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) {
251 &static_cast<const AArch64Subtarget &>(FuncInfo.MF->getSubtarget());
252 Context = &FuncInfo.Fn->getContext();
255 bool fastSelectInstruction(const Instruction *I) override;
257 #include "AArch64GenFastISel.inc"
260 } // end anonymous namespace
262 #include "AArch64GenCallingConv.inc"
264 /// \brief Check if the sign-/zero-extend will be a noop.
265 static bool isIntExtFree(const Instruction *I) {
266 assert((isa<ZExtInst>(I) || isa<SExtInst>(I)) &&
267 "Unexpected integer extend instruction.");
268 assert(!I->getType()->isVectorTy() && I->getType()->isIntegerTy() &&
269 "Unexpected value type.");
270 bool IsZExt = isa<ZExtInst>(I);
272 if (const auto *LI = dyn_cast<LoadInst>(I->getOperand(0)))
276 if (const auto *Arg = dyn_cast<Argument>(I->getOperand(0)))
277 if ((IsZExt && Arg->hasZExtAttr()) || (!IsZExt && Arg->hasSExtAttr()))
283 /// \brief Determine the implicit scale factor that is applied by a memory
284 /// operation for a given value type.
285 static unsigned getImplicitScaleFactor(MVT VT) {
286 switch (VT.SimpleTy) {
289 case MVT::i1: // fall-through
294 case MVT::i32: // fall-through
297 case MVT::i64: // fall-through
303 CCAssignFn *AArch64FastISel::CCAssignFnForCall(CallingConv::ID CC) const {
304 if (CC == CallingConv::WebKit_JS)
305 return CC_AArch64_WebKit_JS;
306 if (CC == CallingConv::GHC)
307 return CC_AArch64_GHC;
308 return Subtarget->isTargetDarwin() ? CC_AArch64_DarwinPCS : CC_AArch64_AAPCS;
311 unsigned AArch64FastISel::fastMaterializeAlloca(const AllocaInst *AI) {
312 assert(TLI.getValueType(AI->getType(), true) == MVT::i64 &&
313 "Alloca should always return a pointer.");
315 // Don't handle dynamic allocas.
316 if (!FuncInfo.StaticAllocaMap.count(AI))
319 DenseMap<const AllocaInst *, int>::iterator SI =
320 FuncInfo.StaticAllocaMap.find(AI);
322 if (SI != FuncInfo.StaticAllocaMap.end()) {
323 unsigned ResultReg = createResultReg(&AArch64::GPR64spRegClass);
324 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri),
326 .addFrameIndex(SI->second)
335 unsigned AArch64FastISel::materializeInt(const ConstantInt *CI, MVT VT) {
340 return fastEmit_i(VT, VT, ISD::Constant, CI->getZExtValue());
342 // Create a copy from the zero register to materialize a "0" value.
343 const TargetRegisterClass *RC = (VT == MVT::i64) ? &AArch64::GPR64RegClass
344 : &AArch64::GPR32RegClass;
345 unsigned ZeroReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
346 unsigned ResultReg = createResultReg(RC);
347 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TargetOpcode::COPY),
348 ResultReg).addReg(ZeroReg, getKillRegState(true));
352 unsigned AArch64FastISel::materializeFP(const ConstantFP *CFP, MVT VT) {
353 // Positive zero (+0.0) has to be materialized with a fmov from the zero
354 // register, because the immediate version of fmov cannot encode zero.
355 if (CFP->isNullValue())
356 return fastMaterializeFloatZero(CFP);
358 if (VT != MVT::f32 && VT != MVT::f64)
361 const APFloat Val = CFP->getValueAPF();
362 bool Is64Bit = (VT == MVT::f64);
363 // This checks to see if we can use FMOV instructions to materialize
364 // a constant, otherwise we have to materialize via the constant pool.
365 if (TLI.isFPImmLegal(Val, VT)) {
367 Is64Bit ? AArch64_AM::getFP64Imm(Val) : AArch64_AM::getFP32Imm(Val);
368 assert((Imm != -1) && "Cannot encode floating-point constant.");
369 unsigned Opc = Is64Bit ? AArch64::FMOVDi : AArch64::FMOVSi;
370 return fastEmitInst_i(Opc, TLI.getRegClassFor(VT), Imm);
373 // For the MachO large code model materialize the FP constant in code.
374 if (Subtarget->isTargetMachO() && TM.getCodeModel() == CodeModel::Large) {
375 unsigned Opc1 = Is64Bit ? AArch64::MOVi64imm : AArch64::MOVi32imm;
376 const TargetRegisterClass *RC = Is64Bit ?
377 &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
379 unsigned TmpReg = createResultReg(RC);
380 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc1), TmpReg)
381 .addImm(CFP->getValueAPF().bitcastToAPInt().getZExtValue());
383 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
384 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
385 TII.get(TargetOpcode::COPY), ResultReg)
386 .addReg(TmpReg, getKillRegState(true));
391 // Materialize via constant pool. MachineConstantPool wants an explicit
393 unsigned Align = DL.getPrefTypeAlignment(CFP->getType());
395 Align = DL.getTypeAllocSize(CFP->getType());
397 unsigned CPI = MCP.getConstantPoolIndex(cast<Constant>(CFP), Align);
398 unsigned ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
399 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
400 ADRPReg).addConstantPoolIndex(CPI, 0, AArch64II::MO_PAGE);
402 unsigned Opc = Is64Bit ? AArch64::LDRDui : AArch64::LDRSui;
403 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
404 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
406 .addConstantPoolIndex(CPI, 0, AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
410 unsigned AArch64FastISel::materializeGV(const GlobalValue *GV) {
411 // We can't handle thread-local variables quickly yet.
412 if (GV->isThreadLocal())
415 // MachO still uses GOT for large code-model accesses, but ELF requires
416 // movz/movk sequences, which FastISel doesn't handle yet.
417 if (TM.getCodeModel() != CodeModel::Small && !Subtarget->isTargetMachO())
420 unsigned char OpFlags = Subtarget->ClassifyGlobalReference(GV, TM);
422 EVT DestEVT = TLI.getValueType(GV->getType(), true);
423 if (!DestEVT.isSimple())
426 unsigned ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
429 if (OpFlags & AArch64II::MO_GOT) {
431 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
433 .addGlobalAddress(GV, 0, AArch64II::MO_GOT | AArch64II::MO_PAGE);
435 ResultReg = createResultReg(&AArch64::GPR64RegClass);
436 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::LDRXui),
439 .addGlobalAddress(GV, 0, AArch64II::MO_GOT | AArch64II::MO_PAGEOFF |
441 } else if (OpFlags & AArch64II::MO_CONSTPOOL) {
442 // We can't handle addresses loaded from a constant pool quickly yet.
446 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
448 .addGlobalAddress(GV, 0, AArch64II::MO_PAGE);
450 ResultReg = createResultReg(&AArch64::GPR64spRegClass);
451 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri),
454 .addGlobalAddress(GV, 0, AArch64II::MO_PAGEOFF | AArch64II::MO_NC)
460 unsigned AArch64FastISel::fastMaterializeConstant(const Constant *C) {
461 EVT CEVT = TLI.getValueType(C->getType(), true);
463 // Only handle simple types.
464 if (!CEVT.isSimple())
466 MVT VT = CEVT.getSimpleVT();
468 if (const auto *CI = dyn_cast<ConstantInt>(C))
469 return materializeInt(CI, VT);
470 else if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
471 return materializeFP(CFP, VT);
472 else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
473 return materializeGV(GV);
478 unsigned AArch64FastISel::fastMaterializeFloatZero(const ConstantFP* CFP) {
479 assert(CFP->isNullValue() &&
480 "Floating-point constant is not a positive zero.");
482 if (!isTypeLegal(CFP->getType(), VT))
485 if (VT != MVT::f32 && VT != MVT::f64)
488 bool Is64Bit = (VT == MVT::f64);
489 unsigned ZReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
490 unsigned Opc = Is64Bit ? AArch64::FMOVXDr : AArch64::FMOVWSr;
491 return fastEmitInst_r(Opc, TLI.getRegClassFor(VT), ZReg, /*IsKill=*/true);
494 /// \brief Check if the multiply is by a power-of-2 constant.
495 static bool isMulPowOf2(const Value *I) {
496 if (const auto *MI = dyn_cast<MulOperator>(I)) {
497 if (const auto *C = dyn_cast<ConstantInt>(MI->getOperand(0)))
498 if (C->getValue().isPowerOf2())
500 if (const auto *C = dyn_cast<ConstantInt>(MI->getOperand(1)))
501 if (C->getValue().isPowerOf2())
507 // Computes the address to get to an object.
508 bool AArch64FastISel::computeAddress(const Value *Obj, Address &Addr, Type *Ty)
510 const User *U = nullptr;
511 unsigned Opcode = Instruction::UserOp1;
512 if (const Instruction *I = dyn_cast<Instruction>(Obj)) {
513 // Don't walk into other basic blocks unless the object is an alloca from
514 // another block, otherwise it may not have a virtual register assigned.
515 if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(Obj)) ||
516 FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
517 Opcode = I->getOpcode();
520 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Obj)) {
521 Opcode = C->getOpcode();
525 if (const PointerType *Ty = dyn_cast<PointerType>(Obj->getType()))
526 if (Ty->getAddressSpace() > 255)
527 // Fast instruction selection doesn't support the special
534 case Instruction::BitCast: {
535 // Look through bitcasts.
536 return computeAddress(U->getOperand(0), Addr, Ty);
538 case Instruction::IntToPtr: {
539 // Look past no-op inttoptrs.
540 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
541 return computeAddress(U->getOperand(0), Addr, Ty);
544 case Instruction::PtrToInt: {
545 // Look past no-op ptrtoints.
546 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
547 return computeAddress(U->getOperand(0), Addr, Ty);
550 case Instruction::GetElementPtr: {
551 Address SavedAddr = Addr;
552 uint64_t TmpOffset = Addr.getOffset();
554 // Iterate through the GEP folding the constants into offsets where
556 gep_type_iterator GTI = gep_type_begin(U);
557 for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e;
559 const Value *Op = *i;
560 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
561 const StructLayout *SL = DL.getStructLayout(STy);
562 unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
563 TmpOffset += SL->getElementOffset(Idx);
565 uint64_t S = DL.getTypeAllocSize(GTI.getIndexedType());
567 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
568 // Constant-offset addressing.
569 TmpOffset += CI->getSExtValue() * S;
572 if (canFoldAddIntoGEP(U, Op)) {
573 // A compatible add with a constant operand. Fold the constant.
575 cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
576 TmpOffset += CI->getSExtValue() * S;
577 // Iterate on the other operand.
578 Op = cast<AddOperator>(Op)->getOperand(0);
582 goto unsupported_gep;
587 // Try to grab the base operand now.
588 Addr.setOffset(TmpOffset);
589 if (computeAddress(U->getOperand(0), Addr, Ty))
592 // We failed, restore everything and try the other options.
598 case Instruction::Alloca: {
599 const AllocaInst *AI = cast<AllocaInst>(Obj);
600 DenseMap<const AllocaInst *, int>::iterator SI =
601 FuncInfo.StaticAllocaMap.find(AI);
602 if (SI != FuncInfo.StaticAllocaMap.end()) {
603 Addr.setKind(Address::FrameIndexBase);
604 Addr.setFI(SI->second);
609 case Instruction::Add: {
610 // Adds of constants are common and easy enough.
611 const Value *LHS = U->getOperand(0);
612 const Value *RHS = U->getOperand(1);
614 if (isa<ConstantInt>(LHS))
617 if (const ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
618 Addr.setOffset(Addr.getOffset() + CI->getSExtValue());
619 return computeAddress(LHS, Addr, Ty);
622 Address Backup = Addr;
623 if (computeAddress(LHS, Addr, Ty) && computeAddress(RHS, Addr, Ty))
629 case Instruction::Sub: {
630 // Subs of constants are common and easy enough.
631 const Value *LHS = U->getOperand(0);
632 const Value *RHS = U->getOperand(1);
634 if (const ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
635 Addr.setOffset(Addr.getOffset() - CI->getSExtValue());
636 return computeAddress(LHS, Addr, Ty);
640 case Instruction::Shl: {
641 if (Addr.getOffsetReg())
644 const auto *CI = dyn_cast<ConstantInt>(U->getOperand(1));
648 unsigned Val = CI->getZExtValue();
649 if (Val < 1 || Val > 3)
652 uint64_t NumBytes = 0;
653 if (Ty && Ty->isSized()) {
654 uint64_t NumBits = DL.getTypeSizeInBits(Ty);
655 NumBytes = NumBits / 8;
656 if (!isPowerOf2_64(NumBits))
660 if (NumBytes != (1ULL << Val))
664 Addr.setExtendType(AArch64_AM::LSL);
666 const Value *Src = U->getOperand(0);
667 if (const auto *I = dyn_cast<Instruction>(Src))
668 if (FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB)
671 // Fold the zext or sext when it won't become a noop.
672 if (const auto *ZE = dyn_cast<ZExtInst>(Src)) {
673 if (!isIntExtFree(ZE) && ZE->getOperand(0)->getType()->isIntegerTy(32)) {
674 Addr.setExtendType(AArch64_AM::UXTW);
675 Src = ZE->getOperand(0);
677 } else if (const auto *SE = dyn_cast<SExtInst>(Src)) {
678 if (!isIntExtFree(SE) && SE->getOperand(0)->getType()->isIntegerTy(32)) {
679 Addr.setExtendType(AArch64_AM::SXTW);
680 Src = SE->getOperand(0);
684 if (const auto *AI = dyn_cast<BinaryOperator>(Src))
685 if (AI->getOpcode() == Instruction::And) {
686 const Value *LHS = AI->getOperand(0);
687 const Value *RHS = AI->getOperand(1);
689 if (const auto *C = dyn_cast<ConstantInt>(LHS))
690 if (C->getValue() == 0xffffffff)
693 if (const auto *C = dyn_cast<ConstantInt>(RHS))
694 if (C->getValue() == 0xffffffff) {
695 Addr.setExtendType(AArch64_AM::UXTW);
696 unsigned Reg = getRegForValue(LHS);
699 bool RegIsKill = hasTrivialKill(LHS);
700 Reg = fastEmitInst_extractsubreg(MVT::i32, Reg, RegIsKill,
702 Addr.setOffsetReg(Reg);
707 unsigned Reg = getRegForValue(Src);
710 Addr.setOffsetReg(Reg);
713 case Instruction::Mul: {
714 if (Addr.getOffsetReg())
720 const Value *LHS = U->getOperand(0);
721 const Value *RHS = U->getOperand(1);
723 // Canonicalize power-of-2 value to the RHS.
724 if (const auto *C = dyn_cast<ConstantInt>(LHS))
725 if (C->getValue().isPowerOf2())
728 assert(isa<ConstantInt>(RHS) && "Expected an ConstantInt.");
729 const auto *C = cast<ConstantInt>(RHS);
730 unsigned Val = C->getValue().logBase2();
731 if (Val < 1 || Val > 3)
734 uint64_t NumBytes = 0;
735 if (Ty && Ty->isSized()) {
736 uint64_t NumBits = DL.getTypeSizeInBits(Ty);
737 NumBytes = NumBits / 8;
738 if (!isPowerOf2_64(NumBits))
742 if (NumBytes != (1ULL << Val))
746 Addr.setExtendType(AArch64_AM::LSL);
748 const Value *Src = LHS;
749 if (const auto *I = dyn_cast<Instruction>(Src))
750 if (FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB)
754 // Fold the zext or sext when it won't become a noop.
755 if (const auto *ZE = dyn_cast<ZExtInst>(Src)) {
756 if (!isIntExtFree(ZE) && ZE->getOperand(0)->getType()->isIntegerTy(32)) {
757 Addr.setExtendType(AArch64_AM::UXTW);
758 Src = ZE->getOperand(0);
760 } else if (const auto *SE = dyn_cast<SExtInst>(Src)) {
761 if (!isIntExtFree(SE) && SE->getOperand(0)->getType()->isIntegerTy(32)) {
762 Addr.setExtendType(AArch64_AM::SXTW);
763 Src = SE->getOperand(0);
767 unsigned Reg = getRegForValue(Src);
770 Addr.setOffsetReg(Reg);
773 case Instruction::And: {
774 if (Addr.getOffsetReg())
777 if (!Ty || DL.getTypeSizeInBits(Ty) != 8)
780 const Value *LHS = U->getOperand(0);
781 const Value *RHS = U->getOperand(1);
783 if (const auto *C = dyn_cast<ConstantInt>(LHS))
784 if (C->getValue() == 0xffffffff)
787 if (const auto *C = dyn_cast<ConstantInt>(RHS))
788 if (C->getValue() == 0xffffffff) {
790 Addr.setExtendType(AArch64_AM::LSL);
791 Addr.setExtendType(AArch64_AM::UXTW);
793 unsigned Reg = getRegForValue(LHS);
796 bool RegIsKill = hasTrivialKill(LHS);
797 Reg = fastEmitInst_extractsubreg(MVT::i32, Reg, RegIsKill,
799 Addr.setOffsetReg(Reg);
804 case Instruction::SExt:
805 case Instruction::ZExt: {
806 if (!Addr.getReg() || Addr.getOffsetReg())
809 const Value *Src = nullptr;
810 // Fold the zext or sext when it won't become a noop.
811 if (const auto *ZE = dyn_cast<ZExtInst>(U)) {
812 if (!isIntExtFree(ZE) && ZE->getOperand(0)->getType()->isIntegerTy(32)) {
813 Addr.setExtendType(AArch64_AM::UXTW);
814 Src = ZE->getOperand(0);
816 } else if (const auto *SE = dyn_cast<SExtInst>(U)) {
817 if (!isIntExtFree(SE) && SE->getOperand(0)->getType()->isIntegerTy(32)) {
818 Addr.setExtendType(AArch64_AM::SXTW);
819 Src = SE->getOperand(0);
827 unsigned Reg = getRegForValue(Src);
830 Addr.setOffsetReg(Reg);
835 if (Addr.isRegBase() && !Addr.getReg()) {
836 unsigned Reg = getRegForValue(Obj);
843 if (!Addr.getOffsetReg()) {
844 unsigned Reg = getRegForValue(Obj);
847 Addr.setOffsetReg(Reg);
854 bool AArch64FastISel::computeCallAddress(const Value *V, Address &Addr) {
855 const User *U = nullptr;
856 unsigned Opcode = Instruction::UserOp1;
859 if (const auto *I = dyn_cast<Instruction>(V)) {
860 Opcode = I->getOpcode();
862 InMBB = I->getParent() == FuncInfo.MBB->getBasicBlock();
863 } else if (const auto *C = dyn_cast<ConstantExpr>(V)) {
864 Opcode = C->getOpcode();
870 case Instruction::BitCast:
871 // Look past bitcasts if its operand is in the same BB.
873 return computeCallAddress(U->getOperand(0), Addr);
875 case Instruction::IntToPtr:
876 // Look past no-op inttoptrs if its operand is in the same BB.
878 TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
879 return computeCallAddress(U->getOperand(0), Addr);
881 case Instruction::PtrToInt:
882 // Look past no-op ptrtoints if its operand is in the same BB.
884 TLI.getValueType(U->getType()) == TLI.getPointerTy())
885 return computeCallAddress(U->getOperand(0), Addr);
889 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
890 Addr.setGlobalValue(GV);
894 // If all else fails, try to materialize the value in a register.
895 if (!Addr.getGlobalValue()) {
896 Addr.setReg(getRegForValue(V));
897 return Addr.getReg() != 0;
904 bool AArch64FastISel::isTypeLegal(Type *Ty, MVT &VT) {
905 EVT evt = TLI.getValueType(Ty, true);
907 // Only handle simple types.
908 if (evt == MVT::Other || !evt.isSimple())
910 VT = evt.getSimpleVT();
912 // This is a legal type, but it's not something we handle in fast-isel.
916 // Handle all other legal types, i.e. a register that will directly hold this
918 return TLI.isTypeLegal(VT);
921 /// \brief Determine if the value type is supported by FastISel.
923 /// FastISel for AArch64 can handle more value types than are legal. This adds
924 /// simple value type such as i1, i8, and i16.
925 bool AArch64FastISel::isTypeSupported(Type *Ty, MVT &VT, bool IsVectorAllowed) {
926 if (Ty->isVectorTy() && !IsVectorAllowed)
929 if (isTypeLegal(Ty, VT))
932 // If this is a type than can be sign or zero-extended to a basic operation
933 // go ahead and accept it now.
