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 isThumb = AFI->isThumbFunction();
105 Context = &funcInfo.Fn->getContext();
108 // Code from FastISel.cpp.
109 virtual unsigned FastEmitInst_(unsigned MachineInstOpcode,
110 const TargetRegisterClass *RC);
111 virtual unsigned FastEmitInst_r(unsigned MachineInstOpcode,
112 const TargetRegisterClass *RC,
113 unsigned Op0, bool Op0IsKill);
114 virtual unsigned FastEmitInst_rr(unsigned MachineInstOpcode,
115 const TargetRegisterClass *RC,
116 unsigned Op0, bool Op0IsKill,
117 unsigned Op1, bool Op1IsKill);
118 virtual unsigned FastEmitInst_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->getSExtValue())) {
555 unsigned Opc = isThumb ? ARM::t2MOVi16 : ARM::MOVi16;
556 unsigned ImmReg = createResultReg(TLI.getRegClassFor(VT));
557 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
558 TII.get(Opc), ImmReg)
559 .addImm(CI->getSExtValue()));
563 // For now 32-bit only.
567 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
569 // MachineConstantPool wants an explicit alignment.
570 unsigned Align = TD.getPrefTypeAlignment(C->getType());
572 // TODO: Figure out if this is correct.
573 Align = TD.getTypeAllocSize(C->getType());
575 unsigned Idx = MCP.getConstantPoolIndex(C, Align);
578 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
579 TII.get(ARM::t2LDRpci), DestReg)
580 .addConstantPoolIndex(Idx));
582 // The extra immediate is for addrmode2.
583 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
584 TII.get(ARM::LDRcp), DestReg)
585 .addConstantPoolIndex(Idx)
591 unsigned ARMFastISel::ARMMaterializeGV(const GlobalValue *GV, EVT VT) {
592 // For now 32-bit only.
593 if (VT != MVT::i32) return 0;
595 Reloc::Model RelocM = TM.getRelocationModel();
597 // TODO: Need more magic for ARM PIC.
598 if (!isThumb && (RelocM == Reloc::PIC_)) return 0;
600 // MachineConstantPool wants an explicit alignment.
601 unsigned Align = TD.getPrefTypeAlignment(GV->getType());
603 // TODO: Figure out if this is correct.
604 Align = TD.getTypeAllocSize(GV->getType());
608 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb() ? 4 : 8);
609 unsigned Id = AFI->createPICLabelUId();
610 ARMConstantPoolValue *CPV = ARMConstantPoolConstant::Create(GV, Id,
613 unsigned Idx = MCP.getConstantPoolIndex(CPV, Align);
616 MachineInstrBuilder MIB;
617 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
619 unsigned Opc = (RelocM != Reloc::PIC_) ? ARM::t2LDRpci : ARM::t2LDRpci_pic;
620 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), DestReg)
621 .addConstantPoolIndex(Idx);
622 if (RelocM == Reloc::PIC_)
625 // The extra immediate is for addrmode2.
626 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(ARM::LDRcp),
628 .addConstantPoolIndex(Idx)
631 AddOptionalDefs(MIB);
633 if (Subtarget->GVIsIndirectSymbol(GV, RelocM)) {
634 unsigned NewDestReg = createResultReg(TLI.getRegClassFor(VT));
636 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
637 TII.get(ARM::t2LDRi12), NewDestReg)
641 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(ARM::LDRi12),
645 DestReg = NewDestReg;
646 AddOptionalDefs(MIB);
652 unsigned ARMFastISel::TargetMaterializeConstant(const Constant *C) {
653 EVT VT = TLI.getValueType(C->getType(), true);
655 // Only handle simple types.
656 if (!VT.isSimple()) return 0;
658 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
659 return ARMMaterializeFP(CFP, VT);
660 else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
661 return ARMMaterializeGV(GV, VT);
662 else if (isa<ConstantInt>(C))
663 return ARMMaterializeInt(C, VT);
668 unsigned ARMFastISel::TargetMaterializeAlloca(const AllocaInst *AI) {
669 // Don't handle dynamic allocas.
670 if (!FuncInfo.StaticAllocaMap.count(AI)) return 0;
673 if (!isLoadTypeLegal(AI->getType(), VT)) return false;
675 DenseMap<const AllocaInst*, int>::iterator SI =
676 FuncInfo.StaticAllocaMap.find(AI);
678 // This will get lowered later into the correct offsets and registers
679 // via rewriteXFrameIndex.
680 if (SI != FuncInfo.StaticAllocaMap.end()) {
681 TargetRegisterClass* RC = TLI.getRegClassFor(VT);
682 unsigned ResultReg = createResultReg(RC);
683 unsigned Opc = isThumb ? ARM::t2ADDri : ARM::ADDri;
684 AddOptionalDefs(BuildMI(*FuncInfo.MBB, *FuncInfo.InsertPt, DL,
685 TII.get(Opc), ResultReg)
686 .addFrameIndex(SI->second)
694 bool ARMFastISel::isTypeLegal(Type *Ty, MVT &VT) {
695 EVT evt = TLI.getValueType(Ty, true);
697 // Only handle simple types.
698 if (evt == MVT::Other || !evt.isSimple()) return false;
699 VT = evt.getSimpleVT();
701 // Handle all legal types, i.e. a register that will directly hold this
703 return TLI.isTypeLegal(VT);
706 bool ARMFastISel::isLoadTypeLegal(Type *Ty, MVT &VT) {
707 if (isTypeLegal(Ty, VT)) return true;
709 // If this is a type than can be sign or zero-extended to a basic operation
710 // go ahead and accept it now.
711 if (VT == MVT::i8 || VT == MVT::i16)
717 // Computes the address to get to an object.
718 bool ARMFastISel::ARMComputeAddress(const Value *Obj, Address &Addr) {
719 // Some boilerplate from the X86 FastISel.
720 const User *U = NULL;
721 unsigned Opcode = Instruction::UserOp1;
722 if (const Instruction *I = dyn_cast<Instruction>(Obj)) {
723 // Don't walk into other basic blocks unless the object is an alloca from
724 // another block, otherwise it may not have a virtual register assigned.
725 if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(Obj)) ||
726 FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
727 Opcode = I->getOpcode();
730 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Obj)) {
731 Opcode = C->getOpcode();
735 if (PointerType *Ty = dyn_cast<PointerType>(Obj->getType()))
736 if (Ty->getAddressSpace() > 255)
737 // Fast instruction selection doesn't support the special
744 case Instruction::BitCast: {
745 // Look through bitcasts.
746 return ARMComputeAddress(U->getOperand(0), Addr);
748 case Instruction::IntToPtr: {
749 // Look past no-op inttoptrs.
750 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
751 return ARMComputeAddress(U->getOperand(0), Addr);
754 case Instruction::PtrToInt: {
755 // Look past no-op ptrtoints.
