1 //===-- PPCISelLowering.h - PPC32 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 PPC uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #ifndef LLVM_LIB_TARGET_POWERPC_PPCISELLOWERING_H
16 #define LLVM_LIB_TARGET_POWERPC_PPCISELLOWERING_H
19 #include "PPCInstrInfo.h"
20 #include "llvm/CodeGen/CallingConvLower.h"
21 #include "llvm/CodeGen/MachineFunction.h"
22 #include "llvm/CodeGen/MachineMemOperand.h"
23 #include "llvm/CodeGen/MachineValueType.h"
24 #include "llvm/CodeGen/SelectionDAG.h"
25 #include "llvm/CodeGen/SelectionDAGNodes.h"
26 #include "llvm/CodeGen/ValueTypes.h"
27 #include "llvm/IR/Attributes.h"
28 #include "llvm/IR/CallingConv.h"
29 #include "llvm/IR/Function.h"
30 #include "llvm/IR/InlineAsm.h"
31 #include "llvm/IR/Metadata.h"
32 #include "llvm/IR/Type.h"
33 #include "llvm/Target/TargetLowering.h"
40 enum NodeType : unsigned {
41 // Start the numbering where the builtin ops and target ops leave off.
42 FIRST_NUMBER = ISD::BUILTIN_OP_END,
44 /// FSEL - Traditional three-operand fsel node.
48 /// FCFID - The FCFID instruction, taking an f64 operand and producing
49 /// and f64 value containing the FP representation of the integer that
50 /// was temporarily in the f64 operand.
53 /// Newer FCFID[US] integer-to-floating-point conversion instructions for
54 /// unsigned integers and single-precision outputs.
55 FCFIDU, FCFIDS, FCFIDUS,
57 /// FCTI[D,W]Z - The FCTIDZ and FCTIWZ instructions, taking an f32 or f64
58 /// operand, producing an f64 value containing the integer representation
62 /// Newer FCTI[D,W]UZ floating-point-to-integer conversion instructions for
63 /// unsigned integers with round toward zero.
66 /// VEXTS, ByteWidth - takes an input in VSFRC and produces an output in
67 /// VSFRC that is sign-extended from ByteWidth to a 64-byte integer.
70 /// Reciprocal estimate instructions (unary FP ops).
73 // VMADDFP, VNMSUBFP - The VMADDFP and VNMSUBFP instructions, taking
74 // three v4f32 operands and producing a v4f32 result.
77 /// VPERM - The PPC VPERM Instruction.
81 /// XXSPLT - The PPC VSX splat instructions
85 /// XXINSERT - The PPC VSX insert instruction
89 /// VECSHL - The PPC VSX shift left instruction
93 /// XXPERMDI - The PPC XXPERMDI instruction
97 /// The CMPB instruction (takes two operands of i32 or i64).
100 /// Hi/Lo - These represent the high and low 16-bit parts of a global
101 /// address respectively. These nodes have two operands, the first of
102 /// which must be a TargetGlobalAddress, and the second of which must be a
103 /// Constant. Selected naively, these turn into 'lis G+C' and 'li G+C',
104 /// though these are usually folded into other nodes.
107 /// The following two target-specific nodes are used for calls through
108 /// function pointers in the 64-bit SVR4 ABI.
110 /// OPRC, CHAIN = DYNALLOC(CHAIN, NEGSIZE, FRAME_INDEX)
111 /// This instruction is lowered in PPCRegisterInfo::eliminateFrameIndex to
112 /// compute an allocation on the stack.
115 /// This instruction is lowered in PPCRegisterInfo::eliminateFrameIndex to
116 /// compute an offset from native SP to the address of the most recent
120 /// GlobalBaseReg - On Darwin, this node represents the result of the mflr
121 /// at function entry, used for PIC code.
124 /// These nodes represent PPC shifts.
126 /// For scalar types, only the last `n + 1` bits of the shift amounts
127 /// are used, where n is log2(sizeof(element) * 8). See sld/slw, etc.
128 /// for exact behaviors.
130 /// For vector types, only the last n bits are used. See vsld.
