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 "MCTargetDesc/ARMAddressingModes.h"
24 #include "llvm/CallingConv.h"
25 #include "llvm/DerivedTypes.h"
26 #include "llvm/GlobalVariable.h"
27 #include "llvm/Instructions.h"
28 #include "llvm/IntrinsicInst.h"
29 #include "llvm/Module.h"
30 #include "llvm/Operator.h"
31 #include "llvm/CodeGen/Analysis.h"
32 #include "llvm/CodeGen/FastISel.h"
33 #include "llvm/CodeGen/FunctionLoweringInfo.h"
34 #include "llvm/CodeGen/MachineInstrBuilder.h"
35 #include "llvm/CodeGen/MachineModuleInfo.h"
36 #include "llvm/CodeGen/MachineConstantPool.h"
37 #include "llvm/CodeGen/MachineFrameInfo.h"
38 #include "llvm/CodeGen/MachineMemOperand.h"
39 #include "llvm/CodeGen/MachineRegisterInfo.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 {
74 // Innocuous defaults for our address.
76 : BaseType(RegBase), Offset(0) {
81 class ARMFastISel : public FastISel {
83 /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
84 /// make the right decision when generating code for different targets.
85 const ARMSubtarget *Subtarget;
86 const TargetMachine &TM;
87 const TargetInstrInfo &TII;
88 const TargetLowering &TLI;
91 // Convenience variables to avoid some queries.
96 explicit ARMFastISel(FunctionLoweringInfo &funcInfo)
98 TM(funcInfo.MF->getTarget()),
99 TII(*TM.getInstrInfo()),
100 TLI(*TM.getTargetLowering()) {
101 Subtarget = &TM.getSubtarget<ARMSubtarget>();
102 AFI = funcInfo.MF->getInfo<ARMFunctionInfo>();
103 isThumb2 = AFI->isThumbFunction();
104 Context = &funcInfo.Fn->getContext();
107 // Code from FastISel.cpp.
108 virtual unsigned FastEmitInst_(unsigned MachineInstOpcode,
109 const TargetRegisterClass *RC);
110 virtual unsigned FastEmitInst_r(unsigned MachineInstOpcode,
111 const TargetRegisterClass *RC,
112 unsigned Op0, bool Op0IsKill);
113 virtual unsigned FastEmitInst_rr(unsigned MachineInstOpcode,
114 const TargetRegisterClass *RC,
115 unsigned Op0, bool Op0IsKill,
116 unsigned Op1, bool Op1IsKill);
117 virtual unsigned FastEmitInst_rrr(unsigned MachineInstOpcode,
118 const TargetRegisterClass *RC,
119 unsigned Op0, bool Op0IsKill,
120 unsigned Op1, bool Op1IsKill,
121 unsigned Op2, bool Op2IsKill);
122 virtual unsigned FastEmitInst_ri(unsigned MachineInstOpcode,
123 const TargetRegisterClass *RC,
124 unsigned Op0, bool Op0IsKill,
126 virtual unsigned FastEmitInst_rf(unsigned MachineInstOpcode,
127 const TargetRegisterClass *RC,
128 unsigned Op0, bool Op0IsKill,
129 const ConstantFP *FPImm);
130 virtual unsigned FastEmitInst_rri(unsigned MachineInstOpcode,
131 const TargetRegisterClass *RC,
132 unsigned Op0, bool Op0IsKill,
133 unsigned Op1, bool Op1IsKill,
135 virtual unsigned FastEmitInst_i(unsigned MachineInstOpcode,
136 const TargetRegisterClass *RC,
138 virtual unsigned FastEmitInst_ii(unsigned MachineInstOpcode,
139 const TargetRegisterClass *RC,
140 uint64_t Imm1, uint64_t Imm2);
142 virtual unsigned FastEmitInst_extractsubreg(MVT RetVT,
143 unsigned Op0, bool Op0IsKill,
146 // Backend specific FastISel code.
147 virtual bool TargetSelectInstruction(const Instruction *I);
148 virtual unsigned TargetMaterializeConstant(const Constant *C);
149 virtual unsigned TargetMaterializeAlloca(const AllocaInst *AI);
150 virtual bool TryToFoldLoad(MachineInstr *MI, unsigned OpNo,
153 #include "ARMGenFastISel.inc"
155 // Instruction selection routines.
157 bool SelectLoad(const Instruction *I);
158 bool SelectStore(const Instruction *I);
159 bool SelectBranch(const Instruction *I);
160 bool SelectCmp(const Instruction *I);
161 bool SelectFPExt(const Instruction *I);
162 bool SelectFPTrunc(const Instruction *I);
163 bool SelectBinaryOp(const Instruction *I, unsigned ISDOpcode);
164 bool SelectSIToFP(const Instruction *I);
165 bool SelectFPToSI(const Instruction *I);
166 bool SelectSDiv(const Instruction *I);
167 bool SelectSRem(const Instruction *I);
168 bool SelectCall(const Instruction *I, const char *IntrMemName);
169 bool SelectIntrinsicCall(const IntrinsicInst &I);
170 bool SelectSelect(const Instruction *I);
171 bool SelectRet(const Instruction *I);
172 bool SelectTrunc(const Instruction *I);
173 bool SelectIntExt(const Instruction *I);
177 bool isTypeLegal(Type *Ty, MVT &VT);
178 bool isLoadTypeLegal(Type *Ty, MVT &VT);
179 bool ARMEmitCmp(const Value *Src1Value, const Value *Src2Value,
181 bool ARMEmitLoad(EVT VT, unsigned &ResultReg, Address &Addr, unsigned Alignment = 0,
182 bool isZExt = true, bool allocReg = true);
184 bool ARMEmitStore(EVT VT, unsigned SrcReg, Address &Addr,
185 unsigned Alignment = 0);
186 bool ARMComputeAddress(const Value *Obj, Address &Addr);
187 void ARMSimplifyAddress(Address &Addr, EVT VT, bool useAM3);
188 bool ARMIsMemCpySmall(uint64_t Len);
189 bool ARMTryEmitSmallMemCpy(Address Dest, Address Src, uint64_t Len);
190 unsigned ARMEmitIntExt(EVT SrcVT, unsigned SrcReg, EVT DestVT, bool isZExt);
191 unsigned ARMMaterializeFP(const ConstantFP *CFP, EVT VT);
192 unsigned ARMMaterializeInt(const Constant *C, EVT VT);
193 unsigned ARMMaterializeGV(const GlobalValue *GV, EVT VT);
194 unsigned ARMMoveToFPReg(EVT VT, unsigned SrcReg);
195 unsigned ARMMoveToIntReg(EVT VT, unsigned SrcReg);
196 unsigned ARMSelectCallOp(const GlobalValue *GV);
198 // Call handling routines.
200 CCAssignFn *CCAssignFnForCall(CallingConv::ID CC, bool Return);
201 bool ProcessCallArgs(SmallVectorImpl<Value*> &Args,
202 SmallVectorImpl<unsigned> &ArgRegs,
203 SmallVectorImpl<MVT> &ArgVTs,
204 SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
205 SmallVectorImpl<unsigned> &RegArgs,
208 bool FinishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
209 const Instruction *I, CallingConv::ID CC,
211 bool ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call);
213 // OptionalDef handling routines.
215 bool isARMNEONPred(const MachineInstr *MI);
216 bool DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR);
217 const MachineInstrBuilder &AddOptionalDefs(const MachineInstrBuilder &MIB);
218 void AddLoadStoreOperands(EVT VT, Address &Addr,
219 const MachineInstrBuilder &MIB,
220 unsigned Flags, bool useAM3);
223 } // end anonymous namespace
225 #include "ARMGenCallingConv.inc"
227 // DefinesOptionalPredicate - This is different from DefinesPredicate in that
228 // we don't care about implicit defs here, just places we'll need to add a
229 // default CCReg argument. Sets CPSR if we're setting CPSR instead of CCR.
230 bool ARMFastISel::DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR) {
231 if (!MI->hasOptionalDef())
234 // Look to see if our OptionalDef is defining CPSR or CCR.
235 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
236 const MachineOperand &MO = MI->getOperand(i);
237 if (!MO.isReg() || !MO.isDef()) continue;
238 if (MO.getReg() == ARM::CPSR)
244 bool ARMFastISel::isARMNEONPred(const MachineInstr *MI) {
245 const MCInstrDesc &MCID = MI->getDesc();
247 // If we're a thumb2 or not NEON function we were handled via isPredicable.
248 if ((MCID.TSFlags & ARMII::DomainMask) != ARMII::DomainNEON ||
249 AFI->isThumb2Function())
252 for (unsigned i = 0, e = MCID.getNumOperands(); i != e; ++i)
253 if (MCID.OpInfo[i].isPredicate())
259 // If the machine is predicable go ahead and add the predicate operands, if
260 // it needs default CC operands add those.
261 // TODO: If we want to support thumb1 then we'll need to deal with optional
262 // CPSR defs that need to be added before the remaining operands. See s_cc_out
263 // for descriptions why.
264 const MachineInstrBuilder &
265 ARMFastISel::AddOptionalDefs(const MachineInstrBuilder &MIB) {
266 MachineInstr *MI = &*MIB;
268 // Do we use a predicate? or...
269 // Are we NEON in ARM mode and have a predicate operand? If so, I know
270 // we're not predicable but add it anyways.
271 if (TII.isPredicable(MI) || isARMNEONPred(MI))
274 // Do we optionally set a predicate? Preds is size > 0 iff the predicate
275 // defines CPSR. All other OptionalDefines in ARM are the CCR register.