756 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
757 return ARMComputeAddress(U->getOperand(0), Addr);
760 case Instruction::GetElementPtr: {
761 Address SavedAddr = Addr;
762 int TmpOffset = Addr.Offset;
764 // Iterate through the GEP folding the constants into offsets where
766 gep_type_iterator GTI = gep_type_begin(U);
767 for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end();
768 i != e; ++i, ++GTI) {
769 const Value *Op = *i;
770 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
771 const StructLayout *SL = TD.getStructLayout(STy);
772 unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
773 TmpOffset += SL->getElementOffset(Idx);
775 uint64_t S = TD.getTypeAllocSize(GTI.getIndexedType());
777 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
778 // Constant-offset addressing.
779 TmpOffset += CI->getSExtValue() * S;
782 if (isa<AddOperator>(Op) &&
783 (!isa<Instruction>(Op) ||
784 FuncInfo.MBBMap[cast<Instruction>(Op)->getParent()]
786 isa<ConstantInt>(cast<AddOperator>(Op)->getOperand(1))) {
787 // An add (in the same block) with a constant operand. Fold the
790 cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
791 TmpOffset += CI->getSExtValue() * S;
792 // Iterate on the other operand.
793 Op = cast<AddOperator>(Op)->getOperand(0);
797 goto unsupported_gep;
802 // Try to grab the base operand now.
803 Addr.Offset = TmpOffset;
804 if (ARMComputeAddress(U->getOperand(0), Addr)) return true;
806 // We failed, restore everything and try the other options.
812 case Instruction::Alloca: {
813 const AllocaInst *AI = cast<AllocaInst>(Obj);
814 DenseMap<const AllocaInst*, int>::iterator SI =
815 FuncInfo.StaticAllocaMap.find(AI);
816 if (SI != FuncInfo.StaticAllocaMap.end()) {
817 Addr.BaseType = Address::FrameIndexBase;
818 Addr.Base.FI = SI->second;
825 // Materialize the global variable's address into a reg which can
826 // then be used later to load the variable.
827 if (const GlobalValue *GV = dyn_cast<GlobalValue>(Obj)) {
828 unsigned Tmp = ARMMaterializeGV(GV, TLI.getValueType(Obj->getType()));
829 if (Tmp == 0) return false;
835 // Try to get this in a register if nothing else has worked.
836 if (Addr.Base.Reg == 0) Addr.Base.Reg = getRegForValue(Obj);
837 return Addr.Base.Reg != 0;
840 void ARMFastISel::ARMSimplifyAddress(Address &Addr, EVT VT) {
842 assert(VT.isSimple() && "Non-simple types are invalid here!");
844 bool needsLowering = false;
845 switch (VT.getSimpleVT().SimpleTy) {
847 assert(false && "Unhandled load/store type!");
852 // Integer loads/stores handle 12-bit offsets.
853 needsLowering = ((Addr.Offset & 0xfff) != Addr.Offset);
857 // Floating point operands handle 8-bit offsets.
858 needsLowering = ((Addr.Offset & 0xff) != Addr.Offset);
862 // If this is a stack pointer and the offset needs to be simplified then
863 // put the alloca address into a register, set the base type back to
864 // register and continue. This should almost never happen.
865 if (needsLowering && Addr.BaseType == Address::FrameIndexBase) {
866 TargetRegisterClass *RC = isThumb ? ARM::tGPRRegisterClass :
867 ARM::GPRRegisterClass;
868 unsigned ResultReg = createResultReg(RC);
869 unsigned Opc = isThumb ? ARM::t2ADDri : ARM::ADDri;
870 AddOptionalDefs(BuildMI(*FuncInfo.MBB, *FuncInfo.InsertPt, DL,
871 TII.get(Opc), ResultReg)
872 .addFrameIndex(Addr.Base.FI)
874 Addr.Base.Reg = ResultReg;
875 Addr.BaseType = Address::RegBase;
878 // Since the offset is too large for the load/store instruction
879 // get the reg+offset into a register.
881 Addr.Base.Reg = FastEmit_ri_(MVT::i32, ISD::ADD, Addr.Base.Reg,
882 /*Op0IsKill*/false, Addr.Offset, MVT::i32);
887 void ARMFastISel::AddLoadStoreOperands(EVT VT, Address &Addr,
888 const MachineInstrBuilder &MIB,
890 // addrmode5 output depends on the selection dag addressing dividing the
891 // offset by 4 that it then later multiplies. Do this here as well.
892 if (VT.getSimpleVT().SimpleTy == MVT::f32 ||
893 VT.getSimpleVT().SimpleTy == MVT::f64)
896 // Frame base works a bit differently. Handle it separately.
897 if (Addr.BaseType == Address::FrameIndexBase) {
898 int FI = Addr.Base.FI;
899 int Offset = Addr.Offset;
900 MachineMemOperand *MMO =
901 FuncInfo.MF->getMachineMemOperand(
902 MachinePointerInfo::getFixedStack(FI, Offset),
904 MFI.getObjectSize(FI),
905 MFI.getObjectAlignment(FI));
906 // Now add the rest of the operands.
907 MIB.addFrameIndex(FI);
909 // ARM halfword load/stores need an additional operand.
910 if (!isThumb && VT.getSimpleVT().SimpleTy == MVT::i16) MIB.addReg(0);
912 MIB.addImm(Addr.Offset);
913 MIB.addMemOperand(MMO);
915 // Now add the rest of the operands.
916 MIB.addReg(Addr.Base.Reg);
918 // ARM halfword load/stores need an additional operand.
919 if (!isThumb && VT.getSimpleVT().SimpleTy == MVT::i16) MIB.addReg(0);
921 MIB.addImm(Addr.Offset);
923 AddOptionalDefs(MIB);
926 bool ARMFastISel::ARMEmitLoad(EVT VT, unsigned &ResultReg, Address &Addr) {
928 assert(VT.isSimple() && "Non-simple types are invalid here!");
930 TargetRegisterClass *RC;
931 switch (VT.getSimpleVT().SimpleTy) {
932 // This is mostly going to be Neon/vector support.
933 default: return false;
935 Opc = isThumb ? ARM::t2LDRHi12 : ARM::LDRH;
936 RC = ARM::GPRRegisterClass;
939 Opc = isThumb ? ARM::t2LDRBi12 : ARM::LDRBi12;
940 RC = ARM::GPRRegisterClass;
943 Opc = isThumb ? ARM::t2LDRi12 : ARM::LDRi12;
944 RC = ARM::GPRRegisterClass;
948 RC = TLI.getRegClassFor(VT);
952 RC = TLI.getRegClassFor(VT);
955 // Simplify this down to something we can handle.
