X86FastISel.cpp revision dce4a407a24b04eebc6a376f8e62b41aaa7b071f
1//===-- X86FastISel.cpp - X86 FastISel implementation ---------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file defines the X86-specific support for the FastISel class. Much 11// of the target-specific code is generated by tablegen in the file 12// X86GenFastISel.inc, which is #included here. 13// 14//===----------------------------------------------------------------------===// 15 16#include "X86.h" 17#include "X86CallingConv.h" 18#include "X86InstrBuilder.h" 19#include "X86MachineFunctionInfo.h" 20#include "X86RegisterInfo.h" 21#include "X86Subtarget.h" 22#include "X86TargetMachine.h" 23#include "llvm/CodeGen/Analysis.h" 24#include "llvm/CodeGen/FastISel.h" 25#include "llvm/CodeGen/FunctionLoweringInfo.h" 26#include "llvm/CodeGen/MachineConstantPool.h" 27#include "llvm/CodeGen/MachineFrameInfo.h" 28#include "llvm/CodeGen/MachineRegisterInfo.h" 29#include "llvm/IR/CallSite.h" 30#include "llvm/IR/CallingConv.h" 31#include "llvm/IR/DerivedTypes.h" 32#include "llvm/IR/GetElementPtrTypeIterator.h" 33#include "llvm/IR/GlobalAlias.h" 34#include "llvm/IR/GlobalVariable.h" 35#include "llvm/IR/Instructions.h" 36#include "llvm/IR/IntrinsicInst.h" 37#include "llvm/IR/Operator.h" 38#include "llvm/Support/ErrorHandling.h" 39#include "llvm/Target/TargetOptions.h" 40using namespace llvm; 41 42namespace { 43 44class X86FastISel final : public FastISel { 45 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can 46 /// make the right decision when generating code for different targets. 47 const X86Subtarget *Subtarget; 48 49 /// X86ScalarSSEf32, X86ScalarSSEf64 - Select between SSE or x87 50 /// floating point ops. 51 /// When SSE is available, use it for f32 operations. 52 /// When SSE2 is available, use it for f64 operations. 53 bool X86ScalarSSEf64; 54 bool X86ScalarSSEf32; 55 56public: 57 explicit X86FastISel(FunctionLoweringInfo &funcInfo, 58 const TargetLibraryInfo *libInfo) 59 : FastISel(funcInfo, libInfo) { 60 Subtarget = &TM.getSubtarget<X86Subtarget>(); 61 X86ScalarSSEf64 = Subtarget->hasSSE2(); 62 X86ScalarSSEf32 = Subtarget->hasSSE1(); 63 } 64 65 bool TargetSelectInstruction(const Instruction *I) override; 66 67 /// \brief The specified machine instr operand is a vreg, and that 68 /// vreg is being provided by the specified load instruction. If possible, 69 /// try to fold the load as an operand to the instruction, returning true if 70 /// possible. 71 bool tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo, 72 const LoadInst *LI) override; 73 74 bool FastLowerArguments() override; 75 76#include "X86GenFastISel.inc" 77 78private: 79 bool X86FastEmitCompare(const Value *LHS, const Value *RHS, EVT VT); 80 81 bool X86FastEmitLoad(EVT VT, const X86AddressMode &AM, unsigned &RR); 82 83 bool X86FastEmitStore(EVT VT, const Value *Val, const X86AddressMode &AM, 84 bool Aligned = false); 85 bool X86FastEmitStore(EVT VT, unsigned ValReg, const X86AddressMode &AM, 86 bool Aligned = false); 87 88 bool X86FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src, EVT SrcVT, 89 unsigned &ResultReg); 90 91 bool X86SelectAddress(const Value *V, X86AddressMode &AM); 92 bool X86SelectCallAddress(const Value *V, X86AddressMode &AM); 93 94 bool X86SelectLoad(const Instruction *I); 95 96 bool X86SelectStore(const Instruction *I); 97 98 bool X86SelectRet(const Instruction *I); 99 100 bool X86SelectCmp(const Instruction *I); 101 102 bool X86SelectZExt(const Instruction *I); 103 104 bool X86SelectBranch(const Instruction *I); 105 106 bool X86SelectShift(const Instruction *I); 107 108 bool X86SelectDivRem(const Instruction *I); 109 110 bool X86SelectSelect(const Instruction *I); 111 112 bool X86SelectTrunc(const Instruction *I); 113 114 bool X86SelectFPExt(const Instruction *I); 115 bool X86SelectFPTrunc(const Instruction *I); 116 117 bool X86VisitIntrinsicCall(const IntrinsicInst &I); 118 bool X86SelectCall(const Instruction *I); 119 120 bool DoSelectCall(const Instruction *I, const char *MemIntName); 121 122 const X86InstrInfo *getInstrInfo() const { 123 return getTargetMachine()->getInstrInfo(); 124 } 125 const X86TargetMachine *getTargetMachine() const { 126 return static_cast<const X86TargetMachine *>(&TM); 127 } 128 129 bool handleConstantAddresses(const Value *V, X86AddressMode &AM); 130 131 unsigned TargetMaterializeConstant(const Constant *C) override; 132 133 unsigned TargetMaterializeAlloca(const AllocaInst *C) override; 134 135 unsigned TargetMaterializeFloatZero(const ConstantFP *CF) override; 136 137 /// isScalarFPTypeInSSEReg - Return true if the specified scalar FP type is 138 /// computed in an SSE register, not on the X87 floating point stack. 139 bool isScalarFPTypeInSSEReg(EVT VT) const { 140 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2 141 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1 142 } 143 144 bool isTypeLegal(Type *Ty, MVT &VT, bool AllowI1 = false); 145 146 bool IsMemcpySmall(uint64_t Len); 147 148 bool TryEmitSmallMemcpy(X86AddressMode DestAM, 149 X86AddressMode SrcAM, uint64_t Len); 150}; 151 152} // end anonymous namespace. 153 154bool X86FastISel::isTypeLegal(Type *Ty, MVT &VT, bool AllowI1) { 155 EVT evt = TLI.getValueType(Ty, /*HandleUnknown=*/true); 156 if (evt == MVT::Other || !evt.isSimple()) 157 // Unhandled type. Halt "fast" selection and bail. 158 return false; 159 160 VT = evt.getSimpleVT(); 161 // For now, require SSE/SSE2 for performing floating-point operations, 162 // since x87 requires additional work. 163 if (VT == MVT::f64 && !X86ScalarSSEf64) 164 return false; 165 if (VT == MVT::f32 && !X86ScalarSSEf32) 166 return false; 167 // Similarly, no f80 support yet. 168 if (VT == MVT::f80) 169 return false; 170 // We only handle legal types. For example, on x86-32 the instruction 171 // selector contains all of the 64-bit instructions from x86-64, 172 // under the assumption that i64 won't be used if the target doesn't 173 // support it. 174 return (AllowI1 && VT == MVT::i1) || TLI.isTypeLegal(VT); 175} 176 177#include "X86GenCallingConv.inc" 178 179/// X86FastEmitLoad - Emit a machine instruction to load a value of type VT. 180/// The address is either pre-computed, i.e. Ptr, or a GlobalAddress, i.e. GV. 181/// Return true and the result register by reference if it is possible. 182bool X86FastISel::X86FastEmitLoad(EVT VT, const X86AddressMode &AM, 183 unsigned &ResultReg) { 184 // Get opcode and regclass of the output for the given load instruction. 185 unsigned Opc = 0; 186 const TargetRegisterClass *RC = nullptr; 187 switch (VT.getSimpleVT().SimpleTy) { 188 default: return false; 189 case MVT::i1: 190 case MVT::i8: 191 Opc = X86::MOV8rm; 192 RC = &X86::GR8RegClass; 193 break; 194 case MVT::i16: 195 Opc = X86::MOV16rm; 196 RC = &X86::GR16RegClass; 197 break; 198 case MVT::i32: 199 Opc = X86::MOV32rm; 200 RC = &X86::GR32RegClass; 201 break; 202 case MVT::i64: 203 // Must be in x86-64 mode. 204 Opc = X86::MOV64rm; 205 RC = &X86::GR64RegClass; 206 break; 207 case MVT::f32: 208 if (X86ScalarSSEf32) { 209 Opc = Subtarget->hasAVX() ? X86::VMOVSSrm : X86::MOVSSrm; 210 RC = &X86::FR32RegClass; 211 } else { 212 Opc = X86::LD_Fp32m; 213 RC = &X86::RFP32RegClass; 214 } 215 break; 216 case MVT::f64: 217 if (X86ScalarSSEf64) { 218 Opc = Subtarget->hasAVX() ? X86::VMOVSDrm : X86::MOVSDrm; 219 RC = &X86::FR64RegClass; 220 } else { 221 Opc = X86::LD_Fp64m; 222 RC = &X86::RFP64RegClass; 223 } 224 break; 225 case MVT::f80: 226 // No f80 support yet. 227 return false; 228 } 229 230 ResultReg = createResultReg(RC); 231 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, 232 DbgLoc, TII.get(Opc), ResultReg), AM); 233 return true; 234} 235 236/// X86FastEmitStore - Emit a machine instruction to store a value Val of 237/// type VT. The address is either pre-computed, consisted of a base ptr, Ptr 238/// and a displacement offset, or a GlobalAddress, 239/// i.e. V. Return true if it is possible. 240bool 241X86FastISel::X86FastEmitStore(EVT VT, unsigned ValReg, 242 const X86AddressMode &AM, bool Aligned) { 243 // Get opcode and regclass of the output for the given store instruction. 244 unsigned Opc = 0; 245 switch (VT.getSimpleVT().SimpleTy) { 246 case MVT::f80: // No f80 support yet. 247 default: return false; 248 case MVT::i1: { 249 // Mask out all but lowest bit. 250 unsigned AndResult = createResultReg(&X86::GR8RegClass); 251 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, 252 TII.get(X86::AND8ri), AndResult).addReg(ValReg).addImm(1); 253 ValReg = AndResult; 254 } 255 // FALLTHROUGH, handling i1 as i8. 256 case MVT::i8: Opc = X86::MOV8mr; break; 257 case MVT::i16: Opc = X86::MOV16mr; break; 258 case MVT::i32: Opc = X86::MOV32mr; break; 259 case MVT::i64: Opc = X86::MOV64mr; break; // Must be in x86-64 mode. 260 case MVT::f32: 261 Opc = X86ScalarSSEf32 ? 262 (Subtarget->hasAVX() ? X86::VMOVSSmr : X86::MOVSSmr) : X86::ST_Fp32m; 263 break; 264 case MVT::f64: 265 Opc = X86ScalarSSEf64 ? 266 (Subtarget->hasAVX() ? X86::VMOVSDmr : X86::MOVSDmr) : X86::ST_Fp64m; 267 break; 268 case MVT::v4f32: 269 if (Aligned) 270 Opc = Subtarget->hasAVX() ? X86::VMOVAPSmr : X86::MOVAPSmr; 271 else 272 Opc = Subtarget->hasAVX() ? X86::VMOVUPSmr : X86::MOVUPSmr; 273 break; 274 case MVT::v2f64: 275 if (Aligned) 276 Opc = Subtarget->hasAVX() ? X86::VMOVAPDmr : X86::MOVAPDmr; 277 else 278 Opc = Subtarget->hasAVX() ? X86::VMOVUPDmr : X86::MOVUPDmr; 279 break; 280 case MVT::v4i32: 281 case MVT::v2i64: 282 case MVT::v8i16: 283 case MVT::v16i8: 284 if (Aligned) 285 Opc = Subtarget->hasAVX() ? X86::VMOVDQAmr : X86::MOVDQAmr; 286 else 287 Opc = Subtarget->hasAVX() ? X86::VMOVDQUmr : X86::MOVDQUmr; 288 break; 289 } 290 291 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, 292 DbgLoc, TII.get(Opc)), AM).addReg(ValReg); 293 return true; 294} 295 296bool X86FastISel::X86FastEmitStore(EVT VT, const Value *Val, 297 const X86AddressMode &AM, bool Aligned) { 298 // Handle 'null' like i32/i64 0. 299 if (isa<ConstantPointerNull>(Val)) 300 Val = Constant::getNullValue(DL.getIntPtrType(Val->getContext())); 301 302 // If this is a store of a simple constant, fold the constant into the store. 303 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Val)) { 304 unsigned Opc = 0; 305 bool Signed = true; 306 switch (VT.getSimpleVT().SimpleTy) { 307 default: break; 308 case MVT::i1: Signed = false; // FALLTHROUGH to handle as i8. 309 case MVT::i8: Opc = X86::MOV8mi; break; 310 case MVT::i16: Opc = X86::MOV16mi; break; 311 case MVT::i32: Opc = X86::MOV32mi; break; 312 case MVT::i64: 313 // Must be a 32-bit sign extended value. 314 if (isInt<32>(CI->getSExtValue())) 315 Opc = X86::MOV64mi32; 316 break; 317 } 318 319 if (Opc) { 320 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, 321 DbgLoc, TII.get(Opc)), AM) 322 .addImm(Signed ? (uint64_t) CI->getSExtValue() : 323 CI->getZExtValue()); 324 return true; 325 } 326 } 327 328 unsigned ValReg = getRegForValue(Val); 329 if (ValReg == 0) 330 return false; 331 332 return X86FastEmitStore(VT, ValReg, AM, Aligned); 333} 334 335/// X86FastEmitExtend - Emit a machine instruction to extend a value Src of 336/// type SrcVT to type DstVT using the specified extension opcode Opc (e.g. 337/// ISD::SIGN_EXTEND). 338bool X86FastISel::X86FastEmitExtend(ISD::NodeType Opc, EVT DstVT, 339 unsigned Src, EVT SrcVT, 340 unsigned &ResultReg) { 341 unsigned RR = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), Opc, 342 Src, /*TODO: Kill=*/false); 343 if (RR == 0) 344 return false; 345 346 ResultReg = RR; 347 return true; 348} 349 350bool X86FastISel::handleConstantAddresses(const Value *V, X86AddressMode &AM) { 351 // Handle constant address. 