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 // All possible address modes, plus some.
58 typedef struct Address {
73 // Innocuous defaults for our address.
75 : BaseType(RegBase), Offset(0), Scale(0), PlusReg(0) {
80 class ARMFastISel : public FastISel {
82 /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
83 /// make the right decision when generating code for different targets.
84 const ARMSubtarget *Subtarget;
85 const TargetMachine &TM;
86 const TargetInstrInfo &TII;
87 const TargetLowering &TLI;
90 // Convenience variables to avoid some queries.
95 explicit ARMFastISel(FunctionLoweringInfo &funcInfo)
97 TM(funcInfo.MF->getTarget()),
98 TII(*TM.getInstrInfo()),
99 TLI(*TM.getTargetLowering()) {
100 Subtarget = &TM.getSubtarget<ARMSubtarget>();
101 AFI = funcInfo.MF->getInfo<ARMFunctionInfo>();
102 isThumb = AFI->isThumbFunction();
103 Context = &funcInfo.Fn->getContext();
106 // Code from FastISel.cpp.
107 virtual unsigned FastEmitInst_(unsigned MachineInstOpcode,
108 const TargetRegisterClass *RC);
109 virtual unsigned FastEmitInst_r(unsigned MachineInstOpcode,
110 const TargetRegisterClass *RC,
111 unsigned Op0, bool Op0IsKill);
112 virtual unsigned FastEmitInst_rr(unsigned MachineInstOpcode,
113 const TargetRegisterClass *RC,
114 unsigned Op0, bool Op0IsKill,
115 unsigned Op1, bool Op1IsKill);
116 virtual unsigned FastEmitInst_ri(unsigned MachineInstOpcode,
117 const TargetRegisterClass *RC,
118 unsigned Op0, bool Op0IsKill,
120 virtual unsigned FastEmitInst_rf(unsigned MachineInstOpcode,
121 const TargetRegisterClass *RC,
122 unsigned Op0, bool Op0IsKill,
123 const ConstantFP *FPImm);
124 virtual unsigned FastEmitInst_i(unsigned MachineInstOpcode,
125 const TargetRegisterClass *RC,
127 virtual unsigned FastEmitInst_rri(unsigned MachineInstOpcode,
128 const TargetRegisterClass *RC,
129 unsigned Op0, bool Op0IsKill,
130 unsigned Op1, bool Op1IsKill,
132 virtual unsigned FastEmitInst_extractsubreg(MVT RetVT,
133 unsigned Op0, bool Op0IsKill,
136 // Backend specific FastISel code.
137 virtual bool TargetSelectInstruction(const Instruction *I);
138 virtual unsigned TargetMaterializeConstant(const Constant *C);
139 virtual unsigned TargetMaterializeAlloca(const AllocaInst *AI);
141 #include "ARMGenFastISel.inc"
143 // Instruction selection routines.
145 bool SelectLoad(const Instruction *I);
146 bool SelectStore(const Instruction *I);
147 bool SelectBranch(const Instruction *I);
148 bool SelectCmp(const Instruction *I);
149 bool SelectFPExt(const Instruction *I);
150 bool SelectFPTrunc(const Instruction *I);
151 bool SelectBinaryOp(const Instruction *I, unsigned ISDOpcode);
152 bool SelectSIToFP(const Instruction *I);
153 bool SelectFPToSI(const Instruction *I);
154 bool SelectSDiv(const Instruction *I);
155 bool SelectSRem(const Instruction *I);
156 bool SelectCall(const Instruction *I);
157 bool SelectSelect(const Instruction *I);
158 bool SelectRet(const Instruction *I);
162 bool isTypeLegal(const Type *Ty, MVT &VT);
163 bool isLoadTypeLegal(const Type *Ty, MVT &VT);
164 bool ARMEmitLoad(EVT VT, unsigned &ResultReg, Address &Addr);
165 bool ARMEmitStore(EVT VT, unsigned SrcReg, Address &Addr);
166 bool ARMComputeAddress(const Value *Obj, Address &Addr);
167 void ARMSimplifyAddress(Address &Addr, EVT VT);
168 unsigned ARMMaterializeFP(const ConstantFP *CFP, EVT VT);
169 unsigned ARMMaterializeInt(const Constant *C, EVT VT);
170 unsigned ARMMaterializeGV(const GlobalValue *GV, EVT VT);
171 unsigned ARMMoveToFPReg(EVT VT, unsigned SrcReg);
172 unsigned ARMMoveToIntReg(EVT VT, unsigned SrcReg);
174 // Call handling routines.
176 bool FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src, EVT SrcVT,
177 unsigned &ResultReg);
178 CCAssignFn *CCAssignFnForCall(CallingConv::ID CC, bool Return);
179 bool ProcessCallArgs(SmallVectorImpl<Value*> &Args,
180 SmallVectorImpl<unsigned> &ArgRegs,
181 SmallVectorImpl<MVT> &ArgVTs,
182 SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
183 SmallVectorImpl<unsigned> &RegArgs,
186 bool FinishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
187 const Instruction *I, CallingConv::ID CC,
189 bool ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call);
191 // OptionalDef handling routines.
193 bool DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR);
194 const MachineInstrBuilder &AddOptionalDefs(const MachineInstrBuilder &MIB);
197 } // end anonymous namespace
199 #include "ARMGenCallingConv.inc"
201 // DefinesOptionalPredicate - This is different from DefinesPredicate in that
202 // we don't care about implicit defs here, just places we'll need to add a
203 // default CCReg argument. Sets CPSR if we're setting CPSR instead of CCR.
204 bool ARMFastISel::DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR) {
205 const TargetInstrDesc &TID = MI->getDesc();
206 if (!TID.hasOptionalDef())
209 // Look to see if our OptionalDef is defining CPSR or CCR.
210 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
211 const MachineOperand &MO = MI->getOperand(i);
212 if (!MO.isReg() || !MO.isDef()) continue;
213 if (MO.getReg() == ARM::CPSR)
219 // If the machine is predicable go ahead and add the predicate operands, if
220 // it needs default CC operands add those.
221 // TODO: If we want to support thumb1 then we'll need to deal with optional
222 // CPSR defs that need to be added before the remaining operands. See s_cc_out
223 // for descriptions why.
224 const MachineInstrBuilder &
225 ARMFastISel::AddOptionalDefs(const MachineInstrBuilder &MIB) {
226 MachineInstr *MI = &*MIB;
228 // Do we use a predicate?
229 if (TII.isPredicable(MI))
232 // Do we optionally set a predicate? Preds is size > 0 iff the predicate
233 // defines CPSR. All other OptionalDefines in ARM are the CCR register.
