X86ISelLowering.h revision 210299
1//===-- X86ISelLowering.h - X86 DAG Lowering Interface ----------*- C++ -*-===// 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 interfaces that X86 uses to lower LLVM code into a 11// selection DAG. 12// 13//===----------------------------------------------------------------------===// 14 15#ifndef X86ISELLOWERING_H 16#define X86ISELLOWERING_H 17 18#include "X86Subtarget.h" 19#include "X86RegisterInfo.h" 20#include "X86MachineFunctionInfo.h" 21#include "llvm/Target/TargetLowering.h" 22#include "llvm/Target/TargetOptions.h" 23#include "llvm/CodeGen/FastISel.h" 24#include "llvm/CodeGen/SelectionDAG.h" 25#include "llvm/CodeGen/CallingConvLower.h" 26 27namespace llvm { 28 namespace X86ISD { 29 // X86 Specific DAG Nodes 30 enum NodeType { 31 // Start the numbering where the builtin ops leave off. 32 FIRST_NUMBER = ISD::BUILTIN_OP_END, 33 34 /// BSF - Bit scan forward. 35 /// BSR - Bit scan reverse. 36 BSF, 37 BSR, 38 39 /// SHLD, SHRD - Double shift instructions. These correspond to 40 /// X86::SHLDxx and X86::SHRDxx instructions. 41 SHLD, 42 SHRD, 43 44 /// FAND - Bitwise logical AND of floating point values. This corresponds 45 /// to X86::ANDPS or X86::ANDPD. 46 FAND, 47 48 /// FOR - Bitwise logical OR of floating point values. This corresponds 49 /// to X86::ORPS or X86::ORPD. 50 FOR, 51 52 /// FXOR - Bitwise logical XOR of floating point values. This corresponds 53 /// to X86::XORPS or X86::XORPD. 54 FXOR, 55 56 /// FSRL - Bitwise logical right shift of floating point values. These 57 /// corresponds to X86::PSRLDQ. 58 FSRL, 59 60 /// FILD, FILD_FLAG - This instruction implements SINT_TO_FP with the 61 /// integer source in memory and FP reg result. This corresponds to the 62 /// X86::FILD*m instructions. It has three inputs (token chain, address, 63 /// and source type) and two outputs (FP value and token chain). FILD_FLAG 64 /// also produces a flag). 65 FILD, 66 FILD_FLAG, 67 68 /// FP_TO_INT*_IN_MEM - This instruction implements FP_TO_SINT with the 69 /// integer destination in memory and a FP reg source. This corresponds 70 /// to the X86::FIST*m instructions and the rounding mode change stuff. It 71 /// has two inputs (token chain and address) and two outputs (int value 72 /// and token chain). 73 FP_TO_INT16_IN_MEM, 74 FP_TO_INT32_IN_MEM, 75 FP_TO_INT64_IN_MEM, 76 77 /// FLD - This instruction implements an extending load to FP stack slots. 78 /// This corresponds to the X86::FLD32m / X86::FLD64m. It takes a chain 79 /// operand, ptr to load from, and a ValueType node indicating the type 80 /// to load to. 81 FLD, 82 83 /// FST - This instruction implements a truncating store to FP stack 84 /// slots. This corresponds to the X86::FST32m / X86::FST64m. It takes a 85 /// chain operand, value to store, address, and a ValueType to store it 86 /// as. 87 FST, 88 89 /// CALL - These operations represent an abstract X86 call 90 /// instruction, which includes a bunch of information. In particular the 91 /// operands of these node are: 92 /// 93 /// #0 - The incoming token chain 94 /// #1 - The callee 95 /// #2 - The number of arg bytes the caller pushes on the stack. 96 /// #3 - The number of arg bytes the callee pops off the stack. 97 /// #4 - The value to pass in AL/AX/EAX (optional) 98 /// #5 - The value to pass in DL/DX/EDX (optional) 99 /// 100 /// The result values of these nodes are: 101 /// 102 /// #0 - The outgoing token chain 103 /// #1 - The first register result value (optional) 104 /// #2 - The second register result value (optional) 105 /// 106 CALL, 107 108 /// RDTSC_DAG - This operation implements the lowering for 109 /// readcyclecounter 110 RDTSC_DAG, 111 112 /// X86 compare and logical compare instructions. 113 CMP, COMI, UCOMI, 114 115 /// X86 bit-test instructions. 116 BT, 117 118 /// X86 SetCC. Operand 0 is condition code, and operand 1 is the flag 119 /// operand produced by a CMP instruction. 120 SETCC, 121 122 // Same as SETCC except it's materialized with a sbb and the value is all 123 // one's or all zero's. 124 SETCC_CARRY, 125 126 /// X86 conditional moves. Operand 0 and operand 1 are the two values 127 /// to select from. Operand 2 is the condition code, and operand 3 is the 128 /// flag operand produced by a CMP or TEST instruction. It also writes a 129 /// flag result. 