1 //===-- ARMFastISel.cpp - ARM 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 ARM-specific support for the FastISel class. Some
11 // of the target-specific code is generated by tablegen in the file
12 // ARMGenFastISel.inc, which is #included here.
14 //===----------------------------------------------------------------------===//
17 #include "ARMBaseInstrInfo.h"
18 #include "ARMBaseRegisterInfo.h"
19 #include "ARMCallingConv.h"
20 #include "ARMConstantPoolValue.h"
21 #include "ARMISelLowering.h"
22 #include "ARMMachineFunctionInfo.h"
23 #include "ARMSubtarget.h"
24 #include "MCTargetDesc/ARMAddressingModes.h"
25 #include "MCTargetDesc/ARMBaseInfo.h"
26 #include "llvm/ADT/APFloat.h"
27 #include "llvm/ADT/APInt.h"
28 #include "llvm/ADT/DenseMap.h"
29 #include "llvm/ADT/STLExtras.h"
30 #include "llvm/ADT/SmallVector.h"
31 #include "llvm/CodeGen/CallingConvLower.h"
32 #include "llvm/CodeGen/FastISel.h"
33 #include "llvm/CodeGen/FunctionLoweringInfo.h"
34 #include "llvm/CodeGen/ISDOpcodes.h"
35 #include "llvm/CodeGen/MachineConstantPool.h"
36 #include "llvm/CodeGen/MachineFrameInfo.h"
37 #include "llvm/CodeGen/MachineInstr.h"
38 #include "llvm/CodeGen/MachineInstrBuilder.h"
39 #include "llvm/CodeGen/MachineMemOperand.h"
40 #include "llvm/CodeGen/MachineOperand.h"
41 #include "llvm/CodeGen/MachineRegisterInfo.h"
42 #include "llvm/CodeGen/MachineValueType.h"
43 #include "llvm/CodeGen/RuntimeLibcalls.h"
44 #include "llvm/CodeGen/ValueTypes.h"
45 #include "llvm/IR/Argument.h"
46 #include "llvm/IR/Attributes.h"
47 #include "llvm/IR/CallSite.h"
48 #include "llvm/IR/CallingConv.h"
49 #include "llvm/IR/Constant.h"
50 #include "llvm/IR/Constants.h"
51 #include "llvm/IR/DataLayout.h"
52 #include "llvm/IR/DerivedTypes.h"
53 #include "llvm/IR/Function.h"
54 #include "llvm/IR/GetElementPtrTypeIterator.h"
55 #include "llvm/IR/GlobalValue.h"
56 #include "llvm/IR/GlobalVariable.h"
57 #include "llvm/IR/InstrTypes.h"
58 #include "llvm/IR/Instruction.h"
59 #include "llvm/IR/Instructions.h"
60 #include "llvm/IR/IntrinsicInst.h"
61 #include "llvm/IR/Module.h"
62 #include "llvm/IR/Operator.h"
63 #include "llvm/IR/Type.h"
64 #include "llvm/IR/User.h"
65 #include "llvm/IR/Value.h"
66 #include "llvm/MC/MCInstrDesc.h"
67 #include "llvm/MC/MCRegisterInfo.h"
68 #include "llvm/Support/Casting.h"
69 #include "llvm/Support/Compiler.h"
70 #include "llvm/Support/ErrorHandling.h"
71 #include "llvm/Support/MathExtras.h"
72 #include "llvm/Target/TargetInstrInfo.h"
73 #include "llvm/Target/TargetLowering.h"
74 #include "llvm/Target/TargetMachine.h"
75 #include "llvm/Target/TargetOptions.h"
84 // All possible address modes, plus some.
85 typedef struct Address {
98 // Innocuous defaults for our address.
104 class ARMFastISel final : public FastISel {
105 /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
106 /// make the right decision when generating code for different targets.
107 const ARMSubtarget *Subtarget;
109 const TargetMachine &TM;
110 const TargetInstrInfo &TII;
111 const TargetLowering &TLI;
112 ARMFunctionInfo *AFI;
114 // Convenience variables to avoid some queries.
116 LLVMContext *Context;
119 explicit ARMFastISel(FunctionLoweringInfo &funcInfo,
120 const TargetLibraryInfo *libInfo)
121 : FastISel(funcInfo, libInfo),
123 &static_cast<const ARMSubtarget &>(funcInfo.MF->getSubtarget())),
124 M(const_cast<Module &>(*funcInfo.Fn->getParent())),
125 TM(funcInfo.MF->getTarget()), TII(*Subtarget->getInstrInfo()),
126 TLI(*Subtarget->getTargetLowering()) {
127 AFI = funcInfo.MF->getInfo<ARMFunctionInfo>();
128 isThumb2 = AFI->isThumbFunction();
129 Context = &funcInfo.Fn->getContext();
133 // Code from FastISel.cpp.
135 unsigned fastEmitInst_r(unsigned MachineInstOpcode,
136 const TargetRegisterClass *RC,
137 unsigned Op0, bool Op0IsKill);
138 unsigned fastEmitInst_rr(unsigned MachineInstOpcode,
139 const TargetRegisterClass *RC,
140 unsigned Op0, bool Op0IsKill,
141 unsigned Op1, bool Op1IsKill);
142 unsigned fastEmitInst_ri(unsigned MachineInstOpcode,
143 const TargetRegisterClass *RC,
144 unsigned Op0, bool Op0IsKill,
146 unsigned fastEmitInst_i(unsigned MachineInstOpcode,
147 const TargetRegisterClass *RC,
150 // Backend specific FastISel code.
152 bool fastSelectInstruction(const Instruction *I) override;
153 unsigned fastMaterializeConstant(const Constant *C) override;
154 unsigned fastMaterializeAlloca(const AllocaInst *AI) override;
155 bool tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo,
156 const LoadInst *LI) override;
157 bool fastLowerArguments() override;
159 #include "ARMGenFastISel.inc"
161 // Instruction selection routines.
163 bool SelectLoad(const Instruction *I);
164 bool SelectStore(const Instruction *I);
165 bool SelectBranch(const Instruction *I);
166 bool SelectIndirectBr(const Instruction *I);
167 bool SelectCmp(const Instruction *I);
168 bool SelectFPExt(const Instruction *I);
169 bool SelectFPTrunc(const Instruction *I);
170 bool SelectBinaryIntOp(const Instruction *I, unsigned ISDOpcode);
171 bool SelectBinaryFPOp(const Instruction *I, unsigned ISDOpcode);
172 bool SelectIToFP(const Instruction *I, bool isSigned);
173 bool SelectFPToI(const Instruction *I, bool isSigned);
174 bool SelectDiv(const Instruction *I, bool isSigned);
175 bool SelectRem(const Instruction *I, bool isSigned);
176 bool SelectCall(const Instruction *I, const char *IntrMemName);
177 bool SelectIntrinsicCall(const IntrinsicInst &I);
178 bool SelectSelect(const Instruction *I);
179 bool SelectRet(const Instruction *I);
180 bool SelectTrunc(const Instruction *I);
181 bool SelectIntExt(const Instruction *I);
182 bool SelectShift(const Instruction *I, ARM_AM::ShiftOpc ShiftTy);
186 bool isPositionIndependent() const;
187 bool isTypeLegal(Type *Ty, MVT &VT);
188 bool isLoadTypeLegal(Type *Ty, MVT &VT);
189 bool ARMEmitCmp(const Value *Src1Value, const Value *Src2Value,
190 bool isZExt, bool isEquality);
191 bool ARMEmitLoad(MVT VT, unsigned &ResultReg, Address &Addr,
192 unsigned Alignment = 0, bool isZExt = true,
193 bool allocReg = true);
194 bool ARMEmitStore(MVT VT, unsigned SrcReg, Address &Addr,
195 unsigned Alignment = 0);
196 bool ARMComputeAddress(const Value *Obj, Address &Addr);
197 void ARMSimplifyAddress(Address &Addr, MVT VT, bool useAM3);
198 bool ARMIsMemCpySmall(uint64_t Len);
199 bool ARMTryEmitSmallMemCpy(Address Dest, Address Src, uint64_t Len,
201 unsigned ARMEmitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, bool isZExt);
202 unsigned ARMMaterializeFP(const ConstantFP *CFP, MVT VT);
203 unsigned ARMMaterializeInt(const Constant *C, MVT VT);
204 unsigned ARMMaterializeGV(const GlobalValue *GV, MVT VT);
205 unsigned ARMMoveToFPReg(MVT VT, unsigned SrcReg);
206 unsigned ARMMoveToIntReg(MVT VT, unsigned SrcReg);
207 unsigned ARMSelectCallOp(bool UseReg);
208 unsigned ARMLowerPICELF(const GlobalValue *GV, unsigned Align, MVT VT);
210 const TargetLowering *getTargetLowering() { return &TLI; }
212 // Call handling routines.
214 CCAssignFn *CCAssignFnForCall(CallingConv::ID CC,
217 bool ProcessCallArgs(SmallVectorImpl<Value*> &Args,
218 SmallVectorImpl<unsigned> &ArgRegs,
219 SmallVectorImpl<MVT> &ArgVTs,
220 SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
221 SmallVectorImpl<unsigned> &RegArgs,
225 unsigned getLibcallReg(const Twine &Name);
226 bool FinishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
227 const Instruction *I, CallingConv::ID CC,
228 unsigned &NumBytes, bool isVarArg);
229 bool ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call);
231 // OptionalDef handling routines.
233 bool isARMNEONPred(const MachineInstr *MI);
234 bool DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR);
235 const MachineInstrBuilder &AddOptionalDefs(const MachineInstrBuilder &MIB);
236 void AddLoadStoreOperands(MVT VT, Address &Addr,
237 const MachineInstrBuilder &MIB,
238 MachineMemOperand::Flags Flags, bool useAM3);
241 } // end anonymous namespace
243 #include "ARMGenCallingConv.inc"
245 // DefinesOptionalPredicate - This is different from DefinesPredicate in that
246 // we don't care about implicit defs here, just places we'll need to add a
247 // default CCReg argument. Sets CPSR if we're setting CPSR instead of CCR.
248 bool ARMFastISel::DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR) {
249 if (!MI->hasOptionalDef())
252 // Look to see if our OptionalDef is defining CPSR or CCR.
253 for (const MachineOperand &MO : MI->operands()) {
254 if (!MO.isReg() || !MO.isDef()) continue;
255 if (MO.getReg() == ARM::CPSR)
261 bool ARMFastISel::isARMNEONPred(const MachineInstr *MI) {
262 const MCInstrDesc &MCID = MI->getDesc();
264 // If we're a thumb2 or not NEON function we'll be handled via isPredicable.
265 if ((MCID.TSFlags & ARMII::DomainMask) != ARMII::DomainNEON ||
266 AFI->isThumb2Function())
267 return MI->isPredicable();
269 for (const MCOperandInfo &opInfo : MCID.operands())
270 if (opInfo.isPredicate())
276 // If the machine is predicable go ahead and add the predicate operands, if
277 // it needs default CC operands add those.
278 // TODO: If we want to support thumb1 then we'll need to deal with optional
279 // CPSR defs that need to be added before the remaining operands. See s_cc_out
280 // for descriptions why.
281 const MachineInstrBuilder &
282 ARMFastISel::AddOptionalDefs(const MachineInstrBuilder &MIB) {
283 MachineInstr *MI = &*MIB;
285 // Do we use a predicate? or...
286 // Are we NEON in ARM mode and have a predicate operand? If so, I know
287 // we're not predicable but add it anyways.
288 if (isARMNEONPred(MI))
289 MIB.add(predOps(ARMCC::AL));
291 // Do we optionally set a predicate? Preds is size > 0 iff the predicate
292 // defines CPSR. All other OptionalDefines in ARM are the CCR register.
294 if (DefinesOptionalPredicate(MI, &CPSR))
295 MIB.add(CPSR ? t1CondCodeOp() : condCodeOp());
299 unsigned ARMFastISel::fastEmitInst_r(unsigned MachineInstOpcode,
300 const TargetRegisterClass *RC,
301 unsigned Op0, bool Op0IsKill) {
302 unsigned ResultReg = createResultReg(RC);
303 const MCInstrDesc &II = TII.get(MachineInstOpcode);
305 // Make sure the input operand is sufficiently constrained to be legal
306 // for this instruction.
307 Op0 = constrainOperandRegClass(II, Op0, 1);
308 if (II.getNumDefs() >= 1) {
309 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II,
310 ResultReg).addReg(Op0, Op0IsKill * RegState::Kill));
312 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
313 .addReg(Op0, Op0IsKill * RegState::Kill));
314 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
315 TII.get(TargetOpcode::COPY), ResultReg)
316 .addReg(II.ImplicitDefs[0]));
321 unsigned ARMFastISel::fastEmitInst_rr(unsigned MachineInstOpcode,
322 const TargetRegisterClass *RC,
323 unsigned Op0, bool Op0IsKill,
324 unsigned Op1, bool Op1IsKill) {
325 unsigned ResultReg = createResultReg(RC);
326 const MCInstrDesc &II = TII.get(MachineInstOpcode);
328 // Make sure the input operands are sufficiently constrained to be legal
329 // for this instruction.
330 Op0 = constrainOperandRegClass(II, Op0, 1);
331 Op1 = constrainOperandRegClass(II, Op1, 2);
333 if (II.getNumDefs() >= 1) {
335 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
336 .addReg(Op0, Op0IsKill * RegState::Kill)
337 .addReg(Op1, Op1IsKill * RegState::Kill));
339 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
340 .addReg(Op0, Op0IsKill * RegState::Kill)
341 .addReg(Op1, Op1IsKill * RegState::Kill));
342 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
343 TII.get(TargetOpcode::COPY), ResultReg)
344 .addReg(II.ImplicitDefs[0]));
349 unsigned ARMFastISel::fastEmitInst_ri(unsigned MachineInstOpcode,
350 const TargetRegisterClass *RC,
351 unsigned Op0, bool Op0IsKill,
353 unsigned ResultReg = createResultReg(RC);
354 const MCInstrDesc &II = TII.get(MachineInstOpcode);
356 // Make sure the input operand is sufficiently constrained to be legal
357 // for this instruction.
