ARMFastISel.cpp revision 022b7fb045973531a1b05f4f0dae28fd836c9890
1//===-- ARMFastISel.cpp - ARM FastISel implementation ---------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 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. 13// 14//===----------------------------------------------------------------------===// 15 16#include "ARM.h" 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" 46using namespace llvm; 47 48static cl::opt<bool> 49EnableARMFastISel("arm-fast-isel", 50 cl::desc("Turn on experimental ARM fast-isel support"), 51 cl::init(false), cl::Hidden); 52 53namespace { 54 55class ARMFastISel : public FastISel { 56 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; 63 ARMFunctionInfo *AFI; 64 65 // Convenience variables to avoid some queries. 66 bool isThumb; 67 LLVMContext *Context; 68 69 public: 70 explicit ARMFastISel(FunctionLoweringInfo &funcInfo) 71 : FastISel(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(); 79 } 80 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, 94 uint64_t Imm); 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, 101 uint64_t Imm); 102 virtual unsigned FastEmitInst_rri(unsigned MachineInstOpcode, 103 const TargetRegisterClass *RC, 104 unsigned Op0, bool Op0IsKill, 105 unsigned Op1, bool Op1IsKill, 106 uint64_t Imm); 107 virtual unsigned FastEmitInst_extractsubreg(MVT RetVT, 108 unsigned Op0, bool Op0IsKill, 109 uint32_t Idx); 110 111 // 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); 115 116 #include "ARMGenFastISel.inc" 117 118 // Instruction selection routines. 119 private: 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); 131 132 // Utility routines. 133 private: 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); 146 147 // Call handling routines. 148 private: 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, 155 CallingConv::ID CC, 156 unsigned &NumBytes); 157 bool FinishCall(EVT RetVT, SmallVectorImpl<unsigned> &UsedRegs, 158 const Instruction *I, CallingConv::ID CC, 159 unsigned &NumBytes); 160 bool ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call); 161 162 // OptionalDef handling routines. 163 private: 164 bool DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR); 165 const MachineInstrBuilder &AddOptionalDefs(const MachineInstrBuilder &MIB); 166}; 167 168} // end anonymous namespace 169 170#include "ARMGenCallingConv.inc" 171 172// 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. 175bool ARMFastISel::DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR) { 176 const TargetInstrDesc &TID = MI->getDesc(); 177 if (!TID.hasOptionalDef()) 178 return false; 179 180 // 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) 185 *CPSR = true; 186 } 187 return true; 188} 189 190// If the machine is predicable go ahead and add the predicate operands, if 191// it needs default CC operands add those. 192const MachineInstrBuilder & 193ARMFastISel::AddOptionalDefs(const MachineInstrBuilder &MIB) { 194 MachineInstr *MI = &*MIB; 195 196 // Do we use a predicate? 197 if (TII.isPredicable(MI)) 198 AddDefaultPred(MIB); 199 200 // 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. 202 bool CPSR = false; 203 if (DefinesOptionalPredicate(MI, &CPSR)) { 204 if (CPSR) 205 AddDefaultT1CC(MIB); 206 else 207 AddDefaultCC(MIB); 208 } 209 return MIB; 210} 211 212unsigned ARMFastISel::FastEmitInst_(unsigned MachineInstOpcode, 213 const TargetRegisterClass* RC) { 214 unsigned ResultReg = createResultReg(RC); 215 const TargetInstrDesc &II = TII.get(MachineInstOpcode); 216 217 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)); 218 return ResultReg; 219} 220 221unsigned 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); 226 227 if (II.getNumDefs() >= 1) 228 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg) 229 .addReg(Op0, Op0IsKill * RegState::Kill)); 230 else { 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])); 236 } 237 return ResultReg; 238} 239 240unsigned 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); 246 247 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)); 251 else { 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])); 258 } 259 return ResultReg; 260} 261 262unsigned ARMFastISel::FastEmitInst_ri(unsigned MachineInstOpcode, 263 const TargetRegisterClass *RC, 264 unsigned Op0, bool Op0IsKill, 265 uint64_t Imm) { 266 unsigned ResultReg = createResultReg(RC); 267 const TargetInstrDesc &II = TII.get(MachineInstOpcode); 268 269 if (II.getNumDefs() >= 1) 270 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg) 271 .addReg(Op0, Op0IsKill * RegState::Kill) 272 .addImm(Imm)); 273 else { 274 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II) 275 .addReg(Op0, Op0IsKill * RegState::Kill) 276 .addImm(Imm)); 277 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, 278 TII.get(TargetOpcode::COPY), ResultReg) 279 .addReg(II.ImplicitDefs[0])); 280 } 281 return ResultReg; 282} 283 284unsigned 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); 290 291 if (II.getNumDefs() >= 1) 292 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg) 293 .addReg(Op0, Op0IsKill * RegState::Kill) 294 .addFPImm(FPImm)); 295 else { 296 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II) 297 .addReg(Op0, Op0IsKill * RegState::Kill) 298 .addFPImm(FPImm)); 299 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, 300 TII.get(TargetOpcode::COPY), ResultReg) 301 .addReg(II.ImplicitDefs[0])); 302 } 303 return ResultReg; 304} 305 306unsigned ARMFastISel::FastEmitInst_rri(unsigned