133 /// The combination of sra[wd]i and addze used to implemented signed
134 /// integer division by a power of 2. The first operand is the dividend,
135 /// and the second is the constant shift amount (representing the
139 /// CALL - A direct function call.
140 /// CALL_NOP is a call with the special NOP which follows 64-bit
144 /// CHAIN,FLAG = MTCTR(VAL, CHAIN[, INFLAG]) - Directly corresponds to a
145 /// MTCTR instruction.
148 /// CHAIN,FLAG = BCTRL(CHAIN, INFLAG) - Directly corresponds to a
149 /// BCTRL instruction.
152 /// CHAIN,FLAG = BCTRL(CHAIN, ADDR, INFLAG) - The combination of a bctrl
153 /// instruction and the TOC reload required on SVR4 PPC64.
156 /// Return with a flag operand, matched by 'blr'
159 /// R32 = MFOCRF(CRREG, INFLAG) - Represents the MFOCRF instruction.
160 /// This copies the bits corresponding to the specified CRREG into the
161 /// resultant GPR. Bits corresponding to other CR regs are undefined.
164 /// Direct move from a VSX register to a GPR
167 /// Direct move from a GPR to a VSX register (algebraic)
170 /// Direct move from a GPR to a VSX register (zero)
173 /// Extract a subvector from signed integer vector and convert to FP.
174 /// It is primarily used to convert a (widened) illegal integer vector
175 /// type to a legal floating point vector type.
176 /// For example v2i32 -> widened to v4i32 -> v2f64
179 /// Extract a subvector from unsigned integer vector and convert to FP.
180 /// As with SINT_VEC_TO_FP, used for converting illegal types.
183 // FIXME: Remove these once the ANDI glue bug is fixed:
184 /// i1 = ANDIo_1_[EQ|GT]_BIT(i32 or i64 x) - Represents the result of the
185 /// eq or gt bit of CR0 after executing andi. x, 1. This is used to
186 /// implement truncation of i32 or i64 to i1.
187 ANDIo_1_EQ_BIT, ANDIo_1_GT_BIT,
189 // READ_TIME_BASE - A read of the 64-bit time-base register on a 32-bit
190 // target (returns (Lo, Hi)). It takes a chain operand.
193 // EH_SJLJ_SETJMP - SjLj exception handling setjmp.
196 // EH_SJLJ_LONGJMP - SjLj exception handling longjmp.
199 /// RESVEC = VCMP(LHS, RHS, OPC) - Represents one of the altivec VCMP*
200 /// instructions. For lack of better number, we use the opcode number
201 /// encoding for the OPC field to identify the compare. For example, 838
205 /// RESVEC, OUTFLAG = VCMPo(LHS, RHS, OPC) - Represents one of the
206 /// altivec VCMP*o instructions. For lack of better number, we use the
207 /// opcode number encoding for the OPC field to identify the compare. For
208 /// example, 838 is VCMPGTSH.
211 /// CHAIN = COND_BRANCH CHAIN, CRRC, OPC, DESTBB [, INFLAG] - This
212 /// corresponds to the COND_BRANCH pseudo instruction. CRRC is the
213 /// condition register to branch on, OPC is the branch opcode to use (e.g.
214 /// PPC::BLE), DESTBB is the destination block to branch to, and INFLAG is
215 /// an optional input flag argument.
218 /// CHAIN = BDNZ CHAIN, DESTBB - These are used to create counter-based
222 /// F8RC = FADDRTZ F8RC, F8RC - This is an FADD done with rounding
223 /// towards zero. Used only as part of the long double-to-int
224 /// conversion sequence.
227 /// F8RC = MFFS - This moves the FPSCR (not modeled) into the register.
230 /// TC_RETURN - A tail call return.
232 /// operand #1 callee (register or absolute)
233 /// operand #2 stack adjustment
234 /// operand #3 optional in flag
237 /// ch, gl = CR6[UN]SET ch, inglue - Toggle CR bit 6 for SVR4 vararg calls
241 /// GPRC = address of _GLOBAL_OFFSET_TABLE_. Used by initial-exec TLS
245 /// GPRC = address of _GLOBAL_OFFSET_TABLE_. Used by general dynamic and
246 /// local dynamic TLS on PPC32.