277 if (DefinesOptionalPredicate(MI, &CPSR)) {
286 unsigned ARMFastISel::FastEmitInst_(unsigned MachineInstOpcode,
287 const TargetRegisterClass* RC) {
288 unsigned ResultReg = createResultReg(RC);
289 const MCInstrDesc &II = TII.get(MachineInstOpcode);
291 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg));
295 unsigned ARMFastISel::FastEmitInst_r(unsigned MachineInstOpcode,
296 const TargetRegisterClass *RC,
297 unsigned Op0, bool Op0IsKill) {
298 unsigned ResultReg = createResultReg(RC);
299 const MCInstrDesc &II = TII.get(MachineInstOpcode);
301 if (II.getNumDefs() >= 1)
302 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
303 .addReg(Op0, Op0IsKill * RegState::Kill));
305 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
306 .addReg(Op0, Op0IsKill * RegState::Kill));
307 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
308 TII.get(TargetOpcode::COPY), ResultReg)
309 .addReg(II.ImplicitDefs[0]));
314 unsigned ARMFastISel::FastEmitInst_rr(unsigned MachineInstOpcode,
315 const TargetRegisterClass *RC,
316 unsigned Op0, bool Op0IsKill,
317 unsigned Op1, bool Op1IsKill) {
318 unsigned ResultReg = createResultReg(RC);
319 const MCInstrDesc &II = TII.get(MachineInstOpcode);
321 if (II.getNumDefs() >= 1)
322 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
323 .addReg(Op0, Op0IsKill * RegState::Kill)
324 .addReg(Op1, Op1IsKill * RegState::Kill));
326 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
327 .addReg(Op0, Op0IsKill * RegState::Kill)
328 .addReg(Op1, Op1IsKill * RegState::Kill));
329 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
330 TII.get(TargetOpcode::COPY), ResultReg)
331 .addReg(II.ImplicitDefs[0]));
336 unsigned ARMFastISel::FastEmitInst_rrr(unsigned MachineInstOpcode,
337 const TargetRegisterClass *RC,
338 unsigned Op0, bool Op0IsKill,
339 unsigned Op1, bool Op1IsKill,
340 unsigned Op2, bool Op2IsKill) {
341 unsigned ResultReg = createResultReg(RC);
342 const MCInstrDesc &II = TII.get(MachineInstOpcode);
344 if (II.getNumDefs() >= 1)
345 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
346 .addReg(Op0, Op0IsKill * RegState::Kill)
347 .addReg(Op1, Op1IsKill * RegState::Kill)
348 .addReg(Op2, Op2IsKill * RegState::Kill));
350 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
351 .addReg(Op0, Op0IsKill * RegState::Kill)
352 .addReg(Op1, Op1IsKill * RegState::Kill)
353 .addReg(Op2, Op2IsKill * RegState::Kill));
354 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
355 TII.get(TargetOpcode::COPY), ResultReg)
356 .addReg(II.ImplicitDefs[0]));
361 unsigned ARMFastISel::FastEmitInst_ri(unsigned MachineInstOpcode,
362 const TargetRegisterClass *RC,
363 unsigned Op0, bool Op0IsKill,
365 unsigned ResultReg = createResultReg(RC);
366 const MCInstrDesc &II = TII.get(MachineInstOpcode);
368 if (II.getNumDefs() >= 1)
369 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
370 .addReg(Op0, Op0IsKill * RegState::Kill)
373 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
374 .addReg(Op0, Op0IsKill * RegState::Kill)
376 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
377 TII.get(TargetOpcode::COPY), ResultReg)
378 .addReg(II.ImplicitDefs[0]));
383 unsigned ARMFastISel::FastEmitInst_rf(unsigned MachineInstOpcode,
384 const TargetRegisterClass *RC,
385 unsigned Op0, bool Op0IsKill,
386 const ConstantFP *FPImm) {
387 unsigned ResultReg = createResultReg(RC);
388 const MCInstrDesc &II = TII.get(MachineInstOpcode);
390 if (II.getNumDefs() >= 1)
391 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
392 .addReg(Op0, Op0IsKill * RegState::Kill)
395 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
396 .addReg(Op0, Op0IsKill * RegState::Kill)
398 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
399 TII.get(TargetOpcode::COPY), ResultReg)
400 .addReg(II.ImplicitDefs[0]));
405 unsigned ARMFastISel::FastEmitInst_rri(unsigned MachineInstOpcode,
406 const TargetRegisterClass *RC,
407 unsigned Op0, bool Op0IsKill,
408 unsigned Op1, bool Op1IsKill,
410 unsigned ResultReg = createResultReg(RC);
411 const MCInstrDesc &II = TII.get(MachineInstOpcode);
413 if (II.getNumDefs() >= 1)
414 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
415 .addReg(Op0, Op0IsKill * RegState::Kill)
416 .addReg(Op1, Op1IsKill * RegState::Kill)
419 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
420 .addReg(Op0, Op0IsKill * RegState::Kill)
421 .addReg(Op1, Op1IsKill * RegState::Kill)
423 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
424 TII.get(TargetOpcode::COPY), ResultReg)
425 .addReg(II.ImplicitDefs[0]));
430 unsigned ARMFastISel::FastEmitInst_i(unsigned MachineInstOpcode,
431 const TargetRegisterClass *RC,
433 unsigned ResultReg = createResultReg(RC);
434 const MCInstrDesc &II = TII.get(MachineInstOpcode);
436 if (II.getNumDefs() >= 1)
437 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
440 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
442 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
443 TII.get(TargetOpcode::COPY), ResultReg)
444 .addReg(II.ImplicitDefs[0]));
449 unsigned ARMFastISel::FastEmitInst_ii(unsigned MachineInstOpcode,
450 const TargetRegisterClass *RC,
451 uint64_t Imm1, uint64_t Imm2) {
452 unsigned ResultReg = createResultReg(RC);
453 const MCInstrDesc &II = TII.get(MachineInstOpcode);
455 if (II.getNumDefs() >= 1)
456 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
457 .addImm(Imm1).addImm(Imm2));
459 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
460 .addImm(Imm1).addImm(Imm2));
461 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
462 TII.get(TargetOpcode::COPY),
464 .addReg(II.ImplicitDefs[0]));
469 unsigned ARMFastISel::FastEmitInst_extractsubreg(MVT RetVT,
470 unsigned Op0, bool Op0IsKill,
472 unsigned ResultReg = createResultReg(TLI.getRegClassFor(RetVT));
473 assert(TargetRegisterInfo::isVirtualRegister(Op0) &&
474 "Cannot yet extract from physregs");
475 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
476 DL, TII.get(TargetOpcode::COPY), ResultReg)
477 .addReg(Op0, getKillRegState(Op0IsKill), Idx));
481 // TODO: Don't worry about 64-bit now, but when this is fixed remove the
482 // checks from the various callers.
483 unsigned ARMFastISel::ARMMoveToFPReg(EVT VT, unsigned SrcReg) {
484 if (VT == MVT::f64) return 0;
486 unsigned MoveReg = createResultReg(TLI.getRegClassFor(VT));
487 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
488 TII.get(ARM::VMOVRS), MoveReg)
493 unsigned ARMFastISel::ARMMoveToIntReg(EVT VT, unsigned SrcReg) {
494 if (VT == MVT::i64) return 0;
496 unsigned MoveReg = createResultReg(TLI.getRegClassFor(VT));
497 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
498 TII.get(ARM::VMOVSR), MoveReg)
503 // For double width floating point we need to materialize two constants
504 // (the high and the low) into integer registers then use a move to get
505 // the combined constant into an FP reg.
506 unsigned ARMFastISel::ARMMaterializeFP(const ConstantFP *CFP, EVT VT) {
507 const APFloat Val = CFP->getValueAPF();
508 bool is64bit = VT == MVT::f64;
510 // This checks to see if we can use VFP3 instructions to materialize
511 // a constant, otherwise we have to go through the constant pool.
512 if (TLI.isFPImmLegal(Val, VT)) {
516 Imm = ARM_AM::getFP64Imm(Val);
519 Imm = ARM_AM::getFP32Imm(Val);
522 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
523 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
529 // Require VFP2 for loading fp constants.
530 if (!Subtarget->hasVFP2()) return false;
532 // MachineConstantPool wants an explicit alignment.
533 unsigned Align = TD.getPrefTypeAlignment(CFP->getType());
535 // TODO: Figure out if this is correct.
536 Align = TD.getTypeAllocSize(CFP->getType());
538 unsigned Idx = MCP.getConstantPoolIndex(cast<Constant>(CFP), Align);
539 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
540 unsigned Opc = is64bit ? ARM::VLDRD : ARM::VLDRS;
542 // The extra reg is for addrmode5.
543 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
545 .addConstantPoolIndex(Idx)
550 unsigned ARMFastISel::ARMMaterializeInt(const Constant *C, EVT VT) {
552 if (VT != MVT::i32 && VT != MVT::i16 && VT != MVT::i8 && VT != MVT::i1)
555 // If we can do this in a single instruction without a constant pool entry
557 const ConstantInt *CI = cast<ConstantInt>(C);
558 if (Subtarget->hasV6T2Ops() && isUInt<16>(CI->getZExtValue())) {
559 unsigned Opc = isThumb2 ? ARM::t2MOVi16 : ARM::MOVi16;
560 unsigned ImmReg = createResultReg(TLI.getRegClassFor(MVT::i32));
561 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
562 TII.get(Opc), ImmReg)
563 .addImm(CI->getZExtValue()));
567 // Use MVN to emit negative constants.
568 if (VT == MVT::i32 && Subtarget->hasV6T2Ops() && CI->isNegative()) {
569 unsigned Imm = (unsigned)~(CI->getSExtValue());
570 bool UseImm = isThumb2 ? (ARM_AM::getT2SOImmVal(Imm) != -1) :
571 (ARM_AM::getSOImmVal(Imm) != -1);
573 unsigned Opc = isThumb2 ? ARM::t2MVNi : ARM::MVNi;
574 unsigned ImmReg = createResultReg(TLI.getRegClassFor(MVT::i32));
575 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
576 TII.get(Opc), ImmReg)
582 // Load from constant pool. For now 32-bit only.
586 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
588 // MachineConstantPool wants an explicit alignment.
589 unsigned Align = TD.getPrefTypeAlignment(C->getType());
591 // TODO: Figure out if this is correct.
592 Align = TD.getTypeAllocSize(C->getType());
594 unsigned Idx = MCP.getConstantPoolIndex(C, Align);
597 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
598 TII.get(ARM::t2LDRpci), DestReg)
599 .addConstantPoolIndex(Idx));
601 // The extra immediate is for addrmode2.
602 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
603 TII.get(ARM::LDRcp), DestReg)
604 .addConstantPoolIndex(Idx)
610 unsigned ARMFastISel::ARMMaterializeGV(const GlobalValue *GV, EVT VT) {
611 // For now 32-bit only.
