegForValue(Src);
689 Addr.setOffsetReg(Reg);
692 case Instruction::Mul: {
693 if (Addr.getOffsetReg())
699 const Value *LHS = U->getOperand(0);
700 const Value *RHS = U->getOperand(1);
702 // Canonicalize power-of-2 value to the RHS.
703 if (const auto *C = dyn_cast<ConstantInt>(LHS))
704 if (C->getValue().isPowerOf2())
707 assert(isa<ConstantInt>(RHS) && "Expected an ConstantInt.");
708 const auto *C = cast<ConstantInt>(RHS);
709 unsigned Val = C->getValue().logBase2();
710 if (Val < 1 || Val > 3)
713 uint64_t NumBytes = 0;
714 if (Ty && Ty->isSized()) {
715 uint64_t NumBits = DL.getTypeSizeInBits(Ty);
716 NumBytes = NumBits / 8;
717 if (!isPowerOf2_64(NumBits))
721 if (NumBytes != (1ULL << Val))
725 Addr.setExtendType(AArch64_AM::LSL);
727 const Value *Src = LHS;
728 if (const auto *I = dyn_cast<Instruction>(Src))
729 if (FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB)
733 // Fold the zext or sext when it won't become a noop.
734 if (const auto *ZE = dyn_cast<ZExtInst>(Src)) {
735 if (!isIntExtFree(ZE) && ZE->getOperand(0)->getType()->isIntegerTy(32)) {
736 Addr.setExtendType(AArch64_AM::UXTW);
737 Src = ZE->getOperand(0);
739 } else if (const auto *SE = dyn_cast<SExtInst>(Src)) {
740 if (!isIntExtFree(SE) && SE->getOperand(0)->getType()->isIntegerTy(32)) {
741 Addr.setExtendType(AArch64_AM::SXTW);
742 Src = SE->getOperand(0);
746 unsigned Reg = getRegForValue(Src);
749 Addr.setOffsetReg(Reg);
752 case Instruction::And: {
753 if (Addr.getOffsetReg())
756 if (DL.getTypeSizeInBits(Ty) != 8)
759 const Value *LHS = U->getOperand(0);
760 const Value *RHS = U->getOperand(1);
762 if (const auto *C = dyn_cast<ConstantInt>(LHS))
763 if (C->getValue() == 0xffffffff)
766 if (const auto *C = dyn_cast<ConstantInt>(RHS))
767 if (C->getValue() == 0xffffffff) {
769 Addr.setExtendType(AArch64_AM::LSL);
770 Addr.setExtendType(AArch64_AM::UXTW);
772 unsigned Reg = getRegForValue(LHS);
775 bool RegIsKill = hasTrivialKill(LHS);
776 Reg = fastEmitInst_extractsubreg(MVT::i32, Reg, RegIsKill,
778 Addr.setOffsetReg(Reg);
783 case Instruction::SExt:
784 case Instruction::ZExt: {
785 if (!Addr.getReg() || Addr.getOffsetReg())
788 const Value *Src = nullptr;
789 // Fold the zext or sext when it won't become a noop.
790 if (const auto *ZE = dyn_cast<ZExtInst>(U)) {
791 if (!isIntExtFree(ZE) && ZE->getOperand(0)->getType()->isIntegerTy(32)) {
792 Addr.setExtendType(AArch64_AM::UXTW);
793 Src = ZE->getOperand(0);
795 } else if (const auto *SE = dyn_cast<SExtInst>(U)) {
796 if (!isIntExtFree(SE) && SE->getOperand(0)->getType()->isIntegerTy(32)) {
797 Addr.setExtendType(AArch64_AM::SXTW);
798 Src = SE->getOperand(0);
806 unsigned Reg = getRegForValue(Src);
809 Addr.setOffsetReg(Reg);
814 if (Addr.isRegBase() && !Addr.getReg()) {
815 unsigned Reg = getRegForValue(Obj);
822 if (!Addr.getOffsetReg()) {
823 unsigned Reg = getRegForValue(Obj);
826 Addr.setOffsetReg(Reg);
833 bool AArch64FastISel::computeCallAddress(const Value *V, Address &Addr) {
834 const User *U = nullptr;
835 unsigned Opcode = Instruction::UserOp1;
838 if (const auto *I = dyn_cast<Instruction>(V)) {
839 Opcode = I->getOpcode();
841 InMBB = I->getParent() == FuncInfo.MBB->getBasicBlock();
842 } else if (const auto *C = dyn_cast<ConstantExpr>(V)) {
843 Opcode = C->getOpcode();
849 case Instruction::BitCast:
850 // Look past bitcasts if its operand is in the same BB.
852 return computeCallAddress(U->getOperand(0), Addr);
854 case Instruction::IntToPtr:
855 // Look past no-op inttoptrs if its operand is in the same BB.
857 TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
858 return computeCallAddress(U->getOperand(0), Addr);
860 case Instruction::PtrToInt:
861 // Look past no-op ptrtoints if its operand is in the same BB.
863 TLI.getValueType(U->getType()) == TLI.getPointerTy())
864 return computeCallAddress(U->getOperand(0), Addr);
868 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
869 Addr.setGlobalValue(GV);
873 // If all else fails, try to materialize the value in a register.
874 if (!Addr.getGlobalValue()) {
875 Addr.setReg(getRegForValue(V));
876 return Addr.getReg() != 0;
883 bool AArch64FastISel::isTypeLegal(Type *Ty, MVT &VT) {
884 EVT evt = TLI.getValueType(Ty, true);
886 // Only handle simple types.
887 if (evt == MVT::Other || !evt.isSimple())
889 VT = evt.getSimpleVT();
891 // This is a legal type, but it's not something we handle in fast-isel.
895 // Handle all other legal types, i.e. a register that will directly hold this
897 return TLI.isTypeLegal(VT);
900 /// \brief Determine if the value type is supported by FastISel.
902 /// FastISel for AArch64 can handle more value types than are legal. This adds
903 /// simple value type such as i1, i8, and i16.
904 bool AArch64FastISel::isTypeSupported(Type *Ty, MVT &VT, bool IsVectorAllowed) {
905 if (Ty->isVectorTy() && !IsVectorAllowed)
908 if (isTypeLegal(Ty, VT))
911 // If this is a type than can be sign or zero-extended to a basic operation
912 // go ahead and accept it now.
913 if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16)
919 bool AArch64FastISel::isValueAvailable(const Value *V) const {
920 if (!isa<Instruction>(V))
923 const auto *I = cast<Instruction>(V);
924 if (FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB)
930 bool AArch64FastISel::simplifyAddress(Address &Addr, MVT VT) {
931 unsigned ScaleFactor = getImplicitScaleFactor(VT);
935 bool ImmediateOffsetNeedsLowering = false;
936 bool RegisterOffsetNeedsLowering = false;
937 int64_t Offset = Addr.getOffset();
938 if (((Offset < 0) || (Offset & (ScaleFactor - 1))) && !isInt<9>(Offset))
939 ImmediateOffsetNeedsLowering = true;
940 else if (Offset > 0 && !(Offset & (ScaleFactor - 1)) &&
941 !isUInt<12>(Offset / ScaleFactor))
942 ImmediateOffsetNeedsLowering = true;
944 // Cannot encode an offset register and an immediate offset in the same
945 // instruction. Fold the immediate offset into the load/store instruction and
946 // emit an additonal add to take care of the offset register.
947 if (!ImmediateOffsetNeedsLowering && Addr.getOffset() && Addr.getOffsetReg())
948 RegisterOffsetNeedsLowering = true;
950 // Cannot encode zero register as base.
951 if (Addr.isRegBase() && Addr.getOffsetReg() && !Addr.getReg())
952 RegisterOffsetNeedsLowering = true;
954 // If this is a stack pointer and the offset needs to be simplified then put
955 // the alloca address into a register, set the base type back to register and
956 // continue. This should almost never happen.
957 if ((ImmediateOffsetNeedsLowering || Addr.getOffsetReg()) && Addr.isFIBase())
959 unsigned ResultReg = createResultReg(&AArch64::GPR64spRegClass);
960 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri),
962 .addFrameIndex(Addr.getFI())
965 Addr.setKind(Address::RegBase);
966 Addr.setReg(ResultReg);
969 if (RegisterOffsetNeedsLowering) {
970 unsigned ResultReg = 0;
972 if (Addr.getExtendType() == AArch64_AM::SXTW ||
973 Addr.getExtendType() == AArch64_AM::UXTW )
974 ResultReg = emitAddSub_rx(/*UseAdd=*/true, MVT::i64, Addr.getReg(),
975 /*TODO:IsKill=*/false, Addr.getOffsetReg(),
976 /*TODO:IsKill=*/false, Addr.getExtendType(),
979 ResultReg = emitAddSub_rs(/*UseAdd=*/true, MVT::i64, Addr.getReg(),
980 /*TODO:IsKill=*/false, Addr.getOffsetReg(),
981 /*TODO:IsKill=*/false, AArch64_AM::LSL,
984 if (Addr.getExtendType() == AArch64_AM::UXTW)
985 ResultReg = emitLSL_ri(MVT::i64, MVT::i32, Addr.getOffsetReg(),
986 /*Op0IsKill=*/false, Addr.getShift(),
988 else if (Addr.getExtendType() == AArch64_AM::SXTW)
989 ResultReg = emitLSL_ri(MVT::i64, MVT::i32, Addr.getOffsetReg(),
990 /*Op0IsKill=*/false, Addr.getShift(),
993 ResultReg = emitLSL_ri(MVT::i64, MVT::i64, Addr.getOffsetReg(),
994 /*Op0IsKill=*/false, Addr.getShift());
999 Addr.setReg(ResultReg);
1000 Addr.setOffsetReg(0);
1002 Addr.setExtendType(AArch64_AM::InvalidShiftExtend);
1005 // Since the offset is too large for the load/store instruction get the
1006 // reg+offset into a register.
1007 if (ImmediateOffsetNeedsLowering) {
1010 // Try to fold the immediate into the add instruction.
1011 ResultReg = emitAdd_ri_(MVT::i64, Addr.getReg(), /*IsKill=*/false, Offset);
1013 ResultReg = fastEmit_i(MVT::i64, MVT::i64, ISD::Constant, Offset);
1017 Addr.setReg(ResultReg);
1023 void AArch64FastISel::addLoadStoreOperands(Address &Addr,
1024 const MachineInstrBuilder &MIB,
1026 unsigned ScaleFactor,
1027 MachineMemOperand *MMO) {
1028 int64_t Offset = Addr.getOffset() / ScaleFactor;
1029 // Frame base works a bit differently. Handle it separately.
1030 if (Addr.isFIBase()) {
1031 int FI = Addr.getFI();
1032 // FIXME: We shouldn't be using getObjectSize/getObjectAlignment. The size
1033 // and alignment should be based on the VT.
1034 MMO = FuncInfo.MF->getMachineMemOperand(
1035 MachinePointerInfo::getFixedStack(FI, Offset), Flags,
1036 MFI.getObjectSize(FI), MFI.getObjectAlignment(FI));
1037 // Now add the rest of the operands.
1038 MIB.addFrameIndex(FI).addImm(Offset);
1040 assert(Addr.isRegBase() && "Unexpected address kind.");
1041 const MCInstrDesc &II = MIB->getDesc();
1042 unsigned Idx = (Flags & MachineMemOperand::MOStore) ? 1 : 0;
1044 constrainOperandRegClass(II, Addr.getReg(), II.getNumDefs()+Idx));
1046 constrainOperandRegClass(II, Addr.getOffsetReg(), II.getNumDefs()+Idx+1));
1047 if (Addr.getOffsetReg()) {
1048 assert(Addr.getOffset() == 0 && "Unexpected offset");
1049 bool IsSigned = Addr.getExtendType() == AArch64_AM::SXTW ||
1050 Addr.getExtendType() == AArch64_AM::SXTX;
1051 MIB.addReg(Addr.getReg());
1052 MIB.addReg(Addr.getOffsetReg());
1053 MIB.addImm(IsSigned);
1054 MIB.addImm(Addr.getShift() != 0);
1056 MIB.addReg(Addr.getReg()).addImm(Offset);
1060 MIB.addMemOperand(MMO);
1063 unsigned AArch64FastISel::emitAddSub(bool UseAdd, MVT RetVT, const Value *LHS,
1064 const Value *RHS, bool SetFlags,
1065 bool WantResult, bool IsZExt) {
1066 AArch64_AM::ShiftExtendType ExtendType = AArch64_AM::InvalidShiftExtend;
1067 bool NeedExtend = false;
1068 switch (RetVT.SimpleTy) {
1076 ExtendType = IsZExt ? AArch64_AM::UXTB : AArch64_AM::SXTB;
1080 ExtendType = IsZExt ? AArch64_AM::UXTH : AArch64_AM::SXTH;
1082 case MVT::i32: // fall-through
1087 RetVT.SimpleTy = std::max(RetVT.SimpleTy, MVT::i32);
1089 // Canonicalize immediates to the RHS first.
1090 if (UseAdd && isa<Constant>(LHS) && !isa<Constant>(RHS))
1091 std::swap(LHS, RHS);
1093 // Canonicalize mul by power of 2 to the RHS.
1094 if (UseAdd && LHS->hasOneUse() && isValueAvailable(LHS))
1095 if (isMulPowOf2(LHS))
1096 std::swap(LHS, RHS);
1098 // Canonicalize shift immediate to the RHS.
1099 if (UseAdd && LHS->hasOneUse() && isValueAvailable(LHS))
1100 if (const auto *SI = dyn_cast<BinaryOperator>(LHS))
1101 if (isa<ConstantInt>(SI->getOperand(1)))
1102 if (SI->getOpcode() == Instruction::Shl ||
1103 SI->getOpcode() == Instruction::LShr ||
1104 SI->getOpcode() == Instruction::AShr )
1105 std::swap(LHS, RHS);
1107 unsigned LHSReg = getRegForValue(LHS);
1110 bool LHSIsKill = hasTrivialKill(LHS);
1113 LHSReg = emitIntExt(SrcVT, LHSReg, RetVT, IsZExt);
1115 unsigned ResultReg = 0;
1116 if (const auto *C = dyn_cast<ConstantInt>(RHS)) {
1117 uint64_t Imm = IsZExt ? C->getZExtValue() : C->getSExtValue();
1118 if (C->isNegative())
1119 ResultReg = emitAddSub_ri(!UseAdd, RetVT, LHSReg, LHSIsKill, -Imm,
1120 SetFlags, WantResult);
1122 ResultReg = emitAddSub_ri(UseAdd, RetVT, LHSReg, LHSIsKill, Imm, SetFlags,
1124 } else if (const auto *C = dyn_cast<Constant>(RHS))
1125 if (C->isNullValue())
1126 ResultReg = emitAddSub_ri(UseAdd, RetVT, LHSReg, LHSIsKill, 0, SetFlags,
1132 // Only extend the RHS within the instruction if there is a valid extend type.
1133 if (ExtendType != AArch64_AM::InvalidShiftExtend && RHS->hasOneUse() &&
1134 isValueAvailable(RHS)) {
1135 if (const auto *SI = dyn_cast<BinaryOperator>(RHS))
1136 if (const auto *C = dyn_cast<ConstantInt>(SI->getOperand(1)))
1137 if ((SI->getOpcode() == Instruction::Shl) && (C->getZExtValue() < 4)) {
1138 unsigned RHSReg = getRegForValue(SI->getOperand(0));
1141 bool RHSIsKill = hasTrivialKill(SI->getOperand(0));
1142 return emitAddSub_rx(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg,
1143 RHSIsKill, ExtendType, C->getZExtValue(),
1144 SetFlags, WantResult);
1146 unsigned RHSReg = getRegForValue(RHS);
1149 bool RHSIsKill = hasTrivialKill(RHS);
1150 return emitAddSub_rx(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg, RHSIsKill,
1151 ExtendType, 0, SetFlags, WantResult);
1154 // Check if the mul can be folded into the instruction.
1155 if (RHS->hasOneUse() && isValueAvailable(RHS))
1156 if (isMulPowOf2(RHS)) {
1157 const Value *MulLHS = cast<MulOperator>(RHS)->getOperand(0);
1158 const Value *MulRHS = cast<MulOperator>(RHS)->getOperand(1);
1160 if (const auto *C = dyn_cast<ConstantInt>(MulLHS))
1161 if (C->getValue().isPowerOf2())
1162 std::swap(MulLHS, MulRHS);
1164 assert(isa<ConstantInt>(MulRHS) && "Expected a ConstantInt.");
1165 uint64_t ShiftVal = cast<ConstantInt>(MulRHS)->getValue().logBase2();
1166 unsigned RHSReg = getRegForValue(MulLHS);
1169 bool RHSIsKill = hasTrivialKill(MulLHS);
1170 return emitAddSub_rs(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg, RHSIsKill,
1171 AArch64_AM::LSL, ShiftVal, SetFlags, WantResult);
1174 // Check if the shift can be folded into the instruction.
