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/MachineRegisterInfo.h"
37 #include "llvm/Support/CallSite.h"
38 #include "llvm/Support/CommandLine.h"
39 #include "llvm/Support/ErrorHandling.h"
40 #include "llvm/Support/GetElementPtrTypeIterator.h"
41 #include "llvm/Target/TargetData.h"
42 #include "llvm/Target/TargetInstrInfo.h"
43 #include "llvm/Target/TargetLowering.h"
44 #include "llvm/Target/TargetMachine.h"
45 #include "llvm/Target/TargetOptions.h"
49 EnableARMFastISel("arm-fast-isel",
50 cl::desc("Turn on experimental ARM fast-isel support"),
51 cl::init(false), cl::Hidden);
55 class ARMFastISel : public FastISel {
57 /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
58 /// make the right decision when generating code for different targets.
59 const ARMSubtarget *Subtarget;
60 const TargetMachine &TM;
61 const TargetInstrInfo &TII;
62 const TargetLowering &TLI;
65 // Convenience variables to avoid some queries.
70 explicit ARMFastISel(FunctionLoweringInfo &funcInfo)
72 TM(funcInfo.MF->getTarget()),
73 TII(*TM.getInstrInfo()),
74 TLI(*TM.getTargetLowering()) {
75 Subtarget = &TM.getSubtarget<ARMSubtarget>();
76 AFI = funcInfo.MF->getInfo<ARMFunctionInfo>();
77 isThumb = AFI->isThumbFunction();
78 Context = &funcInfo.Fn->getContext();
81 // Code from FastISel.cpp.
82 virtual unsigned FastEmitInst_(unsigned MachineInstOpcode,
83 const TargetRegisterClass *RC);
84 virtual unsigned FastEmitInst_r(unsigned MachineInstOpcode,
85 const TargetRegisterClass *RC,
86 unsigned Op0, bool Op0IsKill);
87 virtual unsigned FastEmitInst_rr(unsigned MachineInstOpcode,
88 const TargetRegisterClass *RC,
89 unsigned Op0, bool Op0IsKill,
90 unsigned Op1, bool Op1IsKill);
91 virtual unsigned FastEmitInst_ri(unsigned MachineInstOpcode,
92 const TargetRegisterClass *RC,
93 unsigned Op0, bool Op0IsKill,
95 virtual unsigned FastEmitInst_rf(unsigned MachineInstOpcode,
96 const TargetRegisterClass *RC,
97 unsigned Op0, bool Op0IsKill,
98 const ConstantFP *FPImm);
99 virtual unsigned FastEmitInst_i(unsigned MachineInstOpcode,
100 const TargetRegisterClass *RC,
102 virtual unsigned FastEmitInst_rri(unsigned MachineInstOpcode,
103 const TargetRegisterClass *RC,
104 unsigned Op0, bool Op0IsKill,
105 unsigned Op1, bool Op1IsKill,
107 virtual unsigned FastEmitInst_extractsubreg(MVT RetVT,
108 unsigned Op0, bool Op0IsKill,
111 // Backend specific FastISel code.
112 virtual bool TargetSelectInstruction(const Instruction *I);
113 virtual unsigned TargetMaterializeConstant(const Constant *C);
114 virtual unsigned TargetMaterializeAlloca(const AllocaInst *AI);
116 #include "ARMGenFastISel.inc"
118 // Instruction selection routines.
120 virtual bool SelectLoad(const Instruction *I);
121 virtual bool SelectStore(const Instruction *I);
122 virtual bool SelectBranch(const Instruction *I);
123 virtual bool SelectCmp(const Instruction *I);
124 virtual bool SelectFPExt(const Instruction *I);
125 virtual bool SelectFPTrunc(const Instruction *I);
126 virtual bool SelectBinaryOp(const Instruction *I, unsigned ISDOpcode);
127 virtual bool SelectSIToFP(const Instruction *I);
128 virtual bool SelectFPToSI(const Instruction *I);
129 virtual bool SelectSDiv(const Instruction *I);
130 virtual bool SelectCall(const Instruction *I);
134 bool isTypeLegal(const Type *Ty, EVT &VT);
135 bool isLoadTypeLegal(const Type *Ty, EVT &VT);
136 bool ARMEmitLoad(EVT VT, unsigned &ResultReg, unsigned Reg, int Offset);
137 bool ARMEmitStore(EVT VT, unsigned SrcReg, unsigned Reg, int Offset);
138 bool ARMLoadAlloca(const Instruction *I, EVT VT);
139 bool ARMStoreAlloca(const Instruction *I, unsigned SrcReg, EVT VT);
140 bool ARMComputeRegOffset(const Value *Obj, unsigned &Reg, int &Offset);
141 unsigned ARMMaterializeFP(const ConstantFP *CFP, EVT VT);
142 unsigned ARMMaterializeInt(const Constant *C, EVT VT);
143 unsigned ARMMaterializeGV(const GlobalValue *GV, EVT VT);
144 unsigned ARMMoveToFPReg(EVT VT, unsigned SrcReg);
145 unsigned ARMMoveToIntReg(EVT VT, unsigned SrcReg);
147 // Call handling routines.
149 CCAssignFn *CCAssignFnForCall(CallingConv::ID CC, bool Return);
150 bool ProcessCallArgs(SmallVectorImpl<Value*> &Args,
151 SmallVectorImpl<unsigned> &ArgRegs,
152 SmallVectorImpl<EVT> &ArgVTs,
153 SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
154 SmallVectorImpl<unsigned> &RegArgs,
157 bool FinishCall(EVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
158 const Instruction *I, CallingConv::ID CC,
160 bool ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call);
162 // OptionalDef handling routines.
164 bool DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR);
165 const MachineInstrBuilder &AddOptionalDefs(const MachineInstrBuilder &MIB);
168 } // end anonymous namespace
170 #include "ARMGenCallingConv.inc"
172 // DefinesOptionalPredicate - This is different from DefinesPredicate in that
173 // we don't care about implicit defs here, just places we'll need to add a
174 // default CCReg argument. Sets CPSR if we're setting CPSR instead of CCR.