249 /// G8RC = ADDIS_GOT_TPREL_HA %X2, Symbol - Used by the initial-exec
250 /// TLS model, produces an ADDIS8 instruction that adds the GOT
251 /// base to sym\@got\@tprel\@ha.
254 /// G8RC = LD_GOT_TPREL_L Symbol, G8RReg - Used by the initial-exec
255 /// TLS model, produces a LD instruction with base register G8RReg
256 /// and offset sym\@got\@tprel\@l. This completes the addition that
257 /// finds the offset of "sym" relative to the thread pointer.
260 /// G8RC = ADD_TLS G8RReg, Symbol - Used by the initial-exec TLS
261 /// model, produces an ADD instruction that adds the contents of
262 /// G8RReg to the thread pointer. Symbol contains a relocation
263 /// sym\@tls which is to be replaced by the thread pointer and
264 /// identifies to the linker that the instruction is part of a
268 /// G8RC = ADDIS_TLSGD_HA %X2, Symbol - For the general-dynamic TLS
269 /// model, produces an ADDIS8 instruction that adds the GOT base
270 /// register to sym\@got\@tlsgd\@ha.
273 /// %X3 = ADDI_TLSGD_L G8RReg, Symbol - For the general-dynamic TLS
274 /// model, produces an ADDI8 instruction that adds G8RReg to
275 /// sym\@got\@tlsgd\@l and stores the result in X3. Hidden by
276 /// ADDIS_TLSGD_L_ADDR until after register assignment.
279 /// %X3 = GET_TLS_ADDR %X3, Symbol - For the general-dynamic TLS
280 /// model, produces a call to __tls_get_addr(sym\@tlsgd). Hidden by
281 /// ADDIS_TLSGD_L_ADDR until after register assignment.
284 /// G8RC = ADDI_TLSGD_L_ADDR G8RReg, Symbol, Symbol - Op that
285 /// combines ADDI_TLSGD_L and GET_TLS_ADDR until expansion following
286 /// register assignment.
289 /// G8RC = ADDIS_TLSLD_HA %X2, Symbol - For the local-dynamic TLS
290 /// model, produces an ADDIS8 instruction that adds the GOT base
291 /// register to sym\@got\@tlsld\@ha.
294 /// %X3 = ADDI_TLSLD_L G8RReg, Symbol - For the local-dynamic TLS
295 /// model, produces an ADDI8 instruction that adds G8RReg to
296 /// sym\@got\@tlsld\@l and stores the result in X3. Hidden by
297 /// ADDIS_TLSLD_L_ADDR until after register assignment.
300 /// %X3 = GET_TLSLD_ADDR %X3, Symbol - For the local-dynamic TLS
301 /// model, produces a call to __tls_get_addr(sym\@tlsld). Hidden by
302 /// ADDIS_TLSLD_L_ADDR until after register assignment.
305 /// G8RC = ADDI_TLSLD_L_ADDR G8RReg, Symbol, Symbol - Op that
306 /// combines ADDI_TLSLD_L and GET_TLSLD_ADDR until expansion
307 /// following register assignment.
310 /// G8RC = ADDIS_DTPREL_HA %X3, Symbol - For the local-dynamic TLS
311 /// model, produces an ADDIS8 instruction that adds X3 to
315 /// G8RC = ADDI_DTPREL_L G8RReg, Symbol - For the local-dynamic TLS
316 /// model, produces an ADDI8 instruction that adds G8RReg to
317 /// sym\@got\@dtprel\@l.
320 /// VRRC = VADD_SPLAT Elt, EltSize - Temporary node to be expanded
321 /// during instruction selection to optimize a BUILD_VECTOR into
322 /// operations on splats. This is necessary to avoid losing these
323 /// optimizations due to constant folding.
326 /// CHAIN = SC CHAIN, Imm128 - System call. The 7-bit unsigned
327 /// operand identifies the operating system entry point.
330 /// CHAIN = CLRBHRB CHAIN - Clear branch history rolling buffer.
333 /// GPRC, CHAIN = MFBHRBE CHAIN, Entry, Dummy - Move from branch
334 /// history rolling buffer entry.
337 /// CHAIN = RFEBB CHAIN, State - Return from event-based branch.
