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 "AArch64TargetMachine.h"
18 #include "AArch64Subtarget.h"
19 #include "MCTargetDesc/AArch64AddressingModes.h"
20 #include "llvm/CodeGen/CallingConvLower.h"
21 #include "llvm/CodeGen/FastISel.h"
22 #include "llvm/CodeGen/FunctionLoweringInfo.h"
23 #include "llvm/CodeGen/MachineConstantPool.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.h"
26 #include "llvm/CodeGen/MachineRegisterInfo.h"
27 #include "llvm/IR/CallingConv.h"
28 #include "llvm/IR/DataLayout.h"
29 #include "llvm/IR/DerivedTypes.h"
30 #include "llvm/IR/Function.h"
31 #include "llvm/IR/GetElementPtrTypeIterator.h"
32 #include "llvm/IR/GlobalAlias.h"
33 #include "llvm/IR/GlobalVariable.h"
34 #include "llvm/IR/Instructions.h"
35 #include "llvm/IR/IntrinsicInst.h"
36 #include "llvm/IR/Operator.h"
37 #include "llvm/Support/CommandLine.h"
42 class AArch64FastISel : public FastISel {
60 Address() : Kind(RegBase), Offset(0) { Base.Reg = 0; }
61 void setKind(BaseKind K) { Kind = K; }
62 BaseKind getKind() const { return Kind; }
63 bool isRegBase() const { return Kind == RegBase; }
64 bool isFIBase() const { return Kind == FrameIndexBase; }
65 void setReg(unsigned Reg) {
66 assert(isRegBase() && "Invalid base register access!");
69 unsigned getReg() const {
70 assert(isRegBase() && "Invalid base register access!");
73 void setFI(unsigned FI) {
74 assert(isFIBase() && "Invalid base frame index access!");
77 unsigned getFI() const {
78 assert(isFIBase() && "Invalid base frame index access!");
81 void setOffset(int64_t O) { Offset = O; }
82 int64_t getOffset() { return Offset; }
84 bool isValid() { return isFIBase() || (isRegBase() && getReg() != 0); }
87 /// Subtarget - Keep a pointer to the AArch64Subtarget around so that we can
88 /// make the right decision when generating code for different targets.
89 const AArch64Subtarget *Subtarget;
93 // Selection routines.
94 bool SelectLoad(const Instruction *I);
95 bool SelectStore(const Instruction *I);
96 bool SelectBranch(const Instruction *I);
97 bool SelectIndirectBr(const Instruction *I);
98 bool SelectCmp(const Instruction *I);
99 bool SelectSelect(const Instruction *I);
100 bool SelectFPExt(const Instruction *I);
101 bool SelectFPTrunc(const Instruction *I);
102 bool SelectFPToInt(const Instruction *I, bool Signed);
103 bool SelectIntToFP(const Instruction *I, bool Signed);
104 bool SelectRem(const Instruction *I, unsigned ISDOpcode);
105 bool SelectCall(const Instruction *I, const char *IntrMemName);
106 bool SelectIntrinsicCall(const IntrinsicInst &I);
107 bool SelectRet(const Instruction *I);
108 bool SelectTrunc(const Instruction *I);
109 bool SelectIntExt(const Instruction *I);
110 bool SelectMul(const Instruction *I);
112 // Utility helper routines.
113 bool isTypeLegal(Type *Ty, MVT &VT);
114 bool isLoadStoreTypeLegal(Type *Ty, MVT &VT);
115 bool ComputeAddress(const Value *Obj, Address &Addr);
116 bool SimplifyAddress(Address &Addr, MVT VT, int64_t ScaleFactor,
118 void AddLoadStoreOperands(Address &Addr, const MachineInstrBuilder &MIB,
119 unsigned Flags, bool UseUnscaled);
120 bool IsMemCpySmall(uint64_t Len, unsigned Alignment);
121 bool TryEmitSmallMemCpy(Address Dest, Address Src, uint64_t Len,
124 bool EmitCmp(Value *Src1Value, Value *Src2Value, bool isZExt);
125 bool EmitLoad(MVT VT, unsigned &ResultReg, Address Addr,
126 bool UseUnscaled = false);
127 bool EmitStore(MVT VT, unsigned SrcReg, Address Addr,
128 bool UseUnscaled = false);
129 unsigned EmitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, bool isZExt);
130 unsigned Emiti1Ext(unsigned SrcReg, MVT DestVT, bool isZExt);
132 unsigned AArch64MaterializeFP(const ConstantFP *CFP, MVT VT);
133 unsigned AArch64MaterializeGV(const GlobalValue *GV);
135 // Call handling routines.
137 CCAssignFn *CCAssignFnForCall(CallingConv::ID CC) const;
138 bool ProcessCallArgs(SmallVectorImpl<Value *> &Args,
139 SmallVectorImpl<unsigned> &ArgRegs,
140 SmallVectorImpl<MVT> &ArgVTs,
141 SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
142 SmallVectorImpl<unsigned> &RegArgs, CallingConv::ID CC,
144 bool FinishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
145 const Instruction *I, CallingConv::ID CC, unsigned &NumBytes);
148 // Backend specific FastISel code.
149 unsigned TargetMaterializeAlloca(const AllocaInst *AI) override;
150 unsigned TargetMaterializeConstant(const Constant *C) override;
152 explicit AArch64FastISel(FunctionLoweringInfo &funcInfo,
153 const TargetLibraryInfo *libInfo)
154 : FastISel(funcInfo, libInfo) {
155 Subtarget = &TM.getSubtarget<AArch64Subtarget>();
156 Context = &funcInfo.Fn->getContext();
159 bool TargetSelectInstruction(const Instruction *I) override;
161 #include "AArch64GenFastISel.inc"
164 } // end anonymous namespace
166 #include "AArch64GenCallingConv.inc"
168 CCAssignFn *AArch64FastISel::CCAssignFnForCall(CallingConv::ID CC) const {
169 if (CC == CallingConv::WebKit_JS)
170 return CC_AArch64_WebKit_JS;
171 return Subtarget->isTargetDarwin() ? CC_AArch64_DarwinPCS : CC_AArch64_AAPCS;
174 unsigned AArch64FastISel::TargetMaterializeAlloca(const AllocaInst *AI) {
175 assert(TLI.getValueType(AI->getType(), true) == MVT::i64 &&
176 "Alloca should always return a pointer.");
178 // Don't handle dynamic allocas.
179 if (!FuncInfo.StaticAllocaMap.count(AI))
182 DenseMap<const AllocaInst *, int>::iterator SI =
183 FuncInfo.StaticAllocaMap.find(AI);
185 if (SI != FuncInfo.StaticAllocaMap.end()) {
186 unsigned ResultReg = createResultReg(&AArch64::GPR64RegClass);
187 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri),
189 .addFrameIndex(SI->second)
198 unsigned AArch64FastISel::AArch64MaterializeFP(const ConstantFP *CFP, MVT VT) {
199 if (VT != MVT::f32 && VT != MVT::f64)
202 const APFloat Val = CFP->getValueAPF();
203 bool is64bit = (VT == MVT::f64);
205 // This checks to see if we can use FMOV instructions to materialize
206 // a constant, otherwise we have to materialize via the constant pool.
