FastISel.cpp revision fd06aa7c076f397b307ddd638a2666f4090ee2b1
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/CodeGen/FastISel.h" 47#include "llvm/CodeGen/MachineInstrBuilder.h" 48#include "llvm/CodeGen/MachineModuleInfo.h" 49#include "llvm/CodeGen/MachineRegisterInfo.h" 50#include "llvm/CodeGen/DwarfWriter.h" 51#include "llvm/Analysis/DebugInfo.h" 52#include "llvm/Target/TargetData.h" 53#include "llvm/Target/TargetInstrInfo.h" 54#include "llvm/Target/TargetLowering.h" 55#include "llvm/Target/TargetMachine.h" 56#include "SelectionDAGBuild.h" 57using namespace llvm; 58 59unsigned FastISel::getRegForValue(Value *V) { 60 EVT RealVT = TLI.getValueType(V->getType(), /*AllowUnknown=*/true); 61 // Don't handle non-simple values in FastISel. 62 if (!RealVT.isSimple()) 63 return 0; 64 65 // Ignore illegal types. We must do this before looking up the value 66 // in ValueMap because Arguments are given virtual registers regardless 67 // of whether FastISel can handle them. 68 MVT VT = RealVT.getSimpleVT(); 69 if (!TLI.isTypeLegal(VT)) { 70 // Promote MVT::i1 to a legal type though, because it's common and easy. 71 if (VT == MVT::i1) 72 VT = TLI.getTypeToTransformTo(V->getContext(), VT).getSimpleVT(); 73 else 74 return 0; 75 } 76 77 // Look up the value to see if we already have a register for it. We 78 // cache values defined by Instructions across blocks, and other values 79 // only locally. This is because Instructions already have the SSA 80 // def-dominatess-use requirement enforced. 81 if (ValueMap.count(V)) 82 return ValueMap[V]; 83 unsigned Reg = LocalValueMap[V]; 84 if (Reg != 0) 85 return Reg; 86 87 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 88 if (CI->getValue().getActiveBits() <= 64) 89 Reg = FastEmit_i(VT, VT, ISD::Constant, CI->getZExtValue()); 90 } else if (isa<AllocaInst>(V)) { 91 Reg = TargetMaterializeAlloca(cast<AllocaInst>(V)); 92 } else if (isa<ConstantPointerNull>(V)) { 93 // Translate this as an integer zero so that it can be 94 // local-CSE'd with actual integer zeros. 95 Reg = 96 getRegForValue(Constant::getNullValue(TD.getIntPtrType(V->getContext()))); 97 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(V)) { 98 Reg = FastEmit_f(VT, VT, ISD::ConstantFP, CF); 99 100 if (!Reg) { 101 const APFloat &Flt = CF->getValueAPF(); 102 EVT IntVT = TLI.getPointerTy(); 103 104 uint64_t x[2]; 105 uint32_t IntBitWidth = IntVT.getSizeInBits(); 106 bool isExact; 107 (void) Flt.convertToInteger(x, IntBitWidth, /*isSigned=*/true, 108 APFloat::rmTowardZero, &isExact); 109 if (isExact) { 110 APInt IntVal(IntBitWidth, 2, x); 111 112 unsigned IntegerReg = 113 getRegForValue(ConstantInt::get(V->getContext(), IntVal)); 114 if (IntegerReg != 0) 115 Reg = FastEmit_r(IntVT.getSimpleVT(), VT, ISD::SINT_TO_FP, IntegerReg); 116 } 117 } 118 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) { 119 if (!SelectOperator(CE, CE->getOpcode())) return 0; 120 Reg = LocalValueMap[CE]; 121 } else if (isa<UndefValue>(V)) { 122 Reg = createResultReg(TLI.getRegClassFor(VT)); 123 BuildMI(MBB, DL, TII.get(TargetInstrInfo::IMPLICIT_DEF), Reg); 124 } 125 126 // If target-independent code couldn't handle the value, give target-specific 127 // code a try. 128 if (!Reg && isa<Constant>(V)) 129 Reg = TargetMaterializeConstant(cast<Constant>(V)); 130 131 // Don't cache constant materializations in the general ValueMap. 132 // To do so would require tracking what uses they dominate. 133 if (Reg != 0) 134 LocalValueMap[V] = Reg; 135 return Reg; 136} 137 138unsigned FastISel::lookUpRegForValue(Value *V) { 139 // Look up the value to see if we already have a register for it. We 140 // cache values defined by Instructions across blocks, and other values 141 // only locally. This is because Instructions already have the SSA 142 // def-dominatess-use requirement enforced. 143 if (ValueMap.count(V)) 144 return ValueMap[V]; 145 return LocalValueMap[V]; 146} 147 148/// UpdateValueMap - Update the value map to include the new mapping for this 149/// instruction, or insert an extra copy to get the result in a previous 150/// determined register. 