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 "llvm/CallingConv.h"
24 #include "llvm/DerivedTypes.h"
25 #include "llvm/GlobalVariable.h"
26 #include "llvm/Instructions.h"
27 #include "llvm/IntrinsicInst.h"
28 #include "llvm/Module.h"
29 #include "llvm/CodeGen/Analysis.h"
30 #include "llvm/CodeGen/FastISel.h"
31 #include "llvm/CodeGen/FunctionLoweringInfo.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineModuleInfo.h"
34 #include "llvm/CodeGen/MachineConstantPool.h"
35 #include "llvm/CodeGen/MachineFrameInfo.h"
36 #include "llvm/CodeGen/MachineMemOperand.h"
37 #include "llvm/CodeGen/MachineRegisterInfo.h"
38 #include "llvm/CodeGen/PseudoSourceValue.h"
39 #include "llvm/Support/CallSite.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/ErrorHandling.h"
42 #include "llvm/Support/GetElementPtrTypeIterator.h"
43 #include "llvm/Target/TargetData.h"
44 #include "llvm/Target/TargetInstrInfo.h"
45 #include "llvm/Target/TargetLowering.h"
46 #include "llvm/Target/TargetMachine.h"
47 #include "llvm/Target/TargetOptions.h"
51 DisableARMFastISel("disable-arm-fast-isel",
52 cl::desc("Turn off experimental ARM fast-isel support"),
53 cl::init(false), cl::Hidden);
57 class ARMFastISel : public FastISel {
59 /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
60 /// make the right decision when generating code for different targets.
61 const ARMSubtarget *Subtarget;
62 const TargetMachine &TM;
63 const TargetInstrInfo &TII;
64 const TargetLowering &TLI;
67 // Convenience variables to avoid some queries.
72 explicit ARMFastISel(FunctionLoweringInfo &funcInfo)
74 TM(funcInfo.MF->getTarget()),
75 TII(*TM.getInstrInfo()),
76 TLI(*TM.getTargetLowering()) {
77 Subtarget = &TM.getSubtarget<ARMSubtarget>();
78 AFI = funcInfo.MF->getInfo<ARMFunctionInfo>();
79 isThumb = AFI->isThumbFunction();
80 Context = &funcInfo.Fn->getContext();
83 // Code from FastISel.cpp.
84 virtual unsigned FastEmitInst_(unsigned MachineInstOpcode,
85 const TargetRegisterClass *RC);
86 virtual unsigned FastEmitInst_r(unsigned MachineInstOpcode,
87 const TargetRegisterClass *RC,
88 unsigned Op0, bool Op0IsKill);
89 virtual unsigned FastEmitInst_rr(unsigned MachineInstOpcode,
90 const TargetRegisterClass *RC,
91 unsigned Op0, bool Op0IsKill,
92 unsigned Op1, bool Op1IsKill);
93 virtual unsigned FastEmitInst_ri(unsigned MachineInstOpcode,
94 const TargetRegisterClass *RC,
95 unsigned Op0, bool Op0IsKill,
97 virtual unsigned FastEmitInst_rf(unsigned MachineInstOpcode,
98 const TargetRegisterClass *RC,
99 unsigned Op0, bool Op0IsKill,
100 const ConstantFP *FPImm);
101 virtual unsigned FastEmitInst_i(unsigned MachineInstOpcode,
102 const TargetRegisterClass *RC,
104 virtual unsigned FastEmitInst_rri(unsigned MachineInstOpcode,
105 const TargetRegisterClass *RC,
106 unsigned Op0, bool Op0IsKill,
107 unsigned Op1, bool Op1IsKill,
109 virtual unsigned FastEmitInst_extractsubreg(MVT RetVT,
110 unsigned Op0, bool Op0IsKill,
113 // Backend specific FastISel code.
114 virtual bool TargetSelectInstruction(const Instruction *I);
115 virtual unsigned TargetMaterializeConstant(const Constant *C);
116 virtual unsigned TargetMaterializeAlloca(const AllocaInst *AI);
118 #include "ARMGenFastISel.inc"
120 // Instruction selection routines.
122 bool SelectLoad(const Instruction *I);
123 bool SelectStore(const Instruction *I);
124 bool SelectBranch(const Instruction *I);
125 bool SelectCmp(const Instruction *I);
126 bool SelectFPExt(const Instruction *I);
127 bool SelectFPTrunc(const Instruction *I);
128 bool SelectBinaryOp(const Instruction *I, unsigned ISDOpcode);
129 bool SelectSIToFP(const Instruction *I);
130 bool SelectFPToSI(const Instruction *I);
131 bool SelectSDiv(const Instruction *I);
132 bool SelectSRem(const Instruction *I);
133 bool SelectCall(const Instruction *I);
134 bool SelectSelect(const Instruction *I);
135 bool SelectRet(const Instruction *I);
139 bool isTypeLegal(const Type *Ty, EVT &VT);
140 bool isLoadTypeLegal(const Type *Ty, EVT &VT);
141 bool ARMEmitLoad(EVT VT, unsigned &ResultReg, unsigned Base, int Offset);
142 bool ARMEmitStore(EVT VT, unsigned SrcReg, unsigned Base, int Offset);
143 bool ARMComputeRegOffset(const Value *Obj, unsigned &Base, int &Offset);
144 void ARMSimplifyRegOffset(unsigned &Base, int &Offset, EVT VT);
145 unsigned ARMMaterializeFP(const ConstantFP *CFP, EVT VT);
146 unsigned ARMMaterializeInt(const Constant *C, EVT VT);
147 unsigned ARMMaterializeGV(const GlobalValue *GV, EVT VT);
148 unsigned ARMMoveToFPReg(EVT VT, unsigned SrcReg);
149 unsigned ARMMoveToIntReg(EVT VT, unsigned SrcReg);
151 // Call handling routines.
153 bool FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src, EVT SrcVT,
154 unsigned &ResultReg);
155 CCAssignFn *CCAssignFnForCall(CallingConv::ID CC, bool Return);
156 bool ProcessCallArgs(SmallVectorImpl<Value*> &Args,
157 SmallVectorImpl<unsigned> &ArgRegs,
158 SmallVectorImpl<EVT> &ArgVTs,
159 SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
160 SmallVectorImpl<unsigned> &RegArgs,
163 bool FinishCall(EVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
164 const Instruction *I, CallingConv::ID CC,
166 bool ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call);
168 // OptionalDef handling routines.
170 bool DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR);
171 const MachineInstrBuilder &AddOptionalDefs(const MachineInstrBuilder &MIB);
174 } // end anonymous namespace
176 #include "ARMGenCallingConv.inc"
178 // DefinesOptionalPredicate - This is different from DefinesPredicate in that
179 // we don't care about implicit defs here, just places we'll need to add a
180 // default CCReg argument. Sets CPSR if we're setting CPSR instead of CCR.
181 bool ARMFastISel::DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR) {
182 const TargetInstrDesc &TID = MI->getDesc();
183 if (!TID.hasOptionalDef())
186 // Look to see if our OptionalDef is defining CPSR or CCR.
187 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
188 const MachineOperand &MO = MI->getOperand(i);
189 if (!MO.isReg() || !MO.isDef()) continue;
190 if (MO.getReg() == ARM::CPSR)
196 // If the machine is predicable go ahead and add the predicate operands, if
197 // it needs default CC operands add those.