934 if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16)
940 bool AArch64FastISel::isValueAvailable(const Value *V) const {
941 if (!isa<Instruction>(V))
944 const auto *I = cast<Instruction>(V);
945 if (FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB)
951 bool AArch64FastISel::simplifyAddress(Address &Addr, MVT VT) {
952 unsigned ScaleFactor = getImplicitScaleFactor(VT);
956 bool ImmediateOffsetNeedsLowering = false;
957 bool RegisterOffsetNeedsLowering = false;
958 int64_t Offset = Addr.getOffset();
959 if (((Offset < 0) || (Offset & (ScaleFactor - 1))) && !isInt<9>(Offset))
960 ImmediateOffsetNeedsLowering = true;
961 else if (Offset > 0 && !(Offset & (ScaleFactor - 1)) &&
962 !isUInt<12>(Offset / ScaleFactor))
963 ImmediateOffsetNeedsLowering = true;
965 // Cannot encode an offset register and an immediate offset in the same
966 // instruction. Fold the immediate offset into the load/store instruction and
967 // emit an additonal add to take care of the offset register.
968 if (!ImmediateOffsetNeedsLowering && Addr.getOffset() && Addr.getOffsetReg())
969 RegisterOffsetNeedsLowering = true;
971 // Cannot encode zero register as base.
972 if (Addr.isRegBase() && Addr.getOffsetReg() && !Addr.getReg())
973 RegisterOffsetNeedsLowering = true;
975 // If this is a stack pointer and the offset needs to be simplified then put
976 // the alloca address into a register, set the base type back to register and
977 // continue. This should almost never happen.
978 if ((ImmediateOffsetNeedsLowering || Addr.getOffsetReg()) && Addr.isFIBase())
980 unsigned ResultReg = createResultReg(&AArch64::GPR64spRegClass);
981 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri),
983 .addFrameIndex(Addr.getFI())
986 Addr.setKind(Address::RegBase);
987 Addr.setReg(ResultReg);
990 if (RegisterOffsetNeedsLowering) {
991 unsigned ResultReg = 0;
993 if (Addr.getExtendType() == AArch64_AM::SXTW ||
994 Addr.getExtendType() == AArch64_AM::UXTW )
995 ResultReg = emitAddSub_rx(/*UseAdd=*/true, MVT::i64, Addr.getReg(),
996 /*TODO:IsKill=*/false, Addr.getOffsetReg(),
997 /*TODO:IsKill=*/false, Addr.getExtendType(),
1000 ResultReg = emitAddSub_rs(/*UseAdd=*/true, MVT::i64, Addr.getReg(),
1001 /*TODO:IsKill=*/false, Addr.getOffsetReg(),
1002 /*TODO:IsKill=*/false, AArch64_AM::LSL,
1005 if (Addr.getExtendType() == AArch64_AM::UXTW)
1006 ResultReg = emitLSL_ri(MVT::i64, MVT::i32, Addr.getOffsetReg(),
1007 /*Op0IsKill=*/false, Addr.getShift(),
1009 else if (Addr.getExtendType() == AArch64_AM::SXTW)
1010 ResultReg = emitLSL_ri(MVT::i64, MVT::i32, Addr.getOffsetReg(),
1011 /*Op0IsKill=*/false, Addr.getShift(),
1014 ResultReg = emitLSL_ri(MVT::i64, MVT::i64, Addr.getOffsetReg(),
1015 /*Op0IsKill=*/false, Addr.getShift());
1020 Addr.setReg(ResultReg);
1021 Addr.setOffsetReg(0);
1023 Addr.setExtendType(AArch64_AM::InvalidShiftExtend);
1026 // Since the offset is too large for the load/store instruction get the
1027 // reg+offset into a register.
1028 if (ImmediateOffsetNeedsLowering) {
1031 // Try to fold the immediate into the add instruction.
1032 ResultReg = emitAdd_ri_(MVT::i64, Addr.getReg(), /*IsKill=*/false, Offset);
1034 ResultReg = fastEmit_i(MVT::i64, MVT::i64, ISD::Constant, Offset);
1038 Addr.setReg(ResultReg);
1044 void AArch64FastISel::addLoadStoreOperands(Address &Addr,
1045 const MachineInstrBuilder &MIB,
1047 unsigned ScaleFactor,
1048 MachineMemOperand *MMO) {
1049 int64_t Offset = Addr.getOffset() / ScaleFactor;
1050 // Frame base works a bit differently. Handle it separately.
1051 if (Addr.isFIBase()) {
1052 int FI = Addr.getFI();
1053 // FIXME: We shouldn't be using getObjectSize/getObjectAlignment. The size
1054 // and alignment should be based on the VT.
1055 MMO = FuncInfo.MF->getMachineMemOperand(
1056 MachinePointerInfo::getFixedStack(FI, Offset), Flags,
1057 MFI.getObjectSize(FI), MFI.getObjectAlignment(FI));
1058 // Now add the rest of the operands.
1059 MIB.addFrameIndex(FI).addImm(Offset);
1061 assert(Addr.isRegBase() && "Unexpected address kind.");
1062 const MCInstrDesc &II = MIB->getDesc();
1063 unsigned Idx = (Flags & MachineMemOperand::MOStore) ? 1 : 0;
1065 constrainOperandRegClass(II, Addr.getReg(), II.getNumDefs()+Idx));
1067 constrainOperandRegClass(II, Addr.getOffsetReg(), II.getNumDefs()+Idx+1));
1068 if (Addr.getOffsetReg()) {
1069 assert(Addr.getOffset() == 0 && "Unexpected offset");
1070 bool IsSigned = Addr.getExtendType() == AArch64_AM::SXTW ||
1071 Addr.getExtendType() == AArch64_AM::SXTX;
1072 MIB.addReg(Addr.getReg());
1073 MIB.addReg(Addr.getOffsetReg());
1074 MIB.addImm(IsSigned);
1075 MIB.addImm(Addr.getShift() != 0);
1077 MIB.addReg(Addr.getReg()).addImm(Offset);
1081 MIB.addMemOperand(MMO);
1084 unsigned AArch64FastISel::emitAddSub(bool UseAdd, MVT RetVT, const Value *LHS,
1085 const Value *RHS, bool SetFlags,
1086 bool WantResult, bool IsZExt) {
1087 AArch64_AM::ShiftExtendType ExtendType = AArch64_AM::InvalidShiftExtend;
1088 bool NeedExtend = false;
1089 switch (RetVT.SimpleTy) {
1097 ExtendType = IsZExt ? AArch64_AM::UXTB : AArch64_AM::SXTB;
1101 ExtendType = IsZExt ? AArch64_AM::UXTH : AArch64_AM::SXTH;
1103 case MVT::i32: // fall-through
1108 RetVT.SimpleTy = std::max(RetVT.SimpleTy, MVT::i32);
1110 // Canonicalize immediates to the RHS first.
1111 if (UseAdd && isa<Constant>(LHS) && !isa<Constant>(RHS))
1112 std::swap(LHS, RHS);
1114 // Canonicalize mul by power of 2 to the RHS.
1115 if (UseAdd && LHS->hasOneUse() && isValueAvailable(LHS))
1116 if (isMulPowOf2(LHS))
1117 std::swap(LHS, RHS);
1119 // Canonicalize shift immediate to the RHS.
1120 if (UseAdd && LHS->hasOneUse() && isValueAvailable(LHS))
1121 if (const auto *SI = dyn_cast<BinaryOperator>(LHS))
1122 if (isa<ConstantInt>(SI->getOperand(1)))
1123 if (SI->getOpcode() == Instruction::Shl ||
1124 SI->getOpcode() == Instruction::LShr ||
1125 SI->getOpcode() == Instruction::AShr )
1126 std::swap(LHS, RHS);
1128 unsigned LHSReg = getRegForValue(LHS);
1131 bool LHSIsKill = hasTrivialKill(LHS);
1134 LHSReg = emitIntExt(SrcVT, LHSReg, RetVT, IsZExt);
1136 unsigned ResultReg = 0;
1137 if (const auto *C = dyn_cast<ConstantInt>(RHS)) {
1138 uint64_t Imm = IsZExt ? C->getZExtValue() : C->getSExtValue();
1139 if (C->isNegative())
1140 ResultReg = emitAddSub_ri(!UseAdd, RetVT, LHSReg, LHSIsKill, -Imm,
1141 SetFlags, WantResult);
1143 ResultReg = emitAddSub_ri(UseAdd, RetVT, LHSReg, LHSIsKill, Imm, SetFlags,
1145 } else if (const auto *C = dyn_cast<Constant>(RHS))
1146 if (C->isNullValue())
1147 ResultReg = emitAddSub_ri(UseAdd, RetVT, LHSReg, LHSIsKill, 0, SetFlags,
1153 // Only extend the RHS within the instruction if there is a valid extend type.
1154 if (ExtendType != AArch64_AM::InvalidShiftExtend && RHS->hasOneUse() &&
1155 isValueAvailable(RHS)) {
1156 if (const auto *SI = dyn_cast<BinaryOperator>(RHS))
1157 if (const auto *C = dyn_cast<ConstantInt>(SI->getOperand(1)))
1158 if ((SI->getOpcode() == Instruction::Shl) && (C->getZExtValue() < 4)) {
1159 unsigned RHSReg = getRegForValue(SI->getOperand(0));
1162 bool RHSIsKill = hasTrivialKill(SI->getOperand(0));
1163 return emitAddSub_rx(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg,
1164 RHSIsKill, ExtendType, C->getZExtValue(),
1165 SetFlags, WantResult);
1167 unsigned RHSReg = getRegForValue(RHS);
1170 bool RHSIsKill = hasTrivialKill(RHS);
1171 return emitAddSub_rx(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg, RHSIsKill,
1172 ExtendType, 0, SetFlags, WantResult);
1175 // Check if the mul can be folded into the instruction.
1176 if (RHS->hasOneUse() && isValueAvailable(RHS))
1177 if (isMulPowOf2(RHS)) {
1178 const Value *MulLHS = cast<MulOperator>(RHS)->getOperand(0);
1179 const Value *MulRHS = cast<MulOperator>(RHS)->getOperand(1);
1181 if (const auto *C = dyn_cast<ConstantInt>(MulLHS))
1182 if (C->getValue().isPowerOf2())
1183 std::swap(MulLHS, MulRHS);
1185 assert(isa<ConstantInt>(MulRHS) && "Expected a ConstantInt.");
1186 uint64_t ShiftVal = cast<ConstantInt>(MulRHS)->getValue().logBase2();
1187 unsigned RHSReg = getRegForValue(MulLHS);
1190 bool RHSIsKill = hasTrivialKill(MulLHS);
1191 return emitAddSub_rs(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg, RHSIsKill,
1192 AArch64_AM::LSL, ShiftVal, SetFlags, WantResult);
1195 // Check if the shift can be folded into the instruction.
1196 if (RHS->hasOneUse() && isValueAvailable(RHS))
1197 if (const auto *SI = dyn_cast<BinaryOperator>(RHS)) {
1198 if (const auto *C = dyn_cast<ConstantInt>(SI->getOperand(1))) {
1199 AArch64_AM::ShiftExtendType ShiftType = AArch64_AM::InvalidShiftExtend;
1200 switch (SI->getOpcode()) {
1202 case Instruction::Shl: ShiftType = AArch64_AM::LSL; break;
1203 case Instruction::LShr: ShiftType = AArch64_AM::LSR; break;
1204 case Instruction::AShr: ShiftType = AArch64_AM::ASR; break;
1206 uint64_t ShiftVal = C->getZExtValue();
1207 if (ShiftType != AArch64_AM::InvalidShiftExtend) {
1208 unsigned RHSReg = getRegForValue(SI->getOperand(0));
1211 bool RHSIsKill = hasTrivialKill(SI->getOperand(0));
1212 return emitAddSub_rs(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg,
1213 RHSIsKill, ShiftType, ShiftVal, SetFlags,
1219 unsigned RHSReg = getRegForValue(RHS);
1222 bool RHSIsKill = hasTrivialKill(RHS);
1225 RHSReg = emitIntExt(SrcVT, RHSReg, RetVT, IsZExt);
1227 return emitAddSub_rr(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg, RHSIsKill,
1228 SetFlags, WantResult);
1231 unsigned AArch64FastISel::emitAddSub_rr(bool UseAdd, MVT RetVT, unsigned LHSReg,
1232 bool LHSIsKill, unsigned RHSReg,
1233 bool RHSIsKill, bool SetFlags,
1235 assert(LHSReg && RHSReg && "Invalid register number.");
1237 if (RetVT != MVT::i32 && RetVT != MVT::i64)
1240 static const unsigned OpcTable[2][2][2] = {
1241 { { AArch64::SUBWrr, AArch64::SUBXrr },
1242 { AArch64::ADDWrr, AArch64::ADDXrr } },
1243 { { AArch64::SUBSWrr, AArch64::SUBSXrr },
1244 { AArch64::ADDSWrr, AArch64::ADDSXrr } }
1246 bool Is64Bit = RetVT == MVT::i64;
1247 unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit];
1248 const TargetRegisterClass *RC =
1249 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
1252 ResultReg = createResultReg(RC);
1254 ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
1256 const MCInstrDesc &II = TII.get(Opc);
1257 LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
1258 RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1);
1259 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
1260 .addReg(LHSReg, getKillRegState(LHSIsKill))
1261 .addReg(RHSReg, getKillRegState(RHSIsKill));
1265 unsigned AArch64FastISel::emitAddSub_ri(bool UseAdd, MVT RetVT, unsigned LHSReg,
1266 bool LHSIsKill, uint64_t Imm,
1267 bool SetFlags, bool WantResult) {
1268 assert(LHSReg && "Invalid register number.");
1270 if (RetVT != MVT::i32 && RetVT != MVT::i64)
1274 if (isUInt<12>(Imm))
1276 else if ((Imm & 0xfff000) == Imm) {
1282 static const unsigned OpcTable[2][2][2] = {
1283 { { AArch64::SUBWri, AArch64::SUBXri },
1284 { AArch64::ADDWri, AArch64::ADDXri } },
1285 { { AArch64::SUBSWri, AArch64::SUBSXri },
1286 { AArch64::ADDSWri, AArch64::ADDSXri } }
1288 bool Is64Bit = RetVT == MVT::i64;
1289 unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit];
1290 const TargetRegisterClass *RC;
1292 RC = Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
1294 RC = Is64Bit ? &AArch64::GPR64spRegClass : &AArch64::GPR32spRegClass;
1297 ResultReg = createResultReg(RC);
1299 ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
1301 const MCInstrDesc &II = TII.get(Opc);
1302 LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
1303 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
1304 .addReg(LHSReg, getKillRegState(LHSIsKill))
1306 .addImm(getShifterImm(AArch64_AM::LSL, ShiftImm));
1310 unsigned AArch64FastISel::emitAddSub_rs(bool UseAdd, MVT RetVT, unsigned LHSReg,
1311 bool LHSIsKill, unsigned RHSReg,
1313 AArch64_AM::ShiftExtendType ShiftType,
1314 uint64_t ShiftImm, bool SetFlags,
1316 assert(LHSReg && RHSReg && "Invalid register number.");
1318 if (RetVT != MVT::i32 && RetVT != MVT::i64)
1321 static const unsigned OpcTable[2][2][2] = {
1322 { { AArch64::SUBWrs, AArch64::SUBXrs },
1323 { AArch64::ADDWrs, AArch64::ADDXrs } },
1324 { { AArch64::SUBSWrs, AArch64::SUBSXrs },
1325 { AArch64::ADDSWrs, AArch64::ADDSXrs } }
1327 bool Is64Bit = RetVT == MVT::i64;
1328 unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit];
1329 const TargetRegisterClass *RC =
1330 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
1333 ResultReg = createResultReg(RC);
1335 ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
1337 const MCInstrDesc &II = TII.get(Opc);
1338 LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
1339 RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1);
1340 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
1341 .addReg(LHSReg, getKillRegState(LHSIsKill))
1342 .addReg(RHSReg, getKillRegState(RHSIsKill))
1343 .addImm(getShifterImm(ShiftType, ShiftImm));
1347 unsigned AArch64FastISel::emitAddSub_rx(bool UseAdd, MVT RetVT, unsigned LHSReg,
1348 bool LHSIsKill, unsigned RHSReg,
1350 AArch64_AM::ShiftExtendType ExtType,
1351 uint64_t ShiftImm, bool SetFlags,
1353 assert(LHSReg && RHSReg && "Invalid register number.");
1355 if (RetVT != MVT::i32 && RetVT != MVT::i64)
1358 static const unsigned OpcTable[2][2][2] = {
1359 { { AArch64::SUBWrx, AArch64::SUBXrx },
1360 { AArch64::ADDWrx, AArch64::ADDXrx } },
1361 { { AArch64::SUBSWrx, AArch64::SUBSXrx },
1362 { AArch64::ADDSWrx, AArch64::ADDSXrx } }
1364 bool Is64Bit = RetVT == MVT::i64;
1365 unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit];
1366 const TargetRegisterClass *RC = nullptr;
1368 RC = Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
1370 RC = Is64Bit ? &AArch64::GPR64spRegClass : &AArch64::GPR32spRegClass;
1373 ResultReg = createResultReg(RC);
1375 ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
1377 const MCInstrDesc &II = TII.get(Opc);
1378 LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
1379 RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1);
1380 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
1381 .addReg(LHSReg, getKillRegState(LHSIsKill))
1382 .addReg(RHSReg, getKillRegState(RHSIsKill))
1383 .addImm(getArithExtendImm(ExtType, ShiftImm));
1387 bool AArch64FastISel::emitCmp(const Value *LHS, const Value *RHS, bool IsZExt) {
1388 Type *Ty = LHS->getType();
1389 EVT EVT = TLI.getValueType(Ty, true);
1390 if (!EVT.isSimple())
1392 MVT VT = EVT.getSimpleVT();
1394 switch (VT.SimpleTy) {
1402 return emitICmp(VT, LHS, RHS, IsZExt);
1405 return emitFCmp(VT, LHS, RHS);
1409 bool AArch64FastISel::emitICmp(MVT RetVT, const Value *LHS, const Value *RHS,
1411 return emitSub(RetVT, LHS, RHS, /*SetFlags=*/true, /*WantResult=*/false,
1415 bool AArch64FastISel::emitICmp_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill,
1417 return emitAddSub_ri(/*UseAdd=*/false, RetVT, LHSReg, LHSIsKill, Imm,
1418 /*SetFlags=*/true, /*WantResult=*/false) != 0;
1421 bool AArch64FastISel::emitFCmp(MVT RetVT, const Value *LHS, const Value *RHS) {
1422 if (RetVT != MVT::f32 && RetVT != MVT::f64)
1425 // Check to see if the 2nd operand is a constant that we can encode directly
1427 bool UseImm = false;
1428 if (const auto *CFP = dyn_cast<ConstantFP>(RHS))
1429 if (CFP->isZero() && !CFP->isNegative())
1432 unsigned LHSReg = getRegForValue(LHS);
1435 bool LHSIsKill = hasTrivialKill(LHS);
1438 unsigned Opc = (RetVT == MVT::f64) ? AArch64::FCMPDri : AArch64::FCMPSri;
1439 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc))
1440 .addReg(LHSReg, getKillRegState(LHSIsKill));
1444 unsigned RHSReg = getRegForValue(RHS);
1447 bool RHSIsKill = hasTrivialKill(RHS);
1449 unsigned Opc = (RetVT == MVT::f64) ? AArch64::FCMPDrr : AArch64::FCMPSrr;
1450 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc))
1451 .addReg(LHSReg, getKillRegState(LHSIsKill))
1452 .addReg(RHSReg, getKillRegState(RHSIsKill));
1456 unsigned AArch64FastISel::emitAdd(MVT RetVT, const Value *LHS, const Value *RHS,
1457 bool SetFlags, bool WantResult, bool IsZExt) {
1458 return emitAddSub(/*UseAdd=*/true, RetVT, LHS, RHS, SetFlags, WantResult,
1462 /// \brief This method is a wrapper to simplify add emission.
1464 /// First try to emit an add with an immediate operand using emitAddSub_ri. If
1465 /// that fails, then try to materialize the immediate into a register and use
1466 /// emitAddSub_rr instead.
