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/CodeGen/PseudoSourceValue.h"
41 #include "llvm/Support/CallSite.h"
42 #include "llvm/Support/CommandLine.h"
43 #include "llvm/Support/ErrorHandling.h"
44 #include "llvm/Support/GetElementPtrTypeIterator.h"
45 #include "llvm/Target/TargetData.h"
46 #include "llvm/Target/TargetInstrInfo.h"
47 #include "llvm/Target/TargetLowering.h"
48 #include "llvm/Target/TargetMachine.h"
49 #include "llvm/Target/TargetOptions.h"
53 DisableARMFastISel("disable-arm-fast-isel",
54 cl::desc("Turn off experimental ARM fast-isel support"),
55 cl::init(false), cl::Hidden);
57 extern cl::opt<bool> EnableARMLongCalls;
61 // All possible address modes, plus some.
62 typedef struct Address {
75 // Innocuous defaults for our address.
77 : BaseType(RegBase), Offset(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 isThumb2 = 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_rrr(unsigned MachineInstOpcode,
119 const TargetRegisterClass *RC,
120 unsigned Op0, bool Op0IsKill,
121 unsigned Op1, bool Op1IsKill,
122 unsigned Op2, bool Op2IsKill);
123 virtual unsigned FastEmitInst_ri(unsigned MachineInstOpcode,
124 const TargetRegisterClass *RC,
125 unsigned Op0, bool Op0IsKill,
127 virtual unsigned FastEmitInst_rf(unsigned MachineInstOpcode,
128 const TargetRegisterClass *RC,
129 unsigned Op0, bool Op0IsKill,
130 const ConstantFP *FPImm);
131 virtual unsigned FastEmitInst_rri(unsigned MachineInstOpcode,
132 const TargetRegisterClass *RC,
133 unsigned Op0, bool Op0IsKill,
134 unsigned Op1, bool Op1IsKill,
136 virtual unsigned FastEmitInst_i(unsigned MachineInstOpcode,
137 const TargetRegisterClass *RC,
139 virtual unsigned FastEmitInst_ii(unsigned MachineInstOpcode,
140 const TargetRegisterClass *RC,
141 uint64_t Imm1, uint64_t Imm2);
143 virtual unsigned FastEmitInst_extractsubreg(MVT RetVT,
144 unsigned Op0, bool Op0IsKill,
147 // Backend specific FastISel code.
148 virtual bool TargetSelectInstruction(const Instruction *I);
149 virtual unsigned TargetMaterializeConstant(const Constant *C);
150 virtual unsigned TargetMaterializeAlloca(const AllocaInst *AI);
152 #include "ARMGenFastISel.inc"
154 // Instruction selection routines.
156 bool SelectLoad(const Instruction *I);
157 bool SelectStore(const Instruction *I);
158 bool SelectBranch(const Instruction *I);
159 bool SelectCmp(const Instruction *I);
160 bool SelectFPExt(const Instruction *I);
161 bool SelectFPTrunc(const Instruction *I);
162 bool SelectBinaryOp(const Instruction *I, unsigned ISDOpcode);
163 bool SelectSIToFP(const Instruction *I);
164 bool SelectFPToSI(const Instruction *I);
165 bool SelectSDiv(const Instruction *I);
166 bool SelectSRem(const Instruction *I);
167 bool SelectCall(const Instruction *I);
168 bool SelectSelect(const Instruction *I);
169 bool SelectRet(const Instruction *I);
170 bool SelectTrunc(const Instruction *I);
171 bool SelectIntExt(const Instruction *I);
175 bool isTypeLegal(Type *Ty, MVT &VT);
176 bool isLoadTypeLegal(Type *Ty, MVT &VT);
177 bool ARMEmitCmp(const Value *Src1Value, const Value *Src2Value,
179 bool ARMEmitLoad(EVT VT, unsigned &ResultReg, Address &Addr);
180 bool ARMEmitStore(EVT VT, unsigned SrcReg, Address &Addr);
181 bool ARMComputeAddress(const Value *Obj, Address &Addr);
182 void ARMSimplifyAddress(Address &Addr, EVT VT);
183 unsigned ARMEmitIntExt(EVT SrcVT, unsigned SrcReg, EVT DestVT, bool isZExt);
184 unsigned ARMMaterializeFP(const ConstantFP *CFP, EVT VT);
185 unsigned ARMMaterializeInt(const Constant *C, EVT VT);
186 unsigned ARMMaterializeGV(const GlobalValue *GV, EVT VT);
187 unsigned ARMMoveToFPReg(EVT VT, unsigned SrcReg);
188 unsigned ARMMoveToIntReg(EVT VT, unsigned SrcReg);
189 unsigned ARMSelectCallOp(const GlobalValue *GV);
191 // Call handling routines.
193 bool FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src, EVT SrcVT,
194 unsigned &ResultReg);
195 CCAssignFn *CCAssignFnForCall(CallingConv::ID CC, bool Return);
196 bool ProcessCallArgs(SmallVectorImpl<Value*> &Args,
197 SmallVectorImpl<unsigned> &ArgRegs,
198 SmallVectorImpl<MVT> &ArgVTs,
199 SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
200 SmallVectorImpl<unsigned> &RegArgs,
203 bool FinishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
204 const Instruction *I, CallingConv::ID CC,
206 bool ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call);
208 // OptionalDef handling routines.
210 bool isARMNEONPred(const MachineInstr *MI);
211 bool DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR);
212 const MachineInstrBuilder &AddOptionalDefs(const MachineInstrBuilder &MIB);
213 void AddLoadStoreOperands(EVT VT, Address &Addr,
214 const MachineInstrBuilder &MIB,
218 } // end anonymous namespace
220 #include "ARMGenCallingConv.inc"
222 // DefinesOptionalPredicate - This is different from DefinesPredicate in that
223 // we don't care about implicit defs here, just places we'll need to add a
224 // default CCReg argument. Sets CPSR if we're setting CPSR instead of CCR.
225 bool ARMFastISel::DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR) {
226 const MCInstrDesc &MCID = MI->getDesc();
227 if (!MCID.hasOptionalDef())
230 // Look to see if our OptionalDef is defining CPSR or CCR.
231 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
232 const MachineOperand &MO = MI->getOperand(i);
233 if (!MO.isReg() || !MO.isDef()) continue;
234 if (MO.getReg() == ARM::CPSR)
240 bool ARMFastISel::isARMNEONPred(const MachineInstr *MI) {
241 const MCInstrDesc &MCID = MI->getDesc();
243 // If we're a thumb2 or not NEON function we were handled via isPredicable.
244 if ((MCID.TSFlags & ARMII::DomainMask) != ARMII::DomainNEON ||
245 AFI->isThumb2Function())
248 for (unsigned i = 0, e = MCID.getNumOperands(); i != e; ++i)
249 if (MCID.OpInfo[i].isPredicate())
255 // If the machine is predicable go ahead and add the predicate operands, if
256 // it needs default CC operands add those.
257 // TODO: If we want to support thumb1 then we'll need to deal with optional
258 // CPSR defs that need to be added before the remaining operands. See s_cc_out
259 // for descriptions why.
260 const MachineInstrBuilder &
261 ARMFastISel::AddOptionalDefs(const MachineInstrBuilder &MIB) {
262 MachineInstr *MI = &*MIB;
264 // Do we use a predicate? or...
265 // Are we NEON in ARM mode and have a predicate operand? If so, I know
266 // we're not predicable but add it anyways.
267 if (TII.isPredicable(MI) || isARMNEONPred(MI))
270 // Do we optionally set a predicate? Preds is size > 0 iff the predicate
271 // defines CPSR. All other OptionalDefines in ARM are the CCR register.
