1 //===-- X86ISelLowering.h - X86 DAG Lowering Interface ----------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #ifndef X86ISELLOWERING_H
16 #define X86ISELLOWERING_H
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"
29 // X86 Specific DAG Nodes
31 // Start the numbering where the builtin ops leave off.
32 FIRST_NUMBER = ISD::BUILTIN_OP_END,
34 /// BSF - Bit scan forward.
35 /// BSR - Bit scan reverse.
39 /// SHLD, SHRD - Double shift instructions. These correspond to
40 /// X86::SHLDxx and X86::SHRDxx instructions.
44 /// FAND - Bitwise logical AND of floating point values. This corresponds
45 /// to X86::ANDPS or X86::ANDPD.
48 /// FOR - Bitwise logical OR of floating point values. This corresponds
49 /// to X86::ORPS or X86::ORPD.
52 /// FXOR - Bitwise logical XOR of floating point values. This corresponds
53 /// to X86::XORPS or X86::XORPD.
56 /// FSRL - Bitwise logical right shift of floating point values. These
57 /// corresponds to X86::PSRLDQ.
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).
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
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
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
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:
93 /// #0 - The incoming token chain
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)
100 /// The result values of these nodes are:
102 /// #0 - The outgoing token chain
103 /// #1 - The first register result value (optional)
104 /// #2 - The second register result value (optional)
108 /// RDTSC_DAG - This operation implements the lowering for
112 /// X86 compare and logical compare instructions.
115 /// X86 bit-test instructions.
118 /// X86 SetCC. Operand 0 is condition code, and operand 1 is the flag
119 /// operand produced by a CMP instruction.
122 // Same as SETCC except it's materialized with a sbb and the value is all
123 // one's or all zero's.
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
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.
138 /// Return with a flag operand. Operand 0 is the chain operand, operand
139 /// 1 is the number of bytes of stack to pop.
142 /// REP_STOS - Repeat fill, corresponds to X86::REP_STOSx.
145 /// REP_MOVS - Repeat move, corresponds to X86::REP_MOVSx.
148 /// GlobalBaseReg - On Darwin, this node represents the result of the popl
149 /// at function entry, used for PIC code.
152 /// Wrapper - A wrapper node for TargetConstantPool,
153 /// TargetExternalSymbol, and TargetGlobalAddress.
156 /// WrapperRIP - Special wrapper used under X86-64 PIC mode for RIP
157 /// relative displacements.
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
165 /// PEXTRB - Extract an 8-bit value from a vector and zero extend it to
166 /// i32, corresponds to X86::PEXTRB.
169 /// PEXTRW - Extract a 16-bit value from a vector and zero extend it to
170 /// i32, corresponds to X86::PEXTRW.
173 /// INSERTPS - Insert any element of a 4 x float vector into any element
174 /// of a destination 4 x floatvector.
177 /// PINSRB - Insert the lower 8-bits of a 32-bit value to a vector,
178 /// corresponds to X86::PINSRB.
181 /// PINSRW - Insert the lower 16-bits of a 32-bit value to a vector,
182 /// corresponds to X86::PINSRW.
185 /// PSHUFB - Shuffle 16 8-bit values within a vector.
188 /// FMAX, FMIN - Floating point max and min.
192 /// FRSQRT, FRCP - Floating point reciprocal-sqrt and reciprocal
193 /// approximation. Note that these typically require refinement
194 /// in order to obtain suitable precision.
197 // TLSADDR - Thread Local Storage.
200 // SegmentBaseAddress - The address segment:0
203 // EH_RETURN - Exception Handling helpers.
206 /// TC_RETURN - Tail call return.
208 /// operand #1 callee (register or absolute)
209 /// operand #2 stack adjustment
210 /// operand #3 optional in flag
213 // LCMPXCHG_DAG, LCMPXCHG8_DAG - Compare and swap.
217 // FNSTCW16m - Store FP control world into i16 memory.
220 // VZEXT_MOVL - Vector move low and zero extend.
223 // VZEXT_LOAD - Load, scalar_to_vector, and zero extend.
226 // VSHL, VSRL - Vector logical left / right shift.
229 // CMPPD, CMPPS - Vector double/float comparison.