340 /// VSRC, CHAIN = XXSWAPD CHAIN, VSRC - Occurs only for little
341 /// endian. Maps to an xxswapd instruction that corrects an lxvd2x
342 /// or stxvd2x instruction. The chain is necessary because the
343 /// sequence replaces a load and needs to provide the same number
347 /// An SDNode for swaps that are not associated with any loads/stores
348 /// and thereby have no chain.
351 /// QVFPERM = This corresponds to the QPX qvfperm instruction.
354 /// QVGPCI = This corresponds to the QPX qvgpci instruction.
357 /// QVALIGNI = This corresponds to the QPX qvaligni instruction.
360 /// QVESPLATI = This corresponds to the QPX qvesplati instruction.
363 /// QBFLT = Access the underlying QPX floating-point boolean
367 /// CHAIN = STBRX CHAIN, GPRC, Ptr, Type - This is a
368 /// byte-swapping store instruction. It byte-swaps the low "Type" bits of
369 /// the GPRC input, then stores it through Ptr. Type can be either i16 or
371 STBRX = ISD::FIRST_TARGET_MEMORY_OPCODE,
373 /// GPRC, CHAIN = LBRX CHAIN, Ptr, Type - This is a
374 /// byte-swapping load instruction. It loads "Type" bits, byte swaps it,
375 /// then puts it in the bottom bits of the GPRC. TYPE can be either i16
379 /// STFIWX - The STFIWX instruction. The first operand is an input token
380 /// chain, then an f64 value to store, then an address to store it to.
383 /// GPRC, CHAIN = LFIWAX CHAIN, Ptr - This is a floating-point
384 /// load which sign-extends from a 32-bit integer value into the
385 /// destination 64-bit register.
388 /// GPRC, CHAIN = LFIWZX CHAIN, Ptr - This is a floating-point
389 /// load which zero-extends from a 32-bit integer value into the
390 /// destination 64-bit register.
393 /// GPRC, CHAIN = LXSIZX, CHAIN, Ptr, ByteWidth - This is a load of an
394 /// integer smaller than 64 bits into a VSR. The integer is zero-extended.
395 /// This can be used for converting loaded integers to floating point.
398 /// STXSIX - The STXSI[bh]X instruction. The first operand is an input
399 /// chain, then an f64 value to store, then an address to store it to,
400 /// followed by a byte-width for the store.
403 /// VSRC, CHAIN = LXVD2X_LE CHAIN, Ptr - Occurs only for little endian.
404 /// Maps directly to an lxvd2x instruction that will be followed by
408 /// CHAIN = STXVD2X CHAIN, VSRC, Ptr - Occurs only for little endian.
409 /// Maps directly to an stxvd2x instruction that will be preceded by
413 /// QBRC, CHAIN = QVLFSb CHAIN, Ptr
414 /// The 4xf32 load used for v4i1 constants.
417 /// GPRC = TOC_ENTRY GA, TOC
418 /// Loads the entry for GA from the TOC, where the TOC base is given by
419 /// the last operand.
423 } // end namespace PPCISD
425 /// Define some predicates that are used for node matching.
428 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
429 /// VPKUHUM instruction.
430 bool isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
433 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
434 /// VPKUWUM instruction.
435 bool isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
438 /// isVPKUDUMShuffleMask - Return true if this is the shuffle mask for a
439 /// VPKUDUM instruction.
440 bool isVPKUDUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
443 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
444 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
445 bool isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
446 unsigned ShuffleKind, SelectionDAG &DAG);
448 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
449 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
450 bool isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
451 unsigned ShuffleKind, SelectionDAG &DAG);
453 /// isVMRGEOShuffleMask - Return true if this is a shuffle mask suitable for
454 /// a VMRGEW or VMRGOW instruction
455 bool isVMRGEOShuffleMask(ShuffleVectorSDNode *N, bool CheckEven,
456 unsigned ShuffleKind, SelectionDAG &DAG);
457 /// isXXSLDWIShuffleMask - Return true if this is a shuffle mask suitable
458 /// for a XXSLDWI instruction.