273 if (DefinesOptionalPredicate(MI, &CPSR)) {
282 unsigned ARMFastISel::FastEmitInst_(unsigned MachineInstOpcode,
283 const TargetRegisterClass* RC) {
284 unsigned ResultReg = createResultReg(RC);
285 const MCInstrDesc &II = TII.get(MachineInstOpcode);
287 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg));
291 unsigned ARMFastISel::FastEmitInst_r(unsigned MachineInstOpcode,
292 const TargetRegisterClass *RC,
293 unsigned Op0, bool Op0IsKill) {
294 unsigned ResultReg = createResultReg(RC);
295 const MCInstrDesc &II = TII.get(MachineInstOpcode);
297 if (II.getNumDefs() >= 1)
298 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
299 .addReg(Op0, Op0IsKill * RegState::Kill));
301 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
302 .addReg(Op0, Op0IsKill * RegState::Kill));
303 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
304 TII.get(TargetOpcode::COPY), ResultReg)
305 .addReg(II.ImplicitDefs[0]));
310 unsigned ARMFastISel::FastEmitInst_rr(unsigned MachineInstOpcode,
311 const TargetRegisterClass *RC,
312 unsigned Op0, bool Op0IsKill,
313 unsigned Op1, bool Op1IsKill) {
314 unsigned ResultReg = createResultReg(RC);
315 const MCInstrDesc &II = TII.get(MachineInstOpcode);
317 if (II.getNumDefs() >= 1)
318 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
319 .addReg(Op0, Op0IsKill * RegState::Kill)
320 .addReg(Op1, Op1IsKill * RegState::Kill));
322 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
323 .addReg(Op0, Op0IsKill * RegState::Kill)
324 .addReg(Op1, Op1IsKill * RegState::Kill));
325 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
326 TII.get(TargetOpcode::COPY), ResultReg)
327 .addReg(II.ImplicitDefs[0]));
332 unsigned ARMFastISel::FastEmitInst_rrr(unsigned MachineInstOpcode,
333 const TargetRegisterClass *RC,
334 unsigned Op0, bool Op0IsKill,
335 unsigned Op1, bool Op1IsKill,
336 unsigned Op2, bool Op2IsKill) {
337 unsigned ResultReg = createResultReg(RC);
338 const MCInstrDesc &II = TII.get(MachineInstOpcode);
340 if (II.getNumDefs() >= 1)
341 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
342 .addReg(Op0, Op0IsKill * RegState::Kill)
343 .addReg(Op1, Op1IsKill * RegState::Kill)
344 .addReg(Op2, Op2IsKill * RegState::Kill));
346 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
347 .addReg(Op0, Op0IsKill * RegState::Kill)
348 .addReg(Op1, Op1IsKill * RegState::Kill)
349 .addReg(Op2, Op2IsKill * RegState::Kill));
350 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
351 TII.get(TargetOpcode::COPY), ResultReg)
352 .addReg(II.ImplicitDefs[0]));
357 unsigned ARMFastISel::FastEmitInst_ri(unsigned MachineInstOpcode,
358 const TargetRegisterClass *RC,
359 unsigned Op0, bool Op0IsKill,
361 unsigned ResultReg = createResultReg(RC);
362 const MCInstrDesc &II = TII.get(MachineInstOpcode);
364 if (II.getNumDefs() >= 1)
365 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
366 .addReg(Op0, Op0IsKill * RegState::Kill)
369 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
370 .addReg(Op0, Op0IsKill * RegState::Kill)
372 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
373 TII.get(TargetOpcode::COPY), ResultReg)
374 .addReg(II.ImplicitDefs[0]));
379 unsigned ARMFastISel::FastEmitInst_rf(unsigned MachineInstOpcode,
380 const TargetRegisterClass *RC,
381 unsigned Op0, bool Op0IsKill,
382 const ConstantFP *FPImm) {
383 unsigned ResultReg = createResultReg(RC);
384 const MCInstrDesc &II = TII.get(MachineInstOpcode);
386 if (II.getNumDefs() >= 1)
387 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
388 .addReg(Op0, Op0IsKill * RegState::Kill)
391 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
392 .addReg(Op0, Op0IsKill * RegState::Kill)
394 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
395 TII.get(TargetOpcode::COPY), ResultReg)
396 .addReg(II.ImplicitDefs[0]));
401 unsigned ARMFastISel::FastEmitInst_rri(unsigned MachineInstOpcode,
402 const TargetRegisterClass *RC,
403 unsigned Op0, bool Op0IsKill,
404 unsigned Op1, bool Op1IsKill,
406 unsigned ResultReg = createResultReg(RC);
407 const MCInstrDesc &II = TII.get(MachineInstOpcode);
409 if (II.getNumDefs() >= 1)
410 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
411 .addReg(Op0, Op0IsKill * RegState::Kill)
412 .addReg(Op1, Op1IsKill * RegState::Kill)
415 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
416 .addReg(Op0, Op0IsKill * RegState::Kill)
417 .addReg(Op1, Op1IsKill * RegState::Kill)
419 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
420 TII.get(TargetOpcode::COPY), ResultReg)
421 .addReg(II.ImplicitDefs[0]));
426 unsigned ARMFastISel::FastEmitInst_i(unsigned MachineInstOpcode,
427 const TargetRegisterClass *RC,
429 unsigned ResultReg = createResultReg(RC);
430 const MCInstrDesc &II = TII.get(MachineInstOpcode);
432 if (II.getNumDefs() >= 1)
433 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
436 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
438 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
439 TII.get(TargetOpcode::COPY), ResultReg)
440 .addReg(II.ImplicitDefs[0]));
445 unsigned ARMFastISel::FastEmitInst_ii(unsigned MachineInstOpcode,
446 const TargetRegisterClass *RC,
447 uint64_t Imm1, uint64_t Imm2) {
448 unsigned ResultReg = createResultReg(RC);
449 const MCInstrDesc &II = TII.get(MachineInstOpcode);
451 if (II.getNumDefs() >= 1)
452 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
453 .addImm(Imm1).addImm(Imm2));
455 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
456 .addImm(Imm1).addImm(Imm2));
457 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
458 TII.get(TargetOpcode::COPY),
460 .addReg(II.ImplicitDefs[0]));
465 unsigned ARMFastISel::FastEmitInst_extractsubreg(MVT RetVT,
466 unsigned Op0, bool Op0IsKill,
468 unsigned ResultReg = createResultReg(TLI.getRegClassFor(RetVT));
469 assert(TargetRegisterInfo::isVirtualRegister(Op0) &&
470 "Cannot yet extract from physregs");
471 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
472 DL, TII.get(TargetOpcode::COPY), ResultReg)
473 .addReg(Op0, getKillRegState(Op0IsKill), Idx));
477 // TODO: Don't worry about 64-bit now, but when this is fixed remove the
478 // checks from the various callers.
479 unsigned ARMFastISel::ARMMoveToFPReg(EVT VT, unsigned SrcReg) {
480 if (VT == MVT::f64) return 0;
482 unsigned MoveReg = createResultReg(TLI.getRegClassFor(VT));
483 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
484 TII.get(ARM::VMOVRS), MoveReg)
489 unsigned ARMFastISel::ARMMoveToIntReg(EVT VT, unsigned SrcReg) {
490 if (VT == MVT::i64) return 0;
492 unsigned MoveReg = createResultReg(TLI.getRegClassFor(VT));
493 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
494 TII.get(ARM::VMOVSR), MoveReg)
499 // For double width floating point we need to materialize two constants
500 // (the high and the low) into integer registers then use a move to get
501 // the combined constant into an FP reg.
502 unsigned ARMFastISel::ARMMaterializeFP(const ConstantFP *CFP, EVT VT) {
503 const APFloat Val = CFP->getValueAPF();
504 bool is64bit = VT == MVT::f64;
506 // This checks to see if we can use VFP3 instructions to materialize
507 // a constant, otherwise we have to go through the constant pool.
508 if (TLI.isFPImmLegal(Val, VT)) {
512 Imm = ARM_AM::getFP64Imm(Val);
515 Imm = ARM_AM::getFP32Imm(Val);
518 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
519 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
525 // Require VFP2 for loading fp constants.
526 if (!Subtarget->hasVFP2()) return false;
528 // MachineConstantPool wants an explicit alignment.
529 unsigned Align = TD.getPrefTypeAlignment(CFP->getType());
531 // TODO: Figure out if this is correct.
532 Align = TD.getTypeAllocSize(CFP->getType());
534 unsigned Idx = MCP.getConstantPoolIndex(cast<Constant>(CFP), Align);
535 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
536 unsigned Opc = is64bit ? ARM::VLDRD : ARM::VLDRS;
538 // The extra reg is for addrmode5.
539 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
541 .addConstantPoolIndex(Idx)
546 unsigned ARMFastISel::ARMMaterializeInt(const Constant *C, EVT VT) {
548 if (VT != MVT::i32 && VT != MVT::i16 && VT != MVT::i8 && VT != MVT::i1)
551 // If we can do this in a single instruction without a constant pool entry
553 const ConstantInt *CI = cast<ConstantInt>(C);
554 if (Subtarget->hasV6T2Ops() && isUInt<16>(CI->getZExtValue())) {
555 EVT SrcVT = MVT::i32;
556 unsigned Opc = isThumb2 ? ARM::t2MOVi16 : ARM::MOVi16;
557 unsigned ImmReg = createResultReg(TLI.getRegClassFor(SrcVT));
558 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
559 TII.get(Opc), ImmReg)
560 .addImm(CI->getZExtValue()));
564 // For now 32-bit only.
568 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
570 // MachineConstantPool wants an explicit alignment.
571 unsigned Align = TD.getPrefTypeAlignment(C->getType());
573 // TODO: Figure out if this is correct.
574 Align = TD.getTypeAllocSize(C->getType());
576 unsigned Idx = MCP.getConstantPoolIndex(C, Align);
579 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
580 TII.get(ARM::t2LDRpci), DestReg)
581 .addConstantPoolIndex(Idx));
583 // The extra immediate is for addrmode2.
584 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
585 TII.get(ARM::LDRcp), DestReg)
586 .addConstantPoolIndex(Idx)
592 unsigned ARMFastISel::ARMMaterializeGV(const GlobalValue *GV, EVT VT) {
593 // For now 32-bit only.