130 CMOV, 131 132 /// X86 conditional branches. Operand 0 is the chain operand, operand 1 133 /// is the block to branch if condition is true, operand 2 is the 134 /// condition code, and operand 3 is the flag operand produced by a CMP 135 /// or TEST instruction. 136 BRCOND, 137 138 /// Return with a flag operand. Operand 0 is the chain operand, operand 139 /// 1 is the number of bytes of stack to pop. 140 RET_FLAG, 141 142 /// REP_STOS - Repeat fill, corresponds to X86::REP_STOSx. 143 REP_STOS, 144 145 /// REP_MOVS - Repeat move, corresponds to X86::REP_MOVSx. 146 REP_MOVS, 147 148 /// GlobalBaseReg - On Darwin, this node represents the result of the popl 149 /// at function entry, used for PIC code. 150 GlobalBaseReg, 151 152 /// Wrapper - A wrapper node for TargetConstantPool, 153 /// TargetExternalSymbol, and TargetGlobalAddress. 154 Wrapper, 155 156 /// WrapperRIP - Special wrapper used under X86-64 PIC mode for RIP 157 /// relative displacements. 158 WrapperRIP, 159 160 /// MOVQ2DQ - Copies a 64-bit value from a vector to another vector. 161 /// Can be used to move a vector value from a MMX register to a XMM 162 /// register. 163 MOVQ2DQ, 164 165 /// PEXTRB - Extract an 8-bit value from a vector and zero extend it to 166 /// i32, corresponds to X86::PEXTRB. 167 PEXTRB, 168 169 /// PEXTRW - Extract a 16-bit value from a vector and zero extend it to 170 /// i32, corresponds to X86::PEXTRW. 171 PEXTRW, 172 173 /// INSERTPS - Insert any element of a 4 x float vector into any element 174 /// of a destination 4 x floatvector. 175 INSERTPS, 176 177 /// PINSRB - Insert the lower 8-bits of a 32-bit value to a vector, 178 /// corresponds to X86::PINSRB. 179 PINSRB, 180 181 /// PINSRW - Insert the lower 16-bits of a 32-bit value to a vector, 182 /// corresponds to X86::PINSRW. 183 PINSRW, MMX_PINSRW, 184 185 /// PSHUFB - Shuffle 16 8-bit values within a vector. 186 PSHUFB, 187 188 /// FMAX, FMIN - Floating point max and min. 189 /// 190 FMAX, FMIN, 191 192 /// FRSQRT, FRCP - Floating point reciprocal-sqrt and reciprocal 193 /// approximation. Note that these typically require refinement 194 /// in order to obtain suitable precision. 195 FRSQRT, FRCP, 196 197 // TLSADDR - Thread Local Storage. 198 TLSADDR, 199 200 // TLSCALL - Thread Local Storage. When calling to an OS provided 201 // thunk at the address from an earlier relocation. 202 TLSCALL, 203 204 // SegmentBaseAddress - The address segment:0 205 SegmentBaseAddress, 206 207 // EH_RETURN - Exception Handling helpers. 208 EH_RETURN, 209 210 /// TC_RETURN - Tail call return. 211 /// operand #0 chain 212 /// operand #1 callee (register or absolute) 213 /// operand #2 stack adjustment 214 /// operand #3 optional in flag 215 TC_RETURN, 216 217 // LCMPXCHG_DAG, LCMPXCHG8_DAG - Compare and swap. 218 LCMPXCHG_DAG, 219 LCMPXCHG8_DAG, 220 221 // FNSTCW16m - Store FP control world into i16 memory. 222 FNSTCW16m, 223 224 // VZEXT_MOVL - Vector move low and zero extend. 225 VZEXT_MOVL, 226 227 // VZEXT_LOAD - Load, scalar_to_vector, and zero extend. 228 VZEXT_LOAD, 229 230 // VSHL, VSRL - Vector logical left / right shift. 231 VSHL, VSRL, 232 233 // CMPPD, CMPPS - Vector double/float comparison. 234 // CMPPD, CMPPS - Vector double/float comparison. 235 CMPPD, CMPPS, 236 237 // PCMP* - Vector integer comparisons. 238 PCMPEQB, PCMPEQW, PCMPEQD, PCMPEQQ, 239 PCMPGTB, PCMPGTW, PCMPGTD, PCMPGTQ, 240 241 // ADD, SUB, SMUL, UMUL, etc. - Arithmetic operations with FLAGS results. 242 ADD, SUB, SMUL, UMUL, 243 INC, DEC, OR, XOR, AND, 244 245 // MUL_IMM - X86 specific multiply by immediate. 246 MUL_IMM, 247 248 // PTEST - Vector bitwise comparisons 249 PTEST, 250 251 // VASTART_SAVE_XMM_REGS - Save xmm argument registers to the stack, 252 // according to %al. An operator is needed so that this can be expanded 253 // with control flow. 254 VASTART_SAVE_XMM_REGS, 255 256 // MINGW_ALLOCA - MingW's __alloca call to do stack probing. 257 MINGW_ALLOCA, 258 259 // ATOMADD64_DAG, ATOMSUB64_DAG, ATOMOR64_DAG, ATOMAND64_DAG, 260 // ATOMXOR64_DAG, ATOMNAND64_DAG, ATOMSWAP64_DAG - 261 // Atomic 64-bit binary operations. 