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 SelectIntCast(const Instruction *I);
174 bool isTypeLegal(Type *Ty, MVT &VT);
175 bool isLoadTypeLegal(Type *Ty, MVT &VT);
176 bool ARMEmitCmp(Type *Ty, const Value *Src1Value, const Value *Src2Value);
177 bool ARMEmitLoad(EVT VT, unsigned &ResultReg, Address &Addr);
178 bool ARMEmitStore(EVT VT, unsigned SrcReg, Address &Addr);
179 bool ARMComputeAddress(const Value *Obj, Address &Addr);
180 void ARMSimplifyAddress(Address &Addr, EVT VT);
181 unsigned ARMMaterializeFP(const ConstantFP *CFP, EVT VT);
182 unsigned ARMMaterializeInt(const Constant *C, EVT VT);
183 unsigned ARMMaterializeGV(const GlobalValue *GV, EVT VT);
184 unsigned ARMMoveToFPReg(EVT VT, unsigned SrcReg);
185 unsigned ARMMoveToIntReg(EVT VT, unsigned SrcReg);
186 unsigned ARMSelectCallOp(const GlobalValue *GV);
188 // Call handling routines.
190 bool FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src, EVT SrcVT,
191 unsigned &ResultReg);
192 CCAssignFn *CCAssignFnForCall(CallingConv::ID CC, bool Return);
193 bool ProcessCallArgs(SmallVectorImpl<Value*> &Args,
194 SmallVectorImpl<unsigned> &ArgRegs,
195 SmallVectorImpl<MVT> &ArgVTs,
196 SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
197 SmallVectorImpl<unsigned> &RegArgs,
200 bool FinishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
201 const Instruction *I, CallingConv::ID CC,
203 bool ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call);
205 // OptionalDef handling routines.
207 bool isARMNEONPred(const MachineInstr *MI);
208 bool DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR);
209 const MachineInstrBuilder &AddOptionalDefs(const MachineInstrBuilder &MIB);
210 void AddLoadStoreOperands(EVT VT, Address &Addr,
211 const MachineInstrBuilder &MIB,
215 } // end anonymous namespace
217 #include "ARMGenCallingConv.inc"
219 // DefinesOptionalPredicate - This is different from DefinesPredicate in that
220 // we don't care about implicit defs here, just places we'll need to add a
221 // default CCReg argument. Sets CPSR if we're setting CPSR instead of CCR.
222 bool ARMFastISel::DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR) {
223 const MCInstrDesc &MCID = MI->getDesc();
224 if (!MCID.hasOptionalDef())
227 // Look to see if our OptionalDef is defining CPSR or CCR.
228 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
229 const MachineOperand &MO = MI->getOperand(i);
230 if (!MO.isReg() || !MO.isDef()) continue;
231 if (MO.getReg() == ARM::CPSR)
237 bool ARMFastISel::isARMNEONPred(const MachineInstr *MI) {
238 const MCInstrDesc &MCID = MI->getDesc();
240 // If we're a thumb2 or not NEON function we were handled via isPredicable.
241 if ((MCID.TSFlags & ARMII::DomainMask) != ARMII::DomainNEON ||
242 AFI->isThumb2Function())
245 for (unsigned i = 0, e = MCID.getNumOperands(); i != e; ++i)
246 if (MCID.OpInfo[i].isPredicate())
252 // If the machine is predicable go ahead and add the predicate operands, if
253 // it needs default CC operands add those.
254 // TODO: If we want to support thumb1 then we'll need to deal with optional
255 // CPSR defs that need to be added before the remaining operands. See s_cc_out
256 // for descriptions why.
257 const MachineInstrBuilder &
258 ARMFastISel::AddOptionalDefs(const MachineInstrBuilder &MIB) {
259 MachineInstr *MI = &*MIB;
261 // Do we use a predicate? or...
262 // Are we NEON in ARM mode and have a predicate operand? If so, I know
263 // we're not predicable but add it anyways.
264 if (TII.isPredicable(MI) || isARMNEONPred(MI))
267 // Do we optionally set a predicate? Preds is size > 0 iff the predicate
268 // defines CPSR. All other OptionalDefines in ARM are the CCR register.
270 if (DefinesOptionalPredicate(MI, &CPSR)) {
279 unsigned ARMFastISel::FastEmitInst_(unsigned MachineInstOpcode,
280 const TargetRegisterClass* RC) {
281 unsigned ResultReg = createResultReg(RC);
282 const MCInstrDesc &II = TII.get(MachineInstOpcode);
284 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg));
288 unsigned ARMFastISel::FastEmitInst_r(unsigned MachineInstOpcode,
289 const TargetRegisterClass *RC,
290 unsigned Op0, bool Op0IsKill) {
291 unsigned ResultReg = createResultReg(RC);
292 const MCInstrDesc &II = TII.get(MachineInstOpcode);
294 if (II.getNumDefs() >= 1)
295 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
296 .addReg(Op0, Op0IsKill * RegState::Kill));
298 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
299 .addReg(Op0, Op0IsKill * RegState::Kill));
300 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
301 TII.get(TargetOpcode::COPY), ResultReg)
302 .addReg(II.ImplicitDefs[0]));
307 unsigned ARMFastISel::FastEmitInst_rr(unsigned MachineInstOpcode,
308 const TargetRegisterClass *RC,
309 unsigned Op0, bool Op0IsKill,
310 unsigned Op1, bool Op1IsKill) {
311 unsigned ResultReg = createResultReg(RC);
312 const MCInstrDesc &II = TII.get(MachineInstOpcode);
314 if (II.getNumDefs() >= 1)
315 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
316 .addReg(Op0, Op0IsKill * RegState::Kill)
317 .addReg(Op1, Op1IsKill * RegState::Kill));
319 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
320 .addReg(Op0, Op0IsKill * RegState::Kill)
321 .addReg(Op1, Op1IsKill * RegState::Kill));
322 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
323 TII.get(TargetOpcode::COPY), ResultReg)
324 .addReg(II.ImplicitDefs[0]));
329 unsigned ARMFastISel::FastEmitInst_rrr(unsigned MachineInstOpcode,
330 const TargetRegisterClass *RC,
331 unsigned Op0, bool Op0IsKill,
332 unsigned Op1, bool Op1IsKill,
333 unsigned Op2, bool Op2IsKill) {
334 unsigned ResultReg = createResultReg(RC);
335 const MCInstrDesc &II = TII.get(MachineInstOpcode);
337 if (II.getNumDefs() >= 1)
338 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
339 .addReg(Op0, Op0IsKill * RegState::Kill)
340 .addReg(Op1, Op1IsKill * RegState::Kill)
341 .addReg(Op2, Op2IsKill * RegState::Kill));
343 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
344 .addReg(Op0, Op0IsKill * RegState::Kill)
345 .addReg(Op1, Op1IsKill * RegState::Kill)
346 .addReg(Op2, Op2IsKill * RegState::Kill));
347 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
348 TII.get(TargetOpcode::COPY), ResultReg)
349 .addReg(II.ImplicitDefs[0]));
354 unsigned ARMFastISel::FastEmitInst_ri(unsigned MachineInstOpcode,
355 const TargetRegisterClass *RC,
356 unsigned Op0, bool Op0IsKill,
358 unsigned ResultReg = createResultReg(RC);
359 const MCInstrDesc &II = TII.get(MachineInstOpcode);
361 if (II.getNumDefs() >= 1)
362 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
363 .addReg(Op0, Op0IsKill * RegState::Kill)
366 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
367 .addReg(Op0, Op0IsKill * RegState::Kill)
369 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
370 TII.get(TargetOpcode::COPY), ResultReg)
371 .addReg(II.ImplicitDefs[0]));
376 unsigned ARMFastISel::FastEmitInst_rf(unsigned MachineInstOpcode,
377 const TargetRegisterClass *RC,
378 unsigned Op0, bool Op0IsKill,
379 const ConstantFP *FPImm) {
380 unsigned ResultReg = createResultReg(RC);
381 const MCInstrDesc &II = TII.get(MachineInstOpcode);
383 if (II.getNumDefs() >= 1)
384 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
385 .addReg(Op0, Op0IsKill * RegState::Kill)