352 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) { 353 // Can't handle alternate code models yet. 354 if (TM.getCodeModel() != CodeModel::Small) 355 return false; 356 357 // Can't handle TLS yet. 358 if (const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV)) 359 if (GVar->isThreadLocal()) 360 return false; 361 362 // Can't handle TLS yet, part 2 (this is slightly crazy, but this is how 363 // it works...). 364 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV)) 365 if (const GlobalVariable *GVar = 366 dyn_cast_or_null<GlobalVariable>(GA->getAliasee())) 367 if (GVar->isThreadLocal()) 368 return false; 369 370 // RIP-relative addresses can't have additional register operands, so if 371 // we've already folded stuff into the addressing mode, just force the 372 // global value into its own register, which we can use as the basereg. 373 if (!Subtarget->isPICStyleRIPRel() || 374 (AM.Base.Reg == 0 && AM.IndexReg == 0)) { 375 // Okay, we've committed to selecting this global. Set up the address. 376 AM.GV = GV; 377 378 // Allow the subtarget to classify the global. 379 unsigned char GVFlags = Subtarget->ClassifyGlobalReference(GV, TM); 380 381 // If this reference is relative to the pic base, set it now. 382 if (isGlobalRelativeToPICBase(GVFlags)) { 383 // FIXME: How do we know Base.Reg is free?? 384 AM.Base.Reg = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF); 385 } 386 387 // Unless the ABI requires an extra load, return a direct reference to 388 // the global. 389 if (!isGlobalStubReference(GVFlags)) { 390 if (Subtarget->isPICStyleRIPRel()) { 391 // Use rip-relative addressing if we can. Above we verified that the 392 // base and index registers are unused. 393 assert(AM.Base.Reg == 0 && AM.IndexReg == 0); 394 AM.Base.Reg = X86::RIP; 395 } 396 AM.GVOpFlags = GVFlags; 397 return true; 398 } 399 400 // Ok, we need to do a load from a stub. If we've already loaded from 401 // this stub, reuse the loaded pointer, otherwise emit the load now. 402 DenseMap<const Value*, unsigned>::iterator I = LocalValueMap.find(V); 403 unsigned LoadReg; 404 if (I != LocalValueMap.end() && I->second != 0) { 405 LoadReg = I->second; 406 } else { 407 // Issue load from stub. 408 unsigned Opc = 0; 409 const TargetRegisterClass *RC = nullptr; 410 X86AddressMode StubAM; 411 StubAM.Base.Reg = AM.Base.Reg; 412 StubAM.GV = GV; 413 StubAM.GVOpFlags = GVFlags; 414 415 // Prepare for inserting code in the local-value area. 416 SavePoint SaveInsertPt = enterLocalValueArea(); 417 418 if (TLI.getPointerTy() == MVT::i64) { 419 Opc = X86::MOV64rm; 420 RC = &X86::GR64RegClass; 421 422 if (Subtarget->isPICStyleRIPRel()) 423 StubAM.Base.Reg = X86::RIP; 424 } else { 425 Opc = X86::MOV32rm; 426 RC = &X86::GR32RegClass; 427 } 428 429 LoadReg = createResultReg(RC); 430 MachineInstrBuilder LoadMI = 431 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), LoadReg); 432 addFullAddress(LoadMI, StubAM); 433 434 // Ok, back to normal mode. 435 leaveLocalValueArea(SaveInsertPt); 436 437 // Prevent loading GV stub multiple times in same MBB. 438 LocalValueMap[V] = LoadReg; 439 } 440 441 // Now construct the final address. Note that the Disp, Scale, 442 // and Index values may already be set here. 443 AM.Base.Reg = LoadReg; 444 AM.GV = nullptr; 445 return true; 446 } 447 } 448 449 // If all else fails, try to materialize the value in a register. 450 if (!AM.GV || !Subtarget->isPICStyleRIPRel()) { 451 if (AM.Base.Reg == 0) { 452 AM.Base.Reg = getRegForValue(V); 453 return AM.Base.Reg != 0; 454 } 455 if (AM.IndexReg == 0) { 456 assert(AM.Scale == 1 && "Scale with no index!"); 457 AM.IndexReg = getRegForValue(V); 458 return AM.IndexReg != 0; 459 } 460 } 461 462 return false; 463} 464 465/// X86SelectAddress - Attempt to fill in an address from the given value. 466/// 467bool X86FastISel::X86SelectAddress(const Value *V, X86AddressMode &AM) { 468 SmallVector<const Value *, 32> GEPs; 469redo_gep: 470 const User *U = nullptr; 471 unsigned Opcode = Instruction::UserOp1; 472 if (const Instruction *I = dyn_cast<Instruction>(V)) { 473 // Don't walk into other basic blocks; it's possible we haven't 474 // visited them yet, so the instructions may not yet be assigned 475 // virtual registers. 476 if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(V)) || 477 FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) { 478 Opcode = I->getOpcode(); 479 U = I; 480 } 481 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(V)) { 482 Opcode = C->getOpcode(); 483 U = C; 484 } 485 486 if (PointerType *Ty = dyn_cast<PointerType>(V->getType())) 487 if (Ty->getAddressSpace() > 255) 488 // Fast instruction selection doesn't support the special 489 // address spaces. 490 return false; 491 492 switch (Opcode) { 493 default: break; 494 case Instruction::BitCast: 495 // Look past bitcasts. 496 return X86SelectAddress(U->getOperand(0), AM); 497 498 case Instruction::IntToPtr: 499 // Look past no-op inttoptrs. 500 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy()) 501 return X86SelectAddress(U->getOperand(0), AM); 502 break; 503 504 case Instruction::PtrToInt: 505 // Look past no-op ptrtoints. 506 if (TLI.getValueType(U->getType()) == TLI.getPointerTy()) 507 return X86SelectAddress(U->getOperand(0), AM); 508 break; 509 510 case Instruction::Alloca: { 511 // Do static allocas. 512 const AllocaInst *A = cast<AllocaInst>(V); 513 DenseMap<const AllocaInst*, int>::iterator SI = 514 FuncInfo.StaticAllocaMap.find(A); 515 if (SI != FuncInfo.StaticAllocaMap.end()) { 516 AM.BaseType = X86AddressMode::FrameIndexBase; 517 AM.Base.FrameIndex = SI->second; 518 return true; 519 } 520 break; 521 } 522 523 case Instruction::Add: { 524 // Adds of constants are common and easy enough. 525 if (const ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) { 526 uint64_t Disp = (int32_t)AM.Disp + (uint64_t)CI->getSExtValue(); 527 // They have to fit in the 32-bit signed displacement field though. 528 if (isInt<32>(Disp)) { 529 AM.Disp = (uint32_t)Disp; 530 return X86SelectAddress(U->getOperand(0), AM); 531 } 532 } 533 break; 534 } 535 536 case Instruction::GetElementPtr: { 537 X86AddressMode SavedAM = AM; 538 539 // Pattern-match simple GEPs. 540 uint64_t Disp = (int32_t)AM.Disp; 541 unsigned IndexReg = AM.IndexReg; 542 unsigned Scale = AM.Scale; 543 gep_type_iterator GTI = gep_type_begin(U); 544 // Iterate through the indices, folding what we can. Constants can be 545 // folded, and one dynamic index can be handled, if the scale is supported. 546 for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end(); 547 i != e; ++i, ++GTI) { 548 const Value *Op = *i; 549 if (StructType *STy = dyn_cast<StructType>(*GTI)) { 550 const StructLayout *SL = DL.getStructLayout(STy); 551 Disp += SL->getElementOffset(cast<ConstantInt>(Op)->getZExtValue()); 552 continue; 553 } 554 555 // A array/variable index is always of the form i*S where S is the 556 // constant scale size. See if we can push the scale into immediates. 557 uint64_t S = DL.getTypeAllocSize(GTI.getIndexedType()); 558 for (;;) { 559 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) { 560 // Constant-offset addressing. 561 Disp += CI->getSExtValue() * S; 562 break; 563 } 564 if (canFoldAddIntoGEP(U, Op)) { 565 // A compatible add with a constant operand. Fold the constant. 566 ConstantInt *CI = 567 cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1)); 568 Disp += CI->getSExtValue() * S; 569 // Iterate on the other operand. 570 Op = cast<AddOperator>(Op)->getOperand(0); 571 continue; 572 } 573 if (IndexReg == 0 && 574 (!AM.GV || !Subtarget->isPICStyleRIPRel()) && 575 (S == 1 || S == 2 || S == 4 || S == 8)) { 576 // Scaled-index addressing. 577 Scale = S; 578 IndexReg = getRegForGEPIndex(Op).first; 579 if (IndexReg == 0) 580 return false; 581 break; 582 } 583 // Unsupported. 584 goto unsupported_gep; 585 } 586 } 587 588 // Check for displacement overflow. 589 if (!isInt<32>(Disp)) 590 break; 591 592 AM.IndexReg = IndexReg; 593 AM.Scale = Scale; 594 AM.Disp = (uint32_t)Disp; 595 GEPs.push_back(V); 596 597 if (const GetElementPtrInst *GEP = 598 dyn_cast<GetElementPtrInst>(U->getOperand(0))) { 599 // Ok, the GEP indices were covered by constant-offset and scaled-index 600 // addressing. Update the address state and move on to examining the base. 601 V = GEP; 602 goto redo_gep; 603 } else if (X86SelectAddress(U->getOperand(0), AM)) { 604 return true; 605 } 606 607 // If we couldn't merge the gep value into this addr mode, revert back to 608 // our address and just match the value instead of completely failing. 609 AM = SavedAM; 610 611 for (SmallVectorImpl<const Value *>::reverse_iterator 612 I = GEPs.rbegin(), E = GEPs.rend(); I != E; ++I) 613 if (handleConstantAddresses(*I, AM)) 614 return true; 615 616 return false; 617 unsupported_gep: 618 // Ok, the GEP indices weren't all covered. 619 break; 620 } 621 } 622 623 return handleConstantAddresses(V, AM); 624} 625 626/// X86SelectCallAddress - Attempt to fill in an address from the given value. 627/// 628bool X86FastISel::X86SelectCallAddress(const Value *V, X86AddressMode &AM) { 629 const User *U = nullptr; 630 unsigned Opcode = Instruction::UserOp1; 631 const Instruction *I = dyn_cast<Instruction>(V); 632 // Record if the value is defined in the same basic block. 633 // 634 // This information is crucial to know whether or not folding an 635 // operand is valid. 636 // Indeed, FastISel generates or reuses a virtual register for all 637 // operands of all instructions it selects. Obviously, the definition and 638 // its uses must use the same virtual register otherwise the produced 639 // code is incorrect. 640 // Before instruction selection, FunctionLoweringInfo::set sets the virtual 641 // registers for values that are alive across basic blocks. This ensures 642 // that the values are consistently set between across basic block, even 643 // if different instruction selection mechanisms are used (e.g., a mix of 644 // SDISel and FastISel). 645 // For values local to a basic block, the instruction selection process 646 // generates these virtual registers with whatever method is appropriate 647 // for its needs. In particular, FastISel and SDISel do not share the way 648 // local virtual registers are set. 649 // Therefore, this is impossible (or at least unsafe) to share values 650 // between basic blocks unless they use the same instruction selection 651 // method, which is not guarantee for X86. 652 // Moreover, things like hasOneUse could not be used accurately, if we 653 // allow to reference values across basic blocks whereas they are not 654 // alive across basic blocks initially. 655 bool InMBB = true; 656 if (I) { 657 Opcode = I->getOpcode(); 658 U = I; 659 InMBB = I->getParent() == FuncInfo.MBB->getBasicBlock(); 660 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(V)) { 661 Opcode = C->getOpcode(); 662 U = C; 663 } 664 665 switch (Opcode) { 666 default: break; 667 case Instruction::BitCast: 668 // Look past bitcasts if its operand is in the same BB. 669 if (InMBB) 670 return X86SelectCallAddress(U->getOperand(0), AM); 671 break; 672 673 case Instruction::IntToPtr: 674 // Look past no-op inttoptrs if its operand is in the same BB. 675 if (InMBB && 676 TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy()) 677 return X86SelectCallAddress(U->getOperand(0), AM); 678 break; 679 680 case Instruction::PtrToInt: 681 // Look past no-op ptrtoints if its operand is in the same BB. 682 if (InMBB && 683 TLI.getValueType(U->getType()) == TLI.getPointerTy()) 684 return X86SelectCallAddress(U->getOperand(0), AM); 685 break; 686 } 687 688 // Handle constant address. 