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 "ARMCallingConv.h"
19 #include "ARMConstantPoolValue.h"
20 #include "ARMSubtarget.h"
21 #include "ARMTargetMachine.h"
22 #include "MCTargetDesc/ARMAddressingModes.h"
23 #include "llvm/ADT/STLExtras.h"
24 #include "llvm/CodeGen/Analysis.h"
25 #include "llvm/CodeGen/FastISel.h"
26 #include "llvm/CodeGen/FunctionLoweringInfo.h"
27 #include "llvm/CodeGen/MachineConstantPool.h"
28 #include "llvm/CodeGen/MachineFrameInfo.h"
29 #include "llvm/CodeGen/MachineInstrBuilder.h"
30 #include "llvm/CodeGen/MachineMemOperand.h"
31 #include "llvm/CodeGen/MachineModuleInfo.h"
32 #include "llvm/CodeGen/MachineRegisterInfo.h"
33 #include "llvm/IR/CallingConv.h"
34 #include "llvm/IR/DataLayout.h"
35 #include "llvm/IR/DerivedTypes.h"
36 #include "llvm/IR/GlobalVariable.h"
37 #include "llvm/IR/Instructions.h"
38 #include "llvm/IR/IntrinsicInst.h"
39 #include "llvm/IR/Module.h"
40 #include "llvm/IR/Operator.h"
41 #include "llvm/Support/CallSite.h"
42 #include "llvm/Support/CommandLine.h"
43 #include "llvm/Support/ErrorHandling.h"
44 #include "llvm/Support/GetElementPtrTypeIterator.h"
45 #include "llvm/Target/TargetInstrInfo.h"
46 #include "llvm/Target/TargetLowering.h"
47 #include "llvm/Target/TargetMachine.h"
48 #include "llvm/Target/TargetOptions.h"
51 extern cl::opt<bool> EnableARMLongCalls;
55 // All possible address modes, plus some.
56 typedef struct Address {
69 // Innocuous defaults for our address.
71 : BaseType(RegBase), Offset(0) {
76 class ARMFastISel : public FastISel {
78 /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
79 /// make the right decision when generating code for different targets.
80 const ARMSubtarget *Subtarget;
81 const TargetMachine &TM;
82 const TargetInstrInfo &TII;
83 const TargetLowering &TLI;
86 // Convenience variables to avoid some queries.
91 explicit ARMFastISel(FunctionLoweringInfo &funcInfo,
92 const TargetLibraryInfo *libInfo)
93 : FastISel(funcInfo, libInfo),
94 TM(funcInfo.MF->getTarget()),
95 TII(*TM.getInstrInfo()),
96 TLI(*TM.getTargetLowering()) {
97 Subtarget = &TM.getSubtarget<ARMSubtarget>();
98 AFI = funcInfo.MF->getInfo<ARMFunctionInfo>();
99 isThumb2 = AFI->isThumbFunction();
100 Context = &funcInfo.Fn->getContext();
103 // Code from FastISel.cpp.
105 unsigned FastEmitInst_(unsigned MachineInstOpcode,
106 const TargetRegisterClass *RC);
107 unsigned FastEmitInst_r(unsigned MachineInstOpcode,
108 const TargetRegisterClass *RC,
109 unsigned Op0, bool Op0IsKill);
110 unsigned FastEmitInst_rr(unsigned MachineInstOpcode,
111 const TargetRegisterClass *RC,
112 unsigned Op0, bool Op0IsKill,
113 unsigned Op1, bool Op1IsKill);
114 unsigned FastEmitInst_rrr(unsigned MachineInstOpcode,
115 const TargetRegisterClass *RC,
116 unsigned Op0, bool Op0IsKill,
117 unsigned Op1, bool Op1IsKill,
118 unsigned Op2, bool Op2IsKill);
119 unsigned FastEmitInst_ri(unsigned MachineInstOpcode,
120 const TargetRegisterClass *RC,
121 unsigned Op0, bool Op0IsKill,
123 unsigned FastEmitInst_rf(unsigned MachineInstOpcode,
124 const TargetRegisterClass *RC,
125 unsigned Op0, bool Op0IsKill,
126 const ConstantFP *FPImm);
127 unsigned FastEmitInst_rri(unsigned MachineInstOpcode,
128 const TargetRegisterClass *RC,
129 unsigned Op0, bool Op0IsKill,
130 unsigned Op1, bool Op1IsKill,
132 unsigned FastEmitInst_i(unsigned MachineInstOpcode,
133 const TargetRegisterClass *RC,
135 unsigned FastEmitInst_ii(unsigned MachineInstOpcode,
136 const TargetRegisterClass *RC,
137 uint64_t Imm1, uint64_t Imm2);
139 unsigned FastEmitInst_extractsubreg(MVT RetVT,
140 unsigned Op0, bool Op0IsKill,
143 // Backend specific FastISel code.
145 virtual bool TargetSelectInstruction(const Instruction *I);
146 virtual unsigned TargetMaterializeConstant(const Constant *C);
147 virtual unsigned TargetMaterializeAlloca(const AllocaInst *AI);
148 virtual bool tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo,
150 virtual bool FastLowerArguments();
152 #include "ARMGenFastISel.inc"
154 // Instruction selection routines.
156 bool SelectLoad(const Instruction *I);
157 bool SelectStore(const Instruction *I);
158 bool SelectBranch(const Instruction *I);
159 bool SelectIndirectBr(const Instruction *I);
160 bool SelectCmp(const Instruction *I);
161 bool SelectFPExt(const Instruction *I);
162 bool SelectFPTrunc(const Instruction *I);
163 bool SelectBinaryIntOp(const Instruction *I, unsigned ISDOpcode);
164 bool SelectBinaryFPOp(const Instruction *I, unsigned ISDOpcode);
165 bool SelectIToFP(const Instruction *I, bool isSigned);
166 bool SelectFPToI(const Instruction *I, bool isSigned);
167 bool SelectDiv(const Instruction *I, bool isSigned);
168 bool SelectRem(const Instruction *I, bool isSigned);
169 bool SelectCall(const Instruction *I, const char *IntrMemName);
170 bool SelectIntrinsicCall(const IntrinsicInst &I);
171 bool SelectSelect(const Instruction *I);
172 bool SelectRet(const Instruction *I);
173 bool SelectTrunc(const Instruction *I);
174 bool SelectIntExt(const Instruction *I);
175 bool SelectShift(const Instruction *I, ARM_AM::ShiftOpc ShiftTy);
179 unsigned constrainOperandRegClass(const MCInstrDesc &II, unsigned OpNum,
181 bool isTypeLegal(Type *Ty, MVT &VT);
182 bool isLoadTypeLegal(Type *Ty, MVT &VT);
183 bool ARMEmitCmp(const Value *Src1Value, const Value *Src2Value,
185 bool ARMEmitLoad(MVT VT, unsigned &ResultReg, Address &Addr,
186 unsigned Alignment = 0, bool isZExt = true,
187 bool allocReg = true);
188 bool ARMEmitStore(MVT VT, unsigned SrcReg, Address &Addr,
189 unsigned Alignment = 0);
190 bool ARMComputeAddress(const Value *Obj, Address &Addr);
191 void ARMSimplifyAddress(Address &Addr, MVT VT, bool useAM3);
192 bool ARMIsMemCpySmall(uint64_t Len);
193 bool ARMTryEmitSmallMemCpy(Address Dest, Address Src, uint64_t Len,
195 unsigned ARMEmitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, bool isZExt);
196 unsigned ARMMaterializeFP(const ConstantFP *CFP, MVT VT);
197 unsigned ARMMaterializeInt(const Constant *C, MVT VT);
198 unsigned ARMMaterializeGV(const GlobalValue *GV, MVT VT);
199 unsigned ARMMoveToFPReg(MVT VT, unsigned SrcReg);
200 unsigned ARMMoveToIntReg(MVT VT, unsigned SrcReg);
201 unsigned ARMSelectCallOp(bool UseReg);
202 unsigned ARMLowerPICELF(const GlobalValue *GV, unsigned Align, MVT VT);
204 // Call handling routines.
206 CCAssignFn *CCAssignFnForCall(CallingConv::ID CC,
209 bool ProcessCallArgs(SmallVectorImpl<Value*> &Args,
210 SmallVectorImpl<unsigned> &ArgRegs,
211 SmallVectorImpl<MVT> &ArgVTs,
212 SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
213 SmallVectorImpl<unsigned> &RegArgs,
217 unsigned getLibcallReg(const Twine &Name);
218 bool FinishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
219 const Instruction *I, CallingConv::ID CC,
220 unsigned &NumBytes, bool isVarArg);
221 bool ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call);
223 // OptionalDef handling routines.
225 bool isARMNEONPred(const MachineInstr *MI);
226 bool DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR);
227 const MachineInstrBuilder &AddOptionalDefs(const MachineInstrBuilder &MIB);
228 void AddLoadStoreOperands(MVT VT, Address &Addr,
229 const MachineInstrBuilder &MIB,
230 unsigned Flags, bool useAM3);
233 } // end anonymous namespace
235 #include "ARMGenCallingConv.inc"
237 // DefinesOptionalPredicate - This is different from DefinesPredicate in that
238 // we don't care about implicit defs here, just places we'll need to add a
239 // default CCReg argument. Sets CPSR if we're setting CPSR instead of CCR.
240 bool ARMFastISel::DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR) {
241 if (!MI->hasOptionalDef())
244 // Look to see if our OptionalDef is defining CPSR or CCR.
245 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
246 const MachineOperand &MO = MI->getOperand(i);
247 if (!MO.isReg() || !MO.isDef()) continue;
248 if (MO.getReg() == ARM::CPSR)
254 bool ARMFastISel::isARMNEONPred(const MachineInstr *MI) {
255 const MCInstrDesc &MCID = MI->getDesc();
257 // If we're a thumb2 or not NEON function we'll be handled via isPredicable.
258 if ((MCID.TSFlags & ARMII::DomainMask) != ARMII::DomainNEON ||
259 AFI->isThumb2Function())
260 return MI->isPredicable();
262 for (unsigned i = 0, e = MCID.getNumOperands(); i != e; ++i)
263 if (MCID.OpInfo[i].isPredicate())
269 // If the machine is predicable go ahead and add the predicate operands, if
270 // it needs default CC operands add those.
271 // TODO: If we want to support thumb1 then we'll need to deal with optional
272 // CPSR defs that need to be added before the remaining operands. See s_cc_out
273 // for descriptions why.
274 const MachineInstrBuilder &
275 ARMFastISel::AddOptionalDefs(const MachineInstrBuilder &MIB) {
276 MachineInstr *MI = &*MIB;
278 // Do we use a predicate? or...
279 // Are we NEON in ARM mode and have a predicate operand? If so, I know
280 // we're not predicable but add it anyways.
281 if (isARMNEONPred(MI))
284 // Do we optionally set a predicate? Preds is size > 0 iff the predicate
285 // defines CPSR. All other OptionalDefines in ARM are the CCR register.
287 if (DefinesOptionalPredicate(MI, &CPSR)) {
296 unsigned ARMFastISel::constrainOperandRegClass(const MCInstrDesc &II,
297 unsigned Op, unsigned OpNum) {
298 if (TargetRegisterInfo::isVirtualRegister(Op)) {
299 const TargetRegisterClass *RegClass =
300 TII.getRegClass(II, OpNum, &TRI, *FuncInfo.MF);
301 if (!MRI.constrainRegClass(Op, RegClass)) {
302 // If it's not legal to COPY between the register classes, something
303 // has gone very wrong before we got here.
304 unsigned NewOp = createResultReg(RegClass);
305 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
306 TII.get(TargetOpcode::COPY), NewOp).addReg(Op));
313 unsigned ARMFastISel::FastEmitInst_(unsigned MachineInstOpcode,
314 const TargetRegisterClass* RC) {
315 unsigned ResultReg = createResultReg(RC);
316 const MCInstrDesc &II = TII.get(MachineInstOpcode);
318 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg));
322 unsigned ARMFastISel::FastEmitInst_r(unsigned MachineInstOpcode,
323 const TargetRegisterClass *RC,
324 unsigned Op0, bool Op0IsKill) {
325 unsigned ResultReg = createResultReg(RC);
326 const MCInstrDesc &II = TII.get(MachineInstOpcode);
328 // Make sure the input operand is sufficiently constrained to be legal
329 // for this instruction.
330 Op0 = constrainOperandRegClass(II, Op0, 1);
331 if (II.getNumDefs() >= 1) {
332 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
333 .addReg(Op0, Op0IsKill * RegState::Kill));
335 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
336 .addReg(Op0, Op0IsKill * RegState::Kill));
337 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
338 TII.get(TargetOpcode::COPY), ResultReg)
339 .addReg(II.ImplicitDefs[0]));
344 unsigned ARMFastISel::FastEmitInst_rr(unsigned MachineInstOpcode,
345 const TargetRegisterClass *RC,
346 unsigned Op0, bool Op0IsKill,
347 unsigned Op1, bool Op1IsKill) {
348 unsigned ResultReg = createResultReg(RC);
349 const MCInstrDesc &II = TII.get(MachineInstOpcode);
351 // Make sure the input operands are sufficiently constrained to be legal
352 // for this instruction.
353 Op0 = constrainOperandRegClass(II, Op0, 1);
354 Op1 = constrainOperandRegClass(II, Op1, 2);
356 if (II.getNumDefs() >= 1) {
357 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
358 .addReg(Op0, Op0IsKill * RegState::Kill)
359 .addReg(Op1, Op1IsKill * RegState::Kill));
361 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
362 .addReg(Op0, Op0IsKill * RegState::Kill)
363 .addReg(Op1, Op1IsKill * RegState::Kill));
364 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
365 TII.get(TargetOpcode::COPY), ResultReg)
366 .addReg(II.ImplicitDefs[0]));
371 unsigned ARMFastISel::FastEmitInst_rrr(unsigned MachineInstOpcode,
372 const TargetRegisterClass *RC,
373 unsigned Op0, bool Op0IsKill,
374 unsigned Op1, bool Op1IsKill,
375 unsigned Op2, bool Op2IsKill) {
376 unsigned ResultReg = createResultReg(RC);
377 const MCInstrDesc &II = TII.get(MachineInstOpcode);
379 // Make sure the input operands are sufficiently constrained to be legal
380 // for this instruction.
