1 //===-- FastISel.cpp - Implementation of the FastISel class ---------------===// 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 contains the implementation of the FastISel class. 11 // 12 // "Fast" instruction selection is designed to emit very poor code quickly. 13 // Also, it is not designed to be able to do much lowering, so most illegal 14 // types (e.g. i64 on 32-bit targets) and operations are not supported. It is 15 // also not intended to be able to do much optimization, except in a few cases 16 // where doing optimizations reduces overall compile time. For example, folding 17 // constants into immediate fields is often done, because it's cheap and it 18 // reduces the number of instructions later phases have to examine. 19 // 20 // "Fast" instruction selection is able to fail gracefully and transfer 21 // control to the SelectionDAG selector for operations that it doesn't 22 // support. In many cases, this allows us to avoid duplicating a lot of 23 // the complicated lowering logic that SelectionDAG currently has. 24 // 25 // The intended use for "fast" instruction selection is "-O0" mode 26 // compilation, where the quality of the generated code is irrelevant when 27 // weighed against the speed at which the code can be generated. Also, 28 // at -O0, the LLVM optimizers are not running, and this makes the 29 // compile time of codegen a much higher portion of the overall compile 30 // time. Despite its limitations, "fast" instruction selection is able to 31 // handle enough code on its own to provide noticeable overall speedups 32 // in -O0 compiles. 33 // 34 // Basic operations are supported in a target-independent way, by reading 35 // the same instruction descriptions that the SelectionDAG selector reads, 36 // and identifying simple arithmetic operations that can be directly selected 37 // from simple operators. More complicated operations currently require 38 // target-specific code. 39 // 40 //===----------------------------------------------------------------------===// 41 42 #include "llvm/Function.h" 43 #include "llvm/GlobalVariable.h" 44 #include "llvm/Instructions.h" 45 #include "llvm/IntrinsicInst.h" 46 #include "llvm/Operator.h" 47 #include "llvm/CodeGen/Analysis.h" 48 #include "llvm/CodeGen/FastISel.h" 49 #include "llvm/CodeGen/FunctionLoweringInfo.h" 50 #include "llvm/CodeGen/MachineInstrBuilder.h" 51 #include "llvm/CodeGen/MachineModuleInfo.h" 52 #include "llvm/CodeGen/MachineRegisterInfo.h" 53 #include "llvm/Analysis/DebugInfo.h" 54 #include "llvm/Analysis/Loads.h" 55 #include "llvm/Target/TargetData.h" 56 #include "llvm/Target/TargetInstrInfo.h" 57 #include "llvm/Target/TargetLowering.h" 58 #include "llvm/Target/TargetMachine.h" 59 #include "llvm/Support/ErrorHandling.h" 60 #include "llvm/Support/Debug.h" 61 using namespace llvm; 62 63 /// startNewBlock - Set the current block to which generated machine 64 /// instructions will be appended, and clear the local CSE map. 65 /// 66 void FastISel::startNewBlock() { 67 LocalValueMap.clear(); 68 69 EmitStartPt = 0; 70 71 // Advance the emit start point past any EH_LABEL instructions. 72 MachineBasicBlock::iterator 73 I = FuncInfo.MBB->begin(), E = FuncInfo.MBB->end(); 74 while (I != E && I->getOpcode() == TargetOpcode::EH_LABEL) { 75 EmitStartPt = I; 76 ++I; 77 } 78 LastLocalValue = EmitStartPt; 79 } 80 81 void FastISel::flushLocalValueMap() { 82 LocalValueMap.clear(); 83 LastLocalValue = EmitStartPt; 84 recomputeInsertPt(); 85 } 86 87 bool FastISel::hasTrivialKill(const Value *V) const { 88 // Don't consider constants or arguments to have trivial kills. 89 const Instruction *I = dyn_cast<Instruction>(V); 90 if (!I) 91 return false; 92 93 // No-op casts are trivially coalesced by fast-isel. 94 if (const CastInst *Cast = dyn_cast<CastInst>(I)) 95 if (Cast->isNoopCast(TD.getIntPtrType(Cast->getContext())) && 96 !hasTrivialKill(Cast->getOperand(0))) 97 return false; 98 99 // Only instructions with a single use in the same basic block are considered 100 // to have trivial kills. 101 return I->hasOneUse() && 102 !(I->getOpcode() == Instruction::BitCast || 103 I->getOpcode() == Instruction::PtrToInt || 104 I->getOpcode() == Instruction::IntToPtr) && 105 cast<Instruction>(*I->use_begin())->getParent() == I->getParent(); 106 } 107 108 unsigned FastISel::getRegForValue(const Value *V) { 109 EVT RealVT = TLI.getValueType(V->getType(), /*AllowUnknown=*/true); 110 // Don't handle non-simple values in FastISel. 111 if (!RealVT.isSimple()) 112 return 0; 113 114 // Ignore illegal types. We must do this before looking up the value 115 // in ValueMap because Arguments are given virtual registers regardless 116 // of whether FastISel can handle them. 117 MVT VT = RealVT.getSimpleVT(); 118 if (!TLI.isTypeLegal(VT)) { 119 // Handle integer promotions, though, because they're common and easy. 120 if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16) 121 VT = TLI.getTypeToTransformTo(V->getContext(), VT).getSimpleVT(); 122 else 123 return 0; 124 } 125 126 // Look up the value to see if we already have a register for it. We 127 // cache values defined by Instructions across blocks, and other values 128 // only locally. This is because Instructions already have the SSA 129 // def-dominates-use requirement enforced. 130 DenseMap<const Value *, unsigned>::iterator I = FuncInfo.ValueMap.find(V); 131 if (I != FuncInfo.ValueMap.end()) 132 return I->second; 133 134 unsigned Reg = LocalValueMap[V]; 135 if (Reg != 0) 136 return Reg; 137 138 // In bottom-up mode, just create the virtual register which will be used 139 // to hold the value. It will be materialized later. 140 if (isa<Instruction>(V) && 141 (!isa<AllocaInst>(V) || 142 !FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(V)))) 143 return FuncInfo.InitializeRegForValue(V); 144 145 SavePoint SaveInsertPt = enterLocalValueArea(); 146 147 // Materialize the value in a register. Emit any instructions in the 148 // local value area. 149 Reg = materializeRegForValue(V, VT); 150 151 leaveLocalValueArea(SaveInsertPt); 152 153 return Reg; 154 } 155 156 /// materializeRegForValue - Helper for getRegForValue. This function is 157 /// called when the value isn't already available in a register and must 158 /// be materialized with new instructions. 