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 "ARMRegisterInfo.h"
20 #include "ARMTargetMachine.h"
21 #include "ARMSubtarget.h"
22 #include "ARMConstantPoolValue.h"
23 #include "llvm/CallingConv.h"
24 #include "llvm/DerivedTypes.h"
25 #include "llvm/GlobalVariable.h"
26 #include "llvm/Instructions.h"
27 #include "llvm/IntrinsicInst.h"
28 #include "llvm/Module.h"
29 #include "llvm/Operator.h"
30 #include "llvm/CodeGen/Analysis.h"
31 #include "llvm/CodeGen/FastISel.h"
32 #include "llvm/CodeGen/FunctionLoweringInfo.h"
33 #include "llvm/CodeGen/MachineInstrBuilder.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineConstantPool.h"
36 #include "llvm/CodeGen/MachineFrameInfo.h"
37 #include "llvm/CodeGen/MachineMemOperand.h"
38 #include "llvm/CodeGen/MachineRegisterInfo.h"
39 #include "llvm/CodeGen/PseudoSourceValue.h"
40 #include "llvm/Support/CallSite.h"
41 #include "llvm/Support/CommandLine.h"
42 #include "llvm/Support/ErrorHandling.h"
43 #include "llvm/Support/GetElementPtrTypeIterator.h"
44 #include "llvm/Target/TargetData.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"
52 DisableARMFastISel("disable-arm-fast-isel",
53 cl::desc("Turn off experimental ARM fast-isel support"),
54 cl::init(false), cl::Hidden);
56 extern cl::opt<bool> EnableARMLongCalls;
60 // All possible address modes, plus some.
61 typedef struct Address {
76 // Innocuous defaults for our address.
78 : BaseType(RegBase), Offset(0), Scale(0), PlusReg(0) {
83 class ARMFastISel : public FastISel {
85 /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
86 /// make the right decision when generating code for different targets.
87 const ARMSubtarget *Subtarget;
88 const TargetMachine &TM;
89 const TargetInstrInfo &TII;
90 const TargetLowering &TLI;
93 // Convenience variables to avoid some queries.
98 explicit ARMFastISel(FunctionLoweringInfo &funcInfo)
100 TM(funcInfo.MF->getTarget()),
101 TII(*TM.getInstrInfo()),
102 TLI(*TM.getTargetLowering()) {
103 Subtarget = &TM.getSubtarget<ARMSubtarget>();
104 AFI = funcInfo.MF->getInfo<ARMFunctionInfo>();
105 isThumb = AFI->isThumbFunction();
106 Context = &funcInfo.Fn->getContext();
109 // Code from FastISel.cpp.
110 virtual unsigned FastEmitInst_(unsigned MachineInstOpcode,
111 const TargetRegisterClass *RC);
112 virtual unsigned FastEmitInst_r(unsigned MachineInstOpcode,
113 const TargetRegisterClass *RC,
114 unsigned Op0, bool Op0IsKill);
115 virtual unsigned FastEmitInst_rr(unsigned MachineInstOpcode,
116 const TargetRegisterClass *RC,
117 unsigned Op0, bool Op0IsKill,
118 unsigned Op1, bool Op1IsKill);
119 virtual unsigned FastEmitInst_rrr(unsigned MachineInstOpcode,
120 const TargetRegisterClass *RC,
121 unsigned Op0, bool Op0IsKill,
122 unsigned Op1, bool Op1IsKill,
123 unsigned Op2, bool Op2IsKill);
124 virtual unsigned FastEmitInst_ri(unsigned MachineInstOpcode,
125 const TargetRegisterClass *RC,
126 unsigned Op0, bool Op0IsKill,
128 virtual unsigned FastEmitInst_rf(unsigned MachineInstOpcode,
129 const TargetRegisterClass *RC,
130 unsigned Op0, bool Op0IsKill,
131 const ConstantFP *FPImm);
132 virtual unsigned FastEmitInst_rri(unsigned MachineInstOpcode,
133 const TargetRegisterClass *RC,
134 unsigned Op0, bool Op0IsKill,
135 unsigned Op1, bool Op1IsKill,
137 virtual unsigned FastEmitInst_i(unsigned MachineInstOpcode,
138 const TargetRegisterClass *RC,
141 virtual unsigned FastEmitInst_extractsubreg(MVT RetVT,
142 unsigned Op0, bool Op0IsKill,
145 // Backend specific FastISel code.
146 virtual bool TargetSelectInstruction(const Instruction *I);
147 virtual unsigned TargetMaterializeConstant(const Constant *C);
148 virtual unsigned TargetMaterializeAlloca(const AllocaInst *AI);
150 #include "ARMGenFastISel.inc"
152 // Instruction selection routines.
154 bool SelectLoad(const Instruction *I);
155 bool SelectStore(const Instruction *I);
156 bool SelectBranch(const Instruction *I);
157 bool SelectCmp(const Instruction *I);
158 bool SelectFPExt(const Instruction *I);
159 bool SelectFPTrunc(const Instruction *I);
160 bool SelectBinaryOp(const Instruction *I, unsigned ISDOpcode);
161 bool SelectSIToFP(const Instruction *I);
162 bool SelectFPToSI(const Instruction *I);
163 bool SelectSDiv(const Instruction *I);
164 bool SelectSRem(const Instruction *I);
165 bool SelectCall(const Instruction *I);
166 bool SelectSelect(const Instruction *I);
167 bool SelectRet(const Instruction *I);
171 bool isTypeLegal(const Type *Ty, MVT &VT);
172 bool isLoadTypeLegal(const Type *Ty, MVT &VT);
173 bool ARMEmitLoad(EVT VT, unsigned &ResultReg, Address &Addr);
174 bool ARMEmitStore(EVT VT, unsigned SrcReg, Address &Addr);
175 bool ARMComputeAddress(const Value *Obj, Address &Addr);
176 void ARMSimplifyAddress(Address &Addr, EVT VT);
177 unsigned ARMMaterializeFP(const ConstantFP *CFP, EVT VT);
178 unsigned ARMMaterializeInt(const Constant *C, EVT VT);
179 unsigned ARMMaterializeGV(const GlobalValue *GV, EVT VT);
180 unsigned ARMMoveToFPReg(EVT VT, unsigned SrcReg);
181 unsigned ARMMoveToIntReg(EVT VT, unsigned SrcReg);
182 unsigned ARMSelectCallOp(const GlobalValue *GV);
184 // Call handling routines.
186 bool FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src, EVT SrcVT,
187 unsigned &ResultReg);
188 CCAssignFn *CCAssignFnForCall(CallingConv::ID CC, bool Return);
189 bool ProcessCallArgs(SmallVectorImpl<Value*> &Args,
190 SmallVectorImpl<unsigned> &ArgRegs,
191 SmallVectorImpl<MVT> &ArgVTs,
192 SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
193 SmallVectorImpl<unsigned> &RegArgs,
196 bool FinishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
197 const Instruction *I, CallingConv::ID CC,
199 bool ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call);
201 // OptionalDef handling routines.
203 bool isARMNEONPred(const MachineInstr *MI);
204 bool DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR);
205 const MachineInstrBuilder &AddOptionalDefs(const MachineInstrBuilder &MIB);
206 void AddLoadStoreOperands(EVT VT, Address &Addr,
207 const MachineInstrBuilder &MIB);
210 } // end anonymous namespace
212 #include "ARMGenCallingConv.inc"
214 // DefinesOptionalPredicate - This is different from DefinesPredicate in that
215 // we don't care about implicit defs here, just places we'll need to add a
216 // default CCReg argument. Sets CPSR if we're setting CPSR instead of CCR.
217 bool ARMFastISel::DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR) {
218 const TargetInstrDesc &TID = MI->getDesc();
219 if (!TID.hasOptionalDef())
222 // Look to see if our OptionalDef is defining CPSR or CCR.