956 ARMSimplifyAddress(Addr, VT);
958 // Create the base instruction, then add the operands.
959 ResultReg = createResultReg(RC);
960 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
961 TII.get(Opc), ResultReg);
962 AddLoadStoreOperands(VT, Addr, MIB, MachineMemOperand::MOLoad);
966 bool ARMFastISel::SelectLoad(const Instruction *I) {
967 // Atomic loads need special handling.
968 if (cast<LoadInst>(I)->isAtomic())
971 // Verify we have a legal type before going any further.
973 if (!isLoadTypeLegal(I->getType(), VT))
976 // See if we can handle this address.
978 if (!ARMComputeAddress(I->getOperand(0), Addr)) return false;
981 if (!ARMEmitLoad(VT, ResultReg, Addr)) return false;
982 UpdateValueMap(I, ResultReg);
986 bool ARMFastISel::ARMEmitStore(EVT VT, unsigned SrcReg, Address &Addr) {
988 switch (VT.getSimpleVT().SimpleTy) {
989 // This is mostly going to be Neon/vector support.
990 default: return false;
992 unsigned Res = createResultReg(isThumb ? ARM::tGPRRegisterClass :
993 ARM::GPRRegisterClass);
994 unsigned Opc = isThumb ? ARM::t2ANDri : ARM::ANDri;
995 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
997 .addReg(SrcReg).addImm(1));
999 } // Fallthrough here.
1001 StrOpc = isThumb ? ARM::t2STRBi12 : ARM::STRBi12;
1004 StrOpc = isThumb ? ARM::t2STRHi12 : ARM::STRH;
1007 StrOpc = isThumb ? ARM::t2STRi12 : ARM::STRi12;
1010 if (!Subtarget->hasVFP2()) return false;
1011 StrOpc = ARM::VSTRS;
1014 if (!Subtarget->hasVFP2()) return false;
1015 StrOpc = ARM::VSTRD;
1018 // Simplify this down to something we can handle.
1019 ARMSimplifyAddress(Addr, VT);
1021 // Create the base instruction, then add the operands.
1022 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1024 .addReg(SrcReg, getKillRegState(true));
1025 AddLoadStoreOperands(VT, Addr, MIB, MachineMemOperand::MOStore);
1029 bool ARMFastISel::SelectStore(const Instruction *I) {
1030 Value *Op0 = I->getOperand(0);
1031 unsigned SrcReg = 0;
1033 // Atomic stores need special handling.
1034 if (cast<StoreInst>(I)->isAtomic())
1037 // Verify we have a legal type before going any further.
1039 if (!isLoadTypeLegal(I->getOperand(0)->getType(), VT))
1042 // Get the value to be stored into a register.
1043 SrcReg = getRegForValue(Op0);
1044 if (SrcReg == 0) return false;
1046 // See if we can handle this address.
1048 if (!ARMComputeAddress(I->getOperand(1), Addr))
1051 if (!ARMEmitStore(VT, SrcReg, Addr)) return false;
1055 static ARMCC::CondCodes getComparePred(CmpInst::Predicate Pred) {
1057 // Needs two compares...
1058 case CmpInst::FCMP_ONE:
1059 case CmpInst::FCMP_UEQ:
1061 // AL is our "false" for now. The other two need more compares.
1063 case CmpInst::ICMP_EQ:
1064 case CmpInst::FCMP_OEQ:
1066 case CmpInst::ICMP_SGT:
1067 case CmpInst::FCMP_OGT:
1069 case CmpInst::ICMP_SGE:
1070 case CmpInst::FCMP_OGE:
1072 case CmpInst::ICMP_UGT:
1073 case CmpInst::FCMP_UGT:
1075 case CmpInst::FCMP_OLT:
1077 case CmpInst::ICMP_ULE:
1078 case CmpInst::FCMP_OLE:
1080 case CmpInst::FCMP_ORD:
1082 case CmpInst::FCMP_UNO:
1084 case CmpInst::FCMP_UGE:
1086 case CmpInst::ICMP_SLT:
1087 case CmpInst::FCMP_ULT:
1089 case CmpInst::ICMP_SLE:
1090 case CmpInst::FCMP_ULE:
1092 case CmpInst::FCMP_UNE:
1093 case CmpInst::ICMP_NE:
1095 case CmpInst::ICMP_UGE:
1097 case CmpInst::ICMP_ULT:
1102 bool ARMFastISel::SelectBranch(const Instruction *I) {
1103 const BranchInst *BI = cast<BranchInst>(I);
1104 MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
1105 MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
1107 // Simple branch support.
1109 // If we can, avoid recomputing the compare - redoing it could lead to wonky
1111 if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
1112 if (CI->hasOneUse() && (CI->getParent() == I->getParent())) {
1114 // Get the compare predicate.
1115 // Try to take advantage of fallthrough opportunities.
1116 CmpInst::Predicate Predicate = CI->getPredicate();
1117 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
1118 std::swap(TBB, FBB);
1119 Predicate = CmpInst::getInversePredicate(Predicate);
1122 ARMCC::CondCodes ARMPred = getComparePred(Predicate);
1124 // We may not handle every CC for now.
1125 if (ARMPred == ARMCC::AL) return false;
1127 // Emit the compare.
1128 if (!ARMEmitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
1131 unsigned BrOpc = isThumb ? ARM::t2Bcc : ARM::Bcc;
1132 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
1133 .addMBB(TBB).addImm(ARMPred).addReg(ARM::CPSR);
1134 FastEmitBranch(FBB, DL);
1135 FuncInfo.MBB->addSuccessor(TBB);
1138 } else if (TruncInst *TI = dyn_cast<TruncInst>(BI->getCondition())) {
1140 if (TI->hasOneUse() && TI->getParent() == I->getParent() &&
1141 (isLoadTypeLegal(TI->getOperand(0)->getType(), SourceVT))) {
1142 unsigned TstOpc = isThumb ? ARM::t2TSTri : ARM::TSTri;
1143 unsigned OpReg = getRegForValue(TI->getOperand(0));
1144 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1146 .addReg(OpReg).addImm(1));
1148 unsigned CCMode = ARMCC::NE;
1149 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
1150 std::swap(TBB, FBB);
1154 unsigned BrOpc = isThumb ? ARM::t2Bcc : ARM::Bcc;
1155 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
1156 .addMBB(TBB).addImm(CCMode).addReg(ARM::CPSR);
1158 FastEmitBranch(FBB, DL);
1159 FuncInfo.MBB->addSuccessor(TBB);
1162 } else if (const ConstantInt *CI =
1163 dyn_cast<ConstantInt>(BI->getCondition())) {
1164 uint64_t Imm = CI->getZExtValue();
1165 MachineBasicBlock *Target = (Imm == 0) ? FBB : TBB;
1166 FastEmitBranch(Target, DL);
1170 unsigned CmpReg = getRegForValue(BI->getCondition());
1171 if (CmpReg == 0) return false;
1173 // We've been divorced from our compare! Our block was split, and
1174 // now our compare lives in a predecessor block. We musn't
1175 // re-compare here, as the children of the compare aren't guaranteed
1176 // live across the block boundary (we *could* check for this).