MachineInstOpcode, 307 const TargetRegisterClass *RC, 308 unsigned Op0, bool Op0IsKill, 309 unsigned Op1, bool Op1IsKill, 310 uint64_t Imm) { 311 unsigned ResultReg = createResultReg(RC); 312 const TargetInstrDesc &II = TII.get(MachineInstOpcode); 313 314 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) 318 .addImm(Imm)); 319 else { 320 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II) 321 .addReg(Op0, Op0IsKill * RegState::Kill) 322 .addReg(Op1, Op1IsKill * RegState::Kill) 323 .addImm(Imm)); 324 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, 325 TII.get(TargetOpcode::COPY), ResultReg) 326 .addReg(II.ImplicitDefs[0])); 327 } 328 return ResultReg; 329} 330 331unsigned ARMFastISel::FastEmitInst_i(unsigned MachineInstOpcode, 332 const TargetRegisterClass *RC, 333 uint64_t Imm) { 334 unsigned ResultReg = createResultReg(RC); 335 const TargetInstrDesc &II = TII.get(MachineInstOpcode); 336 337 if (II.getNumDefs() >= 1) 338 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg) 339 .addImm(Imm)); 340 else { 341 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II) 342 .addImm(Imm)); 343 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, 344 TII.get(TargetOpcode::COPY), ResultReg) 345 .addReg(II.ImplicitDefs[0])); 346 } 347 return ResultReg; 348} 349 350unsigned ARMFastISel::FastEmitInst_extractsubreg(MVT RetVT, 351 unsigned Op0, bool Op0IsKill, 352 uint32_t Idx) { 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)); 359 return ResultReg; 360} 361 362// TODO: Don't worry about 64-bit now, but when this is fixed remove the 363// checks from the various callers. 364unsigned ARMFastISel::ARMMoveToFPReg(EVT VT, unsigned SrcReg) { 365 if (VT.getSimpleVT().SimpleTy == MVT::f64) return 0; 366 367 unsigned MoveReg = createResultReg(TLI.getRegClassFor(VT)); 368 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, 369 TII.get(ARM::VMOVRS), MoveReg) 370 .addReg(SrcReg)); 371 return MoveReg; 372} 373 374unsigned ARMFastISel::ARMMoveToIntReg(EVT VT, unsigned SrcReg) { 375 if (VT.getSimpleVT().SimpleTy == MVT::i64) return 0; 376 377 unsigned MoveReg = createResultReg(TLI.getRegClassFor(VT)); 378 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, 379 TII.get(ARM::VMOVSR), MoveReg) 380 .addReg(SrcReg)); 381 return MoveReg; 382} 383 384// 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. 387unsigned ARMFastISel::ARMMaterializeFP(const ConstantFP *CFP, EVT VT) { 388 const APFloat Val = CFP->getValueAPF(); 389 bool is64bit = VT.getSimpleVT().SimpleTy == MVT::f64; 390 391 // 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), 397 DestReg) 398 .addFPImm(CFP)); 399 return DestReg; 400 } 401 402 // Require VFP2 for loading fp constants. 403 if (!Subtarget->hasVFP2()) return false; 404 405 // MachineConstantPool wants an explicit alignment. 406 unsigned Align = TD.getPrefTypeAlignment(CFP->getType()); 407 if (Align == 0) { 408 // TODO: Figure out if this is correct. 409 Align = TD.getTypeAllocSize(CFP->getType()); 410 } 411 unsigned Idx = MCP.getConstantPoolIndex(cast<Constant>(CFP), Align); 412 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT)); 413 unsigned Opc = is64bit ? ARM::VLDRD : ARM::VLDRS; 414 415 // The extra reg is for addrmode5. 416 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), 417 DestReg) 418 .addConstantPoolIndex(Idx) 419 .addReg(0)); 420 return DestReg; 421} 422 423unsigned ARMFastISel::ARMMaterializeInt(const Constant *C, EVT VT) { 424 425 // For now 32-bit only. 426 if (VT.getSimpleVT().SimpleTy != MVT::i32) return false; 427 428 // MachineConstantPool wants an explicit alignment. 429 unsigned Align = TD.getPrefTypeAlignment(C->getType()); 430 if (Align == 0) { 431 // TODO: Figure out if this is correct. 432 Align = TD.getTypeAllocSize(C->getType()); 433 } 434 unsigned Idx = MCP.getConstantPoolIndex(C, Align); 435 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT)); 436 437 if (isThumb) 438 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, 439 TII.get(ARM::t2LDRpci), DestReg) 440 .addConstantPoolIndex(Idx)); 441 else 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)); 447 448 return DestReg; 449} 450 451unsigned ARMFastISel::ARMMaterializeGV(const GlobalValue *GV, EVT VT) { 452 // For now 32-bit only. 453 if (VT.getSimpleVT().SimpleTy != MVT::i32) return 0; 454 455 Reloc::Model RelocM = TM.getRelocationModel(); 456 457 // TODO: No external globals for now. 458 if (Subtarget->GVIsIndirectSymbol(GV, RelocM)) return 0; 459 460 // TODO: Need more magic for ARM PIC. 461 if (!isThumb && (RelocM == Reloc::PIC_)) return 0; 462 463 // MachineConstantPool wants an explicit alignment. 464 unsigned Align = TD.getPrefTypeAlignment(GV->getType()); 465 if (Align == 0) { 466 // TODO: Figure out if this is correct. 467 Align = TD.getTypeAllocSize(GV->getType()); 468 } 469 470 // Grab index. 471 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb() ? 4 : 8); 472 unsigned Id = AFI->createConstPoolEntryUId(); 473 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, Id, 474 ARMCP::CPValue, PCAdj); 475 unsigned Idx = MCP.getConstantPoolIndex(CPV, Align); 476 477 // Load value. 478 MachineInstrBuilder MIB; 479 unsigned DestReg = createResultReg(TLI.getRegClassFor(VT)); 480 if (isThumb) { 481 unsigned Opc = (RelocM != Reloc::PIC_) ? ARM::t2LDRpci : ARM::t2LDRpci_pic; 482 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), DestReg) 483 .addConstantPoolIndex(Idx); 484 if (RelocM == Reloc::PIC_) 485 MIB.addImm(Id); 486 } else { 487 // The extra reg and immediate are for addrmode2. 