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/CodeGen/Analysis.h"
30 #include "llvm/CodeGen/FastISel.h"
31 #include "llvm/CodeGen/FunctionLoweringInfo.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineModuleInfo.h"
34 #include "llvm/CodeGen/MachineConstantPool.h"
35 #include "llvm/CodeGen/MachineFrameInfo.h"
36 #include "llvm/CodeGen/MachineMemOperand.h"
37 #include "llvm/CodeGen/MachineRegisterInfo.h"
38 #include "llvm/CodeGen/PseudoSourceValue.h"
39 #include "llvm/Support/CallSite.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/ErrorHandling.h"
42 #include "llvm/Support/GetElementPtrTypeIterator.h"
43 #include "llvm/Target/TargetData.h"
44 #include "llvm/Target/TargetInstrInfo.h"
45 #include "llvm/Target/TargetLowering.h"
46 #include "llvm/Target/TargetMachine.h"
47 #include "llvm/Target/TargetOptions.h"
51 DisableARMFastISel("disable-arm-fast-isel",
52 cl::desc("Turn off experimental ARM fast-isel support"),
53 cl::init(false), cl::Hidden);
55 extern cl::opt<bool> EnableARMLongCalls;
59 // All possible address modes, plus some.
60 typedef struct Address {
75 // Innocuous defaults for our address.
77 : BaseType(RegBase), Offset(0), Scale(0), PlusReg(0) {
82 class ARMFastISel : public FastISel {
84 /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
85 /// make the right decision when generating code for different targets.
86 const ARMSubtarget *Subtarget;
87 const TargetMachine &TM;
88 const TargetInstrInfo &TII;
89 const TargetLowering &TLI;
92 // Convenience variables to avoid some queries.
97 explicit ARMFastISel(FunctionLoweringInfo &funcInfo)
99 TM(funcInfo.MF->getTarget()),
100 TII(*TM.getInstrInfo()),
101 TLI(*TM.getTargetLowering()) {
102 Subtarget = &TM.getSubtarget<ARMSubtarget>();
103 AFI = funcInfo.MF->getInfo<ARMFunctionInfo>();
104 isThumb = AFI->isThumbFunction();
105 Context = &funcInfo.Fn->getContext();
108 // Code from FastISel.cpp.
109 virtual unsigned FastEmitInst_(unsigned MachineInstOpcode,
110 const TargetRegisterClass *RC);
111 virtual unsigned FastEmitInst_r(unsigned MachineInstOpcode,
112 const TargetRegisterClass *RC,
113 unsigned Op0, bool Op0IsKill);
114 virtual unsigned FastEmitInst_rr(unsigned MachineInstOpcode,
115 const TargetRegisterClass *RC,
116 unsigned Op0, bool Op0IsKill,
117 unsigned Op1, bool Op1IsKill);
118 virtual unsigned FastEmitInst_ri(unsigned MachineInstOpcode,
119 const TargetRegisterClass *RC,
120 unsigned Op0, bool Op0IsKill,
122 virtual unsigned FastEmitInst_rf(unsigned MachineInstOpcode,
123 const TargetRegisterClass *RC,
124 unsigned Op0, bool Op0IsKill,
125 const ConstantFP *FPImm);
126 virtual unsigned FastEmitInst_i(unsigned MachineInstOpcode,
127 const TargetRegisterClass *RC,
129 virtual unsigned FastEmitInst_rri(unsigned MachineInstOpcode,
130 const TargetRegisterClass *RC,
131 unsigned Op0, bool Op0IsKill,
132 unsigned Op1, bool Op1IsKill,
134 virtual unsigned FastEmitInst_extractsubreg(MVT RetVT,
135 unsigned Op0, bool Op0IsKill,
138 // Backend specific FastISel code.
139 virtual bool TargetSelectInstruction(const Instruction *I);
140 virtual unsigned TargetMaterializeConstant(const Constant *C);
141 virtual unsigned TargetMaterializeAlloca(const AllocaInst *AI);
143 #include "ARMGenFastISel.inc"
145 // Instruction selection routines.
147 bool SelectLoad(const Instruction *I);
148 bool SelectStore(const Instruction *I);
149 bool SelectBranch(const Instruction *I);
150 bool SelectCmp(const Instruction *I);
151 bool SelectFPExt(const Instruction *I);
152 bool SelectFPTrunc(const Instruction *I);
153 bool SelectBinaryOp(const Instruction *I, unsigned ISDOpcode);
154 bool SelectSIToFP(const Instruction *I);
155 bool SelectFPToSI(const Instruction *I);
156 bool SelectSDiv(const Instruction *I);
157 bool SelectSRem(const Instruction *I);
158 bool SelectCall(const Instruction *I);
159 bool SelectSelect(const Instruction *I);
160 bool SelectRet(const Instruction *I);
164 bool isTypeLegal(const Type *Ty, MVT &VT);
165 bool isLoadTypeLegal(const Type *Ty, MVT &VT);
166 bool ARMEmitLoad(EVT VT, unsigned &ResultReg, Address &Addr);
167 bool ARMEmitStore(EVT VT, unsigned SrcReg, Address &Addr);
168 bool ARMComputeAddress(const Value *Obj, Address &Addr);
169 void ARMSimplifyAddress(Address &Addr, EVT VT);
170 unsigned ARMMaterializeFP(const ConstantFP *CFP, EVT VT);
171 unsigned ARMMaterializeInt(const Constant *C, EVT VT);
172 unsigned ARMMaterializeGV(const GlobalValue *GV, EVT VT);
173 unsigned ARMMoveToFPReg(EVT VT, unsigned SrcReg);
174 unsigned ARMMoveToIntReg(EVT VT, unsigned SrcReg);
175 unsigned ARMSelectCallOp(const GlobalValue *GV);
177 // Call handling routines.
179 bool FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src, EVT SrcVT,
180 unsigned &ResultReg);
181 CCAssignFn *CCAssignFnForCall(CallingConv::ID CC, bool Return);
182 bool ProcessCallArgs(SmallVectorImpl<Value*> &Args,
183 SmallVectorImpl<unsigned> &ArgRegs,
184 SmallVectorImpl<MVT> &ArgVTs,
185 SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
186 SmallVectorImpl<unsigned> &RegArgs,
189 bool FinishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
190 const Instruction *I, CallingConv::ID CC,
192 bool ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call);
194 // OptionalDef handling routines.
196 bool DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR);
197 const MachineInstrBuilder &AddOptionalDefs(const MachineInstrBuilder &MIB);
198 void AddLoadStoreOperands(EVT VT, Address &Addr,
199 const MachineInstrBuilder &MIB);
202 } // end anonymous namespace
204 #include "ARMGenCallingConv.inc"
206 // DefinesOptionalPredicate - This is different from DefinesPredicate in that
207 // we don't care about implicit defs here, just places we'll need to add a
208 // default CCReg argument. Sets CPSR if we're setting CPSR instead of CCR.
209 bool ARMFastISel::DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR) {
210 const TargetInstrDesc &TID = MI->getDesc();
211 if (!TID.hasOptionalDef())
214 // Look to see if our OptionalDef is defining CPSR or CCR.
215 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
216 const MachineOperand &MO = MI->getOperand(i);
217 if (!MO.isReg() || !MO.isDef()) continue;
218 if (MO.getReg() == ARM::CPSR)
224 // If the machine is predicable go ahead and add the predicate operands, if
225 // it needs default CC operands add those.
226 // TODO: If we want to support thumb1 then we'll need to deal with optional
227 // CPSR defs that need to be added before the remaining operands. See s_cc_out
228 // for descriptions why.
229 const MachineInstrBuilder &
230 ARMFastISel::AddOptionalDefs(const MachineInstrBuilder &MIB) {
231 MachineInstr *MI = &*MIB;
233 // Do we use a predicate?