1467 unsigned AArch64FastISel::emitAdd_ri_(MVT VT, unsigned Op0, bool Op0IsKill,
1471 ResultReg = emitAddSub_ri(false, VT, Op0, Op0IsKill, -Imm);
1473 ResultReg = emitAddSub_ri(true, VT, Op0, Op0IsKill, Imm);
1478 unsigned CReg = fastEmit_i(VT, VT, ISD::Constant, Imm);
1482 ResultReg = emitAddSub_rr(true, VT, Op0, Op0IsKill, CReg, true);
1486 unsigned AArch64FastISel::emitSub(MVT RetVT, const Value *LHS, const Value *RHS,
1487 bool SetFlags, bool WantResult, bool IsZExt) {
1488 return emitAddSub(/*UseAdd=*/false, RetVT, LHS, RHS, SetFlags, WantResult,
1492 unsigned AArch64FastISel::emitSubs_rr(MVT RetVT, unsigned LHSReg,
1493 bool LHSIsKill, unsigned RHSReg,
1494 bool RHSIsKill, bool WantResult) {
1495 return emitAddSub_rr(/*UseAdd=*/false, RetVT, LHSReg, LHSIsKill, RHSReg,
1496 RHSIsKill, /*SetFlags=*/true, WantResult);
1499 unsigned AArch64FastISel::emitSubs_rs(MVT RetVT, unsigned LHSReg,
1500 bool LHSIsKill, unsigned RHSReg,
1502 AArch64_AM::ShiftExtendType ShiftType,
1503 uint64_t ShiftImm, bool WantResult) {
1504 return emitAddSub_rs(/*UseAdd=*/false, RetVT, LHSReg, LHSIsKill, RHSReg,
1505 RHSIsKill, ShiftType, ShiftImm, /*SetFlags=*/true,
1509 unsigned AArch64FastISel::emitLogicalOp(unsigned ISDOpc, MVT RetVT,
1510 const Value *LHS, const Value *RHS) {
1511 // Canonicalize immediates to the RHS first.
1512 if (isa<ConstantInt>(LHS) && !isa<ConstantInt>(RHS))
1513 std::swap(LHS, RHS);
1515 // Canonicalize mul by power-of-2 to the RHS.
1516 if (LHS->hasOneUse() && isValueAvailable(LHS))
1517 if (isMulPowOf2(LHS))
1518 std::swap(LHS, RHS);
1520 // Canonicalize shift immediate to the RHS.
1521 if (LHS->hasOneUse() && isValueAvailable(LHS))
1522 if (const auto *SI = dyn_cast<ShlOperator>(LHS))
1523 if (isa<ConstantInt>(SI->getOperand(1)))
1524 std::swap(LHS, RHS);
1526 unsigned LHSReg = getRegForValue(LHS);
1529 bool LHSIsKill = hasTrivialKill(LHS);
1531 unsigned ResultReg = 0;
1532 if (const auto *C = dyn_cast<ConstantInt>(RHS)) {
1533 uint64_t Imm = C->getZExtValue();
1534 ResultReg = emitLogicalOp_ri(ISDOpc, RetVT, LHSReg, LHSIsKill, Imm);
1539 // Check if the mul can be folded into the instruction.
1540 if (RHS->hasOneUse() && isValueAvailable(RHS))
1541 if (isMulPowOf2(RHS)) {
1542 const Value *MulLHS = cast<MulOperator>(RHS)->getOperand(0);
1543 const Value *MulRHS = cast<MulOperator>(RHS)->getOperand(1);
1545 if (const auto *C = dyn_cast<ConstantInt>(MulLHS))
1546 if (C->getValue().isPowerOf2())
1547 std::swap(MulLHS, MulRHS);
1549 assert(isa<ConstantInt>(MulRHS) && "Expected a ConstantInt.");
1550 uint64_t ShiftVal = cast<ConstantInt>(MulRHS)->getValue().logBase2();
1552 unsigned RHSReg = getRegForValue(MulLHS);
1555 bool RHSIsKill = hasTrivialKill(MulLHS);
1556 return emitLogicalOp_rs(ISDOpc, RetVT, LHSReg, LHSIsKill, RHSReg,
1557 RHSIsKill, ShiftVal);
1560 // Check if the shift can be folded into the instruction.
1561 if (RHS->hasOneUse() && isValueAvailable(RHS))
1562 if (const auto *SI = dyn_cast<ShlOperator>(RHS))
1563 if (const auto *C = dyn_cast<ConstantInt>(SI->getOperand(1))) {
1564 uint64_t ShiftVal = C->getZExtValue();
1565 unsigned RHSReg = getRegForValue(SI->getOperand(0));
1568 bool RHSIsKill = hasTrivialKill(SI->getOperand(0));
1569 return emitLogicalOp_rs(ISDOpc, RetVT, LHSReg, LHSIsKill, RHSReg,
1570 RHSIsKill, ShiftVal);
1573 unsigned RHSReg = getRegForValue(RHS);
1576 bool RHSIsKill = hasTrivialKill(RHS);
1578 MVT VT = std::max(MVT::i32, RetVT.SimpleTy);
1579 ResultReg = fastEmit_rr(VT, VT, ISDOpc, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
1580 if (RetVT >= MVT::i8 && RetVT <= MVT::i16) {
1581 uint64_t Mask = (RetVT == MVT::i8) ? 0xff : 0xffff;
1582 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
1587 unsigned AArch64FastISel::emitLogicalOp_ri(unsigned ISDOpc, MVT RetVT,
1588 unsigned LHSReg, bool LHSIsKill,
1590 assert((ISD::AND + 1 == ISD::OR) && (ISD::AND + 2 == ISD::XOR) &&
1591 "ISD nodes are not consecutive!");
1592 static const unsigned OpcTable[3][2] = {
1593 { AArch64::ANDWri, AArch64::ANDXri },
1594 { AArch64::ORRWri, AArch64::ORRXri },
1595 { AArch64::EORWri, AArch64::EORXri }
1597 const TargetRegisterClass *RC;
1600 switch (RetVT.SimpleTy) {
1607 unsigned Idx = ISDOpc - ISD::AND;
1608 Opc = OpcTable[Idx][0];
1609 RC = &AArch64::GPR32spRegClass;
1614 Opc = OpcTable[ISDOpc - ISD::AND][1];
1615 RC = &AArch64::GPR64spRegClass;
1620 if (!AArch64_AM::isLogicalImmediate(Imm, RegSize))
1623 unsigned ResultReg =
1624 fastEmitInst_ri(Opc, RC, LHSReg, LHSIsKill,
1625 AArch64_AM::encodeLogicalImmediate(Imm, RegSize));
1626 if (RetVT >= MVT::i8 && RetVT <= MVT::i16 && ISDOpc != ISD::AND) {
1627 uint64_t Mask = (RetVT == MVT::i8) ? 0xff : 0xffff;
1628 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
1633 unsigned AArch64FastISel::emitLogicalOp_rs(unsigned ISDOpc, MVT RetVT,
1634 unsigned LHSReg, bool LHSIsKill,
1635 unsigned RHSReg, bool RHSIsKill,
1636 uint64_t ShiftImm) {
1637 assert((ISD::AND + 1 == ISD::OR) && (ISD::AND + 2 == ISD::XOR) &&
1638 "ISD nodes are not consecutive!");
1639 static const unsigned OpcTable[3][2] = {
1640 { AArch64::ANDWrs, AArch64::ANDXrs },
1641 { AArch64::ORRWrs, AArch64::ORRXrs },
1642 { AArch64::EORWrs, AArch64::EORXrs }
1644 const TargetRegisterClass *RC;
1646 switch (RetVT.SimpleTy) {
1653 Opc = OpcTable[ISDOpc - ISD::AND][0];
1654 RC = &AArch64::GPR32RegClass;
1657 Opc = OpcTable[ISDOpc - ISD::AND][1];
1658 RC = &AArch64::GPR64RegClass;
1661 unsigned ResultReg =
1662 fastEmitInst_rri(Opc, RC, LHSReg, LHSIsKill, RHSReg, RHSIsKill,
1663 AArch64_AM::getShifterImm(AArch64_AM::LSL, ShiftImm));
1664 if (RetVT >= MVT::i8 && RetVT <= MVT::i16) {
1665 uint64_t Mask = (RetVT == MVT::i8) ? 0xff : 0xffff;
1666 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
1671 unsigned AArch64FastISel::emitAnd_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill,
1673 return emitLogicalOp_ri(ISD::AND, RetVT, LHSReg, LHSIsKill, Imm);
1676 unsigned AArch64FastISel::emitLoad(MVT VT, MVT RetVT, Address Addr,
1677 bool WantZExt, MachineMemOperand *MMO) {
1678 // Simplify this down to something we can handle.
1679 if (!simplifyAddress(Addr, VT))
1682 unsigned ScaleFactor = getImplicitScaleFactor(VT);
1684 llvm_unreachable("Unexpected value type.");
1686 // Negative offsets require unscaled, 9-bit, signed immediate offsets.
1687 // Otherwise, we try using scaled, 12-bit, unsigned immediate offsets.
1688 bool UseScaled = true;
1689 if ((Addr.getOffset() < 0) || (Addr.getOffset() & (ScaleFactor - 1))) {
1694 static const unsigned GPOpcTable[2][8][4] = {
1696 { { AArch64::LDURSBWi, AArch64::LDURSHWi, AArch64::LDURWi,
1698 { AArch64::LDURSBXi, AArch64::LDURSHXi, AArch64::LDURSWi,
1700 { AArch64::LDRSBWui, AArch64::LDRSHWui, AArch64::LDRWui,
1702 { AArch64::LDRSBXui, AArch64::LDRSHXui, AArch64::LDRSWui,
1704 { AArch64::LDRSBWroX, AArch64::LDRSHWroX, AArch64::LDRWroX,
1706 { AArch64::LDRSBXroX, AArch64::LDRSHXroX, AArch64::LDRSWroX,
1708 { AArch64::LDRSBWroW, AArch64::LDRSHWroW, AArch64::LDRWroW,
1710 { AArch64::LDRSBXroW, AArch64::LDRSHXroW, AArch64::LDRSWroW,
1714 { { AArch64::LDURBBi, AArch64::LDURHHi, AArch64::LDURWi,
1716 { AArch64::LDURBBi, AArch64::LDURHHi, AArch64::LDURWi,
1718 { AArch64::LDRBBui, AArch64::LDRHHui, AArch64::LDRWui,
1720 { AArch64::LDRBBui, AArch64::LDRHHui, AArch64::LDRWui,
1722 { AArch64::LDRBBroX, AArch64::LDRHHroX, AArch64::LDRWroX,
1724 { AArch64::LDRBBroX, AArch64::LDRHHroX, AArch64::LDRWroX,
1726 { AArch64::LDRBBroW, AArch64::LDRHHroW, AArch64::LDRWroW,
1728 { AArch64::LDRBBroW, AArch64::LDRHHroW, AArch64::LDRWroW,
1733 static const unsigned FPOpcTable[4][2] = {
1734 { AArch64::LDURSi, AArch64::LDURDi },
1735 { AArch64::LDRSui, AArch64::LDRDui },
1736 { AArch64::LDRSroX, AArch64::LDRDroX },
1737 { AArch64::LDRSroW, AArch64::LDRDroW }
1741 const TargetRegisterClass *RC;
1742 bool UseRegOffset = Addr.isRegBase() && !Addr.getOffset() && Addr.getReg() &&
1743 Addr.getOffsetReg();
1744 unsigned Idx = UseRegOffset ? 2 : UseScaled ? 1 : 0;
1745 if (Addr.getExtendType() == AArch64_AM::UXTW ||
1746 Addr.getExtendType() == AArch64_AM::SXTW)
1749 bool IsRet64Bit = RetVT == MVT::i64;
1750 switch (VT.SimpleTy) {
1752 llvm_unreachable("Unexpected value type.");
1753 case MVT::i1: // Intentional fall-through.
1755 Opc = GPOpcTable[WantZExt][2 * Idx + IsRet64Bit][0];
1756 RC = (IsRet64Bit && !WantZExt) ?
1757 &AArch64::GPR64RegClass: &AArch64::GPR32RegClass;
1760 Opc = GPOpcTable[WantZExt][2 * Idx + IsRet64Bit][1];
1761 RC = (IsRet64Bit && !WantZExt) ?
1762 &AArch64::GPR64RegClass: &AArch64::GPR32RegClass;
1765 Opc = GPOpcTable[WantZExt][2 * Idx + IsRet64Bit][2];
1766 RC = (IsRet64Bit && !WantZExt) ?
1767 &AArch64::GPR64RegClass: &AArch64::GPR32RegClass;
1770 Opc = GPOpcTable[WantZExt][2 * Idx + IsRet64Bit][3];
1771 RC = &AArch64::GPR64RegClass;
1774 Opc = FPOpcTable[Idx][0];
1775 RC = &AArch64::FPR32RegClass;
1778 Opc = FPOpcTable[Idx][1];
1779 RC = &AArch64::FPR64RegClass;
1783 // Create the base instruction, then add the operands.
1784 unsigned ResultReg = createResultReg(RC);
1785 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1786 TII.get(Opc), ResultReg);
1787 addLoadStoreOperands(Addr, MIB, MachineMemOperand::MOLoad, ScaleFactor, MMO);
1789 // Loading an i1 requires special handling.
1790 if (VT == MVT::i1) {
1791 unsigned ANDReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, 1);
1792 assert(ANDReg && "Unexpected AND instruction emission failure.");
1796 // For zero-extending loads to 64bit we emit a 32bit load and then convert
1797 // the 32bit reg to a 64bit reg.
1798 if (WantZExt && RetVT == MVT::i64 && VT <= MVT::i32) {
1799 unsigned Reg64 = createResultReg(&AArch64::GPR64RegClass);
1800 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1801 TII.get(AArch64::SUBREG_TO_REG), Reg64)
1803 .addReg(ResultReg, getKillRegState(true))
1804 .addImm(AArch64::sub_32);
1810 bool AArch64FastISel::selectAddSub(const Instruction *I) {
1812 if (!isTypeSupported(I->getType(), VT, /*IsVectorAllowed=*/true))
1816 return selectOperator(I, I->getOpcode());
1819 switch (I->getOpcode()) {
1821 llvm_unreachable("Unexpected instruction.");
1822 case Instruction::Add:
1823 ResultReg = emitAdd(VT, I->getOperand(0), I->getOperand(1));
1825 case Instruction::Sub:
1826 ResultReg = emitSub(VT, I->getOperand(0), I->getOperand(1));
1832 updateValueMap(I, ResultReg);
1836 bool AArch64FastISel::selectLogicalOp(const Instruction *I) {
1838 if (!isTypeSupported(I->getType(), VT, /*IsVectorAllowed=*/true))
1842 return selectOperator(I, I->getOpcode());
1845 switch (I->getOpcode()) {
1847 llvm_unreachable("Unexpected instruction.");
1848 case Instruction::And:
1849 ResultReg = emitLogicalOp(ISD::AND, VT, I->getOperand(0), I->getOperand(1));
1851 case Instruction::Or:
1852 ResultReg = emitLogicalOp(ISD::OR, VT, I->getOperand(0), I->getOperand(1));
1854 case Instruction::Xor:
1855 ResultReg = emitLogicalOp(ISD::XOR, VT, I->getOperand(0), I->getOperand(1));
1861 updateValueMap(I, ResultReg);
1865 bool AArch64FastISel::selectLoad(const Instruction *I) {
1867 // Verify we have a legal type before going any further. Currently, we handle
1868 // simple types that will directly fit in a register (i32/f32/i64/f64) or
1869 // those that can be sign or zero-extended to a basic operation (i1/i8/i16).
1870 if (!isTypeSupported(I->getType(), VT, /*IsVectorAllowed=*/true) ||
1871 cast<LoadInst>(I)->isAtomic())
1874 // See if we can handle this address.
1876 if (!computeAddress(I->getOperand(0), Addr, I->getType()))
1879 // Fold the following sign-/zero-extend into the load instruction.
1880 bool WantZExt = true;
1882 const Value *IntExtVal = nullptr;
1883 if (I->hasOneUse()) {
1884 if (const auto *ZE = dyn_cast<ZExtInst>(I->use_begin()->getUser())) {
1885 if (isTypeSupported(ZE->getType(), RetVT))
1889 } else if (const auto *SE = dyn_cast<SExtInst>(I->use_begin()->getUser())) {
1890 if (isTypeSupported(SE->getType(), RetVT))
1898 unsigned ResultReg =
1899 emitLoad(VT, RetVT, Addr, WantZExt, createMachineMemOperandFor(I));
1903 // There are a few different cases we have to handle, because the load or the
1904 // sign-/zero-extend might not be selected by FastISel if we fall-back to
1905 // SelectionDAG. There is also an ordering issue when both instructions are in
1906 // different basic blocks.
1907 // 1.) The load instruction is selected by FastISel, but the integer extend
1908 // not. This usually happens when the integer extend is in a different
1909 // basic block and SelectionDAG took over for that basic block.
1910 // 2.) The load instruction is selected before the integer extend. This only
1911 // happens when the integer extend is in a different basic block.
1912 // 3.) The load instruction is selected by SelectionDAG and the integer extend
1913 // by FastISel. This happens if there are instructions between the load
1914 // and the integer extend that couldn't be selected by FastISel.
1916 // The integer extend hasn't been emitted yet. FastISel or SelectionDAG
1917 // could select it. Emit a copy to subreg if necessary. FastISel will remove
1918 // it when it selects the integer extend.
1919 unsigned Reg = lookUpRegForValue(IntExtVal);
1920 auto *MI = MRI.getUniqueVRegDef(Reg);
1922 if (RetVT == MVT::i64 && VT <= MVT::i32) {
1924 // Delete the last emitted instruction from emitLoad (SUBREG_TO_REG).
1925 std::prev(FuncInfo.InsertPt)->eraseFromParent();
1926 ResultReg = std::prev(FuncInfo.InsertPt)->getOperand(0).getReg();
1928 ResultReg = fastEmitInst_extractsubreg(MVT::i32, ResultReg,
1932 updateValueMap(I, ResultReg);
1936 // The integer extend has already been emitted - delete all the instructions
1937 // that have been emitted by the integer extend lowering code and use the
1938 // result from the load instruction directly.
1941 for (auto &Opnd : MI->uses()) {
1943 Reg = Opnd.getReg();
1947 MI->eraseFromParent();
1950 MI = MRI.getUniqueVRegDef(Reg);
1952 updateValueMap(IntExtVal, ResultReg);
1956 updateValueMap(I, ResultReg);
1960 bool AArch64FastISel::emitStore(MVT VT, unsigned SrcReg, Address Addr,
1961 MachineMemOperand *MMO) {
1962 // Simplify this down to something we can handle.
1963 if (!simplifyAddress(Addr, VT))
1966 unsigned ScaleFactor = getImplicitScaleFactor(VT);
1968 llvm_unreachable("Unexpected value type.");
1970 // Negative offsets require unscaled, 9-bit, signed immediate offsets.
1971 // Otherwise, we try using scaled, 12-bit, unsigned immediate offsets.
1972 bool UseScaled = true;
1973 if ((Addr.getOffset() < 0) || (Addr.getOffset() & (ScaleFactor - 1))) {
1978 static const unsigned OpcTable[4][6] = {
1979 { AArch64::STURBBi, AArch64::STURHHi, AArch64::STURWi, AArch64::STURXi,
1980 AArch64::STURSi, AArch64::STURDi },
1981 { AArch64::STRBBui, AArch64::STRHHui, AArch64::STRWui, AArch64::STRXui,
1982 AArch64::STRSui, AArch64::STRDui },
1983 { AArch64::STRBBroX, AArch64::STRHHroX, AArch64::STRWroX, AArch64::STRXroX,
1984 AArch64::STRSroX, AArch64::STRDroX },
1985 { AArch64::STRBBroW, AArch64::STRHHroW, AArch64::STRWroW, AArch64::STRXroW,
1986 AArch64::STRSroW, AArch64::STRDroW }
1990 bool VTIsi1 = false;
1991 bool UseRegOffset = Addr.isRegBase() && !Addr.getOffset() && Addr.getReg() &&
1992 Addr.getOffsetReg();
1993 unsigned Idx = UseRegOffset ? 2 : UseScaled ? 1 : 0;
1994 if (Addr.getExtendType() == AArch64_AM::UXTW ||
1995 Addr.getExtendType() == AArch64_AM::SXTW)
1998 switch (VT.SimpleTy) {
1999 default: llvm_unreachable("Unexpected value type.");
2000 case MVT::i1: VTIsi1 = true;
2001 case MVT::i8: Opc = OpcTable[Idx][0]; break;
2002 case MVT::i16: Opc = OpcTable[Idx][1]; break;
2003 case MVT::i32: Opc = OpcTable[Idx][2]; break;
2004 case MVT::i64: Opc = OpcTable[Idx][3]; break;
2005 case MVT::f32: Opc = OpcTable[Idx][4]; break;
2006 case MVT::f64: Opc = OpcTable[Idx][5]; break;
2009 // Storing an i1 requires special handling.