235 if (DefinesOptionalPredicate(MI, &CPSR)) {
244 unsigned ARMFastISel::FastEmitInst_(unsigned MachineInstOpcode,
245 const TargetRegisterClass* RC) {
246 unsigned ResultReg = createResultReg(RC);
247 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
249 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg));
253 unsigned ARMFastISel::FastEmitInst_r(unsigned MachineInstOpcode,
254 const TargetRegisterClass *RC,
255 unsigned Op0, bool Op0IsKill) {
256 unsigned ResultReg = createResultReg(RC);
257 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
259 if (II.getNumDefs() >= 1)
260 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
261 .addReg(Op0, Op0IsKill * RegState::Kill));
263 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
264 .addReg(Op0, Op0IsKill * RegState::Kill));
265 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
266 TII.get(TargetOpcode::COPY), ResultReg)
267 .addReg(II.ImplicitDefs[0]));
272 unsigned ARMFastISel::FastEmitInst_rr(unsigned MachineInstOpcode,
273 const TargetRegisterClass *RC,
274 unsigned Op0, bool Op0IsKill,
275 unsigned Op1, bool Op1IsKill) {
276 unsigned ResultReg = createResultReg(RC);
277 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
279 if (II.getNumDefs() >= 1)
280 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
281 .addReg(Op0, Op0IsKill * RegState::Kill)
282 .addReg(Op1, Op1IsKill * RegState::Kill));
284 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
285 .addReg(Op0, Op0IsKill * RegState::Kill)
286 .addReg(Op1, Op1IsKill * RegState::Kill));
287 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
288 TII.get(TargetOpcode::COPY), ResultReg)
289 .addReg(II.ImplicitDefs[0]));
294 unsigned ARMFastISel::FastEmitInst_ri(unsigned MachineInstOpcode,
295 const TargetRegisterClass *RC,
296 unsigned Op0, bool Op0IsKill,
298 unsigned ResultReg = createResultReg(RC);
299 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
301 if (II.getNumDefs() >= 1)
302 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
303 .addReg(Op0, Op0IsKill * RegState::Kill)
306 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
307 .addReg(Op0, Op0IsKill * RegState::Kill)
309 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
310 TII.get(TargetOpcode::COPY), ResultReg)
311 .addReg(II.ImplicitDefs[0]));
316 unsigned ARMFastISel::FastEmitInst_rf(unsigned MachineInstOpcode,
317 const TargetRegisterClass *RC,
318 unsigned Op0, bool Op0IsKill,
319 const ConstantFP *FPImm) {
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)
328 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
329 .addReg(Op0, Op0IsKill * RegState::Kill)
331 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
332 TII.get(TargetOpcode::COPY), ResultReg)
333 .addReg(II.ImplicitDefs[0]));
338 unsigned ARMFastISel::FastEmitInst_rri(unsigned MachineInstOpcode,
339 const TargetRegisterClass *RC,
340 unsigned Op0, bool Op0IsKill,
341 unsigned Op1, bool Op1IsKill,
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)
348 .addReg(Op0, Op0IsKill * RegState::Kill)
349 .addReg(Op1, Op1IsKill * RegState::Kill)
352 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
353 .addReg(Op0, Op0IsKill * RegState::Kill)
354 .addReg(Op1, Op1IsKill * RegState::Kill)
356 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
357 TII.get(TargetOpcode::COPY), ResultReg)
358 .addReg(II.ImplicitDefs[0]));
363 unsigned ARMFastISel::FastEmitInst_i(unsigned MachineInstOpcode,
364 const TargetRegisterClass *RC,
366 unsigned ResultReg = createResultReg(RC);
367 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
369 if (II.getNumDefs() >= 1)
370 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
373 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
375 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
376 TII.get(TargetOpcode::COPY), ResultReg)
377 .addReg(II.ImplicitDefs[0]));
382 unsigned ARMFastISel::FastEmitInst_extractsubreg(MVT RetVT,
383 unsigned Op0, bool Op0IsKill,
385 unsigned ResultReg = createResultReg(TLI.getRegClassFor(RetVT));
386 assert(TargetRegisterInfo::isVirtualRegister(Op0) &&
387 "Cannot yet extract from physregs");
388 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
389 DL, TII.get(TargetOpcode::COPY), ResultReg)
390 .addReg(Op0, getKillRegState(Op0IsKill), Idx));
394 // TODO: Don't worry about 64-bit now, but when this is fixed remove the
395 // checks from the various callers.
396 unsigned ARMFastISel::ARMMoveToFPReg(EVT VT, unsigned SrcReg) {
397 if (VT == MVT::f64) return 0;
399 unsigned MoveReg = createResultReg(TLI.getRegClassFor(VT));
400 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
401 TII.get(ARM::VMOVRS), MoveReg)
406 unsigned ARMFastISel::ARMMoveToIntReg(EVT VT, unsigned SrcReg) {
407 if (VT == MVT::i64) return 0;
409 unsigned MoveReg = createResultReg(TLI.getRegClassFor(VT));
410 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
411 TII.get(ARM::VMOVSR), MoveReg)
416 // For double width floating point we need to materialize two constants
417 // (the high and the low) into integer registers then use a move to get
418 // the combined constant into an FP reg.
419 unsigned ARMFastISel::ARMMaterializeFP(const ConstantFP *CFP, EVT VT) {
420 const APFloat Val = CFP->getValueAPF();
421 bool is64bit = VT == MVT::f64;
423 // This checks to see if we can use VFP3 instructions to materialize
424 // a constant, otherwise we have to go through the constant pool.
425 if (TLI.isFPImmLegal(Val, VT)) {
426 unsigned Opc = is64bit ? ARM::FCONSTD : ARM::FCONSTS;
427 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
428 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
434 // Require VFP2 for loading fp constants.
435 if (!Subtarget->hasVFP2()) return false;
437 // MachineConstantPool wants an explicit alignment.
438 unsigned Align = TD.getPrefTypeAlignment(CFP->getType());
440 // TODO: Figure out if this is correct.
441 Align = TD.getTypeAllocSize(CFP->getType());
443 unsigned Idx = MCP.getConstantPoolIndex(cast<Constant>(CFP), Align);
444 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
445 unsigned Opc = is64bit ? ARM::VLDRD : ARM::VLDRS;
447 // The extra reg is for addrmode5.
448 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
450 .addConstantPoolIndex(Idx)
455 unsigned ARMFastISel::ARMMaterializeInt(const Constant *C, EVT VT) {
457 // For now 32-bit only.
458 if (VT != MVT::i32) return false;
460 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
462 // If we can do this in a single instruction without a constant pool entry
464 const ConstantInt *CI = cast<ConstantInt>(C);
465 if (Subtarget->hasV6T2Ops() && isUInt<16>(CI->getSExtValue())) {
466 unsigned Opc = isThumb ? ARM::t2MOVi16 : ARM::MOVi16;
467 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
468 TII.get(Opc), DestReg)
469 .addImm(CI->getSExtValue()));
473 // MachineConstantPool wants an explicit alignment.
474 unsigned Align = TD.getPrefTypeAlignment(C->getType());
476 // TODO: Figure out if this is correct.
477 Align = TD.getTypeAllocSize(C->getType());
479 unsigned Idx = MCP.getConstantPoolIndex(C, Align);
482 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
483 TII.get(ARM::t2LDRpci), DestReg)
484 .addConstantPoolIndex(Idx));
486 // The extra immediate is for addrmode2.
487 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
488 TII.get(ARM::LDRcp), DestReg)
489 .addConstantPoolIndex(Idx)
495 unsigned ARMFastISel::ARMMaterializeGV(const GlobalValue *GV, EVT VT) {
496 // For now 32-bit only.
497 if (VT != MVT::i32) return 0;
499 Reloc::Model RelocM = TM.getRelocationModel();
501 // TODO: No external globals for now.
502 if (Subtarget->GVIsIndirectSymbol(GV, RelocM)) return 0;
504 // TODO: Need more magic for ARM PIC.
