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 unsigned Opc = isThumb2 ? ARM::t2MOVi16 : ARM::MOVi16;
556 unsigned ImmReg = createResultReg(TLI.getRegClassFor(MVT::i32));
557 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
558 TII.get(Opc), ImmReg)
559 .addImm(CI->getZExtValue()));
563 // Use MVN to emit negative constants.
564 if (VT == MVT::i32 && Subtarget->hasV6T2Ops() && CI->isNegative()) {
565 unsigned Imm = (unsigned)~(CI->getSExtValue());
566 bool EncodeImm = isThumb2 ? (ARM_AM::getT2SOImmVal(Imm) != -1) :
567 (ARM_AM::getSOImmVal(Imm) != -1);
569 unsigned Opc = isThumb2 ? ARM::t2MVNi : ARM::MVNi;
570 unsigned ImmReg = createResultReg(TLI.getRegClassFor(MVT::i32));
571 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
572 TII.get(Opc), ImmReg)
578 // Load from constant pool. For now 32-bit only.
582 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
584 // MachineConstantPool wants an explicit alignment.
585 unsigned Align = TD.getPrefTypeAlignment(C->getType());
587 // TODO: Figure out if this is correct.
588 Align = TD.getTypeAllocSize(C->getType());
590 unsigned Idx = MCP.getConstantPoolIndex(C, Align);
593 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
594 TII.get(ARM::t2LDRpci), DestReg)
595 .addConstantPoolIndex(Idx));
597 // The extra immediate is for addrmode2.
598 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
599 TII.get(ARM::LDRcp), DestReg)
600 .addConstantPoolIndex(Idx)
606 unsigned ARMFastISel::ARMMaterializeGV(const GlobalValue *GV, EVT VT) {
607 // For now 32-bit only.
608 if (VT != MVT::i32) return 0;
610 Reloc::Model RelocM = TM.getRelocationModel();
612 // TODO: Need more magic for ARM PIC.
613 if (!isThumb2 && (RelocM == Reloc::PIC_)) return 0;
615 // MachineConstantPool wants an explicit alignment.
616 unsigned Align = TD.getPrefTypeAlignment(GV->getType());
618 // TODO: Figure out if this is correct.
619 Align = TD.getTypeAllocSize(GV->getType());
623 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb() ? 4 : 8);
624 unsigned Id = AFI->createPICLabelUId();
625 ARMConstantPoolValue *CPV = ARMConstantPoolConstant::Create(GV, Id,
628 unsigned Idx = MCP.getConstantPoolIndex(CPV, Align);
631 MachineInstrBuilder MIB;
632 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
634 unsigned Opc = (RelocM != Reloc::PIC_) ? ARM::t2LDRpci : ARM::t2LDRpci_pic;
635 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), DestReg)
636 .addConstantPoolIndex(Idx);
637 if (RelocM == Reloc::PIC_)
640 // The extra immediate is for addrmode2.
641 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(ARM::LDRcp),
643 .addConstantPoolIndex(Idx)
646 AddOptionalDefs(MIB);
648 if (Subtarget->GVIsIndirectSymbol(GV, RelocM)) {
649 unsigned NewDestReg = createResultReg(TLI.getRegClassFor(VT));
651 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
652 TII.get(ARM::t2LDRi12), NewDestReg)
656 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(ARM::LDRi12),
660 DestReg = NewDestReg;
661 AddOptionalDefs(MIB);
667 unsigned ARMFastISel::TargetMaterializeConstant(const Constant *C) {
668 EVT VT = TLI.getValueType(C->getType(), true);
670 // Only handle simple types.
671 if (!VT.isSimple()) return 0;
673 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
674 return ARMMaterializeFP(CFP, VT);
675 else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
676 return ARMMaterializeGV(GV, VT);
677 else if (isa<ConstantInt>(C))
678 return ARMMaterializeInt(C, VT);
683 unsigned ARMFastISel::TargetMaterializeAlloca(const AllocaInst *AI) {
684 // Don't handle dynamic allocas.
685 if (!FuncInfo.StaticAllocaMap.count(AI)) return 0;
688 if (!isLoadTypeLegal(AI->getType(), VT)) return false;
690 DenseMap<const AllocaInst*, int>::iterator SI =
691 FuncInfo.StaticAllocaMap.find(AI);
693 // This will get lowered later into the correct offsets and registers
694 // via rewriteXFrameIndex.
695 if (SI != FuncInfo.StaticAllocaMap.end()) {
696 TargetRegisterClass* RC = TLI.getRegClassFor(VT);
697 unsigned ResultReg = createResultReg(RC);
698 unsigned Opc = isThumb2 ? ARM::t2ADDri : ARM::ADDri;
699 AddOptionalDefs(BuildMI(*FuncInfo.MBB, *FuncInfo.InsertPt, DL,
700 TII.get(Opc), ResultReg)
701 .addFrameIndex(SI->second)
709 bool ARMFastISel::isTypeLegal(Type *Ty, MVT &VT) {
710 EVT evt = TLI.getValueType(Ty, true);
712 // Only handle simple types.
713 if (evt == MVT::Other || !evt.isSimple()) return false;
714 VT = evt.getSimpleVT();
716 // Handle all legal types, i.e. a register that will directly hold this
718 return TLI.isTypeLegal(VT);
721 bool ARMFastISel::isLoadTypeLegal(Type *Ty, MVT &VT) {
722 if (isTypeLegal(Ty, VT)) return true;
724 // If this is a type than can be sign or zero-extended to a basic operation
725 // go ahead and accept it now.
726 if (VT == MVT::i8 || VT == MVT::i16)
732 // Computes the address to get to an object.
733 bool ARMFastISel::ARMComputeAddress(const Value *Obj, Address &Addr) {
734 // Some boilerplate from the X86 FastISel.
735 const User *U = NULL;
736 unsigned Opcode = Instruction::UserOp1;
737 if (const Instruction *I = dyn_cast<Instruction>(Obj)) {
738 // Don't walk into other basic blocks unless the object is an alloca from
739 // another block, otherwise it may not have a virtual register assigned.
740 if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(Obj)) ||
741 FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
742 Opcode = I->getOpcode();
745 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Obj)) {
746 Opcode = C->getOpcode();
750 if (PointerType *Ty = dyn_cast<PointerType>(Obj->getType()))
751 if (Ty->getAddressSpace() > 255)
752 // Fast instruction selection doesn't support the special
759 case Instruction::BitCast: {
760 // Look through bitcasts.
761 return ARMComputeAddress(U->getOperand(0), Addr);
763 case Instruction::IntToPtr: {
764 // Look past no-op inttoptrs.
765 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
766 return ARMComputeAddress(U->getOperand(0), Addr);
769 case Instruction::PtrToInt: {
770 // Look past no-op ptrtoints.