612 if (VT != MVT::i32) return 0;
614 Reloc::Model RelocM = TM.getRelocationModel();
616 // TODO: Need more magic for ARM PIC.
617 if (!isThumb2 && (RelocM == Reloc::PIC_)) return 0;
619 // MachineConstantPool wants an explicit alignment.
620 unsigned Align = TD.getPrefTypeAlignment(GV->getType());
622 // TODO: Figure out if this is correct.
623 Align = TD.getTypeAllocSize(GV->getType());
627 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb() ? 4 : 8);
628 unsigned Id = AFI->createPICLabelUId();
629 ARMConstantPoolValue *CPV = ARMConstantPoolConstant::Create(GV, Id,
632 unsigned Idx = MCP.getConstantPoolIndex(CPV, Align);
635 MachineInstrBuilder MIB;
636 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
638 unsigned Opc = (RelocM != Reloc::PIC_) ? ARM::t2LDRpci : ARM::t2LDRpci_pic;
639 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), DestReg)
640 .addConstantPoolIndex(Idx);
641 if (RelocM == Reloc::PIC_)
644 // The extra immediate is for addrmode2.
645 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(ARM::LDRcp),
647 .addConstantPoolIndex(Idx)
650 AddOptionalDefs(MIB);
652 if (Subtarget->GVIsIndirectSymbol(GV, RelocM)) {
653 unsigned NewDestReg = createResultReg(TLI.getRegClassFor(VT));
655 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
656 TII.get(ARM::t2LDRi12), NewDestReg)
660 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(ARM::LDRi12),
664 DestReg = NewDestReg;
665 AddOptionalDefs(MIB);
671 unsigned ARMFastISel::TargetMaterializeConstant(const Constant *C) {
672 EVT VT = TLI.getValueType(C->getType(), true);
674 // Only handle simple types.
675 if (!VT.isSimple()) return 0;
677 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
678 return ARMMaterializeFP(CFP, VT);
679 else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
680 return ARMMaterializeGV(GV, VT);
681 else if (isa<ConstantInt>(C))
682 return ARMMaterializeInt(C, VT);
687 // TODO: unsigned ARMFastISel::TargetMaterializeFloatZero(const ConstantFP *CF);
689 unsigned ARMFastISel::TargetMaterializeAlloca(const AllocaInst *AI) {
690 // Don't handle dynamic allocas.
691 if (!FuncInfo.StaticAllocaMap.count(AI)) return 0;
694 if (!isLoadTypeLegal(AI->getType(), VT)) return false;
696 DenseMap<const AllocaInst*, int>::iterator SI =
697 FuncInfo.StaticAllocaMap.find(AI);
699 // This will get lowered later into the correct offsets and registers
700 // via rewriteXFrameIndex.
701 if (SI != FuncInfo.StaticAllocaMap.end()) {
702 TargetRegisterClass* RC = TLI.getRegClassFor(VT);
703 unsigned ResultReg = createResultReg(RC);
704 unsigned Opc = isThumb2 ? ARM::t2ADDri : ARM::ADDri;
705 AddOptionalDefs(BuildMI(*FuncInfo.MBB, *FuncInfo.InsertPt, DL,
706 TII.get(Opc), ResultReg)
707 .addFrameIndex(SI->second)
715 bool ARMFastISel::isTypeLegal(Type *Ty, MVT &VT) {
716 EVT evt = TLI.getValueType(Ty, true);
718 // Only handle simple types.
719 if (evt == MVT::Other || !evt.isSimple()) return false;
720 VT = evt.getSimpleVT();
722 // Handle all legal types, i.e. a register that will directly hold this
724 return TLI.isTypeLegal(VT);
727 bool ARMFastISel::isLoadTypeLegal(Type *Ty, MVT &VT) {
728 if (isTypeLegal(Ty, VT)) return true;
730 // If this is a type than can be sign or zero-extended to a basic operation
731 // go ahead and accept it now.
732 if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16)
738 // Computes the address to get to an object.
739 bool ARMFastISel::ARMComputeAddress(const Value *Obj, Address &Addr) {
740 // Some boilerplate from the X86 FastISel.
741 const User *U = NULL;
742 unsigned Opcode = Instruction::UserOp1;
743 if (const Instruction *I = dyn_cast<Instruction>(Obj)) {
744 // Don't walk into other basic blocks unless the object is an alloca from
745 // another block, otherwise it may not have a virtual register assigned.
746 if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(Obj)) ||
747 FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
748 Opcode = I->getOpcode();
751 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Obj)) {
752 Opcode = C->getOpcode();
756 if (PointerType *Ty = dyn_cast<PointerType>(Obj->getType()))
757 if (Ty->getAddressSpace() > 255)
758 // Fast instruction selection doesn't support the special
765 case Instruction::BitCast: {
766 // Look through bitcasts.
767 return ARMComputeAddress(U->getOperand(0), Addr);
769 case Instruction::IntToPtr: {
770 // Look past no-op inttoptrs.
771 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
772 return ARMComputeAddress(U->getOperand(0), Addr);
775 case Instruction::PtrToInt: {
776 // Look past no-op ptrtoints.
777 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
778 return ARMComputeAddress(U->getOperand(0), Addr);
781 case Instruction::GetElementPtr: {
782 Address SavedAddr = Addr;
783 int TmpOffset = Addr.Offset;
785 // Iterate through the GEP folding the constants into offsets where
787 gep_type_iterator GTI = gep_type_begin(U);
788 for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end();
789 i != e; ++i, ++GTI) {
790 const Value *Op = *i;
791 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
792 const StructLayout *SL = TD.getStructLayout(STy);
793 unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
794 TmpOffset += SL->getElementOffset(Idx);
796 uint64_t S = TD.getTypeAllocSize(GTI.getIndexedType());
798 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
799 // Constant-offset addressing.
800 TmpOffset += CI->getSExtValue() * S;
803 if (isa<AddOperator>(Op) &&
804 (!isa<Instruction>(Op) ||
805 FuncInfo.MBBMap[cast<Instruction>(Op)->getParent()]
807 isa<ConstantInt>(cast<AddOperator>(Op)->getOperand(1))) {
808 // An add (in the same block) with a constant operand. Fold the
811 cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
812 TmpOffset += CI->getSExtValue() * S;
813 // Iterate on the other operand.
814 Op = cast<AddOperator>(Op)->getOperand(0);
818 goto unsupported_gep;
823 // Try to grab the base operand now.
824 Addr.Offset = TmpOffset;
825 if (ARMComputeAddress(U->getOperand(0), Addr)) return true;
827 // We failed, restore everything and try the other options.
833 case Instruction::Alloca: {
834 const AllocaInst *AI = cast<AllocaInst>(Obj);
835 DenseMap<const AllocaInst*, int>::iterator SI =
836 FuncInfo.StaticAllocaMap.find(AI);
837 if (SI != FuncInfo.StaticAllocaMap.end()) {
838 Addr.BaseType = Address::FrameIndexBase;
839 Addr.Base.FI = SI->second;
846 // Materialize the global variable's address into a reg which can
847 // then be used later to load the variable.
848 if (const GlobalValue *GV = dyn_cast<GlobalValue>(Obj)) {
849 unsigned Tmp = ARMMaterializeGV(GV, TLI.getValueType(Obj->getType()));
850 if (Tmp == 0) return false;
856 // Try to get this in a register if nothing else has worked.
857 if (Addr.Base.Reg == 0) Addr.Base.Reg = getRegForValue(Obj);
858 return Addr.Base.Reg != 0;
861 void ARMFastISel::ARMSimplifyAddress(Address &Addr, EVT VT, bool useAM3) {
863 assert(VT.isSimple() && "Non-simple types are invalid here!");
865 bool needsLowering = false;
866 switch (VT.getSimpleVT().SimpleTy) {
868 assert(false && "Unhandled load/store type!");
875 // Integer loads/stores handle 12-bit offsets.
876 needsLowering = ((Addr.Offset & 0xfff) != Addr.Offset);
877 // Handle negative offsets.
878 if (needsLowering && isThumb2)
879 needsLowering = !(Subtarget->hasV6T2Ops() && Addr.Offset < 0 &&
882 // ARM halfword load/stores and signed byte loads use +/-imm8 offsets.
883 needsLowering = (Addr.Offset > 255 || Addr.Offset < -255);
888 // Floating point operands handle 8-bit offsets.
889 needsLowering = ((Addr.Offset & 0xff) != Addr.Offset);
893 // If this is a stack pointer and the offset needs to be simplified then
894 // put the alloca address into a register, set the base type back to
895 // register and continue. This should almost never happen.
896 if (needsLowering && Addr.BaseType == Address::FrameIndexBase) {
897 TargetRegisterClass *RC = isThumb2 ? ARM::tGPRRegisterClass :
898 ARM::GPRRegisterClass;
899 unsigned ResultReg = createResultReg(RC);
900 unsigned Opc = isThumb2 ? ARM::t2ADDri : ARM::ADDri;
901 AddOptionalDefs(BuildMI(*FuncInfo.MBB, *FuncInfo.InsertPt, DL,
902 TII.get(Opc), ResultReg)
903 .addFrameIndex(Addr.Base.FI)
905 Addr.Base.Reg = ResultReg;
906 Addr.BaseType = Address::RegBase;
909 // Since the offset is too large for the load/store instruction
910 // get the reg+offset into a register.
912 Addr.Base.Reg = FastEmit_ri_(MVT::i32, ISD::ADD, Addr.Base.Reg,
913 /*Op0IsKill*/false, Addr.Offset, MVT::i32);
918 void ARMFastISel::AddLoadStoreOperands(EVT VT, Address &Addr,
919 const MachineInstrBuilder &MIB,
920 unsigned Flags, bool useAM3) {
921 // addrmode5 output depends on the selection dag addressing dividing the
922 // offset by 4 that it then later multiplies. Do this here as well.
923 if (VT.getSimpleVT().SimpleTy == MVT::f32 ||
924 VT.getSimpleVT().SimpleTy == MVT::f64)
927 // Frame base works a bit differently. Handle it separately.