159 unsigned FastISel::materializeRegForValue(const Value *V, MVT VT) { 160 unsigned Reg = 0; 161 162 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 163 if (CI->getValue().getActiveBits() <= 64) 164 Reg = FastEmit_i(VT, VT, ISD::Constant, CI->getZExtValue()); 165 } else if (isa<AllocaInst>(V)) { 166 Reg = TargetMaterializeAlloca(cast<AllocaInst>(V)); 167 } else if (isa<ConstantPointerNull>(V)) { 168 // Translate this as an integer zero so that it can be 169 // local-CSE'd with actual integer zeros. 170 Reg = 171 getRegForValue(Constant::getNullValue(TD.getIntPtrType(V->getContext()))); 172 } else if (const ConstantFP *CF = dyn_cast<ConstantFP>(V)) { 173 if (CF->isNullValue()) { 174 Reg = TargetMaterializeFloatZero(CF); 175 } else { 176 // Try to emit the constant directly. 177 Reg = FastEmit_f(VT, VT, ISD::ConstantFP, CF); 178 } 179 180 if (!Reg) { 181 // Try to emit the constant by using an integer constant with a cast. 182 const APFloat &Flt = CF->getValueAPF(); 183 EVT IntVT = TLI.getPointerTy(); 184 185 uint64_t x[2]; 186 uint32_t IntBitWidth = IntVT.getSizeInBits(); 187 bool isExact; 188 (void) Flt.convertToInteger(x, IntBitWidth, /*isSigned=*/true, 189 APFloat::rmTowardZero, &isExact); 190 if (isExact) { 191 APInt IntVal(IntBitWidth, x); 192 193 unsigned IntegerReg = 194 getRegForValue(ConstantInt::get(V->getContext(), IntVal)); 195 if (IntegerReg != 0) 196 Reg = FastEmit_r(IntVT.getSimpleVT(), VT, ISD::SINT_TO_FP, 197 IntegerReg, /*Kill=*/false); 198 } 199 } 200 } else if (const Operator *Op = dyn_cast<Operator>(V)) { 201 if (!SelectOperator(Op, Op->getOpcode())) 202 if (!isa<Instruction>(Op) || 203 !TargetSelectInstruction(cast<Instruction>(Op))) 204 return 0; 205 Reg = lookUpRegForValue(Op); 206 } else if (isa<UndefValue>(V)) { 207 Reg = createResultReg(TLI.getRegClassFor(VT)); 208 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, 209 TII.get(TargetOpcode::IMPLICIT_DEF), Reg); 210 } 211 212 // If target-independent code couldn't handle the value, give target-specific 213 // code a try. 214 if (!Reg && isa<Constant>(V)) 215 Reg = TargetMaterializeConstant(cast<Constant>(V)); 216 217 // Don't cache constant materializations in the general ValueMap. 218 // To do so would require tracking what uses they dominate. 219 if (Reg != 0) { 220 LocalValueMap[V] = Reg; 221 LastLocalValue = MRI.getVRegDef(Reg); 222 } 223 return Reg; 224 } 225 226 unsigned FastISel::lookUpRegForValue(const Value *V) { 227 // Look up the value to see if we already have a register for it. We 228 // cache values defined by Instructions across blocks, and other values 229 // only locally. This is because Instructions already have the SSA 230 // def-dominates-use requirement enforced. 231 DenseMap<const Value *, unsigned>::iterator I = FuncInfo.ValueMap.find(V); 232 if (I != FuncInfo.ValueMap.end()) 233 return I->second; 234 return LocalValueMap[V]; 235 } 236 237 /// UpdateValueMap - Update the value map to include the new mapping for this 238 /// instruction, or insert an extra copy to get the result in a previous 239 /// determined register. 240 /// NOTE: This is only necessary because we might select a block that uses 241 /// a value before we select the block that defines the value. It might be 242 /// possible to fix this by selecting blocks in reverse postorder. 243 void FastISel::UpdateValueMap(const Value *I, unsigned Reg, unsigned NumRegs) { 244 if (!isa<Instruction>(I)) { 245 LocalValueMap[I] = Reg; 246 return; 247 } 248 249 unsigned &AssignedReg = FuncInfo.ValueMap[I]; 250 if (AssignedReg == 0) 251 // Use the new register. 252 AssignedReg = Reg; 253 else if (Reg != AssignedReg) { 254 // Arrange for uses of AssignedReg to be replaced by uses of Reg. 255 for (unsigned i = 0; i < NumRegs; i++) 256 FuncInfo.RegFixups[AssignedReg+i] = Reg+i; 257 258 AssignedReg = Reg; 259 } 260 } 261 262 std::pair<unsigned, bool> FastISel::getRegForGEPIndex(const Value *Idx) { 263 unsigned IdxN = getRegForValue(Idx); 264 if (IdxN == 0) 265 // Unhandled operand. Halt "fast" selection and bail. 266 return std::pair<unsigned, bool>(0, false); 267 268 bool IdxNIsKill = hasTrivialKill(Idx); 269 270 // If the index is smaller or larger than intptr_t, truncate or extend it. 271 MVT PtrVT = TLI.getPointerTy(); 272 EVT IdxVT = EVT::getEVT(Idx->getType(), /*HandleUnknown=*/false); 273 if (IdxVT.bitsLT(PtrVT)) { 274 IdxN = FastEmit_r(IdxVT.getSimpleVT(), PtrVT, ISD::SIGN_EXTEND, 275 IdxN, IdxNIsKill); 276 IdxNIsKill = true; 277 } 278 else if (IdxVT.bitsGT(PtrVT)) { 279 IdxN = FastEmit_r(IdxVT.getSimpleVT(), PtrVT, ISD::TRUNCATE, 280 IdxN, IdxNIsKill); 281 IdxNIsKill = true; 282 } 283 return std::pair<unsigned, bool>(IdxN, IdxNIsKill); 284 } 285 286 void FastISel::recomputeInsertPt() { 287 if (getLastLocalValue()) { 288 FuncInfo.InsertPt = getLastLocalValue(); 289 FuncInfo.MBB = FuncInfo.InsertPt->getParent(); 290 ++FuncInfo.InsertPt; 291 } else 292 FuncInfo.InsertPt = FuncInfo.MBB->getFirstNonPHI(); 293 294 // Now skip past any EH_LABELs, which must remain at the beginning. 295 while (FuncInfo.InsertPt != FuncInfo.MBB->end() && 296 FuncInfo.InsertPt->getOpcode() == TargetOpcode::EH_LABEL) 297 ++FuncInfo.InsertPt; 298 } 299 300 FastISel::SavePoint FastISel::enterLocalValueArea() { 301 MachineBasicBlock::iterator OldInsertPt = FuncInfo.InsertPt; 302 DebugLoc OldDL = DL; 303 recomputeInsertPt(); 304 DL = DebugLoc(); 305 SavePoint SP = { OldInsertPt, OldDL }; 306 return SP; 307 } 308 309 void FastISel::leaveLocalValueArea(SavePoint OldInsertPt) { 310 if (FuncInfo.InsertPt != FuncInfo.MBB->begin()) 311 LastLocalValue = llvm::prior(FuncInfo.InsertPt); 312 313 // Restore the previous insert position. 314 FuncInfo.InsertPt = OldInsertPt.InsertPt; 315 DL = OldInsertPt.DL; 316 } 317 318 /// SelectBinaryOp - Select and emit code for a binary operator instruction, 319 /// which has an opcode which directly corresponds to the given ISD opcode. 