175 bool ARMFastISel::DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR) {
176 const TargetInstrDesc &TID = MI->getDesc();
177 if (!TID.hasOptionalDef())
180 // Look to see if our OptionalDef is defining CPSR or CCR.
181 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
182 const MachineOperand &MO = MI->getOperand(i);
183 if (!MO.isReg() || !MO.isDef()) continue;
184 if (MO.getReg() == ARM::CPSR)
190 // If the machine is predicable go ahead and add the predicate operands, if
191 // it needs default CC operands add those.
192 const MachineInstrBuilder &
193 ARMFastISel::AddOptionalDefs(const MachineInstrBuilder &MIB) {
194 MachineInstr *MI = &*MIB;
196 // Do we use a predicate?
197 if (TII.isPredicable(MI))
200 // Do we optionally set a predicate? Preds is size > 0 iff the predicate
201 // defines CPSR. All other OptionalDefines in ARM are the CCR register.
203 if (DefinesOptionalPredicate(MI, &CPSR)) {
212 unsigned ARMFastISel::FastEmitInst_(unsigned MachineInstOpcode,
213 const TargetRegisterClass* RC) {
214 unsigned ResultReg = createResultReg(RC);
215 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
217 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg));
221 unsigned ARMFastISel::FastEmitInst_r(unsigned MachineInstOpcode,
222 const TargetRegisterClass *RC,
223 unsigned Op0, bool Op0IsKill) {
224 unsigned ResultReg = createResultReg(RC);
225 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
227 if (II.getNumDefs() >= 1)
228 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
229 .addReg(Op0, Op0IsKill * RegState::Kill));
231 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
232 .addReg(Op0, Op0IsKill * RegState::Kill));
233 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
234 TII.get(TargetOpcode::COPY), ResultReg)
235 .addReg(II.ImplicitDefs[0]));
240 unsigned ARMFastISel::FastEmitInst_rr(unsigned MachineInstOpcode,
241 const TargetRegisterClass *RC,
242 unsigned Op0, bool Op0IsKill,
243 unsigned Op1, bool Op1IsKill) {
244 unsigned ResultReg = createResultReg(RC);
245 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
247 if (II.getNumDefs() >= 1)
248 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
249 .addReg(Op0, Op0IsKill * RegState::Kill)
250 .addReg(Op1, Op1IsKill * RegState::Kill));
252 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
253 .addReg(Op0, Op0IsKill * RegState::Kill)
254 .addReg(Op1, Op1IsKill * RegState::Kill));
255 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
256 TII.get(TargetOpcode::COPY), ResultReg)
257 .addReg(II.ImplicitDefs[0]));
262 unsigned ARMFastISel::FastEmitInst_ri(unsigned MachineInstOpcode,
263 const TargetRegisterClass *RC,
264 unsigned Op0, bool Op0IsKill,
266 unsigned ResultReg = createResultReg(RC);
267 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
269 if (II.getNumDefs() >= 1)
270 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
271 .addReg(Op0, Op0IsKill * RegState::Kill)
274 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
275 .addReg(Op0, Op0IsKill * RegState::Kill)
277 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
278 TII.get(TargetOpcode::COPY), ResultReg)
279 .addReg(II.ImplicitDefs[0]));
284 unsigned ARMFastISel::FastEmitInst_rf(unsigned MachineInstOpcode,
285 const TargetRegisterClass *RC,
286 unsigned Op0, bool Op0IsKill,
287 const ConstantFP *FPImm) {
288 unsigned ResultReg = createResultReg(RC);
289 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
291 if (II.getNumDefs() >= 1)
292 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
293 .addReg(Op0, Op0IsKill * RegState::Kill)
296 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
297 .addReg(Op0, Op0IsKill * RegState::Kill)
299 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
300 TII.get(TargetOpcode::COPY), ResultReg)
301 .addReg(II.ImplicitDefs[0]));
306 unsigned ARMFastISel::FastEmitInst_rri(unsigned MachineInstOpcode,
307 const TargetRegisterClass *RC,
308 unsigned Op0, bool Op0IsKill,
309 unsigned Op1, bool Op1IsKill,
311 unsigned ResultReg = createResultReg(RC);
312 const TargetInstrDesc &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)
320 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
321 .addReg(Op0, Op0IsKill * RegState::Kill)
322 .addReg(Op1, Op1IsKill * RegState::Kill)
324 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
325 TII.get(TargetOpcode::COPY), ResultReg)
326 .addReg(II.ImplicitDefs[0]));
331 unsigned ARMFastISel::FastEmitInst_i(unsigned MachineInstOpcode,
332 const TargetRegisterClass *RC,
334 unsigned ResultReg = createResultReg(RC);
335 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
337 if (II.getNumDefs() >= 1)
338 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
341 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
343 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
344 TII.get(TargetOpcode::COPY), ResultReg)
345 .addReg(II.ImplicitDefs[0]));
350 unsigned ARMFastISel::FastEmitInst_extractsubreg(MVT RetVT,
351 unsigned Op0, bool Op0IsKill,
353 unsigned ResultReg = createResultReg(TLI.getRegClassFor(RetVT));
354 assert(TargetRegisterInfo::isVirtualRegister(Op0) &&
355 "Cannot yet extract from physregs");
356 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
357 DL, TII.get(TargetOpcode::COPY), ResultReg)
358 .addReg(Op0, getKillRegState(Op0IsKill), Idx));
362 // TODO: Don't worry about 64-bit now, but when this is fixed remove the
363 // checks from the various callers.
