LLVM API Documentation
00001 //===-- X86ISelLowering.h - X86 DAG Lowering Interface ----------*- C++ -*-===// 00002 // 00003 // The LLVM Compiler Infrastructure 00004 // 00005 // This file is distributed under the University of Illinois Open Source 00006 // License. See LICENSE.TXT for details. 00007 // 00008 //===----------------------------------------------------------------------===// 00009 // 00010 // This file defines the interfaces that X86 uses to lower LLVM code into a 00011 // selection DAG. 00012 // 00013 //===----------------------------------------------------------------------===// 00014 00015 #ifndef X86ISELLOWERING_H 00016 #define X86ISELLOWERING_H 00017 00018 #include "X86Subtarget.h" 00019 #include "X86RegisterInfo.h" 00020 #include "X86MachineFunctionInfo.h" 00021 #include "llvm/Target/TargetLowering.h" 00022 #include "llvm/CodeGen/FastISel.h" 00023 #include "llvm/CodeGen/SelectionDAG.h" 00024 #include "llvm/CodeGen/CallingConvLower.h" 00025 00026 namespace llvm { 00027 namespace X86ISD { 00028 // X86 Specific DAG Nodes 00029 enum NodeType { 00030 // Start the numbering where the builtin ops leave off. 00031 FIRST_NUMBER = ISD::BUILTIN_OP_END, 00032 00033 /// BSF - Bit scan forward. 00034 /// BSR - Bit scan reverse. 00035 BSF, 00036 BSR, 00037 00038 /// SHLD, SHRD - Double shift instructions. These correspond to 00039 /// X86::SHLDxx and X86::SHRDxx instructions. 00040 SHLD, 00041 SHRD, 00042 00043 /// FAND - Bitwise logical AND of floating point values. This corresponds 00044 /// to X86::ANDPS or X86::ANDPD. 00045 FAND, 00046 00047 /// FOR - Bitwise logical OR of floating point values. This corresponds 00048 /// to X86::ORPS or X86::ORPD. 00049 FOR, 00050 00051 /// FXOR - Bitwise logical XOR of floating point values. This corresponds 00052 /// to X86::XORPS or X86::XORPD. 00053 FXOR, 00054 00055 /// FSRL - Bitwise logical right shift of floating point values. These 00056 /// corresponds to X86::PSRLDQ. 00057 FSRL, 00058 00059 /// FILD, FILD_FLAG - This instruction implements SINT_TO_FP with the 00060 /// integer source in memory and FP reg result. This corresponds to the 00061 /// X86::FILD*m instructions. It has three inputs (token chain, address, 00062 /// and source type) and two outputs (FP value and token chain). FILD_FLAG 00063 /// also produces a flag). 00064 FILD, 00065 FILD_FLAG, 00066 00067 /// FP_TO_INT*_IN_MEM - This instruction implements FP_TO_SINT with the 00068 /// integer destination in memory and a FP reg source. This corresponds 00069 /// to the X86::FIST*m instructions and the rounding mode change stuff. It 00070 /// has two inputs (token chain and address) and two outputs (int value 00071 /// and token chain). 00072 FP_TO_INT16_IN_MEM, 00073 FP_TO_INT32_IN_MEM, 00074 FP_TO_INT64_IN_MEM, 00075 00076 /// FLD - This instruction implements an extending load to FP stack slots. 00077 /// This corresponds to the X86::FLD32m / X86::FLD64m. It takes a chain 00078 /// operand, ptr to load from, and a ValueType node indicating the type 00079 /// to load to. 00080 FLD, 00081 00082 /// FST - This instruction implements a truncating store to FP stack 00083 /// slots. This corresponds to the X86::FST32m / X86::FST64m. It takes a 00084 /// chain operand, value to store, address, and a ValueType to store it 00085 /// as. 00086 FST, 00087 00088 /// CALL - These operations represent an abstract X86 call 00089 /// instruction, which includes a bunch of information. In particular the 00090 /// operands of these node are: 00091 /// 00092 /// #0 - The incoming token chain 00093 /// #1 - The callee 00094 /// #2 - The number of arg bytes the caller pushes on the stack. 00095 /// #3 - The number of arg bytes the callee pops off the stack. 00096 /// #4 - The value to pass in AL/AX/EAX (optional) 00097 /// #5 - The value to pass in DL/DX/EDX (optional) 00098 /// 00099 /// The result values of these nodes are: 00100 /// 00101 /// #0 - The outgoing token chain 00102 /// #1 - The first register result value (optional) 00103 /// #2 - The second register result value (optional) 00104 /// 00105 CALL, 00106 00107 /// RDTSC_DAG - This operation implements the lowering for 00108 /// readcyclecounter 00109 RDTSC_DAG, 00110 00111 /// X86 compare and logical compare instructions. 