230 // CMPPD, CMPPS - Vector double/float comparison.
233 // PCMP* - Vector integer comparisons.
234 PCMPEQB, PCMPEQW, PCMPEQD, PCMPEQQ,
235 PCMPGTB, PCMPGTW, PCMPGTD, PCMPGTQ,
237 // Advanced Encryption Standard (AES) Instructions
238 AESIMC, AESENC, AESENCLAST, AESDEC, AESDECLAST,
240 // ADD, SUB, SMUL, UMUL, etc. - Arithmetic operations with FLAGS results.
241 ADD, SUB, SMUL, UMUL,
242 INC, DEC, OR, XOR, AND,
244 // MUL_IMM - X86 specific multiply by immediate.
247 // PTEST - Vector bitwise comparisons
250 // VASTART_SAVE_XMM_REGS - Save xmm argument registers to the stack,
251 // according to %al. An operator is needed so that this can be expanded
252 // with control flow.
253 VASTART_SAVE_XMM_REGS,
255 // MINGW_ALLOCA - MingW's __alloca call to do stack probing.
258 // ATOMADD64_DAG, ATOMSUB64_DAG, ATOMOR64_DAG, ATOMAND64_DAG,
259 // ATOMXOR64_DAG, ATOMNAND64_DAG, ATOMSWAP64_DAG -
260 // Atomic 64-bit binary operations.
261 ATOMADD64_DAG = ISD::FIRST_TARGET_MEMORY_OPCODE,
269 // WARNING: Do not add anything in the end unless you want the node to
270 // have memop! In fact, starting from ATOMADD64_DAG all opcodes will be
271 // thought as target memory ops!
275 /// Define some predicates that are used for node matching.
277 /// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand
278 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
279 bool isPSHUFDMask(ShuffleVectorSDNode *N);
281 /// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand
282 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
283 bool isPSHUFHWMask(ShuffleVectorSDNode *N);
285 /// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand
286 /// specifies a shuffle of elements that is suitable for input to PSHUFD.
287 bool isPSHUFLWMask(ShuffleVectorSDNode *N);
289 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
290 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
291 bool isSHUFPMask(ShuffleVectorSDNode *N);
293 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
294 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
295 bool isMOVHLPSMask(ShuffleVectorSDNode *N);
297 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
298 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
300 bool isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N);
302 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
303 /// specifies a shuffle of elements that is suitable for MOVLP{S|D}.
304 bool isMOVLPMask(ShuffleVectorSDNode *N);
306 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
307 /// specifies a shuffle of elements that is suitable for MOVHP{S|D}.
308 /// as well as MOVLHPS.
309 bool isMOVLHPSMask(ShuffleVectorSDNode *N);
311 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
312 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
313 bool isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat = false);
315 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
316 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
317 bool isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat = false);
319 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
320 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
322 bool isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N);
324 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
325 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
327 bool isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N);
329 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
330 /// specifies a shuffle of elements that is suitable for input to MOVSS,
331 /// MOVSD, and MOVD, i.e. setting the lowest element.
332 bool isMOVLMask(ShuffleVectorSDNode *N);
334 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
335 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
336 bool isMOVSHDUPMask(ShuffleVectorSDNode *N);
338 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
339 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
340 bool isMOVSLDUPMask(ShuffleVectorSDNode *N);
342 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
343 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
344 bool isMOVDDUPMask(ShuffleVectorSDNode *N);
346 /// isPALIGNRMask - Return true if the specified VECTOR_SHUFFLE operand
347 /// specifies a shuffle of elements that is suitable for input to PALIGNR.
348 bool isPALIGNRMask(ShuffleVectorSDNode *N);
350 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
351 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
353 unsigned getShuffleSHUFImmediate(SDNode *N);
355 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
356 /// the specified VECTOR_SHUFFLE mask with PSHUFHW instruction.
357 unsigned getShufflePSHUFHWImmediate(SDNode *N);
359 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
360 /// the specified VECTOR_SHUFFLE mask with PSHUFLW instruction.