689 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) { 690 // Can't handle alternate code models yet. 691 if (TM.getCodeModel() != CodeModel::Small) 692 return false; 693 694 // RIP-relative addresses can't have additional register operands. 695 if (Subtarget->isPICStyleRIPRel() && 696 (AM.Base.Reg != 0 || AM.IndexReg != 0)) 697 return false; 698 699 // Can't handle DbgLocLImport. 700 if (GV->hasDLLImportStorageClass()) 701 return false; 702 703 // Can't handle TLS. 704 if (const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV)) 705 if (GVar->isThreadLocal()) 706 return false; 707 708 // Okay, we've committed to selecting this global. Set up the basic address. 709 AM.GV = GV; 710 711 // No ABI requires an extra load for anything other than DLLImport, which 712 // we rejected above. Return a direct reference to the global. 713 if (Subtarget->isPICStyleRIPRel()) { 714 // Use rip-relative addressing if we can. Above we verified that the 715 // base and index registers are unused. 716 assert(AM.Base.Reg == 0 && AM.IndexReg == 0); 717 AM.Base.Reg = X86::RIP; 718 } else if (Subtarget->isPICStyleStubPIC()) { 719 AM.GVOpFlags = X86II::MO_PIC_BASE_OFFSET; 720 } else if (Subtarget->isPICStyleGOT()) { 721 AM.GVOpFlags = X86II::MO_GOTOFF; 722 } 723 724 return true; 725 } 726 727 // If all else fails, try to materialize the value in a register. 728 if (!AM.GV || !Subtarget->isPICStyleRIPRel()) { 729 if (AM.Base.Reg == 0) { 730 AM.Base.Reg = getRegForValue(V); 731 return AM.Base.Reg != 0; 732 } 733 if (AM.IndexReg == 0) { 734 assert(AM.Scale == 1 && "Scale with no index!"); 735 AM.IndexReg = getRegForValue(V); 736 return AM.IndexReg != 0; 737 } 738 } 739 740 return false; 741} 742 743 744/// X86SelectStore - Select and emit code to implement store instructions. 745bool X86FastISel::X86SelectStore(const Instruction *I) { 746 // Atomic stores need special handling. 747 const StoreInst *S = cast<StoreInst>(I); 748 749 if (S->isAtomic()) 750 return false; 751 752 unsigned SABIAlignment = 753 DL.getABITypeAlignment(S->getValueOperand()->getType()); 754 bool Aligned = S->getAlignment() == 0 || S->getAlignment() >= SABIAlignment; 755 756 MVT VT; 757 if (!isTypeLegal(I->getOperand(0)->getType(), VT, /*AllowI1=*/true)) 758 return false; 759 760 X86AddressMode AM; 761 if (!X86SelectAddress(I->getOperand(1), AM)) 762 return false; 763 764 return X86FastEmitStore(VT, I->getOperand(0), AM, Aligned); 765} 766 767/// X86SelectRet - Select and emit code to implement ret instructions. 768bool X86FastISel::X86SelectRet(const Instruction *I) { 769 const ReturnInst *Ret = cast<ReturnInst>(I); 770 const Function &F = *I->getParent()->getParent(); 771 const X86MachineFunctionInfo *X86MFInfo = 772 FuncInfo.MF->getInfo<X86MachineFunctionInfo>(); 773 774 if (!FuncInfo.CanLowerReturn) 775 return false; 776 777 CallingConv::ID CC = F.getCallingConv(); 778 if (CC != CallingConv::C && 779 CC != CallingConv::Fast && 780 CC != CallingConv::X86_FastCall && 781 CC != CallingConv::X86_64_SysV) 782 return false; 783 784 if (Subtarget->isCallingConvWin64(CC)) 785 return false; 786 787 // Don't handle popping bytes on return for now. 788 if (X86MFInfo->getBytesToPopOnReturn() != 0) 789 return false; 790 791 // fastcc with -tailcallopt is intended to provide a guaranteed 792 // tail call optimization. Fastisel doesn't know how to do that. 793 if (CC == CallingConv::Fast && TM.Options.GuaranteedTailCallOpt) 794 return false; 795 796 // Let SDISel handle vararg functions. 797 if (F.isVarArg()) 798 return false; 799 800 // Build a list of return value registers. 801 SmallVector<unsigned, 4> RetRegs; 802 803 if (Ret->getNumOperands() > 0) { 804 SmallVector<ISD::OutputArg, 4> Outs; 805 GetReturnInfo(F.getReturnType(), F.getAttributes(), Outs, TLI); 806 807 // Analyze operands of the call, assigning locations to each operand. 808 SmallVector<CCValAssign, 16> ValLocs; 809 CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, TM, ValLocs, 810 I->getContext()); 811 CCInfo.AnalyzeReturn(Outs, RetCC_X86); 812 813 const Value *RV = Ret->getOperand(0); 814 unsigned Reg = getRegForValue(RV); 815 if (Reg == 0) 816 return false; 817 818 // Only handle a single return value for now. 819 if (ValLocs.size() != 1) 820 return false; 821 822 CCValAssign &VA = ValLocs[0]; 823 824 // Don't bother handling odd stuff for now. 825 if (VA.getLocInfo() != CCValAssign::Full) 826 return false; 827 // Only handle register returns for now. 828 if (!VA.isRegLoc()) 829 return false; 830 831 // The calling-convention tables for x87 returns don't tell 832 // the whole story. 833 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) 834 return false; 835 836 unsigned SrcReg = Reg + VA.getValNo(); 837 EVT SrcVT = TLI.getValueType(RV->getType()); 838 EVT DstVT = VA.getValVT(); 839 // Special handling for extended integers. 840 if (SrcVT != DstVT) { 841 if (SrcVT != MVT::i1 && SrcVT != MVT::i8 && SrcVT != MVT::i16) 842 return false; 843 844 if (!Outs[0].Flags.isZExt() && !Outs[0].Flags.isSExt()) 845 return false; 846 847 assert(DstVT == MVT::i32 && "X86 should always ext to i32"); 848 849 if (SrcVT == MVT::i1) { 850 if (Outs[0].Flags.isSExt()) 851 return false; 852 SrcReg = FastEmitZExtFromI1(MVT::i8, SrcReg, /*TODO: Kill=*/false); 853 SrcVT = MVT::i8; 854 } 855 unsigned Op = Outs[0].Flags.isZExt() ? ISD::ZERO_EXTEND : 856 ISD::SIGN_EXTEND; 857 SrcReg = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), Op, 858 SrcReg, /*TODO: Kill=*/false); 859 } 860 861 // Make the copy. 862 unsigned DstReg = VA.getLocReg(); 863 const TargetRegisterClass* SrcRC = MRI.getRegClass(SrcReg); 864 // Avoid a cross-class copy. This is very unlikely. 865 if (!SrcRC->contains(DstReg)) 866 return false; 867 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TargetOpcode::COPY), 868 DstReg).addReg(SrcReg); 869 870 // Add register to return instruction. 871 RetRegs.push_back(VA.getLocReg()); 872 } 873 874 // The x86-64 ABI for returning structs by value requires that we copy 875 // the sret argument into %rax for the return. We saved the argument into 876 // a virtual register in the entry block, so now we copy the value out 877 // and into %rax. We also do the same with %eax for Win32. 878 if (F.hasStructRetAttr() && 879 (Subtarget->is64Bit() || Subtarget->isTargetKnownWindowsMSVC())) { 880 unsigned Reg = X86MFInfo->getSRetReturnReg(); 881 assert(Reg && 882 "SRetReturnReg should have been set in LowerFormalArguments()!"); 883 unsigned RetReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX; 884 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TargetOpcode::COPY), 885 RetReg).addReg(Reg); 886 RetRegs.push_back(RetReg); 887 } 888 889 // Now emit the RET. 890 MachineInstrBuilder MIB = 891 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Subtarget->is64Bit() ? X86::RETQ : X86::RETL)); 892 for (unsigned i = 0, e = RetRegs.size(); i != e; ++i) 893 MIB.addReg(RetRegs[i], RegState::Implicit); 894 return true; 895} 896 897/// X86SelectLoad - Select and emit code to implement load instructions. 898/// 899bool X86FastISel::X86SelectLoad(const Instruction *I) { 900 // Atomic loads need special handling. 901 if (cast<LoadInst>(I)->isAtomic()) 902 return false; 903 904 MVT VT; 905 if (!isTypeLegal(I->getType(), VT, /*AllowI1=*/true)) 906 return false; 907 908 X86AddressMode AM; 909 if (!X86SelectAddress(I->getOperand(0), AM)) 910 return false; 911 912 unsigned ResultReg = 0; 913 if (X86FastEmitLoad(VT, AM, ResultReg)) { 914 UpdateValueMap(I, ResultReg); 915 return true; 916 } 917 return false; 918} 919 920static unsigned X86ChooseCmpOpcode(EVT VT, const X86Subtarget *Subtarget) { 921 bool HasAVX = Subtarget->hasAVX(); 922 bool X86ScalarSSEf32 = Subtarget->hasSSE1(); 923 bool X86ScalarSSEf64 = Subtarget->hasSSE2(); 924 925 switch (VT.getSimpleVT().SimpleTy) { 926 default: return 0; 927 case MVT::i8: return X86::CMP8rr; 928 case MVT::i16: return X86::CMP16rr; 929 case MVT::i32: return X86::CMP32rr; 930 case MVT::i64: return X86::CMP64rr; 931 case MVT::f32: 932 return X86ScalarSSEf32 ? (HasAVX ? X86::VUCOMISSrr : X86::UCOMISSrr) : 0; 933 case MVT::f64: 934 return X86ScalarSSEf64 ? (HasAVX ? X86::VUCOMISDrr : X86::UCOMISDrr) : 0; 935 } 936} 937 938/// X86ChooseCmpImmediateOpcode - If we have a comparison with RHS as the RHS 939/// of the comparison, return an opcode that works for the compare (e.g. 940/// CMP32ri) otherwise return 0. 941static unsigned X86ChooseCmpImmediateOpcode(EVT VT, const ConstantInt *RHSC) { 942 switch (VT.getSimpleVT().SimpleTy) { 943 // Otherwise, we can't fold the immediate into this comparison. 944 default: return 0; 945 case MVT::i8: return X86::CMP8ri; 946 case MVT::i16: return X86::CMP16ri; 947 case MVT::i32: return X86::CMP32ri; 948 case MVT::i64: 949 // 64-bit comparisons are only valid if the immediate fits in a 32-bit sext 950 // field. 951 if ((int)RHSC->getSExtValue() == RHSC->getSExtValue()) 952 return X86::CMP64ri32; 953 return 0; 954 } 955} 956 957bool X86FastISel::X86FastEmitCompare(const Value *Op0, const Value *Op1, 958 EVT VT) { 959 unsigned Op0Reg = getRegForValue(Op0); 960 if (Op0Reg == 0) return false; 961 962 // Handle 'null' like i32/i64 0. 963 if (isa<ConstantPointerNull>(Op1)) 964 Op1 = Constant::getNullValue(DL.getIntPtrType(Op0->getContext())); 965 966 // We have two options: compare with register or immediate. If the RHS of 967 // the compare is an immediate that we can fold into this compare, use 968 // CMPri, otherwise use CMPrr. 969 if (const ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) { 970 if (unsigned CompareImmOpc = X86ChooseCmpImmediateOpcode(VT, Op1C)) { 971 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CompareImmOpc)) 972 .addReg(Op0Reg) 973 .addImm(Op1C->getSExtValue()); 974 return true; 975 } 976 } 977 978 unsigned CompareOpc = X86ChooseCmpOpcode(VT, Subtarget); 979 if (CompareOpc == 0) return false; 980 981 unsigned Op1Reg = getRegForValue(Op1); 982 if (Op1Reg == 0) return false; 983 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CompareOpc)) 984 .addReg(Op0Reg) 985 .addReg(Op1Reg); 986 987 return true; 988} 989 990bool X86FastISel::X86SelectCmp(const Instruction *I) { 991 const CmpInst *CI = cast<CmpInst>(I); 992 993 MVT VT; 994 if (!isTypeLegal(I->getOperand(0)->getType(), VT)) 995 return false; 996 997 unsigned ResultReg = createResultReg(&X86::GR8RegClass); 998 unsigned SetCCOpc; 999 bool SwapArgs; // false -> compare Op0, Op1. true -> compare Op1, Op0. 