1 //===-- AArch6464FastISel.cpp - AArch64 FastISel implementation -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the AArch64-specific support for the FastISel class. Some
11 // of the target-specific code is generated by tablegen in the file
12 // AArch64GenFastISel.inc, which is #included here.
14 //===----------------------------------------------------------------------===//
17 #include "AArch64Subtarget.h"
18 #include "AArch64TargetMachine.h"
19 #include "MCTargetDesc/AArch64AddressingModes.h"
20 #include "llvm/Analysis/BranchProbabilityInfo.h"
21 #include "llvm/CodeGen/CallingConvLower.h"
22 #include "llvm/CodeGen/FastISel.h"
23 #include "llvm/CodeGen/FunctionLoweringInfo.h"
24 #include "llvm/CodeGen/MachineConstantPool.h"
25 #include "llvm/CodeGen/MachineFrameInfo.h"
26 #include "llvm/CodeGen/MachineInstrBuilder.h"
27 #include "llvm/CodeGen/MachineRegisterInfo.h"
28 #include "llvm/IR/CallingConv.h"
29 #include "llvm/IR/DataLayout.h"
30 #include "llvm/IR/DerivedTypes.h"
31 #include "llvm/IR/Function.h"
32 #include "llvm/IR/GetElementPtrTypeIterator.h"
33 #include "llvm/IR/GlobalAlias.h"
34 #include "llvm/IR/GlobalVariable.h"
35 #include "llvm/IR/Instructions.h"
36 #include "llvm/IR/IntrinsicInst.h"
37 #include "llvm/IR/Operator.h"
38 #include "llvm/Support/CommandLine.h"
43 class AArch64FastISel final : public FastISel {
53 AArch64_AM::ShiftExtendType ExtType;
61 const GlobalValue *GV;
64 Address() : Kind(RegBase), ExtType(AArch64_AM::InvalidShiftExtend),
65 OffsetReg(0), Shift(0), Offset(0), GV(nullptr) { Base.Reg = 0; }
66 void setKind(BaseKind K) { Kind = K; }
67 BaseKind getKind() const { return Kind; }
68 void setExtendType(AArch64_AM::ShiftExtendType E) { ExtType = E; }
69 AArch64_AM::ShiftExtendType getExtendType() const { return ExtType; }
70 bool isRegBase() const { return Kind == RegBase; }
71 bool isFIBase() const { return Kind == FrameIndexBase; }
72 void setReg(unsigned Reg) {
73 assert(isRegBase() && "Invalid base register access!");
76 unsigned getReg() const {
77 assert(isRegBase() && "Invalid base register access!");
80 void setOffsetReg(unsigned Reg) {
81 assert(isRegBase() && "Invalid offset register access!");
84 unsigned getOffsetReg() const {
85 assert(isRegBase() && "Invalid offset register access!");
88 void setFI(unsigned FI) {
89 assert(isFIBase() && "Invalid base frame index access!");
92 unsigned getFI() const {
93 assert(isFIBase() && "Invalid base frame index access!");
96 void setOffset(int64_t O) { Offset = O; }
97 int64_t getOffset() { return Offset; }
98 void setShift(unsigned S) { Shift = S; }
99 unsigned getShift() { return Shift; }
101 void setGlobalValue(const GlobalValue *G) { GV = G; }
102 const GlobalValue *getGlobalValue() { return GV; }
105 /// Subtarget - Keep a pointer to the AArch64Subtarget around so that we can
106 /// make the right decision when generating code for different targets.
107 const AArch64Subtarget *Subtarget;
108 LLVMContext *Context;
110 bool fastLowerArguments() override;
111 bool fastLowerCall(CallLoweringInfo &CLI) override;
112 bool fastLowerIntrinsicCall(const IntrinsicInst *II) override;
115 // Selection routines.
116 bool selectAddSub(const Instruction *I);
117 bool selectLogicalOp(const Instruction *I);
118 bool selectLoad(const Instruction *I);
119 bool selectStore(const Instruction *I);
120 bool selectBranch(const Instruction *I);
121 bool selectIndirectBr(const Instruction *I);
122 bool selectCmp(const Instruction *I);
123 bool selectSelect(const Instruction *I);
124 bool selectFPExt(const Instruction *I);
125 bool selectFPTrunc(const Instruction *I);
126 bool selectFPToInt(const Instruction *I, bool Signed);
127 bool selectIntToFP(const Instruction *I, bool Signed);
128 bool selectRem(const Instruction *I, unsigned ISDOpcode);
129 bool selectRet(const Instruction *I);
130 bool selectTrunc(const Instruction *I);
131 bool selectIntExt(const Instruction *I);
132 bool selectMul(const Instruction *I);
133 bool selectShift(const Instruction *I);
134 bool selectBitCast(const Instruction *I);
135 bool selectFRem(const Instruction *I);
136 bool selectSDiv(const Instruction *I);
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,
154 // Emit helper routines.
155 unsigned emitAddSub(bool UseAdd, MVT RetVT, const Value *LHS,
156 const Value *RHS, bool SetFlags = false,
157 bool WantResult = true, bool IsZExt = false);
158 unsigned emitAddSub_rr(bool UseAdd, MVT RetVT, unsigned LHSReg,
159 bool LHSIsKill, unsigned RHSReg, bool RHSIsKill,
160 bool SetFlags = false, bool WantResult = true);
161 unsigned emitAddSub_ri(bool UseAdd, MVT RetVT, unsigned LHSReg,
162 bool LHSIsKill, uint64_t Imm, bool SetFlags = false,
163 bool WantResult = true);
164 unsigned emitAddSub_rs(bool UseAdd, MVT RetVT, unsigned LHSReg,
165 bool LHSIsKill, unsigned RHSReg, bool RHSIsKill,
166 AArch64_AM::ShiftExtendType ShiftType,
167 uint64_t ShiftImm, bool SetFlags = false,
168 bool WantResult = true);
169 unsigned emitAddSub_rx(bool UseAdd, MVT RetVT, unsigned LHSReg,
170 bool LHSIsKill, unsigned RHSReg, bool RHSIsKill,
171 AArch64_AM::ShiftExtendType ExtType,
172 uint64_t ShiftImm, bool SetFlags = false,
173 bool WantResult = true);
176 bool emitCmp(const Value *LHS, const Value *RHS, bool IsZExt);
177 bool emitICmp(MVT RetVT, const Value *LHS, const Value *RHS, bool IsZExt);
178 bool emitICmp_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill, uint64_t Imm);
179 bool emitFCmp(MVT RetVT, const Value *LHS, const Value *RHS);
180 bool emitLoad(MVT VT, unsigned &ResultReg, Address Addr,
181 MachineMemOperand *MMO = nullptr);
182 bool emitStore(MVT VT, unsigned SrcReg, Address Addr,
183 MachineMemOperand *MMO = nullptr);
184 unsigned emitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, bool isZExt);
185 unsigned emiti1Ext(unsigned SrcReg, MVT DestVT, bool isZExt);
186 unsigned emitAdd(MVT RetVT, const Value *LHS, const Value *RHS,
187 bool SetFlags = false, bool WantResult = true,
188 bool IsZExt = false);
189 unsigned emitSub(MVT RetVT, const Value *LHS, const Value *RHS,
190 bool SetFlags = false, bool WantResult = true,
191 bool IsZExt = false);
192 unsigned emitSubs_rr(MVT RetVT, unsigned LHSReg, bool LHSIsKill,
193 unsigned RHSReg, bool RHSIsKill, bool WantResult = true);
194 unsigned emitSubs_rs(MVT RetVT, unsigned LHSReg, bool LHSIsKill,
195 unsigned RHSReg, bool RHSIsKill,
196 AArch64_AM::ShiftExtendType ShiftType, uint64_t ShiftImm,
197 bool WantResult = true);
198 unsigned emitLogicalOp(unsigned ISDOpc, MVT RetVT, const Value *LHS,
200 unsigned emitLogicalOp_ri(unsigned ISDOpc, MVT RetVT, unsigned LHSReg,
201 bool LHSIsKill, uint64_t Imm);
202 unsigned emitLogicalOp_rs(unsigned ISDOpc, MVT RetVT, unsigned LHSReg,
203 bool LHSIsKill, unsigned RHSReg, bool RHSIsKill,
205 unsigned emitAnd_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill, uint64_t Imm);
206 unsigned emitMul_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
207 unsigned Op1, bool Op1IsKill);
208 unsigned emitSMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
209 unsigned Op1, bool Op1IsKill);
210 unsigned emitUMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
211 unsigned Op1, bool Op1IsKill);
212 unsigned emitLSL_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
213 unsigned Op1Reg, bool Op1IsKill);
214 unsigned emitLSL_ri(MVT RetVT, MVT SrcVT, unsigned Op0Reg, bool Op0IsKill,
215 uint64_t Imm, bool IsZExt = true);
216 unsigned emitLSR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
217 unsigned Op1Reg, bool Op1IsKill);
218 unsigned emitLSR_ri(MVT RetVT, MVT SrcVT, unsigned Op0Reg, bool Op0IsKill,
219 uint64_t Imm, bool IsZExt = true);
220 unsigned emitASR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
221 unsigned Op1Reg, bool Op1IsKill);
222 unsigned emitASR_ri(MVT RetVT, MVT SrcVT, unsigned Op0Reg, bool Op0IsKill,
223 uint64_t Imm, bool IsZExt = false);
225 unsigned materializeInt(const ConstantInt *CI, MVT VT);
226 unsigned materializeFP(const ConstantFP *CFP, MVT VT);
227 unsigned materializeGV(const GlobalValue *GV);
229 // Call handling routines.
231 CCAssignFn *CCAssignFnForCall(CallingConv::ID CC) const;
232 bool processCallArgs(CallLoweringInfo &CLI, SmallVectorImpl<MVT> &ArgVTs,
234 bool finishCall(CallLoweringInfo &CLI, MVT RetVT, unsigned NumBytes);
237 // Backend specific FastISel code.
238 unsigned fastMaterializeAlloca(const AllocaInst *AI) override;
239 unsigned fastMaterializeConstant(const Constant *C) override;
240 unsigned fastMaterializeFloatZero(const ConstantFP* CF) override;
242 explicit AArch64FastISel(FunctionLoweringInfo &FuncInfo,
243 const TargetLibraryInfo *LibInfo)
244 : FastISel(FuncInfo, LibInfo, /*SkipTargetIndependentISel=*/true) {
245 Subtarget = &TM.getSubtarget<AArch64Subtarget>();
246 Context = &FuncInfo.Fn->getContext();
249 bool fastSelectInstruction(const Instruction *I) override;
251 #include "AArch64GenFastISel.inc"
254 } // end anonymous namespace
256 #include "AArch64GenCallingConv.inc"
258 CCAssignFn *AArch64FastISel::CCAssignFnForCall(CallingConv::ID CC) const {
259 if (CC == CallingConv::WebKit_JS)
260 return CC_AArch64_WebKit_JS;
261 return Subtarget->isTargetDarwin() ? CC_AArch64_DarwinPCS : CC_AArch64_AAPCS;
264 unsigned AArch64FastISel::fastMaterializeAlloca(const AllocaInst *AI) {
265 assert(TLI.getValueType(AI->getType(), true) == MVT::i64 &&
266 "Alloca should always return a pointer.");
268 // Don't handle dynamic allocas.
269 if (!FuncInfo.StaticAllocaMap.count(AI))
272 DenseMap<const AllocaInst *, int>::iterator SI =
273 FuncInfo.StaticAllocaMap.find(AI);
275 if (SI != FuncInfo.StaticAllocaMap.end()) {
276 unsigned ResultReg = createResultReg(&AArch64::GPR64spRegClass);
277 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri),
279 .addFrameIndex(SI->second)
288 unsigned AArch64FastISel::materializeInt(const ConstantInt *CI, MVT VT) {
293 return fastEmit_i(VT, VT, ISD::Constant, CI->getZExtValue());
295 // Create a copy from the zero register to materialize a "0" value.
296 const TargetRegisterClass *RC = (VT == MVT::i64) ? &AArch64::GPR64RegClass
297 : &AArch64::GPR32RegClass;
298 unsigned ZeroReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
299 unsigned ResultReg = createResultReg(RC);
300 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TargetOpcode::COPY),
301 ResultReg).addReg(ZeroReg, getKillRegState(true));
305 unsigned AArch64FastISel::materializeFP(const ConstantFP *CFP, MVT VT) {
306 // Positive zero (+0.0) has to be materialized with a fmov from the zero
307 // register, because the immediate version of fmov cannot encode zero.
308 if (CFP->isNullValue())
309 return fastMaterializeFloatZero(CFP);
311 if (VT != MVT::f32 && VT != MVT::f64)
314 const APFloat Val = CFP->getValueAPF();
315 bool Is64Bit = (VT == MVT::f64);
316 // This checks to see if we can use FMOV instructions to materialize
317 // a constant, otherwise we have to materialize via the constant pool.
318 if (TLI.isFPImmLegal(Val, VT)) {
320 Is64Bit ? AArch64_AM::getFP64Imm(Val) : AArch64_AM::getFP32Imm(Val);
321 assert((Imm != -1) && "Cannot encode floating-point constant.");
322 unsigned Opc = Is64Bit ? AArch64::FMOVDi : AArch64::FMOVSi;
323 return fastEmitInst_i(Opc, TLI.getRegClassFor(VT), Imm);
326 // Materialize via constant pool. MachineConstantPool wants an explicit
328 unsigned Align = DL.getPrefTypeAlignment(CFP->getType());
330 Align = DL.getTypeAllocSize(CFP->getType());
332 unsigned CPI = MCP.getConstantPoolIndex(cast<Constant>(CFP), Align);
333 unsigned ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
334 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
335 ADRPReg).addConstantPoolIndex(CPI, 0, AArch64II::MO_PAGE);
337 unsigned Opc = Is64Bit ? AArch64::LDRDui : AArch64::LDRSui;
338 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
339 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
341 .addConstantPoolIndex(CPI, 0, AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
345 unsigned AArch64FastISel::materializeGV(const GlobalValue *GV) {
346 // We can't handle thread-local variables quickly yet.
347 if (GV->isThreadLocal())
350 // MachO still uses GOT for large code-model accesses, but ELF requires
351 // movz/movk sequences, which FastISel doesn't handle yet.
352 if (TM.getCodeModel() != CodeModel::Small && !Subtarget->isTargetMachO())
355 unsigned char OpFlags = Subtarget->ClassifyGlobalReference(GV, TM);
357 EVT DestEVT = TLI.getValueType(GV->getType(), true);
358 if (!DestEVT.isSimple())
361 unsigned ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
364 if (OpFlags & AArch64II::MO_GOT) {
366 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
368 .addGlobalAddress(GV, 0, AArch64II::MO_GOT | AArch64II::MO_PAGE);
370 ResultReg = createResultReg(&AArch64::GPR64RegClass);
371 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::LDRXui),
374 .addGlobalAddress(GV, 0, AArch64II::MO_GOT | AArch64II::MO_PAGEOFF |
376 } else if (OpFlags & AArch64II::MO_CONSTPOOL) {
377 // We can't handle addresses loaded from a constant pool quickly yet.
381 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
383 .addGlobalAddress(GV, 0, AArch64II::MO_PAGE);
385 ResultReg = createResultReg(&AArch64::GPR64spRegClass);
386 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri),
389 .addGlobalAddress(GV, 0, AArch64II::MO_PAGEOFF | AArch64II::MO_NC)
395 unsigned AArch64FastISel::fastMaterializeConstant(const Constant *C) {
396 EVT CEVT = TLI.getValueType(C->getType(), true);
398 // Only handle simple types.
399 if (!CEVT.isSimple())
401 MVT VT = CEVT.getSimpleVT();
403 if (const auto *CI = dyn_cast<ConstantInt>(C))
404 return materializeInt(CI, VT);
405 else if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
406 return materializeFP(CFP, VT);
407 else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
408 return materializeGV(GV);
413 unsigned AArch64FastISel::fastMaterializeFloatZero(const ConstantFP* CFP) {
414 assert(CFP->isNullValue() &&
415 "Floating-point constant is not a positive zero.");
417 if (!isTypeLegal(CFP->getType(), VT))
420 if (VT != MVT::f32 && VT != MVT::f64)
423 bool Is64Bit = (VT == MVT::f64);
424 unsigned ZReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
425 unsigned Opc = Is64Bit ? AArch64::FMOVXDr : AArch64::FMOVWSr;
426 return fastEmitInst_r(Opc, TLI.getRegClassFor(VT), ZReg, /*IsKill=*/true);
429 /// \brief Check if the multiply is by a power-of-2 constant.
430 static bool isMulPowOf2(const Value *I) {
431 if (const auto *MI = dyn_cast<MulOperator>(I)) {
432 if (const auto *C = dyn_cast<ConstantInt>(MI->getOperand(0)))
433 if (C->getValue().isPowerOf2())
435 if (const auto *C = dyn_cast<ConstantInt>(MI->getOperand(1)))
436 if (C->getValue().isPowerOf2())
442 // Computes the address to get to an object.
443 bool AArch64FastISel::computeAddress(const Value *Obj, Address &Addr, Type *Ty)
445 const User *U = nullptr;
446 unsigned Opcode = Instruction::UserOp1;
447 if (const Instruction *I = dyn_cast<Instruction>(Obj)) {
448 // Don't walk into other basic blocks unless the object is an alloca from
449 // another block, otherwise it may not have a virtual register assigned.
450 if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(Obj)) ||
451 FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
452 Opcode = I->getOpcode();
455 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Obj)) {
456 Opcode = C->getOpcode();
460 if (const PointerType *Ty = dyn_cast<PointerType>(Obj->getType()))
461 if (Ty->getAddressSpace() > 255)
462 // Fast instruction selection doesn't support the special
469 case Instruction::BitCast: {
470 // Look through bitcasts.
471 return computeAddress(U->getOperand(0), Addr, Ty);
473 case Instruction::IntToPtr: {
474 // Look past no-op inttoptrs.
475 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
476 return computeAddress(U->getOperand(0), Addr, Ty);
479 case Instruction::PtrToInt: {
480 // Look past no-op ptrtoints.
481 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
482 return computeAddress(U->getOperand(0), Addr, Ty);
485 case Instruction::GetElementPtr: {
486 Address SavedAddr = Addr;
487 uint64_t TmpOffset = Addr.getOffset();
489 // Iterate through the GEP folding the constants into offsets where
491 gep_type_iterator GTI = gep_type_begin(U);
492 for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e;
494 const Value *Op = *i;
495 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
496 const StructLayout *SL = DL.getStructLayout(STy);
497 unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
498 TmpOffset += SL->getElementOffset(Idx);
500 uint64_t S = DL.getTypeAllocSize(GTI.getIndexedType());
502 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
503 // Constant-offset addressing.