1177 // Regardless, the compare has been done in the predecessor block,
1178 // and it left a value for us in a virtual register. Ergo, we test
1179 // the one-bit value left in the virtual register.
1180 unsigned TstOpc = isThumb ? ARM::t2TSTri : ARM::TSTri;
1181 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TstOpc))
1182 .addReg(CmpReg).addImm(1));
1184 unsigned CCMode = ARMCC::NE;
1185 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
1186 std::swap(TBB, FBB);
1190 unsigned BrOpc = isThumb ? ARM::t2Bcc : ARM::Bcc;
1191 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
1192 .addMBB(TBB).addImm(CCMode).addReg(ARM::CPSR);
1193 FastEmitBranch(FBB, DL);
1194 FuncInfo.MBB->addSuccessor(TBB);
1198 bool ARMFastISel::ARMEmitCmp(const Value *Src1Value, const Value *Src2Value,
1200 Type *Ty = Src1Value->getType();
1201 EVT SrcVT = TLI.getValueType(Ty, true);
1202 if (!SrcVT.isSimple()) return false;
1204 bool isFloat = (Ty->isFloatTy() || Ty->isDoubleTy());
1205 if (isFloat && !Subtarget->hasVFP2())
1209 bool needsExt = false;
1210 switch (SrcVT.getSimpleVT().SimpleTy) {
1211 default: return false;
1212 // TODO: Verify compares.
1214 CmpOpc = ARM::VCMPES;
1217 CmpOpc = ARM::VCMPED;
1223 // Intentional fall-through.
1225 CmpOpc = isThumb ? ARM::t2CMPrr : ARM::CMPrr;
1229 unsigned SrcReg1 = getRegForValue(Src1Value);
1230 if (SrcReg1 == 0) return false;
1232 unsigned SrcReg2 = getRegForValue(Src2Value);
1233 if (SrcReg2 == 0) return false;
1235 // We have i1, i8, or i16, we need to either zero extend or sign extend.
1238 EVT DestVT = MVT::i32;
1239 ResultReg = ARMEmitIntExt(SrcVT, SrcReg1, DestVT, isZExt);
1240 if (ResultReg == 0) return false;
1241 SrcReg1 = ResultReg;
1242 ResultReg = ARMEmitIntExt(SrcVT, SrcReg2, DestVT, isZExt);
1243 if (ResultReg == 0) return false;
1244 SrcReg2 = ResultReg;
1247 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
1248 .addReg(SrcReg1).addReg(SrcReg2));
1250 // For floating point we need to move the result to a comparison register
1251 // that we can then use for branches.
1252 if (Ty->isFloatTy() || Ty->isDoubleTy())
1253 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1254 TII.get(ARM::FMSTAT)));
1258 bool ARMFastISel::SelectCmp(const Instruction *I) {
1259 const CmpInst *CI = cast<CmpInst>(I);
1260 Type *Ty = CI->getOperand(0)->getType();
1262 // Get the compare predicate.
1263 ARMCC::CondCodes ARMPred = getComparePred(CI->getPredicate());
1265 // We may not handle every CC for now.
1266 if (ARMPred == ARMCC::AL) return false;
1268 // Emit the compare.
1269 if (!ARMEmitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
1272 // Now set a register based on the comparison. Explicitly set the predicates
1274 unsigned MovCCOpc = isThumb ? ARM::t2MOVCCi : ARM::MOVCCi;
1275 TargetRegisterClass *RC = isThumb ? ARM::rGPRRegisterClass
1276 : ARM::GPRRegisterClass;
1277 unsigned DestReg = createResultReg(RC);
1278 Constant *Zero = ConstantInt::get(Type::getInt32Ty(*Context), 0);
1279 unsigned ZeroReg = TargetMaterializeConstant(Zero);
1280 bool isFloat = (Ty->isFloatTy() || Ty->isDoubleTy());
1281 unsigned CondReg = isFloat ? ARM::FPSCR : ARM::CPSR;
1282 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), DestReg)
1283 .addReg(ZeroReg).addImm(1)
1284 .addImm(ARMPred).addReg(CondReg);
1286 UpdateValueMap(I, DestReg);
1290 bool ARMFastISel::SelectFPExt(const Instruction *I) {
1291 // Make sure we have VFP and that we're extending float to double.
1292 if (!Subtarget->hasVFP2()) return false;
1294 Value *V = I->getOperand(0);
1295 if (!I->getType()->isDoubleTy() ||
1296 !V->getType()->isFloatTy()) return false;
1298 unsigned Op = getRegForValue(V);
1299 if (Op == 0) return false;
1301 unsigned Result = createResultReg(ARM::DPRRegisterClass);
1302 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1303 TII.get(ARM::VCVTDS), Result)
1305 UpdateValueMap(I, Result);
1309 bool ARMFastISel::SelectFPTrunc(const Instruction *I) {
1310 // Make sure we have VFP and that we're truncating double to float.
1311 if (!Subtarget->hasVFP2()) return false;
1313 Value *V = I->getOperand(0);
1314 if (!(I->getType()->isFloatTy() &&
1315 V->getType()->isDoubleTy())) return false;
1317 unsigned Op = getRegForValue(V);
1318 if (Op == 0) return false;
1320 unsigned Result = createResultReg(ARM::SPRRegisterClass);
1321 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1322 TII.get(ARM::VCVTSD), Result)
1324 UpdateValueMap(I, Result);
1328 bool ARMFastISel::SelectSIToFP(const Instruction *I) {
1329 // Make sure we have VFP.
1330 if (!Subtarget->hasVFP2()) return false;
1333 Type *Ty = I->getType();
1334 if (!isTypeLegal(Ty, DstVT))
1337 Value *Src = I->getOperand(0);
1338 EVT SrcVT = TLI.getValueType(Src->getType(), true);
1339 if (SrcVT != MVT::i32 && SrcVT != MVT::i16 && SrcVT != MVT::i8)
1342 unsigned SrcReg = getRegForValue(Src);
1343 if (SrcReg == 0) return false;
1345 // Handle sign-extension.
1346 if (SrcVT == MVT::i16 || SrcVT == MVT::i8) {
1347 EVT DestVT = MVT::i32;
1348 unsigned ResultReg = ARMEmitIntExt(SrcVT, SrcReg, DestVT, /*isZExt*/ false);
1349 if (ResultReg == 0) return false;
1353 // The conversion routine works on fp-reg to fp-reg and the operand above
1354 // was an integer, move it to the fp registers if possible.