198 // TODO: If we want to support thumb1 then we'll need to deal with optional
199 // CPSR defs that need to be added before the remaining operands. See s_cc_out
200 // for descriptions why.
201 const MachineInstrBuilder &
202 ARMFastISel::AddOptionalDefs(const MachineInstrBuilder &MIB) {
203 MachineInstr *MI = &*MIB;
205 // Do we use a predicate?
206 if (TII.isPredicable(MI))
209 // Do we optionally set a predicate? Preds is size > 0 iff the predicate
210 // defines CPSR. All other OptionalDefines in ARM are the CCR register.
212 if (DefinesOptionalPredicate(MI, &CPSR)) {
221 unsigned ARMFastISel::FastEmitInst_(unsigned MachineInstOpcode,
222 const TargetRegisterClass* RC) {
223 unsigned ResultReg = createResultReg(RC);
224 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
226 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg));
230 unsigned ARMFastISel::FastEmitInst_r(unsigned MachineInstOpcode,
231 const TargetRegisterClass *RC,
232 unsigned Op0, bool Op0IsKill) {
233 unsigned ResultReg = createResultReg(RC);
234 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
236 if (II.getNumDefs() >= 1)
237 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
238 .addReg(Op0, Op0IsKill * RegState::Kill));
240 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
241 .addReg(Op0, Op0IsKill * RegState::Kill));
242 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
243 TII.get(TargetOpcode::COPY), ResultReg)
244 .addReg(II.ImplicitDefs[0]));
249 unsigned ARMFastISel::FastEmitInst_rr(unsigned MachineInstOpcode,
250 const TargetRegisterClass *RC,
251 unsigned Op0, bool Op0IsKill,
252 unsigned Op1, bool Op1IsKill) {
253 unsigned ResultReg = createResultReg(RC);
254 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
256 if (II.getNumDefs() >= 1)
257 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
258 .addReg(Op0, Op0IsKill * RegState::Kill)
259 .addReg(Op1, Op1IsKill * RegState::Kill));
261 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
262 .addReg(Op0, Op0IsKill * RegState::Kill)
263 .addReg(Op1, Op1IsKill * RegState::Kill));
264 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
265 TII.get(TargetOpcode::COPY), ResultReg)
266 .addReg(II.ImplicitDefs[0]));
271 unsigned ARMFastISel::FastEmitInst_ri(unsigned MachineInstOpcode,
272 const TargetRegisterClass *RC,
273 unsigned Op0, bool Op0IsKill,
275 unsigned ResultReg = createResultReg(RC);
276 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
278 if (II.getNumDefs() >= 1)
279 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
280 .addReg(Op0, Op0IsKill * RegState::Kill)
283 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
284 .addReg(Op0, Op0IsKill * RegState::Kill)
286 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
287 TII.get(TargetOpcode::COPY), ResultReg)
288 .addReg(II.ImplicitDefs[0]));
293 unsigned ARMFastISel::FastEmitInst_rf(unsigned MachineInstOpcode,
294 const TargetRegisterClass *RC,
295 unsigned Op0, bool Op0IsKill,
296 const ConstantFP *FPImm) {
297 unsigned ResultReg = createResultReg(RC);
298 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
300 if (II.getNumDefs() >= 1)
301 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
302 .addReg(Op0, Op0IsKill * RegState::Kill)
305 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
306 .addReg(Op0, Op0IsKill * RegState::Kill)
308 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
309 TII.get(TargetOpcode::COPY), ResultReg)
310 .addReg(II.ImplicitDefs[0]));
315 unsigned ARMFastISel::FastEmitInst_rri(unsigned MachineInstOpcode,
316 const TargetRegisterClass *RC,
317 unsigned Op0, bool Op0IsKill,
318 unsigned Op1, bool Op1IsKill,
320 unsigned ResultReg = createResultReg(RC);
321 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
323 if (II.getNumDefs() >= 1)
324 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
325 .addReg(Op0, Op0IsKill * RegState::Kill)
326 .addReg(Op1, Op1IsKill * RegState::Kill)
329 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
330 .addReg(Op0, Op0IsKill * RegState::Kill)
331 .addReg(Op1, Op1IsKill * RegState::Kill)
333 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
334 TII.get(TargetOpcode::COPY), ResultReg)
335 .addReg(II.ImplicitDefs[0]));
340 unsigned ARMFastISel::FastEmitInst_i(unsigned MachineInstOpcode,
341 const TargetRegisterClass *RC,
343 unsigned ResultReg = createResultReg(RC);
344 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
346 if (II.getNumDefs() >= 1)
347 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
350 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
352 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
353 TII.get(TargetOpcode::COPY), ResultReg)
354 .addReg(II.ImplicitDefs[0]));
359 unsigned ARMFastISel::FastEmitInst_extractsubreg(MVT RetVT,
360 unsigned Op0, bool Op0IsKill,
362 unsigned ResultReg = createResultReg(TLI.getRegClassFor(RetVT));
363 assert(TargetRegisterInfo::isVirtualRegister(Op0) &&
364 "Cannot yet extract from physregs");
365 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
366 DL, TII.get(TargetOpcode::COPY), ResultReg)
367 .addReg(Op0, getKillRegState(Op0IsKill), Idx));
371 // TODO: Don't worry about 64-bit now, but when this is fixed remove the
372 // checks from the various callers.
373 unsigned ARMFastISel::ARMMoveToFPReg(EVT VT, unsigned SrcReg) {
374 if (VT.getSimpleVT().SimpleTy == MVT::f64) return 0;
376 unsigned MoveReg = createResultReg(TLI.getRegClassFor(VT));
377 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
378 TII.get(ARM::VMOVRS), MoveReg)
383 unsigned ARMFastISel::ARMMoveToIntReg(EVT VT, unsigned SrcReg) {
384 if (VT.getSimpleVT().SimpleTy == MVT::i64) return 0;
386 unsigned MoveReg = createResultReg(TLI.getRegClassFor(VT));
387 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
388 TII.get(ARM::VMOVSR), MoveReg)
393 // For double width floating point we need to materialize two constants
394 // (the high and the low) into integer registers then use a move to get
395 // the combined constant into an FP reg.
396 unsigned ARMFastISel::ARMMaterializeFP(const ConstantFP *CFP, EVT VT) {
397 const APFloat Val = CFP->getValueAPF();
398 bool is64bit = VT.getSimpleVT().SimpleTy == MVT::f64;
400 // This checks to see if we can use VFP3 instructions to materialize
401 // a constant, otherwise we have to go through the constant pool.
402 if (TLI.isFPImmLegal(Val, VT)) {
403 unsigned Opc = is64bit ? ARM::FCONSTD : ARM::FCONSTS;
404 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
405 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
411 // Require VFP2 for loading fp constants.
412 if (!Subtarget->hasVFP2()) return false;
414 // MachineConstantPool wants an explicit alignment.
415 unsigned Align = TD.getPrefTypeAlignment(CFP->getType());
417 // TODO: Figure out if this is correct.
418 Align = TD.getTypeAllocSize(CFP->getType());
420 unsigned Idx = MCP.getConstantPoolIndex(cast<Constant>(CFP), Align);
421 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
422 unsigned Opc = is64bit ? ARM::VLDRD : ARM::VLDRS;
424 // The extra reg is for addrmode5.
425 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
427 .addConstantPoolIndex(Idx)
432 unsigned ARMFastISel::ARMMaterializeInt(const Constant *C, EVT VT) {
434 // For now 32-bit only.