928 if (Addr.BaseType == Address::FrameIndexBase) {
929 int FI = Addr.Base.FI;
930 int Offset = Addr.Offset;
931 MachineMemOperand *MMO =
932 FuncInfo.MF->getMachineMemOperand(
933 MachinePointerInfo::getFixedStack(FI, Offset),
935 MFI.getObjectSize(FI),
936 MFI.getObjectAlignment(FI));
937 // Now add the rest of the operands.
938 MIB.addFrameIndex(FI);
940 // ARM halfword load/stores and signed byte loads need an additional
943 signed Imm = (Addr.Offset < 0) ? (0x100 | -Addr.Offset) : Addr.Offset;
947 MIB.addImm(Addr.Offset);
949 MIB.addMemOperand(MMO);
951 // Now add the rest of the operands.
952 MIB.addReg(Addr.Base.Reg);
954 // ARM halfword load/stores and signed byte loads need an additional
957 signed Imm = (Addr.Offset < 0) ? (0x100 | -Addr.Offset) : Addr.Offset;
961 MIB.addImm(Addr.Offset);
964 AddOptionalDefs(MIB);
967 bool ARMFastISel::ARMEmitLoad(EVT VT, unsigned &ResultReg, Address &Addr,
968 unsigned Alignment, bool isZExt, bool allocReg) {
969 assert(VT.isSimple() && "Non-simple types are invalid here!");
972 bool needVMOV = false;
973 TargetRegisterClass *RC;
974 switch (VT.getSimpleVT().SimpleTy) {
975 // This is mostly going to be Neon/vector support.
976 default: return false;
980 if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
981 Opc = isZExt ? ARM::t2LDRBi8 : ARM::t2LDRSBi8;
983 Opc = isZExt ? ARM::t2LDRBi12 : ARM::t2LDRSBi12;
992 RC = ARM::GPRRegisterClass;
996 if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
997 Opc = isZExt ? ARM::t2LDRHi8 : ARM::t2LDRSHi8;
999 Opc = isZExt ? ARM::t2LDRHi12 : ARM::t2LDRSHi12;
1001 Opc = isZExt ? ARM::LDRH : ARM::LDRSH;
1004 RC = ARM::GPRRegisterClass;
1008 if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
1011 Opc = ARM::t2LDRi12;
1015 RC = ARM::GPRRegisterClass;
1018 // Unaligned loads need special handling. Floats require word-alignment.
1019 if (Alignment && Alignment < 4) {
1022 Opc = isThumb2 ? ARM::t2LDRi12 : ARM::LDRi12;
1023 RC = ARM::GPRRegisterClass;
1026 RC = TLI.getRegClassFor(VT);
1030 if (Alignment && Alignment < 4) {
1031 // FIXME: Unaligned loads need special handling. Doublewords require
1036 RC = TLI.getRegClassFor(VT);
1039 // Simplify this down to something we can handle.
1040 ARMSimplifyAddress(Addr, VT, useAM3);
1042 // Create the base instruction, then add the operands.
1044 ResultReg = createResultReg(RC);
1045 assert (ResultReg > 255 && "Expected an allocated virtual register.");
1046 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1047 TII.get(Opc), ResultReg);
1048 AddLoadStoreOperands(VT, Addr, MIB, MachineMemOperand::MOLoad, useAM3);
1050 // If we had an unaligned load of a float we've converted it to an regular
1051 // load. Now we must move from the GRP to the FP register.
1053 unsigned MoveReg = createResultReg(TLI.getRegClassFor(MVT::f32));
1054 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1055 TII.get(ARM::VMOVSR), MoveReg)
1056 .addReg(ResultReg));
1057 ResultReg = MoveReg;
1062 bool ARMFastISel::SelectLoad(const Instruction *I) {
1063 // Atomic loads need special handling.
1064 if (cast<LoadInst>(I)->isAtomic())
1067 // Verify we have a legal type before going any further.
1069 if (!isLoadTypeLegal(I->getType(), VT))
1072 // See if we can handle this address.
1074 if (!ARMComputeAddress(I->getOperand(0), Addr)) return false;
1077 if (!ARMEmitLoad(VT, ResultReg, Addr, cast<LoadInst>(I)->getAlignment()))
1079 UpdateValueMap(I, ResultReg);
1083 bool ARMFastISel::ARMEmitStore(EVT VT, unsigned SrcReg, Address &Addr,
1084 unsigned Alignment) {
1086 bool useAM3 = false;
1087 switch (VT.getSimpleVT().SimpleTy) {
1088 // This is mostly going to be Neon/vector support.
1089 default: return false;
1091 unsigned Res = createResultReg(isThumb2 ? ARM::tGPRRegisterClass :
1092 ARM::GPRRegisterClass);
1093 unsigned Opc = isThumb2 ? ARM::t2ANDri : ARM::ANDri;
1094 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1096 .addReg(SrcReg).addImm(1));
1098 } // Fallthrough here.
1101 if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
1102 StrOpc = ARM::t2STRBi8;
1104 StrOpc = ARM::t2STRBi12;
1106 StrOpc = ARM::STRBi12;
1111 if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
1112 StrOpc = ARM::t2STRHi8;
1114 StrOpc = ARM::t2STRHi12;
1122 if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
1123 StrOpc = ARM::t2STRi8;
1125 StrOpc = ARM::t2STRi12;
1127 StrOpc = ARM::STRi12;
1131 if (!Subtarget->hasVFP2()) return false;
1132 StrOpc = ARM::VSTRS;
1133 // Unaligned stores need special handling. Floats require word-alignment.
1134 if (Alignment && Alignment < 4) {
1135 unsigned MoveReg = createResultReg(TLI.getRegClassFor(MVT::i32));
1136 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1137 TII.get(ARM::VMOVRS), MoveReg)
1141 StrOpc = isThumb2 ? ARM::t2STRi12 : ARM::STRi12;
1145 if (!Subtarget->hasVFP2()) return false;
1146 // FIXME: Unaligned stores need special handling. Doublewords require
1148 if (Alignment && Alignment < 4) {
1151 StrOpc = ARM::VSTRD;
1154 // Simplify this down to something we can handle.
1155 ARMSimplifyAddress(Addr, VT, useAM3);
1157 // Create the base instruction, then add the operands.
1158 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1161 AddLoadStoreOperands(VT, Addr, MIB, MachineMemOperand::MOStore, useAM3);
1165 bool ARMFastISel::SelectStore(const Instruction *I) {
1166 Value *Op0 = I->getOperand(0);
1167 unsigned SrcReg = 0;
1169 // Atomic stores need special handling.
1170 if (cast<StoreInst>(I)->isAtomic())
1173 // Verify we have a legal type before going any further.
1175 if (!isLoadTypeLegal(I->getOperand(0)->getType(), VT))
1178 // Get the value to be stored into a register.
1179 SrcReg = getRegForValue(Op0);
1180 if (SrcReg == 0) return false;
1182 // See if we can handle this address.
1184 if (!ARMComputeAddress(I->getOperand(1), Addr))
1187 if (!ARMEmitStore(VT, SrcReg, Addr, cast<StoreInst>(I)->getAlignment()))
1192 static ARMCC::CondCodes getComparePred(CmpInst::Predicate Pred) {
1194 // Needs two compares...
1195 case CmpInst::FCMP_ONE:
1196 case CmpInst::FCMP_UEQ:
1198 // AL is our "false" for now. The other two need more compares.
1200 case CmpInst::ICMP_EQ:
1201 case CmpInst::FCMP_OEQ:
1203 case CmpInst::ICMP_SGT:
1204 case CmpInst::FCMP_OGT:
1206 case CmpInst::ICMP_SGE:
1207 case CmpInst::FCMP_OGE:
1209 case CmpInst::ICMP_UGT:
1210 case CmpInst::FCMP_UGT:
1212 case CmpInst::FCMP_OLT:
1214 case CmpInst::ICMP_ULE:
1215 case CmpInst::FCMP_OLE:
1217 case CmpInst::FCMP_ORD:
1219 case CmpInst::FCMP_UNO:
1221 case CmpInst::FCMP_UGE:
1223 case CmpInst::ICMP_SLT:
1224 case CmpInst::FCMP_ULT:
1226 case CmpInst::ICMP_SLE:
1227 case CmpInst::FCMP_ULE:
1229 case CmpInst::FCMP_UNE:
1230 case CmpInst::ICMP_NE:
1232 case CmpInst::ICMP_UGE:
1234 case CmpInst::ICMP_ULT:
1239 bool ARMFastISel::SelectBranch(const Instruction *I) {
1240 const BranchInst *BI = cast<BranchInst>(I);
1241 MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
1242 MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
1244 // Simple branch support.
1246 // If we can, avoid recomputing the compare - redoing it could lead to wonky
1248 if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
1249 if (CI->hasOneUse() && (CI->getParent() == I->getParent())) {
1251 // Get the compare predicate.
1252 // Try to take advantage of fallthrough opportunities.
1253 CmpInst::Predicate Predicate = CI->getPredicate();
1254 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
1255 std::swap(TBB, FBB);
1256 Predicate = CmpInst::getInversePredicate(Predicate);
1259 ARMCC::CondCodes ARMPred = getComparePred(Predicate);
1261 // We may not handle every CC for now.
1262 if (ARMPred == ARMCC::AL) return false;
1264 // Emit the compare.