505 if (!isThumb && (RelocM == Reloc::PIC_)) return 0;
507 // MachineConstantPool wants an explicit alignment.
508 unsigned Align = TD.getPrefTypeAlignment(GV->getType());
510 // TODO: Figure out if this is correct.
511 Align = TD.getTypeAllocSize(GV->getType());
515 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb() ? 4 : 8);
516 unsigned Id = AFI->createConstPoolEntryUId();
517 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, Id,
518 ARMCP::CPValue, PCAdj);
519 unsigned Idx = MCP.getConstantPoolIndex(CPV, Align);
522 MachineInstrBuilder MIB;
523 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
525 unsigned Opc = (RelocM != Reloc::PIC_) ? ARM::t2LDRpci : ARM::t2LDRpci_pic;
526 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), DestReg)
527 .addConstantPoolIndex(Idx);
528 if (RelocM == Reloc::PIC_)
531 // The extra immediate is for addrmode2.
532 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(ARM::LDRcp),
534 .addConstantPoolIndex(Idx)
537 AddOptionalDefs(MIB);
541 unsigned ARMFastISel::TargetMaterializeConstant(const Constant *C) {
542 EVT VT = TLI.getValueType(C->getType(), true);
544 // Only handle simple types.
545 if (!VT.isSimple()) return 0;
547 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
548 return ARMMaterializeFP(CFP, VT);
549 else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
550 return ARMMaterializeGV(GV, VT);
551 else if (isa<ConstantInt>(C))
552 return ARMMaterializeInt(C, VT);
557 unsigned ARMFastISel::TargetMaterializeAlloca(const AllocaInst *AI) {
558 // Don't handle dynamic allocas.
559 if (!FuncInfo.StaticAllocaMap.count(AI)) return 0;
562 if (!isLoadTypeLegal(AI->getType(), VT)) return false;
564 DenseMap<const AllocaInst*, int>::iterator SI =
565 FuncInfo.StaticAllocaMap.find(AI);
567 // This will get lowered later into the correct offsets and registers
568 // via rewriteXFrameIndex.
569 if (SI != FuncInfo.StaticAllocaMap.end()) {
570 TargetRegisterClass* RC = TLI.getRegClassFor(VT);
571 unsigned ResultReg = createResultReg(RC);
572 unsigned Opc = isThumb ? ARM::t2ADDri : ARM::ADDri;
573 AddOptionalDefs(BuildMI(*FuncInfo.MBB, *FuncInfo.InsertPt, DL,
574 TII.get(Opc), ResultReg)
575 .addFrameIndex(SI->second)
583 bool ARMFastISel::isTypeLegal(const Type *Ty, MVT &VT) {
584 EVT evt = TLI.getValueType(Ty, true);
586 // Only handle simple types.
587 if (evt == MVT::Other || !evt.isSimple()) return false;
588 VT = evt.getSimpleVT();
590 // Handle all legal types, i.e. a register that will directly hold this
592 return TLI.isTypeLegal(VT);
595 bool ARMFastISel::isLoadTypeLegal(const Type *Ty, MVT &VT) {
596 if (isTypeLegal(Ty, VT)) return true;
598 // If this is a type than can be sign or zero-extended to a basic operation
599 // go ahead and accept it now.
600 if (VT == MVT::i8 || VT == MVT::i16)
606 // Computes the address to get to an object.
607 bool ARMFastISel::ARMComputeAddress(const Value *Obj, Address &Addr) {
608 // Some boilerplate from the X86 FastISel.
609 const User *U = NULL;
610 unsigned Opcode = Instruction::UserOp1;
611 if (const Instruction *I = dyn_cast<Instruction>(Obj)) {
612 // Don't walk into other basic blocks; it's possible we haven't
613 // visited them yet, so the instructions may not yet be assigned
614 // virtual registers.
615 if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(Obj)) ||
616 FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
617 Opcode = I->getOpcode();
620 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Obj)) {
621 Opcode = C->getOpcode();
625 if (const PointerType *Ty = dyn_cast<PointerType>(Obj->getType()))
626 if (Ty->getAddressSpace() > 255)
627 // Fast instruction selection doesn't support the special
634 case Instruction::BitCast: {
635 // Look through bitcasts.
636 return ARMComputeAddress(U->getOperand(0), Addr);
638 case Instruction::IntToPtr: {
639 // Look past no-op inttoptrs.
640 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
641 return ARMComputeAddress(U->getOperand(0), Addr);
644 case Instruction::PtrToInt: {
645 // Look past no-op ptrtoints.
646 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
647 return ARMComputeAddress(U->getOperand(0), Addr);
650 case Instruction::GetElementPtr: {
651 int SavedOffset = Addr.Offset;
652 unsigned SavedBase = Addr.Base.Reg;
653 int TmpOffset = Addr.Offset;
655 // Iterate through the GEP folding the constants into offsets where
657 gep_type_iterator GTI = gep_type_begin(U);
658 for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end();
659 i != e; ++i, ++GTI) {
660 const Value *Op = *i;
661 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
662 const StructLayout *SL = TD.getStructLayout(STy);
663 unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
664 TmpOffset += SL->getElementOffset(Idx);
666 uint64_t S = TD.getTypeAllocSize(GTI.getIndexedType());
667 SmallVector<const Value *, 4> Worklist;
668 Worklist.push_back(Op);
670 Op = Worklist.pop_back_val();
671 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
672 // Constant-offset addressing.
673 TmpOffset += CI->getSExtValue() * S;
674 } else if (isa<AddOperator>(Op) &&
675 isa<ConstantInt>(cast<AddOperator>(Op)->getOperand(1))) {
676 // An add with a constant operand. Fold the constant.
678 cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
679 TmpOffset += CI->getSExtValue() * S;
680 // Add the other operand back to the work list.
681 Worklist.push_back(cast<AddOperator>(Op)->getOperand(0));
683 goto unsupported_gep;
684 } while (!Worklist.empty());
688 // Try to grab the base operand now.
689 Addr.Offset = TmpOffset;
690 if (ARMComputeAddress(U->getOperand(0), Addr)) return true;
692 // We failed, restore everything and try the other options.
693 Addr.Offset = SavedOffset;
694 Addr.Base.Reg = SavedBase;
699 case Instruction::Alloca: {
700 const AllocaInst *AI = cast<AllocaInst>(Obj);
701 unsigned Reg = TargetMaterializeAlloca(AI);
703 if (Reg == 0) return false;
710 // Materialize the global variable's address into a reg which can
711 // then be used later to load the variable.
712 if (const GlobalValue *GV = dyn_cast<GlobalValue>(Obj)) {
713 unsigned Tmp = ARMMaterializeGV(GV, TLI.getValueType(Obj->getType()));
714 if (Tmp == 0) return false;
720 // Try to get this in a register if nothing else has worked.
721 if (Addr.Base.Reg == 0) Addr.Base.Reg = getRegForValue(Obj);
722 return Addr.Base.Reg != 0;
725 void ARMFastISel::ARMSimplifyAddress(Address &Addr, EVT VT) {
727 assert(VT.isSimple() && "Non-simple types are invalid here!");
729 bool needsLowering = false;
730 switch (VT.getSimpleVT().SimpleTy) {
732 assert(false && "Unhandled load/store type!");
737 // Integer loads/stores handle 12-bit offsets.