223 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
224 const MachineOperand &MO = MI->getOperand(i);
225 if (!MO.isReg() || !MO.isDef()) continue;
226 if (MO.getReg() == ARM::CPSR)
232 bool ARMFastISel::isARMNEONPred(const MachineInstr *MI) {
233 const TargetInstrDesc &TID = MI->getDesc();
235 // If we're a thumb2 or not NEON function we were handled via isPredicable.
236 if ((TID.TSFlags & ARMII::DomainMask) != ARMII::DomainNEON ||
237 AFI->isThumb2Function())
240 for (unsigned i = 0, e = TID.getNumOperands(); i != e; ++i)
241 if (TID.OpInfo[i].isPredicate())
247 // If the machine is predicable go ahead and add the predicate operands, if
248 // it needs default CC operands add those.
249 // TODO: If we want to support thumb1 then we'll need to deal with optional
250 // CPSR defs that need to be added before the remaining operands. See s_cc_out
251 // for descriptions why.
252 const MachineInstrBuilder &
253 ARMFastISel::AddOptionalDefs(const MachineInstrBuilder &MIB) {
254 MachineInstr *MI = &*MIB;
256 // Do we use a predicate? or...
257 // Are we NEON in ARM mode and have a predicate operand? If so, I know
258 // we're not predicable but add it anyways.
259 if (TII.isPredicable(MI) || isARMNEONPred(MI))
262 // Do we optionally set a predicate? Preds is size > 0 iff the predicate
263 // defines CPSR. All other OptionalDefines in ARM are the CCR register.
265 if (DefinesOptionalPredicate(MI, &CPSR)) {
274 unsigned ARMFastISel::FastEmitInst_(unsigned MachineInstOpcode,
275 const TargetRegisterClass* RC) {
276 unsigned ResultReg = createResultReg(RC);
277 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
279 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg));
283 unsigned ARMFastISel::FastEmitInst_r(unsigned MachineInstOpcode,
284 const TargetRegisterClass *RC,
285 unsigned Op0, bool Op0IsKill) {
286 unsigned ResultReg = createResultReg(RC);
287 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
289 if (II.getNumDefs() >= 1)
290 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
291 .addReg(Op0, Op0IsKill * RegState::Kill));
293 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
294 .addReg(Op0, Op0IsKill * RegState::Kill));
295 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
296 TII.get(TargetOpcode::COPY), ResultReg)
297 .addReg(II.ImplicitDefs[0]));
302 unsigned ARMFastISel::FastEmitInst_rr(unsigned MachineInstOpcode,
303 const TargetRegisterClass *RC,
304 unsigned Op0, bool Op0IsKill,
305 unsigned Op1, bool Op1IsKill) {
306 unsigned ResultReg = createResultReg(RC);
307 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
309 if (II.getNumDefs() >= 1)
310 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
311 .addReg(Op0, Op0IsKill * RegState::Kill)
312 .addReg(Op1, Op1IsKill * RegState::Kill));
314 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
315 .addReg(Op0, Op0IsKill * RegState::Kill)
316 .addReg(Op1, Op1IsKill * RegState::Kill));
317 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
318 TII.get(TargetOpcode::COPY), ResultReg)
319 .addReg(II.ImplicitDefs[0]));
324 unsigned ARMFastISel::FastEmitInst_rrr(unsigned MachineInstOpcode,
325 const TargetRegisterClass *RC,
326 unsigned Op0, bool Op0IsKill,
327 unsigned Op1, bool Op1IsKill,
328 unsigned Op2, bool Op2IsKill) {
329 unsigned ResultReg = createResultReg(RC);
330 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
332 if (II.getNumDefs() >= 1)
333 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
334 .addReg(Op0, Op0IsKill * RegState::Kill)
335 .addReg(Op1, Op1IsKill * RegState::Kill)
336 .addReg(Op2, Op2IsKill * RegState::Kill));
338 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
339 .addReg(Op0, Op0IsKill * RegState::Kill)
340 .addReg(Op1, Op1IsKill * RegState::Kill)
341 .addReg(Op2, Op2IsKill * RegState::Kill));
342 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
343 TII.get(TargetOpcode::COPY), ResultReg)
344 .addReg(II.ImplicitDefs[0]));
349 unsigned ARMFastISel::FastEmitInst_ri(unsigned MachineInstOpcode,
350 const TargetRegisterClass *RC,
351 unsigned Op0, bool Op0IsKill,
353 unsigned ResultReg = createResultReg(RC);
354 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
356 if (II.getNumDefs() >= 1)
357 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
358 .addReg(Op0, Op0IsKill * RegState::Kill)
361 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
362 .addReg(Op0, Op0IsKill * 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_rf(unsigned MachineInstOpcode,
372 const TargetRegisterClass *RC,
373 unsigned Op0, bool Op0IsKill,
374 const ConstantFP *FPImm) {
375 unsigned ResultReg = createResultReg(RC);
376 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
378 if (II.getNumDefs() >= 1)
379 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
380 .addReg(Op0, Op0IsKill * RegState::Kill)
383 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
384 .addReg(Op0, Op0IsKill * RegState::Kill)
386 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
387 TII.get(TargetOpcode::COPY), ResultReg)
388 .addReg(II.ImplicitDefs[0]));
393 unsigned ARMFastISel::FastEmitInst_rri(unsigned MachineInstOpcode,
394 const TargetRegisterClass *RC,
395 unsigned Op0, bool Op0IsKill,
396 unsigned Op1, bool Op1IsKill,
398 unsigned ResultReg = createResultReg(RC);
399 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
401 if (II.getNumDefs() >= 1)
402 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
403 .addReg(Op0, Op0IsKill * RegState::Kill)
404 .addReg(Op1, Op1IsKill * RegState::Kill)
407 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
408 .addReg(Op0, Op0IsKill * RegState::Kill)
409 .addReg(Op1, Op1IsKill * RegState::Kill)
411 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
412 TII.get(TargetOpcode::COPY), ResultReg)
413 .addReg(II.ImplicitDefs[0]));
418 unsigned ARMFastISel::FastEmitInst_i(unsigned MachineInstOpcode,
419 const TargetRegisterClass *RC,
421 unsigned ResultReg = createResultReg(RC);
422 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
424 if (II.getNumDefs() >= 1)
425 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
428 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
430 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
431 TII.get(TargetOpcode::COPY), ResultReg)
432 .addReg(II.ImplicitDefs[0]));
437 unsigned ARMFastISel::FastEmitInst_extractsubreg(MVT RetVT,
438 unsigned Op0, bool Op0IsKill,
440 unsigned ResultReg = createResultReg(TLI.getRegClassFor(RetVT));
441 assert(TargetRegisterInfo::isVirtualRegister(Op0) &&
442 "Cannot yet extract from physregs");
443 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
444 DL, TII.get(TargetOpcode::COPY), ResultReg)
445 .addReg(Op0, getKillRegState(Op0IsKill), Idx));
449 // TODO: Don't worry about 64-bit now, but when this is fixed remove the
450 // checks from the various callers.
451 unsigned ARMFastISel::ARMMoveToFPReg(EVT VT, unsigned SrcReg) {
452 if (VT == MVT::f64) return 0;
454 unsigned MoveReg = createResultReg(TLI.getRegClassFor(VT));
455 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
456 TII.get(ARM::VMOVRS), MoveReg)
461 unsigned ARMFastISel::ARMMoveToIntReg(EVT VT, unsigned SrcReg) {
462 if (VT == MVT::i64) return 0;
464 unsigned MoveReg = createResultReg(TLI.getRegClassFor(VT));
465 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
466 TII.get(ARM::VMOVSR), MoveReg)
471 // For double width floating point we need to materialize two constants
472 // (the high and the low) into integer registers then use a move to get
473 // the combined constant into an FP reg.
474 unsigned ARMFastISel::ARMMaterializeFP(const ConstantFP *CFP, EVT VT) {
475 const APFloat Val = CFP->getValueAPF();
476 bool is64bit = VT == MVT::f64;
478 // This checks to see if we can use VFP3 instructions to materialize
479 // a constant, otherwise we have to go through the constant pool.