1265 if (!ARMEmitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
1268 unsigned BrOpc = isThumb2 ? ARM::t2Bcc : ARM::Bcc;
1269 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
1270 .addMBB(TBB).addImm(ARMPred).addReg(ARM::CPSR);
1271 FastEmitBranch(FBB, DL);
1272 FuncInfo.MBB->addSuccessor(TBB);
1275 } else if (TruncInst *TI = dyn_cast<TruncInst>(BI->getCondition())) {
1277 if (TI->hasOneUse() && TI->getParent() == I->getParent() &&
1278 (isLoadTypeLegal(TI->getOperand(0)->getType(), SourceVT))) {
1279 unsigned TstOpc = isThumb2 ? ARM::t2TSTri : ARM::TSTri;
1280 unsigned OpReg = getRegForValue(TI->getOperand(0));
1281 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1283 .addReg(OpReg).addImm(1));
1285 unsigned CCMode = ARMCC::NE;
1286 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
1287 std::swap(TBB, FBB);
1291 unsigned BrOpc = isThumb2 ? ARM::t2Bcc : ARM::Bcc;
1292 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
1293 .addMBB(TBB).addImm(CCMode).addReg(ARM::CPSR);
1295 FastEmitBranch(FBB, DL);
1296 FuncInfo.MBB->addSuccessor(TBB);
1299 } else if (const ConstantInt *CI =
1300 dyn_cast<ConstantInt>(BI->getCondition())) {
1301 uint64_t Imm = CI->getZExtValue();
1302 MachineBasicBlock *Target = (Imm == 0) ? FBB : TBB;
1303 FastEmitBranch(Target, DL);
1307 unsigned CmpReg = getRegForValue(BI->getCondition());
1308 if (CmpReg == 0) return false;
1310 // We've been divorced from our compare! Our block was split, and
1311 // now our compare lives in a predecessor block. We musn't
1312 // re-compare here, as the children of the compare aren't guaranteed
1313 // live across the block boundary (we *could* check for this).
1314 // Regardless, the compare has been done in the predecessor block,
1315 // and it left a value for us in a virtual register. Ergo, we test
1316 // the one-bit value left in the virtual register.
1317 unsigned TstOpc = isThumb2 ? ARM::t2TSTri : ARM::TSTri;
1318 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TstOpc))
1319 .addReg(CmpReg).addImm(1));
1321 unsigned CCMode = ARMCC::NE;
1322 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
1323 std::swap(TBB, FBB);
1327 unsigned BrOpc = isThumb2 ? ARM::t2Bcc : ARM::Bcc;
1328 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
1329 .addMBB(TBB).addImm(CCMode).addReg(ARM::CPSR);
1330 FastEmitBranch(FBB, DL);
1331 FuncInfo.MBB->addSuccessor(TBB);
1335 bool ARMFastISel::ARMEmitCmp(const Value *Src1Value, const Value *Src2Value,
1337 Type *Ty = Src1Value->getType();
1338 EVT SrcVT = TLI.getValueType(Ty, true);
1339 if (!SrcVT.isSimple()) return false;
1341 bool isFloat = (Ty->isFloatTy() || Ty->isDoubleTy());
1342 if (isFloat && !Subtarget->hasVFP2())
1345 // Check to see if the 2nd operand is a constant that we can encode directly
1348 bool UseImm = false;
1349 bool isNegativeImm = false;
1350 // FIXME: At -O0 we don't have anything that canonicalizes operand order.
1351 // Thus, Src1Value may be a ConstantInt, but we're missing it.
1352 if (const ConstantInt *ConstInt = dyn_cast<ConstantInt>(Src2Value)) {
1353 if (SrcVT == MVT::i32 || SrcVT == MVT::i16 || SrcVT == MVT::i8 ||
1355 const APInt &CIVal = ConstInt->getValue();
1356 Imm = (isZExt) ? (int)CIVal.getZExtValue() : (int)CIVal.getSExtValue();
1358 isNegativeImm = true;
1361 UseImm = isThumb2 ? (ARM_AM::getT2SOImmVal(Imm) != -1) :
1362 (ARM_AM::getSOImmVal(Imm) != -1);
1364 } else if (const ConstantFP *ConstFP = dyn_cast<ConstantFP>(Src2Value)) {
1365 if (SrcVT == MVT::f32 || SrcVT == MVT::f64)
1366 if (ConstFP->isZero() && !ConstFP->isNegative())
1372 bool needsExt = false;
1373 switch (SrcVT.getSimpleVT().SimpleTy) {
1374 default: return false;
1375 // TODO: Verify compares.
1378 CmpOpc = UseImm ? ARM::VCMPEZS : ARM::VCMPES;
1382 CmpOpc = UseImm ? ARM::VCMPEZD : ARM::VCMPED;
1388 // Intentional fall-through.
1392 CmpOpc = ARM::t2CMPrr;
1394 CmpOpc = isNegativeImm ? ARM::t2CMNzri : ARM::t2CMPri;
1397 CmpOpc = ARM::CMPrr;
1399 CmpOpc = isNegativeImm ? ARM::CMNzri : ARM::CMPri;
1404 unsigned SrcReg1 = getRegForValue(Src1Value);
1405 if (SrcReg1 == 0) return false;
1407 unsigned SrcReg2 = 0;
1409 SrcReg2 = getRegForValue(Src2Value);
1410 if (SrcReg2 == 0) return false;
1413 // We have i1, i8, or i16, we need to either zero extend or sign extend.
1416 ResultReg = ARMEmitIntExt(SrcVT, SrcReg1, MVT::i32, isZExt);
1417 if (ResultReg == 0) return false;
1418 SrcReg1 = ResultReg;
1420 ResultReg = ARMEmitIntExt(SrcVT, SrcReg2, MVT::i32, isZExt);
1421 if (ResultReg == 0) return false;
1422 SrcReg2 = ResultReg;
1427 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1429 .addReg(SrcReg1).addReg(SrcReg2));
1431 MachineInstrBuilder MIB;
1432 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
1435 // Only add immediate for icmp as the immediate for fcmp is an implicit 0.0.
1438 AddOptionalDefs(MIB);
1441 // For floating point we need to move the result to a comparison register
1442 // that we can then use for branches.
1443 if (Ty->isFloatTy() || Ty->isDoubleTy())
1444 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1445 TII.get(ARM::FMSTAT)));
1449 bool ARMFastISel::SelectCmp(const Instruction *I) {
1450 const CmpInst *CI = cast<CmpInst>(I);
1451 Type *Ty = CI->getOperand(0)->getType();
1453 // Get the compare predicate.
1454 ARMCC::CondCodes ARMPred = getComparePred(CI->getPredicate());
1456 // We may not handle every CC for now.
1457 if (ARMPred == ARMCC::AL) return false;
1459 // Emit the compare.
1460 if (!ARMEmitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
1463 // Now set a register based on the comparison. Explicitly set the predicates
1465 unsigned MovCCOpc = isThumb2 ? ARM::t2MOVCCi : ARM::MOVCCi;
1466 TargetRegisterClass *RC = isThumb2 ? ARM::rGPRRegisterClass
1467 : ARM::GPRRegisterClass;
1468 unsigned DestReg = createResultReg(RC);
1469 Constant *Zero = ConstantInt::get(Type::getInt32Ty(*Context), 0);
1470 unsigned ZeroReg = TargetMaterializeConstant(Zero);
1471 bool isFloat = (Ty->isFloatTy() || Ty->isDoubleTy());
1472 unsigned CondReg = isFloat ? ARM::FPSCR : ARM::CPSR;
1473 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), DestReg)
1474 .addReg(ZeroReg).addImm(1)
1475 .addImm(ARMPred).addReg(CondReg);
1477 UpdateValueMap(I, DestReg);
1481 bool ARMFastISel::SelectFPExt(const Instruction *I) {
1482 // Make sure we have VFP and that we're extending float to double.
1483 if (!Subtarget->hasVFP2()) return false;
1485 Value *V = I->getOperand(0);
1486 if (!I->getType()->isDoubleTy() ||
1487 !V->getType()->isFloatTy()) return false;
1489 unsigned Op = getRegForValue(V);
1490 if (Op == 0) return false;
1492 unsigned Result = createResultReg(ARM::DPRRegisterClass);
1493 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1494 TII.get(ARM::VCVTDS), Result)
1496 UpdateValueMap(I, Result);
1500 bool ARMFastISel::SelectFPTrunc(const Instruction *I) {
1501 // Make sure we have VFP and that we're truncating double to float.
1502 if (!Subtarget->hasVFP2()) return false;
1504 Value *V = I->getOperand(0);
1505 if (!(I->getType()->isFloatTy() &&
1506 V->getType()->isDoubleTy())) return false;
1508 unsigned Op = getRegForValue(V);
1509 if (Op == 0) return false;
1511 unsigned Result = createResultReg(ARM::SPRRegisterClass);
1512 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1513 TII.get(ARM::VCVTSD), Result)
1515 UpdateValueMap(I, Result);
1519 bool ARMFastISel::SelectSIToFP(const Instruction *I) {
1520 // Make sure we have VFP.
1521 if (!Subtarget->hasVFP2()) return false;
1524 Type *Ty = I->getType();
1525 if (!isTypeLegal(Ty, DstVT))
1528 Value *Src = I->getOperand(0);
1529 EVT SrcVT = TLI.getValueType(Src->getType(), true);
1530 if (SrcVT != MVT::i32 && SrcVT != MVT::i16 && SrcVT != MVT::i8)
1533 unsigned SrcReg = getRegForValue(Src);
1534 if (SrcReg == 0) return false;
1536 // Handle sign-extension.
1537 if (SrcVT == MVT::i16 || SrcVT == MVT::i8) {
1538 EVT DestVT = MVT::i32;
1539 unsigned ResultReg = ARMEmitIntExt(SrcVT, SrcReg, DestVT, /*isZExt*/ false);
1540 if (ResultReg == 0) return false;
1544 // The conversion routine works on fp-reg to fp-reg and the operand above
1545 // was an integer, move it to the fp registers if possible.
1546 unsigned FP = ARMMoveToFPReg(MVT::f32, SrcReg);
1547 if (FP == 0) return false;
1550 if (Ty->isFloatTy()) Opc = ARM::VSITOS;
1551 else if (Ty->isDoubleTy()) Opc = ARM::VSITOD;
1554 unsigned ResultReg = createResultReg(TLI.getRegClassFor(DstVT));
1555 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
1558 UpdateValueMap(I, ResultReg);
1562 bool ARMFastISel::SelectFPToSI(const Instruction *I) {
1563 // Make sure we have VFP.
1564 if (!Subtarget->hasVFP2()) return false;
1567 Type *RetTy = I->getType();
1568 if (!isTypeLegal(RetTy, DstVT))
1571 unsigned Op = getRegForValue(I->getOperand(0));
1572 if (Op == 0) return false;
1575 Type *OpTy = I->getOperand(0)->getType();
1576 if (OpTy->isFloatTy()) Opc = ARM::VTOSIZS;
1577 else if (OpTy->isDoubleTy()) Opc = ARM::VTOSIZD;
1580 // f64->s32 or f32->s32 both need an intermediate f32 reg.