1355 unsigned FP = ARMMoveToFPReg(MVT::f32, SrcReg);
1356 if (FP == 0) return false;
1359 if (Ty->isFloatTy()) Opc = ARM::VSITOS;
1360 else if (Ty->isDoubleTy()) Opc = ARM::VSITOD;
1363 unsigned ResultReg = createResultReg(TLI.getRegClassFor(DstVT));
1364 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
1367 UpdateValueMap(I, ResultReg);
1371 bool ARMFastISel::SelectFPToSI(const Instruction *I) {
1372 // Make sure we have VFP.
1373 if (!Subtarget->hasVFP2()) return false;
1376 Type *RetTy = I->getType();
1377 if (!isTypeLegal(RetTy, DstVT))
1380 unsigned Op = getRegForValue(I->getOperand(0));
1381 if (Op == 0) return false;
1384 Type *OpTy = I->getOperand(0)->getType();
1385 if (OpTy->isFloatTy()) Opc = ARM::VTOSIZS;
1386 else if (OpTy->isDoubleTy()) Opc = ARM::VTOSIZD;
1389 // f64->s32 or f32->s32 both need an intermediate f32 reg.
1390 unsigned ResultReg = createResultReg(TLI.getRegClassFor(MVT::f32));
1391 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
1395 // This result needs to be in an integer register, but the conversion only
1396 // takes place in fp-regs.
1397 unsigned IntReg = ARMMoveToIntReg(DstVT, ResultReg);
1398 if (IntReg == 0) return false;
1400 UpdateValueMap(I, IntReg);
1404 bool ARMFastISel::SelectSelect(const Instruction *I) {
1406 if (!isTypeLegal(I->getType(), VT))
1409 // Things need to be register sized for register moves.
1410 if (VT != MVT::i32) return false;
1411 const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
1413 unsigned CondReg = getRegForValue(I->getOperand(0));
1414 if (CondReg == 0) return false;
1415 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1416 if (Op1Reg == 0) return false;
1417 unsigned Op2Reg = getRegForValue(I->getOperand(2));
1418 if (Op2Reg == 0) return false;
1420 unsigned CmpOpc = isThumb ? ARM::t2TSTri : ARM::TSTri;
1421 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
1422 .addReg(CondReg).addImm(1));
1423 unsigned ResultReg = createResultReg(RC);
1424 unsigned MovCCOpc = isThumb ? ARM::t2MOVCCr : ARM::MOVCCr;
1425 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), ResultReg)
1426 .addReg(Op1Reg).addReg(Op2Reg)
1427 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
1428 UpdateValueMap(I, ResultReg);
1432 bool ARMFastISel::SelectSDiv(const Instruction *I) {
1434 Type *Ty = I->getType();
1435 if (!isTypeLegal(Ty, VT))
1438 // If we have integer div support we should have selected this automagically.
1439 // In case we have a real miss go ahead and return false and we'll pick
1441 if (Subtarget->hasDivide()) return false;
1443 // Otherwise emit a libcall.
1444 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
1446 LC = RTLIB::SDIV_I8;
1447 else if (VT == MVT::i16)
1448 LC = RTLIB::SDIV_I16;
1449 else if (VT == MVT::i32)
1450 LC = RTLIB::SDIV_I32;
1451 else if (VT == MVT::i64)
1452 LC = RTLIB::SDIV_I64;
1453 else if (VT == MVT::i128)
1454 LC = RTLIB::SDIV_I128;
1455 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SDIV!");
1457 return ARMEmitLibcall(I, LC);
1460 bool ARMFastISel::SelectSRem(const Instruction *I) {
1462 Type *Ty = I->getType();
1463 if (!isTypeLegal(Ty, VT))
1466 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
1468 LC = RTLIB::SREM_I8;
1469 else if (VT == MVT::i16)
1470 LC = RTLIB::SREM_I16;
1471 else if (VT == MVT::i32)
1472 LC = RTLIB::SREM_I32;
1473 else if (VT == MVT::i64)
1474 LC = RTLIB::SREM_I64;
1475 else if (VT == MVT::i128)
1476 LC = RTLIB::SREM_I128;
1477 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SREM!");
1479 return ARMEmitLibcall(I, LC);
1482 bool ARMFastISel::SelectBinaryOp(const Instruction *I, unsigned ISDOpcode) {
1483 EVT VT = TLI.getValueType(I->getType(), true);
1485 // We can get here in the case when we want to use NEON for our fp
1486 // operations, but can't figure out how to. Just use the vfp instructions
1488 // FIXME: It'd be nice to use NEON instructions.
1489 Type *Ty = I->getType();
1490 bool isFloat = (Ty->isDoubleTy() || Ty->isFloatTy());
1491 if (isFloat && !Subtarget->hasVFP2())
1494 unsigned Op1 = getRegForValue(I->getOperand(0));
1495 if (Op1 == 0) return false;
1497 unsigned Op2 = getRegForValue(I->getOperand(1));
1498 if (Op2 == 0) return false;
1501 bool is64bit = VT == MVT::f64 || VT == MVT::i64;
1502 switch (ISDOpcode) {
1503 default: return false;
1505 Opc = is64bit ? ARM::VADDD : ARM::VADDS;
1508 Opc = is64bit ? ARM::VSUBD : ARM::VSUBS;
1511 Opc = is64bit ? ARM::VMULD : ARM::VMULS;
1514 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
1515 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1516 TII.get(Opc), ResultReg)
1517 .addReg(Op1).addReg(Op2));
1518 UpdateValueMap(I, ResultReg);
1522 // Call Handling Code
1524 bool ARMFastISel::FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src,
1525 EVT SrcVT, unsigned &ResultReg) {
1526 unsigned RR = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), Opc,
1527 Src, /*TODO: Kill=*/false);
1536 // This is largely taken directly from CCAssignFnForNode - we don't support
1537 // varargs in FastISel so that part has been removed.
1538 // TODO: We may not support all of this.
1539 CCAssignFn *ARMFastISel::CCAssignFnForCall(CallingConv::ID CC, bool Return) {
1542 llvm_unreachable("Unsupported calling convention");
1543 case CallingConv::Fast:
1544 // Ignore fastcc. Silence compiler warnings.
1545 (void)RetFastCC_ARM_APCS;
1546 (void)FastCC_ARM_APCS;
1548 case CallingConv::C:
1549 // Use target triple & subtarget features to do actual dispatch.