358 Op0 = constrainOperandRegClass(II, Op0, 1);
359 if (II.getNumDefs() >= 1) {
361 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
362 .addReg(Op0, Op0IsKill * RegState::Kill)
365 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
366 .addReg(Op0, Op0IsKill * RegState::Kill)
368 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
369 TII.get(TargetOpcode::COPY), ResultReg)
370 .addReg(II.ImplicitDefs[0]));
375 unsigned ARMFastISel::fastEmitInst_i(unsigned MachineInstOpcode,
376 const TargetRegisterClass *RC,
378 unsigned ResultReg = createResultReg(RC);
379 const MCInstrDesc &II = TII.get(MachineInstOpcode);
381 if (II.getNumDefs() >= 1) {
382 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II,
383 ResultReg).addImm(Imm));
385 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
387 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
388 TII.get(TargetOpcode::COPY), ResultReg)
389 .addReg(II.ImplicitDefs[0]));
394 // TODO: Don't worry about 64-bit now, but when this is fixed remove the
395 // checks from the various callers.
396 unsigned ARMFastISel::ARMMoveToFPReg(MVT VT, unsigned SrcReg) {
397 if (VT == MVT::f64) return 0;
399 unsigned MoveReg = createResultReg(TLI.getRegClassFor(VT));
400 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
401 TII.get(ARM::VMOVSR), MoveReg)
406 unsigned ARMFastISel::ARMMoveToIntReg(MVT VT, unsigned SrcReg) {
407 if (VT == MVT::i64) return 0;
409 unsigned MoveReg = createResultReg(TLI.getRegClassFor(VT));
410 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
411 TII.get(ARM::VMOVRS), MoveReg)
416 // For double width floating point we need to materialize two constants
417 // (the high and the low) into integer registers then use a move to get
418 // the combined constant into an FP reg.
419 unsigned ARMFastISel::ARMMaterializeFP(const ConstantFP *CFP, MVT VT) {
420 const APFloat Val = CFP->getValueAPF();
421 bool is64bit = VT == MVT::f64;
423 // This checks to see if we can use VFP3 instructions to materialize
424 // a constant, otherwise we have to go through the constant pool.
425 if (TLI.isFPImmLegal(Val, VT)) {
429 Imm = ARM_AM::getFP64Imm(Val);
432 Imm = ARM_AM::getFP32Imm(Val);
435 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
436 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
437 TII.get(Opc), DestReg).addImm(Imm));
441 // Require VFP2 for loading fp constants.
442 if (!Subtarget->hasVFP2()) return false;
444 // MachineConstantPool wants an explicit alignment.
445 unsigned Align = DL.getPrefTypeAlignment(CFP->getType());
447 // TODO: Figure out if this is correct.
448 Align = DL.getTypeAllocSize(CFP->getType());
450 unsigned Idx = MCP.getConstantPoolIndex(cast<Constant>(CFP), Align);
451 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
452 unsigned Opc = is64bit ? ARM::VLDRD : ARM::VLDRS;
454 // The extra reg is for addrmode5.
456 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), DestReg)
457 .addConstantPoolIndex(Idx)
462 unsigned ARMFastISel::ARMMaterializeInt(const Constant *C, MVT VT) {
463 if (VT != MVT::i32 && VT != MVT::i16 && VT != MVT::i8 && VT != MVT::i1)
466 // If we can do this in a single instruction without a constant pool entry
468 const ConstantInt *CI = cast<ConstantInt>(C);
469 if (Subtarget->hasV6T2Ops() && isUInt<16>(CI->getZExtValue())) {
470 unsigned Opc = isThumb2 ? ARM::t2MOVi16 : ARM::MOVi16;
471 const TargetRegisterClass *RC = isThumb2 ? &ARM::rGPRRegClass :
473 unsigned ImmReg = createResultReg(RC);
474 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
475 TII.get(Opc), ImmReg)
476 .addImm(CI->getZExtValue()));
480 // Use MVN to emit negative constants.
481 if (VT == MVT::i32 && Subtarget->hasV6T2Ops() && CI->isNegative()) {
482 unsigned Imm = (unsigned)~(CI->getSExtValue());
483 bool UseImm = isThumb2 ? (ARM_AM::getT2SOImmVal(Imm) != -1) :
484 (ARM_AM::getSOImmVal(Imm) != -1);
486 unsigned Opc = isThumb2 ? ARM::t2MVNi : ARM::MVNi;
487 const TargetRegisterClass *RC = isThumb2 ? &ARM::rGPRRegClass :
489 unsigned ImmReg = createResultReg(RC);
490 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
491 TII.get(Opc), ImmReg)
497 unsigned ResultReg = 0;
498 if (Subtarget->useMovt(*FuncInfo.MF))
499 ResultReg = fastEmit_i(VT, VT, ISD::Constant, CI->getZExtValue());
504 // Load from constant pool. For now 32-bit only.
508 // MachineConstantPool wants an explicit alignment.
509 unsigned Align = DL.getPrefTypeAlignment(C->getType());
511 // TODO: Figure out if this is correct.
512 Align = DL.getTypeAllocSize(C->getType());
514 unsigned Idx = MCP.getConstantPoolIndex(C, Align);
515 ResultReg = createResultReg(TLI.getRegClassFor(VT));
517 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
518 TII.get(ARM::t2LDRpci), ResultReg)
519 .addConstantPoolIndex(Idx));
521 // The extra immediate is for addrmode2.
522 ResultReg = constrainOperandRegClass(TII.get(ARM::LDRcp), ResultReg, 0);
523 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
524 TII.get(ARM::LDRcp), ResultReg)
525 .addConstantPoolIndex(Idx)
531 bool ARMFastISel::isPositionIndependent() const {
532 return TLI.isPositionIndependent();
535 unsigned ARMFastISel::ARMMaterializeGV(const GlobalValue *GV, MVT VT) {
536 // For now 32-bit only.
537 if (VT != MVT::i32 || GV->isThreadLocal()) return 0;
539 // ROPI/RWPI not currently supported.
540 if (Subtarget->isROPI() || Subtarget->isRWPI())
543 bool IsIndirect = Subtarget->isGVIndirectSymbol(GV);
544 const TargetRegisterClass *RC = isThumb2 ? &ARM::rGPRRegClass
546 unsigned DestReg = createResultReg(RC);
548 // FastISel TLS support on non-MachO is broken, punt to SelectionDAG.
549 const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV);
550 bool IsThreadLocal = GVar && GVar->isThreadLocal();
551 if (!Subtarget->isTargetMachO() && IsThreadLocal) return 0;
553 bool IsPositionIndependent = isPositionIndependent();
554 // Use movw+movt when possible, it avoids constant pool entries.
555 // Non-darwin targets only support static movt relocations in FastISel.
556 if (Subtarget->useMovt(*FuncInfo.MF) &&
557 (Subtarget->isTargetMachO() || !IsPositionIndependent)) {
559 unsigned char TF = 0;
560 if (Subtarget->isTargetMachO())
561 TF = ARMII::MO_NONLAZY;
563 if (IsPositionIndependent)
564 Opc = isThumb2 ? ARM::t2MOV_ga_pcrel : ARM::MOV_ga_pcrel;
566 Opc = isThumb2 ? ARM::t2MOVi32imm : ARM::MOVi32imm;
567 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
568 TII.get(Opc), DestReg).addGlobalAddress(GV, 0, TF));
570 // MachineConstantPool wants an explicit alignment.
571 unsigned Align = DL.getPrefTypeAlignment(GV->getType());
573 // TODO: Figure out if this is correct.
574 Align = DL.getTypeAllocSize(GV->getType());
577 if (Subtarget->isTargetELF() && IsPositionIndependent)
578 return ARMLowerPICELF(GV, Align, VT);
581 unsigned PCAdj = IsPositionIndependent ? (Subtarget->isThumb() ? 4 : 8) : 0;
582 unsigned Id = AFI->createPICLabelUId();
583 ARMConstantPoolValue *CPV = ARMConstantPoolConstant::Create(GV, Id,
586 unsigned Idx = MCP.getConstantPoolIndex(CPV, Align);
589 MachineInstrBuilder MIB;
591 unsigned Opc = IsPositionIndependent ? ARM::t2LDRpci_pic : ARM::t2LDRpci;
592 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc),
593 DestReg).addConstantPoolIndex(Idx);
594 if (IsPositionIndependent)
596 AddOptionalDefs(MIB);
598 // The extra immediate is for addrmode2.
599 DestReg = constrainOperandRegClass(TII.get(ARM::LDRcp), DestReg, 0);
600 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
601 TII.get(ARM::LDRcp), DestReg)
602 .addConstantPoolIndex(Idx)
604 AddOptionalDefs(MIB);
606 if (IsPositionIndependent) {
607 unsigned Opc = IsIndirect ? ARM::PICLDR : ARM::PICADD;
608 unsigned NewDestReg = createResultReg(TLI.getRegClassFor(VT));
610 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
611 DbgLoc, TII.get(Opc), NewDestReg)
614 AddOptionalDefs(MIB);
621 MachineInstrBuilder MIB;
622 unsigned NewDestReg = createResultReg(TLI.getRegClassFor(VT));
624 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
625 TII.get(ARM::t2LDRi12), NewDestReg)
629 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
630 TII.get(ARM::LDRi12), NewDestReg)
633 DestReg = NewDestReg;
634 AddOptionalDefs(MIB);
640 unsigned ARMFastISel::fastMaterializeConstant(const Constant *C) {
641 EVT CEVT = TLI.getValueType(DL, C->getType(), true);
643 // Only handle simple types.
644 if (!CEVT.isSimple()) return 0;
645 MVT VT = CEVT.getSimpleVT();
647 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
648 return ARMMaterializeFP(CFP, VT);
649 else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
650 return ARMMaterializeGV(GV, VT);
651 else if (isa<ConstantInt>(C))
652 return ARMMaterializeInt(C, VT);
657 // TODO: unsigned ARMFastISel::TargetMaterializeFloatZero(const ConstantFP *CF);
659 unsigned ARMFastISel::fastMaterializeAlloca(const AllocaInst *AI) {
660 // Don't handle dynamic allocas.
661 if (!FuncInfo.StaticAllocaMap.count(AI)) return 0;
664 if (!isLoadTypeLegal(AI->getType(), VT)) return 0;
666 DenseMap<const AllocaInst*, int>::iterator SI =
667 FuncInfo.StaticAllocaMap.find(AI);
669 // This will get lowered later into the correct offsets and registers
670 // via rewriteXFrameIndex.
671 if (SI != FuncInfo.StaticAllocaMap.end()) {
672 unsigned Opc = isThumb2 ? ARM::t2ADDri : ARM::ADDri;
673 const TargetRegisterClass* RC = TLI.getRegClassFor(VT);
674 unsigned ResultReg = createResultReg(RC);
675 ResultReg = constrainOperandRegClass(TII.get(Opc), ResultReg, 0);
677 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
678 TII.get(Opc), ResultReg)
679 .addFrameIndex(SI->second)
687 bool ARMFastISel::isTypeLegal(Type *Ty, MVT &VT) {
688 EVT evt = TLI.getValueType(DL, Ty, true);
690 // Only handle simple types.
691 if (evt == MVT::Other || !evt.isSimple()) return false;
692 VT = evt.getSimpleVT();
694 // Handle all legal types, i.e. a register that will directly hold this
696 return TLI.isTypeLegal(VT);
699 bool ARMFastISel::isLoadTypeLegal(Type *Ty, MVT &VT) {
700 if (isTypeLegal(Ty, VT)) return true;
702 // If this is a type than can be sign or zero-extended to a basic operation
703 // go ahead and accept it now.
704 if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16)
710 // Computes the address to get to an object.
711 bool ARMFastISel::ARMComputeAddress(const Value *Obj, Address &Addr) {
712 // Some boilerplate from the X86 FastISel.
713 const User *U = nullptr;
714 unsigned Opcode = Instruction::UserOp1;
715 if (const Instruction *I = dyn_cast<Instruction>(Obj)) {
716 // Don't walk into other basic blocks unless the object is an alloca from
717 // another block, otherwise it may not have a virtual register assigned.
718 if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(Obj)) ||
719 FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
720 Opcode = I->getOpcode();
723 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Obj)) {
724 Opcode = C->getOpcode();
728 if (PointerType *Ty = dyn_cast<PointerType>(Obj->getType()))
729 if (Ty->getAddressSpace() > 255)
730 // Fast instruction selection doesn't support the special
737 case Instruction::BitCast:
738 // Look through bitcasts.
739 return ARMComputeAddress(U->getOperand(0), Addr);
740 case Instruction::IntToPtr:
741 // Look past no-op inttoptrs.
742 if (TLI.getValueType(DL, U->getOperand(0)->getType()) ==
743 TLI.getPointerTy(DL))
744 return ARMComputeAddress(U->getOperand(0), Addr);
746 case Instruction::PtrToInt:
747 // Look past no-op ptrtoints.
748 if (TLI.getValueType(DL, U->getType()) == TLI.getPointerTy(DL))
749 return ARMComputeAddress(U->getOperand(0), Addr);
751 case Instruction::GetElementPtr: {
752 Address SavedAddr = Addr;
753 int TmpOffset = Addr.Offset;
755 // Iterate through the GEP folding the constants into offsets where
757 gep_type_iterator GTI = gep_type_begin(U);
758 for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end();
759 i != e; ++i, ++GTI) {
760 const Value *Op = *i;
761 if (StructType *STy = GTI.getStructTypeOrNull()) {
762 const StructLayout *SL = DL.getStructLayout(STy);
763 unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
764 TmpOffset += SL->getElementOffset(Idx);
766 uint64_t S = DL.getTypeAllocSize(GTI.getIndexedType());
768 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
769 // Constant-offset addressing.