435 if (VT.getSimpleVT().SimpleTy != MVT::i32) return false;
437 // MachineConstantPool wants an explicit alignment.
438 unsigned Align = TD.getPrefTypeAlignment(C->getType());
440 // TODO: Figure out if this is correct.
441 Align = TD.getTypeAllocSize(C->getType());
443 unsigned Idx = MCP.getConstantPoolIndex(C, Align);
444 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
447 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
448 TII.get(ARM::t2LDRpci), DestReg)
449 .addConstantPoolIndex(Idx));
451 // The extra reg and immediate are for addrmode2.
452 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
453 TII.get(ARM::LDRcp), DestReg)
454 .addConstantPoolIndex(Idx)
460 unsigned ARMFastISel::ARMMaterializeGV(const GlobalValue *GV, EVT VT) {
461 // For now 32-bit only.
462 if (VT.getSimpleVT().SimpleTy != MVT::i32) return 0;
464 Reloc::Model RelocM = TM.getRelocationModel();
466 // TODO: No external globals for now.
467 if (Subtarget->GVIsIndirectSymbol(GV, RelocM)) return 0;
469 // TODO: Need more magic for ARM PIC.
470 if (!isThumb && (RelocM == Reloc::PIC_)) return 0;
472 // MachineConstantPool wants an explicit alignment.
473 unsigned Align = TD.getPrefTypeAlignment(GV->getType());
475 // TODO: Figure out if this is correct.
476 Align = TD.getTypeAllocSize(GV->getType());
480 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb() ? 4 : 8);
481 unsigned Id = AFI->createConstPoolEntryUId();
482 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, Id,
483 ARMCP::CPValue, PCAdj);
484 unsigned Idx = MCP.getConstantPoolIndex(CPV, Align);
487 MachineInstrBuilder MIB;
488 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
490 unsigned Opc = (RelocM != Reloc::PIC_) ? ARM::t2LDRpci : ARM::t2LDRpci_pic;
491 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), DestReg)
492 .addConstantPoolIndex(Idx);
493 if (RelocM == Reloc::PIC_)
496 // The extra reg and immediate are for addrmode2.
497 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(ARM::LDRcp),
499 .addConstantPoolIndex(Idx)
500 .addReg(0).addImm(0);
502 AddOptionalDefs(MIB);
506 unsigned ARMFastISel::TargetMaterializeConstant(const Constant *C) {
507 EVT VT = TLI.getValueType(C->getType(), true);
509 // Only handle simple types.
510 if (!VT.isSimple()) return 0;
512 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
513 return ARMMaterializeFP(CFP, VT);
514 else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
515 return ARMMaterializeGV(GV, VT);
516 else if (isa<ConstantInt>(C))
517 return ARMMaterializeInt(C, VT);
522 unsigned ARMFastISel::TargetMaterializeAlloca(const AllocaInst *AI) {
523 // Don't handle dynamic allocas.
524 if (!FuncInfo.StaticAllocaMap.count(AI)) return 0;
527 if (!isLoadTypeLegal(AI->getType(), VT)) return false;
529 DenseMap<const AllocaInst*, int>::iterator SI =
530 FuncInfo.StaticAllocaMap.find(AI);
532 // This will get lowered later into the correct offsets and registers
533 // via rewriteXFrameIndex.
534 if (SI != FuncInfo.StaticAllocaMap.end()) {
535 TargetRegisterClass* RC = TLI.getRegClassFor(VT);
536 unsigned ResultReg = createResultReg(RC);
537 unsigned Opc = isThumb ? ARM::t2ADDri : ARM::ADDri;
538 AddOptionalDefs(BuildMI(*FuncInfo.MBB, *FuncInfo.InsertPt, DL,
539 TII.get(Opc), ResultReg)
540 .addFrameIndex(SI->second)
548 bool ARMFastISel::isTypeLegal(const Type *Ty, EVT &VT) {
549 VT = TLI.getValueType(Ty, true);
551 // Only handle simple types.
552 if (VT == MVT::Other || !VT.isSimple()) return false;
554 // Handle all legal types, i.e. a register that will directly hold this
556 return TLI.isTypeLegal(VT);
559 bool ARMFastISel::isLoadTypeLegal(const Type *Ty, EVT &VT) {
560 if (isTypeLegal(Ty, VT)) return true;
562 // If this is a type than can be sign or zero-extended to a basic operation
563 // go ahead and accept it now.
564 if (VT == MVT::i8 || VT == MVT::i16)
570 // Computes the Reg+Offset to get to an object.
571 bool ARMFastISel::ARMComputeRegOffset(const Value *Obj, unsigned &Base,
573 // Some boilerplate from the X86 FastISel.
574 const User *U = NULL;
575 unsigned Opcode = Instruction::UserOp1;
576 if (const Instruction *I = dyn_cast<Instruction>(Obj)) {
577 // Don't walk into other basic blocks; it's possible we haven't
578 // visited them yet, so the instructions may not yet be assigned
579 // virtual registers.
580 if (FuncInfo.MBBMap[I->getParent()] != FuncInfo.MBB)
582 Opcode = I->getOpcode();
584 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Obj)) {
585 Opcode = C->getOpcode();
589 if (const PointerType *Ty = dyn_cast<PointerType>(Obj->getType()))
590 if (Ty->getAddressSpace() > 255)
591 // Fast instruction selection doesn't support the special
598 case Instruction::BitCast: {
599 // Look through bitcasts.
600 return ARMComputeRegOffset(U->getOperand(0), Base, Offset);
602 case Instruction::IntToPtr: {
603 // Look past no-op inttoptrs.
604 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
605 return ARMComputeRegOffset(U->getOperand(0), Base, Offset);
608 case Instruction::PtrToInt: {
609 // Look past no-op ptrtoints.
610 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
611 return ARMComputeRegOffset(U->getOperand(0), Base, Offset);
614 case Instruction::GetElementPtr: {
615 int SavedOffset = Offset;
616 unsigned SavedBase = Base;
617 int TmpOffset = Offset;
619 // Iterate through the GEP folding the constants into offsets where
621 gep_type_iterator GTI = gep_type_begin(U);
622 for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end();
623 i != e; ++i, ++GTI) {
624 const Value *Op = *i;
625 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
626 const StructLayout *SL = TD.getStructLayout(STy);
627 unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
628 TmpOffset += SL->getElementOffset(Idx);
630 uint64_t S = TD.getTypeAllocSize(GTI.getIndexedType());
631 SmallVector<const Value *, 4> Worklist;
632 Worklist.push_back(Op);
634 Op = Worklist.pop_back_val();
635 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
636 // Constant-offset addressing.
637 TmpOffset += CI->getSExtValue() * S;
638 } else if (isa<AddOperator>(Op) &&
639 isa<ConstantInt>(cast<AddOperator>(Op)->getOperand(1))) {
640 // An add with a constant operand. Fold the constant.
642 cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
643 TmpOffset += CI->getSExtValue() * S;
644 // Add the other operand back to the work list.
645 Worklist.push_back(cast<AddOperator>(Op)->getOperand(0));
647 goto unsupported_gep;
648 } while (!Worklist.empty());
652 // Try to grab the base operand now.
654 if (ARMComputeRegOffset(U->getOperand(0), Base, Offset)) return true;
656 // We failed, restore everything and try the other options.
