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 if (Addr.Offset > 255 || Addr.Offset < -255)
856 needsLowering = true;
861 // Integer loads/stores handle 12-bit offsets.
862 needsLowering = ((Addr.Offset & 0xfff) != Addr.Offset);
866 // Floating point operands handle 8-bit offsets.
867 needsLowering = ((Addr.Offset & 0xff) != Addr.Offset);
871 // If this is a stack pointer and the offset needs to be simplified then
872 // put the alloca address into a register, set the base type back to
873 // register and continue. This should almost never happen.
874 if (needsLowering && Addr.BaseType == Address::FrameIndexBase) {
875 TargetRegisterClass *RC = isThumb2 ? ARM::tGPRRegisterClass :
876 ARM::GPRRegisterClass;
877 unsigned ResultReg = createResultReg(RC);
878 unsigned Opc = isThumb2 ? ARM::t2ADDri : ARM::ADDri;
879 AddOptionalDefs(BuildMI(*FuncInfo.MBB, *FuncInfo.InsertPt, DL,
880 TII.get(Opc), ResultReg)
881 .addFrameIndex(Addr.Base.FI)
883 Addr.Base.Reg = ResultReg;
884 Addr.BaseType = Address::RegBase;
887 // Since the offset is too large for the load/store instruction
888 // get the reg+offset into a register.
890 Addr.Base.Reg = FastEmit_ri_(MVT::i32, ISD::ADD, Addr.Base.Reg,
891 /*Op0IsKill*/false, Addr.Offset, MVT::i32);
896 void ARMFastISel::AddLoadStoreOperands(EVT VT, Address &Addr,
897 const MachineInstrBuilder &MIB,
899 // addrmode5 output depends on the selection dag addressing dividing the
900 // offset by 4 that it then later multiplies. Do this here as well.
901 if (VT.getSimpleVT().SimpleTy == MVT::f32 ||
902 VT.getSimpleVT().SimpleTy == MVT::f64)
905 // Frame base works a bit differently. Handle it separately.
906 if (Addr.BaseType == Address::FrameIndexBase) {
907 int FI = Addr.Base.FI;
908 int Offset = Addr.Offset;
909 MachineMemOperand *MMO =
910 FuncInfo.MF->getMachineMemOperand(
911 MachinePointerInfo::getFixedStack(FI, Offset),
913 MFI.getObjectSize(FI),
914 MFI.getObjectAlignment(FI));
915 // Now add the rest of the operands.
916 MIB.addFrameIndex(FI);
918 // ARM halfword load/stores need an additional operand.
919 if (!isThumb2 && VT.getSimpleVT().SimpleTy == MVT::i16) MIB.addReg(0);
921 MIB.addImm(Addr.Offset);
922 MIB.addMemOperand(MMO);
924 // Now add the rest of the operands.
925 MIB.addReg(Addr.Base.Reg);
927 // ARM halfword load/stores need an additional operand.
928 if (!isThumb2 && VT.getSimpleVT().SimpleTy == MVT::i16) MIB.addReg(0);
930 MIB.addImm(Addr.Offset);
932 AddOptionalDefs(MIB);
935 bool ARMFastISel::ARMEmitLoad(EVT VT, unsigned &ResultReg, Address &Addr) {
937 assert(VT.isSimple() && "Non-simple types are invalid here!");
939 TargetRegisterClass *RC;
940 switch (VT.getSimpleVT().SimpleTy) {
941 // This is mostly going to be Neon/vector support.
942 default: return false;
944 Opc = isThumb2 ? ARM::t2LDRBi12 : ARM::LDRBi12;
945 RC = ARM::GPRRegisterClass;
948 Opc = isThumb2 ? ARM::t2LDRHi12 : ARM::LDRH;
949 RC = ARM::GPRRegisterClass;
952 Opc = isThumb2 ? ARM::t2LDRi12 : ARM::LDRi12;
953 RC = ARM::GPRRegisterClass;
957 RC = TLI.getRegClassFor(VT);
961 RC = TLI.getRegClassFor(VT);
964 // Simplify this down to something we can handle.
965 ARMSimplifyAddress(Addr, VT);
967 // Create the base instruction, then add the operands.
968 ResultReg = createResultReg(RC);
969 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
970 TII.get(Opc), ResultReg);
971 AddLoadStoreOperands(VT, Addr, MIB, MachineMemOperand::MOLoad);
975 bool ARMFastISel::SelectLoad(const Instruction *I) {
976 // Atomic loads need special handling.
977 if (cast<LoadInst>(I)->isAtomic())
980 // Verify we have a legal type before going any further.
982 if (!isLoadTypeLegal(I->getType(), VT))
985 // See if we can handle this address.
987 if (!ARMComputeAddress(I->getOperand(0), Addr)) return false;
990 if (!ARMEmitLoad(VT, ResultReg, Addr)) return false;
991 UpdateValueMap(I, ResultReg);
995 bool ARMFastISel::ARMEmitStore(EVT VT, unsigned SrcReg, Address &Addr) {
997 switch (VT.getSimpleVT().SimpleTy) {
998 // This is mostly going to be Neon/vector support.
999 default: return false;
1001 unsigned Res = createResultReg(isThumb2 ? ARM::tGPRRegisterClass :
1002 ARM::GPRRegisterClass);
1003 unsigned Opc = isThumb2 ? ARM::t2ANDri : ARM::ANDri;
1004 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1006 .addReg(SrcReg).addImm(1));
1008 } // Fallthrough here.
1010 StrOpc = isThumb2 ? ARM::t2STRBi12 : ARM::STRBi12;
1013 StrOpc = isThumb2 ? ARM::t2STRHi12 : ARM::STRH;
1016 StrOpc = isThumb2 ? ARM::t2STRi12 : ARM::STRi12;
1019 if (!Subtarget->hasVFP2()) return false;
1020 StrOpc = ARM::VSTRS;
1023 if (!Subtarget->hasVFP2()) return false;
1024 StrOpc = ARM::VSTRD;
1027 // Simplify this down to something we can handle.