151/// NOTE: This is only necessary because we might select a block that uses 152/// a value before we select the block that defines the value. It might be 153/// possible to fix this by selecting blocks in reverse postorder. 154unsigned FastISel::UpdateValueMap(Value* I, unsigned Reg) { 155 if (!isa<Instruction>(I)) { 156 LocalValueMap[I] = Reg; 157 return Reg; 158 } 159 160 unsigned &AssignedReg = ValueMap[I]; 161 if (AssignedReg == 0) 162 AssignedReg = Reg; 163 else if (Reg != AssignedReg) { 164 const TargetRegisterClass *RegClass = MRI.getRegClass(Reg); 165 TII.copyRegToReg(*MBB, MBB->end(), AssignedReg, 166 Reg, RegClass, RegClass); 167 } 168 return AssignedReg; 169} 170 171unsigned FastISel::getRegForGEPIndex(Value *Idx) { 172 unsigned IdxN = getRegForValue(Idx); 173 if (IdxN == 0) 174 // Unhandled operand. Halt "fast" selection and bail. 175 return 0; 176 177 // If the index is smaller or larger than intptr_t, truncate or extend it. 178 MVT PtrVT = TLI.getPointerTy(); 179 EVT IdxVT = EVT::getEVT(Idx->getType(), /*HandleUnknown=*/false); 180 if (IdxVT.bitsLT(PtrVT)) 181 IdxN = FastEmit_r(IdxVT.getSimpleVT(), PtrVT, ISD::SIGN_EXTEND, IdxN); 182 else if (IdxVT.bitsGT(PtrVT)) 183 IdxN = FastEmit_r(IdxVT.getSimpleVT(), PtrVT, ISD::TRUNCATE, IdxN); 184 return IdxN; 185} 186 187/// SelectBinaryOp - Select and emit code for a binary operator instruction, 188/// which has an opcode which directly corresponds to the given ISD opcode. 189/// 190bool FastISel::SelectBinaryOp(User *I, ISD::NodeType ISDOpcode) { 191 EVT VT = EVT::getEVT(I->getType(), /*HandleUnknown=*/true); 192 if (VT == MVT::Other || !VT.isSimple()) 193 // Unhandled type. Halt "fast" selection and bail. 194 return false; 195 196 // We only handle legal types. For example, on x86-32 the instruction 197 // selector contains all of the 64-bit instructions from x86-64, 198 // under the assumption that i64 won't be used if the target doesn't 199 // support it. 200 if (!TLI.isTypeLegal(VT)) { 201 // MVT::i1 is special. Allow AND, OR, or XOR because they 202 // don't require additional zeroing, which makes them easy. 203 if (VT == MVT::i1 && 204 (ISDOpcode == ISD::AND || ISDOpcode == ISD::OR || 205 ISDOpcode == ISD::XOR)) 206 VT = TLI.getTypeToTransformTo(I->getContext(), VT); 207 else 208 return false; 209 } 210 211 unsigned Op0 = getRegForValue(I->getOperand(0)); 212 if (Op0 == 0) 213 // Unhandled operand. Halt "fast" selection and bail. 214 return false; 215 216 // Check if the second operand is a constant and handle it appropriately. 217 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) { 218 unsigned ResultReg = FastEmit_ri(VT.getSimpleVT(), VT.getSimpleVT(), 219 ISDOpcode, Op0, CI->getZExtValue()); 220 if (ResultReg != 0) { 221 // We successfully emitted code for the given LLVM Instruction. 222 UpdateValueMap(I, ResultReg); 223 return true; 224 } 225 } 226 227 // Check if the second operand is a constant float. 228 if (ConstantFP *CF = dyn_cast<ConstantFP>(I->getOperand(1))) { 229 unsigned ResultReg = FastEmit_rf(VT.getSimpleVT(), VT.getSimpleVT(), 230 ISDOpcode, Op0, CF); 231 if (ResultReg != 0) { 232 // We successfully emitted code for the given LLVM Instruction. 233 UpdateValueMap(I, ResultReg); 234 return true; 235 } 236 } 237 238 unsigned Op1 = getRegForValue(I->getOperand(1)); 239 if (Op1 == 0) 240 // Unhandled operand. Halt "fast" selection and bail. 241 return false; 242 243 // Now we have both operands in registers. Emit the instruction. 244 unsigned ResultReg = FastEmit_rr(VT.getSimpleVT(), VT.getSimpleVT(), 245 ISDOpcode, Op0, Op1); 246 if (ResultReg == 0) 247 // Target-specific code wasn't able to find a machine opcode for 248 // the given ISD opcode and type. Halt "fast" selection and bail. 249 return false; 250 251 // We successfully emitted code for the given LLVM Instruction. 