1175 if (RHS->hasOneUse() && isValueAvailable(RHS))
1176 if (const auto *SI = dyn_cast<BinaryOperator>(RHS)) {
1177 if (const auto *C = dyn_cast<ConstantInt>(SI->getOperand(1))) {
1178 AArch64_AM::ShiftExtendType ShiftType = AArch64_AM::InvalidShiftExtend;
1179 switch (SI->getOpcode()) {
1181 case Instruction::Shl: ShiftType = AArch64_AM::LSL; break;
1182 case Instruction::LShr: ShiftType = AArch64_AM::LSR; break;
1183 case Instruction::AShr: ShiftType = AArch64_AM::ASR; break;
1185 uint64_t ShiftVal = C->getZExtValue();
1186 if (ShiftType != AArch64_AM::InvalidShiftExtend) {
1187 unsigned RHSReg = getRegForValue(SI->getOperand(0));
1190 bool RHSIsKill = hasTrivialKill(SI->getOperand(0));
1191 return emitAddSub_rs(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg,
1192 RHSIsKill, ShiftType, ShiftVal, SetFlags,
1198 unsigned RHSReg = getRegForValue(RHS);
1201 bool RHSIsKill = hasTrivialKill(RHS);
1204 RHSReg = emitIntExt(SrcVT, RHSReg, RetVT, IsZExt);
1206 return emitAddSub_rr(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg, RHSIsKill,
1207 SetFlags, WantResult);
1210 unsigned AArch64FastISel::emitAddSub_rr(bool UseAdd, MVT RetVT, unsigned LHSReg,
1211 bool LHSIsKill, unsigned RHSReg,
1212 bool RHSIsKill, bool SetFlags,
1214 assert(LHSReg && RHSReg && "Invalid register number.");
1216 if (RetVT != MVT::i32 && RetVT != MVT::i64)
1219 static const unsigned OpcTable[2][2][2] = {
1220 { { AArch64::SUBWrr, AArch64::SUBXrr },
1221 { AArch64::ADDWrr, AArch64::ADDXrr } },
1222 { { AArch64::SUBSWrr, AArch64::SUBSXrr },
1223 { AArch64::ADDSWrr, AArch64::ADDSXrr } }
1225 bool Is64Bit = RetVT == MVT::i64;
1226 unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit];
1227 const TargetRegisterClass *RC =
1228 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
1231 ResultReg = createResultReg(RC);
1233 ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
1235 const MCInstrDesc &II = TII.get(Opc);
1236 LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
1237 RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1);
1238 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
1239 .addReg(LHSReg, getKillRegState(LHSIsKill))
1240 .addReg(RHSReg, getKillRegState(RHSIsKill));
1244 unsigned AArch64FastISel::emitAddSub_ri(bool UseAdd, MVT RetVT, unsigned LHSReg,
1245 bool LHSIsKill, uint64_t Imm,
1246 bool SetFlags, bool WantResult) {
1247 assert(LHSReg && "Invalid register number.");
1249 if (RetVT != MVT::i32 && RetVT != MVT::i64)
1253 if (isUInt<12>(Imm))
1255 else if ((Imm & 0xfff000) == Imm) {
1261 static const unsigned OpcTable[2][2][2] = {
1262 { { AArch64::SUBWri, AArch64::SUBXri },
1263 { AArch64::ADDWri, AArch64::ADDXri } },
1264 { { AArch64::SUBSWri, AArch64::SUBSXri },
1265 { AArch64::ADDSWri, AArch64::ADDSXri } }
1267 bool Is64Bit = RetVT == MVT::i64;
1268 unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit];
1269 const TargetRegisterClass *RC;
1271 RC = Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
1273 RC = Is64Bit ? &AArch64::GPR64spRegClass : &AArch64::GPR32spRegClass;
1276 ResultReg = createResultReg(RC);
1278 ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
1280 const MCInstrDesc &II = TII.get(Opc);
1281 LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
1282 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
1283 .addReg(LHSReg, getKillRegState(LHSIsKill))
1285 .addImm(getShifterImm(AArch64_AM::LSL, ShiftImm));
1289 unsigned AArch64FastISel::emitAddSub_rs(bool UseAdd, MVT RetVT, unsigned LHSReg,
1290 bool LHSIsKill, unsigned RHSReg,
1292 AArch64_AM::ShiftExtendType ShiftType,
1293 uint64_t ShiftImm, bool SetFlags,
1295 assert(LHSReg && RHSReg && "Invalid register number.");
1297 if (RetVT != MVT::i32 && RetVT != MVT::i64)
1300 static const unsigned OpcTable[2][2][2] = {
1301 { { AArch64::SUBWrs, AArch64::SUBXrs },
1302 { AArch64::ADDWrs, AArch64::ADDXrs } },
1303 { { AArch64::SUBSWrs, AArch64::SUBSXrs },
1304 { AArch64::ADDSWrs, AArch64::ADDSXrs } }
1306 bool Is64Bit = RetVT == MVT::i64;
1307 unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit];
1308 const TargetRegisterClass *RC =
1309 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
1312 ResultReg = createResultReg(RC);
1314 ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
1316 const MCInstrDesc &II = TII.get(Opc);
1317 LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
1318 RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1);
1319 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
1320 .addReg(LHSReg, getKillRegState(LHSIsKill))
1321 .addReg(RHSReg, getKillRegState(RHSIsKill))
1322 .addImm(getShifterImm(ShiftType, ShiftImm));
1326 unsigned AArch64FastISel::emitAddSub_rx(bool UseAdd, MVT RetVT, unsigned LHSReg,
1327 bool LHSIsKill, unsigned RHSReg,
1329 AArch64_AM::ShiftExtendType ExtType,
1330 uint64_t ShiftImm, bool SetFlags,
1332 assert(LHSReg && RHSReg && "Invalid register number.");
1334 if (RetVT != MVT::i32 && RetVT != MVT::i64)
1337 static const unsigned OpcTable[2][2][2] = {
1338 { { AArch64::SUBWrx, AArch64::SUBXrx },
1339 { AArch64::ADDWrx, AArch64::ADDXrx } },
1340 { { AArch64::SUBSWrx, AArch64::SUBSXrx },
1341 { AArch64::ADDSWrx, AArch64::ADDSXrx } }
1343 bool Is64Bit = RetVT == MVT::i64;
1344 unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit];
1345 const TargetRegisterClass *RC = nullptr;
1347 RC = Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
1349 RC = Is64Bit ? &AArch64::GPR64spRegClass : &AArch64::GPR32spRegClass;
1352 ResultReg = createResultReg(RC);
1354 ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
1356 const MCInstrDesc &II = TII.get(Opc);
1357 LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
1358 RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1);
1359 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
1360 .addReg(LHSReg, getKillRegState(LHSIsKill))
1361 .addReg(RHSReg, getKillRegState(RHSIsKill))
1362 .addImm(getArithExtendImm(ExtType, ShiftImm));
1366 bool AArch64FastISel::emitCmp(const Value *LHS, const Value *RHS, bool IsZExt) {
1367 Type *Ty = LHS->getType();
1368 EVT EVT = TLI.getValueType(Ty, true);
1369 if (!EVT.isSimple())
1371 MVT VT = EVT.getSimpleVT();
1373 switch (VT.SimpleTy) {
1381 return emitICmp(VT, LHS, RHS, IsZExt);
1384 return emitFCmp(VT, LHS, RHS);
1388 bool AArch64FastISel::emitICmp(MVT RetVT, const Value *LHS, const Value *RHS,
1390 return emitSub(RetVT, LHS, RHS, /*SetFlags=*/true, /*WantResult=*/false,
1394 bool AArch64FastISel::emitICmp_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill,
1396 return emitAddSub_ri(/*UseAdd=*/false, RetVT, LHSReg, LHSIsKill, Imm,
1397 /*SetFlags=*/true, /*WantResult=*/false) != 0;
1400 bool AArch64FastISel::emitFCmp(MVT RetVT, const Value *LHS, const Value *RHS) {
1401 if (RetVT != MVT::f32 && RetVT != MVT::f64)
1404 // Check to see if the 2nd operand is a constant that we can encode directly
1406 bool UseImm = false;
1407 if (const auto *CFP = dyn_cast<ConstantFP>(RHS))
1408 if (CFP->isZero() && !CFP->isNegative())
1411 unsigned LHSReg = getRegForValue(LHS);
1414 bool LHSIsKill = hasTrivialKill(LHS);
1417 unsigned Opc = (RetVT == MVT::f64) ? AArch64::FCMPDri : AArch64::FCMPSri;
1418 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc))
1419 .addReg(LHSReg, getKillRegState(LHSIsKill));
1423 unsigned RHSReg = getRegForValue(RHS);
1426 bool RHSIsKill = hasTrivialKill(RHS);
1428 unsigned Opc = (RetVT == MVT::f64) ? AArch64::FCMPDrr : AArch64::FCMPSrr;
1429 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc))
1430 .addReg(LHSReg, getKillRegState(LHSIsKill))
1431 .addReg(RHSReg, getKillRegState(RHSIsKill));
1435 unsigned AArch64FastISel::emitAdd(MVT RetVT, const Value *LHS, const Value *RHS,
1436 bool SetFlags, bool WantResult, bool IsZExt) {
1437 return emitAddSub(/*UseAdd=*/true, RetVT, LHS, RHS, SetFlags, WantResult,
1441 /// \brief This method is a wrapper to simplify add emission.
1443 /// First try to emit an add with an immediate operand using emitAddSub_ri. If
1444 /// that fails, then try to materialize the immediate into a register and use
1445 /// emitAddSub_rr instead.
1446 unsigned AArch64FastISel::emitAdd_ri_(MVT VT, unsigned Op0, bool Op0IsKill,
1450 ResultReg = emitAddSub_ri(false, VT, Op0, Op0IsKill, -Imm);
1452 ResultReg = emitAddSub_ri(true, VT, Op0, Op0IsKill, Imm);
1457 unsigned CReg = fastEmit_i(VT, VT, ISD::Constant, Imm);
1461 ResultReg = emitAddSub_rr(true, VT, Op0, Op0IsKill, CReg, true);
1465 unsigned AArch64FastISel::emitSub(MVT RetVT, const Value *LHS, const Value *RHS,
1466 bool SetFlags, bool WantResult, bool IsZExt) {
1467 return emitAddSub(/*UseAdd=*/false, RetVT, LHS, RHS, SetFlags, WantResult,
1471 unsigned AArch64FastISel::emitSubs_rr(MVT RetVT, unsigned LHSReg,
1472 bool LHSIsKill, unsigned RHSReg,
1473 bool RHSIsKill, bool WantResult) {
1474 return emitAddSub_rr(/*UseAdd=*/false, RetVT, LHSReg, LHSIsKill, RHSReg,
1475 RHSIsKill, /*SetFlags=*/true, WantResult);
1478 unsigned AArch64FastISel::emitSubs_rs(MVT RetVT, unsigned LHSReg,
1479 bool LHSIsKill, unsigned RHSReg,
1481 AArch64_AM::ShiftExtendType ShiftType,
1482 uint64_t ShiftImm, bool WantResult) {
1483 return emitAddSub_rs(/*UseAdd=*/false, RetVT, LHSReg, LHSIsKill, RHSReg,
1484 RHSIsKill, ShiftType, ShiftImm, /*SetFlags=*/true,
1488 unsigned AArch64FastISel::emitLogicalOp(unsigned ISDOpc, MVT RetVT,
1489 const Value *LHS, const Value *RHS) {
1490 // Canonicalize immediates to the RHS first.
1491 if (isa<ConstantInt>(LHS) && !isa<ConstantInt>(RHS))
1492 std::swap(LHS, RHS);
1494 // Canonicalize mul by power-of-2 to the RHS.
1495 if (LHS->hasOneUse() && isValueAvailable(LHS))
1496 if (isMulPowOf2(LHS))
1497 std::swap(LHS, RHS);
1499 // Canonicalize shift immediate to the RHS.
1500 if (LHS->hasOneUse() && isValueAvailable(LHS))
1501 if (const auto *SI = dyn_cast<ShlOperator>(LHS))
1502 if (isa<ConstantInt>(SI->getOperand(1)))
1503 std::swap(LHS, RHS);
1505 unsigned LHSReg = getRegForValue(LHS);
1508 bool LHSIsKill = hasTrivialKill(LHS);
1510 unsigned ResultReg = 0;
1511 if (const auto *C = dyn_cast<ConstantInt>(RHS)) {
1512 uint64_t Imm = C->getZExtValue();
1513 ResultReg = emitLogicalOp_ri(ISDOpc, RetVT, LHSReg, LHSIsKill, Imm);
1518 // Check if the mul can be folded into the instruction.
1519 if (RHS->hasOneUse() && isValueAvailable(RHS))
1520 if (isMulPowOf2(RHS)) {
1521 const Value *MulLHS = cast<MulOperator>(RHS)->getOperand(0);
1522 const Value *MulRHS = cast<MulOperator>(RHS)->getOperand(1);
1524 if (const auto *C = dyn_cast<ConstantInt>(MulLHS))
1525 if (C->getValue().isPowerOf2())
1526 std::swap(MulLHS, MulRHS);
1528 assert(isa<ConstantInt>(MulRHS) && "Expected a ConstantInt.");
1529 uint64_t ShiftVal = cast<ConstantInt>(MulRHS)->getValue().logBase2();
1531 unsigned RHSReg = getRegForValue(MulLHS);
1534 bool RHSIsKill = hasTrivialKill(MulLHS);
1535 return emitLogicalOp_rs(ISDOpc, RetVT, LHSReg, LHSIsKill, RHSReg,
1536 RHSIsKill, ShiftVal);
1539 // Check if the shift can be folded into the instruction.
1540 if (RHS->hasOneUse() && isValueAvailable(RHS))
1541 if (const auto *SI = dyn_cast<ShlOperator>(RHS))
1542 if (const auto *C = dyn_cast<ConstantInt>(SI->getOperand(1))) {
1543 uint64_t ShiftVal = C->getZExtValue();
1544 unsigned RHSReg = getRegForValue(SI->getOperand(0));
1547 bool RHSIsKill = hasTrivialKill(SI->getOperand(0));
1548 return emitLogicalOp_rs(ISDOpc, RetVT, LHSReg, LHSIsKill, RHSReg,
1549 RHSIsKill, ShiftVal);
1552 unsigned RHSReg = getRegForValue(RHS);
1555 bool RHSIsKill = hasTrivialKill(RHS);
1557 MVT VT = std::max(MVT::i32, RetVT.SimpleTy);
1558 ResultReg = fastEmit_rr(VT, VT, ISDOpc, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
1559 if (RetVT >= MVT::i8 && RetVT <= MVT::i16) {
1560 uint64_t Mask = (RetVT == MVT::i8) ? 0xff : 0xffff;
1561 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
1566 unsigned AArch64FastISel::emitLogicalOp_ri(unsigned ISDOpc, MVT RetVT,
1567 unsigned LHSReg, bool LHSIsKill,
1569 assert((ISD::AND + 1 == ISD::OR) && (ISD::AND + 2 == ISD::XOR) &&
1570 "ISD nodes are not consecutive!");
1571 static const unsigned OpcTable[3][2] = {
1572 { AArch64::ANDWri, AArch64::ANDXri },
1573 { AArch64::ORRWri, AArch64::ORRXri },
1574 { AArch64::EORWri, AArch64::EORXri }
1576 const TargetRegisterClass *RC;
1579 switch (RetVT.SimpleTy) {
1586 unsigned Idx = ISDOpc - ISD::AND;
1587 Opc = OpcTable[Idx][0];
1588 RC = &AArch64::GPR32spRegClass;
1593 Opc = OpcTable[ISDOpc - ISD::AND][1];
1594 RC = &AArch64::GPR64spRegClass;
1599 if (!AArch64_AM::isLogicalImmediate(Imm, RegSize))
1602 unsigned ResultReg =
1603 fastEmitInst_ri(Opc, RC, LHSReg, LHSIsKill,
1604 AArch64_AM::encodeLogicalImmediate(Imm, RegSize));
1605 if (RetVT >= MVT::i8 && RetVT <= MVT::i16 && ISDOpc != ISD::AND) {
1606 uint64_t Mask = (RetVT == MVT::i8) ? 0xff : 0xffff;
1607 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
1612 unsigned AArch64FastISel::emitLogicalOp_rs(unsigned ISDOpc, MVT RetVT,
1613 unsigned LHSReg, bool LHSIsKill,
1614 unsigned RHSReg, bool RHSIsKill,
1615 uint64_t ShiftImm) {
1616 assert((ISD::AND + 1 == ISD::OR) && (ISD::AND + 2 == ISD::XOR) &&
1617 "ISD nodes are not consecutive!");
1618 static const unsigned OpcTable[3][2] = {
1619 { AArch64::ANDWrs, AArch64::ANDXrs },
1620 { AArch64::ORRWrs, AArch64::ORRXrs },
1621 { AArch64::EORWrs, AArch64::EORXrs }
1623 const TargetRegisterClass *RC;
1625 switch (RetVT.SimpleTy) {
1632 Opc = OpcTable[ISDOpc - ISD::AND][0];
1633 RC = &AArch64::GPR32RegClass;
1636 Opc = OpcTable[ISDOpc - ISD::AND][1];
1637 RC = &AArch64::GPR64RegClass;
1640 unsigned ResultReg =
1641 fastEmitInst_rri(Opc, RC, LHSReg, LHSIsKill, RHSReg, RHSIsKill,
1642 AArch64_AM::getShifterImm(AArch64_AM::LSL, ShiftImm));
1643 if (RetVT >= MVT::i8 && RetVT <= MVT::i16) {
1644 uint64_t Mask = (RetVT == MVT::i8) ? 0xff : 0xffff;
1645 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
1650 unsigned AArch64FastISel::emitAnd_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill,
1652 return emitLogicalOp_ri(ISD::AND, RetVT, LHSReg, LHSIsKill, Imm);
1655 unsigned AArch64FastISel::emitLoad(MVT VT, MVT RetVT, Address Addr,
1656 bool WantZExt, MachineMemOperand *MMO) {
1657 // Simplify this down to something we can handle.
1658 if (!simplifyAddress(Addr, VT))
1661 unsigned ScaleFactor = getImplicitScaleFactor(VT);
1663 llvm_unreachable("Unexpected value type.");
1665 // Negative offsets require unscaled, 9-bit, signed immediate offsets.
1666 // Otherwise, we try using scaled, 12-bit, unsigned immediate offsets.
1667 bool UseScaled = true;
1668 if ((Addr.getOffset() < 0) || (Addr.getOffset() & (ScaleFactor - 1))) {
1673 static const unsigned GPOpcTable[2][8][4] = {
1675 { { AArch64::LDURSBWi, AArch64::LDURSHWi, AArch64::LDURWi,
1677 { AArch64::LDURSBXi, AArch64::LDURSHXi, AArch64::LDURSWi,
1679 { AArch64::LDRSBWui, AArch64::LDRSHWui, AArch64::LDRWui,
1681 { AArch64::LDRSBXui, AArch64::LDRSHXui, AArch64::LDRSWui,
1683 { AArch64::LDRSBWroX, AArch64::LDRSHWroX, AArch64::LDRWroX,
1685 { AArch64::LDRSBXroX, AArch64::LDRSHXroX, AArch64::LDRSWroX,
1687 { AArch64::LDRSBWroW, AArch64::LDRSHWroW, AArch64::LDRWroW,
1689 { AArch64::LDRSBXroW, AArch64::LDRSHXroW, AArch64::LDRSWroW,
1693 { { AArch64::LDURBBi, AArch64::LDURHHi, AArch64::LDURWi,
1695 { AArch64::LDURBBi, AArch64::LDURHHi, AArch64::LDURWi,
1697 { AArch64::LDRBBui, AArch64::LDRHHui, AArch64::LDRWui,
1699 { AArch64::LDRBBui, AArch64::LDRHHui, AArch64::LDRWui,
1701 { AArch64::LDRBBroX, AArch64::LDRHHroX, AArch64::LDRWroX,
1703 { AArch64::LDRBBroX, AArch64::LDRHHroX, AArch64::LDRWroX,
1705 { AArch64::LDRBBroW, AArch64::LDRHHroW, AArch64::LDRWroW,
1707 { AArch64::LDRBBroW, AArch64::LDRHHroW, AArch64::LDRWroW,
1712 static const unsigned FPOpcTable[4][2] = {
1713 { AArch64::LDURSi, AArch64::LDURDi },
1714 { AArch64::LDRSui, AArch64::LDRDui },
1715 { AArch64::LDRSroX, AArch64::LDRDroX },
1716 { AArch64::LDRSroW, AArch64::LDRDroW }
1720 const TargetRegisterClass *RC;
1721 bool UseRegOffset = Addr.isRegBase() && !Addr.getOffset() && Addr.getReg() &&
1722 Addr.getOffsetReg();
1723 unsigned Idx = UseRegOffset ? 2 : UseScaled ? 1 : 0;
1724 if (Addr.getExtendType() == AArch64_AM::UXTW ||
1725 Addr.getExtendType() == AArch64_AM::SXTW)
1728 bool IsRet64Bit = RetVT == MVT::i64;
1729 switch (VT.SimpleTy) {
1731 llvm_unreachable("Unexpected value type.");
1732 case MVT::i1: // Intentional fall-through.