594 if (VT != MVT::i32) return 0;
596 Reloc::Model RelocM = TM.getRelocationModel();
598 // TODO: Need more magic for ARM PIC.
599 if (!isThumb2 && (RelocM == Reloc::PIC_)) return 0;
601 // MachineConstantPool wants an explicit alignment.
602 unsigned Align = TD.getPrefTypeAlignment(GV->getType());
604 // TODO: Figure out if this is correct.
605 Align = TD.getTypeAllocSize(GV->getType());
609 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb() ? 4 : 8);
610 unsigned Id = AFI->createPICLabelUId();
611 ARMConstantPoolValue *CPV = ARMConstantPoolConstant::Create(GV, Id,
614 unsigned Idx = MCP.getConstantPoolIndex(CPV, Align);
617 MachineInstrBuilder MIB;
618 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
620 unsigned Opc = (RelocM != Reloc::PIC_) ? ARM::t2LDRpci : ARM::t2LDRpci_pic;
621 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), DestReg)
622 .addConstantPoolIndex(Idx);
623 if (RelocM == Reloc::PIC_)
626 // The extra immediate is for addrmode2.
627 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(ARM::LDRcp),
629 .addConstantPoolIndex(Idx)
632 AddOptionalDefs(MIB);
634 if (Subtarget->GVIsIndirectSymbol(GV, RelocM)) {
635 unsigned NewDestReg = createResultReg(TLI.getRegClassFor(VT));
637 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
638 TII.get(ARM::t2LDRi12), NewDestReg)
642 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(ARM::LDRi12),
646 DestReg = NewDestReg;
647 AddOptionalDefs(MIB);
653 unsigned ARMFastISel::TargetMaterializeConstant(const Constant *C) {
654 EVT VT = TLI.getValueType(C->getType(), true);
656 // Only handle simple types.
657 if (!VT.isSimple()) return 0;
659 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
660 return ARMMaterializeFP(CFP, VT);
661 else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
662 return ARMMaterializeGV(GV, VT);
663 else if (isa<ConstantInt>(C))
664 return ARMMaterializeInt(C, VT);
669 unsigned ARMFastISel::TargetMaterializeAlloca(const AllocaInst *AI) {
670 // Don't handle dynamic allocas.
671 if (!FuncInfo.StaticAllocaMap.count(AI)) return 0;
674 if (!isLoadTypeLegal(AI->getType(), VT)) return false;
676 DenseMap<const AllocaInst*, int>::iterator SI =
677 FuncInfo.StaticAllocaMap.find(AI);
679 // This will get lowered later into the correct offsets and registers
680 // via rewriteXFrameIndex.
681 if (SI != FuncInfo.StaticAllocaMap.end()) {
682 TargetRegisterClass* RC = TLI.getRegClassFor(VT);
683 unsigned ResultReg = createResultReg(RC);
684 unsigned Opc = isThumb2 ? ARM::t2ADDri : ARM::ADDri;
685 AddOptionalDefs(BuildMI(*FuncInfo.MBB, *FuncInfo.InsertPt, DL,
686 TII.get(Opc), ResultReg)
687 .addFrameIndex(SI->second)
695 bool ARMFastISel::isTypeLegal(Type *Ty, MVT &VT) {
696 EVT evt = TLI.getValueType(Ty, true);
698 // Only handle simple types.
699 if (evt == MVT::Other || !evt.isSimple()) return false;
700 VT = evt.getSimpleVT();
702 // Handle all legal types, i.e. a register that will directly hold this
704 return TLI.isTypeLegal(VT);
707 bool ARMFastISel::isLoadTypeLegal(Type *Ty, MVT &VT) {
708 if (isTypeLegal(Ty, VT)) return true;
710 // If this is a type than can be sign or zero-extended to a basic operation
711 // go ahead and accept it now.
712 if (VT == MVT::i8 || VT == MVT::i16)
718 // Computes the address to get to an object.
719 bool ARMFastISel::ARMComputeAddress(const Value *Obj, Address &Addr) {
720 // Some boilerplate from the X86 FastISel.
721 const User *U = NULL;
722 unsigned Opcode = Instruction::UserOp1;
723 if (const Instruction *I = dyn_cast<Instruction>(Obj)) {
724 // Don't walk into other basic blocks unless the object is an alloca from
725 // another block, otherwise it may not have a virtual register assigned.
726 if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(Obj)) ||
727 FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
728 Opcode = I->getOpcode();
731 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Obj)) {
732 Opcode = C->getOpcode();
736 if (PointerType *Ty = dyn_cast<PointerType>(Obj->getType()))
737 if (Ty->getAddressSpace() > 255)
738 // Fast instruction selection doesn't support the special
745 case Instruction::BitCast: {
746 // Look through bitcasts.
747 return ARMComputeAddress(U->getOperand(0), Addr);
749 case Instruction::IntToPtr: {
750 // Look past no-op inttoptrs.
751 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
752 return ARMComputeAddress(U->getOperand(0), Addr);
755 case Instruction::PtrToInt: {
756 // Look past no-op ptrtoints.
757 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
758 return ARMComputeAddress(U->getOperand(0), Addr);
761 case Instruction::GetElementPtr: {
762 Address SavedAddr = Addr;
763 int TmpOffset = Addr.Offset;
765 // Iterate through the GEP folding the constants into offsets where
767 gep_type_iterator GTI = gep_type_begin(U);
768 for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end();
769 i != e; ++i, ++GTI) {
770 const Value *Op = *i;
771 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
772 const StructLayout *SL = TD.getStructLayout(STy);
773 unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
774 TmpOffset += SL->getElementOffset(Idx);
776 uint64_t S = TD.getTypeAllocSize(GTI.getIndexedType());
778 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
779 // Constant-offset addressing.
780 TmpOffset += CI->getSExtValue() * S;
783 if (isa<AddOperator>(Op) &&
784 (!isa<Instruction>(Op) ||
785 FuncInfo.MBBMap[cast<Instruction>(Op)->getParent()]
787 isa<ConstantInt>(cast<AddOperator>(Op)->getOperand(1))) {
788 // An add (in the same block) with a constant operand. Fold the
791 cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
792 TmpOffset += CI->getSExtValue() * S;
793 // Iterate on the other operand.
794 Op = cast<AddOperator>(Op)->getOperand(0);
798 goto unsupported_gep;
803 // Try to grab the base operand now.
804 Addr.Offset = TmpOffset;
805 if (ARMComputeAddress(U->getOperand(0), Addr)) return true;
807 // We failed, restore everything and try the other options.
813 case Instruction::Alloca: {
814 const AllocaInst *AI = cast<AllocaInst>(Obj);
815 DenseMap<const AllocaInst*, int>::iterator SI =
816 FuncInfo.StaticAllocaMap.find(AI);
817 if (SI != FuncInfo.StaticAllocaMap.end()) {
818 Addr.BaseType = Address::FrameIndexBase;
819 Addr.Base.FI = SI->second;
826 // Materialize the global variable's address into a reg which can
827 // then be used later to load the variable.
828 if (const GlobalValue *GV = dyn_cast<GlobalValue>(Obj)) {
829 unsigned Tmp = ARMMaterializeGV(GV, TLI.getValueType(Obj->getType()));
830 if (Tmp == 0) return false;
836 // Try to get this in a register if nothing else has worked.
837 if (Addr.Base.Reg == 0) Addr.Base.Reg = getRegForValue(Obj);
838 return Addr.Base.Reg != 0;
841 void ARMFastISel::ARMSimplifyAddress(Address &Addr, EVT VT) {
843 assert(VT.isSimple() && "Non-simple types are invalid here!");
845 bool needsLowering = false;
846 switch (VT.getSimpleVT().SimpleTy) {
848 assert(false && "Unhandled load/store type!");
851 // Integer loads/stores handle 12-bit offsets.
852 needsLowering = ((Addr.Offset & 0xfff) != Addr.Offset);
854 // ARM i16 integer loads/stores handle +/-imm8 offsets.
855 // FIXME: Negative offsets require special handling.
856 if (Addr.Offset > 255 || Addr.Offset < 0)
857 needsLowering = true;
862 // Integer loads/stores handle 12-bit offsets.
863 needsLowering = ((Addr.Offset & 0xfff) != Addr.Offset);
867 // Floating point operands handle 8-bit offsets.
868 needsLowering = ((Addr.Offset & 0xff) != Addr.Offset);
872 // If this is a stack pointer and the offset needs to be simplified then
873 // put the alloca address into a register, set the base type back to
874 // register and continue. This should almost never happen.
875 if (needsLowering && Addr.BaseType == Address::FrameIndexBase) {
876 TargetRegisterClass *RC = isThumb2 ? ARM::tGPRRegisterClass :
877 ARM::GPRRegisterClass;
878 unsigned ResultReg = createResultReg(RC);
879 unsigned Opc = isThumb2 ? ARM::t2ADDri : ARM::ADDri;
880 AddOptionalDefs(BuildMI(*FuncInfo.MBB, *FuncInfo.InsertPt, DL,
881 TII.get(Opc), ResultReg)
882 .addFrameIndex(Addr.Base.FI)
884 Addr.Base.Reg = ResultReg;
885 Addr.BaseType = Address::RegBase;
888 // Since the offset is too large for the load/store instruction
889 // get the reg+offset into a register.