1550 if (Subtarget->isAAPCS_ABI()) {
1551 if (Subtarget->hasVFP2() &&
1552 FloatABIType == FloatABI::Hard)
1553 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
1555 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
1557 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
1558 case CallingConv::ARM_AAPCS_VFP:
1559 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
1560 case CallingConv::ARM_AAPCS:
1561 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
1562 case CallingConv::ARM_APCS:
1563 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
1567 bool ARMFastISel::ProcessCallArgs(SmallVectorImpl<Value*> &Args,
1568 SmallVectorImpl<unsigned> &ArgRegs,
1569 SmallVectorImpl<MVT> &ArgVTs,
1570 SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
1571 SmallVectorImpl<unsigned> &RegArgs,
1573 unsigned &NumBytes) {
1574 SmallVector<CCValAssign, 16> ArgLocs;
1575 CCState CCInfo(CC, false, *FuncInfo.MF, TM, ArgLocs, *Context);
1576 CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CCAssignFnForCall(CC, false));
1578 // Get a count of how many bytes are to be pushed on the stack.
1579 NumBytes = CCInfo.getNextStackOffset();
1581 // Issue CALLSEQ_START
1582 unsigned AdjStackDown = TII.getCallFrameSetupOpcode();
1583 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1584 TII.get(AdjStackDown))
1587 // Process the args.
1588 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1589 CCValAssign &VA = ArgLocs[i];
1590 unsigned Arg = ArgRegs[VA.getValNo()];
1591 MVT ArgVT = ArgVTs[VA.getValNo()];
1593 // We don't handle NEON/vector parameters yet.
1594 if (ArgVT.isVector() || ArgVT.getSizeInBits() > 64)
1597 // Handle arg promotion, etc.
1598 switch (VA.getLocInfo()) {
1599 case CCValAssign::Full: break;
1600 case CCValAssign::SExt: {
1601 bool Emitted = FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
1603 assert(Emitted && "Failed to emit a sext!"); (void)Emitted;
1605 ArgVT = VA.getLocVT();
1608 case CCValAssign::ZExt: {
1609 bool Emitted = FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
1611 assert(Emitted && "Failed to emit a zext!"); (void)Emitted;
1613 ArgVT = VA.getLocVT();
1616 case CCValAssign::AExt: {
1617 bool Emitted = FastEmitExtend(ISD::ANY_EXTEND, VA.getLocVT(),
1620 Emitted = FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
1623 Emitted = FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
1626 assert(Emitted && "Failed to emit a aext!"); (void)Emitted;
1627 ArgVT = VA.getLocVT();
1630 case CCValAssign::BCvt: {
1631 unsigned BC = FastEmit_r(ArgVT, VA.getLocVT(), ISD::BITCAST, Arg,
1632 /*TODO: Kill=*/false);
1633 assert(BC != 0 && "Failed to emit a bitcast!");
1635 ArgVT = VA.getLocVT();
1638 default: llvm_unreachable("Unknown arg promotion!");
1641 // Now copy/store arg to correct locations.
1642 if (VA.isRegLoc() && !VA.needsCustom()) {
1643 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1646 RegArgs.push_back(VA.getLocReg());
1647 } else if (VA.needsCustom()) {
1648 // TODO: We need custom lowering for vector (v2f64) args.
1649 if (VA.getLocVT() != MVT::f64) return false;
1651 CCValAssign &NextVA = ArgLocs[++i];
1653 // TODO: Only handle register args for now.
1654 if(!(VA.isRegLoc() && NextVA.isRegLoc())) return false;
1656 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1657 TII.get(ARM::VMOVRRD), VA.getLocReg())
1658 .addReg(NextVA.getLocReg(), RegState::Define)
1660 RegArgs.push_back(VA.getLocReg());
1661 RegArgs.push_back(NextVA.getLocReg());
1663 assert(VA.isMemLoc());
1664 // Need to store on the stack.
1666 Addr.BaseType = Address::RegBase;
1667 Addr.Base.Reg = ARM::SP;
1668 Addr.Offset = VA.getLocMemOffset();
1670 if (!ARMEmitStore(ArgVT, Arg, Addr)) return false;
1676 bool ARMFastISel::FinishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
1677 const Instruction *I, CallingConv::ID CC,
1678 unsigned &NumBytes) {
1679 // Issue CALLSEQ_END
1680 unsigned AdjStackUp = TII.getCallFrameDestroyOpcode();
1681 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1682 TII.get(AdjStackUp))
1683 .addImm(NumBytes).addImm(0));
1685 // Now the return value.
1686 if (RetVT != MVT::isVoid) {
1687 SmallVector<CCValAssign, 16> RVLocs;
1688 CCState CCInfo(CC, false, *FuncInfo.MF, TM, RVLocs, *Context);
1689 CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC, true));
1691 // Copy all of the result registers out of their specified physreg.
1692 if (RVLocs.size() == 2 && RetVT == MVT::f64) {
1693 // For this move we copy into two registers and then move into the
1694 // double fp reg we want.
1695 EVT DestVT = RVLocs[0].getValVT();
1696 TargetRegisterClass* DstRC = TLI.getRegClassFor(DestVT);
1697 unsigned ResultReg = createResultReg(DstRC);
1698 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1699 TII.get(ARM::VMOVDRR), ResultReg)
1700 .addReg(RVLocs[0].getLocReg())
1701 .addReg(RVLocs[1].getLocReg()));
1703 UsedRegs.push_back(RVLocs[0].getLocReg());
1704 UsedRegs.push_back(RVLocs[1].getLocReg());
1706 // Finally update the result.
1707 UpdateValueMap(I, ResultReg);
1709 assert(RVLocs.size() == 1 &&"Can't handle non-double multi-reg retvals!");
1710 EVT CopyVT = RVLocs[0].getValVT();
1711 TargetRegisterClass* DstRC = TLI.getRegClassFor(CopyVT);
1713 unsigned ResultReg = createResultReg(DstRC);
1714 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1715 ResultReg).addReg(RVLocs[0].getLocReg());
1716 UsedRegs.push_back(RVLocs[0].getLocReg());
1718 // Finally update the result.
1719 UpdateValueMap(I, ResultReg);
1726 bool ARMFastISel::SelectRet(const Instruction *I) {
1727 const ReturnInst *Ret = cast<ReturnInst>(I);
1728 const Function &F = *I->getParent()->getParent();
1730 if (!FuncInfo.CanLowerReturn)
1736 CallingConv::ID CC = F.getCallingConv();
1737 if (Ret->getNumOperands() > 0) {
1738 SmallVector<ISD::OutputArg, 4> Outs;
1739 GetReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(),
1742 // Analyze operands of the call, assigning locations to each operand.
1743 SmallVector<CCValAssign, 16> ValLocs;
1744 CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, TM, ValLocs,I->getContext());
1745 CCInfo.AnalyzeReturn(Outs, CCAssignFnForCall(CC, true /* is Ret */));
1747 const Value *RV = Ret->getOperand(0);
1748 unsigned Reg = getRegForValue(RV);
1752 // Only handle a single return value for now.