388 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
389 .addReg(Op0, Op0IsKill * RegState::Kill)
391 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
392 TII.get(TargetOpcode::COPY), ResultReg)
393 .addReg(II.ImplicitDefs[0]));
398 unsigned ARMFastISel::FastEmitInst_rri(unsigned MachineInstOpcode,
399 const TargetRegisterClass *RC,
400 unsigned Op0, bool Op0IsKill,
401 unsigned Op1, bool Op1IsKill,
403 unsigned ResultReg = createResultReg(RC);
404 const MCInstrDesc &II = TII.get(MachineInstOpcode);
406 if (II.getNumDefs() >= 1)
407 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
408 .addReg(Op0, Op0IsKill * RegState::Kill)
409 .addReg(Op1, Op1IsKill * RegState::Kill)
412 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
413 .addReg(Op0, Op0IsKill * RegState::Kill)
414 .addReg(Op1, Op1IsKill * RegState::Kill)
416 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
417 TII.get(TargetOpcode::COPY), ResultReg)
418 .addReg(II.ImplicitDefs[0]));
423 unsigned ARMFastISel::FastEmitInst_i(unsigned MachineInstOpcode,
424 const TargetRegisterClass *RC,
426 unsigned ResultReg = createResultReg(RC);
427 const MCInstrDesc &II = TII.get(MachineInstOpcode);
429 if (II.getNumDefs() >= 1)
430 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
433 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
435 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
436 TII.get(TargetOpcode::COPY), ResultReg)
437 .addReg(II.ImplicitDefs[0]));
442 unsigned ARMFastISel::FastEmitInst_ii(unsigned MachineInstOpcode,
443 const TargetRegisterClass *RC,
444 uint64_t Imm1, uint64_t Imm2) {
445 unsigned ResultReg = createResultReg(RC);
446 const MCInstrDesc &II = TII.get(MachineInstOpcode);
448 if (II.getNumDefs() >= 1)
449 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
450 .addImm(Imm1).addImm(Imm2));
452 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
453 .addImm(Imm1).addImm(Imm2));
454 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
455 TII.get(TargetOpcode::COPY),
457 .addReg(II.ImplicitDefs[0]));
462 unsigned ARMFastISel::FastEmitInst_extractsubreg(MVT RetVT,
463 unsigned Op0, bool Op0IsKill,
465 unsigned ResultReg = createResultReg(TLI.getRegClassFor(RetVT));
466 assert(TargetRegisterInfo::isVirtualRegister(Op0) &&
467 "Cannot yet extract from physregs");
468 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
469 DL, TII.get(TargetOpcode::COPY), ResultReg)
470 .addReg(Op0, getKillRegState(Op0IsKill), Idx));
474 // TODO: Don't worry about 64-bit now, but when this is fixed remove the
475 // checks from the various callers.
476 unsigned ARMFastISel::ARMMoveToFPReg(EVT VT, unsigned SrcReg) {
477 if (VT == MVT::f64) return 0;
479 unsigned MoveReg = createResultReg(TLI.getRegClassFor(VT));
480 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
481 TII.get(ARM::VMOVRS), MoveReg)
486 unsigned ARMFastISel::ARMMoveToIntReg(EVT VT, unsigned SrcReg) {
487 if (VT == MVT::i64) return 0;
489 unsigned MoveReg = createResultReg(TLI.getRegClassFor(VT));
490 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
491 TII.get(ARM::VMOVSR), MoveReg)
496 // For double width floating point we need to materialize two constants
497 // (the high and the low) into integer registers then use a move to get
498 // the combined constant into an FP reg.
499 unsigned ARMFastISel::ARMMaterializeFP(const ConstantFP *CFP, EVT VT) {
500 const APFloat Val = CFP->getValueAPF();
501 bool is64bit = VT == MVT::f64;
503 // This checks to see if we can use VFP3 instructions to materialize
504 // a constant, otherwise we have to go through the constant pool.
505 if (TLI.isFPImmLegal(Val, VT)) {
509 Imm = ARM_AM::getFP64Imm(Val);
512 Imm = ARM_AM::getFP32Imm(Val);
515 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
516 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
522 // Require VFP2 for loading fp constants.
523 if (!Subtarget->hasVFP2()) return false;
525 // MachineConstantPool wants an explicit alignment.
526 unsigned Align = TD.getPrefTypeAlignment(CFP->getType());
528 // TODO: Figure out if this is correct.
529 Align = TD.getTypeAllocSize(CFP->getType());
531 unsigned Idx = MCP.getConstantPoolIndex(cast<Constant>(CFP), Align);
532 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
533 unsigned Opc = is64bit ? ARM::VLDRD : ARM::VLDRS;
535 // The extra reg is for addrmode5.
536 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
538 .addConstantPoolIndex(Idx)
543 unsigned ARMFastISel::ARMMaterializeInt(const Constant *C, EVT VT) {
545 // For now 32-bit only.
546 if (VT != MVT::i32) return false;
548 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
550 // If we can do this in a single instruction without a constant pool entry
552 const ConstantInt *CI = cast<ConstantInt>(C);
553 if (Subtarget->hasV6T2Ops() && isUInt<16>(CI->getSExtValue())) {
554 unsigned Opc = isThumb ? ARM::t2MOVi16 : ARM::MOVi16;
555 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
556 TII.get(Opc), DestReg)
557 .addImm(CI->getSExtValue()));
561 // MachineConstantPool wants an explicit alignment.
562 unsigned Align = TD.getPrefTypeAlignment(C->getType());
564 // TODO: Figure out if this is correct.
565 Align = TD.getTypeAllocSize(C->getType());
567 unsigned Idx = MCP.getConstantPoolIndex(C, Align);
570 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
571 TII.get(ARM::t2LDRpci), DestReg)
572 .addConstantPoolIndex(Idx));
574 // The extra immediate is for addrmode2.
575 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
576 TII.get(ARM::LDRcp), DestReg)
577 .addConstantPoolIndex(Idx)
583 unsigned ARMFastISel::ARMMaterializeGV(const GlobalValue *GV, EVT VT) {
584 // For now 32-bit only.
585 if (VT != MVT::i32) return 0;
587 Reloc::Model RelocM = TM.getRelocationModel();
589 // TODO: Need more magic for ARM PIC.
590 if (!isThumb && (RelocM == Reloc::PIC_)) return 0;
592 // MachineConstantPool wants an explicit alignment.
593 unsigned Align = TD.getPrefTypeAlignment(GV->getType());
595 // TODO: Figure out if this is correct.
596 Align = TD.getTypeAllocSize(GV->getType());
600 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb() ? 4 : 8);
601 unsigned Id = AFI->createPICLabelUId();
602 ARMConstantPoolValue *CPV = ARMConstantPoolConstant::Create(GV, Id,
605 unsigned Idx = MCP.getConstantPoolIndex(CPV, Align);
608 MachineInstrBuilder MIB;
609 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
611 unsigned Opc = (RelocM != Reloc::PIC_) ? ARM::t2LDRpci : ARM::t2LDRpci_pic;
612 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), DestReg)
613 .addConstantPoolIndex(Idx);
614 if (RelocM == Reloc::PIC_)
617 // The extra immediate is for addrmode2.
618 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(ARM::LDRcp),
620 .addConstantPoolIndex(Idx)
623 AddOptionalDefs(MIB);
625 if (Subtarget->GVIsIndirectSymbol(GV, RelocM)) {
626 unsigned NewDestReg = createResultReg(TLI.getRegClassFor(VT));
628 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
629 TII.get(ARM::t2LDRi12), NewDestReg)
633 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(ARM::LDRi12),
637 DestReg = NewDestReg;
638 AddOptionalDefs(MIB);
644 unsigned ARMFastISel::TargetMaterializeConstant(const Constant *C) {
645 EVT VT = TLI.getValueType(C->getType(), true);
647 // Only handle simple types.