361 unsigned getShufflePSHUFLWImmediate(SDNode *N);
363 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
364 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
365 unsigned getShufflePALIGNRImmediate(SDNode *N);
367 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
369 bool isZeroNode(SDValue Elt);
371 /// isOffsetSuitableForCodeModel - Returns true of the given offset can be
372 /// fit into displacement field of the instruction.
373 bool isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
374 bool hasSymbolicDisplacement = true);
377 //===--------------------------------------------------------------------===//
378 // X86TargetLowering - X86 Implementation of the TargetLowering interface
379 class X86TargetLowering : public TargetLowering {
380 int VarArgsFrameIndex; // FrameIndex for start of varargs area.
381 int RegSaveFrameIndex; // X86-64 vararg func register save area.
382 unsigned VarArgsGPOffset; // X86-64 vararg func int reg offset.
383 unsigned VarArgsFPOffset; // X86-64 vararg func fp reg offset.
384 int BytesToPopOnReturn; // Number of arg bytes ret should pop.
387 explicit X86TargetLowering(X86TargetMachine &TM);
389 /// getPICBaseSymbol - Return the X86-32 PIC base.
390 MCSymbol *getPICBaseSymbol(const MachineFunction *MF, MCContext &Ctx) const;
392 virtual unsigned getJumpTableEncoding() const;
394 virtual const MCExpr *
395 LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
396 const MachineBasicBlock *MBB, unsigned uid,
397 MCContext &Ctx) const;
399 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
401 virtual SDValue getPICJumpTableRelocBase(SDValue Table,
402 SelectionDAG &DAG) const;
403 virtual const MCExpr *
404 getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
405 unsigned JTI, MCContext &Ctx) const;
407 // Return the number of bytes that a function should pop when it returns (in
408 // addition to the space used by the return address).
410 unsigned getBytesToPopOnReturn() const { return BytesToPopOnReturn; }
412 /// getStackPtrReg - Return the stack pointer register we are using: either
414 unsigned getStackPtrReg() const { return X86StackPtr; }
416 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
417 /// function arguments in the caller parameter area. For X86, aggregates
418 /// that contains are placed at 16-byte boundaries while the rest are at
419 /// 4-byte boundaries.
420 virtual unsigned getByValTypeAlignment(const Type *Ty) const;
422 /// getOptimalMemOpType - Returns the target specific optimal type for load
423 /// and store operations as a result of memset, memcpy, and memmove
424 /// lowering. It returns EVT::iAny if SelectionDAG should be responsible for
426 virtual EVT getOptimalMemOpType(uint64_t Size, unsigned Align,
427 bool isSrcConst, bool isSrcStr,
428 SelectionDAG &DAG) const;
430 /// allowsUnalignedMemoryAccesses - Returns true if the target allows
431 /// unaligned memory accesses. of the specified type.
432 virtual bool allowsUnalignedMemoryAccesses(EVT VT) const {
436 /// LowerOperation - Provide custom lowering hooks for some operations.
438 virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG);
440 /// ReplaceNodeResults - Replace the results of node with an illegal result
441 /// type with new values built out of custom code.
443 virtual void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
447 virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const;
449 virtual MachineBasicBlock *EmitInstrWithCustomInserter(MachineInstr *MI,
450 MachineBasicBlock *MBB,
451 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const;
454 /// getTargetNodeName - This method returns the name of a target specific
456 virtual const char *getTargetNodeName(unsigned Opcode) const;
458 /// getSetCCResultType - Return the ISD::SETCC ValueType
459 virtual MVT::SimpleValueType getSetCCResultType(EVT VT) const;
461 /// computeMaskedBitsForTargetNode - Determine which of the bits specified
462 /// in Mask are known to be either zero or one and return them in the
463 /// KnownZero/KnownOne bitsets.
464 virtual void computeMaskedBitsForTargetNode(const SDValue Op,
468 const SelectionDAG &DAG,
469 unsigned Depth = 0) const;
472 isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) const;
474 SDValue getReturnAddressFrameIndex(SelectionDAG &DAG);
476 virtual bool ExpandInlineAsm(CallInst *CI) const;
478 ConstraintType getConstraintType(const std::string &Constraint) const;
480 std::vector<unsigned>
481 getRegClassForInlineAsmConstraint(const std::string &Constraint,
484 virtual const char *LowerXConstraint(EVT ConstraintVT) const;
486 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
487 /// vector. If it is invalid, don't add anything to Ops. If hasMemory is
488 /// true it means one of the asm constraint of the inline asm instruction
489 /// being processed is 'm'.