459 bool isXXSLDWIShuffleMask(ShuffleVectorSDNode *N, unsigned &ShiftElts,
460 bool &Swap, bool IsLE);
461 /// isXXPERMDIShuffleMask - Return true if this is a shuffle mask suitable
462 /// for a XXPERMDI instruction.
463 bool isXXPERMDIShuffleMask(ShuffleVectorSDNode *N, unsigned &ShiftElts,
464 bool &Swap, bool IsLE);
466 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the
467 /// shift amount, otherwise return -1.
468 int isVSLDOIShuffleMask(SDNode *N, unsigned ShuffleKind,
471 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
472 /// specifies a splat of a single element that is suitable for input to
473 /// VSPLTB/VSPLTH/VSPLTW.
474 bool isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize);
476 /// isXXINSERTWMask - Return true if this VECTOR_SHUFFLE can be handled by
477 /// the XXINSERTW instruction introduced in ISA 3.0. This is essentially any
478 /// shuffle of v4f32/v4i32 vectors that just inserts one element from one
479 /// vector into the other. This function will also set a couple of
480 /// output parameters for how much the source vector needs to be shifted and
481 /// what byte number needs to be specified for the instruction to put the
482 /// element in the desired location of the target vector.
483 bool isXXINSERTWMask(ShuffleVectorSDNode *N, unsigned &ShiftElts,
484 unsigned &InsertAtByte, bool &Swap, bool IsLE);
486 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
487 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
488 unsigned getVSPLTImmediate(SDNode *N, unsigned EltSize, SelectionDAG &DAG);
490 /// get_VSPLTI_elt - If this is a build_vector of constants which can be
491 /// formed by using a vspltis[bhw] instruction of the specified element
492 /// size, return the constant being splatted. The ByteSize field indicates
493 /// the number of bytes of each element [124] -> [bhw].
494 SDValue get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG);
496 /// If this is a qvaligni shuffle mask, return the shift
497 /// amount, otherwise return -1.
498 int isQVALIGNIShuffleMask(SDNode *N);
500 } // end namespace PPC
502 class PPCTargetLowering : public TargetLowering {
503 const PPCSubtarget &Subtarget;
506 explicit PPCTargetLowering(const PPCTargetMachine &TM,
507 const PPCSubtarget &STI);
509 /// getTargetNodeName() - This method returns the name of a target specific
511 const char *getTargetNodeName(unsigned Opcode) const override;
513 /// getPreferredVectorAction - The code we generate when vector types are
514 /// legalized by promoting the integer element type is often much worse
515 /// than code we generate if we widen the type for applicable vector types.
516 /// The issue with promoting is that the vector is scalaraized, individual
517 /// elements promoted and then the vector is rebuilt. So say we load a pair
518 /// of v4i8's and shuffle them. This will turn into a mess of 8 extending
519 /// loads, moves back into VSR's (or memory ops if we don't have moves) and
520 /// then the VPERM for the shuffle. All in all a very slow sequence.
521 TargetLoweringBase::LegalizeTypeAction getPreferredVectorAction(EVT VT)
523 if (VT.getScalarSizeInBits() % 8 == 0)
524 return TypeWidenVector;
525 return TargetLoweringBase::getPreferredVectorAction(VT);
528 bool useSoftFloat() const override;
530 MVT getScalarShiftAmountTy(const DataLayout &, EVT) const override {
534 bool isCheapToSpeculateCttz() const override {
538 bool isCheapToSpeculateCtlz() const override {
542 bool isCtlzFast() const override {
546 bool hasAndNotCompare(SDValue) const override {
550 bool convertSetCCLogicToBitwiseLogic(EVT VT) const override {
551 return VT.isScalarInteger();
554 bool supportSplitCSR(MachineFunction *MF) const override {
556 MF->getFunction()->getCallingConv() == CallingConv::CXX_FAST_TLS &&
557 MF->getFunction()->hasFnAttribute(Attribute::NoUnwind);
560 void initializeSplitCSR(MachineBasicBlock *Entry) const override;
562 void insertCopiesSplitCSR(
563 MachineBasicBlock *Entry,
564 const SmallVectorImpl<MachineBasicBlock *> &Exits) const override;
566 /// getSetCCResultType - Return the ISD::SETCC ValueType
567 EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context,
568 EVT VT) const override;
570 /// Return true if target always beneficiates from combining into FMA for a
571 /// given value type. This must typically return false on targets where FMA
572 /// takes more cycles to execute than FADD.