648 if (!VT.isSimple()) return 0;
650 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
651 return ARMMaterializeFP(CFP, VT);
652 else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
653 return ARMMaterializeGV(GV, VT);
654 else if (isa<ConstantInt>(C))
655 return ARMMaterializeInt(C, VT);
660 unsigned ARMFastISel::TargetMaterializeAlloca(const AllocaInst *AI) {
661 // Don't handle dynamic allocas.
662 if (!FuncInfo.StaticAllocaMap.count(AI)) return 0;
665 if (!isLoadTypeLegal(AI->getType(), VT)) return false;
667 DenseMap<const AllocaInst*, int>::iterator SI =
668 FuncInfo.StaticAllocaMap.find(AI);
670 // This will get lowered later into the correct offsets and registers
671 // via rewriteXFrameIndex.
672 if (SI != FuncInfo.StaticAllocaMap.end()) {
673 TargetRegisterClass* RC = TLI.getRegClassFor(VT);
674 unsigned ResultReg = createResultReg(RC);
675 unsigned Opc = isThumb ? ARM::t2ADDri : ARM::ADDri;
676 AddOptionalDefs(BuildMI(*FuncInfo.MBB, *FuncInfo.InsertPt, DL,
677 TII.get(Opc), ResultReg)
678 .addFrameIndex(SI->second)
686 bool ARMFastISel::isTypeLegal(Type *Ty, MVT &VT) {
687 EVT evt = TLI.getValueType(Ty, true);
689 // Only handle simple types.
690 if (evt == MVT::Other || !evt.isSimple()) return false;
691 VT = evt.getSimpleVT();
693 // Handle all legal types, i.e. a register that will directly hold this
695 return TLI.isTypeLegal(VT);
698 bool ARMFastISel::isLoadTypeLegal(Type *Ty, MVT &VT) {
699 if (isTypeLegal(Ty, VT)) return true;
701 // If this is a type than can be sign or zero-extended to a basic operation
702 // go ahead and accept it now.
703 if (VT == MVT::i8 || VT == MVT::i16)
709 // Computes the address to get to an object.
710 bool ARMFastISel::ARMComputeAddress(const Value *Obj, Address &Addr) {
711 // Some boilerplate from the X86 FastISel.
712 const User *U = NULL;
713 unsigned Opcode = Instruction::UserOp1;
714 if (const Instruction *I = dyn_cast<Instruction>(Obj)) {
715 // Don't walk into other basic blocks unless the object is an alloca from
716 // another block, otherwise it may not have a virtual register assigned.
717 if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(Obj)) ||
718 FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
719 Opcode = I->getOpcode();
722 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Obj)) {
723 Opcode = C->getOpcode();
727 if (PointerType *Ty = dyn_cast<PointerType>(Obj->getType()))
728 if (Ty->getAddressSpace() > 255)
729 // Fast instruction selection doesn't support the special
736 case Instruction::BitCast: {
737 // Look through bitcasts.
738 return ARMComputeAddress(U->getOperand(0), Addr);
740 case Instruction::IntToPtr: {
741 // Look past no-op inttoptrs.
742 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
743 return ARMComputeAddress(U->getOperand(0), Addr);
746 case Instruction::PtrToInt: {
747 // Look past no-op ptrtoints.
748 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
749 return ARMComputeAddress(U->getOperand(0), Addr);
752 case Instruction::GetElementPtr: {
753 Address SavedAddr = Addr;
754 int TmpOffset = Addr.Offset;
756 // Iterate through the GEP folding the constants into offsets where
758 gep_type_iterator GTI = gep_type_begin(U);
759 for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end();
760 i != e; ++i, ++GTI) {
761 const Value *Op = *i;
762 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
763 const StructLayout *SL = TD.getStructLayout(STy);
764 unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
765 TmpOffset += SL->getElementOffset(Idx);
767 uint64_t S = TD.getTypeAllocSize(GTI.getIndexedType());
769 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
770 // Constant-offset addressing.
771 TmpOffset += CI->getSExtValue() * S;
774 if (isa<AddOperator>(Op) &&
775 (!isa<Instruction>(Op) ||
776 FuncInfo.MBBMap[cast<Instruction>(Op)->getParent()]
778 isa<ConstantInt>(cast<AddOperator>(Op)->getOperand(1))) {
779 // An add (in the same block) with a constant operand. Fold the
782 cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
783 TmpOffset += CI->getSExtValue() * S;
784 // Iterate on the other operand.
785 Op = cast<AddOperator>(Op)->getOperand(0);
789 goto unsupported_gep;
794 // Try to grab the base operand now.
795 Addr.Offset = TmpOffset;
796 if (ARMComputeAddress(U->getOperand(0), Addr)) return true;
798 // We failed, restore everything and try the other options.
804 case Instruction::Alloca: {
805 const AllocaInst *AI = cast<AllocaInst>(Obj);
806 DenseMap<const AllocaInst*, int>::iterator SI =
807 FuncInfo.StaticAllocaMap.find(AI);
808 if (SI != FuncInfo.StaticAllocaMap.end()) {
809 Addr.BaseType = Address::FrameIndexBase;
810 Addr.Base.FI = SI->second;
817 // Materialize the global variable's address into a reg which can
818 // then be used later to load the variable.
819 if (const GlobalValue *GV = dyn_cast<GlobalValue>(Obj)) {
820 unsigned Tmp = ARMMaterializeGV(GV, TLI.getValueType(Obj->getType()));
821 if (Tmp == 0) return false;
827 // Try to get this in a register if nothing else has worked.
828 if (Addr.Base.Reg == 0) Addr.Base.Reg = getRegForValue(Obj);
829 return Addr.Base.Reg != 0;
832 void ARMFastISel::ARMSimplifyAddress(Address &Addr, EVT VT) {
834 assert(VT.isSimple() && "Non-simple types are invalid here!");
836 bool needsLowering = false;
837 switch (VT.getSimpleVT().SimpleTy) {
839 assert(false && "Unhandled load/store type!");
844 // Integer loads/stores handle 12-bit offsets.
845 needsLowering = ((Addr.Offset & 0xfff) != Addr.Offset);
849 // Floating point operands handle 8-bit offsets.
850 needsLowering = ((Addr.Offset & 0xff) != Addr.Offset);
854 // If this is a stack pointer and the offset needs to be simplified then
855 // put the alloca address into a register, set the base type back to
856 // register and continue. This should almost never happen.
857 if (needsLowering && Addr.BaseType == Address::FrameIndexBase) {
858 TargetRegisterClass *RC = isThumb ? ARM::tGPRRegisterClass :
859 ARM::GPRRegisterClass;
860 unsigned ResultReg = createResultReg(RC);
861 unsigned Opc = isThumb ? ARM::t2ADDri : ARM::ADDri;
862 AddOptionalDefs(BuildMI(*FuncInfo.MBB, *FuncInfo.InsertPt, DL,
863 TII.get(Opc), ResultReg)
864 .addFrameIndex(Addr.Base.FI)
866 Addr.Base.Reg = ResultReg;
867 Addr.BaseType = Address::RegBase;
870 // Since the offset is too large for the load/store instruction
871 // get the reg+offset into a register.
873 Addr.Base.Reg = FastEmit_ri_(MVT::i32, ISD::ADD, Addr.Base.Reg,
874 /*Op0IsKill*/false, Addr.Offset, MVT::i32);
879 void ARMFastISel::AddLoadStoreOperands(EVT VT, Address &Addr,
880 const MachineInstrBuilder &MIB,
882 // addrmode5 output depends on the selection dag addressing dividing the
883 // offset by 4 that it then later multiplies. Do this here as well.
884 if (VT.getSimpleVT().SimpleTy == MVT::f32 ||
885 VT.getSimpleVT().SimpleTy == MVT::f64)
888 // Frame base works a bit differently. Handle it separately.
889 if (Addr.BaseType == Address::FrameIndexBase) {
890 int FI = Addr.Base.FI;
891 int Offset = Addr.Offset;
892 MachineMemOperand *MMO =
893 FuncInfo.MF->getMachineMemOperand(
894 MachinePointerInfo::getFixedStack(FI, Offset),
896 MFI.getObjectSize(FI),
897 MFI.getObjectAlignment(FI));
898 // Now add the rest of the operands.