1734 Opc = GPOpcTable[WantZExt][2 * Idx + IsRet64Bit][0];
1735 RC = (IsRet64Bit && !WantZExt) ?
1736 &AArch64::GPR64RegClass: &AArch64::GPR32RegClass;
1739 Opc = GPOpcTable[WantZExt][2 * Idx + IsRet64Bit][1];
1740 RC = (IsRet64Bit && !WantZExt) ?
1741 &AArch64::GPR64RegClass: &AArch64::GPR32RegClass;
1744 Opc = GPOpcTable[WantZExt][2 * Idx + IsRet64Bit][2];
1745 RC = (IsRet64Bit && !WantZExt) ?
1746 &AArch64::GPR64RegClass: &AArch64::GPR32RegClass;
1749 Opc = GPOpcTable[WantZExt][2 * Idx + IsRet64Bit][3];
1750 RC = &AArch64::GPR64RegClass;
1753 Opc = FPOpcTable[Idx][0];
1754 RC = &AArch64::FPR32RegClass;
1757 Opc = FPOpcTable[Idx][1];
1758 RC = &AArch64::FPR64RegClass;
1762 // Create the base instruction, then add the operands.
1763 unsigned ResultReg = createResultReg(RC);
1764 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1765 TII.get(Opc), ResultReg);
1766 addLoadStoreOperands(Addr, MIB, MachineMemOperand::MOLoad, ScaleFactor, MMO);
1768 // Loading an i1 requires special handling.
1769 if (VT == MVT::i1) {
1770 unsigned ANDReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, 1);
1771 assert(ANDReg && "Unexpected AND instruction emission failure.");
1775 // For zero-extending loads to 64bit we emit a 32bit load and then convert
1776 // the 32bit reg to a 64bit reg.
1777 if (WantZExt && RetVT == MVT::i64 && VT <= MVT::i32) {
1778 unsigned Reg64 = createResultReg(&AArch64::GPR64RegClass);
1779 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1780 TII.get(AArch64::SUBREG_TO_REG), Reg64)
1782 .addReg(ResultReg, getKillRegState(true))
1783 .addImm(AArch64::sub_32);
1789 bool AArch64FastISel::selectAddSub(const Instruction *I) {
1791 if (!isTypeSupported(I->getType(), VT, /*IsVectorAllowed=*/true))
1795 return selectOperator(I, I->getOpcode());
1798 switch (I->getOpcode()) {
1800 llvm_unreachable("Unexpected instruction.");
1801 case Instruction::Add:
1802 ResultReg = emitAdd(VT, I->getOperand(0), I->getOperand(1));
1804 case Instruction::Sub:
1805 ResultReg = emitSub(VT, I->getOperand(0), I->getOperand(1));
1811 updateValueMap(I, ResultReg);
1815 bool AArch64FastISel::selectLogicalOp(const Instruction *I) {
1817 if (!isTypeSupported(I->getType(), VT, /*IsVectorAllowed=*/true))
1821 return selectOperator(I, I->getOpcode());
1824 switch (I->getOpcode()) {
1826 llvm_unreachable("Unexpected instruction.");
1827 case Instruction::And:
1828 ResultReg = emitLogicalOp(ISD::AND, VT, I->getOperand(0), I->getOperand(1));
1830 case Instruction::Or:
1831 ResultReg = emitLogicalOp(ISD::OR, VT, I->getOperand(0), I->getOperand(1));
1833 case Instruction::Xor:
1834 ResultReg = emitLogicalOp(ISD::XOR, VT, I->getOperand(0), I->getOperand(1));
1840 updateValueMap(I, ResultReg);
1844 bool AArch64FastISel::selectLoad(const Instruction *I) {
1846 // Verify we have a legal type before going any further. Currently, we handle
1847 // simple types that will directly fit in a register (i32/f32/i64/f64) or
1848 // those that can be sign or zero-extended to a basic operation (i1/i8/i16).
1849 if (!isTypeSupported(I->getType(), VT, /*IsVectorAllowed=*/true) ||
1850 cast<LoadInst>(I)->isAtomic())
1853 // See if we can handle this address.
1855 if (!computeAddress(I->getOperand(0), Addr, I->getType()))
1858 // Fold the following sign-/zero-extend into the load instruction.
1859 bool WantZExt = true;
1861 const Value *IntExtVal = nullptr;
1862 if (I->hasOneUse()) {
1863 if (const auto *ZE = dyn_cast<ZExtInst>(I->use_begin()->getUser())) {
1864 if (isTypeSupported(ZE->getType(), RetVT))
1868 } else if (const auto *SE = dyn_cast<SExtInst>(I->use_begin()->getUser())) {
1869 if (isTypeSupported(SE->getType(), RetVT))
1877 unsigned ResultReg =
1878 emitLoad(VT, RetVT, Addr, WantZExt, createMachineMemOperandFor(I));
1882 // There are a few different cases we have to handle, because the load or the
1883 // sign-/zero-extend might not be selected by FastISel if we fall-back to
1884 // SelectionDAG. There is also an ordering issue when both instructions are in
1885 // different basic blocks.
1886 // 1.) The load instruction is selected by FastISel, but the integer extend
1887 // not. This usually happens when the integer extend is in a different
1888 // basic block and SelectionDAG took over for that basic block.
1889 // 2.) The load instruction is selected before the integer extend. This only
1890 // happens when the integer extend is in a different basic block.
1891 // 3.) The load instruction is selected by SelectionDAG and the integer extend
1892 // by FastISel. This happens if there are instructions between the load
1893 // and the integer extend that couldn't be selected by FastISel.
1895 // The integer extend hasn't been emitted yet. FastISel or SelectionDAG
1896 // could select it. Emit a copy to subreg if necessary. FastISel will remove
1897 // it when it selects the integer extend.
1898 unsigned Reg = lookUpRegForValue(IntExtVal);
1900 if (RetVT == MVT::i64 && VT <= MVT::i32) {
1902 // Delete the last emitted instruction from emitLoad (SUBREG_TO_REG).
1903 std::prev(FuncInfo.InsertPt)->eraseFromParent();
1904 ResultReg = std::prev(FuncInfo.InsertPt)->getOperand(0).getReg();
1906 ResultReg = fastEmitInst_extractsubreg(MVT::i32, ResultReg,
1910 updateValueMap(I, ResultReg);
1914 // The integer extend has already been emitted - delete all the instructions
1915 // that have been emitted by the integer extend lowering code and use the
1916 // result from the load instruction directly.
1918 auto *MI = MRI.getUniqueVRegDef(Reg);
1922 for (auto &Opnd : MI->uses()) {
1924 Reg = Opnd.getReg();
1928 MI->eraseFromParent();
1930 updateValueMap(IntExtVal, ResultReg);
1934 updateValueMap(I, ResultReg);
1938 bool AArch64FastISel::emitStore(MVT VT, unsigned SrcReg, Address Addr,
1939 MachineMemOperand *MMO) {
1940 // Simplify this down to something we can handle.
1941 if (!simplifyAddress(Addr, VT))
1944 unsigned ScaleFactor = getImplicitScaleFactor(VT);
1946 llvm_unreachable("Unexpected value type.");
1948 // Negative offsets require unscaled, 9-bit, signed immediate offsets.
1949 // Otherwise, we try using scaled, 12-bit, unsigned immediate offsets.
1950 bool UseScaled = true;
1951 if ((Addr.getOffset() < 0) || (Addr.getOffset() & (ScaleFactor - 1))) {
1956 static const unsigned OpcTable[4][6] = {
1957 { AArch64::STURBBi, AArch64::STURHHi, AArch64::STURWi, AArch64::STURXi,
1958 AArch64::STURSi, AArch64::STURDi },
1959 { AArch64::STRBBui, AArch64::STRHHui, AArch64::STRWui, AArch64::STRXui,
1960 AArch64::STRSui, AArch64::STRDui },
1961 { AArch64::STRBBroX, AArch64::STRHHroX, AArch64::STRWroX, AArch64::STRXroX,
1962 AArch64::STRSroX, AArch64::STRDroX },
1963 { AArch64::STRBBroW, AArch64::STRHHroW, AArch64::STRWroW, AArch64::STRXroW,
1964 AArch64::STRSroW, AArch64::STRDroW }
1968 bool VTIsi1 = false;
1969 bool UseRegOffset = Addr.isRegBase() && !Addr.getOffset() && Addr.getReg() &&
1970 Addr.getOffsetReg();
1971 unsigned Idx = UseRegOffset ? 2 : UseScaled ? 1 : 0;
1972 if (Addr.getExtendType() == AArch64_AM::UXTW ||
1973 Addr.getExtendType() == AArch64_AM::SXTW)
1976 switch (VT.SimpleTy) {
1977 default: llvm_unreachable("Unexpected value type.");
1978 case MVT::i1: VTIsi1 = true;
1979 case MVT::i8: Opc = OpcTable[Idx][0]; break;
1980 case MVT::i16: Opc = OpcTable[Idx][1]; break;
1981 case MVT::i32: Opc = OpcTable[Idx][2]; break;
1982 case MVT::i64: Opc = OpcTable[Idx][3]; break;
1983 case MVT::f32: Opc = OpcTable[Idx][4]; break;
1984 case MVT::f64: Opc = OpcTable[Idx][5]; break;
1987 // Storing an i1 requires special handling.
1988 if (VTIsi1 && SrcReg != AArch64::WZR) {
1989 unsigned ANDReg = emitAnd_ri(MVT::i32, SrcReg, /*TODO:IsKill=*/false, 1);
1990 assert(ANDReg && "Unexpected AND instruction emission failure.");
1993 // Create the base instruction, then add the operands.
1994 const MCInstrDesc &II = TII.get(Opc);
1995 SrcReg = constrainOperandRegClass(II, SrcReg, II.getNumDefs());
1996 MachineInstrBuilder MIB =
1997 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II).addReg(SrcReg);
1998 addLoadStoreOperands(Addr, MIB, MachineMemOperand::MOStore, ScaleFactor, MMO);
2003 bool AArch64FastISel::selectStore(const Instruction *I) {
2005 const Value *Op0 = I->getOperand(0);
2006 // Verify we have a legal type before going any further. Currently, we handle
2007 // simple types that will directly fit in a register (i32/f32/i64/f64) or
2008 // those that can be sign or zero-extended to a basic operation (i1/i8/i16).
2009 if (!isTypeSupported(Op0->getType(), VT, /*IsVectorAllowed=*/true) ||
2010 cast<StoreInst>(I)->isAtomic())
2013 // Get the value to be stored into a register. Use the zero register directly
2014 // when possible to avoid an unnecessary copy and a wasted register.
2015 unsigned SrcReg = 0;
2016 if (const auto *CI = dyn_cast<ConstantInt>(Op0)) {
2018 SrcReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
2019 } else if (const auto *CF = dyn_cast<ConstantFP>(Op0)) {
2020 if (CF->isZero() && !CF->isNegative()) {
2021 VT = MVT::getIntegerVT(VT.getSizeInBits());
2022 SrcReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
2027 SrcReg = getRegForValue(Op0);
2032 // See if we can handle this address.
2034 if (!computeAddress(I->getOperand(1), Addr, I->getOperand(0)->getType()))
2037 if (!emitStore(VT, SrcReg, Addr, createMachineMemOperandFor(I)))
2042 static AArch64CC::CondCode getCompareCC(CmpInst::Predicate Pred) {
2044 case CmpInst::FCMP_ONE:
2045 case CmpInst::FCMP_UEQ:
2047 // AL is our "false" for now. The other two need more compares.
2048 return AArch64CC::AL;
2049 case CmpInst::ICMP_EQ:
2050 case CmpInst::FCMP_OEQ:
2051 return AArch64CC::EQ;
2052 case CmpInst::ICMP_SGT:
2053 case CmpInst::FCMP_OGT:
2054 return AArch64CC::GT;
2055 case CmpInst::ICMP_SGE:
2056 case CmpInst::FCMP_OGE:
2057 return AArch64CC::GE;
2058 case CmpInst::ICMP_UGT:
2059 case CmpInst::FCMP_UGT:
2060 return AArch64CC::HI;
2061 case CmpInst::FCMP_OLT:
2062 return AArch64CC::MI;
2063 case CmpInst::ICMP_ULE:
2064 case CmpInst::FCMP_OLE:
2065 return AArch64CC::LS;
2066 case CmpInst::FCMP_ORD:
2067 return AArch64CC::VC;
2068 case CmpInst::FCMP_UNO:
2069 return AArch64CC::VS;
2070 case CmpInst::FCMP_UGE:
2071 return AArch64CC::PL;
2072 case CmpInst::ICMP_SLT:
2073 case CmpInst::FCMP_ULT:
2074 return AArch64CC::LT;
2075 case CmpInst::ICMP_SLE:
2076 case CmpInst::FCMP_ULE:
2077 return AArch64CC::LE;
2078 case CmpInst::FCMP_UNE:
2079 case CmpInst::ICMP_NE:
2080 return AArch64CC::NE;
2081 case CmpInst::ICMP_UGE:
2082 return AArch64CC::HS;
2083 case CmpInst::ICMP_ULT:
2084 return AArch64CC::LO;
2088 /// \brief Try to emit a combined compare-and-branch instruction.
2089 bool AArch64FastISel::emitCompareAndBranch(const BranchInst *BI) {
2090 assert(isa<CmpInst>(BI->getCondition()) && "Expected cmp instruction");
2091 const CmpInst *CI = cast<CmpInst>(BI->getCondition());
2092 CmpInst::Predicate Predicate = optimizeCmpPredicate(CI);
2094 const Value *LHS = CI->getOperand(0);
2095 const Value *RHS = CI->getOperand(1);
2098 if (!isTypeSupported(LHS->getType(), VT))
2101 unsigned BW = VT.getSizeInBits();
2105 MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
2106 MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
2108 // Try to take advantage of fallthrough opportunities.
2109 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
2110 std::swap(TBB, FBB);
2111 Predicate = CmpInst::getInversePredicate(Predicate);
2116 switch (Predicate) {
2119 case CmpInst::ICMP_EQ:
2120 case CmpInst::ICMP_NE:
2121 if (isa<Constant>(LHS) && cast<Constant>(LHS)->isNullValue())
2122 std::swap(LHS, RHS);
2124 if (!isa<Constant>(RHS) || !cast<Constant>(RHS)->isNullValue())
2127 if (const auto *AI = dyn_cast<BinaryOperator>(LHS))
2128 if (AI->getOpcode() == Instruction::And && isValueAvailable(AI)) {
2129 const Value *AndLHS = AI->getOperand(0);
2130 const Value *AndRHS = AI->getOperand(1);
2132 if (const auto *C = dyn_cast<ConstantInt>(AndLHS))
2133 if (C->getValue().isPowerOf2())
2134 std::swap(AndLHS, AndRHS);
2136 if (const auto *C = dyn_cast<ConstantInt>(AndRHS))
2137 if (C->getValue().isPowerOf2()) {
2138 TestBit = C->getValue().logBase2();
2146 IsCmpNE = Predicate == CmpInst::ICMP_NE;
2148 case CmpInst::ICMP_SLT:
2149 case CmpInst::ICMP_SGE:
2150 if (!isa<Constant>(RHS) || !cast<Constant>(RHS)->isNullValue())
2154 IsCmpNE = Predicate == CmpInst::ICMP_SLT;
2156 case CmpInst::ICMP_SGT:
2157 case CmpInst::ICMP_SLE:
2158 if (!isa<ConstantInt>(RHS))
2161 if (cast<ConstantInt>(RHS)->getValue() != APInt(BW, -1, true))
2165 IsCmpNE = Predicate == CmpInst::ICMP_SLE;
2169 static const unsigned OpcTable[2][2][2] = {
2170 { {AArch64::CBZW, AArch64::CBZX },
2171 {AArch64::CBNZW, AArch64::CBNZX} },
2172 { {AArch64::TBZW, AArch64::TBZX },
2173 {AArch64::TBNZW, AArch64::TBNZX} }
2176 bool IsBitTest = TestBit != -1;
2177 bool Is64Bit = BW == 64;
2178 if (TestBit < 32 && TestBit >= 0)
2181 unsigned Opc = OpcTable[IsBitTest][IsCmpNE][Is64Bit];
2182 const MCInstrDesc &II = TII.get(Opc);
2184 unsigned SrcReg = getRegForValue(LHS);
2187 bool SrcIsKill = hasTrivialKill(LHS);
2189 if (BW == 64 && !Is64Bit)
2190 SrcReg = fastEmitInst_extractsubreg(MVT::i32, SrcReg, SrcIsKill,
2193 if ((BW < 32) && !IsBitTest)
2194 SrcReg = emitIntExt(VT, SrcReg, MVT::i32, /*IsZExt=*/true);
2196 // Emit the combined compare and branch instruction.
2197 SrcReg = constrainOperandRegClass(II, SrcReg, II.getNumDefs());
2198 MachineInstrBuilder MIB =
2199 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc))
2200 .addReg(SrcReg, getKillRegState(SrcIsKill));
2202 MIB.addImm(TestBit);
2205 // Obtain the branch weight and add the TrueBB to the successor list.
2206 uint32_t BranchWeight = 0;
2208 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
2209 TBB->getBasicBlock());
2210 FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
2211 fastEmitBranch(FBB, DbgLoc);
2216 bool AArch64FastISel::selectBranch(const Instruction *I) {
2217 const BranchInst *BI = cast<BranchInst>(I);
2218 if (BI->isUnconditional()) {
2219 MachineBasicBlock *MSucc = FuncInfo.MBBMap[BI->getSuccessor(0)];
2220 fastEmitBranch(MSucc, BI->getDebugLoc());
2224 MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
2225 MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
2227 AArch64CC::CondCode CC = AArch64CC::NE;
2228 if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
2229 if (CI->hasOneUse() && isValueAvailable(CI)) {
2230 // Try to optimize or fold the cmp.
