FastISel.cpp revision ccaea7dc0bbc75e5d4ee92fb8cc6af0f928d056f
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#define DEBUG_TYPE "isel" 43#include "llvm/Function.h" 44#include "llvm/GlobalVariable.h" 45#include "llvm/Instructions.h" 46#include "llvm/IntrinsicInst.h" 47#include "llvm/Operator.h" 48#include "llvm/CodeGen/Analysis.h" 49#include "llvm/CodeGen/FastISel.h" 50#include "llvm/CodeGen/FunctionLoweringInfo.h" 51#include "llvm/CodeGen/MachineInstrBuilder.h" 52#include "llvm/CodeGen/MachineModuleInfo.h" 53#include "llvm/CodeGen/MachineRegisterInfo.h" 54#include "llvm/Analysis/DebugInfo.h" 55#include "llvm/Analysis/Loads.h" 56#include "llvm/Target/TargetData.h" 57#include "llvm/Target/TargetInstrInfo.h" 58#include "llvm/Target/TargetLowering.h" 59#include "llvm/Target/TargetMachine.h" 60#include "llvm/Support/ErrorHandling.h" 61#include "llvm/Support/Debug.h" 62#include "llvm/ADT/Statistic.h" 63using namespace llvm; 64 65STATISTIC(NumFastIselSuccessIndependent, "Number of insts selected by " 66 "target-independent selector"); 67STATISTIC(NumFastIselSuccessTarget, "Number of insts selected by " 68 "target-specific selector"); 69STATISTIC(NumFastIselDead, "Number of dead insts removed on failure"); 70 71/// startNewBlock - Set the current block to which generated machine 72/// instructions will be appended, and clear the local CSE map. 73/// 74void FastISel::startNewBlock() { 75 LocalValueMap.clear(); 76 77 EmitStartPt = 0; 78 79 // Advance the emit start point past any EH_LABEL instructions. 80 MachineBasicBlock::iterator 81 I = FuncInfo.MBB->begin(), E = FuncInfo.MBB->end(); 82 while (I != E && I->getOpcode() == TargetOpcode::EH_LABEL) { 83 EmitStartPt = I; 84 ++I; 85 } 86 LastLocalValue = EmitStartPt; 87} 88 89void FastISel::flushLocalValueMap() { 90 LocalValueMap.clear(); 91 LastLocalValue = EmitStartPt; 92 recomputeInsertPt(); 93} 94 95bool FastISel::hasTrivialKill(const Value *V) const { 96 // Don't consider constants or arguments to have trivial kills. 97 const Instruction *I = dyn_cast<Instruction>(V); 98 if (!I) 99 return false; 100 101 // No-op casts are trivially coalesced by fast-isel. 102 if (const CastInst *Cast = dyn_cast<CastInst>(I)) 103 if (Cast->isNoopCast(TD.getIntPtrType(Cast->getContext())) && 104 !hasTrivialKill(Cast->getOperand(0))) 105 return false; 106 107 // GEPs with all zero indices are trivially coalesced by fast-isel. 108 if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) 109 if (GEP->hasAllZeroIndices() && !hasTrivialKill(GEP->getOperand(0))) 110 return false; 111 112 // Only instructions with a single use in the same basic block are considered 113 // to have trivial kills. 114 return I->hasOneUse() && 115 !(I->getOpcode() == Instruction::BitCast || 116 I->getOpcode() == Instruction::PtrToInt || 117 I->getOpcode() == Instruction::IntToPtr) && 118 cast<Instruction>(*I->use_begin())->getParent() == I->getParent(); 119} 120 121unsigned FastISel::getRegForValue(const Value *V) { 122 EVT RealVT = TLI.getValueType(V->getType(), /*AllowUnknown=*/true); 123 // Don't handle non-simple values in FastISel. 124 if (!RealVT.isSimple()) 125 return 0; 126 127 // Ignore illegal types. We must do this before looking up the value 128 // in ValueMap because Arguments are given virtual registers regardless 129 // of whether FastISel can handle them. 130 MVT VT = RealVT.getSimpleVT(); 131 if (!TLI.isTypeLegal(VT)) { 132 // Handle integer promotions, though, because they're common and easy. 133 if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16) 134 VT = TLI.getTypeToTransformTo(V->getContext(), VT).getSimpleVT(); 135 else 136 return 0; 137 } 138 139 // 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-dominates-use requirement enforced. 143 DenseMap<const Value *, unsigned>::iterator I = FuncInfo.ValueMap.find(V); 144 if (I != FuncInfo.ValueMap.end()) 145 return I->second; 146 147 unsigned Reg = LocalValueMap[V]; 148 if (Reg != 0) 149 return Reg; 150 151 // In bottom-up mode, just create the virtual register which will be used 152 // to hold the value. It will be materialized later. 153 if (isa<Instruction>(V) && 154 (!isa<AllocaInst>(V) || 155 !FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(V)))) 156 return FuncInfo.InitializeRegForValue(V); 157 158 SavePoint SaveInsertPt = enterLocalValueArea(); 159 160 // Materialize the value in a register. Emit any instructions in the 161 // local value area. 162 Reg = materializeRegForValue(V, VT); 163 164 leaveLocalValueArea(SaveInsertPt); 165 166 return Reg; 167} 168 169/// materializeRegForValue - Helper for getRegForValue. This function is 170/// called when the value isn't already available in a register and must 171/// be materialized with new instructions. 172unsigned FastISel::materializeRegForValue(const Value *V, MVT VT) { 173 unsigned Reg = 0; 174 175 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 176 if (CI->getValue().getActiveBits() <= 64) 177 Reg = FastEmit_i(VT, VT, ISD::Constant, CI->getZExtValue()); 178 } else if (isa<AllocaInst>(V)) { 179 Reg = TargetMaterializeAlloca(cast<AllocaInst>(V)); 180 } else if (isa<ConstantPointerNull>(V)) { 181 // Translate this as an integer zero so that it can be 182 // local-CSE'd with actual integer zeros. 183 Reg = 184 getRegForValue(Constant::getNullValue(TD.getIntPtrType(V->getContext()))); 185 } else if (const ConstantFP *CF = dyn_cast<ConstantFP>(V)) { 186 if (CF->isNullValue()) { 187 Reg = TargetMaterializeFloatZero(CF); 188 } else { 189 // Try to emit the constant directly. 