738 needsLowering = ((Addr.Offset & 0xfff) != Addr.Offset);
742 // Floating point operands handle 8-bit offsets.
743 needsLowering = ((Addr.Offset & 0xff) != Addr.Offset);
747 // Since the offset is too large for the load/store instruction
748 // get the reg+offset into a register.
750 ARMCC::CondCodes Pred = ARMCC::AL;
751 unsigned PredReg = 0;
753 TargetRegisterClass *RC = isThumb ? ARM::tGPRRegisterClass :
754 ARM::GPRRegisterClass;
755 unsigned BaseReg = createResultReg(RC);
758 emitARMRegPlusImmediate(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
759 BaseReg, Addr.Base.Reg, Addr.Offset,
761 static_cast<const ARMBaseInstrInfo&>(TII));
763 assert(AFI->isThumb2Function());
764 emitT2RegPlusImmediate(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
765 BaseReg, Addr.Base.Reg, Addr.Offset, Pred, PredReg,
766 static_cast<const ARMBaseInstrInfo&>(TII));
769 Addr.Base.Reg = BaseReg;
773 bool ARMFastISel::ARMEmitLoad(EVT VT, unsigned &ResultReg, Address &Addr) {
775 assert(VT.isSimple() && "Non-simple types are invalid here!");
777 TargetRegisterClass *RC;
778 bool isFloat = false;
779 switch (VT.getSimpleVT().SimpleTy) {
781 // This is mostly going to be Neon/vector support.
784 Opc = isThumb ? ARM::t2LDRHi12 : ARM::LDRH;
785 RC = ARM::GPRRegisterClass;
788 Opc = isThumb ? ARM::t2LDRBi12 : ARM::LDRBi12;
789 RC = ARM::GPRRegisterClass;
792 Opc = isThumb ? ARM::t2LDRi12 : ARM::LDRi12;
793 RC = ARM::GPRRegisterClass;
797 RC = TLI.getRegClassFor(VT);
802 RC = TLI.getRegClassFor(VT);
807 ResultReg = createResultReg(RC);
809 ARMSimplifyAddress(Addr, VT);
811 // addrmode5 output depends on the selection dag addressing dividing the
812 // offset by 4 that it then later multiplies. Do this here as well.
816 // LDRH needs an additional operand.
817 if (!isThumb && VT.getSimpleVT().SimpleTy == MVT::i16)
818 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
819 TII.get(Opc), ResultReg)
820 .addReg(Addr.Base.Reg).addReg(0).addImm(Addr.Offset));
822 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
823 TII.get(Opc), ResultReg)
824 .addReg(Addr.Base.Reg).addImm(Addr.Offset));
828 bool ARMFastISel::SelectLoad(const Instruction *I) {
829 // Verify we have a legal type before going any further.
831 if (!isLoadTypeLegal(I->getType(), VT))
834 // Our register and offset with innocuous defaults.
837 // See if we can handle this as Reg + Offset
838 if (!ARMComputeAddress(I->getOperand(0), Addr))
842 if (!ARMEmitLoad(VT, ResultReg, Addr)) return false;
844 UpdateValueMap(I, ResultReg);
848 bool ARMFastISel::ARMEmitStore(EVT VT, unsigned SrcReg, Address &Addr) {
850 bool isFloat = false;
851 bool needReg0Op = false;
852 switch (VT.getSimpleVT().SimpleTy) {
853 default: return false;
855 unsigned Res = createResultReg(isThumb ? ARM::tGPRRegisterClass :
856 ARM::GPRRegisterClass);
857 unsigned Opc = isThumb ? ARM::t2ANDri : ARM::ANDri;
858 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
860 .addReg(SrcReg).addImm(1));
862 } // Fallthrough here.
864 StrOpc = isThumb ? ARM::t2STRBi12 : ARM::STRBi12;
867 StrOpc = isThumb ? ARM::t2STRHi12 : ARM::STRH;
871 StrOpc = isThumb ? ARM::t2STRi12 : ARM::STRi12;
874 if (!Subtarget->hasVFP2()) return false;
879 if (!Subtarget->hasVFP2()) return false;
885 ARMSimplifyAddress(Addr, VT);
887 // addrmode5 output depends on the selection dag addressing dividing the
888 // offset by 4 that it then later multiplies. Do this here as well.
892 // FIXME: The 'needReg0Op' bit goes away once STRH is converted to
893 // not use the mega-addrmode stuff.
895 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
897 .addReg(SrcReg).addReg(Addr.Base.Reg).addImm(Addr.Offset));
899 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
901 .addReg(SrcReg).addReg(Addr.Base.Reg)
902 .addReg(0).addImm(Addr.Offset));
907 bool ARMFastISel::SelectStore(const Instruction *I) {
908 Value *Op0 = I->getOperand(0);
911 // Yay type legalization
913 if (!isLoadTypeLegal(I->getOperand(0)->getType(), VT))
916 // Get the value to be stored into a register.
917 SrcReg = getRegForValue(Op0);
921 // Our register and offset with innocuous defaults.
924 // See if we can handle this as Reg + Offset
925 if (!ARMComputeAddress(I->getOperand(1), Addr))
928 if (!ARMEmitStore(VT, SrcReg, Addr)) return false;
933 static ARMCC::CondCodes getComparePred(CmpInst::Predicate Pred) {
935 // Needs two compares...
936 case CmpInst::FCMP_ONE:
937 case CmpInst::FCMP_UEQ:
939 // AL is our "false" for now. The other two need more compares.
941 case CmpInst::ICMP_EQ:
942 case CmpInst::FCMP_OEQ:
944 case CmpInst::ICMP_SGT:
945 case CmpInst::FCMP_OGT:
947 case CmpInst::ICMP_SGE:
948 case CmpInst::FCMP_OGE:
950 case CmpInst::ICMP_UGT:
951 case CmpInst::FCMP_UGT:
953 case CmpInst::FCMP_OLT:
955 case CmpInst::ICMP_ULE:
956 case CmpInst::FCMP_OLE:
958 case CmpInst::FCMP_ORD:
960 case CmpInst::FCMP_UNO:
962 case CmpInst::FCMP_UGE:
964 case CmpInst::ICMP_SLT:
965 case CmpInst::FCMP_ULT:
967 case CmpInst::ICMP_SLE:
968 case CmpInst::FCMP_ULE:
970 case CmpInst::FCMP_UNE:
971 case CmpInst::ICMP_NE:
973 case CmpInst::ICMP_UGE:
975 case CmpInst::ICMP_ULT:
980 bool ARMFastISel::SelectBranch(const Instruction *I) {
981 const BranchInst *BI = cast<BranchInst>(I);
982 MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
983 MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
985 // Simple branch support.
987 // If we can, avoid recomputing the compare - redoing it could lead to wonky
989 // TODO: Factor this out.
990 if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
991 if (CI->hasOneUse() && (CI->getParent() == I->getParent())) {
993 const Type *Ty = CI->getOperand(0)->getType();
994 if (!isTypeLegal(Ty, VT))
997 bool isFloat = (Ty->isDoubleTy() || Ty->isFloatTy());
998 if (isFloat && !Subtarget->hasVFP2())
1003 switch (VT.SimpleTy) {
1004 default: return false;
1005 // TODO: Verify compares.