771 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
772 return ARMComputeAddress(U->getOperand(0), Addr);
775 case Instruction::GetElementPtr: {
776 Address SavedAddr = Addr;
777 int TmpOffset = Addr.Offset;
779 // Iterate through the GEP folding the constants into offsets where
781 gep_type_iterator GTI = gep_type_begin(U);
782 for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end();
783 i != e; ++i, ++GTI) {
784 const Value *Op = *i;
785 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
786 const StructLayout *SL = TD.getStructLayout(STy);
787 unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
788 TmpOffset += SL->getElementOffset(Idx);
790 uint64_t S = TD.getTypeAllocSize(GTI.getIndexedType());
792 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
793 // Constant-offset addressing.
794 TmpOffset += CI->getSExtValue() * S;
797 if (isa<AddOperator>(Op) &&
798 (!isa<Instruction>(Op) ||
799 FuncInfo.MBBMap[cast<Instruction>(Op)->getParent()]
801 isa<ConstantInt>(cast<AddOperator>(Op)->getOperand(1))) {
802 // An add (in the same block) with a constant operand. Fold the
805 cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
806 TmpOffset += CI->getSExtValue() * S;
807 // Iterate on the other operand.
808 Op = cast<AddOperator>(Op)->getOperand(0);
812 goto unsupported_gep;
817 // Try to grab the base operand now.
818 Addr.Offset = TmpOffset;
819 if (ARMComputeAddress(U->getOperand(0), Addr)) return true;
821 // We failed, restore everything and try the other options.
827 case Instruction::Alloca: {
828 const AllocaInst *AI = cast<AllocaInst>(Obj);
829 DenseMap<const AllocaInst*, int>::iterator SI =
830 FuncInfo.StaticAllocaMap.find(AI);
831 if (SI != FuncInfo.StaticAllocaMap.end()) {
832 Addr.BaseType = Address::FrameIndexBase;
833 Addr.Base.FI = SI->second;
840 // Materialize the global variable's address into a reg which can
841 // then be used later to load the variable.
842 if (const GlobalValue *GV = dyn_cast<GlobalValue>(Obj)) {
843 unsigned Tmp = ARMMaterializeGV(GV, TLI.getValueType(Obj->getType()));
844 if (Tmp == 0) return false;
850 // Try to get this in a register if nothing else has worked.
851 if (Addr.Base.Reg == 0) Addr.Base.Reg = getRegForValue(Obj);
852 return Addr.Base.Reg != 0;
855 void ARMFastISel::ARMSimplifyAddress(Address &Addr, EVT VT) {
857 assert(VT.isSimple() && "Non-simple types are invalid here!");
859 bool needsLowering = false;
860 switch (VT.getSimpleVT().SimpleTy) {
862 assert(false && "Unhandled load/store type!");
865 // Integer loads/stores handle 12-bit offsets.
866 needsLowering = ((Addr.Offset & 0xfff) != Addr.Offset);
868 // ARM i16 integer loads/stores handle +/-imm8 offsets.
869 // FIXME: Negative offsets require special handling.
870 if (Addr.Offset > 255 || Addr.Offset < 0)
871 needsLowering = true;
876 // Integer loads/stores handle 12-bit offsets.
877 needsLowering = ((Addr.Offset & 0xfff) != Addr.Offset);
881 // Floating point operands handle 8-bit offsets.
882 needsLowering = ((Addr.Offset & 0xff) != Addr.Offset);
886 // If this is a stack pointer and the offset needs to be simplified then
887 // put the alloca address into a register, set the base type back to
888 // register and continue. This should almost never happen.
889 if (needsLowering && Addr.BaseType == Address::FrameIndexBase) {
890 TargetRegisterClass *RC = isThumb2 ? ARM::tGPRRegisterClass :
891 ARM::GPRRegisterClass;
892 unsigned ResultReg = createResultReg(RC);
893 unsigned Opc = isThumb2 ? ARM::t2ADDri : ARM::ADDri;
894 AddOptionalDefs(BuildMI(*FuncInfo.MBB, *FuncInfo.InsertPt, DL,
895 TII.get(Opc), ResultReg)
896 .addFrameIndex(Addr.Base.FI)
898 Addr.Base.Reg = ResultReg;
899 Addr.BaseType = Address::RegBase;
902 // Since the offset is too large for the load/store instruction
903 // get the reg+offset into a register.
905 Addr.Base.Reg = FastEmit_ri_(MVT::i32, ISD::ADD, Addr.Base.Reg,
906 /*Op0IsKill*/false, Addr.Offset, MVT::i32);
911 void ARMFastISel::AddLoadStoreOperands(EVT VT, Address &Addr,
912 const MachineInstrBuilder &MIB,
914 // addrmode5 output depends on the selection dag addressing dividing the
915 // offset by 4 that it then later multiplies. Do this here as well.
916 if (VT.getSimpleVT().SimpleTy == MVT::f32 ||
917 VT.getSimpleVT().SimpleTy == MVT::f64)
920 // Frame base works a bit differently. Handle it separately.
921 if (Addr.BaseType == Address::FrameIndexBase) {
922 int FI = Addr.Base.FI;
923 int Offset = Addr.Offset;
924 MachineMemOperand *MMO =
925 FuncInfo.MF->getMachineMemOperand(
926 MachinePointerInfo::getFixedStack(FI, Offset),
928 MFI.getObjectSize(FI),
929 MFI.getObjectAlignment(FI));
930 // Now add the rest of the operands.
931 MIB.addFrameIndex(FI);
933 // ARM halfword load/stores need an additional operand.
934 if (!isThumb2 && VT.getSimpleVT().SimpleTy == MVT::i16) MIB.addReg(0);
936 MIB.addImm(Addr.Offset);
937 MIB.addMemOperand(MMO);
939 // Now add the rest of the operands.
940 MIB.addReg(Addr.Base.Reg);
942 // ARM halfword load/stores need an additional operand.
943 if (!isThumb2 && VT.getSimpleVT().SimpleTy == MVT::i16) MIB.addReg(0);
945 MIB.addImm(Addr.Offset);
947 AddOptionalDefs(MIB);
950 bool ARMFastISel::ARMEmitLoad(EVT VT, unsigned &ResultReg, Address &Addr) {
952 assert(VT.isSimple() && "Non-simple types are invalid here!");
954 TargetRegisterClass *RC;
955 switch (VT.getSimpleVT().SimpleTy) {
956 // This is mostly going to be Neon/vector support.
957 default: return false;
959 Opc = isThumb2 ? ARM::t2LDRBi12 : ARM::LDRBi12;
960 RC = ARM::GPRRegisterClass;
963 Opc = isThumb2 ? ARM::t2LDRHi12 : ARM::LDRH;
964 RC = ARM::GPRRegisterClass;
967 Opc = isThumb2 ? ARM::t2LDRi12 : ARM::LDRi12;
968 RC = ARM::GPRRegisterClass;
972 RC = TLI.getRegClassFor(VT);
976 RC = TLI.getRegClassFor(VT);
979 // Simplify this down to something we can handle.
980 ARMSimplifyAddress(Addr, VT);
982 // Create the base instruction, then add the operands.