770 TmpOffset += CI->getSExtValue() * S;
773 if (canFoldAddIntoGEP(U, Op)) {
774 // A compatible add with a constant operand. Fold the constant.
776 cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
777 TmpOffset += CI->getSExtValue() * S;
778 // Iterate on the other operand.
779 Op = cast<AddOperator>(Op)->getOperand(0);
783 goto unsupported_gep;
788 // Try to grab the base operand now.
789 Addr.Offset = TmpOffset;
790 if (ARMComputeAddress(U->getOperand(0), Addr)) return true;
792 // We failed, restore everything and try the other options.
798 case Instruction::Alloca: {
799 const AllocaInst *AI = cast<AllocaInst>(Obj);
800 DenseMap<const AllocaInst*, int>::iterator SI =
801 FuncInfo.StaticAllocaMap.find(AI);
802 if (SI != FuncInfo.StaticAllocaMap.end()) {
803 Addr.BaseType = Address::FrameIndexBase;
804 Addr.Base.FI = SI->second;
811 // Try to get this in a register if nothing else has worked.
812 if (Addr.Base.Reg == 0) Addr.Base.Reg = getRegForValue(Obj);
813 return Addr.Base.Reg != 0;
816 void ARMFastISel::ARMSimplifyAddress(Address &Addr, MVT VT, bool useAM3) {
817 bool needsLowering = false;
818 switch (VT.SimpleTy) {
819 default: llvm_unreachable("Unhandled load/store type!");
825 // Integer loads/stores handle 12-bit offsets.
826 needsLowering = ((Addr.Offset & 0xfff) != Addr.Offset);
827 // Handle negative offsets.
828 if (needsLowering && isThumb2)
829 needsLowering = !(Subtarget->hasV6T2Ops() && Addr.Offset < 0 &&
832 // ARM halfword load/stores and signed byte loads use +/-imm8 offsets.
833 needsLowering = (Addr.Offset > 255 || Addr.Offset < -255);
838 // Floating point operands handle 8-bit offsets.
839 needsLowering = ((Addr.Offset & 0xff) != Addr.Offset);
843 // If this is a stack pointer and the offset needs to be simplified then
844 // put the alloca address into a register, set the base type back to
845 // register and continue. This should almost never happen.
846 if (needsLowering && Addr.BaseType == Address::FrameIndexBase) {
847 const TargetRegisterClass *RC = isThumb2 ? &ARM::tGPRRegClass
849 unsigned ResultReg = createResultReg(RC);
850 unsigned Opc = isThumb2 ? ARM::t2ADDri : ARM::ADDri;
851 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
852 TII.get(Opc), ResultReg)
853 .addFrameIndex(Addr.Base.FI)
855 Addr.Base.Reg = ResultReg;
856 Addr.BaseType = Address::RegBase;
859 // Since the offset is too large for the load/store instruction
860 // get the reg+offset into a register.
862 Addr.Base.Reg = fastEmit_ri_(MVT::i32, ISD::ADD, Addr.Base.Reg,
863 /*Op0IsKill*/false, Addr.Offset, MVT::i32);
868 void ARMFastISel::AddLoadStoreOperands(MVT VT, Address &Addr,
869 const MachineInstrBuilder &MIB,
870 MachineMemOperand::Flags Flags,
872 // addrmode5 output depends on the selection dag addressing dividing the
873 // offset by 4 that it then later multiplies. Do this here as well.
874 if (VT.SimpleTy == MVT::f32 || VT.SimpleTy == MVT::f64)
877 // Frame base works a bit differently. Handle it separately.
878 if (Addr.BaseType == Address::FrameIndexBase) {
879 int FI = Addr.Base.FI;
880 int Offset = Addr.Offset;
881 MachineMemOperand *MMO = FuncInfo.MF->getMachineMemOperand(
882 MachinePointerInfo::getFixedStack(*FuncInfo.MF, FI, Offset), Flags,
883 MFI.getObjectSize(FI), MFI.getObjectAlignment(FI));
884 // Now add the rest of the operands.
885 MIB.addFrameIndex(FI);
887 // ARM halfword load/stores and signed byte loads need an additional
890 int Imm = (Addr.Offset < 0) ? (0x100 | -Addr.Offset) : Addr.Offset;
894 MIB.addImm(Addr.Offset);
896 MIB.addMemOperand(MMO);
898 // Now add the rest of the operands.
899 MIB.addReg(Addr.Base.Reg);
901 // ARM halfword load/stores and signed byte loads need an additional
904 int Imm = (Addr.Offset < 0) ? (0x100 | -Addr.Offset) : Addr.Offset;
908 MIB.addImm(Addr.Offset);
911 AddOptionalDefs(MIB);
914 bool ARMFastISel::ARMEmitLoad(MVT VT, unsigned &ResultReg, Address &Addr,
915 unsigned Alignment, bool isZExt, bool allocReg) {
918 bool needVMOV = false;
919 const TargetRegisterClass *RC;
920 switch (VT.SimpleTy) {
921 // This is mostly going to be Neon/vector support.
922 default: return false;
926 if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
927 Opc = isZExt ? ARM::t2LDRBi8 : ARM::t2LDRSBi8;
929 Opc = isZExt ? ARM::t2LDRBi12 : ARM::t2LDRSBi12;
938 RC = isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRnopcRegClass;
941 if (Alignment && Alignment < 2 && !Subtarget->allowsUnalignedMem())
945 if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
946 Opc = isZExt ? ARM::t2LDRHi8 : ARM::t2LDRSHi8;
948 Opc = isZExt ? ARM::t2LDRHi12 : ARM::t2LDRSHi12;
950 Opc = isZExt ? ARM::LDRH : ARM::LDRSH;
953 RC = isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRnopcRegClass;
956 if (Alignment && Alignment < 4 && !Subtarget->allowsUnalignedMem())
960 if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
967 RC = isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRnopcRegClass;
970 if (!Subtarget->hasVFP2()) return false;
971 // Unaligned loads need special handling. Floats require word-alignment.
972 if (Alignment && Alignment < 4) {
975 Opc = isThumb2 ? ARM::t2LDRi12 : ARM::LDRi12;
976 RC = isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRnopcRegClass;
979 RC = TLI.getRegClassFor(VT);
983 if (!Subtarget->hasVFP2()) return false;
984 // FIXME: Unaligned loads need special handling. Doublewords require
986 if (Alignment && Alignment < 4)
990 RC = TLI.getRegClassFor(VT);
993 // Simplify this down to something we can handle.
994 ARMSimplifyAddress(Addr, VT, useAM3);
996 // Create the base instruction, then add the operands.
998 ResultReg = createResultReg(RC);
999 assert(ResultReg > 255 && "Expected an allocated virtual register.");
1000 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1001 TII.get(Opc), ResultReg);
1002 AddLoadStoreOperands(VT, Addr, MIB, MachineMemOperand::MOLoad, useAM3);
1004 // If we had an unaligned load of a float we've converted it to an regular
1005 // load. Now we must move from the GRP to the FP register.
1007 unsigned MoveReg = createResultReg(TLI.getRegClassFor(MVT::f32));
1008 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1009 TII.get(ARM::VMOVSR), MoveReg)
1010 .addReg(ResultReg));
1011 ResultReg = MoveReg;
1016 bool ARMFastISel::SelectLoad(const Instruction *I) {
1017 // Atomic loads need special handling.
1018 if (cast<LoadInst>(I)->isAtomic())
1021 const Value *SV = I->getOperand(0);
1022 if (TLI.supportSwiftError()) {
1023 // Swifterror values can come from either a function parameter with
1024 // swifterror attribute or an alloca with swifterror attribute.
1025 if (const Argument *Arg = dyn_cast<Argument>(SV)) {
1026 if (Arg->hasSwiftErrorAttr())
1030 if (const AllocaInst *Alloca = dyn_cast<AllocaInst>(SV)) {
1031 if (Alloca->isSwiftError())
1036 // Verify we have a legal type before going any further.
1038 if (!isLoadTypeLegal(I->getType(), VT))
1041 // See if we can handle this address.
1043 if (!ARMComputeAddress(I->getOperand(0), Addr)) return false;
1046 if (!ARMEmitLoad(VT, ResultReg, Addr, cast<LoadInst>(I)->getAlignment()))
1048 updateValueMap(I, ResultReg);
1052 bool ARMFastISel::ARMEmitStore(MVT VT, unsigned SrcReg, Address &Addr,
1053 unsigned Alignment) {
1055 bool useAM3 = false;
1056 switch (VT.SimpleTy) {
1057 // This is mostly going to be Neon/vector support.
1058 default: return false;
1060 unsigned Res = createResultReg(isThumb2 ? &ARM::tGPRRegClass
1061 : &ARM::GPRRegClass);
1062 unsigned Opc = isThumb2 ? ARM::t2ANDri : ARM::ANDri;
1063 SrcReg = constrainOperandRegClass(TII.get(Opc), SrcReg, 1);
1064 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1066 .addReg(SrcReg).addImm(1));
1072 if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
1073 StrOpc = ARM::t2STRBi8;
1075 StrOpc = ARM::t2STRBi12;
1077 StrOpc = ARM::STRBi12;
1081 if (Alignment && Alignment < 2 && !Subtarget->allowsUnalignedMem())
1085 if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
1086 StrOpc = ARM::t2STRHi8;
1088 StrOpc = ARM::t2STRHi12;
1095 if (Alignment && Alignment < 4 && !Subtarget->allowsUnalignedMem())
1099 if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
1100 StrOpc = ARM::t2STRi8;
1102 StrOpc = ARM::t2STRi12;
1104 StrOpc = ARM::STRi12;
1108 if (!Subtarget->hasVFP2()) return false;
1109 // Unaligned stores need special handling. Floats require word-alignment.
1110 if (Alignment && Alignment < 4) {
1111 unsigned MoveReg = createResultReg(TLI.getRegClassFor(MVT::i32));
1112 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1113 TII.get(ARM::VMOVRS), MoveReg)
1117 StrOpc = isThumb2 ? ARM::t2STRi12 : ARM::STRi12;
1119 StrOpc = ARM::VSTRS;
1123 if (!Subtarget->hasVFP2()) return false;
1124 // FIXME: Unaligned stores need special handling. Doublewords require
1126 if (Alignment && Alignment < 4)
1129 StrOpc = ARM::VSTRD;
1132 // Simplify this down to something we can handle.
1133 ARMSimplifyAddress(Addr, VT, useAM3);
1135 // Create the base instruction, then add the operands.
1136 SrcReg = constrainOperandRegClass(TII.get(StrOpc), SrcReg, 0);
1137 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1140 AddLoadStoreOperands(VT, Addr, MIB, MachineMemOperand::MOStore, useAM3);
1144 bool ARMFastISel::SelectStore(const Instruction *I) {
1145 Value *Op0 = I->getOperand(0);
1146 unsigned SrcReg = 0;
1148 // Atomic stores need special handling.
1149 if (cast<StoreInst>(I)->isAtomic())
1152 const Value *PtrV = I->getOperand(1);
1153 if (TLI.supportSwiftError()) {
1154 // Swifterror values can come from either a function parameter with
1155 // swifterror attribute or an alloca with swifterror attribute.
1156 if (const Argument *Arg = dyn_cast<Argument>(PtrV)) {
1157 if (Arg->hasSwiftErrorAttr())
1161 if (const AllocaInst *Alloca = dyn_cast<AllocaInst>(PtrV)) {
1162 if (Alloca->isSwiftError())
1167 // Verify we have a legal type before going any further.
1169 if (!isLoadTypeLegal(I->getOperand(0)->getType(), VT))
1172 // Get the value to be stored into a register.
1173 SrcReg = getRegForValue(Op0);
1174 if (SrcReg == 0) return false;
1176 // See if we can handle this address.
1178 if (!ARMComputeAddress(I->getOperand(1), Addr))
1181 if (!ARMEmitStore(VT, SrcReg, Addr, cast<StoreInst>(I)->getAlignment()))
1186 static ARMCC::CondCodes getComparePred(CmpInst::Predicate Pred) {
1188 // Needs two compares...
1189 case CmpInst::FCMP_ONE:
1190 case CmpInst::FCMP_UEQ:
1192 // AL is our "false" for now. The other two need more compares.
1194 case CmpInst::ICMP_EQ:
1195 case CmpInst::FCMP_OEQ:
1197 case CmpInst::ICMP_SGT:
1198 case CmpInst::FCMP_OGT:
1200 case CmpInst::ICMP_SGE:
1201 case CmpInst::FCMP_OGE:
1203 case CmpInst::ICMP_UGT:
1204 case CmpInst::FCMP_UGT:
1206 case CmpInst::FCMP_OLT:
1208 case CmpInst::ICMP_ULE:
1209 case CmpInst::FCMP_OLE:
1211 case CmpInst::FCMP_ORD:
1213 case CmpInst::FCMP_UNO:
1215 case CmpInst::FCMP_UGE:
1217 case CmpInst::ICMP_SLT:
1218 case CmpInst::FCMP_ULT:
1220 case CmpInst::ICMP_SLE:
1221 case CmpInst::FCMP_ULE:
1223 case CmpInst::FCMP_UNE:
1224 case CmpInst::ICMP_NE:
1226 case CmpInst::ICMP_UGE:
1228 case CmpInst::ICMP_ULT:
1233 bool ARMFastISel::SelectBranch(const Instruction *I) {
1234 const BranchInst *BI = cast<BranchInst>(I);
1235 MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
1236 MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
1238 // Simple branch support.
1240 // If we can, avoid recomputing the compare - redoing it could lead to wonky
1242 if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
1243 if (CI->hasOneUse() && (CI->getParent() == I->getParent())) {
1244 // Get the compare predicate.
1245 // Try to take advantage of fallthrough opportunities.