252 UpdateValueMap(I, ResultReg); 253 return true; 254} 255 256bool FastISel::SelectGetElementPtr(User *I) { 257 unsigned N = getRegForValue(I->getOperand(0)); 258 if (N == 0) 259 // Unhandled operand. Halt "fast" selection and bail. 260 return false; 261 262 const Type *Ty = I->getOperand(0)->getType(); 263 MVT VT = TLI.getPointerTy(); 264 for (GetElementPtrInst::op_iterator OI = I->op_begin()+1, E = I->op_end(); 265 OI != E; ++OI) { 266 Value *Idx = *OI; 267 if (const StructType *StTy = dyn_cast<StructType>(Ty)) { 268 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue(); 269 if (Field) { 270 // N = N + Offset 271 uint64_t Offs = TD.getStructLayout(StTy)->getElementOffset(Field); 272 // FIXME: This can be optimized by combining the add with a 273 // subsequent one. 274 N = FastEmit_ri_(VT, ISD::ADD, N, Offs, VT); 275 if (N == 0) 276 // Unhandled operand. Halt "fast" selection and bail. 277 return false; 278 } 279 Ty = StTy->getElementType(Field); 280 } else { 281 Ty = cast<SequentialType>(Ty)->getElementType(); 282 283 // If this is a constant subscript, handle it quickly. 284 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) { 285 if (CI->getZExtValue() == 0) continue; 286 uint64_t Offs = 287 TD.getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue(); 288 N = FastEmit_ri_(VT, ISD::ADD, N, Offs, VT); 289 if (N == 0) 290 // Unhandled operand. Halt "fast" selection and bail. 291 return false; 292 continue; 293 } 294 295 // N = N + Idx * ElementSize; 296 uint64_t ElementSize = TD.getTypeAllocSize(Ty); 297 unsigned IdxN = getRegForGEPIndex(Idx); 298 if (IdxN == 0) 299 // Unhandled operand. Halt "fast" selection and bail. 300 return false; 301 302 if (ElementSize != 1) { 303 IdxN = FastEmit_ri_(VT, ISD::MUL, IdxN, ElementSize, VT); 304 if (IdxN == 0) 305 // Unhandled operand. Halt "fast" selection and bail. 306 return false; 307 } 308 N = FastEmit_rr(VT, VT, ISD::ADD, N, IdxN); 309 if (N == 0) 310 // Unhandled operand. Halt "fast" selection and bail. 311 return false; 312 } 313 } 314 315 // We successfully emitted code for the given LLVM Instruction. 316 UpdateValueMap(I, N); 317 return true; 318} 319 320bool FastISel::SelectCall(User *I) { 321 Function *F = cast<CallInst>(I)->getCalledFunction(); 322 if (!F) return false; 323 324 unsigned IID = F->getIntrinsicID(); 325 switch (IID) { 326 default: break; 327 case Intrinsic::dbg_stoppoint: { 328 DbgStopPointInst *SPI = cast<DbgStopPointInst>(I); 329 if (isValidDebugInfoIntrinsic(*SPI, CodeGenOpt::None)) 330 setCurDebugLoc(ExtractDebugLocation(*SPI, MF.getDebugLocInfo())); 331 return true; 332 } 333 case Intrinsic::dbg_region_start: { 334 DbgRegionStartInst *RSI = cast<DbgRegionStartInst>(I); 335 if (isValidDebugInfoIntrinsic(*RSI, CodeGenOpt::None) && DW 336 && DW->ShouldEmitDwarfDebug()) { 337 unsigned ID = 338 DW->RecordRegionStart(RSI->getContext()); 339 const TargetInstrDesc &II = TII.get(TargetInstrInfo::DBG_LABEL); 340 BuildMI(MBB, DL, II).addImm(ID); 341 } 342 return true; 343 } 344 case Intrinsic::dbg_region_end: { 345 DbgRegionEndInst *REI = cast<DbgRegionEndInst>(I); 346 if (isValidDebugInfoIntrinsic(*REI, CodeGenOpt::None) && DW 347 && DW->ShouldEmitDwarfDebug()) { 348 unsigned ID = 0; 349 DISubprogram Subprogram(REI->getContext()); 350 if (isInlinedFnEnd(*REI, MF.getFunction())) { 351 // This is end of an inlined function. 352 const TargetInstrDesc &II = TII.get(TargetInstrInfo::DBG_LABEL); 353 ID = DW->RecordInlinedFnEnd(Subprogram); 354 if (ID) 355 // Returned ID is 0 if this is unbalanced "end of inlined 356 // scope". This could happen if optimizer eats dbg intrinsics 357 // or "beginning of inlined scope" is not recoginized due to 358 // missing location info. In such cases, ignore this region.end. 359 BuildMI(MBB, DL, II).addImm(ID); 360 } else { 361 const TargetInstrDesc &II = TII.get(TargetInstrInfo::DBG_LABEL); 362 ID = DW->RecordRegionEnd(REI->getContext()); 363 BuildMI(MBB, DL, II).addImm(ID); 364 } 365 } 366 return true; 367 } 368 case Intrinsic::dbg_func_start: { 369 DbgFuncStartInst *FSI = cast<DbgFuncStartInst>(I); 370 if (!isValidDebugInfoIntrinsic(*FSI, CodeGenOpt::None) || !DW 371 || !DW->ShouldEmitDwarfDebug()) 372 return true; 373 374 if (isInlinedFnStart(*FSI, MF.getFunction())) { 375 // This is a beginning of an inlined function. 376 377 // If llvm.dbg.func.start is seen in a new block before any 378 // llvm.dbg.stoppoint intrinsic then the location info is unknown. 379 // FIXME : Why DebugLoc is reset at the beginning of each block ? 