1000 switch (CI->getPredicate()) { 1001 case CmpInst::FCMP_OEQ: { 1002 if (!X86FastEmitCompare(CI->getOperand(0), CI->getOperand(1), VT)) 1003 return false; 1004 1005 unsigned EReg = createResultReg(&X86::GR8RegClass); 1006 unsigned NPReg = createResultReg(&X86::GR8RegClass); 1007 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::SETEr), EReg); 1008 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, 1009 TII.get(X86::SETNPr), NPReg); 1010 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, 1011 TII.get(X86::AND8rr), ResultReg).addReg(NPReg).addReg(EReg); 1012 UpdateValueMap(I, ResultReg); 1013 return true; 1014 } 1015 case CmpInst::FCMP_UNE: { 1016 if (!X86FastEmitCompare(CI->getOperand(0), CI->getOperand(1), VT)) 1017 return false; 1018 1019 unsigned NEReg = createResultReg(&X86::GR8RegClass); 1020 unsigned PReg = createResultReg(&X86::GR8RegClass); 1021 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::SETNEr), NEReg); 1022 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::SETPr), PReg); 1023 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::OR8rr),ResultReg) 1024 .addReg(PReg).addReg(NEReg); 1025 UpdateValueMap(I, ResultReg); 1026 return true; 1027 } 1028 case CmpInst::FCMP_OGT: SwapArgs = false; SetCCOpc = X86::SETAr; break; 1029 case CmpInst::FCMP_OGE: SwapArgs = false; SetCCOpc = X86::SETAEr; break; 1030 case CmpInst::FCMP_OLT: SwapArgs = true; SetCCOpc = X86::SETAr; break; 1031 case CmpInst::FCMP_OLE: SwapArgs = true; SetCCOpc = X86::SETAEr; break; 1032 case CmpInst::FCMP_ONE: SwapArgs = false; SetCCOpc = X86::SETNEr; break; 1033 case CmpInst::FCMP_ORD: SwapArgs = false; SetCCOpc = X86::SETNPr; break; 1034 case CmpInst::FCMP_UNO: SwapArgs = false; SetCCOpc = X86::SETPr; break; 1035 case CmpInst::FCMP_UEQ: SwapArgs = false; SetCCOpc = X86::SETEr; break; 1036 case CmpInst::FCMP_UGT: SwapArgs = true; SetCCOpc = X86::SETBr; break; 1037 case CmpInst::FCMP_UGE: SwapArgs = true; SetCCOpc = X86::SETBEr; break; 1038 case CmpInst::FCMP_ULT: SwapArgs = false; SetCCOpc = X86::SETBr; break; 1039 case CmpInst::FCMP_ULE: SwapArgs = false; SetCCOpc = X86::SETBEr; break; 1040 1041 case CmpInst::ICMP_EQ: SwapArgs = false; SetCCOpc = X86::SETEr; break; 1042 case CmpInst::ICMP_NE: SwapArgs = false; SetCCOpc = X86::SETNEr; break; 1043 case CmpInst::ICMP_UGT: SwapArgs = false; SetCCOpc = X86::SETAr; break; 1044 case CmpInst::ICMP_UGE: SwapArgs = false; SetCCOpc = X86::SETAEr; break; 1045 case CmpInst::ICMP_ULT: SwapArgs = false; SetCCOpc = X86::SETBr; break; 1046 case CmpInst::ICMP_ULE: SwapArgs = false; SetCCOpc = X86::SETBEr; break; 1047 case CmpInst::ICMP_SGT: SwapArgs = false; SetCCOpc = X86::SETGr; break; 1048 case CmpInst::ICMP_SGE: SwapArgs = false; SetCCOpc = X86::SETGEr; break; 1049 case CmpInst::ICMP_SLT: SwapArgs = false; SetCCOpc = X86::SETLr; break; 1050 case CmpInst::ICMP_SLE: SwapArgs = false; SetCCOpc = X86::SETLEr; break; 1051 default: 1052 return false; 1053 } 1054 1055 const Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1); 1056 if (SwapArgs) 1057 std::swap(Op0, Op1); 1058 1059 // Emit a compare of Op0/Op1. 1060 if (!X86FastEmitCompare(Op0, Op1, VT)) 1061 return false; 1062 1063 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(SetCCOpc), ResultReg); 1064 UpdateValueMap(I, ResultReg); 1065 return true; 1066} 1067 1068bool X86FastISel::X86SelectZExt(const Instruction *I) { 1069 EVT DstVT = TLI.getValueType(I->getType()); 1070 if (!TLI.isTypeLegal(DstVT)) 1071 return false; 1072 1073 unsigned ResultReg = getRegForValue(I->getOperand(0)); 1074 if (ResultReg == 0) 1075 return false; 1076 1077 // Handle zero-extension from i1 to i8, which is common. 1078 MVT SrcVT = TLI.getSimpleValueType(I->getOperand(0)->getType()); 1079 if (SrcVT.SimpleTy == MVT::i1) { 1080 // Set the high bits to zero. 1081 ResultReg = FastEmitZExtFromI1(MVT::i8, ResultReg, /*TODO: Kill=*/false); 1082 SrcVT = MVT::i8; 1083 1084 if (ResultReg == 0) 1085 return false; 1086 } 1087 1088 if (DstVT == MVT::i64) { 1089 // Handle extension to 64-bits via sub-register shenanigans. 1090 unsigned MovInst; 1091 1092 switch (SrcVT.SimpleTy) { 1093 case MVT::i8: MovInst = X86::MOVZX32rr8; break; 1094 case MVT::i16: MovInst = X86::MOVZX32rr16; break; 1095 case MVT::i32: MovInst = X86::MOV32rr; break; 1096 default: llvm_unreachable("Unexpected zext to i64 source type"); 1097 } 1098 1099 unsigned Result32 = createResultReg(&X86::GR32RegClass); 1100 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(MovInst), Result32) 1101 .addReg(ResultReg); 1102 1103 ResultReg = createResultReg(&X86::GR64RegClass); 1104 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TargetOpcode::SUBREG_TO_REG), 1105 ResultReg) 1106 .addImm(0).addReg(Result32).addImm(X86::sub_32bit); 1107 } else if (DstVT != MVT::i8) { 1108 ResultReg = FastEmit_r(MVT::i8, DstVT.getSimpleVT(), ISD::ZERO_EXTEND, 1109 ResultReg, /*Kill=*/true); 1110 if (ResultReg == 0) 1111 return false; 1112 } 1113 1114 UpdateValueMap(I, ResultReg); 1115 return true; 1116} 1117 1118 1119bool X86FastISel::X86SelectBranch(const Instruction *I) { 1120 // Unconditional branches are selected by tablegen-generated code. 1121 // Handle a conditional branch. 1122 const BranchInst *BI = cast<BranchInst>(I); 1123 MachineBasicBlock *TrueMBB = FuncInfo.MBBMap[BI->getSuccessor(0)]; 1124 MachineBasicBlock *FalseMBB = FuncInfo.MBBMap[BI->getSuccessor(1)]; 1125 1126 // Fold the common case of a conditional branch with a comparison 1127 // in the same block (values defined on other blocks may not have 1128 // initialized registers). 1129 if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) { 1130 if (CI->hasOneUse() && CI->getParent() == I->getParent()) { 1131 EVT VT = TLI.getValueType(CI->getOperand(0)->getType()); 1132 1133 // Try to take advantage of fallthrough opportunities. 1134 CmpInst::Predicate Predicate = CI->getPredicate(); 1135 if (FuncInfo.MBB->isLayoutSuccessor(TrueMBB)) { 1136 std::swap(TrueMBB, FalseMBB); 1137 Predicate = CmpInst::getInversePredicate(Predicate); 1138 } 1139 1140 bool SwapArgs; // false -> compare Op0, Op1. true -> compare Op1, Op0. 1141 unsigned BranchOpc; // Opcode to jump on, e.g. "X86::JA" 1142 1143 switch (Predicate) { 1144 case CmpInst::FCMP_OEQ: 1145 std::swap(TrueMBB, FalseMBB); 1146 Predicate = CmpInst::FCMP_UNE; 1147 // FALL THROUGH 1148 case CmpInst::FCMP_UNE: SwapArgs = false; BranchOpc = X86::JNE_4; break; 1149 case CmpInst::FCMP_OGT: SwapArgs = false; BranchOpc = X86::JA_4; break; 1150 case CmpInst::FCMP_OGE: SwapArgs = false; BranchOpc = X86::JAE_4; break; 1151 case CmpInst::FCMP_OLT: SwapArgs = true; BranchOpc = X86::JA_4; break; 1152 case CmpInst::FCMP_OLE: SwapArgs = true; BranchOpc = X86::JAE_4; break; 1153 case CmpInst::FCMP_ONE: SwapArgs = false; BranchOpc = X86::JNE_4; break; 1154 case CmpInst::FCMP_ORD: SwapArgs = false; BranchOpc = X86::JNP_4; break; 1155 case CmpInst::FCMP_UNO: SwapArgs = false; BranchOpc = X86::JP_4; break; 1156 case CmpInst::FCMP_UEQ: SwapArgs = false; BranchOpc = X86::JE_4; break; 1157 case CmpInst::FCMP_UGT: SwapArgs = true; BranchOpc = X86::JB_4; break; 1158 case CmpInst::FCMP_UGE: SwapArgs = true; BranchOpc = X86::JBE_4; break; 1159 case CmpInst::FCMP_ULT: SwapArgs = false; BranchOpc = X86::JB_4; break; 1160 case CmpInst::FCMP_ULE: SwapArgs = false; BranchOpc = X86::JBE_4; break; 1161 1162 case CmpInst::ICMP_EQ: SwapArgs = false; BranchOpc = X86::JE_4; break; 1163 case CmpInst::ICMP_NE: SwapArgs = false; BranchOpc = X86::JNE_4; break; 1164 case CmpInst::ICMP_UGT: SwapArgs = false; BranchOpc = X86::JA_4; break; 1165 case CmpInst::ICMP_UGE: SwapArgs = false; BranchOpc = X86::JAE_4; break; 1166 case CmpInst::ICMP_ULT: SwapArgs = false; BranchOpc = X86::JB_4; break; 1167 case CmpInst::ICMP_ULE: SwapArgs = false; BranchOpc = X86::JBE_4; break; 1168 case CmpInst::ICMP_SGT: SwapArgs = false; BranchOpc = X86::JG_4; break; 1169 case CmpInst::ICMP_SGE: SwapArgs = false; BranchOpc = X86::JGE_4; break; 1170 case CmpInst::ICMP_SLT: SwapArgs = false; BranchOpc = X86::JL_4; break; 1171 case CmpInst::ICMP_SLE: SwapArgs = false; BranchOpc = X86::JLE_4; break; 1172 default: 1173 return false; 1174 } 1175 1176 const Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1); 1177 if (SwapArgs) 1178 std::swap(Op0, Op1); 1179 1180 // Emit a compare of the LHS and RHS, setting the flags. 1181 if (!X86FastEmitCompare(Op0, Op1, VT)) 1182 return false; 1183 1184 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(BranchOpc)) 1185 .addMBB(TrueMBB); 1186 1187 if (Predicate == CmpInst::FCMP_UNE) { 1188 // X86 requires a second branch to handle UNE (and OEQ, 1189 // which is mapped to UNE above). 1190 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::JP_4)) 1191 .addMBB(TrueMBB); 1192 } 1193 1194 FastEmitBranch(FalseMBB, DbgLoc); 1195 FuncInfo.MBB->addSuccessor(TrueMBB); 1196 return true; 1197 } 1198 } else if (TruncInst *TI = dyn_cast<TruncInst>(BI->getCondition())) { 1199 // Handle things like "%cond = trunc i32 %X to i1 / br i1 %cond", which 1200 // typically happen for _Bool and C++ bools. 1201 MVT SourceVT; 1202 if (TI->hasOneUse() && TI->getParent() == I->getParent() && 1203 isTypeLegal(TI->getOperand(0)->getType(), SourceVT)) { 1204 unsigned TestOpc = 0; 1205 switch (SourceVT.SimpleTy) { 1206 default: break; 1207 case MVT::i8: TestOpc = X86::TEST8ri; break; 1208 case MVT::i16: TestOpc = X86::TEST16ri; break; 1209 case MVT::i32: TestOpc = X86::TEST32ri; break; 1210 case MVT::i64: TestOpc = X86::TEST64ri32; break; 1211 } 1212 if (TestOpc) { 1213 unsigned OpReg = getRegForValue(TI->getOperand(0)); 1214 if (OpReg == 0) return false; 1215 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TestOpc)) 1216 .addReg(OpReg).addImm(1); 1217 1218 unsigned JmpOpc = X86::JNE_4; 1219 if (FuncInfo.MBB->isLayoutSuccessor(TrueMBB)) { 1220 std::swap(TrueMBB, FalseMBB); 1221 JmpOpc = X86::JE_4; 1222 } 1223 1224 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(JmpOpc)) 1225 .addMBB(TrueMBB); 1226 FastEmitBranch(FalseMBB, DbgLoc); 1227 FuncInfo.MBB->addSuccessor(TrueMBB); 1228 return true; 1229 } 1230 } 1231 } 1232 1233 // Otherwise do a clumsy setcc and re-test it. 1234 // Note that i1 essentially gets ANY_EXTEND'ed to i8 where it isn't used 1235 // in an explicit cast, so make sure to handle that correctly. 1236 unsigned OpReg = getRegForValue(BI->getCondition()); 1237 if (OpReg == 0) return false; 1238 1239 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::TEST8ri)) 1240 .addReg(OpReg).addImm(1); 1241 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::JNE_4)) 1242 .addMBB(TrueMBB); 1243 FastEmitBranch(FalseMBB, DbgLoc); 1244 FuncInfo.MBB->addSuccessor(TrueMBB); 1245 return true; 1246} 1247 1248bool X86FastISel::X86SelectShift(const Instruction *I) { 1249 unsigned CReg = 0, OpReg = 0; 1250 const TargetRegisterClass *RC = nullptr; 1251 if (I->getType()->isIntegerTy(8)) { 1252 CReg = X86::CL; 1253 RC = &X86::GR8RegClass; 1254 switch (I->getOpcode()) { 1255 case Instruction::LShr: OpReg = X86::SHR8rCL; break; 1256 case Instruction::AShr: OpReg = X86::SAR8rCL; break; 1257 case Instruction::Shl: OpReg = X86::SHL8rCL; break; 1258 default: return false; 1259 } 1260 } else if (I->getType()->isIntegerTy(16)) { 1261 CReg = X86::CX; 1262 RC = &X86::GR16RegClass; 1263 switch (I->getOpcode()) { 1264 case Instruction::LShr: OpReg = X86::SHR16rCL; break; 1265 case Instruction::AShr: OpReg = X86::SAR16rCL; break; 1266 case Instruction::Shl: OpReg = X86::SHL16rCL; break; 1267 default: return false; 1268 } 1269 } else if (I->getType()->isIntegerTy(32)) { 1270 CReg = X86::ECX; 1271 RC = &X86::GR32RegClass; 1272 switch (I->getOpcode()) { 1273 case Instruction::LShr: OpReg = X86::SHR32rCL; break; 1274 case Instruction::AShr: OpReg = X86::SAR32rCL; break; 1275 case Instruction::Shl: OpReg = X86::SHL32rCL; break; 1276 default: return false; 1277 } 1278 } else if (I->getType()->isIntegerTy(64)) { 1279 CReg = X86::RCX; 1280 RC = &X86::GR64RegClass; 1281 switch (I->getOpcode()) { 1282 case Instruction::LShr: OpReg = X86::SHR64rCL; break; 1283 case Instruction::AShr: OpReg = X86::SAR64rCL; break; 1284 case Instruction::Shl: OpReg = X86::SHL64rCL; break; 1285 default: return false; 1286 } 1287 } else { 1288 return false; 1289 } 1290 1291 MVT VT; 1292 if (!isTypeLegal(I->getType(), VT)) 1293 return false; 1294 1295 unsigned Op0Reg = getRegForValue(I->getOperand(0)); 1296 if (Op0Reg == 0) return false; 1297 1298 unsigned Op1Reg = getRegForValue(I->getOperand(1)); 1299 if (Op1Reg == 0) return false; 1300 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TargetOpcode::COPY), 1301 CReg).addReg(Op1Reg); 1302 1303 // The shift instruction uses X86::CL. If we defined a super-register 1304 // of X86::CL, emit a subreg KILL to precisely describe what we're doing here. 1305 if (CReg != X86::CL) 1306 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, 1307 TII.get(TargetOpcode::KILL), X86::CL) 1308 .addReg(CReg, RegState::Kill); 1309 1310 unsigned ResultReg = createResultReg(RC); 1311 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(OpReg), ResultReg) 1312 .addReg(Op0Reg); 1313 UpdateValueMap(I, ResultReg); 1314 return true; 1315} 1316 1317bool X86FastISel::X86SelectDivRem(const Instruction *I) { 1318 const static unsigned NumTypes = 4; // i8, i16, i32, i64 1319 const static unsigned NumOps = 4; // SDiv, SRem, UDiv, URem 1320 const static bool S = true; // IsSigned 1321 const static bool U = false; // !IsSigned 1322 const static unsigned Copy = TargetOpcode::COPY; 1323 // For the X86 DIV/IDIV instruction, in most cases the dividend 1324 // (numerator) must be in a specific register pair highreg:lowreg, 1325 // producing the quotient in lowreg and the remainder in highreg. 