504 TmpOffset += CI->getSExtValue() * S;
507 if (canFoldAddIntoGEP(U, Op)) {
508 // A compatible add with a constant operand. Fold the constant.
510 cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
511 TmpOffset += CI->getSExtValue() * S;
512 // Iterate on the other operand.
513 Op = cast<AddOperator>(Op)->getOperand(0);
517 goto unsupported_gep;
522 // Try to grab the base operand now.
523 Addr.setOffset(TmpOffset);
524 if (computeAddress(U->getOperand(0), Addr, Ty))
527 // We failed, restore everything and try the other options.
533 case Instruction::Alloca: {
534 const AllocaInst *AI = cast<AllocaInst>(Obj);
535 DenseMap<const AllocaInst *, int>::iterator SI =
536 FuncInfo.StaticAllocaMap.find(AI);
537 if (SI != FuncInfo.StaticAllocaMap.end()) {
538 Addr.setKind(Address::FrameIndexBase);
539 Addr.setFI(SI->second);
544 case Instruction::Add: {
545 // Adds of constants are common and easy enough.
546 const Value *LHS = U->getOperand(0);
547 const Value *RHS = U->getOperand(1);
549 if (isa<ConstantInt>(LHS))
552 if (const ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
553 Addr.setOffset(Addr.getOffset() + (uint64_t)CI->getSExtValue());
554 return computeAddress(LHS, Addr, Ty);
557 Address Backup = Addr;
558 if (computeAddress(LHS, Addr, Ty) && computeAddress(RHS, Addr, Ty))
564 case Instruction::Shl:
565 if (Addr.getOffsetReg())
568 if (const auto *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
569 unsigned Val = CI->getZExtValue();
570 if (Val < 1 || Val > 3)
573 uint64_t NumBytes = 0;
574 if (Ty && Ty->isSized()) {
575 uint64_t NumBits = DL.getTypeSizeInBits(Ty);
576 NumBytes = NumBits / 8;
577 if (!isPowerOf2_64(NumBits))
581 if (NumBytes != (1ULL << Val))
585 Addr.setExtendType(AArch64_AM::LSL);
587 if (const auto *I = dyn_cast<Instruction>(U->getOperand(0)))
588 if (FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB)
591 if (const auto *ZE = dyn_cast<ZExtInst>(U))
592 if (ZE->getOperand(0)->getType()->isIntegerTy(32))
593 Addr.setExtendType(AArch64_AM::UXTW);
595 if (const auto *SE = dyn_cast<SExtInst>(U))
596 if (SE->getOperand(0)->getType()->isIntegerTy(32))
597 Addr.setExtendType(AArch64_AM::SXTW);
599 if (const auto *AI = dyn_cast<BinaryOperator>(U))
600 if (AI->getOpcode() == Instruction::And) {
601 const Value *LHS = AI->getOperand(0);
602 const Value *RHS = AI->getOperand(1);
604 if (const auto *C = dyn_cast<ConstantInt>(LHS))
605 if (C->getValue() == 0xffffffff)
608 if (const auto *C = cast<ConstantInt>(RHS))
609 if (C->getValue() == 0xffffffff) {
610 Addr.setExtendType(AArch64_AM::UXTW);
611 unsigned Reg = getRegForValue(LHS);
614 bool RegIsKill = hasTrivialKill(LHS);
615 Reg = fastEmitInst_extractsubreg(MVT::i32, Reg, RegIsKill,
617 Addr.setOffsetReg(Reg);
622 unsigned Reg = getRegForValue(U->getOperand(0));
625 Addr.setOffsetReg(Reg);
629 case Instruction::Mul: {
630 if (Addr.getOffsetReg())
636 const Value *LHS = U->getOperand(0);
637 const Value *RHS = U->getOperand(1);
639 // Canonicalize power-of-2 value to the RHS.
640 if (const auto *C = dyn_cast<ConstantInt>(LHS))
641 if (C->getValue().isPowerOf2())
644 assert(isa<ConstantInt>(RHS) && "Expected an ConstantInt.");
645 const auto *C = cast<ConstantInt>(RHS);
646 unsigned Val = C->getValue().logBase2();
647 if (Val < 1 || Val > 3)
650 uint64_t NumBytes = 0;
651 if (Ty && Ty->isSized()) {
652 uint64_t NumBits = DL.getTypeSizeInBits(Ty);
653 NumBytes = NumBits / 8;
654 if (!isPowerOf2_64(NumBits))
658 if (NumBytes != (1ULL << Val))
662 Addr.setExtendType(AArch64_AM::LSL);
664 if (const auto *I = dyn_cast<Instruction>(LHS))
665 if (FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB)
668 if (const auto *ZE = dyn_cast<ZExtInst>(U))
669 if (ZE->getOperand(0)->getType()->isIntegerTy(32)) {
670 Addr.setExtendType(AArch64_AM::UXTW);
671 LHS = U->getOperand(0);
674 if (const auto *SE = dyn_cast<SExtInst>(U))
675 if (SE->getOperand(0)->getType()->isIntegerTy(32)) {
676 Addr.setExtendType(AArch64_AM::SXTW);
677 LHS = U->getOperand(0);
680 unsigned Reg = getRegForValue(LHS);
683 Addr.setOffsetReg(Reg);
686 case Instruction::And: {
687 if (Addr.getOffsetReg())
690 if (DL.getTypeSizeInBits(Ty) != 8)
693 const Value *LHS = U->getOperand(0);
694 const Value *RHS = U->getOperand(1);
696 if (const auto *C = dyn_cast<ConstantInt>(LHS))
697 if (C->getValue() == 0xffffffff)
700 if (const auto *C = cast<ConstantInt>(RHS))
701 if (C->getValue() == 0xffffffff) {
703 Addr.setExtendType(AArch64_AM::LSL);
704 Addr.setExtendType(AArch64_AM::UXTW);
706 unsigned Reg = getRegForValue(LHS);
709 bool RegIsKill = hasTrivialKill(LHS);
710 Reg = fastEmitInst_extractsubreg(MVT::i32, Reg, RegIsKill,
712 Addr.setOffsetReg(Reg);
720 if (!Addr.getOffsetReg()) {
721 unsigned Reg = getRegForValue(Obj);
724 Addr.setOffsetReg(Reg);
730 unsigned Reg = getRegForValue(Obj);
737 bool AArch64FastISel::computeCallAddress(const Value *V, Address &Addr) {
738 const User *U = nullptr;
739 unsigned Opcode = Instruction::UserOp1;
742 if (const auto *I = dyn_cast<Instruction>(V)) {
743 Opcode = I->getOpcode();
745 InMBB = I->getParent() == FuncInfo.MBB->getBasicBlock();
746 } else if (const auto *C = dyn_cast<ConstantExpr>(V)) {
747 Opcode = C->getOpcode();
753 case Instruction::BitCast:
754 // Look past bitcasts if its operand is in the same BB.
756 return computeCallAddress(U->getOperand(0), Addr);
758 case Instruction::IntToPtr:
759 // Look past no-op inttoptrs if its operand is in the same BB.
761 TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
762 return computeCallAddress(U->getOperand(0), Addr);
764 case Instruction::PtrToInt:
765 // Look past no-op ptrtoints if its operand is in the same BB.
767 TLI.getValueType(U->getType()) == TLI.getPointerTy())
768 return computeCallAddress(U->getOperand(0), Addr);
772 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
773 Addr.setGlobalValue(GV);
777 // If all else fails, try to materialize the value in a register.
778 if (!Addr.getGlobalValue()) {
779 Addr.setReg(getRegForValue(V));
780 return Addr.getReg() != 0;
787 bool AArch64FastISel::isTypeLegal(Type *Ty, MVT &VT) {
788 EVT evt = TLI.getValueType(Ty, true);
790 // Only handle simple types.
791 if (evt == MVT::Other || !evt.isSimple())
793 VT = evt.getSimpleVT();
795 // This is a legal type, but it's not something we handle in fast-isel.
799 // Handle all other legal types, i.e. a register that will directly hold this
801 return TLI.isTypeLegal(VT);
804 /// \brief Determine if the value type is supported by FastISel.
806 /// FastISel for AArch64 can handle more value types than are legal. This adds
807 /// simple value type such as i1, i8, and i16.
808 bool AArch64FastISel::isTypeSupported(Type *Ty, MVT &VT, bool IsVectorAllowed) {
809 if (Ty->isVectorTy() && !IsVectorAllowed)
812 if (isTypeLegal(Ty, VT))
815 // If this is a type than can be sign or zero-extended to a basic operation
816 // go ahead and accept it now.
817 if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16)
823 bool AArch64FastISel::isValueAvailable(const Value *V) const {
824 if (!isa<Instruction>(V))
827 const auto *I = cast<Instruction>(V);
828 if (FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB)
834 bool AArch64FastISel::simplifyAddress(Address &Addr, MVT VT) {
835 unsigned ScaleFactor;
836 switch (VT.SimpleTy) {
837 default: return false;
838 case MVT::i1: // fall-through
839 case MVT::i8: ScaleFactor = 1; break;
840 case MVT::i16: ScaleFactor = 2; break;
841 case MVT::i32: // fall-through
842 case MVT::f32: ScaleFactor = 4; break;
843 case MVT::i64: // fall-through
844 case MVT::f64: ScaleFactor = 8; break;
847 bool ImmediateOffsetNeedsLowering = false;
848 bool RegisterOffsetNeedsLowering = false;
849 int64_t Offset = Addr.getOffset();
850 if (((Offset < 0) || (Offset & (ScaleFactor - 1))) && !isInt<9>(Offset))
851 ImmediateOffsetNeedsLowering = true;
852 else if (Offset > 0 && !(Offset & (ScaleFactor - 1)) &&
853 !isUInt<12>(Offset / ScaleFactor))
854 ImmediateOffsetNeedsLowering = true;
856 // Cannot encode an offset register and an immediate offset in the same
857 // instruction. Fold the immediate offset into the load/store instruction and
858 // emit an additonal add to take care of the offset register.
859 if (!ImmediateOffsetNeedsLowering && Addr.getOffset() && Addr.isRegBase() &&
861 RegisterOffsetNeedsLowering = true;
863 // Cannot encode zero register as base.
864 if (Addr.isRegBase() && Addr.getOffsetReg() && !Addr.getReg())
865 RegisterOffsetNeedsLowering = true;
867 // If this is a stack pointer and the offset needs to be simplified then put
868 // the alloca address into a register, set the base type back to register and
869 // continue. This should almost never happen.
870 if (ImmediateOffsetNeedsLowering && Addr.isFIBase()) {
871 unsigned ResultReg = createResultReg(&AArch64::GPR64spRegClass);
872 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri),
874 .addFrameIndex(Addr.getFI())
877 Addr.setKind(Address::RegBase);
878 Addr.setReg(ResultReg);
881 if (RegisterOffsetNeedsLowering) {
882 unsigned ResultReg = 0;
884 if (Addr.getExtendType() == AArch64_AM::SXTW ||
885 Addr.getExtendType() == AArch64_AM::UXTW )
886 ResultReg = emitAddSub_rx(/*UseAdd=*/true, MVT::i64, Addr.getReg(),
887 /*TODO:IsKill=*/false, Addr.getOffsetReg(),
888 /*TODO:IsKill=*/false, Addr.getExtendType(),
891 ResultReg = emitAddSub_rs(/*UseAdd=*/true, MVT::i64, Addr.getReg(),
892 /*TODO:IsKill=*/false, Addr.getOffsetReg(),
893 /*TODO:IsKill=*/false, AArch64_AM::LSL,
896 if (Addr.getExtendType() == AArch64_AM::UXTW)
897 ResultReg = emitLSL_ri(MVT::i64, MVT::i32, Addr.getOffsetReg(),
898 /*Op0IsKill=*/false, Addr.getShift(),
900 else if (Addr.getExtendType() == AArch64_AM::SXTW)
901 ResultReg = emitLSL_ri(MVT::i64, MVT::i32, Addr.getOffsetReg(),
902 /*Op0IsKill=*/false, Addr.getShift(),
905 ResultReg = emitLSL_ri(MVT::i64, MVT::i64, Addr.getOffsetReg(),
906 /*Op0IsKill=*/false, Addr.getShift());
911 Addr.setReg(ResultReg);
912 Addr.setOffsetReg(0);
914 Addr.setExtendType(AArch64_AM::InvalidShiftExtend);
917 // Since the offset is too large for the load/store instruction get the
918 // reg+offset into a register.
919 if (ImmediateOffsetNeedsLowering) {
920 unsigned ResultReg = 0;
922 ResultReg = fastEmit_ri_(MVT::i64, ISD::ADD, Addr.getReg(),
923 /*IsKill=*/false, Offset, MVT::i64);
925 ResultReg = fastEmit_i(MVT::i64, MVT::i64, ISD::Constant, Offset);
929 Addr.setReg(ResultReg);
935 void AArch64FastISel::addLoadStoreOperands(Address &Addr,
936 const MachineInstrBuilder &MIB,
938 unsigned ScaleFactor,
939 MachineMemOperand *MMO) {
940 int64_t Offset = Addr.getOffset() / ScaleFactor;
941 // Frame base works a bit differently. Handle it separately.
942 if (Addr.isFIBase()) {
943 int FI = Addr.getFI();
944 // FIXME: We shouldn't be using getObjectSize/getObjectAlignment. The size
945 // and alignment should be based on the VT.
946 MMO = FuncInfo.MF->getMachineMemOperand(
947 MachinePointerInfo::getFixedStack(FI, Offset), Flags,
948 MFI.getObjectSize(FI), MFI.getObjectAlignment(FI));
949 // Now add the rest of the operands.
950 MIB.addFrameIndex(FI).addImm(Offset);
952 assert(Addr.isRegBase() && "Unexpected address kind.");
953 const MCInstrDesc &II = MIB->getDesc();
954 unsigned Idx = (Flags & MachineMemOperand::MOStore) ? 1 : 0;
956 constrainOperandRegClass(II, Addr.getReg(), II.getNumDefs()+Idx));
958 constrainOperandRegClass(II, Addr.getOffsetReg(), II.getNumDefs()+Idx+1));
959 if (Addr.getOffsetReg()) {
960 assert(Addr.getOffset() == 0 && "Unexpected offset");
961 bool IsSigned = Addr.getExtendType() == AArch64_AM::SXTW ||
962 Addr.getExtendType() == AArch64_AM::SXTX;
963 MIB.addReg(Addr.getReg());
964 MIB.addReg(Addr.getOffsetReg());
965 MIB.addImm(IsSigned);
966 MIB.addImm(Addr.getShift() != 0);
968 MIB.addReg(Addr.getReg());
974 MIB.addMemOperand(MMO);
977 unsigned AArch64FastISel::emitAddSub(bool UseAdd, MVT RetVT, const Value *LHS,
978 const Value *RHS, bool SetFlags,
979 bool WantResult, bool IsZExt) {
980 AArch64_AM::ShiftExtendType ExtendType = AArch64_AM::InvalidShiftExtend;
981 bool NeedExtend = false;
982 switch (RetVT.SimpleTy) {
990 ExtendType = IsZExt ? AArch64_AM::UXTB : AArch64_AM::SXTB;
994 ExtendType = IsZExt ? AArch64_AM::UXTH : AArch64_AM::SXTH;
996 case MVT::i32: // fall-through
1001 RetVT.SimpleTy = std::max(RetVT.SimpleTy, MVT::i32);
1003 // Canonicalize immediates to the RHS first.
1004 if (UseAdd && isa<ConstantInt>(LHS) && !isa<ConstantInt>(RHS))
1005 std::swap(LHS, RHS);
1007 // Canonicalize mul by power of 2 to the RHS.
1008 if (UseAdd && LHS->hasOneUse() && isValueAvailable(LHS))
1009 if (isMulPowOf2(LHS))
1010 std::swap(LHS, RHS);
1012 // Canonicalize shift immediate to the RHS.
1013 if (UseAdd && LHS->hasOneUse() && isValueAvailable(LHS))
1014 if (const auto *SI = dyn_cast<BinaryOperator>(LHS))
1015 if (isa<ConstantInt>(SI->getOperand(1)))
1016 if (SI->getOpcode() == Instruction::Shl ||
1017 SI->getOpcode() == Instruction::LShr ||
1018 SI->getOpcode() == Instruction::AShr )
1019 std::swap(LHS, RHS);
1021 unsigned LHSReg = getRegForValue(LHS);
1024 bool LHSIsKill = hasTrivialKill(LHS);
1027 LHSReg = emitIntExt(SrcVT, LHSReg, RetVT, IsZExt);
1029 unsigned ResultReg = 0;
1030 if (const auto *C = dyn_cast<ConstantInt>(RHS)) {
1031 uint64_t Imm = IsZExt ? C->getZExtValue() : C->getSExtValue();
1032 if (C->isNegative())
1033 ResultReg = emitAddSub_ri(!UseAdd, RetVT, LHSReg, LHSIsKill, -Imm,
1034 SetFlags, WantResult);
1036 ResultReg = emitAddSub_ri(UseAdd, RetVT, LHSReg, LHSIsKill, Imm, SetFlags,
1042 // Only extend the RHS within the instruction if there is a valid extend type.
1043 if (ExtendType != AArch64_AM::InvalidShiftExtend && RHS->hasOneUse() &&
1044 isValueAvailable(RHS)) {
1045 if (const auto *SI = dyn_cast<BinaryOperator>(RHS))
1046 if (const auto *C = dyn_cast<ConstantInt>(SI->getOperand(1)))
1047 if ((SI->getOpcode() == Instruction::Shl) && (C->getZExtValue() < 4)) {
1048 unsigned RHSReg = getRegForValue(SI->getOperand(0));
1051 bool RHSIsKill = hasTrivialKill(SI->getOperand(0));
1052 return emitAddSub_rx(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg,
1053 RHSIsKill, ExtendType, C->getZExtValue(),
1054 SetFlags, WantResult);
1056 unsigned RHSReg = getRegForValue(RHS);
1059 bool RHSIsKill = hasTrivialKill(RHS);
1060 return emitAddSub_rx(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg, RHSIsKill,
1061 ExtendType, 0, SetFlags, WantResult);
1064 // Check if the mul can be folded into the instruction.
1065 if (RHS->hasOneUse() && isValueAvailable(RHS))
1066 if (isMulPowOf2(RHS)) {
1067 const Value *MulLHS = cast<MulOperator>(RHS)->getOperand(0);
1068 const Value *MulRHS = cast<MulOperator>(RHS)->getOperand(1);
1070 if (const auto *C = dyn_cast<ConstantInt>(MulLHS))
1071 if (C->getValue().isPowerOf2())
1072 std::swap(MulLHS, MulRHS);
1074 assert(isa<ConstantInt>(MulRHS) && "Expected a ConstantInt.");
1075 uint64_t ShiftVal = cast<ConstantInt>(MulRHS)->getValue().logBase2();
1076 unsigned RHSReg = getRegForValue(MulLHS);
1079 bool RHSIsKill = hasTrivialKill(MulLHS);
1080 return emitAddSub_rs(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg, RHSIsKill,
1081 AArch64_AM::LSL, ShiftVal, SetFlags, WantResult);
1084 // Check if the shift can be folded into the instruction.