207 if (TLI.isFPImmLegal(Val, VT)) {
211 Imm = AArch64_AM::getFP64Imm(Val);
212 Opc = AArch64::FMOVDi;
214 Imm = AArch64_AM::getFP32Imm(Val);
215 Opc = AArch64::FMOVSi;
217 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
218 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
223 // Materialize via constant pool. MachineConstantPool wants an explicit
225 unsigned Align = DL.getPrefTypeAlignment(CFP->getType());
227 Align = DL.getTypeAllocSize(CFP->getType());
229 unsigned Idx = MCP.getConstantPoolIndex(cast<Constant>(CFP), Align);
230 unsigned ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
231 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
232 ADRPReg).addConstantPoolIndex(Idx, 0, AArch64II::MO_PAGE);
234 unsigned Opc = is64bit ? AArch64::LDRDui : AArch64::LDRSui;
235 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
236 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
238 .addConstantPoolIndex(Idx, 0, AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
242 unsigned AArch64FastISel::AArch64MaterializeGV(const GlobalValue *GV) {
243 // We can't handle thread-local variables quickly yet. Unfortunately we have
244 // to peer through any aliases to find out if that rule applies.
245 const GlobalValue *TLSGV = GV;
246 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
247 TLSGV = GA->getAliasee();
249 // MachO still uses GOT for large code-model accesses, but ELF requires
250 // movz/movk sequences, which FastISel doesn't handle yet.
251 if (TM.getCodeModel() != CodeModel::Small && !Subtarget->isTargetMachO())
254 if (const GlobalVariable *GVar = dyn_cast<GlobalVariable>(TLSGV))
255 if (GVar->isThreadLocal())
258 unsigned char OpFlags = Subtarget->ClassifyGlobalReference(GV, TM);
260 EVT DestEVT = TLI.getValueType(GV->getType(), true);
261 if (!DestEVT.isSimple())
264 unsigned ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
267 if (OpFlags & AArch64II::MO_GOT) {
269 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
271 .addGlobalAddress(GV, 0, AArch64II::MO_GOT | AArch64II::MO_PAGE);
273 ResultReg = createResultReg(&AArch64::GPR64RegClass);
274 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::LDRXui),
277 .addGlobalAddress(GV, 0, AArch64II::MO_GOT | AArch64II::MO_PAGEOFF |
281 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADRP),
282 ADRPReg).addGlobalAddress(GV, 0, AArch64II::MO_PAGE);
284 ResultReg = createResultReg(&AArch64::GPR64spRegClass);
285 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri),
288 .addGlobalAddress(GV, 0, AArch64II::MO_PAGEOFF | AArch64II::MO_NC)
294 unsigned AArch64FastISel::TargetMaterializeConstant(const Constant *C) {
295 EVT CEVT = TLI.getValueType(C->getType(), true);
297 // Only handle simple types.
298 if (!CEVT.isSimple())
300 MVT VT = CEVT.getSimpleVT();
302 // FIXME: Handle ConstantInt.
303 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
304 return AArch64MaterializeFP(CFP, VT);
305 else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
306 return AArch64MaterializeGV(GV);
311 // Computes the address to get to an object.
312 bool AArch64FastISel::ComputeAddress(const Value *Obj, Address &Addr) {
313 const User *U = nullptr;
314 unsigned Opcode = Instruction::UserOp1;
315 if (const Instruction *I = dyn_cast<Instruction>(Obj)) {
316 // Don't walk into other basic blocks unless the object is an alloca from
317 // another block, otherwise it may not have a virtual register assigned.
318 if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(Obj)) ||
319 FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
320 Opcode = I->getOpcode();
323 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Obj)) {
324 Opcode = C->getOpcode();
328 if (const PointerType *Ty = dyn_cast<PointerType>(Obj->getType()))
329 if (Ty->getAddressSpace() > 255)
330 // Fast instruction selection doesn't support the special
337 case Instruction::BitCast: {
338 // Look through bitcasts.
339 return ComputeAddress(U->getOperand(0), Addr);
341 case Instruction::IntToPtr: {
342 // Look past no-op inttoptrs.
343 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
344 return ComputeAddress(U->getOperand(0), Addr);
347 case Instruction::PtrToInt: {
348 // Look past no-op ptrtoints.
349 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
350 return ComputeAddress(U->getOperand(0), Addr);
353 case Instruction::GetElementPtr: {
354 Address SavedAddr = Addr;
355 uint64_t TmpOffset = Addr.getOffset();
357 // Iterate through the GEP folding the constants into offsets where
359 gep_type_iterator GTI = gep_type_begin(U);
360 for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e;
362 const Value *Op = *i;
363 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
364 const StructLayout *SL = DL.getStructLayout(STy);
365 unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
366 TmpOffset += SL->getElementOffset(Idx);
368 uint64_t S = DL.getTypeAllocSize(GTI.getIndexedType());
370 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
371 // Constant-offset addressing.
372 TmpOffset += CI->getSExtValue() * S;
375 if (canFoldAddIntoGEP(U, Op)) {
376 // A compatible add with a constant operand. Fold the constant.
378 cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
379 TmpOffset += CI->getSExtValue() * S;
380 // Iterate on the other operand.
381 Op = cast<AddOperator>(Op)->getOperand(0);
385 goto unsupported_gep;
390 // Try to grab the base operand now.
391 Addr.setOffset(TmpOffset);
392 if (ComputeAddress(U->getOperand(0), Addr))
395 // We failed, restore everything and try the other options.
401 case Instruction::Alloca: {
402 const AllocaInst *AI = cast<AllocaInst>(Obj);
403 DenseMap<const AllocaInst *, int>::iterator SI =
404 FuncInfo.StaticAllocaMap.find(AI);
405 if (SI != FuncInfo.StaticAllocaMap.end()) {
406 Addr.setKind(Address::FrameIndexBase);
407 Addr.setFI(SI->second);
414 // Try to get this in a register if nothing else has worked.
416 Addr.setReg(getRegForValue(Obj));
417 return Addr.isValid();
420 bool AArch64FastISel::isTypeLegal(Type *Ty, MVT &VT) {
421 EVT evt = TLI.getValueType(Ty, true);
423 // Only handle simple types.
424 if (evt == MVT::Other || !evt.isSimple())
426 VT = evt.getSimpleVT();
428 // This is a legal type, but it's not something we handle in fast-isel.
432 // Handle all other legal types, i.e. a register that will directly hold this
434 return TLI.isTypeLegal(VT);
437 bool AArch64FastISel::isLoadStoreTypeLegal(Type *Ty, MVT &VT) {
438 if (isTypeLegal(Ty, VT))
441 // If this is a type than can be sign or zero-extended to a basic operation
442 // go ahead and accept it now. For stores, this reflects truncation.
443 if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16)
449 bool AArch64FastISel::SimplifyAddress(Address &Addr, MVT VT,
450 int64_t ScaleFactor, bool UseUnscaled) {
451 bool needsLowering = false;
452 int64_t Offset = Addr.getOffset();
453 switch (VT.SimpleTy) {
464 // Using scaled, 12-bit, unsigned immediate offsets.
465 needsLowering = ((Offset & 0xfff) != Offset);
467 // Using unscaled, 9-bit, signed immediate offsets.
468 needsLowering = (Offset > 256 || Offset < -256);
472 // FIXME: If this is a stack pointer and the offset needs to be simplified
473 // then put the alloca address into a register, set the base type back to
474 // register and continue. This should almost never happen.
475 if (needsLowering && Addr.getKind() == Address::FrameIndexBase) {
479 // Since the offset is too large for the load/store instruction get the
480 // reg+offset into a register.