1007 CmpOpc = ARM::VCMPES;
1008 CondReg = ARM::FPSCR;
1011 CmpOpc = ARM::VCMPED;
1012 CondReg = ARM::FPSCR;
1015 CmpOpc = isThumb ? ARM::t2CMPrr : ARM::CMPrr;
1016 CondReg = ARM::CPSR;
1020 // Get the compare predicate.
1021 ARMCC::CondCodes ARMPred = getComparePred(CI->getPredicate());
1023 // We may not handle every CC for now.
1024 if (ARMPred == ARMCC::AL) return false;
1026 unsigned Arg1 = getRegForValue(CI->getOperand(0));
1027 if (Arg1 == 0) return false;
1029 unsigned Arg2 = getRegForValue(CI->getOperand(1));
1030 if (Arg2 == 0) return false;
1032 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1034 .addReg(Arg1).addReg(Arg2));
1036 // For floating point we need to move the result to a comparison register
1037 // that we can then use for branches.
1039 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1040 TII.get(ARM::FMSTAT)));
1042 unsigned BrOpc = isThumb ? ARM::t2Bcc : ARM::Bcc;
1043 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
1044 .addMBB(TBB).addImm(ARMPred).addReg(ARM::CPSR);
1045 FastEmitBranch(FBB, DL);
1046 FuncInfo.MBB->addSuccessor(TBB);
1051 unsigned CmpReg = getRegForValue(BI->getCondition());
1052 if (CmpReg == 0) return false;
1054 // Re-set the flags just in case.
1055 unsigned CmpOpc = isThumb ? ARM::t2CMPri : ARM::CMPri;
1056 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
1057 .addReg(CmpReg).addImm(0));
1059 unsigned BrOpc = isThumb ? ARM::t2Bcc : ARM::Bcc;
1060 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
1061 .addMBB(TBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
1062 FastEmitBranch(FBB, DL);
1063 FuncInfo.MBB->addSuccessor(TBB);
1067 bool ARMFastISel::SelectCmp(const Instruction *I) {
1068 const CmpInst *CI = cast<CmpInst>(I);
1071 const Type *Ty = CI->getOperand(0)->getType();
1072 if (!isTypeLegal(Ty, VT))
1075 bool isFloat = (Ty->isDoubleTy() || Ty->isFloatTy());
1076 if (isFloat && !Subtarget->hasVFP2())
1081 switch (VT.SimpleTy) {
1082 default: return false;
1083 // TODO: Verify compares.
1085 CmpOpc = ARM::VCMPES;
1086 CondReg = ARM::FPSCR;
1089 CmpOpc = ARM::VCMPED;
1090 CondReg = ARM::FPSCR;
1093 CmpOpc = isThumb ? ARM::t2CMPrr : ARM::CMPrr;
1094 CondReg = ARM::CPSR;
1098 // Get the compare predicate.
1099 ARMCC::CondCodes ARMPred = getComparePred(CI->getPredicate());
1101 // We may not handle every CC for now.
1102 if (ARMPred == ARMCC::AL) return false;
1104 unsigned Arg1 = getRegForValue(CI->getOperand(0));
1105 if (Arg1 == 0) return false;
1107 unsigned Arg2 = getRegForValue(CI->getOperand(1));
1108 if (Arg2 == 0) return false;
1110 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
1111 .addReg(Arg1).addReg(Arg2));
1113 // For floating point we need to move the result to a comparison register
1114 // that we can then use for branches.
1116 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1117 TII.get(ARM::FMSTAT)));
1119 // Now set a register based on the comparison. Explicitly set the predicates
1121 unsigned MovCCOpc = isThumb ? ARM::t2MOVCCi : ARM::MOVCCi;
1122 TargetRegisterClass *RC = isThumb ? ARM::rGPRRegisterClass
1123 : ARM::GPRRegisterClass;
1124 unsigned DestReg = createResultReg(RC);
1126 = ConstantInt::get(Type::getInt32Ty(*Context), 0);
1127 unsigned ZeroReg = TargetMaterializeConstant(Zero);
1128 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), DestReg)
1129 .addReg(ZeroReg).addImm(1)
1130 .addImm(ARMPred).addReg(CondReg);
1132 UpdateValueMap(I, DestReg);
1136 bool ARMFastISel::SelectFPExt(const Instruction *I) {
1137 // Make sure we have VFP and that we're extending float to double.
1138 if (!Subtarget->hasVFP2()) return false;
1140 Value *V = I->getOperand(0);
1141 if (!I->getType()->isDoubleTy() ||
1142 !V->getType()->isFloatTy()) return false;
1144 unsigned Op = getRegForValue(V);
1145 if (Op == 0) return false;
1147 unsigned Result = createResultReg(ARM::DPRRegisterClass);
1148 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1149 TII.get(ARM::VCVTDS), Result)
1151 UpdateValueMap(I, Result);
1155 bool ARMFastISel::SelectFPTrunc(const Instruction *I) {
1156 // Make sure we have VFP and that we're truncating double to float.
1157 if (!Subtarget->hasVFP2()) return false;
1159 Value *V = I->getOperand(0);
1160 if (!(I->getType()->isFloatTy() &&
1161 V->getType()->isDoubleTy())) return false;
1163 unsigned Op = getRegForValue(V);
1164 if (Op == 0) return false;
1166 unsigned Result = createResultReg(ARM::SPRRegisterClass);
1167 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1168 TII.get(ARM::VCVTSD), Result)
1170 UpdateValueMap(I, Result);
1174 bool ARMFastISel::SelectSIToFP(const Instruction *I) {
1175 // Make sure we have VFP.
1176 if (!Subtarget->hasVFP2()) return false;
1179 const Type *Ty = I->getType();
1180 if (!isTypeLegal(Ty, DstVT))
1183 unsigned Op = getRegForValue(I->getOperand(0));
1184 if (Op == 0) return false;
1186 // The conversion routine works on fp-reg to fp-reg and the operand above
1187 // was an integer, move it to the fp registers if possible.
1188 unsigned FP = ARMMoveToFPReg(MVT::f32, Op);
1189 if (FP == 0) return false;
1192 if (Ty->isFloatTy()) Opc = ARM::VSITOS;
1193 else if (Ty->isDoubleTy()) Opc = ARM::VSITOD;
1196 unsigned ResultReg = createResultReg(TLI.getRegClassFor(DstVT));
1197 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
1200 UpdateValueMap(I, ResultReg);
1204 bool ARMFastISel::SelectFPToSI(const Instruction *I) {
1205 // Make sure we have VFP.
1206 if (!Subtarget->hasVFP2()) return false;
1209 const Type *RetTy = I->getType();
1210 if (!isTypeLegal(RetTy, DstVT))
1213 unsigned Op = getRegForValue(I->getOperand(0));
1214 if (Op == 0) return false;
1217 const Type *OpTy = I->getOperand(0)->getType();
1218 if (OpTy->isFloatTy()) Opc = ARM::VTOSIZS;
1219 else if (OpTy->isDoubleTy()) Opc = ARM::VTOSIZD;
1222 // f64->s32 or f32->s32 both need an intermediate f32 reg.
1223 unsigned ResultReg = createResultReg(TLI.getRegClassFor(MVT::f32));
1224 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
1228 // This result needs to be in an integer register, but the conversion only
1229 // takes place in fp-regs.