1326 // For most data types, to set up the instruction, the dividend is 1327 // copied into lowreg, and lowreg is sign-extended or zero-extended 1328 // into highreg. The exception is i8, where the dividend is defined 1329 // as a single register rather than a register pair, and we 1330 // therefore directly sign-extend or zero-extend the dividend into 1331 // lowreg, instead of copying, and ignore the highreg. 1332 const static struct DivRemEntry { 1333 // The following portion depends only on the data type. 1334 const TargetRegisterClass *RC; 1335 unsigned LowInReg; // low part of the register pair 1336 unsigned HighInReg; // high part of the register pair 1337 // The following portion depends on both the data type and the operation. 1338 struct DivRemResult { 1339 unsigned OpDivRem; // The specific DIV/IDIV opcode to use. 1340 unsigned OpSignExtend; // Opcode for sign-extending lowreg into 1341 // highreg, or copying a zero into highreg. 1342 unsigned OpCopy; // Opcode for copying dividend into lowreg, or 1343 // zero/sign-extending into lowreg for i8. 1344 unsigned DivRemResultReg; // Register containing the desired result. 1345 bool IsOpSigned; // Whether to use signed or unsigned form. 1346 } ResultTable[NumOps]; 1347 } OpTable[NumTypes] = { 1348 { &X86::GR8RegClass, X86::AX, 0, { 1349 { X86::IDIV8r, 0, X86::MOVSX16rr8, X86::AL, S }, // SDiv 1350 { X86::IDIV8r, 0, X86::MOVSX16rr8, X86::AH, S }, // SRem 1351 { X86::DIV8r, 0, X86::MOVZX16rr8, X86::AL, U }, // UDiv 1352 { X86::DIV8r, 0, X86::MOVZX16rr8, X86::AH, U }, // URem 1353 } 1354 }, // i8 1355 { &X86::GR16RegClass, X86::AX, X86::DX, { 1356 { X86::IDIV16r, X86::CWD, Copy, X86::AX, S }, // SDiv 1357 { X86::IDIV16r, X86::CWD, Copy, X86::DX, S }, // SRem 1358 { X86::DIV16r, X86::MOV32r0, Copy, X86::AX, U }, // UDiv 1359 { X86::DIV16r, X86::MOV32r0, Copy, X86::DX, U }, // URem 1360 } 1361 }, // i16 1362 { &X86::GR32RegClass, X86::EAX, X86::EDX, { 1363 { X86::IDIV32r, X86::CDQ, Copy, X86::EAX, S }, // SDiv 1364 { X86::IDIV32r, X86::CDQ, Copy, X86::EDX, S }, // SRem 1365 { X86::DIV32r, X86::MOV32r0, Copy, X86::EAX, U }, // UDiv 1366 { X86::DIV32r, X86::MOV32r0, Copy, X86::EDX, U }, // URem 1367 } 1368 }, // i32 1369 { &X86::GR64RegClass, X86::RAX, X86::RDX, { 1370 { X86::IDIV64r, X86::CQO, Copy, X86::RAX, S }, // SDiv 1371 { X86::IDIV64r, X86::CQO, Copy, X86::RDX, S }, // SRem 1372 { X86::DIV64r, X86::MOV32r0, Copy, X86::RAX, U }, // UDiv 1373 { X86::DIV64r, X86::MOV32r0, Copy, X86::RDX, U }, // URem 1374 } 1375 }, // i64 1376 }; 1377 1378 MVT VT; 1379 if (!isTypeLegal(I->getType(), VT)) 1380 return false; 1381 1382 unsigned TypeIndex, OpIndex; 1383 switch (VT.SimpleTy) { 1384 default: return false; 1385 case MVT::i8: TypeIndex = 0; break; 1386 case MVT::i16: TypeIndex = 1; break; 1387 case MVT::i32: TypeIndex = 2; break; 1388 case MVT::i64: TypeIndex = 3; 1389 if (!Subtarget->is64Bit()) 1390 return false; 1391 break; 1392 } 1393 1394 switch (I->getOpcode()) { 1395 default: llvm_unreachable("Unexpected div/rem opcode"); 1396 case Instruction::SDiv: OpIndex = 0; break; 1397 case Instruction::SRem: OpIndex = 1; break; 1398 case Instruction::UDiv: OpIndex = 2; break; 1399 case Instruction::URem: OpIndex = 3; break; 1400 } 1401 1402 const DivRemEntry &TypeEntry = OpTable[TypeIndex]; 1403 const DivRemEntry::DivRemResult &OpEntry = TypeEntry.ResultTable[OpIndex]; 1404 unsigned Op0Reg = getRegForValue(I->getOperand(0)); 1405 if (Op0Reg == 0) 1406 return false; 1407 unsigned Op1Reg = getRegForValue(I->getOperand(1)); 1408 if (Op1Reg == 0) 1409 return false; 1410 1411 // Move op0 into low-order input register. 1412 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, 1413 TII.get(OpEntry.OpCopy), TypeEntry.LowInReg).addReg(Op0Reg); 1414 // Zero-extend or sign-extend into high-order input register. 1415 if (OpEntry.OpSignExtend) { 1416 if (OpEntry.IsOpSigned) 1417 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, 1418 TII.get(OpEntry.OpSignExtend)); 1419 else { 1420 unsigned Zero32 = createResultReg(&X86::GR32RegClass); 1421 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, 1422 TII.get(X86::MOV32r0), Zero32); 1423 1424 // Copy the zero into the appropriate sub/super/identical physical 1425 // register. Unfortunately the operations needed are not uniform enough to 1426 // fit neatly into the table above. 1427 if (VT.SimpleTy == MVT::i16) { 1428 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, 1429 TII.get(Copy), TypeEntry.HighInReg) 1430 .addReg(Zero32, 0, X86::sub_16bit); 1431 } else if (VT.SimpleTy == MVT::i32) { 1432 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, 1433 TII.get(Copy), TypeEntry.HighInReg) 1434 .addReg(Zero32); 1435 } else if (VT.SimpleTy == MVT::i64) { 1436 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, 1437 TII.get(TargetOpcode::SUBREG_TO_REG), TypeEntry.HighInReg) 1438 .addImm(0).addReg(Zero32).addImm(X86::sub_32bit); 1439 } 1440 } 1441 } 1442 // Generate the DIV/IDIV instruction. 1443 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, 1444 TII.get(OpEntry.OpDivRem)).addReg(Op1Reg); 1445 // For i8 remainder, we can't reference AH directly, as we'll end 1446 // up with bogus copies like %R9B = COPY %AH. Reference AX 1447 // instead to prevent AH references in a REX instruction. 1448 // 1449 // The current assumption of the fast register allocator is that isel 1450 // won't generate explicit references to the GPR8_NOREX registers. If 1451 // the allocator and/or the backend get enhanced to be more robust in 1452 // that regard, this can be, and should be, removed. 1453 unsigned ResultReg = 0; 1454 if ((I->getOpcode() == Instruction::SRem || 1455 I->getOpcode() == Instruction::URem) && 1456 OpEntry.DivRemResultReg == X86::AH && Subtarget->is64Bit()) { 1457 unsigned SourceSuperReg = createResultReg(&X86::GR16RegClass); 1458 unsigned ResultSuperReg = createResultReg(&X86::GR16RegClass); 1459 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, 1460 TII.get(Copy), SourceSuperReg).addReg(X86::AX); 1461 1462 // Shift AX right by 8 bits instead of using AH. 1463 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::SHR16ri), 1464 ResultSuperReg).addReg(SourceSuperReg).addImm(8); 1465 1466 // Now reference the 8-bit subreg of the result. 1467 ResultReg = FastEmitInst_extractsubreg(MVT::i8, ResultSuperReg, 1468 /*Kill=*/true, X86::sub_8bit); 1469 } 1470 // Copy the result out of the physreg if we haven't already. 1471 if (!ResultReg) { 1472 ResultReg = createResultReg(TypeEntry.RC); 1473 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Copy), ResultReg) 1474 .addReg(OpEntry.DivRemResultReg); 1475 } 1476 UpdateValueMap(I, ResultReg); 1477 1478 return true; 1479} 1480 1481bool X86FastISel::X86SelectSelect(const Instruction *I) { 1482 MVT VT; 1483 if (!isTypeLegal(I->getType(), VT)) 1484 return false; 1485 1486 // We only use cmov here, if we don't have a cmov instruction bail. 1487 if (!Subtarget->hasCMov()) return false; 1488 1489 unsigned Opc = 0; 1490 const TargetRegisterClass *RC = nullptr; 1491 if (VT == MVT::i16) { 1492 Opc = X86::CMOVE16rr; 1493 RC = &X86::GR16RegClass; 1494 } else if (VT == MVT::i32) { 1495 Opc = X86::CMOVE32rr; 1496 RC = &X86::GR32RegClass; 1497 } else if (VT == MVT::i64) { 1498 Opc = X86::CMOVE64rr; 1499 RC = &X86::GR64RegClass; 1500 } else { 1501 return false; 1502 } 1503 1504 unsigned Op0Reg = getRegForValue(I->getOperand(0)); 1505 if (Op0Reg == 0) return false; 1506 unsigned Op1Reg = getRegForValue(I->getOperand(1)); 1507 if (Op1Reg == 0) return false; 1508 unsigned Op2Reg = getRegForValue(I->getOperand(2)); 1509 if (Op2Reg == 0) return false; 1510 1511 // Selects operate on i1, however, Op0Reg is 8 bits width and may contain 1512 // garbage. Indeed, only the less significant bit is supposed to be accurate. 1513 // If we read more than the lsb, we may see non-zero values whereas lsb 1514 // is zero. Therefore, we have to truncate Op0Reg to i1 for the select. 1515 // This is achieved by performing TEST against 1. 1516 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::TEST8ri)) 1517 .addReg(Op0Reg).addImm(1); 1518 unsigned ResultReg = createResultReg(RC); 1519 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg) 1520 .addReg(Op1Reg).addReg(Op2Reg); 1521 UpdateValueMap(I, ResultReg); 1522 return true; 1523} 1524 1525bool X86FastISel::X86SelectFPExt(const Instruction *I) { 1526 // fpext from float to double. 1527 if (X86ScalarSSEf64 && 1528 I->getType()->isDoubleTy()) { 1529 const Value *V = I->getOperand(0); 1530 if (V->getType()->isFloatTy()) { 1531 unsigned OpReg = getRegForValue(V); 1532 if (OpReg == 0) return false; 1533 unsigned ResultReg = createResultReg(&X86::FR64RegClass); 1534 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, 1535 TII.get(X86::CVTSS2SDrr), ResultReg) 1536 .addReg(OpReg); 1537 UpdateValueMap(I, ResultReg); 1538 return true; 1539 } 1540 } 1541 1542 return false; 1543} 1544 1545bool X86FastISel::X86SelectFPTrunc(const Instruction *I) { 1546 if (X86ScalarSSEf64) { 1547 if (I->getType()->isFloatTy()) { 1548 const Value *V = I->getOperand(0); 1549 if (V->getType()->isDoubleTy()) { 1550 unsigned OpReg = getRegForValue(V); 1551 if (OpReg == 0) return false; 1552 unsigned ResultReg = createResultReg(&X86::FR32RegClass); 1553 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, 1554 TII.get(X86::CVTSD2SSrr), ResultReg) 1555 .addReg(OpReg); 1556 UpdateValueMap(I, ResultReg); 1557 return true; 1558 } 1559 } 1560 } 1561 1562 return false; 1563} 1564 1565bool X86FastISel::X86SelectTrunc(const Instruction *I) { 1566 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType()); 1567 EVT DstVT = TLI.getValueType(I->getType()); 1568 1569 // This code only handles truncation to byte. 1570 if (DstVT != MVT::i8 && DstVT != MVT::i1) 1571 return false; 1572 if (!TLI.isTypeLegal(SrcVT)) 1573 return false; 1574 1575 unsigned InputReg = getRegForValue(I->getOperand(0)); 1576 if (!InputReg) 1577 // Unhandled operand. Halt "fast" selection and bail. 1578 return false; 1579 1580 if (SrcVT == MVT::i8) { 1581 // Truncate from i8 to i1; no code needed. 