2231 CmpInst::Predicate Predicate = optimizeCmpPredicate(CI);
2232 switch (Predicate) {
2235 case CmpInst::FCMP_FALSE:
2236 fastEmitBranch(FBB, DbgLoc);
2238 case CmpInst::FCMP_TRUE:
2239 fastEmitBranch(TBB, DbgLoc);
2243 // Try to emit a combined compare-and-branch first.
2244 if (emitCompareAndBranch(BI))
2247 // Try to take advantage of fallthrough opportunities.
2248 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
2249 std::swap(TBB, FBB);
2250 Predicate = CmpInst::getInversePredicate(Predicate);
2254 if (!emitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
2257 // FCMP_UEQ and FCMP_ONE cannot be checked with a single branch
2259 CC = getCompareCC(Predicate);
2260 AArch64CC::CondCode ExtraCC = AArch64CC::AL;
2261 switch (Predicate) {
2264 case CmpInst::FCMP_UEQ:
2265 ExtraCC = AArch64CC::EQ;
2268 case CmpInst::FCMP_ONE:
2269 ExtraCC = AArch64CC::MI;
2273 assert((CC != AArch64CC::AL) && "Unexpected condition code.");
2275 // Emit the extra branch for FCMP_UEQ and FCMP_ONE.
2276 if (ExtraCC != AArch64CC::AL) {
2277 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
2283 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
2287 // Obtain the branch weight and add the TrueBB to the successor list.
2288 uint32_t BranchWeight = 0;
2290 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
2291 TBB->getBasicBlock());
2292 FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
2294 fastEmitBranch(FBB, DbgLoc);
2297 } else if (TruncInst *TI = dyn_cast<TruncInst>(BI->getCondition())) {
2299 if (TI->hasOneUse() && isValueAvailable(TI) &&
2300 isTypeSupported(TI->getOperand(0)->getType(), SrcVT)) {
2301 unsigned CondReg = getRegForValue(TI->getOperand(0));
2304 bool CondIsKill = hasTrivialKill(TI->getOperand(0));
2306 // Issue an extract_subreg to get the lower 32-bits.
2307 if (SrcVT == MVT::i64) {
2308 CondReg = fastEmitInst_extractsubreg(MVT::i32, CondReg, CondIsKill,
2313 unsigned ANDReg = emitAnd_ri(MVT::i32, CondReg, CondIsKill, 1);
2314 assert(ANDReg && "Unexpected AND instruction emission failure.");
2315 emitICmp_ri(MVT::i32, ANDReg, /*IsKill=*/true, 0);
2317 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
2318 std::swap(TBB, FBB);
2321 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
2325 // Obtain the branch weight and add the TrueBB to the successor list.
2326 uint32_t BranchWeight = 0;
2328 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
2329 TBB->getBasicBlock());
2330 FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
2332 fastEmitBranch(FBB, DbgLoc);
2335 } else if (const auto *CI = dyn_cast<ConstantInt>(BI->getCondition())) {
2336 uint64_t Imm = CI->getZExtValue();
2337 MachineBasicBlock *Target = (Imm == 0) ? FBB : TBB;
2338 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::B))
2341 // Obtain the branch weight and add the target to the successor list.
2342 uint32_t BranchWeight = 0;
2344 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
2345 Target->getBasicBlock());
2346 FuncInfo.MBB->addSuccessor(Target, BranchWeight);
2348 } else if (foldXALUIntrinsic(CC, I, BI->getCondition())) {
2349 // Fake request the condition, otherwise the intrinsic might be completely
2351 unsigned CondReg = getRegForValue(BI->getCondition());
2356 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
2360 // Obtain the branch weight and add the TrueBB to the successor list.
2361 uint32_t BranchWeight = 0;
2363 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
2364 TBB->getBasicBlock());
2365 FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
2367 fastEmitBranch(FBB, DbgLoc);
2371 unsigned CondReg = getRegForValue(BI->getCondition());
2374 bool CondRegIsKill = hasTrivialKill(BI->getCondition());
2376 // We've been divorced from our compare! Our block was split, and
2377 // now our compare lives in a predecessor block. We musn't
2378 // re-compare here, as the children of the compare aren't guaranteed
2379 // live across the block boundary (we *could* check for this).
2380 // Regardless, the compare has been done in the predecessor block,
2381 // and it left a value for us in a virtual register. Ergo, we test
2382 // the one-bit value left in the virtual register.
2383 emitICmp_ri(MVT::i32, CondReg, CondRegIsKill, 0);
2385 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
2386 std::swap(TBB, FBB);
2390 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
2394 // Obtain the branch weight and add the TrueBB to the successor list.
2395 uint32_t BranchWeight = 0;
2397 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
2398 TBB->getBasicBlock());
2399 FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
2401 fastEmitBranch(FBB, DbgLoc);
2405 bool AArch64FastISel::selectIndirectBr(const Instruction *I) {
2406 const IndirectBrInst *BI = cast<IndirectBrInst>(I);
2407 unsigned AddrReg = getRegForValue(BI->getOperand(0));
2411 // Emit the indirect branch.
2412 const MCInstrDesc &II = TII.get(AArch64::BR);
2413 AddrReg = constrainOperandRegClass(II, AddrReg, II.getNumDefs());
2414 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II).addReg(AddrReg);
2416 // Make sure the CFG is up-to-date.
2417 for (unsigned i = 0, e = BI->getNumSuccessors(); i != e; ++i)
2418 FuncInfo.MBB->addSuccessor(FuncInfo.MBBMap[BI->getSuccessor(i)]);
2423 bool AArch64FastISel::selectCmp(const Instruction *I) {
2424 const CmpInst *CI = cast<CmpInst>(I);
2426 // Try to optimize or fold the cmp.
2427 CmpInst::Predicate Predicate = optimizeCmpPredicate(CI);
2428 unsigned ResultReg = 0;
2429 switch (Predicate) {
2432 case CmpInst::FCMP_FALSE:
2433 ResultReg = createResultReg(&AArch64::GPR32RegClass);
2434 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2435 TII.get(TargetOpcode::COPY), ResultReg)
2436 .addReg(AArch64::WZR, getKillRegState(true));
2438 case CmpInst::FCMP_TRUE:
2439 ResultReg = fastEmit_i(MVT::i32, MVT::i32, ISD::Constant, 1);
2444 updateValueMap(I, ResultReg);
2449 if (!emitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
2452 ResultReg = createResultReg(&AArch64::GPR32RegClass);
2454 // FCMP_UEQ and FCMP_ONE cannot be checked with a single instruction. These
2455 // condition codes are inverted, because they are used by CSINC.
2456 static unsigned CondCodeTable[2][2] = {
2457 { AArch64CC::NE, AArch64CC::VC },
2458 { AArch64CC::PL, AArch64CC::LE }
2460 unsigned *CondCodes = nullptr;
2461 switch (Predicate) {
2464 case CmpInst::FCMP_UEQ:
2465 CondCodes = &CondCodeTable[0][0];
2467 case CmpInst::FCMP_ONE:
2468 CondCodes = &CondCodeTable[1][0];
2473 unsigned TmpReg1 = createResultReg(&AArch64::GPR32RegClass);
2474 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::CSINCWr),
2476 .addReg(AArch64::WZR, getKillRegState(true))
2477 .addReg(AArch64::WZR, getKillRegState(true))
2478 .addImm(CondCodes[0]);
2479 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::CSINCWr),
2481 .addReg(TmpReg1, getKillRegState(true))
2482 .addReg(AArch64::WZR, getKillRegState(true))
2483 .addImm(CondCodes[1]);
2485 updateValueMap(I, ResultReg);
2489 // Now set a register based on the comparison.
2490 AArch64CC::CondCode CC = getCompareCC(Predicate);
2491 assert((CC != AArch64CC::AL) && "Unexpected condition code.");
2492 AArch64CC::CondCode invertedCC = getInvertedCondCode(CC);
2493 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::CSINCWr),
2495 .addReg(AArch64::WZR, getKillRegState(true))
2496 .addReg(AArch64::WZR, getKillRegState(true))
2497 .addImm(invertedCC);
2499 updateValueMap(I, ResultReg);
2503 /// \brief Optimize selects of i1 if one of the operands has a 'true' or 'false'
2505 bool AArch64FastISel::optimizeSelect(const SelectInst *SI) {
2506 if (!SI->getType()->isIntegerTy(1))
2509 const Value *Src1Val, *Src2Val;
2511 bool NeedExtraOp = false;
2512 if (auto *CI = dyn_cast<ConstantInt>(SI->getTrueValue())) {
2514 Src1Val = SI->getCondition();
2515 Src2Val = SI->getFalseValue();
2516 Opc = AArch64::ORRWrr;
2518 assert(CI->isZero());
2519 Src1Val = SI->getFalseValue();
2520 Src2Val = SI->getCondition();
2521 Opc = AArch64::BICWrr;
2523 } else if (auto *CI = dyn_cast<ConstantInt>(SI->getFalseValue())) {
2525 Src1Val = SI->getCondition();
2526 Src2Val = SI->getTrueValue();
2527 Opc = AArch64::ORRWrr;
2530 assert(CI->isZero());
2531 Src1Val = SI->getCondition();
2532 Src2Val = SI->getTrueValue();
2533 Opc = AArch64::ANDWrr;
2540 unsigned Src1Reg = getRegForValue(Src1Val);
2543 bool Src1IsKill = hasTrivialKill(Src1Val);
2545 unsigned Src2Reg = getRegForValue(Src2Val);
2548 bool Src2IsKill = hasTrivialKill(Src2Val);
2551 Src1Reg = emitLogicalOp_ri(ISD::XOR, MVT::i32, Src1Reg, Src1IsKill, 1);
2554 unsigned ResultReg = fastEmitInst_rr(Opc, &AArch64::GPR32spRegClass, Src1Reg,
2555 Src1IsKill, Src2Reg, Src2IsKill);
2556 updateValueMap(SI, ResultReg);
2560 bool AArch64FastISel::selectSelect(const Instruction *I) {
2561 assert(isa<SelectInst>(I) && "Expected a select instruction.");
2563 if (!isTypeSupported(I->getType(), VT))
2567 const TargetRegisterClass *RC;
2568 switch (VT.SimpleTy) {
2575 Opc = AArch64::CSELWr;
2576 RC = &AArch64::GPR32RegClass;
2579 Opc = AArch64::CSELXr;
2580 RC = &AArch64::GPR64RegClass;
2583 Opc = AArch64::FCSELSrrr;
2584 RC = &AArch64::FPR32RegClass;
2587 Opc = AArch64::FCSELDrrr;
2588 RC = &AArch64::FPR64RegClass;
2592 const SelectInst *SI = cast<SelectInst>(I);
2593 const Value *Cond = SI->getCondition();
2594 AArch64CC::CondCode CC = AArch64CC::NE;
2595 AArch64CC::CondCode ExtraCC = AArch64CC::AL;
2597 if (optimizeSelect(SI))
2600 // Try to pickup the flags, so we don't have to emit another compare.
2601 if (foldXALUIntrinsic(CC, I, Cond)) {
2602 // Fake request the condition to force emission of the XALU intrinsic.
2603 unsigned CondReg = getRegForValue(Cond);
2606 } else if (isa<CmpInst>(Cond) && cast<CmpInst>(Cond)->hasOneUse() &&
2607 isValueAvailable(Cond)) {
2608 const auto *Cmp = cast<CmpInst>(Cond);
2609 // Try to optimize or fold the cmp.
2610 CmpInst::Predicate Predicate = optimizeCmpPredicate(Cmp);
2611 const Value *FoldSelect = nullptr;
2612 switch (Predicate) {
2615 case CmpInst::FCMP_FALSE:
2616 FoldSelect = SI->getFalseValue();
2618 case CmpInst::FCMP_TRUE:
2619 FoldSelect = SI->getTrueValue();
2624 unsigned SrcReg = getRegForValue(FoldSelect);
2627 unsigned UseReg = lookUpRegForValue(SI);
2629 MRI.clearKillFlags(UseReg);
2631 updateValueMap(I, SrcReg);
2636 if (!emitCmp(Cmp->getOperand(0), Cmp->getOperand(1), Cmp->isUnsigned()))
2639 // FCMP_UEQ and FCMP_ONE cannot be checked with a single select instruction.
2640 CC = getCompareCC(Predicate);
2641 switch (Predicate) {
2644 case CmpInst::FCMP_UEQ:
2645 ExtraCC = AArch64CC::EQ;
2648 case CmpInst::FCMP_ONE:
2649 ExtraCC = AArch64CC::MI;
2653 assert((CC != AArch64CC::AL) && "Unexpected condition code.");
2655 unsigned CondReg = getRegForValue(Cond);
2658 bool CondIsKill = hasTrivialKill(Cond);
2660 // Emit a TST instruction (ANDS wzr, reg, #imm).
2661 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ANDSWri),
2663 .addReg(CondReg, getKillRegState(CondIsKill))
2664 .addImm(AArch64_AM::encodeLogicalImmediate(1, 32));
2667 unsigned Src1Reg = getRegForValue(SI->getTrueValue());
2668 bool Src1IsKill = hasTrivialKill(SI->getTrueValue());
2670 unsigned Src2Reg = getRegForValue(SI->getFalseValue());
2671 bool Src2IsKill = hasTrivialKill(SI->getFalseValue());
2673 if (!Src1Reg || !Src2Reg)
2676 if (ExtraCC != AArch64CC::AL) {
2677 Src2Reg = fastEmitInst_rri(Opc, RC, Src1Reg, Src1IsKill, Src2Reg,
2678 Src2IsKill, ExtraCC);
2681 unsigned ResultReg = fastEmitInst_rri(Opc, RC, Src1Reg, Src1IsKill, Src2Reg,
2683 updateValueMap(I, ResultReg);
2687 bool AArch64FastISel::selectFPExt(const Instruction *I) {
2688 Value *V = I->getOperand(0);
2689 if (!I->getType()->isDoubleTy() || !V->getType()->isFloatTy())
2692 unsigned Op = getRegForValue(V);
2696 unsigned ResultReg = createResultReg(&AArch64::FPR64RegClass);
2697 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::FCVTDSr),
2698 ResultReg).addReg(Op);
2699 updateValueMap(I, ResultReg);
2703 bool AArch64FastISel::selectFPTrunc(const Instruction *I) {
2704 Value *V = I->getOperand(0);
2705 if (!I->getType()->isFloatTy() || !V->getType()->isDoubleTy())
2708 unsigned Op = getRegForValue(V);
2712 unsigned ResultReg = createResultReg(&AArch64::FPR32RegClass);
2713 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::FCVTSDr),
2714 ResultReg).addReg(Op);
2715 updateValueMap(I, ResultReg);
2719 // FPToUI and FPToSI
2720 bool AArch64FastISel::selectFPToInt(const Instruction *I, bool Signed) {
2722 if (!isTypeLegal(I->getType(), DestVT) || DestVT.isVector())
2725 unsigned SrcReg = getRegForValue(I->getOperand(0));
2729 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType(), true);
2730 if (SrcVT == MVT::f128)
2734 if (SrcVT == MVT::f64) {
2736 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZSUWDr : AArch64::FCVTZSUXDr;
2738 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZUUWDr : AArch64::FCVTZUUXDr;
2741 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZSUWSr : AArch64::FCVTZSUXSr;
2743 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZUUWSr : AArch64::FCVTZUUXSr;
2745 unsigned ResultReg = createResultReg(
2746 DestVT == MVT::i32 ? &AArch64::GPR32RegClass : &AArch64::GPR64RegClass);
2747 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
2749 updateValueMap(I, ResultReg);
2753 bool AArch64FastISel::selectIntToFP(const Instruction *I, bool Signed) {
2755 if (!isTypeLegal(I->getType(), DestVT) || DestVT.isVector())
2757 assert ((DestVT == MVT::f32 || DestVT == MVT::f64) &&
2758 "Unexpected value type.");
2760 unsigned SrcReg = getRegForValue(I->getOperand(0));
2763 bool SrcIsKill = hasTrivialKill(I->getOperand(0));
2765 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType(), true);
2767 // Handle sign-extension.
2768 if (SrcVT == MVT::i16 || SrcVT == MVT::i8 || SrcVT == MVT::i1) {
2770 emitIntExt(SrcVT.getSimpleVT(), SrcReg, MVT::i32, /*isZExt*/ !Signed);
2777 if (SrcVT == MVT::i64) {
2779 Opc = (DestVT == MVT::f32) ? AArch64::SCVTFUXSri : AArch64::SCVTFUXDri;
2781 Opc = (DestVT == MVT::f32) ? AArch64::UCVTFUXSri : AArch64::UCVTFUXDri;
2784 Opc = (DestVT == MVT::f32) ? AArch64::SCVTFUWSri : AArch64::SCVTFUWDri;
2786 Opc = (DestVT == MVT::f32) ? AArch64::UCVTFUWSri : AArch64::UCVTFUWDri;
2789 unsigned ResultReg = fastEmitInst_r(Opc, TLI.getRegClassFor(DestVT), SrcReg,
2791 updateValueMap(I, ResultReg);
2795 bool AArch64FastISel::fastLowerArguments() {
2796 if (!FuncInfo.CanLowerReturn)
2799 const Function *F = FuncInfo.Fn;
2803 CallingConv::ID CC = F->getCallingConv();
2804 if (CC != CallingConv::C)
2807 // Only handle simple cases of up to 8 GPR and FPR each.
2808 unsigned GPRCnt = 0;
2809 unsigned FPRCnt = 0;
2811 for (auto const &Arg : F->args()) {
2812 // The first argument is at index 1.