488 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(ARM::LDRcp), 489 DestReg) 490 .addConstantPoolIndex(Idx) 491 .addReg(0).addImm(0); 492 } 493 AddOptionalDefs(MIB); 494 return DestReg; 495} 496 497unsigned ARMFastISel::TargetMaterializeConstant(const Constant *C) { 498 EVT VT = TLI.getValueType(C->getType(), true); 499 500 // Only handle simple types. 501 if (!VT.isSimple()) return 0; 502 503 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C)) 504 return ARMMaterializeFP(CFP, VT); 505 else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C)) 506 return ARMMaterializeGV(GV, VT); 507 else if (isa<ConstantInt>(C)) 508 return ARMMaterializeInt(C, VT); 509 510 return 0; 511} 512 513unsigned ARMFastISel::TargetMaterializeAlloca(const AllocaInst *AI) { 514 // Don't handle dynamic allocas. 515 if (!FuncInfo.StaticAllocaMap.count(AI)) return 0; 516 517 EVT VT; 518 if (!isTypeLegal(AI->getType(), VT)) return false; 519 520 DenseMap<const AllocaInst*, int>::iterator SI = 521 FuncInfo.StaticAllocaMap.find(AI); 522 523 // This will get lowered later into the correct offsets and registers 524 // via rewriteXFrameIndex. 525 if (SI != FuncInfo.StaticAllocaMap.end()) { 526 TargetRegisterClass* RC = TLI.getRegClassFor(VT); 527 unsigned ResultReg = createResultReg(RC); 528 unsigned Opc = isThumb ? ARM::t2ADDri : ARM::ADDri; 529 AddOptionalDefs(BuildMI(*FuncInfo.MBB, *FuncInfo.InsertPt, DL, 530 TII.get(Opc), ResultReg) 531 .addFrameIndex(SI->second) 532 .addImm(0)); 533 return ResultReg; 534 } 535 536 return 0; 537} 538 539bool ARMFastISel::isTypeLegal(const Type *Ty, EVT &VT) { 540 VT = TLI.getValueType(Ty, true); 541 542 // Only handle simple types. 543 if (VT == MVT::Other || !VT.isSimple()) return false; 544 545 // Handle all legal types, i.e. a register that will directly hold this 546 // value. 547 return TLI.isTypeLegal(VT); 548} 549 550bool ARMFastISel::isLoadTypeLegal(const Type *Ty, EVT &VT) { 551 if (isTypeLegal(Ty, VT)) return true; 552 553 // If this is a type than can be sign or zero-extended to a basic operation 554 // go ahead and accept it now. 555 if (VT == MVT::i8 || VT == MVT::i16) 556 return true; 557 558 return false; 559} 560 561// Computes the Reg+Offset to get to an object. 562bool ARMFastISel::ARMComputeRegOffset(const Value *Obj, unsigned &Reg, 563 int &Offset) { 564 // Some boilerplate from the X86 FastISel. 565 const User *U = NULL; 566 unsigned Opcode = Instruction::UserOp1; 567 if (const Instruction *I = dyn_cast<Instruction>(Obj)) { 568 // Don't walk into other basic blocks; it's possible we haven't 569 // visited them yet, so the instructions may not yet be assigned 570 // virtual registers. 571 if (FuncInfo.MBBMap[I->getParent()] != FuncInfo.MBB) 572 return false; 573 Opcode = I->getOpcode(); 574 U = I; 575 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Obj)) { 576 Opcode = C->getOpcode(); 577 U = C; 578 } 579 580 if (const PointerType *Ty = dyn_cast<PointerType>(Obj->getType())) 581 if (Ty->getAddressSpace() > 255) 582 // Fast instruction selection doesn't support the special 583 // address spaces. 584 return false; 585 586 switch (Opcode) { 587 default: 588 break; 589 case Instruction::Alloca: { 590 assert(false && "Alloca should have been handled earlier!"); 591 return false; 592 } 593 } 594 595 // FIXME: Handle global variables. 596 if (const GlobalValue *GV = dyn_cast<GlobalValue>(Obj)) { 597 (void)GV; 598 return false; 599 } 600 601 // Try to get this in a register if nothing else has worked. 602 Reg = getRegForValue(Obj); 603 if (Reg == 0) return false; 604 605 // Since the offset may be too large for the load instruction 606 // get the reg+offset into a register. 607 // TODO: Verify the additions work, otherwise we'll need to add the 608 // offset instead of 0 to the instructions and do all sorts of operand 609 // munging. 610 // TODO: Optimize this somewhat. 611 if (Offset != 0) { 612 ARMCC::CondCodes Pred = ARMCC::AL; 613 unsigned PredReg = 0; 614 615 if (!isThumb) 616 emitARMRegPlusImmediate(*FuncInfo.MBB, FuncInfo.InsertPt, DL, 617 Reg, Reg, Offset, Pred, PredReg, 618 static_cast<const ARMBaseInstrInfo&>(TII)); 619 else { 620 assert(AFI->isThumb2Function()); 621 emitT2RegPlusImmediate(*FuncInfo.MBB, FuncInfo.InsertPt, DL, 622 Reg, Reg, Offset, Pred, PredReg, 623 static_cast<const ARMBaseInstrInfo&>(TII)); 624 } 625 } 626 return true; 627} 628 629bool ARMFastISel::ARMLoadAlloca(const Instruction *I, EVT VT) { 630 Value *Op0 = I->getOperand(0); 631 632 // Verify it's an alloca. 633 if (const AllocaInst *AI = dyn_cast<AllocaInst>(Op0)) { 634 DenseMap<const AllocaInst*, int>::iterator SI = 635 FuncInfo.StaticAllocaMap.find(AI); 636 637 if (SI != FuncInfo.StaticAllocaMap.end()) { 638 TargetRegisterClass* RC = TLI.getRegClassFor(VT); 639 unsigned ResultReg = createResultReg(RC); 640 TII.loadRegFromStackSlot(*FuncInfo.MBB, *FuncInfo.InsertPt, 641 ResultReg, SI->second, RC, 642 TM.getRegisterInfo()); 643 UpdateValueMap(I, ResultReg); 644 return true; 645 } 646 } 647 return false; 648} 649 650bool ARMFastISel::ARMEmitLoad(EVT VT, unsigned &ResultReg, 651 unsigned Reg, int Offset) { 652 653 assert(VT.isSimple() && "Non-simple types are invalid here!"); 654 unsigned Opc; 655 bool isFloat = false; 656 switch (VT.getSimpleVT().SimpleTy) { 657 default: 658 // This is mostly going to be Neon/vector support. 