234 if (TII.isPredicable(MI))
237 // Do we optionally set a predicate? Preds is size > 0 iff the predicate
238 // defines CPSR. All other OptionalDefines in ARM are the CCR register.
240 if (DefinesOptionalPredicate(MI, &CPSR)) {
249 unsigned ARMFastISel::FastEmitInst_(unsigned MachineInstOpcode,
250 const TargetRegisterClass* RC) {
251 unsigned ResultReg = createResultReg(RC);
252 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
254 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg));
258 unsigned ARMFastISel::FastEmitInst_r(unsigned MachineInstOpcode,
259 const TargetRegisterClass *RC,
260 unsigned Op0, bool Op0IsKill) {
261 unsigned ResultReg = createResultReg(RC);
262 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
264 if (II.getNumDefs() >= 1)
265 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
266 .addReg(Op0, Op0IsKill * RegState::Kill));
268 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
269 .addReg(Op0, Op0IsKill * RegState::Kill));
270 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
271 TII.get(TargetOpcode::COPY), ResultReg)
272 .addReg(II.ImplicitDefs[0]));
277 unsigned ARMFastISel::FastEmitInst_rr(unsigned MachineInstOpcode,
278 const TargetRegisterClass *RC,
279 unsigned Op0, bool Op0IsKill,
280 unsigned Op1, bool Op1IsKill) {
281 unsigned ResultReg = createResultReg(RC);
282 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
284 if (II.getNumDefs() >= 1)
285 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
286 .addReg(Op0, Op0IsKill * RegState::Kill)
287 .addReg(Op1, Op1IsKill * RegState::Kill));
289 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
290 .addReg(Op0, Op0IsKill * RegState::Kill)
291 .addReg(Op1, Op1IsKill * RegState::Kill));
292 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
293 TII.get(TargetOpcode::COPY), ResultReg)
294 .addReg(II.ImplicitDefs[0]));
299 unsigned ARMFastISel::FastEmitInst_ri(unsigned MachineInstOpcode,
300 const TargetRegisterClass *RC,
301 unsigned Op0, bool Op0IsKill,
303 unsigned ResultReg = createResultReg(RC);
304 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
306 if (II.getNumDefs() >= 1)
307 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
308 .addReg(Op0, Op0IsKill * RegState::Kill)
311 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
312 .addReg(Op0, Op0IsKill * RegState::Kill)
314 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
315 TII.get(TargetOpcode::COPY), ResultReg)
316 .addReg(II.ImplicitDefs[0]));
321 unsigned ARMFastISel::FastEmitInst_rf(unsigned MachineInstOpcode,
322 const TargetRegisterClass *RC,
323 unsigned Op0, bool Op0IsKill,
324 const ConstantFP *FPImm) {
325 unsigned ResultReg = createResultReg(RC);
326 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
328 if (II.getNumDefs() >= 1)
329 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
330 .addReg(Op0, Op0IsKill * RegState::Kill)
333 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
334 .addReg(Op0, Op0IsKill * RegState::Kill)
336 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
337 TII.get(TargetOpcode::COPY), ResultReg)
338 .addReg(II.ImplicitDefs[0]));
343 unsigned ARMFastISel::FastEmitInst_rri(unsigned MachineInstOpcode,
344 const TargetRegisterClass *RC,
345 unsigned Op0, bool Op0IsKill,
346 unsigned Op1, bool Op1IsKill,
348 unsigned ResultReg = createResultReg(RC);
349 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
351 if (II.getNumDefs() >= 1)
352 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
353 .addReg(Op0, Op0IsKill * RegState::Kill)
354 .addReg(Op1, Op1IsKill * RegState::Kill)
357 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
358 .addReg(Op0, Op0IsKill * RegState::Kill)
359 .addReg(Op1, Op1IsKill * RegState::Kill)
361 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
362 TII.get(TargetOpcode::COPY), ResultReg)
363 .addReg(II.ImplicitDefs[0]));
368 unsigned ARMFastISel::FastEmitInst_i(unsigned MachineInstOpcode,
369 const TargetRegisterClass *RC,
371 unsigned ResultReg = createResultReg(RC);
372 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
374 if (II.getNumDefs() >= 1)
375 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
378 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
380 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
381 TII.get(TargetOpcode::COPY), ResultReg)
382 .addReg(II.ImplicitDefs[0]));
387 unsigned ARMFastISel::FastEmitInst_extractsubreg(MVT RetVT,
388 unsigned Op0, bool Op0IsKill,
390 unsigned ResultReg = createResultReg(TLI.getRegClassFor(RetVT));
391 assert(TargetRegisterInfo::isVirtualRegister(Op0) &&
392 "Cannot yet extract from physregs");
393 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
394 DL, TII.get(TargetOpcode::COPY), ResultReg)
395 .addReg(Op0, getKillRegState(Op0IsKill), Idx));
399 // TODO: Don't worry about 64-bit now, but when this is fixed remove the
400 // checks from the various callers.
401 unsigned ARMFastISel::ARMMoveToFPReg(EVT VT, unsigned SrcReg) {
402 if (VT == MVT::f64) return 0;
404 unsigned MoveReg = createResultReg(TLI.getRegClassFor(VT));
405 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
406 TII.get(ARM::VMOVRS), MoveReg)
411 unsigned ARMFastISel::ARMMoveToIntReg(EVT VT, unsigned SrcReg) {
412 if (VT == MVT::i64) return 0;
414 unsigned MoveReg = createResultReg(TLI.getRegClassFor(VT));
415 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
416 TII.get(ARM::VMOVSR), MoveReg)
421 // For double width floating point we need to materialize two constants
422 // (the high and the low) into integer registers then use a move to get
423 // the combined constant into an FP reg.
424 unsigned ARMFastISel::ARMMaterializeFP(const ConstantFP *CFP, EVT VT) {
425 const APFloat Val = CFP->getValueAPF();
426 bool is64bit = VT == MVT::f64;
428 // This checks to see if we can use VFP3 instructions to materialize
429 // a constant, otherwise we have to go through the constant pool.
430 if (TLI.isFPImmLegal(Val, VT)) {
431 unsigned Opc = is64bit ? ARM::FCONSTD : ARM::FCONSTS;
432 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
433 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
439 // Require VFP2 for loading fp constants.
440 if (!Subtarget->hasVFP2()) return false;
442 // MachineConstantPool wants an explicit alignment.
443 unsigned Align = TD.getPrefTypeAlignment(CFP->getType());
445 // TODO: Figure out if this is correct.
446 Align = TD.getTypeAllocSize(CFP->getType());
448 unsigned Idx = MCP.getConstantPoolIndex(cast<Constant>(CFP), Align);
449 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
450 unsigned Opc = is64bit ? ARM::VLDRD : ARM::VLDRS;
452 // The extra reg is for addrmode5.
453 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
455 .addConstantPoolIndex(Idx)
460 unsigned ARMFastISel::ARMMaterializeInt(const Constant *C, EVT VT) {
462 // For now 32-bit only.
463 if (VT != MVT::i32) return false;
465 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
467 // If we can do this in a single instruction without a constant pool entry
469 const ConstantInt *CI = cast<ConstantInt>(C);
470 if (Subtarget->hasV6T2Ops() && isUInt<16>(CI->getSExtValue())) {
471 unsigned Opc = isThumb ? ARM::t2MOVi16 : ARM::MOVi16;
472 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
473 TII.get(Opc), DestReg)
474 .addImm(CI->getSExtValue()));
478 // MachineConstantPool wants an explicit alignment.
479 unsigned Align = TD.getPrefTypeAlignment(C->getType());
481 // TODO: Figure out if this is correct.
482 Align = TD.getTypeAllocSize(C->getType());
484 unsigned Idx = MCP.getConstantPoolIndex(C, Align);
487 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
488 TII.get(ARM::t2LDRpci), DestReg)
489 .addConstantPoolIndex(Idx));
491 // The extra immediate is for addrmode2.