891 Addr.Base.Reg = FastEmit_ri_(MVT::i32, ISD::ADD, Addr.Base.Reg,
892 /*Op0IsKill*/false, Addr.Offset, MVT::i32);
897 void ARMFastISel::AddLoadStoreOperands(EVT VT, Address &Addr,
898 const MachineInstrBuilder &MIB,
900 // addrmode5 output depends on the selection dag addressing dividing the
901 // offset by 4 that it then later multiplies. Do this here as well.
902 if (VT.getSimpleVT().SimpleTy == MVT::f32 ||
903 VT.getSimpleVT().SimpleTy == MVT::f64)
906 // Frame base works a bit differently. Handle it separately.
907 if (Addr.BaseType == Address::FrameIndexBase) {
908 int FI = Addr.Base.FI;
909 int Offset = Addr.Offset;
910 MachineMemOperand *MMO =
911 FuncInfo.MF->getMachineMemOperand(
912 MachinePointerInfo::getFixedStack(FI, Offset),
914 MFI.getObjectSize(FI),
915 MFI.getObjectAlignment(FI));
916 // Now add the rest of the operands.
917 MIB.addFrameIndex(FI);
919 // ARM halfword load/stores need an additional operand.
920 if (!isThumb2 && VT.getSimpleVT().SimpleTy == MVT::i16) MIB.addReg(0);
922 MIB.addImm(Addr.Offset);
923 MIB.addMemOperand(MMO);
925 // Now add the rest of the operands.
926 MIB.addReg(Addr.Base.Reg);
928 // ARM halfword load/stores need an additional operand.
929 if (!isThumb2 && VT.getSimpleVT().SimpleTy == MVT::i16) MIB.addReg(0);
931 MIB.addImm(Addr.Offset);
933 AddOptionalDefs(MIB);
936 bool ARMFastISel::ARMEmitLoad(EVT VT, unsigned &ResultReg, Address &Addr) {
938 assert(VT.isSimple() && "Non-simple types are invalid here!");
940 TargetRegisterClass *RC;
941 switch (VT.getSimpleVT().SimpleTy) {
942 // This is mostly going to be Neon/vector support.
943 default: return false;
945 Opc = isThumb2 ? ARM::t2LDRBi12 : ARM::LDRBi12;
946 RC = ARM::GPRRegisterClass;
949 Opc = isThumb2 ? ARM::t2LDRHi12 : ARM::LDRH;
950 RC = ARM::GPRRegisterClass;
953 Opc = isThumb2 ? ARM::t2LDRi12 : ARM::LDRi12;
954 RC = ARM::GPRRegisterClass;
958 RC = TLI.getRegClassFor(VT);
962 RC = TLI.getRegClassFor(VT);
965 // Simplify this down to something we can handle.
966 ARMSimplifyAddress(Addr, VT);
968 // Create the base instruction, then add the operands.
969 ResultReg = createResultReg(RC);
970 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
971 TII.get(Opc), ResultReg);
972 AddLoadStoreOperands(VT, Addr, MIB, MachineMemOperand::MOLoad);
976 bool ARMFastISel::SelectLoad(const Instruction *I) {
977 // Atomic loads need special handling.
978 if (cast<LoadInst>(I)->isAtomic())
981 // Verify we have a legal type before going any further.
983 if (!isLoadTypeLegal(I->getType(), VT))
986 // See if we can handle this address.
988 if (!ARMComputeAddress(I->getOperand(0), Addr)) return false;
991 if (!ARMEmitLoad(VT, ResultReg, Addr)) return false;
992 UpdateValueMap(I, ResultReg);
996 bool ARMFastISel::ARMEmitStore(EVT VT, unsigned SrcReg, Address &Addr) {
998 switch (VT.getSimpleVT().SimpleTy) {
999 // This is mostly going to be Neon/vector support.
1000 default: return false;
1002 unsigned Res = createResultReg(isThumb2 ? ARM::tGPRRegisterClass :
1003 ARM::GPRRegisterClass);
1004 unsigned Opc = isThumb2 ? ARM::t2ANDri : ARM::ANDri;
1005 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1007 .addReg(SrcReg).addImm(1));
1009 } // Fallthrough here.
1011 StrOpc = isThumb2 ? ARM::t2STRBi12 : ARM::STRBi12;
1014 StrOpc = isThumb2 ? ARM::t2STRHi12 : ARM::STRH;
1017 StrOpc = isThumb2 ? ARM::t2STRi12 : ARM::STRi12;
1020 if (!Subtarget->hasVFP2()) return false;
1021 StrOpc = ARM::VSTRS;
1024 if (!Subtarget->hasVFP2()) return false;
1025 StrOpc = ARM::VSTRD;
1028 // Simplify this down to something we can handle.
1029 ARMSimplifyAddress(Addr, VT);
1031 // Create the base instruction, then add the operands.
1032 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1034 .addReg(SrcReg, getKillRegState(true));
1035 AddLoadStoreOperands(VT, Addr, MIB, MachineMemOperand::MOStore);
1039 bool ARMFastISel::SelectStore(const Instruction *I) {
1040 Value *Op0 = I->getOperand(0);
1041 unsigned SrcReg = 0;
1043 // Atomic stores need special handling.
1044 if (cast<StoreInst>(I)->isAtomic())
1047 // Verify we have a legal type before going any further.
1049 if (!isLoadTypeLegal(I->getOperand(0)->getType(), VT))
1052 // Get the value to be stored into a register.
1053 SrcReg = getRegForValue(Op0);
1054 if (SrcReg == 0) return false;
1056 // See if we can handle this address.
1058 if (!ARMComputeAddress(I->getOperand(1), Addr))
1061 if (!ARMEmitStore(VT, SrcReg, Addr)) return false;
1065 static ARMCC::CondCodes getComparePred(CmpInst::Predicate Pred) {
1067 // Needs two compares...
1068 case CmpInst::FCMP_ONE:
1069 case CmpInst::FCMP_UEQ:
1071 // AL is our "false" for now. The other two need more compares.
1073 case CmpInst::ICMP_EQ:
1074 case CmpInst::FCMP_OEQ:
1076 case CmpInst::ICMP_SGT:
1077 case CmpInst::FCMP_OGT:
1079 case CmpInst::ICMP_SGE:
1080 case CmpInst::FCMP_OGE:
1082 case CmpInst::ICMP_UGT:
1083 case CmpInst::FCMP_UGT:
1085 case CmpInst::FCMP_OLT:
1087 case CmpInst::ICMP_ULE:
1088 case CmpInst::FCMP_OLE:
1090 case CmpInst::FCMP_ORD:
1092 case CmpInst::FCMP_UNO:
1094 case CmpInst::FCMP_UGE:
1096 case CmpInst::ICMP_SLT:
1097 case CmpInst::FCMP_ULT:
1099 case CmpInst::ICMP_SLE:
1100 case CmpInst::FCMP_ULE:
1102 case CmpInst::FCMP_UNE:
1103 case CmpInst::ICMP_NE:
1105 case CmpInst::ICMP_UGE:
1107 case CmpInst::ICMP_ULT:
1112 bool ARMFastISel::SelectBranch(const Instruction *I) {
1113 const BranchInst *BI = cast<BranchInst>(I);
1114 MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
1115 MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
1117 // Simple branch support.
1119 // If we can, avoid recomputing the compare - redoing it could lead to wonky
1121 if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
1122 if (CI->hasOneUse() && (CI->getParent() == I->getParent())) {
1124 // Get the compare predicate.
1125 // Try to take advantage of fallthrough opportunities.
1126 CmpInst::Predicate Predicate = CI->getPredicate();
1127 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
1128 std::swap(TBB, FBB);
1129 Predicate = CmpInst::getInversePredicate(Predicate);
1132 ARMCC::CondCodes ARMPred = getComparePred(Predicate);
1134 // We may not handle every CC for now.
1135 if (ARMPred == ARMCC::AL) return false;
1137 // Emit the compare.