381 Op0 = constrainOperandRegClass(II, Op0, 1);
382 Op1 = constrainOperandRegClass(II, Op1, 2);
383 Op2 = constrainOperandRegClass(II, Op1, 3);
385 if (II.getNumDefs() >= 1) {
386 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
387 .addReg(Op0, Op0IsKill * RegState::Kill)
388 .addReg(Op1, Op1IsKill * RegState::Kill)
389 .addReg(Op2, Op2IsKill * RegState::Kill));
391 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
392 .addReg(Op0, Op0IsKill * RegState::Kill)
393 .addReg(Op1, Op1IsKill * RegState::Kill)
394 .addReg(Op2, Op2IsKill * RegState::Kill));
395 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
396 TII.get(TargetOpcode::COPY), ResultReg)
397 .addReg(II.ImplicitDefs[0]));
402 unsigned ARMFastISel::FastEmitInst_ri(unsigned MachineInstOpcode,
403 const TargetRegisterClass *RC,
404 unsigned Op0, bool Op0IsKill,
406 unsigned ResultReg = createResultReg(RC);
407 const MCInstrDesc &II = TII.get(MachineInstOpcode);
409 // Make sure the input operand is sufficiently constrained to be legal
410 // for this instruction.
411 Op0 = constrainOperandRegClass(II, Op0, 1);
412 if (II.getNumDefs() >= 1) {
413 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
414 .addReg(Op0, Op0IsKill * RegState::Kill)
417 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
418 .addReg(Op0, Op0IsKill * RegState::Kill)
420 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
421 TII.get(TargetOpcode::COPY), ResultReg)
422 .addReg(II.ImplicitDefs[0]));
427 unsigned ARMFastISel::FastEmitInst_rf(unsigned MachineInstOpcode,
428 const TargetRegisterClass *RC,
429 unsigned Op0, bool Op0IsKill,
430 const ConstantFP *FPImm) {
431 unsigned ResultReg = createResultReg(RC);
432 const MCInstrDesc &II = TII.get(MachineInstOpcode);
434 // Make sure the input operand is sufficiently constrained to be legal
435 // for this instruction.
436 Op0 = constrainOperandRegClass(II, Op0, 1);
437 if (II.getNumDefs() >= 1) {
438 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
439 .addReg(Op0, Op0IsKill * RegState::Kill)
442 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
443 .addReg(Op0, Op0IsKill * RegState::Kill)
445 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
446 TII.get(TargetOpcode::COPY), ResultReg)
447 .addReg(II.ImplicitDefs[0]));
452 unsigned ARMFastISel::FastEmitInst_rri(unsigned MachineInstOpcode,
453 const TargetRegisterClass *RC,
454 unsigned Op0, bool Op0IsKill,
455 unsigned Op1, bool Op1IsKill,
457 unsigned ResultReg = createResultReg(RC);
458 const MCInstrDesc &II = TII.get(MachineInstOpcode);
460 // Make sure the input operands are sufficiently constrained to be legal
461 // for this instruction.
462 Op0 = constrainOperandRegClass(II, Op0, 1);
463 Op1 = constrainOperandRegClass(II, Op1, 2);
464 if (II.getNumDefs() >= 1) {
465 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
466 .addReg(Op0, Op0IsKill * RegState::Kill)
467 .addReg(Op1, Op1IsKill * RegState::Kill)
470 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
471 .addReg(Op0, Op0IsKill * RegState::Kill)
472 .addReg(Op1, Op1IsKill * RegState::Kill)
474 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
475 TII.get(TargetOpcode::COPY), ResultReg)
476 .addReg(II.ImplicitDefs[0]));
481 unsigned ARMFastISel::FastEmitInst_i(unsigned MachineInstOpcode,
482 const TargetRegisterClass *RC,
484 unsigned ResultReg = createResultReg(RC);
485 const MCInstrDesc &II = TII.get(MachineInstOpcode);
487 if (II.getNumDefs() >= 1) {
488 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
491 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
493 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
494 TII.get(TargetOpcode::COPY), ResultReg)
495 .addReg(II.ImplicitDefs[0]));
500 unsigned ARMFastISel::FastEmitInst_ii(unsigned MachineInstOpcode,
501 const TargetRegisterClass *RC,
502 uint64_t Imm1, uint64_t Imm2) {
503 unsigned ResultReg = createResultReg(RC);
504 const MCInstrDesc &II = TII.get(MachineInstOpcode);
506 if (II.getNumDefs() >= 1) {
507 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
508 .addImm(Imm1).addImm(Imm2));
510 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
511 .addImm(Imm1).addImm(Imm2));
512 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
513 TII.get(TargetOpcode::COPY),
515 .addReg(II.ImplicitDefs[0]));
520 unsigned ARMFastISel::FastEmitInst_extractsubreg(MVT RetVT,
521 unsigned Op0, bool Op0IsKill,
523 unsigned ResultReg = createResultReg(TLI.getRegClassFor(RetVT));
524 assert(TargetRegisterInfo::isVirtualRegister(Op0) &&
525 "Cannot yet extract from physregs");
527 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
528 DL, TII.get(TargetOpcode::COPY), ResultReg)
529 .addReg(Op0, getKillRegState(Op0IsKill), Idx));
533 // TODO: Don't worry about 64-bit now, but when this is fixed remove the
534 // checks from the various callers.
535 unsigned ARMFastISel::ARMMoveToFPReg(MVT VT, unsigned SrcReg) {
536 if (VT == MVT::f64) return 0;
538 unsigned MoveReg = createResultReg(TLI.getRegClassFor(VT));
539 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
540 TII.get(ARM::VMOVSR), MoveReg)
545 unsigned ARMFastISel::ARMMoveToIntReg(MVT VT, unsigned SrcReg) {
546 if (VT == MVT::i64) return 0;
548 unsigned MoveReg = createResultReg(TLI.getRegClassFor(VT));
549 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
550 TII.get(ARM::VMOVRS), MoveReg)
555 // For double width floating point we need to materialize two constants
556 // (the high and the low) into integer registers then use a move to get
557 // the combined constant into an FP reg.
558 unsigned ARMFastISel::ARMMaterializeFP(const ConstantFP *CFP, MVT VT) {
559 const APFloat Val = CFP->getValueAPF();
560 bool is64bit = VT == MVT::f64;
562 // This checks to see if we can use VFP3 instructions to materialize
563 // a constant, otherwise we have to go through the constant pool.
564 if (TLI.isFPImmLegal(Val, VT)) {
568 Imm = ARM_AM::getFP64Imm(Val);
571 Imm = ARM_AM::getFP32Imm(Val);
574 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
575 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
581 // Require VFP2 for loading fp constants.
582 if (!Subtarget->hasVFP2()) return false;
584 // MachineConstantPool wants an explicit alignment.
585 unsigned Align = TD.getPrefTypeAlignment(CFP->getType());
587 // TODO: Figure out if this is correct.
588 Align = TD.getTypeAllocSize(CFP->getType());
590 unsigned Idx = MCP.getConstantPoolIndex(cast<Constant>(CFP), Align);
591 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
592 unsigned Opc = is64bit ? ARM::VLDRD : ARM::VLDRS;
594 // The extra reg is for addrmode5.
595 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
597 .addConstantPoolIndex(Idx)
602 unsigned ARMFastISel::ARMMaterializeInt(const Constant *C, MVT VT) {
604 if (VT != MVT::i32 && VT != MVT::i16 && VT != MVT::i8 && VT != MVT::i1)
607 // If we can do this in a single instruction without a constant pool entry
609 const ConstantInt *CI = cast<ConstantInt>(C);
610 if (Subtarget->hasV6T2Ops() && isUInt<16>(CI->getZExtValue())) {
611 unsigned Opc = isThumb2 ? ARM::t2MOVi16 : ARM::MOVi16;
612 const TargetRegisterClass *RC = isThumb2 ? &ARM::rGPRRegClass :
614 unsigned ImmReg = createResultReg(RC);
615 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
616 TII.get(Opc), ImmReg)
617 .addImm(CI->getZExtValue()));
621 // Use MVN to emit negative constants.
622 if (VT == MVT::i32 && Subtarget->hasV6T2Ops() && CI->isNegative()) {
623 unsigned Imm = (unsigned)~(CI->getSExtValue());
624 bool UseImm = isThumb2 ? (ARM_AM::getT2SOImmVal(Imm) != -1) :
625 (ARM_AM::getSOImmVal(Imm) != -1);
627 unsigned Opc = isThumb2 ? ARM::t2MVNi : ARM::MVNi;
628 unsigned ImmReg = createResultReg(TLI.getRegClassFor(MVT::i32));
629 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
630 TII.get(Opc), ImmReg)
636 // Load from constant pool. For now 32-bit only.
640 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
642 // MachineConstantPool wants an explicit alignment.
643 unsigned Align = TD.getPrefTypeAlignment(C->getType());
645 // TODO: Figure out if this is correct.
646 Align = TD.getTypeAllocSize(C->getType());
648 unsigned Idx = MCP.getConstantPoolIndex(C, Align);
651 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
652 TII.get(ARM::t2LDRpci), DestReg)
653 .addConstantPoolIndex(Idx));
655 // The extra immediate is for addrmode2.
656 DestReg = constrainOperandRegClass(TII.get(ARM::LDRcp), DestReg, 0);
657 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
658 TII.get(ARM::LDRcp), DestReg)
659 .addConstantPoolIndex(Idx)
665 unsigned ARMFastISel::ARMMaterializeGV(const GlobalValue *GV, MVT VT) {
666 // For now 32-bit only.
667 if (VT != MVT::i32) return 0;
669 Reloc::Model RelocM = TM.getRelocationModel();
670 bool IsIndirect = Subtarget->GVIsIndirectSymbol(GV, RelocM);
671 const TargetRegisterClass *RC = isThumb2 ?
672 (const TargetRegisterClass*)&ARM::rGPRRegClass :
673 (const TargetRegisterClass*)&ARM::GPRRegClass;
674 unsigned DestReg = createResultReg(RC);
676 // FastISel TLS support on non-Darwin is broken, punt to SelectionDAG.
677 const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV);
678 bool IsThreadLocal = GVar && GVar->isThreadLocal();
679 if (!Subtarget->isTargetDarwin() && IsThreadLocal) return 0;
681 // Use movw+movt when possible, it avoids constant pool entries.
682 // Non-darwin targets only support static movt relocations in FastISel.
683 if (Subtarget->useMovt() &&
684 (Subtarget->isTargetDarwin() || RelocM == Reloc::Static)) {
686 unsigned char TF = 0;
687 if (Subtarget->isTargetDarwin())
688 TF = ARMII::MO_NONLAZY;
692 Opc = isThumb2 ? ARM::t2MOV_ga_pcrel : ARM::MOV_ga_pcrel;
695 Opc = isThumb2 ? ARM::t2MOVi32imm : ARM::MOVi32imm;
698 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
699 DestReg).addGlobalAddress(GV, 0, TF));
701 // MachineConstantPool wants an explicit alignment.
702 unsigned Align = TD.getPrefTypeAlignment(GV->getType());
704 // TODO: Figure out if this is correct.
705 Align = TD.getTypeAllocSize(GV->getType());
708 if (Subtarget->isTargetELF() && RelocM == Reloc::PIC_)
709 return ARMLowerPICELF(GV, Align, VT);
712 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 :
713 (Subtarget->isThumb() ? 4 : 8);
714 unsigned Id = AFI->createPICLabelUId();
715 ARMConstantPoolValue *CPV = ARMConstantPoolConstant::Create(GV, Id,
718 unsigned Idx = MCP.getConstantPoolIndex(CPV, Align);
721 MachineInstrBuilder MIB;
723 unsigned Opc = (RelocM!=Reloc::PIC_) ? ARM::t2LDRpci : ARM::t2LDRpci_pic;
724 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), DestReg)
725 .addConstantPoolIndex(Idx);
726 if (RelocM == Reloc::PIC_)
728 AddOptionalDefs(MIB);
730 // The extra immediate is for addrmode2.
731 DestReg = constrainOperandRegClass(TII.get(ARM::LDRcp), DestReg, 0);
732 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(ARM::LDRcp),
734 .addConstantPoolIndex(Idx)
736 AddOptionalDefs(MIB);
738 if (RelocM == Reloc::PIC_) {
739 unsigned Opc = IsIndirect ? ARM::PICLDR : ARM::PICADD;
740 unsigned NewDestReg = createResultReg(TLI.getRegClassFor(VT));
742 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
743 DL, TII.get(Opc), NewDestReg)
746 AddOptionalDefs(MIB);
753 MachineInstrBuilder MIB;
754 unsigned NewDestReg = createResultReg(TLI.getRegClassFor(VT));
756 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
757 TII.get(ARM::t2LDRi12), NewDestReg)
761 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(ARM::LDRi12),
765 DestReg = NewDestReg;
766 AddOptionalDefs(MIB);
772 unsigned ARMFastISel::TargetMaterializeConstant(const Constant *C) {
773 EVT CEVT = TLI.getValueType(C->getType(), true);
775 // Only handle simple types.
776 if (!CEVT.isSimple()) return 0;
777 MVT VT = CEVT.getSimpleVT();
779 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
780 return ARMMaterializeFP(CFP, VT);
781 else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
782 return ARMMaterializeGV(GV, VT);
783 else if (isa<ConstantInt>(C))
784 return ARMMaterializeInt(C, VT);
789 // TODO: unsigned ARMFastISel::TargetMaterializeFloatZero(const ConstantFP *CF);
791 unsigned ARMFastISel::TargetMaterializeAlloca(const AllocaInst *AI) {
792 // Don't handle dynamic allocas.
793 if (!FuncInfo.StaticAllocaMap.count(AI)) return 0;
796 if (!isLoadTypeLegal(AI->getType(), VT)) return 0;
798 DenseMap<const AllocaInst*, int>::iterator SI =
799 FuncInfo.StaticAllocaMap.find(AI);
801 // This will get lowered later into the correct offsets and registers
802 // via rewriteXFrameIndex.