00112 CMP, COMI, UCOMI, 00113 00114 /// X86 bit-test instructions. 00115 BT, 00116 00117 /// X86 SetCC. Operand 0 is condition code, and operand 1 is the flag 00118 /// operand produced by a CMP instruction. 00119 SETCC, 00120 00121 /// X86 conditional moves. Operand 0 and operand 1 are the two values 00122 /// to select from. Operand 2 is the condition code, and operand 3 is the 00123 /// flag operand produced by a CMP or TEST instruction. It also writes a 00124 /// flag result. 00125 CMOV, 00126 00127 /// X86 conditional branches. Operand 0 is the chain operand, operand 1 00128 /// is the block to branch if condition is true, operand 2 is the 00129 /// condition code, and operand 3 is the flag operand produced by a CMP 00130 /// or TEST instruction. 00131 BRCOND, 00132 00133 /// Return with a flag operand. Operand 0 is the chain operand, operand 00134 /// 1 is the number of bytes of stack to pop. 00135 RET_FLAG, 00136 00137 /// REP_STOS - Repeat fill, corresponds to X86::REP_STOSx. 00138 REP_STOS, 00139 00140 /// REP_MOVS - Repeat move, corresponds to X86::REP_MOVSx. 00141 REP_MOVS, 00142 00143 /// GlobalBaseReg - On Darwin, this node represents the result of the popl 00144 /// at function entry, used for PIC code. 00145 GlobalBaseReg, 00146 00147 /// Wrapper - A wrapper node for TargetConstantPool, 00148 /// TargetExternalSymbol, and TargetGlobalAddress. 00149 Wrapper, 00150 00151 /// WrapperRIP - Special wrapper used under X86-64 PIC mode for RIP 00152 /// relative displacements. 00153 WrapperRIP, 00154 00155 /// PEXTRB - Extract an 8-bit value from a vector and zero extend it to 00156 /// i32, corresponds to X86::PEXTRB. 00157 PEXTRB, 00158 00159 /// PEXTRW - Extract a 16-bit value from a vector and zero extend it to 00160 /// i32, corresponds to X86::PEXTRW. 00161 PEXTRW, 00162 00163 /// INSERTPS - Insert any element of a 4 x float vector into any element 00164 /// of a destination 4 x floatvector. 00165 INSERTPS, 00166 00167 /// PINSRB - Insert the lower 8-bits of a 32-bit value to a vector, 00168 /// corresponds to X86::PINSRB. 00169 PINSRB, 00170 00171 /// PINSRW - Insert the lower 16-bits of a 32-bit value to a vector, 00172 /// corresponds to X86::PINSRW. 00173 PINSRW, 00174 00175 /// PSHUFB - Shuffle 16 8-bit values within a vector. 00176 PSHUFB, 00177 00178 /// FMAX, FMIN - Floating point max and min. 00179 /// 00180 FMAX, FMIN, 00181 00182 /// FRSQRT, FRCP - Floating point reciprocal-sqrt and reciprocal 00183 /// approximation. Note that these typically require refinement 00184 /// in order to obtain suitable precision. 00185 FRSQRT, FRCP, 00186 00187 // TLSADDR - Thread Local Storage. 00188 TLSADDR, 00189 00190 // SegmentBaseAddress - The address segment:0 00191 SegmentBaseAddress, 00192 00193 // EH_RETURN - Exception Handling helpers. 00194 EH_RETURN, 00195 00196 /// TC_RETURN - Tail call return. 00197 /// operand #0 chain 00198 /// operand #1 callee (register or absolute) 00199 /// operand #2 stack adjustment 00200 /// operand #3 optional in flag 00201 TC_RETURN, 00202 00203 // LCMPXCHG_DAG, LCMPXCHG8_DAG - Compare and swap. 00204 LCMPXCHG_DAG, 00205 LCMPXCHG8_DAG, 00206 00207 // ATOMADD64_DAG, ATOMSUB64_DAG, ATOMOR64_DAG, ATOMAND64_DAG, 00208 // ATOMXOR64_DAG, ATOMNAND64_DAG, ATOMSWAP64_DAG - 00209 // Atomic 64-bit binary operations. 00210 ATOMADD64_DAG, 00211 ATOMSUB64_DAG, 00212 ATOMOR64_DAG, 00213 ATOMXOR64_DAG, 00214 ATOMAND64_DAG, 00215 ATOMNAND64_DAG, 00216 ATOMSWAP64_DAG, 00217 00218 // FNSTCW16m - Store FP control world into i16 memory. 00219 FNSTCW16m, 00220 00221 // VZEXT_MOVL - Vector move low and zero extend. 00222 VZEXT_MOVL, 00223 00224 // VZEXT_LOAD - Load, scalar_to_vector, and zero extend. 00225 VZEXT_LOAD, 00226 00227 // VSHL, VSRL - Vector logical left / right shift. 00228 VSHL, VSRL, 00229 00230 // CMPPD, CMPPS - Vector double/float comparison. 