899 MIB.addFrameIndex(FI);
901 // ARM halfword load/stores need an additional operand.
902 if (!isThumb && VT.getSimpleVT().SimpleTy == MVT::i16) MIB.addReg(0);
904 MIB.addImm(Addr.Offset);
905 MIB.addMemOperand(MMO);
907 // Now add the rest of the operands.
908 MIB.addReg(Addr.Base.Reg);
910 // ARM halfword load/stores need an additional operand.
911 if (!isThumb && VT.getSimpleVT().SimpleTy == MVT::i16) MIB.addReg(0);
913 MIB.addImm(Addr.Offset);
915 AddOptionalDefs(MIB);
918 bool ARMFastISel::ARMEmitLoad(EVT VT, unsigned &ResultReg, Address &Addr) {
920 assert(VT.isSimple() && "Non-simple types are invalid here!");
922 TargetRegisterClass *RC;
923 switch (VT.getSimpleVT().SimpleTy) {
924 // This is mostly going to be Neon/vector support.
925 default: return false;
927 Opc = isThumb ? ARM::t2LDRHi12 : ARM::LDRH;
928 RC = ARM::GPRRegisterClass;
931 Opc = isThumb ? ARM::t2LDRBi12 : ARM::LDRBi12;
932 RC = ARM::GPRRegisterClass;
935 Opc = isThumb ? ARM::t2LDRi12 : ARM::LDRi12;
936 RC = ARM::GPRRegisterClass;
940 RC = TLI.getRegClassFor(VT);
944 RC = TLI.getRegClassFor(VT);
947 // Simplify this down to something we can handle.
948 ARMSimplifyAddress(Addr, VT);
950 // Create the base instruction, then add the operands.
951 ResultReg = createResultReg(RC);
952 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
953 TII.get(Opc), ResultReg);
954 AddLoadStoreOperands(VT, Addr, MIB, MachineMemOperand::MOLoad);
958 bool ARMFastISel::SelectLoad(const Instruction *I) {
959 // Atomic loads need special handling.
960 if (cast<LoadInst>(I)->isAtomic())
963 // Verify we have a legal type before going any further.
965 if (!isLoadTypeLegal(I->getType(), VT))
968 // See if we can handle this address.
970 if (!ARMComputeAddress(I->getOperand(0), Addr)) return false;
973 if (!ARMEmitLoad(VT, ResultReg, Addr)) return false;
974 UpdateValueMap(I, ResultReg);
978 bool ARMFastISel::ARMEmitStore(EVT VT, unsigned SrcReg, Address &Addr) {
980 switch (VT.getSimpleVT().SimpleTy) {
981 // This is mostly going to be Neon/vector support.
982 default: return false;
984 unsigned Res = createResultReg(isThumb ? ARM::tGPRRegisterClass :
985 ARM::GPRRegisterClass);
986 unsigned Opc = isThumb ? ARM::t2ANDri : ARM::ANDri;
987 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
989 .addReg(SrcReg).addImm(1));
991 } // Fallthrough here.
993 StrOpc = isThumb ? ARM::t2STRBi12 : ARM::STRBi12;
996 StrOpc = isThumb ? ARM::t2STRHi12 : ARM::STRH;
999 StrOpc = isThumb ? ARM::t2STRi12 : ARM::STRi12;
1002 if (!Subtarget->hasVFP2()) return false;
1003 StrOpc = ARM::VSTRS;
1006 if (!Subtarget->hasVFP2()) return false;
1007 StrOpc = ARM::VSTRD;
1010 // Simplify this down to something we can handle.
1011 ARMSimplifyAddress(Addr, VT);
1013 // Create the base instruction, then add the operands.
1014 MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1016 .addReg(SrcReg, getKillRegState(true));
1017 AddLoadStoreOperands(VT, Addr, MIB, MachineMemOperand::MOStore);
1021 bool ARMFastISel::SelectStore(const Instruction *I) {
1022 Value *Op0 = I->getOperand(0);
1023 unsigned SrcReg = 0;
1025 // Atomic stores need special handling.
1026 if (cast<StoreInst>(I)->isAtomic())
1029 // Verify we have a legal type before going any further.
1031 if (!isLoadTypeLegal(I->getOperand(0)->getType(), VT))
1034 // Get the value to be stored into a register.
1035 SrcReg = getRegForValue(Op0);
1036 if (SrcReg == 0) return false;
1038 // See if we can handle this address.
1040 if (!ARMComputeAddress(I->getOperand(1), Addr))
1043 if (!ARMEmitStore(VT, SrcReg, Addr)) return false;
1047 static ARMCC::CondCodes getComparePred(CmpInst::Predicate Pred) {
1049 // Needs two compares...
1050 case CmpInst::FCMP_ONE:
1051 case CmpInst::FCMP_UEQ:
1053 // AL is our "false" for now. The other two need more compares.
1055 case CmpInst::ICMP_EQ:
1056 case CmpInst::FCMP_OEQ:
1058 case CmpInst::ICMP_SGT:
1059 case CmpInst::FCMP_OGT:
1061 case CmpInst::ICMP_SGE:
1062 case CmpInst::FCMP_OGE:
1064 case CmpInst::ICMP_UGT:
1065 case CmpInst::FCMP_UGT:
1067 case CmpInst::FCMP_OLT:
1069 case CmpInst::ICMP_ULE:
1070 case CmpInst::FCMP_OLE:
1072 case CmpInst::FCMP_ORD:
1074 case CmpInst::FCMP_UNO:
1076 case CmpInst::FCMP_UGE:
1078 case CmpInst::ICMP_SLT:
1079 case CmpInst::FCMP_ULT:
1081 case CmpInst::ICMP_SLE:
1082 case CmpInst::FCMP_ULE:
1084 case CmpInst::FCMP_UNE:
1085 case CmpInst::ICMP_NE:
1087 case CmpInst::ICMP_UGE:
1089 case CmpInst::ICMP_ULT:
1094 bool ARMFastISel::SelectBranch(const Instruction *I) {
1095 const BranchInst *BI = cast<BranchInst>(I);
1096 MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
1097 MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
1099 // Simple branch support.
1101 // If we can, avoid recomputing the compare - redoing it could lead to wonky
1103 // TODO: Factor this out.
1104 if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
1106 Type *Ty = CI->getOperand(0)->getType();
1107 if (CI->hasOneUse() && (CI->getParent() == I->getParent())
1108 && isTypeLegal(Ty, SourceVT)) {
1109 bool isFloat = (Ty->isDoubleTy() || Ty->isFloatTy());
1110 if (isFloat && !Subtarget->hasVFP2())
1114 switch (SourceVT.SimpleTy) {
1115 default: return false;
1116 // TODO: Verify compares.
1118 CmpOpc = ARM::VCMPES;
1121 CmpOpc = ARM::VCMPED;
1124 CmpOpc = isThumb ? ARM::t2CMPrr : ARM::CMPrr;
1128 // Get the compare predicate.
1129 // Try to take advantage of fallthrough opportunities.
1130 CmpInst::Predicate Predicate = CI->getPredicate();
1131 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
1132 std::swap(TBB, FBB);
1133 Predicate = CmpInst::getInversePredicate(Predicate);
1136 ARMCC::CondCodes ARMPred = getComparePred(Predicate);
1138 // We may not handle every CC for now.
1139 if (ARMPred == ARMCC::AL) return false;
1141 unsigned Arg1 = getRegForValue(CI->getOperand(0));
1142 if (Arg1 == 0) return false;
1144 unsigned Arg2 = getRegForValue(CI->getOperand(1));
1145 if (Arg2 == 0) return false;
1147 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1149 .addReg(Arg1).addReg(Arg2));
1151 // For floating point we need to move the result to a comparison register
1152 // that we can then use for branches.