659 return false; 660 case MVT::i16: 661 Opc = isThumb ? ARM::tLDRH : ARM::LDRH; 662 VT = MVT::i32; 663 break; 664 case MVT::i8: 665 Opc = isThumb ? ARM::tLDRB : ARM::LDRB; 666 VT = MVT::i32; 667 break; 668 case MVT::i32: 669 Opc = isThumb ? ARM::tLDR : ARM::LDR; 670 break; 671 case MVT::f32: 672 Opc = ARM::VLDRS; 673 isFloat = true; 674 break; 675 case MVT::f64: 676 Opc = ARM::VLDRD; 677 isFloat = true; 678 break; 679 } 680 681 ResultReg = createResultReg(TLI.getRegClassFor(VT)); 682 683 // TODO: Fix the Addressing modes so that these can share some code. 684 // Since this is a Thumb1 load this will work in Thumb1 or 2 mode. 685 // The thumb addressing mode has operands swapped from the arm addressing 686 // mode, the floating point one only has two operands. 687 if (isFloat) 688 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, 689 TII.get(Opc), ResultReg) 690 .addReg(Reg).addImm(Offset)); 691 else if (isThumb) 692 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, 693 TII.get(Opc), ResultReg) 694 .addReg(Reg).addImm(Offset).addReg(0)); 695 else 696 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, 697 TII.get(Opc), ResultReg) 698 .addReg(Reg).addReg(0).addImm(Offset)); 699 return true; 700} 701 702bool ARMFastISel::SelectLoad(const Instruction *I) { 703 // Verify we have a legal type before going any further. 704 EVT VT; 705 if (!isLoadTypeLegal(I->getType(), VT)) 706 return false; 707 708 // If we're an alloca we know we have a frame index and can emit the load 709 // directly in short order. 710 if (ARMLoadAlloca(I, VT)) 711 return true; 712 713 // Our register and offset with innocuous defaults. 714 unsigned Reg = 0; 715 int Offset = 0; 716 717 // See if we can handle this as Reg + Offset 718 if (!ARMComputeRegOffset(I->getOperand(0), Reg, Offset)) 719 return false; 720 721 unsigned ResultReg; 722 if (!ARMEmitLoad(VT, ResultReg, Reg, Offset /* 0 */)) return false; 723 724 UpdateValueMap(I, ResultReg); 725 return true; 726} 727 728bool ARMFastISel::ARMStoreAlloca(const Instruction *I, unsigned SrcReg, EVT VT){ 729 Value *Op1 = I->getOperand(1); 730 731 // Verify it's an alloca. 732 if (const AllocaInst *AI = dyn_cast<AllocaInst>(Op1)) { 733 DenseMap<const AllocaInst*, int>::iterator SI = 734 FuncInfo.StaticAllocaMap.find(AI); 735 736 if (SI != FuncInfo.StaticAllocaMap.end()) { 737 TargetRegisterClass* RC = TLI.getRegClassFor(VT); 738 assert(SrcReg != 0 && "Nothing to store!"); 739 TII.storeRegToStackSlot(*FuncInfo.MBB, *FuncInfo.InsertPt, 740 SrcReg, true /*isKill*/, SI->second, RC, 741 TM.getRegisterInfo()); 742 return true; 743 } 744 } 745 return false; 746} 747 748bool ARMFastISel::ARMEmitStore(EVT VT, unsigned SrcReg, 749 unsigned DstReg, int Offset) { 750 unsigned StrOpc; 751 bool isFloat = false; 752 switch (VT.getSimpleVT().SimpleTy) { 753 default: return false; 754 case MVT::i1: 755 case MVT::i8: StrOpc = isThumb ? ARM::tSTRB : ARM::STRB; break; 756 case MVT::i16: StrOpc = isThumb ? ARM::tSTRH : ARM::STRH; break; 757 case MVT::i32: StrOpc = isThumb ? ARM::tSTR : ARM::STR; break; 758 case MVT::f32: 759 if (!Subtarget->hasVFP2()) return false; 760 StrOpc = ARM::VSTRS; 761 isFloat = true; 762 break; 763 case MVT::f64: 764 if (!Subtarget->hasVFP2()) return false; 765 StrOpc = ARM::VSTRD; 766 isFloat = true; 767 break; 768 } 769 770 // The thumb addressing mode has operands swapped from the arm addressing 771 // mode, the floating point one only has two operands. 772 if (isFloat) 773 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, 774 TII.get(StrOpc)) 775 .addReg(SrcReg).addReg(DstReg).addImm(Offset)); 776 else if (isThumb) 777 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, 778 TII.get(StrOpc)) 779 .addReg(SrcReg).addReg(DstReg).addImm(Offset).addReg(0)); 780 781 else 782 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, 783 TII.get(StrOpc)) 784 .addReg(SrcReg).addReg(DstReg).addReg(0).addImm(Offset)); 785 786 return true; 787} 788 789bool ARMFastISel::SelectStore(const Instruction *I) { 790 Value *Op0 = I->getOperand(0); 791 unsigned SrcReg = 0; 792 793 // Yay type legalization 794 EVT VT; 795 if (!isLoadTypeLegal(I->getOperand(0)->getType(), VT)) 796 return false; 797 798 // Get the value to be stored into a register. 799 SrcReg = getRegForValue(Op0); 800 if (SrcReg == 0) 801 return false; 802 803 // If we're an alloca we know we have a frame index and can emit the store 804 // quickly. 805 if (ARMStoreAlloca(I, SrcReg, VT)) 806 return true; 807 808 // Our register and offset with innocuous defaults. 809 unsigned Reg = 0; 810 int Offset = 0; 811 812 // See if we can handle this as Reg + Offset 813 if (!ARMComputeRegOffset(I->getOperand(1), Reg, Offset)) 814 return false; 815 816 if (!ARMEmitStore(VT, SrcReg, Reg, Offset /* 0 */)) return false; 817 818 return true; 819} 820 821static ARMCC::CondCodes getComparePred(CmpInst::Predicate Pred) { 822 switch (Pred) { 823 // Needs two compares... 