190 Reg = FastEmit_f(VT, VT, ISD::ConstantFP, CF); 191 } 192 193 if (!Reg) { 194 // Try to emit the constant by using an integer constant with a cast. 195 const APFloat &Flt = CF->getValueAPF(); 196 EVT IntVT = TLI.getPointerTy(); 197 198 uint64_t x[2]; 199 uint32_t IntBitWidth = IntVT.getSizeInBits(); 200 bool isExact; 201 (void) Flt.convertToInteger(x, IntBitWidth, /*isSigned=*/true, 202 APFloat::rmTowardZero, &isExact); 203 if (isExact) { 204 APInt IntVal(IntBitWidth, x); 205 206 unsigned IntegerReg = 207 getRegForValue(ConstantInt::get(V->getContext(), IntVal)); 208 if (IntegerReg != 0) 209 Reg = FastEmit_r(IntVT.getSimpleVT(), VT, ISD::SINT_TO_FP, 210 IntegerReg, /*Kill=*/false); 211 } 212 } 213 } else if (const Operator *Op = dyn_cast<Operator>(V)) { 214 if (!SelectOperator(Op, Op->getOpcode())) 215 if (!isa<Instruction>(Op) || 216 !TargetSelectInstruction(cast<Instruction>(Op))) 217 return 0; 218 Reg = lookUpRegForValue(Op); 219 } else if (isa<UndefValue>(V)) { 220 Reg = createResultReg(TLI.getRegClassFor(VT)); 221 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, 222 TII.get(TargetOpcode::IMPLICIT_DEF), Reg); 223 } 224 225 // If target-independent code couldn't handle the value, give target-specific 226 // code a try. 227 if (!Reg && isa<Constant>(V)) 228 Reg = TargetMaterializeConstant(cast<Constant>(V)); 229 230 // Don't cache constant materializations in the general ValueMap. 231 // To do so would require tracking what uses they dominate. 232 if (Reg != 0) { 233 LocalValueMap[V] = Reg; 234 LastLocalValue = MRI.getVRegDef(Reg); 235 } 236 return Reg; 237} 238 239unsigned FastISel::lookUpRegForValue(const Value *V) { 240 // Look up the value to see if we already have a register for it. We 241 // cache values defined by Instructions across blocks, and other values 242 // only locally. This is because Instructions already have the SSA 243 // def-dominates-use requirement enforced. 244 DenseMap<const Value *, unsigned>::iterator I = FuncInfo.ValueMap.find(V); 245 if (I != FuncInfo.ValueMap.end()) 246 return I->second; 247 return LocalValueMap[V]; 248} 249 250/// UpdateValueMap - Update the value map to include the new mapping for this 251/// instruction, or insert an extra copy to get the result in a previous 252/// determined register. 253/// NOTE: This is only necessary because we might select a block that uses 254/// a value before we select the block that defines the value. It might be 255/// possible to fix this by selecting blocks in reverse postorder. 256void FastISel::UpdateValueMap(const Value *I, unsigned Reg, unsigned NumRegs) { 257 if (!isa<Instruction>(I)) { 258 LocalValueMap[I] = Reg; 259 return; 260 } 261 262 unsigned &AssignedReg = FuncInfo.ValueMap[I]; 263 if (AssignedReg == 0) 264 // Use the new register. 265 AssignedReg = Reg; 266 else if (Reg != AssignedReg) { 267 // Arrange for uses of AssignedReg to be replaced by uses of Reg. 268 for (unsigned i = 0; i < NumRegs; i++) 269 FuncInfo.RegFixups[AssignedReg+i] = Reg+i; 270 271 AssignedReg = Reg; 272 } 273} 274 275std::pair<unsigned, bool> FastISel::getRegForGEPIndex(const Value *Idx) { 276 unsigned IdxN = getRegForValue(Idx); 277 if (IdxN == 0) 278 // Unhandled operand. Halt "fast" selection and bail. 279 return std::pair<unsigned, bool>(0, false); 280 281 bool IdxNIsKill = hasTrivialKill(Idx); 282 283 // If the index is smaller or larger than intptr_t, truncate or extend it. 284 MVT PtrVT = TLI.getPointerTy(); 285 EVT IdxVT = EVT::getEVT(Idx->getType(), /*HandleUnknown=*/false); 286 if (IdxVT.bitsLT(PtrVT)) { 287 IdxN = FastEmit_r(IdxVT.getSimpleVT(), PtrVT, ISD::SIGN_EXTEND, 288 IdxN, IdxNIsKill); 289 IdxNIsKill = true; 290 } 291 else if (IdxVT.bitsGT(PtrVT)) { 292 IdxN = FastEmit_r(IdxVT.getSimpleVT(), PtrVT, ISD::TRUNCATE, 293 IdxN, IdxNIsKill); 294 IdxNIsKill = true; 295 } 296 return std::pair<unsigned, bool>(IdxN, IdxNIsKill); 297} 298 299void FastISel::recomputeInsertPt() { 300 if (getLastLocalValue()) { 301 FuncInfo.InsertPt = getLastLocalValue(); 302 FuncInfo.MBB = FuncInfo.InsertPt->getParent(); 303 ++FuncInfo.InsertPt; 304 } else 305 FuncInfo.InsertPt = FuncInfo.MBB->getFirstNonPHI(); 306 307 // Now skip past any EH_LABELs, which must remain at the beginning. 308 while (FuncInfo.InsertPt != FuncInfo.MBB->end() && 309 FuncInfo.InsertPt->getOpcode() == TargetOpcode::EH_LABEL) 310 ++FuncInfo.InsertPt; 311} 312 313void FastISel::removeDeadCode(MachineBasicBlock::iterator I, 314 MachineBasicBlock::iterator E) { 315 assert (I && E && std::distance(I, E) > 0 && "Invalid iterator!"); 316 while (I != E) { 317 MachineInstr *Dead = &*I; 318 ++I; 319 Dead->eraseFromParent(); 320 ++NumFastIselDead; 321 } 322 recomputeInsertPt(); 323} 324 325FastISel::SavePoint FastISel::enterLocalValueArea() { 326 MachineBasicBlock::iterator OldInsertPt = FuncInfo.InsertPt; 327 DebugLoc OldDL = DL; 328 recomputeInsertPt(); 329 DL = DebugLoc(); 330 SavePoint SP = { OldInsertPt, OldDL }; 331 return SP; 332} 333 334void FastISel::leaveLocalValueArea(SavePoint OldInsertPt) { 335 if (FuncInfo.InsertPt != FuncInfo.MBB->begin()) 336 LastLocalValue = llvm::prior(FuncInfo.InsertPt); 337 338 // Restore the previous insert position. 339 FuncInfo.InsertPt = OldInsertPt.InsertPt; 340 DL = OldInsertPt.DL; 341} 342 343/// SelectBinaryOp - Select and emit code for a binary operator instruction, 344/// which has an opcode which directly corresponds to the given ISD opcode. 345/// 346bool FastISel::SelectBinaryOp(const User *I, unsigned ISDOpcode) { 347 EVT VT = EVT::getEVT(I->getType(), /*HandleUnknown=*/true); 348 if (VT == MVT::Other || !VT.isSimple()) 349 // Unhandled type. Halt "fast" selection and bail. 350 return false; 351 352 // We only handle legal types. For example, on x86-32 the instruction 353 // selector contains all of the 64-bit instructions from x86-64, 354 // under the assumption that i64 won't be used if the target doesn't 355 // support it. 356 if (!TLI.isTypeLegal(VT)) { 357 // MVT::i1 is special. Allow AND, OR, or XOR because they 358 // don't require additional zeroing, which makes them easy. 359 if (VT == MVT::i1 && 360 (ISDOpcode == ISD::AND || ISDOpcode == ISD::OR || 361 ISDOpcode == ISD::XOR)) 362 VT = TLI.getTypeToTransformTo(I->getContext(), VT); 363 else 364 return false; 365 } 366 367 // Check if the first operand is a constant, and handle it as "ri". At -O0, 368 // we don't have anything that canonicalizes operand order. 