657 Offset = SavedOffset;
663 case Instruction::Alloca: {
664 const AllocaInst *AI = cast<AllocaInst>(Obj);
665 unsigned Reg = TargetMaterializeAlloca(AI);
667 if (Reg == 0) return false;
674 // Materialize the global variable's address into a reg which can
675 // then be used later to load the variable.
676 if (const GlobalValue *GV = dyn_cast<GlobalValue>(Obj)) {
677 unsigned Tmp = ARMMaterializeGV(GV, TLI.getValueType(Obj->getType()));
678 if (Tmp == 0) return false;
684 // Try to get this in a register if nothing else has worked.
685 if (Base == 0) Base = getRegForValue(Obj);
689 void ARMFastISel::ARMSimplifyRegOffset(unsigned &Base, int &Offset, EVT VT) {
691 assert(VT.isSimple() && "Non-simple types are invalid here!");
693 bool needsLowering = false;
694 switch (VT.getSimpleVT().SimpleTy) {
696 assert(false && "Unhandled load/store type!");
701 // Integer loads/stores handle 12-bit offsets.
702 needsLowering = ((Offset & 0xfff) != Offset);
706 // Floating point operands handle 8-bit offsets.
707 needsLowering = ((Offset & 0xff) != Offset);
711 // Since the offset is too large for the load/store instruction
712 // get the reg+offset into a register.
714 ARMCC::CondCodes Pred = ARMCC::AL;
715 unsigned PredReg = 0;
717 TargetRegisterClass *RC = isThumb ? ARM::tGPRRegisterClass :
718 ARM::GPRRegisterClass;
719 unsigned BaseReg = createResultReg(RC);
722 emitARMRegPlusImmediate(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
723 BaseReg, Base, Offset, Pred, PredReg,
724 static_cast<const ARMBaseInstrInfo&>(TII));
726 assert(AFI->isThumb2Function());
727 emitT2RegPlusImmediate(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
728 BaseReg, Base, Offset, Pred, PredReg,
729 static_cast<const ARMBaseInstrInfo&>(TII));
736 bool ARMFastISel::ARMEmitLoad(EVT VT, unsigned &ResultReg,
737 unsigned Base, int Offset) {
739 assert(VT.isSimple() && "Non-simple types are invalid here!");
741 TargetRegisterClass *RC;
742 bool isFloat = false;
743 switch (VT.getSimpleVT().SimpleTy) {
745 // This is mostly going to be Neon/vector support.
748 Opc = isThumb ? ARM::t2LDRHi12 : ARM::LDRH;
749 RC = ARM::GPRRegisterClass;
752 Opc = isThumb ? ARM::t2LDRBi12 : ARM::LDRBi12;
753 RC = ARM::GPRRegisterClass;
756 Opc = isThumb ? ARM::t2LDRi12 : ARM::LDRi12;
757 RC = ARM::GPRRegisterClass;
761 RC = TLI.getRegClassFor(VT);
766 RC = TLI.getRegClassFor(VT);
771 ResultReg = createResultReg(RC);
773 ARMSimplifyRegOffset(Base, Offset, VT);
775 // addrmode5 output depends on the selection dag addressing dividing the
776 // offset by 4 that it then later multiplies. Do this here as well.
780 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
781 TII.get(Opc), ResultReg)
782 .addReg(Base).addImm(Offset));
786 bool ARMFastISel::SelectLoad(const Instruction *I) {
787 // Verify we have a legal type before going any further.
789 if (!isLoadTypeLegal(I->getType(), VT))
792 // Our register and offset with innocuous defaults.
796 // See if we can handle this as Reg + Offset
797 if (!ARMComputeRegOffset(I->getOperand(0), Base, Offset))
801 if (!ARMEmitLoad(VT, ResultReg, Base, Offset)) return false;
803 UpdateValueMap(I, ResultReg);
807 bool ARMFastISel::ARMEmitStore(EVT VT, unsigned SrcReg,
808 unsigned Base, int Offset) {
810 bool isFloat = false;
811 bool needReg0Op = false;
812 switch (VT.getSimpleVT().SimpleTy) {
813 default: return false;
816 StrOpc = isThumb ? ARM::t2STRBi12 : ARM::STRBi12;
819 StrOpc = isThumb ? ARM::t2STRHi12 : ARM::STRH;
823 StrOpc = isThumb ? ARM::t2STRi12 : ARM::STRi12;
826 if (!Subtarget->hasVFP2()) return false;
831 if (!Subtarget->hasVFP2()) return false;
837 ARMSimplifyRegOffset(Base, Offset, VT);
839 // addrmode5 output depends on the selection dag addressing dividing the
840 // offset by 4 that it then later multiplies. Do this here as well.
845 // FIXME: The 'needReg0Op' bit goes away once STRH is converted to
846 // not use the mega-addrmode stuff.
848 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
850 .addReg(SrcReg).addReg(Base).addImm(Offset));
852 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
854 .addReg(SrcReg).addReg(Base).addReg(0).addImm(Offset));
859 bool ARMFastISel::SelectStore(const Instruction *I) {
860 Value *Op0 = I->getOperand(0);
863 // Yay type legalization
865 if (!isLoadTypeLegal(I->getOperand(0)->getType(), VT))
868 // Get the value to be stored into a register.
869 SrcReg = getRegForValue(Op0);
873 // Our register and offset with innocuous defaults.
877 // See if we can handle this as Reg + Offset
878 if (!ARMComputeRegOffset(I->getOperand(1), Base, Offset))
881 if (!ARMEmitStore(VT, SrcReg, Base, Offset)) return false;
886 static ARMCC::CondCodes getComparePred(CmpInst::Predicate Pred) {
888 // Needs two compares...
889 case CmpInst::FCMP_ONE:
890 case CmpInst::FCMP_UEQ:
892 // AL is our "false" for now. The other two need more compares.
894 case CmpInst::ICMP_EQ:
895 case CmpInst::FCMP_OEQ:
897 case CmpInst::ICMP_SGT:
898 case CmpInst::FCMP_OGT:
900 case CmpInst::ICMP_SGE:
901 case CmpInst::FCMP_OGE:
903 case CmpInst::ICMP_UGT:
904 case CmpInst::FCMP_UGT:
906 case CmpInst::FCMP_OLT:
908 case CmpInst::ICMP_ULE:
909 case CmpInst::FCMP_OLE:
911 case CmpInst::FCMP_ORD:
913 case CmpInst::FCMP_UNO:
915 case CmpInst::FCMP_UGE:
917 case CmpInst::ICMP_SLT:
918 case CmpInst::FCMP_ULT:
920 case CmpInst::ICMP_SLE:
921 case CmpInst::FCMP_ULE:
923 case CmpInst::FCMP_UNE:
924 case CmpInst::ICMP_NE:
926 case CmpInst::ICMP_UGE:
928 case CmpInst::ICMP_ULT:
933 bool ARMFastISel::SelectBranch(const Instruction *I) {
934 const BranchInst *BI = cast<BranchInst>(I);
935 MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
936 MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
938 // Simple branch support.
940 // If we can, avoid recomputing the compare - redoing it could lead to wonky
942 // TODO: Factor this out.
943 if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
944 if (CI->hasOneUse() && (CI->getParent() == I->getParent())) {
946 const Type *Ty = CI->getOperand(0)->getType();
947 if (!isTypeLegal(Ty, VT))
950 bool isFloat = (Ty->isDoubleTy() || Ty->isFloatTy());
951 if (isFloat && !Subtarget->hasVFP2())
956 switch (VT.getSimpleVT().SimpleTy) {
957 default: return false;
958 // TODO: Verify compares.