2814 if (F->getAttributes().hasAttribute(Idx, Attribute::ByVal) ||
2815 F->getAttributes().hasAttribute(Idx, Attribute::InReg) ||
2816 F->getAttributes().hasAttribute(Idx, Attribute::StructRet) ||
2817 F->getAttributes().hasAttribute(Idx, Attribute::Nest))
2820 Type *ArgTy = Arg.getType();
2821 if (ArgTy->isStructTy() || ArgTy->isArrayTy())
2824 EVT ArgVT = TLI.getValueType(ArgTy);
2825 if (!ArgVT.isSimple())
2828 MVT VT = ArgVT.getSimpleVT().SimpleTy;
2829 if (VT.isFloatingPoint() && !Subtarget->hasFPARMv8())
2832 if (VT.isVector() &&
2833 (!Subtarget->hasNEON() || !Subtarget->isLittleEndian()))
2836 if (VT >= MVT::i1 && VT <= MVT::i64)
2838 else if ((VT >= MVT::f16 && VT <= MVT::f64) || VT.is64BitVector() ||
2839 VT.is128BitVector())
2844 if (GPRCnt > 8 || FPRCnt > 8)
2848 static const MCPhysReg Registers[6][8] = {
2849 { AArch64::W0, AArch64::W1, AArch64::W2, AArch64::W3, AArch64::W4,
2850 AArch64::W5, AArch64::W6, AArch64::W7 },
2851 { AArch64::X0, AArch64::X1, AArch64::X2, AArch64::X3, AArch64::X4,
2852 AArch64::X5, AArch64::X6, AArch64::X7 },
2853 { AArch64::H0, AArch64::H1, AArch64::H2, AArch64::H3, AArch64::H4,
2854 AArch64::H5, AArch64::H6, AArch64::H7 },
2855 { AArch64::S0, AArch64::S1, AArch64::S2, AArch64::S3, AArch64::S4,
2856 AArch64::S5, AArch64::S6, AArch64::S7 },
2857 { AArch64::D0, AArch64::D1, AArch64::D2, AArch64::D3, AArch64::D4,
2858 AArch64::D5, AArch64::D6, AArch64::D7 },
2859 { AArch64::Q0, AArch64::Q1, AArch64::Q2, AArch64::Q3, AArch64::Q4,
2860 AArch64::Q5, AArch64::Q6, AArch64::Q7 }
2863 unsigned GPRIdx = 0;
2864 unsigned FPRIdx = 0;
2865 for (auto const &Arg : F->args()) {
2866 MVT VT = TLI.getSimpleValueType(Arg.getType());
2868 const TargetRegisterClass *RC;
2869 if (VT >= MVT::i1 && VT <= MVT::i32) {
2870 SrcReg = Registers[0][GPRIdx++];
2871 RC = &AArch64::GPR32RegClass;
2873 } else if (VT == MVT::i64) {
2874 SrcReg = Registers[1][GPRIdx++];
2875 RC = &AArch64::GPR64RegClass;
2876 } else if (VT == MVT::f16) {
2877 SrcReg = Registers[2][FPRIdx++];
2878 RC = &AArch64::FPR16RegClass;
2879 } else if (VT == MVT::f32) {
2880 SrcReg = Registers[3][FPRIdx++];
2881 RC = &AArch64::FPR32RegClass;
2882 } else if ((VT == MVT::f64) || VT.is64BitVector()) {
2883 SrcReg = Registers[4][FPRIdx++];
2884 RC = &AArch64::FPR64RegClass;
2885 } else if (VT.is128BitVector()) {
2886 SrcReg = Registers[5][FPRIdx++];
2887 RC = &AArch64::FPR128RegClass;
2889 llvm_unreachable("Unexpected value type.");
2891 unsigned DstReg = FuncInfo.MF->addLiveIn(SrcReg, RC);
2892 // FIXME: Unfortunately it's necessary to emit a copy from the livein copy.
2893 // Without this, EmitLiveInCopies may eliminate the livein if its only
2894 // use is a bitcast (which isn't turned into an instruction).
2895 unsigned ResultReg = createResultReg(RC);
2896 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2897 TII.get(TargetOpcode::COPY), ResultReg)
2898 .addReg(DstReg, getKillRegState(true));
2899 updateValueMap(&Arg, ResultReg);
2904 bool AArch64FastISel::processCallArgs(CallLoweringInfo &CLI,
2905 SmallVectorImpl<MVT> &OutVTs,
2906 unsigned &NumBytes) {
2907 CallingConv::ID CC = CLI.CallConv;
2908 SmallVector<CCValAssign, 16> ArgLocs;
2909 CCState CCInfo(CC, false, *FuncInfo.MF, ArgLocs, *Context);
2910 CCInfo.AnalyzeCallOperands(OutVTs, CLI.OutFlags, CCAssignFnForCall(CC));
2912 // Get a count of how many bytes are to be pushed on the stack.
2913 NumBytes = CCInfo.getNextStackOffset();
2915 // Issue CALLSEQ_START
2916 unsigned AdjStackDown = TII.getCallFrameSetupOpcode();
2917 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackDown))
2920 // Process the args.
2921 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2922 CCValAssign &VA = ArgLocs[i];
2923 const Value *ArgVal = CLI.OutVals[VA.getValNo()];
2924 MVT ArgVT = OutVTs[VA.getValNo()];
2926 unsigned ArgReg = getRegForValue(ArgVal);
2930 // Handle arg promotion: SExt, ZExt, AExt.
2931 switch (VA.getLocInfo()) {
2932 case CCValAssign::Full:
2934 case CCValAssign::SExt: {
2935 MVT DestVT = VA.getLocVT();
2937 ArgReg = emitIntExt(SrcVT, ArgReg, DestVT, /*isZExt=*/false);
2942 case CCValAssign::AExt:
2943 // Intentional fall-through.
2944 case CCValAssign::ZExt: {
2945 MVT DestVT = VA.getLocVT();
2947 ArgReg = emitIntExt(SrcVT, ArgReg, DestVT, /*isZExt=*/true);
2953 llvm_unreachable("Unknown arg promotion!");
2956 // Now copy/store arg to correct locations.
2957 if (VA.isRegLoc() && !VA.needsCustom()) {
2958 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2959 TII.get(TargetOpcode::COPY), VA.getLocReg()).addReg(ArgReg);
2960 CLI.OutRegs.push_back(VA.getLocReg());
2961 } else if (VA.needsCustom()) {
2962 // FIXME: Handle custom args.
2965 assert(VA.isMemLoc() && "Assuming store on stack.");
2967 // Don't emit stores for undef values.
2968 if (isa<UndefValue>(ArgVal))
2971 // Need to store on the stack.
2972 unsigned ArgSize = (ArgVT.getSizeInBits() + 7) / 8;
2974 unsigned BEAlign = 0;
2975 if (ArgSize < 8 && !Subtarget->isLittleEndian())
2976 BEAlign = 8 - ArgSize;
2979 Addr.setKind(Address::RegBase);
2980 Addr.setReg(AArch64::SP);
2981 Addr.setOffset(VA.getLocMemOffset() + BEAlign);
2983 unsigned Alignment = DL.getABITypeAlignment(ArgVal->getType());
2984 MachineMemOperand *MMO = FuncInfo.MF->getMachineMemOperand(
2985 MachinePointerInfo::getStack(Addr.getOffset()),
2986 MachineMemOperand::MOStore, ArgVT.getStoreSize(), Alignment);
2988 if (!emitStore(ArgVT, ArgReg, Addr, MMO))
2995 bool AArch64FastISel::finishCall(CallLoweringInfo &CLI, MVT RetVT,
2996 unsigned NumBytes) {
2997 CallingConv::ID CC = CLI.CallConv;
2999 // Issue CALLSEQ_END
3000 unsigned AdjStackUp = TII.getCallFrameDestroyOpcode();
3001 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackUp))
3002 .addImm(NumBytes).addImm(0);
3004 // Now the return value.
3005 if (RetVT != MVT::isVoid) {
3006 SmallVector<CCValAssign, 16> RVLocs;
3007 CCState CCInfo(CC, false, *FuncInfo.MF, RVLocs, *Context);
3008 CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC));
3010 // Only handle a single return value.
3011 if (RVLocs.size() != 1)
3014 // Copy all of the result registers out of their specified physreg.
3015 MVT CopyVT = RVLocs[0].getValVT();
3016 unsigned ResultReg = createResultReg(TLI.getRegClassFor(CopyVT));
3017 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3018 TII.get(TargetOpcode::COPY), ResultReg)
3019 .addReg(RVLocs[0].getLocReg());
3020 CLI.InRegs.push_back(RVLocs[0].getLocReg());
3022 CLI.ResultReg = ResultReg;
3023 CLI.NumResultRegs = 1;
3029 bool AArch64FastISel::fastLowerCall(CallLoweringInfo &CLI) {
3030 CallingConv::ID CC = CLI.CallConv;
3031 bool IsTailCall = CLI.IsTailCall;
3032 bool IsVarArg = CLI.IsVarArg;
3033 const Value *Callee = CLI.Callee;
3034 const char *SymName = CLI.SymName;
3036 if (!Callee && !SymName)
3039 // Allow SelectionDAG isel to handle tail calls.
3043 CodeModel::Model CM = TM.getCodeModel();
3044 // Only support the small and large code model.
3045 if (CM != CodeModel::Small && CM != CodeModel::Large)
3048 // FIXME: Add large code model support for ELF.
3049 if (CM == CodeModel::Large && !Subtarget->isTargetMachO())
3052 // Let SDISel handle vararg functions.
3056 // FIXME: Only handle *simple* calls for now.
3058 if (CLI.RetTy->isVoidTy())
3059 RetVT = MVT::isVoid;
3060 else if (!isTypeLegal(CLI.RetTy, RetVT))
3063 for (auto Flag : CLI.OutFlags)
3064 if (Flag.isInReg() || Flag.isSRet() || Flag.isNest() || Flag.isByVal())
3067 // Set up the argument vectors.
3068 SmallVector<MVT, 16> OutVTs;
3069 OutVTs.reserve(CLI.OutVals.size());
3071 for (auto *Val : CLI.OutVals) {
3073 if (!isTypeLegal(Val->getType(), VT) &&
3074 !(VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16))
3077 // We don't handle vector parameters yet.
3078 if (VT.isVector() || VT.getSizeInBits() > 64)
3081 OutVTs.push_back(VT);
3085 if (Callee && !computeCallAddress(Callee, Addr))
3088 // Handle the arguments now that we've gotten them.
3090 if (!processCallArgs(CLI, OutVTs, NumBytes))
3094 MachineInstrBuilder MIB;
3095 if (CM == CodeModel::Small) {
3096 const MCInstrDesc &II = TII.get(Addr.getReg() ? AArch64::BLR : AArch64::BL);
3097 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II);
3099 MIB.addExternalSymbol(SymName, 0);
3100 else if (Addr.getGlobalValue())
3101 MIB.addGlobalAddress(Addr.getGlobalValue(), 0, 0);
3102 else if (Addr.getReg()) {
3103 unsigned Reg = constrainOperandRegClass(II, Addr.getReg(), 0);
3108 unsigned CallReg = 0;
3110 unsigned ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
3111 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
3113 .addExternalSymbol(SymName, AArch64II::MO_GOT | AArch64II::MO_PAGE);
3115 CallReg = createResultReg(&AArch64::GPR64RegClass);
3116 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::LDRXui),
3119 .addExternalSymbol(SymName, AArch64II::MO_GOT | AArch64II::MO_PAGEOFF |
3121 } else if (Addr.getGlobalValue())
3122 CallReg = materializeGV(Addr.getGlobalValue());
3123 else if (Addr.getReg())
3124 CallReg = Addr.getReg();
3129 const MCInstrDesc &II = TII.get(AArch64::BLR);
3130 CallReg = constrainOperandRegClass(II, CallReg, 0);
3131 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II).addReg(CallReg);
3134 // Add implicit physical register uses to the call.
3135 for (auto Reg : CLI.OutRegs)
3136 MIB.addReg(Reg, RegState::Implicit);
3138 // Add a register mask with the call-preserved registers.
3139 // Proper defs for return values will be added by setPhysRegsDeadExcept().
3140 MIB.addRegMask(TRI.getCallPreservedMask(CC));
3144 // Finish off the call including any return values.
3145 return finishCall(CLI, RetVT, NumBytes);
3148 bool AArch64FastISel::isMemCpySmall(uint64_t Len, unsigned Alignment) {
3150 return Len / Alignment <= 4;
3155 bool AArch64FastISel::tryEmitSmallMemCpy(Address Dest, Address Src,
3156 uint64_t Len, unsigned Alignment) {
3157 // Make sure we don't bloat code by inlining very large memcpy's.
3158 if (!isMemCpySmall(Len, Alignment))
3161 int64_t UnscaledOffset = 0;
3162 Address OrigDest = Dest;
3163 Address OrigSrc = Src;
3167 if (!Alignment || Alignment >= 8) {
3178 // Bound based on alignment.
3179 if (Len >= 4 && Alignment == 4)
3181 else if (Len >= 2 && Alignment == 2)
3188 unsigned ResultReg = emitLoad(VT, VT, Src);
3192 if (!emitStore(VT, ResultReg, Dest))
3195 int64_t Size = VT.getSizeInBits() / 8;
3197 UnscaledOffset += Size;
3199 // We need to recompute the unscaled offset for each iteration.
3200 Dest.setOffset(OrigDest.getOffset() + UnscaledOffset);
3201 Src.setOffset(OrigSrc.getOffset() + UnscaledOffset);
3207 /// \brief Check if it is possible to fold the condition from the XALU intrinsic
3208 /// into the user. The condition code will only be updated on success.
3209 bool AArch64FastISel::foldXALUIntrinsic(AArch64CC::CondCode &CC,
3210 const Instruction *I,
3211 const Value *Cond) {
3212 if (!isa<ExtractValueInst>(Cond))
3215 const auto *EV = cast<ExtractValueInst>(Cond);
3216 if (!isa<IntrinsicInst>(EV->getAggregateOperand()))
3219 const auto *II = cast<IntrinsicInst>(EV->getAggregateOperand());
3221 const Function *Callee = II->getCalledFunction();
3223 cast<StructType>(Callee->getReturnType())->getTypeAtIndex(0U);
3224 if (!isTypeLegal(RetTy, RetVT))
3227 if (RetVT != MVT::i32 && RetVT != MVT::i64)
3230 const Value *LHS = II->getArgOperand(0);
3231 const Value *RHS = II->getArgOperand(1);
3233 // Canonicalize immediate to the RHS.
3234 if (isa<ConstantInt>(LHS) && !isa<ConstantInt>(RHS) &&
3235 isCommutativeIntrinsic(II))
3236 std::swap(LHS, RHS);
3238 // Simplify multiplies.
3239 unsigned IID = II->getIntrinsicID();
3243 case Intrinsic::smul_with_overflow:
3244 if (const auto *C = dyn_cast<ConstantInt>(RHS))
3245 if (C->getValue() == 2)
3246 IID = Intrinsic::sadd_with_overflow;
3248 case Intrinsic::umul_with_overflow:
3249 if (const auto *C = dyn_cast<ConstantInt>(RHS))
3250 if (C->getValue() == 2)
3251 IID = Intrinsic::uadd_with_overflow;
3255 AArch64CC::CondCode TmpCC;
3259 case Intrinsic::sadd_with_overflow:
3260 case Intrinsic::ssub_with_overflow:
3261 TmpCC = AArch64CC::VS;
3263 case Intrinsic::uadd_with_overflow:
3264 TmpCC = AArch64CC::HS;
3266 case Intrinsic::usub_with_overflow:
3267 TmpCC = AArch64CC::LO;
3269 case Intrinsic::smul_with_overflow:
3270 case Intrinsic::umul_with_overflow:
3271 TmpCC = AArch64CC::NE;
3275 // Check if both instructions are in the same basic block.
3276 if (!isValueAvailable(II))
3279 // Make sure nothing is in the way
3280 BasicBlock::const_iterator Start = I;
3281 BasicBlock::const_iterator End = II;
3282 for (auto Itr = std::prev(Start); Itr != End; --Itr) {
3283 // We only expect extractvalue instructions between the intrinsic and the
3284 // instruction to be selected.
3285 if (!isa<ExtractValueInst>(Itr))
3288 // Check that the extractvalue operand comes from the intrinsic.
3289 const auto *EVI = cast<ExtractValueInst>(Itr);
3290 if (EVI->getAggregateOperand() != II)
3298 bool AArch64FastISel::fastLowerIntrinsicCall(const IntrinsicInst *II) {
3299 // FIXME: Handle more intrinsics.
3300 switch (II->getIntrinsicID()) {
3301 default: return false;
3302 case Intrinsic::frameaddress: {
3303 MachineFrameInfo *MFI = FuncInfo.MF->getFrameInfo();
3304 MFI->setFrameAddressIsTaken(true);
3306 const AArch64RegisterInfo *RegInfo =
3307 static_cast<const AArch64RegisterInfo *>(
3308 TM.getSubtargetImpl()->getRegisterInfo());
3309 unsigned FramePtr = RegInfo->getFrameRegister(*(FuncInfo.MF));
3310 unsigned SrcReg = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
3311 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3312 TII.get(TargetOpcode::COPY), SrcReg).addReg(FramePtr);
3313 // Recursively load frame address
3319 unsigned Depth = cast<ConstantInt>(II->getOperand(0))->getZExtValue();
3321 DestReg = fastEmitInst_ri(AArch64::LDRXui, &AArch64::GPR64RegClass,
3322 SrcReg, /*IsKill=*/true, 0);
3323 assert(DestReg && "Unexpected LDR instruction emission failure.");
3327 updateValueMap(II, SrcReg);
3330 case Intrinsic::memcpy:
3331 case Intrinsic::memmove: {
3332 const auto *MTI = cast<MemTransferInst>(II);
3333 // Don't handle volatile.
3334 if (MTI->isVolatile())
3337 // Disable inlining for memmove before calls to ComputeAddress. Otherwise,
3338 // we would emit dead code because we don't currently handle memmoves.
3339 bool IsMemCpy = (II->getIntrinsicID() == Intrinsic::memcpy);
3340 if (isa<ConstantInt>(MTI->getLength()) && IsMemCpy) {
3341 // Small memcpy's are common enough that we want to do them without a call
3343 uint64_t Len = cast<ConstantInt>(MTI->getLength())->getZExtValue();
3344 unsigned Alignment = MTI->getAlignment();
3345 if (isMemCpySmall(Len, Alignment)) {
3347 if (!computeAddress(MTI->getRawDest(), Dest) ||
3348 !computeAddress(MTI->getRawSource(), Src))
3350 if (tryEmitSmallMemCpy(Dest, Src, Len, Alignment))
3355 if (!MTI->getLength()->getType()->isIntegerTy(64))
3358 if (MTI->getSourceAddressSpace() > 255 || MTI->getDestAddressSpace() > 255)
3359 // Fast instruction selection doesn't support the special
3363 const char *IntrMemName = isa<MemCpyInst>(II) ? "memcpy" : "memmove";
3364 return lowerCallTo(II, IntrMemName, II->getNumArgOperands() - 2);
3366 case Intrinsic::memset: {
3367 const MemSetInst *MSI = cast<MemSetInst>(II);
3368 // Don't handle volatile.