824 case CmpInst::FCMP_ONE: 825 case CmpInst::FCMP_UEQ: 826 default: 827 assert(false && "Unhandled CmpInst::Predicate!"); 828 return ARMCC::AL; 829 case CmpInst::ICMP_EQ: 830 case CmpInst::FCMP_OEQ: 831 return ARMCC::EQ; 832 case CmpInst::ICMP_SGT: 833 case CmpInst::FCMP_OGT: 834 return ARMCC::GT; 835 case CmpInst::ICMP_SGE: 836 case CmpInst::FCMP_OGE: 837 return ARMCC::GE; 838 case CmpInst::ICMP_UGT: 839 case CmpInst::FCMP_UGT: 840 return ARMCC::HI; 841 case CmpInst::FCMP_OLT: 842 return ARMCC::MI; 843 case CmpInst::ICMP_ULE: 844 case CmpInst::FCMP_OLE: 845 return ARMCC::LS; 846 case CmpInst::FCMP_ORD: 847 return ARMCC::VC; 848 case CmpInst::FCMP_UNO: 849 return ARMCC::VS; 850 case CmpInst::FCMP_UGE: 851 return ARMCC::PL; 852 case CmpInst::ICMP_SLT: 853 case CmpInst::FCMP_ULT: 854 return ARMCC::LT; 855 case CmpInst::ICMP_SLE: 856 case CmpInst::FCMP_ULE: 857 return ARMCC::LE; 858 case CmpInst::FCMP_UNE: 859 case CmpInst::ICMP_NE: 860 return ARMCC::NE; 861 case CmpInst::ICMP_UGE: 862 return ARMCC::HS; 863 case CmpInst::ICMP_ULT: 864 return ARMCC::LO; 865 } 866} 867 868bool ARMFastISel::SelectBranch(const Instruction *I) { 869 const BranchInst *BI = cast<BranchInst>(I); 870 MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)]; 871 MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)]; 872 873 // Simple branch support. 874 // TODO: Try to avoid the re-computation in some places. 875 unsigned CondReg = getRegForValue(BI->getCondition()); 876 if (CondReg == 0) return false; 877 878 // Re-set the flags just in case. 879 unsigned CmpOpc = isThumb ? ARM::t2CMPri : ARM::CMPri; 880 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc)) 881 .addReg(CondReg).addImm(1)); 882 883 unsigned BrOpc = isThumb ? ARM::t2Bcc : ARM::Bcc; 884 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc)) 885 .addMBB(TBB).addImm(ARMCC::EQ).addReg(ARM::CPSR); 886 FastEmitBranch(FBB, DL); 887 FuncInfo.MBB->addSuccessor(TBB); 888 return true; 889} 890 891bool ARMFastISel::SelectCmp(const Instruction *I) { 892 const CmpInst *CI = cast<CmpInst>(I); 893 894 EVT VT; 895 const Type *Ty = CI->getOperand(0)->getType(); 896 if (!isTypeLegal(Ty, VT)) 897 return false; 898 899 bool isFloat = (Ty->isDoubleTy() || Ty->isFloatTy()); 900 if (isFloat && !Subtarget->hasVFP2()) 901 return false; 902 903 unsigned CmpOpc; 904 unsigned CondReg; 905 switch (VT.getSimpleVT().SimpleTy) { 906 default: return false; 907 // TODO: Verify compares. 908 case MVT::f32: 909 CmpOpc = ARM::VCMPES; 910 CondReg = ARM::FPSCR; 911 break; 912 case MVT::f64: 913 CmpOpc = ARM::VCMPED; 914 CondReg = ARM::FPSCR; 915 break; 916 case MVT::i32: 917 CmpOpc = isThumb ? ARM::t2CMPrr : ARM::CMPrr; 918 CondReg = ARM::CPSR; 919 break; 920 } 921 922 // Get the compare predicate. 923 ARMCC::CondCodes ARMPred = getComparePred(CI->getPredicate()); 924 925 // We may not handle every CC for now. 926 if (ARMPred == ARMCC::AL) return false; 927 928 unsigned Arg1 = getRegForValue(CI->getOperand(0)); 929 if (Arg1 == 0) return false; 930 931 unsigned Arg2 = getRegForValue(CI->getOperand(1)); 932 if (Arg2 == 0) return false; 933 934 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc)) 935 .addReg(Arg1).addReg(Arg2)); 936 937 // For floating point we need to move the result to a comparison register 938 // that we can then use for branches. 939 if (isFloat) 940 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, 941 TII.get(ARM::FMSTAT))); 942 943 // Now set a register based on the comparison. Explicitly set the predicates 944 // here. 945 unsigned MovCCOpc = isThumb ? ARM::tMOVCCi : ARM::MOVCCi; 946 unsigned DestReg = createResultReg(ARM::GPRRegisterClass); 947 Constant *Zero 948 = ConstantInt::get(Type::getInt32Ty(*Context), 0); 949 unsigned ZeroReg = TargetMaterializeConstant(Zero); 950 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), DestReg) 951 .addReg(ZeroReg).addImm(1) 952 .addImm(ARMPred).addReg(CondReg); 953 954 UpdateValueMap(I, DestReg); 955 return true; 956} 957 958bool ARMFastISel::SelectFPExt(const Instruction *I) { 959 // Make sure we have VFP and that we're extending float to double. 960 if (!Subtarget->hasVFP2()) return false; 961 962 Value *V = I->getOperand(0); 963 if (!I->getType()->isDoubleTy() || 964 !V->getType()->isFloatTy()) return false; 965 966 unsigned Op = getRegForValue(V); 967 if (Op == 0) return false; 968 969 unsigned Result = createResultReg(ARM::DPRRegisterClass); 970 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, 971 TII.get(ARM::VCVTDS), Result) 972 .addReg(Op)); 973 UpdateValueMap(I, Result); 974 return true; 975} 976 977bool ARMFastISel::SelectFPTrunc(const Instruction *I) { 978 // Make sure we have VFP and that we're truncating double to float. 979 if (!Subtarget->hasVFP2()) return false; 980 981 Value *V = I->getOperand(0); 982 if (!(I->getType()->isFloatTy() && 983 V->getType()->isDoubleTy())) return false; 984 985 unsigned Op = getRegForValue(V); 986 if (Op == 0) return false; 987 988 unsigned Result = createResultReg(ARM::SPRRegisterClass); 989 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, 990 TII.get(ARM::VCVTSD), Result) 991 .addReg(Op)); 992 UpdateValueMap(I, Result); 993 return true; 994} 995 996bool ARMFastISel::SelectSIToFP(const Instruction *I) { 997 // Make sure we have VFP. 998 if (!Subtarget->hasVFP2()) return false; 999 1000 EVT DstVT; 1001 const Type *Ty = I->getType(); 1002 if (!isTypeLegal(Ty, DstVT)) 1003 return false; 1004 1005 unsigned Op = getRegForValue(I->getOperand(0)); 1006 if (Op == 0) return false; 1007 1008 // The conversion routine works on fp-reg to fp-reg and the operand above 1009 // was an integer, move it to the fp registers if possible. 