803 if (SI != FuncInfo.StaticAllocaMap.end()) {
804 const TargetRegisterClass* RC = TLI.getRegClassFor(VT);
805 unsigned ResultReg = createResultReg(RC);
806 unsigned Opc = isThumb2 ? ARM::t2ADDri : ARM::ADDri;
807 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
808 TII.get(Opc), ResultReg)
809 .addFrameIndex(SI->second)
817 bool ARMFastISel::isTypeLegal(Type *Ty, MVT &VT) {
818 EVT evt = TLI.getValueType(Ty, true);
820 // Only handle simple types.
821 if (evt == MVT::Other || !evt.isSimple()) return false;
822 VT = evt.getSimpleVT();
824 // Handle all legal types, i.e. a register that will directly hold this
826 return TLI.isTypeLegal(VT);
829 bool ARMFastISel::isLoadTypeLegal(Type *Ty, MVT &VT) {
830 if (isTypeLegal(Ty, VT)) return true;
832 // If this is a type than can be sign or zero-extended to a basic operation
833 // go ahead and accept it now.
834 if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16)
840 // Computes the address to get to an object.
841 bool ARMFastISel::ARMComputeAddress(const Value *Obj, Address &Addr) {
842 // Some boilerplate from the X86 FastISel.
843 const User *U = NULL;
844 unsigned Opcode = Instruction::UserOp1;
845 if (const Instruction *I = dyn_cast<Instruction>(Obj)) {
846 // Don't walk into other basic blocks unless the object is an alloca from
847 // another block, otherwise it may not have a virtual register assigned.
848 if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(Obj)) ||
849 FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
850 Opcode = I->getOpcode();
853 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Obj)) {
854 Opcode = C->getOpcode();
858 if (PointerType *Ty = dyn_cast<PointerType>(Obj->getType()))
859 if (Ty->getAddressSpace() > 255)
860 // Fast instruction selection doesn't support the special
867 case Instruction::BitCast:
868 // Look through bitcasts.
869 return ARMComputeAddress(U->getOperand(0), Addr);
870 case Instruction::IntToPtr:
871 // Look past no-op inttoptrs.
872 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
873 return ARMComputeAddress(U->getOperand(0), Addr);
875 case Instruction::PtrToInt:
876 // Look past no-op ptrtoints.
877 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
878 return ARMComputeAddress(U->getOperand(0), Addr);
880 case Instruction::GetElementPtr: {
881 Address SavedAddr = Addr;
882 int TmpOffset = Addr.Offset;
884 // Iterate through the GEP folding the constants into offsets where
886 gep_type_iterator GTI = gep_type_begin(U);
887 for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end();
888 i != e; ++i, ++GTI) {
889 const Value *Op = *i;
890 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
891 const StructLayout *SL = TD.getStructLayout(STy);
892 unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
893 TmpOffset += SL->getElementOffset(Idx);
895 uint64_t S = TD.getTypeAllocSize(GTI.getIndexedType());
897 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
898 // Constant-offset addressing.
899 TmpOffset += CI->getSExtValue() * S;
902 if (canFoldAddIntoGEP(U, Op)) {
903 // A compatible add with a constant operand. Fold the constant.
905 cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
906 TmpOffset += CI->getSExtValue() * S;
907 // Iterate on the other operand.
908 Op = cast<AddOperator>(Op)->getOperand(0);
912 goto unsupported_gep;
917 // Try to grab the base operand now.
918 Addr.Offset = TmpOffset;
919 if (ARMComputeAddress(U->getOperand(0), Addr)) return true;
921 // We failed, restore everything and try the other options.
927 case Instruction::Alloca: {
928 const AllocaInst *AI = cast<AllocaInst>(Obj);
929 DenseMap<const AllocaInst*, int>::iterator SI =
930 FuncInfo.StaticAllocaMap.find(AI);
931 if (SI != FuncInfo.StaticAllocaMap.end()) {
932 Addr.BaseType = Address::FrameIndexBase;
933 Addr.Base.FI = SI->second;
940 // Try to get this in a register if nothing else has worked.
941 if (Addr.Base.Reg == 0) Addr.Base.Reg = getRegForValue(Obj);
942 return Addr.Base.Reg != 0;
945 void ARMFastISel::ARMSimplifyAddress(Address &Addr, MVT VT, bool useAM3) {
946 bool needsLowering = false;
947 switch (VT.SimpleTy) {
948 default: llvm_unreachable("Unhandled load/store type!");
954 // Integer loads/stores handle 12-bit offsets.
955 needsLowering = ((Addr.Offset & 0xfff) != Addr.Offset);
956 // Handle negative offsets.
957 if (needsLowering && isThumb2)
958 needsLowering = !(Subtarget->hasV6T2Ops() && Addr.Offset < 0 &&
961 // ARM halfword load/stores and signed byte loads use +/-imm8 offsets.
962 needsLowering = (Addr.Offset > 255 || Addr.Offset < -255);
967 // Floating point operands handle 8-bit offsets.
968 needsLowering = ((Addr.Offset & 0xff) != Addr.Offset);
972 // If this is a stack pointer and the offset needs to be simplified then
973 // put the alloca address into a register, set the base type back to
974 // register and continue. This should almost never happen.
975 if (needsLowering && Addr.BaseType == Address::FrameIndexBase) {
976 const TargetRegisterClass *RC = isThumb2 ?
977 (const TargetRegisterClass*)&ARM::tGPRRegClass :
978 (const TargetRegisterClass*)&ARM::GPRRegClass;
979 unsigned ResultReg = createResultReg(RC);
980 unsigned Opc = isThumb2 ? ARM::t2ADDri : ARM::ADDri;
981 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
982 TII.get(Opc), ResultReg)
983 .addFrameIndex(Addr.Base.FI)
985 Addr.Base.Reg = ResultReg;
986 Addr.BaseType = Address::RegBase;
989 // Since the offset is too large for the load/store instruction
990 // get the reg+offset into a register.
992 Addr.Base.Reg = FastEmit_ri_(MVT::i32, ISD::ADD, Addr.Base.Reg,
993 /*Op0IsKill*/false, Addr.Offset, MVT::i32);
998 void ARMFastISel::AddLoadStoreOperands(MVT VT, Address &Addr,
999 const MachineInstrBuilder &MIB,
1000 unsigned Flags, bool useAM3) {
1001 // addrmode5 output depends on the selection dag addressing dividing the
1002 // offset by 4 that it then later multiplies. Do this here as well.
1003 if (VT.SimpleTy == MVT::f32 || VT.SimpleTy == MVT::f64)
1006 // Frame base works a bit differently. Handle it separately.
1007 if (Addr.BaseType == Address::FrameIndexBase) {
1008 int FI = Addr.Base.FI;
1009 int Offset = Addr.Offset;
1010 MachineMemOperand *MMO =
1011 FuncInfo.MF->getMachineMemOperand(
1012 MachinePointerInfo::getFixedStack(FI, Offset),
1014 MFI.getObjectSize(FI),
1015 MFI.getObjectAlignment(FI));
1016 // Now add the rest of the operands.
1017 MIB.addFrameIndex(FI);
1019 // ARM halfword load/stores and signed byte loads need an additional
1022 signed Imm = (Addr.Offset < 0) ? (0x100 | -Addr.Offset) : Addr.Offset;
1026 MIB.addImm(Addr.Offset);
1028 MIB.addMemOperand(MMO);
1030 // Now add the rest of the operands.
1031 MIB.addReg(Addr.Base.Reg);
1033 // ARM halfword load/stores and signed byte loads need an additional
1036 signed Imm = (Addr.Offset < 0) ? (0x100 | -Addr.Offset) : Addr.Offset;
1040 MIB.addImm(Addr.Offset);
1043 AddOptionalDefs(MIB);
1046 bool ARMFastISel::ARMEmitLoad(MVT VT, unsigned &ResultReg, Address &Addr,
1047 unsigned Alignment, bool isZExt, bool allocReg) {
1049 bool useAM3 = false;
1050 bool needVMOV = false;
1051 const TargetRegisterClass *RC;
1052 switch (VT.SimpleTy) {
1053 // This is mostly going to be Neon/vector support.
1054 default: return false;
1058 if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
1059 Opc = isZExt ? ARM::t2LDRBi8 : ARM::t2LDRSBi8;
1061 Opc = isZExt ? ARM::t2LDRBi12 : ARM::t2LDRSBi12;
1070 RC = isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRnopcRegClass;
1073 if (Alignment && Alignment < 2 && !Subtarget->allowsUnalignedMem())
1077 if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
1078 Opc = isZExt ? ARM::t2LDRHi8 : ARM::t2LDRSHi8;
1080 Opc = isZExt ? ARM::t2LDRHi12 : ARM::t2LDRSHi12;
1082 Opc = isZExt ? ARM::LDRH : ARM::LDRSH;
1085 RC = isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRnopcRegClass;
1088 if (Alignment && Alignment < 4 && !Subtarget->allowsUnalignedMem())
1092 if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
1095 Opc = ARM::t2LDRi12;
1099 RC = isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRnopcRegClass;
1102 if (!Subtarget->hasVFP2()) return false;
1103 // Unaligned loads need special handling. Floats require word-alignment.
1104 if (Alignment && Alignment < 4) {
1107 Opc = isThumb2 ? ARM::t2LDRi12 : ARM::LDRi12;
1108 RC = isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRnopcRegClass;
1111 RC = TLI.getRegClassFor(VT);
1115 if (!Subtarget->hasVFP2()) return false;
1116 // FIXME: Unaligned loads need special handling. Doublewords require
1118 if (Alignment && Alignment < 4)
1122 RC = TLI.getRegClassFor(VT);
1125 // Simplify this down to something we can handle.
1126 ARMSimplifyAddress(Addr, VT, useAM3);
1128 // Create the base instruction, then add the operands.
1130 ResultReg = createResultReg(RC);
1131 assert (ResultReg > 255 && "Expected an allocated virtual register.");
1132 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1133 TII.get(Opc), ResultReg);
1134 AddLoadStoreOperands(VT, Addr, MIB, MachineMemOperand::MOLoad, useAM3);
1136 // If we had an unaligned load of a float we've converted it to an regular
1137 // load. Now we must move from the GRP to the FP register.
1139 unsigned MoveReg = createResultReg(TLI.getRegClassFor(MVT::f32));
1140 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1141 TII.get(ARM::VMOVSR), MoveReg)
1142 .addReg(ResultReg));
1143 ResultReg = MoveReg;
1148 bool ARMFastISel::SelectLoad(const Instruction *I) {
1149 // Atomic loads need special handling.
1150 if (cast<LoadInst>(I)->isAtomic())
1153 // Verify we have a legal type before going any further.
1155 if (!isLoadTypeLegal(I->getType(), VT))
1158 // See if we can handle this address.
1160 if (!ARMComputeAddress(I->getOperand(0), Addr)) return false;
1163 if (!ARMEmitLoad(VT, ResultReg, Addr, cast<LoadInst>(I)->getAlignment()))
1165 UpdateValueMap(I, ResultReg);
1169 bool ARMFastISel::ARMEmitStore(MVT VT, unsigned SrcReg, Address &Addr,
1170 unsigned Alignment) {
1172 bool useAM3 = false;
1173 switch (VT.SimpleTy) {
1174 // This is mostly going to be Neon/vector support.
1175 default: return false;
1177 unsigned Res = createResultReg(isThumb2 ?
1178 (const TargetRegisterClass*)&ARM::tGPRRegClass :
1179 (const TargetRegisterClass*)&ARM::GPRRegClass);
1180 unsigned Opc = isThumb2 ? ARM::t2ANDri : ARM::ANDri;
1181 SrcReg = constrainOperandRegClass(TII.get(Opc), SrcReg, 1);
1182 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1184 .addReg(SrcReg).addImm(1));
1186 } // Fallthrough here.
1189 if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
1190 StrOpc = ARM::t2STRBi8;
1192 StrOpc = ARM::t2STRBi12;
1194 StrOpc = ARM::STRBi12;
1198 if (Alignment && Alignment < 2 && !Subtarget->allowsUnalignedMem())
1202 if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
1203 StrOpc = ARM::t2STRHi8;
1205 StrOpc = ARM::t2STRHi12;
1212 if (Alignment && Alignment < 4 && !Subtarget->allowsUnalignedMem())
1216 if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
1217 StrOpc = ARM::t2STRi8;
1219 StrOpc = ARM::t2STRi12;
1221 StrOpc = ARM::STRi12;
1225 if (!Subtarget->hasVFP2()) return false;
1226 // Unaligned stores need special handling. Floats require word-alignment.
1227 if (Alignment && Alignment < 4) {
1228 unsigned MoveReg = createResultReg(TLI.getRegClassFor(MVT::i32));
1229 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1230 TII.get(ARM::VMOVRS), MoveReg)
1234 StrOpc = isThumb2 ? ARM::t2STRi12 : ARM::STRi12;
1236 StrOpc = ARM::VSTRS;
1240 if (!Subtarget->hasVFP2()) return false;
1241 // FIXME: Unaligned stores need special handling. Doublewords require
1243 if (Alignment && Alignment < 4)
1246 StrOpc = ARM::VSTRD;
1249 // Simplify this down to something we can handle.
1250 ARMSimplifyAddress(Addr, VT, useAM3);
1252 // Create the base instruction, then add the operands.
1253 SrcReg = constrainOperandRegClass(TII.get(StrOpc), SrcReg, 0);
1254 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1257 AddLoadStoreOperands(VT, Addr, MIB, MachineMemOperand::MOStore, useAM3);
1261 bool ARMFastISel::SelectStore(const Instruction *I) {
1262 Value *Op0 = I->getOperand(0);
1263 unsigned SrcReg = 0;
1265 // Atomic stores need special handling.
1266 if (cast<StoreInst>(I)->isAtomic())
1269 // Verify we have a legal type before going any further.
1271 if (!isLoadTypeLegal(I->getOperand(0)->getType(), VT))
1274 // Get the value to be stored into a register.