1028 ARMSimplifyAddress(Addr, VT);
1030 // Create the base instruction, then add the operands.
1031 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1033 .addReg(SrcReg, getKillRegState(true));
1034 AddLoadStoreOperands(VT, Addr, MIB, MachineMemOperand::MOStore);
1038 bool ARMFastISel::SelectStore(const Instruction *I) {
1039 Value *Op0 = I->getOperand(0);
1040 unsigned SrcReg = 0;
1042 // Atomic stores need special handling.
1043 if (cast<StoreInst>(I)->isAtomic())
1046 // Verify we have a legal type before going any further.
1048 if (!isLoadTypeLegal(I->getOperand(0)->getType(), VT))
1051 // Get the value to be stored into a register.
1052 SrcReg = getRegForValue(Op0);
1053 if (SrcReg == 0) return false;
1055 // See if we can handle this address.
1057 if (!ARMComputeAddress(I->getOperand(1), Addr))
1060 if (!ARMEmitStore(VT, SrcReg, Addr)) return false;
1064 static ARMCC::CondCodes getComparePred(CmpInst::Predicate Pred) {
1066 // Needs two compares...
1067 case CmpInst::FCMP_ONE:
1068 case CmpInst::FCMP_UEQ:
1070 // AL is our "false" for now. The other two need more compares.
1072 case CmpInst::ICMP_EQ:
1073 case CmpInst::FCMP_OEQ:
1075 case CmpInst::ICMP_SGT:
1076 case CmpInst::FCMP_OGT:
1078 case CmpInst::ICMP_SGE:
1079 case CmpInst::FCMP_OGE:
1081 case CmpInst::ICMP_UGT:
1082 case CmpInst::FCMP_UGT:
1084 case CmpInst::FCMP_OLT:
1086 case CmpInst::ICMP_ULE:
1087 case CmpInst::FCMP_OLE:
1089 case CmpInst::FCMP_ORD:
1091 case CmpInst::FCMP_UNO:
1093 case CmpInst::FCMP_UGE:
1095 case CmpInst::ICMP_SLT:
1096 case CmpInst::FCMP_ULT:
1098 case CmpInst::ICMP_SLE:
1099 case CmpInst::FCMP_ULE:
1101 case CmpInst::FCMP_UNE:
1102 case CmpInst::ICMP_NE:
1104 case CmpInst::ICMP_UGE:
1106 case CmpInst::ICMP_ULT:
1111 bool ARMFastISel::SelectBranch(const Instruction *I) {
1112 const BranchInst *BI = cast<BranchInst>(I);
1113 MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
1114 MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
1116 // Simple branch support.
1118 // If we can, avoid recomputing the compare - redoing it could lead to wonky
1120 if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
1121 if (CI->hasOneUse() && (CI->getParent() == I->getParent())) {
1123 // Get the compare predicate.
1124 // Try to take advantage of fallthrough opportunities.
1125 CmpInst::Predicate Predicate = CI->getPredicate();
1126 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
1127 std::swap(TBB, FBB);
1128 Predicate = CmpInst::getInversePredicate(Predicate);
1131 ARMCC::CondCodes ARMPred = getComparePred(Predicate);
1133 // We may not handle every CC for now.
1134 if (ARMPred == ARMCC::AL) return false;
1136 // Emit the compare.
1137 if (!ARMEmitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
1140 unsigned BrOpc = isThumb2 ? ARM::t2Bcc : ARM::Bcc;
1141 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
1142 .addMBB(TBB).addImm(ARMPred).addReg(ARM::CPSR);
1143 FastEmitBranch(FBB, DL);
1144 FuncInfo.MBB->addSuccessor(TBB);
1147 } else if (TruncInst *TI = dyn_cast<TruncInst>(BI->getCondition())) {
1149 if (TI->hasOneUse() && TI->getParent() == I->getParent() &&
1150 (isLoadTypeLegal(TI->getOperand(0)->getType(), SourceVT))) {
1151 unsigned TstOpc = isThumb2 ? ARM::t2TSTri : ARM::TSTri;
1152 unsigned OpReg = getRegForValue(TI->getOperand(0));
1153 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1155 .addReg(OpReg).addImm(1));
1157 unsigned CCMode = ARMCC::NE;
1158 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
1159 std::swap(TBB, FBB);
1163 unsigned BrOpc = isThumb2 ? ARM::t2Bcc : ARM::Bcc;
1164 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
1165 .addMBB(TBB).addImm(CCMode).addReg(ARM::CPSR);
1167 FastEmitBranch(FBB, DL);
1168 FuncInfo.MBB->addSuccessor(TBB);
1171 } else if (const ConstantInt *CI =
1172 dyn_cast<ConstantInt>(BI->getCondition())) {
1173 uint64_t Imm = CI->getZExtValue();
1174 MachineBasicBlock *Target = (Imm == 0) ? FBB : TBB;
1175 FastEmitBranch(Target, DL);
1179 unsigned CmpReg = getRegForValue(BI->getCondition());
1180 if (CmpReg == 0) return false;
1182 // We've been divorced from our compare! Our block was split, and
1183 // now our compare lives in a predecessor block. We musn't
1184 // re-compare here, as the children of the compare aren't guaranteed
1185 // live across the block boundary (we *could* check for this).
1186 // Regardless, the compare has been done in the predecessor block,
1187 // and it left a value for us in a virtual register. Ergo, we test
1188 // the one-bit value left in the virtual register.