369 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(0))) 370 if (isa<Instruction>(I) && cast<Instruction>(I)->isCommutative()) { 371 unsigned Op1 = getRegForValue(I->getOperand(1)); 372 if (Op1 == 0) return false; 373 374 bool Op1IsKill = hasTrivialKill(I->getOperand(1)); 375 376 unsigned ResultReg = FastEmit_ri_(VT.getSimpleVT(), ISDOpcode, Op1, 377 Op1IsKill, CI->getZExtValue(), 378 VT.getSimpleVT()); 379 if (ResultReg == 0) return false; 380 381 // We successfully emitted code for the given LLVM Instruction. 382 UpdateValueMap(I, ResultReg); 383 return true; 384 } 385 386 387 unsigned Op0 = getRegForValue(I->getOperand(0)); 388 if (Op0 == 0) // Unhandled operand. Halt "fast" selection and bail. 389 return false; 390 391 bool Op0IsKill = hasTrivialKill(I->getOperand(0)); 392 393 // Check if the second operand is a constant and handle it appropriately. 394 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) { 395 uint64_t Imm = CI->getZExtValue(); 396 397 // Transform "sdiv exact X, 8" -> "sra X, 3". 398 if (ISDOpcode == ISD::SDIV && isa<BinaryOperator>(I) && 399 cast<BinaryOperator>(I)->isExact() && 400 isPowerOf2_64(Imm)) { 401 Imm = Log2_64(Imm); 402 ISDOpcode = ISD::SRA; 403 } 404 405 unsigned ResultReg = FastEmit_ri_(VT.getSimpleVT(), ISDOpcode, Op0, 406 Op0IsKill, Imm, VT.getSimpleVT()); 407 if (ResultReg == 0) return false; 408 409 // We successfully emitted code for the given LLVM Instruction. 410 UpdateValueMap(I, ResultReg); 411 return true; 412 } 413 414 // Check if the second operand is a constant float. 415 if (ConstantFP *CF = dyn_cast<ConstantFP>(I->getOperand(1))) { 416 unsigned ResultReg = FastEmit_rf(VT.getSimpleVT(), VT.getSimpleVT(), 417 ISDOpcode, Op0, Op0IsKill, CF); 418 if (ResultReg != 0) { 419 // We successfully emitted code for the given LLVM Instruction. 420 UpdateValueMap(I, ResultReg); 421 return true; 422 } 423 } 424 425 unsigned Op1 = getRegForValue(I->getOperand(1)); 426 if (Op1 == 0) 427 // Unhandled operand. Halt "fast" selection and bail. 428 return false; 429 430 bool Op1IsKill = hasTrivialKill(I->getOperand(1)); 431 432 // Now we have both operands in registers. Emit the instruction. 433 unsigned ResultReg = FastEmit_rr(VT.getSimpleVT(), VT.getSimpleVT(), 434 ISDOpcode, 435 Op0, Op0IsKill, 436 Op1, Op1IsKill); 437 if (ResultReg == 0) 438 // Target-specific code wasn't able to find a machine opcode for 439 // the given ISD opcode and type. Halt "fast" selection and bail. 440 return false; 441 442 // We successfully emitted code for the given LLVM Instruction. 443 UpdateValueMap(I, ResultReg); 444 return true; 445} 446 447bool FastISel::SelectGetElementPtr(const User *I) { 448 unsigned N = getRegForValue(I->getOperand(0)); 449 if (N == 0) 450 // Unhandled operand. Halt "fast" selection and bail. 451 return false; 452 453 bool NIsKill = hasTrivialKill(I->getOperand(0)); 454 455 // Keep a running tab of the total offset to coalesce multiple N = N + Offset 456 // into a single N = N + TotalOffset. 457 uint64_t TotalOffs = 0; 458 // FIXME: What's a good SWAG number for MaxOffs? 459 uint64_t MaxOffs = 2048; 460 Type *Ty = I->getOperand(0)->getType(); 461 MVT VT = TLI.getPointerTy(); 462 for (GetElementPtrInst::const_op_iterator OI = I->op_begin()+1, 463 E = I->op_end(); OI != E; ++OI) { 464 const Value *Idx = *OI; 465 if (StructType *StTy = dyn_cast<StructType>(Ty)) { 466 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue(); 467 if (Field) { 468 // N = N + Offset 469 TotalOffs += TD.getStructLayout(StTy)->getElementOffset(Field); 470 if (TotalOffs >= MaxOffs) { 471 N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, TotalOffs, VT); 472 if (N == 0) 473 // Unhandled operand. Halt "fast" selection and bail. 474 return false; 475 NIsKill = true; 476 TotalOffs = 0; 477 } 478 } 479 Ty = StTy->getElementType(Field); 480 } else { 481 Ty = cast<SequentialType>(Ty)->getElementType(); 482 483 // If this is a constant subscript, handle it quickly. 484 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) { 485 if (CI->isZero()) continue; 486 // N = N + Offset 487 TotalOffs += 488 TD.getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue(); 489 if (TotalOffs >= MaxOffs) { 490 N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, TotalOffs, VT); 491 if (N == 0) 492 // Unhandled operand. Halt "fast" selection and bail. 493 return false; 494 NIsKill = true; 495 TotalOffs = 0; 496 } 497 continue; 498 } 499 if (TotalOffs) { 500 N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, TotalOffs, VT); 501 if (N == 0) 502 // Unhandled operand. Halt "fast" selection and bail. 503 return false; 504 NIsKill = true; 505 TotalOffs = 0; 506 } 507 508 // N = N + Idx * ElementSize; 509 uint64_t ElementSize = TD.getTypeAllocSize(Ty); 510 std::pair<unsigned, bool> Pair = getRegForGEPIndex(Idx); 511 unsigned IdxN = Pair.first; 512 bool IdxNIsKill = Pair.second; 513 if (IdxN == 0) 514 // Unhandled operand. Halt "fast" selection and bail. 515 return false; 516 517 if (ElementSize != 1) { 518 IdxN = FastEmit_ri_(VT, ISD::MUL, IdxN, IdxNIsKill, ElementSize, VT); 519 if (IdxN == 0) 520 // Unhandled operand. Halt "fast" selection and bail. 521 return false; 522 IdxNIsKill = true; 523 } 524 N = FastEmit_rr(VT, VT, ISD::ADD, N, NIsKill, IdxN, IdxNIsKill); 525 if (N == 0) 526 // Unhandled operand. Halt "fast" selection and bail. 527 return false; 528 } 529 } 530 if (TotalOffs) { 531 N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, TotalOffs, VT); 532 if (N == 0) 533 // Unhandled operand. Halt "fast" selection and bail. 534 return false; 535 } 536 537 // We successfully emitted code for the given LLVM Instruction. 