1275 SrcReg = getRegForValue(Op0);
1276 if (SrcReg == 0) return false;
1278 // See if we can handle this address.
1280 if (!ARMComputeAddress(I->getOperand(1), Addr))
1283 if (!ARMEmitStore(VT, SrcReg, Addr, cast<StoreInst>(I)->getAlignment()))
1288 static ARMCC::CondCodes getComparePred(CmpInst::Predicate Pred) {
1290 // Needs two compares...
1291 case CmpInst::FCMP_ONE:
1292 case CmpInst::FCMP_UEQ:
1294 // AL is our "false" for now. The other two need more compares.
1296 case CmpInst::ICMP_EQ:
1297 case CmpInst::FCMP_OEQ:
1299 case CmpInst::ICMP_SGT:
1300 case CmpInst::FCMP_OGT:
1302 case CmpInst::ICMP_SGE:
1303 case CmpInst::FCMP_OGE:
1305 case CmpInst::ICMP_UGT:
1306 case CmpInst::FCMP_UGT:
1308 case CmpInst::FCMP_OLT:
1310 case CmpInst::ICMP_ULE:
1311 case CmpInst::FCMP_OLE:
1313 case CmpInst::FCMP_ORD:
1315 case CmpInst::FCMP_UNO:
1317 case CmpInst::FCMP_UGE:
1319 case CmpInst::ICMP_SLT:
1320 case CmpInst::FCMP_ULT:
1322 case CmpInst::ICMP_SLE:
1323 case CmpInst::FCMP_ULE:
1325 case CmpInst::FCMP_UNE:
1326 case CmpInst::ICMP_NE:
1328 case CmpInst::ICMP_UGE:
1330 case CmpInst::ICMP_ULT:
1335 bool ARMFastISel::SelectBranch(const Instruction *I) {
1336 const BranchInst *BI = cast<BranchInst>(I);
1337 MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
1338 MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
1340 // Simple branch support.
1342 // If we can, avoid recomputing the compare - redoing it could lead to wonky
1344 if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
1345 if (CI->hasOneUse() && (CI->getParent() == I->getParent())) {
1347 // Get the compare predicate.
1348 // Try to take advantage of fallthrough opportunities.
1349 CmpInst::Predicate Predicate = CI->getPredicate();
1350 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
1351 std::swap(TBB, FBB);
1352 Predicate = CmpInst::getInversePredicate(Predicate);
1355 ARMCC::CondCodes ARMPred = getComparePred(Predicate);
1357 // We may not handle every CC for now.
1358 if (ARMPred == ARMCC::AL) return false;
1360 // Emit the compare.
1361 if (!ARMEmitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
1364 unsigned BrOpc = isThumb2 ? ARM::t2Bcc : ARM::Bcc;
1365 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
1366 .addMBB(TBB).addImm(ARMPred).addReg(ARM::CPSR);
1367 FastEmitBranch(FBB, DL);
1368 FuncInfo.MBB->addSuccessor(TBB);
1371 } else if (TruncInst *TI = dyn_cast<TruncInst>(BI->getCondition())) {
1373 if (TI->hasOneUse() && TI->getParent() == I->getParent() &&
1374 (isLoadTypeLegal(TI->getOperand(0)->getType(), SourceVT))) {
1375 unsigned TstOpc = isThumb2 ? ARM::t2TSTri : ARM::TSTri;
1376 unsigned OpReg = getRegForValue(TI->getOperand(0));
1377 OpReg = constrainOperandRegClass(TII.get(TstOpc), OpReg, 0);
1378 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1380 .addReg(OpReg).addImm(1));
1382 unsigned CCMode = ARMCC::NE;
1383 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
1384 std::swap(TBB, FBB);
1388 unsigned BrOpc = isThumb2 ? ARM::t2Bcc : ARM::Bcc;
1389 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
1390 .addMBB(TBB).addImm(CCMode).addReg(ARM::CPSR);
1392 FastEmitBranch(FBB, DL);
1393 FuncInfo.MBB->addSuccessor(TBB);
1396 } else if (const ConstantInt *CI =
1397 dyn_cast<ConstantInt>(BI->getCondition())) {
1398 uint64_t Imm = CI->getZExtValue();
1399 MachineBasicBlock *Target = (Imm == 0) ? FBB : TBB;
1400 FastEmitBranch(Target, DL);
1404 unsigned CmpReg = getRegForValue(BI->getCondition());
1405 if (CmpReg == 0) return false;
1407 // We've been divorced from our compare! Our block was split, and
1408 // now our compare lives in a predecessor block. We musn't
1409 // re-compare here, as the children of the compare aren't guaranteed
1410 // live across the block boundary (we *could* check for this).
1411 // Regardless, the compare has been done in the predecessor block,
1412 // and it left a value for us in a virtual register. Ergo, we test
1413 // the one-bit value left in the virtual register.
1414 unsigned TstOpc = isThumb2 ? ARM::t2TSTri : ARM::TSTri;
1415 CmpReg = constrainOperandRegClass(TII.get(TstOpc), CmpReg, 0);
1416 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TstOpc))
1417 .addReg(CmpReg).addImm(1));
1419 unsigned CCMode = ARMCC::NE;
1420 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
1421 std::swap(TBB, FBB);
1425 unsigned BrOpc = isThumb2 ? ARM::t2Bcc : ARM::Bcc;
1426 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
1427 .addMBB(TBB).addImm(CCMode).addReg(ARM::CPSR);
1428 FastEmitBranch(FBB, DL);
1429 FuncInfo.MBB->addSuccessor(TBB);
1433 bool ARMFastISel::SelectIndirectBr(const Instruction *I) {
1434 unsigned AddrReg = getRegForValue(I->getOperand(0));
1435 if (AddrReg == 0) return false;
1437 unsigned Opc = isThumb2 ? ARM::tBRIND : ARM::BX;
1438 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc))
1441 const IndirectBrInst *IB = cast<IndirectBrInst>(I);
1442 for (unsigned i = 0, e = IB->getNumSuccessors(); i != e; ++i)
1443 FuncInfo.MBB->addSuccessor(FuncInfo.MBBMap[IB->getSuccessor(i)]);
1448 bool ARMFastISel::ARMEmitCmp(const Value *Src1Value, const Value *Src2Value,
1450 Type *Ty = Src1Value->getType();
1451 EVT SrcEVT = TLI.getValueType(Ty, true);
1452 if (!SrcEVT.isSimple()) return false;
1453 MVT SrcVT = SrcEVT.getSimpleVT();
1455 bool isFloat = (Ty->isFloatTy() || Ty->isDoubleTy());
1456 if (isFloat && !Subtarget->hasVFP2())
1459 // Check to see if the 2nd operand is a constant that we can encode directly
1462 bool UseImm = false;
1463 bool isNegativeImm = false;
1464 // FIXME: At -O0 we don't have anything that canonicalizes operand order.
1465 // Thus, Src1Value may be a ConstantInt, but we're missing it.
1466 if (const ConstantInt *ConstInt = dyn_cast<ConstantInt>(Src2Value)) {
1467 if (SrcVT == MVT::i32 || SrcVT == MVT::i16 || SrcVT == MVT::i8 ||
1469 const APInt &CIVal = ConstInt->getValue();
1470 Imm = (isZExt) ? (int)CIVal.getZExtValue() : (int)CIVal.getSExtValue();
1471 // For INT_MIN/LONG_MIN (i.e., 0x80000000) we need to use a cmp, rather
1472 // then a cmn, because there is no way to represent 2147483648 as a
1473 // signed 32-bit int.
1474 if (Imm < 0 && Imm != (int)0x80000000) {
1475 isNegativeImm = true;
1478 UseImm = isThumb2 ? (ARM_AM::getT2SOImmVal(Imm) != -1) :
1479 (ARM_AM::getSOImmVal(Imm) != -1);
1481 } else if (const ConstantFP *ConstFP = dyn_cast<ConstantFP>(Src2Value)) {
1482 if (SrcVT == MVT::f32 || SrcVT == MVT::f64)
1483 if (ConstFP->isZero() && !ConstFP->isNegative())
1489 bool needsExt = false;
1490 switch (SrcVT.SimpleTy) {
1491 default: return false;
1492 // TODO: Verify compares.
1495 CmpOpc = UseImm ? ARM::VCMPEZS : ARM::VCMPES;
1499 CmpOpc = UseImm ? ARM::VCMPEZD : ARM::VCMPED;
1505 // Intentional fall-through.
1509 CmpOpc = ARM::t2CMPrr;
1511 CmpOpc = isNegativeImm ? ARM::t2CMNri : ARM::t2CMPri;
1514 CmpOpc = ARM::CMPrr;
1516 CmpOpc = isNegativeImm ? ARM::CMNri : ARM::CMPri;
1521 unsigned SrcReg1 = getRegForValue(Src1Value);
1522 if (SrcReg1 == 0) return false;
1524 unsigned SrcReg2 = 0;
1526 SrcReg2 = getRegForValue(Src2Value);
1527 if (SrcReg2 == 0) return false;
1530 // We have i1, i8, or i16, we need to either zero extend or sign extend.
1532 SrcReg1 = ARMEmitIntExt(SrcVT, SrcReg1, MVT::i32, isZExt);
1533 if (SrcReg1 == 0) return false;
1535 SrcReg2 = ARMEmitIntExt(SrcVT, SrcReg2, MVT::i32, isZExt);
1536 if (SrcReg2 == 0) return false;
1540 const MCInstrDesc &II = TII.get(CmpOpc);
1541 SrcReg1 = constrainOperandRegClass(II, SrcReg1, 0);
1543 SrcReg2 = constrainOperandRegClass(II, SrcReg2, 1);
1544 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1545 .addReg(SrcReg1).addReg(SrcReg2));
1547 MachineInstrBuilder MIB;
1548 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1551 // Only add immediate for icmp as the immediate for fcmp is an implicit 0.0.
1554 AddOptionalDefs(MIB);
1557 // For floating point we need to move the result to a comparison register
1558 // that we can then use for branches.
1559 if (Ty->isFloatTy() || Ty->isDoubleTy())
1560 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1561 TII.get(ARM::FMSTAT)));
1565 bool ARMFastISel::SelectCmp(const Instruction *I) {
1566 const CmpInst *CI = cast<CmpInst>(I);
1568 // Get the compare predicate.
1569 ARMCC::CondCodes ARMPred = getComparePred(CI->getPredicate());
1571 // We may not handle every CC for now.
1572 if (ARMPred == ARMCC::AL) return false;
1574 // Emit the compare.
1575 if (!ARMEmitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
1578 // Now set a register based on the comparison. Explicitly set the predicates
1580 unsigned MovCCOpc = isThumb2 ? ARM::t2MOVCCi : ARM::MOVCCi;
1581 const TargetRegisterClass *RC = isThumb2 ?
1582 (const TargetRegisterClass*)&ARM::rGPRRegClass :
1583 (const TargetRegisterClass*)&ARM::GPRRegClass;
1584 unsigned DestReg = createResultReg(RC);
1585 Constant *Zero = ConstantInt::get(Type::getInt32Ty(*Context), 0);
1586 unsigned ZeroReg = TargetMaterializeConstant(Zero);
1587 // ARMEmitCmp emits a FMSTAT when necessary, so it's always safe to use CPSR.
1588 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), DestReg)
1589 .addReg(ZeroReg).addImm(1)
1590 .addImm(ARMPred).addReg(ARM::CPSR);
1592 UpdateValueMap(I, DestReg);
1596 bool ARMFastISel::SelectFPExt(const Instruction *I) {
1597 // Make sure we have VFP and that we're extending float to double.
1598 if (!Subtarget->hasVFP2()) return false;
1600 Value *V = I->getOperand(0);
1601 if (!I->getType()->isDoubleTy() ||
1602 !V->getType()->isFloatTy()) return false;
1604 unsigned Op = getRegForValue(V);
1605 if (Op == 0) return false;
1607 unsigned Result = createResultReg(&ARM::DPRRegClass);
1608 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1609 TII.get(ARM::VCVTDS), Result)
1611 UpdateValueMap(I, Result);
1615 bool ARMFastISel::SelectFPTrunc(const Instruction *I) {
1616 // Make sure we have VFP and that we're truncating double to float.
1617 if (!Subtarget->hasVFP2()) return false;
1619 Value *V = I->getOperand(0);
1620 if (!(I->getType()->isFloatTy() &&
1621 V->getType()->isDoubleTy())) return false;
1623 unsigned Op = getRegForValue(V);
1624 if (Op == 0) return false;
1626 unsigned Result = createResultReg(&ARM::SPRRegClass);
1627 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1628 TII.get(ARM::VCVTSD), Result)
1630 UpdateValueMap(I, Result);
1634 bool ARMFastISel::SelectIToFP(const Instruction *I, bool isSigned) {
1635 // Make sure we have VFP.
1636 if (!Subtarget->hasVFP2()) return false;
1639 Type *Ty = I->getType();
1640 if (!isTypeLegal(Ty, DstVT))
1643 Value *Src = I->getOperand(0);
1644 EVT SrcEVT = TLI.getValueType(Src->getType(), true);
1645 if (!SrcEVT.isSimple())
1647 MVT SrcVT = SrcEVT.getSimpleVT();
1648 if (SrcVT != MVT::i32 && SrcVT != MVT::i16 && SrcVT != MVT::i8)
1651 unsigned SrcReg = getRegForValue(Src);
1652 if (SrcReg == 0) return false;
1654 // Handle sign-extension.
1655 if (SrcVT == MVT::i16 || SrcVT == MVT::i8) {
1656 SrcReg = ARMEmitIntExt(SrcVT, SrcReg, MVT::i32,
1657 /*isZExt*/!isSigned);
1658 if (SrcReg == 0) return false;
1661 // The conversion routine works on fp-reg to fp-reg and the operand above
1662 // was an integer, move it to the fp registers if possible.