262 ATOMADD64_DAG = ISD::FIRST_TARGET_MEMORY_OPCODE, 263 ATOMSUB64_DAG, 264 ATOMOR64_DAG, 265 ATOMXOR64_DAG, 266 ATOMAND64_DAG, 267 ATOMNAND64_DAG, 268 ATOMSWAP64_DAG 269 270 // WARNING: Do not add anything in the end unless you want the node to 271 // have memop! In fact, starting from ATOMADD64_DAG all opcodes will be 272 // thought as target memory ops! 273 }; 274 } 275 276 /// Define some predicates that are used for node matching. 277 namespace X86 { 278 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand 279 /// specifies a shuffle of elements that is suitable for input to PSHUFD. 280 bool isPSHUFDMask(ShuffleVectorSDNode *N); 281 282 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand 283 /// specifies a shuffle of elements that is suitable for input to PSHUFD. 284 bool isPSHUFHWMask(ShuffleVectorSDNode *N); 285 286 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand 287 /// specifies a shuffle of elements that is suitable for input to PSHUFD. 288 bool isPSHUFLWMask(ShuffleVectorSDNode *N); 289 290 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand 291 /// specifies a shuffle of elements that is suitable for input to SHUFP*. 292 bool isSHUFPMask(ShuffleVectorSDNode *N); 293 294 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand 295 /// specifies a shuffle of elements that is suitable for input to MOVHLPS. 296 bool isMOVHLPSMask(ShuffleVectorSDNode *N); 297 298 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form 299 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef, 300 /// <2, 3, 2, 3> 301 bool isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N); 302 303 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand 304 /// specifies a shuffle of elements that is suitable for MOVLP{S|D}. 305 bool isMOVLPMask(ShuffleVectorSDNode *N); 306 307 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand 308 /// specifies a shuffle of elements that is suitable for MOVHP{S|D}. 309 /// as well as MOVLHPS. 310 bool isMOVLHPSMask(ShuffleVectorSDNode *N); 311 312 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand 313 /// specifies a shuffle of elements that is suitable for input to UNPCKL. 314 bool isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat = false); 315 316 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand 317 /// specifies a shuffle of elements that is suitable for input to UNPCKH. 318 bool isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat = false); 319 320 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form 321 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef, 322 /// <0, 0, 1, 1> 323 bool isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N); 324 325 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form 326 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef, 327 /// <2, 2, 3, 3> 328 bool isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N); 329 330 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand 331 /// specifies a shuffle of elements that is suitable for input to MOVSS, 332 /// MOVSD, and MOVD, i.e. setting the lowest element. 333 bool isMOVLMask(ShuffleVectorSDNode *N); 334 335 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand 336 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP. 337 bool isMOVSHDUPMask(ShuffleVectorSDNode *N); 338 339 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand 340 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP. 341 bool isMOVSLDUPMask(ShuffleVectorSDNode *N); 342 343 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand 344 /// specifies a shuffle of elements that is suitable for input to MOVDDUP. 345 bool isMOVDDUPMask(ShuffleVectorSDNode *N); 346 347 /// isPALIGNRMask - Return true if the specified VECTOR_SHUFFLE operand 348 /// specifies a shuffle of elements that is suitable for input to PALIGNR. 349 bool isPALIGNRMask(ShuffleVectorSDNode *N); 350 351 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle 352 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP* 353 /// instructions. 354 unsigned getShuffleSHUFImmediate(SDNode *N); 355 356 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle 357 /// the specified VECTOR_SHUFFLE mask with PSHUFHW instruction. 