490 virtual void LowerAsmOperandForConstraint(SDValue Op,
491 char ConstraintLetter,
493 std::vector<SDValue> &Ops,
494 SelectionDAG &DAG) const;
496 /// getRegForInlineAsmConstraint - Given a physical register constraint
497 /// (e.g. {edx}), return the register number and the register class for the
498 /// register. This should only be used for C_Register constraints. On
499 /// error, this returns a register number of 0.
500 std::pair<unsigned, const TargetRegisterClass*>
501 getRegForInlineAsmConstraint(const std::string &Constraint,
504 /// isLegalAddressingMode - Return true if the addressing mode represented
505 /// by AM is legal for this target, for a load/store of the specified type.
506 virtual bool isLegalAddressingMode(const AddrMode &AM, const Type *Ty)const;
508 /// isTruncateFree - Return true if it's free to truncate a value of
509 /// type Ty1 to type Ty2. e.g. On x86 it's free to truncate a i32 value in
510 /// register EAX to i16 by referencing its sub-register AX.
511 virtual bool isTruncateFree(const Type *Ty1, const Type *Ty2) const;
512 virtual bool isTruncateFree(EVT VT1, EVT VT2) const;
514 /// isZExtFree - Return true if any actual instruction that defines a
515 /// value of type Ty1 implicit zero-extends the value to Ty2 in the result
516 /// register. This does not necessarily include registers defined in
517 /// unknown ways, such as incoming arguments, or copies from unknown
518 /// virtual registers. Also, if isTruncateFree(Ty2, Ty1) is true, this
519 /// does not necessarily apply to truncate instructions. e.g. on x86-64,
520 /// all instructions that define 32-bit values implicit zero-extend the
521 /// result out to 64 bits.
522 virtual bool isZExtFree(const Type *Ty1, const Type *Ty2) const;
523 virtual bool isZExtFree(EVT VT1, EVT VT2) const;
525 /// isNarrowingProfitable - Return true if it's profitable to narrow
526 /// operations of type VT1 to VT2. e.g. on x86, it's profitable to narrow
527 /// from i32 to i8 but not from i32 to i16.
528 virtual bool isNarrowingProfitable(EVT VT1, EVT VT2) const;
530 /// isFPImmLegal - Returns true if the target can instruction select the
531 /// specified FP immediate natively. If false, the legalizer will
532 /// materialize the FP immediate as a load from a constant pool.
533 virtual bool isFPImmLegal(const APFloat &Imm, EVT VT) const;
535 /// isShuffleMaskLegal - Targets can use this to indicate that they only
536 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
537 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask
538 /// values are assumed to be legal.
539 virtual bool isShuffleMaskLegal(const SmallVectorImpl<int> &Mask,
542 /// isVectorClearMaskLegal - Similar to isShuffleMaskLegal. This is
543 /// used by Targets can use this to indicate if there is a suitable
544 /// VECTOR_SHUFFLE that can be used to replace a VAND with a constant
546 virtual bool isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
549 /// ShouldShrinkFPConstant - If true, then instruction selection should
550 /// seek to shrink the FP constant of the specified type to a smaller type
551 /// in order to save space and / or reduce runtime.
552 virtual bool ShouldShrinkFPConstant(EVT VT) const {
553 // Don't shrink FP constpool if SSE2 is available since cvtss2sd is more
554 // expensive than a straight movsd. On the other hand, it's important to
555 // shrink long double fp constant since fldt is very slow.
556 return !X86ScalarSSEf64 || VT == MVT::f80;
559 virtual const X86Subtarget* getSubtarget() {
563 /// isScalarFPTypeInSSEReg - Return true if the specified scalar FP type is
564 /// computed in an SSE register, not on the X87 floating point stack.
565 bool isScalarFPTypeInSSEReg(EVT VT) const {
566 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
567 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1
570 /// createFastISel - This method returns a target specific FastISel object,
571 /// or null if the target does not support "fast" ISel.