364 unsigned ARMFastISel::ARMMoveToFPReg(EVT VT, unsigned SrcReg) {
365 if (VT.getSimpleVT().SimpleTy == MVT::f64) return 0;
367 unsigned MoveReg = createResultReg(TLI.getRegClassFor(VT));
368 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
369 TII.get(ARM::VMOVRS), MoveReg)
374 unsigned ARMFastISel::ARMMoveToIntReg(EVT VT, unsigned SrcReg) {
375 if (VT.getSimpleVT().SimpleTy == MVT::i64) return 0;
377 unsigned MoveReg = createResultReg(TLI.getRegClassFor(VT));
378 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
379 TII.get(ARM::VMOVSR), MoveReg)
384 // For double width floating point we need to materialize two constants
385 // (the high and the low) into integer registers then use a move to get
386 // the combined constant into an FP reg.
387 unsigned ARMFastISel::ARMMaterializeFP(const ConstantFP *CFP, EVT VT) {
388 const APFloat Val = CFP->getValueAPF();
389 bool is64bit = VT.getSimpleVT().SimpleTy == MVT::f64;
391 // This checks to see if we can use VFP3 instructions to materialize
392 // a constant, otherwise we have to go through the constant pool.
393 if (TLI.isFPImmLegal(Val, VT)) {
394 unsigned Opc = is64bit ? ARM::FCONSTD : ARM::FCONSTS;
395 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
396 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
402 // Require VFP2 for loading fp constants.
403 if (!Subtarget->hasVFP2()) return false;
405 // MachineConstantPool wants an explicit alignment.
406 unsigned Align = TD.getPrefTypeAlignment(CFP->getType());
408 // TODO: Figure out if this is correct.
409 Align = TD.getTypeAllocSize(CFP->getType());
411 unsigned Idx = MCP.getConstantPoolIndex(cast<Constant>(CFP), Align);
412 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
413 unsigned Opc = is64bit ? ARM::VLDRD : ARM::VLDRS;
415 // The extra reg is for addrmode5.
416 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
418 .addConstantPoolIndex(Idx)
423 unsigned ARMFastISel::ARMMaterializeInt(const Constant *C, EVT VT) {
425 // For now 32-bit only.
426 if (VT.getSimpleVT().SimpleTy != MVT::i32) return false;
428 // MachineConstantPool wants an explicit alignment.
429 unsigned Align = TD.getPrefTypeAlignment(C->getType());
431 // TODO: Figure out if this is correct.
432 Align = TD.getTypeAllocSize(C->getType());
434 unsigned Idx = MCP.getConstantPoolIndex(C, Align);
435 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
438 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
439 TII.get(ARM::t2LDRpci), DestReg)
440 .addConstantPoolIndex(Idx));
442 // The extra reg and immediate are for addrmode2.
443 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
444 TII.get(ARM::LDRcp), DestReg)
445 .addConstantPoolIndex(Idx)
446 .addReg(0).addImm(0));
451 unsigned ARMFastISel::ARMMaterializeGV(const GlobalValue *GV, EVT VT) {
452 // Disable currently...
456 unsigned ARMFastISel::TargetMaterializeConstant(const Constant *C) {
457 EVT VT = TLI.getValueType(C->getType(), true);
459 // Only handle simple types.
460 if (!VT.isSimple()) return 0;
462 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
463 return ARMMaterializeFP(CFP, VT);
464 else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
465 return ARMMaterializeGV(GV, VT);
466 else if (isa<ConstantInt>(C))
467 return ARMMaterializeInt(C, VT);
472 unsigned ARMFastISel::TargetMaterializeAlloca(const AllocaInst *AI) {
473 // Don't handle dynamic allocas.
474 if (!FuncInfo.StaticAllocaMap.count(AI)) return 0;
477 if (!isTypeLegal(AI->getType(), VT)) return false;
479 DenseMap<const AllocaInst*, int>::iterator SI =
480 FuncInfo.StaticAllocaMap.find(AI);
482 // This will get lowered later into the correct offsets and registers
483 // via rewriteXFrameIndex.
484 if (SI != FuncInfo.StaticAllocaMap.end()) {
485 TargetRegisterClass* RC = TLI.getRegClassFor(VT);
486 unsigned ResultReg = createResultReg(RC);
487 unsigned Opc = isThumb ? ARM::t2ADDri : ARM::ADDri;
488 AddOptionalDefs(BuildMI(*FuncInfo.MBB, *FuncInfo.InsertPt, DL,
489 TII.get(Opc), ResultReg)
490 .addFrameIndex(SI->second)
498 bool ARMFastISel::isTypeLegal(const Type *Ty, EVT &VT) {
499 VT = TLI.getValueType(Ty, true);
501 // Only handle simple types.
502 if (VT == MVT::Other || !VT.isSimple()) return false;
504 // Handle all legal types, i.e. a register that will directly hold this
506 return TLI.isTypeLegal(VT);
509 bool ARMFastISel::isLoadTypeLegal(const Type *Ty, EVT &VT) {
510 if (isTypeLegal(Ty, VT)) return true;
512 // If this is a type than can be sign or zero-extended to a basic operation
513 // go ahead and accept it now.
514 if (VT == MVT::i8 || VT == MVT::i16)
520 // Computes the Reg+Offset to get to an object.
521 bool ARMFastISel::ARMComputeRegOffset(const Value *Obj, unsigned &Reg,
523 // Some boilerplate from the X86 FastISel.
524 const User *U = NULL;
525 unsigned Opcode = Instruction::UserOp1;
526 if (const Instruction *I = dyn_cast<Instruction>(Obj)) {
527 // Don't walk into other basic blocks; it's possible we haven't
528 // visited them yet, so the instructions may not yet be assigned
529 // virtual registers.
530 if (FuncInfo.MBBMap[I->getParent()] != FuncInfo.MBB)
532 Opcode = I->getOpcode();
534 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Obj)) {
535 Opcode = C->getOpcode();
539 if (const PointerType *Ty = dyn_cast<PointerType>(Obj->getType()))
540 if (Ty->getAddressSpace() > 255)
541 // Fast instruction selection doesn't support the special
548 case Instruction::Alloca: {
549 assert(false && "Alloca should have been handled earlier!");
554 // FIXME: Handle global variables.