482 uint64_t UnscaledOffset = Addr.getOffset() * ScaleFactor;
483 unsigned ResultReg = FastEmit_ri_(MVT::i64, ISD::ADD, Addr.getReg(), false,
484 UnscaledOffset, MVT::i64);
487 Addr.setReg(ResultReg);
493 void AArch64FastISel::AddLoadStoreOperands(Address &Addr,
494 const MachineInstrBuilder &MIB,
495 unsigned Flags, bool UseUnscaled) {
496 int64_t Offset = Addr.getOffset();
497 // Frame base works a bit differently. Handle it separately.
498 if (Addr.getKind() == Address::FrameIndexBase) {
499 int FI = Addr.getFI();
500 // FIXME: We shouldn't be using getObjectSize/getObjectAlignment. The size
501 // and alignment should be based on the VT.
502 MachineMemOperand *MMO = FuncInfo.MF->getMachineMemOperand(
503 MachinePointerInfo::getFixedStack(FI, Offset), Flags,
504 MFI.getObjectSize(FI), MFI.getObjectAlignment(FI));
505 // Now add the rest of the operands.
506 MIB.addFrameIndex(FI).addImm(Offset).addMemOperand(MMO);
508 // Now add the rest of the operands.
509 MIB.addReg(Addr.getReg());
514 bool AArch64FastISel::EmitLoad(MVT VT, unsigned &ResultReg, Address Addr,
516 // Negative offsets require unscaled, 9-bit, signed immediate offsets.
517 // Otherwise, we try using scaled, 12-bit, unsigned immediate offsets.
518 if (!UseUnscaled && Addr.getOffset() < 0)
522 const TargetRegisterClass *RC;
524 int64_t ScaleFactor = 0;
525 switch (VT.SimpleTy) {
530 // Intentional fall-through.
532 Opc = UseUnscaled ? AArch64::LDURBBi : AArch64::LDRBBui;
533 RC = &AArch64::GPR32RegClass;
537 Opc = UseUnscaled ? AArch64::LDURHHi : AArch64::LDRHHui;
538 RC = &AArch64::GPR32RegClass;
542 Opc = UseUnscaled ? AArch64::LDURWi : AArch64::LDRWui;
543 RC = &AArch64::GPR32RegClass;
547 Opc = UseUnscaled ? AArch64::LDURXi : AArch64::LDRXui;
548 RC = &AArch64::GPR64RegClass;
552 Opc = UseUnscaled ? AArch64::LDURSi : AArch64::LDRSui;
553 RC = TLI.getRegClassFor(VT);
557 Opc = UseUnscaled ? AArch64::LDURDi : AArch64::LDRDui;
558 RC = TLI.getRegClassFor(VT);
564 int64_t Offset = Addr.getOffset();
565 if (Offset & (ScaleFactor - 1))
566 // Retry using an unscaled, 9-bit, signed immediate offset.
567 return EmitLoad(VT, ResultReg, Addr, /*UseUnscaled*/ true);
569 Addr.setOffset(Offset / ScaleFactor);
572 // Simplify this down to something we can handle.
573 if (!SimplifyAddress(Addr, VT, UseUnscaled ? 1 : ScaleFactor, UseUnscaled))
576 // Create the base instruction, then add the operands.
577 ResultReg = createResultReg(RC);
578 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
579 TII.get(Opc), ResultReg);
580 AddLoadStoreOperands(Addr, MIB, MachineMemOperand::MOLoad, UseUnscaled);
582 // Loading an i1 requires special handling.
584 MRI.constrainRegClass(ResultReg, &AArch64::GPR32RegClass);
585 unsigned ANDReg = createResultReg(&AArch64::GPR32spRegClass);
586 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ANDWri),
589 .addImm(AArch64_AM::encodeLogicalImmediate(1, 32));
595 bool AArch64FastISel::SelectLoad(const Instruction *I) {
597 // Verify we have a legal type before going any further. Currently, we handle
598 // simple types that will directly fit in a register (i32/f32/i64/f64) or
599 // those that can be sign or zero-extended to a basic operation (i1/i8/i16).
600 if (!isLoadStoreTypeLegal(I->getType(), VT) || cast<LoadInst>(I)->isAtomic())
603 // See if we can handle this address.
605 if (!ComputeAddress(I->getOperand(0), Addr))
609 if (!EmitLoad(VT, ResultReg, Addr))
612 UpdateValueMap(I, ResultReg);
616 bool AArch64FastISel::EmitStore(MVT VT, unsigned SrcReg, Address Addr,
618 // Negative offsets require unscaled, 9-bit, signed immediate offsets.
619 // Otherwise, we try using scaled, 12-bit, unsigned immediate offsets.
620 if (!UseUnscaled && Addr.getOffset() < 0)
625 int64_t ScaleFactor = 0;
626 // Using scaled, 12-bit, unsigned immediate offsets.
627 switch (VT.SimpleTy) {
633 StrOpc = UseUnscaled ? AArch64::STURBBi : AArch64::STRBBui;
637 StrOpc = UseUnscaled ? AArch64::STURHHi : AArch64::STRHHui;
641 StrOpc = UseUnscaled ? AArch64::STURWi : AArch64::STRWui;
645 StrOpc = UseUnscaled ? AArch64::STURXi : AArch64::STRXui;
649 StrOpc = UseUnscaled ? AArch64::STURSi : AArch64::STRSui;
653 StrOpc = UseUnscaled ? AArch64::STURDi : AArch64::STRDui;
659 int64_t Offset = Addr.getOffset();
660 if (Offset & (ScaleFactor - 1))
661 // Retry using an unscaled, 9-bit, signed immediate offset.
662 return EmitStore(VT, SrcReg, Addr, /*UseUnscaled*/ true);
664 Addr.setOffset(Offset / ScaleFactor);
667 // Simplify this down to something we can handle.
668 if (!SimplifyAddress(Addr, VT, UseUnscaled ? 1 : ScaleFactor, UseUnscaled))
671 // Storing an i1 requires special handling.
673 MRI.constrainRegClass(SrcReg, &AArch64::GPR32RegClass);
674 unsigned ANDReg = createResultReg(&AArch64::GPR32spRegClass);
675 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ANDWri),
678 .addImm(AArch64_AM::encodeLogicalImmediate(1, 32));
681 // Create the base instruction, then add the operands.
682 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
683 TII.get(StrOpc)).addReg(SrcReg);
684 AddLoadStoreOperands(Addr, MIB, MachineMemOperand::MOStore, UseUnscaled);
688 bool AArch64FastISel::SelectStore(const Instruction *I) {
690 Value *Op0 = I->getOperand(0);
691 // Verify we have a legal type before going any further. Currently, we handle
692 // simple types that will directly fit in a register (i32/f32/i64/f64) or
693 // those that can be sign or zero-extended to a basic operation (i1/i8/i16).
694 if (!isLoadStoreTypeLegal(Op0->getType(), VT) ||
695 cast<StoreInst>(I)->isAtomic())
698 // Get the value to be stored into a register.
699 unsigned SrcReg = getRegForValue(Op0);
703 // See if we can handle this address.
705 if (!ComputeAddress(I->getOperand(1), Addr))
708 if (!EmitStore(VT, SrcReg, Addr))
713 static AArch64CC::CondCode getCompareCC(CmpInst::Predicate Pred) {
715 case CmpInst::FCMP_ONE:
716 case CmpInst::FCMP_UEQ:
718 // AL is our "false" for now. The other two need more compares.