2010 if (VTIsi1 && SrcReg != AArch64::WZR) {
2011 unsigned ANDReg = emitAnd_ri(MVT::i32, SrcReg, /*TODO:IsKill=*/false, 1);
2012 assert(ANDReg && "Unexpected AND instruction emission failure.");
2015 // Create the base instruction, then add the operands.
2016 const MCInstrDesc &II = TII.get(Opc);
2017 SrcReg = constrainOperandRegClass(II, SrcReg, II.getNumDefs());
2018 MachineInstrBuilder MIB =
2019 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II).addReg(SrcReg);
2020 addLoadStoreOperands(Addr, MIB, MachineMemOperand::MOStore, ScaleFactor, MMO);
2025 bool AArch64FastISel::selectStore(const Instruction *I) {
2027 const Value *Op0 = I->getOperand(0);
2028 // Verify we have a legal type before going any further. Currently, we handle
2029 // simple types that will directly fit in a register (i32/f32/i64/f64) or
2030 // those that can be sign or zero-extended to a basic operation (i1/i8/i16).
2031 if (!isTypeSupported(Op0->getType(), VT, /*IsVectorAllowed=*/true) ||
2032 cast<StoreInst>(I)->isAtomic())
2035 // Get the value to be stored into a register. Use the zero register directly
2036 // when possible to avoid an unnecessary copy and a wasted register.
2037 unsigned SrcReg = 0;
2038 if (const auto *CI = dyn_cast<ConstantInt>(Op0)) {
2040 SrcReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
2041 } else if (const auto *CF = dyn_cast<ConstantFP>(Op0)) {
2042 if (CF->isZero() && !CF->isNegative()) {
2043 VT = MVT::getIntegerVT(VT.getSizeInBits());
2044 SrcReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
2049 SrcReg = getRegForValue(Op0);
2054 // See if we can handle this address.
2056 if (!computeAddress(I->getOperand(1), Addr, I->getOperand(0)->getType()))
2059 if (!emitStore(VT, SrcReg, Addr, createMachineMemOperandFor(I)))
2064 static AArch64CC::CondCode getCompareCC(CmpInst::Predicate Pred) {
2066 case CmpInst::FCMP_ONE:
2067 case CmpInst::FCMP_UEQ:
2069 // AL is our "false" for now. The other two need more compares.
2070 return AArch64CC::AL;
2071 case CmpInst::ICMP_EQ:
2072 case CmpInst::FCMP_OEQ:
2073 return AArch64CC::EQ;
2074 case CmpInst::ICMP_SGT:
2075 case CmpInst::FCMP_OGT:
2076 return AArch64CC::GT;
2077 case CmpInst::ICMP_SGE:
2078 case CmpInst::FCMP_OGE:
2079 return AArch64CC::GE;
2080 case CmpInst::ICMP_UGT:
2081 case CmpInst::FCMP_UGT:
2082 return AArch64CC::HI;
2083 case CmpInst::FCMP_OLT:
2084 return AArch64CC::MI;
2085 case CmpInst::ICMP_ULE:
2086 case CmpInst::FCMP_OLE:
2087 return AArch64CC::LS;
2088 case CmpInst::FCMP_ORD:
2089 return AArch64CC::VC;
2090 case CmpInst::FCMP_UNO:
2091 return AArch64CC::VS;
2092 case CmpInst::FCMP_UGE:
2093 return AArch64CC::PL;
2094 case CmpInst::ICMP_SLT:
2095 case CmpInst::FCMP_ULT:
2096 return AArch64CC::LT;
2097 case CmpInst::ICMP_SLE:
2098 case CmpInst::FCMP_ULE:
2099 return AArch64CC::LE;
2100 case CmpInst::FCMP_UNE:
2101 case CmpInst::ICMP_NE:
2102 return AArch64CC::NE;
2103 case CmpInst::ICMP_UGE:
2104 return AArch64CC::HS;
2105 case CmpInst::ICMP_ULT:
2106 return AArch64CC::LO;
2110 /// \brief Try to emit a combined compare-and-branch instruction.
2111 bool AArch64FastISel::emitCompareAndBranch(const BranchInst *BI) {
2112 assert(isa<CmpInst>(BI->getCondition()) && "Expected cmp instruction");
2113 const CmpInst *CI = cast<CmpInst>(BI->getCondition());
2114 CmpInst::Predicate Predicate = optimizeCmpPredicate(CI);
2116 const Value *LHS = CI->getOperand(0);
2117 const Value *RHS = CI->getOperand(1);
2120 if (!isTypeSupported(LHS->getType(), VT))
2123 unsigned BW = VT.getSizeInBits();
2127 MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
2128 MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
2130 // Try to take advantage of fallthrough opportunities.
2131 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
2132 std::swap(TBB, FBB);
2133 Predicate = CmpInst::getInversePredicate(Predicate);
2138 switch (Predicate) {
2141 case CmpInst::ICMP_EQ:
2142 case CmpInst::ICMP_NE:
2143 if (isa<Constant>(LHS) && cast<Constant>(LHS)->isNullValue())
2144 std::swap(LHS, RHS);
2146 if (!isa<Constant>(RHS) || !cast<Constant>(RHS)->isNullValue())
2149 if (const auto *AI = dyn_cast<BinaryOperator>(LHS))
2150 if (AI->getOpcode() == Instruction::And && isValueAvailable(AI)) {
2151 const Value *AndLHS = AI->getOperand(0);
2152 const Value *AndRHS = AI->getOperand(1);
2154 if (const auto *C = dyn_cast<ConstantInt>(AndLHS))
2155 if (C->getValue().isPowerOf2())
2156 std::swap(AndLHS, AndRHS);
2158 if (const auto *C = dyn_cast<ConstantInt>(AndRHS))
2159 if (C->getValue().isPowerOf2()) {
2160 TestBit = C->getValue().logBase2();
2168 IsCmpNE = Predicate == CmpInst::ICMP_NE;
2170 case CmpInst::ICMP_SLT:
2171 case CmpInst::ICMP_SGE:
2172 if (!isa<Constant>(RHS) || !cast<Constant>(RHS)->isNullValue())
2176 IsCmpNE = Predicate == CmpInst::ICMP_SLT;
2178 case CmpInst::ICMP_SGT:
2179 case CmpInst::ICMP_SLE:
2180 if (!isa<ConstantInt>(RHS))
2183 if (cast<ConstantInt>(RHS)->getValue() != APInt(BW, -1, true))
2187 IsCmpNE = Predicate == CmpInst::ICMP_SLE;
2191 static const unsigned OpcTable[2][2][2] = {
2192 { {AArch64::CBZW, AArch64::CBZX },
2193 {AArch64::CBNZW, AArch64::CBNZX} },
2194 { {AArch64::TBZW, AArch64::TBZX },
2195 {AArch64::TBNZW, AArch64::TBNZX} }
2198 bool IsBitTest = TestBit != -1;
2199 bool Is64Bit = BW == 64;
2200 if (TestBit < 32 && TestBit >= 0)
2203 unsigned Opc = OpcTable[IsBitTest][IsCmpNE][Is64Bit];
2204 const MCInstrDesc &II = TII.get(Opc);
2206 unsigned SrcReg = getRegForValue(LHS);
2209 bool SrcIsKill = hasTrivialKill(LHS);
2211 if (BW == 64 && !Is64Bit)
2212 SrcReg = fastEmitInst_extractsubreg(MVT::i32, SrcReg, SrcIsKill,
2215 if ((BW < 32) && !IsBitTest)
2216 SrcReg = emitIntExt(VT, SrcReg, MVT::i32, /*IsZExt=*/true);
2218 // Emit the combined compare and branch instruction.
2219 SrcReg = constrainOperandRegClass(II, SrcReg, II.getNumDefs());
2220 MachineInstrBuilder MIB =
2221 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc))
2222 .addReg(SrcReg, getKillRegState(SrcIsKill));
2224 MIB.addImm(TestBit);
2227 // Obtain the branch weight and add the TrueBB to the successor list.
2228 uint32_t BranchWeight = 0;
2230 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
2231 TBB->getBasicBlock());
2232 FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
2233 fastEmitBranch(FBB, DbgLoc);
2238 bool AArch64FastISel::selectBranch(const Instruction *I) {
2239 const BranchInst *BI = cast<BranchInst>(I);
2240 if (BI->isUnconditional()) {
2241 MachineBasicBlock *MSucc = FuncInfo.MBBMap[BI->getSuccessor(0)];
2242 fastEmitBranch(MSucc, BI->getDebugLoc());
2246 MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
2247 MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
2249 AArch64CC::CondCode CC = AArch64CC::NE;
2250 if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
2251 if (CI->hasOneUse() && isValueAvailable(CI)) {
2252 // Try to optimize or fold the cmp.
2253 CmpInst::Predicate Predicate = optimizeCmpPredicate(CI);
2254 switch (Predicate) {
2257 case CmpInst::FCMP_FALSE:
2258 fastEmitBranch(FBB, DbgLoc);
2260 case CmpInst::FCMP_TRUE:
2261 fastEmitBranch(TBB, DbgLoc);
2265 // Try to emit a combined compare-and-branch first.
2266 if (emitCompareAndBranch(BI))
2269 // Try to take advantage of fallthrough opportunities.
2270 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
2271 std::swap(TBB, FBB);
2272 Predicate = CmpInst::getInversePredicate(Predicate);
2276 if (!emitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
2279 // FCMP_UEQ and FCMP_ONE cannot be checked with a single branch
2281 CC = getCompareCC(Predicate);
2282 AArch64CC::CondCode ExtraCC = AArch64CC::AL;
2283 switch (Predicate) {
2286 case CmpInst::FCMP_UEQ:
2287 ExtraCC = AArch64CC::EQ;
2290 case CmpInst::FCMP_ONE:
2291 ExtraCC = AArch64CC::MI;
2295 assert((CC != AArch64CC::AL) && "Unexpected condition code.");
2297 // Emit the extra branch for FCMP_UEQ and FCMP_ONE.
2298 if (ExtraCC != AArch64CC::AL) {
2299 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
2305 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
2309 // Obtain the branch weight and add the TrueBB to the successor list.
2310 uint32_t BranchWeight = 0;
2312 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
2313 TBB->getBasicBlock());
2314 FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
2316 fastEmitBranch(FBB, DbgLoc);
2319 } else if (TruncInst *TI = dyn_cast<TruncInst>(BI->getCondition())) {
2321 if (TI->hasOneUse() && isValueAvailable(TI) &&
2322 isTypeSupported(TI->getOperand(0)->getType(), SrcVT)) {
2323 unsigned CondReg = getRegForValue(TI->getOperand(0));
2326 bool CondIsKill = hasTrivialKill(TI->getOperand(0));
2328 // Issue an extract_subreg to get the lower 32-bits.
2329 if (SrcVT == MVT::i64) {
2330 CondReg = fastEmitInst_extractsubreg(MVT::i32, CondReg, CondIsKill,
2335 unsigned ANDReg = emitAnd_ri(MVT::i32, CondReg, CondIsKill, 1);
2336 assert(ANDReg && "Unexpected AND instruction emission failure.");
2337 emitICmp_ri(MVT::i32, ANDReg, /*IsKill=*/true, 0);
2339 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
2340 std::swap(TBB, FBB);
2343 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
2347 // Obtain the branch weight and add the TrueBB to the successor list.
2348 uint32_t BranchWeight = 0;
2350 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
2351 TBB->getBasicBlock());
2352 FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
2354 fastEmitBranch(FBB, DbgLoc);
2357 } else if (const auto *CI = dyn_cast<ConstantInt>(BI->getCondition())) {
2358 uint64_t Imm = CI->getZExtValue();
2359 MachineBasicBlock *Target = (Imm == 0) ? FBB : TBB;
2360 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::B))
2363 // Obtain the branch weight and add the target to the successor list.
2364 uint32_t BranchWeight = 0;
2366 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
2367 Target->getBasicBlock());
2368 FuncInfo.MBB->addSuccessor(Target, BranchWeight);
2370 } else if (foldXALUIntrinsic(CC, I, BI->getCondition())) {
2371 // Fake request the condition, otherwise the intrinsic might be completely
2373 unsigned CondReg = getRegForValue(BI->getCondition());
2378 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
2382 // Obtain the branch weight and add the TrueBB to the successor list.
2383 uint32_t BranchWeight = 0;
2385 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
2386 TBB->getBasicBlock());
2387 FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
2389 fastEmitBranch(FBB, DbgLoc);
2393 unsigned CondReg = getRegForValue(BI->getCondition());
2396 bool CondRegIsKill = hasTrivialKill(BI->getCondition());
2398 // We've been divorced from our compare! Our block was split, and
2399 // now our compare lives in a predecessor block. We musn't
2400 // re-compare here, as the children of the compare aren't guaranteed
2401 // live across the block boundary (we *could* check for this).
2402 // Regardless, the compare has been done in the predecessor block,
2403 // and it left a value for us in a virtual register. Ergo, we test
2404 // the one-bit value left in the virtual register.
2405 emitICmp_ri(MVT::i32, CondReg, CondRegIsKill, 0);
2407 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
2408 std::swap(TBB, FBB);
2412 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
2416 // Obtain the branch weight and add the TrueBB to the successor list.
2417 uint32_t BranchWeight = 0;
2419 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
2420 TBB->getBasicBlock());
2421 FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
2423 fastEmitBranch(FBB, DbgLoc);
2427 bool AArch64FastISel::selectIndirectBr(const Instruction *I) {
2428 const IndirectBrInst *BI = cast<IndirectBrInst>(I);
2429 unsigned AddrReg = getRegForValue(BI->getOperand(0));
2433 // Emit the indirect branch.
2434 const MCInstrDesc &II = TII.get(AArch64::BR);
2435 AddrReg = constrainOperandRegClass(II, AddrReg, II.getNumDefs());
2436 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II).addReg(AddrReg);
2438 // Make sure the CFG is up-to-date.
2439 for (unsigned i = 0, e = BI->getNumSuccessors(); i != e; ++i)
2440 FuncInfo.MBB->addSuccessor(FuncInfo.MBBMap[BI->getSuccessor(i)]);
2445 bool AArch64FastISel::selectCmp(const Instruction *I) {
2446 const CmpInst *CI = cast<CmpInst>(I);
2448 // Try to optimize or fold the cmp.
2449 CmpInst::Predicate Predicate = optimizeCmpPredicate(CI);
2450 unsigned ResultReg = 0;
2451 switch (Predicate) {
2454 case CmpInst::FCMP_FALSE:
2455 ResultReg = createResultReg(&AArch64::GPR32RegClass);
2456 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2457 TII.get(TargetOpcode::COPY), ResultReg)
2458 .addReg(AArch64::WZR, getKillRegState(true));
2460 case CmpInst::FCMP_TRUE:
2461 ResultReg = fastEmit_i(MVT::i32, MVT::i32, ISD::Constant, 1);
2466 updateValueMap(I, ResultReg);
2471 if (!emitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
2474 ResultReg = createResultReg(&AArch64::GPR32RegClass);
2476 // FCMP_UEQ and FCMP_ONE cannot be checked with a single instruction. These
2477 // condition codes are inverted, because they are used by CSINC.
2478 static unsigned CondCodeTable[2][2] = {
2479 { AArch64CC::NE, AArch64CC::VC },
2480 { AArch64CC::PL, AArch64CC::LE }
2482 unsigned *CondCodes = nullptr;
2483 switch (Predicate) {
2486 case CmpInst::FCMP_UEQ:
2487 CondCodes = &CondCodeTable[0][0];
2489 case CmpInst::FCMP_ONE:
2490 CondCodes = &CondCodeTable[1][0];
2495 unsigned TmpReg1 = createResultReg(&AArch64::GPR32RegClass);
2496 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::CSINCWr),
2498 .addReg(AArch64::WZR, getKillRegState(true))
2499 .addReg(AArch64::WZR, getKillRegState(true))
2500 .addImm(CondCodes[0]);
2501 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::CSINCWr),
2503 .addReg(TmpReg1, getKillRegState(true))
2504 .addReg(AArch64::WZR, getKillRegState(true))
2505 .addImm(CondCodes[1]);
2507 updateValueMap(I, ResultReg);
2511 // Now set a register based on the comparison.
2512 AArch64CC::CondCode CC = getCompareCC(Predicate);
2513 assert((CC != AArch64CC::AL) && "Unexpected condition code.");
2514 AArch64CC::CondCode invertedCC = getInvertedCondCode(CC);
2515 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::CSINCWr),
2517 .addReg(AArch64::WZR, getKillRegState(true))
2518 .addReg(AArch64::WZR, getKillRegState(true))
2519 .addImm(invertedCC);
2521 updateValueMap(I, ResultReg);
2525 /// \brief Optimize selects of i1 if one of the operands has a 'true' or 'false'
2527 bool AArch64FastISel::optimizeSelect(const SelectInst *SI) {
2528 if (!SI->getType()->isIntegerTy(1))
2531 const Value *Src1Val, *Src2Val;
2533 bool NeedExtraOp = false;
2534 if (auto *CI = dyn_cast<ConstantInt>(SI->getTrueValue())) {
2536 Src1Val = SI->getCondition();
2537 Src2Val = SI->getFalseValue();
2538 Opc = AArch64::ORRWrr;
2540 assert(CI->isZero());
2541 Src1Val = SI->getFalseValue();
2542 Src2Val = SI->getCondition();
2543 Opc = AArch64::BICWrr;
2545 } else if (auto *CI = dyn_cast<ConstantInt>(SI->getFalseValue())) {
2547 Src1Val = SI->getCondition();
2548 Src2Val = SI->getTrueValue();
2549 Opc = AArch64::ORRWrr;
2552 assert(CI->isZero());
2553 Src1Val = SI->getCondition();
2554 Src2Val = SI->getTrueValue();
2555 Opc = AArch64::ANDWrr;
2562 unsigned Src1Reg = getRegForValue(Src1Val);
2565 bool Src1IsKill = hasTrivialKill(Src1Val);
2567 unsigned Src2Reg = getRegForValue(Src2Val);
2570 bool Src2IsKill = hasTrivialKill(Src2Val);
2573 Src1Reg = emitLogicalOp_ri(ISD::XOR, MVT::i32, Src1Reg, Src1IsKill, 1);
2576 unsigned ResultReg = fastEmitInst_rr(Opc, &AArch64::GPR32spRegClass, Src1Reg,
2577 Src1IsKill, Src2Reg, Src2IsKill);
2578 updateValueMap(SI, ResultReg);
2582 bool AArch64FastISel::selectSelect(const Instruction *I) {
2583 assert(isa<SelectInst>(I) && "Expected a select instruction.");
2585 if (!isTypeSupported(I->getType(), VT))
2589 const TargetRegisterClass *RC;
2590 switch (VT.SimpleTy) {
2597 Opc = AArch64::CSELWr;
2598 RC = &AArch64::GPR32RegClass;
2601 Opc = AArch64::CSELXr;
2602 RC = &AArch64::GPR64RegClass;
2605 Opc = AArch64::FCSELSrrr;
2606 RC = &AArch64::FPR32RegClass;
2609 Opc = AArch64::FCSELDrrr;
2610 RC = &AArch64::FPR64RegClass;
2614 const SelectInst *SI = cast<SelectInst>(I);
2615 const Value *Cond = SI->getCondition();
2616 AArch64CC::CondCode CC = AArch64CC::NE;
2617 AArch64CC::CondCode ExtraCC = AArch64CC::AL;
2619 if (optimizeSelect(SI))
2622 // Try to pickup the flags, so we don't have to emit another compare.
2623 if (foldXALUIntrinsic(CC, I, Cond)) {
2624 // Fake request the condition to force emission of the XALU intrinsic.
2625 unsigned CondReg = getRegForValue(Cond);
2628 } else if (isa<CmpInst>(Cond) && cast<CmpInst>(Cond)->hasOneUse() &&
2629 isValueAvailable(Cond)) {
2630 const auto *Cmp = cast<CmpInst>(Cond);
2631 // Try to optimize or fold the cmp.