320 /// 321 bool FastISel::SelectBinaryOp(const User *I, unsigned ISDOpcode) { 322 EVT VT = EVT::getEVT(I->getType(), /*HandleUnknown=*/true); 323 if (VT == MVT::Other || !VT.isSimple()) 324 // Unhandled type. Halt "fast" selection and bail. 325 return false; 326 327 // We only handle legal types. For example, on x86-32 the instruction 328 // selector contains all of the 64-bit instructions from x86-64, 329 // under the assumption that i64 won't be used if the target doesn't 330 // support it. 331 if (!TLI.isTypeLegal(VT)) { 332 // MVT::i1 is special. Allow AND, OR, or XOR because they 333 // don't require additional zeroing, which makes them easy. 334 if (VT == MVT::i1 && 335 (ISDOpcode == ISD::AND || ISDOpcode == ISD::OR || 336 ISDOpcode == ISD::XOR)) 337 VT = TLI.getTypeToTransformTo(I->getContext(), VT); 338 else 339 return false; 340 } 341 342 // Check if the first operand is a constant, and handle it as "ri". At -O0, 343 // we don't have anything that canonicalizes operand order. 344 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(0))) 345 if (isa<Instruction>(I) && cast<Instruction>(I)->isCommutative()) { 346 unsigned Op1 = getRegForValue(I->getOperand(1)); 347 if (Op1 == 0) return false; 348 349 bool Op1IsKill = hasTrivialKill(I->getOperand(1)); 350 351 unsigned ResultReg = FastEmit_ri_(VT.getSimpleVT(), ISDOpcode, Op1, 352 Op1IsKill, CI->getZExtValue(), 353 VT.getSimpleVT()); 354 if (ResultReg == 0) return false; 355 356 // We successfully emitted code for the given LLVM Instruction. 357 UpdateValueMap(I, ResultReg); 358 return true; 359 } 360 361 362 unsigned Op0 = getRegForValue(I->getOperand(0)); 363 if (Op0 == 0) // Unhandled operand. Halt "fast" selection and bail. 364 return false; 365 366 bool Op0IsKill = hasTrivialKill(I->getOperand(0)); 367 368 // Check if the second operand is a constant and handle it appropriately. 369 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) { 370 uint64_t Imm = CI->getZExtValue(); 371 372 // Transform "sdiv exact X, 8" -> "sra X, 3". 373 if (ISDOpcode == ISD::SDIV && isa<BinaryOperator>(I) && 374 cast<BinaryOperator>(I)->isExact() && 375 isPowerOf2_64(Imm)) { 376 Imm = Log2_64(Imm); 377 ISDOpcode = ISD::SRA; 378 } 379 380 unsigned ResultReg = FastEmit_ri_(VT.getSimpleVT(), ISDOpcode, Op0, 381 Op0IsKill, Imm, VT.getSimpleVT()); 382 if (ResultReg == 0) return false; 383 384 // We successfully emitted code for the given LLVM Instruction. 385 UpdateValueMap(I, ResultReg); 386 return true; 387 } 388 389 // Check if the second operand is a constant float. 390 if (ConstantFP *CF = dyn_cast<ConstantFP>(I->getOperand(1))) { 391 unsigned ResultReg = FastEmit_rf(VT.getSimpleVT(), VT.getSimpleVT(), 392 ISDOpcode, Op0, Op0IsKill, CF); 393 if (ResultReg != 0) { 394 // We successfully emitted code for the given LLVM Instruction. 395 UpdateValueMap(I, ResultReg); 396 return true; 397 } 398 } 399 400 unsigned Op1 = getRegForValue(I->getOperand(1)); 401 if (Op1 == 0) 402 // Unhandled operand. Halt "fast" selection and bail. 403 return false; 404 405 bool Op1IsKill = hasTrivialKill(I->getOperand(1)); 406 407 // Now we have both operands in registers. Emit the instruction. 408 unsigned ResultReg = FastEmit_rr(VT.getSimpleVT(), VT.getSimpleVT(), 409 ISDOpcode, 410 Op0, Op0IsKill, 411 Op1, Op1IsKill); 412 if (ResultReg == 0) 413 // Target-specific code wasn't able to find a machine opcode for 414 // the given ISD opcode and type. Halt "fast" selection and bail. 415 return false; 416 417 // We successfully emitted code for the given LLVM Instruction. 418 UpdateValueMap(I, ResultReg); 419 return true; 420 } 421 422 bool FastISel::SelectGetElementPtr(const User *I) { 423 unsigned N = getRegForValue(I->getOperand(0)); 424 if (N == 0) 425 // Unhandled operand. Halt "fast" selection and bail. 426 return false; 427 428 bool NIsKill = hasTrivialKill(I->getOperand(0)); 429 430 Type *Ty = I->getOperand(0)->getType(); 431 MVT VT = TLI.getPointerTy(); 432 for (GetElementPtrInst::const_op_iterator OI = I->op_begin()+1, 433 E = I->op_end(); OI != E; ++OI) { 434 const Value *Idx = *OI; 435 if (StructType *StTy = dyn_cast<StructType>(Ty)) { 436 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue(); 437 if (Field) { 438 // N = N + Offset 439 uint64_t Offs = TD.getStructLayout(StTy)->getElementOffset(Field); 440 // FIXME: This can be optimized by combining the add with a 441 // subsequent one. 442 N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, Offs, VT); 443 if (N == 0) 444 // Unhandled operand. Halt "fast" selection and bail. 445 return false; 446 NIsKill = true; 447 } 448 Ty = StTy->getElementType(Field); 449 } else { 450 Ty = cast<SequentialType>(Ty)->getElementType(); 451 452 // If this is a constant subscript, handle it quickly. 453 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) { 454 if (CI->isZero()) continue; 455 uint64_t Offs = 456 TD.getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue(); 457 N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, Offs, VT); 458 if (N == 0) 459 // Unhandled operand. Halt "fast" selection and bail. 460 return false; 461 NIsKill = true; 462 continue; 463 } 464 465 // N = N + Idx * ElementSize; 466 uint64_t ElementSize = TD.getTypeAllocSize(Ty); 467 std::pair<unsigned, bool> Pair = getRegForGEPIndex(Idx); 468 unsigned IdxN = Pair.first; 469 bool IdxNIsKill = Pair.second; 470 if (IdxN == 0) 471 // Unhandled operand. Halt "fast" selection and bail. 472 return false; 473 474 if (ElementSize != 1) { 475 IdxN = FastEmit_ri_(VT, ISD::MUL, IdxN, IdxNIsKill, ElementSize, VT); 476 if (IdxN == 0) 477 // Unhandled operand. Halt "fast" selection and bail. 478 return false; 479 IdxNIsKill = true; 480 } 481 N = FastEmit_rr(VT, VT, ISD::ADD, N, NIsKill, IdxN, IdxNIsKill); 482 if (N == 0) 483 // Unhandled operand. Halt "fast" selection and bail. 484 return false; 485 } 486 } 487 488 // We successfully emitted code for the given LLVM Instruction. 489 UpdateValueMap(I, N); 490 return true; 491 } 492 493 bool FastISel::SelectCall(const User *I) { 494 const CallInst *Call = cast<CallInst>(I); 495 496 // Handle simple inline asms. 497 if (const InlineAsm *IA = dyn_cast<InlineAsm>(Call->getCalledValue())) { 498 // Don't attempt to handle constraints. 