960 CmpOpc = ARM::VCMPES;
961 CondReg = ARM::FPSCR;
964 CmpOpc = ARM::VCMPED;
965 CondReg = ARM::FPSCR;
968 CmpOpc = isThumb ? ARM::t2CMPrr : ARM::CMPrr;
973 // Get the compare predicate.
974 ARMCC::CondCodes ARMPred = getComparePred(CI->getPredicate());
976 // We may not handle every CC for now.
977 if (ARMPred == ARMCC::AL) return false;
979 unsigned Arg1 = getRegForValue(CI->getOperand(0));
980 if (Arg1 == 0) return false;
982 unsigned Arg2 = getRegForValue(CI->getOperand(1));
983 if (Arg2 == 0) return false;
985 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
987 .addReg(Arg1).addReg(Arg2));
989 // For floating point we need to move the result to a comparison register
990 // that we can then use for branches.
992 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
993 TII.get(ARM::FMSTAT)));
995 unsigned BrOpc = isThumb ? ARM::t2Bcc : ARM::Bcc;
996 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
997 .addMBB(TBB).addImm(ARMPred).addReg(ARM::CPSR);
998 FastEmitBranch(FBB, DL);
999 FuncInfo.MBB->addSuccessor(TBB);
1004 unsigned CmpReg = getRegForValue(BI->getCondition());
1005 if (CmpReg == 0) return false;
1007 // Re-set the flags just in case.
1008 unsigned CmpOpc = isThumb ? ARM::t2CMPri : ARM::CMPri;
1009 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
1010 .addReg(CmpReg).addImm(1));
1012 unsigned BrOpc = isThumb ? ARM::t2Bcc : ARM::Bcc;
1013 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
1014 .addMBB(TBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
1015 FastEmitBranch(FBB, DL);
1016 FuncInfo.MBB->addSuccessor(TBB);
1020 bool ARMFastISel::SelectCmp(const Instruction *I) {
1021 const CmpInst *CI = cast<CmpInst>(I);
1024 const Type *Ty = CI->getOperand(0)->getType();
1025 if (!isTypeLegal(Ty, VT))
1028 bool isFloat = (Ty->isDoubleTy() || Ty->isFloatTy());
1029 if (isFloat && !Subtarget->hasVFP2())
1034 switch (VT.getSimpleVT().SimpleTy) {
1035 default: return false;
1036 // TODO: Verify compares.
1038 CmpOpc = ARM::VCMPES;
1039 CondReg = ARM::FPSCR;
1042 CmpOpc = ARM::VCMPED;
1043 CondReg = ARM::FPSCR;
1046 CmpOpc = isThumb ? ARM::t2CMPrr : ARM::CMPrr;
1047 CondReg = ARM::CPSR;
1051 // Get the compare predicate.
1052 ARMCC::CondCodes ARMPred = getComparePred(CI->getPredicate());
1054 // We may not handle every CC for now.
1055 if (ARMPred == ARMCC::AL) return false;
1057 unsigned Arg1 = getRegForValue(CI->getOperand(0));
1058 if (Arg1 == 0) return false;
1060 unsigned Arg2 = getRegForValue(CI->getOperand(1));
1061 if (Arg2 == 0) return false;
1063 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
1064 .addReg(Arg1).addReg(Arg2));
1066 // For floating point we need to move the result to a comparison register
1067 // that we can then use for branches.
1069 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1070 TII.get(ARM::FMSTAT)));
1072 // Now set a register based on the comparison. Explicitly set the predicates
1074 unsigned MovCCOpc = isThumb ? ARM::t2MOVCCi : ARM::MOVCCi;
1075 TargetRegisterClass *RC = isThumb ? ARM::rGPRRegisterClass
1076 : ARM::GPRRegisterClass;
1077 unsigned DestReg = createResultReg(RC);
1079 = ConstantInt::get(Type::getInt32Ty(*Context), 0);
1080 unsigned ZeroReg = TargetMaterializeConstant(Zero);
1081 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), DestReg)
1082 .addReg(ZeroReg).addImm(1)
1083 .addImm(ARMPred).addReg(CondReg);
1085 UpdateValueMap(I, DestReg);
1089 bool ARMFastISel::SelectFPExt(const Instruction *I) {
1090 // Make sure we have VFP and that we're extending float to double.
1091 if (!Subtarget->hasVFP2()) return false;
1093 Value *V = I->getOperand(0);
1094 if (!I->getType()->isDoubleTy() ||
1095 !V->getType()->isFloatTy()) return false;
1097 unsigned Op = getRegForValue(V);
1098 if (Op == 0) return false;
1100 unsigned Result = createResultReg(ARM::DPRRegisterClass);
1101 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1102 TII.get(ARM::VCVTDS), Result)
1104 UpdateValueMap(I, Result);
1108 bool ARMFastISel::SelectFPTrunc(const Instruction *I) {
1109 // Make sure we have VFP and that we're truncating double to float.
1110 if (!Subtarget->hasVFP2()) return false;
1112 Value *V = I->getOperand(0);
1113 if (!(I->getType()->isFloatTy() &&
1114 V->getType()->isDoubleTy())) return false;
1116 unsigned Op = getRegForValue(V);
1117 if (Op == 0) return false;
1119 unsigned Result = createResultReg(ARM::SPRRegisterClass);
1120 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1121 TII.get(ARM::VCVTSD), Result)
1123 UpdateValueMap(I, Result);
1127 bool ARMFastISel::SelectSIToFP(const Instruction *I) {
1128 // Make sure we have VFP.
1129 if (!Subtarget->hasVFP2()) return false;
1132 const Type *Ty = I->getType();
1133 if (!isTypeLegal(Ty, DstVT))
1136 unsigned Op = getRegForValue(I->getOperand(0));
1137 if (Op == 0) return false;
1139 // The conversion routine works on fp-reg to fp-reg and the operand above
1140 // was an integer, move it to the fp registers if possible.
1141 unsigned FP = ARMMoveToFPReg(MVT::f32, Op);
1142 if (FP == 0) return false;
1145 if (Ty->isFloatTy()) Opc = ARM::VSITOS;
1146 else if (Ty->isDoubleTy()) Opc = ARM::VSITOD;
1149 unsigned ResultReg = createResultReg(TLI.getRegClassFor(DstVT));
1150 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
1153 UpdateValueMap(I, ResultReg);
1157 bool ARMFastISel::SelectFPToSI(const Instruction *I) {
1158 // Make sure we have VFP.
1159 if (!Subtarget->hasVFP2()) return false;
1162 const Type *RetTy = I->getType();
1163 if (!isTypeLegal(RetTy, DstVT))
1166 unsigned Op = getRegForValue(I->getOperand(0));
1167 if (Op == 0) return false;
1170 const Type *OpTy = I->getOperand(0)->getType();
1171 if (OpTy->isFloatTy()) Opc = ARM::VTOSIZS;
1172 else if (OpTy->isDoubleTy()) Opc = ARM::VTOSIZD;
1175 // f64->s32 or f32->s32 both need an intermediate f32 reg.
1176 unsigned ResultReg = createResultReg(TLI.getRegClassFor(MVT::f32));
1177 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
1181 // This result needs to be in an integer register, but the conversion only
1182 // takes place in fp-regs.