3369 if (MSI->isVolatile())
3372 if (!MSI->getLength()->getType()->isIntegerTy(64))
3375 if (MSI->getDestAddressSpace() > 255)
3376 // Fast instruction selection doesn't support the special
3380 return lowerCallTo(II, "memset", II->getNumArgOperands() - 2);
3382 case Intrinsic::sin:
3383 case Intrinsic::cos:
3384 case Intrinsic::pow: {
3386 if (!isTypeLegal(II->getType(), RetVT))
3389 if (RetVT != MVT::f32 && RetVT != MVT::f64)
3392 static const RTLIB::Libcall LibCallTable[3][2] = {
3393 { RTLIB::SIN_F32, RTLIB::SIN_F64 },
3394 { RTLIB::COS_F32, RTLIB::COS_F64 },
3395 { RTLIB::POW_F32, RTLIB::POW_F64 }
3398 bool Is64Bit = RetVT == MVT::f64;
3399 switch (II->getIntrinsicID()) {
3401 llvm_unreachable("Unexpected intrinsic.");
3402 case Intrinsic::sin:
3403 LC = LibCallTable[0][Is64Bit];
3405 case Intrinsic::cos:
3406 LC = LibCallTable[1][Is64Bit];
3408 case Intrinsic::pow:
3409 LC = LibCallTable[2][Is64Bit];
3414 Args.reserve(II->getNumArgOperands());
3416 // Populate the argument list.
3417 for (auto &Arg : II->arg_operands()) {
3420 Entry.Ty = Arg->getType();
3421 Args.push_back(Entry);
3424 CallLoweringInfo CLI;
3425 CLI.setCallee(TLI.getLibcallCallingConv(LC), II->getType(),
3426 TLI.getLibcallName(LC), std::move(Args));
3427 if (!lowerCallTo(CLI))
3429 updateValueMap(II, CLI.ResultReg);
3432 case Intrinsic::fabs: {
3434 if (!isTypeLegal(II->getType(), VT))
3438 switch (VT.SimpleTy) {
3442 Opc = AArch64::FABSSr;
3445 Opc = AArch64::FABSDr;
3448 unsigned SrcReg = getRegForValue(II->getOperand(0));
3451 bool SrcRegIsKill = hasTrivialKill(II->getOperand(0));
3452 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
3453 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
3454 .addReg(SrcReg, getKillRegState(SrcRegIsKill));
3455 updateValueMap(II, ResultReg);
3458 case Intrinsic::trap: {
3459 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::BRK))
3463 case Intrinsic::sqrt: {
3464 Type *RetTy = II->getCalledFunction()->getReturnType();
3467 if (!isTypeLegal(RetTy, VT))
3470 unsigned Op0Reg = getRegForValue(II->getOperand(0));
3473 bool Op0IsKill = hasTrivialKill(II->getOperand(0));
3475 unsigned ResultReg = fastEmit_r(VT, VT, ISD::FSQRT, Op0Reg, Op0IsKill);
3479 updateValueMap(II, ResultReg);
3482 case Intrinsic::sadd_with_overflow:
3483 case Intrinsic::uadd_with_overflow:
3484 case Intrinsic::ssub_with_overflow:
3485 case Intrinsic::usub_with_overflow:
3486 case Intrinsic::smul_with_overflow:
3487 case Intrinsic::umul_with_overflow: {
3488 // This implements the basic lowering of the xalu with overflow intrinsics.
3489 const Function *Callee = II->getCalledFunction();
3490 auto *Ty = cast<StructType>(Callee->getReturnType());
3491 Type *RetTy = Ty->getTypeAtIndex(0U);
3494 if (!isTypeLegal(RetTy, VT))
3497 if (VT != MVT::i32 && VT != MVT::i64)
3500 const Value *LHS = II->getArgOperand(0);
3501 const Value *RHS = II->getArgOperand(1);
3502 // Canonicalize immediate to the RHS.
3503 if (isa<ConstantInt>(LHS) && !isa<ConstantInt>(RHS) &&
3504 isCommutativeIntrinsic(II))
3505 std::swap(LHS, RHS);
3507 // Simplify multiplies.
3508 unsigned IID = II->getIntrinsicID();
3512 case Intrinsic::smul_with_overflow:
3513 if (const auto *C = dyn_cast<ConstantInt>(RHS))
3514 if (C->getValue() == 2) {
3515 IID = Intrinsic::sadd_with_overflow;
3519 case Intrinsic::umul_with_overflow:
3520 if (const auto *C = dyn_cast<ConstantInt>(RHS))
3521 if (C->getValue() == 2) {
3522 IID = Intrinsic::uadd_with_overflow;
3528 unsigned ResultReg1 = 0, ResultReg2 = 0, MulReg = 0;
3529 AArch64CC::CondCode CC = AArch64CC::Invalid;
3531 default: llvm_unreachable("Unexpected intrinsic!");
3532 case Intrinsic::sadd_with_overflow:
3533 ResultReg1 = emitAdd(VT, LHS, RHS, /*SetFlags=*/true);
3536 case Intrinsic::uadd_with_overflow:
3537 ResultReg1 = emitAdd(VT, LHS, RHS, /*SetFlags=*/true);
3540 case Intrinsic::ssub_with_overflow:
3541 ResultReg1 = emitSub(VT, LHS, RHS, /*SetFlags=*/true);
3544 case Intrinsic::usub_with_overflow:
3545 ResultReg1 = emitSub(VT, LHS, RHS, /*SetFlags=*/true);
3548 case Intrinsic::smul_with_overflow: {
3550 unsigned LHSReg = getRegForValue(LHS);
3553 bool LHSIsKill = hasTrivialKill(LHS);
3555 unsigned RHSReg = getRegForValue(RHS);
3558 bool RHSIsKill = hasTrivialKill(RHS);
3560 if (VT == MVT::i32) {
3561 MulReg = emitSMULL_rr(MVT::i64, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
3562 unsigned ShiftReg = emitLSR_ri(MVT::i64, MVT::i64, MulReg,
3563 /*IsKill=*/false, 32);
3564 MulReg = fastEmitInst_extractsubreg(VT, MulReg, /*IsKill=*/true,
3566 ShiftReg = fastEmitInst_extractsubreg(VT, ShiftReg, /*IsKill=*/true,
3568 emitSubs_rs(VT, ShiftReg, /*IsKill=*/true, MulReg, /*IsKill=*/false,
3569 AArch64_AM::ASR, 31, /*WantResult=*/false);
3571 assert(VT == MVT::i64 && "Unexpected value type.");
3572 MulReg = emitMul_rr(VT, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
3573 unsigned SMULHReg = fastEmit_rr(VT, VT, ISD::MULHS, LHSReg, LHSIsKill,
3575 emitSubs_rs(VT, SMULHReg, /*IsKill=*/true, MulReg, /*IsKill=*/false,
3576 AArch64_AM::ASR, 63, /*WantResult=*/false);
3580 case Intrinsic::umul_with_overflow: {
3582 unsigned LHSReg = getRegForValue(LHS);
3585 bool LHSIsKill = hasTrivialKill(LHS);
3587 unsigned RHSReg = getRegForValue(RHS);
3590 bool RHSIsKill = hasTrivialKill(RHS);
3592 if (VT == MVT::i32) {
3593 MulReg = emitUMULL_rr(MVT::i64, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
3594 emitSubs_rs(MVT::i64, AArch64::XZR, /*IsKill=*/true, MulReg,
3595 /*IsKill=*/false, AArch64_AM::LSR, 32,
3596 /*WantResult=*/false);
3597 MulReg = fastEmitInst_extractsubreg(VT, MulReg, /*IsKill=*/true,
3600 assert(VT == MVT::i64 && "Unexpected value type.");
3601 MulReg = emitMul_rr(VT, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
3602 unsigned UMULHReg = fastEmit_rr(VT, VT, ISD::MULHU, LHSReg, LHSIsKill,
3604 emitSubs_rr(VT, AArch64::XZR, /*IsKill=*/true, UMULHReg,
3605 /*IsKill=*/false, /*WantResult=*/false);
3612 ResultReg1 = createResultReg(TLI.getRegClassFor(VT));
3613 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3614 TII.get(TargetOpcode::COPY), ResultReg1).addReg(MulReg);
3617 ResultReg2 = fastEmitInst_rri(AArch64::CSINCWr, &AArch64::GPR32RegClass,
3618 AArch64::WZR, /*IsKill=*/true, AArch64::WZR,
3619 /*IsKill=*/true, getInvertedCondCode(CC));
3621 assert((ResultReg1 + 1) == ResultReg2 &&
3622 "Nonconsecutive result registers.");
3623 updateValueMap(II, ResultReg1, 2);
3630 bool AArch64FastISel::selectRet(const Instruction *I) {
3631 const ReturnInst *Ret = cast<ReturnInst>(I);
3632 const Function &F = *I->getParent()->getParent();
3634 if (!FuncInfo.CanLowerReturn)
3640 // Build a list of return value registers.
3641 SmallVector<unsigned, 4> RetRegs;
3643 if (Ret->getNumOperands() > 0) {
3644 CallingConv::ID CC = F.getCallingConv();
3645 SmallVector<ISD::OutputArg, 4> Outs;
3646 GetReturnInfo(F.getReturnType(), F.getAttributes(), Outs, TLI);
3648 // Analyze operands of the call, assigning locations to each operand.
3649 SmallVector<CCValAssign, 16> ValLocs;
3650 CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, ValLocs, I->getContext());
3651 CCAssignFn *RetCC = CC == CallingConv::WebKit_JS ? RetCC_AArch64_WebKit_JS
3652 : RetCC_AArch64_AAPCS;
3653 CCInfo.AnalyzeReturn(Outs, RetCC);
3655 // Only handle a single return value for now.
3656 if (ValLocs.size() != 1)
3659 CCValAssign &VA = ValLocs[0];
3660 const Value *RV = Ret->getOperand(0);
3662 // Don't bother handling odd stuff for now.
3663 if ((VA.getLocInfo() != CCValAssign::Full) &&
3664 (VA.getLocInfo() != CCValAssign::BCvt))
3667 // Only handle register returns for now.
3671 unsigned Reg = getRegForValue(RV);
3675 unsigned SrcReg = Reg + VA.getValNo();
3676 unsigned DestReg = VA.getLocReg();
3677 // Avoid a cross-class copy. This is very unlikely.
3678 if (!MRI.getRegClass(SrcReg)->contains(DestReg))
3681 EVT RVEVT = TLI.getValueType(RV->getType());
3682 if (!RVEVT.isSimple())
3685 // Vectors (of > 1 lane) in big endian need tricky handling.
3686 if (RVEVT.isVector() && RVEVT.getVectorNumElements() > 1 &&
3687 !Subtarget->isLittleEndian())
3690 MVT RVVT = RVEVT.getSimpleVT();
3691 if (RVVT == MVT::f128)
3694 MVT DestVT = VA.getValVT();
3695 // Special handling for extended integers.
3696 if (RVVT != DestVT) {
3697 if (RVVT != MVT::i1 && RVVT != MVT::i8 && RVVT != MVT::i16)
3700 if (!Outs[0].Flags.isZExt() && !Outs[0].Flags.isSExt())
3703 bool IsZExt = Outs[0].Flags.isZExt();
3704 SrcReg = emitIntExt(RVVT, SrcReg, DestVT, IsZExt);
3710 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3711 TII.get(TargetOpcode::COPY), DestReg).addReg(SrcReg);
3713 // Add register to return instruction.
3714 RetRegs.push_back(VA.getLocReg());
3717 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3718 TII.get(AArch64::RET_ReallyLR));
3719 for (unsigned i = 0, e = RetRegs.size(); i != e; ++i)
3720 MIB.addReg(RetRegs[i], RegState::Implicit);
3724 bool AArch64FastISel::selectTrunc(const Instruction *I) {
3725 Type *DestTy = I->getType();
3726 Value *Op = I->getOperand(0);
3727 Type *SrcTy = Op->getType();
3729 EVT SrcEVT = TLI.getValueType(SrcTy, true);
3730 EVT DestEVT = TLI.getValueType(DestTy, true);
3731 if (!SrcEVT.isSimple())
3733 if (!DestEVT.isSimple())
3736 MVT SrcVT = SrcEVT.getSimpleVT();
3737 MVT DestVT = DestEVT.getSimpleVT();
3739 if (SrcVT != MVT::i64 && SrcVT != MVT::i32 && SrcVT != MVT::i16 &&
3742 if (DestVT != MVT::i32 && DestVT != MVT::i16 && DestVT != MVT::i8 &&
3746 unsigned SrcReg = getRegForValue(Op);
3749 bool SrcIsKill = hasTrivialKill(Op);
3751 // If we're truncating from i64 to a smaller non-legal type then generate an
3752 // AND. Otherwise, we know the high bits are undefined and a truncate only
3753 // generate a COPY. We cannot mark the source register also as result
3754 // register, because this can incorrectly transfer the kill flag onto the
3757 if (SrcVT == MVT::i64) {
3759 switch (DestVT.SimpleTy) {
3761 // Trunc i64 to i32 is handled by the target-independent fast-isel.
3773 // Issue an extract_subreg to get the lower 32-bits.
3774 unsigned Reg32 = fastEmitInst_extractsubreg(MVT::i32, SrcReg, SrcIsKill,
3776 // Create the AND instruction which performs the actual truncation.
3777 ResultReg = emitAnd_ri(MVT::i32, Reg32, /*IsKill=*/true, Mask);
3778 assert(ResultReg && "Unexpected AND instruction emission failure.");
3780 ResultReg = createResultReg(&AArch64::GPR32RegClass);
3781 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3782 TII.get(TargetOpcode::COPY), ResultReg)
3783 .addReg(SrcReg, getKillRegState(SrcIsKill));
3786 updateValueMap(I, ResultReg);
3790 unsigned AArch64FastISel::emiti1Ext(unsigned SrcReg, MVT DestVT, bool IsZExt) {
3791 assert((DestVT == MVT::i8 || DestVT == MVT::i16 || DestVT == MVT::i32 ||
3792 DestVT == MVT::i64) &&
3793 "Unexpected value type.");
3794 // Handle i8 and i16 as i32.
3795 if (DestVT == MVT::i8 || DestVT == MVT::i16)
3799 unsigned ResultReg = emitAnd_ri(MVT::i32, SrcReg, /*TODO:IsKill=*/false, 1);
3800 assert(ResultReg && "Unexpected AND instruction emission failure.");
3801 if (DestVT == MVT::i64) {
3802 // We're ZExt i1 to i64. The ANDWri Wd, Ws, #1 implicitly clears the
3803 // upper 32 bits. Emit a SUBREG_TO_REG to extend from Wd to Xd.
3804 unsigned Reg64 = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
3805 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3806 TII.get(AArch64::SUBREG_TO_REG), Reg64)
3809 .addImm(AArch64::sub_32);
3814 if (DestVT == MVT::i64) {
3815 // FIXME: We're SExt i1 to i64.
3818 return fastEmitInst_rii(AArch64::SBFMWri, &AArch64::GPR32RegClass, SrcReg,
3819 /*TODO:IsKill=*/false, 0, 0);
3823 unsigned AArch64FastISel::emitMul_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
3824 unsigned Op1, bool Op1IsKill) {
3826 switch (RetVT.SimpleTy) {
3832 Opc = AArch64::MADDWrrr; ZReg = AArch64::WZR; break;
3834 Opc = AArch64::MADDXrrr; ZReg = AArch64::XZR; break;
3837 const TargetRegisterClass *RC =
3838 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3839 return fastEmitInst_rrr(Opc, RC, Op0, Op0IsKill, Op1, Op1IsKill,
3840 /*IsKill=*/ZReg, true);
3843 unsigned AArch64FastISel::emitSMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
3844 unsigned Op1, bool Op1IsKill) {
3845 if (RetVT != MVT::i64)
3848 return fastEmitInst_rrr(AArch64::SMADDLrrr, &AArch64::GPR64RegClass,
3849 Op0, Op0IsKill, Op1, Op1IsKill,
3850 AArch64::XZR, /*IsKill=*/true);
3853 unsigned AArch64FastISel::emitUMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
3854 unsigned Op1, bool Op1IsKill) {
3855 if (RetVT != MVT::i64)
3858 return fastEmitInst_rrr(AArch64::UMADDLrrr, &AArch64::GPR64RegClass,
3859 Op0, Op0IsKill, Op1, Op1IsKill,
3860 AArch64::XZR, /*IsKill=*/true);
3863 unsigned AArch64FastISel::emitLSL_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
3864 unsigned Op1Reg, bool Op1IsKill) {
3866 bool NeedTrunc = false;
3868 switch (RetVT.SimpleTy) {
3870 case MVT::i8: Opc = AArch64::LSLVWr; NeedTrunc = true; Mask = 0xff; break;
3871 case MVT::i16: Opc = AArch64::LSLVWr; NeedTrunc = true; Mask = 0xffff; break;
3872 case MVT::i32: Opc = AArch64::LSLVWr; break;
3873 case MVT::i64: Opc = AArch64::LSLVXr; break;
3876 const TargetRegisterClass *RC =
3877 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3879 Op1Reg = emitAnd_ri(MVT::i32, Op1Reg, Op1IsKill, Mask);
3882 unsigned ResultReg = fastEmitInst_rr(Opc, RC, Op0Reg, Op0IsKill, Op1Reg,
3885 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
3889 unsigned AArch64FastISel::emitLSL_ri(MVT RetVT, MVT SrcVT, unsigned Op0,
3890 bool Op0IsKill, uint64_t Shift,
3892 assert(RetVT.SimpleTy >= SrcVT.SimpleTy &&
3893 "Unexpected source/return type pair.");
3894 assert((SrcVT == MVT::i1 || SrcVT == MVT::i8 || SrcVT == MVT::i16 ||
3895 SrcVT == MVT::i32 || SrcVT == MVT::i64) &&
3896 "Unexpected source value type.");
3897 assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 ||
3898 RetVT == MVT::i64) && "Unexpected return value type.");
3900 bool Is64Bit = (RetVT == MVT::i64);
3901 unsigned RegSize = Is64Bit ? 64 : 32;
3902 unsigned DstBits = RetVT.getSizeInBits();
3903 unsigned SrcBits = SrcVT.getSizeInBits();
3904 const TargetRegisterClass *RC =
3905 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3907 // Just emit a copy for "zero" shifts.
3909 if (RetVT == SrcVT) {
3910 unsigned ResultReg = createResultReg(RC);
3911 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3912 TII.get(TargetOpcode::COPY), ResultReg)
3913 .addReg(Op0, getKillRegState(Op0IsKill));
3916 return emitIntExt(SrcVT, Op0, RetVT, IsZExt);
3919 // Don't deal with undefined shifts.