2632 CmpInst::Predicate Predicate = optimizeCmpPredicate(Cmp);
2633 const Value *FoldSelect = nullptr;
2634 switch (Predicate) {
2637 case CmpInst::FCMP_FALSE:
2638 FoldSelect = SI->getFalseValue();
2640 case CmpInst::FCMP_TRUE:
2641 FoldSelect = SI->getTrueValue();
2646 unsigned SrcReg = getRegForValue(FoldSelect);
2649 unsigned UseReg = lookUpRegForValue(SI);
2651 MRI.clearKillFlags(UseReg);
2653 updateValueMap(I, SrcReg);
2658 if (!emitCmp(Cmp->getOperand(0), Cmp->getOperand(1), Cmp->isUnsigned()))
2661 // FCMP_UEQ and FCMP_ONE cannot be checked with a single select instruction.
2662 CC = getCompareCC(Predicate);
2663 switch (Predicate) {
2666 case CmpInst::FCMP_UEQ:
2667 ExtraCC = AArch64CC::EQ;
2670 case CmpInst::FCMP_ONE:
2671 ExtraCC = AArch64CC::MI;
2675 assert((CC != AArch64CC::AL) && "Unexpected condition code.");
2677 unsigned CondReg = getRegForValue(Cond);
2680 bool CondIsKill = hasTrivialKill(Cond);
2682 // Emit a TST instruction (ANDS wzr, reg, #imm).
2683 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ANDSWri),
2685 .addReg(CondReg, getKillRegState(CondIsKill))
2686 .addImm(AArch64_AM::encodeLogicalImmediate(1, 32));
2689 unsigned Src1Reg = getRegForValue(SI->getTrueValue());
2690 bool Src1IsKill = hasTrivialKill(SI->getTrueValue());
2692 unsigned Src2Reg = getRegForValue(SI->getFalseValue());
2693 bool Src2IsKill = hasTrivialKill(SI->getFalseValue());
2695 if (!Src1Reg || !Src2Reg)
2698 if (ExtraCC != AArch64CC::AL) {
2699 Src2Reg = fastEmitInst_rri(Opc, RC, Src1Reg, Src1IsKill, Src2Reg,
2700 Src2IsKill, ExtraCC);
2703 unsigned ResultReg = fastEmitInst_rri(Opc, RC, Src1Reg, Src1IsKill, Src2Reg,
2705 updateValueMap(I, ResultReg);
2709 bool AArch64FastISel::selectFPExt(const Instruction *I) {
2710 Value *V = I->getOperand(0);
2711 if (!I->getType()->isDoubleTy() || !V->getType()->isFloatTy())
2714 unsigned Op = getRegForValue(V);
2718 unsigned ResultReg = createResultReg(&AArch64::FPR64RegClass);
2719 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::FCVTDSr),
2720 ResultReg).addReg(Op);
2721 updateValueMap(I, ResultReg);
2725 bool AArch64FastISel::selectFPTrunc(const Instruction *I) {
2726 Value *V = I->getOperand(0);
2727 if (!I->getType()->isFloatTy() || !V->getType()->isDoubleTy())
2730 unsigned Op = getRegForValue(V);
2734 unsigned ResultReg = createResultReg(&AArch64::FPR32RegClass);
2735 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::FCVTSDr),
2736 ResultReg).addReg(Op);
2737 updateValueMap(I, ResultReg);
2741 // FPToUI and FPToSI
2742 bool AArch64FastISel::selectFPToInt(const Instruction *I, bool Signed) {
2744 if (!isTypeLegal(I->getType(), DestVT) || DestVT.isVector())
2747 unsigned SrcReg = getRegForValue(I->getOperand(0));
2751 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType(), true);
2752 if (SrcVT == MVT::f128)
2756 if (SrcVT == MVT::f64) {
2758 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZSUWDr : AArch64::FCVTZSUXDr;
2760 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZUUWDr : AArch64::FCVTZUUXDr;
2763 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZSUWSr : AArch64::FCVTZSUXSr;
2765 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZUUWSr : AArch64::FCVTZUUXSr;
2767 unsigned ResultReg = createResultReg(
2768 DestVT == MVT::i32 ? &AArch64::GPR32RegClass : &AArch64::GPR64RegClass);
2769 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
2771 updateValueMap(I, ResultReg);
2775 bool AArch64FastISel::selectIntToFP(const Instruction *I, bool Signed) {
2777 if (!isTypeLegal(I->getType(), DestVT) || DestVT.isVector())
2779 assert ((DestVT == MVT::f32 || DestVT == MVT::f64) &&
2780 "Unexpected value type.");
2782 unsigned SrcReg = getRegForValue(I->getOperand(0));
2785 bool SrcIsKill = hasTrivialKill(I->getOperand(0));
2787 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType(), true);
2789 // Handle sign-extension.
2790 if (SrcVT == MVT::i16 || SrcVT == MVT::i8 || SrcVT == MVT::i1) {
2792 emitIntExt(SrcVT.getSimpleVT(), SrcReg, MVT::i32, /*isZExt*/ !Signed);
2799 if (SrcVT == MVT::i64) {
2801 Opc = (DestVT == MVT::f32) ? AArch64::SCVTFUXSri : AArch64::SCVTFUXDri;
2803 Opc = (DestVT == MVT::f32) ? AArch64::UCVTFUXSri : AArch64::UCVTFUXDri;
2806 Opc = (DestVT == MVT::f32) ? AArch64::SCVTFUWSri : AArch64::SCVTFUWDri;
2808 Opc = (DestVT == MVT::f32) ? AArch64::UCVTFUWSri : AArch64::UCVTFUWDri;
2811 unsigned ResultReg = fastEmitInst_r(Opc, TLI.getRegClassFor(DestVT), SrcReg,
2813 updateValueMap(I, ResultReg);
2817 bool AArch64FastISel::fastLowerArguments() {
2818 if (!FuncInfo.CanLowerReturn)
2821 const Function *F = FuncInfo.Fn;
2825 CallingConv::ID CC = F->getCallingConv();
2826 if (CC != CallingConv::C)
2829 // Only handle simple cases of up to 8 GPR and FPR each.
2830 unsigned GPRCnt = 0;
2831 unsigned FPRCnt = 0;
2833 for (auto const &Arg : F->args()) {
2834 // The first argument is at index 1.
2836 if (F->getAttributes().hasAttribute(Idx, Attribute::ByVal) ||
2837 F->getAttributes().hasAttribute(Idx, Attribute::InReg) ||
2838 F->getAttributes().hasAttribute(Idx, Attribute::StructRet) ||
2839 F->getAttributes().hasAttribute(Idx, Attribute::Nest))
2842 Type *ArgTy = Arg.getType();
2843 if (ArgTy->isStructTy() || ArgTy->isArrayTy())
2846 EVT ArgVT = TLI.getValueType(ArgTy);
2847 if (!ArgVT.isSimple())
2850 MVT VT = ArgVT.getSimpleVT().SimpleTy;
2851 if (VT.isFloatingPoint() && !Subtarget->hasFPARMv8())
2854 if (VT.isVector() &&
2855 (!Subtarget->hasNEON() || !Subtarget->isLittleEndian()))
2858 if (VT >= MVT::i1 && VT <= MVT::i64)
2860 else if ((VT >= MVT::f16 && VT <= MVT::f64) || VT.is64BitVector() ||
2861 VT.is128BitVector())
2866 if (GPRCnt > 8 || FPRCnt > 8)
2870 static const MCPhysReg Registers[6][8] = {
2871 { AArch64::W0, AArch64::W1, AArch64::W2, AArch64::W3, AArch64::W4,
2872 AArch64::W5, AArch64::W6, AArch64::W7 },
2873 { AArch64::X0, AArch64::X1, AArch64::X2, AArch64::X3, AArch64::X4,
2874 AArch64::X5, AArch64::X6, AArch64::X7 },
2875 { AArch64::H0, AArch64::H1, AArch64::H2, AArch64::H3, AArch64::H4,
2876 AArch64::H5, AArch64::H6, AArch64::H7 },
2877 { AArch64::S0, AArch64::S1, AArch64::S2, AArch64::S3, AArch64::S4,
2878 AArch64::S5, AArch64::S6, AArch64::S7 },
2879 { AArch64::D0, AArch64::D1, AArch64::D2, AArch64::D3, AArch64::D4,
2880 AArch64::D5, AArch64::D6, AArch64::D7 },
2881 { AArch64::Q0, AArch64::Q1, AArch64::Q2, AArch64::Q3, AArch64::Q4,
2882 AArch64::Q5, AArch64::Q6, AArch64::Q7 }
2885 unsigned GPRIdx = 0;
2886 unsigned FPRIdx = 0;
2887 for (auto const &Arg : F->args()) {
2888 MVT VT = TLI.getSimpleValueType(Arg.getType());
2890 const TargetRegisterClass *RC;
2891 if (VT >= MVT::i1 && VT <= MVT::i32) {
2892 SrcReg = Registers[0][GPRIdx++];
2893 RC = &AArch64::GPR32RegClass;
2895 } else if (VT == MVT::i64) {
2896 SrcReg = Registers[1][GPRIdx++];
2897 RC = &AArch64::GPR64RegClass;
2898 } else if (VT == MVT::f16) {
2899 SrcReg = Registers[2][FPRIdx++];
2900 RC = &AArch64::FPR16RegClass;
2901 } else if (VT == MVT::f32) {
2902 SrcReg = Registers[3][FPRIdx++];
2903 RC = &AArch64::FPR32RegClass;
2904 } else if ((VT == MVT::f64) || VT.is64BitVector()) {
2905 SrcReg = Registers[4][FPRIdx++];
2906 RC = &AArch64::FPR64RegClass;
2907 } else if (VT.is128BitVector()) {
2908 SrcReg = Registers[5][FPRIdx++];
2909 RC = &AArch64::FPR128RegClass;
2911 llvm_unreachable("Unexpected value type.");
2913 unsigned DstReg = FuncInfo.MF->addLiveIn(SrcReg, RC);
2914 // FIXME: Unfortunately it's necessary to emit a copy from the livein copy.
2915 // Without this, EmitLiveInCopies may eliminate the livein if its only
2916 // use is a bitcast (which isn't turned into an instruction).
2917 unsigned ResultReg = createResultReg(RC);
2918 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2919 TII.get(TargetOpcode::COPY), ResultReg)
2920 .addReg(DstReg, getKillRegState(true));
2921 updateValueMap(&Arg, ResultReg);
2926 bool AArch64FastISel::processCallArgs(CallLoweringInfo &CLI,
2927 SmallVectorImpl<MVT> &OutVTs,
2928 unsigned &NumBytes) {
2929 CallingConv::ID CC = CLI.CallConv;
2930 SmallVector<CCValAssign, 16> ArgLocs;
2931 CCState CCInfo(CC, false, *FuncInfo.MF, ArgLocs, *Context);
2932 CCInfo.AnalyzeCallOperands(OutVTs, CLI.OutFlags, CCAssignFnForCall(CC));
2934 // Get a count of how many bytes are to be pushed on the stack.
2935 NumBytes = CCInfo.getNextStackOffset();
2937 // Issue CALLSEQ_START
2938 unsigned AdjStackDown = TII.getCallFrameSetupOpcode();
2939 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackDown))
2942 // Process the args.
2943 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2944 CCValAssign &VA = ArgLocs[i];
2945 const Value *ArgVal = CLI.OutVals[VA.getValNo()];
2946 MVT ArgVT = OutVTs[VA.getValNo()];
2948 unsigned ArgReg = getRegForValue(ArgVal);
2952 // Handle arg promotion: SExt, ZExt, AExt.
2953 switch (VA.getLocInfo()) {
2954 case CCValAssign::Full:
2956 case CCValAssign::SExt: {
2957 MVT DestVT = VA.getLocVT();
2959 ArgReg = emitIntExt(SrcVT, ArgReg, DestVT, /*isZExt=*/false);
2964 case CCValAssign::AExt:
2965 // Intentional fall-through.
2966 case CCValAssign::ZExt: {
2967 MVT DestVT = VA.getLocVT();
2969 ArgReg = emitIntExt(SrcVT, ArgReg, DestVT, /*isZExt=*/true);
2975 llvm_unreachable("Unknown arg promotion!");
2978 // Now copy/store arg to correct locations.
2979 if (VA.isRegLoc() && !VA.needsCustom()) {
2980 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2981 TII.get(TargetOpcode::COPY), VA.getLocReg()).addReg(ArgReg);
2982 CLI.OutRegs.push_back(VA.getLocReg());
2983 } else if (VA.needsCustom()) {
2984 // FIXME: Handle custom args.
2987 assert(VA.isMemLoc() && "Assuming store on stack.");
2989 // Don't emit stores for undef values.
2990 if (isa<UndefValue>(ArgVal))
2993 // Need to store on the stack.
2994 unsigned ArgSize = (ArgVT.getSizeInBits() + 7) / 8;
2996 unsigned BEAlign = 0;
2997 if (ArgSize < 8 && !Subtarget->isLittleEndian())
2998 BEAlign = 8 - ArgSize;
3001 Addr.setKind(Address::RegBase);
3002 Addr.setReg(AArch64::SP);
3003 Addr.setOffset(VA.getLocMemOffset() + BEAlign);
3005 unsigned Alignment = DL.getABITypeAlignment(ArgVal->getType());
3006 MachineMemOperand *MMO = FuncInfo.MF->getMachineMemOperand(
3007 MachinePointerInfo::getStack(Addr.getOffset()),
3008 MachineMemOperand::MOStore, ArgVT.getStoreSize(), Alignment);
3010 if (!emitStore(ArgVT, ArgReg, Addr, MMO))
3017 bool AArch64FastISel::finishCall(CallLoweringInfo &CLI, MVT RetVT,
3018 unsigned NumBytes) {
3019 CallingConv::ID CC = CLI.CallConv;
3021 // Issue CALLSEQ_END
3022 unsigned AdjStackUp = TII.getCallFrameDestroyOpcode();
3023 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackUp))
3024 .addImm(NumBytes).addImm(0);
3026 // Now the return value.
3027 if (RetVT != MVT::isVoid) {
3028 SmallVector<CCValAssign, 16> RVLocs;
3029 CCState CCInfo(CC, false, *FuncInfo.MF, RVLocs, *Context);
3030 CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC));
3032 // Only handle a single return value.
3033 if (RVLocs.size() != 1)
3036 // Copy all of the result registers out of their specified physreg.
3037 MVT CopyVT = RVLocs[0].getValVT();
3039 // TODO: Handle big-endian results
3040 if (CopyVT.isVector() && !Subtarget->isLittleEndian())
3043 unsigned ResultReg = createResultReg(TLI.getRegClassFor(CopyVT));
3044 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3045 TII.get(TargetOpcode::COPY), ResultReg)
3046 .addReg(RVLocs[0].getLocReg());
3047 CLI.InRegs.push_back(RVLocs[0].getLocReg());
3049 CLI.ResultReg = ResultReg;
3050 CLI.NumResultRegs = 1;
3056 bool AArch64FastISel::fastLowerCall(CallLoweringInfo &CLI) {
3057 CallingConv::ID CC = CLI.CallConv;
3058 bool IsTailCall = CLI.IsTailCall;
3059 bool IsVarArg = CLI.IsVarArg;
3060 const Value *Callee = CLI.Callee;
3061 const char *SymName = CLI.SymName;
3063 if (!Callee && !SymName)
3066 // Allow SelectionDAG isel to handle tail calls.
3070 CodeModel::Model CM = TM.getCodeModel();
3071 // Only support the small and large code model.
3072 if (CM != CodeModel::Small && CM != CodeModel::Large)
3075 // FIXME: Add large code model support for ELF.
3076 if (CM == CodeModel::Large && !Subtarget->isTargetMachO())
3079 // Let SDISel handle vararg functions.
3083 // FIXME: Only handle *simple* calls for now.
3085 if (CLI.RetTy->isVoidTy())
3086 RetVT = MVT::isVoid;
3087 else if (!isTypeLegal(CLI.RetTy, RetVT))
3090 for (auto Flag : CLI.OutFlags)
3091 if (Flag.isInReg() || Flag.isSRet() || Flag.isNest() || Flag.isByVal())
3094 // Set up the argument vectors.
3095 SmallVector<MVT, 16> OutVTs;
3096 OutVTs.reserve(CLI.OutVals.size());
3098 for (auto *Val : CLI.OutVals) {
3100 if (!isTypeLegal(Val->getType(), VT) &&
3101 !(VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16))
3104 // We don't handle vector parameters yet.
3105 if (VT.isVector() || VT.getSizeInBits() > 64)
3108 OutVTs.push_back(VT);
3112 if (Callee && !computeCallAddress(Callee, Addr))
3115 // Handle the arguments now that we've gotten them.
3117 if (!processCallArgs(CLI, OutVTs, NumBytes))
3121 MachineInstrBuilder MIB;
3122 if (CM == CodeModel::Small) {
3123 const MCInstrDesc &II = TII.get(Addr.getReg() ? AArch64::BLR : AArch64::BL);
3124 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II);
3126 MIB.addExternalSymbol(SymName, 0);
3127 else if (Addr.getGlobalValue())
3128 MIB.addGlobalAddress(Addr.getGlobalValue(), 0, 0);
3129 else if (Addr.getReg()) {
3130 unsigned Reg = constrainOperandRegClass(II, Addr.getReg(), 0);
3135 unsigned CallReg = 0;
3137 unsigned ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
3138 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
3140 .addExternalSymbol(SymName, AArch64II::MO_GOT | AArch64II::MO_PAGE);
3142 CallReg = createResultReg(&AArch64::GPR64RegClass);
3143 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::LDRXui),
3146 .addExternalSymbol(SymName, AArch64II::MO_GOT | AArch64II::MO_PAGEOFF |
3148 } else if (Addr.getGlobalValue())
3149 CallReg = materializeGV(Addr.getGlobalValue());
3150 else if (Addr.getReg())
3151 CallReg = Addr.getReg();
3156 const MCInstrDesc &II = TII.get(AArch64::BLR);
3157 CallReg = constrainOperandRegClass(II, CallReg, 0);
3158 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II).addReg(CallReg);
3161 // Add implicit physical register uses to the call.
3162 for (auto Reg : CLI.OutRegs)
3163 MIB.addReg(Reg, RegState::Implicit);
3165 // Add a register mask with the call-preserved registers.
3166 // Proper defs for return values will be added by setPhysRegsDeadExcept().
3167 MIB.addRegMask(TRI.getCallPreservedMask(*FuncInfo.MF, CC));
3171 // Finish off the call including any return values.
3172 return finishCall(CLI, RetVT, NumBytes);
3175 bool AArch64FastISel::isMemCpySmall(uint64_t Len, unsigned Alignment) {
3177 return Len / Alignment <= 4;
3182 bool AArch64FastISel::tryEmitSmallMemCpy(Address Dest, Address Src,
3183 uint64_t Len, unsigned Alignment) {
3184 // Make sure we don't bloat code by inlining very large memcpy's.
3185 if (!isMemCpySmall(Len, Alignment))
3188 int64_t UnscaledOffset = 0;
3189 Address OrigDest = Dest;
3190 Address OrigSrc = Src;
3194 if (!Alignment || Alignment >= 8) {
3205 // Bound based on alignment.
3206 if (Len >= 4 && Alignment == 4)
3208 else if (Len >= 2 && Alignment == 2)
3215 unsigned ResultReg = emitLoad(VT, VT, Src);
3219 if (!emitStore(VT, ResultReg, Dest))
3222 int64_t Size = VT.getSizeInBits() / 8;
3224 UnscaledOffset += Size;
3226 // We need to recompute the unscaled offset for each iteration.
3227 Dest.setOffset(OrigDest.getOffset() + UnscaledOffset);
3228 Src.setOffset(OrigSrc.getOffset() + UnscaledOffset);
3234 /// \brief Check if it is possible to fold the condition from the XALU intrinsic
3235 /// into the user. The condition code will only be updated on success.