499 if (!IA->getConstraintString().empty()) 500 return false; 501 502 unsigned ExtraInfo = 0; 503 if (IA->hasSideEffects()) 504 ExtraInfo |= InlineAsm::Extra_HasSideEffects; 505 if (IA->isAlignStack()) 506 ExtraInfo |= InlineAsm::Extra_IsAlignStack; 507 508 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, 509 TII.get(TargetOpcode::INLINEASM)) 510 .addExternalSymbol(IA->getAsmString().c_str()) 511 .addImm(ExtraInfo); 512 return true; 513 } 514 515 const Function *F = Call->getCalledFunction(); 516 if (!F) return false; 517 518 // Handle selected intrinsic function calls. 519 switch (F->getIntrinsicID()) { 520 default: break; 521 case Intrinsic::dbg_declare: { 522 const DbgDeclareInst *DI = cast<DbgDeclareInst>(Call); 523 if (!DIVariable(DI->getVariable()).Verify() || 524 !FuncInfo.MF->getMMI().hasDebugInfo()) 525 return true; 526 527 const Value *Address = DI->getAddress(); 528 if (!Address || isa<UndefValue>(Address) || isa<AllocaInst>(Address)) 529 return true; 530 531 unsigned Reg = 0; 532 unsigned Offset = 0; 533 if (const Argument *Arg = dyn_cast<Argument>(Address)) { 534 // Some arguments' frame index is recorded during argument lowering. 535 Offset = FuncInfo.getArgumentFrameIndex(Arg); 536 if (Offset) 537 Reg = TRI.getFrameRegister(*FuncInfo.MF); 538 } 539 if (!Reg) 540 Reg = getRegForValue(Address); 541 542 if (Reg) 543 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, 544 TII.get(TargetOpcode::DBG_VALUE)) 545 .addReg(Reg, RegState::Debug).addImm(Offset) 546 .addMetadata(DI->getVariable()); 547 return true; 548 } 549 case Intrinsic::dbg_value: { 550 // This form of DBG_VALUE is target-independent. 551 const DbgValueInst *DI = cast<DbgValueInst>(Call); 552 const MCInstrDesc &II = TII.get(TargetOpcode::DBG_VALUE); 553 const Value *V = DI->getValue(); 554 if (!V) { 555 // Currently the optimizer can produce this; insert an undef to 556 // help debugging. Probably the optimizer should not do this. 557 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II) 558 .addReg(0U).addImm(DI->getOffset()) 559 .addMetadata(DI->getVariable()); 560 } else if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 561 if (CI->getBitWidth() > 64) 562 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II) 563 .addCImm(CI).addImm(DI->getOffset()) 564 .addMetadata(DI->getVariable()); 565 else 566 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II) 567 .addImm(CI->getZExtValue()).addImm(DI->getOffset()) 568 .addMetadata(DI->getVariable()); 569 } else if (const ConstantFP *CF = dyn_cast<ConstantFP>(V)) { 570 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II) 571 .addFPImm(CF).addImm(DI->getOffset()) 572 .addMetadata(DI->getVariable()); 573 } else if (unsigned Reg = lookUpRegForValue(V)) { 574 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II) 575 .addReg(Reg, RegState::Debug).addImm(DI->getOffset()) 576 .addMetadata(DI->getVariable()); 577 } else { 578 // We can't yet handle anything else here because it would require 579 // generating code, thus altering codegen because of debug info. 580 DEBUG(dbgs() << "Dropping debug info for " << DI); 581 } 582 return true; 583 } 584 case Intrinsic::eh_exception: { 585 EVT VT = TLI.getValueType(Call->getType()); 586 if (TLI.getOperationAction(ISD::EXCEPTIONADDR, VT)!=TargetLowering::Expand) 587 break; 588 589 assert(FuncInfo.MBB->isLandingPad() && 590 "Call to eh.exception not in landing pad!"); 591 unsigned Reg = TLI.getExceptionAddressRegister(); 592 const TargetRegisterClass *RC = TLI.getRegClassFor(VT); 593 unsigned ResultReg = createResultReg(RC); 594 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), 595 ResultReg).addReg(Reg); 596 UpdateValueMap(Call, ResultReg); 597 return true; 598 } 599 case Intrinsic::eh_selector: { 600 EVT VT = TLI.getValueType(Call->getType()); 601 if (TLI.getOperationAction(ISD::EHSELECTION, VT) != TargetLowering::Expand) 602 break; 603 if (FuncInfo.MBB->isLandingPad()) 604 AddCatchInfo(*Call, &FuncInfo.MF->getMMI(), FuncInfo.MBB); 605 else { 606 #ifndef NDEBUG 607 FuncInfo.CatchInfoLost.insert(Call); 608 #endif 609 // FIXME: Mark exception selector register as live in. Hack for PR1508. 610 unsigned Reg = TLI.getExceptionSelectorRegister(); 611 if (Reg) FuncInfo.MBB->addLiveIn(Reg); 612 } 613 614 unsigned Reg = TLI.getExceptionSelectorRegister(); 615 EVT SrcVT = TLI.getPointerTy(); 616 const TargetRegisterClass *RC = TLI.getRegClassFor(SrcVT); 617 unsigned ResultReg = createResultReg(RC); 618 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), 619 ResultReg).addReg(Reg); 620 621 bool ResultRegIsKill = hasTrivialKill(Call); 622 623 // Cast the register to the type of the selector. 624 if (SrcVT.bitsGT(MVT::i32)) 625 ResultReg = FastEmit_r(SrcVT.getSimpleVT(), MVT::i32, ISD::TRUNCATE, 626 ResultReg, ResultRegIsKill); 627 else if (SrcVT.bitsLT(MVT::i32)) 628 ResultReg = FastEmit_r(SrcVT.getSimpleVT(), MVT::i32, 629 ISD::SIGN_EXTEND, ResultReg, ResultRegIsKill); 630 if (ResultReg == 0) 631 // Unhandled operand. Halt "fast" selection and bail. 632 return false; 633 634 UpdateValueMap(Call, ResultReg); 635 636 return true; 637 } 638 case Intrinsic::objectsize: { 639 ConstantInt *CI = cast<ConstantInt>(Call->getArgOperand(1)); 640 unsigned long long Res = CI->isZero() ? -1ULL : 0; 641 Constant *ResCI = ConstantInt::get(Call->getType(), Res); 642 unsigned ResultReg = getRegForValue(ResCI); 643 if (ResultReg == 0) 644 return false; 645 UpdateValueMap(Call, ResultReg); 646 return true; 647 } 648 } 649 650 // Usually, it does not make sense to initialize a value, 651 // make an unrelated function call and use the value, because 652 // it tends to be spilled on the stack. So, we move the pointer 653 // to the last local value to the beginning of the block, so that 654 // all the values which have already been materialized, 655 // appear after the call. It also makes sense to skip intrinsics 656 // since they tend to be inlined. 