1085 if (RHS->hasOneUse() && isValueAvailable(RHS))
1086 if (const auto *SI = dyn_cast<BinaryOperator>(RHS)) {
1087 if (const auto *C = dyn_cast<ConstantInt>(SI->getOperand(1))) {
1088 AArch64_AM::ShiftExtendType ShiftType = AArch64_AM::InvalidShiftExtend;
1089 switch (SI->getOpcode()) {
1091 case Instruction::Shl: ShiftType = AArch64_AM::LSL; break;
1092 case Instruction::LShr: ShiftType = AArch64_AM::LSR; break;
1093 case Instruction::AShr: ShiftType = AArch64_AM::ASR; break;
1095 uint64_t ShiftVal = C->getZExtValue();
1096 if (ShiftType != AArch64_AM::InvalidShiftExtend) {
1097 unsigned RHSReg = getRegForValue(SI->getOperand(0));
1100 bool RHSIsKill = hasTrivialKill(SI->getOperand(0));
1101 return emitAddSub_rs(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg,
1102 RHSIsKill, ShiftType, ShiftVal, SetFlags,
1108 unsigned RHSReg = getRegForValue(RHS);
1111 bool RHSIsKill = hasTrivialKill(RHS);
1114 RHSReg = emitIntExt(SrcVT, RHSReg, RetVT, IsZExt);
1116 return emitAddSub_rr(UseAdd, RetVT, LHSReg, LHSIsKill, RHSReg, RHSIsKill,
1117 SetFlags, WantResult);
1120 unsigned AArch64FastISel::emitAddSub_rr(bool UseAdd, MVT RetVT, unsigned LHSReg,
1121 bool LHSIsKill, unsigned RHSReg,
1122 bool RHSIsKill, bool SetFlags,
1124 assert(LHSReg && RHSReg && "Invalid register number.");
1126 if (RetVT != MVT::i32 && RetVT != MVT::i64)
1129 static const unsigned OpcTable[2][2][2] = {
1130 { { AArch64::SUBWrr, AArch64::SUBXrr },
1131 { AArch64::ADDWrr, AArch64::ADDXrr } },
1132 { { AArch64::SUBSWrr, AArch64::SUBSXrr },
1133 { AArch64::ADDSWrr, AArch64::ADDSXrr } }
1135 bool Is64Bit = RetVT == MVT::i64;
1136 unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit];
1137 const TargetRegisterClass *RC =
1138 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
1141 ResultReg = createResultReg(RC);
1143 ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
1145 const MCInstrDesc &II = TII.get(Opc);
1146 LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
1147 RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1);
1148 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
1149 .addReg(LHSReg, getKillRegState(LHSIsKill))
1150 .addReg(RHSReg, getKillRegState(RHSIsKill));
1154 unsigned AArch64FastISel::emitAddSub_ri(bool UseAdd, MVT RetVT, unsigned LHSReg,
1155 bool LHSIsKill, uint64_t Imm,
1156 bool SetFlags, bool WantResult) {
1157 assert(LHSReg && "Invalid register number.");
1159 if (RetVT != MVT::i32 && RetVT != MVT::i64)
1163 if (isUInt<12>(Imm))
1165 else if ((Imm & 0xfff000) == Imm) {
1171 static const unsigned OpcTable[2][2][2] = {
1172 { { AArch64::SUBWri, AArch64::SUBXri },
1173 { AArch64::ADDWri, AArch64::ADDXri } },
1174 { { AArch64::SUBSWri, AArch64::SUBSXri },
1175 { AArch64::ADDSWri, AArch64::ADDSXri } }
1177 bool Is64Bit = RetVT == MVT::i64;
1178 unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit];
1179 const TargetRegisterClass *RC;
1181 RC = Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
1183 RC = Is64Bit ? &AArch64::GPR64spRegClass : &AArch64::GPR32spRegClass;
1186 ResultReg = createResultReg(RC);
1188 ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
1190 const MCInstrDesc &II = TII.get(Opc);
1191 LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
1192 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
1193 .addReg(LHSReg, getKillRegState(LHSIsKill))
1195 .addImm(getShifterImm(AArch64_AM::LSL, ShiftImm));
1199 unsigned AArch64FastISel::emitAddSub_rs(bool UseAdd, MVT RetVT, unsigned LHSReg,
1200 bool LHSIsKill, unsigned RHSReg,
1202 AArch64_AM::ShiftExtendType ShiftType,
1203 uint64_t ShiftImm, bool SetFlags,
1205 assert(LHSReg && RHSReg && "Invalid register number.");
1207 if (RetVT != MVT::i32 && RetVT != MVT::i64)
1210 static const unsigned OpcTable[2][2][2] = {
1211 { { AArch64::SUBWrs, AArch64::SUBXrs },
1212 { AArch64::ADDWrs, AArch64::ADDXrs } },
1213 { { AArch64::SUBSWrs, AArch64::SUBSXrs },
1214 { AArch64::ADDSWrs, AArch64::ADDSXrs } }
1216 bool Is64Bit = RetVT == MVT::i64;
1217 unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit];
1218 const TargetRegisterClass *RC =
1219 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
1222 ResultReg = createResultReg(RC);
1224 ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
1226 const MCInstrDesc &II = TII.get(Opc);
1227 LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
1228 RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1);
1229 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
1230 .addReg(LHSReg, getKillRegState(LHSIsKill))
1231 .addReg(RHSReg, getKillRegState(RHSIsKill))
1232 .addImm(getShifterImm(ShiftType, ShiftImm));
1236 unsigned AArch64FastISel::emitAddSub_rx(bool UseAdd, MVT RetVT, unsigned LHSReg,
1237 bool LHSIsKill, unsigned RHSReg,
1239 AArch64_AM::ShiftExtendType ExtType,
1240 uint64_t ShiftImm, bool SetFlags,
1242 assert(LHSReg && RHSReg && "Invalid register number.");
1244 if (RetVT != MVT::i32 && RetVT != MVT::i64)
1247 static const unsigned OpcTable[2][2][2] = {
1248 { { AArch64::SUBWrx, AArch64::SUBXrx },
1249 { AArch64::ADDWrx, AArch64::ADDXrx } },
1250 { { AArch64::SUBSWrx, AArch64::SUBSXrx },
1251 { AArch64::ADDSWrx, AArch64::ADDSXrx } }
1253 bool Is64Bit = RetVT == MVT::i64;
1254 unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit];
1255 const TargetRegisterClass *RC = nullptr;
1257 RC = Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
1259 RC = Is64Bit ? &AArch64::GPR64spRegClass : &AArch64::GPR32spRegClass;
1262 ResultReg = createResultReg(RC);
1264 ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
1266 const MCInstrDesc &II = TII.get(Opc);
1267 LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
1268 RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1);
1269 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
1270 .addReg(LHSReg, getKillRegState(LHSIsKill))
1271 .addReg(RHSReg, getKillRegState(RHSIsKill))
1272 .addImm(getArithExtendImm(ExtType, ShiftImm));
1276 bool AArch64FastISel::emitCmp(const Value *LHS, const Value *RHS, bool IsZExt) {
1277 Type *Ty = LHS->getType();
1278 EVT EVT = TLI.getValueType(Ty, true);
1279 if (!EVT.isSimple())
1281 MVT VT = EVT.getSimpleVT();
1283 switch (VT.SimpleTy) {
1291 return emitICmp(VT, LHS, RHS, IsZExt);
1294 return emitFCmp(VT, LHS, RHS);
1298 bool AArch64FastISel::emitICmp(MVT RetVT, const Value *LHS, const Value *RHS,
1300 return emitSub(RetVT, LHS, RHS, /*SetFlags=*/true, /*WantResult=*/false,
1304 bool AArch64FastISel::emitICmp_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill,
1306 return emitAddSub_ri(/*UseAdd=*/false, RetVT, LHSReg, LHSIsKill, Imm,
1307 /*SetFlags=*/true, /*WantResult=*/false) != 0;
1310 bool AArch64FastISel::emitFCmp(MVT RetVT, const Value *LHS, const Value *RHS) {
1311 if (RetVT != MVT::f32 && RetVT != MVT::f64)
1314 // Check to see if the 2nd operand is a constant that we can encode directly
1316 bool UseImm = false;
1317 if (const auto *CFP = dyn_cast<ConstantFP>(RHS))
1318 if (CFP->isZero() && !CFP->isNegative())
1321 unsigned LHSReg = getRegForValue(LHS);
1324 bool LHSIsKill = hasTrivialKill(LHS);
1327 unsigned Opc = (RetVT == MVT::f64) ? AArch64::FCMPDri : AArch64::FCMPSri;
1328 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc))
1329 .addReg(LHSReg, getKillRegState(LHSIsKill));
1333 unsigned RHSReg = getRegForValue(RHS);
1336 bool RHSIsKill = hasTrivialKill(RHS);
1338 unsigned Opc = (RetVT == MVT::f64) ? AArch64::FCMPDrr : AArch64::FCMPSrr;
1339 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc))
1340 .addReg(LHSReg, getKillRegState(LHSIsKill))
1341 .addReg(RHSReg, getKillRegState(RHSIsKill));
1345 unsigned AArch64FastISel::emitAdd(MVT RetVT, const Value *LHS, const Value *RHS,
1346 bool SetFlags, bool WantResult, bool IsZExt) {
1347 return emitAddSub(/*UseAdd=*/true, RetVT, LHS, RHS, SetFlags, WantResult,
1351 unsigned AArch64FastISel::emitSub(MVT RetVT, const Value *LHS, const Value *RHS,
1352 bool SetFlags, bool WantResult, bool IsZExt) {
1353 return emitAddSub(/*UseAdd=*/false, RetVT, LHS, RHS, SetFlags, WantResult,
1357 unsigned AArch64FastISel::emitSubs_rr(MVT RetVT, unsigned LHSReg,
1358 bool LHSIsKill, unsigned RHSReg,
1359 bool RHSIsKill, bool WantResult) {
1360 return emitAddSub_rr(/*UseAdd=*/false, RetVT, LHSReg, LHSIsKill, RHSReg,
1361 RHSIsKill, /*SetFlags=*/true, WantResult);
1364 unsigned AArch64FastISel::emitSubs_rs(MVT RetVT, unsigned LHSReg,
1365 bool LHSIsKill, unsigned RHSReg,
1367 AArch64_AM::ShiftExtendType ShiftType,
1368 uint64_t ShiftImm, bool WantResult) {
1369 return emitAddSub_rs(/*UseAdd=*/false, RetVT, LHSReg, LHSIsKill, RHSReg,
1370 RHSIsKill, ShiftType, ShiftImm, /*SetFlags=*/true,
1374 unsigned AArch64FastISel::emitLogicalOp(unsigned ISDOpc, MVT RetVT,
1375 const Value *LHS, const Value *RHS) {
1376 // Canonicalize immediates to the RHS first.
1377 if (isa<ConstantInt>(LHS) && !isa<ConstantInt>(RHS))
1378 std::swap(LHS, RHS);
1380 // Canonicalize mul by power-of-2 to the RHS.
1381 if (LHS->hasOneUse() && isValueAvailable(LHS))
1382 if (isMulPowOf2(LHS))
1383 std::swap(LHS, RHS);
1385 // Canonicalize shift immediate to the RHS.
1386 if (LHS->hasOneUse() && isValueAvailable(LHS))
1387 if (const auto *SI = dyn_cast<ShlOperator>(LHS))
1388 if (isa<ConstantInt>(SI->getOperand(1)))
1389 std::swap(LHS, RHS);
1391 unsigned LHSReg = getRegForValue(LHS);
1394 bool LHSIsKill = hasTrivialKill(LHS);
1396 unsigned ResultReg = 0;
1397 if (const auto *C = dyn_cast<ConstantInt>(RHS)) {
1398 uint64_t Imm = C->getZExtValue();
1399 ResultReg = emitLogicalOp_ri(ISDOpc, RetVT, LHSReg, LHSIsKill, Imm);
1404 // Check if the mul can be folded into the instruction.
1405 if (RHS->hasOneUse() && isValueAvailable(RHS))
1406 if (isMulPowOf2(RHS)) {
1407 const Value *MulLHS = cast<MulOperator>(RHS)->getOperand(0);
1408 const Value *MulRHS = cast<MulOperator>(RHS)->getOperand(1);
1410 if (const auto *C = dyn_cast<ConstantInt>(MulLHS))
1411 if (C->getValue().isPowerOf2())
1412 std::swap(MulLHS, MulRHS);
1414 assert(isa<ConstantInt>(MulRHS) && "Expected a ConstantInt.");
1415 uint64_t ShiftVal = cast<ConstantInt>(MulRHS)->getValue().logBase2();
1417 unsigned RHSReg = getRegForValue(MulLHS);
1420 bool RHSIsKill = hasTrivialKill(MulLHS);
1421 return emitLogicalOp_rs(ISDOpc, RetVT, LHSReg, LHSIsKill, RHSReg,
1422 RHSIsKill, ShiftVal);
1425 // Check if the shift can be folded into the instruction.
1426 if (RHS->hasOneUse() && isValueAvailable(RHS))
1427 if (const auto *SI = dyn_cast<ShlOperator>(RHS))
1428 if (const auto *C = dyn_cast<ConstantInt>(SI->getOperand(1))) {
1429 uint64_t ShiftVal = C->getZExtValue();
1430 unsigned RHSReg = getRegForValue(SI->getOperand(0));
1433 bool RHSIsKill = hasTrivialKill(SI->getOperand(0));
1434 return emitLogicalOp_rs(ISDOpc, RetVT, LHSReg, LHSIsKill, RHSReg,
1435 RHSIsKill, ShiftVal);
1438 unsigned RHSReg = getRegForValue(RHS);
1441 bool RHSIsKill = hasTrivialKill(RHS);
1443 MVT VT = std::max(MVT::i32, RetVT.SimpleTy);
1444 ResultReg = fastEmit_rr(VT, VT, ISDOpc, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
1445 if (RetVT >= MVT::i8 && RetVT <= MVT::i16) {
1446 uint64_t Mask = (RetVT == MVT::i8) ? 0xff : 0xffff;
1447 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
1452 unsigned AArch64FastISel::emitLogicalOp_ri(unsigned ISDOpc, MVT RetVT,
1453 unsigned LHSReg, bool LHSIsKill,
1455 assert((ISD::AND + 1 == ISD::OR) && (ISD::AND + 2 == ISD::XOR) &&
1456 "ISD nodes are not consecutive!");
1457 static const unsigned OpcTable[3][2] = {
1458 { AArch64::ANDWri, AArch64::ANDXri },
1459 { AArch64::ORRWri, AArch64::ORRXri },
1460 { AArch64::EORWri, AArch64::EORXri }
1462 const TargetRegisterClass *RC;
1465 switch (RetVT.SimpleTy) {
1472 unsigned Idx = ISDOpc - ISD::AND;
1473 Opc = OpcTable[Idx][0];
1474 RC = &AArch64::GPR32spRegClass;
1479 Opc = OpcTable[ISDOpc - ISD::AND][1];
1480 RC = &AArch64::GPR64spRegClass;
1485 if (!AArch64_AM::isLogicalImmediate(Imm, RegSize))
1488 unsigned ResultReg =
1489 fastEmitInst_ri(Opc, RC, LHSReg, LHSIsKill,
1490 AArch64_AM::encodeLogicalImmediate(Imm, RegSize));
1491 if (RetVT >= MVT::i8 && RetVT <= MVT::i16 && ISDOpc != ISD::AND) {
1492 uint64_t Mask = (RetVT == MVT::i8) ? 0xff : 0xffff;
1493 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
1498 unsigned AArch64FastISel::emitLogicalOp_rs(unsigned ISDOpc, MVT RetVT,
1499 unsigned LHSReg, bool LHSIsKill,
1500 unsigned RHSReg, bool RHSIsKill,
1501 uint64_t ShiftImm) {
1502 assert((ISD::AND + 1 == ISD::OR) && (ISD::AND + 2 == ISD::XOR) &&
1503 "ISD nodes are not consecutive!");
1504 static const unsigned OpcTable[3][2] = {
1505 { AArch64::ANDWrs, AArch64::ANDXrs },
1506 { AArch64::ORRWrs, AArch64::ORRXrs },
1507 { AArch64::EORWrs, AArch64::EORXrs }
1509 const TargetRegisterClass *RC;
1511 switch (RetVT.SimpleTy) {
1518 Opc = OpcTable[ISDOpc - ISD::AND][0];
1519 RC = &AArch64::GPR32RegClass;
1522 Opc = OpcTable[ISDOpc - ISD::AND][1];
1523 RC = &AArch64::GPR64RegClass;
1526 unsigned ResultReg =
1527 fastEmitInst_rri(Opc, RC, LHSReg, LHSIsKill, RHSReg, RHSIsKill,
1528 AArch64_AM::getShifterImm(AArch64_AM::LSL, ShiftImm));
1529 if (RetVT >= MVT::i8 && RetVT <= MVT::i16) {
1530 uint64_t Mask = (RetVT == MVT::i8) ? 0xff : 0xffff;
1531 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
1536 unsigned AArch64FastISel::emitAnd_ri(MVT RetVT, unsigned LHSReg, bool LHSIsKill,
1538 return emitLogicalOp_ri(ISD::AND, RetVT, LHSReg, LHSIsKill, Imm);
1541 bool AArch64FastISel::emitLoad(MVT VT, unsigned &ResultReg, Address Addr,
1542 MachineMemOperand *MMO) {
1543 // Simplify this down to something we can handle.
1544 if (!simplifyAddress(Addr, VT))
1547 unsigned ScaleFactor;
1548 switch (VT.SimpleTy) {
1549 default: llvm_unreachable("Unexpected value type.");
1550 case MVT::i1: // fall-through
1551 case MVT::i8: ScaleFactor = 1; break;
1552 case MVT::i16: ScaleFactor = 2; break;
1553 case MVT::i32: // fall-through
1554 case MVT::f32: ScaleFactor = 4; break;
1555 case MVT::i64: // fall-through
1556 case MVT::f64: ScaleFactor = 8; break;
1559 // Negative offsets require unscaled, 9-bit, signed immediate offsets.
1560 // Otherwise, we try using scaled, 12-bit, unsigned immediate offsets.