1246 CmpInst::Predicate Predicate = CI->getPredicate();
1247 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
1248 std::swap(TBB, FBB);
1249 Predicate = CmpInst::getInversePredicate(Predicate);
1252 ARMCC::CondCodes ARMPred = getComparePred(Predicate);
1254 // We may not handle every CC for now.
1255 if (ARMPred == ARMCC::AL) return false;
1257 // Emit the compare.
1258 if (!ARMEmitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned(),
1262 unsigned BrOpc = isThumb2 ? ARM::t2Bcc : ARM::Bcc;
1263 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(BrOpc))
1264 .addMBB(TBB).addImm(ARMPred).addReg(ARM::CPSR);
1265 finishCondBranch(BI->getParent(), TBB, FBB);
1268 } else if (TruncInst *TI = dyn_cast<TruncInst>(BI->getCondition())) {
1270 if (TI->hasOneUse() && TI->getParent() == I->getParent() &&
1271 (isLoadTypeLegal(TI->getOperand(0)->getType(), SourceVT))) {
1272 unsigned TstOpc = isThumb2 ? ARM::t2TSTri : ARM::TSTri;
1273 unsigned OpReg = getRegForValue(TI->getOperand(0));
1274 OpReg = constrainOperandRegClass(TII.get(TstOpc), OpReg, 0);
1275 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1277 .addReg(OpReg).addImm(1));
1279 unsigned CCMode = ARMCC::NE;
1280 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
1281 std::swap(TBB, FBB);
1285 unsigned BrOpc = isThumb2 ? ARM::t2Bcc : ARM::Bcc;
1286 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(BrOpc))
1287 .addMBB(TBB).addImm(CCMode).addReg(ARM::CPSR);
1289 finishCondBranch(BI->getParent(), TBB, FBB);
1292 } else if (const ConstantInt *CI =
1293 dyn_cast<ConstantInt>(BI->getCondition())) {
1294 uint64_t Imm = CI->getZExtValue();
1295 MachineBasicBlock *Target = (Imm == 0) ? FBB : TBB;
1296 fastEmitBranch(Target, DbgLoc);
1300 unsigned CmpReg = getRegForValue(BI->getCondition());
1301 if (CmpReg == 0) return false;
1303 // We've been divorced from our compare! Our block was split, and
1304 // now our compare lives in a predecessor block. We musn't
1305 // re-compare here, as the children of the compare aren't guaranteed
1306 // live across the block boundary (we *could* check for this).
1307 // Regardless, the compare has been done in the predecessor block,
1308 // and it left a value for us in a virtual register. Ergo, we test
1309 // the one-bit value left in the virtual register.
1310 unsigned TstOpc = isThumb2 ? ARM::t2TSTri : ARM::TSTri;
1311 CmpReg = constrainOperandRegClass(TII.get(TstOpc), CmpReg, 0);
1313 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TstOpc))
1317 unsigned CCMode = ARMCC::NE;
1318 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
1319 std::swap(TBB, FBB);
1323 unsigned BrOpc = isThumb2 ? ARM::t2Bcc : ARM::Bcc;
1324 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(BrOpc))
1325 .addMBB(TBB).addImm(CCMode).addReg(ARM::CPSR);
1326 finishCondBranch(BI->getParent(), TBB, FBB);
1330 bool ARMFastISel::SelectIndirectBr(const Instruction *I) {
1331 unsigned AddrReg = getRegForValue(I->getOperand(0));
1332 if (AddrReg == 0) return false;
1334 unsigned Opc = isThumb2 ? ARM::tBRIND : ARM::BX;
1335 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1336 TII.get(Opc)).addReg(AddrReg));
1338 const IndirectBrInst *IB = cast<IndirectBrInst>(I);
1339 for (const BasicBlock *SuccBB : IB->successors())
1340 FuncInfo.MBB->addSuccessor(FuncInfo.MBBMap[SuccBB]);
1345 bool ARMFastISel::ARMEmitCmp(const Value *Src1Value, const Value *Src2Value,
1346 bool isZExt, bool isEquality) {
1347 Type *Ty = Src1Value->getType();
1348 EVT SrcEVT = TLI.getValueType(DL, Ty, true);
1349 if (!SrcEVT.isSimple()) return false;
1350 MVT SrcVT = SrcEVT.getSimpleVT();
1352 if (Ty->isFloatTy() && !Subtarget->hasVFP2())
1355 if (Ty->isDoubleTy() && (!Subtarget->hasVFP2() || Subtarget->isFPOnlySP()))
1358 // Check to see if the 2nd operand is a constant that we can encode directly
1361 bool UseImm = false;
1362 bool isNegativeImm = false;
1363 // FIXME: At -O0 we don't have anything that canonicalizes operand order.
1364 // Thus, Src1Value may be a ConstantInt, but we're missing it.
1365 if (const ConstantInt *ConstInt = dyn_cast<ConstantInt>(Src2Value)) {
1366 if (SrcVT == MVT::i32 || SrcVT == MVT::i16 || SrcVT == MVT::i8 ||
1368 const APInt &CIVal = ConstInt->getValue();
1369 Imm = (isZExt) ? (int)CIVal.getZExtValue() : (int)CIVal.getSExtValue();
1370 // For INT_MIN/LONG_MIN (i.e., 0x80000000) we need to use a cmp, rather
1371 // then a cmn, because there is no way to represent 2147483648 as a
1372 // signed 32-bit int.
1373 if (Imm < 0 && Imm != (int)0x80000000) {
1374 isNegativeImm = true;
1377 UseImm = isThumb2 ? (ARM_AM::getT2SOImmVal(Imm) != -1) :
1378 (ARM_AM::getSOImmVal(Imm) != -1);
1380 } else if (const ConstantFP *ConstFP = dyn_cast<ConstantFP>(Src2Value)) {
1381 if (SrcVT == MVT::f32 || SrcVT == MVT::f64)
1382 if (ConstFP->isZero() && !ConstFP->isNegative())
1388 bool needsExt = false;
1389 switch (SrcVT.SimpleTy) {
1390 default: return false;
1391 // TODO: Verify compares.
1394 // Equality comparisons shouldn't raise Invalid on uordered inputs.
1396 CmpOpc = UseImm ? ARM::VCMPZS : ARM::VCMPS;
1398 CmpOpc = UseImm ? ARM::VCMPEZS : ARM::VCMPES;
1402 // Equality comparisons shouldn't raise Invalid on uordered inputs.
1404 CmpOpc = UseImm ? ARM::VCMPZD : ARM::VCMPD;
1406 CmpOpc = UseImm ? ARM::VCMPEZD : ARM::VCMPED;
1412 // Intentional fall-through.
1416 CmpOpc = ARM::t2CMPrr;
1418 CmpOpc = isNegativeImm ? ARM::t2CMNri : ARM::t2CMPri;
1421 CmpOpc = ARM::CMPrr;
1423 CmpOpc = isNegativeImm ? ARM::CMNri : ARM::CMPri;
1428 unsigned SrcReg1 = getRegForValue(Src1Value);
1429 if (SrcReg1 == 0) return false;
1431 unsigned SrcReg2 = 0;
1433 SrcReg2 = getRegForValue(Src2Value);
1434 if (SrcReg2 == 0) return false;
1437 // We have i1, i8, or i16, we need to either zero extend or sign extend.
1439 SrcReg1 = ARMEmitIntExt(SrcVT, SrcReg1, MVT::i32, isZExt);
1440 if (SrcReg1 == 0) return false;
1442 SrcReg2 = ARMEmitIntExt(SrcVT, SrcReg2, MVT::i32, isZExt);
1443 if (SrcReg2 == 0) return false;
1447 const MCInstrDesc &II = TII.get(CmpOpc);
1448 SrcReg1 = constrainOperandRegClass(II, SrcReg1, 0);
1450 SrcReg2 = constrainOperandRegClass(II, SrcReg2, 1);
1451 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
1452 .addReg(SrcReg1).addReg(SrcReg2));
1454 MachineInstrBuilder MIB;
1455 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
1458 // Only add immediate for icmp as the immediate for fcmp is an implicit 0.0.
1461 AddOptionalDefs(MIB);
1464 // For floating point we need to move the result to a comparison register
1465 // that we can then use for branches.
1466 if (Ty->isFloatTy() || Ty->isDoubleTy())
1467 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1468 TII.get(ARM::FMSTAT)));
1472 bool ARMFastISel::SelectCmp(const Instruction *I) {
1473 const CmpInst *CI = cast<CmpInst>(I);
1475 // Get the compare predicate.
1476 ARMCC::CondCodes ARMPred = getComparePred(CI->getPredicate());
1478 // We may not handle every CC for now.
1479 if (ARMPred == ARMCC::AL) return false;
1481 // Emit the compare.
1482 if (!ARMEmitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned(),
1486 // Now set a register based on the comparison. Explicitly set the predicates
1488 unsigned MovCCOpc = isThumb2 ? ARM::t2MOVCCi : ARM::MOVCCi;
1489 const TargetRegisterClass *RC = isThumb2 ? &ARM::rGPRRegClass
1490 : &ARM::GPRRegClass;
1491 unsigned DestReg = createResultReg(RC);
1492 Constant *Zero = ConstantInt::get(Type::getInt32Ty(*Context), 0);
1493 unsigned ZeroReg = fastMaterializeConstant(Zero);
1494 // ARMEmitCmp emits a FMSTAT when necessary, so it's always safe to use CPSR.
1495 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(MovCCOpc), DestReg)
1496 .addReg(ZeroReg).addImm(1)
1497 .addImm(ARMPred).addReg(ARM::CPSR);
1499 updateValueMap(I, DestReg);
1503 bool ARMFastISel::SelectFPExt(const Instruction *I) {
1504 // Make sure we have VFP and that we're extending float to double.
1505 if (!Subtarget->hasVFP2() || Subtarget->isFPOnlySP()) return false;
1507 Value *V = I->getOperand(0);
1508 if (!I->getType()->isDoubleTy() ||
1509 !V->getType()->isFloatTy()) return false;
1511 unsigned Op = getRegForValue(V);
1512 if (Op == 0) return false;
1514 unsigned Result = createResultReg(&ARM::DPRRegClass);
1515 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1516 TII.get(ARM::VCVTDS), Result)
1518 updateValueMap(I, Result);
1522 bool ARMFastISel::SelectFPTrunc(const Instruction *I) {
1523 // Make sure we have VFP and that we're truncating double to float.
1524 if (!Subtarget->hasVFP2() || Subtarget->isFPOnlySP()) return false;
1526 Value *V = I->getOperand(0);
1527 if (!(I->getType()->isFloatTy() &&
1528 V->getType()->isDoubleTy())) return false;
1530 unsigned Op = getRegForValue(V);
1531 if (Op == 0) return false;
1533 unsigned Result = createResultReg(&ARM::SPRRegClass);
1534 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1535 TII.get(ARM::VCVTSD), Result)
1537 updateValueMap(I, Result);
1541 bool ARMFastISel::SelectIToFP(const Instruction *I, bool isSigned) {
1542 // Make sure we have VFP.
1543 if (!Subtarget->hasVFP2()) return false;
1546 Type *Ty = I->getType();
1547 if (!isTypeLegal(Ty, DstVT))
1550 Value *Src = I->getOperand(0);
1551 EVT SrcEVT = TLI.getValueType(DL, Src->getType(), true);
1552 if (!SrcEVT.isSimple())
1554 MVT SrcVT = SrcEVT.getSimpleVT();
1555 if (SrcVT != MVT::i32 && SrcVT != MVT::i16 && SrcVT != MVT::i8)
1558 unsigned SrcReg = getRegForValue(Src);
1559 if (SrcReg == 0) return false;
1561 // Handle sign-extension.
1562 if (SrcVT == MVT::i16 || SrcVT == MVT::i8) {
1563 SrcReg = ARMEmitIntExt(SrcVT, SrcReg, MVT::i32,
1564 /*isZExt*/!isSigned);
1565 if (SrcReg == 0) return false;
1568 // The conversion routine works on fp-reg to fp-reg and the operand above
1569 // was an integer, move it to the fp registers if possible.
1570 unsigned FP = ARMMoveToFPReg(MVT::f32, SrcReg);
1571 if (FP == 0) return false;
1574 if (Ty->isFloatTy()) Opc = isSigned ? ARM::VSITOS : ARM::VUITOS;
1575 else if (Ty->isDoubleTy() && !Subtarget->isFPOnlySP())
1576 Opc = isSigned ? ARM::VSITOD : ARM::VUITOD;
1579 unsigned ResultReg = createResultReg(TLI.getRegClassFor(DstVT));
1580 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1581 TII.get(Opc), ResultReg).addReg(FP));
1582 updateValueMap(I, ResultReg);
1586 bool ARMFastISel::SelectFPToI(const Instruction *I, bool isSigned) {
1587 // Make sure we have VFP.
1588 if (!Subtarget->hasVFP2()) return false;
1591 Type *RetTy = I->getType();
1592 if (!isTypeLegal(RetTy, DstVT))
1595 unsigned Op = getRegForValue(I->getOperand(0));
1596 if (Op == 0) return false;
1599 Type *OpTy = I->getOperand(0)->getType();
1600 if (OpTy->isFloatTy()) Opc = isSigned ? ARM::VTOSIZS : ARM::VTOUIZS;
1601 else if (OpTy->isDoubleTy() && !Subtarget->isFPOnlySP())
1602 Opc = isSigned ? ARM::VTOSIZD : ARM::VTOUIZD;
1605 // f64->s32/u32 or f32->s32/u32 both need an intermediate f32 reg.
1606 unsigned ResultReg = createResultReg(TLI.getRegClassFor(MVT::f32));
1607 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1608 TII.get(Opc), ResultReg).addReg(Op));
1610 // This result needs to be in an integer register, but the conversion only
1611 // takes place in fp-regs.