538 UpdateValueMap(I, N); 539 return true; 540} 541 542bool FastISel::SelectCall(const User *I) { 543 const CallInst *Call = cast<CallInst>(I); 544 545 // Handle simple inline asms. 546 if (const InlineAsm *IA = dyn_cast<InlineAsm>(Call->getCalledValue())) { 547 // Don't attempt to handle constraints. 548 if (!IA->getConstraintString().empty()) 549 return false; 550 551 unsigned ExtraInfo = 0; 552 if (IA->hasSideEffects()) 553 ExtraInfo |= InlineAsm::Extra_HasSideEffects; 554 if (IA->isAlignStack()) 555 ExtraInfo |= InlineAsm::Extra_IsAlignStack; 556 557 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, 558 TII.get(TargetOpcode::INLINEASM)) 559 .addExternalSymbol(IA->getAsmString().c_str()) 560 .addImm(ExtraInfo); 561 return true; 562 } 563 564 MachineModuleInfo &MMI = FuncInfo.MF->getMMI(); 565 ComputeUsesVAFloatArgument(*Call, &MMI); 566 567 const Function *F = Call->getCalledFunction(); 568 if (!F) return false; 569 570 // Handle selected intrinsic function calls. 571 switch (F->getIntrinsicID()) { 572 default: break; 573 // At -O0 we don't care about the lifetime intrinsics. 574 case Intrinsic::lifetime_start: 575 case Intrinsic::lifetime_end: 576 return true; 577 case Intrinsic::dbg_declare: { 578 const DbgDeclareInst *DI = cast<DbgDeclareInst>(Call); 579 if (!DIVariable(DI->getVariable()).Verify() || 580 !FuncInfo.MF->getMMI().hasDebugInfo()) { 581 DEBUG(dbgs() << "Dropping debug info for " << *DI << "\n"); 582 return true; 583 } 584 585 const Value *Address = DI->getAddress(); 586 if (!Address || isa<UndefValue>(Address)) { 587 DEBUG(dbgs() << "Dropping debug info for " << *DI << "\n"); 588 return true; 589 } 590 591 unsigned Reg = 0; 592 unsigned Offset = 0; 593 if (const Argument *Arg = dyn_cast<Argument>(Address)) { 594 // Some arguments' frame index is recorded during argument lowering. 595 Offset = FuncInfo.getArgumentFrameIndex(Arg); 596 if (Offset) 597 Reg = TRI.getFrameRegister(*FuncInfo.MF); 598 } 599 if (!Reg) 600 Reg = getRegForValue(Address); 601 602 if (Reg) 603 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, 604 TII.get(TargetOpcode::DBG_VALUE)) 605 .addReg(Reg, RegState::Debug).addImm(Offset) 606 .addMetadata(DI->getVariable()); 607 return true; 608 } 609 case Intrinsic::dbg_value: { 610 // This form of DBG_VALUE is target-independent. 611 const DbgValueInst *DI = cast<DbgValueInst>(Call); 612 const MCInstrDesc &II = TII.get(TargetOpcode::DBG_VALUE); 613 const Value *V = DI->getValue(); 614 if (!V) { 615 // Currently the optimizer can produce this; insert an undef to 616 // help debugging. Probably the optimizer should not do this. 617 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II) 618 .addReg(0U).addImm(DI->getOffset()) 619 .addMetadata(DI->getVariable()); 620 } else if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 621 if (CI->getBitWidth() > 64) 622 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II) 623 .addCImm(CI).addImm(DI->getOffset()) 624 .addMetadata(DI->getVariable()); 625 else 626 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II) 627 .addImm(CI->getZExtValue()).addImm(DI->getOffset()) 628 .addMetadata(DI->getVariable()); 629 } else if (const ConstantFP *CF = dyn_cast<ConstantFP>(V)) { 630 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II) 631 .addFPImm(CF).addImm(DI->getOffset()) 632 .addMetadata(DI->getVariable()); 633 } else if (unsigned Reg = lookUpRegForValue(V)) { 634 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II) 635 .addReg(Reg, RegState::Debug).addImm(DI->getOffset()) 636 .addMetadata(DI->getVariable()); 637 } else { 638 // We can't yet handle anything else here because it would require 639 // generating code, thus altering codegen because of debug info. 640 DEBUG(dbgs() << "Dropping debug info for " << DI); 641 } 642 return true; 643 } 644 case Intrinsic::objectsize: { 645 ConstantInt *CI = cast<ConstantInt>(Call->getArgOperand(1)); 646 unsigned long long Res = CI->isZero() ? -1ULL : 0; 647 Constant *ResCI = ConstantInt::get(Call->getType(), Res); 648 unsigned ResultReg = getRegForValue(ResCI); 649 if (ResultReg == 0) 650 return false; 651 UpdateValueMap(Call, ResultReg); 652 return true; 653 } 654 } 655 656 // Usually, it does not make sense to initialize a value, 657 // make an unrelated function call and use the value, because 658 // it tends to be spilled on the stack. So, we move the pointer 659 // to the last local value to the beginning of the block, so that 660 // all the values which have already been materialized, 661 // appear after the call. It also makes sense to skip intrinsics 662 // since they tend to be inlined. 663 if (!isa<IntrinsicInst>(F)) 664 flushLocalValueMap(); 665 666 // An arbitrary call. Bail. 667 return false; 668} 669 670bool FastISel::SelectCast(const User *I, unsigned Opcode) { 671 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType()); 672 EVT DstVT = TLI.getValueType(I->getType()); 673 674 if (SrcVT == MVT::Other || !SrcVT.isSimple() || 675 DstVT == MVT::Other || !DstVT.isSimple()) 676 // Unhandled type. Halt "fast" selection and bail. 677 return false; 678 679 // Check if the destination type is legal. 680 if (!TLI.isTypeLegal(DstVT)) 681 return false; 682 683 // Check if the source operand is legal. 684 if (!TLI.isTypeLegal(SrcVT)) 685 return false; 686 687 unsigned InputReg = getRegForValue(I->getOperand(0)); 688 if (!InputReg) 689 // Unhandled operand. Halt "fast" selection and bail. 690 return false; 691 692 bool InputRegIsKill = hasTrivialKill(I->getOperand(0)); 693 694 unsigned ResultReg = FastEmit_r(SrcVT.getSimpleVT(), 695 DstVT.getSimpleVT(), 696 Opcode, 697 InputReg, InputRegIsKill); 698 if (!ResultReg) 699 return false; 700 701 UpdateValueMap(I, ResultReg); 702 return true; 703} 704 705bool FastISel::SelectBitCast(const User *I) { 706 // If the bitcast doesn't change the type, just use the operand value. 