1183 unsigned IntReg = ARMMoveToIntReg(DstVT, ResultReg);
1184 if (IntReg == 0) return false;
1186 UpdateValueMap(I, IntReg);
1190 bool ARMFastISel::SelectSelect(const Instruction *I) {
1191 EVT VT = TLI.getValueType(I->getType(), /*HandleUnknown=*/true);
1192 if (VT == MVT::Other || !isTypeLegal(I->getType(), VT))
1195 // Things need to be register sized for register moves.
1196 if (VT.getSimpleVT().SimpleTy != MVT::i32) return false;
1197 const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
1199 unsigned CondReg = getRegForValue(I->getOperand(0));
1200 if (CondReg == 0) return false;
1201 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1202 if (Op1Reg == 0) return false;
1203 unsigned Op2Reg = getRegForValue(I->getOperand(2));
1204 if (Op2Reg == 0) return false;
1206 unsigned CmpOpc = isThumb ? ARM::t2TSTri : ARM::TSTri;
1207 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
1208 .addReg(CondReg).addImm(1));
1209 unsigned ResultReg = createResultReg(RC);
1210 unsigned MovCCOpc = isThumb ? ARM::t2MOVCCr : ARM::MOVCCr;
1211 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), ResultReg)
1212 .addReg(Op1Reg).addReg(Op2Reg)
1213 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
1214 UpdateValueMap(I, ResultReg);
1218 bool ARMFastISel::SelectSDiv(const Instruction *I) {
1220 const Type *Ty = I->getType();
1221 if (!isTypeLegal(Ty, VT))
1224 // If we have integer div support we should have selected this automagically.
1225 // In case we have a real miss go ahead and return false and we'll pick
1227 if (Subtarget->hasDivide()) return false;
1229 // Otherwise emit a libcall.
1230 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
1232 LC = RTLIB::SDIV_I8;
1233 else if (VT == MVT::i16)
1234 LC = RTLIB::SDIV_I16;
1235 else if (VT == MVT::i32)
1236 LC = RTLIB::SDIV_I32;
1237 else if (VT == MVT::i64)
1238 LC = RTLIB::SDIV_I64;
1239 else if (VT == MVT::i128)
1240 LC = RTLIB::SDIV_I128;
1241 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SDIV!");
1243 return ARMEmitLibcall(I, LC);
1246 bool ARMFastISel::SelectSRem(const Instruction *I) {
1248 const Type *Ty = I->getType();
1249 if (!isTypeLegal(Ty, VT))
1252 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
1254 LC = RTLIB::SREM_I8;
1255 else if (VT == MVT::i16)
1256 LC = RTLIB::SREM_I16;
1257 else if (VT == MVT::i32)
1258 LC = RTLIB::SREM_I32;
1259 else if (VT == MVT::i64)
1260 LC = RTLIB::SREM_I64;
1261 else if (VT == MVT::i128)
1262 LC = RTLIB::SREM_I128;
1263 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SREM!");
1265 return ARMEmitLibcall(I, LC);
1268 bool ARMFastISel::SelectBinaryOp(const Instruction *I, unsigned ISDOpcode) {
1269 EVT VT = TLI.getValueType(I->getType(), true);
1271 // We can get here in the case when we want to use NEON for our fp
1272 // operations, but can't figure out how to. Just use the vfp instructions
1274 // FIXME: It'd be nice to use NEON instructions.
1275 const Type *Ty = I->getType();
1276 bool isFloat = (Ty->isDoubleTy() || Ty->isFloatTy());
1277 if (isFloat && !Subtarget->hasVFP2())
1280 unsigned Op1 = getRegForValue(I->getOperand(0));
1281 if (Op1 == 0) return false;
1283 unsigned Op2 = getRegForValue(I->getOperand(1));
1284 if (Op2 == 0) return false;
1287 bool is64bit = VT.getSimpleVT().SimpleTy == MVT::f64 ||
1288 VT.getSimpleVT().SimpleTy == MVT::i64;
1289 switch (ISDOpcode) {
1290 default: return false;
1292 Opc = is64bit ? ARM::VADDD : ARM::VADDS;
1295 Opc = is64bit ? ARM::VSUBD : ARM::VSUBS;
1298 Opc = is64bit ? ARM::VMULD : ARM::VMULS;
1301 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
1302 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1303 TII.get(Opc), ResultReg)
1304 .addReg(Op1).addReg(Op2));
1305 UpdateValueMap(I, ResultReg);
1309 // Call Handling Code
1311 bool ARMFastISel::FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src,
1312 EVT SrcVT, unsigned &ResultReg) {
1313 unsigned RR = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), Opc,
1314 Src, /*TODO: Kill=*/false);
1323 // This is largely taken directly from CCAssignFnForNode - we don't support
1324 // varargs in FastISel so that part has been removed.
1325 // TODO: We may not support all of this.
1326 CCAssignFn *ARMFastISel::CCAssignFnForCall(CallingConv::ID CC, bool Return) {
1329 llvm_unreachable("Unsupported calling convention");
1330 case CallingConv::Fast:
1331 // Ignore fastcc. Silence compiler warnings.
1332 (void)RetFastCC_ARM_APCS;
1333 (void)FastCC_ARM_APCS;
1335 case CallingConv::C:
1336 // Use target triple & subtarget features to do actual dispatch.
1337 if (Subtarget->isAAPCS_ABI()) {
1338 if (Subtarget->hasVFP2() &&
1339 FloatABIType == FloatABI::Hard)
1340 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
1342 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
1344 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
1345 case CallingConv::ARM_AAPCS_VFP:
1346 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
1347 case CallingConv::ARM_AAPCS:
1348 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
1349 case CallingConv::ARM_APCS:
1350 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
1354 bool ARMFastISel::ProcessCallArgs(SmallVectorImpl<Value*> &Args,
1355 SmallVectorImpl<unsigned> &ArgRegs,
1356 SmallVectorImpl<EVT> &ArgVTs,
1357 SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
1358 SmallVectorImpl<unsigned> &RegArgs,
1360 unsigned &NumBytes) {
1361 SmallVector<CCValAssign, 16> ArgLocs;
1362 CCState CCInfo(CC, false, TM, ArgLocs, *Context);
1363 CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CCAssignFnForCall(CC, false));
1365 // Get a count of how many bytes are to be pushed on the stack.
1366 NumBytes = CCInfo.getNextStackOffset();
1368 // Issue CALLSEQ_START
1369 unsigned AdjStackDown = TM.getRegisterInfo()->getCallFrameSetupOpcode();
1370 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1371 TII.get(AdjStackDown))
1374 // Process the args.
1375 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1376 CCValAssign &VA = ArgLocs[i];
1377 unsigned Arg = ArgRegs[VA.getValNo()];
1378 EVT ArgVT = ArgVTs[VA.getValNo()];
1380 // We don't handle NEON parameters yet.
1381 if (VA.getLocVT().isVector() && VA.getLocVT().getSizeInBits() > 64)
1384 // Handle arg promotion, etc.