3920 if (Shift >= DstBits)
3923 // For immediate shifts we can fold the zero-/sign-extension into the shift.
3924 // {S|U}BFM Wd, Wn, #r, #s
3925 // Wd<32+s-r,32-r> = Wn<s:0> when r > s
3927 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3928 // %2 = shl i16 %1, 4
3929 // Wd<32+7-28,32-28> = Wn<7:0> <- clamp s to 7
3930 // 0b1111_1111_1111_1111__1111_1010_1010_0000 sext
3931 // 0b0000_0000_0000_0000__0000_0101_0101_0000 sext | zext
3932 // 0b0000_0000_0000_0000__0000_1010_1010_0000 zext
3934 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3935 // %2 = shl i16 %1, 8
3936 // Wd<32+7-24,32-24> = Wn<7:0>
3937 // 0b1111_1111_1111_1111__1010_1010_0000_0000 sext
3938 // 0b0000_0000_0000_0000__0101_0101_0000_0000 sext | zext
3939 // 0b0000_0000_0000_0000__1010_1010_0000_0000 zext
3941 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3942 // %2 = shl i16 %1, 12
3943 // Wd<32+3-20,32-20> = Wn<3:0>
3944 // 0b1111_1111_1111_1111__1010_0000_0000_0000 sext
3945 // 0b0000_0000_0000_0000__0101_0000_0000_0000 sext | zext
3946 // 0b0000_0000_0000_0000__1010_0000_0000_0000 zext
3948 unsigned ImmR = RegSize - Shift;
3949 // Limit the width to the length of the source type.
3950 unsigned ImmS = std::min<unsigned>(SrcBits - 1, DstBits - 1 - Shift);
3951 static const unsigned OpcTable[2][2] = {
3952 {AArch64::SBFMWri, AArch64::SBFMXri},
3953 {AArch64::UBFMWri, AArch64::UBFMXri}
3955 unsigned Opc = OpcTable[IsZExt][Is64Bit];
3956 if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) {
3957 unsigned TmpReg = MRI.createVirtualRegister(RC);
3958 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3959 TII.get(AArch64::SUBREG_TO_REG), TmpReg)
3961 .addReg(Op0, getKillRegState(Op0IsKill))
3962 .addImm(AArch64::sub_32);
3966 return fastEmitInst_rii(Opc, RC, Op0, Op0IsKill, ImmR, ImmS);
3969 unsigned AArch64FastISel::emitLSR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
3970 unsigned Op1Reg, bool Op1IsKill) {
3972 bool NeedTrunc = false;
3974 switch (RetVT.SimpleTy) {
3976 case MVT::i8: Opc = AArch64::LSRVWr; NeedTrunc = true; Mask = 0xff; break;
3977 case MVT::i16: Opc = AArch64::LSRVWr; NeedTrunc = true; Mask = 0xffff; break;
3978 case MVT::i32: Opc = AArch64::LSRVWr; break;
3979 case MVT::i64: Opc = AArch64::LSRVXr; break;
3982 const TargetRegisterClass *RC =
3983 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3985 Op0Reg = emitAnd_ri(MVT::i32, Op0Reg, Op0IsKill, Mask);
3986 Op1Reg = emitAnd_ri(MVT::i32, Op1Reg, Op1IsKill, Mask);
3987 Op0IsKill = Op1IsKill = true;
3989 unsigned ResultReg = fastEmitInst_rr(Opc, RC, Op0Reg, Op0IsKill, Op1Reg,
3992 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
3996 unsigned AArch64FastISel::emitLSR_ri(MVT RetVT, MVT SrcVT, unsigned Op0,
3997 bool Op0IsKill, uint64_t Shift,
3999 assert(RetVT.SimpleTy >= SrcVT.SimpleTy &&
4000 "Unexpected source/return type pair.");
4001 assert((SrcVT == MVT::i1 || SrcVT == MVT::i8 || SrcVT == MVT::i16 ||
4002 SrcVT == MVT::i32 || SrcVT == MVT::i64) &&
4003 "Unexpected source value type.");
4004 assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 ||
4005 RetVT == MVT::i64) && "Unexpected return value type.");
4007 bool Is64Bit = (RetVT == MVT::i64);
4008 unsigned RegSize = Is64Bit ? 64 : 32;
4009 unsigned DstBits = RetVT.getSizeInBits();
4010 unsigned SrcBits = SrcVT.getSizeInBits();
4011 const TargetRegisterClass *RC =
4012 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
4014 // Just emit a copy for "zero" shifts.
4016 if (RetVT == SrcVT) {
4017 unsigned ResultReg = createResultReg(RC);
4018 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
4019 TII.get(TargetOpcode::COPY), ResultReg)
4020 .addReg(Op0, getKillRegState(Op0IsKill));
4023 return emitIntExt(SrcVT, Op0, RetVT, IsZExt);
4026 // Don't deal with undefined shifts.
4027 if (Shift >= DstBits)
4030 // For immediate shifts we can fold the zero-/sign-extension into the shift.
4031 // {S|U}BFM Wd, Wn, #r, #s
4032 // Wd<s-r:0> = Wn<s:r> when r <= s
4034 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
4035 // %2 = lshr i16 %1, 4
4036 // Wd<7-4:0> = Wn<7:4>
4037 // 0b0000_0000_0000_0000__0000_1111_1111_1010 sext
4038 // 0b0000_0000_0000_0000__0000_0000_0000_0101 sext | zext
4039 // 0b0000_0000_0000_0000__0000_0000_0000_1010 zext
4041 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
4042 // %2 = lshr i16 %1, 8
4043 // Wd<7-7,0> = Wn<7:7>
4044 // 0b0000_0000_0000_0000__0000_0000_1111_1111 sext
4045 // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
4046 // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
4048 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
4049 // %2 = lshr i16 %1, 12
4050 // Wd<7-7,0> = Wn<7:7> <- clamp r to 7
4051 // 0b0000_0000_0000_0000__0000_0000_0000_1111 sext
4052 // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
4053 // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
4055 if (Shift >= SrcBits && IsZExt)
4056 return materializeInt(ConstantInt::get(*Context, APInt(RegSize, 0)), RetVT);
4058 // It is not possible to fold a sign-extend into the LShr instruction. In this
4059 // case emit a sign-extend.
4061 Op0 = emitIntExt(SrcVT, Op0, RetVT, IsZExt);
4066 SrcBits = SrcVT.getSizeInBits();
4070 unsigned ImmR = std::min<unsigned>(SrcBits - 1, Shift);
4071 unsigned ImmS = SrcBits - 1;
4072 static const unsigned OpcTable[2][2] = {
4073 {AArch64::SBFMWri, AArch64::SBFMXri},
4074 {AArch64::UBFMWri, AArch64::UBFMXri}
4076 unsigned Opc = OpcTable[IsZExt][Is64Bit];
4077 if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) {
4078 unsigned TmpReg = MRI.createVirtualRegister(RC);
4079 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
4080 TII.get(AArch64::SUBREG_TO_REG), TmpReg)
4082 .addReg(Op0, getKillRegState(Op0IsKill))
4083 .addImm(AArch64::sub_32);
4087 return fastEmitInst_rii(Opc, RC, Op0, Op0IsKill, ImmR, ImmS);
4090 unsigned AArch64FastISel::emitASR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
4091 unsigned Op1Reg, bool Op1IsKill) {
4093 bool NeedTrunc = false;
4095 switch (RetVT.SimpleTy) {
4097 case MVT::i8: Opc = AArch64::ASRVWr; NeedTrunc = true; Mask = 0xff; break;
4098 case MVT::i16: Opc = AArch64::ASRVWr; NeedTrunc = true; Mask = 0xffff; break;
4099 case MVT::i32: Opc = AArch64::ASRVWr; break;
4100 case MVT::i64: Opc = AArch64::ASRVXr; break;
4103 const TargetRegisterClass *RC =
4104 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
4106 Op0Reg = emitIntExt(RetVT, Op0Reg, MVT::i32, /*IsZExt=*/false);
4107 Op1Reg = emitAnd_ri(MVT::i32, Op1Reg, Op1IsKill, Mask);
4108 Op0IsKill = Op1IsKill = true;
4110 unsigned ResultReg = fastEmitInst_rr(Opc, RC, Op0Reg, Op0IsKill, Op1Reg,
4113 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
4117 unsigned AArch64FastISel::emitASR_ri(MVT RetVT, MVT SrcVT, unsigned Op0,
4118 bool Op0IsKill, uint64_t Shift,
4120 assert(RetVT.SimpleTy >= SrcVT.SimpleTy &&
4121 "Unexpected source/return type pair.");
4122 assert((SrcVT == MVT::i1 || SrcVT == MVT::i8 || SrcVT == MVT::i16 ||
4123 SrcVT == MVT::i32 || SrcVT == MVT::i64) &&
4124 "Unexpected source value type.");
4125 assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 ||
4126 RetVT == MVT::i64) && "Unexpected return value type.");
4128 bool Is64Bit = (RetVT == MVT::i64);
4129 unsigned RegSize = Is64Bit ? 64 : 32;
4130 unsigned DstBits = RetVT.getSizeInBits();
4131 unsigned SrcBits = SrcVT.getSizeInBits();
4132 const TargetRegisterClass *RC =
4133 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
4135 // Just emit a copy for "zero" shifts.
4137 if (RetVT == SrcVT) {
4138 unsigned ResultReg = createResultReg(RC);
4139 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
4140 TII.get(TargetOpcode::COPY), ResultReg)
4141 .addReg(Op0, getKillRegState(Op0IsKill));
4144 return emitIntExt(SrcVT, Op0, RetVT, IsZExt);
4147 // Don't deal with undefined shifts.
4148 if (Shift >= DstBits)
4151 // For immediate shifts we can fold the zero-/sign-extension into the shift.
4152 // {S|U}BFM Wd, Wn, #r, #s
4153 // Wd<s-r:0> = Wn<s:r> when r <= s
4155 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
4156 // %2 = ashr i16 %1, 4
4157 // Wd<7-4:0> = Wn<7:4>
4158 // 0b1111_1111_1111_1111__1111_1111_1111_1010 sext
4159 // 0b0000_0000_0000_0000__0000_0000_0000_0101 sext | zext
4160 // 0b0000_0000_0000_0000__0000_0000_0000_1010 zext
4162 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
4163 // %2 = ashr i16 %1, 8
4164 // Wd<7-7,0> = Wn<7:7>
4165 // 0b1111_1111_1111_1111__1111_1111_1111_1111 sext
4166 // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
4167 // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
4169 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
4170 // %2 = ashr i16 %1, 12
4171 // Wd<7-7,0> = Wn<7:7> <- clamp r to 7
4172 // 0b1111_1111_1111_1111__1111_1111_1111_1111 sext
4173 // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
4174 // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
4176 if (Shift >= SrcBits && IsZExt)
4177 return materializeInt(ConstantInt::get(*Context, APInt(RegSize, 0)), RetVT);
4179 unsigned ImmR = std::min<unsigned>(SrcBits - 1, Shift);
4180 unsigned ImmS = SrcBits - 1;
4181 static const unsigned OpcTable[2][2] = {
4182 {AArch64::SBFMWri, AArch64::SBFMXri},
4183 {AArch64::UBFMWri, AArch64::UBFMXri}
4185 unsigned Opc = OpcTable[IsZExt][Is64Bit];
4186 if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) {
4187 unsigned TmpReg = MRI.createVirtualRegister(RC);
4188 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
4189 TII.get(AArch64::SUBREG_TO_REG), TmpReg)
4191 .addReg(Op0, getKillRegState(Op0IsKill))
4192 .addImm(AArch64::sub_32);
4196 return fastEmitInst_rii(Opc, RC, Op0, Op0IsKill, ImmR, ImmS);
4199 unsigned AArch64FastISel::emitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT,
4201 assert(DestVT != MVT::i1 && "ZeroExt/SignExt an i1?");
4203 // FastISel does not have plumbing to deal with extensions where the SrcVT or
4204 // DestVT are odd things, so test to make sure that they are both types we can
4205 // handle (i1/i8/i16/i32 for SrcVT and i8/i16/i32/i64 for DestVT), otherwise
4206 // bail out to SelectionDAG.
4207 if (((DestVT != MVT::i8) && (DestVT != MVT::i16) &&
4208 (DestVT != MVT::i32) && (DestVT != MVT::i64)) ||
4209 ((SrcVT != MVT::i1) && (SrcVT != MVT::i8) &&
4210 (SrcVT != MVT::i16) && (SrcVT != MVT::i32)))
4216 switch (SrcVT.SimpleTy) {
4220 return emiti1Ext(SrcReg, DestVT, IsZExt);
4222 if (DestVT == MVT::i64)
4223 Opc = IsZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
4225 Opc = IsZExt ? AArch64::UBFMWri : AArch64::SBFMWri;
4229 if (DestVT == MVT::i64)
4230 Opc = IsZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
4232 Opc = IsZExt ? AArch64::UBFMWri : AArch64::SBFMWri;
4236 assert(DestVT == MVT::i64 && "IntExt i32 to i32?!?");
4237 Opc = IsZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
4242 // Handle i8 and i16 as i32.
4243 if (DestVT == MVT::i8 || DestVT == MVT::i16)
4245 else if (DestVT == MVT::i64) {
4246 unsigned Src64 = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
4247 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
4248 TII.get(AArch64::SUBREG_TO_REG), Src64)
4251 .addImm(AArch64::sub_32);
4255 const TargetRegisterClass *RC =
4256 (DestVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
4257 return fastEmitInst_rii(Opc, RC, SrcReg, /*TODO:IsKill=*/false, 0, Imm);
4260 static bool isZExtLoad(const MachineInstr *LI) {
4261 switch (LI->getOpcode()) {
4264 case AArch64::LDURBBi:
4265 case AArch64::LDURHHi:
4266 case AArch64::LDURWi:
4267 case AArch64::LDRBBui:
4268 case AArch64::LDRHHui:
4269 case AArch64::LDRWui:
4270 case AArch64::LDRBBroX:
4271 case AArch64::LDRHHroX:
4272 case AArch64::LDRWroX:
4273 case AArch64::LDRBBroW:
4274 case AArch64::LDRHHroW:
4275 case AArch64::LDRWroW:
4280 static bool isSExtLoad(const MachineInstr *LI) {
4281 switch (LI->getOpcode()) {
4284 case AArch64::LDURSBWi:
4285 case AArch64::LDURSHWi:
4286 case AArch64::LDURSBXi:
4287 case AArch64::LDURSHXi:
4288 case AArch64::LDURSWi:
4289 case AArch64::LDRSBWui:
4290 case AArch64::LDRSHWui:
4291 case AArch64::LDRSBXui:
4292 case AArch64::LDRSHXui:
4293 case AArch64::LDRSWui:
4294 case AArch64::LDRSBWroX:
4295 case AArch64::LDRSHWroX:
4296 case AArch64::LDRSBXroX:
4297 case AArch64::LDRSHXroX:
4298 case AArch64::LDRSWroX:
4299 case AArch64::LDRSBWroW:
4300 case AArch64::LDRSHWroW:
4301 case AArch64::LDRSBXroW:
4302 case AArch64::LDRSHXroW:
4303 case AArch64::LDRSWroW:
4308 bool AArch64FastISel::optimizeIntExtLoad(const Instruction *I, MVT RetVT,
4310 const auto *LI = dyn_cast<LoadInst>(I->getOperand(0));
4311 if (!LI || !LI->hasOneUse())
4314 // Check if the load instruction has already been selected.
4315 unsigned Reg = lookUpRegForValue(LI);
4319 MachineInstr *MI = MRI.getUniqueVRegDef(Reg);
4323 // Check if the correct load instruction has been emitted - SelectionDAG might
4324 // have emitted a zero-extending load, but we need a sign-extending load.
4325 bool IsZExt = isa<ZExtInst>(I);
4326 const auto *LoadMI = MI;
4327 if (LoadMI->getOpcode() == TargetOpcode::COPY &&
4328 LoadMI->getOperand(1).getSubReg() == AArch64::sub_32) {
4329 unsigned LoadReg = MI->getOperand(1).getReg();
4330 LoadMI = MRI.getUniqueVRegDef(LoadReg);
4331 assert(LoadMI && "Expected valid instruction");
4333 if (!(IsZExt && isZExtLoad(LoadMI)) && !(!IsZExt && isSExtLoad(LoadMI)))
4336 // Nothing to be done.
4337 if (RetVT != MVT::i64 || SrcVT > MVT::i32) {
4338 updateValueMap(I, Reg);
4343 unsigned Reg64 = createResultReg(&AArch64::GPR64RegClass);
4344 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
4345 TII.get(AArch64::SUBREG_TO_REG), Reg64)
4347 .addReg(Reg, getKillRegState(true))
4348 .addImm(AArch64::sub_32);
4351 assert((MI->getOpcode() == TargetOpcode::COPY &&
4352 MI->getOperand(1).getSubReg() == AArch64::sub_32) &&
4353 "Expected copy instruction");
4354 Reg = MI->getOperand(1).getReg();
4355 MI->eraseFromParent();
4357 updateValueMap(I, Reg);
4361 bool AArch64FastISel::selectIntExt(const Instruction *I) {
4362 assert((isa<ZExtInst>(I) || isa<SExtInst>(I)) &&
4363 "Unexpected integer extend instruction.");
4366 if (!isTypeSupported(I->getType(), RetVT))
4369 if (!isTypeSupported(I->getOperand(0)->getType(), SrcVT))
4372 // Try to optimize already sign-/zero-extended values from load instructions.
4373 if (optimizeIntExtLoad(I, RetVT, SrcVT))
4376 unsigned SrcReg = getRegForValue(I->getOperand(0));
4379 bool SrcIsKill = hasTrivialKill(I->getOperand(0));
4381 // Try to optimize already sign-/zero-extended values from function arguments.
4382 bool IsZExt = isa<ZExtInst>(I);
4383 if (const auto *Arg = dyn_cast<Argument>(I->getOperand(0))) {
4384 if ((IsZExt && Arg->hasZExtAttr()) || (!IsZExt && Arg->hasSExtAttr())) {
4385 if (RetVT == MVT::i64 && SrcVT != MVT::i64) {
4386 unsigned ResultReg = createResultReg(&AArch64::GPR64RegClass);
4387 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
4388 TII.get(AArch64::SUBREG_TO_REG), ResultReg)
4390 .addReg(SrcReg, getKillRegState(SrcIsKill))
4391 .addImm(AArch64::sub_32);
4394 // Conservatively clear all kill flags from all uses, because we are
4395 // replacing a sign-/zero-extend instruction at IR level with a nop at MI
4396 // level. The result of the instruction at IR level might have been
4397 // trivially dead, which is now not longer true.