1663 unsigned FP = ARMMoveToFPReg(MVT::f32, SrcReg);
1664 if (FP == 0) return false;
1667 if (Ty->isFloatTy()) Opc = isSigned ? ARM::VSITOS : ARM::VUITOS;
1668 else if (Ty->isDoubleTy()) Opc = isSigned ? ARM::VSITOD : ARM::VUITOD;
1671 unsigned ResultReg = createResultReg(TLI.getRegClassFor(DstVT));
1672 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
1675 UpdateValueMap(I, ResultReg);
1679 bool ARMFastISel::SelectFPToI(const Instruction *I, bool isSigned) {
1680 // Make sure we have VFP.
1681 if (!Subtarget->hasVFP2()) return false;
1684 Type *RetTy = I->getType();
1685 if (!isTypeLegal(RetTy, DstVT))
1688 unsigned Op = getRegForValue(I->getOperand(0));
1689 if (Op == 0) return false;
1692 Type *OpTy = I->getOperand(0)->getType();
1693 if (OpTy->isFloatTy()) Opc = isSigned ? ARM::VTOSIZS : ARM::VTOUIZS;
1694 else if (OpTy->isDoubleTy()) Opc = isSigned ? ARM::VTOSIZD : ARM::VTOUIZD;
1697 // f64->s32/u32 or f32->s32/u32 both need an intermediate f32 reg.
1698 unsigned ResultReg = createResultReg(TLI.getRegClassFor(MVT::f32));
1699 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
1703 // This result needs to be in an integer register, but the conversion only
1704 // takes place in fp-regs.
1705 unsigned IntReg = ARMMoveToIntReg(DstVT, ResultReg);
1706 if (IntReg == 0) return false;
1708 UpdateValueMap(I, IntReg);
1712 bool ARMFastISel::SelectSelect(const Instruction *I) {
1714 if (!isTypeLegal(I->getType(), VT))
1717 // Things need to be register sized for register moves.
1718 if (VT != MVT::i32) return false;
1720 unsigned CondReg = getRegForValue(I->getOperand(0));
1721 if (CondReg == 0) return false;
1722 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1723 if (Op1Reg == 0) return false;
1725 // Check to see if we can use an immediate in the conditional move.
1727 bool UseImm = false;
1728 bool isNegativeImm = false;
1729 if (const ConstantInt *ConstInt = dyn_cast<ConstantInt>(I->getOperand(2))) {
1730 assert (VT == MVT::i32 && "Expecting an i32.");
1731 Imm = (int)ConstInt->getValue().getZExtValue();
1733 isNegativeImm = true;
1736 UseImm = isThumb2 ? (ARM_AM::getT2SOImmVal(Imm) != -1) :
1737 (ARM_AM::getSOImmVal(Imm) != -1);
1740 unsigned Op2Reg = 0;
1742 Op2Reg = getRegForValue(I->getOperand(2));
1743 if (Op2Reg == 0) return false;
1746 unsigned CmpOpc = isThumb2 ? ARM::t2CMPri : ARM::CMPri;
1747 CondReg = constrainOperandRegClass(TII.get(CmpOpc), CondReg, 0);
1748 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
1749 .addReg(CondReg).addImm(0));
1752 const TargetRegisterClass *RC;
1754 RC = isThumb2 ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
1755 MovCCOpc = isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr;
1757 RC = isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRRegClass;
1759 MovCCOpc = isThumb2 ? ARM::t2MOVCCi : ARM::MOVCCi;
1761 MovCCOpc = isThumb2 ? ARM::t2MVNCCi : ARM::MVNCCi;
1763 unsigned ResultReg = createResultReg(RC);
1765 Op2Reg = constrainOperandRegClass(TII.get(MovCCOpc), Op2Reg, 1);
1766 Op1Reg = constrainOperandRegClass(TII.get(MovCCOpc), Op1Reg, 2);
1767 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), ResultReg)
1768 .addReg(Op2Reg).addReg(Op1Reg).addImm(ARMCC::NE).addReg(ARM::CPSR);
1770 Op1Reg = constrainOperandRegClass(TII.get(MovCCOpc), Op1Reg, 1);
1771 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), ResultReg)
1772 .addReg(Op1Reg).addImm(Imm).addImm(ARMCC::EQ).addReg(ARM::CPSR);
1774 UpdateValueMap(I, ResultReg);
1778 bool ARMFastISel::SelectDiv(const Instruction *I, bool isSigned) {
1780 Type *Ty = I->getType();
1781 if (!isTypeLegal(Ty, VT))
1784 // If we have integer div support we should have selected this automagically.
1785 // In case we have a real miss go ahead and return false and we'll pick
1787 if (Subtarget->hasDivide()) return false;
1789 // Otherwise emit a libcall.
1790 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
1792 LC = isSigned ? RTLIB::SDIV_I8 : RTLIB::UDIV_I8;
1793 else if (VT == MVT::i16)
1794 LC = isSigned ? RTLIB::SDIV_I16 : RTLIB::UDIV_I16;
1795 else if (VT == MVT::i32)
1796 LC = isSigned ? RTLIB::SDIV_I32 : RTLIB::UDIV_I32;
1797 else if (VT == MVT::i64)
1798 LC = isSigned ? RTLIB::SDIV_I64 : RTLIB::UDIV_I64;
1799 else if (VT == MVT::i128)
1800 LC = isSigned ? RTLIB::SDIV_I128 : RTLIB::UDIV_I128;
1801 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SDIV!");
1803 return ARMEmitLibcall(I, LC);
1806 bool ARMFastISel::SelectRem(const Instruction *I, bool isSigned) {
1808 Type *Ty = I->getType();
1809 if (!isTypeLegal(Ty, VT))
1812 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
1814 LC = isSigned ? RTLIB::SREM_I8 : RTLIB::UREM_I8;
1815 else if (VT == MVT::i16)
1816 LC = isSigned ? RTLIB::SREM_I16 : RTLIB::UREM_I16;
1817 else if (VT == MVT::i32)
1818 LC = isSigned ? RTLIB::SREM_I32 : RTLIB::UREM_I32;
1819 else if (VT == MVT::i64)
1820 LC = isSigned ? RTLIB::SREM_I64 : RTLIB::UREM_I64;
1821 else if (VT == MVT::i128)
1822 LC = isSigned ? RTLIB::SREM_I128 : RTLIB::UREM_I128;
1823 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SREM!");
1825 return ARMEmitLibcall(I, LC);
1828 bool ARMFastISel::SelectBinaryIntOp(const Instruction *I, unsigned ISDOpcode) {
1829 EVT DestVT = TLI.getValueType(I->getType(), true);
1831 // We can get here in the case when we have a binary operation on a non-legal
1832 // type and the target independent selector doesn't know how to handle it.
1833 if (DestVT != MVT::i16 && DestVT != MVT::i8 && DestVT != MVT::i1)
1837 switch (ISDOpcode) {
1838 default: return false;
1840 Opc = isThumb2 ? ARM::t2ADDrr : ARM::ADDrr;
1843 Opc = isThumb2 ? ARM::t2ORRrr : ARM::ORRrr;
1846 Opc = isThumb2 ? ARM::t2SUBrr : ARM::SUBrr;
1850 unsigned SrcReg1 = getRegForValue(I->getOperand(0));
1851 if (SrcReg1 == 0) return false;
1853 // TODO: Often the 2nd operand is an immediate, which can be encoded directly
1854 // in the instruction, rather then materializing the value in a register.
1855 unsigned SrcReg2 = getRegForValue(I->getOperand(1));
1856 if (SrcReg2 == 0) return false;
1858 unsigned ResultReg = createResultReg(&ARM::GPRnopcRegClass);
1859 SrcReg1 = constrainOperandRegClass(TII.get(Opc), SrcReg1, 1);
1860 SrcReg2 = constrainOperandRegClass(TII.get(Opc), SrcReg2, 2);
1861 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1862 TII.get(Opc), ResultReg)
1863 .addReg(SrcReg1).addReg(SrcReg2));
1864 UpdateValueMap(I, ResultReg);
1868 bool ARMFastISel::SelectBinaryFPOp(const Instruction *I, unsigned ISDOpcode) {
1869 EVT FPVT = TLI.getValueType(I->getType(), true);
1870 if (!FPVT.isSimple()) return false;
1871 MVT VT = FPVT.getSimpleVT();
1873 // We can get here in the case when we want to use NEON for our fp
1874 // operations, but can't figure out how to. Just use the vfp instructions
1876 // FIXME: It'd be nice to use NEON instructions.
1877 Type *Ty = I->getType();
1878 bool isFloat = (Ty->isDoubleTy() || Ty->isFloatTy());
1879 if (isFloat && !Subtarget->hasVFP2())
1883 bool is64bit = VT == MVT::f64 || VT == MVT::i64;
1884 switch (ISDOpcode) {
1885 default: return false;
1887 Opc = is64bit ? ARM::VADDD : ARM::VADDS;
1890 Opc = is64bit ? ARM::VSUBD : ARM::VSUBS;
1893 Opc = is64bit ? ARM::VMULD : ARM::VMULS;
1896 unsigned Op1 = getRegForValue(I->getOperand(0));
1897 if (Op1 == 0) return false;
1899 unsigned Op2 = getRegForValue(I->getOperand(1));
1900 if (Op2 == 0) return false;
1902 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT.SimpleTy));
1903 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1904 TII.get(Opc), ResultReg)
1905 .addReg(Op1).addReg(Op2));
1906 UpdateValueMap(I, ResultReg);
1910 // Call Handling Code
1912 // This is largely taken directly from CCAssignFnForNode
1913 // TODO: We may not support all of this.
1914 CCAssignFn *ARMFastISel::CCAssignFnForCall(CallingConv::ID CC,
1919 llvm_unreachable("Unsupported calling convention");
1920 case CallingConv::Fast:
1921 if (Subtarget->hasVFP2() && !isVarArg) {
1922 if (!Subtarget->isAAPCS_ABI())
1923 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1924 // For AAPCS ABI targets, just use VFP variant of the calling convention.
1925 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1928 case CallingConv::C:
1929 // Use target triple & subtarget features to do actual dispatch.
1930 if (Subtarget->isAAPCS_ABI()) {
1931 if (Subtarget->hasVFP2() &&
1932 TM.Options.FloatABIType == FloatABI::Hard && !isVarArg)
1933 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
1935 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
1937 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
1938 case CallingConv::ARM_AAPCS_VFP:
1940 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
1941 // Fall through to soft float variant, variadic functions don't
1942 // use hard floating point ABI.
1943 case CallingConv::ARM_AAPCS:
1944 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
1945 case CallingConv::ARM_APCS:
1946 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
1947 case CallingConv::GHC:
1949 llvm_unreachable("Can't return in GHC call convention");
1951 return CC_ARM_APCS_GHC;
1955 bool ARMFastISel::ProcessCallArgs(SmallVectorImpl<Value*> &Args,
1956 SmallVectorImpl<unsigned> &ArgRegs,
1957 SmallVectorImpl<MVT> &ArgVTs,
1958 SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
1959 SmallVectorImpl<unsigned> &RegArgs,
1963 SmallVector<CCValAssign, 16> ArgLocs;
1964 CCState CCInfo(CC, isVarArg, *FuncInfo.MF, TM, ArgLocs, *Context);
1965 CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags,
1966 CCAssignFnForCall(CC, false, isVarArg));
1968 // Check that we can handle all of the arguments. If we can't, then bail out
1969 // now before we add code to the MBB.
1970 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1971 CCValAssign &VA = ArgLocs[i];
1972 MVT ArgVT = ArgVTs[VA.getValNo()];
1974 // We don't handle NEON/vector parameters yet.
1975 if (ArgVT.isVector() || ArgVT.getSizeInBits() > 64)
1978 // Now copy/store arg to correct locations.
1979 if (VA.isRegLoc() && !VA.needsCustom()) {
1981 } else if (VA.needsCustom()) {
1982 // TODO: We need custom lowering for vector (v2f64) args.
1983 if (VA.getLocVT() != MVT::f64 ||
1984 // TODO: Only handle register args for now.
1985 !VA.isRegLoc() || !ArgLocs[++i].isRegLoc())
1988 switch (ArgVT.SimpleTy) {
1997 if (!Subtarget->hasVFP2())
2001 if (!Subtarget->hasVFP2())
2008 // At the point, we are able to handle the call's arguments in fast isel.
2010 // Get a count of how many bytes are to be pushed on the stack.
2011 NumBytes = CCInfo.getNextStackOffset();
2013 // Issue CALLSEQ_START
2014 unsigned AdjStackDown = TII.getCallFrameSetupOpcode();
2015 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
2016 TII.get(AdjStackDown))
2019 // Process the args.
2020 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2021 CCValAssign &VA = ArgLocs[i];
2022 unsigned Arg = ArgRegs[VA.getValNo()];
2023 MVT ArgVT = ArgVTs[VA.getValNo()];
2025 assert((!ArgVT.isVector() && ArgVT.getSizeInBits() <= 64) &&
2026 "We don't handle NEON/vector parameters yet.");
2028 // Handle arg promotion, etc.
2029 switch (VA.getLocInfo()) {
2030 case CCValAssign::Full: break;
2031 case CCValAssign::SExt: {
2032 MVT DestVT = VA.getLocVT();
2033 Arg = ARMEmitIntExt(ArgVT, Arg, DestVT, /*isZExt*/false);
2034 assert (Arg != 0 && "Failed to emit a sext");
2038 case CCValAssign::AExt:
2039 // Intentional fall-through. Handle AExt and ZExt.
2040 case CCValAssign::ZExt: {
2041 MVT DestVT = VA.getLocVT();
2042 Arg = ARMEmitIntExt(ArgVT, Arg, DestVT, /*isZExt*/true);
2043 assert (Arg != 0 && "Failed to emit a zext");
2047 case CCValAssign::BCvt: {
2048 unsigned BC = FastEmit_r(ArgVT, VA.getLocVT(), ISD::BITCAST, Arg,
2049 /*TODO: Kill=*/false);
2050 assert(BC != 0 && "Failed to emit a bitcast!");
2052 ArgVT = VA.getLocVT();
2055 default: llvm_unreachable("Unknown arg promotion!");
2058 // Now copy/store arg to correct locations.
2059 if (VA.isRegLoc() && !VA.needsCustom()) {
2060 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
2063 RegArgs.push_back(VA.getLocReg());
2064 } else if (VA.needsCustom()) {
2065 // TODO: We need custom lowering for vector (v2f64) args.