657 if (!isa<IntrinsicInst>(F)) 658 flushLocalValueMap(); 659 660 // An arbitrary call. Bail. 661 return false; 662 } 663 664 bool FastISel::SelectCast(const User *I, unsigned Opcode) { 665 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType()); 666 EVT DstVT = TLI.getValueType(I->getType()); 667 668 if (SrcVT == MVT::Other || !SrcVT.isSimple() || 669 DstVT == MVT::Other || !DstVT.isSimple()) 670 // Unhandled type. Halt "fast" selection and bail. 671 return false; 672 673 // Check if the destination type is legal. 674 if (!TLI.isTypeLegal(DstVT)) 675 return false; 676 677 // Check if the source operand is legal. 678 if (!TLI.isTypeLegal(SrcVT)) 679 return false; 680 681 unsigned InputReg = getRegForValue(I->getOperand(0)); 682 if (!InputReg) 683 // Unhandled operand. Halt "fast" selection and bail. 684 return false; 685 686 bool InputRegIsKill = hasTrivialKill(I->getOperand(0)); 687 688 unsigned ResultReg = FastEmit_r(SrcVT.getSimpleVT(), 689 DstVT.getSimpleVT(), 690 Opcode, 691 InputReg, InputRegIsKill); 692 if (!ResultReg) 693 return false; 694 695 UpdateValueMap(I, ResultReg); 696 return true; 697 } 698 699 bool FastISel::SelectBitCast(const User *I) { 700 // If the bitcast doesn't change the type, just use the operand value. 701 if (I->getType() == I->getOperand(0)->getType()) { 702 unsigned Reg = getRegForValue(I->getOperand(0)); 703 if (Reg == 0) 704 return false; 705 UpdateValueMap(I, Reg); 706 return true; 707 } 708 709 // Bitcasts of other values become reg-reg copies or BITCAST operators. 710 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType()); 711 EVT DstVT = TLI.getValueType(I->getType()); 712 713 if (SrcVT == MVT::Other || !SrcVT.isSimple() || 714 DstVT == MVT::Other || !DstVT.isSimple() || 715 !TLI.isTypeLegal(SrcVT) || !TLI.isTypeLegal(DstVT)) 716 // Unhandled type. Halt "fast" selection and bail. 717 return false; 718 719 unsigned Op0 = getRegForValue(I->getOperand(0)); 720 if (Op0 == 0) 721 // Unhandled operand. Halt "fast" selection and bail. 722 return false; 723 724 bool Op0IsKill = hasTrivialKill(I->getOperand(0)); 725 726 // First, try to perform the bitcast by inserting a reg-reg copy. 727 unsigned ResultReg = 0; 728 if (SrcVT.getSimpleVT() == DstVT.getSimpleVT()) { 729 TargetRegisterClass* SrcClass = TLI.getRegClassFor(SrcVT); 730 TargetRegisterClass* DstClass = TLI.getRegClassFor(DstVT); 731 // Don't attempt a cross-class copy. It will likely fail. 732 if (SrcClass == DstClass) { 733 ResultReg = createResultReg(DstClass); 734 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), 735 ResultReg).addReg(Op0); 736 } 737 } 738 739 // If the reg-reg copy failed, select a BITCAST opcode. 740 if (!ResultReg) 741 ResultReg = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), 742 ISD::BITCAST, Op0, Op0IsKill); 743 744 if (!ResultReg) 745 return false; 746 747 UpdateValueMap(I, ResultReg); 748 return true; 749 } 750 751 bool 752 FastISel::SelectInstruction(const Instruction *I) { 753 // Just before the terminator instruction, insert instructions to 754 // feed PHI nodes in successor blocks. 755 if (isa<TerminatorInst>(I)) 756 if (!HandlePHINodesInSuccessorBlocks(I->getParent())) 757 return false; 758 759 DL = I->getDebugLoc(); 760 761 // First, try doing target-independent selection. 762 if (SelectOperator(I, I->getOpcode())) { 763 DL = DebugLoc(); 764 return true; 765 } 766 767 // Next, try calling the target to attempt to handle the instruction. 768 if (TargetSelectInstruction(I)) { 769 DL = DebugLoc(); 770 return true; 771 } 772 773 DL = DebugLoc(); 774 return false; 775 } 776 777 /// FastEmitBranch - Emit an unconditional branch to the given block, 778 /// unless it is the immediate (fall-through) successor, and update 779 /// the CFG. 780 void 781 FastISel::FastEmitBranch(MachineBasicBlock *MSucc, DebugLoc DL) { 782 if (FuncInfo.MBB->isLayoutSuccessor(MSucc)) { 783 // The unconditional fall-through case, which needs no instructions. 784 } else { 785 // The unconditional branch case. 786 TII.InsertBranch(*FuncInfo.MBB, MSucc, NULL, 787 SmallVector<MachineOperand, 0>(), DL); 788 } 789 FuncInfo.MBB->addSuccessor(MSucc); 790 } 791 792 /// SelectFNeg - Emit an FNeg operation. 793 /// 794 bool 795 FastISel::SelectFNeg(const User *I) { 796 unsigned OpReg = getRegForValue(BinaryOperator::getFNegArgument(I)); 797 if (OpReg == 0) return false; 798 799 bool OpRegIsKill = hasTrivialKill(I); 800 801 // If the target has ISD::FNEG, use it. 802 EVT VT = TLI.getValueType(I->getType()); 803 unsigned ResultReg = FastEmit_r(VT.getSimpleVT(), VT.getSimpleVT(), 804 ISD::FNEG, OpReg, OpRegIsKill); 805 if (ResultReg != 0) { 806 UpdateValueMap(I, ResultReg); 807 return true; 808 } 809 810 // Bitcast the value to integer, twiddle the sign bit with xor, 811 // and then bitcast it back to floating-point. 812 if (VT.getSizeInBits() > 64) return false; 813 EVT IntVT = EVT::getIntegerVT(I->getContext(), VT.getSizeInBits()); 814 if (!TLI.isTypeLegal(IntVT)) 815 return false; 816 817 unsigned IntReg = FastEmit_r(VT.getSimpleVT(), IntVT.getSimpleVT(), 818 ISD::BITCAST, OpReg, OpRegIsKill); 819 if (IntReg == 0) 820 return false; 821 822 unsigned IntResultReg = FastEmit_ri_(IntVT.getSimpleVT(), ISD::XOR, 823 IntReg, /*Kill=*/true, 824 UINT64_C(1) << (VT.getSizeInBits()-1), 825 IntVT.getSimpleVT()); 826 if (IntResultReg == 0) 827 return false; 828 829 ResultReg = FastEmit_r(IntVT.getSimpleVT(), VT.getSimpleVT(), 830 ISD::BITCAST, IntResultReg, /*Kill=*/true); 831 if (ResultReg == 0) 832 return false; 833 834 UpdateValueMap(I, ResultReg); 835 return true; 836 } 837 838 bool 839 FastISel::SelectExtractValue(const User *U) { 840 const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(U); 841 if (!EVI) 842 return false; 843 844 // Make sure we only try to handle extracts with a legal result. But also 845 // allow i1 because it's easy. 846 EVT RealVT = TLI.getValueType(EVI->getType(), /*AllowUnknown=*/true); 847 if (!RealVT.isSimple()) 848 return false; 849 MVT VT = RealVT.getSimpleVT(); 850 if (!TLI.isTypeLegal(VT) && VT != MVT::i1) 851 return false; 852 853 const Value *Op0 = EVI->getOperand(0); 854 Type *AggTy = Op0->getType(); 855 856 // Get the base result register. 