1385 switch (VA.getLocInfo()) {
1386 case CCValAssign::Full: break;
1387 case CCValAssign::SExt: {
1388 bool Emitted = FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
1390 assert(Emitted && "Failed to emit a sext!"); Emitted=Emitted;
1392 ArgVT = VA.getLocVT();
1395 case CCValAssign::ZExt: {
1396 bool Emitted = FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
1398 assert(Emitted && "Failed to emit a zext!"); Emitted=Emitted;
1400 ArgVT = VA.getLocVT();
1403 case CCValAssign::AExt: {
1404 bool Emitted = FastEmitExtend(ISD::ANY_EXTEND, VA.getLocVT(),
1407 Emitted = FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
1410 Emitted = FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
1413 assert(Emitted && "Failed to emit a aext!"); Emitted=Emitted;
1414 ArgVT = VA.getLocVT();
1417 case CCValAssign::BCvt: {
1418 unsigned BC = FastEmit_r(ArgVT.getSimpleVT(),
1419 VA.getLocVT().getSimpleVT(),
1420 ISD::BIT_CONVERT, Arg, /*TODO: Kill=*/false);
1421 assert(BC != 0 && "Failed to emit a bitcast!");
1423 ArgVT = VA.getLocVT();
1426 default: llvm_unreachable("Unknown arg promotion!");
1429 // Now copy/store arg to correct locations.
1430 if (VA.isRegLoc() && !VA.needsCustom()) {
1431 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1434 RegArgs.push_back(VA.getLocReg());
1435 } else if (VA.needsCustom()) {
1436 // TODO: We need custom lowering for vector (v2f64) args.
1437 if (VA.getLocVT() != MVT::f64) return false;
1439 CCValAssign &NextVA = ArgLocs[++i];
1441 // TODO: Only handle register args for now.
1442 if(!(VA.isRegLoc() && NextVA.isRegLoc())) return false;
1444 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1445 TII.get(ARM::VMOVRRD), VA.getLocReg())
1446 .addReg(NextVA.getLocReg(), RegState::Define)
1448 RegArgs.push_back(VA.getLocReg());
1449 RegArgs.push_back(NextVA.getLocReg());
1451 assert(VA.isMemLoc());
1452 // Need to store on the stack.
1453 unsigned Base = ARM::SP;
1454 int Offset = VA.getLocMemOffset();
1456 if (!ARMEmitStore(ArgVT, Arg, Base, Offset)) return false;
1462 bool ARMFastISel::FinishCall(EVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
1463 const Instruction *I, CallingConv::ID CC,
1464 unsigned &NumBytes) {
1465 // Issue CALLSEQ_END
1466 unsigned AdjStackUp = TM.getRegisterInfo()->getCallFrameDestroyOpcode();
1467 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1468 TII.get(AdjStackUp))
1469 .addImm(NumBytes).addImm(0));
1471 // Now the return value.
1472 if (RetVT.getSimpleVT().SimpleTy != MVT::isVoid) {
1473 SmallVector<CCValAssign, 16> RVLocs;
1474 CCState CCInfo(CC, false, TM, RVLocs, *Context);
1475 CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC, true));
1477 // Copy all of the result registers out of their specified physreg.
1478 if (RVLocs.size() == 2 && RetVT.getSimpleVT().SimpleTy == MVT::f64) {
1479 // For this move we copy into two registers and then move into the
1480 // double fp reg we want.
1481 EVT DestVT = RVLocs[0].getValVT();
1482 TargetRegisterClass* DstRC = TLI.getRegClassFor(DestVT);
1483 unsigned ResultReg = createResultReg(DstRC);
1484 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1485 TII.get(ARM::VMOVDRR), ResultReg)
1486 .addReg(RVLocs[0].getLocReg())
1487 .addReg(RVLocs[1].getLocReg()));
1489 UsedRegs.push_back(RVLocs[0].getLocReg());
1490 UsedRegs.push_back(RVLocs[1].getLocReg());
1492 // Finally update the result.
1493 UpdateValueMap(I, ResultReg);
1495 assert(RVLocs.size() == 1 &&"Can't handle non-double multi-reg retvals!");
1496 EVT CopyVT = RVLocs[0].getValVT();
1497 TargetRegisterClass* DstRC = TLI.getRegClassFor(CopyVT);
1499 unsigned ResultReg = createResultReg(DstRC);
1500 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1501 ResultReg).addReg(RVLocs[0].getLocReg());
1502 UsedRegs.push_back(RVLocs[0].getLocReg());
1504 // Finally update the result.
1505 UpdateValueMap(I, ResultReg);
1512 bool ARMFastISel::SelectRet(const Instruction *I) {
1513 const ReturnInst *Ret = cast<ReturnInst>(I);
1514 const Function &F = *I->getParent()->getParent();
1516 if (!FuncInfo.CanLowerReturn)
1522 CallingConv::ID CC = F.getCallingConv();
1523 if (Ret->getNumOperands() > 0) {
1524 SmallVector<ISD::OutputArg, 4> Outs;
1525 GetReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(),
1528 // Analyze operands of the call, assigning locations to each operand.
1529 SmallVector<CCValAssign, 16> ValLocs;
1530 CCState CCInfo(CC, F.isVarArg(), TM, ValLocs, I->getContext());
1531 CCInfo.AnalyzeReturn(Outs, CCAssignFnForCall(CC, true /* is Ret */));
1533 const Value *RV = Ret->getOperand(0);
1534 unsigned Reg = getRegForValue(RV);
1538 // Only handle a single return value for now.
1539 if (ValLocs.size() != 1)
1542 CCValAssign &VA = ValLocs[0];
1544 // Don't bother handling odd stuff for now.
1545 if (VA.getLocInfo() != CCValAssign::Full)
1547 // Only handle register returns for now.
1550 // TODO: For now, don't try to handle cases where getLocInfo()
1551 // says Full but the types don't match.
1552 if (VA.getValVT() != TLI.getValueType(RV->getType()))
1556 unsigned SrcReg = Reg + VA.getValNo();
1557 unsigned DstReg = VA.getLocReg();
1558 const TargetRegisterClass* SrcRC = MRI.getRegClass(SrcReg);
1559 // Avoid a cross-class copy. This is very unlikely.
1560 if (!SrcRC->contains(DstReg))
1562 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1563 DstReg).addReg(SrcReg);
1565 // Mark the register as live out of the function.
1566 MRI.addLiveOut(VA.getLocReg());
1569 unsigned RetOpc = isThumb ? ARM::tBX_RET : ARM::BX_RET;
1570 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1575 // A quick function that will emit a call for a named libcall in F with the
1576 // vector of passed arguments for the Instruction in I. We can assume that we
1577 // can emit a call for any libcall we can produce. This is an abridged version
1578 // of the full call infrastructure since we won't need to worry about things
1579 // like computed function pointers or strange arguments at call sites.
1580 // TODO: Try to unify this and the normal call bits for ARM, then try to unify
1582 bool ARMFastISel::ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call) {
1583 CallingConv::ID CC = TLI.getLibcallCallingConv(Call);
1585 // Handle *simple* calls for now.
1586 const Type *RetTy = I->getType();
1588 if (RetTy->isVoidTy())
1589 RetVT = MVT::isVoid;
1590 else if (!isTypeLegal(RetTy, RetVT))
1593 // For now we're using BLX etc on the assumption that we have v5t ops.
1594 if (!Subtarget->hasV5TOps()) return false;
1596 // Set up the argument vectors.