573 createFastISel(MachineFunction &mf,
574 MachineModuleInfo *mmi, DwarfWriter *dw,
575 DenseMap<const Value *, unsigned> &,
576 DenseMap<const BasicBlock *, MachineBasicBlock *> &,
577 DenseMap<const AllocaInst *, int> &
579 , SmallSet<Instruction*, 8> &
583 /// getFunctionAlignment - Return the Log2 alignment of this function.
584 virtual unsigned getFunctionAlignment(const Function *F) const;
587 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
588 /// make the right decision when generating code for different targets.
589 const X86Subtarget *Subtarget;
590 const X86RegisterInfo *RegInfo;
591 const TargetData *TD;
593 /// X86StackPtr - X86 physical register used as stack ptr.
594 unsigned X86StackPtr;
596 /// X86ScalarSSEf32, X86ScalarSSEf64 - Select between SSE or x87
597 /// floating point ops.
598 /// When SSE is available, use it for f32 operations.
599 /// When SSE2 is available, use it for f64 operations.
600 bool X86ScalarSSEf32;
601 bool X86ScalarSSEf64;
603 /// LegalFPImmediates - A list of legal fp immediates.
604 std::vector<APFloat> LegalFPImmediates;
606 /// addLegalFPImmediate - Indicate that this x86 target can instruction
607 /// select the specified FP immediate natively.
608 void addLegalFPImmediate(const APFloat& Imm) {
609 LegalFPImmediates.push_back(Imm);
612 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
613 CallingConv::ID CallConv, bool isVarArg,
614 const SmallVectorImpl<ISD::InputArg> &Ins,
615 DebugLoc dl, SelectionDAG &DAG,
616 SmallVectorImpl<SDValue> &InVals);
617 SDValue LowerMemArgument(SDValue Chain,
618 CallingConv::ID CallConv,
619 const SmallVectorImpl<ISD::InputArg> &ArgInfo,
620 DebugLoc dl, SelectionDAG &DAG,
621 const CCValAssign &VA, MachineFrameInfo *MFI,
623 SDValue LowerMemOpCallTo(SDValue Chain, SDValue StackPtr, SDValue Arg,
624 DebugLoc dl, SelectionDAG &DAG,
625 const CCValAssign &VA,
626 ISD::ArgFlagsTy Flags);
628 // Call lowering helpers.
630 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
631 /// for tail call optimization. Targets which want to do tail call
632 /// optimization should implement this function.
633 bool IsEligibleForTailCallOptimization(SDValue Callee,
634 CallingConv::ID CalleeCC,
636 bool isCalleeStructRet,
637 bool isCallerStructRet,
638 const SmallVectorImpl<ISD::OutputArg> &Outs,
639 const SmallVectorImpl<ISD::InputArg> &Ins,
640 SelectionDAG& DAG) const;
641 bool IsCalleePop(bool isVarArg, CallingConv::ID CallConv);
642 SDValue EmitTailCallLoadRetAddr(SelectionDAG &DAG, SDValue &OutRetAddr,
643 SDValue Chain, bool IsTailCall, bool Is64Bit,
644 int FPDiff, DebugLoc dl);
646 CCAssignFn *CCAssignFnForNode(CallingConv::ID CallConv) const;
647 unsigned GetAlignedArgumentStackSize(unsigned StackSize, SelectionDAG &DAG);
649 std::pair<SDValue,SDValue> FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
652 SDValue LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
654 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG);
655 SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG);
656 SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG);
657 SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG);
658 SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG);
659 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG);
660 SDValue LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG);
661 SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG);
662 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG);
663 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG);
664 SDValue LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
665 int64_t Offset, SelectionDAG &DAG) const;
666 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG);
667 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG);
668 SDValue LowerExternalSymbol(SDValue Op, SelectionDAG &DAG);
669 SDValue LowerShift(SDValue Op, SelectionDAG &DAG);
670 SDValue BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain, SDValue StackSlot,
672 SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG);
673 SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG);
674 SDValue LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG);
675 SDValue LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG);
676 SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG);
677 SDValue LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG);
678 SDValue LowerFABS(SDValue Op, SelectionDAG &DAG);
679 SDValue LowerFNEG(SDValue Op, SelectionDAG &DAG);