1230 unsigned IntReg = ARMMoveToIntReg(DstVT, ResultReg);
1231 if (IntReg == 0) return false;
1233 UpdateValueMap(I, IntReg);
1237 bool ARMFastISel::SelectSelect(const Instruction *I) {
1239 if (!isTypeLegal(I->getType(), VT))
1242 // Things need to be register sized for register moves.
1243 if (VT != MVT::i32) return false;
1244 const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
1246 unsigned CondReg = getRegForValue(I->getOperand(0));
1247 if (CondReg == 0) return false;
1248 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1249 if (Op1Reg == 0) return false;
1250 unsigned Op2Reg = getRegForValue(I->getOperand(2));
1251 if (Op2Reg == 0) return false;
1253 unsigned CmpOpc = isThumb ? ARM::t2TSTri : ARM::TSTri;
1254 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
1255 .addReg(CondReg).addImm(1));
1256 unsigned ResultReg = createResultReg(RC);
1257 unsigned MovCCOpc = isThumb ? ARM::t2MOVCCr : ARM::MOVCCr;
1258 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), ResultReg)
1259 .addReg(Op1Reg).addReg(Op2Reg)
1260 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
1261 UpdateValueMap(I, ResultReg);
1265 bool ARMFastISel::SelectSDiv(const Instruction *I) {
1267 const Type *Ty = I->getType();
1268 if (!isTypeLegal(Ty, VT))
1271 // If we have integer div support we should have selected this automagically.
1272 // In case we have a real miss go ahead and return false and we'll pick
1274 if (Subtarget->hasDivide()) return false;
1276 // Otherwise emit a libcall.
1277 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
1279 LC = RTLIB::SDIV_I8;
1280 else if (VT == MVT::i16)
1281 LC = RTLIB::SDIV_I16;
1282 else if (VT == MVT::i32)
1283 LC = RTLIB::SDIV_I32;
1284 else if (VT == MVT::i64)
1285 LC = RTLIB::SDIV_I64;
1286 else if (VT == MVT::i128)
1287 LC = RTLIB::SDIV_I128;
1288 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SDIV!");
1290 return ARMEmitLibcall(I, LC);
1293 bool ARMFastISel::SelectSRem(const Instruction *I) {
1295 const Type *Ty = I->getType();
1296 if (!isTypeLegal(Ty, VT))
1299 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
1301 LC = RTLIB::SREM_I8;
1302 else if (VT == MVT::i16)
1303 LC = RTLIB::SREM_I16;
1304 else if (VT == MVT::i32)
1305 LC = RTLIB::SREM_I32;
1306 else if (VT == MVT::i64)
1307 LC = RTLIB::SREM_I64;
1308 else if (VT == MVT::i128)
1309 LC = RTLIB::SREM_I128;
1310 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SREM!");
1312 return ARMEmitLibcall(I, LC);
1315 bool ARMFastISel::SelectBinaryOp(const Instruction *I, unsigned ISDOpcode) {
1316 EVT VT = TLI.getValueType(I->getType(), true);
1318 // We can get here in the case when we want to use NEON for our fp
1319 // operations, but can't figure out how to. Just use the vfp instructions
1321 // FIXME: It'd be nice to use NEON instructions.
1322 const Type *Ty = I->getType();
1323 bool isFloat = (Ty->isDoubleTy() || Ty->isFloatTy());
1324 if (isFloat && !Subtarget->hasVFP2())
1327 unsigned Op1 = getRegForValue(I->getOperand(0));
1328 if (Op1 == 0) return false;
1330 unsigned Op2 = getRegForValue(I->getOperand(1));
1331 if (Op2 == 0) return false;
1334 bool is64bit = VT == MVT::f64 || VT == MVT::i64;
1335 switch (ISDOpcode) {
1336 default: return false;
1338 Opc = is64bit ? ARM::VADDD : ARM::VADDS;
1341 Opc = is64bit ? ARM::VSUBD : ARM::VSUBS;
1344 Opc = is64bit ? ARM::VMULD : ARM::VMULS;
1347 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
1348 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1349 TII.get(Opc), ResultReg)
1350 .addReg(Op1).addReg(Op2));
1351 UpdateValueMap(I, ResultReg);
1355 // Call Handling Code
1357 bool ARMFastISel::FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src,
1358 EVT SrcVT, unsigned &ResultReg) {
1359 unsigned RR = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), Opc,
1360 Src, /*TODO: Kill=*/false);
1369 // This is largely taken directly from CCAssignFnForNode - we don't support
1370 // varargs in FastISel so that part has been removed.
1371 // TODO: We may not support all of this.
1372 CCAssignFn *ARMFastISel::CCAssignFnForCall(CallingConv::ID CC, bool Return) {
1375 llvm_unreachable("Unsupported calling convention");
1376 case CallingConv::Fast:
1377 // Ignore fastcc. Silence compiler warnings.
1378 (void)RetFastCC_ARM_APCS;
1379 (void)FastCC_ARM_APCS;
1381 case CallingConv::C:
1382 // Use target triple & subtarget features to do actual dispatch.
1383 if (Subtarget->isAAPCS_ABI()) {
1384 if (Subtarget->hasVFP2() &&
1385 FloatABIType == FloatABI::Hard)
1386 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
1388 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
1390 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
1391 case CallingConv::ARM_AAPCS_VFP:
1392 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
1393 case CallingConv::ARM_AAPCS:
1394 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
1395 case CallingConv::ARM_APCS:
1396 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
1400 bool ARMFastISel::ProcessCallArgs(SmallVectorImpl<Value*> &Args,
1401 SmallVectorImpl<unsigned> &ArgRegs,
1402 SmallVectorImpl<MVT> &ArgVTs,
1403 SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
1404 SmallVectorImpl<unsigned> &RegArgs,
1406 unsigned &NumBytes) {
1407 SmallVector<CCValAssign, 16> ArgLocs;
1408 CCState CCInfo(CC, false, TM, ArgLocs, *Context);
1409 CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CCAssignFnForCall(CC, false));
1411 // Get a count of how many bytes are to be pushed on the stack.
1412 NumBytes = CCInfo.getNextStackOffset();
1414 // Issue CALLSEQ_START
1415 unsigned AdjStackDown = TM.getRegisterInfo()->getCallFrameSetupOpcode();
1416 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1417 TII.get(AdjStackDown))
1420 // Process the args.
1421 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1422 CCValAssign &VA = ArgLocs[i];
1423 unsigned Arg = ArgRegs[VA.getValNo()];
1424 MVT ArgVT = ArgVTs[VA.getValNo()];
1426 // We don't handle NEON parameters yet.
1427 if (VA.getLocVT().isVector() && VA.getLocVT().getSizeInBits() > 64)
1430 // Handle arg promotion, etc.