1612 unsigned IntReg = ARMMoveToIntReg(DstVT, ResultReg);
1613 if (IntReg == 0) return false;
1615 updateValueMap(I, IntReg);
1619 bool ARMFastISel::SelectSelect(const Instruction *I) {
1621 if (!isTypeLegal(I->getType(), VT))
1624 // Things need to be register sized for register moves.
1625 if (VT != MVT::i32) return false;
1627 unsigned CondReg = getRegForValue(I->getOperand(0));
1628 if (CondReg == 0) return false;
1629 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1630 if (Op1Reg == 0) return false;
1632 // Check to see if we can use an immediate in the conditional move.
1634 bool UseImm = false;
1635 bool isNegativeImm = false;
1636 if (const ConstantInt *ConstInt = dyn_cast<ConstantInt>(I->getOperand(2))) {
1637 assert(VT == MVT::i32 && "Expecting an i32.");
1638 Imm = (int)ConstInt->getValue().getZExtValue();
1640 isNegativeImm = true;
1643 UseImm = isThumb2 ? (ARM_AM::getT2SOImmVal(Imm) != -1) :
1644 (ARM_AM::getSOImmVal(Imm) != -1);
1647 unsigned Op2Reg = 0;
1649 Op2Reg = getRegForValue(I->getOperand(2));
1650 if (Op2Reg == 0) return false;
1653 unsigned TstOpc = isThumb2 ? ARM::t2TSTri : ARM::TSTri;
1654 CondReg = constrainOperandRegClass(TII.get(TstOpc), CondReg, 0);
1656 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TstOpc))
1661 const TargetRegisterClass *RC;
1663 RC = isThumb2 ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
1664 MovCCOpc = isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr;
1666 RC = isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRRegClass;
1668 MovCCOpc = isThumb2 ? ARM::t2MOVCCi : ARM::MOVCCi;
1670 MovCCOpc = isThumb2 ? ARM::t2MVNCCi : ARM::MVNCCi;
1672 unsigned ResultReg = createResultReg(RC);
1674 Op2Reg = constrainOperandRegClass(TII.get(MovCCOpc), Op2Reg, 1);
1675 Op1Reg = constrainOperandRegClass(TII.get(MovCCOpc), Op1Reg, 2);
1676 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(MovCCOpc),
1683 Op1Reg = constrainOperandRegClass(TII.get(MovCCOpc), Op1Reg, 1);
1684 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(MovCCOpc),
1691 updateValueMap(I, ResultReg);
1695 bool ARMFastISel::SelectDiv(const Instruction *I, bool isSigned) {
1697 Type *Ty = I->getType();
1698 if (!isTypeLegal(Ty, VT))
1701 // If we have integer div support we should have selected this automagically.
1702 // In case we have a real miss go ahead and return false and we'll pick
1704 if (Subtarget->hasDivideInThumbMode())
1707 // Otherwise emit a libcall.
1708 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
1710 LC = isSigned ? RTLIB::SDIV_I8 : RTLIB::UDIV_I8;
1711 else if (VT == MVT::i16)
1712 LC = isSigned ? RTLIB::SDIV_I16 : RTLIB::UDIV_I16;
1713 else if (VT == MVT::i32)
1714 LC = isSigned ? RTLIB::SDIV_I32 : RTLIB::UDIV_I32;
1715 else if (VT == MVT::i64)
1716 LC = isSigned ? RTLIB::SDIV_I64 : RTLIB::UDIV_I64;
1717 else if (VT == MVT::i128)
1718 LC = isSigned ? RTLIB::SDIV_I128 : RTLIB::UDIV_I128;
1719 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SDIV!");
1721 return ARMEmitLibcall(I, LC);
1724 bool ARMFastISel::SelectRem(const Instruction *I, bool isSigned) {
1726 Type *Ty = I->getType();
1727 if (!isTypeLegal(Ty, VT))
1730 // Many ABIs do not provide a libcall for standalone remainder, so we need to
1731 // use divrem (see the RTABI 4.3.1). Since FastISel can't handle non-double
1732 // multi-reg returns, we'll have to bail out.
1733 if (!TLI.hasStandaloneRem(VT)) {
1737 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
1739 LC = isSigned ? RTLIB::SREM_I8 : RTLIB::UREM_I8;
1740 else if (VT == MVT::i16)
1741 LC = isSigned ? RTLIB::SREM_I16 : RTLIB::UREM_I16;
1742 else if (VT == MVT::i32)
1743 LC = isSigned ? RTLIB::SREM_I32 : RTLIB::UREM_I32;
1744 else if (VT == MVT::i64)
1745 LC = isSigned ? RTLIB::SREM_I64 : RTLIB::UREM_I64;
1746 else if (VT == MVT::i128)
1747 LC = isSigned ? RTLIB::SREM_I128 : RTLIB::UREM_I128;
1748 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SREM!");
1750 return ARMEmitLibcall(I, LC);
1753 bool ARMFastISel::SelectBinaryIntOp(const Instruction *I, unsigned ISDOpcode) {
1754 EVT DestVT = TLI.getValueType(DL, I->getType(), true);
1756 // We can get here in the case when we have a binary operation on a non-legal
1757 // type and the target independent selector doesn't know how to handle it.
1758 if (DestVT != MVT::i16 && DestVT != MVT::i8 && DestVT != MVT::i1)
1762 switch (ISDOpcode) {
1763 default: return false;
1765 Opc = isThumb2 ? ARM::t2ADDrr : ARM::ADDrr;
1768 Opc = isThumb2 ? ARM::t2ORRrr : ARM::ORRrr;
1771 Opc = isThumb2 ? ARM::t2SUBrr : ARM::SUBrr;
1775 unsigned SrcReg1 = getRegForValue(I->getOperand(0));
1776 if (SrcReg1 == 0) return false;
1778 // TODO: Often the 2nd operand is an immediate, which can be encoded directly
1779 // in the instruction, rather then materializing the value in a register.
1780 unsigned SrcReg2 = getRegForValue(I->getOperand(1));
1781 if (SrcReg2 == 0) return false;
1783 unsigned ResultReg = createResultReg(&ARM::GPRnopcRegClass);
1784 SrcReg1 = constrainOperandRegClass(TII.get(Opc), SrcReg1, 1);
1785 SrcReg2 = constrainOperandRegClass(TII.get(Opc), SrcReg2, 2);
1786 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1787 TII.get(Opc), ResultReg)
1788 .addReg(SrcReg1).addReg(SrcReg2));
1789 updateValueMap(I, ResultReg);
1793 bool ARMFastISel::SelectBinaryFPOp(const Instruction *I, unsigned ISDOpcode) {
1794 EVT FPVT = TLI.getValueType(DL, I->getType(), true);
1795 if (!FPVT.isSimple()) return false;
1796 MVT VT = FPVT.getSimpleVT();
1798 // FIXME: Support vector types where possible.
1802 // We can get here in the case when we want to use NEON for our fp
1803 // operations, but can't figure out how to. Just use the vfp instructions
1805 // FIXME: It'd be nice to use NEON instructions.
1806 Type *Ty = I->getType();
1807 if (Ty->isFloatTy() && !Subtarget->hasVFP2())
1809 if (Ty->isDoubleTy() && (!Subtarget->hasVFP2() || Subtarget->isFPOnlySP()))
1813 bool is64bit = VT == MVT::f64 || VT == MVT::i64;
1814 switch (ISDOpcode) {
1815 default: return false;
1817 Opc = is64bit ? ARM::VADDD : ARM::VADDS;
1820 Opc = is64bit ? ARM::VSUBD : ARM::VSUBS;
1823 Opc = is64bit ? ARM::VMULD : ARM::VMULS;
1826 unsigned Op1 = getRegForValue(I->getOperand(0));
1827 if (Op1 == 0) return false;
1829 unsigned Op2 = getRegForValue(I->getOperand(1));
1830 if (Op2 == 0) return false;
1832 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT.SimpleTy));
1833 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1834 TII.get(Opc), ResultReg)
1835 .addReg(Op1).addReg(Op2));
1836 updateValueMap(I, ResultReg);
1840 // Call Handling Code
1842 // This is largely taken directly from CCAssignFnForNode
1843 // TODO: We may not support all of this.
1844 CCAssignFn *ARMFastISel::CCAssignFnForCall(CallingConv::ID CC,
1849 llvm_unreachable("Unsupported calling convention");
1850 case CallingConv::Fast:
1851 if (Subtarget->hasVFP2() && !isVarArg) {
1852 if (!Subtarget->isAAPCS_ABI())
1853 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1854 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1855 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1858 case CallingConv::C:
1859 case CallingConv::CXX_FAST_TLS:
1860 // Use target triple & subtarget features to do actual dispatch.
1861 if (Subtarget->isAAPCS_ABI()) {
1862 if (Subtarget->hasVFP2() &&
1863 TM.Options.FloatABIType == FloatABI::Hard && !isVarArg)
1864 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
1866 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
1868 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
1870 case CallingConv::ARM_AAPCS_VFP:
1871 case CallingConv::Swift:
1873 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
1874 // Fall through to soft float variant, variadic functions don't
1875 // use hard floating point ABI.
1877 case CallingConv::ARM_AAPCS:
1878 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
1879 case CallingConv::ARM_APCS:
1880 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
1881 case CallingConv::GHC:
1883 llvm_unreachable("Can't return in GHC call convention");
1885 return CC_ARM_APCS_GHC;
1889 bool ARMFastISel::ProcessCallArgs(SmallVectorImpl<Value*> &Args,
1890 SmallVectorImpl<unsigned> &ArgRegs,
1891 SmallVectorImpl<MVT> &ArgVTs,
1892 SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
1893 SmallVectorImpl<unsigned> &RegArgs,
1897 SmallVector<CCValAssign, 16> ArgLocs;
1898 CCState CCInfo(CC, isVarArg, *FuncInfo.MF, ArgLocs, *Context);
1899 CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags,
1900 CCAssignFnForCall(CC, false, isVarArg));
1902 // Check that we can handle all of the arguments. If we can't, then bail out
1903 // now before we add code to the MBB.
1904 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1905 CCValAssign &VA = ArgLocs[i];
1906 MVT ArgVT = ArgVTs[VA.getValNo()];
1908 // We don't handle NEON/vector parameters yet.
1909 if (ArgVT.isVector() || ArgVT.getSizeInBits() > 64)
1912 // Now copy/store arg to correct locations.
1913 if (VA.isRegLoc() && !VA.needsCustom()) {
1915 } else if (VA.needsCustom()) {
1916 // TODO: We need custom lowering for vector (v2f64) args.
1917 if (VA.getLocVT() != MVT::f64 ||
1918 // TODO: Only handle register args for now.
1919 !VA.isRegLoc() || !ArgLocs[++i].isRegLoc())
1922 switch (ArgVT.SimpleTy) {
1931 if (!Subtarget->hasVFP2())
1935 if (!Subtarget->hasVFP2())
1942 // At the point, we are able to handle the call's arguments in fast isel.
1944 // Get a count of how many bytes are to be pushed on the stack.
1945 NumBytes = CCInfo.getNextStackOffset();
1947 // Issue CALLSEQ_START
1948 unsigned AdjStackDown = TII.getCallFrameSetupOpcode();
1949 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1950 TII.get(AdjStackDown))
1951 .addImm(NumBytes).addImm(0));
1953 // Process the args.
1954 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1955 CCValAssign &VA = ArgLocs[i];
1956 const Value *ArgVal = Args[VA.getValNo()];
1957 unsigned Arg = ArgRegs[VA.getValNo()];
1958 MVT ArgVT = ArgVTs[VA.getValNo()];
1960 assert((!ArgVT.isVector() && ArgVT.getSizeInBits() <= 64) &&
1961 "We don't handle NEON/vector parameters yet.");
1963 // Handle arg promotion, etc.
1964 switch (VA.getLocInfo()) {
1965 case CCValAssign::Full: break;
1966 case CCValAssign::SExt: {
1967 MVT DestVT = VA.getLocVT();
1968 Arg = ARMEmitIntExt(ArgVT, Arg, DestVT, /*isZExt*/false);
1969 assert(Arg != 0 && "Failed to emit a sext");
1973 case CCValAssign::AExt:
1974 // Intentional fall-through. Handle AExt and ZExt.
1975 case CCValAssign::ZExt: {
1976 MVT DestVT = VA.getLocVT();
1977 Arg = ARMEmitIntExt(ArgVT, Arg, DestVT, /*isZExt*/true);
1978 assert(Arg != 0 && "Failed to emit a zext");
1982 case CCValAssign::BCvt: {
1983 unsigned BC = fastEmit_r(ArgVT, VA.getLocVT(), ISD::BITCAST, Arg,
1984 /*TODO: Kill=*/false);
1985 assert(BC != 0 && "Failed to emit a bitcast!");
1987 ArgVT = VA.getLocVT();
1990 default: llvm_unreachable("Unknown arg promotion!");
1993 // Now copy/store arg to correct locations.
1994 if (VA.isRegLoc() && !VA.needsCustom()) {
1995 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1996 TII.get(TargetOpcode::COPY), VA.getLocReg()).addReg(Arg);
1997 RegArgs.push_back(VA.getLocReg());
1998 } else if (VA.needsCustom()) {
1999 // TODO: We need custom lowering for vector (v2f64) args.
2000 assert(VA.getLocVT() == MVT::f64 &&
2001 "Custom lowering for v2f64 args not available");
2003 // FIXME: ArgLocs[++i] may extend beyond ArgLocs.size()
2004 CCValAssign &NextVA = ArgLocs[++i];
2006 assert(VA.isRegLoc() && NextVA.isRegLoc() &&
2007 "We only handle register args!");
2009 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2010 TII.get(ARM::VMOVRRD), VA.getLocReg())
2011 .addReg(NextVA.getLocReg(), RegState::Define)
2013 RegArgs.push_back(VA.getLocReg());
2014 RegArgs.push_back(NextVA.getLocReg());
2016 assert(VA.isMemLoc());
2017 // Need to store on the stack.