1581 unsigned ResultReg = createResultReg(TLI.getRegClassFor(MVT::f32));
1582 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
1586 // This result needs to be in an integer register, but the conversion only
1587 // takes place in fp-regs.
1588 unsigned IntReg = ARMMoveToIntReg(DstVT, ResultReg);
1589 if (IntReg == 0) return false;
1591 UpdateValueMap(I, IntReg);
1595 bool ARMFastISel::SelectSelect(const Instruction *I) {
1597 if (!isTypeLegal(I->getType(), VT))
1600 // Things need to be register sized for register moves.
1601 if (VT != MVT::i32) return false;
1602 const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
1604 unsigned CondReg = getRegForValue(I->getOperand(0));
1605 if (CondReg == 0) return false;
1606 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1607 if (Op1Reg == 0) return false;
1609 // Check to see if we can use an immediate in the conditional move.
1611 bool UseImm = false;
1612 bool isNegativeImm = false;
1613 if (const ConstantInt *ConstInt = dyn_cast<ConstantInt>(I->getOperand(2))) {
1614 assert (VT == MVT::i32 && "Expecting an i32.");
1615 Imm = (int)ConstInt->getValue().getZExtValue();
1617 isNegativeImm = true;
1620 UseImm = isThumb2 ? (ARM_AM::getT2SOImmVal(Imm) != -1) :
1621 (ARM_AM::getSOImmVal(Imm) != -1);
1624 unsigned Op2Reg = 0;
1626 Op2Reg = getRegForValue(I->getOperand(2));
1627 if (Op2Reg == 0) return false;
1630 unsigned CmpOpc = isThumb2 ? ARM::t2CMPri : ARM::CMPri;
1631 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
1632 .addReg(CondReg).addImm(0));
1636 MovCCOpc = isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr;
1638 if (!isNegativeImm) {
1639 MovCCOpc = isThumb2 ? ARM::t2MOVCCi : ARM::MOVCCi;
1641 MovCCOpc = isThumb2 ? ARM::t2MVNCCi : ARM::MVNCCi;
1644 unsigned ResultReg = createResultReg(RC);
1646 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), ResultReg)
1647 .addReg(Op2Reg).addReg(Op1Reg).addImm(ARMCC::NE).addReg(ARM::CPSR);
1649 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), ResultReg)
1650 .addReg(Op1Reg).addImm(Imm).addImm(ARMCC::EQ).addReg(ARM::CPSR);
1651 UpdateValueMap(I, ResultReg);
1655 bool ARMFastISel::SelectSDiv(const Instruction *I) {
1657 Type *Ty = I->getType();
1658 if (!isTypeLegal(Ty, VT))
1661 // If we have integer div support we should have selected this automagically.
1662 // In case we have a real miss go ahead and return false and we'll pick
1664 if (Subtarget->hasDivide()) return false;
1666 // Otherwise emit a libcall.
1667 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
1669 LC = RTLIB::SDIV_I8;
1670 else if (VT == MVT::i16)
1671 LC = RTLIB::SDIV_I16;
1672 else if (VT == MVT::i32)
1673 LC = RTLIB::SDIV_I32;
1674 else if (VT == MVT::i64)
1675 LC = RTLIB::SDIV_I64;
1676 else if (VT == MVT::i128)
1677 LC = RTLIB::SDIV_I128;
1678 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SDIV!");
1680 return ARMEmitLibcall(I, LC);
1683 bool ARMFastISel::SelectSRem(const Instruction *I) {
1685 Type *Ty = I->getType();
1686 if (!isTypeLegal(Ty, VT))
1689 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
1691 LC = RTLIB::SREM_I8;
1692 else if (VT == MVT::i16)
1693 LC = RTLIB::SREM_I16;
1694 else if (VT == MVT::i32)
1695 LC = RTLIB::SREM_I32;
1696 else if (VT == MVT::i64)
1697 LC = RTLIB::SREM_I64;
1698 else if (VT == MVT::i128)
1699 LC = RTLIB::SREM_I128;
1700 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SREM!");
1702 return ARMEmitLibcall(I, LC);
1705 bool ARMFastISel::SelectBinaryOp(const Instruction *I, unsigned ISDOpcode) {
1706 EVT VT = TLI.getValueType(I->getType(), true);
1708 // We can get here in the case when we want to use NEON for our fp
1709 // operations, but can't figure out how to. Just use the vfp instructions
1711 // FIXME: It'd be nice to use NEON instructions.
1712 Type *Ty = I->getType();
1713 bool isFloat = (Ty->isDoubleTy() || Ty->isFloatTy());
1714 if (isFloat && !Subtarget->hasVFP2())
1718 bool is64bit = VT == MVT::f64 || VT == MVT::i64;
1719 switch (ISDOpcode) {
1720 default: return false;
1722 Opc = is64bit ? ARM::VADDD : ARM::VADDS;
1725 Opc = is64bit ? ARM::VSUBD : ARM::VSUBS;
1728 Opc = is64bit ? ARM::VMULD : ARM::VMULS;
1731 unsigned Op1 = getRegForValue(I->getOperand(0));
1732 if (Op1 == 0) return false;
1734 unsigned Op2 = getRegForValue(I->getOperand(1));
1735 if (Op2 == 0) return false;
1737 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
1738 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1739 TII.get(Opc), ResultReg)
1740 .addReg(Op1).addReg(Op2));
1741 UpdateValueMap(I, ResultReg);
1745 // Call Handling Code
1747 // This is largely taken directly from CCAssignFnForNode - we don't support
1748 // varargs in FastISel so that part has been removed.
1749 // TODO: We may not support all of this.
1750 CCAssignFn *ARMFastISel::CCAssignFnForCall(CallingConv::ID CC, bool Return) {
1753 llvm_unreachable("Unsupported calling convention");
1754 case CallingConv::Fast:
1755 // Ignore fastcc. Silence compiler warnings.
1756 (void)RetFastCC_ARM_APCS;
1757 (void)FastCC_ARM_APCS;
1759 case CallingConv::C:
1760 // Use target triple & subtarget features to do actual dispatch.
1761 if (Subtarget->isAAPCS_ABI()) {
1762 if (Subtarget->hasVFP2() &&
1763 TM.Options.FloatABIType == FloatABI::Hard)
1764 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
1766 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
1768 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
1769 case CallingConv::ARM_AAPCS_VFP:
1770 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
1771 case CallingConv::ARM_AAPCS:
1772 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
1773 case CallingConv::ARM_APCS:
1774 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
1778 bool ARMFastISel::ProcessCallArgs(SmallVectorImpl<Value*> &Args,
1779 SmallVectorImpl<unsigned> &ArgRegs,
1780 SmallVectorImpl<MVT> &ArgVTs,
1781 SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
1782 SmallVectorImpl<unsigned> &RegArgs,
1784 unsigned &NumBytes) {
1785 SmallVector<CCValAssign, 16> ArgLocs;
1786 CCState CCInfo(CC, false, *FuncInfo.MF, TM, ArgLocs, *Context);
1787 CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CCAssignFnForCall(CC, false));
1789 // Get a count of how many bytes are to be pushed on the stack.
1790 NumBytes = CCInfo.getNextStackOffset();
1792 // Issue CALLSEQ_START
1793 unsigned AdjStackDown = TII.getCallFrameSetupOpcode();
1794 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1795 TII.get(AdjStackDown))
1798 // Process the args.
1799 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1800 CCValAssign &VA = ArgLocs[i];
1801 unsigned Arg = ArgRegs[VA.getValNo()];
1802 MVT ArgVT = ArgVTs[VA.getValNo()];
1804 // We don't handle NEON/vector parameters yet.
1805 if (ArgVT.isVector() || ArgVT.getSizeInBits() > 64)
1808 // Handle arg promotion, etc.
1809 switch (VA.getLocInfo()) {
1810 case CCValAssign::Full: break;
1811 case CCValAssign::SExt: {
1812 MVT DestVT = VA.getLocVT();
1813 unsigned ResultReg = ARMEmitIntExt(ArgVT, Arg, DestVT,
1815 assert (ResultReg != 0 && "Failed to emit a sext");
1820 case CCValAssign::AExt:
1821 // Intentional fall-through. Handle AExt and ZExt.
1822 case CCValAssign::ZExt: {
1823 MVT DestVT = VA.getLocVT();
1824 unsigned ResultReg = ARMEmitIntExt(ArgVT, Arg, DestVT,
1826 assert (ResultReg != 0 && "Failed to emit a sext");
1831 case CCValAssign::BCvt: {
1832 unsigned BC = FastEmit_r(ArgVT, VA.getLocVT(), ISD::BITCAST, Arg,
1833 /*TODO: Kill=*/false);
1834 assert(BC != 0 && "Failed to emit a bitcast!");
1836 ArgVT = VA.getLocVT();
1839 default: llvm_unreachable("Unknown arg promotion!");
1842 // Now copy/store arg to correct locations.
1843 if (VA.isRegLoc() && !VA.needsCustom()) {
1844 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1847 RegArgs.push_back(VA.getLocReg());
1848 } else if (VA.needsCustom()) {
1849 // TODO: We need custom lowering for vector (v2f64) args.
1850 if (VA.getLocVT() != MVT::f64) return false;
1852 CCValAssign &NextVA = ArgLocs[++i];
1854 // TODO: Only handle register args for now.
1855 if(!(VA.isRegLoc() && NextVA.isRegLoc())) return false;
1857 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1858 TII.get(ARM::VMOVRRD), VA.getLocReg())
1859 .addReg(NextVA.getLocReg(), RegState::Define)
1861 RegArgs.push_back(VA.getLocReg());
1862 RegArgs.push_back(NextVA.getLocReg());
1864 assert(VA.isMemLoc());
1865 // Need to store on the stack.
1867 Addr.BaseType = Address::RegBase;
1868 Addr.Base.Reg = ARM::SP;
1869 Addr.Offset = VA.getLocMemOffset();
1871 if (!ARMEmitStore(ArgVT, Arg, Addr)) return false;
1877 bool ARMFastISel::FinishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
1878 const Instruction *I, CallingConv::ID CC,
1879 unsigned &NumBytes) {
1880 // Issue CALLSEQ_END
1881 unsigned AdjStackUp = TII.getCallFrameDestroyOpcode();
1882 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1883 TII.get(AdjStackUp))
1884 .addImm(NumBytes).addImm(0));
1886 // Now the return value.