1010 unsigned FP = ARMMoveToFPReg(MVT::f32, Op); 1011 if (FP == 0) return false; 1012 1013 unsigned Opc; 1014 if (Ty->isFloatTy()) Opc = ARM::VSITOS; 1015 else if (Ty->isDoubleTy()) Opc = ARM::VSITOD; 1016 else return 0; 1017 1018 unsigned ResultReg = createResultReg(TLI.getRegClassFor(DstVT)); 1019 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), 1020 ResultReg) 1021 .addReg(FP)); 1022 UpdateValueMap(I, ResultReg); 1023 return true; 1024} 1025 1026bool ARMFastISel::SelectFPToSI(const Instruction *I) { 1027 // Make sure we have VFP. 1028 if (!Subtarget->hasVFP2()) return false; 1029 1030 EVT DstVT; 1031 const Type *RetTy = I->getType(); 1032 if (!isTypeLegal(RetTy, DstVT)) 1033 return false; 1034 1035 unsigned Op = getRegForValue(I->getOperand(0)); 1036 if (Op == 0) return false; 1037 1038 unsigned Opc; 1039 const Type *OpTy = I->getOperand(0)->getType(); 1040 if (OpTy->isFloatTy()) Opc = ARM::VTOSIZS; 1041 else if (OpTy->isDoubleTy()) Opc = ARM::VTOSIZD; 1042 else return 0; 1043 1044 // f64->s32 or f32->s32 both need an intermediate f32 reg. 1045 unsigned ResultReg = createResultReg(TLI.getRegClassFor(MVT::f32)); 1046 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), 1047 ResultReg) 1048 .addReg(Op)); 1049 1050 // This result needs to be in an integer register, but the conversion only 1051 // takes place in fp-regs. 1052 unsigned IntReg = ARMMoveToIntReg(DstVT, ResultReg); 1053 if (IntReg == 0) return false; 1054 1055 UpdateValueMap(I, IntReg); 1056 return true; 1057} 1058 1059bool ARMFastISel::SelectSDiv(const Instruction *I) { 1060 EVT VT; 1061 const Type *Ty = I->getType(); 1062 if (!isTypeLegal(Ty, VT)) 1063 return false; 1064 1065 // If we have integer div support we should have selected this automagically. 1066 // In case we have a real miss go ahead and return false and we'll pick 1067 // it up later. 1068 if (Subtarget->hasDivide()) return false; 1069 1070 // Otherwise emit a libcall. 1071 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL; 1072 if (VT == MVT::i16) 1073 LC = RTLIB::SDIV_I16; 1074 else if (VT == MVT::i32) 1075 LC = RTLIB::SDIV_I32; 1076 else if (VT == MVT::i64) 1077 LC = RTLIB::SDIV_I64; 1078 else if (VT == MVT::i128) 1079 LC = RTLIB::SDIV_I128; 1080 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SDIV!"); 1081 1082 return ARMEmitLibcall(I, LC); 1083} 1084 1085bool ARMFastISel::SelectBinaryOp(const Instruction *I, unsigned ISDOpcode) { 1086 EVT VT = TLI.getValueType(I->getType(), true); 1087 1088 // We can get here in the case when we want to use NEON for our fp 1089 // operations, but can't figure out how to. Just use the vfp instructions 1090 // if we have them. 1091 // FIXME: It'd be nice to use NEON instructions. 1092 const Type *Ty = I->getType(); 1093 bool isFloat = (Ty->isDoubleTy() || Ty->isFloatTy()); 1094 if (isFloat && !Subtarget->hasVFP2()) 1095 return false; 1096 1097 unsigned Op1 = getRegForValue(I->getOperand(0)); 1098 if (Op1 == 0) return false; 1099 1100 unsigned Op2 = getRegForValue(I->getOperand(1)); 1101 if (Op2 == 0) return false; 1102 1103 unsigned Opc; 1104 bool is64bit = VT.getSimpleVT().SimpleTy == MVT::f64 || 1105 VT.getSimpleVT().SimpleTy == MVT::i64; 1106 switch (ISDOpcode) { 1107 default: return false; 1108 case ISD::FADD: 1109 Opc = is64bit ? ARM::VADDD : ARM::VADDS; 1110 break; 1111 case ISD::FSUB: 1112 Opc = is64bit ? ARM::VSUBD : ARM::VSUBS; 1113 break; 1114 case ISD::FMUL: 1115 Opc = is64bit ? ARM::VMULD : ARM::VMULS; 1116 break; 1117 } 1118 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT)); 1119 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, 1120 TII.get(Opc), ResultReg) 1121 .addReg(Op1).addReg(Op2)); 1122 UpdateValueMap(I, ResultReg); 1123 return true; 1124} 1125 1126// Call Handling Code 1127 1128// This is largely taken directly from CCAssignFnForNode - we don't support 1129// varargs in FastISel so that part has been removed. 1130// TODO: We may not support all of this. 1131CCAssignFn *ARMFastISel::CCAssignFnForCall(CallingConv::ID CC, bool Return) { 1132 switch (CC) { 1133 default: 1134 llvm_unreachable("Unsupported calling convention"); 1135 case CallingConv::C: 1136 case CallingConv::Fast: 1137 // Use target triple & subtarget features to do actual dispatch. 1138 if (Subtarget->isAAPCS_ABI()) { 1139 if (Subtarget->hasVFP2() && 1140 FloatABIType == FloatABI::Hard) 1141 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP); 1142 else 1143 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS); 1144 } else 1145 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS); 1146 case CallingConv::ARM_AAPCS_VFP: 1147 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP); 1148 case CallingConv::ARM_AAPCS: 1149 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS); 1150 case CallingConv::ARM_APCS: 1151 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS); 1152 } 1153} 1154 1155bool ARMFastISel::ProcessCallArgs(SmallVectorImpl<Value*> &Args, 1156 SmallVectorImpl<unsigned> &ArgRegs, 1157 SmallVectorImpl<EVT> &ArgVTs, 1158 SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags, 1159 SmallVectorImpl<unsigned> &RegArgs, 1160 CallingConv::ID CC, 1161 unsigned &NumBytes) { 1162 SmallVector<CCValAssign, 16> ArgLocs; 1163 CCState CCInfo(CC, false, TM, ArgLocs, *Context); 1164 CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CCAssignFnForCall(CC, false)); 1165 1166 // Get a count of how many bytes are to be pushed on the stack. 1167 NumBytes = CCInfo.getNextStackOffset(); 1168 1169 // Issue CALLSEQ_START 1170 unsigned AdjStackDown = TM.getRegisterInfo()->getCallFrameSetupOpcode(); 1171 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(AdjStackDown)) 1172 .addImm(NumBytes); 1173 1174 // Process the args. 1175 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 1176 CCValAssign &VA = ArgLocs[i]; 1177 unsigned Arg = ArgRegs[VA.getValNo()]; 1178 EVT ArgVT = ArgVTs[VA.getValNo()]; 1179 1180 // Handle arg promotion, etc. 1181 switch (VA.getLocInfo()) { 1182 case CCValAssign::Full: break; 1183 default: 1184 assert(false && "Handle arg promotion."); 1185 return false; 1186 } 1187 1188 // Now copy/store arg to correct locations. 