492 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
493 TII.get(ARM::LDRcp), DestReg)
494 .addConstantPoolIndex(Idx)
500 unsigned ARMFastISel::ARMMaterializeGV(const GlobalValue *GV, EVT VT) {
501 // For now 32-bit only.
502 if (VT != MVT::i32) return 0;
504 Reloc::Model RelocM = TM.getRelocationModel();
506 // TODO: No external globals for now.
507 if (Subtarget->GVIsIndirectSymbol(GV, RelocM)) return 0;
509 // TODO: Need more magic for ARM PIC.
510 if (!isThumb && (RelocM == Reloc::PIC_)) return 0;
512 // MachineConstantPool wants an explicit alignment.
513 unsigned Align = TD.getPrefTypeAlignment(GV->getType());
515 // TODO: Figure out if this is correct.
516 Align = TD.getTypeAllocSize(GV->getType());
520 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb() ? 4 : 8);
521 unsigned Id = AFI->createPICLabelUId();
522 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, Id,
523 ARMCP::CPValue, PCAdj);
524 unsigned Idx = MCP.getConstantPoolIndex(CPV, Align);
527 MachineInstrBuilder MIB;
528 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
530 unsigned Opc = (RelocM != Reloc::PIC_) ? ARM::t2LDRpci : ARM::t2LDRpci_pic;
531 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), DestReg)
532 .addConstantPoolIndex(Idx);
533 if (RelocM == Reloc::PIC_)
536 // The extra immediate is for addrmode2.
537 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(ARM::LDRcp),
539 .addConstantPoolIndex(Idx)
542 AddOptionalDefs(MIB);
546 unsigned ARMFastISel::TargetMaterializeConstant(const Constant *C) {
547 EVT VT = TLI.getValueType(C->getType(), true);
549 // Only handle simple types.
550 if (!VT.isSimple()) return 0;
552 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
553 return ARMMaterializeFP(CFP, VT);
554 else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
555 return ARMMaterializeGV(GV, VT);
556 else if (isa<ConstantInt>(C))
557 return ARMMaterializeInt(C, VT);
562 unsigned ARMFastISel::TargetMaterializeAlloca(const AllocaInst *AI) {
563 // Don't handle dynamic allocas.
564 if (!FuncInfo.StaticAllocaMap.count(AI)) return 0;
567 if (!isLoadTypeLegal(AI->getType(), VT)) return false;
569 DenseMap<const AllocaInst*, int>::iterator SI =
570 FuncInfo.StaticAllocaMap.find(AI);
572 // This will get lowered later into the correct offsets and registers
573 // via rewriteXFrameIndex.
574 if (SI != FuncInfo.StaticAllocaMap.end()) {
575 TargetRegisterClass* RC = TLI.getRegClassFor(VT);
576 unsigned ResultReg = createResultReg(RC);
577 unsigned Opc = isThumb ? ARM::t2ADDri : ARM::ADDri;
578 AddOptionalDefs(BuildMI(*FuncInfo.MBB, *FuncInfo.InsertPt, DL,
579 TII.get(Opc), ResultReg)
580 .addFrameIndex(SI->second)
588 bool ARMFastISel::isTypeLegal(const Type *Ty, MVT &VT) {
589 EVT evt = TLI.getValueType(Ty, true);
591 // Only handle simple types.
592 if (evt == MVT::Other || !evt.isSimple()) return false;
593 VT = evt.getSimpleVT();
595 // Handle all legal types, i.e. a register that will directly hold this
597 return TLI.isTypeLegal(VT);
600 bool ARMFastISel::isLoadTypeLegal(const Type *Ty, MVT &VT) {
601 if (isTypeLegal(Ty, VT)) return true;
603 // If this is a type than can be sign or zero-extended to a basic operation
604 // go ahead and accept it now.
605 if (VT == MVT::i8 || VT == MVT::i16)
611 // Computes the address to get to an object.
612 bool ARMFastISel::ARMComputeAddress(const Value *Obj, Address &Addr) {
613 // Some boilerplate from the X86 FastISel.
614 const User *U = NULL;
615 unsigned Opcode = Instruction::UserOp1;
616 if (const Instruction *I = dyn_cast<Instruction>(Obj)) {
617 // Don't walk into other basic blocks unless the object is an alloca from
618 // another block, otherwise it may not have a virtual register assigned.
619 if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(Obj)) ||
620 FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
621 Opcode = I->getOpcode();
624 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Obj)) {
625 Opcode = C->getOpcode();
629 if (const PointerType *Ty = dyn_cast<PointerType>(Obj->getType()))
630 if (Ty->getAddressSpace() > 255)
631 // Fast instruction selection doesn't support the special
638 case Instruction::BitCast: {
639 // Look through bitcasts.
640 return ARMComputeAddress(U->getOperand(0), Addr);
642 case Instruction::IntToPtr: {
643 // Look past no-op inttoptrs.
644 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
645 return ARMComputeAddress(U->getOperand(0), Addr);
648 case Instruction::PtrToInt: {
649 // Look past no-op ptrtoints.
650 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
651 return ARMComputeAddress(U->getOperand(0), Addr);
654 case Instruction::GetElementPtr: {
655 Address SavedAddr = Addr;
656 int TmpOffset = Addr.Offset;
658 // Iterate through the GEP folding the constants into offsets where
660 gep_type_iterator GTI = gep_type_begin(U);
661 for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end();
662 i != e; ++i, ++GTI) {
663 const Value *Op = *i;
664 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
665 const StructLayout *SL = TD.getStructLayout(STy);
666 unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
667 TmpOffset += SL->getElementOffset(Idx);
669 uint64_t S = TD.getTypeAllocSize(GTI.getIndexedType());
670 SmallVector<const Value *, 4> Worklist;
671 Worklist.push_back(Op);
673 Op = Worklist.pop_back_val();
674 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
675 // Constant-offset addressing.
676 TmpOffset += CI->getSExtValue() * S;
677 } else if (isa<AddOperator>(Op) &&
678 isa<ConstantInt>(cast<AddOperator>(Op)->getOperand(1))) {
679 // An add with a constant operand. Fold the constant.
681 cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
682 TmpOffset += CI->getSExtValue() * S;
683 // Add the other operand back to the work list.
684 Worklist.push_back(cast<AddOperator>(Op)->getOperand(0));
686 goto unsupported_gep;
687 } while (!Worklist.empty());
691 // Try to grab the base operand now.
692 Addr.Offset = TmpOffset;
693 if (ARMComputeAddress(U->getOperand(0), Addr)) return true;
695 // We failed, restore everything and try the other options.
701 case Instruction::Alloca: {
702 const AllocaInst *AI = cast<AllocaInst>(Obj);
703 DenseMap<const AllocaInst*, int>::iterator SI =
704 FuncInfo.StaticAllocaMap.find(AI);
705 if (SI != FuncInfo.StaticAllocaMap.end()) {
706 Addr.BaseType = Address::FrameIndexBase;
707 Addr.Base.FI = SI->second;
714 // Materialize the global variable's address into a reg which can
715 // then be used later to load the variable.
716 if (const GlobalValue *GV = dyn_cast<GlobalValue>(Obj)) {
717 unsigned Tmp = ARMMaterializeGV(GV, TLI.getValueType(Obj->getType()));
718 if (Tmp == 0) return false;
724 // Try to get this in a register if nothing else has worked.
725 if (Addr.Base.Reg == 0) Addr.Base.Reg = getRegForValue(Obj);
726 return Addr.Base.Reg != 0;
729 void ARMFastISel::ARMSimplifyAddress(Address &Addr, EVT VT) {
731 assert(VT.isSimple() && "Non-simple types are invalid here!");
733 bool needsLowering = false;
734 switch (VT.getSimpleVT().SimpleTy) {
736 assert(false && "Unhandled load/store type!");
741 // Integer loads/stores handle 12-bit offsets.