358 unsigned getShufflePSHUFHWImmediate(SDNode *N); 359 360 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle 361 /// the specified VECTOR_SHUFFLE mask with PSHUFLW instruction. 362 unsigned getShufflePSHUFLWImmediate(SDNode *N); 363 364 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle 365 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction. 366 unsigned getShufflePALIGNRImmediate(SDNode *N); 367 368 /// isZeroNode - Returns true if Elt is a constant zero or a floating point 369 /// constant +0.0. 370 bool isZeroNode(SDValue Elt); 371 372 /// isOffsetSuitableForCodeModel - Returns true of the given offset can be 373 /// fit into displacement field of the instruction. 374 bool isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M, 375 bool hasSymbolicDisplacement = true); 376 } 377 378 //===--------------------------------------------------------------------===// 379 // X86TargetLowering - X86 Implementation of the TargetLowering interface 380 class X86TargetLowering : public TargetLowering { 381 public: 382 explicit X86TargetLowering(X86TargetMachine &TM); 383 384 /// getPICBaseSymbol - Return the X86-32 PIC base. 385 MCSymbol *getPICBaseSymbol(const MachineFunction *MF, MCContext &Ctx) const; 386 387 virtual unsigned getJumpTableEncoding() const; 388 389 virtual const MCExpr * 390 LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI, 391 const MachineBasicBlock *MBB, unsigned uid, 392 MCContext &Ctx) const; 393 394 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC 395 /// jumptable. 396 virtual SDValue getPICJumpTableRelocBase(SDValue Table, 397 SelectionDAG &DAG) const; 398 virtual const MCExpr * 399 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, 400 unsigned JTI, MCContext &Ctx) const; 401 402 /// getStackPtrReg - Return the stack pointer register we are using: either 403 /// ESP or RSP. 404 unsigned getStackPtrReg() const { return X86StackPtr; } 405 406 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate 407 /// function arguments in the caller parameter area. For X86, aggregates 408 /// that contains are placed at 16-byte boundaries while the rest are at 409 /// 4-byte boundaries. 410 virtual unsigned getByValTypeAlignment(const Type *Ty) const; 411 412 /// getOptimalMemOpType - Returns the target specific optimal type for load 413 /// and store operations as a result of memset, memcpy, and memmove 414 /// lowering. If DstAlign is zero that means it's safe to destination 415 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it 416 /// means there isn't a need to check it against alignment requirement, 417 /// probably because the source does not need to be loaded. If 418 /// 'NonScalarIntSafe' is true, that means it's safe to return a 419 /// non-scalar-integer type, e.g. empty string source, constant, or loaded 420 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is 421 /// constant so it does not need to be loaded. 422 /// It returns EVT::Other if the type should be determined using generic 423 /// target-independent logic. 424 virtual EVT 425 getOptimalMemOpType(uint64_t Size, unsigned DstAlign, unsigned SrcAlign, 426 bool NonScalarIntSafe, bool MemcpyStrSrc, 427 MachineFunction &MF) const; 428 429 /// allowsUnalignedMemoryAccesses - Returns true if the target allows 430 /// unaligned memory accesses. of the specified type. 431 virtual bool allowsUnalignedMemoryAccesses(EVT VT) const { 432 return true; 433 } 434 435 /// LowerOperation - Provide custom lowering hooks for some operations. 436 /// 437 virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const; 438 439 /// ReplaceNodeResults - Replace the results of node with an illegal result 440 /// type with new values built out of custom code. 441 /// 442 virtual void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results, 443 SelectionDAG &DAG) const; 444 445 446 virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const; 447 448 /// isTypeDesirableForOp - Return true if the target has native support for 449 /// the specified value type and it is 'desirable' to use the type for the 450 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16 451 /// instruction encodings are longer and some i16 instructions are slow. 452 virtual bool isTypeDesirableForOp(unsigned Opc, EVT VT) const; 453 454 /// isTypeDesirable - Return true if the target has native support for the 455 /// specified value type and it is 'desirable' to use the type. e.g. On x86 456 /// i16 is legal, but undesirable since i16 instruction encodings are longer 457 /// and some i16 instructions are slow. 458 virtual bool IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const; 459 460 virtual MachineBasicBlock * 461 EmitInstrWithCustomInserter(MachineInstr *MI, 462 MachineBasicBlock *MBB) const; 463 464 465 /// getTargetNodeName - This method returns the name of a target specific 466 /// DAG node. 467 virtual const char *getTargetNodeName(unsigned Opcode) const; 468 469 /// getSetCCResultType - Return the ISD::SETCC ValueType 470 virtual MVT::SimpleValueType getSetCCResultType(EVT VT) const; 471 472 /// computeMaskedBitsForTargetNode - Determine which of the bits specified 473 /// in Mask are known to be either zero or one and return them in the 474 /// KnownZero/KnownOne bitsets. 475 virtual void computeMaskedBitsForTargetNode(const SDValue Op, 476 const APInt &Mask, 477 APInt &KnownZero, 478 APInt &KnownOne, 479 const SelectionDAG &DAG, 480 unsigned Depth = 0) const; 481 482 virtual bool 483 isGAPlusOffset(SDNode *N, const GlobalValue* &GA, int64_t &Offset) const; 484 485 SDValue getReturnAddressFrameIndex(SelectionDAG &DAG) const; 486 487 virtual bool ExpandInlineAsm(CallInst *CI) const; 488 489 ConstraintType getConstraintType(const std::string &Constraint) const; 490 491 std::vector<unsigned> 492 getRegClassForInlineAsmConstraint(const std::string &Constraint, 493 EVT VT) const; 494 495 virtual const char *LowerXConstraint(EVT ConstraintVT) const; 496 497 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops 498 /// vector. If it is invalid, don't add anything to Ops. If hasMemory is 499 /// true it means one of the asm constraint of the inline asm instruction 500 /// being processed is 'm'. 501 virtual void LowerAsmOperandForConstraint(SDValue Op, 502 char ConstraintLetter, 503 std::vector<SDValue> &Ops, 504 SelectionDAG &DAG) const; 505 506 /// getRegForInlineAsmConstraint - Given a physical register constraint 507 /// (e.g. {edx}), return the register number and the register class for the 508 /// register. This should only be used for C_Register constraints. On 509 /// error, this returns a register number of 0. 510 std::pair<unsigned, const TargetRegisterClass*> 511 getRegForInlineAsmConstraint(const std::string &Constraint, 512 EVT VT) const; 513 514 /// isLegalAddressingMode - Return true if the addressing mode represented 515 /// by AM is legal for this target, for a load/store of the specified type. 516 virtual bool isLegalAddressingMode(const AddrMode &AM, const Type *Ty)const; 517 518 /// isTruncateFree - Return true if it's free to truncate a value of 519 /// type Ty1 to type Ty2. e.g. On x86 it's free to truncate a i32 value in 520 /// register EAX to i16 by referencing its sub-register AX. 521 virtual bool isTruncateFree(const Type *Ty1, const Type *Ty2) const; 522 virtual bool isTruncateFree(EVT VT1, EVT VT2) const; 523 524 /// isZExtFree - Return true if any actual instruction that defines a 525 /// value of type Ty1 implicit zero-extends the value to Ty2 in the result 526 /// register. This does not necessarily include registers defined in 527 /// unknown ways, such as incoming arguments, or copies from unknown 528 /// virtual registers. Also, if isTruncateFree(Ty2, Ty1) is true, this 529 /// does not necessarily apply to truncate instructions. e.g. on x86-64, 530 /// all instructions that define 32-bit values implicit zero-extend the 531 /// result out to 64 bits. 532 virtual bool isZExtFree(const Type *Ty1, const Type *Ty2) const; 533 virtual bool isZExtFree(EVT VT1, EVT VT2) const; 534 535 /// isNarrowingProfitable - Return true if it's profitable to narrow 536 /// operations of type VT1 to VT2. e.g. on x86, it's profitable to narrow 537 /// from i32 to i8 but not from i32 to i16. 