707 if (I->getType() == I->getOperand(0)->getType()) { 708 unsigned Reg = getRegForValue(I->getOperand(0)); 709 if (Reg == 0) 710 return false; 711 UpdateValueMap(I, Reg); 712 return true; 713 } 714 715 // Bitcasts of other values become reg-reg copies or BITCAST operators. 716 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType()); 717 EVT DstVT = TLI.getValueType(I->getType()); 718 719 if (SrcVT == MVT::Other || !SrcVT.isSimple() || 720 DstVT == MVT::Other || !DstVT.isSimple() || 721 !TLI.isTypeLegal(SrcVT) || !TLI.isTypeLegal(DstVT)) 722 // Unhandled type. Halt "fast" selection and bail. 723 return false; 724 725 unsigned Op0 = getRegForValue(I->getOperand(0)); 726 if (Op0 == 0) 727 // Unhandled operand. Halt "fast" selection and bail. 728 return false; 729 730 bool Op0IsKill = hasTrivialKill(I->getOperand(0)); 731 732 // First, try to perform the bitcast by inserting a reg-reg copy. 733 unsigned ResultReg = 0; 734 if (SrcVT.getSimpleVT() == DstVT.getSimpleVT()) { 735 const TargetRegisterClass* SrcClass = TLI.getRegClassFor(SrcVT); 736 const TargetRegisterClass* DstClass = TLI.getRegClassFor(DstVT); 737 // Don't attempt a cross-class copy. It will likely fail. 738 if (SrcClass == DstClass) { 739 ResultReg = createResultReg(DstClass); 740 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), 741 ResultReg).addReg(Op0); 742 } 743 } 744 745 // If the reg-reg copy failed, select a BITCAST opcode. 746 if (!ResultReg) 747 ResultReg = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), 748 ISD::BITCAST, Op0, Op0IsKill); 749 750 if (!ResultReg) 751 return false; 752 753 UpdateValueMap(I, ResultReg); 754 return true; 755} 756 757bool 758FastISel::SelectInstruction(const Instruction *I) { 759 // Just before the terminator instruction, insert instructions to 760 // feed PHI nodes in successor blocks. 761 if (isa<TerminatorInst>(I)) 762 if (!HandlePHINodesInSuccessorBlocks(I->getParent())) 763 return false; 764 765 DL = I->getDebugLoc(); 766 767 MachineBasicBlock::iterator SavedInsertPt = FuncInfo.InsertPt; 768 769 // First, try doing target-independent selection. 770 if (SelectOperator(I, I->getOpcode())) { 771 ++NumFastIselSuccessIndependent; 772 DL = DebugLoc(); 773 return true; 774 } 775 // Remove dead code. However, ignore call instructions since we've flushed 776 // the local value map and recomputed the insert point. 777 if (!isa<CallInst>(I)) { 778 recomputeInsertPt(); 779 if (SavedInsertPt != FuncInfo.InsertPt) 780 removeDeadCode(FuncInfo.InsertPt, SavedInsertPt); 781 } 782 783 // Next, try calling the target to attempt to handle the instruction. 784 SavedInsertPt = FuncInfo.InsertPt; 785 if (TargetSelectInstruction(I)) { 786 ++NumFastIselSuccessTarget; 787 DL = DebugLoc(); 788 return true; 789 } 790 // Check for dead code and remove as necessary. 791 recomputeInsertPt(); 792 if (SavedInsertPt != FuncInfo.InsertPt) 793 removeDeadCode(FuncInfo.InsertPt, SavedInsertPt); 794 795 DL = DebugLoc(); 796 return false; 797} 798 799/// FastEmitBranch - Emit an unconditional branch to the given block, 800/// unless it is the immediate (fall-through) successor, and update 801/// the CFG. 802void 803FastISel::FastEmitBranch(MachineBasicBlock *MSucc, DebugLoc DL) { 804 if (FuncInfo.MBB->isLayoutSuccessor(MSucc)) { 805 // The unconditional fall-through case, which needs no instructions. 806 } else { 807 // The unconditional branch case. 808 TII.InsertBranch(*FuncInfo.MBB, MSucc, NULL, 809 SmallVector<MachineOperand, 0>(), DL); 810 } 811 FuncInfo.MBB->addSuccessor(MSucc); 812} 813 814/// SelectFNeg - Emit an FNeg operation. 815/// 816bool 817FastISel::SelectFNeg(const User *I) { 818 unsigned OpReg = getRegForValue(BinaryOperator::getFNegArgument(I)); 819 if (OpReg == 0) return false; 820 821 bool OpRegIsKill = hasTrivialKill(I); 822 823 // If the target has ISD::FNEG, use it. 824 EVT VT = TLI.getValueType(I->getType()); 825 unsigned ResultReg = FastEmit_r(VT.getSimpleVT(), VT.getSimpleVT(), 826 ISD::FNEG, OpReg, OpRegIsKill); 827 if (ResultReg != 0) { 828 UpdateValueMap(I, ResultReg); 829 return true; 830 } 831 832 // Bitcast the value to integer, twiddle the sign bit with xor, 833 // and then bitcast it back to floating-point. 834 if (VT.getSizeInBits() > 64) return false; 835 EVT IntVT = EVT::getIntegerVT(I->getContext(), VT.getSizeInBits()); 836 if (!TLI.isTypeLegal(IntVT)) 837 return false; 838 839 unsigned IntReg = FastEmit_r(VT.getSimpleVT(), IntVT.getSimpleVT(), 840 ISD::BITCAST, OpReg, OpRegIsKill); 841 if (IntReg == 0) 842 return false; 843 844 unsigned IntResultReg = FastEmit_ri_(IntVT.getSimpleVT(), ISD::XOR, 845 IntReg, /*Kill=*/true, 846 UINT64_C(1) << (VT.getSizeInBits()-1), 847 IntVT.getSimpleVT()); 848 if (IntResultReg == 0) 849 return false; 850 851 ResultReg = FastEmit_r(IntVT.getSimpleVT(), VT.getSimpleVT(), 852 ISD::BITCAST, IntResultReg, /*Kill=*/true); 853 if (ResultReg == 0) 854 return false; 855 856 UpdateValueMap(I, ResultReg); 857 return true; 858} 859 860bool 861FastISel::SelectExtractValue(const User *U) { 862 const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(U); 863 if (!EVI) 864 return false; 865 866 // Make sure we only try to handle extracts with a legal result. But also 867 // allow i1 because it's easy. 868 EVT RealVT = TLI.getValueType(EVI->getType(), /*AllowUnknown=*/true); 869 if (!RealVT.isSimple()) 870 return false; 871 MVT VT = RealVT.getSimpleVT(); 872 if (!TLI.isTypeLegal(VT) && VT != MVT::i1) 873 return false; 874 875 const Value *Op0 = EVI->getOperand(0); 876 Type *AggTy = Op0->getType(); 877 878 // Get the base result register. 879 unsigned ResultReg; 880 DenseMap<const Value *, unsigned>::iterator I = FuncInfo.ValueMap.find(Op0); 881 if (I != FuncInfo.ValueMap.end()) 882 ResultReg = I->second; 883 else if (isa<Instruction>(Op0)) 884 ResultReg = FuncInfo.InitializeRegForValue(Op0); 885 else 886 return false; // fast-isel can't handle aggregate constants at the moment 887 888 // Get the actual result register, which is an offset from the base register. 