480 if (TLI.isFPImmLegal(Val, VT)) {
481 unsigned Opc = is64bit ? ARM::FCONSTD : ARM::FCONSTS;
482 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
483 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
489 // Require VFP2 for loading fp constants.
490 if (!Subtarget->hasVFP2()) return false;
492 // MachineConstantPool wants an explicit alignment.
493 unsigned Align = TD.getPrefTypeAlignment(CFP->getType());
495 // TODO: Figure out if this is correct.
496 Align = TD.getTypeAllocSize(CFP->getType());
498 unsigned Idx = MCP.getConstantPoolIndex(cast<Constant>(CFP), Align);
499 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
500 unsigned Opc = is64bit ? ARM::VLDRD : ARM::VLDRS;
502 // The extra reg is for addrmode5.
503 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
505 .addConstantPoolIndex(Idx)
510 unsigned ARMFastISel::ARMMaterializeInt(const Constant *C, EVT VT) {
512 // For now 32-bit only.
513 if (VT != MVT::i32) return false;
515 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
517 // If we can do this in a single instruction without a constant pool entry
519 const ConstantInt *CI = cast<ConstantInt>(C);
520 if (Subtarget->hasV6T2Ops() && isUInt<16>(CI->getSExtValue())) {
521 unsigned Opc = isThumb ? ARM::t2MOVi16 : ARM::MOVi16;
522 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
523 TII.get(Opc), DestReg)
524 .addImm(CI->getSExtValue()));
528 // MachineConstantPool wants an explicit alignment.
529 unsigned Align = TD.getPrefTypeAlignment(C->getType());
531 // TODO: Figure out if this is correct.
532 Align = TD.getTypeAllocSize(C->getType());
534 unsigned Idx = MCP.getConstantPoolIndex(C, Align);
537 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
538 TII.get(ARM::t2LDRpci), DestReg)
539 .addConstantPoolIndex(Idx));
541 // The extra immediate is for addrmode2.
542 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
543 TII.get(ARM::LDRcp), DestReg)
544 .addConstantPoolIndex(Idx)
550 unsigned ARMFastISel::ARMMaterializeGV(const GlobalValue *GV, EVT VT) {
551 // For now 32-bit only.
552 if (VT != MVT::i32) return 0;
554 Reloc::Model RelocM = TM.getRelocationModel();
556 // TODO: No external globals for now.
557 if (Subtarget->GVIsIndirectSymbol(GV, RelocM)) return 0;
559 // TODO: Need more magic for ARM PIC.
560 if (!isThumb && (RelocM == Reloc::PIC_)) return 0;
562 // MachineConstantPool wants an explicit alignment.
563 unsigned Align = TD.getPrefTypeAlignment(GV->getType());
565 // TODO: Figure out if this is correct.
566 Align = TD.getTypeAllocSize(GV->getType());
570 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb() ? 4 : 8);
571 unsigned Id = AFI->createPICLabelUId();
572 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, Id,
573 ARMCP::CPValue, PCAdj);
574 unsigned Idx = MCP.getConstantPoolIndex(CPV, Align);
577 MachineInstrBuilder MIB;
578 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
580 unsigned Opc = (RelocM != Reloc::PIC_) ? ARM::t2LDRpci : ARM::t2LDRpci_pic;
581 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), DestReg)
582 .addConstantPoolIndex(Idx);
583 if (RelocM == Reloc::PIC_)
586 // The extra immediate is for addrmode2.
587 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(ARM::LDRcp),
589 .addConstantPoolIndex(Idx)
592 AddOptionalDefs(MIB);
596 unsigned ARMFastISel::TargetMaterializeConstant(const Constant *C) {
597 EVT VT = TLI.getValueType(C->getType(), true);
599 // Only handle simple types.
600 if (!VT.isSimple()) return 0;
602 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
603 return ARMMaterializeFP(CFP, VT);
604 else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
605 return ARMMaterializeGV(GV, VT);
606 else if (isa<ConstantInt>(C))
607 return ARMMaterializeInt(C, VT);
612 unsigned ARMFastISel::TargetMaterializeAlloca(const AllocaInst *AI) {
613 // Don't handle dynamic allocas.
614 if (!FuncInfo.StaticAllocaMap.count(AI)) return 0;
617 if (!isLoadTypeLegal(AI->getType(), VT)) return false;
619 DenseMap<const AllocaInst*, int>::iterator SI =
620 FuncInfo.StaticAllocaMap.find(AI);
622 // This will get lowered later into the correct offsets and registers
623 // via rewriteXFrameIndex.
624 if (SI != FuncInfo.StaticAllocaMap.end()) {
625 TargetRegisterClass* RC = TLI.getRegClassFor(VT);
626 unsigned ResultReg = createResultReg(RC);
627 unsigned Opc = isThumb ? ARM::t2ADDri : ARM::ADDri;
628 AddOptionalDefs(BuildMI(*FuncInfo.MBB, *FuncInfo.InsertPt, DL,
629 TII.get(Opc), ResultReg)
630 .addFrameIndex(SI->second)
638 bool ARMFastISel::isTypeLegal(const Type *Ty, MVT &VT) {
639 EVT evt = TLI.getValueType(Ty, true);
641 // Only handle simple types.
642 if (evt == MVT::Other || !evt.isSimple()) return false;
643 VT = evt.getSimpleVT();
645 // Handle all legal types, i.e. a register that will directly hold this
647 return TLI.isTypeLegal(VT);
650 bool ARMFastISel::isLoadTypeLegal(const Type *Ty, MVT &VT) {
651 if (isTypeLegal(Ty, VT)) return true;
653 // If this is a type than can be sign or zero-extended to a basic operation
654 // go ahead and accept it now.
655 if (VT == MVT::i8 || VT == MVT::i16)
661 // Computes the address to get to an object.
662 bool ARMFastISel::ARMComputeAddress(const Value *Obj, Address &Addr) {
663 // Some boilerplate from the X86 FastISel.
664 const User *U = NULL;
665 unsigned Opcode = Instruction::UserOp1;
666 if (const Instruction *I = dyn_cast<Instruction>(Obj)) {
667 // Don't walk into other basic blocks unless the object is an alloca from
668 // another block, otherwise it may not have a virtual register assigned.
669 if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(Obj)) ||
670 FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
671 Opcode = I->getOpcode();
674 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Obj)) {
675 Opcode = C->getOpcode();
679 if (const PointerType *Ty = dyn_cast<PointerType>(Obj->getType()))
680 if (Ty->getAddressSpace() > 255)
681 // Fast instruction selection doesn't support the special
688 case Instruction::BitCast: {
689 // Look through bitcasts.
690 return ARMComputeAddress(U->getOperand(0), Addr);
692 case Instruction::IntToPtr: {
693 // Look past no-op inttoptrs.
694 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
695 return ARMComputeAddress(U->getOperand(0), Addr);
698 case Instruction::PtrToInt: {
699 // Look past no-op ptrtoints.
700 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
701 return ARMComputeAddress(U->getOperand(0), Addr);
704 case Instruction::GetElementPtr: {
705 Address SavedAddr = Addr;
706 int TmpOffset = Addr.Offset;
708 // Iterate through the GEP folding the constants into offsets where
710 gep_type_iterator GTI = gep_type_begin(U);
711 for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end();
712 i != e; ++i, ++GTI) {
713 const Value *Op = *i;
714 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
715 const StructLayout *SL = TD.getStructLayout(STy);
716 unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
717 TmpOffset += SL->getElementOffset(Idx);
719 uint64_t S = TD.getTypeAllocSize(GTI.getIndexedType());
721 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
722 // Constant-offset addressing.
723 TmpOffset += CI->getSExtValue() * S;
726 if (isa<AddOperator>(Op) &&
727 (!isa<Instruction>(Op) ||
728 FuncInfo.MBBMap[cast<Instruction>(Op)->getParent()]
730 isa<ConstantInt>(cast<AddOperator>(Op)->getOperand(1))) {
731 // An add (in the same block) with a constant operand. Fold the
734 cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
735 TmpOffset += CI->getSExtValue() * S;
736 // Iterate on the other operand.