1189 if (VA.isRegLoc()) { 1190 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), 1191 VA.getLocReg()) 1192 .addReg(Arg); 1193 RegArgs.push_back(VA.getLocReg()); 1194 } else { 1195 // Need to store 1196 return false; 1197 } 1198 } 1199 1200 return true; 1201} 1202 1203bool ARMFastISel::FinishCall(EVT RetVT, SmallVectorImpl<unsigned> &UsedRegs, 1204 const Instruction *I, CallingConv::ID CC, 1205 unsigned &NumBytes) { 1206 // Issue CALLSEQ_END 1207 unsigned AdjStackUp = TM.getRegisterInfo()->getCallFrameDestroyOpcode(); 1208 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(AdjStackUp)) 1209 .addImm(NumBytes).addImm(0); 1210 1211 // Now the return value. 1212 if (RetVT.getSimpleVT().SimpleTy != MVT::isVoid) { 1213 SmallVector<CCValAssign, 16> RVLocs; 1214 CCState CCInfo(CC, false, TM, RVLocs, *Context); 1215 CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC, true)); 1216 1217 // Copy all of the result registers out of their specified physreg. 1218 if (RVLocs.size() == 2 && RetVT.getSimpleVT().SimpleTy == MVT::f64) { 1219 // For this move we copy into two registers and then move into the 1220 // double fp reg we want. 1221 // TODO: Are the copies necessary? 1222 TargetRegisterClass *CopyRC = TLI.getRegClassFor(MVT::i32); 1223 unsigned Copy1 = createResultReg(CopyRC); 1224 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), 1225 Copy1).addReg(RVLocs[0].getLocReg()); 1226 UsedRegs.push_back(RVLocs[0].getLocReg()); 1227 1228 unsigned Copy2 = createResultReg(CopyRC); 1229 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), 1230 Copy2).addReg(RVLocs[1].getLocReg()); 1231 UsedRegs.push_back(RVLocs[1].getLocReg()); 1232 1233 EVT DestVT = RVLocs[0].getValVT(); 1234 TargetRegisterClass* DstRC = TLI.getRegClassFor(DestVT); 1235 unsigned ResultReg = createResultReg(DstRC); 1236 AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, 1237 TII.get(ARM::VMOVDRR), ResultReg) 1238 .addReg(Copy1).addReg(Copy2)); 1239 1240 // Finally update the result. 1241 UpdateValueMap(I, ResultReg); 1242 } else { 1243 assert(RVLocs.size() == 1 && "Can't handle non-double multi-reg retvals!"); 1244 EVT CopyVT = RVLocs[0].getValVT(); 1245 TargetRegisterClass* DstRC = TLI.getRegClassFor(CopyVT); 1246 1247 unsigned ResultReg = createResultReg(DstRC); 1248 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), 1249 ResultReg).addReg(RVLocs[0].getLocReg()); 1250 UsedRegs.push_back(RVLocs[0].getLocReg()); 1251 1252 // Finally update the result. 1253 UpdateValueMap(I, ResultReg); 1254 } 1255 } 1256 1257 return true; 1258} 1259 1260// A quick function that will emit a call for a named libcall in F with the 1261// vector of passed arguments for the Instruction in I. We can assume that we 1262// can emit a call for any libcall we can produce. This is an abridged version 1263// of the full call infrastructure since we won't need to worry about things 1264// like computed function pointers or strange arguments at call sites. 1265// TODO: Try to unify this and the normal call bits for ARM, then try to unify 1266// with X86. 1267bool ARMFastISel::ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call) { 1268 CallingConv::ID CC = TLI.getLibcallCallingConv(Call); 1269 1270 // Handle *simple* calls for now. 1271 const Type *RetTy = I->getType(); 1272 EVT RetVT; 1273 if (RetTy->isVoidTy()) 1274 RetVT = MVT::isVoid; 1275 else if (!isTypeLegal(RetTy, RetVT)) 1276 return false; 1277 1278 // For now we're using BLX etc on the assumption that we have v5t ops. 1279 if (!Subtarget->hasV5TOps()) return false; 1280 1281 // Set up the argument vectors. 1282 SmallVector<Value*, 8> Args; 1283 SmallVector<unsigned, 8> ArgRegs; 1284 SmallVector<EVT, 8> ArgVTs; 1285 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags; 1286 Args.reserve(I->getNumOperands()); 1287 ArgRegs.reserve(I->getNumOperands()); 1288 ArgVTs.reserve(I->getNumOperands()); 1289 ArgFlags.reserve(I->getNumOperands()); 1290 for (unsigned i = 0; i < I->getNumOperands(); ++i) { 1291 Value *Op = I->getOperand(i); 1292 unsigned Arg = getRegForValue(Op); 1293 if (Arg == 0) return false; 1294 1295 const Type *ArgTy = Op->getType(); 1296 EVT ArgVT; 1297 if (!isTypeLegal(ArgTy, ArgVT)) return false; 1298 1299 ISD::ArgFlagsTy Flags; 1300 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy); 1301 Flags.setOrigAlign(OriginalAlignment); 1302 1303 Args.push_back(Op); 1304 ArgRegs.push_back(Arg); 1305 ArgVTs.push_back(ArgVT); 1306 ArgFlags.push_back(Flags); 1307 } 1308 1309 // Handle the arguments now that we've gotten them. 1310 SmallVector<unsigned, 4> RegArgs; 1311 unsigned NumBytes; 1312 if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes)) 1313 return false; 1314 1315 // Issue the call, BLXr9 for darwin, BLX otherwise. This uses V5 ops. 1316 // TODO: Turn this into the table of arm call ops. 1317 MachineInstrBuilder MIB; 1318 unsigned CallOpc; 1319 if(isThumb) 1320 CallOpc = Subtarget->isTargetDarwin() ? ARM::tBLXi_r9 : ARM::tBLXi; 1321 else 1322 CallOpc = Subtarget->isTargetDarwin() ? ARM::BLr9 : ARM::BL; 1323 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CallOpc)) 1324 .addExternalSymbol(TLI.getLibcallName(Call)); 1325 1326 // Add implicit physical register uses to the call. 1327 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i) 1328 MIB.addReg(RegArgs[i]); 1329 1330 // Finish off the call including any return values. 