1561 bool UseScaled = true;
1562 if ((Addr.getOffset() < 0) || (Addr.getOffset() & (ScaleFactor - 1))) {
1567 static const unsigned OpcTable[4][6] = {
1568 { AArch64::LDURBBi, AArch64::LDURHHi, AArch64::LDURWi, AArch64::LDURXi,
1569 AArch64::LDURSi, AArch64::LDURDi },
1570 { AArch64::LDRBBui, AArch64::LDRHHui, AArch64::LDRWui, AArch64::LDRXui,
1571 AArch64::LDRSui, AArch64::LDRDui },
1572 { AArch64::LDRBBroX, AArch64::LDRHHroX, AArch64::LDRWroX, AArch64::LDRXroX,
1573 AArch64::LDRSroX, AArch64::LDRDroX },
1574 { AArch64::LDRBBroW, AArch64::LDRHHroW, AArch64::LDRWroW, AArch64::LDRXroW,
1575 AArch64::LDRSroW, AArch64::LDRDroW }
1579 const TargetRegisterClass *RC;
1580 bool VTIsi1 = false;
1581 bool UseRegOffset = Addr.isRegBase() && !Addr.getOffset() && Addr.getReg() &&
1582 Addr.getOffsetReg();
1583 unsigned Idx = UseRegOffset ? 2 : UseScaled ? 1 : 0;
1584 if (Addr.getExtendType() == AArch64_AM::UXTW ||
1585 Addr.getExtendType() == AArch64_AM::SXTW)
1588 switch (VT.SimpleTy) {
1589 default: llvm_unreachable("Unexpected value type.");
1590 case MVT::i1: VTIsi1 = true; // Intentional fall-through.
1591 case MVT::i8: Opc = OpcTable[Idx][0]; RC = &AArch64::GPR32RegClass; break;
1592 case MVT::i16: Opc = OpcTable[Idx][1]; RC = &AArch64::GPR32RegClass; break;
1593 case MVT::i32: Opc = OpcTable[Idx][2]; RC = &AArch64::GPR32RegClass; break;
1594 case MVT::i64: Opc = OpcTable[Idx][3]; RC = &AArch64::GPR64RegClass; break;
1595 case MVT::f32: Opc = OpcTable[Idx][4]; RC = &AArch64::FPR32RegClass; break;
1596 case MVT::f64: Opc = OpcTable[Idx][5]; RC = &AArch64::FPR64RegClass; break;
1599 // Create the base instruction, then add the operands.
1600 ResultReg = createResultReg(RC);
1601 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1602 TII.get(Opc), ResultReg);
1603 addLoadStoreOperands(Addr, MIB, MachineMemOperand::MOLoad, ScaleFactor, MMO);
1605 // Loading an i1 requires special handling.
1607 unsigned ANDReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, 1);
1608 assert(ANDReg && "Unexpected AND instruction emission failure.");
1614 bool AArch64FastISel::selectAddSub(const Instruction *I) {
1616 if (!isTypeSupported(I->getType(), VT, /*IsVectorAllowed=*/true))
1620 return selectOperator(I, I->getOpcode());
1623 switch (I->getOpcode()) {
1625 llvm_unreachable("Unexpected instruction.");
1626 case Instruction::Add:
1627 ResultReg = emitAdd(VT, I->getOperand(0), I->getOperand(1));
1629 case Instruction::Sub:
1630 ResultReg = emitSub(VT, I->getOperand(0), I->getOperand(1));
1636 updateValueMap(I, ResultReg);
1640 bool AArch64FastISel::selectLogicalOp(const Instruction *I) {
1642 if (!isTypeSupported(I->getType(), VT, /*IsVectorAllowed=*/true))
1646 return selectOperator(I, I->getOpcode());
1649 switch (I->getOpcode()) {
1651 llvm_unreachable("Unexpected instruction.");
1652 case Instruction::And:
1653 ResultReg = emitLogicalOp(ISD::AND, VT, I->getOperand(0), I->getOperand(1));
1655 case Instruction::Or:
1656 ResultReg = emitLogicalOp(ISD::OR, VT, I->getOperand(0), I->getOperand(1));
1658 case Instruction::Xor:
1659 ResultReg = emitLogicalOp(ISD::XOR, VT, I->getOperand(0), I->getOperand(1));
1665 updateValueMap(I, ResultReg);
1669 bool AArch64FastISel::selectLoad(const Instruction *I) {
1671 // Verify we have a legal type before going any further. Currently, we handle
1672 // simple types that will directly fit in a register (i32/f32/i64/f64) or
1673 // those that can be sign or zero-extended to a basic operation (i1/i8/i16).
1674 if (!isTypeSupported(I->getType(), VT, /*IsVectorAllowed=*/true) ||
1675 cast<LoadInst>(I)->isAtomic())
1678 // See if we can handle this address.
1680 if (!computeAddress(I->getOperand(0), Addr, I->getType()))
1684 if (!emitLoad(VT, ResultReg, Addr, createMachineMemOperandFor(I)))
1687 updateValueMap(I, ResultReg);
1691 bool AArch64FastISel::emitStore(MVT VT, unsigned SrcReg, Address Addr,
1692 MachineMemOperand *MMO) {
1693 // Simplify this down to something we can handle.
1694 if (!simplifyAddress(Addr, VT))
1697 unsigned ScaleFactor;
1698 switch (VT.SimpleTy) {
1699 default: llvm_unreachable("Unexpected value type.");
1700 case MVT::i1: // fall-through
1701 case MVT::i8: ScaleFactor = 1; break;
1702 case MVT::i16: ScaleFactor = 2; break;
1703 case MVT::i32: // fall-through
1704 case MVT::f32: ScaleFactor = 4; break;
1705 case MVT::i64: // fall-through
1706 case MVT::f64: ScaleFactor = 8; break;
1709 // Negative offsets require unscaled, 9-bit, signed immediate offsets.
1710 // Otherwise, we try using scaled, 12-bit, unsigned immediate offsets.
1711 bool UseScaled = true;
1712 if ((Addr.getOffset() < 0) || (Addr.getOffset() & (ScaleFactor - 1))) {
1718 static const unsigned OpcTable[4][6] = {
1719 { AArch64::STURBBi, AArch64::STURHHi, AArch64::STURWi, AArch64::STURXi,
1720 AArch64::STURSi, AArch64::STURDi },
1721 { AArch64::STRBBui, AArch64::STRHHui, AArch64::STRWui, AArch64::STRXui,
1722 AArch64::STRSui, AArch64::STRDui },
1723 { AArch64::STRBBroX, AArch64::STRHHroX, AArch64::STRWroX, AArch64::STRXroX,
1724 AArch64::STRSroX, AArch64::STRDroX },
1725 { AArch64::STRBBroW, AArch64::STRHHroW, AArch64::STRWroW, AArch64::STRXroW,
1726 AArch64::STRSroW, AArch64::STRDroW }
1731 bool VTIsi1 = false;
1732 bool UseRegOffset = Addr.isRegBase() && !Addr.getOffset() && Addr.getReg() &&
1733 Addr.getOffsetReg();
1734 unsigned Idx = UseRegOffset ? 2 : UseScaled ? 1 : 0;
1735 if (Addr.getExtendType() == AArch64_AM::UXTW ||
1736 Addr.getExtendType() == AArch64_AM::SXTW)
1739 switch (VT.SimpleTy) {
1740 default: llvm_unreachable("Unexpected value type.");
1741 case MVT::i1: VTIsi1 = true;
1742 case MVT::i8: Opc = OpcTable[Idx][0]; break;
1743 case MVT::i16: Opc = OpcTable[Idx][1]; break;
1744 case MVT::i32: Opc = OpcTable[Idx][2]; break;
1745 case MVT::i64: Opc = OpcTable[Idx][3]; break;
1746 case MVT::f32: Opc = OpcTable[Idx][4]; break;
1747 case MVT::f64: Opc = OpcTable[Idx][5]; break;
1750 // Storing an i1 requires special handling.
1751 if (VTIsi1 && SrcReg != AArch64::WZR) {
1752 unsigned ANDReg = emitAnd_ri(MVT::i32, SrcReg, /*TODO:IsKill=*/false, 1);
1753 assert(ANDReg && "Unexpected AND instruction emission failure.");
1756 // Create the base instruction, then add the operands.
1757 const MCInstrDesc &II = TII.get(Opc);
1758 SrcReg = constrainOperandRegClass(II, SrcReg, II.getNumDefs());
1759 MachineInstrBuilder MIB =
1760 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II).addReg(SrcReg);
1761 addLoadStoreOperands(Addr, MIB, MachineMemOperand::MOStore, ScaleFactor, MMO);
1766 bool AArch64FastISel::selectStore(const Instruction *I) {
1768 const Value *Op0 = I->getOperand(0);
1769 // Verify we have a legal type before going any further. Currently, we handle
1770 // simple types that will directly fit in a register (i32/f32/i64/f64) or
1771 // those that can be sign or zero-extended to a basic operation (i1/i8/i16).
1772 if (!isTypeSupported(Op0->getType(), VT, /*IsVectorAllowed=*/true) ||
1773 cast<StoreInst>(I)->isAtomic())
1776 // Get the value to be stored into a register. Use the zero register directly
1777 // when possible to avoid an unnecessary copy and a wasted register.
1778 unsigned SrcReg = 0;
1779 if (const auto *CI = dyn_cast<ConstantInt>(Op0)) {
1781 SrcReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
1782 } else if (const auto *CF = dyn_cast<ConstantFP>(Op0)) {
1783 if (CF->isZero() && !CF->isNegative()) {
1784 VT = MVT::getIntegerVT(VT.getSizeInBits());
1785 SrcReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
1790 SrcReg = getRegForValue(Op0);
1795 // See if we can handle this address.
1797 if (!computeAddress(I->getOperand(1), Addr, I->getOperand(0)->getType()))
1800 if (!emitStore(VT, SrcReg, Addr, createMachineMemOperandFor(I)))
1805 static AArch64CC::CondCode getCompareCC(CmpInst::Predicate Pred) {
1807 case CmpInst::FCMP_ONE:
1808 case CmpInst::FCMP_UEQ:
1810 // AL is our "false" for now. The other two need more compares.
1811 return AArch64CC::AL;
1812 case CmpInst::ICMP_EQ:
1813 case CmpInst::FCMP_OEQ:
1814 return AArch64CC::EQ;
1815 case CmpInst::ICMP_SGT:
1816 case CmpInst::FCMP_OGT:
1817 return AArch64CC::GT;
1818 case CmpInst::ICMP_SGE:
1819 case CmpInst::FCMP_OGE:
1820 return AArch64CC::GE;
1821 case CmpInst::ICMP_UGT:
1822 case CmpInst::FCMP_UGT:
1823 return AArch64CC::HI;
1824 case CmpInst::FCMP_OLT:
1825 return AArch64CC::MI;
1826 case CmpInst::ICMP_ULE:
1827 case CmpInst::FCMP_OLE:
1828 return AArch64CC::LS;
1829 case CmpInst::FCMP_ORD:
1830 return AArch64CC::VC;
1831 case CmpInst::FCMP_UNO:
1832 return AArch64CC::VS;
1833 case CmpInst::FCMP_UGE:
1834 return AArch64CC::PL;
1835 case CmpInst::ICMP_SLT:
1836 case CmpInst::FCMP_ULT:
1837 return AArch64CC::LT;
1838 case CmpInst::ICMP_SLE:
1839 case CmpInst::FCMP_ULE:
1840 return AArch64CC::LE;
1841 case CmpInst::FCMP_UNE:
1842 case CmpInst::ICMP_NE:
1843 return AArch64CC::NE;
1844 case CmpInst::ICMP_UGE:
1845 return AArch64CC::HS;
1846 case CmpInst::ICMP_ULT:
1847 return AArch64CC::LO;
1851 /// \brief Check if the comparison against zero and the following branch can be
1852 /// folded into a single instruction (CBZ or CBNZ).
1853 static bool canFoldZeroCheckIntoBranch(const CmpInst *CI) {
1854 CmpInst::Predicate Predicate = CI->getPredicate();
1855 if ((Predicate != CmpInst::ICMP_EQ) && (Predicate != CmpInst::ICMP_NE))
1858 Type *Ty = CI->getOperand(0)->getType();
1859 if (!Ty->isIntegerTy())
1862 unsigned BW = cast<IntegerType>(Ty)->getBitWidth();
1863 if (BW != 1 && BW != 8 && BW != 16 && BW != 32 && BW != 64)
1866 if (const auto *C = dyn_cast<ConstantInt>(CI->getOperand(0)))
1867 if (C->isNullValue())
1870 if (const auto *C = dyn_cast<ConstantInt>(CI->getOperand(1)))
1871 if (C->isNullValue())
1877 bool AArch64FastISel::selectBranch(const Instruction *I) {
1878 const BranchInst *BI = cast<BranchInst>(I);
1879 if (BI->isUnconditional()) {
1880 MachineBasicBlock *MSucc = FuncInfo.MBBMap[BI->getSuccessor(0)];
1881 fastEmitBranch(MSucc, BI->getDebugLoc());
1885 MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
1886 MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
1888 AArch64CC::CondCode CC = AArch64CC::NE;
1889 if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
1890 if (CI->hasOneUse() && isValueAvailable(CI)) {
1891 // Try to optimize or fold the cmp.
1892 CmpInst::Predicate Predicate = optimizeCmpPredicate(CI);
1893 switch (Predicate) {
1896 case CmpInst::FCMP_FALSE:
1897 fastEmitBranch(FBB, DbgLoc);
1899 case CmpInst::FCMP_TRUE:
1900 fastEmitBranch(TBB, DbgLoc);
1904 // Try to take advantage of fallthrough opportunities.
1905 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
1906 std::swap(TBB, FBB);
1907 Predicate = CmpInst::getInversePredicate(Predicate);
1910 // Try to optimize comparisons against zero.
1911 if (canFoldZeroCheckIntoBranch(CI)) {
1912 const Value *LHS = CI->getOperand(0);
1913 const Value *RHS = CI->getOperand(1);
1915 // Canonicalize zero values to the RHS.
1916 if (const auto *C = dyn_cast<ConstantInt>(LHS))
1917 if (C->isNullValue())
1918 std::swap(LHS, RHS);
1921 if (const auto *AI = dyn_cast<BinaryOperator>(LHS))
1922 if (AI->getOpcode() == Instruction::And) {
1923 const Value *AndLHS = AI->getOperand(0);
1924 const Value *AndRHS = AI->getOperand(1);
1926 if (const auto *C = dyn_cast<ConstantInt>(AndLHS))
1927 if (C->getValue().isPowerOf2())
1928 std::swap(AndLHS, AndRHS);
1930 if (const auto *C = dyn_cast<ConstantInt>(AndRHS))
1931 if (C->getValue().isPowerOf2()) {
1932 TestBit = C->getValue().logBase2();
1937 static const unsigned OpcTable[2][2][2] = {
1938 { {AArch64::CBZW, AArch64::CBZX },
1939 {AArch64::CBNZW, AArch64::CBNZX} },
1940 { {AArch64::TBZW, AArch64::TBZX },
1941 {AArch64::TBNZW, AArch64::TBNZX} }
1943 bool IsBitTest = TestBit != -1;
1944 bool IsCmpNE = Predicate == CmpInst::ICMP_NE;
1945 bool Is64Bit = LHS->getType()->isIntegerTy(64);
1946 unsigned Opc = OpcTable[IsBitTest][IsCmpNE][Is64Bit];
1948 unsigned SrcReg = getRegForValue(LHS);
1951 bool SrcIsKill = hasTrivialKill(LHS);
1953 // Emit the combined compare and branch instruction.
1954 MachineInstrBuilder MIB =
1955 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc))
1956 .addReg(SrcReg, getKillRegState(SrcIsKill));
1958 MIB.addImm(TestBit);
1961 // Obtain the branch weight and add the TrueBB to the successor list.
1962 uint32_t BranchWeight = 0;
1964 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
1965 TBB->getBasicBlock());
1966 FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
1968 fastEmitBranch(FBB, DbgLoc);
1973 if (!emitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
1976 // FCMP_UEQ and FCMP_ONE cannot be checked with a single branch
1978 CC = getCompareCC(Predicate);
1979 AArch64CC::CondCode ExtraCC = AArch64CC::AL;
1980 switch (Predicate) {
1983 case CmpInst::FCMP_UEQ:
1984 ExtraCC = AArch64CC::EQ;
1987 case CmpInst::FCMP_ONE:
1988 ExtraCC = AArch64CC::MI;
1992 assert((CC != AArch64CC::AL) && "Unexpected condition code.");
1994 // Emit the extra branch for FCMP_UEQ and FCMP_ONE.
1995 if (ExtraCC != AArch64CC::AL) {
1996 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
2002 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
2006 // Obtain the branch weight and add the TrueBB to the successor list.
2007 uint32_t BranchWeight = 0;
2009 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
2010 TBB->getBasicBlock());
2011 FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
2013 fastEmitBranch(FBB, DbgLoc);
2016 } else if (TruncInst *TI = dyn_cast<TruncInst>(BI->getCondition())) {
2018 if (TI->hasOneUse() && isValueAvailable(TI) &&
2019 isTypeSupported(TI->getOperand(0)->getType(), SrcVT)) {
2020 unsigned CondReg = getRegForValue(TI->getOperand(0));
2023 bool CondIsKill = hasTrivialKill(TI->getOperand(0));
2025 // Issue an extract_subreg to get the lower 32-bits.
2026 if (SrcVT == MVT::i64) {
2027 CondReg = fastEmitInst_extractsubreg(MVT::i32, CondReg, CondIsKill,
2032 unsigned ANDReg = emitAnd_ri(MVT::i32, CondReg, CondIsKill, 1);
2033 assert(ANDReg && "Unexpected AND instruction emission failure.");
2034 emitICmp_ri(MVT::i32, ANDReg, /*IsKill=*/true, 0);
2036 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
2037 std::swap(TBB, FBB);
2040 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
2044 // Obtain the branch weight and add the TrueBB to the successor list.
2045 uint32_t BranchWeight = 0;
2047 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
2048 TBB->getBasicBlock());
2049 FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
2051 fastEmitBranch(FBB, DbgLoc);
2054 } else if (const auto *CI = dyn_cast<ConstantInt>(BI->getCondition())) {
2055 uint64_t Imm = CI->getZExtValue();
2056 MachineBasicBlock *Target = (Imm == 0) ? FBB : TBB;
2057 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::B))
2060 // Obtain the branch weight and add the target to the successor list.
2061 uint32_t BranchWeight = 0;
2063 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
2064 Target->getBasicBlock());
2065 FuncInfo.MBB->addSuccessor(Target, BranchWeight);
2067 } else if (foldXALUIntrinsic(CC, I, BI->getCondition())) {
2068 // Fake request the condition, otherwise the intrinsic might be completely
2070 unsigned CondReg = getRegForValue(BI->getCondition());
2075 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
2079 // Obtain the branch weight and add the TrueBB to the successor list.