737 Op = cast<AddOperator>(Op)->getOperand(0);
741 goto unsupported_gep;
746 // Try to grab the base operand now.
747 Addr.Offset = TmpOffset;
748 if (ARMComputeAddress(U->getOperand(0), Addr)) return true;
750 // We failed, restore everything and try the other options.
756 case Instruction::Alloca: {
757 const AllocaInst *AI = cast<AllocaInst>(Obj);
758 DenseMap<const AllocaInst*, int>::iterator SI =
759 FuncInfo.StaticAllocaMap.find(AI);
760 if (SI != FuncInfo.StaticAllocaMap.end()) {
761 Addr.BaseType = Address::FrameIndexBase;
762 Addr.Base.FI = SI->second;
769 // Materialize the global variable's address into a reg which can
770 // then be used later to load the variable.
771 if (const GlobalValue *GV = dyn_cast<GlobalValue>(Obj)) {
772 unsigned Tmp = ARMMaterializeGV(GV, TLI.getValueType(Obj->getType()));
773 if (Tmp == 0) return false;
779 // Try to get this in a register if nothing else has worked.
780 if (Addr.Base.Reg == 0) Addr.Base.Reg = getRegForValue(Obj);
781 return Addr.Base.Reg != 0;
784 void ARMFastISel::ARMSimplifyAddress(Address &Addr, EVT VT) {
786 assert(VT.isSimple() && "Non-simple types are invalid here!");
788 bool needsLowering = false;
789 switch (VT.getSimpleVT().SimpleTy) {
791 assert(false && "Unhandled load/store type!");
796 // Integer loads/stores handle 12-bit offsets.
797 needsLowering = ((Addr.Offset & 0xfff) != Addr.Offset);
801 // Floating point operands handle 8-bit offsets.
802 needsLowering = ((Addr.Offset & 0xff) != Addr.Offset);
806 // If this is a stack pointer and the offset needs to be simplified then
807 // put the alloca address into a register, set the base type back to
808 // register and continue. This should almost never happen.
809 if (needsLowering && Addr.BaseType == Address::FrameIndexBase) {
810 TargetRegisterClass *RC = isThumb ? ARM::tGPRRegisterClass :
811 ARM::GPRRegisterClass;
812 unsigned ResultReg = createResultReg(RC);
813 unsigned Opc = isThumb ? ARM::t2ADDri : ARM::ADDri;
814 AddOptionalDefs(BuildMI(*FuncInfo.MBB, *FuncInfo.InsertPt, DL,
815 TII.get(Opc), ResultReg)
816 .addFrameIndex(Addr.Base.FI)
818 Addr.Base.Reg = ResultReg;
819 Addr.BaseType = Address::RegBase;
822 // Since the offset is too large for the load/store instruction
823 // get the reg+offset into a register.
825 Addr.Base.Reg = FastEmit_ri_(MVT::i32, ISD::ADD, Addr.Base.Reg,
826 /*Op0IsKill*/false, Addr.Offset, MVT::i32);
831 void ARMFastISel::AddLoadStoreOperands(EVT VT, Address &Addr,
832 const MachineInstrBuilder &MIB) {
833 // addrmode5 output depends on the selection dag addressing dividing the
834 // offset by 4 that it then later multiplies. Do this here as well.
835 if (VT.getSimpleVT().SimpleTy == MVT::f32 ||
836 VT.getSimpleVT().SimpleTy == MVT::f64)
839 // Frame base works a bit differently. Handle it separately.
840 if (Addr.BaseType == Address::FrameIndexBase) {
841 int FI = Addr.Base.FI;
842 int Offset = Addr.Offset;
843 MachineMemOperand *MMO =
844 FuncInfo.MF->getMachineMemOperand(
845 MachinePointerInfo::getFixedStack(FI, Offset),
846 MachineMemOperand::MOLoad,
847 MFI.getObjectSize(FI),
848 MFI.getObjectAlignment(FI));
849 // Now add the rest of the operands.
850 MIB.addFrameIndex(FI);
852 // ARM halfword load/stores need an additional operand.
853 if (!isThumb && VT.getSimpleVT().SimpleTy == MVT::i16) MIB.addReg(0);
855 MIB.addImm(Addr.Offset);
856 MIB.addMemOperand(MMO);
858 // Now add the rest of the operands.
859 MIB.addReg(Addr.Base.Reg);
861 // ARM halfword load/stores need an additional operand.
862 if (!isThumb && VT.getSimpleVT().SimpleTy == MVT::i16) MIB.addReg(0);
864 MIB.addImm(Addr.Offset);
866 AddOptionalDefs(MIB);
869 bool ARMFastISel::ARMEmitLoad(EVT VT, unsigned &ResultReg, Address &Addr) {
871 assert(VT.isSimple() && "Non-simple types are invalid here!");
873 TargetRegisterClass *RC;
874 switch (VT.getSimpleVT().SimpleTy) {
875 // This is mostly going to be Neon/vector support.
876 default: return false;
878 Opc = isThumb ? ARM::t2LDRHi12 : ARM::LDRH;
879 RC = ARM::GPRRegisterClass;
882 Opc = isThumb ? ARM::t2LDRBi12 : ARM::LDRBi12;
883 RC = ARM::GPRRegisterClass;
886 Opc = isThumb ? ARM::t2LDRi12 : ARM::LDRi12;
887 RC = ARM::GPRRegisterClass;
891 RC = TLI.getRegClassFor(VT);
895 RC = TLI.getRegClassFor(VT);
898 // Simplify this down to something we can handle.
899 ARMSimplifyAddress(Addr, VT);
901 // Create the base instruction, then add the operands.
902 ResultReg = createResultReg(RC);
903 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
904 TII.get(Opc), ResultReg);
905 AddLoadStoreOperands(VT, Addr, MIB);
909 bool ARMFastISel::SelectLoad(const Instruction *I) {
910 // Verify we have a legal type before going any further.
912 if (!isLoadTypeLegal(I->getType(), VT))
915 // See if we can handle this address.
917 if (!ARMComputeAddress(I->getOperand(0), Addr)) return false;
920 if (!ARMEmitLoad(VT, ResultReg, Addr)) return false;
921 UpdateValueMap(I, ResultReg);
925 bool ARMFastISel::ARMEmitStore(EVT VT, unsigned SrcReg, Address &Addr) {
927 switch (VT.getSimpleVT().SimpleTy) {
928 // This is mostly going to be Neon/vector support.
929 default: return false;
931 unsigned Res = createResultReg(isThumb ? ARM::tGPRRegisterClass :
932 ARM::GPRRegisterClass);
933 unsigned Opc = isThumb ? ARM::t2ANDri : ARM::ANDri;
934 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
936 .addReg(SrcReg).addImm(1));
938 } // Fallthrough here.
940 StrOpc = isThumb ? ARM::t2STRBi12 : ARM::STRBi12;
943 StrOpc = isThumb ? ARM::t2STRHi12 : ARM::STRH;
946 StrOpc = isThumb ? ARM::t2STRi12 : ARM::STRi12;
949 if (!Subtarget->hasVFP2()) return false;
953 if (!Subtarget->hasVFP2()) return false;
957 // Simplify this down to something we can handle.
958 ARMSimplifyAddress(Addr, VT);
960 // Create the base instruction, then add the operands.
961 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
963 .addReg(SrcReg, getKillRegState(true));
964 AddLoadStoreOperands(VT, Addr, MIB);
968 bool ARMFastISel::SelectStore(const Instruction *I) {
969 Value *Op0 = I->getOperand(0);
972 // Verify we have a legal type before going any further.
974 if (!isLoadTypeLegal(I->getOperand(0)->getType(), VT))
977 // Get the value to be stored into a register.
978 SrcReg = getRegForValue(Op0);
979 if (SrcReg == 0) return false;
981 // See if we can handle this address.
983 if (!ARMComputeAddress(I->getOperand(1), Addr))
986 if (!ARMEmitStore(VT, SrcReg, Addr)) return false;
990 static ARMCC::CondCodes getComparePred(CmpInst::Predicate Pred) {
992 // Needs two compares...