00231 // CMPPD, CMPPS - Vector double/float comparison. 00232 CMPPD, CMPPS, 00233 00234 // PCMP* - Vector integer comparisons. 00235 PCMPEQB, PCMPEQW, PCMPEQD, PCMPEQQ, 00236 PCMPGTB, PCMPGTW, PCMPGTD, PCMPGTQ, 00237 00238 // ADD, SUB, SMUL, UMUL, etc. - Arithmetic operations with FLAGS results. 00239 ADD, SUB, SMUL, UMUL, 00240 INC, DEC, 00241 00242 // MUL_IMM - X86 specific multiply by immediate. 00243 MUL_IMM, 00244 00245 // PTEST - Vector bitwise comparisons 00246 PTEST, 00247 00248 // VASTART_SAVE_XMM_REGS - Save xmm argument registers to the stack, 00249 // according to %al. An operator is needed so that this can be expanded 00250 // with control flow. 00251 VASTART_SAVE_XMM_REGS 00252 }; 00253 } 00254 00255 /// Define some predicates that are used for node matching. 00256 namespace X86 { 00257 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand 00258 /// specifies a shuffle of elements that is suitable for input to PSHUFD. 00259 bool isPSHUFDMask(ShuffleVectorSDNode *N); 00260 00261 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand 00262 /// specifies a shuffle of elements that is suitable for input to PSHUFD. 00263 bool isPSHUFHWMask(ShuffleVectorSDNode *N); 00264 00265 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand 00266 /// specifies a shuffle of elements that is suitable for input to PSHUFD. 00267 bool isPSHUFLWMask(ShuffleVectorSDNode *N); 00268 00269 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand 00270 /// specifies a shuffle of elements that is suitable for input to SHUFP*. 00271 bool isSHUFPMask(ShuffleVectorSDNode *N); 00272 00273 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand 00274 /// specifies a shuffle of elements that is suitable for input to MOVHLPS. 00275 bool isMOVHLPSMask(ShuffleVectorSDNode *N); 00276 00277 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form 00278 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef, 00279 /// <2, 3, 2, 3> 00280 bool isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N); 00281 00282 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand 00283 /// specifies a shuffle of elements that is suitable for MOVLP{S|D}. 00284 bool isMOVLPMask(ShuffleVectorSDNode *N); 00285 00286 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand 00287 /// specifies a shuffle of elements that is suitable for MOVHP{S|D}. 00288 /// as well as MOVLHPS. 00289 bool isMOVHPMask(ShuffleVectorSDNode *N); 00290 00291 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand 00292 /// specifies a shuffle of elements that is suitable for input to UNPCKL. 00293 bool isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat = false); 00294 00295 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand 00296 /// specifies a shuffle of elements that is suitable for input to UNPCKH. 00297 bool isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat = false); 00298 00299 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form 00300 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef, 00301 /// <0, 0, 1, 1> 00302 bool isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N); 00303 00304 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form 00305 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef, 00306 /// <2, 2, 3, 3> 00307 bool isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N); 00308 00309 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand 00310 /// specifies a shuffle of elements that is suitable for input to MOVSS, 00311 /// MOVSD, and MOVD, i.e. setting the lowest element. 00312 bool isMOVLMask(ShuffleVectorSDNode *N); 00313 00314 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand 00315 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP. 00316 bool isMOVSHDUPMask(ShuffleVectorSDNode *N); 00317 00318 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand 00319 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP. 00320 bool isMOVSLDUPMask(ShuffleVectorSDNode *N); 00321 00322 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand 00323 /// specifies a shuffle of elements that is suitable for input to MOVDDUP. 