983 ResultReg = createResultReg(RC);
984 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
985 TII.get(Opc), ResultReg);
986 AddLoadStoreOperands(VT, Addr, MIB, MachineMemOperand::MOLoad);
990 bool ARMFastISel::SelectLoad(const Instruction *I) {
991 // Atomic loads need special handling.
992 if (cast<LoadInst>(I)->isAtomic())
995 // Verify we have a legal type before going any further.
997 if (!isLoadTypeLegal(I->getType(), VT))
1000 // See if we can handle this address.
1002 if (!ARMComputeAddress(I->getOperand(0), Addr)) return false;
1005 if (!ARMEmitLoad(VT, ResultReg, Addr)) return false;
1006 UpdateValueMap(I, ResultReg);
1010 bool ARMFastISel::ARMEmitStore(EVT VT, unsigned SrcReg, Address &Addr) {
1012 switch (VT.getSimpleVT().SimpleTy) {
1013 // This is mostly going to be Neon/vector support.
1014 default: return false;
1016 unsigned Res = createResultReg(isThumb2 ? ARM::tGPRRegisterClass :
1017 ARM::GPRRegisterClass);
1018 unsigned Opc = isThumb2 ? ARM::t2ANDri : ARM::ANDri;
1019 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1021 .addReg(SrcReg).addImm(1));
1023 } // Fallthrough here.
1025 StrOpc = isThumb2 ? ARM::t2STRBi12 : ARM::STRBi12;
1028 StrOpc = isThumb2 ? ARM::t2STRHi12 : ARM::STRH;
1031 StrOpc = isThumb2 ? ARM::t2STRi12 : ARM::STRi12;
1034 if (!Subtarget->hasVFP2()) return false;
1035 StrOpc = ARM::VSTRS;
1038 if (!Subtarget->hasVFP2()) return false;
1039 StrOpc = ARM::VSTRD;
1042 // Simplify this down to something we can handle.
1043 ARMSimplifyAddress(Addr, VT);
1045 // Create the base instruction, then add the operands.
1046 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1048 .addReg(SrcReg, getKillRegState(true));
1049 AddLoadStoreOperands(VT, Addr, MIB, MachineMemOperand::MOStore);
1053 bool ARMFastISel::SelectStore(const Instruction *I) {
1054 Value *Op0 = I->getOperand(0);
1055 unsigned SrcReg = 0;
1057 // Atomic stores need special handling.
1058 if (cast<StoreInst>(I)->isAtomic())
1061 // Verify we have a legal type before going any further.
1063 if (!isLoadTypeLegal(I->getOperand(0)->getType(), VT))
1066 // Get the value to be stored into a register.
1067 SrcReg = getRegForValue(Op0);
1068 if (SrcReg == 0) return false;
1070 // See if we can handle this address.
1072 if (!ARMComputeAddress(I->getOperand(1), Addr))
1075 if (!ARMEmitStore(VT, SrcReg, Addr)) return false;
1079 static ARMCC::CondCodes getComparePred(CmpInst::Predicate Pred) {
1081 // Needs two compares...
1082 case CmpInst::FCMP_ONE:
1083 case CmpInst::FCMP_UEQ:
1085 // AL is our "false" for now. The other two need more compares.
1087 case CmpInst::ICMP_EQ:
1088 case CmpInst::FCMP_OEQ:
1090 case CmpInst::ICMP_SGT:
1091 case CmpInst::FCMP_OGT:
1093 case CmpInst::ICMP_SGE:
1094 case CmpInst::FCMP_OGE:
1096 case CmpInst::ICMP_UGT:
1097 case CmpInst::FCMP_UGT:
1099 case CmpInst::FCMP_OLT:
1101 case CmpInst::ICMP_ULE:
1102 case CmpInst::FCMP_OLE:
1104 case CmpInst::FCMP_ORD:
1106 case CmpInst::FCMP_UNO:
1108 case CmpInst::FCMP_UGE:
1110 case CmpInst::ICMP_SLT:
1111 case CmpInst::FCMP_ULT:
1113 case CmpInst::ICMP_SLE:
1114 case CmpInst::FCMP_ULE:
1116 case CmpInst::FCMP_UNE:
1117 case CmpInst::ICMP_NE:
1119 case CmpInst::ICMP_UGE:
1121 case CmpInst::ICMP_ULT:
1126 bool ARMFastISel::SelectBranch(const Instruction *I) {
1127 const BranchInst *BI = cast<BranchInst>(I);
1128 MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
1129 MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
1131 // Simple branch support.
1133 // If we can, avoid recomputing the compare - redoing it could lead to wonky
1135 if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
1136 if (CI->hasOneUse() && (CI->getParent() == I->getParent())) {
1138 // Get the compare predicate.
1139 // Try to take advantage of fallthrough opportunities.
1140 CmpInst::Predicate Predicate = CI->getPredicate();
1141 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
1142 std::swap(TBB, FBB);
1143 Predicate = CmpInst::getInversePredicate(Predicate);
1146 ARMCC::CondCodes ARMPred = getComparePred(Predicate);
1148 // We may not handle every CC for now.
1149 if (ARMPred == ARMCC::AL) return false;
1151 // Emit the compare.
1152 if (!ARMEmitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
1155 unsigned BrOpc = isThumb2 ? ARM::t2Bcc : ARM::Bcc;
1156 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
1157 .addMBB(TBB).addImm(ARMPred).addReg(ARM::CPSR);
1158 FastEmitBranch(FBB, DL);
1159 FuncInfo.MBB->addSuccessor(TBB);
1162 } else if (TruncInst *TI = dyn_cast<TruncInst>(BI->getCondition())) {
1164 if (TI->hasOneUse() && TI->getParent() == I->getParent() &&
1165 (isLoadTypeLegal(TI->getOperand(0)->getType(), SourceVT))) {
1166 unsigned TstOpc = isThumb2 ? ARM::t2TSTri : ARM::TSTri;
1167 unsigned OpReg = getRegForValue(TI->getOperand(0));
1168 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1170 .addReg(OpReg).addImm(1));
1172 unsigned CCMode = ARMCC::NE;
1173 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
1174 std::swap(TBB, FBB);
1178 unsigned BrOpc = isThumb2 ? ARM::t2Bcc : ARM::Bcc;
1179 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
1180 .addMBB(TBB).addImm(CCMode).addReg(ARM::CPSR);
1182 FastEmitBranch(FBB, DL);
1183 FuncInfo.MBB->addSuccessor(TBB);
1186 } else if (const ConstantInt *CI =
1187 dyn_cast<ConstantInt>(BI->getCondition())) {
1188 uint64_t Imm = CI->getZExtValue();
1189 MachineBasicBlock *Target = (Imm == 0) ? FBB : TBB;
1190 FastEmitBranch(Target, DL);
1194 unsigned CmpReg = getRegForValue(BI->getCondition());
1195 if (CmpReg == 0) return false;
1197 // We've been divorced from our compare! Our block was split, and
1198 // now our compare lives in a predecessor block. We musn't
1199 // re-compare here, as the children of the compare aren't guaranteed
1200 // live across the block boundary (we *could* check for this).