2019 // Don't emit stores for undef values.
2020 if (isa<UndefValue>(ArgVal))
2024 Addr.BaseType = Address::RegBase;
2025 Addr.Base.Reg = ARM::SP;
2026 Addr.Offset = VA.getLocMemOffset();
2028 bool EmitRet = ARMEmitStore(ArgVT, Arg, Addr); (void)EmitRet;
2029 assert(EmitRet && "Could not emit a store for argument!");
2036 bool ARMFastISel::FinishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
2037 const Instruction *I, CallingConv::ID CC,
2038 unsigned &NumBytes, bool isVarArg) {
2039 // Issue CALLSEQ_END
2040 unsigned AdjStackUp = TII.getCallFrameDestroyOpcode();
2041 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2042 TII.get(AdjStackUp))
2043 .addImm(NumBytes).addImm(0));
2045 // Now the return value.
2046 if (RetVT != MVT::isVoid) {
2047 SmallVector<CCValAssign, 16> RVLocs;
2048 CCState CCInfo(CC, isVarArg, *FuncInfo.MF, RVLocs, *Context);
2049 CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC, true, isVarArg));
2051 // Copy all of the result registers out of their specified physreg.
2052 if (RVLocs.size() == 2 && RetVT == MVT::f64) {
2053 // For this move we copy into two registers and then move into the
2054 // double fp reg we want.
2055 MVT DestVT = RVLocs[0].getValVT();
2056 const TargetRegisterClass* DstRC = TLI.getRegClassFor(DestVT);
2057 unsigned ResultReg = createResultReg(DstRC);
2058 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2059 TII.get(ARM::VMOVDRR), ResultReg)
2060 .addReg(RVLocs[0].getLocReg())
2061 .addReg(RVLocs[1].getLocReg()));
2063 UsedRegs.push_back(RVLocs[0].getLocReg());
2064 UsedRegs.push_back(RVLocs[1].getLocReg());
2066 // Finally update the result.
2067 updateValueMap(I, ResultReg);
2069 assert(RVLocs.size() == 1 &&"Can't handle non-double multi-reg retvals!");
2070 MVT CopyVT = RVLocs[0].getValVT();
2072 // Special handling for extended integers.
2073 if (RetVT == MVT::i1 || RetVT == MVT::i8 || RetVT == MVT::i16)
2076 const TargetRegisterClass* DstRC = TLI.getRegClassFor(CopyVT);
2078 unsigned ResultReg = createResultReg(DstRC);
2079 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2080 TII.get(TargetOpcode::COPY),
2081 ResultReg).addReg(RVLocs[0].getLocReg());
2082 UsedRegs.push_back(RVLocs[0].getLocReg());
2084 // Finally update the result.
2085 updateValueMap(I, ResultReg);
2092 bool ARMFastISel::SelectRet(const Instruction *I) {
2093 const ReturnInst *Ret = cast<ReturnInst>(I);
2094 const Function &F = *I->getParent()->getParent();
2096 if (!FuncInfo.CanLowerReturn)
2099 if (TLI.supportSwiftError() &&
2100 F.getAttributes().hasAttrSomewhere(Attribute::SwiftError))
2103 if (TLI.supportSplitCSR(FuncInfo.MF))
2106 // Build a list of return value registers.
2107 SmallVector<unsigned, 4> RetRegs;
2109 CallingConv::ID CC = F.getCallingConv();
2110 if (Ret->getNumOperands() > 0) {
2111 SmallVector<ISD::OutputArg, 4> Outs;
2112 GetReturnInfo(F.getReturnType(), F.getAttributes(), Outs, TLI, DL);
2114 // Analyze operands of the call, assigning locations to each operand.
2115 SmallVector<CCValAssign, 16> ValLocs;
2116 CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, ValLocs, I->getContext());
2117 CCInfo.AnalyzeReturn(Outs, CCAssignFnForCall(CC, true /* is Ret */,
2120 const Value *RV = Ret->getOperand(0);
2121 unsigned Reg = getRegForValue(RV);
2125 // Only handle a single return value for now.
2126 if (ValLocs.size() != 1)
2129 CCValAssign &VA = ValLocs[0];
2131 // Don't bother handling odd stuff for now.
2132 if (VA.getLocInfo() != CCValAssign::Full)
2134 // Only handle register returns for now.
2138 unsigned SrcReg = Reg + VA.getValNo();
2139 EVT RVEVT = TLI.getValueType(DL, RV->getType());
2140 if (!RVEVT.isSimple()) return false;
2141 MVT RVVT = RVEVT.getSimpleVT();
2142 MVT DestVT = VA.getValVT();
2143 // Special handling for extended integers.
2144 if (RVVT != DestVT) {
2145 if (RVVT != MVT::i1 && RVVT != MVT::i8 && RVVT != MVT::i16)
2148 assert(DestVT == MVT::i32 && "ARM should always ext to i32");
2150 // Perform extension if flagged as either zext or sext. Otherwise, do
2152 if (Outs[0].Flags.isZExt() || Outs[0].Flags.isSExt()) {
2153 SrcReg = ARMEmitIntExt(RVVT, SrcReg, DestVT, Outs[0].Flags.isZExt());
2154 if (SrcReg == 0) return false;
2159 unsigned DstReg = VA.getLocReg();
2160 const TargetRegisterClass* SrcRC = MRI.getRegClass(SrcReg);
2161 // Avoid a cross-class copy. This is very unlikely.
2162 if (!SrcRC->contains(DstReg))
2164 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2165 TII.get(TargetOpcode::COPY), DstReg).addReg(SrcReg);
2167 // Add register to return instruction.
2168 RetRegs.push_back(VA.getLocReg());
2171 unsigned RetOpc = isThumb2 ? ARM::tBX_RET : ARM::BX_RET;
2172 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2174 AddOptionalDefs(MIB);
2175 for (unsigned R : RetRegs)
2176 MIB.addReg(R, RegState::Implicit);
2180 unsigned ARMFastISel::ARMSelectCallOp(bool UseReg) {
2182 return isThumb2 ? ARM::tBLXr : ARM::BLX;
2184 return isThumb2 ? ARM::tBL : ARM::BL;
2187 unsigned ARMFastISel::getLibcallReg(const Twine &Name) {
2188 // Manually compute the global's type to avoid building it when unnecessary.
2189 Type *GVTy = Type::getInt32PtrTy(*Context, /*AS=*/0);
2190 EVT LCREVT = TLI.getValueType(DL, GVTy);
2191 if (!LCREVT.isSimple()) return 0;
2193 GlobalValue *GV = new GlobalVariable(M, Type::getInt32Ty(*Context), false,
2194 GlobalValue::ExternalLinkage, nullptr,
2196 assert(GV->getType() == GVTy && "We miscomputed the type for the global!");
2197 return ARMMaterializeGV(GV, LCREVT.getSimpleVT());
2200 // A quick function that will emit a call for a named libcall in F with the
2201 // vector of passed arguments for the Instruction in I. We can assume that we
2202 // can emit a call for any libcall we can produce. This is an abridged version
2203 // of the full call infrastructure since we won't need to worry about things
2204 // like computed function pointers or strange arguments at call sites.
2205 // TODO: Try to unify this and the normal call bits for ARM, then try to unify
2207 bool ARMFastISel::ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call) {
2208 CallingConv::ID CC = TLI.getLibcallCallingConv(Call);
2210 // Handle *simple* calls for now.
2211 Type *RetTy = I->getType();
2213 if (RetTy->isVoidTy())
2214 RetVT = MVT::isVoid;
2215 else if (!isTypeLegal(RetTy, RetVT))
2218 // Can't handle non-double multi-reg retvals.
2219 if (RetVT != MVT::isVoid && RetVT != MVT::i32) {
2220 SmallVector<CCValAssign, 16> RVLocs;
2221 CCState CCInfo(CC, false, *FuncInfo.MF, RVLocs, *Context);
2222 CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC, true, false));
2223 if (RVLocs.size() >= 2 && RetVT != MVT::f64)
2227 // Set up the argument vectors.
2228 SmallVector<Value*, 8> Args;
2229 SmallVector<unsigned, 8> ArgRegs;
2230 SmallVector<MVT, 8> ArgVTs;
2231 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
2232 Args.reserve(I->getNumOperands());
2233 ArgRegs.reserve(I->getNumOperands());
2234 ArgVTs.reserve(I->getNumOperands());
2235 ArgFlags.reserve(I->getNumOperands());
2236 for (Value *Op : I->operands()) {
2237 unsigned Arg = getRegForValue(Op);
2238 if (Arg == 0) return false;
2240 Type *ArgTy = Op->getType();
2242 if (!isTypeLegal(ArgTy, ArgVT)) return false;
2244 ISD::ArgFlagsTy Flags;
2245 unsigned OriginalAlignment = DL.getABITypeAlignment(ArgTy);
2246 Flags.setOrigAlign(OriginalAlignment);
2249 ArgRegs.push_back(Arg);
2250 ArgVTs.push_back(ArgVT);
2251 ArgFlags.push_back(Flags);
2254 // Handle the arguments now that we've gotten them.
2255 SmallVector<unsigned, 4> RegArgs;
2257 if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags,
2258 RegArgs, CC, NumBytes, false))
2261 unsigned CalleeReg = 0;
2262 if (Subtarget->genLongCalls()) {
2263 CalleeReg = getLibcallReg(TLI.getLibcallName(Call));
2264 if (CalleeReg == 0) return false;
2268 unsigned CallOpc = ARMSelectCallOp(Subtarget->genLongCalls());
2269 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
2270 DbgLoc, TII.get(CallOpc));
2271 // BL / BLX don't take a predicate, but tBL / tBLX do.
2273 MIB.add(predOps(ARMCC::AL));
2274 if (Subtarget->genLongCalls())
2275 MIB.addReg(CalleeReg);
2277 MIB.addExternalSymbol(TLI.getLibcallName(Call));
2279 // Add implicit physical register uses to the call.
2280 for (unsigned R : RegArgs)
2281 MIB.addReg(R, RegState::Implicit);
2283 // Add a register mask with the call-preserved registers.
2284 // Proper defs for return values will be added by setPhysRegsDeadExcept().
2285 MIB.addRegMask(TRI.getCallPreservedMask(*FuncInfo.MF, CC));
2287 // Finish off the call including any return values.
2288 SmallVector<unsigned, 4> UsedRegs;
2289 if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes, false)) return false;
2291 // Set all unused physreg defs as dead.
2292 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
2297 bool ARMFastISel::SelectCall(const Instruction *I,
2298 const char *IntrMemName = nullptr) {
2299 const CallInst *CI = cast<CallInst>(I);
2300 const Value *Callee = CI->getCalledValue();
2302 // Can't handle inline asm.
2303 if (isa<InlineAsm>(Callee)) return false;
2305 // Allow SelectionDAG isel to handle tail calls.
2306 if (CI->isTailCall()) return false;
2308 // Check the calling convention.
2309 ImmutableCallSite CS(CI);
2310 CallingConv::ID CC = CS.getCallingConv();
2312 // TODO: Avoid some calling conventions?
2314 FunctionType *FTy = CS.getFunctionType();
2315 bool isVarArg = FTy->isVarArg();
2317 // Handle *simple* calls for now.
2318 Type *RetTy = I->getType();
2320 if (RetTy->isVoidTy())
2321 RetVT = MVT::isVoid;
2322 else if (!isTypeLegal(RetTy, RetVT) && RetVT != MVT::i16 &&
2323 RetVT != MVT::i8 && RetVT != MVT::i1)
2326 // Can't handle non-double multi-reg retvals.
2327 if (RetVT != MVT::isVoid && RetVT != MVT::i1 && RetVT != MVT::i8 &&
2328 RetVT != MVT::i16 && RetVT != MVT::i32) {
2329 SmallVector<CCValAssign, 16> RVLocs;
2330 CCState CCInfo(CC, isVarArg, *FuncInfo.MF, RVLocs, *Context);
2331 CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC, true, isVarArg));
2332 if (RVLocs.size() >= 2 && RetVT != MVT::f64)
2336 // Set up the argument vectors.
2337 SmallVector<Value*, 8> Args;
2338 SmallVector<unsigned, 8> ArgRegs;
2339 SmallVector<MVT, 8> ArgVTs;
2340 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
2341 unsigned arg_size = CS.arg_size();
2342 Args.reserve(arg_size);
2343 ArgRegs.reserve(arg_size);
2344 ArgVTs.reserve(arg_size);
2345 ArgFlags.reserve(arg_size);
2346 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
2348 // If we're lowering a memory intrinsic instead of a regular call, skip the
2349 // last two arguments, which shouldn't be passed to the underlying function.
2350 if (IntrMemName && e-i <= 2)
2353 ISD::ArgFlagsTy Flags;
2354 unsigned ArgIdx = i - CS.arg_begin();
2355 if (CS.paramHasAttr(ArgIdx, Attribute::SExt))
2357 if (CS.paramHasAttr(ArgIdx, Attribute::ZExt))
2360 // FIXME: Only handle *easy* calls for now.
2361 if (CS.paramHasAttr(ArgIdx, Attribute::InReg) ||
2362 CS.paramHasAttr(ArgIdx, Attribute::StructRet) ||
2363 CS.paramHasAttr(ArgIdx, Attribute::SwiftSelf) ||
2364 CS.paramHasAttr(ArgIdx, Attribute::SwiftError) ||
2365 CS.paramHasAttr(ArgIdx, Attribute::Nest) ||
2366 CS.paramHasAttr(ArgIdx, Attribute::ByVal))
2369 Type *ArgTy = (*i)->getType();
2371 if (!isTypeLegal(ArgTy, ArgVT) && ArgVT != MVT::i16 && ArgVT != MVT::i8 &&
2375 unsigned Arg = getRegForValue(*i);
2379 unsigned OriginalAlignment = DL.getABITypeAlignment(ArgTy);
2380 Flags.setOrigAlign(OriginalAlignment);
2383 ArgRegs.push_back(Arg);
2384 ArgVTs.push_back(ArgVT);
2385 ArgFlags.push_back(Flags);
2388 // Handle the arguments now that we've gotten them.