00324 bool isMOVDDUPMask(ShuffleVectorSDNode *N); 00325 00326 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle 00327 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP* 00328 /// instructions. 00329 unsigned getShuffleSHUFImmediate(SDNode *N); 00330 00331 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle 00332 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW 00333 /// instructions. 00334 unsigned getShufflePSHUFHWImmediate(SDNode *N); 00335 00336 /// getShufflePSHUFKWImmediate - Return the appropriate immediate to shuffle 00337 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW 00338 /// instructions. 00339 unsigned getShufflePSHUFLWImmediate(SDNode *N); 00340 00341 /// isZeroNode - Returns true if Elt is a constant zero or a floating point 00342 /// constant +0.0. 00343 bool isZeroNode(SDValue Elt); 00344 00345 /// isOffsetSuitableForCodeModel - Returns true of the given offset can be 00346 /// fit into displacement field of the instruction. 00347 bool isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M, 00348 bool hasSymbolicDisplacement = true); 00349 } 00350 00351 //===--------------------------------------------------------------------===// 00352 // X86TargetLowering - X86 Implementation of the TargetLowering interface 00353 class X86TargetLowering : public TargetLowering { 00354 int VarArgsFrameIndex; // FrameIndex for start of varargs area. 00355 int RegSaveFrameIndex; // X86-64 vararg func register save area. 00356 unsigned VarArgsGPOffset; // X86-64 vararg func int reg offset. 00357 unsigned VarArgsFPOffset; // X86-64 vararg func fp reg offset. 00358 int BytesToPopOnReturn; // Number of arg bytes ret should pop. 00359 int BytesCallerReserves; // Number of arg bytes caller makes. 00360 00361 public: 00362 explicit X86TargetLowering(X86TargetMachine &TM); 00363 00364 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC 00365 /// jumptable. 00366 SDValue getPICJumpTableRelocBase(SDValue Table, 00367 SelectionDAG &DAG) const; 00368 00369 // Return the number of bytes that a function should pop when it returns (in 00370 // addition to the space used by the return address). 00371 // 00372 unsigned getBytesToPopOnReturn() const { return BytesToPopOnReturn; } 00373 00374 // Return the number of bytes that the caller reserves for arguments passed 00375 // to this function. 00376 unsigned getBytesCallerReserves() const { return BytesCallerReserves; } 00377 00378 /// getStackPtrReg - Return the stack pointer register we are using: either 00379 /// ESP or RSP. 00380 unsigned getStackPtrReg() const { return X86StackPtr; } 00381 00382 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate 00383 /// function arguments in the caller parameter area. For X86, aggregates 00384 /// that contains are placed at 16-byte boundaries while the rest are at 00385 /// 4-byte boundaries. 00386 virtual unsigned getByValTypeAlignment(const Type *Ty) const; 00387 00388 /// getOptimalMemOpType - Returns the target specific optimal type for load 00389 /// and store operations as a result of memset, memcpy, and memmove 00390 /// lowering. It returns EVT::iAny if SelectionDAG should be responsible for 00391 /// determining it. 00392 virtual EVT getOptimalMemOpType(uint64_t Size, unsigned Align, 00393 bool isSrcConst, bool isSrcStr, 00394 SelectionDAG &DAG) const; 00395 00396 /// allowsUnalignedMemoryAccesses - Returns true if the target allows 00397 /// unaligned memory accesses. of the specified type. 00398 virtual bool allowsUnalignedMemoryAccesses(EVT VT) const { 00399 return true; 00400 } 00401 00402 /// LowerOperation - Provide custom lowering hooks for some operations. 