1138 if (!ARMEmitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
1141 unsigned BrOpc = isThumb2 ? ARM::t2Bcc : ARM::Bcc;
1142 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
1143 .addMBB(TBB).addImm(ARMPred).addReg(ARM::CPSR);
1144 FastEmitBranch(FBB, DL);
1145 FuncInfo.MBB->addSuccessor(TBB);
1148 } else if (TruncInst *TI = dyn_cast<TruncInst>(BI->getCondition())) {
1150 if (TI->hasOneUse() && TI->getParent() == I->getParent() &&
1151 (isLoadTypeLegal(TI->getOperand(0)->getType(), SourceVT))) {
1152 unsigned TstOpc = isThumb2 ? ARM::t2TSTri : ARM::TSTri;
1153 unsigned OpReg = getRegForValue(TI->getOperand(0));
1154 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1156 .addReg(OpReg).addImm(1));
1158 unsigned CCMode = ARMCC::NE;
1159 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
1160 std::swap(TBB, FBB);
1164 unsigned BrOpc = isThumb2 ? ARM::t2Bcc : ARM::Bcc;
1165 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
1166 .addMBB(TBB).addImm(CCMode).addReg(ARM::CPSR);
1168 FastEmitBranch(FBB, DL);
1169 FuncInfo.MBB->addSuccessor(TBB);
1172 } else if (const ConstantInt *CI =
1173 dyn_cast<ConstantInt>(BI->getCondition())) {
1174 uint64_t Imm = CI->getZExtValue();
1175 MachineBasicBlock *Target = (Imm == 0) ? FBB : TBB;
1176 FastEmitBranch(Target, DL);
1180 unsigned CmpReg = getRegForValue(BI->getCondition());
1181 if (CmpReg == 0) return false;
1183 // We've been divorced from our compare! Our block was split, and
1184 // now our compare lives in a predecessor block. We musn't
1185 // re-compare here, as the children of the compare aren't guaranteed
1186 // live across the block boundary (we *could* check for this).
1187 // Regardless, the compare has been done in the predecessor block,
1188 // and it left a value for us in a virtual register. Ergo, we test
1189 // the one-bit value left in the virtual register.
1190 unsigned TstOpc = isThumb2 ? ARM::t2TSTri : ARM::TSTri;
1191 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TstOpc))
1192 .addReg(CmpReg).addImm(1));
1194 unsigned CCMode = ARMCC::NE;
1195 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
1196 std::swap(TBB, FBB);
1200 unsigned BrOpc = isThumb2 ? ARM::t2Bcc : ARM::Bcc;
1201 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
1202 .addMBB(TBB).addImm(CCMode).addReg(ARM::CPSR);
1203 FastEmitBranch(FBB, DL);
1204 FuncInfo.MBB->addSuccessor(TBB);
1208 bool ARMFastISel::ARMEmitCmp(const Value *Src1Value, const Value *Src2Value,
1210 Type *Ty = Src1Value->getType();
1211 EVT SrcVT = TLI.getValueType(Ty, true);
1212 if (!SrcVT.isSimple()) return false;
1214 bool isFloat = (Ty->isFloatTy() || Ty->isDoubleTy());
1215 if (isFloat && !Subtarget->hasVFP2())
1218 // Check to see if the 2nd operand is a constant that we can encode directly
1221 bool EncodeImm = false;
1222 bool isNegativeImm = false;
1223 if (const ConstantInt *ConstInt = dyn_cast<ConstantInt>(Src2Value)) {
1224 if (SrcVT == MVT::i32 || SrcVT == MVT::i16 || SrcVT == MVT::i8 ||
1226 const APInt &CIVal = ConstInt->getValue();
1227 EncodedImm = (isZExt) ? (int)CIVal.getZExtValue() : (int)CIVal.getSExtValue();
1228 if (EncodedImm < 0) {
1229 isNegativeImm = true;
1230 EncodedImm = -EncodedImm;
1232 EncodeImm = isThumb2 ? (ARM_AM::getT2SOImmVal(EncodedImm) != -1) :
1233 (ARM_AM::getSOImmVal(EncodedImm) != -1);
1235 } else if (const ConstantFP *ConstFP = dyn_cast<ConstantFP>(Src2Value)) {
1236 if (SrcVT == MVT::f32 || SrcVT == MVT::f64)
1237 if (ConstFP->isZero() && !ConstFP->isNegative())
1243 bool needsExt = false;
1244 switch (SrcVT.getSimpleVT().SimpleTy) {
1245 default: return false;
1246 // TODO: Verify compares.
1249 CmpOpc = EncodeImm ? ARM::VCMPEZS : ARM::VCMPES;
1253 CmpOpc = EncodeImm ? ARM::VCMPEZD : ARM::VCMPED;
1259 // Intentional fall-through.
1263 CmpOpc = ARM::t2CMPrr;
1265 CmpOpc = isNegativeImm ? ARM::t2CMNzri : ARM::t2CMPri;
1268 CmpOpc = ARM::CMPrr;
1270 CmpOpc = isNegativeImm ? ARM::CMNzri : ARM::CMPri;
1275 unsigned SrcReg1 = getRegForValue(Src1Value);
1276 if (SrcReg1 == 0) return false;
1280 SrcReg2 = getRegForValue(Src2Value);
1281 if (SrcReg2 == 0) return false;
1284 // We have i1, i8, or i16, we need to either zero extend or sign extend.
1287 ResultReg = ARMEmitIntExt(SrcVT, SrcReg1, MVT::i32, isZExt);
1288 if (ResultReg == 0) return false;
1289 SrcReg1 = ResultReg;
1291 ResultReg = ARMEmitIntExt(SrcVT, SrcReg2, MVT::i32, isZExt);
1292 if (ResultReg == 0) return false;
1293 SrcReg2 = ResultReg;
1298 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1300 .addReg(SrcReg1).addReg(SrcReg2));
1302 MachineInstrBuilder MIB;
1303 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
1306 // Only add immediate for icmp as the immediate for fcmp is an implicit 0.0.
1308 MIB.addImm(EncodedImm);
1309 AddOptionalDefs(MIB);
1312 // For floating point we need to move the result to a comparison register
1313 // that we can then use for branches.
1314 if (Ty->isFloatTy() || Ty->isDoubleTy())
1315 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1316 TII.get(ARM::FMSTAT)));
1320 bool ARMFastISel::SelectCmp(const Instruction *I) {
1321 const CmpInst *CI = cast<CmpInst>(I);
1322 Type *Ty = CI->getOperand(0)->getType();
1324 // Get the compare predicate.
1325 ARMCC::CondCodes ARMPred = getComparePred(CI->getPredicate());
1327 // We may not handle every CC for now.
1328 if (ARMPred == ARMCC::AL) return false;
1330 // Emit the compare.
1331 if (!ARMEmitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
1334 // Now set a register based on the comparison. Explicitly set the predicates
1336 unsigned MovCCOpc = isThumb2 ? ARM::t2MOVCCi : ARM::MOVCCi;
1337 TargetRegisterClass *RC = isThumb2 ? ARM::rGPRRegisterClass
1338 : ARM::GPRRegisterClass;
1339 unsigned DestReg = createResultReg(RC);
1340 Constant *Zero = ConstantInt::get(Type::getInt32Ty(*Context), 0);
1341 unsigned ZeroReg = TargetMaterializeConstant(Zero);
1342 bool isFloat = (Ty->isFloatTy() || Ty->isDoubleTy());
1343 unsigned CondReg = isFloat ? ARM::FPSCR : ARM::CPSR;
1344 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), DestReg)
1345 .addReg(ZeroReg).addImm(1)
1346 .addImm(ARMPred).addReg(CondReg);
1348 UpdateValueMap(I, DestReg);
1352 bool ARMFastISel::SelectFPExt(const Instruction *I) {
1353 // Make sure we have VFP and that we're extending float to double.
1354 if (!Subtarget->hasVFP2()) return false;
1356 Value *V = I->getOperand(0);
1357 if (!I->getType()->isDoubleTy() ||
1358 !V->getType()->isFloatTy()) return false;
1360 unsigned Op = getRegForValue(V);
1361 if (Op == 0) return false;
1363 unsigned Result = createResultReg(ARM::DPRRegisterClass);
1364 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1365 TII.get(ARM::VCVTDS), Result)
1367 UpdateValueMap(I, Result);
1371 bool ARMFastISel::SelectFPTrunc(const Instruction *I) {
1372 // Make sure we have VFP and that we're truncating double to float.
1373 if (!Subtarget->hasVFP2()) return false;
1375 Value *V = I->getOperand(0);
1376 if (!(I->getType()->isFloatTy() &&
1377 V->getType()->isDoubleTy())) return false;
1379 unsigned Op = getRegForValue(V);
1380 if (Op == 0) return false;
1382 unsigned Result = createResultReg(ARM::SPRRegisterClass);
1383 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1384 TII.get(ARM::VCVTSD), Result)
1386 UpdateValueMap(I, Result);
1390 bool ARMFastISel::SelectSIToFP(const Instruction *I) {
1391 // Make sure we have VFP.
1392 if (!Subtarget->hasVFP2()) return false;
1395 Type *Ty = I->getType();
1396 if (!isTypeLegal(Ty, DstVT))
1399 Value *Src = I->getOperand(0);
1400 EVT SrcVT = TLI.getValueType(Src->getType(), true);
1401 if (SrcVT != MVT::i32 && SrcVT != MVT::i16 && SrcVT != MVT::i8)
1404 unsigned SrcReg = getRegForValue(Src);
1405 if (SrcReg == 0) return false;
1407 // Handle sign-extension.