1189 unsigned TstOpc = isThumb2 ? ARM::t2TSTri : ARM::TSTri;
1190 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TstOpc))
1191 .addReg(CmpReg).addImm(1));
1193 unsigned CCMode = ARMCC::NE;
1194 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
1195 std::swap(TBB, FBB);
1199 unsigned BrOpc = isThumb2 ? ARM::t2Bcc : ARM::Bcc;
1200 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
1201 .addMBB(TBB).addImm(CCMode).addReg(ARM::CPSR);
1202 FastEmitBranch(FBB, DL);
1203 FuncInfo.MBB->addSuccessor(TBB);
1207 bool ARMFastISel::ARMEmitCmp(const Value *Src1Value, const Value *Src2Value,
1209 Type *Ty = Src1Value->getType();
1210 EVT SrcVT = TLI.getValueType(Ty, true);
1211 if (!SrcVT.isSimple()) return false;
1213 bool isFloat = (Ty->isFloatTy() || Ty->isDoubleTy());
1214 if (isFloat && !Subtarget->hasVFP2())
1217 // 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();
1228 isNegativeImm = CIVal.isNegative();
1229 Imm = (isNegativeImm) ? (-CIVal).getZExtValue() : CIVal.getZExtValue();
1230 EncodedImm = (int)Imm;
1231 EncodeImm = isThumb2 ? (ARM_AM::getT2SOImmVal(EncodedImm) != -1) :
1232 (ARM_AM::getSOImmVal(EncodedImm) != -1);
1234 } else if (const ConstantFP *ConstFP = dyn_cast<ConstantFP>(Src2Value)) {
1235 if (SrcVT == MVT::f32 || SrcVT == MVT::f64)
1236 if (ConstFP->isZero() && !ConstFP->isNegative())
1242 bool needsExt = false;
1243 switch (SrcVT.getSimpleVT().SimpleTy) {
1244 default: return false;
1245 // TODO: Verify compares.
1248 CmpOpc = EncodeImm ? ARM::VCMPEZS : ARM::VCMPES;
1252 CmpOpc = EncodeImm ? ARM::VCMPEZD : ARM::VCMPED;
1258 // Intentional fall-through.
1262 CmpOpc = ARM::t2CMPrr;
1264 CmpOpc = isNegativeImm ? ARM::t2CMNzri : ARM::t2CMPri;
1267 CmpOpc = ARM::CMPrr;
1269 CmpOpc = isNegativeImm ? ARM::CMNzri : ARM::CMPri;
1274 unsigned SrcReg1 = getRegForValue(Src1Value);
1275 if (SrcReg1 == 0) return false;
1279 SrcReg2 = getRegForValue(Src2Value);
1280 if (SrcReg2 == 0) return false;
1283 // We have i1, i8, or i16, we need to either zero extend or sign extend.
1286 ResultReg = ARMEmitIntExt(SrcVT, SrcReg1, MVT::i32, isZExt);
1287 if (ResultReg == 0) return false;
1288 SrcReg1 = ResultReg;
1290 ResultReg = ARMEmitIntExt(SrcVT, SrcReg2, MVT::i32, isZExt);
1291 if (ResultReg == 0) return false;
1292 SrcReg2 = ResultReg;
1297 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1299 .addReg(SrcReg1).addReg(SrcReg2));
1301 MachineInstrBuilder MIB;
1302 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
1305 // Only add immediate for icmp as the immediate for fcmp is an implicit 0.0.
1307 MIB.addImm(EncodedImm);
1308 AddOptionalDefs(MIB);
1311 // For floating point we need to move the result to a comparison register
1312 // that we can then use for branches.
1313 if (Ty->isFloatTy() || Ty->isDoubleTy())
1314 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1315 TII.get(ARM::FMSTAT)));
1319 bool ARMFastISel::SelectCmp(const Instruction *I) {
1320 const CmpInst *CI = cast<CmpInst>(I);
1321 Type *Ty = CI->getOperand(0)->getType();
1323 // Get the compare predicate.
1324 ARMCC::CondCodes ARMPred = getComparePred(CI->getPredicate());
1326 // We may not handle every CC for now.
1327 if (ARMPred == ARMCC::AL) return false;
1329 // Emit the compare.
1330 if (!ARMEmitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
1333 // Now set a register based on the comparison. Explicitly set the predicates
1335 unsigned MovCCOpc = isThumb2 ? ARM::t2MOVCCi : ARM::MOVCCi;
1336 TargetRegisterClass *RC = isThumb2 ? ARM::rGPRRegisterClass
1337 : ARM::GPRRegisterClass;
1338 unsigned DestReg = createResultReg(RC);
1339 Constant *Zero = ConstantInt::get(Type::getInt32Ty(*Context), 0);
1340 unsigned ZeroReg = TargetMaterializeConstant(Zero);
1341 bool isFloat = (Ty->isFloatTy() || Ty->isDoubleTy());
1342 unsigned CondReg = isFloat ? ARM::FPSCR : ARM::CPSR;
1343 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), DestReg)
1344 .addReg(ZeroReg).addImm(1)
1345 .addImm(ARMPred).addReg(CondReg);
1347 UpdateValueMap(I, DestReg);
1351 bool ARMFastISel::SelectFPExt(const Instruction *I) {
1352 // Make sure we have VFP and that we're extending float to double.
1353 if (!Subtarget->hasVFP2()) return false;
1355 Value *V = I->getOperand(0);
1356 if (!I->getType()->isDoubleTy() ||
1357 !V->getType()->isFloatTy()) return false;
1359 unsigned Op = getRegForValue(V);
1360 if (Op == 0) return false;
1362 unsigned Result = createResultReg(ARM::DPRRegisterClass);
1363 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1364 TII.get(ARM::VCVTDS), Result)
1366 UpdateValueMap(I, Result);
1370 bool ARMFastISel::SelectFPTrunc(const Instruction *I) {
1371 // Make sure we have VFP and that we're truncating double to float.