2080 uint32_t BranchWeight = 0;
2082 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
2083 TBB->getBasicBlock());
2084 FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
2086 fastEmitBranch(FBB, DbgLoc);
2090 unsigned CondReg = getRegForValue(BI->getCondition());
2093 bool CondRegIsKill = hasTrivialKill(BI->getCondition());
2095 // We've been divorced from our compare! Our block was split, and
2096 // now our compare lives in a predecessor block. We musn't
2097 // re-compare here, as the children of the compare aren't guaranteed
2098 // live across the block boundary (we *could* check for this).
2099 // Regardless, the compare has been done in the predecessor block,
2100 // and it left a value for us in a virtual register. Ergo, we test
2101 // the one-bit value left in the virtual register.
2102 emitICmp_ri(MVT::i32, CondReg, CondRegIsKill, 0);
2104 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
2105 std::swap(TBB, FBB);
2109 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
2113 // Obtain the branch weight and add the TrueBB to the successor list.
2114 uint32_t BranchWeight = 0;
2116 BranchWeight = FuncInfo.BPI->getEdgeWeight(BI->getParent(),
2117 TBB->getBasicBlock());
2118 FuncInfo.MBB->addSuccessor(TBB, BranchWeight);
2120 fastEmitBranch(FBB, DbgLoc);
2124 bool AArch64FastISel::selectIndirectBr(const Instruction *I) {
2125 const IndirectBrInst *BI = cast<IndirectBrInst>(I);
2126 unsigned AddrReg = getRegForValue(BI->getOperand(0));
2130 // Emit the indirect branch.
2131 const MCInstrDesc &II = TII.get(AArch64::BR);
2132 AddrReg = constrainOperandRegClass(II, AddrReg, II.getNumDefs());
2133 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II).addReg(AddrReg);
2135 // Make sure the CFG is up-to-date.
2136 for (unsigned i = 0, e = BI->getNumSuccessors(); i != e; ++i)
2137 FuncInfo.MBB->addSuccessor(FuncInfo.MBBMap[BI->getSuccessor(i)]);
2142 bool AArch64FastISel::selectCmp(const Instruction *I) {
2143 const CmpInst *CI = cast<CmpInst>(I);
2145 // Try to optimize or fold the cmp.
2146 CmpInst::Predicate Predicate = optimizeCmpPredicate(CI);
2147 unsigned ResultReg = 0;
2148 switch (Predicate) {
2151 case CmpInst::FCMP_FALSE:
2152 ResultReg = createResultReg(&AArch64::GPR32RegClass);
2153 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2154 TII.get(TargetOpcode::COPY), ResultReg)
2155 .addReg(AArch64::WZR, getKillRegState(true));
2157 case CmpInst::FCMP_TRUE:
2158 ResultReg = fastEmit_i(MVT::i32, MVT::i32, ISD::Constant, 1);
2163 updateValueMap(I, ResultReg);
2168 if (!emitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
2171 ResultReg = createResultReg(&AArch64::GPR32RegClass);
2173 // FCMP_UEQ and FCMP_ONE cannot be checked with a single instruction. These
2174 // condition codes are inverted, because they are used by CSINC.
2175 static unsigned CondCodeTable[2][2] = {
2176 { AArch64CC::NE, AArch64CC::VC },
2177 { AArch64CC::PL, AArch64CC::LE }
2179 unsigned *CondCodes = nullptr;
2180 switch (Predicate) {
2183 case CmpInst::FCMP_UEQ:
2184 CondCodes = &CondCodeTable[0][0];
2186 case CmpInst::FCMP_ONE:
2187 CondCodes = &CondCodeTable[1][0];
2192 unsigned TmpReg1 = createResultReg(&AArch64::GPR32RegClass);
2193 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::CSINCWr),
2195 .addReg(AArch64::WZR, getKillRegState(true))
2196 .addReg(AArch64::WZR, getKillRegState(true))
2197 .addImm(CondCodes[0]);
2198 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::CSINCWr),
2200 .addReg(TmpReg1, getKillRegState(true))
2201 .addReg(AArch64::WZR, getKillRegState(true))
2202 .addImm(CondCodes[1]);
2204 updateValueMap(I, ResultReg);
2208 // Now set a register based on the comparison.
2209 AArch64CC::CondCode CC = getCompareCC(Predicate);
2210 assert((CC != AArch64CC::AL) && "Unexpected condition code.");
2211 AArch64CC::CondCode invertedCC = getInvertedCondCode(CC);
2212 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::CSINCWr),
2214 .addReg(AArch64::WZR, getKillRegState(true))
2215 .addReg(AArch64::WZR, getKillRegState(true))
2216 .addImm(invertedCC);
2218 updateValueMap(I, ResultReg);
2222 bool AArch64FastISel::selectSelect(const Instruction *I) {
2223 const SelectInst *SI = cast<SelectInst>(I);
2225 EVT DestEVT = TLI.getValueType(SI->getType(), true);
2226 if (!DestEVT.isSimple())
2229 MVT DestVT = DestEVT.getSimpleVT();
2230 if (DestVT != MVT::i32 && DestVT != MVT::i64 && DestVT != MVT::f32 &&
2235 const TargetRegisterClass *RC = nullptr;
2236 switch (DestVT.SimpleTy) {
2237 default: return false;
2239 SelectOpc = AArch64::CSELWr; RC = &AArch64::GPR32RegClass; break;
2241 SelectOpc = AArch64::CSELXr; RC = &AArch64::GPR64RegClass; break;
2243 SelectOpc = AArch64::FCSELSrrr; RC = &AArch64::FPR32RegClass; break;
2245 SelectOpc = AArch64::FCSELDrrr; RC = &AArch64::FPR64RegClass; break;
2248 const Value *Cond = SI->getCondition();
2249 bool NeedTest = true;
2250 AArch64CC::CondCode CC = AArch64CC::NE;
2251 if (foldXALUIntrinsic(CC, I, Cond))
2254 unsigned CondReg = getRegForValue(Cond);
2257 bool CondIsKill = hasTrivialKill(Cond);
2260 unsigned ANDReg = emitAnd_ri(MVT::i32, CondReg, CondIsKill, 1);
2261 assert(ANDReg && "Unexpected AND instruction emission failure.");
2262 emitICmp_ri(MVT::i32, ANDReg, /*IsKill=*/true, 0);
2265 unsigned TrueReg = getRegForValue(SI->getTrueValue());
2266 bool TrueIsKill = hasTrivialKill(SI->getTrueValue());
2268 unsigned FalseReg = getRegForValue(SI->getFalseValue());
2269 bool FalseIsKill = hasTrivialKill(SI->getFalseValue());
2271 if (!TrueReg || !FalseReg)
2274 unsigned ResultReg = fastEmitInst_rri(SelectOpc, RC, TrueReg, TrueIsKill,
2275 FalseReg, FalseIsKill, CC);
2276 updateValueMap(I, ResultReg);
2280 bool AArch64FastISel::selectFPExt(const Instruction *I) {
2281 Value *V = I->getOperand(0);
2282 if (!I->getType()->isDoubleTy() || !V->getType()->isFloatTy())
2285 unsigned Op = getRegForValue(V);
2289 unsigned ResultReg = createResultReg(&AArch64::FPR64RegClass);
2290 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::FCVTDSr),
2291 ResultReg).addReg(Op);
2292 updateValueMap(I, ResultReg);
2296 bool AArch64FastISel::selectFPTrunc(const Instruction *I) {
2297 Value *V = I->getOperand(0);
2298 if (!I->getType()->isFloatTy() || !V->getType()->isDoubleTy())
2301 unsigned Op = getRegForValue(V);
2305 unsigned ResultReg = createResultReg(&AArch64::FPR32RegClass);
2306 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::FCVTSDr),
2307 ResultReg).addReg(Op);
2308 updateValueMap(I, ResultReg);
2312 // FPToUI and FPToSI
2313 bool AArch64FastISel::selectFPToInt(const Instruction *I, bool Signed) {
2315 if (!isTypeLegal(I->getType(), DestVT) || DestVT.isVector())
2318 unsigned SrcReg = getRegForValue(I->getOperand(0));
2322 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType(), true);
2323 if (SrcVT == MVT::f128)
2327 if (SrcVT == MVT::f64) {
2329 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZSUWDr : AArch64::FCVTZSUXDr;
2331 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZUUWDr : AArch64::FCVTZUUXDr;
2334 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZSUWSr : AArch64::FCVTZSUXSr;
2336 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZUUWSr : AArch64::FCVTZUUXSr;
2338 unsigned ResultReg = createResultReg(
2339 DestVT == MVT::i32 ? &AArch64::GPR32RegClass : &AArch64::GPR64RegClass);
2340 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
2342 updateValueMap(I, ResultReg);
2346 bool AArch64FastISel::selectIntToFP(const Instruction *I, bool Signed) {
2348 if (!isTypeLegal(I->getType(), DestVT) || DestVT.isVector())
2350 assert ((DestVT == MVT::f32 || DestVT == MVT::f64) &&
2351 "Unexpected value type.");
2353 unsigned SrcReg = getRegForValue(I->getOperand(0));
2356 bool SrcIsKill = hasTrivialKill(I->getOperand(0));
2358 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType(), true);
2360 // Handle sign-extension.
2361 if (SrcVT == MVT::i16 || SrcVT == MVT::i8 || SrcVT == MVT::i1) {
2363 emitIntExt(SrcVT.getSimpleVT(), SrcReg, MVT::i32, /*isZExt*/ !Signed);
2370 if (SrcVT == MVT::i64) {
2372 Opc = (DestVT == MVT::f32) ? AArch64::SCVTFUXSri : AArch64::SCVTFUXDri;
2374 Opc = (DestVT == MVT::f32) ? AArch64::UCVTFUXSri : AArch64::UCVTFUXDri;
2377 Opc = (DestVT == MVT::f32) ? AArch64::SCVTFUWSri : AArch64::SCVTFUWDri;
2379 Opc = (DestVT == MVT::f32) ? AArch64::UCVTFUWSri : AArch64::UCVTFUWDri;
2382 unsigned ResultReg = fastEmitInst_r(Opc, TLI.getRegClassFor(DestVT), SrcReg,
2384 updateValueMap(I, ResultReg);
2388 bool AArch64FastISel::fastLowerArguments() {
2389 if (!FuncInfo.CanLowerReturn)
2392 const Function *F = FuncInfo.Fn;
2396 CallingConv::ID CC = F->getCallingConv();
2397 if (CC != CallingConv::C)
2400 // Only handle simple cases of up to 8 GPR and FPR each.
2401 unsigned GPRCnt = 0;
2402 unsigned FPRCnt = 0;
2404 for (auto const &Arg : F->args()) {
2405 // The first argument is at index 1.
2407 if (F->getAttributes().hasAttribute(Idx, Attribute::ByVal) ||
2408 F->getAttributes().hasAttribute(Idx, Attribute::InReg) ||
2409 F->getAttributes().hasAttribute(Idx, Attribute::StructRet) ||
2410 F->getAttributes().hasAttribute(Idx, Attribute::Nest))
2413 Type *ArgTy = Arg.getType();
2414 if (ArgTy->isStructTy() || ArgTy->isArrayTy())
2417 EVT ArgVT = TLI.getValueType(ArgTy);
2418 if (!ArgVT.isSimple())
2421 MVT VT = ArgVT.getSimpleVT().SimpleTy;
2422 if (VT.isFloatingPoint() && !Subtarget->hasFPARMv8())
2425 if (VT.isVector() &&
2426 (!Subtarget->hasNEON() || !Subtarget->isLittleEndian()))
2429 if (VT >= MVT::i1 && VT <= MVT::i64)
2431 else if ((VT >= MVT::f16 && VT <= MVT::f64) || VT.is64BitVector() ||
2432 VT.is128BitVector())
2437 if (GPRCnt > 8 || FPRCnt > 8)
2441 static const MCPhysReg Registers[6][8] = {
2442 { AArch64::W0, AArch64::W1, AArch64::W2, AArch64::W3, AArch64::W4,
2443 AArch64::W5, AArch64::W6, AArch64::W7 },
2444 { AArch64::X0, AArch64::X1, AArch64::X2, AArch64::X3, AArch64::X4,
2445 AArch64::X5, AArch64::X6, AArch64::X7 },
2446 { AArch64::H0, AArch64::H1, AArch64::H2, AArch64::H3, AArch64::H4,
2447 AArch64::H5, AArch64::H6, AArch64::H7 },
2448 { AArch64::S0, AArch64::S1, AArch64::S2, AArch64::S3, AArch64::S4,
2449 AArch64::S5, AArch64::S6, AArch64::S7 },
2450 { AArch64::D0, AArch64::D1, AArch64::D2, AArch64::D3, AArch64::D4,
2451 AArch64::D5, AArch64::D6, AArch64::D7 },
2452 { AArch64::Q0, AArch64::Q1, AArch64::Q2, AArch64::Q3, AArch64::Q4,
2453 AArch64::Q5, AArch64::Q6, AArch64::Q7 }
2456 unsigned GPRIdx = 0;
2457 unsigned FPRIdx = 0;
2458 for (auto const &Arg : F->args()) {
2459 MVT VT = TLI.getSimpleValueType(Arg.getType());
2461 const TargetRegisterClass *RC;
2462 if (VT >= MVT::i1 && VT <= MVT::i32) {
2463 SrcReg = Registers[0][GPRIdx++];
2464 RC = &AArch64::GPR32RegClass;
2466 } else if (VT == MVT::i64) {
2467 SrcReg = Registers[1][GPRIdx++];
2468 RC = &AArch64::GPR64RegClass;
2469 } else if (VT == MVT::f16) {
2470 SrcReg = Registers[2][FPRIdx++];
2471 RC = &AArch64::FPR16RegClass;
2472 } else if (VT == MVT::f32) {
2473 SrcReg = Registers[3][FPRIdx++];
2474 RC = &AArch64::FPR32RegClass;
2475 } else if ((VT == MVT::f64) || VT.is64BitVector()) {
2476 SrcReg = Registers[4][FPRIdx++];
2477 RC = &AArch64::FPR64RegClass;
2478 } else if (VT.is128BitVector()) {
2479 SrcReg = Registers[5][FPRIdx++];
2480 RC = &AArch64::FPR128RegClass;
2482 llvm_unreachable("Unexpected value type.");
2484 unsigned DstReg = FuncInfo.MF->addLiveIn(SrcReg, RC);
2485 // FIXME: Unfortunately it's necessary to emit a copy from the livein copy.
2486 // Without this, EmitLiveInCopies may eliminate the livein if its only
2487 // use is a bitcast (which isn't turned into an instruction).
2488 unsigned ResultReg = createResultReg(RC);
2489 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2490 TII.get(TargetOpcode::COPY), ResultReg)
2491 .addReg(DstReg, getKillRegState(true));
2492 updateValueMap(&Arg, ResultReg);
2497 bool AArch64FastISel::processCallArgs(CallLoweringInfo &CLI,
2498 SmallVectorImpl<MVT> &OutVTs,
2499 unsigned &NumBytes) {
2500 CallingConv::ID CC = CLI.CallConv;
2501 SmallVector<CCValAssign, 16> ArgLocs;
2502 CCState CCInfo(CC, false, *FuncInfo.MF, ArgLocs, *Context);
2503 CCInfo.AnalyzeCallOperands(OutVTs, CLI.OutFlags, CCAssignFnForCall(CC));
2505 // Get a count of how many bytes are to be pushed on the stack.
2506 NumBytes = CCInfo.getNextStackOffset();
2508 // Issue CALLSEQ_START
2509 unsigned AdjStackDown = TII.getCallFrameSetupOpcode();
2510 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackDown))
2513 // Process the args.
2514 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2515 CCValAssign &VA = ArgLocs[i];
2516 const Value *ArgVal = CLI.OutVals[VA.getValNo()];
2517 MVT ArgVT = OutVTs[VA.getValNo()];
2519 unsigned ArgReg = getRegForValue(ArgVal);
2523 // Handle arg promotion: SExt, ZExt, AExt.
2524 switch (VA.getLocInfo()) {
2525 case CCValAssign::Full:
2527 case CCValAssign::SExt: {
2528 MVT DestVT = VA.getLocVT();
2530 ArgReg = emitIntExt(SrcVT, ArgReg, DestVT, /*isZExt=*/false);
2535 case CCValAssign::AExt:
2536 // Intentional fall-through.
2537 case CCValAssign::ZExt: {
2538 MVT DestVT = VA.getLocVT();
2540 ArgReg = emitIntExt(SrcVT, ArgReg, DestVT, /*isZExt=*/true);
2546 llvm_unreachable("Unknown arg promotion!");
2549 // Now copy/store arg to correct locations.
2550 if (VA.isRegLoc() && !VA.needsCustom()) {
2551 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2552 TII.get(TargetOpcode::COPY), VA.getLocReg()).addReg(ArgReg);
2553 CLI.OutRegs.push_back(VA.getLocReg());
2554 } else if (VA.needsCustom()) {
2555 // FIXME: Handle custom args.
2558 assert(VA.isMemLoc() && "Assuming store on stack.");
2560 // Don't emit stores for undef values.
2561 if (isa<UndefValue>(ArgVal))
2564 // Need to store on the stack.
2565 unsigned ArgSize = (ArgVT.getSizeInBits() + 7) / 8;
2567 unsigned BEAlign = 0;
2568 if (ArgSize < 8 && !Subtarget->isLittleEndian())
2569 BEAlign = 8 - ArgSize;
2572 Addr.setKind(Address::RegBase);
2573 Addr.setReg(AArch64::SP);
2574 Addr.setOffset(VA.getLocMemOffset() + BEAlign);
2576 unsigned Alignment = DL.getABITypeAlignment(ArgVal->getType());
2577 MachineMemOperand *MMO = FuncInfo.MF->getMachineMemOperand(
2578 MachinePointerInfo::getStack(Addr.getOffset()),
2579 MachineMemOperand::MOStore, ArgVT.getStoreSize(), Alignment);
2581 if (!emitStore(ArgVT, ArgReg, Addr, MMO))
2588 bool AArch64FastISel::finishCall(CallLoweringInfo &CLI, MVT RetVT,
2589 unsigned NumBytes) {
2590 CallingConv::ID CC = CLI.CallConv;
2592 // Issue CALLSEQ_END
2593 unsigned AdjStackUp = TII.getCallFrameDestroyOpcode();
2594 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackUp))
2595 .addImm(NumBytes).addImm(0);
2597 // Now the return value.
2598 if (RetVT != MVT::isVoid) {
2599 SmallVector<CCValAssign, 16> RVLocs;
2600 CCState CCInfo(CC, false, *FuncInfo.MF, RVLocs, *Context);
2601 CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC));
2603 // Only handle a single return value.
2604 if (RVLocs.size() != 1)
2607 // Copy all of the result registers out of their specified physreg.