857 unsigned ResultReg; 858 DenseMap<const Value *, unsigned>::iterator I = FuncInfo.ValueMap.find(Op0); 859 if (I != FuncInfo.ValueMap.end()) 860 ResultReg = I->second; 861 else if (isa<Instruction>(Op0)) 862 ResultReg = FuncInfo.InitializeRegForValue(Op0); 863 else 864 return false; // fast-isel can't handle aggregate constants at the moment 865 866 // Get the actual result register, which is an offset from the base register. 867 unsigned VTIndex = ComputeLinearIndex(AggTy, EVI->getIndices()); 868 869 SmallVector<EVT, 4> AggValueVTs; 870 ComputeValueVTs(TLI, AggTy, AggValueVTs); 871 872 for (unsigned i = 0; i < VTIndex; i++) 873 ResultReg += TLI.getNumRegisters(FuncInfo.Fn->getContext(), AggValueVTs[i]); 874 875 UpdateValueMap(EVI, ResultReg); 876 return true; 877 } 878 879 bool 880 FastISel::SelectOperator(const User *I, unsigned Opcode) { 881 switch (Opcode) { 882 case Instruction::Add: 883 return SelectBinaryOp(I, ISD::ADD); 884 case Instruction::FAdd: 885 return SelectBinaryOp(I, ISD::FADD); 886 case Instruction::Sub: 887 return SelectBinaryOp(I, ISD::SUB); 888 case Instruction::FSub: 889 // FNeg is currently represented in LLVM IR as a special case of FSub. 890 if (BinaryOperator::isFNeg(I)) 891 return SelectFNeg(I); 892 return SelectBinaryOp(I, ISD::FSUB); 893 case Instruction::Mul: 894 return SelectBinaryOp(I, ISD::MUL); 895 case Instruction::FMul: 896 return SelectBinaryOp(I, ISD::FMUL); 897 case Instruction::SDiv: 898 return SelectBinaryOp(I, ISD::SDIV); 899 case Instruction::UDiv: 900 return SelectBinaryOp(I, ISD::UDIV); 901 case Instruction::FDiv: 902 return SelectBinaryOp(I, ISD::FDIV); 903 case Instruction::SRem: 904 return SelectBinaryOp(I, ISD::SREM); 905 case Instruction::URem: 906 return SelectBinaryOp(I, ISD::UREM); 907 case Instruction::FRem: 908 return SelectBinaryOp(I, ISD::FREM); 909 case Instruction::Shl: 910 return SelectBinaryOp(I, ISD::SHL); 911 case Instruction::LShr: 912 return SelectBinaryOp(I, ISD::SRL); 913 case Instruction::AShr: 914 return SelectBinaryOp(I, ISD::SRA); 915 case Instruction::And: 916 return SelectBinaryOp(I, ISD::AND); 917 case Instruction::Or: 918 return SelectBinaryOp(I, ISD::OR); 919 case Instruction::Xor: 920 return SelectBinaryOp(I, ISD::XOR); 921 922 case Instruction::GetElementPtr: 923 return SelectGetElementPtr(I); 924 925 case Instruction::Br: { 926 const BranchInst *BI = cast<BranchInst>(I); 927 928 if (BI->isUnconditional()) { 929 const BasicBlock *LLVMSucc = BI->getSuccessor(0); 930 MachineBasicBlock *MSucc = FuncInfo.MBBMap[LLVMSucc]; 931 FastEmitBranch(MSucc, BI->getDebugLoc()); 932 return true; 933 } 934 935 // Conditional branches are not handed yet. 936 // Halt "fast" selection and bail. 937 return false; 938 } 939 940 case Instruction::Unreachable: 941 // Nothing to emit. 942 return true; 943 944 case Instruction::Alloca: 945 // FunctionLowering has the static-sized case covered. 946 if (FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(I))) 947 return true; 948 949 // Dynamic-sized alloca is not handled yet. 950 return false; 951 952 case Instruction::Call: 953 return SelectCall(I); 954 955 case Instruction::BitCast: 956 return SelectBitCast(I); 957 958 case Instruction::FPToSI: 959 return SelectCast(I, ISD::FP_TO_SINT); 960 case Instruction::ZExt: 961 return SelectCast(I, ISD::ZERO_EXTEND); 962 case Instruction::SExt: 963 return SelectCast(I, ISD::SIGN_EXTEND); 964 case Instruction::Trunc: 965 return SelectCast(I, ISD::TRUNCATE); 966 case Instruction::SIToFP: 967 return SelectCast(I, ISD::SINT_TO_FP); 968 969 case Instruction::IntToPtr: // Deliberate fall-through. 970 case Instruction::PtrToInt: { 971 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType()); 972 EVT DstVT = TLI.getValueType(I->getType()); 973 if (DstVT.bitsGT(SrcVT)) 974 return SelectCast(I, ISD::ZERO_EXTEND); 975 if (DstVT.bitsLT(SrcVT)) 976 return SelectCast(I, ISD::TRUNCATE); 977 unsigned Reg = getRegForValue(I->getOperand(0)); 978 if (Reg == 0) return false; 979 UpdateValueMap(I, Reg); 980 return true; 981 } 982 983 case Instruction::ExtractValue: 984 return SelectExtractValue(I); 985 986 case Instruction::PHI: 987 llvm_unreachable("FastISel shouldn't visit PHI nodes!"); 988 989 default: 990 // Unhandled instruction. Halt "fast" selection and bail. 991 return false; 992 } 993 } 994 995 FastISel::FastISel(FunctionLoweringInfo &funcInfo) 996 : FuncInfo(funcInfo), 997 MRI(FuncInfo.MF->getRegInfo()), 998 MFI(*FuncInfo.MF->getFrameInfo()), 999 MCP(*FuncInfo.MF->getConstantPool()), 1000 TM(FuncInfo.MF->getTarget()), 1001 TD(*TM.getTargetData()), 1002 TII(*TM.getInstrInfo()), 1003 TLI(*TM.getTargetLowering()), 1004 TRI(*TM.getRegisterInfo()) { 1005 } 1006 1007 FastISel::~FastISel() {} 1008 1009 unsigned FastISel::FastEmit_(MVT, MVT, 1010 unsigned) { 1011 return 0; 1012 } 1013 1014 unsigned FastISel::FastEmit_r(MVT, MVT, 1015 unsigned, 1016 unsigned /*Op0*/, bool /*Op0IsKill*/) { 1017 return 0; 1018 } 1019 1020 unsigned FastISel::FastEmit_rr(MVT, MVT, 1021 unsigned, 1022 unsigned /*Op0*/, bool /*Op0IsKill*/, 1023 unsigned /*Op1*/, bool /*Op1IsKill*/) { 1024 return 0; 1025 } 1026 1027 unsigned FastISel::FastEmit_i(MVT, MVT, unsigned, uint64_t /*Imm*/) { 1028 return 0; 1029 } 1030 1031 unsigned FastISel::FastEmit_f(MVT, MVT, 1032 unsigned, const ConstantFP * /*FPImm*/) { 1033 return 0; 1034 } 1035 1036 unsigned FastISel::FastEmit_ri(MVT, MVT, 1037 unsigned, 1038 unsigned /*Op0*/, bool /*Op0IsKill*/, 1039 uint64_t /*Imm*/) { 1040 return 0; 1041 } 1042 1043 unsigned FastISel::FastEmit_rf(MVT, MVT, 1044 unsigned, 1045 unsigned /*Op0*/, bool /*Op0IsKill*/, 1046 const ConstantFP * /*FPImm*/) { 1047 return 0; 1048 } 1049 1050 unsigned FastISel::FastEmit_rri(MVT, MVT, 1051 unsigned, 1052 unsigned /*Op0*/, bool /*Op0IsKill*/, 1053 unsigned /*Op1*/, bool /*Op1IsKill*/, 1054 uint64_t /*Imm*/) { 1055 return 0; 1056 } 1057 1058 /// FastEmit_ri_ - This method is a wrapper of FastEmit_ri. It first tries 1059 /// to emit an instruction with an immediate operand using FastEmit_ri. 1060 /// If that fails, it materializes the immediate into a register and try 1061 /// FastEmit_rr instead. 