1582 UpdateValueMap(I, InputReg); 1583 return true; 1584 } 1585 1586 if (!Subtarget->is64Bit()) { 1587 // If we're on x86-32; we can't extract an i8 from a general register. 1588 // First issue a copy to GR16_ABCD or GR32_ABCD. 1589 const TargetRegisterClass *CopyRC = (SrcVT == MVT::i16) ? 1590 (const TargetRegisterClass*)&X86::GR16_ABCDRegClass : 1591 (const TargetRegisterClass*)&X86::GR32_ABCDRegClass; 1592 unsigned CopyReg = createResultReg(CopyRC); 1593 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TargetOpcode::COPY), 1594 CopyReg).addReg(InputReg); 1595 InputReg = CopyReg; 1596 } 1597 1598 // Issue an extract_subreg. 1599 unsigned ResultReg = FastEmitInst_extractsubreg(MVT::i8, 1600 InputReg, /*Kill=*/true, 1601 X86::sub_8bit); 1602 if (!ResultReg) 1603 return false; 1604 1605 UpdateValueMap(I, ResultReg); 1606 return true; 1607} 1608 1609bool X86FastISel::IsMemcpySmall(uint64_t Len) { 1610 return Len <= (Subtarget->is64Bit() ? 32 : 16); 1611} 1612 1613bool X86FastISel::TryEmitSmallMemcpy(X86AddressMode DestAM, 1614 X86AddressMode SrcAM, uint64_t Len) { 1615 1616 // Make sure we don't bloat code by inlining very large memcpy's. 1617 if (!IsMemcpySmall(Len)) 1618 return false; 1619 1620 bool i64Legal = Subtarget->is64Bit(); 1621 1622 // We don't care about alignment here since we just emit integer accesses. 1623 while (Len) { 1624 MVT VT; 1625 if (Len >= 8 && i64Legal) 1626 VT = MVT::i64; 1627 else if (Len >= 4) 1628 VT = MVT::i32; 1629 else if (Len >= 2) 1630 VT = MVT::i16; 1631 else { 1632 VT = MVT::i8; 1633 } 1634 1635 unsigned Reg; 1636 bool RV = X86FastEmitLoad(VT, SrcAM, Reg); 1637 RV &= X86FastEmitStore(VT, Reg, DestAM); 1638 assert(RV && "Failed to emit load or store??"); 1639 1640 unsigned Size = VT.getSizeInBits()/8; 1641 Len -= Size; 1642 DestAM.Disp += Size; 1643 SrcAM.Disp += Size; 1644 } 1645 1646 return true; 1647} 1648 1649bool X86FastISel::X86VisitIntrinsicCall(const IntrinsicInst &I) { 1650 // FIXME: Handle more intrinsics. 1651 switch (I.getIntrinsicID()) { 1652 default: return false; 1653 case Intrinsic::memcpy: { 1654 const MemCpyInst &MCI = cast<MemCpyInst>(I); 1655 // Don't handle volatile or variable length memcpys. 1656 if (MCI.isVolatile()) 1657 return false; 1658 1659 if (isa<ConstantInt>(MCI.getLength())) { 1660 // Small memcpy's are common enough that we want to do them 1661 // without a call if possible. 1662 uint64_t Len = cast<ConstantInt>(MCI.getLength())->getZExtValue(); 1663 if (IsMemcpySmall(Len)) { 1664 X86AddressMode DestAM, SrcAM; 1665 if (!X86SelectAddress(MCI.getRawDest(), DestAM) || 1666 !X86SelectAddress(MCI.getRawSource(), SrcAM)) 1667 return false; 1668 TryEmitSmallMemcpy(DestAM, SrcAM, Len); 1669 return true; 1670 } 1671 } 1672 1673 unsigned SizeWidth = Subtarget->is64Bit() ? 64 : 32; 1674 if (!MCI.getLength()->getType()->isIntegerTy(SizeWidth)) 1675 return false; 1676 1677 if (MCI.getSourceAddressSpace() > 255 || MCI.getDestAddressSpace() > 255) 1678 return false; 1679 1680 return DoSelectCall(&I, "memcpy"); 1681 } 1682 case Intrinsic::memset: { 1683 const MemSetInst &MSI = cast<MemSetInst>(I); 1684 1685 if (MSI.isVolatile()) 1686 return false; 1687 1688 unsigned SizeWidth = Subtarget->is64Bit() ? 64 : 32; 1689 if (!MSI.getLength()->getType()->isIntegerTy(SizeWidth)) 1690 return false; 1691 1692 if (MSI.getDestAddressSpace() > 255) 1693 return false; 1694 1695 return DoSelectCall(&I, "memset"); 1696 } 1697 case Intrinsic::stackprotector: { 1698 // Emit code to store the stack guard onto the stack. 1699 EVT PtrTy = TLI.getPointerTy(); 1700 1701 const Value *Op1 = I.getArgOperand(0); // The guard's value. 1702 const AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1)); 1703 1704 MFI.setStackProtectorIndex(FuncInfo.StaticAllocaMap[Slot]); 1705 1706 // Grab the frame index. 1707 X86AddressMode AM; 1708 if (!X86SelectAddress(Slot, AM)) return false; 1709 if (!X86FastEmitStore(PtrTy, Op1, AM)) return false; 1710 return true; 1711 } 1712 case Intrinsic::dbg_declare: { 1713 const DbgDeclareInst *DI = cast<DbgDeclareInst>(&I); 1714 X86AddressMode AM; 1715 assert(DI->getAddress() && "Null address should be checked earlier!"); 1716 if (!X86SelectAddress(DI->getAddress(), AM)) 1717 return false; 1718 const MCInstrDesc &II = TII.get(TargetOpcode::DBG_VALUE); 1719 // FIXME may need to add RegState::Debug to any registers produced, 1720 // although ESP/EBP should be the only ones at the moment. 1721 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II), AM). 1722 addImm(0).addMetadata(DI->getVariable()); 1723 return true; 1724 } 1725 case Intrinsic::trap: { 1726 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::TRAP)); 1727 return true; 1728 } 1729 case Intrinsic::sadd_with_overflow: 1730 case Intrinsic::uadd_with_overflow: { 1731 // FIXME: Should fold immediates. 1732 1733 // Replace "add with overflow" intrinsics with an "add" instruction followed 1734 // by a seto/setc instruction. 1735 const Function *Callee = I.getCalledFunction(); 1736 Type *RetTy = 1737 cast<StructType>(Callee->getReturnType())->getTypeAtIndex(unsigned(0)); 1738 1739 MVT VT; 1740 if (!isTypeLegal(RetTy, VT)) 1741 return false; 1742 1743 const Value *Op1 = I.getArgOperand(0); 1744 const Value *Op2 = I.getArgOperand(1); 1745 unsigned Reg1 = getRegForValue(Op1); 1746 unsigned Reg2 = getRegForValue(Op2); 1747 1748 if (Reg1 == 0 || Reg2 == 0) 1749 // FIXME: Handle values *not* in registers. 1750 return false; 1751 1752 unsigned OpC = 0; 1753 if (VT == MVT::i32) 1754 OpC = X86::ADD32rr; 1755 else if (VT == MVT::i64) 1756 OpC = X86::ADD64rr; 1757 else 1758 return false; 1759 1760 // The call to CreateRegs builds two sequential registers, to store the 1761 // both the returned values. 1762 unsigned ResultReg = FuncInfo.CreateRegs(I.getType()); 1763 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(OpC), ResultReg) 1764 .addReg(Reg1).addReg(Reg2); 1765 1766 unsigned Opc = X86::SETBr; 1767 if (I.getIntrinsicID() == Intrinsic::sadd_with_overflow) 1768 Opc = X86::SETOr; 1769 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), 1770 ResultReg + 1); 1771 1772 UpdateValueMap(&I, ResultReg, 2); 1773 return true; 1774 } 1775 } 1776} 1777 1778bool X86FastISel::FastLowerArguments() { 1779 if (!FuncInfo.CanLowerReturn) 1780 return false; 1781 1782 const Function *F = FuncInfo.Fn; 1783 if (F->isVarArg()) 1784 return false; 1785 1786 CallingConv::ID CC = F->getCallingConv(); 1787 if (CC != CallingConv::C) 1788 return false; 1789 1790 if (Subtarget->isCallingConvWin64(CC)) 1791 return false; 1792 1793 if (!Subtarget->is64Bit()) 1794 return false; 1795 1796 // Only handle simple cases. i.e. Up to 6 i32/i64 scalar arguments. 1797 unsigned Idx = 1; 1798 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end(); 1799 I != E; ++I, ++Idx) { 1800 if (Idx > 6) 1801 return false; 1802 1803 if (F->getAttributes().hasAttribute(Idx, Attribute::ByVal) || 1804 F->getAttributes().hasAttribute(Idx, Attribute::InReg) || 1805 F->getAttributes().hasAttribute(Idx, Attribute::StructRet) || 1806 F->getAttributes().hasAttribute(Idx, Attribute::Nest)) 1807 return false; 1808 1809 Type *ArgTy = I->getType(); 1810 if (ArgTy->isStructTy() || ArgTy->isArrayTy() || ArgTy->isVectorTy()) 1811 return false; 1812 1813 EVT ArgVT = TLI.getValueType(ArgTy); 1814 if (!ArgVT.isSimple()) return false; 1815 switch (ArgVT.getSimpleVT().SimpleTy) { 1816 case MVT::i32: 1817 case MVT::i64: 1818 break; 1819 default: 1820 return false; 1821 } 1822 } 1823 1824 static const MCPhysReg GPR32ArgRegs[] = { 1825 X86::EDI, X86::ESI, X86::EDX, X86::ECX, X86::R8D, X86::R9D 1826 }; 1827 static const MCPhysReg GPR64ArgRegs[] = { 1828 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8 , X86::R9 1829 }; 1830 1831 Idx = 0; 1832 const TargetRegisterClass *RC32 = TLI.getRegClassFor(MVT::i32); 1833 const TargetRegisterClass *RC64 = TLI.getRegClassFor(MVT::i64); 1834 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end(); 1835 I != E; ++I, ++Idx) { 1836 bool is32Bit = TLI.getValueType(I->getType()) == MVT::i32; 1837 const TargetRegisterClass *RC = is32Bit ? RC32 : RC64; 1838 unsigned SrcReg = is32Bit ? GPR32ArgRegs[Idx] : GPR64ArgRegs[Idx]; 1839 unsigned DstReg = FuncInfo.MF->addLiveIn(SrcReg, RC); 1840 // FIXME: Unfortunately it's necessary to emit a copy from the livein copy. 1841 // Without this, EmitLiveInCopies may eliminate the livein if its only 1842 // use is a bitcast (which isn't turned into an instruction). 1843 unsigned ResultReg = createResultReg(RC); 1844 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, 1845 TII.get(TargetOpcode::COPY), 1846 ResultReg).addReg(DstReg, getKillRegState(true)); 1847 UpdateValueMap(I, ResultReg); 1848 } 1849 return true; 1850} 1851 1852bool X86FastISel::X86SelectCall(const Instruction *I) { 1853 const CallInst *CI = cast<CallInst>(I); 1854 const Value *Callee = CI->getCalledValue(); 1855 1856 // Can't handle inline asm yet. 1857 if (isa<InlineAsm>(Callee)) 1858 return false; 1859 1860 // Handle intrinsic calls. 1861 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) 1862 return X86VisitIntrinsicCall(*II); 1863 1864 // Allow SelectionDAG isel to handle tail calls. 1865 if (cast<CallInst>(I)->isTailCall()) 1866 return false; 1867 1868 return DoSelectCall(I, nullptr); 1869} 1870 1871static unsigned computeBytesPoppedByCallee(const X86Subtarget &Subtarget, 1872 const ImmutableCallSite &CS) { 1873 if (Subtarget.is64Bit()) 1874 return 0; 1875 if (Subtarget.getTargetTriple().isOSMSVCRT()) 1876 return 0; 1877 CallingConv::ID CC = CS.getCallingConv(); 1878 if (CC == CallingConv::Fast || CC == CallingConv::GHC) 1879 return 0; 1880 if (!CS.paramHasAttr(1, Attribute::StructRet)) 1881 return 0; 1882 if (CS.paramHasAttr(1, Attribute::InReg)) 1883 return 0; 1884 return 4; 1885} 1886 1887// Select either a call, or an llvm.memcpy/memmove/memset intrinsic 1888bool X86FastISel::DoSelectCall(const Instruction *I, const char *MemIntName) { 1889 const CallInst *CI = cast<CallInst>(I); 1890 const Value *Callee = CI->getCalledValue(); 1891 1892 // Handle only C and fastcc calling conventions for now. 1893 ImmutableCallSite CS(CI); 1894 CallingConv::ID CC = CS.getCallingConv(); 1895 bool isWin64 = Subtarget->isCallingConvWin64(CC); 1896 if (CC != CallingConv::C && CC != CallingConv::Fast && 1897 CC != CallingConv::X86_FastCall && CC != CallingConv::X86_64_Win64 && 1898 CC != CallingConv::X86_64_SysV) 1899 return false; 1900 1901 // fastcc with -tailcallopt is intended to provide a guaranteed 1902 // tail call optimization. Fastisel doesn't know how to do that. 1903 if (CC == CallingConv::Fast && TM.Options.GuaranteedTailCallOpt) 1904 return false; 1905 1906 PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType()); 1907 FunctionType *FTy = cast<FunctionType>(PT->getElementType()); 1908 bool isVarArg = FTy->isVarArg(); 1909 1910 // Don't know how to handle Win64 varargs yet. Nothing special needed for 1911 // x86-32. Special handling for x86-64 is implemented. 1912 if (isVarArg && isWin64) 1913 return false; 1914 1915 // Don't know about inalloca yet. 1916 if (CS.hasInAllocaArgument()) 1917 return false; 1918 1919 // Fast-isel doesn't know about callee-pop yet. 1920 if (X86::isCalleePop(CC, Subtarget->is64Bit(), isVarArg, 1921 TM.Options.GuaranteedTailCallOpt)) 1922 return false; 1923 1924 // Check whether the function can return without sret-demotion. 