00403 /// 00404 virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG); 00405 00406 /// ReplaceNodeResults - Replace the results of node with an illegal result 00407 /// type with new values built out of custom code. 00408 /// 00409 virtual void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results, 00410 SelectionDAG &DAG); 00411 00412 00413 virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const; 00414 00415 virtual MachineBasicBlock *EmitInstrWithCustomInserter(MachineInstr *MI, 00416 MachineBasicBlock *MBB) const; 00417 00418 00419 /// getTargetNodeName - This method returns the name of a target specific 00420 /// DAG node. 00421 virtual const char *getTargetNodeName(unsigned Opcode) const; 00422 00423 /// getSetCCResultType - Return the ISD::SETCC ValueType 00424 virtual MVT::SimpleValueType getSetCCResultType(EVT VT) const; 00425 00426 /// computeMaskedBitsForTargetNode - Determine which of the bits specified 00427 /// in Mask are known to be either zero or one and return them in the 00428 /// KnownZero/KnownOne bitsets. 00429 virtual void computeMaskedBitsForTargetNode(const SDValue Op, 00430 const APInt &Mask, 00431 APInt &KnownZero, 00432 APInt &KnownOne, 00433 const SelectionDAG &DAG, 00434 unsigned Depth = 0) const; 00435 00436 virtual bool 00437 isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) const; 00438 00439 SDValue getReturnAddressFrameIndex(SelectionDAG &DAG); 00440 00441 virtual bool ExpandInlineAsm(CallInst *CI) const; 00442 00443 ConstraintType getConstraintType(const std::string &Constraint) const; 00444 00445 std::vector<unsigned> 00446 getRegClassForInlineAsmConstraint(const std::string &Constraint, 00447 EVT VT) const; 00448 00449 virtual const char *LowerXConstraint(EVT ConstraintVT) const; 00450 00451 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops 00452 /// vector. If it is invalid, don't add anything to Ops. If hasMemory is 00453 /// true it means one of the asm constraint of the inline asm instruction 00454 /// being processed is 'm'. 00455 virtual void LowerAsmOperandForConstraint(SDValue Op, 00456 char ConstraintLetter, 00457 bool hasMemory, 00458 std::vector<SDValue> &Ops, 00459 SelectionDAG &DAG) const; 00460 00461 /// getRegForInlineAsmConstraint - Given a physical register constraint 00462 /// (e.g. {edx}), return the register number and the register class for the 00463 /// register. This should only be used for C_Register constraints. On 00464 /// error, this returns a register number of 0. 00465 std::pair<unsigned, const TargetRegisterClass*> 00466 getRegForInlineAsmConstraint(const std::string &Constraint, 00467 EVT VT) const; 00468 00469 /// isLegalAddressingMode - Return true if the addressing mode represented 00470 /// by AM is legal for this target, for a load/store of the specified type. 00471 virtual bool isLegalAddressingMode(const AddrMode &AM, const Type *Ty)const; 00472 00473 /// isTruncateFree - Return true if it's free to truncate a value of 00474 /// type Ty1 to type Ty2. e.g. On x86 it's free to truncate a i32 value in 00475 /// register EAX to i16 by referencing its sub-register AX. 00476 virtual bool isTruncateFree(const Type *Ty1, const Type *Ty2) const; 00477 virtual bool isTruncateFree(EVT VT1, EVT VT2) const; 00478 00479 /// isZExtFree - Return true if any actual instruction that defines a 00480 /// value of type Ty1 implicit zero-extends the value to Ty2 in the result 00481 /// register. This does not necessarily include registers defined in 00482 /// unknown ways, such as incoming arguments, or copies from unknown 00483 /// virtual registers. Also, if isTruncateFree(Ty2, Ty1) is true, this 00484 /// does not necessarily apply to truncate instructions. e.g. on x86-64, 00485 /// all instructions that define 32-bit values implicit zero-extend the 00486 /// result out to 64 bits. 00487 virtual bool isZExtFree(const Type *Ty1, const Type *Ty2) const; 00488 virtual bool isZExtFree(EVT VT1, EVT VT2) const; 00489 00490 /// isNarrowingProfitable - Return true if it's profitable to narrow 00491 /// operations of type VT1 to VT2. e.g. on x86, it's profitable to narrow 00492 /// from i32 to i8 but not from i32 to i16. 00493 virtual bool isNarrowingProfitable(EVT VT1, EVT VT2) const; 00494 00495 /// isShuffleMaskLegal - Targets can use this to indicate that they only 00496 /// support *some* VECTOR_SHUFFLE operations, those with specific masks. 