742 needsLowering = ((Addr.Offset & 0xfff) != Addr.Offset);
746 // Floating point operands handle 8-bit offsets.
747 needsLowering = ((Addr.Offset & 0xff) != Addr.Offset);
751 // If this is a stack pointer and the offset needs to be simplified then
752 // put the alloca address into a register, set the base type back to
753 // register and continue. This should almost never happen.
754 if (needsLowering && Addr.BaseType == Address::FrameIndexBase) {
755 TargetRegisterClass *RC = isThumb ? ARM::tGPRRegisterClass :
756 ARM::GPRRegisterClass;
757 unsigned ResultReg = createResultReg(RC);
758 unsigned Opc = isThumb ? ARM::t2ADDri : ARM::ADDri;
759 AddOptionalDefs(BuildMI(*FuncInfo.MBB, *FuncInfo.InsertPt, DL,
760 TII.get(Opc), ResultReg)
761 .addFrameIndex(Addr.Base.FI)
763 Addr.Base.Reg = ResultReg;
764 Addr.BaseType = Address::RegBase;
767 // Since the offset is too large for the load/store instruction
768 // get the reg+offset into a register.
770 ARMCC::CondCodes Pred = ARMCC::AL;
771 unsigned PredReg = 0;
773 TargetRegisterClass *RC = isThumb ? ARM::tGPRRegisterClass :
774 ARM::GPRRegisterClass;
775 unsigned BaseReg = createResultReg(RC);
778 emitARMRegPlusImmediate(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
779 BaseReg, Addr.Base.Reg, Addr.Offset,
781 static_cast<const ARMBaseInstrInfo&>(TII));
783 assert(AFI->isThumb2Function());
784 emitT2RegPlusImmediate(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
785 BaseReg, Addr.Base.Reg, Addr.Offset, Pred, PredReg,
786 static_cast<const ARMBaseInstrInfo&>(TII));
789 Addr.Base.Reg = BaseReg;
793 void ARMFastISel::AddLoadStoreOperands(EVT VT, Address &Addr,
794 const MachineInstrBuilder &MIB) {
795 // addrmode5 output depends on the selection dag addressing dividing the
796 // offset by 4 that it then later multiplies. Do this here as well.
797 if (VT.getSimpleVT().SimpleTy == MVT::f32 ||
798 VT.getSimpleVT().SimpleTy == MVT::f64)
801 // Frame base works a bit differently. Handle it separately.
802 if (Addr.BaseType == Address::FrameIndexBase) {
803 int FI = Addr.Base.FI;
804 int Offset = Addr.Offset;
805 MachineMemOperand *MMO =
806 FuncInfo.MF->getMachineMemOperand(
807 MachinePointerInfo::getFixedStack(FI, Offset),
808 MachineMemOperand::MOLoad,
809 MFI.getObjectSize(FI),
810 MFI.getObjectAlignment(FI));
811 // Now add the rest of the operands.
812 MIB.addFrameIndex(FI);
814 // ARM halfword load/stores need an additional operand.
815 if (!isThumb && VT.getSimpleVT().SimpleTy == MVT::i16) MIB.addReg(0);
817 MIB.addImm(Addr.Offset);
818 MIB.addMemOperand(MMO);
820 // Now add the rest of the operands.
821 MIB.addReg(Addr.Base.Reg);
823 // ARM halfword load/stores need an additional operand.
824 if (!isThumb && VT.getSimpleVT().SimpleTy == MVT::i16) MIB.addReg(0);
826 MIB.addImm(Addr.Offset);
828 AddOptionalDefs(MIB);
831 bool ARMFastISel::ARMEmitLoad(EVT VT, unsigned &ResultReg, Address &Addr) {
833 assert(VT.isSimple() && "Non-simple types are invalid here!");
835 TargetRegisterClass *RC;
836 switch (VT.getSimpleVT().SimpleTy) {
837 // This is mostly going to be Neon/vector support.
838 default: return false;
840 Opc = isThumb ? ARM::t2LDRHi12 : ARM::LDRH;
841 RC = ARM::GPRRegisterClass;
844 Opc = isThumb ? ARM::t2LDRBi12 : ARM::LDRBi12;
845 RC = ARM::GPRRegisterClass;
848 Opc = isThumb ? ARM::t2LDRi12 : ARM::LDRi12;
849 RC = ARM::GPRRegisterClass;
853 RC = TLI.getRegClassFor(VT);
857 RC = TLI.getRegClassFor(VT);
860 // Simplify this down to something we can handle.
861 ARMSimplifyAddress(Addr, VT);
863 // Create the base instruction, then add the operands.
864 ResultReg = createResultReg(RC);
865 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
866 TII.get(Opc), ResultReg);
867 AddLoadStoreOperands(VT, Addr, MIB);
871 bool ARMFastISel::SelectLoad(const Instruction *I) {
872 // Verify we have a legal type before going any further.
874 if (!isLoadTypeLegal(I->getType(), VT))
877 // See if we can handle this address.
879 if (!ARMComputeAddress(I->getOperand(0), Addr)) return false;
882 if (!ARMEmitLoad(VT, ResultReg, Addr)) return false;
883 UpdateValueMap(I, ResultReg);
887 bool ARMFastISel::ARMEmitStore(EVT VT, unsigned SrcReg, Address &Addr) {
889 switch (VT.getSimpleVT().SimpleTy) {
890 // This is mostly going to be Neon/vector support.
891 default: return false;
893 unsigned Res = createResultReg(isThumb ? ARM::tGPRRegisterClass :
894 ARM::GPRRegisterClass);
895 unsigned Opc = isThumb ? ARM::t2ANDri : ARM::ANDri;
896 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
898 .addReg(SrcReg).addImm(1));
900 } // Fallthrough here.
902 StrOpc = isThumb ? ARM::t2STRBi12 : ARM::STRBi12;
905 StrOpc = isThumb ? ARM::t2STRHi12 : ARM::STRH;
908 StrOpc = isThumb ? ARM::t2STRi12 : ARM::STRi12;
911 if (!Subtarget->hasVFP2()) return false;
915 if (!Subtarget->hasVFP2()) return false;
919 // Simplify this down to something we can handle.
920 ARMSimplifyAddress(Addr, VT);
922 // Create the base instruction, then add the operands.
923 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
925 .addReg(SrcReg, getKillRegState(true));
926 AddLoadStoreOperands(VT, Addr, MIB);
930 bool ARMFastISel::SelectStore(const Instruction *I) {
931 Value *Op0 = I->getOperand(0);
934 // Verify we have a legal type before going any further.
936 if (!isLoadTypeLegal(I->getOperand(0)->getType(), VT))
939 // Get the value to be stored into a register.
940 SrcReg = getRegForValue(Op0);
941 if (SrcReg == 0) return false;
943 // See if we can handle this address.
945 if (!ARMComputeAddress(I->getOperand(1), Addr))
948 if (!ARMEmitStore(VT, SrcReg, Addr)) return false;
952 static ARMCC::CondCodes getComparePred(CmpInst::Predicate Pred) {
954 // Needs two compares...
955 case CmpInst::FCMP_ONE:
956 case CmpInst::FCMP_UEQ:
958 // AL is our "false" for now. The other two need more compares.