1431 switch (VA.getLocInfo()) {
1432 case CCValAssign::Full: break;
1433 case CCValAssign::SExt: {
1434 bool Emitted = FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
1436 assert(Emitted && "Failed to emit a sext!"); Emitted=Emitted;
1438 ArgVT = VA.getLocVT();
1441 case CCValAssign::ZExt: {
1442 bool Emitted = FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
1444 assert(Emitted && "Failed to emit a zext!"); Emitted=Emitted;
1446 ArgVT = VA.getLocVT();
1449 case CCValAssign::AExt: {
1450 bool Emitted = FastEmitExtend(ISD::ANY_EXTEND, VA.getLocVT(),
1453 Emitted = FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
1456 Emitted = FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
1459 assert(Emitted && "Failed to emit a aext!"); Emitted=Emitted;
1460 ArgVT = VA.getLocVT();
1463 case CCValAssign::BCvt: {
1464 unsigned BC = FastEmit_r(ArgVT, VA.getLocVT(), ISD::BIT_CONVERT, Arg,
1465 /*TODO: Kill=*/false);
1466 assert(BC != 0 && "Failed to emit a bitcast!");
1468 ArgVT = VA.getLocVT();
1471 default: llvm_unreachable("Unknown arg promotion!");
1474 // Now copy/store arg to correct locations.
1475 if (VA.isRegLoc() && !VA.needsCustom()) {
1476 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1479 RegArgs.push_back(VA.getLocReg());
1480 } else if (VA.needsCustom()) {
1481 // TODO: We need custom lowering for vector (v2f64) args.
1482 if (VA.getLocVT() != MVT::f64) return false;
1484 CCValAssign &NextVA = ArgLocs[++i];
1486 // TODO: Only handle register args for now.
1487 if(!(VA.isRegLoc() && NextVA.isRegLoc())) return false;
1489 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1490 TII.get(ARM::VMOVRRD), VA.getLocReg())
1491 .addReg(NextVA.getLocReg(), RegState::Define)
1493 RegArgs.push_back(VA.getLocReg());
1494 RegArgs.push_back(NextVA.getLocReg());
1496 assert(VA.isMemLoc());
1497 // Need to store on the stack.
1499 Addr.BaseType = Address::RegBase;
1500 Addr.Base.Reg = ARM::SP;
1501 Addr.Offset = VA.getLocMemOffset();
1503 if (!ARMEmitStore(ArgVT, Arg, Addr)) return false;
1509 bool ARMFastISel::FinishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
1510 const Instruction *I, CallingConv::ID CC,
1511 unsigned &NumBytes) {
1512 // Issue CALLSEQ_END
1513 unsigned AdjStackUp = TM.getRegisterInfo()->getCallFrameDestroyOpcode();
1514 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1515 TII.get(AdjStackUp))
1516 .addImm(NumBytes).addImm(0));
1518 // Now the return value.
1519 if (RetVT != MVT::isVoid) {
1520 SmallVector<CCValAssign, 16> RVLocs;
1521 CCState CCInfo(CC, false, TM, RVLocs, *Context);
1522 CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC, true));
1524 // Copy all of the result registers out of their specified physreg.
1525 if (RVLocs.size() == 2 && RetVT == MVT::f64) {
1526 // For this move we copy into two registers and then move into the
1527 // double fp reg we want.
1528 EVT DestVT = RVLocs[0].getValVT();
1529 TargetRegisterClass* DstRC = TLI.getRegClassFor(DestVT);
1530 unsigned ResultReg = createResultReg(DstRC);
1531 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1532 TII.get(ARM::VMOVDRR), ResultReg)
1533 .addReg(RVLocs[0].getLocReg())
1534 .addReg(RVLocs[1].getLocReg()));
1536 UsedRegs.push_back(RVLocs[0].getLocReg());
1537 UsedRegs.push_back(RVLocs[1].getLocReg());
1539 // Finally update the result.
1540 UpdateValueMap(I, ResultReg);
1542 assert(RVLocs.size() == 1 &&"Can't handle non-double multi-reg retvals!");
1543 EVT CopyVT = RVLocs[0].getValVT();
1544 TargetRegisterClass* DstRC = TLI.getRegClassFor(CopyVT);
1546 unsigned ResultReg = createResultReg(DstRC);
1547 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1548 ResultReg).addReg(RVLocs[0].getLocReg());
1549 UsedRegs.push_back(RVLocs[0].getLocReg());
1551 // Finally update the result.
1552 UpdateValueMap(I, ResultReg);
1559 bool ARMFastISel::SelectRet(const Instruction *I) {
1560 const ReturnInst *Ret = cast<ReturnInst>(I);
1561 const Function &F = *I->getParent()->getParent();
1563 if (!FuncInfo.CanLowerReturn)
1569 CallingConv::ID CC = F.getCallingConv();
1570 if (Ret->getNumOperands() > 0) {
1571 SmallVector<ISD::OutputArg, 4> Outs;
1572 GetReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(),
1575 // Analyze operands of the call, assigning locations to each operand.
1576 SmallVector<CCValAssign, 16> ValLocs;
1577 CCState CCInfo(CC, F.isVarArg(), TM, ValLocs, I->getContext());
1578 CCInfo.AnalyzeReturn(Outs, CCAssignFnForCall(CC, true /* is Ret */));
1580 const Value *RV = Ret->getOperand(0);
1581 unsigned Reg = getRegForValue(RV);
1585 // Only handle a single return value for now.
1586 if (ValLocs.size() != 1)
1589 CCValAssign &VA = ValLocs[0];
1591 // Don't bother handling odd stuff for now.
1592 if (VA.getLocInfo() != CCValAssign::Full)
1594 // Only handle register returns for now.
1597 // TODO: For now, don't try to handle cases where getLocInfo()
1598 // says Full but the types don't match.
1599 if (TLI.getValueType(RV->getType()) != VA.getValVT())
1603 unsigned SrcReg = Reg + VA.getValNo();
1604 unsigned DstReg = VA.getLocReg();
1605 const TargetRegisterClass* SrcRC = MRI.getRegClass(SrcReg);
1606 // Avoid a cross-class copy. This is very unlikely.
1607 if (!SrcRC->contains(DstReg))
1609 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1610 DstReg).addReg(SrcReg);
1612 // Mark the register as live out of the function.
1613 MRI.addLiveOut(VA.getLocReg());
1616 unsigned RetOpc = isThumb ? ARM::tBX_RET : ARM::BX_RET;
1617 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1622 // A quick function that will emit a call for a named libcall in F with the
1623 // vector of passed arguments for the Instruction in I. We can assume that we
1624 // can emit a call for any libcall we can produce. This is an abridged version
1625 // of the full call infrastructure since we won't need to worry about things
1626 // like computed function pointers or strange arguments at call sites.
1627 // TODO: Try to unify this and the normal call bits for ARM, then try to unify
1629 bool ARMFastISel::ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call) {
1630 CallingConv::ID CC = TLI.getLibcallCallingConv(Call);
1632 // Handle *simple* calls for now.
1633 const Type *RetTy = I->getType();
1635 if (RetTy->isVoidTy())
1636 RetVT = MVT::isVoid;
1637 else if (!isTypeLegal(RetTy, RetVT))
1640 // For now we're using BLX etc on the assumption that we have v5t ops.
1641 if (!Subtarget->hasV5TOps()) return false;
1643 // Set up the argument vectors.