555 if (const GlobalValue *GV = dyn_cast<GlobalValue>(Obj)) {
560 // Try to get this in a register if nothing else has worked.
561 Reg = getRegForValue(Obj);
562 if (Reg == 0) return false;
564 // Since the offset may be too large for the load instruction
565 // get the reg+offset into a register.
566 // TODO: Verify the additions work, otherwise we'll need to add the
567 // offset instead of 0 to the instructions and do all sorts of operand
569 // TODO: Optimize this somewhat.
571 ARMCC::CondCodes Pred = ARMCC::AL;
572 unsigned PredReg = 0;
575 emitARMRegPlusImmediate(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
576 Reg, Reg, Offset, Pred, PredReg,
577 static_cast<const ARMBaseInstrInfo&>(TII));
579 assert(AFI->isThumb2Function());
580 emitT2RegPlusImmediate(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
581 Reg, Reg, Offset, Pred, PredReg,
582 static_cast<const ARMBaseInstrInfo&>(TII));
588 bool ARMFastISel::ARMLoadAlloca(const Instruction *I, EVT VT) {
589 Value *Op0 = I->getOperand(0);
591 // Verify it's an alloca.
592 if (const AllocaInst *AI = dyn_cast<AllocaInst>(Op0)) {
593 DenseMap<const AllocaInst*, int>::iterator SI =
594 FuncInfo.StaticAllocaMap.find(AI);
596 if (SI != FuncInfo.StaticAllocaMap.end()) {
597 TargetRegisterClass* RC = TLI.getRegClassFor(VT);
598 unsigned ResultReg = createResultReg(RC);
599 TII.loadRegFromStackSlot(*FuncInfo.MBB, *FuncInfo.InsertPt,
600 ResultReg, SI->second, RC,
601 TM.getRegisterInfo());
602 UpdateValueMap(I, ResultReg);
609 bool ARMFastISel::ARMEmitLoad(EVT VT, unsigned &ResultReg,
610 unsigned Reg, int Offset) {
612 assert(VT.isSimple() && "Non-simple types are invalid here!");
614 bool isFloat = false;
615 switch (VT.getSimpleVT().SimpleTy) {
617 // This is mostly going to be Neon/vector support.
620 Opc = isThumb ? ARM::tLDRH : ARM::LDRH;
624 Opc = isThumb ? ARM::tLDRB : ARM::LDRB;
628 Opc = isThumb ? ARM::tLDR : ARM::LDR;
640 ResultReg = createResultReg(TLI.getRegClassFor(VT));
642 // TODO: Fix the Addressing modes so that these can share some code.
643 // Since this is a Thumb1 load this will work in Thumb1 or 2 mode.
644 // The thumb addressing mode has operands swapped from the arm addressing
645 // mode, the floating point one only has two operands.
647 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
648 TII.get(Opc), ResultReg)
649 .addReg(Reg).addImm(Offset));
651 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
652 TII.get(Opc), ResultReg)
653 .addReg(Reg).addImm(Offset).addReg(0));
655 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
656 TII.get(Opc), ResultReg)
657 .addReg(Reg).addReg(0).addImm(Offset));
661 bool ARMFastISel::SelectLoad(const Instruction *I) {
662 // Verify we have a legal type before going any further.
664 if (!isLoadTypeLegal(I->getType(), VT))
667 // If we're an alloca we know we have a frame index and can emit the load
668 // directly in short order.
669 if (ARMLoadAlloca(I, VT))
672 // Our register and offset with innocuous defaults.
676 // See if we can handle this as Reg + Offset
677 if (!ARMComputeRegOffset(I->getOperand(0), Reg, Offset))
681 if (!ARMEmitLoad(VT, ResultReg, Reg, Offset /* 0 */)) return false;
683 UpdateValueMap(I, ResultReg);
687 bool ARMFastISel::ARMStoreAlloca(const Instruction *I, unsigned SrcReg, EVT VT){
688 Value *Op1 = I->getOperand(1);
690 // Verify it's an alloca.
691 if (const AllocaInst *AI = dyn_cast<AllocaInst>(Op1)) {
692 DenseMap<const AllocaInst*, int>::iterator SI =
693 FuncInfo.StaticAllocaMap.find(AI);
695 if (SI != FuncInfo.StaticAllocaMap.end()) {
696 TargetRegisterClass* RC = TLI.getRegClassFor(VT);
697 assert(SrcReg != 0 && "Nothing to store!");
698 TII.storeRegToStackSlot(*FuncInfo.MBB, *FuncInfo.InsertPt,
699 SrcReg, true /*isKill*/, SI->second, RC,
700 TM.getRegisterInfo());
707 bool ARMFastISel::ARMEmitStore(EVT VT, unsigned SrcReg,
708 unsigned DstReg, int Offset) {
710 bool isFloat = false;
711 switch (VT.getSimpleVT().SimpleTy) {
712 default: return false;
714 case MVT::i8: StrOpc = isThumb ? ARM::tSTRB : ARM::STRB; break;
715 case MVT::i16: StrOpc = isThumb ? ARM::tSTRH : ARM::STRH; break;
716 case MVT::i32: StrOpc = isThumb ? ARM::tSTR : ARM::STR; break;
718 if (!Subtarget->hasVFP2()) return false;
723 if (!Subtarget->hasVFP2()) return false;
729 // The thumb addressing mode has operands swapped from the arm addressing
730 // mode, the floating point one only has two operands.