3236 bool AArch64FastISel::foldXALUIntrinsic(AArch64CC::CondCode &CC,
3237 const Instruction *I,
3238 const Value *Cond) {
3239 if (!isa<ExtractValueInst>(Cond))
3242 const auto *EV = cast<ExtractValueInst>(Cond);
3243 if (!isa<IntrinsicInst>(EV->getAggregateOperand()))
3246 const auto *II = cast<IntrinsicInst>(EV->getAggregateOperand());
3248 const Function *Callee = II->getCalledFunction();
3250 cast<StructType>(Callee->getReturnType())->getTypeAtIndex(0U);
3251 if (!isTypeLegal(RetTy, RetVT))
3254 if (RetVT != MVT::i32 && RetVT != MVT::i64)
3257 const Value *LHS = II->getArgOperand(0);
3258 const Value *RHS = II->getArgOperand(1);
3260 // Canonicalize immediate to the RHS.
3261 if (isa<ConstantInt>(LHS) && !isa<ConstantInt>(RHS) &&
3262 isCommutativeIntrinsic(II))
3263 std::swap(LHS, RHS);
3265 // Simplify multiplies.
3266 unsigned IID = II->getIntrinsicID();
3270 case Intrinsic::smul_with_overflow:
3271 if (const auto *C = dyn_cast<ConstantInt>(RHS))
3272 if (C->getValue() == 2)
3273 IID = Intrinsic::sadd_with_overflow;
3275 case Intrinsic::umul_with_overflow:
3276 if (const auto *C = dyn_cast<ConstantInt>(RHS))
3277 if (C->getValue() == 2)
3278 IID = Intrinsic::uadd_with_overflow;
3282 AArch64CC::CondCode TmpCC;
3286 case Intrinsic::sadd_with_overflow:
3287 case Intrinsic::ssub_with_overflow:
3288 TmpCC = AArch64CC::VS;
3290 case Intrinsic::uadd_with_overflow:
3291 TmpCC = AArch64CC::HS;
3293 case Intrinsic::usub_with_overflow:
3294 TmpCC = AArch64CC::LO;
3296 case Intrinsic::smul_with_overflow:
3297 case Intrinsic::umul_with_overflow:
3298 TmpCC = AArch64CC::NE;
3302 // Check if both instructions are in the same basic block.
3303 if (!isValueAvailable(II))
3306 // Make sure nothing is in the way
3307 BasicBlock::const_iterator Start = I;
3308 BasicBlock::const_iterator End = II;
3309 for (auto Itr = std::prev(Start); Itr != End; --Itr) {
3310 // We only expect extractvalue instructions between the intrinsic and the
3311 // instruction to be selected.
3312 if (!isa<ExtractValueInst>(Itr))
3315 // Check that the extractvalue operand comes from the intrinsic.
3316 const auto *EVI = cast<ExtractValueInst>(Itr);
3317 if (EVI->getAggregateOperand() != II)
3325 bool AArch64FastISel::fastLowerIntrinsicCall(const IntrinsicInst *II) {
3326 // FIXME: Handle more intrinsics.
3327 switch (II->getIntrinsicID()) {
3328 default: return false;
3329 case Intrinsic::frameaddress: {
3330 MachineFrameInfo *MFI = FuncInfo.MF->getFrameInfo();
3331 MFI->setFrameAddressIsTaken(true);
3333 const AArch64RegisterInfo *RegInfo =
3334 static_cast<const AArch64RegisterInfo *>(Subtarget->getRegisterInfo());
3335 unsigned FramePtr = RegInfo->getFrameRegister(*(FuncInfo.MF));
3336 unsigned SrcReg = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
3337 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3338 TII.get(TargetOpcode::COPY), SrcReg).addReg(FramePtr);
3339 // Recursively load frame address
3345 unsigned Depth = cast<ConstantInt>(II->getOperand(0))->getZExtValue();
3347 DestReg = fastEmitInst_ri(AArch64::LDRXui, &AArch64::GPR64RegClass,
3348 SrcReg, /*IsKill=*/true, 0);
3349 assert(DestReg && "Unexpected LDR instruction emission failure.");
3353 updateValueMap(II, SrcReg);
3356 case Intrinsic::memcpy:
3357 case Intrinsic::memmove: {
3358 const auto *MTI = cast<MemTransferInst>(II);
3359 // Don't handle volatile.
3360 if (MTI->isVolatile())
3363 // Disable inlining for memmove before calls to ComputeAddress. Otherwise,
3364 // we would emit dead code because we don't currently handle memmoves.
3365 bool IsMemCpy = (II->getIntrinsicID() == Intrinsic::memcpy);
3366 if (isa<ConstantInt>(MTI->getLength()) && IsMemCpy) {
3367 // Small memcpy's are common enough that we want to do them without a call
3369 uint64_t Len = cast<ConstantInt>(MTI->getLength())->getZExtValue();
3370 unsigned Alignment = MTI->getAlignment();
3371 if (isMemCpySmall(Len, Alignment)) {
3373 if (!computeAddress(MTI->getRawDest(), Dest) ||
3374 !computeAddress(MTI->getRawSource(), Src))
3376 if (tryEmitSmallMemCpy(Dest, Src, Len, Alignment))
3381 if (!MTI->getLength()->getType()->isIntegerTy(64))
3384 if (MTI->getSourceAddressSpace() > 255 || MTI->getDestAddressSpace() > 255)
3385 // Fast instruction selection doesn't support the special
3389 const char *IntrMemName = isa<MemCpyInst>(II) ? "memcpy" : "memmove";
3390 return lowerCallTo(II, IntrMemName, II->getNumArgOperands() - 2);
3392 case Intrinsic::memset: {
3393 const MemSetInst *MSI = cast<MemSetInst>(II);
3394 // Don't handle volatile.
3395 if (MSI->isVolatile())
3398 if (!MSI->getLength()->getType()->isIntegerTy(64))
3401 if (MSI->getDestAddressSpace() > 255)
3402 // Fast instruction selection doesn't support the special
3406 return lowerCallTo(II, "memset", II->getNumArgOperands() - 2);
3408 case Intrinsic::sin:
3409 case Intrinsic::cos:
3410 case Intrinsic::pow: {
3412 if (!isTypeLegal(II->getType(), RetVT))
3415 if (RetVT != MVT::f32 && RetVT != MVT::f64)
3418 static const RTLIB::Libcall LibCallTable[3][2] = {
3419 { RTLIB::SIN_F32, RTLIB::SIN_F64 },
3420 { RTLIB::COS_F32, RTLIB::COS_F64 },
3421 { RTLIB::POW_F32, RTLIB::POW_F64 }
3424 bool Is64Bit = RetVT == MVT::f64;
3425 switch (II->getIntrinsicID()) {
3427 llvm_unreachable("Unexpected intrinsic.");
3428 case Intrinsic::sin:
3429 LC = LibCallTable[0][Is64Bit];
3431 case Intrinsic::cos:
3432 LC = LibCallTable[1][Is64Bit];
3434 case Intrinsic::pow:
3435 LC = LibCallTable[2][Is64Bit];
3440 Args.reserve(II->getNumArgOperands());
3442 // Populate the argument list.
3443 for (auto &Arg : II->arg_operands()) {
3446 Entry.Ty = Arg->getType();
3447 Args.push_back(Entry);
3450 CallLoweringInfo CLI;
3451 CLI.setCallee(TLI.getLibcallCallingConv(LC), II->getType(),
3452 TLI.getLibcallName(LC), std::move(Args));
3453 if (!lowerCallTo(CLI))
3455 updateValueMap(II, CLI.ResultReg);
3458 case Intrinsic::fabs: {
3460 if (!isTypeLegal(II->getType(), VT))
3464 switch (VT.SimpleTy) {
3468 Opc = AArch64::FABSSr;
3471 Opc = AArch64::FABSDr;
3474 unsigned SrcReg = getRegForValue(II->getOperand(0));
3477 bool SrcRegIsKill = hasTrivialKill(II->getOperand(0));
3478 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
3479 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
3480 .addReg(SrcReg, getKillRegState(SrcRegIsKill));
3481 updateValueMap(II, ResultReg);
3484 case Intrinsic::trap: {
3485 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::BRK))
3489 case Intrinsic::sqrt: {
3490 Type *RetTy = II->getCalledFunction()->getReturnType();
3493 if (!isTypeLegal(RetTy, VT))
3496 unsigned Op0Reg = getRegForValue(II->getOperand(0));
3499 bool Op0IsKill = hasTrivialKill(II->getOperand(0));
3501 unsigned ResultReg = fastEmit_r(VT, VT, ISD::FSQRT, Op0Reg, Op0IsKill);
3505 updateValueMap(II, ResultReg);
3508 case Intrinsic::sadd_with_overflow:
3509 case Intrinsic::uadd_with_overflow:
3510 case Intrinsic::ssub_with_overflow:
3511 case Intrinsic::usub_with_overflow:
3512 case Intrinsic::smul_with_overflow:
3513 case Intrinsic::umul_with_overflow: {
3514 // This implements the basic lowering of the xalu with overflow intrinsics.
3515 const Function *Callee = II->getCalledFunction();
3516 auto *Ty = cast<StructType>(Callee->getReturnType());
3517 Type *RetTy = Ty->getTypeAtIndex(0U);
3520 if (!isTypeLegal(RetTy, VT))
3523 if (VT != MVT::i32 && VT != MVT::i64)
3526 const Value *LHS = II->getArgOperand(0);
3527 const Value *RHS = II->getArgOperand(1);
3528 // Canonicalize immediate to the RHS.
3529 if (isa<ConstantInt>(LHS) && !isa<ConstantInt>(RHS) &&
3530 isCommutativeIntrinsic(II))
3531 std::swap(LHS, RHS);
3533 // Simplify multiplies.
3534 unsigned IID = II->getIntrinsicID();
3538 case Intrinsic::smul_with_overflow:
3539 if (const auto *C = dyn_cast<ConstantInt>(RHS))
3540 if (C->getValue() == 2) {
3541 IID = Intrinsic::sadd_with_overflow;
3545 case Intrinsic::umul_with_overflow:
3546 if (const auto *C = dyn_cast<ConstantInt>(RHS))
3547 if (C->getValue() == 2) {
3548 IID = Intrinsic::uadd_with_overflow;
3554 unsigned ResultReg1 = 0, ResultReg2 = 0, MulReg = 0;
3555 AArch64CC::CondCode CC = AArch64CC::Invalid;
3557 default: llvm_unreachable("Unexpected intrinsic!");
3558 case Intrinsic::sadd_with_overflow:
3559 ResultReg1 = emitAdd(VT, LHS, RHS, /*SetFlags=*/true);
3562 case Intrinsic::uadd_with_overflow:
3563 ResultReg1 = emitAdd(VT, LHS, RHS, /*SetFlags=*/true);
3566 case Intrinsic::ssub_with_overflow:
3567 ResultReg1 = emitSub(VT, LHS, RHS, /*SetFlags=*/true);
3570 case Intrinsic::usub_with_overflow:
3571 ResultReg1 = emitSub(VT, LHS, RHS, /*SetFlags=*/true);
3574 case Intrinsic::smul_with_overflow: {
3576 unsigned LHSReg = getRegForValue(LHS);
3579 bool LHSIsKill = hasTrivialKill(LHS);
3581 unsigned RHSReg = getRegForValue(RHS);
3584 bool RHSIsKill = hasTrivialKill(RHS);
3586 if (VT == MVT::i32) {
3587 MulReg = emitSMULL_rr(MVT::i64, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
3588 unsigned ShiftReg = emitLSR_ri(MVT::i64, MVT::i64, MulReg,
3589 /*IsKill=*/false, 32);
3590 MulReg = fastEmitInst_extractsubreg(VT, MulReg, /*IsKill=*/true,
3592 ShiftReg = fastEmitInst_extractsubreg(VT, ShiftReg, /*IsKill=*/true,
3594 emitSubs_rs(VT, ShiftReg, /*IsKill=*/true, MulReg, /*IsKill=*/false,
3595 AArch64_AM::ASR, 31, /*WantResult=*/false);
3597 assert(VT == MVT::i64 && "Unexpected value type.");
3598 MulReg = emitMul_rr(VT, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
3599 unsigned SMULHReg = fastEmit_rr(VT, VT, ISD::MULHS, LHSReg, LHSIsKill,
3601 emitSubs_rs(VT, SMULHReg, /*IsKill=*/true, MulReg, /*IsKill=*/false,
3602 AArch64_AM::ASR, 63, /*WantResult=*/false);
3606 case Intrinsic::umul_with_overflow: {
3608 unsigned LHSReg = getRegForValue(LHS);
3611 bool LHSIsKill = hasTrivialKill(LHS);
3613 unsigned RHSReg = getRegForValue(RHS);
3616 bool RHSIsKill = hasTrivialKill(RHS);
3618 if (VT == MVT::i32) {
3619 MulReg = emitUMULL_rr(MVT::i64, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
3620 emitSubs_rs(MVT::i64, AArch64::XZR, /*IsKill=*/true, MulReg,
3621 /*IsKill=*/false, AArch64_AM::LSR, 32,
3622 /*WantResult=*/false);
3623 MulReg = fastEmitInst_extractsubreg(VT, MulReg, /*IsKill=*/true,
3626 assert(VT == MVT::i64 && "Unexpected value type.");
3627 MulReg = emitMul_rr(VT, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
3628 unsigned UMULHReg = fastEmit_rr(VT, VT, ISD::MULHU, LHSReg, LHSIsKill,
3630 emitSubs_rr(VT, AArch64::XZR, /*IsKill=*/true, UMULHReg,
3631 /*IsKill=*/false, /*WantResult=*/false);
3638 ResultReg1 = createResultReg(TLI.getRegClassFor(VT));
3639 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3640 TII.get(TargetOpcode::COPY), ResultReg1).addReg(MulReg);
3643 ResultReg2 = fastEmitInst_rri(AArch64::CSINCWr, &AArch64::GPR32RegClass,
3644 AArch64::WZR, /*IsKill=*/true, AArch64::WZR,
3645 /*IsKill=*/true, getInvertedCondCode(CC));
3647 assert((ResultReg1 + 1) == ResultReg2 &&
3648 "Nonconsecutive result registers.");
3649 updateValueMap(II, ResultReg1, 2);
3656 bool AArch64FastISel::selectRet(const Instruction *I) {
3657 const ReturnInst *Ret = cast<ReturnInst>(I);
3658 const Function &F = *I->getParent()->getParent();
3660 if (!FuncInfo.CanLowerReturn)
3666 // Build a list of return value registers.
3667 SmallVector<unsigned, 4> RetRegs;
3669 if (Ret->getNumOperands() > 0) {
3670 CallingConv::ID CC = F.getCallingConv();
3671 SmallVector<ISD::OutputArg, 4> Outs;
3672 GetReturnInfo(F.getReturnType(), F.getAttributes(), Outs, TLI);
3674 // Analyze operands of the call, assigning locations to each operand.
3675 SmallVector<CCValAssign, 16> ValLocs;
3676 CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, ValLocs, I->getContext());
3677 CCAssignFn *RetCC = CC == CallingConv::WebKit_JS ? RetCC_AArch64_WebKit_JS
3678 : RetCC_AArch64_AAPCS;
3679 CCInfo.AnalyzeReturn(Outs, RetCC);
3681 // Only handle a single return value for now.
3682 if (ValLocs.size() != 1)
3685 CCValAssign &VA = ValLocs[0];
3686 const Value *RV = Ret->getOperand(0);
3688 // Don't bother handling odd stuff for now.
3689 if ((VA.getLocInfo() != CCValAssign::Full) &&
3690 (VA.getLocInfo() != CCValAssign::BCvt))
3693 // Only handle register returns for now.
3697 unsigned Reg = getRegForValue(RV);
3701 unsigned SrcReg = Reg + VA.getValNo();
3702 unsigned DestReg = VA.getLocReg();
3703 // Avoid a cross-class copy. This is very unlikely.
3704 if (!MRI.getRegClass(SrcReg)->contains(DestReg))
3707 EVT RVEVT = TLI.getValueType(RV->getType());
3708 if (!RVEVT.isSimple())
3711 // Vectors (of > 1 lane) in big endian need tricky handling.
3712 if (RVEVT.isVector() && RVEVT.getVectorNumElements() > 1 &&
3713 !Subtarget->isLittleEndian())
3716 MVT RVVT = RVEVT.getSimpleVT();
3717 if (RVVT == MVT::f128)
3720 MVT DestVT = VA.getValVT();
3721 // Special handling for extended integers.
3722 if (RVVT != DestVT) {
3723 if (RVVT != MVT::i1 && RVVT != MVT::i8 && RVVT != MVT::i16)
3726 if (!Outs[0].Flags.isZExt() && !Outs[0].Flags.isSExt())
3729 bool IsZExt = Outs[0].Flags.isZExt();
3730 SrcReg = emitIntExt(RVVT, SrcReg, DestVT, IsZExt);
3736 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3737 TII.get(TargetOpcode::COPY), DestReg).addReg(SrcReg);
3739 // Add register to return instruction.
3740 RetRegs.push_back(VA.getLocReg());
3743 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3744 TII.get(AArch64::RET_ReallyLR));
3745 for (unsigned i = 0, e = RetRegs.size(); i != e; ++i)
3746 MIB.addReg(RetRegs[i], RegState::Implicit);
3750 bool AArch64FastISel::selectTrunc(const Instruction *I) {
3751 Type *DestTy = I->getType();
3752 Value *Op = I->getOperand(0);
3753 Type *SrcTy = Op->getType();
3755 EVT SrcEVT = TLI.getValueType(SrcTy, true);
3756 EVT DestEVT = TLI.getValueType(DestTy, true);
3757 if (!SrcEVT.isSimple())
3759 if (!DestEVT.isSimple())
3762 MVT SrcVT = SrcEVT.getSimpleVT();
3763 MVT DestVT = DestEVT.getSimpleVT();
3765 if (SrcVT != MVT::i64 && SrcVT != MVT::i32 && SrcVT != MVT::i16 &&
3768 if (DestVT != MVT::i32 && DestVT != MVT::i16 && DestVT != MVT::i8 &&
3772 unsigned SrcReg = getRegForValue(Op);
3775 bool SrcIsKill = hasTrivialKill(Op);
3777 // If we're truncating from i64 to a smaller non-legal type then generate an
3778 // AND. Otherwise, we know the high bits are undefined and a truncate only
3779 // generate a COPY. We cannot mark the source register also as result
3780 // register, because this can incorrectly transfer the kill flag onto the
3783 if (SrcVT == MVT::i64) {
3785 switch (DestVT.SimpleTy) {
3787 // Trunc i64 to i32 is handled by the target-independent fast-isel.
3799 // Issue an extract_subreg to get the lower 32-bits.
3800 unsigned Reg32 = fastEmitInst_extractsubreg(MVT::i32, SrcReg, SrcIsKill,
3802 // Create the AND instruction which performs the actual truncation.
3803 ResultReg = emitAnd_ri(MVT::i32, Reg32, /*IsKill=*/true, Mask);
3804 assert(ResultReg && "Unexpected AND instruction emission failure.");
3806 ResultReg = createResultReg(&AArch64::GPR32RegClass);
3807 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3808 TII.get(TargetOpcode::COPY), ResultReg)
3809 .addReg(SrcReg, getKillRegState(SrcIsKill));
3812 updateValueMap(I, ResultReg);
3816 unsigned AArch64FastISel::emiti1Ext(unsigned SrcReg, MVT DestVT, bool IsZExt) {
3817 assert((DestVT == MVT::i8 || DestVT == MVT::i16 || DestVT == MVT::i32 ||
3818 DestVT == MVT::i64) &&
3819 "Unexpected value type.");
3820 // Handle i8 and i16 as i32.
3821 if (DestVT == MVT::i8 || DestVT == MVT::i16)
3825 unsigned ResultReg = emitAnd_ri(MVT::i32, SrcReg, /*TODO:IsKill=*/false, 1);
3826 assert(ResultReg && "Unexpected AND instruction emission failure.");
3827 if (DestVT == MVT::i64) {
3828 // We're ZExt i1 to i64. The ANDWri Wd, Ws, #1 implicitly clears the
3829 // upper 32 bits. Emit a SUBREG_TO_REG to extend from Wd to Xd.
3830 unsigned Reg64 = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
3831 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3832 TII.get(AArch64::SUBREG_TO_REG), Reg64)
3835 .addImm(AArch64::sub_32);
3840 if (DestVT == MVT::i64) {
3841 // FIXME: We're SExt i1 to i64.