1372 if (!Subtarget->hasVFP2()) return false;
1374 Value *V = I->getOperand(0);
1375 if (!(I->getType()->isFloatTy() &&
1376 V->getType()->isDoubleTy())) return false;
1378 unsigned Op = getRegForValue(V);
1379 if (Op == 0) return false;
1381 unsigned Result = createResultReg(ARM::SPRRegisterClass);
1382 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1383 TII.get(ARM::VCVTSD), Result)
1385 UpdateValueMap(I, Result);
1389 bool ARMFastISel::SelectSIToFP(const Instruction *I) {
1390 // Make sure we have VFP.
1391 if (!Subtarget->hasVFP2()) return false;
1394 Type *Ty = I->getType();
1395 if (!isTypeLegal(Ty, DstVT))
1398 Value *Src = I->getOperand(0);
1399 EVT SrcVT = TLI.getValueType(Src->getType(), true);
1400 if (SrcVT != MVT::i32 && SrcVT != MVT::i16 && SrcVT != MVT::i8)
1403 unsigned SrcReg = getRegForValue(Src);
1404 if (SrcReg == 0) return false;
1406 // Handle sign-extension.
1407 if (SrcVT == MVT::i16 || SrcVT == MVT::i8) {
1408 EVT DestVT = MVT::i32;
1409 unsigned ResultReg = ARMEmitIntExt(SrcVT, SrcReg, DestVT, /*isZExt*/ false);
1410 if (ResultReg == 0) return false;
1414 // The conversion routine works on fp-reg to fp-reg and the operand above
1415 // was an integer, move it to the fp registers if possible.
1416 unsigned FP = ARMMoveToFPReg(MVT::f32, SrcReg);
1417 if (FP == 0) return false;
1420 if (Ty->isFloatTy()) Opc = ARM::VSITOS;
1421 else if (Ty->isDoubleTy()) Opc = ARM::VSITOD;
1424 unsigned ResultReg = createResultReg(TLI.getRegClassFor(DstVT));
1425 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
1428 UpdateValueMap(I, ResultReg);
1432 bool ARMFastISel::SelectFPToSI(const Instruction *I) {
1433 // Make sure we have VFP.
1434 if (!Subtarget->hasVFP2()) return false;
1437 Type *RetTy = I->getType();
1438 if (!isTypeLegal(RetTy, DstVT))
1441 unsigned Op = getRegForValue(I->getOperand(0));
1442 if (Op == 0) return false;
1445 Type *OpTy = I->getOperand(0)->getType();
1446 if (OpTy->isFloatTy()) Opc = ARM::VTOSIZS;
1447 else if (OpTy->isDoubleTy()) Opc = ARM::VTOSIZD;
1450 // f64->s32 or f32->s32 both need an intermediate f32 reg.
1451 unsigned ResultReg = createResultReg(TLI.getRegClassFor(MVT::f32));
1452 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
1456 // This result needs to be in an integer register, but the conversion only
1457 // takes place in fp-regs.
1458 unsigned IntReg = ARMMoveToIntReg(DstVT, ResultReg);
1459 if (IntReg == 0) return false;
1461 UpdateValueMap(I, IntReg);
1465 bool ARMFastISel::SelectSelect(const Instruction *I) {
1467 if (!isTypeLegal(I->getType(), VT))
1470 // Things need to be register sized for register moves.
1471 if (VT != MVT::i32) return false;
1472 const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
1474 unsigned CondReg = getRegForValue(I->getOperand(0));
1475 if (CondReg == 0) return false;
1476 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1477 if (Op1Reg == 0) return false;
1478 unsigned Op2Reg = getRegForValue(I->getOperand(2));
1479 if (Op2Reg == 0) return false;
1481 unsigned CmpOpc = isThumb2 ? ARM::t2TSTri : ARM::TSTri;
1482 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
1483 .addReg(CondReg).addImm(1));
1484 unsigned ResultReg = createResultReg(RC);
1485 unsigned MovCCOpc = isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr;
1486 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), ResultReg)
1487 .addReg(Op1Reg).addReg(Op2Reg)
1488 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
1489 UpdateValueMap(I, ResultReg);
1493 bool ARMFastISel::SelectSDiv(const Instruction *I) {
1495 Type *Ty = I->getType();
1496 if (!isTypeLegal(Ty, VT))
1499 // If we have integer div support we should have selected this automagically.
1500 // In case we have a real miss go ahead and return false and we'll pick
1502 if (Subtarget->hasDivide()) return false;
1504 // Otherwise emit a libcall.
1505 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
1507 LC = RTLIB::SDIV_I8;
1508 else if (VT == MVT::i16)
1509 LC = RTLIB::SDIV_I16;
1510 else if (VT == MVT::i32)
1511 LC = RTLIB::SDIV_I32;
1512 else if (VT == MVT::i64)
1513 LC = RTLIB::SDIV_I64;
1514 else if (VT == MVT::i128)
1515 LC = RTLIB::SDIV_I128;
1516 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SDIV!");
1518 return ARMEmitLibcall(I, LC);
1521 bool ARMFastISel::SelectSRem(const Instruction *I) {
1523 Type *Ty = I->getType();
1524 if (!isTypeLegal(Ty, VT))
1527 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
1529 LC = RTLIB::SREM_I8;
1530 else if (VT == MVT::i16)
1531 LC = RTLIB::SREM_I16;
1532 else if (VT == MVT::i32)
1533 LC = RTLIB::SREM_I32;
1534 else if (VT == MVT::i64)
1535 LC = RTLIB::SREM_I64;
1536 else if (VT == MVT::i128)
1537 LC = RTLIB::SREM_I128;
1538 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SREM!");
1540 return ARMEmitLibcall(I, LC);
1543 bool ARMFastISel::SelectBinaryOp(const Instruction *I, unsigned ISDOpcode) {
1544 EVT VT = TLI.getValueType(I->getType(), true);
1546 // We can get here in the case when we want to use NEON for our fp
1547 // operations, but can't figure out how to. Just use the vfp instructions
1549 // FIXME: It'd be nice to use NEON instructions.