573 bool enableAggressiveFMAFusion(EVT VT) const override;
575 /// getPreIndexedAddressParts - returns true by value, base pointer and
576 /// offset pointer and addressing mode by reference if the node's address
577 /// can be legally represented as pre-indexed load / store address.
578 bool getPreIndexedAddressParts(SDNode *N, SDValue &Base,
580 ISD::MemIndexedMode &AM,
581 SelectionDAG &DAG) const override;
583 /// SelectAddressRegReg - Given the specified addressed, check to see if it
584 /// can be represented as an indexed [r+r] operation. Returns false if it
585 /// can be more efficiently represented with [r+imm].
586 bool SelectAddressRegReg(SDValue N, SDValue &Base, SDValue &Index,
587 SelectionDAG &DAG) const;
589 /// SelectAddressRegImm - Returns true if the address N can be represented
590 /// by a base register plus a signed 16-bit displacement [r+imm], and if it
591 /// is not better represented as reg+reg. If Aligned is true, only accept
592 /// displacements suitable for STD and friends, i.e. multiples of 4.
593 bool SelectAddressRegImm(SDValue N, SDValue &Disp, SDValue &Base,
594 SelectionDAG &DAG, bool Aligned) const;
596 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
597 /// represented as an indexed [r+r] operation.
598 bool SelectAddressRegRegOnly(SDValue N, SDValue &Base, SDValue &Index,
599 SelectionDAG &DAG) const;
601 Sched::Preference getSchedulingPreference(SDNode *N) const override;
603 /// LowerOperation - Provide custom lowering hooks for some operations.
605 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
607 /// ReplaceNodeResults - Replace the results of node with an illegal result
608 /// type with new values built out of custom code.
610 void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
611 SelectionDAG &DAG) const override;
613 SDValue expandVSXLoadForLE(SDNode *N, DAGCombinerInfo &DCI) const;
614 SDValue expandVSXStoreForLE(SDNode *N, DAGCombinerInfo &DCI) const;
616 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
618 SDValue BuildSDIVPow2(SDNode *N, const APInt &Divisor, SelectionDAG &DAG,
619 std::vector<SDNode *> *Created) const override;
621 unsigned getRegisterByName(const char* RegName, EVT VT,
622 SelectionDAG &DAG) const override;
624 void computeKnownBitsForTargetNode(const SDValue Op,
626 const APInt &DemandedElts,
627 const SelectionDAG &DAG,
628 unsigned Depth = 0) const override;
630 unsigned getPrefLoopAlignment(MachineLoop *ML) const override;
632 bool shouldInsertFencesForAtomic(const Instruction *I) const override {
636 Instruction *emitLeadingFence(IRBuilder<> &Builder, Instruction *Inst,
637 AtomicOrdering Ord) const override;
638 Instruction *emitTrailingFence(IRBuilder<> &Builder, Instruction *Inst,
639 AtomicOrdering Ord) const override;
642 EmitInstrWithCustomInserter(MachineInstr &MI,
643 MachineBasicBlock *MBB) const override;
644 MachineBasicBlock *EmitAtomicBinary(MachineInstr &MI,
645 MachineBasicBlock *MBB,
648 unsigned CmpOpcode = 0,
649 unsigned CmpPred = 0) const;
650 MachineBasicBlock *EmitPartwordAtomicBinary(MachineInstr &MI,
651 MachineBasicBlock *MBB,
654 unsigned CmpOpcode = 0,
655 unsigned CmpPred = 0) const;
657 MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr &MI,
658 MachineBasicBlock *MBB) const;
660 MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr &MI,
661 MachineBasicBlock *MBB) const;
663 ConstraintType getConstraintType(StringRef Constraint) const override;
665 /// Examine constraint string and operand type and determine a weight value.
666 /// The operand object must already have been set up with the operand type.