3844 return fastEmitInst_rii(AArch64::SBFMWri, &AArch64::GPR32RegClass, SrcReg,
3845 /*TODO:IsKill=*/false, 0, 0);
3849 unsigned AArch64FastISel::emitMul_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
3850 unsigned Op1, bool Op1IsKill) {
3852 switch (RetVT.SimpleTy) {
3858 Opc = AArch64::MADDWrrr; ZReg = AArch64::WZR; break;
3860 Opc = AArch64::MADDXrrr; ZReg = AArch64::XZR; break;
3863 const TargetRegisterClass *RC =
3864 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3865 return fastEmitInst_rrr(Opc, RC, Op0, Op0IsKill, Op1, Op1IsKill,
3866 /*IsKill=*/ZReg, true);
3869 unsigned AArch64FastISel::emitSMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
3870 unsigned Op1, bool Op1IsKill) {
3871 if (RetVT != MVT::i64)
3874 return fastEmitInst_rrr(AArch64::SMADDLrrr, &AArch64::GPR64RegClass,
3875 Op0, Op0IsKill, Op1, Op1IsKill,
3876 AArch64::XZR, /*IsKill=*/true);
3879 unsigned AArch64FastISel::emitUMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
3880 unsigned Op1, bool Op1IsKill) {
3881 if (RetVT != MVT::i64)
3884 return fastEmitInst_rrr(AArch64::UMADDLrrr, &AArch64::GPR64RegClass,
3885 Op0, Op0IsKill, Op1, Op1IsKill,
3886 AArch64::XZR, /*IsKill=*/true);
3889 unsigned AArch64FastISel::emitLSL_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
3890 unsigned Op1Reg, bool Op1IsKill) {
3892 bool NeedTrunc = false;
3894 switch (RetVT.SimpleTy) {
3896 case MVT::i8: Opc = AArch64::LSLVWr; NeedTrunc = true; Mask = 0xff; break;
3897 case MVT::i16: Opc = AArch64::LSLVWr; NeedTrunc = true; Mask = 0xffff; break;
3898 case MVT::i32: Opc = AArch64::LSLVWr; break;
3899 case MVT::i64: Opc = AArch64::LSLVXr; break;
3902 const TargetRegisterClass *RC =
3903 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3905 Op1Reg = emitAnd_ri(MVT::i32, Op1Reg, Op1IsKill, Mask);
3908 unsigned ResultReg = fastEmitInst_rr(Opc, RC, Op0Reg, Op0IsKill, Op1Reg,
3911 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
3915 unsigned AArch64FastISel::emitLSL_ri(MVT RetVT, MVT SrcVT, unsigned Op0,
3916 bool Op0IsKill, uint64_t Shift,
3918 assert(RetVT.SimpleTy >= SrcVT.SimpleTy &&
3919 "Unexpected source/return type pair.");
3920 assert((SrcVT == MVT::i1 || SrcVT == MVT::i8 || SrcVT == MVT::i16 ||
3921 SrcVT == MVT::i32 || SrcVT == MVT::i64) &&
3922 "Unexpected source value type.");
3923 assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 ||
3924 RetVT == MVT::i64) && "Unexpected return value type.");
3926 bool Is64Bit = (RetVT == MVT::i64);
3927 unsigned RegSize = Is64Bit ? 64 : 32;
3928 unsigned DstBits = RetVT.getSizeInBits();
3929 unsigned SrcBits = SrcVT.getSizeInBits();
3930 const TargetRegisterClass *RC =
3931 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3933 // Just emit a copy for "zero" shifts.
3935 if (RetVT == SrcVT) {
3936 unsigned ResultReg = createResultReg(RC);
3937 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3938 TII.get(TargetOpcode::COPY), ResultReg)
3939 .addReg(Op0, getKillRegState(Op0IsKill));
3942 return emitIntExt(SrcVT, Op0, RetVT, IsZExt);
3945 // Don't deal with undefined shifts.
3946 if (Shift >= DstBits)
3949 // For immediate shifts we can fold the zero-/sign-extension into the shift.
3950 // {S|U}BFM Wd, Wn, #r, #s
3951 // Wd<32+s-r,32-r> = Wn<s:0> when r > s
3953 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3954 // %2 = shl i16 %1, 4
3955 // Wd<32+7-28,32-28> = Wn<7:0> <- clamp s to 7
3956 // 0b1111_1111_1111_1111__1111_1010_1010_0000 sext
3957 // 0b0000_0000_0000_0000__0000_0101_0101_0000 sext | zext
3958 // 0b0000_0000_0000_0000__0000_1010_1010_0000 zext
3960 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3961 // %2 = shl i16 %1, 8
3962 // Wd<32+7-24,32-24> = Wn<7:0>
3963 // 0b1111_1111_1111_1111__1010_1010_0000_0000 sext
3964 // 0b0000_0000_0000_0000__0101_0101_0000_0000 sext | zext
3965 // 0b0000_0000_0000_0000__1010_1010_0000_0000 zext
3967 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3968 // %2 = shl i16 %1, 12
3969 // Wd<32+3-20,32-20> = Wn<3:0>
3970 // 0b1111_1111_1111_1111__1010_0000_0000_0000 sext
3971 // 0b0000_0000_0000_0000__0101_0000_0000_0000 sext | zext
3972 // 0b0000_0000_0000_0000__1010_0000_0000_0000 zext
3974 unsigned ImmR = RegSize - Shift;
3975 // Limit the width to the length of the source type.
3976 unsigned ImmS = std::min<unsigned>(SrcBits - 1, DstBits - 1 - Shift);
3977 static const unsigned OpcTable[2][2] = {
3978 {AArch64::SBFMWri, AArch64::SBFMXri},
3979 {AArch64::UBFMWri, AArch64::UBFMXri}
3981 unsigned Opc = OpcTable[IsZExt][Is64Bit];
3982 if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) {
3983 unsigned TmpReg = MRI.createVirtualRegister(RC);
3984 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3985 TII.get(AArch64::SUBREG_TO_REG), TmpReg)
3987 .addReg(Op0, getKillRegState(Op0IsKill))
3988 .addImm(AArch64::sub_32);
3992 return fastEmitInst_rii(Opc, RC, Op0, Op0IsKill, ImmR, ImmS);
3995 unsigned AArch64FastISel::emitLSR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
3996 unsigned Op1Reg, bool Op1IsKill) {
3998 bool NeedTrunc = false;
4000 switch (RetVT.SimpleTy) {
4002 case MVT::i8: Opc = AArch64::LSRVWr; NeedTrunc = true; Mask = 0xff; break;
4003 case MVT::i16: Opc = AArch64::LSRVWr; NeedTrunc = true; Mask = 0xffff; break;
4004 case MVT::i32: Opc = AArch64::LSRVWr; break;
4005 case MVT::i64: Opc = AArch64::LSRVXr; break;
4008 const TargetRegisterClass *RC =
4009 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
4011 Op0Reg = emitAnd_ri(MVT::i32, Op0Reg, Op0IsKill, Mask);
4012 Op1Reg = emitAnd_ri(MVT::i32, Op1Reg, Op1IsKill, Mask);
4013 Op0IsKill = Op1IsKill = true;
4015 unsigned ResultReg = fastEmitInst_rr(Opc, RC, Op0Reg, Op0IsKill, Op1Reg,
4018 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
4022 unsigned AArch64FastISel::emitLSR_ri(MVT RetVT, MVT SrcVT, unsigned Op0,
4023 bool Op0IsKill, uint64_t Shift,
4025 assert(RetVT.SimpleTy >= SrcVT.SimpleTy &&
4026 "Unexpected source/return type pair.");
4027 assert((SrcVT == MVT::i1 || SrcVT == MVT::i8 || SrcVT == MVT::i16 ||
4028 SrcVT == MVT::i32 || SrcVT == MVT::i64) &&
4029 "Unexpected source value type.");
4030 assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 ||
4031 RetVT == MVT::i64) && "Unexpected return value type.");
4033 bool Is64Bit = (RetVT == MVT::i64);
4034 unsigned RegSize = Is64Bit ? 64 : 32;
4035 unsigned DstBits = RetVT.getSizeInBits();
4036 unsigned SrcBits = SrcVT.getSizeInBits();
4037 const TargetRegisterClass *RC =
4038 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
4040 // Just emit a copy for "zero" shifts.
4042 if (RetVT == SrcVT) {
4043 unsigned ResultReg = createResultReg(RC);
4044 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
4045 TII.get(TargetOpcode::COPY), ResultReg)
4046 .addReg(Op0, getKillRegState(Op0IsKill));
4049 return emitIntExt(SrcVT, Op0, RetVT, IsZExt);
4052 // Don't deal with undefined shifts.
4053 if (Shift >= DstBits)
4056 // For immediate shifts we can fold the zero-/sign-extension into the shift.
4057 // {S|U}BFM Wd, Wn, #r, #s
4058 // Wd<s-r:0> = Wn<s:r> when r <= s
4060 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
4061 // %2 = lshr i16 %1, 4
4062 // Wd<7-4:0> = Wn<7:4>
4063 // 0b0000_0000_0000_0000__0000_1111_1111_1010 sext
4064 // 0b0000_0000_0000_0000__0000_0000_0000_0101 sext | zext
4065 // 0b0000_0000_0000_0000__0000_0000_0000_1010 zext
4067 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
4068 // %2 = lshr i16 %1, 8
4069 // Wd<7-7,0> = Wn<7:7>
4070 // 0b0000_0000_0000_0000__0000_0000_1111_1111 sext
4071 // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
4072 // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
4074 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
4075 // %2 = lshr i16 %1, 12
4076 // Wd<7-7,0> = Wn<7:7> <- clamp r to 7
4077 // 0b0000_0000_0000_0000__0000_0000_0000_1111 sext
4078 // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
4079 // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
4081 if (Shift >= SrcBits && IsZExt)
4082 return materializeInt(ConstantInt::get(*Context, APInt(RegSize, 0)), RetVT);
4084 // It is not possible to fold a sign-extend into the LShr instruction. In this
4085 // case emit a sign-extend.
4087 Op0 = emitIntExt(SrcVT, Op0, RetVT, IsZExt);
4092 SrcBits = SrcVT.getSizeInBits();
4096 unsigned ImmR = std::min<unsigned>(SrcBits - 1, Shift);
4097 unsigned ImmS = SrcBits - 1;
4098 static const unsigned OpcTable[2][2] = {
4099 {AArch64::SBFMWri, AArch64::SBFMXri},
4100 {AArch64::UBFMWri, AArch64::UBFMXri}
4102 unsigned Opc = OpcTable[IsZExt][Is64Bit];
4103 if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) {
4104 unsigned TmpReg = MRI.createVirtualRegister(RC);
4105 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
4106 TII.get(AArch64::SUBREG_TO_REG), TmpReg)
4108 .addReg(Op0, getKillRegState(Op0IsKill))
4109 .addImm(AArch64::sub_32);
4113 return fastEmitInst_rii(Opc, RC, Op0, Op0IsKill, ImmR, ImmS);
4116 unsigned AArch64FastISel::emitASR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
4117 unsigned Op1Reg, bool Op1IsKill) {
4119 bool NeedTrunc = false;
4121 switch (RetVT.SimpleTy) {
4123 case MVT::i8: Opc = AArch64::ASRVWr; NeedTrunc = true; Mask = 0xff; break;
4124 case MVT::i16: Opc = AArch64::ASRVWr; NeedTrunc = true; Mask = 0xffff; break;
4125 case MVT::i32: Opc = AArch64::ASRVWr; break;
4126 case MVT::i64: Opc = AArch64::ASRVXr; break;
4129 const TargetRegisterClass *RC =
4130 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
4132 Op0Reg = emitIntExt(RetVT, Op0Reg, MVT::i32, /*IsZExt=*/false);
4133 Op1Reg = emitAnd_ri(MVT::i32, Op1Reg, Op1IsKill, Mask);
4134 Op0IsKill = Op1IsKill = true;
4136 unsigned ResultReg = fastEmitInst_rr(Opc, RC, Op0Reg, Op0IsKill, Op1Reg,
4139 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
4143 unsigned AArch64FastISel::emitASR_ri(MVT RetVT, MVT SrcVT, unsigned Op0,
4144 bool Op0IsKill, uint64_t Shift,
4146 assert(RetVT.SimpleTy >= SrcVT.SimpleTy &&
4147 "Unexpected source/return type pair.");
4148 assert((SrcVT == MVT::i1 || SrcVT == MVT::i8 || SrcVT == MVT::i16 ||
4149 SrcVT == MVT::i32 || SrcVT == MVT::i64) &&
4150 "Unexpected source value type.");
4151 assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 ||
4152 RetVT == MVT::i64) && "Unexpected return value type.");
4154 bool Is64Bit = (RetVT == MVT::i64);
4155 unsigned RegSize = Is64Bit ? 64 : 32;
4156 unsigned DstBits = RetVT.getSizeInBits();
4157 unsigned SrcBits = SrcVT.getSizeInBits();
4158 const TargetRegisterClass *RC =
4159 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
4161 // Just emit a copy for "zero" shifts.
4163 if (RetVT == SrcVT) {
4164 unsigned ResultReg = createResultReg(RC);
4165 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
4166 TII.get(TargetOpcode::COPY), ResultReg)
4167 .addReg(Op0, getKillRegState(Op0IsKill));
4170 return emitIntExt(SrcVT, Op0, RetVT, IsZExt);
4173 // Don't deal with undefined shifts.
4174 if (Shift >= DstBits)
4177 // For immediate shifts we can fold the zero-/sign-extension into the shift.
4178 // {S|U}BFM Wd, Wn, #r, #s
4179 // Wd<s-r:0> = Wn<s:r> when r <= s
4181 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
4182 // %2 = ashr i16 %1, 4
4183 // Wd<7-4:0> = Wn<7:4>
4184 // 0b1111_1111_1111_1111__1111_1111_1111_1010 sext
4185 // 0b0000_0000_0000_0000__0000_0000_0000_0101 sext | zext
4186 // 0b0000_0000_0000_0000__0000_0000_0000_1010 zext
4188 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
4189 // %2 = ashr i16 %1, 8
4190 // Wd<7-7,0> = Wn<7:7>
4191 // 0b1111_1111_1111_1111__1111_1111_1111_1111 sext
4192 // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
4193 // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
4195 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
4196 // %2 = ashr i16 %1, 12
4197 // Wd<7-7,0> = Wn<7:7> <- clamp r to 7
4198 // 0b1111_1111_1111_1111__1111_1111_1111_1111 sext
4199 // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
4200 // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
4202 if (Shift >= SrcBits && IsZExt)
4203 return materializeInt(ConstantInt::get(*Context, APInt(RegSize, 0)), RetVT);
4205 unsigned ImmR = std::min<unsigned>(SrcBits - 1, Shift);
4206 unsigned ImmS = SrcBits - 1;
4207 static const unsigned OpcTable[2][2] = {
4208 {AArch64::SBFMWri, AArch64::SBFMXri},
4209 {AArch64::UBFMWri, AArch64::UBFMXri}
4211 unsigned Opc = OpcTable[IsZExt][Is64Bit];
4212 if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) {
4213 unsigned TmpReg = MRI.createVirtualRegister(RC);
4214 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
4215 TII.get(AArch64::SUBREG_TO_REG), TmpReg)
4217 .addReg(Op0, getKillRegState(Op0IsKill))
4218 .addImm(AArch64::sub_32);
4222 return fastEmitInst_rii(Opc, RC, Op0, Op0IsKill, ImmR, ImmS);
4225 unsigned AArch64FastISel::emitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT,
4227 assert(DestVT != MVT::i1 && "ZeroExt/SignExt an i1?");
4229 // FastISel does not have plumbing to deal with extensions where the SrcVT or
4230 // DestVT are odd things, so test to make sure that they are both types we can
4231 // handle (i1/i8/i16/i32 for SrcVT and i8/i16/i32/i64 for DestVT), otherwise
4232 // bail out to SelectionDAG.
4233 if (((DestVT != MVT::i8) && (DestVT != MVT::i16) &&
4234 (DestVT != MVT::i32) && (DestVT != MVT::i64)) ||
4235 ((SrcVT != MVT::i1) && (SrcVT != MVT::i8) &&
4236 (SrcVT != MVT::i16) && (SrcVT != MVT::i32)))
4242 switch (SrcVT.SimpleTy) {
4246 return emiti1Ext(SrcReg, DestVT, IsZExt);
4248 if (DestVT == MVT::i64)
4249 Opc = IsZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
4251 Opc = IsZExt ? AArch64::UBFMWri : AArch64::SBFMWri;
4255 if (DestVT == MVT::i64)
4256 Opc = IsZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
4258 Opc = IsZExt ? AArch64::UBFMWri : AArch64::SBFMWri;
4262 assert(DestVT == MVT::i64 && "IntExt i32 to i32?!?");
4263 Opc = IsZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
4268 // Handle i8 and i16 as i32.
4269 if (DestVT == MVT::i8 || DestVT == MVT::i16)
4271 else if (DestVT == MVT::i64) {
4272 unsigned Src64 = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
4273 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
4274 TII.get(AArch64::SUBREG_TO_REG), Src64)
4277 .addImm(AArch64::sub_32);
4281 const TargetRegisterClass *RC =
4282 (DestVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
4283 return fastEmitInst_rii(Opc, RC, SrcReg, /*TODO:IsKill=*/false, 0, Imm);
4286 static bool isZExtLoad(const MachineInstr *LI) {
4287 switch (LI->getOpcode()) {
4290 case AArch64::LDURBBi:
4291 case AArch64::LDURHHi:
4292 case AArch64::LDURWi:
4293 case AArch64::LDRBBui:
4294 case AArch64::LDRHHui:
4295 case AArch64::LDRWui:
4296 case AArch64::LDRBBroX:
4297 case AArch64::LDRHHroX:
4298 case AArch64::LDRWroX:
4299 case AArch64::LDRBBroW:
4300 case AArch64::LDRHHroW:
4301 case AArch64::LDRWroW:
4306 static bool isSExtLoad(const MachineInstr *LI) {
4307 switch (LI->getOpcode()) {
4310 case AArch64::LDURSBWi:
4311 case AArch64::LDURSHWi:
4312 case AArch64::LDURSBXi:
4313 case AArch64::LDURSHXi:
4314 case AArch64::LDURSWi:
4315 case AArch64::LDRSBWui:
4316 case AArch64::LDRSHWui:
4317 case AArch64::LDRSBXui:
4318 case AArch64::LDRSHXui:
4319 case AArch64::LDRSWui:
4320 case AArch64::LDRSBWroX:
4321 case AArch64::LDRSHWroX:
4322 case AArch64::LDRSBXroX:
4323 case AArch64::LDRSHXroX:
4324 case AArch64::LDRSWroX:
4325 case AArch64::LDRSBWroW:
4326 case AArch64::LDRSHWroW:
4327 case AArch64::LDRSBXroW:
4328 case AArch64::LDRSHXroW:
4329 case AArch64::LDRSWroW:
4334 bool AArch64FastISel::optimizeIntExtLoad(const Instruction *I, MVT RetVT,
4336 const auto *LI = dyn_cast<LoadInst>(I->getOperand(0));
4337 if (!LI || !LI->hasOneUse())
4340 // Check if the load instruction has already been selected.
4341 unsigned Reg = lookUpRegForValue(LI);
4345 MachineInstr *MI = MRI.getUniqueVRegDef(Reg);
4349 // Check if the correct load instruction has been emitted - SelectionDAG might
4350 // have emitted a zero-extending load, but we need a sign-extending load.
4351 bool IsZExt = isa<ZExtInst>(I);
4352 const auto *LoadMI = MI;
4353 if (LoadMI->getOpcode() == TargetOpcode::COPY &&
4354 LoadMI->getOperand(1).getSubReg() == AArch64::sub_32) {
4355 unsigned LoadReg = MI->getOperand(1).getReg();
4356 LoadMI = MRI.getUniqueVRegDef(LoadReg);
4357 assert(LoadMI && "Expected valid instruction");
4359 if (!(IsZExt && isZExtLoad(LoadMI)) && !(!IsZExt && isSExtLoad(LoadMI)))
4362 // Nothing to be done.