719 return AArch64CC::AL;
720 case CmpInst::ICMP_EQ:
721 case CmpInst::FCMP_OEQ:
722 return AArch64CC::EQ;
723 case CmpInst::ICMP_SGT:
724 case CmpInst::FCMP_OGT:
725 return AArch64CC::GT;
726 case CmpInst::ICMP_SGE:
727 case CmpInst::FCMP_OGE:
728 return AArch64CC::GE;
729 case CmpInst::ICMP_UGT:
730 case CmpInst::FCMP_UGT:
731 return AArch64CC::HI;
732 case CmpInst::FCMP_OLT:
733 return AArch64CC::MI;
734 case CmpInst::ICMP_ULE:
735 case CmpInst::FCMP_OLE:
736 return AArch64CC::LS;
737 case CmpInst::FCMP_ORD:
738 return AArch64CC::VC;
739 case CmpInst::FCMP_UNO:
740 return AArch64CC::VS;
741 case CmpInst::FCMP_UGE:
742 return AArch64CC::PL;
743 case CmpInst::ICMP_SLT:
744 case CmpInst::FCMP_ULT:
745 return AArch64CC::LT;
746 case CmpInst::ICMP_SLE:
747 case CmpInst::FCMP_ULE:
748 return AArch64CC::LE;
749 case CmpInst::FCMP_UNE:
750 case CmpInst::ICMP_NE:
751 return AArch64CC::NE;
752 case CmpInst::ICMP_UGE:
753 return AArch64CC::HS;
754 case CmpInst::ICMP_ULT:
755 return AArch64CC::LO;
759 bool AArch64FastISel::SelectBranch(const Instruction *I) {
760 const BranchInst *BI = cast<BranchInst>(I);
761 MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
762 MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
764 if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
765 if (CI->hasOneUse() && (CI->getParent() == I->getParent())) {
766 // We may not handle every CC for now.
767 AArch64CC::CondCode CC = getCompareCC(CI->getPredicate());
768 if (CC == AArch64CC::AL)
772 if (!EmitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
776 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
779 FuncInfo.MBB->addSuccessor(TBB);
781 FastEmitBranch(FBB, DbgLoc);
784 } else if (TruncInst *TI = dyn_cast<TruncInst>(BI->getCondition())) {
786 if (TI->hasOneUse() && TI->getParent() == I->getParent() &&
787 (isLoadStoreTypeLegal(TI->getOperand(0)->getType(), SrcVT))) {
788 unsigned CondReg = getRegForValue(TI->getOperand(0));
792 // Issue an extract_subreg to get the lower 32-bits.
793 if (SrcVT == MVT::i64)
794 CondReg = FastEmitInst_extractsubreg(MVT::i32, CondReg, /*Kill=*/true,
797 MRI.constrainRegClass(CondReg, &AArch64::GPR32RegClass);
798 unsigned ANDReg = createResultReg(&AArch64::GPR32spRegClass);
799 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
800 TII.get(AArch64::ANDWri), ANDReg)
802 .addImm(AArch64_AM::encodeLogicalImmediate(1, 32));
803 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
804 TII.get(AArch64::SUBSWri))
810 unsigned CC = AArch64CC::NE;
811 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
815 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
818 FuncInfo.MBB->addSuccessor(TBB);
819 FastEmitBranch(FBB, DbgLoc);
822 } else if (const ConstantInt *CI =
823 dyn_cast<ConstantInt>(BI->getCondition())) {
824 uint64_t Imm = CI->getZExtValue();
825 MachineBasicBlock *Target = (Imm == 0) ? FBB : TBB;
826 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::B))
828 FuncInfo.MBB->addSuccessor(Target);
832 unsigned CondReg = getRegForValue(BI->getCondition());
836 // We've been divorced from our compare! Our block was split, and
837 // now our compare lives in a predecessor block. We musn't
838 // re-compare here, as the children of the compare aren't guaranteed
839 // live across the block boundary (we *could* check for this).
840 // Regardless, the compare has been done in the predecessor block,
841 // and it left a value for us in a virtual register. Ergo, we test
842 // the one-bit value left in the virtual register.
843 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::SUBSWri),
849 unsigned CC = AArch64CC::NE;
850 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
855 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::Bcc))
858 FuncInfo.MBB->addSuccessor(TBB);
859 FastEmitBranch(FBB, DbgLoc);
863 bool AArch64FastISel::SelectIndirectBr(const Instruction *I) {
864 const IndirectBrInst *BI = cast<IndirectBrInst>(I);
865 unsigned AddrReg = getRegForValue(BI->getOperand(0));
869 // Emit the indirect branch.
870 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::BR))
873 // Make sure the CFG is up-to-date.
874 for (unsigned i = 0, e = BI->getNumSuccessors(); i != e; ++i)
875 FuncInfo.MBB->addSuccessor(FuncInfo.MBBMap[BI->getSuccessor(i)]);
880 bool AArch64FastISel::EmitCmp(Value *Src1Value, Value *Src2Value, bool isZExt) {
881 Type *Ty = Src1Value->getType();
882 EVT SrcEVT = TLI.getValueType(Ty, true);
883 if (!SrcEVT.isSimple())
885 MVT SrcVT = SrcEVT.getSimpleVT();
887 // Check to see if the 2nd operand is a constant that we can encode directly
891 bool isNegativeImm = false;
892 if (const ConstantInt *ConstInt = dyn_cast<ConstantInt>(Src2Value)) {
893 if (SrcVT == MVT::i64 || SrcVT == MVT::i32 || SrcVT == MVT::i16 ||
894 SrcVT == MVT::i8 || SrcVT == MVT::i1) {
895 const APInt &CIVal = ConstInt->getValue();
897 Imm = (isZExt) ? CIVal.getZExtValue() : CIVal.getSExtValue();
898 if (CIVal.isNegative()) {
899 isNegativeImm = true;
902 // FIXME: We can handle more immediates using shifts.
903 UseImm = ((Imm & 0xfff) == Imm);
905 } else if (const ConstantFP *ConstFP = dyn_cast<ConstantFP>(Src2Value)) {
906 if (SrcVT == MVT::f32 || SrcVT == MVT::f64)
907 if (ConstFP->isZero() && !ConstFP->isNegative())
914 bool needsExt = false;
915 switch (SrcVT.SimpleTy) {
922 // Intentional fall-through.
926 CmpOpc = isNegativeImm ? AArch64::ADDSWri : AArch64::SUBSWri;
928 CmpOpc = AArch64::SUBSWrr;
933 CmpOpc = isNegativeImm ? AArch64::ADDSXri : AArch64::SUBSXri;
935 CmpOpc = AArch64::SUBSXrr;
939 CmpOpc = UseImm ? AArch64::FCMPSri : AArch64::FCMPSrr;
943 CmpOpc = UseImm ? AArch64::FCMPDri : AArch64::FCMPDrr;
947 unsigned SrcReg1 = getRegForValue(Src1Value);
953 SrcReg2 = getRegForValue(Src2Value);
958 // We have i1, i8, or i16, we need to either zero extend or sign extend.
960 SrcReg1 = EmitIntExt(SrcVT, SrcReg1, MVT::i32, isZExt);
964 SrcReg2 = EmitIntExt(SrcVT, SrcReg2, MVT::i32, isZExt);
972 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CmpOpc))
978 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CmpOpc))
984 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CmpOpc))
987 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CmpOpc))
994 bool AArch64FastISel::SelectCmp(const Instruction *I) {
995 const CmpInst *CI = cast<CmpInst>(I);
997 // We may not handle every CC for now.