960 case CmpInst::ICMP_EQ:
961 case CmpInst::FCMP_OEQ:
963 case CmpInst::ICMP_SGT:
964 case CmpInst::FCMP_OGT:
966 case CmpInst::ICMP_SGE:
967 case CmpInst::FCMP_OGE:
969 case CmpInst::ICMP_UGT:
970 case CmpInst::FCMP_UGT:
972 case CmpInst::FCMP_OLT:
974 case CmpInst::ICMP_ULE:
975 case CmpInst::FCMP_OLE:
977 case CmpInst::FCMP_ORD:
979 case CmpInst::FCMP_UNO:
981 case CmpInst::FCMP_UGE:
983 case CmpInst::ICMP_SLT:
984 case CmpInst::FCMP_ULT:
986 case CmpInst::ICMP_SLE:
987 case CmpInst::FCMP_ULE:
989 case CmpInst::FCMP_UNE:
990 case CmpInst::ICMP_NE:
992 case CmpInst::ICMP_UGE:
994 case CmpInst::ICMP_ULT:
999 bool ARMFastISel::SelectBranch(const Instruction *I) {
1000 const BranchInst *BI = cast<BranchInst>(I);
1001 MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
1002 MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
1004 // Simple branch support.
1006 // If we can, avoid recomputing the compare - redoing it could lead to wonky
1008 // TODO: Factor this out.
1009 if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
1010 if (CI->hasOneUse() && (CI->getParent() == I->getParent())) {
1012 const Type *Ty = CI->getOperand(0)->getType();
1013 if (!isTypeLegal(Ty, VT))
1016 bool isFloat = (Ty->isDoubleTy() || Ty->isFloatTy());
1017 if (isFloat && !Subtarget->hasVFP2())
1021 switch (VT.SimpleTy) {
1022 default: return false;
1023 // TODO: Verify compares.
1025 CmpOpc = ARM::VCMPES;
1028 CmpOpc = ARM::VCMPED;
1031 CmpOpc = isThumb ? ARM::t2CMPrr : ARM::CMPrr;
1035 // Get the compare predicate.
1036 ARMCC::CondCodes ARMPred = getComparePred(CI->getPredicate());
1038 // We may not handle every CC for now.
1039 if (ARMPred == ARMCC::AL) return false;
1041 unsigned Arg1 = getRegForValue(CI->getOperand(0));
1042 if (Arg1 == 0) return false;
1044 unsigned Arg2 = getRegForValue(CI->getOperand(1));
1045 if (Arg2 == 0) return false;
1047 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1049 .addReg(Arg1).addReg(Arg2));
1051 // For floating point we need to move the result to a comparison register
1052 // that we can then use for branches.
1054 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1055 TII.get(ARM::FMSTAT)));
1057 unsigned BrOpc = isThumb ? ARM::t2Bcc : ARM::Bcc;
1058 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
1059 .addMBB(TBB).addImm(ARMPred).addReg(ARM::CPSR);
1060 FastEmitBranch(FBB, DL);
1061 FuncInfo.MBB->addSuccessor(TBB);
1066 unsigned CmpReg = getRegForValue(BI->getCondition());
1067 if (CmpReg == 0) return false;
1069 // Re-set the flags just in case.
1070 unsigned CmpOpc = isThumb ? ARM::t2CMPri : ARM::CMPri;
1071 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
1072 .addReg(CmpReg).addImm(0));
1074 unsigned BrOpc = isThumb ? ARM::t2Bcc : ARM::Bcc;
1075 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
1076 .addMBB(TBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
1077 FastEmitBranch(FBB, DL);
1078 FuncInfo.MBB->addSuccessor(TBB);
1082 bool ARMFastISel::SelectCmp(const Instruction *I) {
1083 const CmpInst *CI = cast<CmpInst>(I);
1086 const Type *Ty = CI->getOperand(0)->getType();
1087 if (!isTypeLegal(Ty, VT))
1090 bool isFloat = (Ty->isDoubleTy() || Ty->isFloatTy());
1091 if (isFloat && !Subtarget->hasVFP2())
1096 switch (VT.SimpleTy) {
1097 default: return false;
1098 // TODO: Verify compares.
1100 CmpOpc = ARM::VCMPES;
1101 CondReg = ARM::FPSCR;
1104 CmpOpc = ARM::VCMPED;
1105 CondReg = ARM::FPSCR;
1108 CmpOpc = isThumb ? ARM::t2CMPrr : ARM::CMPrr;
1109 CondReg = ARM::CPSR;
1113 // Get the compare predicate.
1114 ARMCC::CondCodes ARMPred = getComparePred(CI->getPredicate());
1116 // We may not handle every CC for now.
1117 if (ARMPred == ARMCC::AL) return false;
1119 unsigned Arg1 = getRegForValue(CI->getOperand(0));
1120 if (Arg1 == 0) return false;
1122 unsigned Arg2 = getRegForValue(CI->getOperand(1));
1123 if (Arg2 == 0) return false;
1125 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
1126 .addReg(Arg1).addReg(Arg2));
1128 // For floating point we need to move the result to a comparison register
1129 // that we can then use for branches.
1131 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1132 TII.get(ARM::FMSTAT)));
1134 // Now set a register based on the comparison. Explicitly set the predicates
1136 unsigned MovCCOpc = isThumb ? ARM::t2MOVCCi : ARM::MOVCCi;
1137 TargetRegisterClass *RC = isThumb ? ARM::rGPRRegisterClass
1138 : ARM::GPRRegisterClass;
1139 unsigned DestReg = createResultReg(RC);
1141 = ConstantInt::get(Type::getInt32Ty(*Context), 0);
1142 unsigned ZeroReg = TargetMaterializeConstant(Zero);
1143 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), DestReg)
1144 .addReg(ZeroReg).addImm(1)
1145 .addImm(ARMPred).addReg(CondReg);
1147 UpdateValueMap(I, DestReg);
1151 bool ARMFastISel::SelectFPExt(const Instruction *I) {
1152 // Make sure we have VFP and that we're extending float to double.
1153 if (!Subtarget->hasVFP2()) return false;
1155 Value *V = I->getOperand(0);
1156 if (!I->getType()->isDoubleTy() ||
1157 !V->getType()->isFloatTy()) return false;
1159 unsigned Op = getRegForValue(V);
1160 if (Op == 0) return false;
1162 unsigned Result = createResultReg(ARM::DPRRegisterClass);
1163 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1164 TII.get(ARM::VCVTDS), Result)
1166 UpdateValueMap(I, Result);
1170 bool ARMFastISel::SelectFPTrunc(const Instruction *I) {
1171 // Make sure we have VFP and that we're truncating double to float.
1172 if (!Subtarget->hasVFP2()) return false;
1174 Value *V = I->getOperand(0);
1175 if (!(I->getType()->isFloatTy() &&
1176 V->getType()->isDoubleTy())) return false;
1178 unsigned Op = getRegForValue(V);
1179 if (Op == 0) return false;
1181 unsigned Result = createResultReg(ARM::SPRRegisterClass);
1182 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1183 TII.get(ARM::VCVTSD), Result)
1185 UpdateValueMap(I, Result);
1189 bool ARMFastISel::SelectSIToFP(const Instruction *I) {
1190 // Make sure we have VFP.
1191 if (!Subtarget->hasVFP2()) return false;
1194 const Type *Ty = I->getType();
1195 if (!isTypeLegal(Ty, DstVT))
1198 unsigned Op = getRegForValue(I->getOperand(0));
1199 if (Op == 0) return false;
1201 // The conversion routine works on fp-reg to fp-reg and the operand above
1202 // was an integer, move it to the fp registers if possible.
1203 unsigned FP = ARMMoveToFPReg(MVT::f32, Op);
1204 if (FP == 0) return false;
1207 if (Ty->isFloatTy()) Opc = ARM::VSITOS;
1208 else if (Ty->isDoubleTy()) Opc = ARM::VSITOD;
1211 unsigned ResultReg = createResultReg(TLI.getRegClassFor(DstVT));
1212 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
1215 UpdateValueMap(I, ResultReg);
1219 bool ARMFastISel::SelectFPToSI(const Instruction *I) {
1220 // Make sure we have VFP.