1408 if (SrcVT == MVT::i16 || SrcVT == MVT::i8) {
1409 EVT DestVT = MVT::i32;
1410 unsigned ResultReg = ARMEmitIntExt(SrcVT, SrcReg, DestVT, /*isZExt*/ false);
1411 if (ResultReg == 0) return false;
1415 // The conversion routine works on fp-reg to fp-reg and the operand above
1416 // was an integer, move it to the fp registers if possible.
1417 unsigned FP = ARMMoveToFPReg(MVT::f32, SrcReg);
1418 if (FP == 0) return false;
1421 if (Ty->isFloatTy()) Opc = ARM::VSITOS;
1422 else if (Ty->isDoubleTy()) Opc = ARM::VSITOD;
1425 unsigned ResultReg = createResultReg(TLI.getRegClassFor(DstVT));
1426 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
1429 UpdateValueMap(I, ResultReg);
1433 bool ARMFastISel::SelectFPToSI(const Instruction *I) {
1434 // Make sure we have VFP.
1435 if (!Subtarget->hasVFP2()) return false;
1438 Type *RetTy = I->getType();
1439 if (!isTypeLegal(RetTy, DstVT))
1442 unsigned Op = getRegForValue(I->getOperand(0));
1443 if (Op == 0) return false;
1446 Type *OpTy = I->getOperand(0)->getType();
1447 if (OpTy->isFloatTy()) Opc = ARM::VTOSIZS;
1448 else if (OpTy->isDoubleTy()) Opc = ARM::VTOSIZD;
1451 // f64->s32 or f32->s32 both need an intermediate f32 reg.
1452 unsigned ResultReg = createResultReg(TLI.getRegClassFor(MVT::f32));
1453 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
1457 // This result needs to be in an integer register, but the conversion only
1458 // takes place in fp-regs.
1459 unsigned IntReg = ARMMoveToIntReg(DstVT, ResultReg);
1460 if (IntReg == 0) return false;
1462 UpdateValueMap(I, IntReg);
1466 bool ARMFastISel::SelectSelect(const Instruction *I) {
1468 if (!isTypeLegal(I->getType(), VT))
1471 // Things need to be register sized for register moves.
1472 if (VT != MVT::i32) return false;
1473 const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
1475 unsigned CondReg = getRegForValue(I->getOperand(0));
1476 if (CondReg == 0) return false;
1477 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1478 if (Op1Reg == 0) return false;
1479 unsigned Op2Reg = getRegForValue(I->getOperand(2));
1480 if (Op2Reg == 0) return false;
1482 unsigned CmpOpc = isThumb2 ? ARM::t2TSTri : ARM::TSTri;
1483 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
1484 .addReg(CondReg).addImm(1));
1485 unsigned ResultReg = createResultReg(RC);
1486 unsigned MovCCOpc = isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr;
1487 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), ResultReg)
1488 .addReg(Op1Reg).addReg(Op2Reg)
1489 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
1490 UpdateValueMap(I, ResultReg);
1494 bool ARMFastISel::SelectSDiv(const Instruction *I) {
1496 Type *Ty = I->getType();
1497 if (!isTypeLegal(Ty, VT))
1500 // If we have integer div support we should have selected this automagically.
1501 // In case we have a real miss go ahead and return false and we'll pick
1503 if (Subtarget->hasDivide()) return false;
1505 // Otherwise emit a libcall.
1506 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
1508 LC = RTLIB::SDIV_I8;
1509 else if (VT == MVT::i16)
1510 LC = RTLIB::SDIV_I16;
1511 else if (VT == MVT::i32)
1512 LC = RTLIB::SDIV_I32;
1513 else if (VT == MVT::i64)
1514 LC = RTLIB::SDIV_I64;
1515 else if (VT == MVT::i128)
1516 LC = RTLIB::SDIV_I128;
1517 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SDIV!");
1519 return ARMEmitLibcall(I, LC);
1522 bool ARMFastISel::SelectSRem(const Instruction *I) {
1524 Type *Ty = I->getType();
1525 if (!isTypeLegal(Ty, VT))
1528 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
1530 LC = RTLIB::SREM_I8;
1531 else if (VT == MVT::i16)
1532 LC = RTLIB::SREM_I16;
1533 else if (VT == MVT::i32)
1534 LC = RTLIB::SREM_I32;
1535 else if (VT == MVT::i64)
1536 LC = RTLIB::SREM_I64;
1537 else if (VT == MVT::i128)
1538 LC = RTLIB::SREM_I128;
1539 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SREM!");
1541 return ARMEmitLibcall(I, LC);
1544 bool ARMFastISel::SelectBinaryOp(const Instruction *I, unsigned ISDOpcode) {
1545 EVT VT = TLI.getValueType(I->getType(), true);
1547 // We can get here in the case when we want to use NEON for our fp
1548 // operations, but can't figure out how to. Just use the vfp instructions
1550 // FIXME: It'd be nice to use NEON instructions.
1551 Type *Ty = I->getType();
1552 bool isFloat = (Ty->isDoubleTy() || Ty->isFloatTy());
1553 if (isFloat && !Subtarget->hasVFP2())
1556 unsigned Op1 = getRegForValue(I->getOperand(0));
1557 if (Op1 == 0) return false;
1559 unsigned Op2 = getRegForValue(I->getOperand(1));
1560 if (Op2 == 0) return false;
1563 bool is64bit = VT == MVT::f64 || VT == MVT::i64;
1564 switch (ISDOpcode) {
1565 default: return false;
1567 Opc = is64bit ? ARM::VADDD : ARM::VADDS;
1570 Opc = is64bit ? ARM::VSUBD : ARM::VSUBS;
1573 Opc = is64bit ? ARM::VMULD : ARM::VMULS;
1576 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
1577 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1578 TII.get(Opc), ResultReg)
1579 .addReg(Op1).addReg(Op2));
1580 UpdateValueMap(I, ResultReg);
1584 // Call Handling Code
1586 bool ARMFastISel::FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src,
1587 EVT SrcVT, unsigned &ResultReg) {
1588 unsigned RR = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), Opc,
1589 Src, /*TODO: Kill=*/false);
1598 // This is largely taken directly from CCAssignFnForNode - we don't support
1599 // varargs in FastISel so that part has been removed.
1600 // TODO: We may not support all of this.
1601 CCAssignFn *ARMFastISel::CCAssignFnForCall(CallingConv::ID CC, bool Return) {
1604 llvm_unreachable("Unsupported calling convention");
1605 case CallingConv::Fast:
1606 // Ignore fastcc. Silence compiler warnings.
1607 (void)RetFastCC_ARM_APCS;
1608 (void)FastCC_ARM_APCS;
1610 case CallingConv::C:
1611 // Use target triple & subtarget features to do actual dispatch.
1612 if (Subtarget->isAAPCS_ABI()) {
1613 if (Subtarget->hasVFP2() &&
1614 FloatABIType == FloatABI::Hard)
1615 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
1617 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
1619 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
1620 case CallingConv::ARM_AAPCS_VFP:
1621 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
1622 case CallingConv::ARM_AAPCS:
1623 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
1624 case CallingConv::ARM_APCS:
1625 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
1629 bool ARMFastISel::ProcessCallArgs(SmallVectorImpl<Value*> &Args,
1630 SmallVectorImpl<unsigned> &ArgRegs,
1631 SmallVectorImpl<MVT> &ArgVTs,
1632 SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
1633 SmallVectorImpl<unsigned> &RegArgs,
1635 unsigned &NumBytes) {
1636 SmallVector<CCValAssign, 16> ArgLocs;
1637 CCState CCInfo(CC, false, *FuncInfo.MF, TM, ArgLocs, *Context);
1638 CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CCAssignFnForCall(CC, false));
1640 // Get a count of how many bytes are to be pushed on the stack.
1641 NumBytes = CCInfo.getNextStackOffset();
1643 // Issue CALLSEQ_START
1644 unsigned AdjStackDown = TII.getCallFrameSetupOpcode();
1645 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1646 TII.get(AdjStackDown))
1649 // Process the args.
1650 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1651 CCValAssign &VA = ArgLocs[i];
1652 unsigned Arg = ArgRegs[VA.getValNo()];
1653 MVT ArgVT = ArgVTs[VA.getValNo()];
1655 // We don't handle NEON/vector parameters yet.
1656 if (ArgVT.isVector() || ArgVT.getSizeInBits() > 64)
1659 // Handle arg promotion, etc.
1660 switch (VA.getLocInfo()) {
1661 case CCValAssign::Full: break;
1662 case CCValAssign::SExt: {
1663 EVT DestVT = VA.getLocVT();
1664 unsigned ResultReg = ARMEmitIntExt(ArgVT, Arg, DestVT,
1666 assert (ResultReg != 0 && "Failed to emit a sext");
1670 case CCValAssign::AExt:
1671 // Intentional fall-through. Handle AExt and ZExt.