680 SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG);
681 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG);
682 SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG);
683 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG);
684 SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG);
685 SDValue LowerMEMSET(SDValue Op, SelectionDAG &DAG);
686 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG);
687 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG);
688 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG);
689 SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG);
690 SDValue LowerVACOPY(SDValue Op, SelectionDAG &DAG);
691 SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG);
692 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG);
693 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG);
694 SDValue LowerFRAME_TO_ARGS_OFFSET(SDValue Op, SelectionDAG &DAG);
695 SDValue LowerEH_RETURN(SDValue Op, SelectionDAG &DAG);
696 SDValue LowerTRAMPOLINE(SDValue Op, SelectionDAG &DAG);
697 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG);
698 SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG);
699 SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG);
700 SDValue LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG);
701 SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG);
703 SDValue LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG);
704 SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG);
705 SDValue LowerREADCYCLECOUNTER(SDValue Op, SelectionDAG &DAG);
708 LowerFormalArguments(SDValue Chain,
709 CallingConv::ID CallConv, bool isVarArg,
710 const SmallVectorImpl<ISD::InputArg> &Ins,
711 DebugLoc dl, SelectionDAG &DAG,
712 SmallVectorImpl<SDValue> &InVals);
714 LowerCall(SDValue Chain, SDValue Callee,
715 CallingConv::ID CallConv, bool isVarArg, bool &isTailCall,
716 const SmallVectorImpl<ISD::OutputArg> &Outs,
717 const SmallVectorImpl<ISD::InputArg> &Ins,
718 DebugLoc dl, SelectionDAG &DAG,
719 SmallVectorImpl<SDValue> &InVals);
722 LowerReturn(SDValue Chain,
723 CallingConv::ID CallConv, bool isVarArg,
724 const SmallVectorImpl<ISD::OutputArg> &Outs,
725 DebugLoc dl, SelectionDAG &DAG);
728 CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
729 const SmallVectorImpl<EVT> &OutTys,
730 const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
733 void ReplaceATOMIC_BINARY_64(SDNode *N, SmallVectorImpl<SDValue> &Results,
734 SelectionDAG &DAG, unsigned NewOp);
736 SDValue EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
738 SDValue Dst, SDValue Src,
739 SDValue Size, unsigned Align,
740 const Value *DstSV, uint64_t DstSVOff);
741 SDValue EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
743 SDValue Dst, SDValue Src,
744 SDValue Size, unsigned Align,
746 const Value *DstSV, uint64_t DstSVOff,
747 const Value *SrcSV, uint64_t SrcSVOff);
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;
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,
770 TargetRegisterClass *RC,
771 bool invSrc = false) const;
773 MachineBasicBlock *EmitAtomicBit6432WithCustomInserter(
774 MachineInstr *BInstr,
775 MachineBasicBlock *BB,
780 bool invSrc = false) const;
782 /// Utility function to emit atomic min and max. It takes the min/max
783 /// instruction to expand, the associated basic block, and the associated
784 /// cmov opcode for moving the min or max value.
785 MachineBasicBlock *EmitAtomicMinMaxWithCustomInserter(MachineInstr *BInstr,
786 MachineBasicBlock *BB,
787 unsigned cmovOpc) const;
789 /// Utility function to emit the xmm reg save portion of va_start.
790 MachineBasicBlock *EmitVAStartSaveXMMRegsWithCustomInserter(
791 MachineInstr *BInstr,
792 MachineBasicBlock *BB) const;
794 MachineBasicBlock *EmitLoweredSelect(MachineInstr *I,
795 MachineBasicBlock *BB,
796 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const;
798 MachineBasicBlock *EmitLoweredMingwAlloca(MachineInstr *MI,
799 MachineBasicBlock *BB,
800 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const;
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);
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,
813 FastISel *createFastISel(MachineFunction &mf,
814 MachineModuleInfo *mmi, DwarfWriter *dw,
815 DenseMap<const Value *, unsigned> &,
816 DenseMap<const BasicBlock *, MachineBasicBlock *> &,
817 DenseMap<const AllocaInst *, int> &
819 , SmallSet<Instruction*, 8> &
825 #endif // X86ISELLOWERING_H