993 case CmpInst::FCMP_ONE:
994 case CmpInst::FCMP_UEQ:
996 // AL is our "false" for now. The other two need more compares.
998 case CmpInst::ICMP_EQ:
999 case CmpInst::FCMP_OEQ:
1001 case CmpInst::ICMP_SGT:
1002 case CmpInst::FCMP_OGT:
1004 case CmpInst::ICMP_SGE:
1005 case CmpInst::FCMP_OGE:
1007 case CmpInst::ICMP_UGT:
1008 case CmpInst::FCMP_UGT:
1010 case CmpInst::FCMP_OLT:
1012 case CmpInst::ICMP_ULE:
1013 case CmpInst::FCMP_OLE:
1015 case CmpInst::FCMP_ORD:
1017 case CmpInst::FCMP_UNO:
1019 case CmpInst::FCMP_UGE:
1021 case CmpInst::ICMP_SLT:
1022 case CmpInst::FCMP_ULT:
1024 case CmpInst::ICMP_SLE:
1025 case CmpInst::FCMP_ULE:
1027 case CmpInst::FCMP_UNE:
1028 case CmpInst::ICMP_NE:
1030 case CmpInst::ICMP_UGE:
1032 case CmpInst::ICMP_ULT:
1037 bool ARMFastISel::SelectBranch(const Instruction *I) {
1038 const BranchInst *BI = cast<BranchInst>(I);
1039 MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
1040 MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
1042 // Simple branch support.
1044 // If we can, avoid recomputing the compare - redoing it could lead to wonky
1046 // TODO: Factor this out.
1047 if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
1048 if (CI->hasOneUse() && (CI->getParent() == I->getParent())) {
1050 const Type *Ty = CI->getOperand(0)->getType();
1051 if (!isTypeLegal(Ty, VT))
1054 bool isFloat = (Ty->isDoubleTy() || Ty->isFloatTy());
1055 if (isFloat && !Subtarget->hasVFP2())
1059 switch (VT.SimpleTy) {
1060 default: return false;
1061 // TODO: Verify compares.
1063 CmpOpc = ARM::VCMPES;
1066 CmpOpc = ARM::VCMPED;
1069 CmpOpc = isThumb ? ARM::t2CMPrr : ARM::CMPrr;
1073 // Get the compare predicate.
1074 ARMCC::CondCodes ARMPred = getComparePred(CI->getPredicate());
1076 // We may not handle every CC for now.
1077 if (ARMPred == ARMCC::AL) return false;
1079 unsigned Arg1 = getRegForValue(CI->getOperand(0));
1080 if (Arg1 == 0) return false;
1082 unsigned Arg2 = getRegForValue(CI->getOperand(1));
1083 if (Arg2 == 0) return false;
1085 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1087 .addReg(Arg1).addReg(Arg2));
1089 // For floating point we need to move the result to a comparison register
1090 // that we can then use for branches.
1092 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1093 TII.get(ARM::FMSTAT)));
1095 unsigned BrOpc = isThumb ? ARM::t2Bcc : ARM::Bcc;
1096 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
1097 .addMBB(TBB).addImm(ARMPred).addReg(ARM::CPSR);
1098 FastEmitBranch(FBB, DL);
1099 FuncInfo.MBB->addSuccessor(TBB);
1104 unsigned CmpReg = getRegForValue(BI->getCondition());
1105 if (CmpReg == 0) return false;
1107 // We've been divorced from our compare! Our block was split, and
1108 // now our compare lives in a predecessor block. We musn't
1109 // re-compare here, as the children of the compare aren't guaranteed
1110 // live across the block boundary (we *could* check for this).
1111 // Regardless, the compare has been done in the predecessor block,
1112 // and it left a value for us in a virtual register. Ergo, we test
1113 // the one-bit value left in the virtual register.
1114 unsigned TstOpc = isThumb ? ARM::t2TSTri : ARM::TSTri;
1115 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TstOpc))
1116 .addReg(CmpReg).addImm(1));
1119 unsigned CCMode = ARMCC::NE;
1120 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
1121 std::swap(TBB, FBB);
1125 unsigned BrOpc = isThumb ? ARM::t2Bcc : ARM::Bcc;
1126 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
1127 .addMBB(TBB).addImm(CCMode).addReg(ARM::CPSR);
1128 FastEmitBranch(FBB, DL);
1129 FuncInfo.MBB->addSuccessor(TBB);
1133 bool ARMFastISel::SelectCmp(const Instruction *I) {
1134 const CmpInst *CI = cast<CmpInst>(I);
1137 const Type *Ty = CI->getOperand(0)->getType();
1138 if (!isTypeLegal(Ty, VT))
1141 bool isFloat = (Ty->isDoubleTy() || Ty->isFloatTy());
1142 if (isFloat && !Subtarget->hasVFP2())
1147 switch (VT.SimpleTy) {
1148 default: return false;
1149 // TODO: Verify compares.
1151 CmpOpc = ARM::VCMPES;
1152 CondReg = ARM::FPSCR;
1155 CmpOpc = ARM::VCMPED;
1156 CondReg = ARM::FPSCR;
1159 CmpOpc = isThumb ? ARM::t2CMPrr : ARM::CMPrr;
1160 CondReg = ARM::CPSR;
1164 // Get the compare predicate.
1165 ARMCC::CondCodes ARMPred = getComparePred(CI->getPredicate());
1167 // We may not handle every CC for now.
1168 if (ARMPred == ARMCC::AL) return false;
1170 unsigned Arg1 = getRegForValue(CI->getOperand(0));
1171 if (Arg1 == 0) return false;
1173 unsigned Arg2 = getRegForValue(CI->getOperand(1));
1174 if (Arg2 == 0) return false;
1176 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
1177 .addReg(Arg1).addReg(Arg2));
1179 // For floating point we need to move the result to a comparison register
1180 // that we can then use for branches.
1182 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1183 TII.get(ARM::FMSTAT)));
1185 // Now set a register based on the comparison. Explicitly set the predicates
1187 unsigned MovCCOpc = isThumb ? ARM::t2MOVCCi : ARM::MOVCCi;
1188 TargetRegisterClass *RC = isThumb ? ARM::rGPRRegisterClass
1189 : ARM::GPRRegisterClass;
1190 unsigned DestReg = createResultReg(RC);
1192 = ConstantInt::get(Type::getInt32Ty(*Context), 0);
1193 unsigned ZeroReg = TargetMaterializeConstant(Zero);
1194 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), DestReg)
1195 .addReg(ZeroReg).addImm(1)
1196 .addImm(ARMPred).addReg(CondReg);
1198 UpdateValueMap(I, DestReg);
1202 bool ARMFastISel::SelectFPExt(const Instruction *I) {
1203 // Make sure we have VFP and that we're extending float to double.
1204 if (!Subtarget->hasVFP2()) return false;
1206 Value *V = I->getOperand(0);
1207 if (!I->getType()->isDoubleTy() ||
1208 !V->getType()->isFloatTy()) return false;
1210 unsigned Op = getRegForValue(V);
1211 if (Op == 0) return false;
1213 unsigned Result = createResultReg(ARM::DPRRegisterClass);
1214 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1215 TII.get(ARM::VCVTDS), Result)
1217 UpdateValueMap(I, Result);
1221 bool ARMFastISel::SelectFPTrunc(const Instruction *I) {
1222 // Make sure we have VFP and that we're truncating double to float.
1223 if (!Subtarget->hasVFP2()) return false;
1225 Value *V = I->getOperand(0);
1226 if (!(I->getType()->isFloatTy() &&
1227 V->getType()->isDoubleTy())) return false;
1229 unsigned Op = getRegForValue(V);
1230 if (Op == 0) return false;
1232 unsigned Result = createResultReg(ARM::SPRRegisterClass);
1233 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1234 TII.get(ARM::VCVTSD), Result)
1236 UpdateValueMap(I, Result);
1240 bool ARMFastISel::SelectSIToFP(const Instruction *I) {
1241 // Make sure we have VFP.