1062 unsigned FastISel::FastEmit_ri_(MVT VT, unsigned Opcode, 1063 unsigned Op0, bool Op0IsKill, 1064 uint64_t Imm, MVT ImmType) { 1065 // If this is a multiply by a power of two, emit this as a shift left. 1066 if (Opcode == ISD::MUL && isPowerOf2_64(Imm)) { 1067 Opcode = ISD::SHL; 1068 Imm = Log2_64(Imm); 1069 } else if (Opcode == ISD::UDIV && isPowerOf2_64(Imm)) { 1070 // div x, 8 -> srl x, 3 1071 Opcode = ISD::SRL; 1072 Imm = Log2_64(Imm); 1073 } 1074 1075 // Horrible hack (to be removed), check to make sure shift amounts are 1076 // in-range. 1077 if ((Opcode == ISD::SHL || Opcode == ISD::SRA || Opcode == ISD::SRL) && 1078 Imm >= VT.getSizeInBits()) 1079 return 0; 1080 1081 // First check if immediate type is legal. If not, we can't use the ri form. 1082 unsigned ResultReg = FastEmit_ri(VT, VT, Opcode, Op0, Op0IsKill, Imm); 1083 if (ResultReg != 0) 1084 return ResultReg; 1085 unsigned MaterialReg = FastEmit_i(ImmType, ImmType, ISD::Constant, Imm); 1086 if (MaterialReg == 0) { 1087 // This is a bit ugly/slow, but failing here means falling out of 1088 // fast-isel, which would be very slow. 1089 IntegerType *ITy = IntegerType::get(FuncInfo.Fn->getContext(), 1090 VT.getSizeInBits()); 1091 MaterialReg = getRegForValue(ConstantInt::get(ITy, Imm)); 1092 } 1093 return FastEmit_rr(VT, VT, Opcode, 1094 Op0, Op0IsKill, 1095 MaterialReg, /*Kill=*/true); 1096 } 1097 1098 unsigned FastISel::createResultReg(const TargetRegisterClass* RC) { 1099 return MRI.createVirtualRegister(RC); 1100 } 1101 1102 unsigned FastISel::FastEmitInst_(unsigned MachineInstOpcode, 1103 const TargetRegisterClass* RC) { 1104 unsigned ResultReg = createResultReg(RC); 1105 const MCInstrDesc &II = TII.get(MachineInstOpcode); 1106 1107 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg); 1108 return ResultReg; 1109 } 1110 1111 unsigned FastISel::FastEmitInst_r(unsigned MachineInstOpcode, 1112 const TargetRegisterClass *RC, 1113 unsigned Op0, bool Op0IsKill) { 1114 unsigned ResultReg = createResultReg(RC); 1115 const MCInstrDesc &II = TII.get(MachineInstOpcode); 1116 1117 if (II.getNumDefs() >= 1) 1118 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg) 1119 .addReg(Op0, Op0IsKill * RegState::Kill); 1120 else { 1121 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II) 1122 .addReg(Op0, Op0IsKill * RegState::Kill); 1123 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), 1124 ResultReg).addReg(II.ImplicitDefs[0]); 1125 } 1126 1127 return ResultReg; 1128 } 1129 1130 unsigned FastISel::FastEmitInst_rr(unsigned MachineInstOpcode, 1131 const TargetRegisterClass *RC, 1132 unsigned Op0, bool Op0IsKill, 1133 unsigned Op1, bool Op1IsKill) { 1134 unsigned ResultReg = createResultReg(RC); 1135 const MCInstrDesc &II = TII.get(MachineInstOpcode); 1136 1137 if (II.getNumDefs() >= 1) 1138 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg) 1139 .addReg(Op0, Op0IsKill * RegState::Kill) 1140 .addReg(Op1, Op1IsKill * RegState::Kill); 1141 else { 1142 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II) 1143 .addReg(Op0, Op0IsKill * RegState::Kill) 1144 .addReg(Op1, Op1IsKill * RegState::Kill); 1145 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), 1146 ResultReg).addReg(II.ImplicitDefs[0]); 1147 } 1148 return ResultReg; 1149 } 1150 1151 unsigned FastISel::FastEmitInst_rrr(unsigned MachineInstOpcode, 1152 const TargetRegisterClass *RC, 1153 unsigned Op0, bool Op0IsKill, 1154 unsigned Op1, bool Op1IsKill, 1155 unsigned Op2, bool Op2IsKill) { 1156 unsigned ResultReg = createResultReg(RC); 1157 const MCInstrDesc &II = TII.get(MachineInstOpcode); 1158 1159 if (II.getNumDefs() >= 1) 1160 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg) 1161 .addReg(Op0, Op0IsKill * RegState::Kill) 1162 .addReg(Op1, Op1IsKill * RegState::Kill) 1163 .addReg(Op2, Op2IsKill * RegState::Kill); 1164 else { 1165 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II) 1166 .addReg(Op0, Op0IsKill * RegState::Kill) 1167 .addReg(Op1, Op1IsKill * RegState::Kill) 1168 .addReg(Op2, Op2IsKill * RegState::Kill); 1169 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), 1170 ResultReg).addReg(II.ImplicitDefs[0]); 1171 } 1172 return ResultReg; 1173 } 1174 1175 unsigned FastISel::FastEmitInst_ri(unsigned MachineInstOpcode, 1176 const TargetRegisterClass *RC, 1177 unsigned Op0, bool Op0IsKill, 1178 uint64_t Imm) { 1179 unsigned ResultReg = createResultReg(RC); 1180 const MCInstrDesc &II = TII.get(MachineInstOpcode); 1181 1182 if (II.getNumDefs() >= 1) 1183 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg) 1184 .addReg(Op0, Op0IsKill * RegState::Kill) 1185 .addImm(Imm); 1186 else { 1187 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II) 1188 .addReg(Op0, Op0IsKill * RegState::Kill) 1189 .addImm(Imm); 1190 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), 1191 ResultReg).addReg(II.ImplicitDefs[0]); 1192 } 1193 return ResultReg; 1194 } 1195 1196 unsigned FastISel::FastEmitInst_rii(unsigned MachineInstOpcode, 1197 const TargetRegisterClass *RC, 1198 unsigned Op0, bool Op0IsKill, 1199 uint64_t Imm1, uint64_t Imm2) { 1200 unsigned ResultReg = createResultReg(RC); 1201 const MCInstrDesc &II = TII.get(MachineInstOpcode); 1202 1203 if (II.getNumDefs() >= 1) 1204 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg) 1205 .addReg(Op0, Op0IsKill * RegState::Kill) 1206 .addImm(Imm1) 1207 .addImm(Imm2); 1208 else { 1209 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II) 1210 .addReg(Op0, Op0IsKill * RegState::Kill) 1211 .addImm(Imm1) 1212 .addImm(Imm2); 1213 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), 1214 ResultReg).addReg(II.ImplicitDefs[0]); 1215 } 1216 return ResultReg; 1217 } 1218 1219 unsigned FastISel::FastEmitInst_rf(unsigned MachineInstOpcode, 1220 const TargetRegisterClass *RC, 1221 unsigned Op0, bool Op0IsKill, 1222 const ConstantFP *FPImm) { 1223 unsigned ResultReg = createResultReg(RC); 1224 const MCInstrDesc &II = TII.get(MachineInstOpcode); 1225 1226 if (II.getNumDefs() >= 1) 1227 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg) 1228 .addReg(Op0, Op0IsKill * RegState::Kill) 1229 .addFPImm(FPImm); 1230 else { 1231 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II) 