1331 SmallVector<unsigned, 4> UsedRegs; 1332 if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false; 1333 1334 // Set all unused physreg defs as dead. 1335 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI); 1336 1337 return true; 1338} 1339 1340bool ARMFastISel::SelectCall(const Instruction *I) { 1341 const CallInst *CI = cast<CallInst>(I); 1342 const Value *Callee = CI->getCalledValue(); 1343 1344 // Can't handle inline asm or worry about intrinsics yet. 1345 if (isa<InlineAsm>(Callee) || isa<IntrinsicInst>(CI)) return false; 1346 1347 // Only handle global variable Callees that are direct calls. 1348 const GlobalValue *GV = dyn_cast<GlobalValue>(Callee); 1349 if (!GV || Subtarget->GVIsIndirectSymbol(GV, TM.getRelocationModel())) 1350 return false; 1351 1352 // Check the calling convention. 1353 ImmutableCallSite CS(CI); 1354 CallingConv::ID CC = CS.getCallingConv(); 1355 // TODO: Avoid some calling conventions? 1356 if (CC != CallingConv::C) { 1357 errs() << "Can't handle calling convention: " << CC << "\n"; 1358 return false; 1359 } 1360 1361 // Let SDISel handle vararg functions. 1362 const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType()); 1363 const FunctionType *FTy = cast<FunctionType>(PT->getElementType()); 1364 if (FTy->isVarArg()) 1365 return false; 1366 1367 // Handle *simple* calls for now. 1368 const Type *RetTy = I->getType(); 1369 EVT RetVT; 1370 if (RetTy->isVoidTy()) 1371 RetVT = MVT::isVoid; 1372 else if (!isTypeLegal(RetTy, RetVT)) 1373 return false; 1374 1375 // For now we're using BLX etc on the assumption that we have v5t ops. 1376 // TODO: Maybe? 1377 if (!Subtarget->hasV5TOps()) return false; 1378 1379 // Set up the argument vectors. 1380 SmallVector<Value*, 8> Args; 1381 SmallVector<unsigned, 8> ArgRegs; 1382 SmallVector<EVT, 8> ArgVTs; 1383 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags; 1384 Args.reserve(CS.arg_size()); 1385 ArgRegs.reserve(CS.arg_size()); 1386 ArgVTs.reserve(CS.arg_size()); 1387 ArgFlags.reserve(CS.arg_size()); 1388 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end(); 1389 i != e; ++i) { 1390 unsigned Arg = getRegForValue(*i); 1391 1392 if (Arg == 0) 1393 return false; 1394 ISD::ArgFlagsTy Flags; 1395 unsigned AttrInd = i - CS.arg_begin() + 1; 1396 if (CS.paramHasAttr(AttrInd, Attribute::SExt)) 1397 Flags.setSExt(); 1398 if (CS.paramHasAttr(AttrInd, Attribute::ZExt)) 1399 Flags.setZExt(); 1400 1401 // FIXME: Only handle *easy* calls for now. 1402 if (CS.paramHasAttr(AttrInd, Attribute::InReg) || 1403 CS.paramHasAttr(AttrInd, Attribute::StructRet) || 1404 CS.paramHasAttr(AttrInd, Attribute::Nest) || 1405 CS.paramHasAttr(AttrInd, Attribute::ByVal)) 1406 return false; 1407 1408 const Type *ArgTy = (*i)->getType(); 1409 EVT ArgVT; 1410 if (!isTypeLegal(ArgTy, ArgVT)) 1411 return false; 1412 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy); 1413 Flags.setOrigAlign(OriginalAlignment); 1414 1415 Args.push_back(*i); 1416 ArgRegs.push_back(Arg); 1417 ArgVTs.push_back(ArgVT); 1418 ArgFlags.push_back(Flags); 1419 } 1420 1421 // Handle the arguments now that we've gotten them. 1422 SmallVector<unsigned, 4> RegArgs; 1423 unsigned NumBytes; 1424 if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes)) 1425 return false; 1426 1427 // Issue the call, BLXr9 for darwin, BLX otherwise. This uses V5 ops. 1428 // TODO: Turn this into the table of arm call ops. 1429 MachineInstrBuilder MIB; 1430 unsigned CallOpc; 1431 if(isThumb) 1432 CallOpc = Subtarget->isTargetDarwin() ? ARM::tBLXi_r9 : ARM::tBLXi; 1433 else 1434 CallOpc = Subtarget->isTargetDarwin() ? ARM::BLr9 : ARM::BL; 1435 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CallOpc)) 1436 .addGlobalAddress(GV, 0, 0); 1437 1438 // Add implicit physical register uses to the call. 1439 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i) 1440 MIB.addReg(RegArgs[i]); 1441 1442 // Finish off the call including any return values. 1443 SmallVector<unsigned, 4> UsedRegs; 1444 if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false; 1445 1446 // Set all unused physreg defs as dead. 1447 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI); 1448 1449 return true; 1450 1451} 1452 1453// TODO: SoftFP support. 1454bool ARMFastISel::TargetSelectInstruction(const Instruction *I) { 1455 // No Thumb-1 for now. 1456 if (isThumb && !AFI->isThumb2Function()) return false; 1457 1458 switch (I->getOpcode()) { 1459 case Instruction::Load: 1460 return SelectLoad(I); 1461 case Instruction::Store: 1462 return SelectStore(I); 1463 case Instruction::Br: 1464 return SelectBranch(I); 1465 case Instruction::ICmp: 1466 case Instruction::FCmp: 1467 return SelectCmp(I); 1468 case Instruction::FPExt: 1469 return SelectFPExt(I); 1470 case Instruction::FPTrunc: 1471 return SelectFPTrunc(I); 1472 case Instruction::SIToFP: 1473 return SelectSIToFP(I); 1474 case Instruction::FPToSI: 1475 return SelectFPToSI(I); 1476 case Instruction::FAdd: 1477 return SelectBinaryOp(I, ISD::FADD); 1478 case Instruction::FSub: 1479 return SelectBinaryOp(I, ISD::FSUB); 1480 case Instruction::FMul: 1481 return SelectBinaryOp(I, ISD::FMUL); 1482 case Instruction::SDiv: 1483 return SelectSDiv(I); 1484 case Instruction::Call: 1485 return SelectCall(I); 1486 default: break; 1487 } 1488 return false; 1489} 1490 1491namespace llvm { 1492 llvm::FastISel *ARM::createFastISel(FunctionLoweringInfo &funcInfo) { 1493 if (EnableARMFastISel) return new ARMFastISel(funcInfo); 1494 return 0; 1495 } 1496} 1497