1925 SmallVector<ISD::OutputArg, 4> Outs; 1926 GetReturnInfo(I->getType(), CS.getAttributes(), Outs, TLI); 1927 bool CanLowerReturn = TLI.CanLowerReturn(CS.getCallingConv(), 1928 *FuncInfo.MF, FTy->isVarArg(), 1929 Outs, FTy->getContext()); 1930 if (!CanLowerReturn) 1931 return false; 1932 1933 // Materialize callee address in a register. FIXME: GV address can be 1934 // handled with a CALLpcrel32 instead. 1935 X86AddressMode CalleeAM; 1936 if (!X86SelectCallAddress(Callee, CalleeAM)) 1937 return false; 1938 unsigned CalleeOp = 0; 1939 const GlobalValue *GV = nullptr; 1940 if (CalleeAM.GV != nullptr) { 1941 GV = CalleeAM.GV; 1942 } else if (CalleeAM.Base.Reg != 0) { 1943 CalleeOp = CalleeAM.Base.Reg; 1944 } else 1945 return false; 1946 1947 // Deal with call operands first. 1948 SmallVector<const Value *, 8> ArgVals; 1949 SmallVector<unsigned, 8> Args; 1950 SmallVector<MVT, 8> ArgVTs; 1951 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags; 1952 unsigned arg_size = CS.arg_size(); 1953 Args.reserve(arg_size); 1954 ArgVals.reserve(arg_size); 1955 ArgVTs.reserve(arg_size); 1956 ArgFlags.reserve(arg_size); 1957 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end(); 1958 i != e; ++i) { 1959 // If we're lowering a mem intrinsic instead of a regular call, skip the 1960 // last two arguments, which should not passed to the underlying functions. 1961 if (MemIntName && e-i <= 2) 1962 break; 1963 Value *ArgVal = *i; 1964 ISD::ArgFlagsTy Flags; 1965 unsigned AttrInd = i - CS.arg_begin() + 1; 1966 if (CS.paramHasAttr(AttrInd, Attribute::SExt)) 1967 Flags.setSExt(); 1968 if (CS.paramHasAttr(AttrInd, Attribute::ZExt)) 1969 Flags.setZExt(); 1970 1971 if (CS.paramHasAttr(AttrInd, Attribute::ByVal)) { 1972 PointerType *Ty = cast<PointerType>(ArgVal->getType()); 1973 Type *ElementTy = Ty->getElementType(); 1974 unsigned FrameSize = DL.getTypeAllocSize(ElementTy); 1975 unsigned FrameAlign = CS.getParamAlignment(AttrInd); 1976 if (!FrameAlign) 1977 FrameAlign = TLI.getByValTypeAlignment(ElementTy); 1978 Flags.setByVal(); 1979 Flags.setByValSize(FrameSize); 1980 Flags.setByValAlign(FrameAlign); 1981 if (!IsMemcpySmall(FrameSize)) 1982 return false; 1983 } 1984 1985 if (CS.paramHasAttr(AttrInd, Attribute::InReg)) 1986 Flags.setInReg(); 1987 if (CS.paramHasAttr(AttrInd, Attribute::Nest)) 1988 Flags.setNest(); 1989 1990 // If this is an i1/i8/i16 argument, promote to i32 to avoid an extra 1991 // instruction. This is safe because it is common to all fastisel supported 1992 // calling conventions on x86. 1993 if (ConstantInt *CI = dyn_cast<ConstantInt>(ArgVal)) { 1994 if (CI->getBitWidth() == 1 || CI->getBitWidth() == 8 || 1995 CI->getBitWidth() == 16) { 1996 if (Flags.isSExt()) 1997 ArgVal = ConstantExpr::getSExt(CI,Type::getInt32Ty(CI->getContext())); 1998 else 1999 ArgVal = ConstantExpr::getZExt(CI,Type::getInt32Ty(CI->getContext())); 2000 } 2001 } 2002 2003 unsigned ArgReg; 2004 2005 // Passing bools around ends up doing a trunc to i1 and passing it. 2006 // Codegen this as an argument + "and 1". 2007 if (ArgVal->getType()->isIntegerTy(1) && isa<TruncInst>(ArgVal) && 2008 cast<TruncInst>(ArgVal)->getParent() == I->getParent() && 2009 ArgVal->hasOneUse()) { 2010 ArgVal = cast<TruncInst>(ArgVal)->getOperand(0); 2011 ArgReg = getRegForValue(ArgVal); 2012 if (ArgReg == 0) return false; 2013 2014 MVT ArgVT; 2015 if (!isTypeLegal(ArgVal->getType(), ArgVT)) return false; 2016 2017 ArgReg = FastEmit_ri(ArgVT, ArgVT, ISD::AND, ArgReg, 2018 ArgVal->hasOneUse(), 1); 2019 } else { 2020 ArgReg = getRegForValue(ArgVal); 2021 } 2022 2023 if (ArgReg == 0) return false; 2024 2025 Type *ArgTy = ArgVal->getType(); 2026 MVT ArgVT; 2027 if (!isTypeLegal(ArgTy, ArgVT)) 2028 return false; 2029 if (ArgVT == MVT::x86mmx) 2030 return false; 2031 unsigned OriginalAlignment = DL.getABITypeAlignment(ArgTy); 2032 Flags.setOrigAlign(OriginalAlignment); 2033 2034 Args.push_back(ArgReg); 2035 ArgVals.push_back(ArgVal); 2036 ArgVTs.push_back(ArgVT); 2037 ArgFlags.push_back(Flags); 2038 } 2039 2040 // Analyze operands of the call, assigning locations to each operand. 2041 SmallVector<CCValAssign, 16> ArgLocs; 2042 CCState CCInfo(CC, isVarArg, *FuncInfo.MF, TM, ArgLocs, 2043 I->getParent()->getContext()); 2044 2045 // Allocate shadow area for Win64 2046 if (isWin64) 2047 CCInfo.AllocateStack(32, 8); 2048 2049 CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CC_X86); 2050 2051 // Get a count of how many bytes are to be pushed on the stack. 2052 unsigned NumBytes = CCInfo.getNextStackOffset(); 2053 2054 // Issue CALLSEQ_START 2055 unsigned AdjStackDown = TII.getCallFrameSetupOpcode(); 2056 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackDown)) 2057 .addImm(NumBytes); 2058 2059 // Process argument: walk the register/memloc assignments, inserting 2060 // copies / loads. 2061 SmallVector<unsigned, 4> RegArgs; 2062 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 2063 CCValAssign &VA = ArgLocs[i]; 2064 unsigned Arg = Args[VA.getValNo()]; 2065 EVT ArgVT = ArgVTs[VA.getValNo()]; 2066 2067 // Promote the value if needed. 2068 switch (VA.getLocInfo()) { 2069 case CCValAssign::Full: break; 2070 case CCValAssign::SExt: { 2071 assert(VA.getLocVT().isInteger() && !VA.getLocVT().isVector() && 2072 "Unexpected extend"); 2073 bool Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(), 2074 Arg, ArgVT, Arg); 2075 assert(Emitted && "Failed to emit a sext!"); (void)Emitted; 2076 ArgVT = VA.getLocVT(); 2077 break; 2078 } 2079 case CCValAssign::ZExt: { 2080 assert(VA.getLocVT().isInteger() && !VA.getLocVT().isVector() && 2081 "Unexpected extend"); 2082 bool Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(), 2083 Arg, ArgVT, Arg); 2084 assert(Emitted && "Failed to emit a zext!"); (void)Emitted; 2085 ArgVT = VA.getLocVT(); 2086 break; 2087 } 2088 case CCValAssign::AExt: { 2089 assert(VA.getLocVT().isInteger() && !VA.getLocVT().isVector() && 2090 "Unexpected extend"); 2091 bool Emitted = X86FastEmitExtend(ISD::ANY_EXTEND, VA.getLocVT(), 2092 Arg, ArgVT, Arg); 2093 if (!Emitted) 2094 Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(), 2095 Arg, ArgVT, Arg); 2096 if (!Emitted) 2097 Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(), 2098 Arg, ArgVT, Arg); 2099 2100 assert(Emitted && "Failed to emit a aext!"); (void)Emitted; 2101 ArgVT = VA.getLocVT(); 2102 break; 2103 } 2104 case CCValAssign::BCvt: { 2105 unsigned BC = FastEmit_r(ArgVT.getSimpleVT(), VA.getLocVT(), 2106 ISD::BITCAST, Arg, /*TODO: Kill=*/false); 2107 assert(BC != 0 && "Failed to emit a bitcast!"); 2108 Arg = BC; 2109 ArgVT = VA.getLocVT(); 2110 break; 2111 } 2112 case CCValAssign::VExt: 2113 // VExt has not been implemented, so this should be impossible to reach 2114 // for now. However, fallback to Selection DAG isel once implemented. 2115 return false; 2116 case CCValAssign::Indirect: 2117 // FIXME: Indirect doesn't need extending, but fast-isel doesn't fully 2118 // support this. 2119 return false; 2120 case CCValAssign::FPExt: 2121 llvm_unreachable("Unexpected loc info!"); 2122 } 2123 2124 if (VA.isRegLoc()) { 2125 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, 2126 TII.get(TargetOpcode::COPY), VA.getLocReg()).addReg(Arg); 2127 RegArgs.push_back(VA.getLocReg()); 2128 } else { 2129 unsigned LocMemOffset = VA.getLocMemOffset(); 2130 X86AddressMode AM; 2131 const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo*>( 2132 getTargetMachine()->getRegisterInfo()); 2133 AM.Base.Reg = RegInfo->getStackRegister(); 2134 AM.Disp = LocMemOffset; 2135 const Value *ArgVal = ArgVals[VA.getValNo()]; 2136 ISD::ArgFlagsTy Flags = ArgFlags[VA.getValNo()]; 2137 2138 if (Flags.isByVal()) { 2139 X86AddressMode SrcAM; 2140 SrcAM.Base.Reg = Arg; 2141 bool Res = TryEmitSmallMemcpy(AM, SrcAM, Flags.getByValSize()); 2142 assert(Res && "memcpy length already checked!"); (void)Res; 2143 } else if (isa<ConstantInt>(ArgVal) || isa<ConstantPointerNull>(ArgVal)) { 2144 // If this is a really simple value, emit this with the Value* version 2145 // of X86FastEmitStore. If it isn't simple, we don't want to do this, 2146 // as it can cause us to reevaluate the argument. 2147 if (!X86FastEmitStore(ArgVT, ArgVal, AM)) 2148 return false; 2149 } else { 2150 if (!X86FastEmitStore(ArgVT, Arg, AM)) 2151 return false; 2152 } 2153 } 2154 } 2155 2156 // ELF / PIC requires GOT in the EBX register before function calls via PLT 2157 // GOT pointer. 2158 if (Subtarget->isPICStyleGOT()) { 2159 unsigned Base = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF); 2160 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, 2161 TII.get(TargetOpcode::COPY), X86::EBX).addReg(Base); 2162 } 2163 2164 if (Subtarget->is64Bit() && isVarArg && !isWin64) { 2165 // Count the number of XMM registers allocated. 2166 static const MCPhysReg XMMArgRegs[] = { 2167 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3, 2168 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7 2169 }; 2170 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8); 2171 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::MOV8ri), 2172 X86::AL).addImm(NumXMMRegs); 2173 } 2174 2175 // Issue the call. 2176 MachineInstrBuilder MIB; 2177 if (CalleeOp) { 2178 // Register-indirect call. 2179 unsigned CallOpc; 2180 if (Subtarget->is64Bit()) 2181 CallOpc = X86::CALL64r; 2182 else 2183 CallOpc = X86::CALL32r; 2184 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CallOpc)) 2185 .addReg(CalleeOp); 2186 2187 } else { 2188 // Direct call. 2189 assert(GV && "Not a direct call"); 2190 unsigned CallOpc; 2191 if (Subtarget->is64Bit()) 2192 CallOpc = X86::CALL64pcrel32; 2193 else 2194 CallOpc = X86::CALLpcrel32; 2195 2196 // See if we need any target-specific flags on the GV operand. 2197 unsigned char OpFlags = 0; 2198 2199 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to 2200 // external symbols most go through the PLT in PIC mode. If the symbol 2201 // has hidden or protected visibility, or if it is static or local, then 2202 // we don't need to use the PLT - we can directly call it. 2203 if (Subtarget->isTargetELF() && 2204 TM.getRelocationModel() == Reloc::PIC_ && 2205 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) { 2206 OpFlags = X86II::MO_PLT; 2207 } else if (Subtarget->isPICStyleStubAny() && 2208 (GV->isDeclaration() || GV->isWeakForLinker()) && 2209 (!Subtarget->getTargetTriple().isMacOSX() || 2210 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) { 2211 // PC-relative references to external symbols should go through $stub, 2212 // unless we're building with the leopard linker or later, which 2213 // automatically synthesizes these stubs. 2214 OpFlags = X86II::MO_DARWIN_STUB; 2215 } 2216 2217 2218 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CallOpc)); 2219 if (MemIntName) 2220 MIB.addExternalSymbol(MemIntName, OpFlags); 2221 else 2222 MIB.addGlobalAddress(GV, 0, OpFlags); 2223 } 2224 2225 // Add a register mask with the call-preserved registers. 2226 // Proper defs for return values will be added by setPhysRegsDeadExcept(). 