1221 if (!Subtarget->hasVFP2()) return false;
1224 const Type *RetTy = I->getType();
1225 if (!isTypeLegal(RetTy, DstVT))
1228 unsigned Op = getRegForValue(I->getOperand(0));
1229 if (Op == 0) return false;
1232 const Type *OpTy = I->getOperand(0)->getType();
1233 if (OpTy->isFloatTy()) Opc = ARM::VTOSIZS;
1234 else if (OpTy->isDoubleTy()) Opc = ARM::VTOSIZD;
1237 // f64->s32 or f32->s32 both need an intermediate f32 reg.
1238 unsigned ResultReg = createResultReg(TLI.getRegClassFor(MVT::f32));
1239 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
1243 // This result needs to be in an integer register, but the conversion only
1244 // takes place in fp-regs.
1245 unsigned IntReg = ARMMoveToIntReg(DstVT, ResultReg);
1246 if (IntReg == 0) return false;
1248 UpdateValueMap(I, IntReg);
1252 bool ARMFastISel::SelectSelect(const Instruction *I) {
1254 if (!isTypeLegal(I->getType(), VT))
1257 // Things need to be register sized for register moves.
1258 if (VT != MVT::i32) return false;
1259 const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
1261 unsigned CondReg = getRegForValue(I->getOperand(0));
1262 if (CondReg == 0) return false;
1263 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1264 if (Op1Reg == 0) return false;
1265 unsigned Op2Reg = getRegForValue(I->getOperand(2));
1266 if (Op2Reg == 0) return false;
1268 unsigned CmpOpc = isThumb ? ARM::t2TSTri : ARM::TSTri;
1269 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
1270 .addReg(CondReg).addImm(1));
1271 unsigned ResultReg = createResultReg(RC);
1272 unsigned MovCCOpc = isThumb ? ARM::t2MOVCCr : ARM::MOVCCr;
1273 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), ResultReg)
1274 .addReg(Op1Reg).addReg(Op2Reg)
1275 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
1276 UpdateValueMap(I, ResultReg);
1280 bool ARMFastISel::SelectSDiv(const Instruction *I) {
1282 const Type *Ty = I->getType();
1283 if (!isTypeLegal(Ty, VT))
1286 // If we have integer div support we should have selected this automagically.
1287 // In case we have a real miss go ahead and return false and we'll pick
1289 if (Subtarget->hasDivide()) return false;
1291 // Otherwise emit a libcall.
1292 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
1294 LC = RTLIB::SDIV_I8;
1295 else if (VT == MVT::i16)
1296 LC = RTLIB::SDIV_I16;
1297 else if (VT == MVT::i32)
1298 LC = RTLIB::SDIV_I32;
1299 else if (VT == MVT::i64)
1300 LC = RTLIB::SDIV_I64;
1301 else if (VT == MVT::i128)
1302 LC = RTLIB::SDIV_I128;
1303 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SDIV!");
1305 return ARMEmitLibcall(I, LC);
1308 bool ARMFastISel::SelectSRem(const Instruction *I) {
1310 const Type *Ty = I->getType();
1311 if (!isTypeLegal(Ty, VT))
1314 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
1316 LC = RTLIB::SREM_I8;
1317 else if (VT == MVT::i16)
1318 LC = RTLIB::SREM_I16;
1319 else if (VT == MVT::i32)
1320 LC = RTLIB::SREM_I32;
1321 else if (VT == MVT::i64)
1322 LC = RTLIB::SREM_I64;
1323 else if (VT == MVT::i128)
1324 LC = RTLIB::SREM_I128;
1325 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SREM!");
1327 return ARMEmitLibcall(I, LC);
1330 bool ARMFastISel::SelectBinaryOp(const Instruction *I, unsigned ISDOpcode) {
1331 EVT VT = TLI.getValueType(I->getType(), true);
1333 // We can get here in the case when we want to use NEON for our fp
1334 // operations, but can't figure out how to. Just use the vfp instructions
1336 // FIXME: It'd be nice to use NEON instructions.
1337 const Type *Ty = I->getType();
1338 bool isFloat = (Ty->isDoubleTy() || Ty->isFloatTy());
1339 if (isFloat && !Subtarget->hasVFP2())
1342 unsigned Op1 = getRegForValue(I->getOperand(0));
1343 if (Op1 == 0) return false;
1345 unsigned Op2 = getRegForValue(I->getOperand(1));
1346 if (Op2 == 0) return false;
1349 bool is64bit = VT == MVT::f64 || VT == MVT::i64;
1350 switch (ISDOpcode) {
1351 default: return false;
1353 Opc = is64bit ? ARM::VADDD : ARM::VADDS;
1356 Opc = is64bit ? ARM::VSUBD : ARM::VSUBS;
1359 Opc = is64bit ? ARM::VMULD : ARM::VMULS;
1362 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
1363 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1364 TII.get(Opc), ResultReg)
1365 .addReg(Op1).addReg(Op2));
1366 UpdateValueMap(I, ResultReg);
1370 // Call Handling Code
1372 bool ARMFastISel::FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src,
1373 EVT SrcVT, unsigned &ResultReg) {
1374 unsigned RR = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), Opc,
1375 Src, /*TODO: Kill=*/false);
1384 // This is largely taken directly from CCAssignFnForNode - we don't support
1385 // varargs in FastISel so that part has been removed.
1386 // TODO: We may not support all of this.
1387 CCAssignFn *ARMFastISel::CCAssignFnForCall(CallingConv::ID CC, bool Return) {
1390 llvm_unreachable("Unsupported calling convention");
1391 case CallingConv::Fast:
1392 // Ignore fastcc. Silence compiler warnings.
1393 (void)RetFastCC_ARM_APCS;
1394 (void)FastCC_ARM_APCS;
1396 case CallingConv::C:
1397 // Use target triple & subtarget features to do actual dispatch.
1398 if (Subtarget->isAAPCS_ABI()) {
1399 if (Subtarget->hasVFP2() &&
1400 FloatABIType == FloatABI::Hard)
1401 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
1403 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
1405 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
1406 case CallingConv::ARM_AAPCS_VFP:
1407 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
1408 case CallingConv::ARM_AAPCS:
1409 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
1410 case CallingConv::ARM_APCS:
1411 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
1415 bool ARMFastISel::ProcessCallArgs(SmallVectorImpl<Value*> &Args,
1416 SmallVectorImpl<unsigned> &ArgRegs,
1417 SmallVectorImpl<MVT> &ArgVTs,
1418 SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
1419 SmallVectorImpl<unsigned> &RegArgs,
1421 unsigned &NumBytes) {
1422 SmallVector<CCValAssign, 16> ArgLocs;
1423 CCState CCInfo(CC, false, TM, ArgLocs, *Context);
1424 CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CCAssignFnForCall(CC, false));
1426 // Get a count of how many bytes are to be pushed on the stack.
1427 NumBytes = CCInfo.getNextStackOffset();
1429 // Issue CALLSEQ_START
1430 unsigned AdjStackDown = TM.getRegisterInfo()->getCallFrameSetupOpcode();
1431 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1432 TII.get(AdjStackDown))
1435 // Process the args.
1436 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1437 CCValAssign &VA = ArgLocs[i];
1438 unsigned Arg = ArgRegs[VA.getValNo()];
1439 MVT ArgVT = ArgVTs[VA.getValNo()];
1441 // We don't handle NEON/vector parameters yet.
1442 if (ArgVT.isVector() || ArgVT.getSizeInBits() > 64)
1445 // Handle arg promotion, etc.