2389 SmallVector<unsigned, 4> RegArgs;
2391 if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags,
2392 RegArgs, CC, NumBytes, isVarArg))
2395 bool UseReg = false;
2396 const GlobalValue *GV = dyn_cast<GlobalValue>(Callee);
2397 if (!GV || Subtarget->genLongCalls()) UseReg = true;
2399 unsigned CalleeReg = 0;
2402 CalleeReg = getLibcallReg(IntrMemName);
2404 CalleeReg = getRegForValue(Callee);
2406 if (CalleeReg == 0) return false;
2410 unsigned CallOpc = ARMSelectCallOp(UseReg);
2411 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
2412 DbgLoc, TII.get(CallOpc));
2414 // ARM calls don't take a predicate, but tBL / tBLX do.
2416 MIB.add(predOps(ARMCC::AL));
2418 MIB.addReg(CalleeReg);
2419 else if (!IntrMemName)
2420 MIB.addGlobalAddress(GV, 0, 0);
2422 MIB.addExternalSymbol(IntrMemName, 0);
2424 // Add implicit physical register uses to the call.
2425 for (unsigned R : RegArgs)
2426 MIB.addReg(R, RegState::Implicit);
2428 // Add a register mask with the call-preserved registers.
2429 // Proper defs for return values will be added by setPhysRegsDeadExcept().
2430 MIB.addRegMask(TRI.getCallPreservedMask(*FuncInfo.MF, CC));
2432 // Finish off the call including any return values.
2433 SmallVector<unsigned, 4> UsedRegs;
2434 if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes, isVarArg))
2437 // Set all unused physreg defs as dead.
2438 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
2443 bool ARMFastISel::ARMIsMemCpySmall(uint64_t Len) {
2447 bool ARMFastISel::ARMTryEmitSmallMemCpy(Address Dest, Address Src,
2448 uint64_t Len, unsigned Alignment) {
2449 // Make sure we don't bloat code by inlining very large memcpy's.
2450 if (!ARMIsMemCpySmall(Len))
2455 if (!Alignment || Alignment >= 4) {
2461 assert(Len == 1 && "Expected a length of 1!");
2465 // Bound based on alignment.
2466 if (Len >= 2 && Alignment == 2)
2475 RV = ARMEmitLoad(VT, ResultReg, Src);
2476 assert(RV && "Should be able to handle this load.");
2477 RV = ARMEmitStore(VT, ResultReg, Dest);
2478 assert(RV && "Should be able to handle this store.");
2481 unsigned Size = VT.getSizeInBits()/8;
2483 Dest.Offset += Size;
2490 bool ARMFastISel::SelectIntrinsicCall(const IntrinsicInst &I) {
2491 // FIXME: Handle more intrinsics.
2492 switch (I.getIntrinsicID()) {
2493 default: return false;
2494 case Intrinsic::frameaddress: {
2495 MachineFrameInfo &MFI = FuncInfo.MF->getFrameInfo();
2496 MFI.setFrameAddressIsTaken(true);
2498 unsigned LdrOpc = isThumb2 ? ARM::t2LDRi12 : ARM::LDRi12;
2499 const TargetRegisterClass *RC = isThumb2 ? &ARM::tGPRRegClass
2500 : &ARM::GPRRegClass;
2502 const ARMBaseRegisterInfo *RegInfo =
2503 static_cast<const ARMBaseRegisterInfo *>(Subtarget->getRegisterInfo());
2504 unsigned FramePtr = RegInfo->getFrameRegister(*(FuncInfo.MF));
2505 unsigned SrcReg = FramePtr;
2507 // Recursively load frame address
2513 unsigned Depth = cast<ConstantInt>(I.getOperand(0))->getZExtValue();
2515 DestReg = createResultReg(RC);
2516 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2517 TII.get(LdrOpc), DestReg)
2518 .addReg(SrcReg).addImm(0));
2521 updateValueMap(&I, SrcReg);
2524 case Intrinsic::memcpy:
2525 case Intrinsic::memmove: {
2526 const MemTransferInst &MTI = cast<MemTransferInst>(I);
2527 // Don't handle volatile.
2528 if (MTI.isVolatile())
2531 // Disable inlining for memmove before calls to ComputeAddress. Otherwise,
2532 // we would emit dead code because we don't currently handle memmoves.
2533 bool isMemCpy = (I.getIntrinsicID() == Intrinsic::memcpy);
2534 if (isa<ConstantInt>(MTI.getLength()) && isMemCpy) {
2535 // Small memcpy's are common enough that we want to do them without a call
2537 uint64_t Len = cast<ConstantInt>(MTI.getLength())->getZExtValue();
2538 if (ARMIsMemCpySmall(Len)) {
2540 if (!ARMComputeAddress(MTI.getRawDest(), Dest) ||
2541 !ARMComputeAddress(MTI.getRawSource(), Src))
2543 unsigned Alignment = MTI.getAlignment();
2544 if (ARMTryEmitSmallMemCpy(Dest, Src, Len, Alignment))
2549 if (!MTI.getLength()->getType()->isIntegerTy(32))
2552 if (MTI.getSourceAddressSpace() > 255 || MTI.getDestAddressSpace() > 255)
2555 const char *IntrMemName = isa<MemCpyInst>(I) ? "memcpy" : "memmove";
2556 return SelectCall(&I, IntrMemName);
2558 case Intrinsic::memset: {
2559 const MemSetInst &MSI = cast<MemSetInst>(I);
2560 // Don't handle volatile.
2561 if (MSI.isVolatile())
2564 if (!MSI.getLength()->getType()->isIntegerTy(32))
2567 if (MSI.getDestAddressSpace() > 255)
2570 return SelectCall(&I, "memset");
2572 case Intrinsic::trap: {
2573 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(
2574 Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP));
2580 bool ARMFastISel::SelectTrunc(const Instruction *I) {
2581 // The high bits for a type smaller than the register size are assumed to be
2583 Value *Op = I->getOperand(0);
2586 SrcVT = TLI.getValueType(DL, Op->getType(), true);
2587 DestVT = TLI.getValueType(DL, I->getType(), true);
2589 if (SrcVT != MVT::i32 && SrcVT != MVT::i16 && SrcVT != MVT::i8)
2591 if (DestVT != MVT::i16 && DestVT != MVT::i8 && DestVT != MVT::i1)
2594 unsigned SrcReg = getRegForValue(Op);
2595 if (!SrcReg) return false;
2597 // Because the high bits are undefined, a truncate doesn't generate
2599 updateValueMap(I, SrcReg);
2603 unsigned ARMFastISel::ARMEmitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT,
2605 if (DestVT != MVT::i32 && DestVT != MVT::i16 && DestVT != MVT::i8)
2607 if (SrcVT != MVT::i16 && SrcVT != MVT::i8 && SrcVT != MVT::i1)
2610 // Table of which combinations can be emitted as a single instruction,
2611 // and which will require two.
2612 static const uint8_t isSingleInstrTbl[3][2][2][2] = {
2614 // !hasV6Ops hasV6Ops !hasV6Ops hasV6Ops
2615 // ext: s z s z s z s z
2616 /* 1 */ { { { 0, 1 }, { 0, 1 } }, { { 0, 0 }, { 0, 1 } } },
2617 /* 8 */ { { { 0, 1 }, { 1, 1 } }, { { 0, 0 }, { 1, 1 } } },
2618 /* 16 */ { { { 0, 0 }, { 1, 1 } }, { { 0, 0 }, { 1, 1 } } }
2621 // Target registers for:
2622 // - For ARM can never be PC.
2623 // - For 16-bit Thumb are restricted to lower 8 registers.
2624 // - For 32-bit Thumb are restricted to non-SP and non-PC.
2625 static const TargetRegisterClass *RCTbl[2][2] = {
2626 // Instructions: Two Single
2627 /* ARM */ { &ARM::GPRnopcRegClass, &ARM::GPRnopcRegClass },
2628 /* Thumb */ { &ARM::tGPRRegClass, &ARM::rGPRRegClass }
2631 // Table governing the instruction(s) to be emitted.
2632 static const struct InstructionTable {
2634 uint32_t hasS : 1; // Some instructions have an S bit, always set it to 0.
2635 uint32_t Shift : 7; // For shift operand addressing mode, used by MOVsi.
2636 uint32_t Imm : 8; // All instructions have either a shift or a mask.
2637 } IT[2][2][3][2] = {
2638 { // Two instructions (first is left shift, second is in this table).
2639 { // ARM Opc S Shift Imm
2640 /* 1 bit sext */ { { ARM::MOVsi , 1, ARM_AM::asr , 31 },
2641 /* 1 bit zext */ { ARM::MOVsi , 1, ARM_AM::lsr , 31 } },
2642 /* 8 bit sext */ { { ARM::MOVsi , 1, ARM_AM::asr , 24 },
2643 /* 8 bit zext */ { ARM::MOVsi , 1, ARM_AM::lsr , 24 } },
2644 /* 16 bit sext */ { { ARM::MOVsi , 1, ARM_AM::asr , 16 },
2645 /* 16 bit zext */ { ARM::MOVsi , 1, ARM_AM::lsr , 16 } }
2647 { // Thumb Opc S Shift Imm
2648 /* 1 bit sext */ { { ARM::tASRri , 0, ARM_AM::no_shift, 31 },
2649 /* 1 bit zext */ { ARM::tLSRri , 0, ARM_AM::no_shift, 31 } },
2650 /* 8 bit sext */ { { ARM::tASRri , 0, ARM_AM::no_shift, 24 },
2651 /* 8 bit zext */ { ARM::tLSRri , 0, ARM_AM::no_shift, 24 } },
2652 /* 16 bit sext */ { { ARM::tASRri , 0, ARM_AM::no_shift, 16 },
2653 /* 16 bit zext */ { ARM::tLSRri , 0, ARM_AM::no_shift, 16 } }
2656 { // Single instruction.
2657 { // ARM Opc S Shift Imm
2658 /* 1 bit sext */ { { ARM::KILL , 0, ARM_AM::no_shift, 0 },
2659 /* 1 bit zext */ { ARM::ANDri , 1, ARM_AM::no_shift, 1 } },
2660 /* 8 bit sext */ { { ARM::SXTB , 0, ARM_AM::no_shift, 0 },
2661 /* 8 bit zext */ { ARM::ANDri , 1, ARM_AM::no_shift, 255 } },
2662 /* 16 bit sext */ { { ARM::SXTH , 0, ARM_AM::no_shift, 0 },
2663 /* 16 bit zext */ { ARM::UXTH , 0, ARM_AM::no_shift, 0 } }
2665 { // Thumb Opc S Shift Imm
2666 /* 1 bit sext */ { { ARM::KILL , 0, ARM_AM::no_shift, 0 },
2667 /* 1 bit zext */ { ARM::t2ANDri, 1, ARM_AM::no_shift, 1 } },
2668 /* 8 bit sext */ { { ARM::t2SXTB , 0, ARM_AM::no_shift, 0 },
2669 /* 8 bit zext */ { ARM::t2ANDri, 1, ARM_AM::no_shift, 255 } },
2670 /* 16 bit sext */ { { ARM::t2SXTH , 0, ARM_AM::no_shift, 0 },
2671 /* 16 bit zext */ { ARM::t2UXTH , 0, ARM_AM::no_shift, 0 } }
2676 unsigned SrcBits = SrcVT.getSizeInBits();
2677 unsigned DestBits = DestVT.getSizeInBits();
2679 assert((SrcBits < DestBits) && "can only extend to larger types");
2680 assert((DestBits == 32 || DestBits == 16 || DestBits == 8) &&
2681 "other sizes unimplemented");
2682 assert((SrcBits == 16 || SrcBits == 8 || SrcBits == 1) &&
2683 "other sizes unimplemented");
2685 bool hasV6Ops = Subtarget->hasV6Ops();
2686 unsigned Bitness = SrcBits / 8; // {1,8,16}=>{0,1,2}
2687 assert((Bitness < 3) && "sanity-check table bounds");
2689 bool isSingleInstr = isSingleInstrTbl[Bitness][isThumb2][hasV6Ops][isZExt];
2690 const TargetRegisterClass *RC = RCTbl[isThumb2][isSingleInstr];
2691 const InstructionTable *ITP = &IT[isSingleInstr][isThumb2][Bitness][isZExt];
2692 unsigned Opc = ITP->Opc;
2693 assert(ARM::KILL != Opc && "Invalid table entry");
2694 unsigned hasS = ITP->hasS;
2695 ARM_AM::ShiftOpc Shift = (ARM_AM::ShiftOpc) ITP->Shift;
2696 assert(((Shift == ARM_AM::no_shift) == (Opc != ARM::MOVsi)) &&
2697 "only MOVsi has shift operand addressing mode");
2698 unsigned Imm = ITP->Imm;
2700 // 16-bit Thumb instructions always set CPSR (unless they're in an IT block).
2701 bool setsCPSR = &ARM::tGPRRegClass == RC;
2702 unsigned LSLOpc = isThumb2 ? ARM::tLSLri : ARM::MOVsi;
2704 // MOVsi encodes shift and immediate in shift operand addressing mode.
2705 // The following condition has the same value when emitting two
2706 // instruction sequences: both are shifts.
2707 bool ImmIsSO = (Shift != ARM_AM::no_shift);
2709 // Either one or two instructions are emitted.
2710 // They're always of the form:
2712 // CPSR is set only by 16-bit Thumb instructions.
2713 // Predicate, if any, is AL.
2714 // S bit, if available, is always 0.
2715 // When two are emitted the first's result will feed as the second's input,
2716 // that value is then dead.