1232 .addReg(Op0, Op0IsKill * RegState::Kill) 1233 .addFPImm(FPImm); 1234 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), 1235 ResultReg).addReg(II.ImplicitDefs[0]); 1236 } 1237 return ResultReg; 1238 } 1239 1240 unsigned FastISel::FastEmitInst_rri(unsigned MachineInstOpcode, 1241 const TargetRegisterClass *RC, 1242 unsigned Op0, bool Op0IsKill, 1243 unsigned Op1, bool Op1IsKill, 1244 uint64_t Imm) { 1245 unsigned ResultReg = createResultReg(RC); 1246 const MCInstrDesc &II = TII.get(MachineInstOpcode); 1247 1248 if (II.getNumDefs() >= 1) 1249 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg) 1250 .addReg(Op0, Op0IsKill * RegState::Kill) 1251 .addReg(Op1, Op1IsKill * RegState::Kill) 1252 .addImm(Imm); 1253 else { 1254 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II) 1255 .addReg(Op0, Op0IsKill * RegState::Kill) 1256 .addReg(Op1, Op1IsKill * RegState::Kill) 1257 .addImm(Imm); 1258 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), 1259 ResultReg).addReg(II.ImplicitDefs[0]); 1260 } 1261 return ResultReg; 1262 } 1263 1264 unsigned FastISel::FastEmitInst_i(unsigned MachineInstOpcode, 1265 const TargetRegisterClass *RC, 1266 uint64_t Imm) { 1267 unsigned ResultReg = createResultReg(RC); 1268 const MCInstrDesc &II = TII.get(MachineInstOpcode); 1269 1270 if (II.getNumDefs() >= 1) 1271 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg).addImm(Imm); 1272 else { 1273 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II).addImm(Imm); 1274 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), 1275 ResultReg).addReg(II.ImplicitDefs[0]); 1276 } 1277 return ResultReg; 1278 } 1279 1280 unsigned FastISel::FastEmitInst_ii(unsigned MachineInstOpcode, 1281 const TargetRegisterClass *RC, 1282 uint64_t Imm1, uint64_t Imm2) { 1283 unsigned ResultReg = createResultReg(RC); 1284 const MCInstrDesc &II = TII.get(MachineInstOpcode); 1285 1286 if (II.getNumDefs() >= 1) 1287 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg) 1288 .addImm(Imm1).addImm(Imm2); 1289 else { 1290 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II).addImm(Imm1).addImm(Imm2); 1291 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), 1292 ResultReg).addReg(II.ImplicitDefs[0]); 1293 } 1294 return ResultReg; 1295 } 1296 1297 unsigned FastISel::FastEmitInst_extractsubreg(MVT RetVT, 1298 unsigned Op0, bool Op0IsKill, 1299 uint32_t Idx) { 1300 unsigned ResultReg = createResultReg(TLI.getRegClassFor(RetVT)); 1301 assert(TargetRegisterInfo::isVirtualRegister(Op0) && 1302 "Cannot yet extract from physregs"); 1303 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, 1304 DL, TII.get(TargetOpcode::COPY), ResultReg) 1305 .addReg(Op0, getKillRegState(Op0IsKill), Idx); 1306 return ResultReg; 1307 } 1308 1309 /// FastEmitZExtFromI1 - Emit MachineInstrs to compute the value of Op 1310 /// with all but the least significant bit set to zero. 1311 unsigned FastISel::FastEmitZExtFromI1(MVT VT, unsigned Op0, bool Op0IsKill) { 1312 return FastEmit_ri(VT, VT, ISD::AND, Op0, Op0IsKill, 1); 1313 } 1314 1315 /// HandlePHINodesInSuccessorBlocks - Handle PHI nodes in successor blocks. 1316 /// Emit code to ensure constants are copied into registers when needed. 1317 /// Remember the virtual registers that need to be added to the Machine PHI 1318 /// nodes as input. We cannot just directly add them, because expansion 1319 /// might result in multiple MBB's for one BB. As such, the start of the 1320 /// BB might correspond to a different MBB than the end. 1321 bool FastISel::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) { 1322 const TerminatorInst *TI = LLVMBB->getTerminator(); 1323 1324 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled; 1325 unsigned OrigNumPHINodesToUpdate = FuncInfo.PHINodesToUpdate.size(); 1326 1327 // Check successor nodes' PHI nodes that expect a constant to be available 1328 // from this block. 1329 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) { 1330 const BasicBlock *SuccBB = TI->getSuccessor(succ); 1331 if (!isa<PHINode>(SuccBB->begin())) continue; 1332 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB]; 1333 1334 // If this terminator has multiple identical successors (common for 1335 // switches), only handle each succ once. 1336 if (!SuccsHandled.insert(SuccMBB)) continue; 1337 1338 MachineBasicBlock::iterator MBBI = SuccMBB->begin(); 1339 1340 // At this point we know that there is a 1-1 correspondence between LLVM PHI 1341 // nodes and Machine PHI nodes, but the incoming operands have not been 1342 // emitted yet. 1343 for (BasicBlock::const_iterator I = SuccBB->begin(); 1344 const PHINode *PN = dyn_cast<PHINode>(I); ++I) { 1345 1346 // Ignore dead phi's. 1347 if (PN->use_empty()) continue; 1348 1349 // Only handle legal types. Two interesting things to note here. First, 1350 // by bailing out early, we may leave behind some dead instructions, 1351 // since SelectionDAG's HandlePHINodesInSuccessorBlocks will insert its 1352 // own moves. Second, this check is necessary because FastISel doesn't 1353 // use CreateRegs to create registers, so it always creates 1354 // exactly one register for each non-void instruction. 1355 EVT VT = TLI.getValueType(PN->getType(), /*AllowUnknown=*/true); 1356 if (VT == MVT::Other || !TLI.isTypeLegal(VT)) { 1357 // Promote MVT::i1. 1358 if (VT == MVT::i1) 1359 VT = TLI.getTypeToTransformTo(LLVMBB->getContext(), VT); 1360 else { 1361 FuncInfo.PHINodesToUpdate.resize(OrigNumPHINodesToUpdate); 1362 return false; 1363 } 1364 } 1365 1366 const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB); 1367 1368 // Set the DebugLoc for the copy. Prefer the location of the operand 1369 // if there is one; use the location of the PHI otherwise. 1370 DL = PN->getDebugLoc(); 1371 if (const Instruction *Inst = dyn_cast<Instruction>(PHIOp)) 1372 DL = Inst->getDebugLoc(); 1373 1374 unsigned Reg = getRegForValue(PHIOp); 1375 if (Reg == 0) { 1376 FuncInfo.PHINodesToUpdate.resize(OrigNumPHINodesToUpdate); 1377 return false; 1378 } 1379 FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg)); 1380 DL = DebugLoc(); 1381 } 1382 } 1383 1384 return true; 1385 } 1386