2608 MVT CopyVT = RVLocs[0].getValVT();
2609 unsigned ResultReg = createResultReg(TLI.getRegClassFor(CopyVT));
2610 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2611 TII.get(TargetOpcode::COPY), ResultReg)
2612 .addReg(RVLocs[0].getLocReg());
2613 CLI.InRegs.push_back(RVLocs[0].getLocReg());
2615 CLI.ResultReg = ResultReg;
2616 CLI.NumResultRegs = 1;
2622 bool AArch64FastISel::fastLowerCall(CallLoweringInfo &CLI) {
2623 CallingConv::ID CC = CLI.CallConv;
2624 bool IsTailCall = CLI.IsTailCall;
2625 bool IsVarArg = CLI.IsVarArg;
2626 const Value *Callee = CLI.Callee;
2627 const char *SymName = CLI.SymName;
2629 if (!Callee && !SymName)
2632 // Allow SelectionDAG isel to handle tail calls.
2636 CodeModel::Model CM = TM.getCodeModel();
2637 // Only support the small and large code model.
2638 if (CM != CodeModel::Small && CM != CodeModel::Large)
2641 // FIXME: Add large code model support for ELF.
2642 if (CM == CodeModel::Large && !Subtarget->isTargetMachO())
2645 // Let SDISel handle vararg functions.
2649 // FIXME: Only handle *simple* calls for now.
2651 if (CLI.RetTy->isVoidTy())
2652 RetVT = MVT::isVoid;
2653 else if (!isTypeLegal(CLI.RetTy, RetVT))
2656 for (auto Flag : CLI.OutFlags)
2657 if (Flag.isInReg() || Flag.isSRet() || Flag.isNest() || Flag.isByVal())
2660 // Set up the argument vectors.
2661 SmallVector<MVT, 16> OutVTs;
2662 OutVTs.reserve(CLI.OutVals.size());
2664 for (auto *Val : CLI.OutVals) {
2666 if (!isTypeLegal(Val->getType(), VT) &&
2667 !(VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16))
2670 // We don't handle vector parameters yet.
2671 if (VT.isVector() || VT.getSizeInBits() > 64)
2674 OutVTs.push_back(VT);
2678 if (Callee && !computeCallAddress(Callee, Addr))
2681 // Handle the arguments now that we've gotten them.
2683 if (!processCallArgs(CLI, OutVTs, NumBytes))
2687 MachineInstrBuilder MIB;
2688 if (CM == CodeModel::Small) {
2689 const MCInstrDesc &II = TII.get(Addr.getReg() ? AArch64::BLR : AArch64::BL);
2690 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II);
2692 MIB.addExternalSymbol(SymName, 0);
2693 else if (Addr.getGlobalValue())
2694 MIB.addGlobalAddress(Addr.getGlobalValue(), 0, 0);
2695 else if (Addr.getReg()) {
2696 unsigned Reg = constrainOperandRegClass(II, Addr.getReg(), 0);
2701 unsigned CallReg = 0;
2703 unsigned ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
2704 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
2706 .addExternalSymbol(SymName, AArch64II::MO_GOT | AArch64II::MO_PAGE);
2708 CallReg = createResultReg(&AArch64::GPR64RegClass);
2709 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::LDRXui),
2712 .addExternalSymbol(SymName, AArch64II::MO_GOT | AArch64II::MO_PAGEOFF |
2714 } else if (Addr.getGlobalValue())
2715 CallReg = materializeGV(Addr.getGlobalValue());
2716 else if (Addr.getReg())
2717 CallReg = Addr.getReg();
2722 const MCInstrDesc &II = TII.get(AArch64::BLR);
2723 CallReg = constrainOperandRegClass(II, CallReg, 0);
2724 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II).addReg(CallReg);
2727 // Add implicit physical register uses to the call.
2728 for (auto Reg : CLI.OutRegs)
2729 MIB.addReg(Reg, RegState::Implicit);
2731 // Add a register mask with the call-preserved registers.
2732 // Proper defs for return values will be added by setPhysRegsDeadExcept().
2733 MIB.addRegMask(TRI.getCallPreservedMask(CC));
2737 // Finish off the call including any return values.
2738 return finishCall(CLI, RetVT, NumBytes);
2741 bool AArch64FastISel::isMemCpySmall(uint64_t Len, unsigned Alignment) {
2743 return Len / Alignment <= 4;
2748 bool AArch64FastISel::tryEmitSmallMemCpy(Address Dest, Address Src,
2749 uint64_t Len, unsigned Alignment) {
2750 // Make sure we don't bloat code by inlining very large memcpy's.
2751 if (!isMemCpySmall(Len, Alignment))
2754 int64_t UnscaledOffset = 0;
2755 Address OrigDest = Dest;
2756 Address OrigSrc = Src;
2760 if (!Alignment || Alignment >= 8) {
2771 // Bound based on alignment.
2772 if (Len >= 4 && Alignment == 4)
2774 else if (Len >= 2 && Alignment == 2)
2783 RV = emitLoad(VT, ResultReg, Src);
2787 RV = emitStore(VT, ResultReg, Dest);
2791 int64_t Size = VT.getSizeInBits() / 8;
2793 UnscaledOffset += Size;
2795 // We need to recompute the unscaled offset for each iteration.
2796 Dest.setOffset(OrigDest.getOffset() + UnscaledOffset);
2797 Src.setOffset(OrigSrc.getOffset() + UnscaledOffset);
2803 /// \brief Check if it is possible to fold the condition from the XALU intrinsic
2804 /// into the user. The condition code will only be updated on success.
2805 bool AArch64FastISel::foldXALUIntrinsic(AArch64CC::CondCode &CC,
2806 const Instruction *I,
2807 const Value *Cond) {
2808 if (!isa<ExtractValueInst>(Cond))
2811 const auto *EV = cast<ExtractValueInst>(Cond);
2812 if (!isa<IntrinsicInst>(EV->getAggregateOperand()))
2815 const auto *II = cast<IntrinsicInst>(EV->getAggregateOperand());
2817 const Function *Callee = II->getCalledFunction();
2819 cast<StructType>(Callee->getReturnType())->getTypeAtIndex(0U);
2820 if (!isTypeLegal(RetTy, RetVT))
2823 if (RetVT != MVT::i32 && RetVT != MVT::i64)
2826 AArch64CC::CondCode TmpCC;
2827 switch (II->getIntrinsicID()) {
2828 default: return false;
2829 case Intrinsic::sadd_with_overflow:
2830 case Intrinsic::ssub_with_overflow: TmpCC = AArch64CC::VS; break;
2831 case Intrinsic::uadd_with_overflow: TmpCC = AArch64CC::HS; break;
2832 case Intrinsic::usub_with_overflow: TmpCC = AArch64CC::LO; break;
2833 case Intrinsic::smul_with_overflow:
2834 case Intrinsic::umul_with_overflow: TmpCC = AArch64CC::NE; break;
2837 // Check if both instructions are in the same basic block.
2838 if (!isValueAvailable(II))
2841 // Make sure nothing is in the way
2842 BasicBlock::const_iterator Start = I;
2843 BasicBlock::const_iterator End = II;
2844 for (auto Itr = std::prev(Start); Itr != End; --Itr) {
2845 // We only expect extractvalue instructions between the intrinsic and the
2846 // instruction to be selected.
2847 if (!isa<ExtractValueInst>(Itr))
2850 // Check that the extractvalue operand comes from the intrinsic.
2851 const auto *EVI = cast<ExtractValueInst>(Itr);
2852 if (EVI->getAggregateOperand() != II)
2860 bool AArch64FastISel::fastLowerIntrinsicCall(const IntrinsicInst *II) {
2861 // FIXME: Handle more intrinsics.
2862 switch (II->getIntrinsicID()) {
2863 default: return false;
2864 case Intrinsic::frameaddress: {
2865 MachineFrameInfo *MFI = FuncInfo.MF->getFrameInfo();
2866 MFI->setFrameAddressIsTaken(true);
2868 const AArch64RegisterInfo *RegInfo =
2869 static_cast<const AArch64RegisterInfo *>(
2870 TM.getSubtargetImpl()->getRegisterInfo());
2871 unsigned FramePtr = RegInfo->getFrameRegister(*(FuncInfo.MF));
2872 unsigned SrcReg = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
2873 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2874 TII.get(TargetOpcode::COPY), SrcReg).addReg(FramePtr);
2875 // Recursively load frame address
2881 unsigned Depth = cast<ConstantInt>(II->getOperand(0))->getZExtValue();
2883 DestReg = fastEmitInst_ri(AArch64::LDRXui, &AArch64::GPR64RegClass,
2884 SrcReg, /*IsKill=*/true, 0);
2885 assert(DestReg && "Unexpected LDR instruction emission failure.");
2889 updateValueMap(II, SrcReg);
2892 case Intrinsic::memcpy:
2893 case Intrinsic::memmove: {
2894 const auto *MTI = cast<MemTransferInst>(II);
2895 // Don't handle volatile.
2896 if (MTI->isVolatile())
2899 // Disable inlining for memmove before calls to ComputeAddress. Otherwise,
2900 // we would emit dead code because we don't currently handle memmoves.
2901 bool IsMemCpy = (II->getIntrinsicID() == Intrinsic::memcpy);
2902 if (isa<ConstantInt>(MTI->getLength()) && IsMemCpy) {
2903 // Small memcpy's are common enough that we want to do them without a call
2905 uint64_t Len = cast<ConstantInt>(MTI->getLength())->getZExtValue();
2906 unsigned Alignment = MTI->getAlignment();
2907 if (isMemCpySmall(Len, Alignment)) {
2909 if (!computeAddress(MTI->getRawDest(), Dest) ||
2910 !computeAddress(MTI->getRawSource(), Src))
2912 if (tryEmitSmallMemCpy(Dest, Src, Len, Alignment))
2917 if (!MTI->getLength()->getType()->isIntegerTy(64))
2920 if (MTI->getSourceAddressSpace() > 255 || MTI->getDestAddressSpace() > 255)
2921 // Fast instruction selection doesn't support the special
2925 const char *IntrMemName = isa<MemCpyInst>(II) ? "memcpy" : "memmove";
2926 return lowerCallTo(II, IntrMemName, II->getNumArgOperands() - 2);
2928 case Intrinsic::memset: {
2929 const MemSetInst *MSI = cast<MemSetInst>(II);
2930 // Don't handle volatile.
2931 if (MSI->isVolatile())
2934 if (!MSI->getLength()->getType()->isIntegerTy(64))
2937 if (MSI->getDestAddressSpace() > 255)
2938 // Fast instruction selection doesn't support the special
2942 return lowerCallTo(II, "memset", II->getNumArgOperands() - 2);
2944 case Intrinsic::sin:
2945 case Intrinsic::cos:
2946 case Intrinsic::pow: {
2948 if (!isTypeLegal(II->getType(), RetVT))
2951 if (RetVT != MVT::f32 && RetVT != MVT::f64)
2954 static const RTLIB::Libcall LibCallTable[3][2] = {
2955 { RTLIB::SIN_F32, RTLIB::SIN_F64 },
2956 { RTLIB::COS_F32, RTLIB::COS_F64 },
2957 { RTLIB::POW_F32, RTLIB::POW_F64 }
2960 bool Is64Bit = RetVT == MVT::f64;
2961 switch (II->getIntrinsicID()) {
2963 llvm_unreachable("Unexpected intrinsic.");
2964 case Intrinsic::sin:
2965 LC = LibCallTable[0][Is64Bit];
2967 case Intrinsic::cos:
2968 LC = LibCallTable[1][Is64Bit];
2970 case Intrinsic::pow:
2971 LC = LibCallTable[2][Is64Bit];
2976 Args.reserve(II->getNumArgOperands());
2978 // Populate the argument list.
2979 for (auto &Arg : II->arg_operands()) {
2982 Entry.Ty = Arg->getType();
2983 Args.push_back(Entry);
2986 CallLoweringInfo CLI;
2987 CLI.setCallee(TLI.getLibcallCallingConv(LC), II->getType(),
2988 TLI.getLibcallName(LC), std::move(Args));
2989 if (!lowerCallTo(CLI))
2991 updateValueMap(II, CLI.ResultReg);
2994 case Intrinsic::trap: {
2995 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::BRK))
2999 case Intrinsic::sqrt: {
3000 Type *RetTy = II->getCalledFunction()->getReturnType();
3003 if (!isTypeLegal(RetTy, VT))
3006 unsigned Op0Reg = getRegForValue(II->getOperand(0));
3009 bool Op0IsKill = hasTrivialKill(II->getOperand(0));
3011 unsigned ResultReg = fastEmit_r(VT, VT, ISD::FSQRT, Op0Reg, Op0IsKill);
3015 updateValueMap(II, ResultReg);
3018 case Intrinsic::sadd_with_overflow:
3019 case Intrinsic::uadd_with_overflow:
3020 case Intrinsic::ssub_with_overflow:
3021 case Intrinsic::usub_with_overflow:
3022 case Intrinsic::smul_with_overflow:
3023 case Intrinsic::umul_with_overflow: {
3024 // This implements the basic lowering of the xalu with overflow intrinsics.
3025 const Function *Callee = II->getCalledFunction();
3026 auto *Ty = cast<StructType>(Callee->getReturnType());
3027 Type *RetTy = Ty->getTypeAtIndex(0U);
3030 if (!isTypeLegal(RetTy, VT))
3033 if (VT != MVT::i32 && VT != MVT::i64)
3036 const Value *LHS = II->getArgOperand(0);
3037 const Value *RHS = II->getArgOperand(1);
3038 // Canonicalize immediate to the RHS.
3039 if (isa<ConstantInt>(LHS) && !isa<ConstantInt>(RHS) &&
3040 isCommutativeIntrinsic(II))
3041 std::swap(LHS, RHS);
3043 unsigned ResultReg1 = 0, ResultReg2 = 0, MulReg = 0;
3044 AArch64CC::CondCode CC = AArch64CC::Invalid;
3045 switch (II->getIntrinsicID()) {
3046 default: llvm_unreachable("Unexpected intrinsic!");
3047 case Intrinsic::sadd_with_overflow:
3048 ResultReg1 = emitAdd(VT, LHS, RHS, /*SetFlags=*/true);
3051 case Intrinsic::uadd_with_overflow:
3052 ResultReg1 = emitAdd(VT, LHS, RHS, /*SetFlags=*/true);
3055 case Intrinsic::ssub_with_overflow:
3056 ResultReg1 = emitSub(VT, LHS, RHS, /*SetFlags=*/true);
3059 case Intrinsic::usub_with_overflow:
3060 ResultReg1 = emitSub(VT, LHS, RHS, /*SetFlags=*/true);
3063 case Intrinsic::smul_with_overflow: {
3065 unsigned LHSReg = getRegForValue(LHS);
3068 bool LHSIsKill = hasTrivialKill(LHS);
3070 unsigned RHSReg = getRegForValue(RHS);
3073 bool RHSIsKill = hasTrivialKill(RHS);
3075 if (VT == MVT::i32) {
3076 MulReg = emitSMULL_rr(MVT::i64, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
3077 unsigned ShiftReg = emitLSR_ri(MVT::i64, MVT::i64, MulReg,
3078 /*IsKill=*/false, 32);
3079 MulReg = fastEmitInst_extractsubreg(VT, MulReg, /*IsKill=*/true,
3081 ShiftReg = fastEmitInst_extractsubreg(VT, ShiftReg, /*IsKill=*/true,
3083 emitSubs_rs(VT, ShiftReg, /*IsKill=*/true, MulReg, /*IsKill=*/false,
3084 AArch64_AM::ASR, 31, /*WantResult=*/false);
3086 assert(VT == MVT::i64 && "Unexpected value type.");
3087 MulReg = emitMul_rr(VT, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
3088 unsigned SMULHReg = fastEmit_rr(VT, VT, ISD::MULHS, LHSReg, LHSIsKill,
3090 emitSubs_rs(VT, SMULHReg, /*IsKill=*/true, MulReg, /*IsKill=*/false,
3091 AArch64_AM::ASR, 63, /*WantResult=*/false);
3095 case Intrinsic::umul_with_overflow: {
3097 unsigned LHSReg = getRegForValue(LHS);
3100 bool LHSIsKill = hasTrivialKill(LHS);
3102 unsigned RHSReg = getRegForValue(RHS);
3105 bool RHSIsKill = hasTrivialKill(RHS);
3107 if (VT == MVT::i32) {
3108 MulReg = emitUMULL_rr(MVT::i64, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
3109 emitSubs_rs(MVT::i64, AArch64::XZR, /*IsKill=*/true, MulReg,
3110 /*IsKill=*/false, AArch64_AM::LSR, 32,
3111 /*WantResult=*/false);
3112 MulReg = fastEmitInst_extractsubreg(VT, MulReg, /*IsKill=*/true,
3115 assert(VT == MVT::i64 && "Unexpected value type.");
3116 MulReg = emitMul_rr(VT, LHSReg, LHSIsKill, RHSReg, RHSIsKill);
3117 unsigned UMULHReg = fastEmit_rr(VT, VT, ISD::MULHU, LHSReg, LHSIsKill,
3119 emitSubs_rr(VT, AArch64::XZR, /*IsKill=*/true, UMULHReg,
3120 /*IsKill=*/false, /*WantResult=*/false);
3127 ResultReg1 = createResultReg(TLI.getRegClassFor(VT));
3128 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3129 TII.get(TargetOpcode::COPY), ResultReg1).addReg(MulReg);
3132 ResultReg2 = fastEmitInst_rri(AArch64::CSINCWr, &AArch64::GPR32RegClass,
3133 AArch64::WZR, /*IsKill=*/true, AArch64::WZR,
3134 /*IsKill=*/true, getInvertedCondCode(CC));
3135 assert((ResultReg1 + 1) == ResultReg2 &&
3136 "Nonconsecutive result registers.");
3137 updateValueMap(II, ResultReg1, 2);
3144 bool AArch64FastISel::selectRet(const Instruction *I) {
3145 const ReturnInst *Ret = cast<ReturnInst>(I);
3146 const Function &F = *I->getParent()->getParent();
3148 if (!FuncInfo.CanLowerReturn)
3154 // Build a list of return value registers.
3155 SmallVector<unsigned, 4> RetRegs;
3157 if (Ret->getNumOperands() > 0) {
3158 CallingConv::ID CC = F.getCallingConv();
3159 SmallVector<ISD::OutputArg, 4> Outs;
3160 GetReturnInfo(F.getReturnType(), F.getAttributes(), Outs, TLI);
3162 // Analyze operands of the call, assigning locations to each operand.
3163 SmallVector<CCValAssign, 16> ValLocs;
3164 CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, ValLocs, I->getContext());
3165 CCAssignFn *RetCC = CC == CallingConv::WebKit_JS ? RetCC_AArch64_WebKit_JS
3166 : RetCC_AArch64_AAPCS;
3167 CCInfo.AnalyzeReturn(Outs, RetCC);
3169 // Only handle a single return value for now.
3170 if (ValLocs.size() != 1)
3173 CCValAssign &VA = ValLocs[0];
3174 const Value *RV = Ret->getOperand(0);
3176 // Don't bother handling odd stuff for now.