732 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
734 .addReg(SrcReg).addReg(DstReg).addImm(Offset));
736 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
738 .addReg(SrcReg).addReg(DstReg).addImm(Offset).addReg(0));
741 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
743 .addReg(SrcReg).addReg(DstReg).addReg(0).addImm(Offset));
748 bool ARMFastISel::SelectStore(const Instruction *I) {
749 Value *Op0 = I->getOperand(0);
752 // Yay type legalization
754 if (!isLoadTypeLegal(I->getOperand(0)->getType(), VT))
757 // Get the value to be stored into a register.
758 SrcReg = getRegForValue(Op0);
762 // If we're an alloca we know we have a frame index and can emit the store
764 if (ARMStoreAlloca(I, SrcReg, VT))
767 // Our register and offset with innocuous defaults.
771 // See if we can handle this as Reg + Offset
772 if (!ARMComputeRegOffset(I->getOperand(1), Reg, Offset))
775 if (!ARMEmitStore(VT, SrcReg, Reg, Offset /* 0 */)) return false;
780 static ARMCC::CondCodes getComparePred(CmpInst::Predicate Pred) {
782 // Needs two compares...
783 case CmpInst::FCMP_ONE:
784 case CmpInst::FCMP_UEQ:
786 assert(false && "Unhandled CmpInst::Predicate!");
788 case CmpInst::ICMP_EQ:
789 case CmpInst::FCMP_OEQ:
791 case CmpInst::ICMP_SGT:
792 case CmpInst::FCMP_OGT:
794 case CmpInst::ICMP_SGE:
795 case CmpInst::FCMP_OGE:
797 case CmpInst::ICMP_UGT:
798 case CmpInst::FCMP_UGT:
800 case CmpInst::FCMP_OLT:
802 case CmpInst::ICMP_ULE:
803 case CmpInst::FCMP_OLE:
805 case CmpInst::FCMP_ORD:
807 case CmpInst::FCMP_UNO:
809 case CmpInst::FCMP_UGE:
811 case CmpInst::ICMP_SLT:
812 case CmpInst::FCMP_ULT:
814 case CmpInst::ICMP_SLE:
815 case CmpInst::FCMP_ULE:
817 case CmpInst::FCMP_UNE:
818 case CmpInst::ICMP_NE:
820 case CmpInst::ICMP_UGE:
822 case CmpInst::ICMP_ULT:
827 bool ARMFastISel::SelectBranch(const Instruction *I) {
828 const BranchInst *BI = cast<BranchInst>(I);
829 MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
830 MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
832 // Simple branch support.
833 // TODO: Try to avoid the re-computation in some places.
834 unsigned CondReg = getRegForValue(BI->getCondition());
835 if (CondReg == 0) return false;
837 // Re-set the flags just in case.
838 unsigned CmpOpc = isThumb ? ARM::t2CMPri : ARM::CMPri;
839 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
840 .addReg(CondReg).addImm(1));
842 unsigned BrOpc = isThumb ? ARM::t2Bcc : ARM::Bcc;
843 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
844 .addMBB(TBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
845 FastEmitBranch(FBB, DL);
846 FuncInfo.MBB->addSuccessor(TBB);
850 bool ARMFastISel::SelectCmp(const Instruction *I) {
851 const CmpInst *CI = cast<CmpInst>(I);
854 const Type *Ty = CI->getOperand(0)->getType();
855 if (!isTypeLegal(Ty, VT))
858 bool isFloat = (Ty->isDoubleTy() || Ty->isFloatTy());
859 if (isFloat && !Subtarget->hasVFP2())
864 switch (VT.getSimpleVT().SimpleTy) {
865 default: return false;
866 // TODO: Verify compares.
868 CmpOpc = ARM::VCMPES;
869 CondReg = ARM::FPSCR;
872 CmpOpc = ARM::VCMPED;
873 CondReg = ARM::FPSCR;
876 CmpOpc = isThumb ? ARM::t2CMPrr : ARM::CMPrr;
881 // Get the compare predicate.
882 ARMCC::CondCodes ARMPred = getComparePred(CI->getPredicate());
884 // We may not handle every CC for now.
885 if (ARMPred == ARMCC::AL) return false;
887 unsigned Arg1 = getRegForValue(CI->getOperand(0));
888 if (Arg1 == 0) return false;
890 unsigned Arg2 = getRegForValue(CI->getOperand(1));
891 if (Arg2 == 0) return false;
893 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
894 .addReg(Arg1).addReg(Arg2));
896 // For floating point we need to move the result to a comparison register
897 // that we can then use for branches.
899 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
900 TII.get(ARM::FMSTAT)));
902 // Now set a register based on the comparison. Explicitly set the predicates
904 unsigned MovCCOpc = isThumb ? ARM::tMOVCCi : ARM::MOVCCi;
905 unsigned DestReg = createResultReg(ARM::GPRRegisterClass);
907 = ConstantInt::get(Type::getInt32Ty(*Context), 0);
908 unsigned ZeroReg = TargetMaterializeConstant(Zero);
909 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), DestReg)
910 .addReg(ZeroReg).addImm(1)
911 .addImm(ARMPred).addReg(CondReg);
913 UpdateValueMap(I, DestReg);
917 bool ARMFastISel::SelectFPExt(const Instruction *I) {
918 // Make sure we have VFP and that we're extending float to double.
919 if (!Subtarget->hasVFP2()) return false;
921 Value *V = I->getOperand(0);
922 if (!I->getType()->isDoubleTy() ||
923 !V->getType()->isFloatTy()) return false;
925 unsigned Op = getRegForValue(V);
926 if (Op == 0) return false;
928 unsigned Result = createResultReg(ARM::DPRRegisterClass);
929 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
930 TII.get(ARM::VCVTDS), Result)
932 UpdateValueMap(I, Result);
936 bool ARMFastISel::SelectFPTrunc(const Instruction *I) {
937 // Make sure we have VFP and that we're truncating double to float.