1550 Type *Ty = I->getType();
1551 bool isFloat = (Ty->isDoubleTy() || Ty->isFloatTy());
1552 if (isFloat && !Subtarget->hasVFP2())
1555 unsigned Op1 = getRegForValue(I->getOperand(0));
1556 if (Op1 == 0) return false;
1558 unsigned Op2 = getRegForValue(I->getOperand(1));
1559 if (Op2 == 0) return false;
1562 bool is64bit = VT == MVT::f64 || VT == MVT::i64;
1563 switch (ISDOpcode) {
1564 default: return false;
1566 Opc = is64bit ? ARM::VADDD : ARM::VADDS;
1569 Opc = is64bit ? ARM::VSUBD : ARM::VSUBS;
1572 Opc = is64bit ? ARM::VMULD : ARM::VMULS;
1575 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
1576 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1577 TII.get(Opc), ResultReg)
1578 .addReg(Op1).addReg(Op2));
1579 UpdateValueMap(I, ResultReg);
1583 // Call Handling Code
1585 bool ARMFastISel::FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src,
1586 EVT SrcVT, unsigned &ResultReg) {
1587 unsigned RR = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), Opc,
1588 Src, /*TODO: Kill=*/false);
1597 // This is largely taken directly from CCAssignFnForNode - we don't support
1598 // varargs in FastISel so that part has been removed.
1599 // TODO: We may not support all of this.
1600 CCAssignFn *ARMFastISel::CCAssignFnForCall(CallingConv::ID CC, bool Return) {
1603 llvm_unreachable("Unsupported calling convention");
1604 case CallingConv::Fast:
1605 // Ignore fastcc. Silence compiler warnings.
1606 (void)RetFastCC_ARM_APCS;
1607 (void)FastCC_ARM_APCS;
1609 case CallingConv::C:
1610 // Use target triple & subtarget features to do actual dispatch.
1611 if (Subtarget->isAAPCS_ABI()) {
1612 if (Subtarget->hasVFP2() &&
1613 FloatABIType == FloatABI::Hard)
1614 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
1616 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
1618 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
1619 case CallingConv::ARM_AAPCS_VFP:
1620 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
1621 case CallingConv::ARM_AAPCS:
1622 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
1623 case CallingConv::ARM_APCS:
1624 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
1628 bool ARMFastISel::ProcessCallArgs(SmallVectorImpl<Value*> &Args,
1629 SmallVectorImpl<unsigned> &ArgRegs,
1630 SmallVectorImpl<MVT> &ArgVTs,
1631 SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
1632 SmallVectorImpl<unsigned> &RegArgs,
1634 unsigned &NumBytes) {
1635 SmallVector<CCValAssign, 16> ArgLocs;
1636 CCState CCInfo(CC, false, *FuncInfo.MF, TM, ArgLocs, *Context);
1637 CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CCAssignFnForCall(CC, false));
1639 // Get a count of how many bytes are to be pushed on the stack.
1640 NumBytes = CCInfo.getNextStackOffset();
1642 // Issue CALLSEQ_START
1643 unsigned AdjStackDown = TII.getCallFrameSetupOpcode();
1644 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1645 TII.get(AdjStackDown))
1648 // Process the args.
1649 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1650 CCValAssign &VA = ArgLocs[i];
1651 unsigned Arg = ArgRegs[VA.getValNo()];
1652 MVT ArgVT = ArgVTs[VA.getValNo()];
1654 // We don't handle NEON/vector parameters yet.
1655 if (ArgVT.isVector() || ArgVT.getSizeInBits() > 64)
1658 // Handle arg promotion, etc.
1659 switch (VA.getLocInfo()) {
1660 case CCValAssign::Full: break;
1661 case CCValAssign::SExt: {
1662 EVT DestVT = VA.getLocVT();
1663 unsigned ResultReg = ARMEmitIntExt(ArgVT, Arg, DestVT,
1665 assert (ResultReg != 0 && "Failed to emit a sext");
1669 case CCValAssign::AExt:
1670 // Intentional fall-through. Handle AExt and ZExt.
1671 case CCValAssign::ZExt: {
1672 EVT DestVT = VA.getLocVT();
1673 unsigned ResultReg = ARMEmitIntExt(ArgVT, Arg, DestVT,
1675 assert (ResultReg != 0 && "Failed to emit a sext");
1679 case CCValAssign::BCvt: {
1680 unsigned BC = FastEmit_r(ArgVT, VA.getLocVT(), ISD::BITCAST, Arg,
1681 /*TODO: Kill=*/false);
1682 assert(BC != 0 && "Failed to emit a bitcast!");
1684 ArgVT = VA.getLocVT();
1687 default: llvm_unreachable("Unknown arg promotion!");
1690 // Now copy/store arg to correct locations.
1691 if (VA.isRegLoc() && !VA.needsCustom()) {
1692 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1695 RegArgs.push_back(VA.getLocReg());
1696 } else if (VA.needsCustom()) {
1697 // TODO: We need custom lowering for vector (v2f64) args.
1698 if (VA.getLocVT() != MVT::f64) return false;
1700 CCValAssign &NextVA = ArgLocs[++i];
1702 // TODO: Only handle register args for now.