1242 if (!Subtarget->hasVFP2()) return false;
1245 const Type *Ty = I->getType();
1246 if (!isTypeLegal(Ty, DstVT))
1249 unsigned Op = getRegForValue(I->getOperand(0));
1250 if (Op == 0) return false;
1252 // The conversion routine works on fp-reg to fp-reg and the operand above
1253 // was an integer, move it to the fp registers if possible.
1254 unsigned FP = ARMMoveToFPReg(MVT::f32, Op);
1255 if (FP == 0) return false;
1258 if (Ty->isFloatTy()) Opc = ARM::VSITOS;
1259 else if (Ty->isDoubleTy()) Opc = ARM::VSITOD;
1262 unsigned ResultReg = createResultReg(TLI.getRegClassFor(DstVT));
1263 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
1266 UpdateValueMap(I, ResultReg);
1270 bool ARMFastISel::SelectFPToSI(const Instruction *I) {
1271 // Make sure we have VFP.
1272 if (!Subtarget->hasVFP2()) return false;
1275 const Type *RetTy = I->getType();
1276 if (!isTypeLegal(RetTy, DstVT))
1279 unsigned Op = getRegForValue(I->getOperand(0));
1280 if (Op == 0) return false;
1283 const Type *OpTy = I->getOperand(0)->getType();
1284 if (OpTy->isFloatTy()) Opc = ARM::VTOSIZS;
1285 else if (OpTy->isDoubleTy()) Opc = ARM::VTOSIZD;
1288 // f64->s32 or f32->s32 both need an intermediate f32 reg.
1289 unsigned ResultReg = createResultReg(TLI.getRegClassFor(MVT::f32));
1290 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
1294 // This result needs to be in an integer register, but the conversion only
1295 // takes place in fp-regs.
1296 unsigned IntReg = ARMMoveToIntReg(DstVT, ResultReg);
1297 if (IntReg == 0) return false;
1299 UpdateValueMap(I, IntReg);
1303 bool ARMFastISel::SelectSelect(const Instruction *I) {
1305 if (!isTypeLegal(I->getType(), VT))
1308 // Things need to be register sized for register moves.
1309 if (VT != MVT::i32) return false;
1310 const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
1312 unsigned CondReg = getRegForValue(I->getOperand(0));
1313 if (CondReg == 0) return false;
1314 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1315 if (Op1Reg == 0) return false;
1316 unsigned Op2Reg = getRegForValue(I->getOperand(2));
1317 if (Op2Reg == 0) return false;
1319 unsigned CmpOpc = isThumb ? ARM::t2TSTri : ARM::TSTri;
1320 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
1321 .addReg(CondReg).addImm(1));
1322 unsigned ResultReg = createResultReg(RC);
1323 unsigned MovCCOpc = isThumb ? ARM::t2MOVCCr : ARM::MOVCCr;
1324 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), ResultReg)
1325 .addReg(Op1Reg).addReg(Op2Reg)
1326 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
1327 UpdateValueMap(I, ResultReg);
1331 bool ARMFastISel::SelectSDiv(const Instruction *I) {
1333 const Type *Ty = I->getType();
1334 if (!isTypeLegal(Ty, VT))
1337 // If we have integer div support we should have selected this automagically.
1338 // In case we have a real miss go ahead and return false and we'll pick
1340 if (Subtarget->hasDivide()) return false;
1342 // Otherwise emit a libcall.
1343 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
1345 LC = RTLIB::SDIV_I8;
1346 else if (VT == MVT::i16)
1347 LC = RTLIB::SDIV_I16;
1348 else if (VT == MVT::i32)
1349 LC = RTLIB::SDIV_I32;
1350 else if (VT == MVT::i64)
1351 LC = RTLIB::SDIV_I64;
1352 else if (VT == MVT::i128)
1353 LC = RTLIB::SDIV_I128;
1354 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SDIV!");
1356 return ARMEmitLibcall(I, LC);
1359 bool ARMFastISel::SelectSRem(const Instruction *I) {
1361 const Type *Ty = I->getType();
1362 if (!isTypeLegal(Ty, VT))
1365 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
1367 LC = RTLIB::SREM_I8;
1368 else if (VT == MVT::i16)
1369 LC = RTLIB::SREM_I16;
1370 else if (VT == MVT::i32)
1371 LC = RTLIB::SREM_I32;
1372 else if (VT == MVT::i64)
1373 LC = RTLIB::SREM_I64;
1374 else if (VT == MVT::i128)
1375 LC = RTLIB::SREM_I128;
1376 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SREM!");
1378 return ARMEmitLibcall(I, LC);
1381 bool ARMFastISel::SelectBinaryOp(const Instruction *I, unsigned ISDOpcode) {
1382 EVT VT = TLI.getValueType(I->getType(), true);
1384 // We can get here in the case when we want to use NEON for our fp
1385 // operations, but can't figure out how to. Just use the vfp instructions
1387 // FIXME: It'd be nice to use NEON instructions.
1388 const Type *Ty = I->getType();
1389 bool isFloat = (Ty->isDoubleTy() || Ty->isFloatTy());
1390 if (isFloat && !Subtarget->hasVFP2())
1393 unsigned Op1 = getRegForValue(I->getOperand(0));
1394 if (Op1 == 0) return false;
1396 unsigned Op2 = getRegForValue(I->getOperand(1));
1397 if (Op2 == 0) return false;
1400 bool is64bit = VT == MVT::f64 || VT == MVT::i64;
1401 switch (ISDOpcode) {
1402 default: return false;
1404 Opc = is64bit ? ARM::VADDD : ARM::VADDS;
1407 Opc = is64bit ? ARM::VSUBD : ARM::VSUBS;
1410 Opc = is64bit ? ARM::VMULD : ARM::VMULS;
1413 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
1414 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1415 TII.get(Opc), ResultReg)
1416 .addReg(Op1).addReg(Op2));
1417 UpdateValueMap(I, ResultReg);
1421 // Call Handling Code
1423 bool ARMFastISel::FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src,
1424 EVT SrcVT, unsigned &ResultReg) {
1425 unsigned RR = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), Opc,
1426 Src, /*TODO: Kill=*/false);
1435 // This is largely taken directly from CCAssignFnForNode - we don't support
1436 // varargs in FastISel so that part has been removed.
1437 // TODO: We may not support all of this.
1438 CCAssignFn *ARMFastISel::CCAssignFnForCall(CallingConv::ID CC, bool Return) {
1441 llvm_unreachable("Unsupported calling convention");
1442 case CallingConv::Fast:
1443 // Ignore fastcc. Silence compiler warnings.
1444 (void)RetFastCC_ARM_APCS;
1445 (void)FastCC_ARM_APCS;
1447 case CallingConv::C:
1448 // Use target triple & subtarget features to do actual dispatch.
1449 if (Subtarget->isAAPCS_ABI()) {
1450 if (Subtarget->hasVFP2() &&
1451 FloatABIType == FloatABI::Hard)
1452 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
1454 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
1456 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
1457 case CallingConv::ARM_AAPCS_VFP:
1458 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
1459 case CallingConv::ARM_AAPCS:
1460 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
1461 case CallingConv::ARM_APCS:
1462 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
1466 bool ARMFastISel::ProcessCallArgs(SmallVectorImpl<Value*> &Args,
1467 SmallVectorImpl<unsigned> &ArgRegs,
1468 SmallVectorImpl<MVT> &ArgVTs,
1469 SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
1470 SmallVectorImpl<unsigned> &RegArgs,
1472 unsigned &NumBytes) {
1473 SmallVector<CCValAssign, 16> ArgLocs;
1474 CCState CCInfo(CC, false, TM, ArgLocs, *Context);
1475 CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CCAssignFnForCall(CC, false));
1477 // Get a count of how many bytes are to be pushed on the stack.
1478 NumBytes = CCInfo.getNextStackOffset();
1480 // Issue CALLSEQ_START
1481 unsigned AdjStackDown = TM.getRegisterInfo()->getCallFrameSetupOpcode();
1482 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1483 TII.get(AdjStackDown))
1486 // Process the args.