1201 // Regardless, the compare has been done in the predecessor block,
1202 // and it left a value for us in a virtual register. Ergo, we test
1203 // the one-bit value left in the virtual register.
1204 unsigned TstOpc = isThumb2 ? ARM::t2TSTri : ARM::TSTri;
1205 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TstOpc))
1206 .addReg(CmpReg).addImm(1));
1208 unsigned CCMode = ARMCC::NE;
1209 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
1210 std::swap(TBB, FBB);
1214 unsigned BrOpc = isThumb2 ? ARM::t2Bcc : ARM::Bcc;
1215 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
1216 .addMBB(TBB).addImm(CCMode).addReg(ARM::CPSR);
1217 FastEmitBranch(FBB, DL);
1218 FuncInfo.MBB->addSuccessor(TBB);
1222 bool ARMFastISel::ARMEmitCmp(const Value *Src1Value, const Value *Src2Value,
1224 Type *Ty = Src1Value->getType();
1225 EVT SrcVT = TLI.getValueType(Ty, true);
1226 if (!SrcVT.isSimple()) return false;
1228 bool isFloat = (Ty->isFloatTy() || Ty->isDoubleTy());
1229 if (isFloat && !Subtarget->hasVFP2())
1232 // Check to see if the 2nd operand is a constant that we can encode directly
1235 bool EncodeImm = false;
1236 bool isNegativeImm = false;
1237 if (const ConstantInt *ConstInt = dyn_cast<ConstantInt>(Src2Value)) {
1238 if (SrcVT == MVT::i32 || SrcVT == MVT::i16 || SrcVT == MVT::i8 ||
1240 const APInt &CIVal = ConstInt->getValue();
1241 EncodedImm = (isZExt) ? (int)CIVal.getZExtValue() : (int)CIVal.getSExtValue();
1242 if (EncodedImm < 0) {
1243 isNegativeImm = true;
1244 EncodedImm = -EncodedImm;
1246 EncodeImm = isThumb2 ? (ARM_AM::getT2SOImmVal(EncodedImm) != -1) :
1247 (ARM_AM::getSOImmVal(EncodedImm) != -1);
1249 } else if (const ConstantFP *ConstFP = dyn_cast<ConstantFP>(Src2Value)) {
1250 if (SrcVT == MVT::f32 || SrcVT == MVT::f64)
1251 if (ConstFP->isZero() && !ConstFP->isNegative())
1257 bool needsExt = false;
1258 switch (SrcVT.getSimpleVT().SimpleTy) {
1259 default: return false;
1260 // TODO: Verify compares.
1263 CmpOpc = EncodeImm ? ARM::VCMPEZS : ARM::VCMPES;
1267 CmpOpc = EncodeImm ? ARM::VCMPEZD : ARM::VCMPED;
1273 // Intentional fall-through.
1277 CmpOpc = ARM::t2CMPrr;
1279 CmpOpc = isNegativeImm ? ARM::t2CMNzri : ARM::t2CMPri;
1282 CmpOpc = ARM::CMPrr;
1284 CmpOpc = isNegativeImm ? ARM::CMNzri : ARM::CMPri;
1289 unsigned SrcReg1 = getRegForValue(Src1Value);
1290 if (SrcReg1 == 0) return false;
1294 SrcReg2 = getRegForValue(Src2Value);
1295 if (SrcReg2 == 0) return false;
1298 // We have i1, i8, or i16, we need to either zero extend or sign extend.
1301 ResultReg = ARMEmitIntExt(SrcVT, SrcReg1, MVT::i32, isZExt);
1302 if (ResultReg == 0) return false;
1303 SrcReg1 = ResultReg;
1305 ResultReg = ARMEmitIntExt(SrcVT, SrcReg2, MVT::i32, isZExt);
1306 if (ResultReg == 0) return false;
1307 SrcReg2 = ResultReg;
1312 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1314 .addReg(SrcReg1).addReg(SrcReg2));
1316 MachineInstrBuilder MIB;
1317 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
1320 // Only add immediate for icmp as the immediate for fcmp is an implicit 0.0.
1322 MIB.addImm(EncodedImm);
1323 AddOptionalDefs(MIB);
1326 // For floating point we need to move the result to a comparison register
1327 // that we can then use for branches.
1328 if (Ty->isFloatTy() || Ty->isDoubleTy())
1329 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1330 TII.get(ARM::FMSTAT)));
1334 bool ARMFastISel::SelectCmp(const Instruction *I) {
1335 const CmpInst *CI = cast<CmpInst>(I);
1336 Type *Ty = CI->getOperand(0)->getType();
1338 // Get the compare predicate.
1339 ARMCC::CondCodes ARMPred = getComparePred(CI->getPredicate());
1341 // We may not handle every CC for now.
1342 if (ARMPred == ARMCC::AL) return false;
1344 // Emit the compare.
1345 if (!ARMEmitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
1348 // Now set a register based on the comparison. Explicitly set the predicates
1350 unsigned MovCCOpc = isThumb2 ? ARM::t2MOVCCi : ARM::MOVCCi;
1351 TargetRegisterClass *RC = isThumb2 ? ARM::rGPRRegisterClass
1352 : ARM::GPRRegisterClass;
1353 unsigned DestReg = createResultReg(RC);
1354 Constant *Zero = ConstantInt::get(Type::getInt32Ty(*Context), 0);
1355 unsigned ZeroReg = TargetMaterializeConstant(Zero);
1356 bool isFloat = (Ty->isFloatTy() || Ty->isDoubleTy());
1357 unsigned CondReg = isFloat ? ARM::FPSCR : ARM::CPSR;
1358 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), DestReg)
1359 .addReg(ZeroReg).addImm(1)
1360 .addImm(ARMPred).addReg(CondReg);
1362 UpdateValueMap(I, DestReg);
1366 bool ARMFastISel::SelectFPExt(const Instruction *I) {
1367 // Make sure we have VFP and that we're extending float to double.
1368 if (!Subtarget->hasVFP2()) return false;
1370 Value *V = I->getOperand(0);
1371 if (!I->getType()->isDoubleTy() ||
1372 !V->getType()->isFloatTy()) return false;
1374 unsigned Op = getRegForValue(V);
1375 if (Op == 0) return false;
1377 unsigned Result = createResultReg(ARM::DPRRegisterClass);
1378 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1379 TII.get(ARM::VCVTDS), Result)
1381 UpdateValueMap(I, Result);
1385 bool ARMFastISel::SelectFPTrunc(const Instruction *I) {
1386 // Make sure we have VFP and that we're truncating double to float.
1387 if (!Subtarget->hasVFP2()) return false;
1389 Value *V = I->getOperand(0);
1390 if (!(I->getType()->isFloatTy() &&
1391 V->getType()->isDoubleTy())) return false;
1393 unsigned Op = getRegForValue(V);
1394 if (Op == 0) return false;
1396 unsigned Result = createResultReg(ARM::SPRRegisterClass);
1397 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1398 TII.get(ARM::VCVTSD), Result)
1400 UpdateValueMap(I, Result);
1404 bool ARMFastISel::SelectSIToFP(const Instruction *I) {
1405 // Make sure we have VFP.