3177 if ((VA.getLocInfo() != CCValAssign::Full) &&
3178 (VA.getLocInfo() != CCValAssign::BCvt))
3181 // Only handle register returns for now.
3185 unsigned Reg = getRegForValue(RV);
3189 unsigned SrcReg = Reg + VA.getValNo();
3190 unsigned DestReg = VA.getLocReg();
3191 // Avoid a cross-class copy. This is very unlikely.
3192 if (!MRI.getRegClass(SrcReg)->contains(DestReg))
3195 EVT RVEVT = TLI.getValueType(RV->getType());
3196 if (!RVEVT.isSimple())
3199 // Vectors (of > 1 lane) in big endian need tricky handling.
3200 if (RVEVT.isVector() && RVEVT.getVectorNumElements() > 1 &&
3201 !Subtarget->isLittleEndian())
3204 MVT RVVT = RVEVT.getSimpleVT();
3205 if (RVVT == MVT::f128)
3208 MVT DestVT = VA.getValVT();
3209 // Special handling for extended integers.
3210 if (RVVT != DestVT) {
3211 if (RVVT != MVT::i1 && RVVT != MVT::i8 && RVVT != MVT::i16)
3214 if (!Outs[0].Flags.isZExt() && !Outs[0].Flags.isSExt())
3217 bool IsZExt = Outs[0].Flags.isZExt();
3218 SrcReg = emitIntExt(RVVT, SrcReg, DestVT, IsZExt);
3224 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3225 TII.get(TargetOpcode::COPY), DestReg).addReg(SrcReg);
3227 // Add register to return instruction.
3228 RetRegs.push_back(VA.getLocReg());
3231 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3232 TII.get(AArch64::RET_ReallyLR));
3233 for (unsigned i = 0, e = RetRegs.size(); i != e; ++i)
3234 MIB.addReg(RetRegs[i], RegState::Implicit);
3238 bool AArch64FastISel::selectTrunc(const Instruction *I) {
3239 Type *DestTy = I->getType();
3240 Value *Op = I->getOperand(0);
3241 Type *SrcTy = Op->getType();
3243 EVT SrcEVT = TLI.getValueType(SrcTy, true);
3244 EVT DestEVT = TLI.getValueType(DestTy, true);
3245 if (!SrcEVT.isSimple())
3247 if (!DestEVT.isSimple())
3250 MVT SrcVT = SrcEVT.getSimpleVT();
3251 MVT DestVT = DestEVT.getSimpleVT();
3253 if (SrcVT != MVT::i64 && SrcVT != MVT::i32 && SrcVT != MVT::i16 &&
3256 if (DestVT != MVT::i32 && DestVT != MVT::i16 && DestVT != MVT::i8 &&
3260 unsigned SrcReg = getRegForValue(Op);
3263 bool SrcIsKill = hasTrivialKill(Op);
3265 // If we're truncating from i64 to a smaller non-legal type then generate an
3266 // AND. Otherwise, we know the high bits are undefined and a truncate only
3267 // generate a COPY. We cannot mark the source register also as result
3268 // register, because this can incorrectly transfer the kill flag onto the
3271 if (SrcVT == MVT::i64) {
3273 switch (DestVT.SimpleTy) {
3275 // Trunc i64 to i32 is handled by the target-independent fast-isel.
3287 // Issue an extract_subreg to get the lower 32-bits.
3288 unsigned Reg32 = fastEmitInst_extractsubreg(MVT::i32, SrcReg, SrcIsKill,
3290 // Create the AND instruction which performs the actual truncation.
3291 ResultReg = emitAnd_ri(MVT::i32, Reg32, /*IsKill=*/true, Mask);
3292 assert(ResultReg && "Unexpected AND instruction emission failure.");
3294 ResultReg = createResultReg(&AArch64::GPR32RegClass);
3295 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3296 TII.get(TargetOpcode::COPY), ResultReg)
3297 .addReg(SrcReg, getKillRegState(SrcIsKill));
3300 updateValueMap(I, ResultReg);
3304 unsigned AArch64FastISel::emiti1Ext(unsigned SrcReg, MVT DestVT, bool IsZExt) {
3305 assert((DestVT == MVT::i8 || DestVT == MVT::i16 || DestVT == MVT::i32 ||
3306 DestVT == MVT::i64) &&
3307 "Unexpected value type.");
3308 // Handle i8 and i16 as i32.
3309 if (DestVT == MVT::i8 || DestVT == MVT::i16)
3313 unsigned ResultReg = emitAnd_ri(MVT::i32, SrcReg, /*TODO:IsKill=*/false, 1);
3314 assert(ResultReg && "Unexpected AND instruction emission failure.");
3315 if (DestVT == MVT::i64) {
3316 // We're ZExt i1 to i64. The ANDWri Wd, Ws, #1 implicitly clears the
3317 // upper 32 bits. Emit a SUBREG_TO_REG to extend from Wd to Xd.
3318 unsigned Reg64 = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
3319 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3320 TII.get(AArch64::SUBREG_TO_REG), Reg64)
3323 .addImm(AArch64::sub_32);
3328 if (DestVT == MVT::i64) {
3329 // FIXME: We're SExt i1 to i64.
3332 return fastEmitInst_rii(AArch64::SBFMWri, &AArch64::GPR32RegClass, SrcReg,
3333 /*TODO:IsKill=*/false, 0, 0);
3337 unsigned AArch64FastISel::emitMul_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
3338 unsigned Op1, bool Op1IsKill) {
3340 switch (RetVT.SimpleTy) {
3346 Opc = AArch64::MADDWrrr; ZReg = AArch64::WZR; break;
3348 Opc = AArch64::MADDXrrr; ZReg = AArch64::XZR; break;
3351 const TargetRegisterClass *RC =
3352 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3353 return fastEmitInst_rrr(Opc, RC, Op0, Op0IsKill, Op1, Op1IsKill,
3354 /*IsKill=*/ZReg, true);
3357 unsigned AArch64FastISel::emitSMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
3358 unsigned Op1, bool Op1IsKill) {
3359 if (RetVT != MVT::i64)
3362 return fastEmitInst_rrr(AArch64::SMADDLrrr, &AArch64::GPR64RegClass,
3363 Op0, Op0IsKill, Op1, Op1IsKill,
3364 AArch64::XZR, /*IsKill=*/true);
3367 unsigned AArch64FastISel::emitUMULL_rr(MVT RetVT, unsigned Op0, bool Op0IsKill,
3368 unsigned Op1, bool Op1IsKill) {
3369 if (RetVT != MVT::i64)
3372 return fastEmitInst_rrr(AArch64::UMADDLrrr, &AArch64::GPR64RegClass,
3373 Op0, Op0IsKill, Op1, Op1IsKill,
3374 AArch64::XZR, /*IsKill=*/true);
3377 unsigned AArch64FastISel::emitLSL_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
3378 unsigned Op1Reg, bool Op1IsKill) {
3380 bool NeedTrunc = false;
3382 switch (RetVT.SimpleTy) {
3384 case MVT::i8: Opc = AArch64::LSLVWr; NeedTrunc = true; Mask = 0xff; break;
3385 case MVT::i16: Opc = AArch64::LSLVWr; NeedTrunc = true; Mask = 0xffff; break;
3386 case MVT::i32: Opc = AArch64::LSLVWr; break;
3387 case MVT::i64: Opc = AArch64::LSLVXr; break;
3390 const TargetRegisterClass *RC =
3391 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3393 Op1Reg = emitAnd_ri(MVT::i32, Op1Reg, Op1IsKill, Mask);
3396 unsigned ResultReg = fastEmitInst_rr(Opc, RC, Op0Reg, Op0IsKill, Op1Reg,
3399 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
3403 unsigned AArch64FastISel::emitLSL_ri(MVT RetVT, MVT SrcVT, unsigned Op0,
3404 bool Op0IsKill, uint64_t Shift,
3406 assert(RetVT.SimpleTy >= SrcVT.SimpleTy &&
3407 "Unexpected source/return type pair.");
3408 assert((SrcVT == MVT::i8 || SrcVT == MVT::i16 || SrcVT == MVT::i32 ||
3409 SrcVT == MVT::i64) && "Unexpected source value type.");
3410 assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 ||
3411 RetVT == MVT::i64) && "Unexpected return value type.");
3413 bool Is64Bit = (RetVT == MVT::i64);
3414 unsigned RegSize = Is64Bit ? 64 : 32;
3415 unsigned DstBits = RetVT.getSizeInBits();
3416 unsigned SrcBits = SrcVT.getSizeInBits();
3418 // Don't deal with undefined shifts.
3419 if (Shift >= DstBits)
3422 // For immediate shifts we can fold the zero-/sign-extension into the shift.
3423 // {S|U}BFM Wd, Wn, #r, #s
3424 // Wd<32+s-r,32-r> = Wn<s:0> when r > s
3426 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3427 // %2 = shl i16 %1, 4
3428 // Wd<32+7-28,32-28> = Wn<7:0> <- clamp s to 7
3429 // 0b1111_1111_1111_1111__1111_1010_1010_0000 sext
3430 // 0b0000_0000_0000_0000__0000_0101_0101_0000 sext | zext
3431 // 0b0000_0000_0000_0000__0000_1010_1010_0000 zext
3433 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3434 // %2 = shl i16 %1, 8
3435 // Wd<32+7-24,32-24> = Wn<7:0>
3436 // 0b1111_1111_1111_1111__1010_1010_0000_0000 sext
3437 // 0b0000_0000_0000_0000__0101_0101_0000_0000 sext | zext
3438 // 0b0000_0000_0000_0000__1010_1010_0000_0000 zext
3440 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3441 // %2 = shl i16 %1, 12
3442 // Wd<32+3-20,32-20> = Wn<3:0>
3443 // 0b1111_1111_1111_1111__1010_0000_0000_0000 sext
3444 // 0b0000_0000_0000_0000__0101_0000_0000_0000 sext | zext
3445 // 0b0000_0000_0000_0000__1010_0000_0000_0000 zext
3447 unsigned ImmR = RegSize - Shift;
3448 // Limit the width to the length of the source type.
3449 unsigned ImmS = std::min<unsigned>(SrcBits - 1, DstBits - 1 - Shift);
3450 static const unsigned OpcTable[2][2] = {
3451 {AArch64::SBFMWri, AArch64::SBFMXri},
3452 {AArch64::UBFMWri, AArch64::UBFMXri}
3454 unsigned Opc = OpcTable[IsZext][Is64Bit];
3455 const TargetRegisterClass *RC =
3456 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3457 if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) {
3458 unsigned TmpReg = MRI.createVirtualRegister(RC);
3459 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3460 TII.get(AArch64::SUBREG_TO_REG), TmpReg)
3462 .addReg(Op0, getKillRegState(Op0IsKill))
3463 .addImm(AArch64::sub_32);
3467 return fastEmitInst_rii(Opc, RC, Op0, Op0IsKill, ImmR, ImmS);
3470 unsigned AArch64FastISel::emitLSR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
3471 unsigned Op1Reg, bool Op1IsKill) {
3473 bool NeedTrunc = false;
3475 switch (RetVT.SimpleTy) {
3477 case MVT::i8: Opc = AArch64::LSRVWr; NeedTrunc = true; Mask = 0xff; break;
3478 case MVT::i16: Opc = AArch64::LSRVWr; NeedTrunc = true; Mask = 0xffff; break;
3479 case MVT::i32: Opc = AArch64::LSRVWr; break;
3480 case MVT::i64: Opc = AArch64::LSRVXr; break;
3483 const TargetRegisterClass *RC =
3484 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3486 Op0Reg = emitAnd_ri(MVT::i32, Op0Reg, Op0IsKill, Mask);
3487 Op1Reg = emitAnd_ri(MVT::i32, Op1Reg, Op1IsKill, Mask);
3488 Op0IsKill = Op1IsKill = true;
3490 unsigned ResultReg = fastEmitInst_rr(Opc, RC, Op0Reg, Op0IsKill, Op1Reg,
3493 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
3497 unsigned AArch64FastISel::emitLSR_ri(MVT RetVT, MVT SrcVT, unsigned Op0,
3498 bool Op0IsKill, uint64_t Shift,
3500 assert(RetVT.SimpleTy >= SrcVT.SimpleTy &&
3501 "Unexpected source/return type pair.");
3502 assert((SrcVT == MVT::i8 || SrcVT == MVT::i16 || SrcVT == MVT::i32 ||
3503 SrcVT == MVT::i64) && "Unexpected source value type.");
3504 assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 ||
3505 RetVT == MVT::i64) && "Unexpected return value type.");
3507 bool Is64Bit = (RetVT == MVT::i64);
3508 unsigned RegSize = Is64Bit ? 64 : 32;
3509 unsigned DstBits = RetVT.getSizeInBits();
3510 unsigned SrcBits = SrcVT.getSizeInBits();
3512 // Don't deal with undefined shifts.
3513 if (Shift >= DstBits)
3516 // For immediate shifts we can fold the zero-/sign-extension into the shift.
3517 // {S|U}BFM Wd, Wn, #r, #s
3518 // Wd<s-r:0> = Wn<s:r> when r <= s
3520 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3521 // %2 = lshr i16 %1, 4
3522 // Wd<7-4:0> = Wn<7:4>
3523 // 0b0000_0000_0000_0000__0000_1111_1111_1010 sext
3524 // 0b0000_0000_0000_0000__0000_0000_0000_0101 sext | zext
3525 // 0b0000_0000_0000_0000__0000_0000_0000_1010 zext
3527 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3528 // %2 = lshr i16 %1, 8
3529 // Wd<7-7,0> = Wn<7:7>
3530 // 0b0000_0000_0000_0000__0000_0000_1111_1111 sext
3531 // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
3532 // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
3534 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3535 // %2 = lshr i16 %1, 12
3536 // Wd<7-7,0> = Wn<7:7> <- clamp r to 7
3537 // 0b0000_0000_0000_0000__0000_0000_0000_1111 sext
3538 // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
3539 // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
3541 if (Shift >= SrcBits && IsZExt)
3542 return materializeInt(ConstantInt::get(*Context, APInt(RegSize, 0)), RetVT);
3544 // It is not possible to fold a sign-extend into the LShr instruction. In this
3545 // case emit a sign-extend.
3547 Op0 = emitIntExt(SrcVT, Op0, RetVT, IsZExt);
3552 SrcBits = SrcVT.getSizeInBits();
3556 unsigned ImmR = std::min<unsigned>(SrcBits - 1, Shift);
3557 unsigned ImmS = SrcBits - 1;
3558 static const unsigned OpcTable[2][2] = {
3559 {AArch64::SBFMWri, AArch64::SBFMXri},
3560 {AArch64::UBFMWri, AArch64::UBFMXri}
3562 unsigned Opc = OpcTable[IsZExt][Is64Bit];
3563 const TargetRegisterClass *RC =
3564 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3565 if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) {
3566 unsigned TmpReg = MRI.createVirtualRegister(RC);
3567 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3568 TII.get(AArch64::SUBREG_TO_REG), TmpReg)
3570 .addReg(Op0, getKillRegState(Op0IsKill))
3571 .addImm(AArch64::sub_32);
3575 return fastEmitInst_rii(Opc, RC, Op0, Op0IsKill, ImmR, ImmS);
3578 unsigned AArch64FastISel::emitASR_rr(MVT RetVT, unsigned Op0Reg, bool Op0IsKill,
3579 unsigned Op1Reg, bool Op1IsKill) {
3581 bool NeedTrunc = false;
3583 switch (RetVT.SimpleTy) {
3585 case MVT::i8: Opc = AArch64::ASRVWr; NeedTrunc = true; Mask = 0xff; break;
3586 case MVT::i16: Opc = AArch64::ASRVWr; NeedTrunc = true; Mask = 0xffff; break;
3587 case MVT::i32: Opc = AArch64::ASRVWr; break;
3588 case MVT::i64: Opc = AArch64::ASRVXr; break;
3591 const TargetRegisterClass *RC =
3592 (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3594 Op0Reg = emitIntExt(RetVT, Op0Reg, MVT::i32, /*IsZExt=*/false);
3595 Op1Reg = emitAnd_ri(MVT::i32, Op1Reg, Op1IsKill, Mask);
3596 Op0IsKill = Op1IsKill = true;
3598 unsigned ResultReg = fastEmitInst_rr(Opc, RC, Op0Reg, Op0IsKill, Op1Reg,
3601 ResultReg = emitAnd_ri(MVT::i32, ResultReg, /*IsKill=*/true, Mask);
3605 unsigned AArch64FastISel::emitASR_ri(MVT RetVT, MVT SrcVT, unsigned Op0,
3606 bool Op0IsKill, uint64_t Shift,
3608 assert(RetVT.SimpleTy >= SrcVT.SimpleTy &&
3609 "Unexpected source/return type pair.");
3610 assert((SrcVT == MVT::i8 || SrcVT == MVT::i16 || SrcVT == MVT::i32 ||
3611 SrcVT == MVT::i64) && "Unexpected source value type.");
3612 assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 ||
3613 RetVT == MVT::i64) && "Unexpected return value type.");
3615 bool Is64Bit = (RetVT == MVT::i64);
3616 unsigned RegSize = Is64Bit ? 64 : 32;
3617 unsigned DstBits = RetVT.getSizeInBits();
3618 unsigned SrcBits = SrcVT.getSizeInBits();
3620 // Don't deal with undefined shifts.
3621 if (Shift >= DstBits)
3624 // For immediate shifts we can fold the zero-/sign-extension into the shift.
3625 // {S|U}BFM Wd, Wn, #r, #s
3626 // Wd<s-r:0> = Wn<s:r> when r <= s
3628 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3629 // %2 = ashr i16 %1, 4
3630 // Wd<7-4:0> = Wn<7:4>
3631 // 0b1111_1111_1111_1111__1111_1111_1111_1010 sext
3632 // 0b0000_0000_0000_0000__0000_0000_0000_0101 sext | zext
3633 // 0b0000_0000_0000_0000__0000_0000_0000_1010 zext
3635 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3636 // %2 = ashr i16 %1, 8
3637 // Wd<7-7,0> = Wn<7:7>
3638 // 0b1111_1111_1111_1111__1111_1111_1111_1111 sext
3639 // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
3640 // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
3642 // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
3643 // %2 = ashr i16 %1, 12
3644 // Wd<7-7,0> = Wn<7:7> <- clamp r to 7
3645 // 0b1111_1111_1111_1111__1111_1111_1111_1111 sext
3646 // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
3647 // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext
3649 if (Shift >= SrcBits && IsZExt)
3650 return materializeInt(ConstantInt::get(*Context, APInt(RegSize, 0)), RetVT);
3652 unsigned ImmR = std::min<unsigned>(SrcBits - 1, Shift);
3653 unsigned ImmS = SrcBits - 1;
3654 static const unsigned OpcTable[2][2] = {
3655 {AArch64::SBFMWri, AArch64::SBFMXri},
3656 {AArch64::UBFMWri, AArch64::UBFMXri}
3658 unsigned Opc = OpcTable[IsZExt][Is64Bit];
3659 const TargetRegisterClass *RC =
3660 Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3661 if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) {
3662 unsigned TmpReg = MRI.createVirtualRegister(RC);
3663 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3664 TII.get(AArch64::SUBREG_TO_REG), TmpReg)
3666 .addReg(Op0, getKillRegState(Op0IsKill))
3667 .addImm(AArch64::sub_32);
3671 return fastEmitInst_rii(Opc, RC, Op0, Op0IsKill, ImmR, ImmS);
3674 unsigned AArch64FastISel::emitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT,
3676 assert(DestVT != MVT::i1 && "ZeroExt/SignExt an i1?");
3678 // FastISel does not have plumbing to deal with extensions where the SrcVT or
3679 // DestVT are odd things, so test to make sure that they are both types we can
3680 // handle (i1/i8/i16/i32 for SrcVT and i8/i16/i32/i64 for DestVT), otherwise
3681 // bail out to SelectionDAG.