380 DebugLoc PrevLoc = DL; 381 if (PrevLoc.isUnknown()) 382 return true; 383 // Record the source line. 384 setCurDebugLoc(ExtractDebugLocation(*FSI, MF.getDebugLocInfo())); 385 386 DebugLocTuple PrevLocTpl = MF.getDebugLocTuple(PrevLoc); 387 DISubprogram SP(FSI->getSubprogram()); 388 unsigned LabelID = DW->RecordInlinedFnStart(SP, 389 DICompileUnit(PrevLocTpl.CompileUnit), 390 PrevLocTpl.Line, 391 PrevLocTpl.Col); 392 const TargetInstrDesc &II = TII.get(TargetInstrInfo::DBG_LABEL); 393 BuildMI(MBB, DL, II).addImm(LabelID); 394 return true; 395 } 396 397 // This is a beginning of a new function. 398 MF.setDefaultDebugLoc(ExtractDebugLocation(*FSI, MF.getDebugLocInfo())); 399 400 // llvm.dbg.func_start also defines beginning of function scope. 401 DW->RecordRegionStart(FSI->getSubprogram()); 402 return true; 403 } 404 case Intrinsic::dbg_declare: { 405 DbgDeclareInst *DI = cast<DbgDeclareInst>(I); 406 if (!isValidDebugInfoIntrinsic(*DI, CodeGenOpt::None) || !DW 407 || !DW->ShouldEmitDwarfDebug()) 408 return true; 409 410 Value *Variable = DI->getVariable(); 411 Value *Address = DI->getAddress(); 412 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Address)) 413 Address = BCI->getOperand(0); 414 AllocaInst *AI = dyn_cast<AllocaInst>(Address); 415 // Don't handle byval struct arguments or VLAs, for example. 416 if (!AI) break; 417 DenseMap<const AllocaInst*, int>::iterator SI = 418 StaticAllocaMap.find(AI); 419 if (SI == StaticAllocaMap.end()) break; // VLAs. 420 int FI = SI->second; 421 422 DW->RecordVariable(cast<MDNode>(Variable), FI); 423 return true; 424 } 425 case Intrinsic::eh_exception: { 426 EVT VT = TLI.getValueType(I->getType()); 427 switch (TLI.getOperationAction(ISD::EXCEPTIONADDR, VT)) { 428 default: break; 429 case TargetLowering::Expand: { 430 assert(MBB->isLandingPad() && "Call to eh.exception not in landing pad!"); 431 unsigned Reg = TLI.getExceptionAddressRegister(); 432 const TargetRegisterClass *RC = TLI.getRegClassFor(VT); 433 unsigned ResultReg = createResultReg(RC); 434 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg, 435 Reg, RC, RC); 436 assert(InsertedCopy && "Can't copy address registers!"); 437 InsertedCopy = InsertedCopy; 438 UpdateValueMap(I, ResultReg); 439 return true; 440 } 441 } 442 break; 443 } 444 case Intrinsic::eh_selector_i32: 445 case Intrinsic::eh_selector_i64: { 446 EVT VT = TLI.getValueType(I->getType()); 447 switch (TLI.getOperationAction(ISD::EHSELECTION, VT)) { 448 default: break; 449 case TargetLowering::Expand: { 450 EVT VT = (IID == Intrinsic::eh_selector_i32 ? 451 MVT::i32 : MVT::i64); 452 453 if (MMI) { 454 if (MBB->isLandingPad()) 455 AddCatchInfo(*cast<CallInst>(I), MMI, MBB); 456 else { 457#ifndef NDEBUG 458 CatchInfoLost.insert(cast<CallInst>(I)); 459#endif 460 // FIXME: Mark exception selector register as live in. Hack for PR1508. 461 unsigned Reg = TLI.getExceptionSelectorRegister(); 462 if (Reg) MBB->addLiveIn(Reg); 463 } 464 465 unsigned Reg = TLI.getExceptionSelectorRegister(); 466 const TargetRegisterClass *RC = TLI.getRegClassFor(VT); 467 unsigned ResultReg = createResultReg(RC); 468 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg, 469 Reg, RC, RC); 470 assert(InsertedCopy && "Can't copy address registers!"); 471 InsertedCopy = InsertedCopy; 472 UpdateValueMap(I, ResultReg); 473 } else { 474 unsigned ResultReg = 475 getRegForValue(Constant::getNullValue(I->getType())); 476 UpdateValueMap(I, ResultReg); 477 } 478 return true; 479 } 480 } 481 break; 482 } 483 } 484 return false; 485} 486 487bool FastISel::SelectCast(User *I, ISD::NodeType Opcode) { 488 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType()); 489 EVT DstVT = TLI.getValueType(I->getType()); 490 491 if (SrcVT == MVT::Other || !SrcVT.isSimple() || 492 DstVT == MVT::Other || !DstVT.isSimple()) 493 // Unhandled type. Halt "fast" selection and bail. 494 return false; 495 496 // Check if the destination type is legal. Or as a special case, 497 // it may be i1 if we're doing a truncate because that's 498 // easy and somewhat common. 499 if (!TLI.isTypeLegal(DstVT)) 500 if (DstVT != MVT::i1 || Opcode != ISD::TRUNCATE) 501 // Unhandled type. Halt "fast" selection and bail. 502 return false; 503 504 // Check if the source operand is legal. Or as a special case, 505 // it may be i1 if we're doing zero-extension because that's 506 // easy and somewhat common. 