1154 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1155 TII.get(ARM::FMSTAT)));
1157 unsigned BrOpc = isThumb ? ARM::t2Bcc : ARM::Bcc;
1158 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
1159 .addMBB(TBB).addImm(ARMPred).addReg(ARM::CPSR);
1160 FastEmitBranch(FBB, DL);
1161 FuncInfo.MBB->addSuccessor(TBB);
1164 } else if (TruncInst *TI = dyn_cast<TruncInst>(BI->getCondition())) {
1166 if (TI->hasOneUse() && TI->getParent() == I->getParent() &&
1167 (isLoadTypeLegal(TI->getOperand(0)->getType(), SourceVT))) {
1168 unsigned TstOpc = isThumb ? ARM::t2TSTri : ARM::TSTri;
1169 unsigned OpReg = getRegForValue(TI->getOperand(0));
1170 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1172 .addReg(OpReg).addImm(1));
1174 unsigned CCMode = ARMCC::NE;
1175 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
1176 std::swap(TBB, FBB);
1180 unsigned BrOpc = isThumb ? ARM::t2Bcc : ARM::Bcc;
1181 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
1182 .addMBB(TBB).addImm(CCMode).addReg(ARM::CPSR);
1184 FastEmitBranch(FBB, DL);
1185 FuncInfo.MBB->addSuccessor(TBB);
1190 unsigned CmpReg = getRegForValue(BI->getCondition());
1191 if (CmpReg == 0) return false;
1193 // We've been divorced from our compare! Our block was split, and
1194 // now our compare lives in a predecessor block. We musn't
1195 // re-compare here, as the children of the compare aren't guaranteed
1196 // live across the block boundary (we *could* check for this).
1197 // Regardless, the compare has been done in the predecessor block,
1198 // and it left a value for us in a virtual register. Ergo, we test
1199 // the one-bit value left in the virtual register.
1200 unsigned TstOpc = isThumb ? ARM::t2TSTri : ARM::TSTri;
1201 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TstOpc))
1202 .addReg(CmpReg).addImm(1));
1204 unsigned CCMode = ARMCC::NE;
1205 if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
1206 std::swap(TBB, FBB);
1210 unsigned BrOpc = isThumb ? ARM::t2Bcc : ARM::Bcc;
1211 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
1212 .addMBB(TBB).addImm(CCMode).addReg(ARM::CPSR);
1213 FastEmitBranch(FBB, DL);
1214 FuncInfo.MBB->addSuccessor(TBB);
1218 bool ARMFastISel::ARMEmitCmp(Type *Ty, const Value *Src1Value,
1219 const Value *Src2Value) {
1221 if (!isTypeLegal(Ty, VT))
1224 if ((Ty->isFloatTy() || Ty->isDoubleTy()) && !Subtarget->hasVFP2())
1228 switch (VT.SimpleTy) {
1229 default: return false;
1230 // TODO: Verify compares.
1232 CmpOpc = ARM::VCMPES;
1235 CmpOpc = ARM::VCMPED;
1238 CmpOpc = isThumb ? ARM::t2CMPrr : ARM::CMPrr;
1242 unsigned Src1 = getRegForValue(Src1Value);
1243 if (Src1 == 0) return false;
1245 unsigned Src2 = getRegForValue(Src2Value);
1246 if (Src2 == 0) return false;
1248 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
1249 .addReg(Src1).addReg(Src2));
1253 bool ARMFastISel::SelectCmp(const Instruction *I) {
1254 const CmpInst *CI = cast<CmpInst>(I);
1256 // Get the compare predicate.
1257 ARMCC::CondCodes ARMPred = getComparePred(CI->getPredicate());
1259 // We may not handle every CC for now.
1260 if (ARMPred == ARMCC::AL) return false;
1262 // Emit the compare.
1263 Type *Ty = CI->getOperand(0)->getType();
1264 if (!ARMEmitCmp(Ty, CI->getOperand(0), CI->getOperand(1)))
1267 // For floating point we need to move the result to a comparison register
1268 // that we can then use for branches.
1269 bool isFloat = Ty->isFloatTy() || Ty->isDoubleTy();
1271 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1272 TII.get(ARM::FMSTAT)));
1274 // Now set a register based on the comparison. Explicitly set the predicates
1276 unsigned MovCCOpc = isThumb ? ARM::t2MOVCCi : ARM::MOVCCi;
1277 TargetRegisterClass *RC = isThumb ? ARM::rGPRRegisterClass
1278 : ARM::GPRRegisterClass;
1279 unsigned DestReg = createResultReg(RC);
1281 = ConstantInt::get(Type::getInt32Ty(*Context), 0);
1282 unsigned ZeroReg = TargetMaterializeConstant(Zero);
1283 unsigned CondReg = isFloat ? ARM::FPSCR : ARM::CPSR;
1284 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), DestReg)
1285 .addReg(ZeroReg).addImm(1)
1286 .addImm(ARMPred).addReg(CondReg);
1288 UpdateValueMap(I, DestReg);
1292 bool ARMFastISel::SelectFPExt(const Instruction *I) {
1293 // Make sure we have VFP and that we're extending float to double.
1294 if (!Subtarget->hasVFP2()) return false;
1296 Value *V = I->getOperand(0);
1297 if (!I->getType()->isDoubleTy() ||
1298 !V->getType()->isFloatTy()) return false;
1300 unsigned Op = getRegForValue(V);
1301 if (Op == 0) return false;
1303 unsigned Result = createResultReg(ARM::DPRRegisterClass);
1304 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1305 TII.get(ARM::VCVTDS), Result)
1307 UpdateValueMap(I, Result);
1311 bool ARMFastISel::SelectFPTrunc(const Instruction *I) {
1312 // Make sure we have VFP and that we're truncating double to float.
1313 if (!Subtarget->hasVFP2()) return false;
1315 Value *V = I->getOperand(0);
1316 if (!(I->getType()->isFloatTy() &&
1317 V->getType()->isDoubleTy())) return false;
1319 unsigned Op = getRegForValue(V);
1320 if (Op == 0) return false;
1322 unsigned Result = createResultReg(ARM::SPRRegisterClass);
1323 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1324 TII.get(ARM::VCVTSD), Result)
1326 UpdateValueMap(I, Result);
1330 bool ARMFastISel::SelectSIToFP(const Instruction *I) {
1331 // Make sure we have VFP.
1332 if (!Subtarget->hasVFP2()) return false;
1335 Type *Ty = I->getType();
1336 if (!isTypeLegal(Ty, DstVT))
1339 // FIXME: Handle sign-extension where necessary.
1340 if (!I->getOperand(0)->getType()->isIntegerTy(32))
1343 unsigned Op = getRegForValue(I->getOperand(0));
1344 if (Op == 0) return false;
1346 // The conversion routine works on fp-reg to fp-reg and the operand above
1347 // was an integer, move it to the fp registers if possible.
1348 unsigned FP = ARMMoveToFPReg(MVT::f32, Op);
1349 if (FP == 0) return false;
1352 if (Ty->isFloatTy()) Opc = ARM::VSITOS;
1353 else if (Ty->isDoubleTy()) Opc = ARM::VSITOD;
1356 unsigned ResultReg = createResultReg(TLI.getRegClassFor(DstVT));
1357 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
1360 UpdateValueMap(I, ResultReg);
1364 bool ARMFastISel::SelectFPToSI(const Instruction *I) {
1365 // Make sure we have VFP.
1366 if (!Subtarget->hasVFP2()) return false;
1369 Type *RetTy = I->getType();
1370 if (!isTypeLegal(RetTy, DstVT))
1373 unsigned Op = getRegForValue(I->getOperand(0));
1374 if (Op == 0) return false;
1377 Type *OpTy = I->getOperand(0)->getType();
1378 if (OpTy->isFloatTy()) Opc = ARM::VTOSIZS;
1379 else if (OpTy->isDoubleTy()) Opc = ARM::VTOSIZD;
1382 // f64->s32 or f32->s32 both need an intermediate f32 reg.
1383 unsigned ResultReg = createResultReg(TLI.getRegClassFor(MVT::f32));
1384 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
1388 // This result needs to be in an integer register, but the conversion only
1389 // takes place in fp-regs.
1390 unsigned IntReg = ARMMoveToIntReg(DstVT, ResultReg);
1391 if (IntReg == 0) return false;
1393 UpdateValueMap(I, IntReg);
1397 bool ARMFastISel::SelectSelect(const Instruction *I) {
1399 if (!isTypeLegal(I->getType(), VT))
1402 // Things need to be register sized for register moves.