998 AArch64CC::CondCode CC = getCompareCC(CI->getPredicate());
999 if (CC == AArch64CC::AL)
1003 if (!EmitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
1006 // Now set a register based on the comparison.
1007 AArch64CC::CondCode invertedCC = getInvertedCondCode(CC);
1008 unsigned ResultReg = createResultReg(&AArch64::GPR32RegClass);
1009 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::CSINCWr),
1011 .addReg(AArch64::WZR)
1012 .addReg(AArch64::WZR)
1013 .addImm(invertedCC);
1015 UpdateValueMap(I, ResultReg);
1019 bool AArch64FastISel::SelectSelect(const Instruction *I) {
1020 const SelectInst *SI = cast<SelectInst>(I);
1022 EVT DestEVT = TLI.getValueType(SI->getType(), true);
1023 if (!DestEVT.isSimple())
1026 MVT DestVT = DestEVT.getSimpleVT();
1027 if (DestVT != MVT::i32 && DestVT != MVT::i64 && DestVT != MVT::f32 &&
1031 unsigned CondReg = getRegForValue(SI->getCondition());
1034 unsigned TrueReg = getRegForValue(SI->getTrueValue());
1037 unsigned FalseReg = getRegForValue(SI->getFalseValue());
1042 MRI.constrainRegClass(CondReg, &AArch64::GPR32RegClass);
1043 unsigned ANDReg = createResultReg(&AArch64::GPR32spRegClass);
1044 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ANDWri),
1047 .addImm(AArch64_AM::encodeLogicalImmediate(1, 32));
1049 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::SUBSWri))
1056 switch (DestVT.SimpleTy) {
1060 SelectOpc = AArch64::CSELWr;
1063 SelectOpc = AArch64::CSELXr;
1066 SelectOpc = AArch64::FCSELSrrr;
1069 SelectOpc = AArch64::FCSELDrrr;
1073 unsigned ResultReg = createResultReg(TLI.getRegClassFor(DestVT));
1074 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(SelectOpc),
1078 .addImm(AArch64CC::NE);
1080 UpdateValueMap(I, ResultReg);
1084 bool AArch64FastISel::SelectFPExt(const Instruction *I) {
1085 Value *V = I->getOperand(0);
1086 if (!I->getType()->isDoubleTy() || !V->getType()->isFloatTy())
1089 unsigned Op = getRegForValue(V);
1093 unsigned ResultReg = createResultReg(&AArch64::FPR64RegClass);
1094 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::FCVTDSr),
1095 ResultReg).addReg(Op);
1096 UpdateValueMap(I, ResultReg);
1100 bool AArch64FastISel::SelectFPTrunc(const Instruction *I) {
1101 Value *V = I->getOperand(0);
1102 if (!I->getType()->isFloatTy() || !V->getType()->isDoubleTy())
1105 unsigned Op = getRegForValue(V);
1109 unsigned ResultReg = createResultReg(&AArch64::FPR32RegClass);
1110 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::FCVTSDr),
1111 ResultReg).addReg(Op);
1112 UpdateValueMap(I, ResultReg);
1116 // FPToUI and FPToSI
1117 bool AArch64FastISel::SelectFPToInt(const Instruction *I, bool Signed) {
1119 if (!isTypeLegal(I->getType(), DestVT) || DestVT.isVector())
1122 unsigned SrcReg = getRegForValue(I->getOperand(0));
1126 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType(), true);
1127 if (SrcVT == MVT::f128)
1131 if (SrcVT == MVT::f64) {
1133 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZSUWDr : AArch64::FCVTZSUXDr;
1135 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZUUWDr : AArch64::FCVTZUUXDr;
1138 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZSUWSr : AArch64::FCVTZSUXSr;
1140 Opc = (DestVT == MVT::i32) ? AArch64::FCVTZUUWSr : AArch64::FCVTZUUXSr;
1142 unsigned ResultReg = createResultReg(
1143 DestVT == MVT::i32 ? &AArch64::GPR32RegClass : &AArch64::GPR64RegClass);
1144 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
1146 UpdateValueMap(I, ResultReg);
1150 bool AArch64FastISel::SelectIntToFP(const Instruction *I, bool Signed) {
1152 if (!isTypeLegal(I->getType(), DestVT) || DestVT.isVector())
1154 assert ((DestVT == MVT::f32 || DestVT == MVT::f64) &&
1155 "Unexpected value type.");
1157 unsigned SrcReg = getRegForValue(I->getOperand(0));
1161 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType(), true);
1163 // Handle sign-extension.
1164 if (SrcVT == MVT::i16 || SrcVT == MVT::i8 || SrcVT == MVT::i1) {
1166 EmitIntExt(SrcVT.getSimpleVT(), SrcReg, MVT::i32, /*isZExt*/ !Signed);
1171 MRI.constrainRegClass(SrcReg, SrcVT == MVT::i64 ? &AArch64::GPR64RegClass
1172 : &AArch64::GPR32RegClass);
1175 if (SrcVT == MVT::i64) {
1177 Opc = (DestVT == MVT::f32) ? AArch64::SCVTFUXSri : AArch64::SCVTFUXDri;
1179 Opc = (DestVT == MVT::f32) ? AArch64::UCVTFUXSri : AArch64::UCVTFUXDri;
1182 Opc = (DestVT == MVT::f32) ? AArch64::SCVTFUWSri : AArch64::SCVTFUWDri;
1184 Opc = (DestVT == MVT::f32) ? AArch64::UCVTFUWSri : AArch64::UCVTFUWDri;
1187 unsigned ResultReg = createResultReg(TLI.getRegClassFor(DestVT));
1188 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
1190 UpdateValueMap(I, ResultReg);
1194 bool AArch64FastISel::ProcessCallArgs(
1195 SmallVectorImpl<Value *> &Args, SmallVectorImpl<unsigned> &ArgRegs,
1196 SmallVectorImpl<MVT> &ArgVTs, SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
1197 SmallVectorImpl<unsigned> &RegArgs, CallingConv::ID CC,
1198 unsigned &NumBytes) {
1199 SmallVector<CCValAssign, 16> ArgLocs;
1200 CCState CCInfo(CC, false, *FuncInfo.MF, TM, ArgLocs, *Context);
1201 CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CCAssignFnForCall(CC));
1203 // Get a count of how many bytes are to be pushed on the stack.
1204 NumBytes = CCInfo.getNextStackOffset();
1206 // Issue CALLSEQ_START
1207 unsigned AdjStackDown = TII.getCallFrameSetupOpcode();
1208 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackDown))
1211 // Process the args.
1212 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1213 CCValAssign &VA = ArgLocs[i];
1214 unsigned Arg = ArgRegs[VA.getValNo()];
1215 MVT ArgVT = ArgVTs[VA.getValNo()];
1217 // Handle arg promotion: SExt, ZExt, AExt.
1218 switch (VA.getLocInfo()) {
1219 case CCValAssign::Full:
1221 case CCValAssign::SExt: {
1222 MVT DestVT = VA.getLocVT();
1224 Arg = EmitIntExt(SrcVT, Arg, DestVT, /*isZExt*/ false);
1230 case CCValAssign::AExt:
1231 // Intentional fall-through.
1232 case CCValAssign::ZExt: {
1233 MVT DestVT = VA.getLocVT();
1235 Arg = EmitIntExt(SrcVT, Arg, DestVT, /*isZExt*/ true);
1242 llvm_unreachable("Unknown arg promotion!");
1245 // Now copy/store arg to correct locations.