889 unsigned VTIndex = ComputeLinearIndex(AggTy, EVI->getIndices()); 890 891 SmallVector<EVT, 4> AggValueVTs; 892 ComputeValueVTs(TLI, AggTy, AggValueVTs); 893 894 for (unsigned i = 0; i < VTIndex; i++) 895 ResultReg += TLI.getNumRegisters(FuncInfo.Fn->getContext(), AggValueVTs[i]); 896 897 UpdateValueMap(EVI, ResultReg); 898 return true; 899} 900 901bool 902FastISel::SelectOperator(const User *I, unsigned Opcode) { 903 switch (Opcode) { 904 case Instruction::Add: 905 return SelectBinaryOp(I, ISD::ADD); 906 case Instruction::FAdd: 907 return SelectBinaryOp(I, ISD::FADD); 908 case Instruction::Sub: 909 return SelectBinaryOp(I, ISD::SUB); 910 case Instruction::FSub: 911 // FNeg is currently represented in LLVM IR as a special case of FSub. 912 if (BinaryOperator::isFNeg(I)) 913 return SelectFNeg(I); 914 return SelectBinaryOp(I, ISD::FSUB); 915 case Instruction::Mul: 916 return SelectBinaryOp(I, ISD::MUL); 917 case Instruction::FMul: 918 return SelectBinaryOp(I, ISD::FMUL); 919 case Instruction::SDiv: 920 return SelectBinaryOp(I, ISD::SDIV); 921 case Instruction::UDiv: 922 return SelectBinaryOp(I, ISD::UDIV); 923 case Instruction::FDiv: 924 return SelectBinaryOp(I, ISD::FDIV); 925 case Instruction::SRem: 926 return SelectBinaryOp(I, ISD::SREM); 927 case Instruction::URem: 928 return SelectBinaryOp(I, ISD::UREM); 929 case Instruction::FRem: 930 return SelectBinaryOp(I, ISD::FREM); 931 case Instruction::Shl: 932 return SelectBinaryOp(I, ISD::SHL); 933 case Instruction::LShr: 934 return SelectBinaryOp(I, ISD::SRL); 935 case Instruction::AShr: 936 return SelectBinaryOp(I, ISD::SRA); 937 case Instruction::And: 938 return SelectBinaryOp(I, ISD::AND); 939 case Instruction::Or: 940 return SelectBinaryOp(I, ISD::OR); 941 case Instruction::Xor: 942 return SelectBinaryOp(I, ISD::XOR); 943 944 case Instruction::GetElementPtr: 945 return SelectGetElementPtr(I); 946 947 case Instruction::Br: { 948 const BranchInst *BI = cast<BranchInst>(I); 949 950 if (BI->isUnconditional()) { 951 const BasicBlock *LLVMSucc = BI->getSuccessor(0); 952 MachineBasicBlock *MSucc = FuncInfo.MBBMap[LLVMSucc]; 953 FastEmitBranch(MSucc, BI->getDebugLoc()); 954 return true; 955 } 956 957 // Conditional branches are not handed yet. 958 // Halt "fast" selection and bail. 959 return false; 960 } 961 962 case Instruction::Unreachable: 963 // Nothing to emit. 964 return true; 965 966 case Instruction::Alloca: 967 // FunctionLowering has the static-sized case covered. 968 if (FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(I))) 969 return true; 970 971 // Dynamic-sized alloca is not handled yet. 972 return false; 973 974 case Instruction::Call: 975 return SelectCall(I); 976 977 case Instruction::BitCast: 978 return SelectBitCast(I); 979 980 case Instruction::FPToSI: 981 return SelectCast(I, ISD::FP_TO_SINT); 982 case Instruction::ZExt: 983 return SelectCast(I, ISD::ZERO_EXTEND); 984 case Instruction::SExt: 985 return SelectCast(I, ISD::SIGN_EXTEND); 986 case Instruction::Trunc: 987 return SelectCast(I, ISD::TRUNCATE); 988 case Instruction::SIToFP: 989 return SelectCast(I, ISD::SINT_TO_FP); 990 991 case Instruction::IntToPtr: // Deliberate fall-through. 992 case Instruction::PtrToInt: { 993 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType()); 994 EVT DstVT = TLI.getValueType(I->getType()); 995 if (DstVT.bitsGT(SrcVT)) 996 return SelectCast(I, ISD::ZERO_EXTEND); 997 if (DstVT.bitsLT(SrcVT)) 998 return SelectCast(I, ISD::TRUNCATE); 999 unsigned Reg = getRegForValue(I->getOperand(0)); 1000 if (Reg == 0) return false; 1001 UpdateValueMap(I, Reg); 1002 return true; 1003 } 1004 1005 case Instruction::ExtractValue: 1006 return SelectExtractValue(I); 1007 1008 case Instruction::PHI: 1009 llvm_unreachable("FastISel shouldn't visit PHI nodes!"); 1010 1011 default: 1012 // Unhandled instruction. Halt "fast" selection and bail. 1013 return false; 1014 } 1015} 1016 1017FastISel::FastISel(FunctionLoweringInfo &funcInfo) 1018 : FuncInfo(funcInfo), 1019 MRI(FuncInfo.MF->getRegInfo()), 1020 MFI(*FuncInfo.MF->getFrameInfo()), 1021 MCP(*FuncInfo.MF->getConstantPool()), 1022 TM(FuncInfo.MF->getTarget()), 1023 TD(*TM.getTargetData()), 1024 TII(*TM.getInstrInfo()), 1025 TLI(*TM.getTargetLowering()), 1026 TRI(*TM.getRegisterInfo()) { 1027} 1028 1029FastISel::~FastISel() {} 1030 1031unsigned FastISel::FastEmit_(MVT, MVT, 1032 unsigned) { 1033 return 0; 1034} 1035 1036unsigned FastISel::FastEmit_r(MVT, MVT, 1037 unsigned, 1038 unsigned /*Op0*/, bool /*Op0IsKill*/) { 1039 return 0; 1040} 1041 1042unsigned FastISel::FastEmit_rr(MVT, MVT, 1043 unsigned, 1044 unsigned /*Op0*/, bool /*Op0IsKill*/, 1045 unsigned /*Op1*/, bool /*Op1IsKill*/) { 1046 return 0; 1047} 1048 1049unsigned FastISel::FastEmit_i(MVT, MVT, unsigned, uint64_t /*Imm*/) { 1050 return 0; 1051} 1052 1053unsigned FastISel::FastEmit_f(MVT, MVT, 1054 unsigned, const ConstantFP * /*FPImm*/) { 1055 return 0; 1056} 1057 1058unsigned FastISel::FastEmit_ri(MVT, MVT, 1059 unsigned, 1060 unsigned /*Op0*/, bool /*Op0IsKill*/, 1061 uint64_t /*Imm*/) { 1062 return 0; 1063} 1064 1065unsigned FastISel::FastEmit_rf(MVT, MVT, 1066 unsigned, 1067 unsigned /*Op0*/, bool /*Op0IsKill*/, 1068 const ConstantFP * /*FPImm*/) { 1069 return 0; 1070} 1071 1072unsigned FastISel::FastEmit_rri(MVT, MVT, 1073 unsigned, 1074 unsigned /*Op0*/, bool /*Op0IsKill*/, 1075 unsigned /*Op1*/, bool /*Op1IsKill*/, 1076 uint64_t /*Imm*/) { 1077 return 0; 1078} 1079 1080/// FastEmit_ri_ - This method is a wrapper of FastEmit_ri. It first tries 1081/// to emit an instruction with an immediate operand using FastEmit_ri. 1082/// If that fails, it materializes the immediate into a register and try 1083/// FastEmit_rr instead. 