507 if (!TLI.isTypeLegal(SrcVT)) 508 if (SrcVT != MVT::i1 || Opcode != ISD::ZERO_EXTEND) 509 // Unhandled type. Halt "fast" selection and bail. 510 return false; 511 512 unsigned InputReg = getRegForValue(I->getOperand(0)); 513 if (!InputReg) 514 // Unhandled operand. Halt "fast" selection and bail. 515 return false; 516 517 // If the operand is i1, arrange for the high bits in the register to be zero. 518 if (SrcVT == MVT::i1) { 519 SrcVT = TLI.getTypeToTransformTo(I->getContext(), SrcVT); 520 InputReg = FastEmitZExtFromI1(SrcVT.getSimpleVT(), InputReg); 521 if (!InputReg) 522 return false; 523 } 524 // If the result is i1, truncate to the target's type for i1 first. 525 if (DstVT == MVT::i1) 526 DstVT = TLI.getTypeToTransformTo(I->getContext(), DstVT); 527 528 unsigned ResultReg = FastEmit_r(SrcVT.getSimpleVT(), 529 DstVT.getSimpleVT(), 530 Opcode, 531 InputReg); 532 if (!ResultReg) 533 return false; 534 535 UpdateValueMap(I, ResultReg); 536 return true; 537} 538 539bool FastISel::SelectBitCast(User *I) { 540 // If the bitcast doesn't change the type, just use the operand value. 541 if (I->getType() == I->getOperand(0)->getType()) { 542 unsigned Reg = getRegForValue(I->getOperand(0)); 543 if (Reg == 0) 544 return false; 545 UpdateValueMap(I, Reg); 546 return true; 547 } 548 549 // Bitcasts of other values become reg-reg copies or BIT_CONVERT operators. 550 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType()); 551 EVT DstVT = TLI.getValueType(I->getType()); 552 553 if (SrcVT == MVT::Other || !SrcVT.isSimple() || 554 DstVT == MVT::Other || !DstVT.isSimple() || 555 !TLI.isTypeLegal(SrcVT) || !TLI.isTypeLegal(DstVT)) 556 // Unhandled type. Halt "fast" selection and bail. 557 return false; 558 559 unsigned Op0 = getRegForValue(I->getOperand(0)); 560 if (Op0 == 0) 561 // Unhandled operand. Halt "fast" selection and bail. 562 return false; 563 564 // First, try to perform the bitcast by inserting a reg-reg copy. 565 unsigned ResultReg = 0; 566 if (SrcVT.getSimpleVT() == DstVT.getSimpleVT()) { 567 TargetRegisterClass* SrcClass = TLI.getRegClassFor(SrcVT); 568 TargetRegisterClass* DstClass = TLI.getRegClassFor(DstVT); 569 ResultReg = createResultReg(DstClass); 570 571 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg, 572 Op0, DstClass, SrcClass); 573 if (!InsertedCopy) 574 ResultReg = 0; 575 } 576 577 // If the reg-reg copy failed, select a BIT_CONVERT opcode. 578 if (!ResultReg) 579 ResultReg = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), 580 ISD::BIT_CONVERT, Op0); 581 582 if (!ResultReg) 583 return false; 584 585 UpdateValueMap(I, ResultReg); 586 return true; 587} 588 589bool 590FastISel::SelectInstruction(Instruction *I) { 591 return SelectOperator(I, I->getOpcode()); 592} 593 594/// FastEmitBranch - Emit an unconditional branch to the given block, 595/// unless it is the immediate (fall-through) successor, and update 596/// the CFG. 597void 598FastISel::FastEmitBranch(MachineBasicBlock *MSucc) { 599 MachineFunction::iterator NextMBB = 600 next(MachineFunction::iterator(MBB)); 601 602 if (MBB->isLayoutSuccessor(MSucc)) { 603 // The unconditional fall-through case, which needs no instructions. 604 } else { 605 // The unconditional branch case. 606 TII.InsertBranch(*MBB, MSucc, NULL, SmallVector<MachineOperand, 0>()); 607 } 608 MBB->addSuccessor(MSucc); 609} 610 611/// SelectFNeg - Emit an FNeg operation. 612/// 613bool 614FastISel::SelectFNeg(User *I) { 615 unsigned OpReg = getRegForValue(BinaryOperator::getFNegArgument(I)); 616 if (OpReg == 0) return false; 617 618 // Bitcast the value to integer, twiddle the sign bit with xor, 619 // and then bitcast it back to floating-point. 