00497 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask 00498 /// values are assumed to be legal. 00499 virtual bool isShuffleMaskLegal(const SmallVectorImpl<int> &Mask, 00500 EVT VT) const; 00501 00502 /// isVectorClearMaskLegal - Similar to isShuffleMaskLegal. This is 00503 /// used by Targets can use this to indicate if there is a suitable 00504 /// VECTOR_SHUFFLE that can be used to replace a VAND with a constant 00505 /// pool entry. 00506 virtual bool isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask, 00507 EVT VT) const; 00508 00509 /// ShouldShrinkFPConstant - If true, then instruction selection should 00510 /// seek to shrink the FP constant of the specified type to a smaller type 00511 /// in order to save space and / or reduce runtime. 00512 virtual bool ShouldShrinkFPConstant(EVT VT) const { 00513 // Don't shrink FP constpool if SSE2 is available since cvtss2sd is more 00514 // expensive than a straight movsd. On the other hand, it's important to 00515 // shrink long double fp constant since fldt is very slow. 00516 return !X86ScalarSSEf64 || VT == MVT::f80; 00517 } 00518 00519 /// IsEligibleForTailCallOptimization - Check whether the call is eligible 00520 /// for tail call optimization. Targets which want to do tail call 00521 /// optimization should implement this function. 00522 virtual bool 00523 IsEligibleForTailCallOptimization(SDValue Callee, 00524 unsigned CalleeCC, 00525 bool isVarArg, 00526 const SmallVectorImpl<ISD::InputArg> &Ins, 00527 SelectionDAG& DAG) const; 00528 00529 virtual const X86Subtarget* getSubtarget() { 00530 return Subtarget; 00531 } 00532 00533 /// isScalarFPTypeInSSEReg - Return true if the specified scalar FP type is 00534 /// computed in an SSE register, not on the X87 floating point stack. 00535 bool isScalarFPTypeInSSEReg(EVT VT) const { 00536 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2 00537 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1 00538 } 00539 00540 /// getWidenVectorType: given a vector type, returns the type to widen 00541 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself. 00542 /// If there is no vector type that we want to widen to, returns EVT::Other 00543 /// When and were to widen is target dependent based on the cost of 00544 /// scalarizing vs using the wider vector type. 00545 virtual EVT getWidenVectorType(EVT VT) const; 00546 00547 /// createFastISel - This method returns a target specific FastISel object, 00548 /// or null if the target does not support "fast" ISel. 00549 virtual FastISel * 00550 createFastISel(MachineFunction &mf, 00551 MachineModuleInfo *mmi, DwarfWriter *dw, 00552 DenseMap<const Value *, unsigned> &, 00553 DenseMap<const BasicBlock *, MachineBasicBlock *> &, 00554 DenseMap<const AllocaInst *, int> & 00555 #ifndef NDEBUG 00556 , SmallSet<Instruction*, 8> & 00557 #endif 00558 ); 00559 00560 /// getFunctionAlignment - Return the Log2 alignment of this function. 00561 virtual unsigned getFunctionAlignment(const Function *F) const; 00562 00563 private: 00564 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can 00565 /// make the right decision when generating code for different targets. 00566 const X86Subtarget *Subtarget; 00567 const X86RegisterInfo *RegInfo; 00568 const TargetData *TD; 00569 00570 /// X86StackPtr - X86 physical register used as stack ptr. 00571 unsigned X86StackPtr; 00572 00573 /// X86ScalarSSEf32, X86ScalarSSEf64 - Select between SSE or x87 00574 /// floating point ops. 00575 /// When SSE is available, use it for f32 operations. 00576 /// When SSE2 is available, use it for f64 operations. 00577 bool X86ScalarSSEf32; 00578 bool X86ScalarSSEf64; 00579 00580 SDValue LowerCallResult(SDValue Chain, SDValue InFlag, 00581 unsigned CallConv, bool isVarArg, 00582 const SmallVectorImpl<ISD::InputArg> &Ins, 00583 DebugLoc dl, SelectionDAG &DAG, 00584 SmallVectorImpl<SDValue> &InVals); 00585 SDValue LowerMemArgument(SDValue Chain, 00586 unsigned CallConv, 00587 const SmallVectorImpl<ISD::InputArg> &ArgInfo, 00588 DebugLoc dl, SelectionDAG &DAG, 00589 const CCValAssign &VA, MachineFrameInfo *MFI, 00590 unsigned i); 00591 SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg, 00592 DebugLoc dl, SelectionDAG &DAG, 00593 const CCValAssign &VA, 00594 ISD::ArgFlagsTy Flags); 00595 00596 // Call lowering helpers. 