1753 if (ValLocs.size() != 1)
1756 CCValAssign &VA = ValLocs[0];
1758 // Don't bother handling odd stuff for now.
1759 if (VA.getLocInfo() != CCValAssign::Full)
1761 // Only handle register returns for now.
1765 unsigned SrcReg = Reg + VA.getValNo();
1766 EVT RVVT = TLI.getValueType(RV->getType());
1767 EVT DestVT = VA.getValVT();
1768 // Special handling for extended integers.
1769 if (RVVT != DestVT) {
1770 if (RVVT != MVT::i1 && RVVT != MVT::i8 && RVVT != MVT::i16)
1773 if (!Outs[0].Flags.isZExt() && !Outs[0].Flags.isSExt())
1776 assert(DestVT == MVT::i32 && "ARM should always ext to i32");
1778 bool isZExt = Outs[0].Flags.isZExt();
1779 unsigned ResultReg = ARMEmitIntExt(RVVT, SrcReg, DestVT, isZExt);
1780 if (ResultReg == 0) return false;
1785 unsigned DstReg = VA.getLocReg();
1786 const TargetRegisterClass* SrcRC = MRI.getRegClass(SrcReg);
1787 // Avoid a cross-class copy. This is very unlikely.
1788 if (!SrcRC->contains(DstReg))
1790 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1791 DstReg).addReg(SrcReg);
1793 // Mark the register as live out of the function.
1794 MRI.addLiveOut(VA.getLocReg());
1797 unsigned RetOpc = isThumb ? ARM::tBX_RET : ARM::BX_RET;
1798 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1803 unsigned ARMFastISel::ARMSelectCallOp(const GlobalValue *GV) {
1805 // Darwin needs the r9 versions of the opcodes.
1806 bool isDarwin = Subtarget->isTargetDarwin();
1808 return isDarwin ? ARM::tBLr9 : ARM::tBL;
1810 return isDarwin ? ARM::BLr9 : ARM::BL;
1814 // A quick function that will emit a call for a named libcall in F with the
1815 // vector of passed arguments for the Instruction in I. We can assume that we
1816 // can emit a call for any libcall we can produce. This is an abridged version
1817 // of the full call infrastructure since we won't need to worry about things
1818 // like computed function pointers or strange arguments at call sites.
1819 // TODO: Try to unify this and the normal call bits for ARM, then try to unify
1821 bool ARMFastISel::ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call) {
1822 CallingConv::ID CC = TLI.getLibcallCallingConv(Call);
1824 // Handle *simple* calls for now.
1825 Type *RetTy = I->getType();
1827 if (RetTy->isVoidTy())
1828 RetVT = MVT::isVoid;
1829 else if (!isTypeLegal(RetTy, RetVT))
1832 // TODO: For now if we have long calls specified we don't handle the call.
1833 if (EnableARMLongCalls) return false;
1835 // Set up the argument vectors.
1836 SmallVector<Value*, 8> Args;
1837 SmallVector<unsigned, 8> ArgRegs;
1838 SmallVector<MVT, 8> ArgVTs;
1839 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
1840 Args.reserve(I->getNumOperands());
1841 ArgRegs.reserve(I->getNumOperands());
1842 ArgVTs.reserve(I->getNumOperands());
1843 ArgFlags.reserve(I->getNumOperands());
1844 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
1845 Value *Op = I->getOperand(i);
1846 unsigned Arg = getRegForValue(Op);
1847 if (Arg == 0) return false;
1849 Type *ArgTy = Op->getType();
1851 if (!isTypeLegal(ArgTy, ArgVT)) return false;
1853 ISD::ArgFlagsTy Flags;
1854 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
1855 Flags.setOrigAlign(OriginalAlignment);
1858 ArgRegs.push_back(Arg);
1859 ArgVTs.push_back(ArgVT);
1860 ArgFlags.push_back(Flags);
1863 // Handle the arguments now that we've gotten them.
1864 SmallVector<unsigned, 4> RegArgs;
1866 if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
1869 // Issue the call, BLr9 for darwin, BL otherwise.
1870 // TODO: Turn this into the table of arm call ops.
1871 MachineInstrBuilder MIB;
1872 unsigned CallOpc = ARMSelectCallOp(NULL);
1874 // Explicitly adding the predicate here.
1875 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1877 .addExternalSymbol(TLI.getLibcallName(Call));
1879 // Explicitly adding the predicate here.
1880 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1882 .addExternalSymbol(TLI.getLibcallName(Call)));
1884 // Add implicit physical register uses to the call.
1885 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
1886 MIB.addReg(RegArgs[i]);
1888 // Finish off the call including any return values.
1889 SmallVector<unsigned, 4> UsedRegs;
1890 if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false;
1892 // Set all unused physreg defs as dead.
1893 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
1898 bool ARMFastISel::SelectCall(const Instruction *I) {
1899 const CallInst *CI = cast<CallInst>(I);
1900 const Value *Callee = CI->getCalledValue();
1902 // Can't handle inline asm or worry about intrinsics yet.
1903 if (isa<InlineAsm>(Callee) || isa<IntrinsicInst>(CI)) return false;
1905 // Only handle global variable Callees.
1906 const GlobalValue *GV = dyn_cast<GlobalValue>(Callee);
1910 // Check the calling convention.
1911 ImmutableCallSite CS(CI);
1912 CallingConv::ID CC = CS.getCallingConv();
1914 // TODO: Avoid some calling conventions?
1916 // Let SDISel handle vararg functions.
1917 PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
1918 FunctionType *FTy = cast<FunctionType>(PT->getElementType());
1919 if (FTy->isVarArg())
1922 // Handle *simple* calls for now.
1923 Type *RetTy = I->getType();
1925 if (RetTy->isVoidTy())
1926 RetVT = MVT::isVoid;
1927 else if (!isTypeLegal(RetTy, RetVT))
1930 // TODO: For now if we have long calls specified we don't handle the call.
1931 if (EnableARMLongCalls) return false;
1933 // Set up the argument vectors.
1934 SmallVector<Value*, 8> Args;
1935 SmallVector<unsigned, 8> ArgRegs;
1936 SmallVector<MVT, 8> ArgVTs;
1937 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
1938 Args.reserve(CS.arg_size());
1939 ArgRegs.reserve(CS.arg_size());
1940 ArgVTs.reserve(CS.arg_size());
1941 ArgFlags.reserve(CS.arg_size());
1942 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
1944 unsigned Arg = getRegForValue(*i);
1948 ISD::ArgFlagsTy Flags;
1949 unsigned AttrInd = i - CS.arg_begin() + 1;
1950 if (CS.paramHasAttr(AttrInd, Attribute::SExt))
1952 if (CS.paramHasAttr(AttrInd, Attribute::ZExt))
1955 // FIXME: Only handle *easy* calls for now.