2066 assert(VA.getLocVT() == MVT::f64 &&
2067 "Custom lowering for v2f64 args not available");
2069 CCValAssign &NextVA = ArgLocs[++i];
2071 assert(VA.isRegLoc() && NextVA.isRegLoc() &&
2072 "We only handle register args!");
2074 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
2075 TII.get(ARM::VMOVRRD), VA.getLocReg())
2076 .addReg(NextVA.getLocReg(), RegState::Define)
2078 RegArgs.push_back(VA.getLocReg());
2079 RegArgs.push_back(NextVA.getLocReg());
2081 assert(VA.isMemLoc());
2082 // Need to store on the stack.
2084 Addr.BaseType = Address::RegBase;
2085 Addr.Base.Reg = ARM::SP;
2086 Addr.Offset = VA.getLocMemOffset();
2088 bool EmitRet = ARMEmitStore(ArgVT, Arg, Addr); (void)EmitRet;
2089 assert(EmitRet && "Could not emit a store for argument!");
2096 bool ARMFastISel::FinishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
2097 const Instruction *I, CallingConv::ID CC,
2098 unsigned &NumBytes, bool isVarArg) {
2099 // Issue CALLSEQ_END
2100 unsigned AdjStackUp = TII.getCallFrameDestroyOpcode();
2101 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
2102 TII.get(AdjStackUp))
2103 .addImm(NumBytes).addImm(0));
2105 // Now the return value.
2106 if (RetVT != MVT::isVoid) {
2107 SmallVector<CCValAssign, 16> RVLocs;
2108 CCState CCInfo(CC, isVarArg, *FuncInfo.MF, TM, RVLocs, *Context);
2109 CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC, true, isVarArg));
2111 // Copy all of the result registers out of their specified physreg.
2112 if (RVLocs.size() == 2 && RetVT == MVT::f64) {
2113 // For this move we copy into two registers and then move into the
2114 // double fp reg we want.
2115 MVT DestVT = RVLocs[0].getValVT();
2116 const TargetRegisterClass* DstRC = TLI.getRegClassFor(DestVT);
2117 unsigned ResultReg = createResultReg(DstRC);
2118 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
2119 TII.get(ARM::VMOVDRR), ResultReg)
2120 .addReg(RVLocs[0].getLocReg())
2121 .addReg(RVLocs[1].getLocReg()));
2123 UsedRegs.push_back(RVLocs[0].getLocReg());
2124 UsedRegs.push_back(RVLocs[1].getLocReg());
2126 // Finally update the result.
2127 UpdateValueMap(I, ResultReg);
2129 assert(RVLocs.size() == 1 &&"Can't handle non-double multi-reg retvals!");
2130 MVT CopyVT = RVLocs[0].getValVT();
2132 // Special handling for extended integers.
2133 if (RetVT == MVT::i1 || RetVT == MVT::i8 || RetVT == MVT::i16)
2136 const TargetRegisterClass* DstRC = TLI.getRegClassFor(CopyVT);
2138 unsigned ResultReg = createResultReg(DstRC);
2139 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
2140 ResultReg).addReg(RVLocs[0].getLocReg());
2141 UsedRegs.push_back(RVLocs[0].getLocReg());
2143 // Finally update the result.
2144 UpdateValueMap(I, ResultReg);
2151 bool ARMFastISel::SelectRet(const Instruction *I) {
2152 const ReturnInst *Ret = cast<ReturnInst>(I);
2153 const Function &F = *I->getParent()->getParent();
2155 if (!FuncInfo.CanLowerReturn)
2158 // Build a list of return value registers.
2159 SmallVector<unsigned, 4> RetRegs;
2161 CallingConv::ID CC = F.getCallingConv();
2162 if (Ret->getNumOperands() > 0) {
2163 SmallVector<ISD::OutputArg, 4> Outs;
2164 GetReturnInfo(F.getReturnType(), F.getAttributes(), Outs, TLI);
2166 // Analyze operands of the call, assigning locations to each operand.
2167 SmallVector<CCValAssign, 16> ValLocs;
2168 CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, TM, ValLocs,I->getContext());
2169 CCInfo.AnalyzeReturn(Outs, CCAssignFnForCall(CC, true /* is Ret */,
2172 const Value *RV = Ret->getOperand(0);
2173 unsigned Reg = getRegForValue(RV);
2177 // Only handle a single return value for now.
2178 if (ValLocs.size() != 1)
2181 CCValAssign &VA = ValLocs[0];
2183 // Don't bother handling odd stuff for now.
2184 if (VA.getLocInfo() != CCValAssign::Full)
2186 // Only handle register returns for now.
2190 unsigned SrcReg = Reg + VA.getValNo();
2191 EVT RVEVT = TLI.getValueType(RV->getType());
2192 if (!RVEVT.isSimple()) return false;
2193 MVT RVVT = RVEVT.getSimpleVT();
2194 MVT DestVT = VA.getValVT();
2195 // Special handling for extended integers.
2196 if (RVVT != DestVT) {
2197 if (RVVT != MVT::i1 && RVVT != MVT::i8 && RVVT != MVT::i16)
2200 assert(DestVT == MVT::i32 && "ARM should always ext to i32");
2202 // Perform extension if flagged as either zext or sext. Otherwise, do
2204 if (Outs[0].Flags.isZExt() || Outs[0].Flags.isSExt()) {
2205 SrcReg = ARMEmitIntExt(RVVT, SrcReg, DestVT, Outs[0].Flags.isZExt());
2206 if (SrcReg == 0) return false;
2211 unsigned DstReg = VA.getLocReg();
2212 const TargetRegisterClass* SrcRC = MRI.getRegClass(SrcReg);
2213 // Avoid a cross-class copy. This is very unlikely.
2214 if (!SrcRC->contains(DstReg))
2216 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
2217 DstReg).addReg(SrcReg);
2219 // Add register to return instruction.
2220 RetRegs.push_back(VA.getLocReg());
2223 unsigned RetOpc = isThumb2 ? ARM::tBX_RET : ARM::BX_RET;
2224 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
2226 AddOptionalDefs(MIB);
2227 for (unsigned i = 0, e = RetRegs.size(); i != e; ++i)
2228 MIB.addReg(RetRegs[i], RegState::Implicit);
2232 unsigned ARMFastISel::ARMSelectCallOp(bool UseReg) {
2234 return isThumb2 ? ARM::tBLXr : ARM::BLX;
2236 return isThumb2 ? ARM::tBL : ARM::BL;
2239 unsigned ARMFastISel::getLibcallReg(const Twine &Name) {
2240 // Manually compute the global's type to avoid building it when unnecessary.
2241 Type *GVTy = Type::getInt32PtrTy(*Context, /*AS=*/0);
2242 EVT LCREVT = TLI.getValueType(GVTy);
2243 if (!LCREVT.isSimple()) return 0;
2245 GlobalValue *GV = new GlobalVariable(Type::getInt32Ty(*Context), false,
2246 GlobalValue::ExternalLinkage, 0, Name);
2247 assert(GV->getType() == GVTy && "We miscomputed the type for the global!");
2248 return ARMMaterializeGV(GV, LCREVT.getSimpleVT());
2251 // A quick function that will emit a call for a named libcall in F with the
2252 // vector of passed arguments for the Instruction in I. We can assume that we
2253 // can emit a call for any libcall we can produce. This is an abridged version
2254 // of the full call infrastructure since we won't need to worry about things
2255 // like computed function pointers or strange arguments at call sites.
2256 // TODO: Try to unify this and the normal call bits for ARM, then try to unify
2258 bool ARMFastISel::ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call) {
2259 CallingConv::ID CC = TLI.getLibcallCallingConv(Call);
2261 // Handle *simple* calls for now.
2262 Type *RetTy = I->getType();
2264 if (RetTy->isVoidTy())
2265 RetVT = MVT::isVoid;
2266 else if (!isTypeLegal(RetTy, RetVT))
2269 // Can't handle non-double multi-reg retvals.
2270 if (RetVT != MVT::isVoid && RetVT != MVT::i32) {
2271 SmallVector<CCValAssign, 16> RVLocs;
2272 CCState CCInfo(CC, false, *FuncInfo.MF, TM, RVLocs, *Context);
2273 CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC, true, false));
2274 if (RVLocs.size() >= 2 && RetVT != MVT::f64)
2278 // Set up the argument vectors.
2279 SmallVector<Value*, 8> Args;
2280 SmallVector<unsigned, 8> ArgRegs;
2281 SmallVector<MVT, 8> ArgVTs;
2282 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
2283 Args.reserve(I->getNumOperands());
2284 ArgRegs.reserve(I->getNumOperands());
2285 ArgVTs.reserve(I->getNumOperands());
2286 ArgFlags.reserve(I->getNumOperands());
2287 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
2288 Value *Op = I->getOperand(i);
2289 unsigned Arg = getRegForValue(Op);
2290 if (Arg == 0) return false;
2292 Type *ArgTy = Op->getType();
2294 if (!isTypeLegal(ArgTy, ArgVT)) return false;
2296 ISD::ArgFlagsTy Flags;
2297 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
2298 Flags.setOrigAlign(OriginalAlignment);
2301 ArgRegs.push_back(Arg);
2302 ArgVTs.push_back(ArgVT);
2303 ArgFlags.push_back(Flags);
2306 // Handle the arguments now that we've gotten them.
2307 SmallVector<unsigned, 4> RegArgs;
2309 if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags,
2310 RegArgs, CC, NumBytes, false))
2313 unsigned CalleeReg = 0;
2314 if (EnableARMLongCalls) {
2315 CalleeReg = getLibcallReg(TLI.getLibcallName(Call));
2316 if (CalleeReg == 0) return false;
2320 unsigned CallOpc = ARMSelectCallOp(EnableARMLongCalls);
2321 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
2322 DL, TII.get(CallOpc));
2323 // BL / BLX don't take a predicate, but tBL / tBLX do.
2325 AddDefaultPred(MIB);
2326 if (EnableARMLongCalls)
2327 MIB.addReg(CalleeReg);
2329 MIB.addExternalSymbol(TLI.getLibcallName(Call));
2331 // Add implicit physical register uses to the call.
2332 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
2333 MIB.addReg(RegArgs[i], RegState::Implicit);
2335 // Add a register mask with the call-preserved registers.
2336 // Proper defs for return values will be added by setPhysRegsDeadExcept().
2337 MIB.addRegMask(TRI.getCallPreservedMask(CC));
2339 // Finish off the call including any return values.
2340 SmallVector<unsigned, 4> UsedRegs;
2341 if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes, false)) return false;
2343 // Set all unused physreg defs as dead.
2344 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
2349 bool ARMFastISel::SelectCall(const Instruction *I,
2350 const char *IntrMemName = 0) {
2351 const CallInst *CI = cast<CallInst>(I);
2352 const Value *Callee = CI->getCalledValue();
2354 // Can't handle inline asm.
2355 if (isa<InlineAsm>(Callee)) return false;
2357 // Allow SelectionDAG isel to handle tail calls.
2358 if (CI->isTailCall()) return false;
2360 // Check the calling convention.
2361 ImmutableCallSite CS(CI);
2362 CallingConv::ID CC = CS.getCallingConv();
2364 // TODO: Avoid some calling conventions?
2366 PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
2367 FunctionType *FTy = cast<FunctionType>(PT->getElementType());
2368 bool isVarArg = FTy->isVarArg();
2370 // Handle *simple* calls for now.
2371 Type *RetTy = I->getType();
2373 if (RetTy->isVoidTy())
2374 RetVT = MVT::isVoid;
2375 else if (!isTypeLegal(RetTy, RetVT) && RetVT != MVT::i16 &&
2376 RetVT != MVT::i8 && RetVT != MVT::i1)
2379 // Can't handle non-double multi-reg retvals.
2380 if (RetVT != MVT::isVoid && RetVT != MVT::i1 && RetVT != MVT::i8 &&
2381 RetVT != MVT::i16 && RetVT != MVT::i32) {
2382 SmallVector<CCValAssign, 16> RVLocs;
2383 CCState CCInfo(CC, isVarArg, *FuncInfo.MF, TM, RVLocs, *Context);
2384 CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC, true, isVarArg));
2385 if (RVLocs.size() >= 2 && RetVT != MVT::f64)
2389 // Set up the argument vectors.
2390 SmallVector<Value*, 8> Args;
2391 SmallVector<unsigned, 8> ArgRegs;
2392 SmallVector<MVT, 8> ArgVTs;
2393 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
2394 unsigned arg_size = CS.arg_size();
2395 Args.reserve(arg_size);
2396 ArgRegs.reserve(arg_size);
2397 ArgVTs.reserve(arg_size);
2398 ArgFlags.reserve(arg_size);
2399 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
2401 // If we're lowering a memory intrinsic instead of a regular call, skip the
2402 // last two arguments, which shouldn't be passed to the underlying function.
2403 if (IntrMemName && e-i <= 2)
2406 ISD::ArgFlagsTy Flags;
2407 unsigned AttrInd = i - CS.arg_begin() + 1;
2408 if (CS.paramHasAttr(AttrInd, Attribute::SExt))
2410 if (CS.paramHasAttr(AttrInd, Attribute::ZExt))
2413 // FIXME: Only handle *easy* calls for now.
2414 if (CS.paramHasAttr(AttrInd, Attribute::InReg) ||
2415 CS.paramHasAttr(AttrInd, Attribute::StructRet) ||
2416 CS.paramHasAttr(AttrInd, Attribute::Nest) ||
2417 CS.paramHasAttr(AttrInd, Attribute::ByVal))
2420 Type *ArgTy = (*i)->getType();
2422 if (!isTypeLegal(ArgTy, ArgVT) && ArgVT != MVT::i16 && ArgVT != MVT::i8 &&
2426 unsigned Arg = getRegForValue(*i);
2430 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
2431 Flags.setOrigAlign(OriginalAlignment);
2434 ArgRegs.push_back(Arg);
2435 ArgVTs.push_back(ArgVT);
2436 ArgFlags.push_back(Flags);
2439 // Handle the arguments now that we've gotten them.