938 if (!Subtarget->hasVFP2()) return false;
940 Value *V = I->getOperand(0);
941 if (!I->getType()->isFloatTy() ||
942 !V->getType()->isDoubleTy()) return false;
944 unsigned Op = getRegForValue(V);
945 if (Op == 0) return false;
947 unsigned Result = createResultReg(ARM::SPRRegisterClass);
948 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
949 TII.get(ARM::VCVTSD), Result)
951 UpdateValueMap(I, Result);
955 bool ARMFastISel::SelectSIToFP(const Instruction *I) {
956 // Make sure we have VFP.
957 if (!Subtarget->hasVFP2()) return false;
960 const Type *Ty = I->getType();
961 if (!isTypeLegal(Ty, DstVT))
964 unsigned Op = getRegForValue(I->getOperand(0));
965 if (Op == 0) return false;
967 // The conversion routine works on fp-reg to fp-reg and the operand above
968 // was an integer, move it to the fp registers if possible.
969 unsigned FP = ARMMoveToFPReg(DstVT, Op);
970 if (FP == 0) return false;
973 if (Ty->isFloatTy()) Opc = ARM::VSITOS;
974 else if (Ty->isDoubleTy()) Opc = ARM::VSITOD;
977 unsigned ResultReg = createResultReg(TLI.getRegClassFor(DstVT));
978 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
981 UpdateValueMap(I, ResultReg);
985 bool ARMFastISel::SelectFPToSI(const Instruction *I) {
986 // Make sure we have VFP.
987 if (!Subtarget->hasVFP2()) return false;
990 const Type *RetTy = I->getType();
991 if (!isTypeLegal(RetTy, DstVT))
994 unsigned Op = getRegForValue(I->getOperand(0));
995 if (Op == 0) return false;
998 const Type *OpTy = I->getOperand(0)->getType();
999 if (OpTy->isFloatTy()) Opc = ARM::VTOSIZS;
1000 else if (OpTy->isDoubleTy()) Opc = ARM::VTOSIZD;
1002 EVT OpVT = TLI.getValueType(OpTy, true);
1004 unsigned ResultReg = createResultReg(TLI.getRegClassFor(OpVT));
1005 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
1009 // This result needs to be in an integer register, but the conversion only
1010 // takes place in fp-regs.
1011 unsigned IntReg = ARMMoveToIntReg(DstVT, ResultReg);
1012 if (IntReg == 0) return false;
1014 UpdateValueMap(I, IntReg);
1018 bool ARMFastISel::SelectSDiv(const Instruction *I) {
1020 const Type *Ty = I->getType();
1021 if (!isTypeLegal(Ty, VT))
1024 // If we have integer div support we should have selected this automagically.
1025 // In case we have a real miss go ahead and return false and we'll pick
1027 if (Subtarget->hasDivide()) return false;
1029 // Otherwise emit a libcall.
1030 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
1032 LC = RTLIB::SDIV_I16;
1033 else if (VT == MVT::i32)
1034 LC = RTLIB::SDIV_I32;
1035 else if (VT == MVT::i64)
1036 LC = RTLIB::SDIV_I64;
1037 else if (VT == MVT::i128)
1038 LC = RTLIB::SDIV_I128;
1039 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SDIV!");
1041 return ARMEmitLibcall(I, LC);
1044 bool ARMFastISel::SelectBinaryOp(const Instruction *I, unsigned ISDOpcode) {
1045 EVT VT = TLI.getValueType(I->getType(), true);
1047 // We can get here in the case when we want to use NEON for our fp
1048 // operations, but can't figure out how to. Just use the vfp instructions
1050 // FIXME: It'd be nice to use NEON instructions.
1051 const Type *Ty = I->getType();
1052 bool isFloat = (Ty->isDoubleTy() || Ty->isFloatTy());
1053 if (isFloat && !Subtarget->hasVFP2())
1056 unsigned Op1 = getRegForValue(I->getOperand(0));
1057 if (Op1 == 0) return false;
1059 unsigned Op2 = getRegForValue(I->getOperand(1));
1060 if (Op2 == 0) return false;
1063 bool is64bit = VT.getSimpleVT().SimpleTy == MVT::f64 ||
1064 VT.getSimpleVT().SimpleTy == MVT::i64;
1065 switch (ISDOpcode) {
1066 default: return false;
1068 Opc = is64bit ? ARM::VADDD : ARM::VADDS;
1071 Opc = is64bit ? ARM::VSUBD : ARM::VSUBS;
1074 Opc = is64bit ? ARM::VMULD : ARM::VMULS;
1077 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
1078 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1079 TII.get(Opc), ResultReg)
1080 .addReg(Op1).addReg(Op2));
1081 UpdateValueMap(I, ResultReg);
1085 // Call Handling Code
1087 // This is largely taken directly from CCAssignFnForNode - we don't support
1088 // varargs in FastISel so that part has been removed.
1089 // TODO: We may not support all of this.
1090 CCAssignFn *ARMFastISel::CCAssignFnForCall(CallingConv::ID CC, bool Return) {
1093 llvm_unreachable("Unsupported calling convention");
1094 case CallingConv::C:
1095 case CallingConv::Fast:
1096 // Use target triple & subtarget features to do actual dispatch.
1097 if (Subtarget->isAAPCS_ABI()) {
1098 if (Subtarget->hasVFP2() &&
1099 FloatABIType == FloatABI::Hard)
1100 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
1102 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
1104 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
1105 case CallingConv::ARM_AAPCS_VFP:
1106 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
1107 case CallingConv::ARM_AAPCS:
1108 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
1109 case CallingConv::ARM_APCS:
1110 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
1114 bool ARMFastISel::ProcessCallArgs(SmallVectorImpl<Value*> &Args,
1115 SmallVectorImpl<unsigned> &ArgRegs,
1116 SmallVectorImpl<EVT> &ArgVTs,
1117 SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
1118 SmallVectorImpl<unsigned> &RegArgs,
1120 unsigned &NumBytes) {
1121 SmallVector<CCValAssign, 16> ArgLocs;
1122 CCState CCInfo(CC, false, TM, ArgLocs, *Context);
1123 CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CCAssignFnForCall(CC, false));
1125 // Get a count of how many bytes are to be pushed on the stack.