1084unsigned FastISel::FastEmit_ri_(MVT VT, unsigned Opcode, 1085 unsigned Op0, bool Op0IsKill, 1086 uint64_t Imm, MVT ImmType) { 1087 // If this is a multiply by a power of two, emit this as a shift left. 1088 if (Opcode == ISD::MUL && isPowerOf2_64(Imm)) { 1089 Opcode = ISD::SHL; 1090 Imm = Log2_64(Imm); 1091 } else if (Opcode == ISD::UDIV && isPowerOf2_64(Imm)) { 1092 // div x, 8 -> srl x, 3 1093 Opcode = ISD::SRL; 1094 Imm = Log2_64(Imm); 1095 } 1096 1097 // Horrible hack (to be removed), check to make sure shift amounts are 1098 // in-range. 1099 if ((Opcode == ISD::SHL || Opcode == ISD::SRA || Opcode == ISD::SRL) && 1100 Imm >= VT.getSizeInBits()) 1101 return 0; 1102 1103 // First check if immediate type is legal. If not, we can't use the ri form. 1104 unsigned ResultReg = FastEmit_ri(VT, VT, Opcode, Op0, Op0IsKill, Imm); 1105 if (ResultReg != 0) 1106 return ResultReg; 1107 unsigned MaterialReg = FastEmit_i(ImmType, ImmType, ISD::Constant, Imm); 1108 if (MaterialReg == 0) { 1109 // This is a bit ugly/slow, but failing here means falling out of 1110 // fast-isel, which would be very slow. 1111 IntegerType *ITy = IntegerType::get(FuncInfo.Fn->getContext(), 1112 VT.getSizeInBits()); 1113 MaterialReg = getRegForValue(ConstantInt::get(ITy, Imm)); 1114 } 1115 return FastEmit_rr(VT, VT, Opcode, 1116 Op0, Op0IsKill, 1117 MaterialReg, /*Kill=*/true); 1118} 1119 1120unsigned FastISel::createResultReg(const TargetRegisterClass* RC) { 1121 return MRI.createVirtualRegister(RC); 1122} 1123 1124unsigned FastISel::FastEmitInst_(unsigned MachineInstOpcode, 1125 const TargetRegisterClass* RC) { 1126 unsigned ResultReg = createResultReg(RC); 1127 const MCInstrDesc &II = TII.get(MachineInstOpcode); 1128 1129 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg); 1130 return ResultReg; 1131} 1132 1133unsigned FastISel::FastEmitInst_r(unsigned MachineInstOpcode, 1134 const TargetRegisterClass *RC, 1135 unsigned Op0, bool Op0IsKill) { 1136 unsigned ResultReg = createResultReg(RC); 1137 const MCInstrDesc &II = TII.get(MachineInstOpcode); 1138 1139 if (II.getNumDefs() >= 1) 1140 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg) 1141 .addReg(Op0, Op0IsKill * RegState::Kill); 1142 else { 1143 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II) 1144 .addReg(Op0, Op0IsKill * RegState::Kill); 1145 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), 1146 ResultReg).addReg(II.ImplicitDefs[0]); 1147 } 1148 1149 return ResultReg; 1150} 1151 1152unsigned FastISel::FastEmitInst_rr(unsigned MachineInstOpcode, 1153 const TargetRegisterClass *RC, 1154 unsigned Op0, bool Op0IsKill, 1155 unsigned Op1, bool Op1IsKill) { 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 else { 1164 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II) 1165 .addReg(Op0, Op0IsKill * RegState::Kill) 1166 .addReg(Op1, Op1IsKill * RegState::Kill); 1167 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), 1168 ResultReg).addReg(II.ImplicitDefs[0]); 1169 } 1170 return ResultReg; 1171} 1172 1173unsigned FastISel::FastEmitInst_rrr(unsigned MachineInstOpcode, 1174 const TargetRegisterClass *RC, 1175 unsigned Op0, bool Op0IsKill, 1176 unsigned Op1, bool Op1IsKill, 1177 unsigned Op2, bool Op2IsKill) { 1178 unsigned ResultReg = createResultReg(RC); 1179 const MCInstrDesc &II = TII.get(MachineInstOpcode); 1180 1181 if (II.getNumDefs() >= 1) 1182 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg) 1183 .addReg(Op0, Op0IsKill * RegState::Kill) 1184 .addReg(Op1, Op1IsKill * RegState::Kill) 1185 .addReg(Op2, Op2IsKill * RegState::Kill); 1186 else { 1187 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II) 1188 .addReg(Op0, Op0IsKill * RegState::Kill) 1189 .addReg(Op1, Op1IsKill * RegState::Kill) 1190 .addReg(Op2, Op2IsKill * RegState::Kill); 1191 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), 1192 ResultReg).addReg(II.ImplicitDefs[0]); 1193 } 1194 return ResultReg; 1195} 1196 1197unsigned FastISel::FastEmitInst_ri(unsigned MachineInstOpcode, 1198 const TargetRegisterClass *RC, 1199 unsigned Op0, bool Op0IsKill, 1200 uint64_t Imm) { 1201 unsigned ResultReg = createResultReg(RC); 1202 const MCInstrDesc &II = TII.get(MachineInstOpcode); 1203 1204 if (II.getNumDefs() >= 1) 1205 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg) 1206 .addReg(Op0, Op0IsKill * RegState::Kill) 1207 .addImm(Imm); 1208 else { 1209 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II) 1210 .addReg(Op0, Op0IsKill * RegState::Kill) 1211 .addImm(Imm); 1212 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), 1213 ResultReg).addReg(II.ImplicitDefs[0]); 1214 } 1215 return ResultReg; 1216} 1217 1218unsigned FastISel::FastEmitInst_rii(unsigned MachineInstOpcode, 1219 const TargetRegisterClass *RC, 1220 unsigned Op0, bool Op0IsKill, 1221 uint64_t Imm1, uint64_t Imm2) { 1222 unsigned ResultReg = createResultReg(RC); 1223 const MCInstrDesc &II = TII.get(MachineInstOpcode); 1224 1225 if (II.getNumDefs() >= 1) 1226 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg) 1227 .addReg(Op0, Op0IsKill * RegState::Kill) 1228 .addImm(Imm1) 1229 .addImm(Imm2); 1230 else { 1231 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II) 1232 .addReg(Op0, Op0IsKill * RegState::Kill) 1233 .addImm(Imm1) 1234 .addImm(Imm2); 1235 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), 1236 ResultReg).addReg(II.ImplicitDefs[0]); 1237 } 1238 return ResultReg; 1239} 1240 1241unsigned FastISel::FastEmitInst_rf(unsigned MachineInstOpcode, 1242 const TargetRegisterClass *RC, 1243 unsigned Op0, bool Op0IsKill, 1244 const ConstantFP *FPImm) { 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 .addFPImm(FPImm); 1252 else { 1253 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II) 1254 .addReg(Op0, Op0IsKill * RegState::Kill) 1255 .addFPImm(FPImm); 1256 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), 1257 ResultReg).addReg(II.ImplicitDefs[0]); 