1597 SmallVector<Value*, 8> Args;
1598 SmallVector<unsigned, 8> ArgRegs;
1599 SmallVector<EVT, 8> ArgVTs;
1600 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
1601 Args.reserve(I->getNumOperands());
1602 ArgRegs.reserve(I->getNumOperands());
1603 ArgVTs.reserve(I->getNumOperands());
1604 ArgFlags.reserve(I->getNumOperands());
1605 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
1606 Value *Op = I->getOperand(i);
1607 unsigned Arg = getRegForValue(Op);
1608 if (Arg == 0) return false;
1610 const Type *ArgTy = Op->getType();
1612 if (!isTypeLegal(ArgTy, ArgVT)) return false;
1614 ISD::ArgFlagsTy Flags;
1615 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
1616 Flags.setOrigAlign(OriginalAlignment);
1619 ArgRegs.push_back(Arg);
1620 ArgVTs.push_back(ArgVT);
1621 ArgFlags.push_back(Flags);
1624 // Handle the arguments now that we've gotten them.
1625 SmallVector<unsigned, 4> RegArgs;
1627 if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
1630 // Issue the call, BLXr9 for darwin, BLX otherwise. This uses V5 ops.
1631 // TODO: Turn this into the table of arm call ops.
1632 MachineInstrBuilder MIB;
1635 CallOpc = Subtarget->isTargetDarwin() ? ARM::tBLXi_r9 : ARM::tBLXi;
1637 CallOpc = Subtarget->isTargetDarwin() ? ARM::BLr9 : ARM::BL;
1638 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CallOpc))
1639 .addExternalSymbol(TLI.getLibcallName(Call));
1641 // Add implicit physical register uses to the call.
1642 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
1643 MIB.addReg(RegArgs[i]);
1645 // Finish off the call including any return values.
1646 SmallVector<unsigned, 4> UsedRegs;
1647 if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false;
1649 // Set all unused physreg defs as dead.
1650 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
1655 bool ARMFastISel::SelectCall(const Instruction *I) {
1656 const CallInst *CI = cast<CallInst>(I);
1657 const Value *Callee = CI->getCalledValue();
1659 // Can't handle inline asm or worry about intrinsics yet.
1660 if (isa<InlineAsm>(Callee) || isa<IntrinsicInst>(CI)) return false;
1662 // Only handle global variable Callees that are direct calls.
1663 const GlobalValue *GV = dyn_cast<GlobalValue>(Callee);
1664 if (!GV || Subtarget->GVIsIndirectSymbol(GV, TM.getRelocationModel()))
1667 // Check the calling convention.
1668 ImmutableCallSite CS(CI);
1669 CallingConv::ID CC = CS.getCallingConv();
1671 // TODO: Avoid some calling conventions?
1673 // Let SDISel handle vararg functions.
1674 const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
1675 const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
1676 if (FTy->isVarArg())
1679 // Handle *simple* calls for now.
1680 const Type *RetTy = I->getType();
1682 if (RetTy->isVoidTy())
1683 RetVT = MVT::isVoid;
1684 else if (!isTypeLegal(RetTy, RetVT))
1687 // For now we're using BLX etc on the assumption that we have v5t ops.
1689 if (!Subtarget->hasV5TOps()) return false;
1691 // Set up the argument vectors.
1692 SmallVector<Value*, 8> Args;
1693 SmallVector<unsigned, 8> ArgRegs;
1694 SmallVector<EVT, 8> ArgVTs;
1695 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
1696 Args.reserve(CS.arg_size());
1697 ArgRegs.reserve(CS.arg_size());
1698 ArgVTs.reserve(CS.arg_size());
1699 ArgFlags.reserve(CS.arg_size());
1700 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
1702 unsigned Arg = getRegForValue(*i);
1706 ISD::ArgFlagsTy Flags;
1707 unsigned AttrInd = i - CS.arg_begin() + 1;
1708 if (CS.paramHasAttr(AttrInd, Attribute::SExt))
1710 if (CS.paramHasAttr(AttrInd, Attribute::ZExt))
1713 // FIXME: Only handle *easy* calls for now.
1714 if (CS.paramHasAttr(AttrInd, Attribute::InReg) ||
1715 CS.paramHasAttr(AttrInd, Attribute::StructRet) ||
1716 CS.paramHasAttr(AttrInd, Attribute::Nest) ||
1717 CS.paramHasAttr(AttrInd, Attribute::ByVal))
1720 const Type *ArgTy = (*i)->getType();
1722 if (!isTypeLegal(ArgTy, ArgVT))
1724 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
1725 Flags.setOrigAlign(OriginalAlignment);
1728 ArgRegs.push_back(Arg);
1729 ArgVTs.push_back(ArgVT);
1730 ArgFlags.push_back(Flags);
1733 // Handle the arguments now that we've gotten them.
1734 SmallVector<unsigned, 4> RegArgs;
1736 if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
1739 // Issue the call, BLXr9 for darwin, BLX otherwise. This uses V5 ops.
1740 // TODO: Turn this into the table of arm call ops.
1741 MachineInstrBuilder MIB;
1744 CallOpc = Subtarget->isTargetDarwin() ? ARM::tBLXi_r9 : ARM::tBLXi;
1746 CallOpc = Subtarget->isTargetDarwin() ? ARM::BLr9 : ARM::BL;
1747 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CallOpc))
1748 .addGlobalAddress(GV, 0, 0);
1750 // Add implicit physical register uses to the call.
1751 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
1752 MIB.addReg(RegArgs[i]);
1754 // Finish off the call including any return values.
1755 SmallVector<unsigned, 4> UsedRegs;
1756 if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false;
1758 // Set all unused physreg defs as dead.
1759 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
1765 // TODO: SoftFP support.
1766 bool ARMFastISel::TargetSelectInstruction(const Instruction *I) {
1768 switch (I->getOpcode()) {
1769 case Instruction::Load:
1770 return SelectLoad(I);
1771 case Instruction::Store:
1772 return SelectStore(I);
1773 case Instruction::Br:
1774 return SelectBranch(I);
1775 case Instruction::ICmp:
1776 case Instruction::FCmp:
1777 return SelectCmp(I);
1778 case Instruction::FPExt:
1779 return SelectFPExt(I);
1780 case Instruction::FPTrunc:
1781 return SelectFPTrunc(I);
1782 case Instruction::SIToFP:
1783 return SelectSIToFP(I);
1784 case Instruction::FPToSI:
1785 return SelectFPToSI(I);
1786 case Instruction::FAdd:
1787 return SelectBinaryOp(I, ISD::FADD);
1788 case Instruction::FSub:
1789 return SelectBinaryOp(I, ISD::FSUB);
1790 case Instruction::FMul:
1791 return SelectBinaryOp(I, ISD::FMUL);
1792 case Instruction::SDiv:
1793 return SelectSDiv(I);
1794 case Instruction::SRem:
1795 return SelectSRem(I);
1796 case Instruction::Call:
1797 return SelectCall(I);
1798 case Instruction::Select:
1799 return SelectSelect(I);
1800 case Instruction::Ret:
1801 return SelectRet(I);
1808 llvm::FastISel *ARM::createFastISel(FunctionLoweringInfo &funcInfo) {
1809 // Completely untested on non-darwin.
1810 const TargetMachine &TM = funcInfo.MF->getTarget();
1812 // Darwin and thumb1 only for now.
1813 const ARMSubtarget *Subtarget = &TM.getSubtarget<ARMSubtarget>();
1814 if (Subtarget->isTargetDarwin() && !Subtarget->isThumb1Only() &&
1815 !DisableARMFastISel)
1816 return new ARMFastISel(funcInfo);