1246 if (VA.isRegLoc() && !VA.needsCustom()) {
1247 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1248 TII.get(TargetOpcode::COPY), VA.getLocReg()).addReg(Arg);
1249 RegArgs.push_back(VA.getLocReg());
1250 } else if (VA.needsCustom()) {
1251 // FIXME: Handle custom args.
1254 assert(VA.isMemLoc() && "Assuming store on stack.");
1256 // Need to store on the stack.
1257 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
1259 unsigned BEAlign = 0;
1260 if (ArgSize < 8 && !Subtarget->isLittleEndian())
1261 BEAlign = 8 - ArgSize;
1264 Addr.setKind(Address::RegBase);
1265 Addr.setReg(AArch64::SP);
1266 Addr.setOffset(VA.getLocMemOffset() + BEAlign);
1268 if (!EmitStore(ArgVT, Arg, Addr))
1275 bool AArch64FastISel::FinishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
1276 const Instruction *I, CallingConv::ID CC,
1277 unsigned &NumBytes) {
1278 // Issue CALLSEQ_END
1279 unsigned AdjStackUp = TII.getCallFrameDestroyOpcode();
1280 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackUp))
1284 // Now the return value.
1285 if (RetVT != MVT::isVoid) {
1286 SmallVector<CCValAssign, 16> RVLocs;
1287 CCState CCInfo(CC, false, *FuncInfo.MF, TM, RVLocs, *Context);
1288 CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC));
1290 // Only handle a single return value.
1291 if (RVLocs.size() != 1)
1294 // Copy all of the result registers out of their specified physreg.
1295 MVT CopyVT = RVLocs[0].getValVT();
1296 unsigned ResultReg = createResultReg(TLI.getRegClassFor(CopyVT));
1297 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1298 TII.get(TargetOpcode::COPY),
1299 ResultReg).addReg(RVLocs[0].getLocReg());
1300 UsedRegs.push_back(RVLocs[0].getLocReg());
1302 // Finally update the result.
1303 UpdateValueMap(I, ResultReg);
1309 bool AArch64FastISel::SelectCall(const Instruction *I,
1310 const char *IntrMemName = nullptr) {
1311 const CallInst *CI = cast<CallInst>(I);
1312 const Value *Callee = CI->getCalledValue();
1314 // Don't handle inline asm or intrinsics.
1315 if (isa<InlineAsm>(Callee))
1318 // Only handle global variable Callees.
1319 const GlobalValue *GV = dyn_cast<GlobalValue>(Callee);
1323 // Check the calling convention.
1324 ImmutableCallSite CS(CI);
1325 CallingConv::ID CC = CS.getCallingConv();
1327 // Let SDISel handle vararg functions.
1328 PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
1329 FunctionType *FTy = cast<FunctionType>(PT->getElementType());
1330 if (FTy->isVarArg())
1333 // Handle *simple* calls for now.
1335 Type *RetTy = I->getType();
1336 if (RetTy->isVoidTy())
1337 RetVT = MVT::isVoid;
1338 else if (!isTypeLegal(RetTy, RetVT))
1341 // Set up the argument vectors.
1342 SmallVector<Value *, 8> Args;
1343 SmallVector<unsigned, 8> ArgRegs;
1344 SmallVector<MVT, 8> ArgVTs;
1345 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
1346 Args.reserve(CS.arg_size());
1347 ArgRegs.reserve(CS.arg_size());
1348 ArgVTs.reserve(CS.arg_size());
1349 ArgFlags.reserve(CS.arg_size());
1351 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
1353 // If we're lowering a memory intrinsic instead of a regular call, skip the
1354 // last two arguments, which shouldn't be passed to the underlying function.
1355 if (IntrMemName && e - i <= 2)
1358 unsigned Arg = getRegForValue(*i);
1362 ISD::ArgFlagsTy Flags;
1363 unsigned AttrInd = i - CS.arg_begin() + 1;
1364 if (CS.paramHasAttr(AttrInd, Attribute::SExt))
1366 if (CS.paramHasAttr(AttrInd, Attribute::ZExt))
1369 // FIXME: Only handle *easy* calls for now.
1370 if (CS.paramHasAttr(AttrInd, Attribute::InReg) ||
1371 CS.paramHasAttr(AttrInd, Attribute::StructRet) ||
1372 CS.paramHasAttr(AttrInd, Attribute::Nest) ||
1373 CS.paramHasAttr(AttrInd, Attribute::ByVal))
1377 Type *ArgTy = (*i)->getType();
1378 if (!isTypeLegal(ArgTy, ArgVT) &&
1379 !(ArgVT == MVT::i1 || ArgVT == MVT::i8 || ArgVT == MVT::i16))
1382 // We don't handle vector parameters yet.
1383 if (ArgVT.isVector() || ArgVT.getSizeInBits() > 64)
1386 unsigned OriginalAlignment = DL.getABITypeAlignment(ArgTy);
1387 Flags.setOrigAlign(OriginalAlignment);
1390 ArgRegs.push_back(Arg);
1391 ArgVTs.push_back(ArgVT);
1392 ArgFlags.push_back(Flags);
1395 // Handle the arguments now that we've gotten them.
1396 SmallVector<unsigned, 4> RegArgs;
1398 if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
1402 MachineInstrBuilder MIB;
1403 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::BL));
1405 MIB.addGlobalAddress(GV, 0, 0);
1407 MIB.addExternalSymbol(IntrMemName, 0);
1409 // Add implicit physical register uses to the call.
1410 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
1411 MIB.addReg(RegArgs[i], RegState::Implicit);
1413 // Add a register mask with the call-preserved registers.
1414 // Proper defs for return values will be added by setPhysRegsDeadExcept().
1415 MIB.addRegMask(TRI.getCallPreservedMask(CS.getCallingConv()));
1417 // Finish off the call including any return values.
1418 SmallVector<unsigned, 4> UsedRegs;
1419 if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes))
1422 // Set all unused physreg defs as dead.
1423 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
1428 bool AArch64FastISel::IsMemCpySmall(uint64_t Len, unsigned Alignment) {
1430 return Len / Alignment <= 4;
1435 bool AArch64FastISel::TryEmitSmallMemCpy(Address Dest, Address Src,
1436 uint64_t Len, unsigned Alignment) {
1437 // Make sure we don't bloat code by inlining very large memcpy's.
1438 if (!IsMemCpySmall(Len, Alignment))
1441 int64_t UnscaledOffset = 0;
1442 Address OrigDest = Dest;
1443 Address OrigSrc = Src;
1447 if (!Alignment || Alignment >= 8) {
1458 // Bound based on alignment.
1459 if (Len >= 4 && Alignment == 4)
1461 else if (Len >= 2 && Alignment == 2)
1470 RV = EmitLoad(VT, ResultReg, Src);
1471 assert(RV == true && "Should be able to handle this load.");
1472 RV = EmitStore(VT, ResultReg, Dest);
1473 assert(RV == true && "Should be able to handle this store.");
1476 int64_t Size = VT.getSizeInBits() / 8;
1478 UnscaledOffset += Size;
1480 // We need to recompute the unscaled offset for each iteration.
1481 Dest.setOffset(OrigDest.getOffset() + UnscaledOffset);
1482 Src.setOffset(OrigSrc.getOffset() + UnscaledOffset);
1488 bool AArch64FastISel::SelectIntrinsicCall(const IntrinsicInst &I) {
1489 // FIXME: Handle more intrinsics.