3682 if (((DestVT != MVT::i8) && (DestVT != MVT::i16) &&
3683 (DestVT != MVT::i32) && (DestVT != MVT::i64)) ||
3684 ((SrcVT != MVT::i1) && (SrcVT != MVT::i8) &&
3685 (SrcVT != MVT::i16) && (SrcVT != MVT::i32)))
3691 switch (SrcVT.SimpleTy) {
3695 return emiti1Ext(SrcReg, DestVT, IsZExt);
3697 if (DestVT == MVT::i64)
3698 Opc = IsZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
3700 Opc = IsZExt ? AArch64::UBFMWri : AArch64::SBFMWri;
3704 if (DestVT == MVT::i64)
3705 Opc = IsZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
3707 Opc = IsZExt ? AArch64::UBFMWri : AArch64::SBFMWri;
3711 assert(DestVT == MVT::i64 && "IntExt i32 to i32?!?");
3712 Opc = IsZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
3717 // Handle i8 and i16 as i32.
3718 if (DestVT == MVT::i8 || DestVT == MVT::i16)
3720 else if (DestVT == MVT::i64) {
3721 unsigned Src64 = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
3722 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3723 TII.get(AArch64::SUBREG_TO_REG), Src64)
3726 .addImm(AArch64::sub_32);
3730 const TargetRegisterClass *RC =
3731 (DestVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3732 return fastEmitInst_rii(Opc, RC, SrcReg, /*TODO:IsKill=*/false, 0, Imm);
3735 bool AArch64FastISel::selectIntExt(const Instruction *I) {
3736 // On ARM, in general, integer casts don't involve legal types; this code
3737 // handles promotable integers. The high bits for a type smaller than
3738 // the register size are assumed to be undefined.
3739 Type *DestTy = I->getType();
3740 Value *Src = I->getOperand(0);
3741 Type *SrcTy = Src->getType();
3743 unsigned SrcReg = getRegForValue(Src);
3747 EVT SrcEVT = TLI.getValueType(SrcTy, true);
3748 EVT DestEVT = TLI.getValueType(DestTy, true);
3749 if (!SrcEVT.isSimple())
3751 if (!DestEVT.isSimple())
3754 MVT SrcVT = SrcEVT.getSimpleVT();
3755 MVT DestVT = DestEVT.getSimpleVT();
3756 unsigned ResultReg = 0;
3758 bool IsZExt = isa<ZExtInst>(I);
3759 // Check if it is an argument and if it is already zero/sign-extended.
3760 if (const auto *Arg = dyn_cast<Argument>(Src)) {
3761 if ((IsZExt && Arg->hasZExtAttr()) || (!IsZExt && Arg->hasSExtAttr())) {
3762 if (DestVT == MVT::i64) {
3763 ResultReg = createResultReg(TLI.getRegClassFor(DestVT));
3764 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3765 TII.get(AArch64::SUBREG_TO_REG), ResultReg)
3768 .addImm(AArch64::sub_32);
3775 ResultReg = emitIntExt(SrcVT, SrcReg, DestVT, IsZExt);
3780 updateValueMap(I, ResultReg);
3784 bool AArch64FastISel::selectRem(const Instruction *I, unsigned ISDOpcode) {
3785 EVT DestEVT = TLI.getValueType(I->getType(), true);
3786 if (!DestEVT.isSimple())
3789 MVT DestVT = DestEVT.getSimpleVT();
3790 if (DestVT != MVT::i64 && DestVT != MVT::i32)
3794 bool Is64bit = (DestVT == MVT::i64);
3795 switch (ISDOpcode) {
3799 DivOpc = Is64bit ? AArch64::SDIVXr : AArch64::SDIVWr;
3802 DivOpc = Is64bit ? AArch64::UDIVXr : AArch64::UDIVWr;
3805 unsigned MSubOpc = Is64bit ? AArch64::MSUBXrrr : AArch64::MSUBWrrr;
3806 unsigned Src0Reg = getRegForValue(I->getOperand(0));
3809 bool Src0IsKill = hasTrivialKill(I->getOperand(0));
3811 unsigned Src1Reg = getRegForValue(I->getOperand(1));
3814 bool Src1IsKill = hasTrivialKill(I->getOperand(1));
3816 const TargetRegisterClass *RC =
3817 (DestVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
3818 unsigned QuotReg = fastEmitInst_rr(DivOpc, RC, Src0Reg, /*IsKill=*/false,
3819 Src1Reg, /*IsKill=*/false);
3820 assert(QuotReg && "Unexpected DIV instruction emission failure.");
3821 // The remainder is computed as numerator - (quotient * denominator) using the
3822 // MSUB instruction.
3823 unsigned ResultReg = fastEmitInst_rrr(MSubOpc, RC, QuotReg, /*IsKill=*/true,
3824 Src1Reg, Src1IsKill, Src0Reg,
3826 updateValueMap(I, ResultReg);
3830 bool AArch64FastISel::selectMul(const Instruction *I) {
3832 if (!isTypeSupported(I->getType(), VT, /*IsVectorAllowed=*/true))
3836 return selectBinaryOp(I, ISD::MUL);
3838 const Value *Src0 = I->getOperand(0);
3839 const Value *Src1 = I->getOperand(1);
3840 if (const auto *C = dyn_cast<ConstantInt>(Src0))
3841 if (C->getValue().isPowerOf2())
3842 std::swap(Src0, Src1);
3844 // Try to simplify to a shift instruction.
3845 if (const auto *C = dyn_cast<ConstantInt>(Src1))
3846 if (C->getValue().isPowerOf2()) {
3847 uint64_t ShiftVal = C->getValue().logBase2();
3850 if (const auto *ZExt = dyn_cast<ZExtInst>(Src0)) {
3852 if (isValueAvailable(ZExt) && isTypeSupported(ZExt->getSrcTy(), VT)) {
3855 Src0 = ZExt->getOperand(0);
3857 } else if (const auto *SExt = dyn_cast<SExtInst>(Src0)) {
3859 if (isValueAvailable(SExt) && isTypeSupported(SExt->getSrcTy(), VT)) {
3862 Src0 = SExt->getOperand(0);
3866 unsigned Src0Reg = getRegForValue(Src0);
3869 bool Src0IsKill = hasTrivialKill(Src0);
3871 unsigned ResultReg =
3872 emitLSL_ri(VT, SrcVT, Src0Reg, Src0IsKill, ShiftVal, IsZExt);
3875 updateValueMap(I, ResultReg);
3880 unsigned Src0Reg = getRegForValue(I->getOperand(0));
3883 bool Src0IsKill = hasTrivialKill(I->getOperand(0));
3885 unsigned Src1Reg = getRegForValue(I->getOperand(1));
3888 bool Src1IsKill = hasTrivialKill(I->getOperand(1));
3890 unsigned ResultReg = emitMul_rr(VT, Src0Reg, Src0IsKill, Src1Reg, Src1IsKill);
3895 updateValueMap(I, ResultReg);
3899 bool AArch64FastISel::selectShift(const Instruction *I) {
3901 if (!isTypeSupported(I->getType(), RetVT, /*IsVectorAllowed=*/true))
3904 if (RetVT.isVector())
3905 return selectOperator(I, I->getOpcode());
3907 if (const auto *C = dyn_cast<ConstantInt>(I->getOperand(1))) {
3908 unsigned ResultReg = 0;
3909 uint64_t ShiftVal = C->getZExtValue();
3911 bool IsZExt = (I->getOpcode() == Instruction::AShr) ? false : true;
3912 const Value *Op0 = I->getOperand(0);
3913 if (const auto *ZExt = dyn_cast<ZExtInst>(Op0)) {
3915 if (isValueAvailable(ZExt) && isTypeSupported(ZExt->getSrcTy(), TmpVT)) {
3918 Op0 = ZExt->getOperand(0);
3920 } else if (const auto *SExt = dyn_cast<SExtInst>(Op0)) {
3922 if (isValueAvailable(SExt) && isTypeSupported(SExt->getSrcTy(), TmpVT)) {
3925 Op0 = SExt->getOperand(0);
3929 unsigned Op0Reg = getRegForValue(Op0);
3932 bool Op0IsKill = hasTrivialKill(Op0);
3934 switch (I->getOpcode()) {
3935 default: llvm_unreachable("Unexpected instruction.");
3936 case Instruction::Shl:
3937 ResultReg = emitLSL_ri(RetVT, SrcVT, Op0Reg, Op0IsKill, ShiftVal, IsZExt);
3939 case Instruction::AShr:
3940 ResultReg = emitASR_ri(RetVT, SrcVT, Op0Reg, Op0IsKill, ShiftVal, IsZExt);
3942 case Instruction::LShr:
3943 ResultReg = emitLSR_ri(RetVT, SrcVT, Op0Reg, Op0IsKill, ShiftVal, IsZExt);
3949 updateValueMap(I, ResultReg);
3953 unsigned Op0Reg = getRegForValue(I->getOperand(0));
3956 bool Op0IsKill = hasTrivialKill(I->getOperand(0));
3958 unsigned Op1Reg = getRegForValue(I->getOperand(1));
3961 bool Op1IsKill = hasTrivialKill(I->getOperand(1));
3963 unsigned ResultReg = 0;
3964 switch (I->getOpcode()) {
3965 default: llvm_unreachable("Unexpected instruction.");
3966 case Instruction::Shl:
3967 ResultReg = emitLSL_rr(RetVT, Op0Reg, Op0IsKill, Op1Reg, Op1IsKill);
3969 case Instruction::AShr:
3970 ResultReg = emitASR_rr(RetVT, Op0Reg, Op0IsKill, Op1Reg, Op1IsKill);
3972 case Instruction::LShr:
3973 ResultReg = emitLSR_rr(RetVT, Op0Reg, Op0IsKill, Op1Reg, Op1IsKill);
3980 updateValueMap(I, ResultReg);
3984 bool AArch64FastISel::selectBitCast(const Instruction *I) {
3987 if (!isTypeLegal(I->getOperand(0)->getType(), SrcVT))
3989 if (!isTypeLegal(I->getType(), RetVT))
3993 if (RetVT == MVT::f32 && SrcVT == MVT::i32)
3994 Opc = AArch64::FMOVWSr;
3995 else if (RetVT == MVT::f64 && SrcVT == MVT::i64)
3996 Opc = AArch64::FMOVXDr;
3997 else if (RetVT == MVT::i32 && SrcVT == MVT::f32)
3998 Opc = AArch64::FMOVSWr;
3999 else if (RetVT == MVT::i64 && SrcVT == MVT::f64)
4000 Opc = AArch64::FMOVDXr;
4004 const TargetRegisterClass *RC = nullptr;
4005 switch (RetVT.SimpleTy) {
4006 default: llvm_unreachable("Unexpected value type.");
4007 case MVT::i32: RC = &AArch64::GPR32RegClass; break;
4008 case MVT::i64: RC = &AArch64::GPR64RegClass; break;
4009 case MVT::f32: RC = &AArch64::FPR32RegClass; break;
4010 case MVT::f64: RC = &AArch64::FPR64RegClass; break;
4012 unsigned Op0Reg = getRegForValue(I->getOperand(0));
4015 bool Op0IsKill = hasTrivialKill(I->getOperand(0));
4016 unsigned ResultReg = fastEmitInst_r(Opc, RC, Op0Reg, Op0IsKill);
4021 updateValueMap(I, ResultReg);
4025 bool AArch64FastISel::selectFRem(const Instruction *I) {
4027 if (!isTypeLegal(I->getType(), RetVT))
4031 switch (RetVT.SimpleTy) {
4035 LC = RTLIB::REM_F32;
4038 LC = RTLIB::REM_F64;
4043 Args.reserve(I->getNumOperands());
4045 // Populate the argument list.
4046 for (auto &Arg : I->operands()) {
4049 Entry.Ty = Arg->getType();
4050 Args.push_back(Entry);
4053 CallLoweringInfo CLI;
4054 CLI.setCallee(TLI.getLibcallCallingConv(LC), I->getType(),
4055 TLI.getLibcallName(LC), std::move(Args));
4056 if (!lowerCallTo(CLI))
4058 updateValueMap(I, CLI.ResultReg);
4062 bool AArch64FastISel::selectSDiv(const Instruction *I) {
4064 if (!isTypeLegal(I->getType(), VT))
4067 if (!isa<ConstantInt>(I->getOperand(1)))
4068 return selectBinaryOp(I, ISD::SDIV);
4070 const APInt &C = cast<ConstantInt>(I->getOperand(1))->getValue();
4071 if ((VT != MVT::i32 && VT != MVT::i64) || !C ||
4072 !(C.isPowerOf2() || (-C).isPowerOf2()))
4073 return selectBinaryOp(I, ISD::SDIV);
4075 unsigned Lg2 = C.countTrailingZeros();
4076 unsigned Src0Reg = getRegForValue(I->getOperand(0));
4079 bool Src0IsKill = hasTrivialKill(I->getOperand(0));
4081 if (cast<BinaryOperator>(I)->isExact()) {
4082 unsigned ResultReg = emitASR_ri(VT, VT, Src0Reg, Src0IsKill, Lg2);
4085 updateValueMap(I, ResultReg);
4089 unsigned Pow2MinusOne = (1 << Lg2) - 1;
4090 unsigned AddReg = emitAddSub_ri(/*UseAdd=*/true, VT, Src0Reg,
4091 /*IsKill=*/false, Pow2MinusOne);
4095 // (Src0 < 0) ? Pow2 - 1 : 0;
4096 if (!emitICmp_ri(VT, Src0Reg, /*IsKill=*/false, 0))
4100 const TargetRegisterClass *RC;
4101 if (VT == MVT::i64) {
4102 SelectOpc = AArch64::CSELXr;
4103 RC = &AArch64::GPR64RegClass;
4105 SelectOpc = AArch64::CSELWr;
4106 RC = &AArch64::GPR32RegClass;
4108 unsigned SelectReg =
4109 fastEmitInst_rri(SelectOpc, RC, AddReg, /*IsKill=*/true, Src0Reg,
4110 Src0IsKill, AArch64CC::LT);
4114 // Divide by Pow2 --> ashr. If we're dividing by a negative value we must also
4115 // negate the result.
4116 unsigned ZeroReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
4119 ResultReg = emitAddSub_rs(/*UseAdd=*/false, VT, ZeroReg, /*IsKill=*/true,
4120 SelectReg, /*IsKill=*/true, AArch64_AM::ASR, Lg2);
4122 ResultReg = emitASR_ri(VT, VT, SelectReg, /*IsKill=*/true, Lg2);
4127 updateValueMap(I, ResultReg);
4131 bool AArch64FastISel::fastSelectInstruction(const Instruction *I) {
4132 switch (I->getOpcode()) {
4135 case Instruction::Add:
4136 case Instruction::Sub:
4137 return selectAddSub(I);
4138 case Instruction::Mul:
4139 return selectMul(I);
4140 case Instruction::SDiv:
4141 return selectSDiv(I);
4142 case Instruction::SRem:
4143 if (!selectBinaryOp(I, ISD::SREM))
4144 return selectRem(I, ISD::SREM);
4146 case Instruction::URem:
4147 if (!selectBinaryOp(I, ISD::UREM))
4148 return selectRem(I, ISD::UREM);
4150 case Instruction::Shl:
4151 case Instruction::LShr:
4152 case Instruction::AShr:
4153 return selectShift(I);
4154 case Instruction::And:
4155 case Instruction::Or:
4156 case Instruction::Xor:
4157 return selectLogicalOp(I);
4158 case Instruction::Br:
4159 return selectBranch(I);
4160 case Instruction::IndirectBr:
4161 return selectIndirectBr(I);
4162 case Instruction::BitCast:
4163 if (!FastISel::selectBitCast(I))
4164 return selectBitCast(I);
4166 case Instruction::FPToSI:
4167 if (!selectCast(I, ISD::FP_TO_SINT))
4168 return selectFPToInt(I, /*Signed=*/true);
4170 case Instruction::FPToUI:
4171 return selectFPToInt(I, /*Signed=*/false);
4172 case Instruction::ZExt:
4173 if (!selectCast(I, ISD::ZERO_EXTEND))
4174 return selectIntExt(I);
4176 case Instruction::SExt:
4177 if (!selectCast(I, ISD::SIGN_EXTEND))
4178 return selectIntExt(I);
4180 case Instruction::Trunc:
4181 if (!selectCast(I, ISD::TRUNCATE))
4182 return selectTrunc(I);
4184 case Instruction::FPExt:
4185 return selectFPExt(I);
4186 case Instruction::FPTrunc:
4187 return selectFPTrunc(I);
4188 case Instruction::SIToFP:
4189 if (!selectCast(I, ISD::SINT_TO_FP))
4190 return selectIntToFP(I, /*Signed=*/true);
4192 case Instruction::UIToFP:
4193 return selectIntToFP(I, /*Signed=*/false);
4194 case Instruction::Load:
4195 return selectLoad(I);
4196 case Instruction::Store:
4197 return selectStore(I);
4198 case Instruction::FCmp:
4199 case Instruction::ICmp:
4200 return selectCmp(I);
4201 case Instruction::Select:
4202 return selectSelect(I);
4203 case Instruction::Ret:
4204 return selectRet(I);
4205 case Instruction::FRem:
4206 return selectFRem(I);
4209 // fall-back to target-independent instruction selection.
4210 return selectOperator(I, I->getOpcode());
4211 // Silence warnings.
4212 (void)&CC_AArch64_DarwinPCS_VarArg;
4216 llvm::FastISel *AArch64::createFastISel(FunctionLoweringInfo &FuncInfo,
4217 const TargetLibraryInfo *LibInfo) {
4218 return new AArch64FastISel(FuncInfo, LibInfo);