1403 if (VT != MVT::i32) return false;
1404 const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
1406 unsigned CondReg = getRegForValue(I->getOperand(0));
1407 if (CondReg == 0) return false;
1408 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1409 if (Op1Reg == 0) return false;
1410 unsigned Op2Reg = getRegForValue(I->getOperand(2));
1411 if (Op2Reg == 0) return false;
1413 unsigned CmpOpc = isThumb ? ARM::t2TSTri : ARM::TSTri;
1414 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
1415 .addReg(CondReg).addImm(1));
1416 unsigned ResultReg = createResultReg(RC);
1417 unsigned MovCCOpc = isThumb ? ARM::t2MOVCCr : ARM::MOVCCr;
1418 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), ResultReg)
1419 .addReg(Op1Reg).addReg(Op2Reg)
1420 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
1421 UpdateValueMap(I, ResultReg);
1425 bool ARMFastISel::SelectSDiv(const Instruction *I) {
1427 Type *Ty = I->getType();
1428 if (!isTypeLegal(Ty, VT))
1431 // If we have integer div support we should have selected this automagically.
1432 // In case we have a real miss go ahead and return false and we'll pick
1434 if (Subtarget->hasDivide()) return false;
1436 // Otherwise emit a libcall.
1437 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
1439 LC = RTLIB::SDIV_I8;
1440 else if (VT == MVT::i16)
1441 LC = RTLIB::SDIV_I16;
1442 else if (VT == MVT::i32)
1443 LC = RTLIB::SDIV_I32;
1444 else if (VT == MVT::i64)
1445 LC = RTLIB::SDIV_I64;
1446 else if (VT == MVT::i128)
1447 LC = RTLIB::SDIV_I128;
1448 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SDIV!");
1450 return ARMEmitLibcall(I, LC);
1453 bool ARMFastISel::SelectSRem(const Instruction *I) {
1455 Type *Ty = I->getType();
1456 if (!isTypeLegal(Ty, VT))
1459 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
1461 LC = RTLIB::SREM_I8;
1462 else if (VT == MVT::i16)
1463 LC = RTLIB::SREM_I16;
1464 else if (VT == MVT::i32)
1465 LC = RTLIB::SREM_I32;
1466 else if (VT == MVT::i64)
1467 LC = RTLIB::SREM_I64;
1468 else if (VT == MVT::i128)
1469 LC = RTLIB::SREM_I128;
1470 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SREM!");
1472 return ARMEmitLibcall(I, LC);
1475 bool ARMFastISel::SelectBinaryOp(const Instruction *I, unsigned ISDOpcode) {
1476 EVT VT = TLI.getValueType(I->getType(), true);
1478 // We can get here in the case when we want to use NEON for our fp
1479 // operations, but can't figure out how to. Just use the vfp instructions
1481 // FIXME: It'd be nice to use NEON instructions.
1482 Type *Ty = I->getType();
1483 bool isFloat = (Ty->isDoubleTy() || Ty->isFloatTy());
1484 if (isFloat && !Subtarget->hasVFP2())
1487 unsigned Op1 = getRegForValue(I->getOperand(0));
1488 if (Op1 == 0) return false;
1490 unsigned Op2 = getRegForValue(I->getOperand(1));
1491 if (Op2 == 0) return false;
1494 bool is64bit = VT == MVT::f64 || VT == MVT::i64;
1495 switch (ISDOpcode) {
1496 default: return false;
1498 Opc = is64bit ? ARM::VADDD : ARM::VADDS;
1501 Opc = is64bit ? ARM::VSUBD : ARM::VSUBS;
1504 Opc = is64bit ? ARM::VMULD : ARM::VMULS;
1507 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
1508 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1509 TII.get(Opc), ResultReg)
1510 .addReg(Op1).addReg(Op2));
1511 UpdateValueMap(I, ResultReg);
1515 // Call Handling Code
1517 bool ARMFastISel::FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src,
1518 EVT SrcVT, unsigned &ResultReg) {
1519 unsigned RR = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), Opc,
1520 Src, /*TODO: Kill=*/false);
1529 // This is largely taken directly from CCAssignFnForNode - we don't support
1530 // varargs in FastISel so that part has been removed.
1531 // TODO: We may not support all of this.
1532 CCAssignFn *ARMFastISel::CCAssignFnForCall(CallingConv::ID CC, bool Return) {
1535 llvm_unreachable("Unsupported calling convention");
1536 case CallingConv::Fast:
1537 // Ignore fastcc. Silence compiler warnings.
1538 (void)RetFastCC_ARM_APCS;
1539 (void)FastCC_ARM_APCS;
1541 case CallingConv::C:
1542 // Use target triple & subtarget features to do actual dispatch.
1543 if (Subtarget->isAAPCS_ABI()) {
1544 if (Subtarget->hasVFP2() &&
1545 FloatABIType == FloatABI::Hard)
1546 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
1548 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
1550 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
1551 case CallingConv::ARM_AAPCS_VFP:
1552 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
1553 case CallingConv::ARM_AAPCS:
1554 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
1555 case CallingConv::ARM_APCS:
1556 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
1560 bool ARMFastISel::ProcessCallArgs(SmallVectorImpl<Value*> &Args,
1561 SmallVectorImpl<unsigned> &ArgRegs,
1562 SmallVectorImpl<MVT> &ArgVTs,
1563 SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
1564 SmallVectorImpl<unsigned> &RegArgs,
1566 unsigned &NumBytes) {
1567 SmallVector<CCValAssign, 16> ArgLocs;
1568 CCState CCInfo(CC, false, *FuncInfo.MF, TM, ArgLocs, *Context);
1569 CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CCAssignFnForCall(CC, false));
1571 // Get a count of how many bytes are to be pushed on the stack.
1572 NumBytes = CCInfo.getNextStackOffset();
1574 // Issue CALLSEQ_START
1575 unsigned AdjStackDown = TII.getCallFrameSetupOpcode();
1576 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1577 TII.get(AdjStackDown))
1580 // Process the args.
1581 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1582 CCValAssign &VA = ArgLocs[i];
1583 unsigned Arg = ArgRegs[VA.getValNo()];
1584 MVT ArgVT = ArgVTs[VA.getValNo()];
1586 // We don't handle NEON/vector parameters yet.
1587 if (ArgVT.isVector() || ArgVT.getSizeInBits() > 64)
1590 // Handle arg promotion, etc.
1591 switch (VA.getLocInfo()) {
1592 case CCValAssign::Full: break;
1593 case CCValAssign::SExt: {
1594 bool Emitted = FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
1596 assert(Emitted && "Failed to emit a sext!"); (void)Emitted;
1598 ArgVT = VA.getLocVT();
1601 case CCValAssign::ZExt: {
1602 bool Emitted = FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
1604 assert(Emitted && "Failed to emit a zext!"); (void)Emitted;
1606 ArgVT = VA.getLocVT();
1609 case CCValAssign::AExt: {
1610 bool Emitted = FastEmitExtend(ISD::ANY_EXTEND, VA.getLocVT(),
1613 Emitted = FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
1616 Emitted = FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
1619 assert(Emitted && "Failed to emit a aext!"); (void)Emitted;
1620 ArgVT = VA.getLocVT();
1623 case CCValAssign::BCvt: {
1624 unsigned BC = FastEmit_r(ArgVT, VA.getLocVT(), ISD::BITCAST, Arg,
1625 /*TODO: Kill=*/false);
1626 assert(BC != 0 && "Failed to emit a bitcast!");
1628 ArgVT = VA.getLocVT();
1631 default: llvm_unreachable("Unknown arg promotion!");
1634 // Now copy/store arg to correct locations.
1635 if (VA.isRegLoc() && !VA.needsCustom()) {
1636 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1639 RegArgs.push_back(VA.getLocReg());
1640 } else if (VA.needsCustom()) {
1641 // TODO: We need custom lowering for vector (v2f64) args.
1642 if (VA.getLocVT() != MVT::f64) return false;
1644 CCValAssign &NextVA = ArgLocs[++i];
1646 // TODO: Only handle register args for now.