1644 SmallVector<Value*, 8> Args;
1645 SmallVector<unsigned, 8> ArgRegs;
1646 SmallVector<MVT, 8> ArgVTs;
1647 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
1648 Args.reserve(I->getNumOperands());
1649 ArgRegs.reserve(I->getNumOperands());
1650 ArgVTs.reserve(I->getNumOperands());
1651 ArgFlags.reserve(I->getNumOperands());
1652 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
1653 Value *Op = I->getOperand(i);
1654 unsigned Arg = getRegForValue(Op);
1655 if (Arg == 0) return false;
1657 const Type *ArgTy = Op->getType();
1659 if (!isTypeLegal(ArgTy, ArgVT)) return false;
1661 ISD::ArgFlagsTy Flags;
1662 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
1663 Flags.setOrigAlign(OriginalAlignment);
1666 ArgRegs.push_back(Arg);
1667 ArgVTs.push_back(ArgVT);
1668 ArgFlags.push_back(Flags);
1671 // Handle the arguments now that we've gotten them.
1672 SmallVector<unsigned, 4> RegArgs;
1674 if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
1677 // Issue the call, BLXr9 for darwin, BLX otherwise. This uses V5 ops.
1678 // TODO: Turn this into the table of arm call ops.
1679 MachineInstrBuilder MIB;
1682 CallOpc = Subtarget->isTargetDarwin() ? ARM::tBLXi_r9 : ARM::tBLXi;
1684 CallOpc = Subtarget->isTargetDarwin() ? ARM::BLr9 : ARM::BL;
1685 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CallOpc))
1686 .addExternalSymbol(TLI.getLibcallName(Call));
1688 // Add implicit physical register uses to the call.
1689 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
1690 MIB.addReg(RegArgs[i]);
1692 // Finish off the call including any return values.
1693 SmallVector<unsigned, 4> UsedRegs;
1694 if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false;
1696 // Set all unused physreg defs as dead.
1697 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
1702 bool ARMFastISel::SelectCall(const Instruction *I) {
1703 const CallInst *CI = cast<CallInst>(I);
1704 const Value *Callee = CI->getCalledValue();
1706 // Can't handle inline asm or worry about intrinsics yet.
1707 if (isa<InlineAsm>(Callee) || isa<IntrinsicInst>(CI)) return false;
1709 // Only handle global variable Callees that are direct calls.
1710 const GlobalValue *GV = dyn_cast<GlobalValue>(Callee);
1711 if (!GV || Subtarget->GVIsIndirectSymbol(GV, TM.getRelocationModel()))
1714 // Check the calling convention.
1715 ImmutableCallSite CS(CI);
1716 CallingConv::ID CC = CS.getCallingConv();
1718 // TODO: Avoid some calling conventions?
1720 // Let SDISel handle vararg functions.
1721 const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
1722 const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
1723 if (FTy->isVarArg())
1726 // Handle *simple* calls for now.
1727 const Type *RetTy = I->getType();
1729 if (RetTy->isVoidTy())
1730 RetVT = MVT::isVoid;
1731 else if (!isTypeLegal(RetTy, RetVT))
1734 // For now we're using BLX etc on the assumption that we have v5t ops.
1736 if (!Subtarget->hasV5TOps()) return false;
1738 // Set up the argument vectors.
1739 SmallVector<Value*, 8> Args;
1740 SmallVector<unsigned, 8> ArgRegs;
1741 SmallVector<MVT, 8> ArgVTs;
1742 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
1743 Args.reserve(CS.arg_size());
1744 ArgRegs.reserve(CS.arg_size());
1745 ArgVTs.reserve(CS.arg_size());
1746 ArgFlags.reserve(CS.arg_size());
1747 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
1749 unsigned Arg = getRegForValue(*i);
1753 ISD::ArgFlagsTy Flags;
1754 unsigned AttrInd = i - CS.arg_begin() + 1;
1755 if (CS.paramHasAttr(AttrInd, Attribute::SExt))
1757 if (CS.paramHasAttr(AttrInd, Attribute::ZExt))
1760 // FIXME: Only handle *easy* calls for now.
1761 if (CS.paramHasAttr(AttrInd, Attribute::InReg) ||
1762 CS.paramHasAttr(AttrInd, Attribute::StructRet) ||
1763 CS.paramHasAttr(AttrInd, Attribute::Nest) ||
1764 CS.paramHasAttr(AttrInd, Attribute::ByVal))
1767 const Type *ArgTy = (*i)->getType();
1769 if (!isTypeLegal(ArgTy, ArgVT))
1771 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
1772 Flags.setOrigAlign(OriginalAlignment);
1775 ArgRegs.push_back(Arg);
1776 ArgVTs.push_back(ArgVT);
1777 ArgFlags.push_back(Flags);
1780 // Handle the arguments now that we've gotten them.
1781 SmallVector<unsigned, 4> RegArgs;
1783 if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
1786 // Issue the call, BLXr9 for darwin, BLX otherwise. This uses V5 ops.
1787 // TODO: Turn this into the table of arm call ops.
1788 MachineInstrBuilder MIB;
1791 CallOpc = Subtarget->isTargetDarwin() ? ARM::tBLXi_r9 : ARM::tBLXi;
1793 CallOpc = Subtarget->isTargetDarwin() ? ARM::BLr9 : ARM::BL;
1794 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CallOpc))
1795 .addGlobalAddress(GV, 0, 0);
1797 // Add implicit physical register uses to the call.
1798 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
1799 MIB.addReg(RegArgs[i]);
1801 // Finish off the call including any return values.
1802 SmallVector<unsigned, 4> UsedRegs;
1803 if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false;
1805 // Set all unused physreg defs as dead.
1806 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
1812 // TODO: SoftFP support.
1813 bool ARMFastISel::TargetSelectInstruction(const Instruction *I) {
1815 switch (I->getOpcode()) {
1816 case Instruction::Load:
1817 return SelectLoad(I);
1818 case Instruction::Store:
1819 return SelectStore(I);
1820 case Instruction::Br:
1821 return SelectBranch(I);
1822 case Instruction::ICmp:
1823 case Instruction::FCmp:
1824 return SelectCmp(I);
1825 case Instruction::FPExt:
1826 return SelectFPExt(I);
1827 case Instruction::FPTrunc:
1828 return SelectFPTrunc(I);
1829 case Instruction::SIToFP:
1830 return SelectSIToFP(I);
1831 case Instruction::FPToSI:
1832 return SelectFPToSI(I);
1833 case Instruction::FAdd:
1834 return SelectBinaryOp(I, ISD::FADD);
1835 case Instruction::FSub:
1836 return SelectBinaryOp(I, ISD::FSUB);
1837 case Instruction::FMul:
1838 return SelectBinaryOp(I, ISD::FMUL);
1839 case Instruction::SDiv:
1840 return SelectSDiv(I);
1841 case Instruction::SRem:
1842 return SelectSRem(I);
1843 case Instruction::Call:
1844 return SelectCall(I);
1845 case Instruction::Select:
1846 return SelectSelect(I);
1847 case Instruction::Ret:
1848 return SelectRet(I);
1855 llvm::FastISel *ARM::createFastISel(FunctionLoweringInfo &funcInfo) {
1856 // Completely untested on non-darwin.
1857 const TargetMachine &TM = funcInfo.MF->getTarget();
1859 // Darwin and thumb1 only for now.
1860 const ARMSubtarget *Subtarget = &TM.getSubtarget<ARMSubtarget>();
1861 if (Subtarget->isTargetDarwin() && !Subtarget->isThumb1Only() &&
1862 !DisableARMFastISel)
1863 return new ARMFastISel(funcInfo);