1446 switch (VA.getLocInfo()) {
1447 case CCValAssign::Full: break;
1448 case CCValAssign::SExt: {
1449 bool Emitted = FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
1451 assert(Emitted && "Failed to emit a sext!"); (void)Emitted;
1453 ArgVT = VA.getLocVT();
1456 case CCValAssign::ZExt: {
1457 bool Emitted = FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
1459 assert(Emitted && "Failed to emit a zext!"); (void)Emitted;
1461 ArgVT = VA.getLocVT();
1464 case CCValAssign::AExt: {
1465 bool Emitted = FastEmitExtend(ISD::ANY_EXTEND, VA.getLocVT(),
1468 Emitted = FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
1471 Emitted = FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
1474 assert(Emitted && "Failed to emit a aext!"); (void)Emitted;
1475 ArgVT = VA.getLocVT();
1478 case CCValAssign::BCvt: {
1479 unsigned BC = FastEmit_r(ArgVT, VA.getLocVT(), ISD::BITCAST, Arg,
1480 /*TODO: Kill=*/false);
1481 assert(BC != 0 && "Failed to emit a bitcast!");
1483 ArgVT = VA.getLocVT();
1486 default: llvm_unreachable("Unknown arg promotion!");
1489 // Now copy/store arg to correct locations.
1490 if (VA.isRegLoc() && !VA.needsCustom()) {
1491 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1494 RegArgs.push_back(VA.getLocReg());
1495 } else if (VA.needsCustom()) {
1496 // TODO: We need custom lowering for vector (v2f64) args.
1497 if (VA.getLocVT() != MVT::f64) return false;
1499 CCValAssign &NextVA = ArgLocs[++i];
1501 // TODO: Only handle register args for now.
1502 if(!(VA.isRegLoc() && NextVA.isRegLoc())) return false;
1504 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1505 TII.get(ARM::VMOVRRD), VA.getLocReg())
1506 .addReg(NextVA.getLocReg(), RegState::Define)
1508 RegArgs.push_back(VA.getLocReg());
1509 RegArgs.push_back(NextVA.getLocReg());
1511 assert(VA.isMemLoc());
1512 // Need to store on the stack.
1514 Addr.BaseType = Address::RegBase;
1515 Addr.Base.Reg = ARM::SP;
1516 Addr.Offset = VA.getLocMemOffset();
1518 if (!ARMEmitStore(ArgVT, Arg, Addr)) return false;
1524 bool ARMFastISel::FinishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
1525 const Instruction *I, CallingConv::ID CC,
1526 unsigned &NumBytes) {
1527 // Issue CALLSEQ_END
1528 unsigned AdjStackUp = TM.getRegisterInfo()->getCallFrameDestroyOpcode();
1529 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1530 TII.get(AdjStackUp))
1531 .addImm(NumBytes).addImm(0));
1533 // Now the return value.
1534 if (RetVT != MVT::isVoid) {
1535 SmallVector<CCValAssign, 16> RVLocs;
1536 CCState CCInfo(CC, false, TM, RVLocs, *Context);
1537 CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC, true));
1539 // Copy all of the result registers out of their specified physreg.
1540 if (RVLocs.size() == 2 && RetVT == MVT::f64) {
1541 // For this move we copy into two registers and then move into the
1542 // double fp reg we want.
1543 EVT DestVT = RVLocs[0].getValVT();
1544 TargetRegisterClass* DstRC = TLI.getRegClassFor(DestVT);
1545 unsigned ResultReg = createResultReg(DstRC);
1546 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1547 TII.get(ARM::VMOVDRR), ResultReg)
1548 .addReg(RVLocs[0].getLocReg())
1549 .addReg(RVLocs[1].getLocReg()));
1551 UsedRegs.push_back(RVLocs[0].getLocReg());
1552 UsedRegs.push_back(RVLocs[1].getLocReg());
1554 // Finally update the result.
1555 UpdateValueMap(I, ResultReg);
1557 assert(RVLocs.size() == 1 &&"Can't handle non-double multi-reg retvals!");
1558 EVT CopyVT = RVLocs[0].getValVT();
1559 TargetRegisterClass* DstRC = TLI.getRegClassFor(CopyVT);
1561 unsigned ResultReg = createResultReg(DstRC);
1562 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1563 ResultReg).addReg(RVLocs[0].getLocReg());
1564 UsedRegs.push_back(RVLocs[0].getLocReg());
1566 // Finally update the result.
1567 UpdateValueMap(I, ResultReg);
1574 bool ARMFastISel::SelectRet(const Instruction *I) {
1575 const ReturnInst *Ret = cast<ReturnInst>(I);
1576 const Function &F = *I->getParent()->getParent();
1578 if (!FuncInfo.CanLowerReturn)
1584 CallingConv::ID CC = F.getCallingConv();
1585 if (Ret->getNumOperands() > 0) {
1586 SmallVector<ISD::OutputArg, 4> Outs;
1587 GetReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(),
1590 // Analyze operands of the call, assigning locations to each operand.
1591 SmallVector<CCValAssign, 16> ValLocs;
1592 CCState CCInfo(CC, F.isVarArg(), TM, ValLocs, I->getContext());
1593 CCInfo.AnalyzeReturn(Outs, CCAssignFnForCall(CC, true /* is Ret */));
1595 const Value *RV = Ret->getOperand(0);
1596 unsigned Reg = getRegForValue(RV);
1600 // Only handle a single return value for now.
1601 if (ValLocs.size() != 1)
1604 CCValAssign &VA = ValLocs[0];
1606 // Don't bother handling odd stuff for now.
1607 if (VA.getLocInfo() != CCValAssign::Full)
1609 // Only handle register returns for now.
1612 // TODO: For now, don't try to handle cases where getLocInfo()
1613 // says Full but the types don't match.
1614 if (TLI.getValueType(RV->getType()) != VA.getValVT())
1618 unsigned SrcReg = Reg + VA.getValNo();
1619 unsigned DstReg = VA.getLocReg();
1620 const TargetRegisterClass* SrcRC = MRI.getRegClass(SrcReg);
1621 // Avoid a cross-class copy. This is very unlikely.
1622 if (!SrcRC->contains(DstReg))
1624 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1625 DstReg).addReg(SrcReg);
1627 // Mark the register as live out of the function.
1628 MRI.addLiveOut(VA.getLocReg());
1631 unsigned RetOpc = isThumb ? ARM::tBX_RET : ARM::BX_RET;
1632 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1637 unsigned ARMFastISel::ARMSelectCallOp(const GlobalValue *GV) {
1639 // Depend our opcode for thumb on whether or not we're targeting an
1640 // externally callable function. For libcalls we'll just pass a NULL GV
1642 bool isExternal = false;
1643 if (!GV || GV->hasExternalLinkage()) isExternal = true;
1645 // Darwin needs the r9 versions of the opcodes.
1646 bool isDarwin = Subtarget->isTargetDarwin();
1647 if (isThumb && isExternal) {
1648 return isDarwin ? ARM::tBLXi_r9 : ARM::tBLXi;
1649 } else if (isThumb) {
1650 return isDarwin ? ARM::tBLr9 : ARM::tBL;
1652 return isDarwin ? ARM::BLr9 : ARM::BL;
1656 // A quick function that will emit a call for a named libcall in F with the
1657 // vector of passed arguments for the Instruction in I. We can assume that we
1658 // can emit a call for any libcall we can produce. This is an abridged version
1659 // of the full call infrastructure since we won't need to worry about things
1660 // like computed function pointers or strange arguments at call sites.
1661 // TODO: Try to unify this and the normal call bits for ARM, then try to unify
1663 bool ARMFastISel::ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call) {
1664 CallingConv::ID CC = TLI.getLibcallCallingConv(Call);
1666 // Handle *simple* calls for now.
1667 const Type *RetTy = I->getType();
1669 if (RetTy->isVoidTy())
1670 RetVT = MVT::isVoid;
1671 else if (!isTypeLegal(RetTy, RetVT))
1674 // For now we're using BLX etc on the assumption that we have v5t ops.
1675 if (!Subtarget->hasV5TOps()) return false;
1677 // TODO: For now if we have long calls specified we don't handle the call.