2717 unsigned NumInstrsEmitted = isSingleInstr ? 1 : 2;
2718 for (unsigned Instr = 0; Instr != NumInstrsEmitted; ++Instr) {
2719 ResultReg = createResultReg(RC);
2720 bool isLsl = (0 == Instr) && !isSingleInstr;
2721 unsigned Opcode = isLsl ? LSLOpc : Opc;
2722 ARM_AM::ShiftOpc ShiftAM = isLsl ? ARM_AM::lsl : Shift;
2723 unsigned ImmEnc = ImmIsSO ? ARM_AM::getSORegOpc(ShiftAM, Imm) : Imm;
2724 bool isKill = 1 == Instr;
2725 MachineInstrBuilder MIB = BuildMI(
2726 *FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opcode), ResultReg);
2728 MIB.addReg(ARM::CPSR, RegState::Define);
2729 SrcReg = constrainOperandRegClass(TII.get(Opcode), SrcReg, 1 + setsCPSR);
2730 MIB.addReg(SrcReg, isKill * RegState::Kill)
2732 .add(predOps(ARMCC::AL));
2734 MIB.add(condCodeOp());
2735 // Second instruction consumes the first's result.
2742 bool ARMFastISel::SelectIntExt(const Instruction *I) {
2743 // On ARM, in general, integer casts don't involve legal types; this code
2744 // handles promotable integers.
2745 Type *DestTy = I->getType();
2746 Value *Src = I->getOperand(0);
2747 Type *SrcTy = Src->getType();
2749 bool isZExt = isa<ZExtInst>(I);
2750 unsigned SrcReg = getRegForValue(Src);
2751 if (!SrcReg) return false;
2753 EVT SrcEVT, DestEVT;
2754 SrcEVT = TLI.getValueType(DL, SrcTy, true);
2755 DestEVT = TLI.getValueType(DL, DestTy, true);
2756 if (!SrcEVT.isSimple()) return false;
2757 if (!DestEVT.isSimple()) return false;
2759 MVT SrcVT = SrcEVT.getSimpleVT();
2760 MVT DestVT = DestEVT.getSimpleVT();
2761 unsigned ResultReg = ARMEmitIntExt(SrcVT, SrcReg, DestVT, isZExt);
2762 if (ResultReg == 0) return false;
2763 updateValueMap(I, ResultReg);
2767 bool ARMFastISel::SelectShift(const Instruction *I,
2768 ARM_AM::ShiftOpc ShiftTy) {
2769 // We handle thumb2 mode by target independent selector
2770 // or SelectionDAG ISel.
2774 // Only handle i32 now.
2775 EVT DestVT = TLI.getValueType(DL, I->getType(), true);
2776 if (DestVT != MVT::i32)
2779 unsigned Opc = ARM::MOVsr;
2781 Value *Src2Value = I->getOperand(1);
2782 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Src2Value)) {
2783 ShiftImm = CI->getZExtValue();
2785 // Fall back to selection DAG isel if the shift amount
2786 // is zero or greater than the width of the value type.
2787 if (ShiftImm == 0 || ShiftImm >=32)
2793 Value *Src1Value = I->getOperand(0);
2794 unsigned Reg1 = getRegForValue(Src1Value);
2795 if (Reg1 == 0) return false;
2798 if (Opc == ARM::MOVsr) {
2799 Reg2 = getRegForValue(Src2Value);
2800 if (Reg2 == 0) return false;
2803 unsigned ResultReg = createResultReg(&ARM::GPRnopcRegClass);
2804 if(ResultReg == 0) return false;
2806 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2807 TII.get(Opc), ResultReg)
2810 if (Opc == ARM::MOVsi)
2811 MIB.addImm(ARM_AM::getSORegOpc(ShiftTy, ShiftImm));
2812 else if (Opc == ARM::MOVsr) {
2814 MIB.addImm(ARM_AM::getSORegOpc(ShiftTy, 0));
2817 AddOptionalDefs(MIB);
2818 updateValueMap(I, ResultReg);
2822 // TODO: SoftFP support.
2823 bool ARMFastISel::fastSelectInstruction(const Instruction *I) {
2824 switch (I->getOpcode()) {
2825 case Instruction::Load:
2826 return SelectLoad(I);
2827 case Instruction::Store:
2828 return SelectStore(I);
2829 case Instruction::Br:
2830 return SelectBranch(I);
2831 case Instruction::IndirectBr:
2832 return SelectIndirectBr(I);
2833 case Instruction::ICmp:
2834 case Instruction::FCmp:
2835 return SelectCmp(I);
2836 case Instruction::FPExt:
2837 return SelectFPExt(I);
2838 case Instruction::FPTrunc:
2839 return SelectFPTrunc(I);
2840 case Instruction::SIToFP:
2841 return SelectIToFP(I, /*isSigned*/ true);
2842 case Instruction::UIToFP:
2843 return SelectIToFP(I, /*isSigned*/ false);
2844 case Instruction::FPToSI:
2845 return SelectFPToI(I, /*isSigned*/ true);
2846 case Instruction::FPToUI:
2847 return SelectFPToI(I, /*isSigned*/ false);
2848 case Instruction::Add:
2849 return SelectBinaryIntOp(I, ISD::ADD);
2850 case Instruction::Or:
2851 return SelectBinaryIntOp(I, ISD::OR);
2852 case Instruction::Sub:
2853 return SelectBinaryIntOp(I, ISD::SUB);
2854 case Instruction::FAdd:
2855 return SelectBinaryFPOp(I, ISD::FADD);
2856 case Instruction::FSub:
2857 return SelectBinaryFPOp(I, ISD::FSUB);
2858 case Instruction::FMul:
2859 return SelectBinaryFPOp(I, ISD::FMUL);
2860 case Instruction::SDiv:
2861 return SelectDiv(I, /*isSigned*/ true);
2862 case Instruction::UDiv:
2863 return SelectDiv(I, /*isSigned*/ false);
2864 case Instruction::SRem:
2865 return SelectRem(I, /*isSigned*/ true);
2866 case Instruction::URem:
2867 return SelectRem(I, /*isSigned*/ false);
2868 case Instruction::Call:
2869 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2870 return SelectIntrinsicCall(*II);
2871 return SelectCall(I);
2872 case Instruction::Select:
2873 return SelectSelect(I);
2874 case Instruction::Ret:
2875 return SelectRet(I);
2876 case Instruction::Trunc:
2877 return SelectTrunc(I);
2878 case Instruction::ZExt:
2879 case Instruction::SExt:
2880 return SelectIntExt(I);
2881 case Instruction::Shl:
2882 return SelectShift(I, ARM_AM::lsl);
2883 case Instruction::LShr:
2884 return SelectShift(I, ARM_AM::lsr);
2885 case Instruction::AShr:
2886 return SelectShift(I, ARM_AM::asr);
2894 // This table describes sign- and zero-extend instructions which can be
2895 // folded into a preceding load. All of these extends have an immediate
2896 // (sometimes a mask and sometimes a shift) that's applied after
2898 const struct FoldableLoadExtendsStruct {
2899 uint16_t Opc[2]; // ARM, Thumb.
2900 uint8_t ExpectedImm;
2902 uint8_t ExpectedVT : 7;
2903 } FoldableLoadExtends[] = {
2904 { { ARM::SXTH, ARM::t2SXTH }, 0, 0, MVT::i16 },
2905 { { ARM::UXTH, ARM::t2UXTH }, 0, 1, MVT::i16 },
2906 { { ARM::ANDri, ARM::t2ANDri }, 255, 1, MVT::i8 },
2907 { { ARM::SXTB, ARM::t2SXTB }, 0, 0, MVT::i8 },
2908 { { ARM::UXTB, ARM::t2UXTB }, 0, 1, MVT::i8 }
2911 } // end anonymous namespace
2913 /// \brief The specified machine instr operand is a vreg, and that
2914 /// vreg is being provided by the specified load instruction. If possible,
2915 /// try to fold the load as an operand to the instruction, returning true if
2917 bool ARMFastISel::tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo,
2918 const LoadInst *LI) {
2919 // Verify we have a legal type before going any further.
2921 if (!isLoadTypeLegal(LI->getType(), VT))
2924 // Combine load followed by zero- or sign-extend.
2925 // ldrb r1, [r0] ldrb r1, [r0]
2927 // mov r3, r2 mov r3, r1
2928 if (MI->getNumOperands() < 3 || !MI->getOperand(2).isImm())
2930 const uint64_t Imm = MI->getOperand(2).getImm();
2934 for (const FoldableLoadExtendsStruct &FLE : FoldableLoadExtends) {
2935 if (FLE.Opc[isThumb2] == MI->getOpcode() &&
2936 (uint64_t)FLE.ExpectedImm == Imm &&
2937 MVT((MVT::SimpleValueType)FLE.ExpectedVT) == VT) {
2939 isZExt = FLE.isZExt;
2942 if (!Found) return false;
2944 // See if we can handle this address.
2946 if (!ARMComputeAddress(LI->getOperand(0), Addr)) return false;
2948 unsigned ResultReg = MI->getOperand(0).getReg();
2949 if (!ARMEmitLoad(VT, ResultReg, Addr, LI->getAlignment(), isZExt, false))
2951 MI->eraseFromParent();
2955 unsigned ARMFastISel::ARMLowerPICELF(const GlobalValue *GV,
2956 unsigned Align, MVT VT) {
2957 bool UseGOT_PREL = !TM.shouldAssumeDSOLocal(*GV->getParent(), GV);
2959 LLVMContext *Context = &MF->getFunction()->getContext();
2960 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2961 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2962 ARMConstantPoolValue *CPV = ARMConstantPoolConstant::Create(
2963 GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj,
2964 UseGOT_PREL ? ARMCP::GOT_PREL : ARMCP::no_modifier,
2965 /*AddCurrentAddress=*/UseGOT_PREL);
2967 unsigned ConstAlign =
2968 MF->getDataLayout().getPrefTypeAlignment(Type::getInt32PtrTy(*Context));
2969 unsigned Idx = MF->getConstantPool()->getConstantPoolIndex(CPV, ConstAlign);
2971 unsigned TempReg = MF->getRegInfo().createVirtualRegister(&ARM::rGPRRegClass);
2972 unsigned Opc = isThumb2 ? ARM::t2LDRpci : ARM::LDRcp;
2973 MachineInstrBuilder MIB =
2974 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), TempReg)
2975 .addConstantPoolIndex(Idx);
2976 if (Opc == ARM::LDRcp)
2978 MIB.add(predOps(ARMCC::AL));
2980 // Fix the address by adding pc.
2981 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
2982 Opc = Subtarget->isThumb() ? ARM::tPICADD : UseGOT_PREL ? ARM::PICLDR
2984 DestReg = constrainOperandRegClass(TII.get(Opc), DestReg, 0);
2985 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), DestReg)
2987 .addImm(ARMPCLabelIndex);
2988 if (!Subtarget->isThumb())
2989 MIB.add(predOps(ARMCC::AL));
2991 if (UseGOT_PREL && Subtarget->isThumb()) {
2992 unsigned NewDestReg = createResultReg(TLI.getRegClassFor(VT));
2993 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2994 TII.get(ARM::t2LDRi12), NewDestReg)
2997 DestReg = NewDestReg;
2998 AddOptionalDefs(MIB);
3003 bool ARMFastISel::fastLowerArguments() {
3004 if (!FuncInfo.CanLowerReturn)
3007 const Function *F = FuncInfo.Fn;
3011 CallingConv::ID CC = F->getCallingConv();
3015 case CallingConv::Fast:
3016 case CallingConv::C:
3017 case CallingConv::ARM_AAPCS_VFP:
3018 case CallingConv::ARM_AAPCS:
3019 case CallingConv::ARM_APCS:
3020 case CallingConv::Swift:
3024 // Only handle simple cases. i.e. Up to 4 i8/i16/i32 scalar arguments
3025 // which are passed in r0 - r3.
3026 for (const Argument &Arg : F->args()) {
3027 if (Arg.getArgNo() >= 4)
3030 if (Arg.hasAttribute(Attribute::InReg) ||
3031 Arg.hasAttribute(Attribute::StructRet) ||
3032 Arg.hasAttribute(Attribute::SwiftSelf) ||
3033 Arg.hasAttribute(Attribute::SwiftError) ||
3034 Arg.hasAttribute(Attribute::ByVal))
3037 Type *ArgTy = Arg.getType();
3038 if (ArgTy->isStructTy() || ArgTy->isArrayTy() || ArgTy->isVectorTy())
3041 EVT ArgVT = TLI.getValueType(DL, ArgTy);
3042 if (!ArgVT.isSimple()) return false;
3043 switch (ArgVT.getSimpleVT().SimpleTy) {
3053 static const MCPhysReg GPRArgRegs[] = {
3054 ARM::R0, ARM::R1, ARM::R2, ARM::R3
3057 const TargetRegisterClass *RC = &ARM::rGPRRegClass;
3058 for (const Argument &Arg : F->args()) {
3059 unsigned ArgNo = Arg.getArgNo();
3060 unsigned SrcReg = GPRArgRegs[ArgNo];
3061 unsigned DstReg = FuncInfo.MF->addLiveIn(SrcReg, RC);
3062 // FIXME: Unfortunately it's necessary to emit a copy from the livein copy.
3063 // Without this, EmitLiveInCopies may eliminate the livein if its only
3064 // use is a bitcast (which isn't turned into an instruction).
3065 unsigned ResultReg = createResultReg(RC);
3066 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
3067 TII.get(TargetOpcode::COPY),
3068 ResultReg).addReg(DstReg, getKillRegState(true));
3069 updateValueMap(&Arg, ResultReg);
3077 FastISel *ARM::createFastISel(FunctionLoweringInfo &funcInfo,
3078 const TargetLibraryInfo *libInfo) {
3079 if (funcInfo.MF->getSubtarget<ARMSubtarget>().useFastISel())
3080 return new ARMFastISel(funcInfo, libInfo);
3085 } // end namespace llvm