2440 SmallVector<unsigned, 4> RegArgs;
2442 if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags,
2443 RegArgs, CC, NumBytes, isVarArg))
2446 bool UseReg = false;
2447 const GlobalValue *GV = dyn_cast<GlobalValue>(Callee);
2448 if (!GV || EnableARMLongCalls) UseReg = true;
2450 unsigned CalleeReg = 0;
2453 CalleeReg = getLibcallReg(IntrMemName);
2455 CalleeReg = getRegForValue(Callee);
2457 if (CalleeReg == 0) return false;
2461 unsigned CallOpc = ARMSelectCallOp(UseReg);
2462 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
2463 DL, TII.get(CallOpc));
2465 unsigned char OpFlags = 0;
2467 // Add MO_PLT for global address or external symbol in the PIC relocation
2469 if (Subtarget->isTargetELF() && TM.getRelocationModel() == Reloc::PIC_)
2470 OpFlags = ARMII::MO_PLT;
2472 // ARM calls don't take a predicate, but tBL / tBLX do.
2474 AddDefaultPred(MIB);
2476 MIB.addReg(CalleeReg);
2477 else if (!IntrMemName)
2478 MIB.addGlobalAddress(GV, 0, OpFlags);
2480 MIB.addExternalSymbol(IntrMemName, OpFlags);
2482 // Add implicit physical register uses to the call.
2483 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
2484 MIB.addReg(RegArgs[i], RegState::Implicit);
2486 // Add a register mask with the call-preserved registers.
2487 // Proper defs for return values will be added by setPhysRegsDeadExcept().
2488 MIB.addRegMask(TRI.getCallPreservedMask(CC));
2490 // Finish off the call including any return values.
2491 SmallVector<unsigned, 4> UsedRegs;
2492 if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes, isVarArg))
2495 // Set all unused physreg defs as dead.
2496 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
2501 bool ARMFastISel::ARMIsMemCpySmall(uint64_t Len) {
2505 bool ARMFastISel::ARMTryEmitSmallMemCpy(Address Dest, Address Src,
2506 uint64_t Len, unsigned Alignment) {
2507 // Make sure we don't bloat code by inlining very large memcpy's.
2508 if (!ARMIsMemCpySmall(Len))
2513 if (!Alignment || Alignment >= 4) {
2519 assert (Len == 1 && "Expected a length of 1!");
2523 // Bound based on alignment.
2524 if (Len >= 2 && Alignment == 2)
2533 RV = ARMEmitLoad(VT, ResultReg, Src);
2534 assert (RV == true && "Should be able to handle this load.");
2535 RV = ARMEmitStore(VT, ResultReg, Dest);
2536 assert (RV == true && "Should be able to handle this store.");
2539 unsigned Size = VT.getSizeInBits()/8;
2541 Dest.Offset += Size;
2548 bool ARMFastISel::SelectIntrinsicCall(const IntrinsicInst &I) {
2549 // FIXME: Handle more intrinsics.
2550 switch (I.getIntrinsicID()) {
2551 default: return false;
2552 case Intrinsic::frameaddress: {
2553 MachineFrameInfo *MFI = FuncInfo.MF->getFrameInfo();
2554 MFI->setFrameAddressIsTaken(true);
2557 const TargetRegisterClass *RC;
2559 LdrOpc = ARM::t2LDRi12;
2560 RC = (const TargetRegisterClass*)&ARM::tGPRRegClass;
2562 LdrOpc = ARM::LDRi12;
2563 RC = (const TargetRegisterClass*)&ARM::GPRRegClass;
2566 const ARMBaseRegisterInfo *RegInfo =
2567 static_cast<const ARMBaseRegisterInfo*>(TM.getRegisterInfo());
2568 unsigned FramePtr = RegInfo->getFrameRegister(*(FuncInfo.MF));
2569 unsigned SrcReg = FramePtr;
2571 // Recursively load frame address
2577 unsigned Depth = cast<ConstantInt>(I.getOperand(0))->getZExtValue();
2579 DestReg = createResultReg(RC);
2580 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
2581 TII.get(LdrOpc), DestReg)
2582 .addReg(SrcReg).addImm(0));
2585 UpdateValueMap(&I, SrcReg);
2588 case Intrinsic::memcpy:
2589 case Intrinsic::memmove: {
2590 const MemTransferInst &MTI = cast<MemTransferInst>(I);
2591 // Don't handle volatile.
2592 if (MTI.isVolatile())
2595 // Disable inlining for memmove before calls to ComputeAddress. Otherwise,
2596 // we would emit dead code because we don't currently handle memmoves.
2597 bool isMemCpy = (I.getIntrinsicID() == Intrinsic::memcpy);
2598 if (isa<ConstantInt>(MTI.getLength()) && isMemCpy) {
2599 // Small memcpy's are common enough that we want to do them without a call
2601 uint64_t Len = cast<ConstantInt>(MTI.getLength())->getZExtValue();
2602 if (ARMIsMemCpySmall(Len)) {
2604 if (!ARMComputeAddress(MTI.getRawDest(), Dest) ||
2605 !ARMComputeAddress(MTI.getRawSource(), Src))
2607 unsigned Alignment = MTI.getAlignment();
2608 if (ARMTryEmitSmallMemCpy(Dest, Src, Len, Alignment))
2613 if (!MTI.getLength()->getType()->isIntegerTy(32))
2616 if (MTI.getSourceAddressSpace() > 255 || MTI.getDestAddressSpace() > 255)
2619 const char *IntrMemName = isa<MemCpyInst>(I) ? "memcpy" : "memmove";
2620 return SelectCall(&I, IntrMemName);
2622 case Intrinsic::memset: {
2623 const MemSetInst &MSI = cast<MemSetInst>(I);
2624 // Don't handle volatile.
2625 if (MSI.isVolatile())
2628 if (!MSI.getLength()->getType()->isIntegerTy(32))
2631 if (MSI.getDestAddressSpace() > 255)
2634 return SelectCall(&I, "memset");
2636 case Intrinsic::trap: {
2637 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(
2638 Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP));
2644 bool ARMFastISel::SelectTrunc(const Instruction *I) {
2645 // The high bits for a type smaller than the register size are assumed to be
2647 Value *Op = I->getOperand(0);
2650 SrcVT = TLI.getValueType(Op->getType(), true);
2651 DestVT = TLI.getValueType(I->getType(), true);
2653 if (SrcVT != MVT::i32 && SrcVT != MVT::i16 && SrcVT != MVT::i8)
2655 if (DestVT != MVT::i16 && DestVT != MVT::i8 && DestVT != MVT::i1)
2658 unsigned SrcReg = getRegForValue(Op);
2659 if (!SrcReg) return false;
2661 // Because the high bits are undefined, a truncate doesn't generate
2663 UpdateValueMap(I, SrcReg);
2667 unsigned ARMFastISel::ARMEmitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT,
2669 if (DestVT != MVT::i32 && DestVT != MVT::i16 && DestVT != MVT::i8)
2671 if (SrcVT != MVT::i16 && SrcVT != MVT::i8 && SrcVT != MVT::i1)
2674 // Table of which combinations can be emitted as a single instruction,
2675 // and which will require two.
2676 static const uint8_t isSingleInstrTbl[3][2][2][2] = {
2678 // !hasV6Ops hasV6Ops !hasV6Ops hasV6Ops
2679 // ext: s z s z s z s z
2680 /* 1 */ { { { 0, 1 }, { 0, 1 } }, { { 0, 0 }, { 0, 1 } } },
2681 /* 8 */ { { { 0, 1 }, { 1, 1 } }, { { 0, 0 }, { 1, 1 } } },
2682 /* 16 */ { { { 0, 0 }, { 1, 1 } }, { { 0, 0 }, { 1, 1 } } }
2685 // Target registers for:
2686 // - For ARM can never be PC.
2687 // - For 16-bit Thumb are restricted to lower 8 registers.
2688 // - For 32-bit Thumb are restricted to non-SP and non-PC.
2689 static const TargetRegisterClass *RCTbl[2][2] = {
2690 // Instructions: Two Single
2691 /* ARM */ { &ARM::GPRnopcRegClass, &ARM::GPRnopcRegClass },
2692 /* Thumb */ { &ARM::tGPRRegClass, &ARM::rGPRRegClass }
2695 // Table governing the instruction(s) to be emitted.
2696 static const struct InstructionTable {
2698 uint32_t hasS : 1; // Some instructions have an S bit, always set it to 0.
2699 uint32_t Shift : 7; // For shift operand addressing mode, used by MOVsi.
2700 uint32_t Imm : 8; // All instructions have either a shift or a mask.
2701 } IT[2][2][3][2] = {
2702 { // Two instructions (first is left shift, second is in this table).
2703 { // ARM Opc S Shift Imm
2704 /* 1 bit sext */ { { ARM::MOVsi , 1, ARM_AM::asr , 31 },
2705 /* 1 bit zext */ { ARM::MOVsi , 1, ARM_AM::lsr , 31 } },
2706 /* 8 bit sext */ { { ARM::MOVsi , 1, ARM_AM::asr , 24 },
2707 /* 8 bit zext */ { ARM::MOVsi , 1, ARM_AM::lsr , 24 } },
2708 /* 16 bit sext */ { { ARM::MOVsi , 1, ARM_AM::asr , 16 },
2709 /* 16 bit zext */ { ARM::MOVsi , 1, ARM_AM::lsr , 16 } }
2711 { // Thumb Opc S Shift Imm
2712 /* 1 bit sext */ { { ARM::tASRri , 0, ARM_AM::no_shift, 31 },
2713 /* 1 bit zext */ { ARM::tLSRri , 0, ARM_AM::no_shift, 31 } },
2714 /* 8 bit sext */ { { ARM::tASRri , 0, ARM_AM::no_shift, 24 },
2715 /* 8 bit zext */ { ARM::tLSRri , 0, ARM_AM::no_shift, 24 } },
2716 /* 16 bit sext */ { { ARM::tASRri , 0, ARM_AM::no_shift, 16 },
2717 /* 16 bit zext */ { ARM::tLSRri , 0, ARM_AM::no_shift, 16 } }
2720 { // Single instruction.
2721 { // ARM Opc S Shift Imm
2722 /* 1 bit sext */ { { ARM::KILL , 0, ARM_AM::no_shift, 0 },
2723 /* 1 bit zext */ { ARM::ANDri , 1, ARM_AM::no_shift, 1 } },
2724 /* 8 bit sext */ { { ARM::SXTB , 0, ARM_AM::no_shift, 0 },
2725 /* 8 bit zext */ { ARM::ANDri , 1, ARM_AM::no_shift, 255 } },
2726 /* 16 bit sext */ { { ARM::SXTH , 0, ARM_AM::no_shift, 0 },
2727 /* 16 bit zext */ { ARM::UXTH , 0, ARM_AM::no_shift, 0 } }
2729 { // Thumb Opc S Shift Imm
2730 /* 1 bit sext */ { { ARM::KILL , 0, ARM_AM::no_shift, 0 },
2731 /* 1 bit zext */ { ARM::t2ANDri, 1, ARM_AM::no_shift, 1 } },
2732 /* 8 bit sext */ { { ARM::t2SXTB , 0, ARM_AM::no_shift, 0 },
2733 /* 8 bit zext */ { ARM::t2ANDri, 1, ARM_AM::no_shift, 255 } },
2734 /* 16 bit sext */ { { ARM::t2SXTH , 0, ARM_AM::no_shift, 0 },
2735 /* 16 bit zext */ { ARM::t2UXTH , 0, ARM_AM::no_shift, 0 } }
2740 unsigned SrcBits = SrcVT.getSizeInBits();
2741 unsigned DestBits = DestVT.getSizeInBits();
2743 assert((SrcBits < DestBits) && "can only extend to larger types");
2744 assert((DestBits == 32 || DestBits == 16 || DestBits == 8) &&
2745 "other sizes unimplemented");
2746 assert((SrcBits == 16 || SrcBits == 8 || SrcBits == 1) &&
2747 "other sizes unimplemented");
2749 bool hasV6Ops = Subtarget->hasV6Ops();
2750 unsigned Bitness = SrcBits / 8; // {1,8,16}=>{0,1,2}
2751 assert((Bitness < 3) && "sanity-check table bounds");
2753 bool isSingleInstr = isSingleInstrTbl[Bitness][isThumb2][hasV6Ops][isZExt];
2754 const TargetRegisterClass *RC = RCTbl[isThumb2][isSingleInstr];
2755 const InstructionTable *ITP = &IT[isSingleInstr][isThumb2][Bitness][isZExt];
2756 unsigned Opc = ITP->Opc;
2757 assert(ARM::KILL != Opc && "Invalid table entry");
2758 unsigned hasS = ITP->hasS;
2759 ARM_AM::ShiftOpc Shift = (ARM_AM::ShiftOpc) ITP->Shift;
2760 assert(((Shift == ARM_AM::no_shift) == (Opc != ARM::MOVsi)) &&
2761 "only MOVsi has shift operand addressing mode");
2762 unsigned Imm = ITP->Imm;
2764 // 16-bit Thumb instructions always set CPSR (unless they're in an IT block).
2765 bool setsCPSR = &ARM::tGPRRegClass == RC;
2766 unsigned LSLOpc = isThumb2 ? ARM::tLSLri : ARM::MOVsi;
2768 // MOVsi encodes shift and immediate in shift operand addressing mode.
2769 // The following condition has the same value when emitting two
2770 // instruction sequences: both are shifts.
2771 bool ImmIsSO = (Shift != ARM_AM::no_shift);
2773 // Either one or two instructions are emitted.
2774 // They're always of the form:
2776 // CPSR is set only by 16-bit Thumb instructions.
2777 // Predicate, if any, is AL.
2778 // S bit, if available, is always 0.
2779 // When two are emitted the first's result will feed as the second's input,
2780 // that value is then dead.