1678 if (EnableARMLongCalls) return false;
1680 // Set up the argument vectors.
1681 SmallVector<Value*, 8> Args;
1682 SmallVector<unsigned, 8> ArgRegs;
1683 SmallVector<MVT, 8> ArgVTs;
1684 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
1685 Args.reserve(I->getNumOperands());
1686 ArgRegs.reserve(I->getNumOperands());
1687 ArgVTs.reserve(I->getNumOperands());
1688 ArgFlags.reserve(I->getNumOperands());
1689 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
1690 Value *Op = I->getOperand(i);
1691 unsigned Arg = getRegForValue(Op);
1692 if (Arg == 0) return false;
1694 const Type *ArgTy = Op->getType();
1696 if (!isTypeLegal(ArgTy, ArgVT)) return false;
1698 ISD::ArgFlagsTy Flags;
1699 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
1700 Flags.setOrigAlign(OriginalAlignment);
1703 ArgRegs.push_back(Arg);
1704 ArgVTs.push_back(ArgVT);
1705 ArgFlags.push_back(Flags);
1708 // Handle the arguments now that we've gotten them.
1709 SmallVector<unsigned, 4> RegArgs;
1711 if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
1714 // Issue the call, BLXr9 for darwin, BLX otherwise. This uses V5 ops.
1715 // TODO: Turn this into the table of arm call ops.
1716 MachineInstrBuilder MIB;
1717 unsigned CallOpc = ARMSelectCallOp(NULL);
1719 // Explicitly adding the predicate here.
1720 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1722 .addExternalSymbol(TLI.getLibcallName(Call));
1724 // Explicitly adding the predicate here.
1725 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1727 .addExternalSymbol(TLI.getLibcallName(Call)));
1729 // Add implicit physical register uses to the call.
1730 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
1731 MIB.addReg(RegArgs[i]);
1733 // Finish off the call including any return values.
1734 SmallVector<unsigned, 4> UsedRegs;
1735 if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false;
1737 // Set all unused physreg defs as dead.
1738 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
1743 bool ARMFastISel::SelectCall(const Instruction *I) {
1744 const CallInst *CI = cast<CallInst>(I);
1745 const Value *Callee = CI->getCalledValue();
1747 // Can't handle inline asm or worry about intrinsics yet.
1748 if (isa<InlineAsm>(Callee) || isa<IntrinsicInst>(CI)) return false;
1750 // Only handle global variable Callees that are direct calls.
1751 const GlobalValue *GV = dyn_cast<GlobalValue>(Callee);
1752 if (!GV || Subtarget->GVIsIndirectSymbol(GV, TM.getRelocationModel()))
1755 // Check the calling convention.
1756 ImmutableCallSite CS(CI);
1757 CallingConv::ID CC = CS.getCallingConv();
1759 // TODO: Avoid some calling conventions?
1761 // Let SDISel handle vararg functions.
1762 const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
1763 const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
1764 if (FTy->isVarArg())
1767 // Handle *simple* calls for now.
1768 const Type *RetTy = I->getType();
1770 if (RetTy->isVoidTy())
1771 RetVT = MVT::isVoid;
1772 else if (!isTypeLegal(RetTy, RetVT))
1775 // For now we're using BLX etc on the assumption that we have v5t ops.
1777 if (!Subtarget->hasV5TOps()) return false;
1779 // TODO: For now if we have long calls specified we don't handle the call.
1780 if (EnableARMLongCalls) return false;
1782 // Set up the argument vectors.
1783 SmallVector<Value*, 8> Args;
1784 SmallVector<unsigned, 8> ArgRegs;
1785 SmallVector<MVT, 8> ArgVTs;
1786 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
1787 Args.reserve(CS.arg_size());
1788 ArgRegs.reserve(CS.arg_size());
1789 ArgVTs.reserve(CS.arg_size());
1790 ArgFlags.reserve(CS.arg_size());
1791 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
1793 unsigned Arg = getRegForValue(*i);
1797 ISD::ArgFlagsTy Flags;
1798 unsigned AttrInd = i - CS.arg_begin() + 1;
1799 if (CS.paramHasAttr(AttrInd, Attribute::SExt))
1801 if (CS.paramHasAttr(AttrInd, Attribute::ZExt))
1804 // FIXME: Only handle *easy* calls for now.
1805 if (CS.paramHasAttr(AttrInd, Attribute::InReg) ||
1806 CS.paramHasAttr(AttrInd, Attribute::StructRet) ||
1807 CS.paramHasAttr(AttrInd, Attribute::Nest) ||
1808 CS.paramHasAttr(AttrInd, Attribute::ByVal))
1811 const Type *ArgTy = (*i)->getType();
1813 if (!isTypeLegal(ArgTy, ArgVT))
1815 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
1816 Flags.setOrigAlign(OriginalAlignment);
1819 ArgRegs.push_back(Arg);
1820 ArgVTs.push_back(ArgVT);
1821 ArgFlags.push_back(Flags);
1824 // Handle the arguments now that we've gotten them.
1825 SmallVector<unsigned, 4> RegArgs;
1827 if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
1830 // Issue the call, BLXr9 for darwin, BLX otherwise. This uses V5 ops.
1831 // TODO: Turn this into the table of arm call ops.
1832 MachineInstrBuilder MIB;
1833 unsigned CallOpc = ARMSelectCallOp(GV);
1834 // Explicitly adding the predicate here.
1836 // Explicitly adding the predicate here.
1837 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1839 .addGlobalAddress(GV, 0, 0);
1841 // Explicitly adding the predicate here.
1842 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1844 .addGlobalAddress(GV, 0, 0));
1846 // Add implicit physical register uses to the call.
1847 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
1848 MIB.addReg(RegArgs[i]);
1850 // Finish off the call including any return values.
1851 SmallVector<unsigned, 4> UsedRegs;
1852 if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false;
1854 // Set all unused physreg defs as dead.
1855 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
1861 // TODO: SoftFP support.
1862 bool ARMFastISel::TargetSelectInstruction(const Instruction *I) {
1864 switch (I->getOpcode()) {
1865 case Instruction::Load:
1866 return SelectLoad(I);
1867 case Instruction::Store:
1868 return SelectStore(I);
1869 case Instruction::Br:
1870 return SelectBranch(I);
1871 case Instruction::ICmp:
1872 case Instruction::FCmp:
1873 return SelectCmp(I);
1874 case Instruction::FPExt:
1875 return SelectFPExt(I);
1876 case Instruction::FPTrunc:
1877 return SelectFPTrunc(I);
1878 case Instruction::SIToFP:
1879 return SelectSIToFP(I);
1880 case Instruction::FPToSI:
1881 return SelectFPToSI(I);
1882 case Instruction::FAdd:
1883 return SelectBinaryOp(I, ISD::FADD);
1884 case Instruction::FSub:
1885 return SelectBinaryOp(I, ISD::FSUB);
1886 case Instruction::FMul:
1887 return SelectBinaryOp(I, ISD::FMUL);
1888 case Instruction::SDiv:
1889 return SelectSDiv(I);
1890 case Instruction::SRem:
1891 return SelectSRem(I);
1892 case Instruction::Call:
1893 return SelectCall(I);
1894 case Instruction::Select:
1895 return SelectSelect(I);
1896 case Instruction::Ret:
1897 return SelectRet(I);
1904 llvm::FastISel *ARM::createFastISel(FunctionLoweringInfo &funcInfo) {
1905 // Completely untested on non-darwin.
1906 const TargetMachine &TM = funcInfo.MF->getTarget();
1908 // Darwin and thumb1 only for now.
1909 const ARMSubtarget *Subtarget = &TM.getSubtarget<ARMSubtarget>();
1910 if (Subtarget->isTargetDarwin() && !Subtarget->isThumb1Only() &&
1911 !DisableARMFastISel)
1912 return new ARMFastISel(funcInfo);