2227 MIB.addRegMask(TRI.getCallPreservedMask(CS.getCallingConv())); 2228 2229 // Add an implicit use GOT pointer in EBX. 2230 if (Subtarget->isPICStyleGOT()) 2231 MIB.addReg(X86::EBX, RegState::Implicit); 2232 2233 if (Subtarget->is64Bit() && isVarArg && !isWin64) 2234 MIB.addReg(X86::AL, RegState::Implicit); 2235 2236 // Add implicit physical register uses to the call. 2237 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i) 2238 MIB.addReg(RegArgs[i], RegState::Implicit); 2239 2240 // Issue CALLSEQ_END 2241 unsigned AdjStackUp = TII.getCallFrameDestroyOpcode(); 2242 const unsigned NumBytesCallee = computeBytesPoppedByCallee(*Subtarget, CS); 2243 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackUp)) 2244 .addImm(NumBytes).addImm(NumBytesCallee); 2245 2246 // Build info for return calling conv lowering code. 2247 // FIXME: This is practically a copy-paste from TargetLowering::LowerCallTo. 2248 SmallVector<ISD::InputArg, 32> Ins; 2249 SmallVector<EVT, 4> RetTys; 2250 ComputeValueVTs(TLI, I->getType(), RetTys); 2251 for (unsigned i = 0, e = RetTys.size(); i != e; ++i) { 2252 EVT VT = RetTys[i]; 2253 MVT RegisterVT = TLI.getRegisterType(I->getParent()->getContext(), VT); 2254 unsigned NumRegs = TLI.getNumRegisters(I->getParent()->getContext(), VT); 2255 for (unsigned j = 0; j != NumRegs; ++j) { 2256 ISD::InputArg MyFlags; 2257 MyFlags.VT = RegisterVT; 2258 MyFlags.Used = !CS.getInstruction()->use_empty(); 2259 if (CS.paramHasAttr(0, Attribute::SExt)) 2260 MyFlags.Flags.setSExt(); 2261 if (CS.paramHasAttr(0, Attribute::ZExt)) 2262 MyFlags.Flags.setZExt(); 2263 if (CS.paramHasAttr(0, Attribute::InReg)) 2264 MyFlags.Flags.setInReg(); 2265 Ins.push_back(MyFlags); 2266 } 2267 } 2268 2269 // Now handle call return values. 2270 SmallVector<unsigned, 4> UsedRegs; 2271 SmallVector<CCValAssign, 16> RVLocs; 2272 CCState CCRetInfo(CC, false, *FuncInfo.MF, TM, RVLocs, 2273 I->getParent()->getContext()); 2274 unsigned ResultReg = FuncInfo.CreateRegs(I->getType()); 2275 CCRetInfo.AnalyzeCallResult(Ins, RetCC_X86); 2276 for (unsigned i = 0; i != RVLocs.size(); ++i) { 2277 EVT CopyVT = RVLocs[i].getValVT(); 2278 unsigned CopyReg = ResultReg + i; 2279 2280 // If this is a call to a function that returns an fp value on the x87 fp 2281 // stack, but where we prefer to use the value in xmm registers, copy it 2282 // out as F80 and use a truncate to move it from fp stack reg to xmm reg. 2283 if ((RVLocs[i].getLocReg() == X86::ST0 || 2284 RVLocs[i].getLocReg() == X86::ST1)) { 2285 if (isScalarFPTypeInSSEReg(RVLocs[i].getValVT())) { 2286 CopyVT = MVT::f80; 2287 CopyReg = createResultReg(&X86::RFP80RegClass); 2288 } 2289 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, 2290 TII.get(X86::FpPOP_RETVAL), CopyReg); 2291 } else { 2292 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, 2293 TII.get(TargetOpcode::COPY), 2294 CopyReg).addReg(RVLocs[i].getLocReg()); 2295 UsedRegs.push_back(RVLocs[i].getLocReg()); 2296 } 2297 2298 if (CopyVT != RVLocs[i].getValVT()) { 2299 // Round the F80 the right size, which also moves to the appropriate xmm 2300 // register. This is accomplished by storing the F80 value in memory and 2301 // then loading it back. Ewww... 2302 EVT ResVT = RVLocs[i].getValVT(); 2303 unsigned Opc = ResVT == MVT::f32 ? X86::ST_Fp80m32 : X86::ST_Fp80m64; 2304 unsigned MemSize = ResVT.getSizeInBits()/8; 2305 int FI = MFI.CreateStackObject(MemSize, MemSize, false); 2306 addFrameReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, 2307 TII.get(Opc)), FI) 2308 .addReg(CopyReg); 2309 Opc = ResVT == MVT::f32 ? X86::MOVSSrm : X86::MOVSDrm; 2310 addFrameReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, 2311 TII.get(Opc), ResultReg + i), FI); 2312 } 2313 } 2314 2315 if (RVLocs.size()) 2316 UpdateValueMap(I, ResultReg, RVLocs.size()); 2317 2318 // Set all unused physreg defs as dead. 2319 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI); 2320 2321 return true; 2322} 2323 2324 2325bool 2326X86FastISel::TargetSelectInstruction(const Instruction *I) { 2327 switch (I->getOpcode()) { 2328 default: break; 2329 case Instruction::Load: 2330 return X86SelectLoad(I); 2331 case Instruction::Store: 2332 return X86SelectStore(I); 2333 case Instruction::Ret: 2334 return X86SelectRet(I); 2335 case Instruction::ICmp: 2336 case Instruction::FCmp: 2337 return X86SelectCmp(I); 2338 case Instruction::ZExt: 2339 return X86SelectZExt(I); 2340 case Instruction::Br: 2341 return X86SelectBranch(I); 2342 case Instruction::Call: 2343 return X86SelectCall(I); 2344 case Instruction::LShr: 2345 case Instruction::AShr: 2346 case Instruction::Shl: 2347 return X86SelectShift(I); 2348 case Instruction::SDiv: 2349 case Instruction::UDiv: 2350 case Instruction::SRem: 2351 case Instruction::URem: 2352 return X86SelectDivRem(I); 2353 case Instruction::Select: 2354 return X86SelectSelect(I); 2355 case Instruction::Trunc: 2356 return X86SelectTrunc(I); 2357 case Instruction::FPExt: 2358 return X86SelectFPExt(I); 2359 case Instruction::FPTrunc: 2360 return X86SelectFPTrunc(I); 2361 case Instruction::IntToPtr: // Deliberate fall-through. 2362 case Instruction::PtrToInt: { 2363 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType()); 2364 EVT DstVT = TLI.getValueType(I->getType()); 2365 if (DstVT.bitsGT(SrcVT)) 2366 return X86SelectZExt(I); 2367 if (DstVT.bitsLT(SrcVT)) 2368 return X86SelectTrunc(I); 2369 unsigned Reg = getRegForValue(I->getOperand(0)); 2370 if (Reg == 0) return false; 2371 UpdateValueMap(I, Reg); 2372 return true; 2373 } 2374 } 2375 2376 return false; 2377} 2378 2379unsigned X86FastISel::TargetMaterializeConstant(const Constant *C) { 2380 MVT VT; 2381 if (!isTypeLegal(C->getType(), VT)) 2382 return 0; 2383 2384 // Can't handle alternate code models yet. 2385 if (TM.getCodeModel() != CodeModel::Small) 2386 return 0; 2387 2388 // Get opcode and regclass of the output for the given load instruction. 2389 unsigned Opc = 0; 2390 const TargetRegisterClass *RC = nullptr; 2391 switch (VT.SimpleTy) { 2392 default: return 0; 2393 case MVT::i8: 2394 Opc = X86::MOV8rm; 2395 RC = &X86::GR8RegClass; 2396 break; 2397 case MVT::i16: 2398 Opc = X86::MOV16rm; 2399 RC = &X86::GR16RegClass; 2400 break; 2401 case MVT::i32: 2402 Opc = X86::MOV32rm; 2403 RC = &X86::GR32RegClass; 2404 break; 2405 case MVT::i64: 2406 // Must be in x86-64 mode. 2407 Opc = X86::MOV64rm; 2408 RC = &X86::GR64RegClass; 2409 break; 2410 case MVT::f32: 2411 if (X86ScalarSSEf32) { 2412 Opc = Subtarget->hasAVX() ? X86::VMOVSSrm : X86::MOVSSrm; 2413 RC = &X86::FR32RegClass; 2414 } else { 2415 Opc = X86::LD_Fp32m; 2416 RC = &X86::RFP32RegClass; 2417 } 2418 break; 2419 case MVT::f64: 2420 if (X86ScalarSSEf64) { 2421 Opc = Subtarget->hasAVX() ? X86::VMOVSDrm : X86::MOVSDrm; 2422 RC = &X86::FR64RegClass; 2423 } else { 2424 Opc = X86::LD_Fp64m; 2425 RC = &X86::RFP64RegClass; 2426 } 2427 break; 2428 case MVT::f80: 2429 // No f80 support yet. 2430 return 0; 2431 } 2432 2433 // Materialize addresses with LEA instructions. 2434 if (isa<GlobalValue>(C)) { 2435 X86AddressMode AM; 2436 if (X86SelectAddress(C, AM)) { 2437 // If the expression is just a basereg, then we're done, otherwise we need 2438 // to emit an LEA. 2439 if (AM.BaseType == X86AddressMode::RegBase && 2440 AM.IndexReg == 0 && AM.Disp == 0 && AM.GV == nullptr) 2441 return AM.Base.Reg; 2442 2443 Opc = TLI.getPointerTy() == MVT::i32 ? X86::LEA32r : X86::LEA64r; 2444 unsigned ResultReg = createResultReg(RC); 2445 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, 2446 TII.get(Opc), ResultReg), AM); 2447 return ResultReg; 2448 } 2449 return 0; 2450 } 2451 2452 // MachineConstantPool wants an explicit alignment. 2453 unsigned Align = DL.getPrefTypeAlignment(C->getType()); 2454 if (Align == 0) { 2455 // Alignment of vector types. FIXME! 2456 Align = DL.getTypeAllocSize(C->getType()); 2457 } 2458 2459 // x86-32 PIC requires a PIC base register for constant pools. 2460 unsigned PICBase = 0; 2461 unsigned char OpFlag = 0; 2462 if (Subtarget->isPICStyleStubPIC()) { // Not dynamic-no-pic 2463 OpFlag = X86II::MO_PIC_BASE_OFFSET; 2464 PICBase = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF); 2465 } else if (Subtarget->isPICStyleGOT()) { 2466 OpFlag = X86II::MO_GOTOFF; 2467 PICBase = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF); 2468 } else if (Subtarget->isPICStyleRIPRel() && 2469 TM.getCodeModel() == CodeModel::Small) { 2470 PICBase = X86::RIP; 2471 } 2472 2473 // Create the load from the constant pool. 2474 unsigned MCPOffset = MCP.getConstantPoolIndex(C, Align); 2475 unsigned ResultReg = createResultReg(RC); 2476 addConstantPoolReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, 2477 TII.get(Opc), ResultReg), 2478 MCPOffset, PICBase, OpFlag); 2479 2480 return ResultReg; 2481} 2482 2483unsigned X86FastISel::TargetMaterializeAlloca(const AllocaInst *C) { 2484 // Fail on dynamic allocas. At this point, getRegForValue has already 2485 // checked its CSE maps, so if we're here trying to handle a dynamic 2486 // alloca, we're not going to succeed. X86SelectAddress has a 2487 // check for dynamic allocas, because it's called directly from 2488 // various places, but TargetMaterializeAlloca also needs a check 2489 // in order to avoid recursion between getRegForValue, 2490 // X86SelectAddrss, and TargetMaterializeAlloca. 2491 if (!FuncInfo.StaticAllocaMap.count(C)) 2492 return 0; 2493 assert(C->isStaticAlloca() && "dynamic alloca in the static alloca map?"); 2494 2495 X86AddressMode AM; 2496 if (!X86SelectAddress(C, AM)) 2497 return 0; 2498 unsigned Opc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r; 2499 const TargetRegisterClass* RC = TLI.getRegClassFor(TLI.getPointerTy()); 2500 unsigned ResultReg = createResultReg(RC); 2501 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, 2502 TII.get(Opc), ResultReg), AM); 2503 return ResultReg; 2504} 2505 2506unsigned X86FastISel::TargetMaterializeFloatZero(const ConstantFP *CF) { 2507 MVT VT; 2508 if (!isTypeLegal(CF->getType(), VT)) 2509 return 0; 2510 2511 // Get opcode and regclass for the given zero. 2512 unsigned Opc = 0; 2513 const TargetRegisterClass *RC = nullptr; 2514 switch (VT.SimpleTy) { 2515 default: return 0; 2516 case MVT::f32: 2517 if (X86ScalarSSEf32) { 2518 Opc = X86::FsFLD0SS; 2519 RC = &X86::FR32RegClass; 2520 } else { 2521 Opc = X86::LD_Fp032; 2522 RC = &X86::RFP32RegClass; 2523 } 2524 break; 2525 case MVT::f64: 2526 if (X86ScalarSSEf64) { 2527 Opc = X86::FsFLD0SD; 2528 RC = &X86::FR64RegClass; 2529 } else { 2530 Opc = X86::LD_Fp064; 2531 RC = &X86::RFP64RegClass; 2532 } 2533 break; 2534 case MVT::f80: 2535 // No f80 support yet. 2536 return 0; 2537 } 2538 2539 unsigned ResultReg = createResultReg(RC); 2540 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg); 2541 return ResultReg; 2542} 2543 2544 2545bool X86FastISel::tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo, 2546 const LoadInst *LI) { 2547 X86AddressMode AM; 2548 if (!X86SelectAddress(LI->getOperand(0), AM)) 2549 return false; 2550 2551 const X86InstrInfo &XII = (const X86InstrInfo&)TII; 2552 2553 unsigned Size = DL.getTypeAllocSize(LI->getType()); 2554 unsigned Alignment = LI->getAlignment(); 2555 2556 SmallVector<MachineOperand, 8> AddrOps; 2557 AM.getFullAddress(AddrOps); 2558 2559 MachineInstr *Result = 2560 XII.foldMemoryOperandImpl(*FuncInfo.MF, MI, OpNo, AddrOps, Size, Alignment); 2561 if (!Result) return false; 2562 2563 FuncInfo.MBB->insert(FuncInfo.InsertPt, Result); 2564 MI->eraseFromParent(); 2565 return true; 2566} 2567 2568 2569namespace llvm { 2570 FastISel *X86::createFastISel(FunctionLoweringInfo &funcInfo, 2571 const TargetLibraryInfo *libInfo) { 2572 return new X86FastISel(funcInfo, libInfo); 2573 } 2574} 2575