1887 if (RetVT != MVT::isVoid) {
1888 SmallVector<CCValAssign, 16> RVLocs;
1889 CCState CCInfo(CC, false, *FuncInfo.MF, TM, RVLocs, *Context);
1890 CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC, true));
1892 // Copy all of the result registers out of their specified physreg.
1893 if (RVLocs.size() == 2 && RetVT == MVT::f64) {
1894 // For this move we copy into two registers and then move into the
1895 // double fp reg we want.
1896 EVT DestVT = RVLocs[0].getValVT();
1897 TargetRegisterClass* DstRC = TLI.getRegClassFor(DestVT);
1898 unsigned ResultReg = createResultReg(DstRC);
1899 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1900 TII.get(ARM::VMOVDRR), ResultReg)
1901 .addReg(RVLocs[0].getLocReg())
1902 .addReg(RVLocs[1].getLocReg()));
1904 UsedRegs.push_back(RVLocs[0].getLocReg());
1905 UsedRegs.push_back(RVLocs[1].getLocReg());
1907 // Finally update the result.
1908 UpdateValueMap(I, ResultReg);
1910 assert(RVLocs.size() == 1 &&"Can't handle non-double multi-reg retvals!");
1911 EVT CopyVT = RVLocs[0].getValVT();
1913 // Special handling for extended integers.
1914 if (RetVT == MVT::i1 || RetVT == MVT::i8 || RetVT == MVT::i16)
1917 TargetRegisterClass* DstRC = TLI.getRegClassFor(CopyVT);
1919 unsigned ResultReg = createResultReg(DstRC);
1920 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1921 ResultReg).addReg(RVLocs[0].getLocReg());
1922 UsedRegs.push_back(RVLocs[0].getLocReg());
1924 // Finally update the result.
1925 UpdateValueMap(I, ResultReg);
1932 bool ARMFastISel::SelectRet(const Instruction *I) {
1933 const ReturnInst *Ret = cast<ReturnInst>(I);
1934 const Function &F = *I->getParent()->getParent();
1936 if (!FuncInfo.CanLowerReturn)
1942 CallingConv::ID CC = F.getCallingConv();
1943 if (Ret->getNumOperands() > 0) {
1944 SmallVector<ISD::OutputArg, 4> Outs;
1945 GetReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(),
1948 // Analyze operands of the call, assigning locations to each operand.
1949 SmallVector<CCValAssign, 16> ValLocs;
1950 CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, TM, ValLocs,I->getContext());
1951 CCInfo.AnalyzeReturn(Outs, CCAssignFnForCall(CC, true /* is Ret */));
1953 const Value *RV = Ret->getOperand(0);
1954 unsigned Reg = getRegForValue(RV);
1958 // Only handle a single return value for now.
1959 if (ValLocs.size() != 1)
1962 CCValAssign &VA = ValLocs[0];
1964 // Don't bother handling odd stuff for now.
1965 if (VA.getLocInfo() != CCValAssign::Full)
1967 // Only handle register returns for now.
1971 unsigned SrcReg = Reg + VA.getValNo();
1972 EVT RVVT = TLI.getValueType(RV->getType());
1973 EVT DestVT = VA.getValVT();
1974 // Special handling for extended integers.
1975 if (RVVT != DestVT) {
1976 if (RVVT != MVT::i1 && RVVT != MVT::i8 && RVVT != MVT::i16)
1979 if (!Outs[0].Flags.isZExt() && !Outs[0].Flags.isSExt())
1982 assert(DestVT == MVT::i32 && "ARM should always ext to i32");
1984 bool isZExt = Outs[0].Flags.isZExt();
1985 unsigned ResultReg = ARMEmitIntExt(RVVT, SrcReg, DestVT, isZExt);
1986 if (ResultReg == 0) return false;
1991 unsigned DstReg = VA.getLocReg();
1992 const TargetRegisterClass* SrcRC = MRI.getRegClass(SrcReg);
1993 // Avoid a cross-class copy. This is very unlikely.
1994 if (!SrcRC->contains(DstReg))
1996 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1997 DstReg).addReg(SrcReg);
1999 // Mark the register as live out of the function.
2000 MRI.addLiveOut(VA.getLocReg());
2003 unsigned RetOpc = isThumb2 ? ARM::tBX_RET : ARM::BX_RET;
2004 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
2009 unsigned ARMFastISel::ARMSelectCallOp(const GlobalValue *GV) {
2011 // Darwin needs the r9 versions of the opcodes.
2012 bool isDarwin = Subtarget->isTargetDarwin();
2014 return isDarwin ? ARM::tBLr9 : ARM::tBL;
2016 return isDarwin ? ARM::BLr9 : ARM::BL;
2020 // A quick function that will emit a call for a named libcall in F with the
2021 // vector of passed arguments for the Instruction in I. We can assume that we
2022 // can emit a call for any libcall we can produce. This is an abridged version
2023 // of the full call infrastructure since we won't need to worry about things
2024 // like computed function pointers or strange arguments at call sites.
2025 // TODO: Try to unify this and the normal call bits for ARM, then try to unify
2027 bool ARMFastISel::ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call) {
2028 CallingConv::ID CC = TLI.getLibcallCallingConv(Call);
2030 // Handle *simple* calls for now.
2031 Type *RetTy = I->getType();
2033 if (RetTy->isVoidTy())
2034 RetVT = MVT::isVoid;
2035 else if (!isTypeLegal(RetTy, RetVT))
2038 // TODO: For now if we have long calls specified we don't handle the call.
2039 if (EnableARMLongCalls) return false;
2041 // Set up the argument vectors.
2042 SmallVector<Value*, 8> Args;
2043 SmallVector<unsigned, 8> ArgRegs;
2044 SmallVector<MVT, 8> ArgVTs;
2045 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
2046 Args.reserve(I->getNumOperands());
2047 ArgRegs.reserve(I->getNumOperands());
2048 ArgVTs.reserve(I->getNumOperands());
2049 ArgFlags.reserve(I->getNumOperands());
2050 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
2051 Value *Op = I->getOperand(i);
2052 unsigned Arg = getRegForValue(Op);
2053 if (Arg == 0) return false;
2055 Type *ArgTy = Op->getType();
2057 if (!isTypeLegal(ArgTy, ArgVT)) return false;
2059 ISD::ArgFlagsTy Flags;
2060 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
2061 Flags.setOrigAlign(OriginalAlignment);
2064 ArgRegs.push_back(Arg);
2065 ArgVTs.push_back(ArgVT);
2066 ArgFlags.push_back(Flags);
2069 // Handle the arguments now that we've gotten them.
2070 SmallVector<unsigned, 4> RegArgs;
2072 if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
2075 // Issue the call, BLr9 for darwin, BL otherwise.
2076 // TODO: Turn this into the table of arm call ops.
2077 MachineInstrBuilder MIB;
2078 unsigned CallOpc = ARMSelectCallOp(NULL);
2080 // Explicitly adding the predicate here.
2081 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
2083 .addExternalSymbol(TLI.getLibcallName(Call));
2085 // Explicitly adding the predicate here.
2086 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
2088 .addExternalSymbol(TLI.getLibcallName(Call)));
2090 // Add implicit physical register uses to the call.
2091 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
2092 MIB.addReg(RegArgs[i]);
2094 // Finish off the call including any return values.
2095 SmallVector<unsigned, 4> UsedRegs;
2096 if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false;
2098 // Set all unused physreg defs as dead.
2099 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
2104 bool ARMFastISel::SelectCall(const Instruction *I,
2105 const char *IntrMemName = 0) {
2106 const CallInst *CI = cast<CallInst>(I);
2107 const Value *Callee = CI->getCalledValue();
2109 // Can't handle inline asm.
2110 if (isa<InlineAsm>(Callee)) return false;
2112 // Only handle global variable Callees.
2113 const GlobalValue *GV = dyn_cast<GlobalValue>(Callee);
2117 // Check the calling convention.
2118 ImmutableCallSite CS(CI);
2119 CallingConv::ID CC = CS.getCallingConv();
2121 // TODO: Avoid some calling conventions?
2123 // Let SDISel handle vararg functions.
2124 PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
2125 FunctionType *FTy = cast<FunctionType>(PT->getElementType());
2126 if (FTy->isVarArg())
2129 // Handle *simple* calls for now.
2130 Type *RetTy = I->getType();
2132 if (RetTy->isVoidTy())
2133 RetVT = MVT::isVoid;
2134 else if (!isTypeLegal(RetTy, RetVT) && RetVT != MVT::i16 &&
2135 RetVT != MVT::i8 && RetVT != MVT::i1)
2138 // TODO: For now if we have long calls specified we don't handle the call.
2139 if (EnableARMLongCalls) return false;
2141 // Set up the argument vectors.
2142 SmallVector<Value*, 8> Args;
2143 SmallVector<unsigned, 8> ArgRegs;
2144 SmallVector<MVT, 8> ArgVTs;
2145 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
2146 Args.reserve(CS.arg_size());
2147 ArgRegs.reserve(CS.arg_size());
2148 ArgVTs.reserve(CS.arg_size());
2149 ArgFlags.reserve(CS.arg_size());
2150 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
2152 // If we're lowering a memory intrinsic instead of a regular call, skip the
2153 // last two arguments, which shouldn't be passed to the underlying function.
2154 if (IntrMemName && e-i <= 2)
2157 ISD::ArgFlagsTy Flags;
2158 unsigned AttrInd = i - CS.arg_begin() + 1;
2159 if (CS.paramHasAttr(AttrInd, Attribute::SExt))
2161 if (CS.paramHasAttr(AttrInd, Attribute::ZExt))
2164 // FIXME: Only handle *easy* calls for now.