00597 bool IsCalleePop(bool isVarArg, unsigned CallConv); 00598 SDValue EmitTailCallLoadRetAddr(SelectionDAG &DAG, SDValue &OutRetAddr, 00599 SDValue Chain, bool IsTailCall, bool Is64Bit, 00600 int FPDiff, DebugLoc dl); 00601 00602 CCAssignFn *CCAssignFnForNode(unsigned CallConv) const; 00603 NameDecorationStyle NameDecorationForCallConv(unsigned CallConv); 00604 unsigned GetAlignedArgumentStackSize(unsigned StackSize, SelectionDAG &DAG); 00605 00606 std::pair<SDValue,SDValue> FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, 00607 bool isSigned); 00608 00609 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG); 00610 SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG); 00611 SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG); 00612 SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG); 00613 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG); 00614 SDValue LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG); 00615 SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG); 00616 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG); 00617 SDValue LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl, 00618 int64_t Offset, SelectionDAG &DAG) const; 00619 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG); 00620 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG); 00621 SDValue LowerExternalSymbol(SDValue Op, SelectionDAG &DAG); 00622 SDValue LowerShift(SDValue Op, SelectionDAG &DAG); 00623 SDValue BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain, SDValue StackSlot, 00624 SelectionDAG &DAG); 00625 SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG); 00626 SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG); 00627 SDValue LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG); 00628 SDValue LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG); 00629 SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG); 00630 SDValue LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG); 00631 SDValue LowerFABS(SDValue Op, SelectionDAG &DAG); 00632 SDValue LowerFNEG(SDValue Op, SelectionDAG &DAG); 00633 SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG); 00634 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG); 00635 SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG); 00636 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG); 00637 SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG); 00638 SDValue LowerMEMSET(SDValue Op, SelectionDAG &DAG); 00639 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG); 00640 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG); 00641 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG); 00642 SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG); 00643 SDValue LowerVACOPY(SDValue Op, SelectionDAG &DAG); 00644 SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG); 00645 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG); 00646 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG); 00647 SDValue LowerFRAME_TO_ARGS_OFFSET(SDValue Op, SelectionDAG &DAG); 00648 SDValue LowerEH_RETURN(SDValue Op, SelectionDAG &DAG); 00649 SDValue LowerTRAMPOLINE(SDValue Op, SelectionDAG &DAG); 00650 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG); 00651 SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG); 00652 SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG); 00653 SDValue LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG); 00654 SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG); 00655 00656 SDValue LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG); 00657 SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG); 00658 SDValue LowerREADCYCLECOUNTER(SDValue Op, SelectionDAG &DAG); 