1487 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1488 CCValAssign &VA = ArgLocs[i];
1489 unsigned Arg = ArgRegs[VA.getValNo()];
1490 MVT ArgVT = ArgVTs[VA.getValNo()];
1492 // We don't handle NEON/vector parameters yet.
1493 if (ArgVT.isVector() || ArgVT.getSizeInBits() > 64)
1496 // Handle arg promotion, etc.
1497 switch (VA.getLocInfo()) {
1498 case CCValAssign::Full: break;
1499 case CCValAssign::SExt: {
1500 bool Emitted = FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
1502 assert(Emitted && "Failed to emit a sext!"); (void)Emitted;
1504 ArgVT = VA.getLocVT();
1507 case CCValAssign::ZExt: {
1508 bool Emitted = FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
1510 assert(Emitted && "Failed to emit a zext!"); (void)Emitted;
1512 ArgVT = VA.getLocVT();
1515 case CCValAssign::AExt: {
1516 bool Emitted = FastEmitExtend(ISD::ANY_EXTEND, VA.getLocVT(),
1519 Emitted = FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
1522 Emitted = FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
1525 assert(Emitted && "Failed to emit a aext!"); (void)Emitted;
1526 ArgVT = VA.getLocVT();
1529 case CCValAssign::BCvt: {
1530 unsigned BC = FastEmit_r(ArgVT, VA.getLocVT(), ISD::BITCAST, Arg,
1531 /*TODO: Kill=*/false);
1532 assert(BC != 0 && "Failed to emit a bitcast!");
1534 ArgVT = VA.getLocVT();
1537 default: llvm_unreachable("Unknown arg promotion!");
1540 // Now copy/store arg to correct locations.
1541 if (VA.isRegLoc() && !VA.needsCustom()) {
1542 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1545 RegArgs.push_back(VA.getLocReg());
1546 } else if (VA.needsCustom()) {
1547 // TODO: We need custom lowering for vector (v2f64) args.
1548 if (VA.getLocVT() != MVT::f64) return false;
1550 CCValAssign &NextVA = ArgLocs[++i];
1552 // TODO: Only handle register args for now.
1553 if(!(VA.isRegLoc() && NextVA.isRegLoc())) return false;
1555 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1556 TII.get(ARM::VMOVRRD), VA.getLocReg())
1557 .addReg(NextVA.getLocReg(), RegState::Define)
1559 RegArgs.push_back(VA.getLocReg());
1560 RegArgs.push_back(NextVA.getLocReg());
1562 assert(VA.isMemLoc());
1563 // Need to store on the stack.
1565 Addr.BaseType = Address::RegBase;
1566 Addr.Base.Reg = ARM::SP;
1567 Addr.Offset = VA.getLocMemOffset();
1569 if (!ARMEmitStore(ArgVT, Arg, Addr)) return false;
1575 bool ARMFastISel::FinishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
1576 const Instruction *I, CallingConv::ID CC,
1577 unsigned &NumBytes) {
1578 // Issue CALLSEQ_END
1579 unsigned AdjStackUp = TM.getRegisterInfo()->getCallFrameDestroyOpcode();
1580 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1581 TII.get(AdjStackUp))
1582 .addImm(NumBytes).addImm(0));
1584 // Now the return value.
1585 if (RetVT != MVT::isVoid) {
1586 SmallVector<CCValAssign, 16> RVLocs;
1587 CCState CCInfo(CC, false, TM, RVLocs, *Context);
1588 CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC, true));
1590 // Copy all of the result registers out of their specified physreg.
1591 if (RVLocs.size() == 2 && RetVT == MVT::f64) {
1592 // For this move we copy into two registers and then move into the
1593 // double fp reg we want.
1594 EVT DestVT = RVLocs[0].getValVT();
1595 TargetRegisterClass* DstRC = TLI.getRegClassFor(DestVT);
1596 unsigned ResultReg = createResultReg(DstRC);
1597 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1598 TII.get(ARM::VMOVDRR), ResultReg)
1599 .addReg(RVLocs[0].getLocReg())
1600 .addReg(RVLocs[1].getLocReg()));
1602 UsedRegs.push_back(RVLocs[0].getLocReg());
1603 UsedRegs.push_back(RVLocs[1].getLocReg());
1605 // Finally update the result.
1606 UpdateValueMap(I, ResultReg);
1608 assert(RVLocs.size() == 1 &&"Can't handle non-double multi-reg retvals!");
1609 EVT CopyVT = RVLocs[0].getValVT();
1610 TargetRegisterClass* DstRC = TLI.getRegClassFor(CopyVT);
1612 unsigned ResultReg = createResultReg(DstRC);
1613 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1614 ResultReg).addReg(RVLocs[0].getLocReg());
1615 UsedRegs.push_back(RVLocs[0].getLocReg());
1617 // Finally update the result.
1618 UpdateValueMap(I, ResultReg);
1625 bool ARMFastISel::SelectRet(const Instruction *I) {
1626 const ReturnInst *Ret = cast<ReturnInst>(I);
1627 const Function &F = *I->getParent()->getParent();
1629 if (!FuncInfo.CanLowerReturn)
1635 CallingConv::ID CC = F.getCallingConv();
1636 if (Ret->getNumOperands() > 0) {
1637 SmallVector<ISD::OutputArg, 4> Outs;
1638 GetReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(),
1641 // Analyze operands of the call, assigning locations to each operand.
1642 SmallVector<CCValAssign, 16> ValLocs;
1643 CCState CCInfo(CC, F.isVarArg(), TM, ValLocs, I->getContext());
1644 CCInfo.AnalyzeReturn(Outs, CCAssignFnForCall(CC, true /* is Ret */));
1646 const Value *RV = Ret->getOperand(0);
1647 unsigned Reg = getRegForValue(RV);
1651 // Only handle a single return value for now.
1652 if (ValLocs.size() != 1)
1655 CCValAssign &VA = ValLocs[0];
1657 // Don't bother handling odd stuff for now.
1658 if (VA.getLocInfo() != CCValAssign::Full)
1660 // Only handle register returns for now.
1663 // TODO: For now, don't try to handle cases where getLocInfo()
1664 // says Full but the types don't match.
1665 if (TLI.getValueType(RV->getType()) != VA.getValVT())
1669 unsigned SrcReg = Reg + VA.getValNo();
1670 unsigned DstReg = VA.getLocReg();
1671 const TargetRegisterClass* SrcRC = MRI.getRegClass(SrcReg);
1672 // Avoid a cross-class copy. This is very unlikely.
1673 if (!SrcRC->contains(DstReg))
1675 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1676 DstReg).addReg(SrcReg);
1678 // Mark the register as live out of the function.
1679 MRI.addLiveOut(VA.getLocReg());
1682 unsigned RetOpc = isThumb ? ARM::tBX_RET : ARM::BX_RET;
1683 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1688 unsigned ARMFastISel::ARMSelectCallOp(const GlobalValue *GV) {
1690 // Darwin needs the r9 versions of the opcodes.
1691 bool isDarwin = Subtarget->isTargetDarwin();
1693 return isDarwin ? ARM::tBLr9 : ARM::tBL;
1695 return isDarwin ? ARM::BLr9 : ARM::BL;
1699 // A quick function that will emit a call for a named libcall in F with the
1700 // vector of passed arguments for the Instruction in I. We can assume that we
1701 // can emit a call for any libcall we can produce. This is an abridged version
1702 // of the full call infrastructure since we won't need to worry about things
1703 // like computed function pointers or strange arguments at call sites.
1704 // TODO: Try to unify this and the normal call bits for ARM, then try to unify
1706 bool ARMFastISel::ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call) {
1707 CallingConv::ID CC = TLI.getLibcallCallingConv(Call);
1709 // Handle *simple* calls for now.
1710 const Type *RetTy = I->getType();
1712 if (RetTy->isVoidTy())
1713 RetVT = MVT::isVoid;
1714 else if (!isTypeLegal(RetTy, RetVT))
1717 // For now we're using BLX etc on the assumption that we have v5t ops.