1703 if(!(VA.isRegLoc() && NextVA.isRegLoc())) return false;
1705 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1706 TII.get(ARM::VMOVRRD), VA.getLocReg())
1707 .addReg(NextVA.getLocReg(), RegState::Define)
1709 RegArgs.push_back(VA.getLocReg());
1710 RegArgs.push_back(NextVA.getLocReg());
1712 assert(VA.isMemLoc());
1713 // Need to store on the stack.
1715 Addr.BaseType = Address::RegBase;
1716 Addr.Base.Reg = ARM::SP;
1717 Addr.Offset = VA.getLocMemOffset();
1719 if (!ARMEmitStore(ArgVT, Arg, Addr)) return false;
1725 bool ARMFastISel::FinishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
1726 const Instruction *I, CallingConv::ID CC,
1727 unsigned &NumBytes) {
1728 // Issue CALLSEQ_END
1729 unsigned AdjStackUp = TII.getCallFrameDestroyOpcode();
1730 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1731 TII.get(AdjStackUp))
1732 .addImm(NumBytes).addImm(0));
1734 // Now the return value.
1735 if (RetVT != MVT::isVoid) {
1736 SmallVector<CCValAssign, 16> RVLocs;
1737 CCState CCInfo(CC, false, *FuncInfo.MF, TM, RVLocs, *Context);
1738 CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC, true));
1740 // Copy all of the result registers out of their specified physreg.
1741 if (RVLocs.size() == 2 && RetVT == MVT::f64) {
1742 // For this move we copy into two registers and then move into the
1743 // double fp reg we want.
1744 EVT DestVT = RVLocs[0].getValVT();
1745 TargetRegisterClass* DstRC = TLI.getRegClassFor(DestVT);
1746 unsigned ResultReg = createResultReg(DstRC);
1747 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1748 TII.get(ARM::VMOVDRR), ResultReg)
1749 .addReg(RVLocs[0].getLocReg())
1750 .addReg(RVLocs[1].getLocReg()));
1752 UsedRegs.push_back(RVLocs[0].getLocReg());
1753 UsedRegs.push_back(RVLocs[1].getLocReg());
1755 // Finally update the result.
1756 UpdateValueMap(I, ResultReg);
1758 assert(RVLocs.size() == 1 &&"Can't handle non-double multi-reg retvals!");
1759 EVT CopyVT = RVLocs[0].getValVT();
1761 // Special handling for extended integers.
1762 if (RetVT == MVT::i1 || RetVT == MVT::i8 || RetVT == MVT::i16)
1765 TargetRegisterClass* DstRC = TLI.getRegClassFor(CopyVT);
1767 unsigned ResultReg = createResultReg(DstRC);
1768 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1769 ResultReg).addReg(RVLocs[0].getLocReg());
1770 UsedRegs.push_back(RVLocs[0].getLocReg());
1772 // Finally update the result.
1773 UpdateValueMap(I, ResultReg);
1780 bool ARMFastISel::SelectRet(const Instruction *I) {
1781 const ReturnInst *Ret = cast<ReturnInst>(I);
1782 const Function &F = *I->getParent()->getParent();
1784 if (!FuncInfo.CanLowerReturn)
1790 CallingConv::ID CC = F.getCallingConv();
1791 if (Ret->getNumOperands() > 0) {
1792 SmallVector<ISD::OutputArg, 4> Outs;
1793 GetReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(),
1796 // Analyze operands of the call, assigning locations to each operand.
1797 SmallVector<CCValAssign, 16> ValLocs;
1798 CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, TM, ValLocs,I->getContext());
1799 CCInfo.AnalyzeReturn(Outs, CCAssignFnForCall(CC, true /* is Ret */));
1801 const Value *RV = Ret->getOperand(0);
1802 unsigned Reg = getRegForValue(RV);
1806 // Only handle a single return value for now.
1807 if (ValLocs.size() != 1)
1810 CCValAssign &VA = ValLocs[0];
1812 // Don't bother handling odd stuff for now.
1813 if (VA.getLocInfo() != CCValAssign::Full)
1815 // Only handle register returns for now.
1819 unsigned SrcReg = Reg + VA.getValNo();
1820 EVT RVVT = TLI.getValueType(RV->getType());
1821 EVT DestVT = VA.getValVT();
1822 // Special handling for extended integers.
1823 if (RVVT != DestVT) {
1824 if (RVVT != MVT::i1 && RVVT != MVT::i8 && RVVT != MVT::i16)
1827 if (!Outs[0].Flags.isZExt() && !Outs[0].Flags.isSExt())
1830 assert(DestVT == MVT::i32 && "ARM should always ext to i32");
1832 bool isZExt = Outs[0].Flags.isZExt();
1833 unsigned ResultReg = ARMEmitIntExt(RVVT, SrcReg, DestVT, isZExt);
1834 if (ResultReg == 0) return false;
1839 unsigned DstReg = VA.getLocReg();
1840 const TargetRegisterClass* SrcRC = MRI.getRegClass(SrcReg);
1841 // Avoid a cross-class copy. This is very unlikely.
1842 if (!SrcRC->contains(DstReg))
1844 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1845 DstReg).addReg(SrcReg);
1847 // Mark the register as live out of the function.
1848 MRI.addLiveOut(VA.getLocReg());
1851 unsigned RetOpc = isThumb2 ? ARM::tBX_RET : ARM::BX_RET;
1852 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1857 unsigned ARMFastISel::ARMSelectCallOp(const GlobalValue *GV) {
1859 // Darwin needs the r9 versions of the opcodes.
1860 bool isDarwin = Subtarget->isTargetDarwin();
1862 return isDarwin ? ARM::tBLr9 : ARM::tBL;
1864 return isDarwin ? ARM::BLr9 : ARM::BL;
1868 // A quick function that will emit a call for a named libcall in F with the
1869 // vector of passed arguments for the Instruction in I. We can assume that we
1870 // can emit a call for any libcall we can produce. This is an abridged version
1871 // of the full call infrastructure since we won't need to worry about things
1872 // like computed function pointers or strange arguments at call sites.