620 EVT VT = TLI.getValueType(I->getType()); 621 if (VT.getSizeInBits() > 64) return false; 622 EVT IntVT = EVT::getIntegerVT(I->getContext(), VT.getSizeInBits()); 623 624 unsigned IntReg = FastEmit_r(VT.getSimpleVT(), IntVT.getSimpleVT(), 625 ISD::BIT_CONVERT, OpReg); 626 if (IntReg == 0) 627 return false; 628 629 unsigned IntResultReg = FastEmit_ri_(IntVT.getSimpleVT(), ISD::XOR, IntReg, 630 UINT64_C(1) << (VT.getSizeInBits()-1), 631 IntVT.getSimpleVT()); 632 if (IntResultReg == 0) 633 return false; 634 635 unsigned ResultReg = FastEmit_r(IntVT.getSimpleVT(), VT.getSimpleVT(), 636 ISD::BIT_CONVERT, IntResultReg); 637 if (ResultReg == 0) 638 return false; 639 640 UpdateValueMap(I, ResultReg); 641 return true; 642} 643 644bool 645FastISel::SelectOperator(User *I, unsigned Opcode) { 646 switch (Opcode) { 647 case Instruction::Add: 648 return SelectBinaryOp(I, ISD::ADD); 649 case Instruction::FAdd: 650 return SelectBinaryOp(I, ISD::FADD); 651 case Instruction::Sub: 652 return SelectBinaryOp(I, ISD::SUB); 653 case Instruction::FSub: 654 // FNeg is currently represented in LLVM IR as a special case of FSub. 655 if (BinaryOperator::isFNeg(I)) 656 return SelectFNeg(I); 657 return SelectBinaryOp(I, ISD::FSUB); 658 case Instruction::Mul: 659 return SelectBinaryOp(I, ISD::MUL); 660 case Instruction::FMul: 661 return SelectBinaryOp(I, ISD::FMUL); 662 case Instruction::SDiv: 663 return SelectBinaryOp(I, ISD::SDIV); 664 case Instruction::UDiv: 665 return SelectBinaryOp(I, ISD::UDIV); 666 case Instruction::FDiv: 667 return SelectBinaryOp(I, ISD::FDIV); 668 case Instruction::SRem: 669 return SelectBinaryOp(I, ISD::SREM); 670 case Instruction::URem: 671 return SelectBinaryOp(I, ISD::UREM); 672 case Instruction::FRem: 673 return SelectBinaryOp(I, ISD::FREM); 674 case Instruction::Shl: 675 return SelectBinaryOp(I, ISD::SHL); 676 case Instruction::LShr: 677 return SelectBinaryOp(I, ISD::SRL); 678 case Instruction::AShr: 679 return SelectBinaryOp(I, ISD::SRA); 680 case Instruction::And: 681 return SelectBinaryOp(I, ISD::AND); 682 case Instruction::Or: 683 return SelectBinaryOp(I, ISD::OR); 684 case Instruction::Xor: 685 return SelectBinaryOp(I, ISD::XOR); 686 687 case Instruction::GetElementPtr: 688 return SelectGetElementPtr(I); 689 690 case Instruction::Br: { 691 BranchInst *BI = cast<BranchInst>(I); 692 693 if (BI->isUnconditional()) { 694 BasicBlock *LLVMSucc = BI->getSuccessor(0); 695 MachineBasicBlock *MSucc = MBBMap[LLVMSucc]; 696 FastEmitBranch(MSucc); 697 return true; 698 } 699 700 // Conditional branches are not handed yet. 701 // Halt "fast" selection and bail. 702 return false; 703 } 704 705 case Instruction::Unreachable: 706 // Nothing to emit. 707 return true; 708 709 case Instruction::PHI: 710 // PHI nodes are already emitted. 711 return true; 712 713 case Instruction::Alloca: 714 // FunctionLowering has the static-sized case covered. 715 if (StaticAllocaMap.count(cast<AllocaInst>(I))) 716 return true; 717 718 // Dynamic-sized alloca is not handled yet. 719 return false; 720 721 case Instruction::Call: 722 return SelectCall(I); 723 724 case Instruction::BitCast: 725 return SelectBitCast(I); 726 727 case Instruction::FPToSI: 728 return SelectCast(I, ISD::FP_TO_SINT); 729 case Instruction::ZExt: 730 return SelectCast(I, ISD::ZERO_EXTEND); 731 case Instruction::SExt: 732 return SelectCast(I, ISD::SIGN_EXTEND); 733 case Instruction::Trunc: 734 return SelectCast(I, ISD::TRUNCATE); 735 case Instruction::SIToFP: 736 return SelectCast(I, ISD::SINT_TO_FP); 737 738 case Instruction::IntToPtr: // Deliberate fall-through. 739 case Instruction::PtrToInt: { 740 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType()); 741 EVT DstVT = TLI.getValueType(I->getType()); 742 if (DstVT.bitsGT(SrcVT)) 743 return SelectCast(I, ISD::ZERO_EXTEND); 744 if (DstVT.bitsLT(SrcVT)) 745 return SelectCast(I, ISD::TRUNCATE); 746 unsigned Reg = getRegForValue(I->getOperand(0)); 747 if (Reg == 0) return false; 748 UpdateValueMap(I, Reg); 749 return true; 750 } 751 752 default: 753 // Unhandled instruction. Halt "fast" selection and bail. 754 return false; 755 } 756} 757 758FastISel::FastISel(MachineFunction &mf, 759 MachineModuleInfo *mmi, 760 DwarfWriter *dw, 761 DenseMap<const Value *, unsigned> &vm, 762 DenseMap<const BasicBlock *, MachineBasicBlock *> &bm, 763 DenseMap<const AllocaInst *, int> &am 764#ifndef NDEBUG 765 , SmallSet<Instruction*, 8> &cil 766#endif 767 ) 768 : MBB(0), 769 ValueMap(vm), 770 MBBMap(bm), 771 StaticAllocaMap(am), 772#ifndef NDEBUG 773 CatchInfoLost(cil), 774#endif 775 MF(mf), 776 MMI(mmi), 777 DW(dw), 778 MRI(MF.getRegInfo()), 779 MFI(*MF.getFrameInfo()), 780 MCP(*MF.getConstantPool()), 781 TM(MF.getTarget()), 782 TD(*TM.getTargetData()), 783 TII(*TM.getInstrInfo()), 