1406 if (!Subtarget->hasVFP2()) return false;
1409 Type *Ty = I->getType();
1410 if (!isTypeLegal(Ty, DstVT))
1413 Value *Src = I->getOperand(0);
1414 EVT SrcVT = TLI.getValueType(Src->getType(), true);
1415 if (SrcVT != MVT::i32 && SrcVT != MVT::i16 && SrcVT != MVT::i8)
1418 unsigned SrcReg = getRegForValue(Src);
1419 if (SrcReg == 0) return false;
1421 // Handle sign-extension.
1422 if (SrcVT == MVT::i16 || SrcVT == MVT::i8) {
1423 EVT DestVT = MVT::i32;
1424 unsigned ResultReg = ARMEmitIntExt(SrcVT, SrcReg, DestVT, /*isZExt*/ false);
1425 if (ResultReg == 0) return false;
1429 // The conversion routine works on fp-reg to fp-reg and the operand above
1430 // was an integer, move it to the fp registers if possible.
1431 unsigned FP = ARMMoveToFPReg(MVT::f32, SrcReg);
1432 if (FP == 0) return false;
1435 if (Ty->isFloatTy()) Opc = ARM::VSITOS;
1436 else if (Ty->isDoubleTy()) Opc = ARM::VSITOD;
1439 unsigned ResultReg = createResultReg(TLI.getRegClassFor(DstVT));
1440 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
1443 UpdateValueMap(I, ResultReg);
1447 bool ARMFastISel::SelectFPToSI(const Instruction *I) {
1448 // Make sure we have VFP.
1449 if (!Subtarget->hasVFP2()) return false;
1452 Type *RetTy = I->getType();
1453 if (!isTypeLegal(RetTy, DstVT))
1456 unsigned Op = getRegForValue(I->getOperand(0));
1457 if (Op == 0) return false;
1460 Type *OpTy = I->getOperand(0)->getType();
1461 if (OpTy->isFloatTy()) Opc = ARM::VTOSIZS;
1462 else if (OpTy->isDoubleTy()) Opc = ARM::VTOSIZD;
1465 // f64->s32 or f32->s32 both need an intermediate f32 reg.
1466 unsigned ResultReg = createResultReg(TLI.getRegClassFor(MVT::f32));
1467 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
1471 // This result needs to be in an integer register, but the conversion only
1472 // takes place in fp-regs.
1473 unsigned IntReg = ARMMoveToIntReg(DstVT, ResultReg);
1474 if (IntReg == 0) return false;
1476 UpdateValueMap(I, IntReg);
1480 bool ARMFastISel::SelectSelect(const Instruction *I) {
1482 if (!isTypeLegal(I->getType(), VT))
1485 // Things need to be register sized for register moves.
1486 if (VT != MVT::i32) return false;
1487 const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
1489 unsigned CondReg = getRegForValue(I->getOperand(0));
1490 if (CondReg == 0) return false;
1491 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1492 if (Op1Reg == 0) return false;
1493 unsigned Op2Reg = getRegForValue(I->getOperand(2));
1494 if (Op2Reg == 0) return false;
1496 unsigned CmpOpc = isThumb2 ? ARM::t2TSTri : ARM::TSTri;
1497 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
1498 .addReg(CondReg).addImm(1));
1499 unsigned ResultReg = createResultReg(RC);
1500 unsigned MovCCOpc = isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr;
1501 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), ResultReg)
1502 .addReg(Op1Reg).addReg(Op2Reg)
1503 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
1504 UpdateValueMap(I, ResultReg);
1508 bool ARMFastISel::SelectSDiv(const Instruction *I) {
1510 Type *Ty = I->getType();
1511 if (!isTypeLegal(Ty, VT))
1514 // If we have integer div support we should have selected this automagically.
1515 // In case we have a real miss go ahead and return false and we'll pick
1517 if (Subtarget->hasDivide()) return false;
1519 // Otherwise emit a libcall.
1520 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
1522 LC = RTLIB::SDIV_I8;
1523 else if (VT == MVT::i16)
1524 LC = RTLIB::SDIV_I16;
1525 else if (VT == MVT::i32)
1526 LC = RTLIB::SDIV_I32;
1527 else if (VT == MVT::i64)
1528 LC = RTLIB::SDIV_I64;
1529 else if (VT == MVT::i128)
1530 LC = RTLIB::SDIV_I128;
1531 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SDIV!");
1533 return ARMEmitLibcall(I, LC);
1536 bool ARMFastISel::SelectSRem(const Instruction *I) {
1538 Type *Ty = I->getType();
1539 if (!isTypeLegal(Ty, VT))
1542 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
1544 LC = RTLIB::SREM_I8;
1545 else if (VT == MVT::i16)
1546 LC = RTLIB::SREM_I16;
1547 else if (VT == MVT::i32)
1548 LC = RTLIB::SREM_I32;
1549 else if (VT == MVT::i64)
1550 LC = RTLIB::SREM_I64;
1551 else if (VT == MVT::i128)
1552 LC = RTLIB::SREM_I128;
1553 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SREM!");
1555 return ARMEmitLibcall(I, LC);
1558 bool ARMFastISel::SelectBinaryOp(const Instruction *I, unsigned ISDOpcode) {
1559 EVT VT = TLI.getValueType(I->getType(), true);
1561 // We can get here in the case when we want to use NEON for our fp
1562 // operations, but can't figure out how to. Just use the vfp instructions
1564 // FIXME: It'd be nice to use NEON instructions.
1565 Type *Ty = I->getType();
1566 bool isFloat = (Ty->isDoubleTy() || Ty->isFloatTy());
1567 if (isFloat && !Subtarget->hasVFP2())
1570 unsigned Op1 = getRegForValue(I->getOperand(0));
1571 if (Op1 == 0) return false;
1573 unsigned Op2 = getRegForValue(I->getOperand(1));
1574 if (Op2 == 0) return false;
1577 bool is64bit = VT == MVT::f64 || VT == MVT::i64;
1578 switch (ISDOpcode) {
1579 default: return false;
1581 Opc = is64bit ? ARM::VADDD : ARM::VADDS;
1584 Opc = is64bit ? ARM::VSUBD : ARM::VSUBS;
1587 Opc = is64bit ? ARM::VMULD : ARM::VMULS;
1590 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
1591 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1592 TII.get(Opc), ResultReg)
1593 .addReg(Op1).addReg(Op2));
1594 UpdateValueMap(I, ResultReg);
1598 // Call Handling Code
1600 bool ARMFastISel::FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src,
1601 EVT SrcVT, unsigned &ResultReg) {
1602 unsigned RR = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), Opc,
1603 Src, /*TODO: Kill=*/false);
1612 // This is largely taken directly from CCAssignFnForNode - we don't support
1613 // varargs in FastISel so that part has been removed.
1614 // TODO: We may not support all of this.