1956 if (CS.paramHasAttr(AttrInd, Attribute::InReg) ||
1957 CS.paramHasAttr(AttrInd, Attribute::StructRet) ||
1958 CS.paramHasAttr(AttrInd, Attribute::Nest) ||
1959 CS.paramHasAttr(AttrInd, Attribute::ByVal))
1962 Type *ArgTy = (*i)->getType();
1964 // FIXME: Should be able to handle i1, i8, and/or i16 parameters.
1965 if (!isTypeLegal(ArgTy, ArgVT))
1967 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
1968 Flags.setOrigAlign(OriginalAlignment);
1971 ArgRegs.push_back(Arg);
1972 ArgVTs.push_back(ArgVT);
1973 ArgFlags.push_back(Flags);
1976 // Handle the arguments now that we've gotten them.
1977 SmallVector<unsigned, 4> RegArgs;
1979 if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
1982 // Issue the call, BLr9 for darwin, BL otherwise.
1983 // TODO: Turn this into the table of arm call ops.
1984 MachineInstrBuilder MIB;
1985 unsigned CallOpc = ARMSelectCallOp(GV);
1986 // Explicitly adding the predicate here.
1988 // Explicitly adding the predicate here.
1989 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1991 .addGlobalAddress(GV, 0, 0);
1993 // Explicitly adding the predicate here.
1994 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1996 .addGlobalAddress(GV, 0, 0));
1998 // Add implicit physical register uses to the call.
1999 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
2000 MIB.addReg(RegArgs[i]);
2002 // Finish off the call including any return values.
2003 SmallVector<unsigned, 4> UsedRegs;
2004 if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false;
2006 // Set all unused physreg defs as dead.
2007 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
2012 bool ARMFastISel::SelectTrunc(const Instruction *I) {
2013 // The high bits for a type smaller than the register size are assumed to be
2015 Value *Op = I->getOperand(0);
2018 SrcVT = TLI.getValueType(Op->getType(), true);
2019 DestVT = TLI.getValueType(I->getType(), true);
2021 if (SrcVT != MVT::i32 && SrcVT != MVT::i16 && SrcVT != MVT::i8)
2023 if (DestVT != MVT::i16 && DestVT != MVT::i8 && DestVT != MVT::i1)
2026 unsigned SrcReg = getRegForValue(Op);
2027 if (!SrcReg) return false;
2029 // Because the high bits are undefined, a truncate doesn't generate
2031 UpdateValueMap(I, SrcReg);
2035 unsigned ARMFastISel::ARMEmitIntExt(EVT SrcVT, unsigned SrcReg, EVT DestVT,
2037 if (DestVT != MVT::i32 && DestVT != MVT::i16 && DestVT != MVT::i8)
2041 bool isBoolZext = false;
2042 if (!SrcVT.isSimple()) return 0;
2043 switch (SrcVT.getSimpleVT().SimpleTy) {
2046 if (!Subtarget->hasV6Ops()) return 0;
2048 Opc = isThumb ? ARM::t2UXTH : ARM::UXTH;
2050 Opc = isThumb ? ARM::t2SXTH : ARM::SXTH;
2053 if (!Subtarget->hasV6Ops()) return 0;
2055 Opc = isThumb ? ARM::t2UXTB : ARM::UXTB;
2057 Opc = isThumb ? ARM::t2SXTB : ARM::SXTB;
2061 Opc = isThumb ? ARM::t2ANDri : ARM::ANDri;
2068 unsigned ResultReg = createResultReg(TLI.getRegClassFor(MVT::i32));
2069 MachineInstrBuilder MIB;
2070 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), ResultReg)
2076 AddOptionalDefs(MIB);
2080 bool ARMFastISel::SelectIntExt(const Instruction *I) {
2081 // On ARM, in general, integer casts don't involve legal types; this code
2082 // handles promotable integers.
2083 // FIXME: We could save an instruction in many cases by special-casing
2084 // load instructions.
2085 Type *DestTy = I->getType();
2086 Value *Src = I->getOperand(0);
2087 Type *SrcTy = Src->getType();
2090 SrcVT = TLI.getValueType(SrcTy, true);
2091 DestVT = TLI.getValueType(DestTy, true);
2093 bool isZExt = isa<ZExtInst>(I);
2094 unsigned SrcReg = getRegForValue(Src);
2095 if (!SrcReg) return false;
2097 unsigned ResultReg = ARMEmitIntExt(SrcVT, SrcReg, DestVT, isZExt);
2098 if (ResultReg == 0) return false;
2099 UpdateValueMap(I, ResultReg);
2103 // TODO: SoftFP support.
2104 bool ARMFastISel::TargetSelectInstruction(const Instruction *I) {
2106 switch (I->getOpcode()) {
2107 case Instruction::Load:
2108 return SelectLoad(I);
2109 case Instruction::Store:
2110 return SelectStore(I);
2111 case Instruction::Br:
2112 return SelectBranch(I);
2113 case Instruction::ICmp:
2114 case Instruction::FCmp:
2115 return SelectCmp(I);
2116 case Instruction::FPExt:
2117 return SelectFPExt(I);
2118 case Instruction::FPTrunc:
2119 return SelectFPTrunc(I);
2120 case Instruction::SIToFP:
2121 return SelectSIToFP(I);
2122 case Instruction::FPToSI:
2123 return SelectFPToSI(I);
2124 case Instruction::FAdd:
2125 return SelectBinaryOp(I, ISD::FADD);
2126 case Instruction::FSub:
2127 return SelectBinaryOp(I, ISD::FSUB);
2128 case Instruction::FMul:
2129 return SelectBinaryOp(I, ISD::FMUL);
2130 case Instruction::SDiv:
2131 return SelectSDiv(I);
2132 case Instruction::SRem:
2133 return SelectSRem(I);
2134 case Instruction::Call:
2135 return SelectCall(I);
2136 case Instruction::Select:
2137 return SelectSelect(I);
2138 case Instruction::Ret:
2139 return SelectRet(I);
2140 case Instruction::Trunc:
2141 return SelectTrunc(I);
2142 case Instruction::ZExt:
2143 case Instruction::SExt:
2144 return SelectIntExt(I);
2151 llvm::FastISel *ARM::createFastISel(FunctionLoweringInfo &funcInfo) {
2152 // Completely untested on non-darwin.
2153 const TargetMachine &TM = funcInfo.MF->getTarget();
2155 // Darwin and thumb1 only for now.
2156 const ARMSubtarget *Subtarget = &TM.getSubtarget<ARMSubtarget>();
2157 if (Subtarget->isTargetDarwin() && !Subtarget->isThumb1Only() &&
2158 !DisableARMFastISel)
2159 return new ARMFastISel(funcInfo);