4363 if (RetVT != MVT::i64 || SrcVT > MVT::i32) {
4364 updateValueMap(I, Reg);
4369 unsigned Reg64 = createResultReg(&AArch64::GPR64RegClass);
4370 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
4371 TII.get(AArch64::SUBREG_TO_REG), Reg64)
4373 .addReg(Reg, getKillRegState(true))
4374 .addImm(AArch64::sub_32);
4377 assert((MI->getOpcode() == TargetOpcode::COPY &&
4378 MI->getOperand(1).getSubReg() == AArch64::sub_32) &&
4379 "Expected copy instruction");
4380 Reg = MI->getOperand(1).getReg();
4381 MI->eraseFromParent();
4383 updateValueMap(I, Reg);
4387 bool AArch64FastISel::selectIntExt(const Instruction *I) {
4388 assert((isa<ZExtInst>(I) || isa<SExtInst>(I)) &&
4389 "Unexpected integer extend instruction.");
4392 if (!isTypeSupported(I->getType(), RetVT))
4395 if (!isTypeSupported(I->getOperand(0)->getType(), SrcVT))
4398 // Try to optimize already sign-/zero-extended values from load instructions.
4399 if (optimizeIntExtLoad(I, RetVT, SrcVT))
4402 unsigned SrcReg = getRegForValue(I->getOperand(0));
4405 bool SrcIsKill = hasTrivialKill(I->getOperand(0));
4407 // Try to optimize already sign-/zero-extended values from function arguments.
4408 bool IsZExt = isa<ZExtInst>(I);
4409 if (const auto *Arg = dyn_cast<Argument>(I->getOperand(0))) {
4410 if ((IsZExt && Arg->hasZExtAttr()) || (!IsZExt && Arg->hasSExtAttr())) {
4411 if (RetVT == MVT::i64 && SrcVT != MVT::i64) {
4412 unsigned ResultReg = createResultReg(&AArch64::GPR64RegClass);
4413 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
4414 TII.get(AArch64::SUBREG_TO_REG), ResultReg)
4416 .addReg(SrcReg, getKillRegState(SrcIsKill))
4417 .addImm(AArch64::sub_32);
4420 // Conservatively clear all kill flags from all uses, because we are
4421 // replacing a sign-/zero-extend instruction at IR level with a nop at MI
4422 // level. The result of the instruction at IR level might have been
4423 // trivially dead, which is now not longer true.
4424 unsigned UseReg = lookUpRegForValue(I);
4426 MRI.clearKillFlags(UseReg);
4428 updateValueMap(I, SrcReg);
4433 unsigned ResultReg = emitIntExt(SrcVT, SrcReg, RetVT, IsZExt);
4437 updateValueMap(I, ResultReg);
4441 bool AArch64FastISel::selectRem(const Instruction *I, unsigned ISDOpcode) {
4442 EVT DestEVT = TLI.getValueType(I->getType(), true);
4443 if (!DestEVT.isSimple())
4446 MVT DestVT = DestEVT.getSimpleVT();
4447 if (DestVT != MVT::i64 && DestVT != MVT::i32)
4451 bool Is64bit = (DestVT == MVT::i64);
4452 switch (ISDOpcode) {
4456 DivOpc = Is64bit ? AArch64::SDIVXr : AArch64::SDIVWr;
4459 DivOpc = Is64bit ? AArch64::UDIVXr : AArch64::UDIVWr;
4462 unsigned MSubOpc = Is64bit ? AArch64::MSUBXrrr : AArch64::MSUBWrrr;
4463 unsigned Src0Reg = getRegForValue(I->getOperand(0));
4466 bool Src0IsKill = hasTrivialKill(I->getOperand(0));
4468 unsigned Src1Reg = getRegForValue(I->getOperand(1));
4471 bool Src1IsKill = hasTrivialKill(I->getOperand(1));
4473 const TargetRegisterClass *RC =
4474 (DestVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
4475 unsigned QuotReg = fastEmitInst_rr(DivOpc, RC, Src0Reg, /*IsKill=*/false,
4476 Src1Reg, /*IsKill=*/false);
4477 assert(QuotReg && "Unexpected DIV instruction emission failure.");
4478 // The remainder is computed as numerator - (quotient * denominator) using the
4479 // MSUB instruction.
4480 unsigned ResultReg = fastEmitInst_rrr(MSubOpc, RC, QuotReg, /*IsKill=*/true,
4481 Src1Reg, Src1IsKill, Src0Reg,
4483 updateValueMap(I, ResultReg);
4487 bool AArch64FastISel::selectMul(const Instruction *I) {
4489 if (!isTypeSupported(I->getType(), VT, /*IsVectorAllowed=*/true))
4493 return selectBinaryOp(I, ISD::MUL);
4495 const Value *Src0 = I->getOperand(0);
4496 const Value *Src1 = I->getOperand(1);
4497 if (const auto *C = dyn_cast<ConstantInt>(Src0))
4498 if (C->getValue().isPowerOf2())
4499 std::swap(Src0, Src1);
4501 // Try to simplify to a shift instruction.
4502 if (const auto *C = dyn_cast<ConstantInt>(Src1))
4503 if (C->getValue().isPowerOf2()) {
4504 uint64_t ShiftVal = C->getValue().logBase2();
4507 if (const auto *ZExt = dyn_cast<ZExtInst>(Src0)) {
4508 if (!isIntExtFree(ZExt)) {
4510 if (isValueAvailable(ZExt) && isTypeSupported(ZExt->getSrcTy(), VT)) {
4513 Src0 = ZExt->getOperand(0);
4516 } else if (const auto *SExt = dyn_cast<SExtInst>(Src0)) {
4517 if (!isIntExtFree(SExt)) {
4519 if (isValueAvailable(SExt) && isTypeSupported(SExt->getSrcTy(), VT)) {
4522 Src0 = SExt->getOperand(0);
4527 unsigned Src0Reg = getRegForValue(Src0);
4530 bool Src0IsKill = hasTrivialKill(Src0);
4532 unsigned ResultReg =
4533 emitLSL_ri(VT, SrcVT, Src0Reg, Src0IsKill, ShiftVal, IsZExt);
4536 updateValueMap(I, ResultReg);
4541 unsigned Src0Reg = getRegForValue(I->getOperand(0));
4544 bool Src0IsKill = hasTrivialKill(I->getOperand(0));
4546 unsigned Src1Reg = getRegForValue(I->getOperand(1));
4549 bool Src1IsKill = hasTrivialKill(I->getOperand(1));
4551 unsigned ResultReg = emitMul_rr(VT, Src0Reg, Src0IsKill, Src1Reg, Src1IsKill);
4556 updateValueMap(I, ResultReg);
4560 bool AArch64FastISel::selectShift(const Instruction *I) {
4562 if (!isTypeSupported(I->getType(), RetVT, /*IsVectorAllowed=*/true))
4565 if (RetVT.isVector())
4566 return selectOperator(I, I->getOpcode());
4568 if (const auto *C = dyn_cast<ConstantInt>(I->getOperand(1))) {
4569 unsigned ResultReg = 0;
4570 uint64_t ShiftVal = C->getZExtValue();
4572 bool IsZExt = I->getOpcode() != Instruction::AShr;
4573 const Value *Op0 = I->getOperand(0);
4574 if (const auto *ZExt = dyn_cast<ZExtInst>(Op0)) {
4575 if (!isIntExtFree(ZExt)) {
4577 if (isValueAvailable(ZExt) && isTypeSupported(ZExt->getSrcTy(), TmpVT)) {
4580 Op0 = ZExt->getOperand(0);
4583 } else if (const auto *SExt = dyn_cast<SExtInst>(Op0)) {
4584 if (!isIntExtFree(SExt)) {
4586 if (isValueAvailable(SExt) && isTypeSupported(SExt->getSrcTy(), TmpVT)) {
4589 Op0 = SExt->getOperand(0);
4594 unsigned Op0Reg = getRegForValue(Op0);
4597 bool Op0IsKill = hasTrivialKill(Op0);
4599 switch (I->getOpcode()) {
4600 default: llvm_unreachable("Unexpected instruction.");
4601 case Instruction::Shl:
4602 ResultReg = emitLSL_ri(RetVT, SrcVT, Op0Reg, Op0IsKill, ShiftVal, IsZExt);
4604 case Instruction::AShr:
4605 ResultReg = emitASR_ri(RetVT, SrcVT, Op0Reg, Op0IsKill, ShiftVal, IsZExt);
4607 case Instruction::LShr:
4608 ResultReg = emitLSR_ri(RetVT, SrcVT, Op0Reg, Op0IsKill, ShiftVal, IsZExt);
4614 updateValueMap(I, ResultReg);
4618 unsigned Op0Reg = getRegForValue(I->getOperand(0));
4621 bool Op0IsKill = hasTrivialKill(I->getOperand(0));
4623 unsigned Op1Reg = getRegForValue(I->getOperand(1));
4626 bool Op1IsKill = hasTrivialKill(I->getOperand(1));
4628 unsigned ResultReg = 0;
4629 switch (I->getOpcode()) {
4630 default: llvm_unreachable("Unexpected instruction.");
4631 case Instruction::Shl:
4632 ResultReg = emitLSL_rr(RetVT, Op0Reg, Op0IsKill, Op1Reg, Op1IsKill);
4634 case Instruction::AShr:
4635 ResultReg = emitASR_rr(RetVT, Op0Reg, Op0IsKill, Op1Reg, Op1IsKill);
4637 case Instruction::LShr:
4638 ResultReg = emitLSR_rr(RetVT, Op0Reg, Op0IsKill, Op1Reg, Op1IsKill);
4645 updateValueMap(I, ResultReg);
4649 bool AArch64FastISel::selectBitCast(const Instruction *I) {
4652 if (!isTypeLegal(I->getOperand(0)->getType(), SrcVT))
4654 if (!isTypeLegal(I->getType(), RetVT))
4658 if (RetVT == MVT::f32 && SrcVT == MVT::i32)
4659 Opc = AArch64::FMOVWSr;
4660 else if (RetVT == MVT::f64 && SrcVT == MVT::i64)
4661 Opc = AArch64::FMOVXDr;
4662 else if (RetVT == MVT::i32 && SrcVT == MVT::f32)
4663 Opc = AArch64::FMOVSWr;
4664 else if (RetVT == MVT::i64 && SrcVT == MVT::f64)
4665 Opc = AArch64::FMOVDXr;
4669 const TargetRegisterClass *RC = nullptr;
4670 switch (RetVT.SimpleTy) {
4671 default: llvm_unreachable("Unexpected value type.");
4672 case MVT::i32: RC = &AArch64::GPR32RegClass; break;
4673 case MVT::i64: RC = &AArch64::GPR64RegClass; break;
4674 case MVT::f32: RC = &AArch64::FPR32RegClass; break;
4675 case MVT::f64: RC = &AArch64::FPR64RegClass; break;
4677 unsigned Op0Reg = getRegForValue(I->getOperand(0));
4680 bool Op0IsKill = hasTrivialKill(I->getOperand(0));
4681 unsigned ResultReg = fastEmitInst_r(Opc, RC, Op0Reg, Op0IsKill);
4686 updateValueMap(I, ResultReg);
4690 bool AArch64FastISel::selectFRem(const Instruction *I) {
4692 if (!isTypeLegal(I->getType(), RetVT))
4696 switch (RetVT.SimpleTy) {
4700 LC = RTLIB::REM_F32;
4703 LC = RTLIB::REM_F64;
4708 Args.reserve(I->getNumOperands());
4710 // Populate the argument list.
4711 for (auto &Arg : I->operands()) {
4714 Entry.Ty = Arg->getType();
4715 Args.push_back(Entry);
4718 CallLoweringInfo CLI;
4719 CLI.setCallee(TLI.getLibcallCallingConv(LC), I->getType(),
4720 TLI.getLibcallName(LC), std::move(Args));
4721 if (!lowerCallTo(CLI))
4723 updateValueMap(I, CLI.ResultReg);
4727 bool AArch64FastISel::selectSDiv(const Instruction *I) {
4729 if (!isTypeLegal(I->getType(), VT))
4732 if (!isa<ConstantInt>(I->getOperand(1)))
4733 return selectBinaryOp(I, ISD::SDIV);
4735 const APInt &C = cast<ConstantInt>(I->getOperand(1))->getValue();
4736 if ((VT != MVT::i32 && VT != MVT::i64) || !C ||
4737 !(C.isPowerOf2() || (-C).isPowerOf2()))
4738 return selectBinaryOp(I, ISD::SDIV);
4740 unsigned Lg2 = C.countTrailingZeros();
4741 unsigned Src0Reg = getRegForValue(I->getOperand(0));
4744 bool Src0IsKill = hasTrivialKill(I->getOperand(0));
4746 if (cast<BinaryOperator>(I)->isExact()) {
4747 unsigned ResultReg = emitASR_ri(VT, VT, Src0Reg, Src0IsKill, Lg2);
4750 updateValueMap(I, ResultReg);
4754 int64_t Pow2MinusOne = (1ULL << Lg2) - 1;
4755 unsigned AddReg = emitAdd_ri_(VT, Src0Reg, /*IsKill=*/false, Pow2MinusOne);
4759 // (Src0 < 0) ? Pow2 - 1 : 0;
4760 if (!emitICmp_ri(VT, Src0Reg, /*IsKill=*/false, 0))
4764 const TargetRegisterClass *RC;
4765 if (VT == MVT::i64) {
4766 SelectOpc = AArch64::CSELXr;
4767 RC = &AArch64::GPR64RegClass;
4769 SelectOpc = AArch64::CSELWr;
4770 RC = &AArch64::GPR32RegClass;
4772 unsigned SelectReg =
4773 fastEmitInst_rri(SelectOpc, RC, AddReg, /*IsKill=*/true, Src0Reg,
4774 Src0IsKill, AArch64CC::LT);
4778 // Divide by Pow2 --> ashr. If we're dividing by a negative value we must also
4779 // negate the result.
4780 unsigned ZeroReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
4783 ResultReg = emitAddSub_rs(/*UseAdd=*/false, VT, ZeroReg, /*IsKill=*/true,
4784 SelectReg, /*IsKill=*/true, AArch64_AM::ASR, Lg2);
4786 ResultReg = emitASR_ri(VT, VT, SelectReg, /*IsKill=*/true, Lg2);
4791 updateValueMap(I, ResultReg);
4795 /// This is mostly a copy of the existing FastISel getRegForGEPIndex code. We
4796 /// have to duplicate it for AArch64, because otherwise we would fail during the
4797 /// sign-extend emission.
4798 std::pair<unsigned, bool> AArch64FastISel::getRegForGEPIndex(const Value *Idx) {
4799 unsigned IdxN = getRegForValue(Idx);
4801 // Unhandled operand. Halt "fast" selection and bail.
4802 return std::pair<unsigned, bool>(0, false);
4804 bool IdxNIsKill = hasTrivialKill(Idx);
4806 // If the index is smaller or larger than intptr_t, truncate or extend it.
4807 MVT PtrVT = TLI.getPointerTy();
4808 EVT IdxVT = EVT::getEVT(Idx->getType(), /*HandleUnknown=*/false);
4809 if (IdxVT.bitsLT(PtrVT)) {
4810 IdxN = emitIntExt(IdxVT.getSimpleVT(), IdxN, PtrVT, /*IsZExt=*/false);
4812 } else if (IdxVT.bitsGT(PtrVT))
4813 llvm_unreachable("AArch64 FastISel doesn't support types larger than i64");
4814 return std::pair<unsigned, bool>(IdxN, IdxNIsKill);
4817 /// This is mostly a copy of the existing FastISel GEP code, but we have to
4818 /// duplicate it for AArch64, because otherwise we would bail out even for
4819 /// simple cases. This is because the standard fastEmit functions don't cover
4820 /// MUL at all and ADD is lowered very inefficientily.
4821 bool AArch64FastISel::selectGetElementPtr(const Instruction *I) {
4822 unsigned N = getRegForValue(I->getOperand(0));
4825 bool NIsKill = hasTrivialKill(I->getOperand(0));
4827 // Keep a running tab of the total offset to coalesce multiple N = N + Offset
4828 // into a single N = N + TotalOffset.
4829 uint64_t TotalOffs = 0;
4830 Type *Ty = I->getOperand(0)->getType();
4831 MVT VT = TLI.getPointerTy();
4832 for (auto OI = std::next(I->op_begin()), E = I->op_end(); OI != E; ++OI) {
4833 const Value *Idx = *OI;
4834 if (auto *StTy = dyn_cast<StructType>(Ty)) {
4835 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
4838 TotalOffs += DL.getStructLayout(StTy)->getElementOffset(Field);
4839 Ty = StTy->getElementType(Field);
4841 Ty = cast<SequentialType>(Ty)->getElementType();
4842 // If this is a constant subscript, handle it quickly.
4843 if (const auto *CI = dyn_cast<ConstantInt>(Idx)) {
4848 DL.getTypeAllocSize(Ty) * cast<ConstantInt>(CI)->getSExtValue();
4852 N = emitAdd_ri_(VT, N, NIsKill, TotalOffs);
4859 // N = N + Idx * ElementSize;
4860 uint64_t ElementSize = DL.getTypeAllocSize(Ty);
4861 std::pair<unsigned, bool> Pair = getRegForGEPIndex(Idx);
4862 unsigned IdxN = Pair.first;
4863 bool IdxNIsKill = Pair.second;
4867 if (ElementSize != 1) {
4868 unsigned C = fastEmit_i(VT, VT, ISD::Constant, ElementSize);
4871 IdxN = emitMul_rr(VT, IdxN, IdxNIsKill, C, true);
4876 N = fastEmit_rr(VT, VT, ISD::ADD, N, NIsKill, IdxN, IdxNIsKill);
4882 N = emitAdd_ri_(VT, N, NIsKill, TotalOffs);
4886 updateValueMap(I, N);
4890 bool AArch64FastISel::fastSelectInstruction(const Instruction *I) {
4891 switch (I->getOpcode()) {
4894 case Instruction::Add:
4895 case Instruction::Sub:
4896 return selectAddSub(I);
4897 case Instruction::Mul:
4898 return selectMul(I);
4899 case Instruction::SDiv:
4900 return selectSDiv(I);
4901 case Instruction::SRem:
4902 if (!selectBinaryOp(I, ISD::SREM))
4903 return selectRem(I, ISD::SREM);
4905 case Instruction::URem:
4906 if (!selectBinaryOp(I, ISD::UREM))
4907 return selectRem(I, ISD::UREM);
4909 case Instruction::Shl:
4910 case Instruction::LShr:
4911 case Instruction::AShr:
4912 return selectShift(I);
4913 case Instruction::And:
4914 case Instruction::Or:
4915 case Instruction::Xor:
4916 return selectLogicalOp(I);
4917 case Instruction::Br:
4918 return selectBranch(I);
4919 case Instruction::IndirectBr:
4920 return selectIndirectBr(I);
4921 case Instruction::BitCast:
4922 if (!FastISel::selectBitCast(I))
4923 return selectBitCast(I);
4925 case Instruction::FPToSI:
4926 if (!selectCast(I, ISD::FP_TO_SINT))
4927 return selectFPToInt(I, /*Signed=*/true);
4929 case Instruction::FPToUI:
4930 return selectFPToInt(I, /*Signed=*/false);
4931 case Instruction::ZExt:
4932 case Instruction::SExt:
4933 return selectIntExt(I);
4934 case Instruction::Trunc:
4935 if (!selectCast(I, ISD::TRUNCATE))
4936 return selectTrunc(I);
4938 case Instruction::FPExt:
4939 return selectFPExt(I);
4940 case Instruction::FPTrunc:
4941 return selectFPTrunc(I);
4942 case Instruction::SIToFP:
4943 if (!selectCast(I, ISD::SINT_TO_FP))
4944 return selectIntToFP(I, /*Signed=*/true);
4946 case Instruction::UIToFP:
4947 return selectIntToFP(I, /*Signed=*/false);
4948 case Instruction::Load:
4949 return selectLoad(I);
4950 case Instruction::Store:
4951 return selectStore(I);
4952 case Instruction::FCmp:
4953 case Instruction::ICmp:
4954 return selectCmp(I);
4955 case Instruction::Select:
4956 return selectSelect(I);
4957 case Instruction::Ret:
4958 return selectRet(I);
4959 case Instruction::FRem:
4960 return selectFRem(I);
4961 case Instruction::GetElementPtr:
4962 return selectGetElementPtr(I);
4965 // fall-back to target-independent instruction selection.
4966 return selectOperator(I, I->getOpcode());
4967 // Silence warnings.
4968 (void)&CC_AArch64_DarwinPCS_VarArg;
4972 llvm::FastISel *AArch64::createFastISel(FunctionLoweringInfo &FuncInfo,
4973 const TargetLibraryInfo *LibInfo) {
4974 return new AArch64FastISel(FuncInfo, LibInfo);