00659 00660 virtual SDValue 00661 LowerFormalArguments(SDValue Chain, 00662 unsigned CallConv, bool isVarArg, 00663 const SmallVectorImpl<ISD::InputArg> &Ins, 00664 DebugLoc dl, SelectionDAG &DAG, 00665 SmallVectorImpl<SDValue> &InVals); 00666 virtual SDValue 00667 LowerCall(SDValue Chain, SDValue Callee, 00668 unsigned CallConv, bool isVarArg, bool isTailCall, 00669 const SmallVectorImpl<ISD::OutputArg> &Outs, 00670 const SmallVectorImpl<ISD::InputArg> &Ins, 00671 DebugLoc dl, SelectionDAG &DAG, 00672 SmallVectorImpl<SDValue> &InVals); 00673 00674 virtual SDValue 00675 LowerReturn(SDValue Chain, 00676 unsigned CallConv, bool isVarArg, 00677 const SmallVectorImpl<ISD::OutputArg> &Outs, 00678 DebugLoc dl, SelectionDAG &DAG); 00679 00680 void ReplaceATOMIC_BINARY_64(SDNode *N, SmallVectorImpl<SDValue> &Results, 00681 SelectionDAG &DAG, unsigned NewOp); 00682 00683 SDValue EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl, 00684 SDValue Chain, 00685 SDValue Dst, SDValue Src, 00686 SDValue Size, unsigned Align, 00687 const Value *DstSV, uint64_t DstSVOff); 00688 SDValue EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl, 00689 SDValue Chain, 00690 SDValue Dst, SDValue Src, 00691 SDValue Size, unsigned Align, 00692 bool AlwaysInline, 00693 const Value *DstSV, uint64_t DstSVOff, 00694 const Value *SrcSV, uint64_t SrcSVOff); 00695 00696 /// Utility function to emit string processing sse4.2 instructions 00697 /// that return in xmm0. 00698 // This takes the instruction to expand, the associated machine basic 00699 // block, the number of args, and whether or not the second arg is 00700 // in memory or not. 00701 MachineBasicBlock *EmitPCMP(MachineInstr *BInstr, MachineBasicBlock *BB, 00702 unsigned argNum, bool inMem) const; 00703 00704 /// Utility function to emit atomic bitwise operations (and, or, xor). 00705 // It takes the bitwise instruction to expand, the associated machine basic 00706 // block, and the associated X86 opcodes for reg/reg and reg/imm. 00707 MachineBasicBlock *EmitAtomicBitwiseWithCustomInserter( 00708 MachineInstr *BInstr, 00709 MachineBasicBlock *BB, 00710 unsigned regOpc, 00711 unsigned immOpc, 00712 unsigned loadOpc, 00713 unsigned cxchgOpc, 00714 unsigned copyOpc, 00715 unsigned notOpc, 00716 unsigned EAXreg, 00717 TargetRegisterClass *RC, 00718 bool invSrc = false) const; 00719 00720 MachineBasicBlock *EmitAtomicBit6432WithCustomInserter( 00721 MachineInstr *BInstr, 00722 MachineBasicBlock *BB, 00723 unsigned regOpcL, 00724 unsigned regOpcH, 00725 unsigned immOpcL, 00726 unsigned immOpcH, 00727 bool invSrc = false) const; 00728 00729 /// Utility function to emit atomic min and max. It takes the min/max 00730 /// instruction to expand, the associated basic block, and the associated 00731 /// cmov opcode for moving the min or max value. 00732 MachineBasicBlock *EmitAtomicMinMaxWithCustomInserter(MachineInstr *BInstr, 00733 MachineBasicBlock *BB, 00734 unsigned cmovOpc) const; 00735 00736 /// Utility function to emit the xmm reg save portion of va_start. 00737 MachineBasicBlock *EmitVAStartSaveXMMRegsWithCustomInserter( 00738 MachineInstr *BInstr, 00739 MachineBasicBlock *BB) const; 00740 00741 /// Emit nodes that will be selected as "test Op0,Op0", or something 00742 /// equivalent, for use with the given x86 condition code. 00743 SDValue EmitTest(SDValue Op0, unsigned X86CC, SelectionDAG &DAG); 00744 00745 /// Emit nodes that will be selected as "cmp Op0,Op1", or something 00746 /// equivalent, for use with the given x86 condition code. 00747 SDValue EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC, 00748 SelectionDAG &DAG); 00749 }; 00750 00751 namespace X86 { 00752 FastISel *createFastISel(MachineFunction &mf, 00753 MachineModuleInfo *mmi, DwarfWriter *dw, 00754 DenseMap<const Value *, unsigned> &, 00755 DenseMap<const BasicBlock *, MachineBasicBlock *> &, 00756 DenseMap<const AllocaInst *, int> & 00757 #ifndef NDEBUG 00758 , SmallSet<Instruction*, 8> & 00759 #endif 00760 ); 00761 } 00762 } 00763 00764 #endif // X86ISELLOWERING_H