1615 CCAssignFn *ARMFastISel::CCAssignFnForCall(CallingConv::ID CC, bool Return) {
1618 llvm_unreachable("Unsupported calling convention");
1619 case CallingConv::Fast:
1620 // Ignore fastcc. Silence compiler warnings.
1621 (void)RetFastCC_ARM_APCS;
1622 (void)FastCC_ARM_APCS;
1624 case CallingConv::C:
1625 // Use target triple & subtarget features to do actual dispatch.
1626 if (Subtarget->isAAPCS_ABI()) {
1627 if (Subtarget->hasVFP2() &&
1628 FloatABIType == FloatABI::Hard)
1629 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
1631 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
1633 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
1634 case CallingConv::ARM_AAPCS_VFP:
1635 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
1636 case CallingConv::ARM_AAPCS:
1637 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
1638 case CallingConv::ARM_APCS:
1639 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
1643 bool ARMFastISel::ProcessCallArgs(SmallVectorImpl<Value*> &Args,
1644 SmallVectorImpl<unsigned> &ArgRegs,
1645 SmallVectorImpl<MVT> &ArgVTs,
1646 SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
1647 SmallVectorImpl<unsigned> &RegArgs,
1649 unsigned &NumBytes) {
1650 SmallVector<CCValAssign, 16> ArgLocs;
1651 CCState CCInfo(CC, false, *FuncInfo.MF, TM, ArgLocs, *Context);
1652 CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CCAssignFnForCall(CC, false));
1654 // Get a count of how many bytes are to be pushed on the stack.
1655 NumBytes = CCInfo.getNextStackOffset();
1657 // Issue CALLSEQ_START
1658 unsigned AdjStackDown = TII.getCallFrameSetupOpcode();
1659 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1660 TII.get(AdjStackDown))
1663 // Process the args.
1664 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1665 CCValAssign &VA = ArgLocs[i];
1666 unsigned Arg = ArgRegs[VA.getValNo()];
1667 MVT ArgVT = ArgVTs[VA.getValNo()];
1669 // We don't handle NEON/vector parameters yet.
1670 if (ArgVT.isVector() || ArgVT.getSizeInBits() > 64)
1673 // Handle arg promotion, etc.
1674 switch (VA.getLocInfo()) {
1675 case CCValAssign::Full: break;
1676 case CCValAssign::SExt: {
1677 EVT DestVT = VA.getLocVT();
1678 unsigned ResultReg = ARMEmitIntExt(ArgVT, Arg, DestVT,
1680 assert (ResultReg != 0 && "Failed to emit a sext");
1684 case CCValAssign::AExt:
1685 // Intentional fall-through. Handle AExt and ZExt.
1686 case CCValAssign::ZExt: {
1687 EVT DestVT = VA.getLocVT();
1688 unsigned ResultReg = ARMEmitIntExt(ArgVT, Arg, DestVT,
1690 assert (ResultReg != 0 && "Failed to emit a sext");
1694 case CCValAssign::BCvt: {
1695 unsigned BC = FastEmit_r(ArgVT, VA.getLocVT(), ISD::BITCAST, Arg,
1696 /*TODO: Kill=*/false);
1697 assert(BC != 0 && "Failed to emit a bitcast!");
1699 ArgVT = VA.getLocVT();
1702 default: llvm_unreachable("Unknown arg promotion!");
1705 // Now copy/store arg to correct locations.
1706 if (VA.isRegLoc() && !VA.needsCustom()) {
1707 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1710 RegArgs.push_back(VA.getLocReg());
1711 } else if (VA.needsCustom()) {
1712 // TODO: We need custom lowering for vector (v2f64) args.
1713 if (VA.getLocVT() != MVT::f64) return false;
1715 CCValAssign &NextVA = ArgLocs[++i];
1717 // TODO: Only handle register args for now.
1718 if(!(VA.isRegLoc() && NextVA.isRegLoc())) return false;
1720 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1721 TII.get(ARM::VMOVRRD), VA.getLocReg())
1722 .addReg(NextVA.getLocReg(), RegState::Define)
1724 RegArgs.push_back(VA.getLocReg());
1725 RegArgs.push_back(NextVA.getLocReg());
1727 assert(VA.isMemLoc());
1728 // Need to store on the stack.
1730 Addr.BaseType = Address::RegBase;
1731 Addr.Base.Reg = ARM::SP;
1732 Addr.Offset = VA.getLocMemOffset();
1734 if (!ARMEmitStore(ArgVT, Arg, Addr)) return false;
1740 bool ARMFastISel::FinishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
1741 const Instruction *I, CallingConv::ID CC,
1742 unsigned &NumBytes) {
1743 // Issue CALLSEQ_END
1744 unsigned AdjStackUp = TII.getCallFrameDestroyOpcode();
1745 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1746 TII.get(AdjStackUp))
1747 .addImm(NumBytes).addImm(0));
1749 // Now the return value.
1750 if (RetVT != MVT::isVoid) {
1751 SmallVector<CCValAssign, 16> RVLocs;
1752 CCState CCInfo(CC, false, *FuncInfo.MF, TM, RVLocs, *Context);
1753 CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC, true));
1755 // Copy all of the result registers out of their specified physreg.
1756 if (RVLocs.size() == 2 && RetVT == MVT::f64) {
1757 // For this move we copy into two registers and then move into the
1758 // double fp reg we want.
1759 EVT DestVT = RVLocs[0].getValVT();
1760 TargetRegisterClass* DstRC = TLI.getRegClassFor(DestVT);
1761 unsigned ResultReg = createResultReg(DstRC);
1762 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1763 TII.get(ARM::VMOVDRR), ResultReg)
1764 .addReg(RVLocs[0].getLocReg())
1765 .addReg(RVLocs[1].getLocReg()));
1767 UsedRegs.push_back(RVLocs[0].getLocReg());
1768 UsedRegs.push_back(RVLocs[1].getLocReg());
1770 // Finally update the result.
1771 UpdateValueMap(I, ResultReg);
1773 assert(RVLocs.size() == 1 &&"Can't handle non-double multi-reg retvals!");
1774 EVT CopyVT = RVLocs[0].getValVT();
1776 // Special handling for extended integers.
1777 if (RetVT == MVT::i1 || RetVT == MVT::i8 || RetVT == MVT::i16)
1780 TargetRegisterClass* DstRC = TLI.getRegClassFor(CopyVT);
1782 unsigned ResultReg = createResultReg(DstRC);
1783 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1784 ResultReg).addReg(RVLocs[0].getLocReg());
1785 UsedRegs.push_back(RVLocs[0].getLocReg());
1787 // Finally update the result.
1788 UpdateValueMap(I, ResultReg);
1795 bool ARMFastISel::SelectRet(const Instruction *I) {
1796 const ReturnInst *Ret = cast<ReturnInst>(I);
1797 const Function &F = *I->getParent()->getParent();
1799 if (!FuncInfo.CanLowerReturn)
1805 CallingConv::ID CC = F.getCallingConv();
1806 if (Ret->getNumOperands() > 0) {
1807 SmallVector<ISD::OutputArg, 4> Outs;
1808 GetReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(),
1811 // Analyze operands of the call, assigning locations to each operand.