1126 NumBytes = CCInfo.getNextStackOffset();
1128 // Issue CALLSEQ_START
1129 unsigned AdjStackDown = TM.getRegisterInfo()->getCallFrameSetupOpcode();
1130 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(AdjStackDown))
1133 // Process the args.
1134 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1135 CCValAssign &VA = ArgLocs[i];
1136 unsigned Arg = ArgRegs[VA.getValNo()];
1137 EVT ArgVT = ArgVTs[VA.getValNo()];
1139 // Handle arg promotion, etc.
1140 switch (VA.getLocInfo()) {
1141 case CCValAssign::Full: break;
1143 assert(false && "Handle arg promotion.");
1147 // Now copy/store arg to correct locations.
1148 if (VA.isRegLoc()) {
1149 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1152 RegArgs.push_back(VA.getLocReg());
1162 bool ARMFastISel::FinishCall(EVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
1163 const Instruction *I, CallingConv::ID CC,
1164 unsigned &NumBytes) {
1165 // Issue CALLSEQ_END
1166 unsigned AdjStackUp = TM.getRegisterInfo()->getCallFrameDestroyOpcode();
1167 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(AdjStackUp))
1168 .addImm(NumBytes).addImm(0);
1170 // Now the return value.
1171 if (RetVT.getSimpleVT().SimpleTy != MVT::isVoid) {
1172 SmallVector<CCValAssign, 16> RVLocs;
1173 CCState CCInfo(CC, false, TM, RVLocs, *Context);
1174 CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC, true));
1176 // Copy all of the result registers out of their specified physreg.
1177 if (RVLocs.size() == 2 && RetVT.getSimpleVT().SimpleTy == MVT::f64) {
1178 // For this move we copy into two registers and then move into the
1179 // double fp reg we want.
1180 // TODO: Are the copies necessary?
1181 TargetRegisterClass *CopyRC = TLI.getRegClassFor(MVT::i32);
1182 unsigned Copy1 = createResultReg(CopyRC);
1183 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1184 Copy1).addReg(RVLocs[0].getLocReg());
1185 UsedRegs.push_back(RVLocs[0].getLocReg());
1187 unsigned Copy2 = createResultReg(CopyRC);
1188 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1189 Copy2).addReg(RVLocs[1].getLocReg());
1190 UsedRegs.push_back(RVLocs[1].getLocReg());
1192 EVT DestVT = RVLocs[0].getValVT();
1193 TargetRegisterClass* DstRC = TLI.getRegClassFor(DestVT);
1194 unsigned ResultReg = createResultReg(DstRC);
1195 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1196 TII.get(ARM::VMOVDRR), ResultReg)
1197 .addReg(Copy1).addReg(Copy2));
1199 // Finally update the result.
1200 UpdateValueMap(I, ResultReg);
1202 assert(RVLocs.size() == 1 && "Can't handle non-double multi-reg retvals!");
1203 EVT CopyVT = RVLocs[0].getValVT();
1204 TargetRegisterClass* DstRC = TLI.getRegClassFor(CopyVT);
1206 unsigned ResultReg = createResultReg(DstRC);
1207 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1208 ResultReg).addReg(RVLocs[0].getLocReg());
1209 UsedRegs.push_back(RVLocs[0].getLocReg());
1211 // Finally update the result.
1212 UpdateValueMap(I, ResultReg);
1219 // A quick function that will emit a call for a named libcall in F with the
1220 // vector of passed arguments for the Instruction in I. We can assume that we
1221 // can emit a call for any libcall we can produce. This is an abridged version
1222 // of the full call infrastructure since we won't need to worry about things
1223 // like computed function pointers or strange arguments at call sites.
1224 // TODO: Try to unify this and the normal call bits for ARM, then try to unify
1226 bool ARMFastISel::ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call) {
1227 CallingConv::ID CC = TLI.getLibcallCallingConv(Call);
1229 // Handle *simple* calls for now.
1230 const Type *RetTy = I->getType();
1232 if (RetTy->isVoidTy())
1233 RetVT = MVT::isVoid;
1234 else if (!isTypeLegal(RetTy, RetVT))
1237 // For now we're using BLX etc on the assumption that we have v5t ops.
1238 if (!Subtarget->hasV5TOps()) return false;
1240 // Set up the argument vectors.
1241 SmallVector<Value*, 8> Args;
1242 SmallVector<unsigned, 8> ArgRegs;
1243 SmallVector<EVT, 8> ArgVTs;
1244 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
1245 Args.reserve(I->getNumOperands());
1246 ArgRegs.reserve(I->getNumOperands());
1247 ArgVTs.reserve(I->getNumOperands());
1248 ArgFlags.reserve(I->getNumOperands());
1249 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
1250 Value *Op = I->getOperand(i);
1251 unsigned Arg = getRegForValue(Op);
1252 if (Arg == 0) return false;
1254 const Type *ArgTy = Op->getType();
1256 if (!isTypeLegal(ArgTy, ArgVT)) return false;
1258 ISD::ArgFlagsTy Flags;
1259 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
1260 Flags.setOrigAlign(OriginalAlignment);
1263 ArgRegs.push_back(Arg);
1264 ArgVTs.push_back(ArgVT);
1265 ArgFlags.push_back(Flags);
1268 // Handle the arguments now that we've gotten them.
1269 SmallVector<unsigned, 4> RegArgs;
1271 if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
1274 // Issue the call, BLXr9 for darwin, BLX otherwise. This uses V5 ops.
1275 // TODO: Turn this into the table of arm call ops.