784 TLI(*TM.getTargetLowering()) { 785} 786 787FastISel::~FastISel() {} 788 789unsigned FastISel::FastEmit_(MVT, MVT, 790 ISD::NodeType) { 791 return 0; 792} 793 794unsigned FastISel::FastEmit_r(MVT, MVT, 795 ISD::NodeType, unsigned /*Op0*/) { 796 return 0; 797} 798 799unsigned FastISel::FastEmit_rr(MVT, MVT, 800 ISD::NodeType, unsigned /*Op0*/, 801 unsigned /*Op0*/) { 802 return 0; 803} 804 805unsigned FastISel::FastEmit_i(MVT, MVT, ISD::NodeType, uint64_t /*Imm*/) { 806 return 0; 807} 808 809unsigned FastISel::FastEmit_f(MVT, MVT, 810 ISD::NodeType, ConstantFP * /*FPImm*/) { 811 return 0; 812} 813 814unsigned FastISel::FastEmit_ri(MVT, MVT, 815 ISD::NodeType, unsigned /*Op0*/, 816 uint64_t /*Imm*/) { 817 return 0; 818} 819 820unsigned FastISel::FastEmit_rf(MVT, MVT, 821 ISD::NodeType, unsigned /*Op0*/, 822 ConstantFP * /*FPImm*/) { 823 return 0; 824} 825 826unsigned FastISel::FastEmit_rri(MVT, MVT, 827 ISD::NodeType, 828 unsigned /*Op0*/, unsigned /*Op1*/, 829 uint64_t /*Imm*/) { 830 return 0; 831} 832 833/// FastEmit_ri_ - This method is a wrapper of FastEmit_ri. It first tries 834/// to emit an instruction with an immediate operand using FastEmit_ri. 835/// If that fails, it materializes the immediate into a register and try 836/// FastEmit_rr instead. 837unsigned FastISel::FastEmit_ri_(MVT VT, ISD::NodeType Opcode, 838 unsigned Op0, uint64_t Imm, 839 MVT ImmType) { 840 // First check if immediate type is legal. If not, we can't use the ri form. 841 unsigned ResultReg = FastEmit_ri(VT, VT, Opcode, Op0, Imm); 842 if (ResultReg != 0) 843 return ResultReg; 844 unsigned MaterialReg = FastEmit_i(ImmType, ImmType, ISD::Constant, Imm); 845 if (MaterialReg == 0) 846 return 0; 847 return FastEmit_rr(VT, VT, Opcode, Op0, MaterialReg); 848} 849 850/// FastEmit_rf_ - This method is a wrapper of FastEmit_ri. It first tries 851/// to emit an instruction with a floating-point immediate operand using 852/// FastEmit_rf. If that fails, it materializes the immediate into a register 853/// and try FastEmit_rr instead. 854unsigned FastISel::FastEmit_rf_(MVT VT, ISD::NodeType Opcode, 855 unsigned Op0, ConstantFP *FPImm, 856 MVT ImmType) { 857 // First check if immediate type is legal. If not, we can't use the rf form. 858 unsigned ResultReg = FastEmit_rf(VT, VT, Opcode, Op0, FPImm); 859 if (ResultReg != 0) 860 return ResultReg; 861 862 // Materialize the constant in a register. 863 unsigned MaterialReg = FastEmit_f(ImmType, ImmType, ISD::ConstantFP, FPImm); 864 if (MaterialReg == 0) { 865 // If the target doesn't have a way to directly enter a floating-point 866 // value into a register, use an alternate approach. 867 // TODO: The current approach only supports floating-point constants 868 // that can be constructed by conversion from integer values. This should 869 // be replaced by code that creates a load from a constant-pool entry, 870 // which will require some target-specific work. 871 const APFloat &Flt = FPImm->getValueAPF(); 872 EVT IntVT = TLI.getPointerTy(); 873 874 uint64_t x[2]; 875 uint32_t IntBitWidth = IntVT.getSizeInBits(); 876 bool isExact; 877 (void) Flt.convertToInteger(x, IntBitWidth, /*isSigned=*/true, 878 APFloat::rmTowardZero, &isExact); 879 if (!isExact) 880 return 0; 881 APInt IntVal(IntBitWidth, 2, x); 882 883 unsigned IntegerReg = FastEmit_i(IntVT.getSimpleVT(), IntVT.getSimpleVT(), 884 ISD::Constant, IntVal.getZExtValue()); 885 if (IntegerReg == 0) 886 return 0; 887 MaterialReg = FastEmit_r(IntVT.getSimpleVT(), VT, 888 ISD::SINT_TO_FP, IntegerReg); 889 if (MaterialReg == 0) 890 return 0; 891 } 892 return FastEmit_rr(VT, VT, Opcode, Op0, MaterialReg); 893} 894 895unsigned FastISel::createResultReg(const TargetRegisterClass* RC) { 896 return MRI.createVirtualRegister(RC); 897} 898 899unsigned FastISel::FastEmitInst_(unsigned MachineInstOpcode, 900 const TargetRegisterClass* RC) { 901 unsigned ResultReg = createResultReg(RC); 902 const TargetInstrDesc &II = TII.get(MachineInstOpcode); 903 904 BuildMI(MBB, DL, II, ResultReg); 905 return ResultReg; 906} 907 908unsigned FastISel::FastEmitInst_r(unsigned MachineInstOpcode, 909 const TargetRegisterClass *RC, 910 unsigned Op0) { 911 unsigned ResultReg = createResultReg(RC); 