1672 case CCValAssign::ZExt: {
1673 EVT DestVT = VA.getLocVT();
1674 unsigned ResultReg = ARMEmitIntExt(ArgVT, Arg, DestVT,
1676 assert (ResultReg != 0 && "Failed to emit a sext");
1680 case CCValAssign::BCvt: {
1681 unsigned BC = FastEmit_r(ArgVT, VA.getLocVT(), ISD::BITCAST, Arg,
1682 /*TODO: Kill=*/false);
1683 assert(BC != 0 && "Failed to emit a bitcast!");
1685 ArgVT = VA.getLocVT();
1688 default: llvm_unreachable("Unknown arg promotion!");
1691 // Now copy/store arg to correct locations.
1692 if (VA.isRegLoc() && !VA.needsCustom()) {
1693 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1696 RegArgs.push_back(VA.getLocReg());
1697 } else if (VA.needsCustom()) {
1698 // TODO: We need custom lowering for vector (v2f64) args.
1699 if (VA.getLocVT() != MVT::f64) return false;
1701 CCValAssign &NextVA = ArgLocs[++i];
1703 // TODO: Only handle register args for now.
1704 if(!(VA.isRegLoc() && NextVA.isRegLoc())) return false;
1706 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1707 TII.get(ARM::VMOVRRD), VA.getLocReg())
1708 .addReg(NextVA.getLocReg(), RegState::Define)
1710 RegArgs.push_back(VA.getLocReg());
1711 RegArgs.push_back(NextVA.getLocReg());
1713 assert(VA.isMemLoc());
1714 // Need to store on the stack.
1716 Addr.BaseType = Address::RegBase;
1717 Addr.Base.Reg = ARM::SP;
1718 Addr.Offset = VA.getLocMemOffset();
1720 if (!ARMEmitStore(ArgVT, Arg, Addr)) return false;
1726 bool ARMFastISel::FinishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
1727 const Instruction *I, CallingConv::ID CC,
1728 unsigned &NumBytes) {
1729 // Issue CALLSEQ_END
1730 unsigned AdjStackUp = TII.getCallFrameDestroyOpcode();
1731 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1732 TII.get(AdjStackUp))
1733 .addImm(NumBytes).addImm(0));
1735 // Now the return value.
1736 if (RetVT != MVT::isVoid) {
1737 SmallVector<CCValAssign, 16> RVLocs;
1738 CCState CCInfo(CC, false, *FuncInfo.MF, TM, RVLocs, *Context);
1739 CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC, true));
1741 // Copy all of the result registers out of their specified physreg.
1742 if (RVLocs.size() == 2 && RetVT == MVT::f64) {
1743 // For this move we copy into two registers and then move into the
1744 // double fp reg we want.
1745 EVT DestVT = RVLocs[0].getValVT();
1746 TargetRegisterClass* DstRC = TLI.getRegClassFor(DestVT);
1747 unsigned ResultReg = createResultReg(DstRC);
1748 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1749 TII.get(ARM::VMOVDRR), ResultReg)
1750 .addReg(RVLocs[0].getLocReg())
1751 .addReg(RVLocs[1].getLocReg()));
1753 UsedRegs.push_back(RVLocs[0].getLocReg());
1754 UsedRegs.push_back(RVLocs[1].getLocReg());
1756 // Finally update the result.
1757 UpdateValueMap(I, ResultReg);
1759 assert(RVLocs.size() == 1 &&"Can't handle non-double multi-reg retvals!");
1760 EVT CopyVT = RVLocs[0].getValVT();
1762 // Special handling for extended integers.
1763 if (RetVT == MVT::i1 || RetVT == MVT::i8 || RetVT == MVT::i16)
1766 TargetRegisterClass* DstRC = TLI.getRegClassFor(CopyVT);
1768 unsigned ResultReg = createResultReg(DstRC);
1769 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1770 ResultReg).addReg(RVLocs[0].getLocReg());
1771 UsedRegs.push_back(RVLocs[0].getLocReg());
1773 // Finally update the result.
1774 UpdateValueMap(I, ResultReg);
1781 bool ARMFastISel::SelectRet(const Instruction *I) {
1782 const ReturnInst *Ret = cast<ReturnInst>(I);
1783 const Function &F = *I->getParent()->getParent();
1785 if (!FuncInfo.CanLowerReturn)
1791 CallingConv::ID CC = F.getCallingConv();
1792 if (Ret->getNumOperands() > 0) {
1793 SmallVector<ISD::OutputArg, 4> Outs;
1794 GetReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(),
1797 // Analyze operands of the call, assigning locations to each operand.
1798 SmallVector<CCValAssign, 16> ValLocs;
1799 CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, TM, ValLocs,I->getContext());
1800 CCInfo.AnalyzeReturn(Outs, CCAssignFnForCall(CC, true /* is Ret */));
1802 const Value *RV = Ret->getOperand(0);
1803 unsigned Reg = getRegForValue(RV);
1807 // Only handle a single return value for now.
1808 if (ValLocs.size() != 1)
1811 CCValAssign &VA = ValLocs[0];
1813 // Don't bother handling odd stuff for now.
1814 if (VA.getLocInfo() != CCValAssign::Full)
1816 // Only handle register returns for now.
1820 unsigned SrcReg = Reg + VA.getValNo();
1821 EVT RVVT = TLI.getValueType(RV->getType());
1822 EVT DestVT = VA.getValVT();
1823 // Special handling for extended integers.
1824 if (RVVT != DestVT) {
1825 if (RVVT != MVT::i1 && RVVT != MVT::i8 && RVVT != MVT::i16)
1828 if (!Outs[0].Flags.isZExt() && !Outs[0].Flags.isSExt())
1831 assert(DestVT == MVT::i32 && "ARM should always ext to i32");
1833 bool isZExt = Outs[0].Flags.isZExt();
1834 unsigned ResultReg = ARMEmitIntExt(RVVT, SrcReg, DestVT, isZExt);
1835 if (ResultReg == 0) return false;
1840 unsigned DstReg = VA.getLocReg();
1841 const TargetRegisterClass* SrcRC = MRI.getRegClass(SrcReg);
1842 // Avoid a cross-class copy. This is very unlikely.
1843 if (!SrcRC->contains(DstReg))
1845 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1846 DstReg).addReg(SrcReg);
1848 // Mark the register as live out of the function.
1849 MRI.addLiveOut(VA.getLocReg());
1852 unsigned RetOpc = isThumb2 ? ARM::tBX_RET : ARM::BX_RET;
1853 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1858 unsigned ARMFastISel::ARMSelectCallOp(const GlobalValue *GV) {
1860 // Darwin needs the r9 versions of the opcodes.
1861 bool isDarwin = Subtarget->isTargetDarwin();
1863 return isDarwin ? ARM::tBLr9 : ARM::tBL;
1865 return isDarwin ? ARM::BLr9 : ARM::BL;
1869 // A quick function that will emit a call for a named libcall in F with the
1870 // vector of passed arguments for the Instruction in I. We can assume that we
1871 // can emit a call for any libcall we can produce. This is an abridged version
1872 // of the full call infrastructure since we won't need to worry about things
1873 // like computed function pointers or strange arguments at call sites.
1874 // TODO: Try to unify this and the normal call bits for ARM, then try to unify
1876 bool ARMFastISel::ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call) {
1877 CallingConv::ID CC = TLI.getLibcallCallingConv(Call);
1879 // Handle *simple* calls for now.
1880 Type *RetTy = I->getType();
1882 if (RetTy->isVoidTy())
1883 RetVT = MVT::isVoid;
1884 else if (!isTypeLegal(RetTy, RetVT))
1887 // TODO: For now if we have long calls specified we don't handle the call.
1888 if (EnableARMLongCalls) return false;
1890 // Set up the argument vectors.
1891 SmallVector<Value*, 8> Args;
1892 SmallVector<unsigned, 8> ArgRegs;
1893 SmallVector<MVT, 8> ArgVTs;
1894 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
1895 Args.reserve(I->getNumOperands());
1896 ArgRegs.reserve(I->getNumOperands());
1897 ArgVTs.reserve(I->getNumOperands());
1898 ArgFlags.reserve(I->getNumOperands());
1899 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
1900 Value *Op = I->getOperand(i);
1901 unsigned Arg = getRegForValue(Op);
1902 if (Arg == 0) return false;
1904 Type *ArgTy = Op->getType();
1906 if (!isTypeLegal(ArgTy, ArgVT)) return false;
1908 ISD::ArgFlagsTy Flags;
1909 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
1910 Flags.setOrigAlign(OriginalAlignment);
1913 ArgRegs.push_back(Arg);
1914 ArgVTs.push_back(ArgVT);
1915 ArgFlags.push_back(Flags);
1918 // Handle the arguments now that we've gotten them.
1919 SmallVector<unsigned, 4> RegArgs;
1921 if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
1924 // Issue the call, BLr9 for darwin, BL otherwise.
1925 // TODO: Turn this into the table of arm call ops.
1926 MachineInstrBuilder MIB;
1927 unsigned CallOpc = ARMSelectCallOp(NULL);
1929 // Explicitly adding the predicate here.
1930 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1932 .addExternalSymbol(TLI.getLibcallName(Call));
1934 // Explicitly adding the predicate here.
1935 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1937 .addExternalSymbol(TLI.getLibcallName(Call)));
1939 // Add implicit physical register uses to the call.