1647 if(!(VA.isRegLoc() && NextVA.isRegLoc())) return false;
1649 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1650 TII.get(ARM::VMOVRRD), VA.getLocReg())
1651 .addReg(NextVA.getLocReg(), RegState::Define)
1653 RegArgs.push_back(VA.getLocReg());
1654 RegArgs.push_back(NextVA.getLocReg());
1656 assert(VA.isMemLoc());
1657 // Need to store on the stack.
1659 Addr.BaseType = Address::RegBase;
1660 Addr.Base.Reg = ARM::SP;
1661 Addr.Offset = VA.getLocMemOffset();
1663 if (!ARMEmitStore(ArgVT, Arg, Addr)) return false;
1669 bool ARMFastISel::FinishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
1670 const Instruction *I, CallingConv::ID CC,
1671 unsigned &NumBytes) {
1672 // Issue CALLSEQ_END
1673 unsigned AdjStackUp = TII.getCallFrameDestroyOpcode();
1674 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1675 TII.get(AdjStackUp))
1676 .addImm(NumBytes).addImm(0));
1678 // Now the return value.
1679 if (RetVT != MVT::isVoid) {
1680 SmallVector<CCValAssign, 16> RVLocs;
1681 CCState CCInfo(CC, false, *FuncInfo.MF, TM, RVLocs, *Context);
1682 CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC, true));
1684 // Copy all of the result registers out of their specified physreg.
1685 if (RVLocs.size() == 2 && RetVT == MVT::f64) {
1686 // For this move we copy into two registers and then move into the
1687 // double fp reg we want.
1688 EVT DestVT = RVLocs[0].getValVT();
1689 TargetRegisterClass* DstRC = TLI.getRegClassFor(DestVT);
1690 unsigned ResultReg = createResultReg(DstRC);
1691 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1692 TII.get(ARM::VMOVDRR), ResultReg)
1693 .addReg(RVLocs[0].getLocReg())
1694 .addReg(RVLocs[1].getLocReg()));
1696 UsedRegs.push_back(RVLocs[0].getLocReg());
1697 UsedRegs.push_back(RVLocs[1].getLocReg());
1699 // Finally update the result.
1700 UpdateValueMap(I, ResultReg);
1702 assert(RVLocs.size() == 1 &&"Can't handle non-double multi-reg retvals!");
1703 EVT CopyVT = RVLocs[0].getValVT();
1704 TargetRegisterClass* DstRC = TLI.getRegClassFor(CopyVT);
1706 unsigned ResultReg = createResultReg(DstRC);
1707 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1708 ResultReg).addReg(RVLocs[0].getLocReg());
1709 UsedRegs.push_back(RVLocs[0].getLocReg());
1711 // Finally update the result.
1712 UpdateValueMap(I, ResultReg);
1719 bool ARMFastISel::SelectRet(const Instruction *I) {
1720 const ReturnInst *Ret = cast<ReturnInst>(I);
1721 const Function &F = *I->getParent()->getParent();
1723 if (!FuncInfo.CanLowerReturn)
1729 CallingConv::ID CC = F.getCallingConv();
1730 if (Ret->getNumOperands() > 0) {
1731 SmallVector<ISD::OutputArg, 4> Outs;
1732 GetReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(),
1735 // Analyze operands of the call, assigning locations to each operand.
1736 SmallVector<CCValAssign, 16> ValLocs;
1737 CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, TM, ValLocs,I->getContext());
1738 CCInfo.AnalyzeReturn(Outs, CCAssignFnForCall(CC, true /* is Ret */));
1740 const Value *RV = Ret->getOperand(0);
1741 unsigned Reg = getRegForValue(RV);
1745 // Only handle a single return value for now.
1746 if (ValLocs.size() != 1)
1749 CCValAssign &VA = ValLocs[0];
1751 // Don't bother handling odd stuff for now.
1752 // FIXME: Should be able to handle i1, i8, and/or i16 return types.
1753 if (VA.getLocInfo() != CCValAssign::Full)
1755 // Only handle register returns for now.
1758 // TODO: For now, don't try to handle cases where getLocInfo()
1759 // says Full but the types don't match.
1760 if (TLI.getValueType(RV->getType()) != VA.getValVT())
1764 unsigned SrcReg = Reg + VA.getValNo();
1765 unsigned DstReg = VA.getLocReg();
1766 const TargetRegisterClass* SrcRC = MRI.getRegClass(SrcReg);
1767 // Avoid a cross-class copy. This is very unlikely.
1768 if (!SrcRC->contains(DstReg))
1770 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1771 DstReg).addReg(SrcReg);
1773 // Mark the register as live out of the function.
1774 MRI.addLiveOut(VA.getLocReg());
1777 unsigned RetOpc = isThumb ? ARM::tBX_RET : ARM::BX_RET;
1778 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1783 unsigned ARMFastISel::ARMSelectCallOp(const GlobalValue *GV) {
1785 // Darwin needs the r9 versions of the opcodes.
1786 bool isDarwin = Subtarget->isTargetDarwin();
1788 return isDarwin ? ARM::tBLr9 : ARM::tBL;
1790 return isDarwin ? ARM::BLr9 : ARM::BL;
1794 // A quick function that will emit a call for a named libcall in F with the
1795 // vector of passed arguments for the Instruction in I. We can assume that we
1796 // can emit a call for any libcall we can produce. This is an abridged version
1797 // of the full call infrastructure since we won't need to worry about things
1798 // like computed function pointers or strange arguments at call sites.
1799 // TODO: Try to unify this and the normal call bits for ARM, then try to unify
1801 bool ARMFastISel::ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call) {
1802 CallingConv::ID CC = TLI.getLibcallCallingConv(Call);
1804 // Handle *simple* calls for now.
1805 Type *RetTy = I->getType();
1807 if (RetTy->isVoidTy())
1808 RetVT = MVT::isVoid;
1809 else if (!isTypeLegal(RetTy, RetVT))
1812 // TODO: For now if we have long calls specified we don't handle the call.
1813 if (EnableARMLongCalls) return false;
1815 // Set up the argument vectors.
1816 SmallVector<Value*, 8> Args;
1817 SmallVector<unsigned, 8> ArgRegs;
1818 SmallVector<MVT, 8> ArgVTs;
1819 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
1820 Args.reserve(I->getNumOperands());
1821 ArgRegs.reserve(I->getNumOperands());
1822 ArgVTs.reserve(I->getNumOperands());
1823 ArgFlags.reserve(I->getNumOperands());
1824 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
1825 Value *Op = I->getOperand(i);
1826 unsigned Arg = getRegForValue(Op);
1827 if (Arg == 0) return false;
1829 Type *ArgTy = Op->getType();
1831 if (!isTypeLegal(ArgTy, ArgVT)) return false;
1833 ISD::ArgFlagsTy Flags;
1834 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
1835 Flags.setOrigAlign(OriginalAlignment);
1838 ArgRegs.push_back(Arg);
1839 ArgVTs.push_back(ArgVT);
1840 ArgFlags.push_back(Flags);
1843 // Handle the arguments now that we've gotten them.
1844 SmallVector<unsigned, 4> RegArgs;
1846 if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
1849 // Issue the call, BLr9 for darwin, BL otherwise.
1850 // TODO: Turn this into the table of arm call ops.
1851 MachineInstrBuilder MIB;
1852 unsigned CallOpc = ARMSelectCallOp(NULL);
1854 // Explicitly adding the predicate here.
1855 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1857 .addExternalSymbol(TLI.getLibcallName(Call));
1859 // Explicitly adding the predicate here.
1860 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1862 .addExternalSymbol(TLI.getLibcallName(Call)));
1864 // Add implicit physical register uses to the call.
1865 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
1866 MIB.addReg(RegArgs[i]);
1868 // Finish off the call including any return values.
1869 SmallVector<unsigned, 4> UsedRegs;
1870 if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false;
1872 // Set all unused physreg defs as dead.