2165 if (CS.paramHasAttr(AttrInd, Attribute::InReg) ||
2166 CS.paramHasAttr(AttrInd, Attribute::StructRet) ||
2167 CS.paramHasAttr(AttrInd, Attribute::Nest) ||
2168 CS.paramHasAttr(AttrInd, Attribute::ByVal))
2171 Type *ArgTy = (*i)->getType();
2173 if (!isTypeLegal(ArgTy, ArgVT) && ArgVT != MVT::i16 && ArgVT != MVT::i8 &&
2177 unsigned Arg = getRegForValue(*i);
2181 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
2182 Flags.setOrigAlign(OriginalAlignment);
2185 ArgRegs.push_back(Arg);
2186 ArgVTs.push_back(ArgVT);
2187 ArgFlags.push_back(Flags);
2190 // Handle the arguments now that we've gotten them.
2191 SmallVector<unsigned, 4> RegArgs;
2193 if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
2196 // Issue the call, BLr9 for darwin, BL otherwise.
2197 // TODO: Turn this into the table of arm call ops.
2198 MachineInstrBuilder MIB;
2199 unsigned CallOpc = ARMSelectCallOp(GV);
2200 // Explicitly adding the predicate here.
2202 // Explicitly adding the predicate here.
2203 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
2206 MIB.addGlobalAddress(GV, 0, 0);
2208 MIB.addExternalSymbol(IntrMemName, 0);
2211 // Explicitly adding the predicate here.
2212 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
2214 .addGlobalAddress(GV, 0, 0));
2216 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
2218 .addExternalSymbol(IntrMemName, 0));
2221 // Add implicit physical register uses to the call.
2222 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
2223 MIB.addReg(RegArgs[i]);
2225 // Finish off the call including any return values.
2226 SmallVector<unsigned, 4> UsedRegs;
2227 if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false;
2229 // Set all unused physreg defs as dead.
2230 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
2235 bool ARMFastISel::ARMIsMemCpySmall(uint64_t Len) {
2239 bool ARMFastISel::ARMTryEmitSmallMemCpy(Address Dest, Address Src, uint64_t Len) {
2240 // Make sure we don't bloat code by inlining very large memcpy's.
2241 if (!ARMIsMemCpySmall(Len))
2244 // We don't care about alignment here since we just emit integer accesses.
2258 RV = ARMEmitLoad(VT, ResultReg, Src);
2259 assert (RV = true && "Should be able to handle this load.");
2260 RV = ARMEmitStore(VT, ResultReg, Dest);
2261 assert (RV = true && "Should be able to handle this store.");
2263 unsigned Size = VT.getSizeInBits()/8;
2265 Dest.Offset += Size;
2272 bool ARMFastISel::SelectIntrinsicCall(const IntrinsicInst &I) {
2273 // FIXME: Handle more intrinsics.
2274 switch (I.getIntrinsicID()) {
2275 default: return false;
2276 case Intrinsic::memcpy:
2277 case Intrinsic::memmove: {
2278 const MemTransferInst &MTI = cast<MemTransferInst>(I);
2279 // Don't handle volatile.
2280 if (MTI.isVolatile())
2283 // Disable inlining for memmove before calls to ComputeAddress. Otherwise,
2284 // we would emit dead code because we don't currently handle memmoves.
2285 bool isMemCpy = (I.getIntrinsicID() == Intrinsic::memcpy);
2286 if (isa<ConstantInt>(MTI.getLength()) && isMemCpy) {
2287 // Small memcpy's are common enough that we want to do them without a call
2289 uint64_t Len = cast<ConstantInt>(MTI.getLength())->getZExtValue();
2290 if (ARMIsMemCpySmall(Len)) {
2292 if (!ARMComputeAddress(MTI.getRawDest(), Dest) ||
2293 !ARMComputeAddress(MTI.getRawSource(), Src))
2295 if (ARMTryEmitSmallMemCpy(Dest, Src, Len))
2300 if (!MTI.getLength()->getType()->isIntegerTy(32))
2303 if (MTI.getSourceAddressSpace() > 255 || MTI.getDestAddressSpace() > 255)
2306 const char *IntrMemName = isa<MemCpyInst>(I) ? "memcpy" : "memmove";
2307 return SelectCall(&I, IntrMemName);
2309 case Intrinsic::memset: {
2310 const MemSetInst &MSI = cast<MemSetInst>(I);
2311 // Don't handle volatile.
2312 if (MSI.isVolatile())
2315 if (!MSI.getLength()->getType()->isIntegerTy(32))
2318 if (MSI.getDestAddressSpace() > 255)
2321 return SelectCall(&I, "memset");
2327 bool ARMFastISel::SelectTrunc(const Instruction *I) {
2328 // The high bits for a type smaller than the register size are assumed to be
2330 Value *Op = I->getOperand(0);
2333 SrcVT = TLI.getValueType(Op->getType(), true);
2334 DestVT = TLI.getValueType(I->getType(), true);
2336 if (SrcVT != MVT::i32 && SrcVT != MVT::i16 && SrcVT != MVT::i8)
2338 if (DestVT != MVT::i16 && DestVT != MVT::i8 && DestVT != MVT::i1)
2341 unsigned SrcReg = getRegForValue(Op);
2342 if (!SrcReg) return false;
2344 // Because the high bits are undefined, a truncate doesn't generate
2346 UpdateValueMap(I, SrcReg);
2350 unsigned ARMFastISel::ARMEmitIntExt(EVT SrcVT, unsigned SrcReg, EVT DestVT,
2352 if (DestVT != MVT::i32 && DestVT != MVT::i16 && DestVT != MVT::i8)
2356 bool isBoolZext = false;
2357 if (!SrcVT.isSimple()) return 0;
2358 switch (SrcVT.getSimpleVT().SimpleTy) {
2361 if (!Subtarget->hasV6Ops()) return 0;
2363 Opc = isThumb2 ? ARM::t2UXTH : ARM::UXTH;
2365 Opc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
2368 if (!Subtarget->hasV6Ops()) return 0;
2370 Opc = isThumb2 ? ARM::t2UXTB : ARM::UXTB;
2372 Opc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
2376 Opc = isThumb2 ? ARM::t2ANDri : ARM::ANDri;
2383 unsigned ResultReg = createResultReg(TLI.getRegClassFor(MVT::i32));
2384 MachineInstrBuilder MIB;
2385 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), ResultReg)
2391 AddOptionalDefs(MIB);
2395 bool ARMFastISel::SelectIntExt(const Instruction *I) {
2396 // On ARM, in general, integer casts don't involve legal types; this code
2397 // handles promotable integers.
2398 Type *DestTy = I->getType();
2399 Value *Src = I->getOperand(0);
2400 Type *SrcTy = Src->getType();
2403 SrcVT = TLI.getValueType(SrcTy, true);
2404 DestVT = TLI.getValueType(DestTy, true);
2406 bool isZExt = isa<ZExtInst>(I);
2407 unsigned SrcReg = getRegForValue(Src);
2408 if (!SrcReg) return false;
2410 unsigned ResultReg = ARMEmitIntExt(SrcVT, SrcReg, DestVT, isZExt);
2411 if (ResultReg == 0) return false;
2412 UpdateValueMap(I, ResultReg);
2416 // TODO: SoftFP support.
2417 bool ARMFastISel::TargetSelectInstruction(const Instruction *I) {
2419 switch (I->getOpcode()) {
2420 case Instruction::Load:
2421 return SelectLoad(I);
2422 case Instruction::Store:
2423 return SelectStore(I);
2424 case Instruction::Br:
2425 return SelectBranch(I);
2426 case Instruction::ICmp:
2427 case Instruction::FCmp:
2428 return SelectCmp(I);
2429 case Instruction::FPExt:
2430 return SelectFPExt(I);
2431 case Instruction::FPTrunc:
2432 return SelectFPTrunc(I);
2433 case Instruction::SIToFP:
2434 return SelectSIToFP(I);
2435 case Instruction::FPToSI:
2436 return SelectFPToSI(I);
2437 case Instruction::FAdd:
2438 return SelectBinaryOp(I, ISD::FADD);
2439 case Instruction::FSub:
2440 return SelectBinaryOp(I, ISD::FSUB);
2441 case Instruction::FMul:
2442 return SelectBinaryOp(I, ISD::FMUL);
2443 case Instruction::SDiv:
2444 return SelectSDiv(I);
2445 case Instruction::SRem:
2446 return SelectSRem(I);
2447 case Instruction::Call:
2448 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2449 return SelectIntrinsicCall(*II);
2450 return SelectCall(I);
2451 case Instruction::Select:
2452 return SelectSelect(I);
2453 case Instruction::Ret:
2454 return SelectRet(I);
2455 case Instruction::Trunc:
2456 return SelectTrunc(I);
2457 case Instruction::ZExt:
2458 case Instruction::SExt:
2459 return SelectIntExt(I);
2465 /// TryToFoldLoad - The specified machine instr operand is a vreg, and that
2466 /// vreg is being provided by the specified load instruction. If possible,
2467 /// try to fold the load as an operand to the instruction, returning true if
2469 bool ARMFastISel::TryToFoldLoad(MachineInstr *MI, unsigned OpNo,
2470 const LoadInst *LI) {
2471 // Verify we have a legal type before going any further.
2473 if (!isLoadTypeLegal(LI->getType(), VT))
2476 // Combine load followed by zero- or sign-extend.
2477 // ldrb r1, [r0] ldrb r1, [r0]
2479 // mov r3, r2 mov r3, r1
2481 switch(MI->getOpcode()) {
2482 default: return false;
2500 // See if we can handle this address.
2502 if (!ARMComputeAddress(LI->getOperand(0), Addr)) return false;
2504 unsigned ResultReg = MI->getOperand(0).getReg();
2505 if (!ARMEmitLoad(VT, ResultReg, Addr, LI->getAlignment(), isZExt, false))
2507 MI->eraseFromParent();
2512 llvm::FastISel *ARM::createFastISel(FunctionLoweringInfo &funcInfo) {
2513 // Completely untested on non-darwin.
2514 const TargetMachine &TM = funcInfo.MF->getTarget();
2516 // Darwin and thumb1 only for now.
2517 const ARMSubtarget *Subtarget = &TM.getSubtarget<ARMSubtarget>();
2518 if (Subtarget->isTargetDarwin() && !Subtarget->isThumb1Only() &&
2519 !DisableARMFastISel)
2520 return new ARMFastISel(funcInfo);