538 virtual bool isNarrowingProfitable(EVT VT1, EVT VT2) const; 539 540 /// isFPImmLegal - Returns true if the target can instruction select the 541 /// specified FP immediate natively. If false, the legalizer will 542 /// materialize the FP immediate as a load from a constant pool. 543 virtual bool isFPImmLegal(const APFloat &Imm, EVT VT) const; 544 545 /// isShuffleMaskLegal - Targets can use this to indicate that they only 546 /// support *some* VECTOR_SHUFFLE operations, those with specific masks. 547 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask 548 /// values are assumed to be legal. 549 virtual bool isShuffleMaskLegal(const SmallVectorImpl<int> &Mask, 550 EVT VT) const; 551 552 /// isVectorClearMaskLegal - Similar to isShuffleMaskLegal. This is 553 /// used by Targets can use this to indicate if there is a suitable 554 /// VECTOR_SHUFFLE that can be used to replace a VAND with a constant 555 /// pool entry. 556 virtual bool isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask, 557 EVT VT) const; 558 559 /// ShouldShrinkFPConstant - If true, then instruction selection should 560 /// seek to shrink the FP constant of the specified type to a smaller type 561 /// in order to save space and / or reduce runtime. 562 virtual bool ShouldShrinkFPConstant(EVT VT) const { 563 // Don't shrink FP constpool if SSE2 is available since cvtss2sd is more 564 // expensive than a straight movsd. On the other hand, it's important to 565 // shrink long double fp constant since fldt is very slow. 566 return !X86ScalarSSEf64 || VT == MVT::f80; 567 } 568 569 const X86Subtarget* getSubtarget() const { 570 return Subtarget; 571 } 572 573 /// isScalarFPTypeInSSEReg - Return true if the specified scalar FP type is 574 /// computed in an SSE register, not on the X87 floating point stack. 575 bool isScalarFPTypeInSSEReg(EVT VT) const { 576 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2 577 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1 578 } 579 580 /// createFastISel - This method returns a target specific FastISel object, 581 /// or null if the target does not support "fast" ISel. 582 virtual FastISel *createFastISel(FunctionLoweringInfo &funcInfo) const; 583 584 /// getFunctionAlignment - Return the Log2 alignment of this function. 585 virtual unsigned getFunctionAlignment(const Function *F) const; 586 587 /// getStackCookieLocation - Return true if the target stores stack 588 /// protector cookies at a fixed offset in some non-standard address 589 /// space, and populates the address space and offset as 590 /// appropriate. 591 virtual bool getStackCookieLocation(unsigned &AddressSpace, unsigned &Offset) const; 592 593 private: 594 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can 595 /// make the right decision when generating code for different targets. 596 const X86Subtarget *Subtarget; 597 const X86RegisterInfo *RegInfo; 598 const TargetData *TD; 599 600 /// X86StackPtr - X86 physical register used as stack ptr. 601 unsigned X86StackPtr; 602 603 /// X86ScalarSSEf32, X86ScalarSSEf64 - Select between SSE or x87 604 /// floating point ops. 605 /// When SSE is available, use it for f32 operations. 606 /// When SSE2 is available, use it for f64 operations. 607 bool X86ScalarSSEf32; 608 bool X86ScalarSSEf64; 609 610 /// LegalFPImmediates - A list of legal fp immediates. 611 std::vector<APFloat> LegalFPImmediates; 612 613 /// addLegalFPImmediate - Indicate that this x86 target can instruction 614 /// select the specified FP immediate natively. 615 void addLegalFPImmediate(const APFloat& Imm) { 616 LegalFPImmediates.push_back(Imm); 617 } 618 619 SDValue LowerCallResult(SDValue Chain, SDValue InFlag, 620 CallingConv::ID CallConv, bool isVarArg, 621 const SmallVectorImpl<ISD::InputArg> &Ins, 622 DebugLoc dl, SelectionDAG &DAG, 623 SmallVectorImpl<SDValue> &InVals) const; 624 SDValue LowerMemArgument(SDValue Chain, 625 CallingConv::ID CallConv, 626 const SmallVectorImpl<ISD::InputArg> &ArgInfo, 627 DebugLoc dl, SelectionDAG &DAG, 628 const CCValAssign &VA, MachineFrameInfo *MFI, 629 unsigned i) const; 630 SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg, 631 DebugLoc dl, SelectionDAG &DAG, 632 const CCValAssign &VA, 633 ISD::ArgFlagsTy Flags) const; 634 635 // Call lowering helpers. 