1812 SmallVector<CCValAssign, 16> ValLocs;
1813 CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, TM, ValLocs,I->getContext());
1814 CCInfo.AnalyzeReturn(Outs, CCAssignFnForCall(CC, true /* is Ret */));
1816 const Value *RV = Ret->getOperand(0);
1817 unsigned Reg = getRegForValue(RV);
1821 // Only handle a single return value for now.
1822 if (ValLocs.size() != 1)
1825 CCValAssign &VA = ValLocs[0];
1827 // Don't bother handling odd stuff for now.
1828 if (VA.getLocInfo() != CCValAssign::Full)
1830 // Only handle register returns for now.
1834 unsigned SrcReg = Reg + VA.getValNo();
1835 EVT RVVT = TLI.getValueType(RV->getType());
1836 EVT DestVT = VA.getValVT();
1837 // Special handling for extended integers.
1838 if (RVVT != DestVT) {
1839 if (RVVT != MVT::i1 && RVVT != MVT::i8 && RVVT != MVT::i16)
1842 if (!Outs[0].Flags.isZExt() && !Outs[0].Flags.isSExt())
1845 assert(DestVT == MVT::i32 && "ARM should always ext to i32");
1847 bool isZExt = Outs[0].Flags.isZExt();
1848 unsigned ResultReg = ARMEmitIntExt(RVVT, SrcReg, DestVT, isZExt);
1849 if (ResultReg == 0) return false;
1854 unsigned DstReg = VA.getLocReg();
1855 const TargetRegisterClass* SrcRC = MRI.getRegClass(SrcReg);
1856 // Avoid a cross-class copy. This is very unlikely.
1857 if (!SrcRC->contains(DstReg))
1859 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1860 DstReg).addReg(SrcReg);
1862 // Mark the register as live out of the function.
1863 MRI.addLiveOut(VA.getLocReg());
1866 unsigned RetOpc = isThumb2 ? ARM::tBX_RET : ARM::BX_RET;
1867 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1872 unsigned ARMFastISel::ARMSelectCallOp(const GlobalValue *GV) {
1874 // Darwin needs the r9 versions of the opcodes.
1875 bool isDarwin = Subtarget->isTargetDarwin();
1877 return isDarwin ? ARM::tBLr9 : ARM::tBL;
1879 return isDarwin ? ARM::BLr9 : ARM::BL;
1883 // A quick function that will emit a call for a named libcall in F with the
1884 // vector of passed arguments for the Instruction in I. We can assume that we
1885 // can emit a call for any libcall we can produce. This is an abridged version
1886 // of the full call infrastructure since we won't need to worry about things
1887 // like computed function pointers or strange arguments at call sites.
1888 // TODO: Try to unify this and the normal call bits for ARM, then try to unify
1890 bool ARMFastISel::ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call) {
1891 CallingConv::ID CC = TLI.getLibcallCallingConv(Call);
1893 // Handle *simple* calls for now.
1894 Type *RetTy = I->getType();
1896 if (RetTy->isVoidTy())
1897 RetVT = MVT::isVoid;
1898 else if (!isTypeLegal(RetTy, RetVT))
1901 // TODO: For now if we have long calls specified we don't handle the call.
1902 if (EnableARMLongCalls) return false;
1904 // Set up the argument vectors.
1905 SmallVector<Value*, 8> Args;
1906 SmallVector<unsigned, 8> ArgRegs;
1907 SmallVector<MVT, 8> ArgVTs;
1908 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
1909 Args.reserve(I->getNumOperands());
1910 ArgRegs.reserve(I->getNumOperands());
1911 ArgVTs.reserve(I->getNumOperands());
1912 ArgFlags.reserve(I->getNumOperands());
1913 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
1914 Value *Op = I->getOperand(i);
1915 unsigned Arg = getRegForValue(Op);
1916 if (Arg == 0) return false;
1918 Type *ArgTy = Op->getType();
1920 if (!isTypeLegal(ArgTy, ArgVT)) return false;
1922 ISD::ArgFlagsTy Flags;
1923 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
1924 Flags.setOrigAlign(OriginalAlignment);
1927 ArgRegs.push_back(Arg);
1928 ArgVTs.push_back(ArgVT);
1929 ArgFlags.push_back(Flags);
1932 // Handle the arguments now that we've gotten them.
1933 SmallVector<unsigned, 4> RegArgs;
1935 if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
1938 // Issue the call, BLr9 for darwin, BL otherwise.
1939 // TODO: Turn this into the table of arm call ops.
1940 MachineInstrBuilder MIB;
1941 unsigned CallOpc = ARMSelectCallOp(NULL);
1943 // Explicitly adding the predicate here.
1944 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1946 .addExternalSymbol(TLI.getLibcallName(Call));
1948 // Explicitly adding the predicate here.
1949 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1951 .addExternalSymbol(TLI.getLibcallName(Call)));
1953 // Add implicit physical register uses to the call.
1954 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
1955 MIB.addReg(RegArgs[i]);
1957 // Finish off the call including any return values.
1958 SmallVector<unsigned, 4> UsedRegs;
1959 if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false;
1961 // Set all unused physreg defs as dead.
1962 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
1967 bool ARMFastISel::SelectCall(const Instruction *I) {
1968 const CallInst *CI = cast<CallInst>(I);
1969 const Value *Callee = CI->getCalledValue();
1971 // Can't handle inline asm or worry about intrinsics yet.
1972 if (isa<InlineAsm>(Callee) || isa<IntrinsicInst>(CI)) return false;
1974 // Only handle global variable Callees.
1975 const GlobalValue *GV = dyn_cast<GlobalValue>(Callee);
1979 // Check the calling convention.
1980 ImmutableCallSite CS(CI);
1981 CallingConv::ID CC = CS.getCallingConv();
1983 // TODO: Avoid some calling conventions?
1985 // Let SDISel handle vararg functions.
1986 PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
1987 FunctionType *FTy = cast<FunctionType>(PT->getElementType());
1988 if (FTy->isVarArg())
1991 // Handle *simple* calls for now.
1992 Type *RetTy = I->getType();
1994 if (RetTy->isVoidTy())
1995 RetVT = MVT::isVoid;
1996 else if (!isTypeLegal(RetTy, RetVT) && RetVT != MVT::i16 &&
1997 RetVT != MVT::i8 && RetVT != MVT::i1)
2000 // TODO: For now if we have long calls specified we don't handle the call.
2001 if (EnableARMLongCalls) return false;
2003 // Set up the argument vectors.