1258 } 1259 return ResultReg; 1260} 1261 1262unsigned FastISel::FastEmitInst_rri(unsigned MachineInstOpcode, 1263 const TargetRegisterClass *RC, 1264 unsigned Op0, bool Op0IsKill, 1265 unsigned Op1, bool Op1IsKill, 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) 1272 .addReg(Op0, Op0IsKill * RegState::Kill) 1273 .addReg(Op1, Op1IsKill * RegState::Kill) 1274 .addImm(Imm); 1275 else { 1276 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II) 1277 .addReg(Op0, Op0IsKill * RegState::Kill) 1278 .addReg(Op1, Op1IsKill * RegState::Kill) 1279 .addImm(Imm); 1280 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), 1281 ResultReg).addReg(II.ImplicitDefs[0]); 1282 } 1283 return ResultReg; 1284} 1285 1286unsigned FastISel::FastEmitInst_i(unsigned MachineInstOpcode, 1287 const TargetRegisterClass *RC, 1288 uint64_t Imm) { 1289 unsigned ResultReg = createResultReg(RC); 1290 const MCInstrDesc &II = TII.get(MachineInstOpcode); 1291 1292 if (II.getNumDefs() >= 1) 1293 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg).addImm(Imm); 1294 else { 1295 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II).addImm(Imm); 1296 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), 1297 ResultReg).addReg(II.ImplicitDefs[0]); 1298 } 1299 return ResultReg; 1300} 1301 1302unsigned FastISel::FastEmitInst_ii(unsigned MachineInstOpcode, 1303 const TargetRegisterClass *RC, 1304 uint64_t Imm1, uint64_t Imm2) { 1305 unsigned ResultReg = createResultReg(RC); 1306 const MCInstrDesc &II = TII.get(MachineInstOpcode); 1307 1308 if (II.getNumDefs() >= 1) 1309 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg) 1310 .addImm(Imm1).addImm(Imm2); 1311 else { 1312 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II).addImm(Imm1).addImm(Imm2); 1313 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY), 1314 ResultReg).addReg(II.ImplicitDefs[0]); 1315 } 1316 return ResultReg; 1317} 1318 1319unsigned FastISel::FastEmitInst_extractsubreg(MVT RetVT, 1320 unsigned Op0, bool Op0IsKill, 1321 uint32_t Idx) { 1322 unsigned ResultReg = createResultReg(TLI.getRegClassFor(RetVT)); 1323 assert(TargetRegisterInfo::isVirtualRegister(Op0) && 1324 "Cannot yet extract from physregs"); 1325 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, 1326 DL, TII.get(TargetOpcode::COPY), ResultReg) 1327 .addReg(Op0, getKillRegState(Op0IsKill), Idx); 1328 return ResultReg; 1329} 1330 1331/// FastEmitZExtFromI1 - Emit MachineInstrs to compute the value of Op 1332/// with all but the least significant bit set to zero. 1333unsigned FastISel::FastEmitZExtFromI1(MVT VT, unsigned Op0, bool Op0IsKill) { 1334 return FastEmit_ri(VT, VT, ISD::AND, Op0, Op0IsKill, 1); 1335} 1336 1337/// HandlePHINodesInSuccessorBlocks - Handle PHI nodes in successor blocks. 1338/// Emit code to ensure constants are copied into registers when needed. 1339/// Remember the virtual registers that need to be added to the Machine PHI 1340/// nodes as input. We cannot just directly add them, because expansion 1341/// might result in multiple MBB's for one BB. As such, the start of the 1342/// BB might correspond to a different MBB than the end. 1343bool FastISel::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) { 1344 const TerminatorInst *TI = LLVMBB->getTerminator(); 1345 1346 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled; 1347 unsigned OrigNumPHINodesToUpdate = FuncInfo.PHINodesToUpdate.size(); 1348 1349 // Check successor nodes' PHI nodes that expect a constant to be available 1350 // from this block. 1351 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) { 1352 const BasicBlock *SuccBB = TI->getSuccessor(succ); 1353 if (!isa<PHINode>(SuccBB->begin())) continue; 1354 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB]; 1355 1356 // If this terminator has multiple identical successors (common for 1357 // switches), only handle each succ once. 1358 if (!SuccsHandled.insert(SuccMBB)) continue; 1359 1360 MachineBasicBlock::iterator MBBI = SuccMBB->begin(); 1361 1362 // At this point we know that there is a 1-1 correspondence between LLVM PHI 1363 // nodes and Machine PHI nodes, but the incoming operands have not been 1364 // emitted yet. 1365 for (BasicBlock::const_iterator I = SuccBB->begin(); 1366 const PHINode *PN = dyn_cast<PHINode>(I); ++I) { 1367 1368 // Ignore dead phi's. 1369 if (PN->use_empty()) continue; 1370 1371 // Only handle legal types. Two interesting things to note here. First, 1372 // by bailing out early, we may leave behind some dead instructions, 1373 // since SelectionDAG's HandlePHINodesInSuccessorBlocks will insert its 1374 // own moves. Second, this check is necessary because FastISel doesn't 1375 // use CreateRegs to create registers, so it always creates 1376 // exactly one register for each non-void instruction. 1377 EVT VT = TLI.getValueType(PN->getType(), /*AllowUnknown=*/true); 1378 if (VT == MVT::Other || !TLI.isTypeLegal(VT)) { 1379 // Handle integer promotions, though, because they're common and easy. 1380 if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16) 1381 VT = TLI.getTypeToTransformTo(LLVMBB->getContext(), VT); 1382 else { 1383 FuncInfo.PHINodesToUpdate.resize(OrigNumPHINodesToUpdate); 1384 return false; 1385 } 1386 } 1387 1388 const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB); 1389 1390 // Set the DebugLoc for the copy. Prefer the location of the operand 1391 // if there is one; use the location of the PHI otherwise. 1392 DL = PN->getDebugLoc(); 1393 if (const Instruction *Inst = dyn_cast<Instruction>(PHIOp)) 1394 DL = Inst->getDebugLoc(); 1395 1396 unsigned Reg = getRegForValue(PHIOp); 1397 if (Reg == 0) { 1398 FuncInfo.PHINodesToUpdate.resize(OrigNumPHINodesToUpdate); 1399 return false; 1400 } 1401 FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg)); 1402 DL = DebugLoc(); 1403 } 1404 } 1405 1406 return true; 1407} 1408