636 637 /// IsEligibleForTailCallOptimization - Check whether the call is eligible 638 /// for tail call optimization. Targets which want to do tail call 639 /// optimization should implement this function. 640 bool IsEligibleForTailCallOptimization(SDValue Callee, 641 CallingConv::ID CalleeCC, 642 bool isVarArg, 643 bool isCalleeStructRet, 644 bool isCallerStructRet, 645 const SmallVectorImpl<ISD::OutputArg> &Outs, 646 const SmallVectorImpl<SDValue> &OutVals, 647 const SmallVectorImpl<ISD::InputArg> &Ins, 648 SelectionDAG& DAG) const; 649 bool IsCalleePop(bool isVarArg, CallingConv::ID CallConv) const; 650 SDValue EmitTailCallLoadRetAddr(SelectionDAG &DAG, SDValue &OutRetAddr, 651 SDValue Chain, bool IsTailCall, bool Is64Bit, 652 int FPDiff, DebugLoc dl) const; 653 654 CCAssignFn *CCAssignFnForNode(CallingConv::ID CallConv) const; 655 unsigned GetAlignedArgumentStackSize(unsigned StackSize, 656 SelectionDAG &DAG) const; 657 658 std::pair<SDValue,SDValue> FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, 659 bool isSigned) const; 660 661 SDValue LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl, 662 SelectionDAG &DAG) const; 663 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const; 664 SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const; 665 SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const; 666 SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const; 667 SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) const; 668 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const; 669 SDValue LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) const; 670 SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const; 671 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const; 672 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const; 673 SDValue LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl, 674 int64_t Offset, SelectionDAG &DAG) const; 675 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const; 676 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const; 677 SDValue LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const; 678 SDValue LowerShift(SDValue Op, SelectionDAG &DAG) const; 679 SDValue BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain, SDValue StackSlot, 680 SelectionDAG &DAG) const; 681 SDValue LowerBIT_CONVERT(SDValue op, SelectionDAG &DAG) const; 682 SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) const; 683 SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const; 684 SDValue LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) const; 685 SDValue LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) const; 686 SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) const; 687 SDValue LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) const; 688 SDValue LowerFABS(SDValue Op, SelectionDAG &DAG) const; 689 SDValue LowerFNEG(SDValue Op, SelectionDAG &DAG) const; 690 SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const; 691 SDValue LowerToBT(SDValue And, ISD::CondCode CC, 692 DebugLoc dl, SelectionDAG &DAG) const; 693 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const; 694 SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) const; 695 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const; 696 SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const; 697 SDValue LowerMEMSET(SDValue Op, SelectionDAG &DAG) const; 698 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const; 699 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const; 700 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const; 701 SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const; 702 SDValue LowerVACOPY(SDValue Op, SelectionDAG &DAG) const; 703 SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const; 704 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const; 705 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const; 706 SDValue LowerFRAME_TO_ARGS_OFFSET(SDValue Op, SelectionDAG &DAG) const; 707 SDValue LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const; 708 SDValue LowerTRAMPOLINE(SDValue Op, SelectionDAG &DAG) const; 709 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const; 710 SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG) const; 711 SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) const; 712 SDValue LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const; 713 SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) const; 714 715 SDValue LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const; 716 SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const; 717 SDValue LowerREADCYCLECOUNTER(SDValue Op, SelectionDAG &DAG) const; 718 719 virtual SDValue 720 LowerFormalArguments(SDValue Chain, 721 CallingConv::ID CallConv, bool isVarArg, 722 const SmallVectorImpl<ISD::InputArg> &Ins, 723 DebugLoc dl, SelectionDAG &DAG, 724 SmallVectorImpl<SDValue> &InVals) const; 725 virtual SDValue 726 LowerCall(SDValue Chain, SDValue Callee, 727 CallingConv::ID CallConv, bool isVarArg, bool &isTailCall, 728 const SmallVectorImpl<ISD::OutputArg> &Outs, 729 const SmallVectorImpl<SDValue> &OutVals, 730 const SmallVectorImpl<ISD::InputArg> &Ins, 731 DebugLoc dl, SelectionDAG &DAG, 732 SmallVectorImpl<SDValue> &InVals) const; 733 734 virtual SDValue 735 LowerReturn(SDValue Chain, 736 CallingConv::ID CallConv, bool isVarArg, 737 const SmallVectorImpl<ISD::OutputArg> &Outs, 738 const SmallVectorImpl<SDValue> &OutVals, 739 DebugLoc dl, SelectionDAG &DAG) const; 740 741 virtual bool 742 CanLowerReturn(CallingConv::ID CallConv, bool isVarArg, 743 const SmallVectorImpl<ISD::OutputArg> &Outs, 744 LLVMContext &Context) const; 745 746 void ReplaceATOMIC_BINARY_64(SDNode *N, SmallVectorImpl<SDValue> &Results, 747 SelectionDAG &DAG, unsigned NewOp) const; 748 749 /// Utility function to emit string processing sse4.2 instructions 750 /// that return in xmm0. 751 /// This takes the instruction to expand, the associated machine basic 752 /// block, the number of args, and whether or not the second arg is 753 /// in memory or not. 754 MachineBasicBlock *EmitPCMP(MachineInstr *BInstr, MachineBasicBlock *BB, 755 unsigned argNum, bool inMem) const; 756 757 /// Utility function to emit atomic bitwise operations (and, or, xor). 758 /// It takes the bitwise instruction to expand, the associated machine basic 759 /// block, and the associated X86 opcodes for reg/reg and reg/imm. 760 MachineBasicBlock *EmitAtomicBitwiseWithCustomInserter( 761 MachineInstr *BInstr, 762 MachineBasicBlock *BB, 763 unsigned regOpc, 764 unsigned immOpc, 765 unsigned loadOpc, 766 unsigned cxchgOpc, 767 unsigned notOpc, 768 unsigned EAXreg, 769 TargetRegisterClass *RC, 770 bool invSrc = false) const; 771 772 MachineBasicBlock *EmitAtomicBit6432WithCustomInserter( 773 MachineInstr *BInstr, 774 MachineBasicBlock *BB, 775 unsigned regOpcL, 776 unsigned regOpcH, 777 unsigned immOpcL, 778 unsigned immOpcH, 779 bool invSrc = false) const; 780 781 /// Utility function to emit atomic min and max. It takes the min/max 782 /// instruction to expand, the associated basic block, and the associated 783 /// cmov opcode for moving the min or max value. 784 MachineBasicBlock *EmitAtomicMinMaxWithCustomInserter(MachineInstr *BInstr, 785 MachineBasicBlock *BB, 786 unsigned cmovOpc) const; 787 788 /// Utility function to emit the xmm reg save portion of va_start. 789 MachineBasicBlock *EmitVAStartSaveXMMRegsWithCustomInserter( 790 MachineInstr *BInstr, 791 MachineBasicBlock *BB) const; 792 793 MachineBasicBlock *EmitLoweredSelect(MachineInstr *I, 794 MachineBasicBlock *BB) const; 795 796 MachineBasicBlock *EmitLoweredMingwAlloca(MachineInstr *MI, 797 MachineBasicBlock *BB) const; 798 799 MachineBasicBlock *EmitLoweredTLSCall(MachineInstr *MI, 800 MachineBasicBlock *BB) const; 801 802 /// Emit nodes that will be selected as "test Op0,Op0", or something 803 /// equivalent, for use with the given x86 condition code. 804 SDValue EmitTest(SDValue Op0, unsigned X86CC, SelectionDAG &DAG) const; 805 806 /// Emit nodes that will be selected as "cmp Op0,Op1", or something 807 /// equivalent, for use with the given x86 condition code. 808 SDValue EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC, 809 SelectionDAG &DAG) const; 810 }; 811 812 namespace X86 { 813 FastISel *createFastISel(FunctionLoweringInfo &funcInfo); 814 } 815} 816 817#endif // X86ISELLOWERING_H 818