2004 SmallVector<Value*, 8> Args;
2005 SmallVector<unsigned, 8> ArgRegs;
2006 SmallVector<MVT, 8> ArgVTs;
2007 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
2008 Args.reserve(CS.arg_size());
2009 ArgRegs.reserve(CS.arg_size());
2010 ArgVTs.reserve(CS.arg_size());
2011 ArgFlags.reserve(CS.arg_size());
2012 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
2014 unsigned Arg = getRegForValue(*i);
2018 ISD::ArgFlagsTy Flags;
2019 unsigned AttrInd = i - CS.arg_begin() + 1;
2020 if (CS.paramHasAttr(AttrInd, Attribute::SExt))
2022 if (CS.paramHasAttr(AttrInd, Attribute::ZExt))
2025 // FIXME: Only handle *easy* calls for now.
2026 if (CS.paramHasAttr(AttrInd, Attribute::InReg) ||
2027 CS.paramHasAttr(AttrInd, Attribute::StructRet) ||
2028 CS.paramHasAttr(AttrInd, Attribute::Nest) ||
2029 CS.paramHasAttr(AttrInd, Attribute::ByVal))
2032 Type *ArgTy = (*i)->getType();
2034 if (!isTypeLegal(ArgTy, ArgVT) && ArgVT != MVT::i16 && ArgVT != MVT::i8 &&
2037 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
2038 Flags.setOrigAlign(OriginalAlignment);
2041 ArgRegs.push_back(Arg);
2042 ArgVTs.push_back(ArgVT);
2043 ArgFlags.push_back(Flags);
2046 // Handle the arguments now that we've gotten them.
2047 SmallVector<unsigned, 4> RegArgs;
2049 if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
2052 // Issue the call, BLr9 for darwin, BL otherwise.
2053 // TODO: Turn this into the table of arm call ops.
2054 MachineInstrBuilder MIB;
2055 unsigned CallOpc = ARMSelectCallOp(GV);
2056 // Explicitly adding the predicate here.
2058 // Explicitly adding the predicate here.
2059 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
2061 .addGlobalAddress(GV, 0, 0);
2063 // Explicitly adding the predicate here.
2064 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
2066 .addGlobalAddress(GV, 0, 0));
2068 // Add implicit physical register uses to the call.
2069 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
2070 MIB.addReg(RegArgs[i]);
2072 // Finish off the call including any return values.
2073 SmallVector<unsigned, 4> UsedRegs;
2074 if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false;
2076 // Set all unused physreg defs as dead.
2077 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
2082 bool ARMFastISel::SelectTrunc(const Instruction *I) {
2083 // The high bits for a type smaller than the register size are assumed to be
2085 Value *Op = I->getOperand(0);
2088 SrcVT = TLI.getValueType(Op->getType(), true);
2089 DestVT = TLI.getValueType(I->getType(), true);
2091 if (SrcVT != MVT::i32 && SrcVT != MVT::i16 && SrcVT != MVT::i8)
2093 if (DestVT != MVT::i16 && DestVT != MVT::i8 && DestVT != MVT::i1)
2096 unsigned SrcReg = getRegForValue(Op);
2097 if (!SrcReg) return false;
2099 // Because the high bits are undefined, a truncate doesn't generate
2101 UpdateValueMap(I, SrcReg);
2105 unsigned ARMFastISel::ARMEmitIntExt(EVT SrcVT, unsigned SrcReg, EVT DestVT,
2107 if (DestVT != MVT::i32 && DestVT != MVT::i16 && DestVT != MVT::i8)
2111 bool isBoolZext = false;
2112 if (!SrcVT.isSimple()) return 0;
2113 switch (SrcVT.getSimpleVT().SimpleTy) {
2116 if (!Subtarget->hasV6Ops()) return 0;
2118 Opc = isThumb2 ? ARM::t2UXTH : ARM::UXTH;
2120 Opc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
2123 if (!Subtarget->hasV6Ops()) return 0;
2125 Opc = isThumb2 ? ARM::t2UXTB : ARM::UXTB;
2127 Opc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
2131 Opc = isThumb2 ? ARM::t2ANDri : ARM::ANDri;
2138 unsigned ResultReg = createResultReg(TLI.getRegClassFor(MVT::i32));
2139 MachineInstrBuilder MIB;
2140 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), ResultReg)
2146 AddOptionalDefs(MIB);
2150 bool ARMFastISel::SelectIntExt(const Instruction *I) {
2151 // On ARM, in general, integer casts don't involve legal types; this code
2152 // handles promotable integers.
2153 // FIXME: We could save an instruction in many cases by special-casing
2154 // load instructions.
2155 Type *DestTy = I->getType();
2156 Value *Src = I->getOperand(0);
2157 Type *SrcTy = Src->getType();
2160 SrcVT = TLI.getValueType(SrcTy, true);
2161 DestVT = TLI.getValueType(DestTy, true);
2163 bool isZExt = isa<ZExtInst>(I);
2164 unsigned SrcReg = getRegForValue(Src);
2165 if (!SrcReg) return false;
2167 unsigned ResultReg = ARMEmitIntExt(SrcVT, SrcReg, DestVT, isZExt);
2168 if (ResultReg == 0) return false;
2169 UpdateValueMap(I, ResultReg);
2173 // TODO: SoftFP support.
2174 bool ARMFastISel::TargetSelectInstruction(const Instruction *I) {
2176 switch (I->getOpcode()) {
2177 case Instruction::Load:
2178 return SelectLoad(I);
2179 case Instruction::Store:
2180 return SelectStore(I);
2181 case Instruction::Br:
2182 return SelectBranch(I);
2183 case Instruction::ICmp:
2184 case Instruction::FCmp:
2185 return SelectCmp(I);
2186 case Instruction::FPExt:
2187 return SelectFPExt(I);
2188 case Instruction::FPTrunc:
2189 return SelectFPTrunc(I);
2190 case Instruction::SIToFP:
2191 return SelectSIToFP(I);
2192 case Instruction::FPToSI:
2193 return SelectFPToSI(I);
2194 case Instruction::FAdd:
2195 return SelectBinaryOp(I, ISD::FADD);
2196 case Instruction::FSub:
2197 return SelectBinaryOp(I, ISD::FSUB);
2198 case Instruction::FMul:
2199 return SelectBinaryOp(I, ISD::FMUL);
2200 case Instruction::SDiv:
2201 return SelectSDiv(I);
2202 case Instruction::SRem:
2203 return SelectSRem(I);
2204 case Instruction::Call:
2205 return SelectCall(I);
2206 case Instruction::Select:
2207 return SelectSelect(I);
2208 case Instruction::Ret:
2209 return SelectRet(I);
2210 case Instruction::Trunc:
2211 return SelectTrunc(I);
2212 case Instruction::ZExt:
2213 case Instruction::SExt:
2214 return SelectIntExt(I);
2221 llvm::FastISel *ARM::createFastISel(FunctionLoweringInfo &funcInfo) {
2222 // Completely untested on non-darwin.
2223 const TargetMachine &TM = funcInfo.MF->getTarget();
2225 // Darwin and thumb1 only for now.
2226 const ARMSubtarget *Subtarget = &TM.getSubtarget<ARMSubtarget>();
2227 if (Subtarget->isTargetDarwin() && !Subtarget->isThumb1Only() &&
2228 !DisableARMFastISel)
2229 return new ARMFastISel(funcInfo);