2781 unsigned NumInstrsEmitted = isSingleInstr ? 1 : 2;
2782 for (unsigned Instr = 0; Instr != NumInstrsEmitted; ++Instr) {
2783 ResultReg = createResultReg(RC);
2784 bool isLsl = (0 == Instr) && !isSingleInstr;
2785 unsigned Opcode = isLsl ? LSLOpc : Opc;
2786 ARM_AM::ShiftOpc ShiftAM = isLsl ? ARM_AM::lsl : Shift;
2787 unsigned ImmEnc = ImmIsSO ? ARM_AM::getSORegOpc(ShiftAM, Imm) : Imm;
2788 bool isKill = 1 == Instr;
2789 MachineInstrBuilder MIB = BuildMI(
2790 *FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opcode), ResultReg);
2792 MIB.addReg(ARM::CPSR, RegState::Define);
2793 SrcReg = constrainOperandRegClass(TII.get(Opcode), SrcReg, 1 + setsCPSR);
2794 AddDefaultPred(MIB.addReg(SrcReg, isKill * RegState::Kill).addImm(ImmEnc));
2797 // Second instruction consumes the first's result.
2804 bool ARMFastISel::SelectIntExt(const Instruction *I) {
2805 // On ARM, in general, integer casts don't involve legal types; this code
2806 // handles promotable integers.
2807 Type *DestTy = I->getType();
2808 Value *Src = I->getOperand(0);
2809 Type *SrcTy = Src->getType();
2811 bool isZExt = isa<ZExtInst>(I);
2812 unsigned SrcReg = getRegForValue(Src);
2813 if (!SrcReg) return false;
2815 EVT SrcEVT, DestEVT;
2816 SrcEVT = TLI.getValueType(SrcTy, true);
2817 DestEVT = TLI.getValueType(DestTy, true);
2818 if (!SrcEVT.isSimple()) return false;
2819 if (!DestEVT.isSimple()) return false;
2821 MVT SrcVT = SrcEVT.getSimpleVT();
2822 MVT DestVT = DestEVT.getSimpleVT();
2823 unsigned ResultReg = ARMEmitIntExt(SrcVT, SrcReg, DestVT, isZExt);
2824 if (ResultReg == 0) return false;
2825 UpdateValueMap(I, ResultReg);
2829 bool ARMFastISel::SelectShift(const Instruction *I,
2830 ARM_AM::ShiftOpc ShiftTy) {
2831 // We handle thumb2 mode by target independent selector
2832 // or SelectionDAG ISel.
2836 // Only handle i32 now.
2837 EVT DestVT = TLI.getValueType(I->getType(), true);
2838 if (DestVT != MVT::i32)
2841 unsigned Opc = ARM::MOVsr;
2843 Value *Src2Value = I->getOperand(1);
2844 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Src2Value)) {
2845 ShiftImm = CI->getZExtValue();
2847 // Fall back to selection DAG isel if the shift amount
2848 // is zero or greater than the width of the value type.
2849 if (ShiftImm == 0 || ShiftImm >=32)
2855 Value *Src1Value = I->getOperand(0);
2856 unsigned Reg1 = getRegForValue(Src1Value);
2857 if (Reg1 == 0) return false;
2860 if (Opc == ARM::MOVsr) {
2861 Reg2 = getRegForValue(Src2Value);
2862 if (Reg2 == 0) return false;
2865 unsigned ResultReg = createResultReg(&ARM::GPRnopcRegClass);
2866 if(ResultReg == 0) return false;
2868 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
2869 TII.get(Opc), ResultReg)
2872 if (Opc == ARM::MOVsi)
2873 MIB.addImm(ARM_AM::getSORegOpc(ShiftTy, ShiftImm));
2874 else if (Opc == ARM::MOVsr) {
2876 MIB.addImm(ARM_AM::getSORegOpc(ShiftTy, 0));
2879 AddOptionalDefs(MIB);
2880 UpdateValueMap(I, ResultReg);
2884 // TODO: SoftFP support.
2885 bool ARMFastISel::TargetSelectInstruction(const Instruction *I) {
2887 switch (I->getOpcode()) {
2888 case Instruction::Load:
2889 return SelectLoad(I);
2890 case Instruction::Store:
2891 return SelectStore(I);
2892 case Instruction::Br:
2893 return SelectBranch(I);
2894 case Instruction::IndirectBr:
2895 return SelectIndirectBr(I);
2896 case Instruction::ICmp:
2897 case Instruction::FCmp:
2898 return SelectCmp(I);
2899 case Instruction::FPExt:
2900 return SelectFPExt(I);
2901 case Instruction::FPTrunc:
2902 return SelectFPTrunc(I);
2903 case Instruction::SIToFP:
2904 return SelectIToFP(I, /*isSigned*/ true);
2905 case Instruction::UIToFP:
2906 return SelectIToFP(I, /*isSigned*/ false);
2907 case Instruction::FPToSI:
2908 return SelectFPToI(I, /*isSigned*/ true);
2909 case Instruction::FPToUI:
2910 return SelectFPToI(I, /*isSigned*/ false);
2911 case Instruction::Add:
2912 return SelectBinaryIntOp(I, ISD::ADD);
2913 case Instruction::Or:
2914 return SelectBinaryIntOp(I, ISD::OR);
2915 case Instruction::Sub:
2916 return SelectBinaryIntOp(I, ISD::SUB);
2917 case Instruction::FAdd:
2918 return SelectBinaryFPOp(I, ISD::FADD);
2919 case Instruction::FSub:
2920 return SelectBinaryFPOp(I, ISD::FSUB);
2921 case Instruction::FMul:
2922 return SelectBinaryFPOp(I, ISD::FMUL);
2923 case Instruction::SDiv:
2924 return SelectDiv(I, /*isSigned*/ true);
2925 case Instruction::UDiv:
2926 return SelectDiv(I, /*isSigned*/ false);
2927 case Instruction::SRem:
2928 return SelectRem(I, /*isSigned*/ true);
2929 case Instruction::URem:
2930 return SelectRem(I, /*isSigned*/ false);
2931 case Instruction::Call:
2932 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2933 return SelectIntrinsicCall(*II);
2934 return SelectCall(I);
2935 case Instruction::Select:
2936 return SelectSelect(I);
2937 case Instruction::Ret:
2938 return SelectRet(I);
2939 case Instruction::Trunc:
2940 return SelectTrunc(I);
2941 case Instruction::ZExt:
2942 case Instruction::SExt:
2943 return SelectIntExt(I);
2944 case Instruction::Shl:
2945 return SelectShift(I, ARM_AM::lsl);
2946 case Instruction::LShr:
2947 return SelectShift(I, ARM_AM::lsr);
2948 case Instruction::AShr:
2949 return SelectShift(I, ARM_AM::asr);
2956 // This table describes sign- and zero-extend instructions which can be
2957 // folded into a preceding load. All of these extends have an immediate
2958 // (sometimes a mask and sometimes a shift) that's applied after
2960 const struct FoldableLoadExtendsStruct {
2961 uint16_t Opc[2]; // ARM, Thumb.
2962 uint8_t ExpectedImm;
2964 uint8_t ExpectedVT : 7;
2965 } FoldableLoadExtends[] = {
2966 { { ARM::SXTH, ARM::t2SXTH }, 0, 0, MVT::i16 },
2967 { { ARM::UXTH, ARM::t2UXTH }, 0, 1, MVT::i16 },
2968 { { ARM::ANDri, ARM::t2ANDri }, 255, 1, MVT::i8 },
2969 { { ARM::SXTB, ARM::t2SXTB }, 0, 0, MVT::i8 },
2970 { { ARM::UXTB, ARM::t2UXTB }, 0, 1, MVT::i8 }
2974 /// \brief The specified machine instr operand is a vreg, and that
2975 /// vreg is being provided by the specified load instruction. If possible,
2976 /// try to fold the load as an operand to the instruction, returning true if
2978 bool ARMFastISel::tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo,
2979 const LoadInst *LI) {
2980 // Verify we have a legal type before going any further.
2982 if (!isLoadTypeLegal(LI->getType(), VT))
2985 // Combine load followed by zero- or sign-extend.
2986 // ldrb r1, [r0] ldrb r1, [r0]
2988 // mov r3, r2 mov r3, r1
2989 if (MI->getNumOperands() < 3 || !MI->getOperand(2).isImm())
2991 const uint64_t Imm = MI->getOperand(2).getImm();
2995 for (unsigned i = 0, e = array_lengthof(FoldableLoadExtends);
2997 if (FoldableLoadExtends[i].Opc[isThumb2] == MI->getOpcode() &&
2998 (uint64_t)FoldableLoadExtends[i].ExpectedImm == Imm &&
2999 MVT((MVT::SimpleValueType)FoldableLoadExtends[i].ExpectedVT) == VT) {
3001 isZExt = FoldableLoadExtends[i].isZExt;
3004 if (!Found) return false;
3006 // See if we can handle this address.
3008 if (!ARMComputeAddress(LI->getOperand(0), Addr)) return false;
3010 unsigned ResultReg = MI->getOperand(0).getReg();
3011 if (!ARMEmitLoad(VT, ResultReg, Addr, LI->getAlignment(), isZExt, false))
3013 MI->eraseFromParent();
3017 unsigned ARMFastISel::ARMLowerPICELF(const GlobalValue *GV,
3018 unsigned Align, MVT VT) {
3019 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
3020 ARMConstantPoolConstant *CPV =
3021 ARMConstantPoolConstant::Create(GV, UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
3022 unsigned Idx = MCP.getConstantPoolIndex(CPV, Align);
3025 unsigned DestReg1 = createResultReg(TLI.getRegClassFor(VT));
3028 DestReg1 = constrainOperandRegClass(TII.get(ARM::t2LDRpci), DestReg1, 0);
3029 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
3030 TII.get(ARM::t2LDRpci), DestReg1)
3031 .addConstantPoolIndex(Idx));
3032 Opc = UseGOTOFF ? ARM::t2ADDrr : ARM::t2LDRs;
3034 // The extra immediate is for addrmode2.
3035 DestReg1 = constrainOperandRegClass(TII.get(ARM::LDRcp), DestReg1, 0);
3036 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
3037 DL, TII.get(ARM::LDRcp), DestReg1)
3038 .addConstantPoolIndex(Idx).addImm(0));
3039 Opc = UseGOTOFF ? ARM::ADDrr : ARM::LDRrs;
3042 unsigned GlobalBaseReg = AFI->getGlobalBaseReg();
3043 if (GlobalBaseReg == 0) {
3044 GlobalBaseReg = MRI.createVirtualRegister(TLI.getRegClassFor(VT));
3045 AFI->setGlobalBaseReg(GlobalBaseReg);
3048 unsigned DestReg2 = createResultReg(TLI.getRegClassFor(VT));
3049 DestReg2 = constrainOperandRegClass(TII.get(Opc), DestReg2, 0);
3050 DestReg1 = constrainOperandRegClass(TII.get(Opc), DestReg1, 1);
3051 GlobalBaseReg = constrainOperandRegClass(TII.get(Opc), GlobalBaseReg, 2);
3052 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
3053 DL, TII.get(Opc), DestReg2)
3055 .addReg(GlobalBaseReg);
3058 AddOptionalDefs(MIB);
3063 bool ARMFastISel::FastLowerArguments() {
3064 if (!FuncInfo.CanLowerReturn)
3067 const Function *F = FuncInfo.Fn;
3071 CallingConv::ID CC = F->getCallingConv();
3075 case CallingConv::Fast:
3076 case CallingConv::C:
3077 case CallingConv::ARM_AAPCS_VFP:
3078 case CallingConv::ARM_AAPCS:
3079 case CallingConv::ARM_APCS:
3083 // Only handle simple cases. i.e. Up to 4 i8/i16/i32 scalar arguments
3084 // which are passed in r0 - r3.
3086 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
3087 I != E; ++I, ++Idx) {
3091 if (F->getAttributes().hasAttribute(Idx, Attribute::InReg) ||
3092 F->getAttributes().hasAttribute(Idx, Attribute::StructRet) ||
3093 F->getAttributes().hasAttribute(Idx, Attribute::ByVal))
3096 Type *ArgTy = I->getType();
3097 if (ArgTy->isStructTy() || ArgTy->isArrayTy() || ArgTy->isVectorTy())
3100 EVT ArgVT = TLI.getValueType(ArgTy);
3101 if (!ArgVT.isSimple()) return false;
3102 switch (ArgVT.getSimpleVT().SimpleTy) {
3113 static const uint16_t GPRArgRegs[] = {
3114 ARM::R0, ARM::R1, ARM::R2, ARM::R3
3117 const TargetRegisterClass *RC = &ARM::rGPRRegClass;
3119 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
3120 I != E; ++I, ++Idx) {
3121 unsigned SrcReg = GPRArgRegs[Idx];
3122 unsigned DstReg = FuncInfo.MF->addLiveIn(SrcReg, RC);
3123 // FIXME: Unfortunately it's necessary to emit a copy from the livein copy.
3124 // Without this, EmitLiveInCopies may eliminate the livein if its only
3125 // use is a bitcast (which isn't turned into an instruction).
3126 unsigned ResultReg = createResultReg(RC);
3127 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
3128 ResultReg).addReg(DstReg, getKillRegState(true));
3129 UpdateValueMap(I, ResultReg);
3136 FastISel *ARM::createFastISel(FunctionLoweringInfo &funcInfo,
3137 const TargetLibraryInfo *libInfo) {
3138 const TargetMachine &TM = funcInfo.MF->getTarget();
3140 const ARMSubtarget *Subtarget = &TM.getSubtarget<ARMSubtarget>();
3141 // Thumb2 support on iOS; ARM support on iOS, Linux and NaCl.
3142 bool UseFastISel = false;
3143 UseFastISel |= Subtarget->isTargetIOS() && !Subtarget->isThumb1Only();
3144 UseFastISel |= Subtarget->isTargetLinux() && !Subtarget->isThumb();
3145 UseFastISel |= Subtarget->isTargetNaCl() && !Subtarget->isThumb();
3148 // iOS always has a FP for backtracking, force other targets
3149 // to keep their FP when doing FastISel. The emitted code is
3150 // currently superior, and in cases like test-suite's lencod
3151 // FastISel isn't quite correct when FP is eliminated.
3152 TM.Options.NoFramePointerElim = true;
3153 return new ARMFastISel(funcInfo, libInfo);