912 const TargetInstrDesc &II = TII.get(MachineInstOpcode); 913 914 if (II.getNumDefs() >= 1) 915 BuildMI(MBB, DL, II, ResultReg).addReg(Op0); 916 else { 917 BuildMI(MBB, DL, II).addReg(Op0); 918 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg, 919 II.ImplicitDefs[0], RC, RC); 920 if (!InsertedCopy) 921 ResultReg = 0; 922 } 923 924 return ResultReg; 925} 926 927unsigned FastISel::FastEmitInst_rr(unsigned MachineInstOpcode, 928 const TargetRegisterClass *RC, 929 unsigned Op0, unsigned Op1) { 930 unsigned ResultReg = createResultReg(RC); 931 const TargetInstrDesc &II = TII.get(MachineInstOpcode); 932 933 if (II.getNumDefs() >= 1) 934 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addReg(Op1); 935 else { 936 BuildMI(MBB, DL, II).addReg(Op0).addReg(Op1); 937 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg, 938 II.ImplicitDefs[0], RC, RC); 939 if (!InsertedCopy) 940 ResultReg = 0; 941 } 942 return ResultReg; 943} 944 945unsigned FastISel::FastEmitInst_ri(unsigned MachineInstOpcode, 946 const TargetRegisterClass *RC, 947 unsigned Op0, uint64_t Imm) { 948 unsigned ResultReg = createResultReg(RC); 949 const TargetInstrDesc &II = TII.get(MachineInstOpcode); 950 951 if (II.getNumDefs() >= 1) 952 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addImm(Imm); 953 else { 954 BuildMI(MBB, DL, II).addReg(Op0).addImm(Imm); 955 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg, 956 II.ImplicitDefs[0], RC, RC); 957 if (!InsertedCopy) 958 ResultReg = 0; 959 } 960 return ResultReg; 961} 962 963unsigned FastISel::FastEmitInst_rf(unsigned MachineInstOpcode, 964 const TargetRegisterClass *RC, 965 unsigned Op0, ConstantFP *FPImm) { 966 unsigned ResultReg = createResultReg(RC); 967 const TargetInstrDesc &II = TII.get(MachineInstOpcode); 968 969 if (II.getNumDefs() >= 1) 970 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addFPImm(FPImm); 971 else { 972 BuildMI(MBB, DL, II).addReg(Op0).addFPImm(FPImm); 973 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg, 974 II.ImplicitDefs[0], RC, RC); 975 if (!InsertedCopy) 976 ResultReg = 0; 977 } 978 return ResultReg; 979} 980 981unsigned FastISel::FastEmitInst_rri(unsigned MachineInstOpcode, 982 const TargetRegisterClass *RC, 983 unsigned Op0, unsigned Op1, uint64_t Imm) { 984 unsigned ResultReg = createResultReg(RC); 985 const TargetInstrDesc &II = TII.get(MachineInstOpcode); 986 987 if (II.getNumDefs() >= 1) 988 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addReg(Op1).addImm(Imm); 989 else { 990 BuildMI(MBB, DL, II).addReg(Op0).addReg(Op1).addImm(Imm); 991 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg, 992 II.ImplicitDefs[0], RC, RC); 993 if (!InsertedCopy) 994 ResultReg = 0; 995 } 996 return ResultReg; 997} 998 999unsigned FastISel::FastEmitInst_i(unsigned MachineInstOpcode, 1000 const TargetRegisterClass *RC, 1001 uint64_t Imm) { 1002 unsigned ResultReg = createResultReg(RC); 1003 const TargetInstrDesc &II = TII.get(MachineInstOpcode); 1004 1005 if (II.getNumDefs() >= 1) 1006 BuildMI(MBB, DL, II, ResultReg).addImm(Imm); 1007 else { 1008 BuildMI(MBB, DL, II).addImm(Imm); 1009 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg, 1010 II.ImplicitDefs[0], RC, RC); 1011 if (!InsertedCopy) 1012 ResultReg = 0; 1013 } 1014 return ResultReg; 1015} 1016 1017unsigned FastISel::FastEmitInst_extractsubreg(MVT RetVT, 1018 unsigned Op0, uint32_t Idx) { 1019 const TargetRegisterClass* RC = MRI.getRegClass(Op0); 1020 1021 unsigned ResultReg = createResultReg(TLI.getRegClassFor(RetVT)); 1022 const TargetInstrDesc &II = TII.get(TargetInstrInfo::EXTRACT_SUBREG); 1023 1024 if (II.getNumDefs() >= 1) 1025 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addImm(Idx); 1026 else { 1027 BuildMI(MBB, DL, II).addReg(Op0).addImm(Idx); 1028 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg, 1029 II.ImplicitDefs[0], RC, RC); 1030 if (!InsertedCopy) 1031 ResultReg = 0; 1032 } 1033 return ResultReg; 1034} 1035 1036/// FastEmitZExtFromI1 - Emit MachineInstrs to compute the value of Op 1037/// with all but the least significant bit set to zero. 1038unsigned FastISel::FastEmitZExtFromI1(MVT VT, unsigned Op) { 1039 return FastEmit_ri(VT, VT, ISD::AND, Op, 1); 1040} 1041