1873 // TODO: Try to unify this and the normal call bits for ARM, then try to unify
1875 bool ARMFastISel::ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call) {
1876 CallingConv::ID CC = TLI.getLibcallCallingConv(Call);
1878 // Handle *simple* calls for now.
1879 Type *RetTy = I->getType();
1881 if (RetTy->isVoidTy())
1882 RetVT = MVT::isVoid;
1883 else if (!isTypeLegal(RetTy, RetVT))
1886 // TODO: For now if we have long calls specified we don't handle the call.
1887 if (EnableARMLongCalls) return false;
1889 // Set up the argument vectors.
1890 SmallVector<Value*, 8> Args;
1891 SmallVector<unsigned, 8> ArgRegs;
1892 SmallVector<MVT, 8> ArgVTs;
1893 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
1894 Args.reserve(I->getNumOperands());
1895 ArgRegs.reserve(I->getNumOperands());
1896 ArgVTs.reserve(I->getNumOperands());
1897 ArgFlags.reserve(I->getNumOperands());
1898 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
1899 Value *Op = I->getOperand(i);
1900 unsigned Arg = getRegForValue(Op);
1901 if (Arg == 0) return false;
1903 Type *ArgTy = Op->getType();
1905 if (!isTypeLegal(ArgTy, ArgVT)) return false;
1907 ISD::ArgFlagsTy Flags;
1908 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
1909 Flags.setOrigAlign(OriginalAlignment);
1912 ArgRegs.push_back(Arg);
1913 ArgVTs.push_back(ArgVT);
1914 ArgFlags.push_back(Flags);
1917 // Handle the arguments now that we've gotten them.
1918 SmallVector<unsigned, 4> RegArgs;
1920 if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
1923 // Issue the call, BLr9 for darwin, BL otherwise.
1924 // TODO: Turn this into the table of arm call ops.
1925 MachineInstrBuilder MIB;
1926 unsigned CallOpc = ARMSelectCallOp(NULL);
1928 // Explicitly adding the predicate here.
1929 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1931 .addExternalSymbol(TLI.getLibcallName(Call));
1933 // Explicitly adding the predicate here.
1934 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1936 .addExternalSymbol(TLI.getLibcallName(Call)));
1938 // Add implicit physical register uses to the call.
1939 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
1940 MIB.addReg(RegArgs[i]);
1942 // Finish off the call including any return values.
1943 SmallVector<unsigned, 4> UsedRegs;
1944 if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false;
1946 // Set all unused physreg defs as dead.
1947 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
1952 bool ARMFastISel::SelectCall(const Instruction *I) {
1953 const CallInst *CI = cast<CallInst>(I);
1954 const Value *Callee = CI->getCalledValue();
1956 // Can't handle inline asm or worry about intrinsics yet.
1957 if (isa<InlineAsm>(Callee) || isa<IntrinsicInst>(CI)) return false;
1959 // Only handle global variable Callees.
1960 const GlobalValue *GV = dyn_cast<GlobalValue>(Callee);
1964 // Check the calling convention.
1965 ImmutableCallSite CS(CI);
1966 CallingConv::ID CC = CS.getCallingConv();
1968 // TODO: Avoid some calling conventions?
1970 // Let SDISel handle vararg functions.
1971 PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
1972 FunctionType *FTy = cast<FunctionType>(PT->getElementType());
1973 if (FTy->isVarArg())
1976 // Handle *simple* calls for now.
1977 Type *RetTy = I->getType();
1979 if (RetTy->isVoidTy())
1980 RetVT = MVT::isVoid;
1981 else if (!isTypeLegal(RetTy, RetVT) && RetVT != MVT::i16 &&
1982 RetVT != MVT::i8 && RetVT != MVT::i1)
1985 // TODO: For now if we have long calls specified we don't handle the call.
1986 if (EnableARMLongCalls) return false;
1988 // Set up the argument vectors.
1989 SmallVector<Value*, 8> Args;
1990 SmallVector<unsigned, 8> ArgRegs;
1991 SmallVector<MVT, 8> ArgVTs;
1992 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
1993 Args.reserve(CS.arg_size());
1994 ArgRegs.reserve(CS.arg_size());
1995 ArgVTs.reserve(CS.arg_size());
1996 ArgFlags.reserve(CS.arg_size());
1997 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
1999 unsigned Arg = getRegForValue(*i);
2003 ISD::ArgFlagsTy Flags;
2004 unsigned AttrInd = i - CS.arg_begin() + 1;
2005 if (CS.paramHasAttr(AttrInd, Attribute::SExt))
2007 if (CS.paramHasAttr(AttrInd, Attribute::ZExt))
2010 // FIXME: Only handle *easy* calls for now.
2011 if (CS.paramHasAttr(AttrInd, Attribute::InReg) ||
2012 CS.paramHasAttr(AttrInd, Attribute::StructRet) ||
2013 CS.paramHasAttr(AttrInd, Attribute::Nest) ||
2014 CS.paramHasAttr(AttrInd, Attribute::ByVal))
2017 Type *ArgTy = (*i)->getType();
2019 if (!isTypeLegal(ArgTy, ArgVT) && ArgVT != MVT::i16 && ArgVT != MVT::i8 &&
2022 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
2023 Flags.setOrigAlign(OriginalAlignment);
2026 ArgRegs.push_back(Arg);
2027 ArgVTs.push_back(ArgVT);
2028 ArgFlags.push_back(Flags);
2031 // Handle the arguments now that we've gotten them.