4398 unsigned UseReg = lookUpRegForValue(I);
4400 MRI.clearKillFlags(UseReg);
4402 updateValueMap(I, SrcReg);
4407 unsigned ResultReg = emitIntExt(SrcVT, SrcReg, RetVT, IsZExt);
4411 updateValueMap(I, ResultReg);
4415 bool AArch64FastISel::selectRem(const Instruction *I, unsigned ISDOpcode) {
4416 EVT DestEVT = TLI.getValueType(I->getType(), true);
4417 if (!DestEVT.isSimple())
4420 MVT DestVT = DestEVT.getSimpleVT();
4421 if (DestVT != MVT::i64 && DestVT != MVT::i32)
4425 bool Is64bit = (DestVT == MVT::i64);
4426 switch (ISDOpcode) {
4430 DivOpc = Is64bit ? AArch64::SDIVXr : AArch64::SDIVWr;
4433 DivOpc = Is64bit ? AArch64::UDIVXr : AArch64::UDIVWr;
4436 unsigned MSubOpc = Is64bit ? AArch64::MSUBXrrr : AArch64::MSUBWrrr;
4437 unsigned Src0Reg = getRegForValue(I->getOperand(0));
4440 bool Src0IsKill = hasTrivialKill(I->getOperand(0));
4442 unsigned Src1Reg = getRegForValue(I->getOperand(1));
4445 bool Src1IsKill = hasTrivialKill(I->getOperand(1));
4447 const TargetRegisterClass *RC =
4448 (DestVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
4449 unsigned QuotReg = fastEmitInst_rr(DivOpc, RC, Src0Reg, /*IsKill=*/false,
4450 Src1Reg, /*IsKill=*/false);
4451 assert(QuotReg && "Unexpected DIV instruction emission failure.");
4452 // The remainder is computed as numerator - (quotient * denominator) using the
4453 // MSUB instruction.
4454 unsigned ResultReg = fastEmitInst_rrr(MSubOpc, RC, QuotReg, /*IsKill=*/true,
4455 Src1Reg, Src1IsKill, Src0Reg,
4457 updateValueMap(I, ResultReg);
4461 bool AArch64FastISel::selectMul(const Instruction *I) {
4463 if (!isTypeSupported(I->getType(), VT, /*IsVectorAllowed=*/true))
4467 return selectBinaryOp(I, ISD::MUL);
4469 const Value *Src0 = I->getOperand(0);
4470 const Value *Src1 = I->getOperand(1);
4471 if (const auto *C = dyn_cast<ConstantInt>(Src0))
4472 if (C->getValue().isPowerOf2())
4473 std::swap(Src0, Src1);
4475 // Try to simplify to a shift instruction.
4476 if (const auto *C = dyn_cast<ConstantInt>(Src1))
4477 if (C->getValue().isPowerOf2()) {
4478 uint64_t ShiftVal = C->getValue().logBase2();
4481 if (const auto *ZExt = dyn_cast<ZExtInst>(Src0)) {
4482 if (!isIntExtFree(ZExt)) {
4484 if (isValueAvailable(ZExt) && isTypeSupported(ZExt->getSrcTy(), VT)) {
4487 Src0 = ZExt->getOperand(0);
4490 } else if (const auto *SExt = dyn_cast<SExtInst>(Src0)) {
4491 if (!isIntExtFree(SExt)) {
4493 if (isValueAvailable(SExt) && isTypeSupported(SExt->getSrcTy(), VT)) {
4496 Src0 = SExt->getOperand(0);
4501 unsigned Src0Reg = getRegForValue(Src0);
4504 bool Src0IsKill = hasTrivialKill(Src0);
4506 unsigned ResultReg =
4507 emitLSL_ri(VT, SrcVT, Src0Reg, Src0IsKill, ShiftVal, IsZExt);
4510 updateValueMap(I, ResultReg);
4515 unsigned Src0Reg = getRegForValue(I->getOperand(0));
4518 bool Src0IsKill = hasTrivialKill(I->getOperand(0));
4520 unsigned Src1Reg = getRegForValue(I->getOperand(1));
4523 bool Src1IsKill = hasTrivialKill(I->getOperand(1));
4525 unsigned ResultReg = emitMul_rr(VT, Src0Reg, Src0IsKill, Src1Reg, Src1IsKill);
4530 updateValueMap(I, ResultReg);
4534 bool AArch64FastISel::selectShift(const Instruction *I) {
4536 if (!isTypeSupported(I->getType(), RetVT, /*IsVectorAllowed=*/true))
4539 if (RetVT.isVector())
4540 return selectOperator(I, I->getOpcode());
4542 if (const auto *C = dyn_cast<ConstantInt>(I->getOperand(1))) {
4543 unsigned ResultReg = 0;
4544 uint64_t ShiftVal = C->getZExtValue();
4546 bool IsZExt = (I->getOpcode() == Instruction::AShr) ? false : true;
4547 const Value *Op0 = I->getOperand(0);
4548 if (const auto *ZExt = dyn_cast<ZExtInst>(Op0)) {
4549 if (!isIntExtFree(ZExt)) {
4551 if (isValueAvailable(ZExt) && isTypeSupported(ZExt->getSrcTy(), TmpVT)) {
4554 Op0 = ZExt->getOperand(0);
4557 } else if (const auto *SExt = dyn_cast<SExtInst>(Op0)) {
4558 if (!isIntExtFree(SExt)) {
4560 if (isValueAvailable(SExt) && isTypeSupported(SExt->getSrcTy(), TmpVT)) {
4563 Op0 = SExt->getOperand(0);
4568 unsigned Op0Reg = getRegForValue(Op0);
4571 bool Op0IsKill = hasTrivialKill(Op0);
4573 switch (I->getOpcode()) {
4574 default: llvm_unreachable("Unexpected instruction.");
4575 case Instruction::Shl:
4576 ResultReg = emitLSL_ri(RetVT, SrcVT, Op0Reg, Op0IsKill, ShiftVal, IsZExt);
4578 case Instruction::AShr:
4579 ResultReg = emitASR_ri(RetVT, SrcVT, Op0Reg, Op0IsKill, ShiftVal, IsZExt);
4581 case Instruction::LShr:
4582 ResultReg = emitLSR_ri(RetVT, SrcVT, Op0Reg, Op0IsKill, ShiftVal, IsZExt);
4588 updateValueMap(I, ResultReg);
4592 unsigned Op0Reg = getRegForValue(I->getOperand(0));
4595 bool Op0IsKill = hasTrivialKill(I->getOperand(0));
4597 unsigned Op1Reg = getRegForValue(I->getOperand(1));
4600 bool Op1IsKill = hasTrivialKill(I->getOperand(1));
4602 unsigned ResultReg = 0;
4603 switch (I->getOpcode()) {
4604 default: llvm_unreachable("Unexpected instruction.");
4605 case Instruction::Shl:
4606 ResultReg = emitLSL_rr(RetVT, Op0Reg, Op0IsKill, Op1Reg, Op1IsKill);
4608 case Instruction::AShr:
4609 ResultReg = emitASR_rr(RetVT, Op0Reg, Op0IsKill, Op1Reg, Op1IsKill);
4611 case Instruction::LShr:
4612 ResultReg = emitLSR_rr(RetVT, Op0Reg, Op0IsKill, Op1Reg, Op1IsKill);
4619 updateValueMap(I, ResultReg);
4623 bool AArch64FastISel::selectBitCast(const Instruction *I) {
4626 if (!isTypeLegal(I->getOperand(0)->getType(), SrcVT))
4628 if (!isTypeLegal(I->getType(), RetVT))
4632 if (RetVT == MVT::f32 && SrcVT == MVT::i32)
4633 Opc = AArch64::FMOVWSr;
4634 else if (RetVT == MVT::f64 && SrcVT == MVT::i64)
4635 Opc = AArch64::FMOVXDr;
4636 else if (RetVT == MVT::i32 && SrcVT == MVT::f32)
4637 Opc = AArch64::FMOVSWr;
4638 else if (RetVT == MVT::i64 && SrcVT == MVT::f64)
4639 Opc = AArch64::FMOVDXr;
4643 const TargetRegisterClass *RC = nullptr;
4644 switch (RetVT.SimpleTy) {
4645 default: llvm_unreachable("Unexpected value type.");
4646 case MVT::i32: RC = &AArch64::GPR32RegClass; break;
4647 case MVT::i64: RC = &AArch64::GPR64RegClass; break;
4648 case MVT::f32: RC = &AArch64::FPR32RegClass; break;
4649 case MVT::f64: RC = &AArch64::FPR64RegClass; break;
4651 unsigned Op0Reg = getRegForValue(I->getOperand(0));
4654 bool Op0IsKill = hasTrivialKill(I->getOperand(0));
4655 unsigned ResultReg = fastEmitInst_r(Opc, RC, Op0Reg, Op0IsKill);
4660 updateValueMap(I, ResultReg);
4664 bool AArch64FastISel::selectFRem(const Instruction *I) {
4666 if (!isTypeLegal(I->getType(), RetVT))
4670 switch (RetVT.SimpleTy) {
4674 LC = RTLIB::REM_F32;
4677 LC = RTLIB::REM_F64;
4682 Args.reserve(I->getNumOperands());
4684 // Populate the argument list.
4685 for (auto &Arg : I->operands()) {
4688 Entry.Ty = Arg->getType();
4689 Args.push_back(Entry);
4692 CallLoweringInfo CLI;
4693 CLI.setCallee(TLI.getLibcallCallingConv(LC), I->getType(),
4694 TLI.getLibcallName(LC), std::move(Args));
4695 if (!lowerCallTo(CLI))
4697 updateValueMap(I, CLI.ResultReg);
4701 bool AArch64FastISel::selectSDiv(const Instruction *I) {
4703 if (!isTypeLegal(I->getType(), VT))
4706 if (!isa<ConstantInt>(I->getOperand(1)))
4707 return selectBinaryOp(I, ISD::SDIV);
4709 const APInt &C = cast<ConstantInt>(I->getOperand(1))->getValue();
4710 if ((VT != MVT::i32 && VT != MVT::i64) || !C ||
4711 !(C.isPowerOf2() || (-C).isPowerOf2()))
4712 return selectBinaryOp(I, ISD::SDIV);
4714 unsigned Lg2 = C.countTrailingZeros();
4715 unsigned Src0Reg = getRegForValue(I->getOperand(0));
4718 bool Src0IsKill = hasTrivialKill(I->getOperand(0));
4720 if (cast<BinaryOperator>(I)->isExact()) {
4721 unsigned ResultReg = emitASR_ri(VT, VT, Src0Reg, Src0IsKill, Lg2);
4724 updateValueMap(I, ResultReg);
4728 int64_t Pow2MinusOne = (1ULL << Lg2) - 1;
4729 unsigned AddReg = emitAdd_ri_(VT, Src0Reg, /*IsKill=*/false, Pow2MinusOne);
4733 // (Src0 < 0) ? Pow2 - 1 : 0;
4734 if (!emitICmp_ri(VT, Src0Reg, /*IsKill=*/false, 0))
4738 const TargetRegisterClass *RC;
4739 if (VT == MVT::i64) {
4740 SelectOpc = AArch64::CSELXr;
4741 RC = &AArch64::GPR64RegClass;
4743 SelectOpc = AArch64::CSELWr;
4744 RC = &AArch64::GPR32RegClass;
4746 unsigned SelectReg =
4747 fastEmitInst_rri(SelectOpc, RC, AddReg, /*IsKill=*/true, Src0Reg,
4748 Src0IsKill, AArch64CC::LT);
4752 // Divide by Pow2 --> ashr. If we're dividing by a negative value we must also
4753 // negate the result.
4754 unsigned ZeroReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
4757 ResultReg = emitAddSub_rs(/*UseAdd=*/false, VT, ZeroReg, /*IsKill=*/true,
4758 SelectReg, /*IsKill=*/true, AArch64_AM::ASR, Lg2);
4760 ResultReg = emitASR_ri(VT, VT, SelectReg, /*IsKill=*/true, Lg2);
4765 updateValueMap(I, ResultReg);
4769 /// This is mostly a copy of the existing FastISel getRegForGEPIndex code. We
4770 /// have to duplicate it for AArch64, because otherwise we would fail during the
4771 /// sign-extend emission.
4772 std::pair<unsigned, bool> AArch64FastISel::getRegForGEPIndex(const Value *Idx) {
4773 unsigned IdxN = getRegForValue(Idx);
4775 // Unhandled operand. Halt "fast" selection and bail.
4776 return std::pair<unsigned, bool>(0, false);
4778 bool IdxNIsKill = hasTrivialKill(Idx);
4780 // If the index is smaller or larger than intptr_t, truncate or extend it.
4781 MVT PtrVT = TLI.getPointerTy();
4782 EVT IdxVT = EVT::getEVT(Idx->getType(), /*HandleUnknown=*/false);
4783 if (IdxVT.bitsLT(PtrVT)) {
4784 IdxN = emitIntExt(IdxVT.getSimpleVT(), IdxN, PtrVT, /*IsZExt=*/false);
4786 } else if (IdxVT.bitsGT(PtrVT))
4787 llvm_unreachable("AArch64 FastISel doesn't support types larger than i64");
4788 return std::pair<unsigned, bool>(IdxN, IdxNIsKill);
4791 /// This is mostly a copy of the existing FastISel GEP code, but we have to
4792 /// duplicate it for AArch64, because otherwise we would bail out even for
4793 /// simple cases. This is because the standard fastEmit functions don't cover
4794 /// MUL at all and ADD is lowered very inefficientily.
4795 bool AArch64FastISel::selectGetElementPtr(const Instruction *I) {
4796 unsigned N = getRegForValue(I->getOperand(0));
4799 bool NIsKill = hasTrivialKill(I->getOperand(0));
4801 // Keep a running tab of the total offset to coalesce multiple N = N + Offset
4802 // into a single N = N + TotalOffset.
4803 uint64_t TotalOffs = 0;
4804 Type *Ty = I->getOperand(0)->getType();
4805 MVT VT = TLI.getPointerTy();
4806 for (auto OI = std::next(I->op_begin()), E = I->op_end(); OI != E; ++OI) {
4807 const Value *Idx = *OI;
4808 if (auto *StTy = dyn_cast<StructType>(Ty)) {
4809 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
4812 TotalOffs += DL.getStructLayout(StTy)->getElementOffset(Field);
4813 Ty = StTy->getElementType(Field);
4815 Ty = cast<SequentialType>(Ty)->getElementType();
4816 // If this is a constant subscript, handle it quickly.
4817 if (const auto *CI = dyn_cast<ConstantInt>(Idx)) {
4822 DL.getTypeAllocSize(Ty) * cast<ConstantInt>(CI)->getSExtValue();
4826 N = emitAdd_ri_(VT, N, NIsKill, TotalOffs);
4833 // N = N + Idx * ElementSize;
4834 uint64_t ElementSize = DL.getTypeAllocSize(Ty);
4835 std::pair<unsigned, bool> Pair = getRegForGEPIndex(Idx);
4836 unsigned IdxN = Pair.first;
4837 bool IdxNIsKill = Pair.second;
4841 if (ElementSize != 1) {
4842 unsigned C = fastEmit_i(VT, VT, ISD::Constant, ElementSize);
4845 IdxN = emitMul_rr(VT, IdxN, IdxNIsKill, C, true);
4850 N = fastEmit_rr(VT, VT, ISD::ADD, N, NIsKill, IdxN, IdxNIsKill);
4856 N = emitAdd_ri_(VT, N, NIsKill, TotalOffs);
4860 updateValueMap(I, N);
4864 bool AArch64FastISel::fastSelectInstruction(const Instruction *I) {
4865 switch (I->getOpcode()) {
4868 case Instruction::Add:
4869 case Instruction::Sub:
4870 return selectAddSub(I);
4871 case Instruction::Mul:
4872 return selectMul(I);
4873 case Instruction::SDiv:
4874 return selectSDiv(I);
4875 case Instruction::SRem:
4876 if (!selectBinaryOp(I, ISD::SREM))
4877 return selectRem(I, ISD::SREM);
4879 case Instruction::URem:
4880 if (!selectBinaryOp(I, ISD::UREM))
4881 return selectRem(I, ISD::UREM);
4883 case Instruction::Shl:
4884 case Instruction::LShr:
4885 case Instruction::AShr:
4886 return selectShift(I);
4887 case Instruction::And:
4888 case Instruction::Or:
4889 case Instruction::Xor:
4890 return selectLogicalOp(I);
4891 case Instruction::Br:
4892 return selectBranch(I);
4893 case Instruction::IndirectBr:
4894 return selectIndirectBr(I);
4895 case Instruction::BitCast:
4896 if (!FastISel::selectBitCast(I))
4897 return selectBitCast(I);
4899 case Instruction::FPToSI:
4900 if (!selectCast(I, ISD::FP_TO_SINT))
4901 return selectFPToInt(I, /*Signed=*/true);
4903 case Instruction::FPToUI:
4904 return selectFPToInt(I, /*Signed=*/false);
4905 case Instruction::ZExt:
4906 case Instruction::SExt:
4907 return selectIntExt(I);
4908 case Instruction::Trunc:
4909 if (!selectCast(I, ISD::TRUNCATE))
4910 return selectTrunc(I);
4912 case Instruction::FPExt:
4913 return selectFPExt(I);
4914 case Instruction::FPTrunc:
4915 return selectFPTrunc(I);
4916 case Instruction::SIToFP:
4917 if (!selectCast(I, ISD::SINT_TO_FP))
4918 return selectIntToFP(I, /*Signed=*/true);
4920 case Instruction::UIToFP:
4921 return selectIntToFP(I, /*Signed=*/false);
4922 case Instruction::Load:
4923 return selectLoad(I);
4924 case Instruction::Store:
4925 return selectStore(I);
4926 case Instruction::FCmp:
4927 case Instruction::ICmp:
4928 return selectCmp(I);
4929 case Instruction::Select:
4930 return selectSelect(I);
4931 case Instruction::Ret:
4932 return selectRet(I);
4933 case Instruction::FRem:
4934 return selectFRem(I);
4935 case Instruction::GetElementPtr:
4936 return selectGetElementPtr(I);
4939 // fall-back to target-independent instruction selection.
4940 return selectOperator(I, I->getOpcode());
4941 // Silence warnings.
4942 (void)&CC_AArch64_DarwinPCS_VarArg;
4946 llvm::FastISel *AArch64::createFastISel(FunctionLoweringInfo &FuncInfo,
4947 const TargetLibraryInfo *LibInfo) {
4948 return new AArch64FastISel(FuncInfo, LibInfo);