1276 MachineInstrBuilder MIB;
1279 CallOpc = Subtarget->isTargetDarwin() ? ARM::tBLXi_r9 : ARM::tBLXi;
1281 CallOpc = Subtarget->isTargetDarwin() ? ARM::BLr9 : ARM::BL;
1282 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CallOpc))
1283 .addExternalSymbol(TLI.getLibcallName(Call));
1285 // Add implicit physical register uses to the call.
1286 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
1287 MIB.addReg(RegArgs[i]);
1289 // Finish off the call including any return values.
1290 SmallVector<unsigned, 4> UsedRegs;
1291 if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false;
1293 // Set all unused physreg defs as dead.
1294 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
1299 bool ARMFastISel::SelectCall(const Instruction *I) {
1300 const CallInst *CI = cast<CallInst>(I);
1301 const Value *Callee = CI->getCalledValue();
1303 // Can't handle inline asm or worry about intrinsics yet.
1304 if (isa<InlineAsm>(Callee) || isa<IntrinsicInst>(CI)) return false;
1306 // Only handle global variable Callees that are direct calls.
1307 const GlobalValue *GV = dyn_cast<GlobalValue>(Callee);
1308 if (!GV || Subtarget->GVIsIndirectSymbol(GV, TM.getRelocationModel()))
1311 // Check the calling convention.
1312 ImmutableCallSite CS(CI);
1313 CallingConv::ID CC = CS.getCallingConv();
1314 // TODO: Avoid some calling conventions?
1315 if (CC != CallingConv::C) {
1316 errs() << "Can't handle calling convention: " << CC << "\n";
1320 // Let SDISel handle vararg functions.
1321 const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
1322 const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
1323 if (FTy->isVarArg())
1326 // Handle *simple* calls for now.
1327 const Type *RetTy = I->getType();
1329 if (RetTy->isVoidTy())
1330 RetVT = MVT::isVoid;
1331 else if (!isTypeLegal(RetTy, RetVT))
1334 // For now we're using BLX etc on the assumption that we have v5t ops.
1336 if (!Subtarget->hasV5TOps()) return false;
1338 // Set up the argument vectors.
1339 SmallVector<Value*, 8> Args;
1340 SmallVector<unsigned, 8> ArgRegs;
1341 SmallVector<EVT, 8> ArgVTs;
1342 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
1343 Args.reserve(CS.arg_size());
1344 ArgRegs.reserve(CS.arg_size());
1345 ArgVTs.reserve(CS.arg_size());
1346 ArgFlags.reserve(CS.arg_size());
1347 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
1349 unsigned Arg = getRegForValue(*i);
1353 ISD::ArgFlagsTy Flags;
1354 unsigned AttrInd = i - CS.arg_begin() + 1;
1355 if (CS.paramHasAttr(AttrInd, Attribute::SExt))
1357 if (CS.paramHasAttr(AttrInd, Attribute::ZExt))
1360 // FIXME: Only handle *easy* calls for now.
1361 if (CS.paramHasAttr(AttrInd, Attribute::InReg) ||
1362 CS.paramHasAttr(AttrInd, Attribute::StructRet) ||
1363 CS.paramHasAttr(AttrInd, Attribute::Nest) ||
1364 CS.paramHasAttr(AttrInd, Attribute::ByVal))
1367 const Type *ArgTy = (*i)->getType();
1369 if (!isTypeLegal(ArgTy, ArgVT))
1371 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
1372 Flags.setOrigAlign(OriginalAlignment);
1375 ArgRegs.push_back(Arg);
1376 ArgVTs.push_back(ArgVT);
1377 ArgFlags.push_back(Flags);
1380 // Handle the arguments now that we've gotten them.
1381 SmallVector<unsigned, 4> RegArgs;
1383 if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
1386 // Issue the call, BLXr9 for darwin, BLX otherwise. This uses V5 ops.
1387 // TODO: Turn this into the table of arm call ops.
1388 MachineInstrBuilder MIB;
1391 CallOpc = Subtarget->isTargetDarwin() ? ARM::tBLXi_r9 : ARM::tBLXi;
1393 CallOpc = Subtarget->isTargetDarwin() ? ARM::BLr9 : ARM::BL;
1394 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CallOpc))
1395 .addGlobalAddress(GV, 0, 0);
1397 // Add implicit physical register uses to the call.
1398 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
1399 MIB.addReg(RegArgs[i]);
1401 // Finish off the call including any return values.
1402 SmallVector<unsigned, 4> UsedRegs;
1403 if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false;
1405 // Set all unused physreg defs as dead.
1406 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
1412 // TODO: SoftFP support.
1413 bool ARMFastISel::TargetSelectInstruction(const Instruction *I) {
1414 // No Thumb-1 for now.
1415 if (isThumb && !AFI->isThumb2Function()) return false;
1417 switch (I->getOpcode()) {
1418 case Instruction::Load:
1419 return SelectLoad(I);
1420 case Instruction::Store:
1421 return SelectStore(I);
1422 case Instruction::Br:
1423 return SelectBranch(I);
1424 case Instruction::ICmp:
1425 case Instruction::FCmp:
1426 return SelectCmp(I);
1427 case Instruction::FPExt:
1428 return SelectFPExt(I);
1429 case Instruction::FPTrunc:
1430 return SelectFPTrunc(I);
1431 case Instruction::SIToFP:
1432 return SelectSIToFP(I);
1433 case Instruction::FPToSI:
1434 return SelectFPToSI(I);
1435 case Instruction::FAdd:
1436 return SelectBinaryOp(I, ISD::FADD);
1437 case Instruction::FSub:
1438 return SelectBinaryOp(I, ISD::FSUB);
1439 case Instruction::FMul:
1440 return SelectBinaryOp(I, ISD::FMUL);
1441 case Instruction::SDiv:
1442 return SelectSDiv(I);
1443 case Instruction::Call:
1444 return SelectCall(I);
1451 llvm::FastISel *ARM::createFastISel(FunctionLoweringInfo &funcInfo) {
1452 if (EnableARMFastISel) return new ARMFastISel(funcInfo);