2032 SmallVector<unsigned, 4> RegArgs;
2034 if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
2037 // Issue the call, BLr9 for darwin, BL otherwise.
2038 // TODO: Turn this into the table of arm call ops.
2039 MachineInstrBuilder MIB;
2040 unsigned CallOpc = ARMSelectCallOp(GV);
2041 // Explicitly adding the predicate here.
2043 // Explicitly adding the predicate here.
2044 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
2046 .addGlobalAddress(GV, 0, 0);
2048 // Explicitly adding the predicate here.
2049 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
2051 .addGlobalAddress(GV, 0, 0));
2053 // Add implicit physical register uses to the call.
2054 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
2055 MIB.addReg(RegArgs[i]);
2057 // Finish off the call including any return values.
2058 SmallVector<unsigned, 4> UsedRegs;
2059 if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false;
2061 // Set all unused physreg defs as dead.
2062 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
2067 bool ARMFastISel::SelectTrunc(const Instruction *I) {
2068 // The high bits for a type smaller than the register size are assumed to be
2070 Value *Op = I->getOperand(0);
2073 SrcVT = TLI.getValueType(Op->getType(), true);
2074 DestVT = TLI.getValueType(I->getType(), true);
2076 if (SrcVT != MVT::i32 && SrcVT != MVT::i16 && SrcVT != MVT::i8)
2078 if (DestVT != MVT::i16 && DestVT != MVT::i8 && DestVT != MVT::i1)
2081 unsigned SrcReg = getRegForValue(Op);
2082 if (!SrcReg) return false;
2084 // Because the high bits are undefined, a truncate doesn't generate
2086 UpdateValueMap(I, SrcReg);
2090 unsigned ARMFastISel::ARMEmitIntExt(EVT SrcVT, unsigned SrcReg, EVT DestVT,
2092 if (DestVT != MVT::i32 && DestVT != MVT::i16 && DestVT != MVT::i8)
2096 bool isBoolZext = false;
2097 if (!SrcVT.isSimple()) return 0;
2098 switch (SrcVT.getSimpleVT().SimpleTy) {
2101 if (!Subtarget->hasV6Ops()) return 0;
2103 Opc = isThumb2 ? ARM::t2UXTH : ARM::UXTH;
2105 Opc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
2108 if (!Subtarget->hasV6Ops()) return 0;
2110 Opc = isThumb2 ? ARM::t2UXTB : ARM::UXTB;
2112 Opc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
2116 Opc = isThumb2 ? ARM::t2ANDri : ARM::ANDri;
2123 unsigned ResultReg = createResultReg(TLI.getRegClassFor(MVT::i32));
2124 MachineInstrBuilder MIB;
2125 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), ResultReg)
2131 AddOptionalDefs(MIB);
2135 bool ARMFastISel::SelectIntExt(const Instruction *I) {
2136 // On ARM, in general, integer casts don't involve legal types; this code
2137 // handles promotable integers.
2138 // FIXME: We could save an instruction in many cases by special-casing
2139 // load instructions.
2140 Type *DestTy = I->getType();
2141 Value *Src = I->getOperand(0);
2142 Type *SrcTy = Src->getType();
2145 SrcVT = TLI.getValueType(SrcTy, true);
2146 DestVT = TLI.getValueType(DestTy, true);
2148 bool isZExt = isa<ZExtInst>(I);
2149 unsigned SrcReg = getRegForValue(Src);
2150 if (!SrcReg) return false;
2152 unsigned ResultReg = ARMEmitIntExt(SrcVT, SrcReg, DestVT, isZExt);
2153 if (ResultReg == 0) return false;
2154 UpdateValueMap(I, ResultReg);
2158 // TODO: SoftFP support.
2159 bool ARMFastISel::TargetSelectInstruction(const Instruction *I) {
2161 switch (I->getOpcode()) {
2162 case Instruction::Load:
2163 return SelectLoad(I);
2164 case Instruction::Store:
2165 return SelectStore(I);
2166 case Instruction::Br:
2167 return SelectBranch(I);
2168 case Instruction::ICmp:
2169 case Instruction::FCmp:
2170 return SelectCmp(I);
2171 case Instruction::FPExt:
2172 return SelectFPExt(I);
2173 case Instruction::FPTrunc:
2174 return SelectFPTrunc(I);
2175 case Instruction::SIToFP:
2176 return SelectSIToFP(I);
2177 case Instruction::FPToSI:
2178 return SelectFPToSI(I);
2179 case Instruction::FAdd:
2180 return SelectBinaryOp(I, ISD::FADD);
2181 case Instruction::FSub:
2182 return SelectBinaryOp(I, ISD::FSUB);
2183 case Instruction::FMul:
2184 return SelectBinaryOp(I, ISD::FMUL);
2185 case Instruction::SDiv:
2186 return SelectSDiv(I);
2187 case Instruction::SRem:
2188 return SelectSRem(I);
2189 case Instruction::Call:
2190 return SelectCall(I);
2191 case Instruction::Select:
2192 return SelectSelect(I);
2193 case Instruction::Ret:
2194 return SelectRet(I);
2195 case Instruction::Trunc:
2196 return SelectTrunc(I);
2197 case Instruction::ZExt:
2198 case Instruction::SExt:
2199 return SelectIntExt(I);
2206 llvm::FastISel *ARM::createFastISel(FunctionLoweringInfo &funcInfo) {
2207 // Completely untested on non-darwin.
2208 const TargetMachine &TM = funcInfo.MF->getTarget();
2210 // Darwin and thumb1 only for now.
2211 const ARMSubtarget *Subtarget = &TM.getSubtarget<ARMSubtarget>();
2212 if (Subtarget->isTargetDarwin() && !Subtarget->isThumb1Only() &&
2213 !DisableARMFastISel)
2214 return new ARMFastISel(funcInfo);