1940 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
1941 MIB.addReg(RegArgs[i]);
1943 // Finish off the call including any return values.
1944 SmallVector<unsigned, 4> UsedRegs;
1945 if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false;
1947 // Set all unused physreg defs as dead.
1948 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
1953 bool ARMFastISel::SelectCall(const Instruction *I) {
1954 const CallInst *CI = cast<CallInst>(I);
1955 const Value *Callee = CI->getCalledValue();
1957 // Can't handle inline asm or worry about intrinsics yet.
1958 if (isa<InlineAsm>(Callee) || isa<IntrinsicInst>(CI)) return false;
1960 // Only handle global variable Callees.
1961 const GlobalValue *GV = dyn_cast<GlobalValue>(Callee);
1965 // Check the calling convention.
1966 ImmutableCallSite CS(CI);
1967 CallingConv::ID CC = CS.getCallingConv();
1969 // TODO: Avoid some calling conventions?
1971 // Let SDISel handle vararg functions.
1972 PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
1973 FunctionType *FTy = cast<FunctionType>(PT->getElementType());
1974 if (FTy->isVarArg())
1977 // Handle *simple* calls for now.
1978 Type *RetTy = I->getType();
1980 if (RetTy->isVoidTy())
1981 RetVT = MVT::isVoid;
1982 else if (!isTypeLegal(RetTy, RetVT) && RetVT != MVT::i16 &&
1983 RetVT != MVT::i8 && RetVT != MVT::i1)
1986 // TODO: For now if we have long calls specified we don't handle the call.
1987 if (EnableARMLongCalls) return false;
1989 // Set up the argument vectors.
1990 SmallVector<Value*, 8> Args;
1991 SmallVector<unsigned, 8> ArgRegs;
1992 SmallVector<MVT, 8> ArgVTs;
1993 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
1994 Args.reserve(CS.arg_size());
1995 ArgRegs.reserve(CS.arg_size());
1996 ArgVTs.reserve(CS.arg_size());
1997 ArgFlags.reserve(CS.arg_size());
1998 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
2000 unsigned Arg = getRegForValue(*i);
2004 ISD::ArgFlagsTy Flags;
2005 unsigned AttrInd = i - CS.arg_begin() + 1;
2006 if (CS.paramHasAttr(AttrInd, Attribute::SExt))
2008 if (CS.paramHasAttr(AttrInd, Attribute::ZExt))
2011 // FIXME: Only handle *easy* calls for now.
2012 if (CS.paramHasAttr(AttrInd, Attribute::InReg) ||
2013 CS.paramHasAttr(AttrInd, Attribute::StructRet) ||
2014 CS.paramHasAttr(AttrInd, Attribute::Nest) ||
2015 CS.paramHasAttr(AttrInd, Attribute::ByVal))
2018 Type *ArgTy = (*i)->getType();
2020 if (!isTypeLegal(ArgTy, ArgVT) && ArgVT != MVT::i16 && ArgVT != MVT::i8 &&
2023 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
2024 Flags.setOrigAlign(OriginalAlignment);
2027 ArgRegs.push_back(Arg);
2028 ArgVTs.push_back(ArgVT);
2029 ArgFlags.push_back(Flags);
2032 // Handle the arguments now that we've gotten them.
2033 SmallVector<unsigned, 4> RegArgs;
2035 if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
2038 // Issue the call, BLr9 for darwin, BL otherwise.
2039 // TODO: Turn this into the table of arm call ops.
2040 MachineInstrBuilder MIB;
2041 unsigned CallOpc = ARMSelectCallOp(GV);
2042 // Explicitly adding the predicate here.
2044 // Explicitly adding the predicate here.
2045 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
2047 .addGlobalAddress(GV, 0, 0);
2049 // Explicitly adding the predicate here.
2050 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
2052 .addGlobalAddress(GV, 0, 0));
2054 // Add implicit physical register uses to the call.
2055 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
2056 MIB.addReg(RegArgs[i]);
2058 // Finish off the call including any return values.
2059 SmallVector<unsigned, 4> UsedRegs;
2060 if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false;
2062 // Set all unused physreg defs as dead.
2063 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
2068 bool ARMFastISel::SelectTrunc(const Instruction *I) {
2069 // The high bits for a type smaller than the register size are assumed to be
2071 Value *Op = I->getOperand(0);
2074 SrcVT = TLI.getValueType(Op->getType(), true);
2075 DestVT = TLI.getValueType(I->getType(), true);
2077 if (SrcVT != MVT::i32 && SrcVT != MVT::i16 && SrcVT != MVT::i8)
2079 if (DestVT != MVT::i16 && DestVT != MVT::i8 && DestVT != MVT::i1)
2082 unsigned SrcReg = getRegForValue(Op);
2083 if (!SrcReg) return false;
2085 // Because the high bits are undefined, a truncate doesn't generate
2087 UpdateValueMap(I, SrcReg);
2091 unsigned ARMFastISel::ARMEmitIntExt(EVT SrcVT, unsigned SrcReg, EVT DestVT,
2093 if (DestVT != MVT::i32 && DestVT != MVT::i16 && DestVT != MVT::i8)
2097 bool isBoolZext = false;
2098 if (!SrcVT.isSimple()) return 0;
2099 switch (SrcVT.getSimpleVT().SimpleTy) {
2102 if (!Subtarget->hasV6Ops()) return 0;
2104 Opc = isThumb2 ? ARM::t2UXTH : ARM::UXTH;
2106 Opc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
2109 if (!Subtarget->hasV6Ops()) return 0;
2111 Opc = isThumb2 ? ARM::t2UXTB : ARM::UXTB;
2113 Opc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
2117 Opc = isThumb2 ? ARM::t2ANDri : ARM::ANDri;
2124 unsigned ResultReg = createResultReg(TLI.getRegClassFor(MVT::i32));
2125 MachineInstrBuilder MIB;
2126 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), ResultReg)
2132 AddOptionalDefs(MIB);
2136 bool ARMFastISel::SelectIntExt(const Instruction *I) {
2137 // On ARM, in general, integer casts don't involve legal types; this code
2138 // handles promotable integers.
2139 // FIXME: We could save an instruction in many cases by special-casing
2140 // load instructions.
2141 Type *DestTy = I->getType();
2142 Value *Src = I->getOperand(0);
2143 Type *SrcTy = Src->getType();
2146 SrcVT = TLI.getValueType(SrcTy, true);
2147 DestVT = TLI.getValueType(DestTy, true);
2149 bool isZExt = isa<ZExtInst>(I);
2150 unsigned SrcReg = getRegForValue(Src);
2151 if (!SrcReg) return false;
2153 unsigned ResultReg = ARMEmitIntExt(SrcVT, SrcReg, DestVT, isZExt);
2154 if (ResultReg == 0) return false;
2155 UpdateValueMap(I, ResultReg);
2159 // TODO: SoftFP support.
2160 bool ARMFastISel::TargetSelectInstruction(const Instruction *I) {
2162 switch (I->getOpcode()) {
2163 case Instruction::Load:
2164 return SelectLoad(I);
2165 case Instruction::Store:
2166 return SelectStore(I);
2167 case Instruction::Br:
2168 return SelectBranch(I);
2169 case Instruction::ICmp:
2170 case Instruction::FCmp:
2171 return SelectCmp(I);
2172 case Instruction::FPExt:
2173 return SelectFPExt(I);
2174 case Instruction::FPTrunc:
2175 return SelectFPTrunc(I);
2176 case Instruction::SIToFP:
2177 return SelectSIToFP(I);
2178 case Instruction::FPToSI:
2179 return SelectFPToSI(I);
2180 case Instruction::FAdd:
2181 return SelectBinaryOp(I, ISD::FADD);
2182 case Instruction::FSub:
2183 return SelectBinaryOp(I, ISD::FSUB);
2184 case Instruction::FMul:
2185 return SelectBinaryOp(I, ISD::FMUL);
2186 case Instruction::SDiv:
2187 return SelectSDiv(I);
2188 case Instruction::SRem:
2189 return SelectSRem(I);
2190 case Instruction::Call:
2191 return SelectCall(I);
2192 case Instruction::Select:
2193 return SelectSelect(I);
2194 case Instruction::Ret:
2195 return SelectRet(I);
2196 case Instruction::Trunc:
2197 return SelectTrunc(I);
2198 case Instruction::ZExt:
2199 case Instruction::SExt:
2200 return SelectIntExt(I);
2207 llvm::FastISel *ARM::createFastISel(FunctionLoweringInfo &funcInfo) {
2208 // Completely untested on non-darwin.
2209 const TargetMachine &TM = funcInfo.MF->getTarget();
2211 // Darwin and thumb1 only for now.
2212 const ARMSubtarget *Subtarget = &TM.getSubtarget<ARMSubtarget>();
2213 if (Subtarget->isTargetDarwin() && !Subtarget->isThumb1Only() &&
2214 !DisableARMFastISel)
2215 return new ARMFastISel(funcInfo);