1490 switch (I.getIntrinsicID()) {
1493 case Intrinsic::memcpy:
1494 case Intrinsic::memmove: {
1495 const MemTransferInst &MTI = cast<MemTransferInst>(I);
1496 // Don't handle volatile.
1497 if (MTI.isVolatile())
1500 // Disable inlining for memmove before calls to ComputeAddress. Otherwise,
1501 // we would emit dead code because we don't currently handle memmoves.
1502 bool isMemCpy = (I.getIntrinsicID() == Intrinsic::memcpy);
1503 if (isa<ConstantInt>(MTI.getLength()) && isMemCpy) {
1504 // Small memcpy's are common enough that we want to do them without a call
1506 uint64_t Len = cast<ConstantInt>(MTI.getLength())->getZExtValue();
1507 unsigned Alignment = MTI.getAlignment();
1508 if (IsMemCpySmall(Len, Alignment)) {
1510 if (!ComputeAddress(MTI.getRawDest(), Dest) ||
1511 !ComputeAddress(MTI.getRawSource(), Src))
1513 if (TryEmitSmallMemCpy(Dest, Src, Len, Alignment))
1518 if (!MTI.getLength()->getType()->isIntegerTy(64))
1521 if (MTI.getSourceAddressSpace() > 255 || MTI.getDestAddressSpace() > 255)
1522 // Fast instruction selection doesn't support the special
1526 const char *IntrMemName = isa<MemCpyInst>(I) ? "memcpy" : "memmove";
1527 return SelectCall(&I, IntrMemName);
1529 case Intrinsic::memset: {
1530 const MemSetInst &MSI = cast<MemSetInst>(I);
1531 // Don't handle volatile.
1532 if (MSI.isVolatile())
1535 if (!MSI.getLength()->getType()->isIntegerTy(64))
1538 if (MSI.getDestAddressSpace() > 255)
1539 // Fast instruction selection doesn't support the special
1543 return SelectCall(&I, "memset");
1545 case Intrinsic::trap: {
1546 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::BRK))
1554 bool AArch64FastISel::SelectRet(const Instruction *I) {
1555 const ReturnInst *Ret = cast<ReturnInst>(I);
1556 const Function &F = *I->getParent()->getParent();
1558 if (!FuncInfo.CanLowerReturn)
1564 // Build a list of return value registers.
1565 SmallVector<unsigned, 4> RetRegs;
1567 if (Ret->getNumOperands() > 0) {
1568 CallingConv::ID CC = F.getCallingConv();
1569 SmallVector<ISD::OutputArg, 4> Outs;
1570 GetReturnInfo(F.getReturnType(), F.getAttributes(), Outs, TLI);
1572 // Analyze operands of the call, assigning locations to each operand.
1573 SmallVector<CCValAssign, 16> ValLocs;
1574 CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, TM, ValLocs,
1576 CCAssignFn *RetCC = CC == CallingConv::WebKit_JS ? RetCC_AArch64_WebKit_JS
1577 : RetCC_AArch64_AAPCS;
1578 CCInfo.AnalyzeReturn(Outs, RetCC);
1580 // Only handle a single return value for now.
1581 if (ValLocs.size() != 1)
1584 CCValAssign &VA = ValLocs[0];
1585 const Value *RV = Ret->getOperand(0);
1587 // Don't bother handling odd stuff for now.
1588 if (VA.getLocInfo() != CCValAssign::Full)
1590 // Only handle register returns for now.
1593 unsigned Reg = getRegForValue(RV);
1597 unsigned SrcReg = Reg + VA.getValNo();
1598 unsigned DestReg = VA.getLocReg();
1599 // Avoid a cross-class copy. This is very unlikely.
1600 if (!MRI.getRegClass(SrcReg)->contains(DestReg))
1603 EVT RVEVT = TLI.getValueType(RV->getType());
1604 if (!RVEVT.isSimple())
1607 // Vectors (of > 1 lane) in big endian need tricky handling.
1608 if (RVEVT.isVector() && RVEVT.getVectorNumElements() > 1)
1611 MVT RVVT = RVEVT.getSimpleVT();
1612 if (RVVT == MVT::f128)
1614 MVT DestVT = VA.getValVT();
1615 // Special handling for extended integers.
1616 if (RVVT != DestVT) {
1617 if (RVVT != MVT::i1 && RVVT != MVT::i8 && RVVT != MVT::i16)
1620 if (!Outs[0].Flags.isZExt() && !Outs[0].Flags.isSExt())
1623 bool isZExt = Outs[0].Flags.isZExt();
1624 SrcReg = EmitIntExt(RVVT, SrcReg, DestVT, isZExt);
1630 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1631 TII.get(TargetOpcode::COPY), DestReg).addReg(SrcReg);
1633 // Add register to return instruction.
1634 RetRegs.push_back(VA.getLocReg());
1637 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1638 TII.get(AArch64::RET_ReallyLR));
1639 for (unsigned i = 0, e = RetRegs.size(); i != e; ++i)
1640 MIB.addReg(RetRegs[i], RegState::Implicit);
1644 bool AArch64FastISel::SelectTrunc(const Instruction *I) {
1645 Type *DestTy = I->getType();
1646 Value *Op = I->getOperand(0);
1647 Type *SrcTy = Op->getType();
1649 EVT SrcEVT = TLI.getValueType(SrcTy, true);
1650 EVT DestEVT = TLI.getValueType(DestTy, true);
1651 if (!SrcEVT.isSimple())
1653 if (!DestEVT.isSimple())
1656 MVT SrcVT = SrcEVT.getSimpleVT();
1657 MVT DestVT = DestEVT.getSimpleVT();
1659 if (SrcVT != MVT::i64 && SrcVT != MVT::i32 && SrcVT != MVT::i16 &&
1662 if (DestVT != MVT::i32 && DestVT != MVT::i16 && DestVT != MVT::i8 &&
1666 unsigned SrcReg = getRegForValue(Op);
1670 // If we're truncating from i64 to a smaller non-legal type then generate an
1671 // AND. Otherwise, we know the high bits are undefined and a truncate doesn't
1672 // generate any code.
1673 if (SrcVT == MVT::i64) {
1675 switch (DestVT.SimpleTy) {
1677 // Trunc i64 to i32 is handled by the target-independent fast-isel.
1689 // Issue an extract_subreg to get the lower 32-bits.
1690 unsigned Reg32 = FastEmitInst_extractsubreg(MVT::i32, SrcReg, /*Kill=*/true,
1692 MRI.constrainRegClass(Reg32, &AArch64::GPR32RegClass);
1693 // Create the AND instruction which performs the actual truncation.
1694 unsigned ANDReg = createResultReg(&AArch64::GPR32spRegClass);
1695 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ANDWri),
1698 .addImm(AArch64_AM::encodeLogicalImmediate(Mask, 32));
1702 UpdateValueMap(I, SrcReg);
1706 unsigned AArch64FastISel::Emiti1Ext(unsigned SrcReg, MVT DestVT, bool isZExt) {
1707 assert((DestVT == MVT::i8 || DestVT == MVT::i16 || DestVT == MVT::i32 ||
1708 DestVT == MVT::i64) &&
1709 "Unexpected value type.");
1710 // Handle i8 and i16 as i32.
1711 if (DestVT == MVT::i8 || DestVT == MVT::i16)
1715 MRI.constrainRegClass(SrcReg, &AArch64::GPR32RegClass);
1716 unsigned ResultReg = createResultReg(&AArch64::GPR32spRegClass);
1717 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ANDWri),
1720 .addImm(AArch64_AM::encodeLogicalImmediate(1, 32));
1722 if (DestVT == MVT::i64) {
1723 // We're ZExt i1 to i64. The ANDWri Wd, Ws, #1 implicitly clears the
1724 // upper 32 bits. Emit a SUBREG_TO_REG to extend from Wd to Xd.