1873 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
1878 bool ARMFastISel::SelectCall(const Instruction *I) {
1879 const CallInst *CI = cast<CallInst>(I);
1880 const Value *Callee = CI->getCalledValue();
1882 // Can't handle inline asm or worry about intrinsics yet.
1883 if (isa<InlineAsm>(Callee) || isa<IntrinsicInst>(CI)) return false;
1885 // Only handle global variable Callees.
1886 const GlobalValue *GV = dyn_cast<GlobalValue>(Callee);
1890 // Check the calling convention.
1891 ImmutableCallSite CS(CI);
1892 CallingConv::ID CC = CS.getCallingConv();
1894 // TODO: Avoid some calling conventions?
1896 // Let SDISel handle vararg functions.
1897 PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
1898 FunctionType *FTy = cast<FunctionType>(PT->getElementType());
1899 if (FTy->isVarArg())
1902 // Handle *simple* calls for now.
1903 Type *RetTy = I->getType();
1905 if (RetTy->isVoidTy())
1906 RetVT = MVT::isVoid;
1907 else if (!isTypeLegal(RetTy, RetVT))
1910 // TODO: For now if we have long calls specified we don't handle the call.
1911 if (EnableARMLongCalls) return false;
1913 // Set up the argument vectors.
1914 SmallVector<Value*, 8> Args;
1915 SmallVector<unsigned, 8> ArgRegs;
1916 SmallVector<MVT, 8> ArgVTs;
1917 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
1918 Args.reserve(CS.arg_size());
1919 ArgRegs.reserve(CS.arg_size());
1920 ArgVTs.reserve(CS.arg_size());
1921 ArgFlags.reserve(CS.arg_size());
1922 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
1924 unsigned Arg = getRegForValue(*i);
1928 ISD::ArgFlagsTy Flags;
1929 unsigned AttrInd = i - CS.arg_begin() + 1;
1930 if (CS.paramHasAttr(AttrInd, Attribute::SExt))
1932 if (CS.paramHasAttr(AttrInd, Attribute::ZExt))
1935 // FIXME: Only handle *easy* calls for now.
1936 if (CS.paramHasAttr(AttrInd, Attribute::InReg) ||
1937 CS.paramHasAttr(AttrInd, Attribute::StructRet) ||
1938 CS.paramHasAttr(AttrInd, Attribute::Nest) ||
1939 CS.paramHasAttr(AttrInd, Attribute::ByVal))
1942 Type *ArgTy = (*i)->getType();
1944 // FIXME: Should be able to handle i1, i8, and/or i16 parameters.
1945 if (!isTypeLegal(ArgTy, ArgVT))
1947 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
1948 Flags.setOrigAlign(OriginalAlignment);
1951 ArgRegs.push_back(Arg);
1952 ArgVTs.push_back(ArgVT);
1953 ArgFlags.push_back(Flags);
1956 // Handle the arguments now that we've gotten them.
1957 SmallVector<unsigned, 4> RegArgs;
1959 if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
1962 // Issue the call, BLr9 for darwin, BL otherwise.
1963 // TODO: Turn this into the table of arm call ops.
1964 MachineInstrBuilder MIB;
1965 unsigned CallOpc = ARMSelectCallOp(GV);
1966 // Explicitly adding the predicate here.
1968 // Explicitly adding the predicate here.
1969 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1971 .addGlobalAddress(GV, 0, 0);
1973 // Explicitly adding the predicate here.
1974 MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1976 .addGlobalAddress(GV, 0, 0));
1978 // Add implicit physical register uses to the call.
1979 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
1980 MIB.addReg(RegArgs[i]);
1982 // Finish off the call including any return values.
1983 SmallVector<unsigned, 4> UsedRegs;
1984 if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false;
1986 // Set all unused physreg defs as dead.
1987 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
1993 bool ARMFastISel::SelectIntCast(const Instruction *I) {
1994 // On ARM, in general, integer casts don't involve legal types; this code
1995 // handles promotable integers. The high bits for a type smaller than
1996 // the register size are assumed to be undefined.
1997 Type *DestTy = I->getType();
1998 Value *Op = I->getOperand(0);
1999 Type *SrcTy = Op->getType();
2002 SrcVT = TLI.getValueType(SrcTy, true);
2003 DestVT = TLI.getValueType(DestTy, true);
2005 if (isa<TruncInst>(I)) {
2006 if (SrcVT != MVT::i32 && SrcVT != MVT::i16 && SrcVT != MVT::i8)
2008 if (DestVT != MVT::i16 && DestVT != MVT::i8 && DestVT != MVT::i1)
2011 unsigned SrcReg = getRegForValue(Op);
2012 if (!SrcReg) return false;
2014 // Because the high bits are undefined, a truncate doesn't generate
2016 UpdateValueMap(I, SrcReg);
2019 if (DestVT != MVT::i32 && DestVT != MVT::i16 && DestVT != MVT::i8)
2023 bool isZext = isa<ZExtInst>(I);
2024 bool isBoolZext = false;
2025 if (!SrcVT.isSimple())
2027 switch (SrcVT.getSimpleVT().SimpleTy) {
2028 default: return false;
2030 if (!Subtarget->hasV6Ops()) return false;
2032 Opc = isThumb ? ARM::t2UXTH : ARM::UXTH;
2034 Opc = isThumb ? ARM::t2SXTH : ARM::SXTH;
2037 if (!Subtarget->hasV6Ops()) return false;
2039 Opc = isThumb ? ARM::t2UXTB : ARM::UXTB;
2041 Opc = isThumb ? ARM::t2SXTB : ARM::SXTB;
2045 Opc = isThumb ? ARM::t2ANDri : ARM::ANDri;
2052 // FIXME: We could save an instruction in many cases by special-casing
2053 // load instructions.
2054 unsigned SrcReg = getRegForValue(Op);
2055 if (!SrcReg) return false;
2057 unsigned DestReg = createResultReg(TLI.getRegClassFor(MVT::i32));
2058 MachineInstrBuilder MIB;
2059 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), DestReg)
2065 AddOptionalDefs(MIB);
2066 UpdateValueMap(I, DestReg);
2070 // TODO: SoftFP support.
2071 bool ARMFastISel::TargetSelectInstruction(const Instruction *I) {
2073 switch (I->getOpcode()) {
2074 case Instruction::Load:
2075 return SelectLoad(I);
2076 case Instruction::Store:
2077 return SelectStore(I);
2078 case Instruction::Br:
2079 return SelectBranch(I);
2080 case Instruction::ICmp:
2081 case Instruction::FCmp:
2082 return SelectCmp(I);
2083 case Instruction::FPExt:
2084 return SelectFPExt(I);
2085 case Instruction::FPTrunc:
2086 return SelectFPTrunc(I);
2087 case Instruction::SIToFP:
2088 return SelectSIToFP(I);
2089 case Instruction::FPToSI:
2090 return SelectFPToSI(I);
2091 case Instruction::FAdd:
2092 return SelectBinaryOp(I, ISD::FADD);
2093 case Instruction::FSub:
2094 return SelectBinaryOp(I, ISD::FSUB);
2095 case Instruction::FMul:
2096 return SelectBinaryOp(I, ISD::FMUL);
2097 case Instruction::SDiv:
2098 return SelectSDiv(I);
2099 case Instruction::SRem:
2100 return SelectSRem(I);
2101 case Instruction::Call:
2102 return SelectCall(I);
2103 case Instruction::Select:
2104 return SelectSelect(I);
2105 case Instruction::Ret:
2106 return SelectRet(I);
2107 case Instruction::Trunc:
2108 case Instruction::ZExt:
2109 case Instruction::SExt:
2110 return SelectIntCast(I);
2117 llvm::FastISel *ARM::createFastISel(FunctionLoweringInfo &funcInfo) {
2118 // Completely untested on non-darwin.
2119 const TargetMachine &TM = funcInfo.MF->getTarget();
2121 // Darwin and thumb1 only for now.
2122 const ARMSubtarget *Subtarget = &TM.getSubtarget<ARMSubtarget>();
2123 if (Subtarget->isTargetDarwin() && !Subtarget->isThumb1Only() &&
2124 !DisableARMFastISel)
2125 return new ARMFastISel(funcInfo);