1718 if (!Subtarget->hasV5TOps()) return false;
1720 // TODO: For now if we have long calls specified we don't handle the call.
1721 if (EnableARMLongCalls) return false;
1723 // Set up the argument vectors.
1724 SmallVector<Value*, 8> Args;
1725 SmallVector<unsigned, 8> ArgRegs;
1726 SmallVector<MVT, 8> ArgVTs;
1727 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
1728 Args.reserve(I->getNumOperands());
1729 ArgRegs.reserve(I->getNumOperands());
1730 ArgVTs.reserve(I->getNumOperands());
1731 ArgFlags.reserve(I->getNumOperands());
1732 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
1733 Value *Op = I->getOperand(i);
1734 unsigned Arg = getRegForValue(Op);
1735 if (Arg == 0) return false;
1737 const Type *ArgTy = Op->getType();
1739 if (!isTypeLegal(ArgTy, ArgVT)) return false;
1741 ISD::ArgFlagsTy Flags;
1742 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
1743 Flags.setOrigAlign(OriginalAlignment);
1746 ArgRegs.push_back(Arg);
1747 ArgVTs.push_back(ArgVT);
1748 ArgFlags.push_back(Flags);
1751 // Handle the arguments now that we've gotten them.
1752 SmallVector<unsigned, 4> RegArgs;
1754 if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
1757 // Issue the call, BLXr9 for darwin, BLX otherwise. This uses V5 ops.
1758 // TODO: Turn this into the table of arm call ops.
1759 MachineInstrBuilder MIB;
1760 unsigned CallOpc = ARMSelectCallOp(NULL);
1762 // Explicitly adding the predicate here.
1763 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1765 .addExternalSymbol(TLI.getLibcallName(Call));
1767 // Explicitly adding the predicate here.
1768 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1770 .addExternalSymbol(TLI.getLibcallName(Call)));
1772 // Add implicit physical register uses to the call.
1773 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
1774 MIB.addReg(RegArgs[i]);
1776 // Finish off the call including any return values.
1777 SmallVector<unsigned, 4> UsedRegs;
1778 if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false;
1780 // Set all unused physreg defs as dead.
1781 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
1786 bool ARMFastISel::SelectCall(const Instruction *I) {
1787 const CallInst *CI = cast<CallInst>(I);
1788 const Value *Callee = CI->getCalledValue();
1790 // Can't handle inline asm or worry about intrinsics yet.
1791 if (isa<InlineAsm>(Callee) || isa<IntrinsicInst>(CI)) return false;
1793 // Only handle global variable Callees that are direct calls.
1794 const GlobalValue *GV = dyn_cast<GlobalValue>(Callee);
1795 if (!GV || Subtarget->GVIsIndirectSymbol(GV, TM.getRelocationModel()))
1798 // Check the calling convention.
1799 ImmutableCallSite CS(CI);
1800 CallingConv::ID CC = CS.getCallingConv();
1802 // TODO: Avoid some calling conventions?
1804 // Let SDISel handle vararg functions.
1805 const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
1806 const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
1807 if (FTy->isVarArg())
1810 // Handle *simple* calls for now.
1811 const Type *RetTy = I->getType();
1813 if (RetTy->isVoidTy())
1814 RetVT = MVT::isVoid;
1815 else if (!isTypeLegal(RetTy, RetVT))
1818 // For now we're using BLX etc on the assumption that we have v5t ops.
1820 if (!Subtarget->hasV5TOps()) return false;
1822 // TODO: For now if we have long calls specified we don't handle the call.
1823 if (EnableARMLongCalls) return false;
1825 // Set up the argument vectors.
1826 SmallVector<Value*, 8> Args;
1827 SmallVector<unsigned, 8> ArgRegs;
1828 SmallVector<MVT, 8> ArgVTs;
1829 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
1830 Args.reserve(CS.arg_size());
1831 ArgRegs.reserve(CS.arg_size());
1832 ArgVTs.reserve(CS.arg_size());
1833 ArgFlags.reserve(CS.arg_size());
1834 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
1836 unsigned Arg = getRegForValue(*i);
1840 ISD::ArgFlagsTy Flags;
1841 unsigned AttrInd = i - CS.arg_begin() + 1;
1842 if (CS.paramHasAttr(AttrInd, Attribute::SExt))
1844 if (CS.paramHasAttr(AttrInd, Attribute::ZExt))
1847 // FIXME: Only handle *easy* calls for now.
1848 if (CS.paramHasAttr(AttrInd, Attribute::InReg) ||
1849 CS.paramHasAttr(AttrInd, Attribute::StructRet) ||
1850 CS.paramHasAttr(AttrInd, Attribute::Nest) ||
1851 CS.paramHasAttr(AttrInd, Attribute::ByVal))
1854 const Type *ArgTy = (*i)->getType();
1856 if (!isTypeLegal(ArgTy, ArgVT))
1858 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
1859 Flags.setOrigAlign(OriginalAlignment);
1862 ArgRegs.push_back(Arg);
1863 ArgVTs.push_back(ArgVT);
1864 ArgFlags.push_back(Flags);
1867 // Handle the arguments now that we've gotten them.
1868 SmallVector<unsigned, 4> RegArgs;
1870 if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
1873 // Issue the call, BLXr9 for darwin, BLX otherwise. This uses V5 ops.
1874 // TODO: Turn this into the table of arm call ops.
1875 MachineInstrBuilder MIB;
1876 unsigned CallOpc = ARMSelectCallOp(GV);
1877 // Explicitly adding the predicate here.
1879 // Explicitly adding the predicate here.
1880 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1882 .addGlobalAddress(GV, 0, 0);
1884 // Explicitly adding the predicate here.
1885 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1887 .addGlobalAddress(GV, 0, 0));
1889 // Add implicit physical register uses to the call.
1890 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
1891 MIB.addReg(RegArgs[i]);
1893 // Finish off the call including any return values.
1894 SmallVector<unsigned, 4> UsedRegs;
1895 if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false;
1897 // Set all unused physreg defs as dead.
1898 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
1904 // TODO: SoftFP support.
1905 bool ARMFastISel::TargetSelectInstruction(const Instruction *I) {
1907 switch (I->getOpcode()) {
1908 case Instruction::Load:
1909 return SelectLoad(I);
1910 case Instruction::Store:
1911 return SelectStore(I);
1912 case Instruction::Br:
1913 return SelectBranch(I);
1914 case Instruction::ICmp:
1915 case Instruction::FCmp:
1916 return SelectCmp(I);
1917 case Instruction::FPExt:
1918 return SelectFPExt(I);
1919 case Instruction::FPTrunc:
1920 return SelectFPTrunc(I);
1921 case Instruction::SIToFP:
1922 return SelectSIToFP(I);
1923 case Instruction::FPToSI:
1924 return SelectFPToSI(I);
1925 case Instruction::FAdd:
1926 return SelectBinaryOp(I, ISD::FADD);
1927 case Instruction::FSub:
1928 return SelectBinaryOp(I, ISD::FSUB);
1929 case Instruction::FMul:
1930 return SelectBinaryOp(I, ISD::FMUL);
1931 case Instruction::SDiv:
1932 return SelectSDiv(I);
1933 case Instruction::SRem:
1934 return SelectSRem(I);
1935 case Instruction::Call:
1936 return SelectCall(I);
1937 case Instruction::Select:
1938 return SelectSelect(I);
1939 case Instruction::Ret:
1940 return SelectRet(I);
1947 llvm::FastISel *ARM::createFastISel(FunctionLoweringInfo &funcInfo) {
1948 // Completely untested on non-darwin.
1949 const TargetMachine &TM = funcInfo.MF->getTarget();
1951 // Darwin and thumb1 only for now.
1952 const ARMSubtarget *Subtarget = &TM.getSubtarget<ARMSubtarget>();
1953 if (Subtarget->isTargetDarwin() && !Subtarget->isThumb1Only() &&
1954 !DisableARMFastISel)
1955 return new ARMFastISel(funcInfo);