1725 unsigned Reg64 = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
1726 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1727 TII.get(AArch64::SUBREG_TO_REG), Reg64)
1730 .addImm(AArch64::sub_32);
1735 if (DestVT == MVT::i64) {
1736 // FIXME: We're SExt i1 to i64.
1739 unsigned ResultReg = createResultReg(&AArch64::GPR32RegClass);
1740 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::SBFMWri),
1749 unsigned AArch64FastISel::EmitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT,
1751 assert(DestVT != MVT::i1 && "ZeroExt/SignExt an i1?");
1755 switch (SrcVT.SimpleTy) {
1759 return Emiti1Ext(SrcReg, DestVT, isZExt);
1761 if (DestVT == MVT::i64)
1762 Opc = isZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
1764 Opc = isZExt ? AArch64::UBFMWri : AArch64::SBFMWri;
1768 if (DestVT == MVT::i64)
1769 Opc = isZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
1771 Opc = isZExt ? AArch64::UBFMWri : AArch64::SBFMWri;
1775 assert(DestVT == MVT::i64 && "IntExt i32 to i32?!?");
1776 Opc = isZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
1781 // Handle i8 and i16 as i32.
1782 if (DestVT == MVT::i8 || DestVT == MVT::i16)
1784 else if (DestVT == MVT::i64) {
1785 unsigned Src64 = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
1786 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1787 TII.get(AArch64::SUBREG_TO_REG), Src64)
1790 .addImm(AArch64::sub_32);
1794 unsigned ResultReg = createResultReg(TLI.getRegClassFor(DestVT));
1795 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
1803 bool AArch64FastISel::SelectIntExt(const Instruction *I) {
1804 // On ARM, in general, integer casts don't involve legal types; this code
1805 // handles promotable integers. The high bits for a type smaller than
1806 // the register size are assumed to be undefined.
1807 Type *DestTy = I->getType();
1808 Value *Src = I->getOperand(0);
1809 Type *SrcTy = Src->getType();
1811 bool isZExt = isa<ZExtInst>(I);
1812 unsigned SrcReg = getRegForValue(Src);
1816 EVT SrcEVT = TLI.getValueType(SrcTy, true);
1817 EVT DestEVT = TLI.getValueType(DestTy, true);
1818 if (!SrcEVT.isSimple())
1820 if (!DestEVT.isSimple())
1823 MVT SrcVT = SrcEVT.getSimpleVT();
1824 MVT DestVT = DestEVT.getSimpleVT();
1825 unsigned ResultReg = EmitIntExt(SrcVT, SrcReg, DestVT, isZExt);
1828 UpdateValueMap(I, ResultReg);
1832 bool AArch64FastISel::SelectRem(const Instruction *I, unsigned ISDOpcode) {
1833 EVT DestEVT = TLI.getValueType(I->getType(), true);
1834 if (!DestEVT.isSimple())
1837 MVT DestVT = DestEVT.getSimpleVT();
1838 if (DestVT != MVT::i64 && DestVT != MVT::i32)
1842 bool is64bit = (DestVT == MVT::i64);
1843 switch (ISDOpcode) {
1847 DivOpc = is64bit ? AArch64::SDIVXr : AArch64::SDIVWr;
1850 DivOpc = is64bit ? AArch64::UDIVXr : AArch64::UDIVWr;
1853 unsigned MSubOpc = is64bit ? AArch64::MSUBXrrr : AArch64::MSUBWrrr;
1854 unsigned Src0Reg = getRegForValue(I->getOperand(0));
1858 unsigned Src1Reg = getRegForValue(I->getOperand(1));
1862 unsigned QuotReg = createResultReg(TLI.getRegClassFor(DestVT));
1863 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(DivOpc), QuotReg)
1866 // The remainder is computed as numerator - (quotient * denominator) using the
1867 // MSUB instruction.
1868 unsigned ResultReg = createResultReg(TLI.getRegClassFor(DestVT));
1869 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(MSubOpc), ResultReg)
1873 UpdateValueMap(I, ResultReg);
1877 bool AArch64FastISel::SelectMul(const Instruction *I) {
1878 EVT SrcEVT = TLI.getValueType(I->getOperand(0)->getType(), true);
1879 if (!SrcEVT.isSimple())
1881 MVT SrcVT = SrcEVT.getSimpleVT();
1883 // Must be simple value type. Don't handle vectors.
1884 if (SrcVT != MVT::i64 && SrcVT != MVT::i32 && SrcVT != MVT::i16 &&
1890 switch (SrcVT.SimpleTy) {
1896 ZReg = AArch64::WZR;
1897 Opc = AArch64::MADDWrrr;
1900 ZReg = AArch64::XZR;
1901 Opc = AArch64::MADDXrrr;
1905 unsigned Src0Reg = getRegForValue(I->getOperand(0));
1909 unsigned Src1Reg = getRegForValue(I->getOperand(1));
1913 // Create the base instruction, then add the operands.
1914 unsigned ResultReg = createResultReg(TLI.getRegClassFor(SrcVT));
1915 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
1919 UpdateValueMap(I, ResultReg);
1923 bool AArch64FastISel::TargetSelectInstruction(const Instruction *I) {
1924 switch (I->getOpcode()) {
1927 case Instruction::Load:
1928 return SelectLoad(I);
1929 case Instruction::Store:
1930 return SelectStore(I);
1931 case Instruction::Br:
1932 return SelectBranch(I);
1933 case Instruction::IndirectBr:
1934 return SelectIndirectBr(I);
1935 case Instruction::FCmp:
1936 case Instruction::ICmp:
1937 return SelectCmp(I);
1938 case Instruction::Select:
1939 return SelectSelect(I);
1940 case Instruction::FPExt:
1941 return SelectFPExt(I);
1942 case Instruction::FPTrunc:
1943 return SelectFPTrunc(I);
1944 case Instruction::FPToSI:
1945 return SelectFPToInt(I, /*Signed=*/true);
1946 case Instruction::FPToUI:
1947 return SelectFPToInt(I, /*Signed=*/false);
1948 case Instruction::SIToFP:
1949 return SelectIntToFP(I, /*Signed=*/true);
1950 case Instruction::UIToFP:
1951 return SelectIntToFP(I, /*Signed=*/false);
1952 case Instruction::SRem:
1953 return SelectRem(I, ISD::SREM);
1954 case Instruction::URem:
1955 return SelectRem(I, ISD::UREM);
1956 case Instruction::Call:
1957 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1958 return SelectIntrinsicCall(*II);
1959 return SelectCall(I);
1960 case Instruction::Ret:
1961 return SelectRet(I);
1962 case Instruction::Trunc:
1963 return SelectTrunc(I);
1964 case Instruction::ZExt:
1965 case Instruction::SExt:
1966 return SelectIntExt(I);
1967 case Instruction::Mul:
1968 // FIXME: This really should be handled by the target-independent selector.
1969 return SelectMul(I);
1972 // Silence warnings.
1973 (void)&CC_AArch64_DarwinPCS_VarArg;
1977 llvm::FastISel *AArch64::createFastISel(FunctionLoweringInfo &funcInfo,
1978 const TargetLibraryInfo *libInfo) {
1979 return new AArch64FastISel(funcInfo, libInfo);