667 ConstraintWeight getSingleConstraintMatchWeight(
668 AsmOperandInfo &info, const char *constraint) const override;
670 std::pair<unsigned, const TargetRegisterClass *>
671 getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
672 StringRef Constraint, MVT VT) const override;
674 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
675 /// function arguments in the caller parameter area. This is the actual
676 /// alignment, not its logarithm.
677 unsigned getByValTypeAlignment(Type *Ty,
678 const DataLayout &DL) const override;
680 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
681 /// vector. If it is invalid, don't add anything to Ops.
682 void LowerAsmOperandForConstraint(SDValue Op,
683 std::string &Constraint,
684 std::vector<SDValue> &Ops,
685 SelectionDAG &DAG) const override;
688 getInlineAsmMemConstraint(StringRef ConstraintCode) const override {
689 if (ConstraintCode == "es")
690 return InlineAsm::Constraint_es;
691 else if (ConstraintCode == "o")
692 return InlineAsm::Constraint_o;
693 else if (ConstraintCode == "Q")
694 return InlineAsm::Constraint_Q;
695 else if (ConstraintCode == "Z")
696 return InlineAsm::Constraint_Z;
697 else if (ConstraintCode == "Zy")
698 return InlineAsm::Constraint_Zy;
699 return TargetLowering::getInlineAsmMemConstraint(ConstraintCode);
702 /// isLegalAddressingMode - Return true if the addressing mode represented
703 /// by AM is legal for this target, for a load/store of the specified type.
704 bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM,
705 Type *Ty, unsigned AS) const override;
707 /// isLegalICmpImmediate - Return true if the specified immediate is legal
708 /// icmp immediate, that is the target has icmp instructions which can
709 /// compare a register against the immediate without having to materialize
710 /// the immediate into a register.
711 bool isLegalICmpImmediate(int64_t Imm) const override;
713 /// isLegalAddImmediate - Return true if the specified immediate is legal
714 /// add immediate, that is the target has add instructions which can
715 /// add a register and the immediate without having to materialize
716 /// the immediate into a register.
717 bool isLegalAddImmediate(int64_t Imm) const override;
719 /// isTruncateFree - Return true if it's free to truncate a value of
720 /// type Ty1 to type Ty2. e.g. On PPC it's free to truncate a i64 value in
721 /// register X1 to i32 by referencing its sub-register R1.
722 bool isTruncateFree(Type *Ty1, Type *Ty2) const override;
723 bool isTruncateFree(EVT VT1, EVT VT2) const override;
725 bool isZExtFree(SDValue Val, EVT VT2) const override;
727 bool isFPExtFree(EVT VT) const override;
729 /// \brief Returns true if it is beneficial to convert a load of a constant
730 /// to just the constant itself.
731 bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
732 Type *Ty) const override;
734 bool convertSelectOfConstantsToMath() const override {
738 bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const override;
740 bool getTgtMemIntrinsic(IntrinsicInfo &Info,
742 unsigned Intrinsic) const override;
744 /// getOptimalMemOpType - Returns the target specific optimal type for load
745 /// and store operations as a result of memset, memcpy, and memmove
746 /// lowering. If DstAlign is zero that means it's safe to destination
747 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
748 /// means there isn't a need to check it against alignment requirement,
749 /// probably because the source does not need to be loaded. If 'IsMemset' is
750 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
751 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
752 /// source is constant so it does not need to be loaded.
753 /// It returns EVT::Other if the type should be determined using generic
754 /// target-independent logic.
756 getOptimalMemOpType(uint64_t Size, unsigned DstAlign, unsigned SrcAlign,
757 bool IsMemset, bool ZeroMemset, bool MemcpyStrSrc,
758 MachineFunction &MF) const override;
760 /// Is unaligned memory access allowed for the given type, and is it fast
761 /// relative to software emulation.
762 bool allowsMisalignedMemoryAccesses(EVT VT,
765 bool *Fast = nullptr) const override;
767 /// isFMAFasterThanFMulAndFAdd - Return true if an FMA operation is faster
768 /// than a pair of fmul and fadd instructions. fmuladd intrinsics will be
769 /// expanded to FMAs when this method returns true, otherwise fmuladd is
770 /// expanded to fmul + fadd.
771 bool isFMAFasterThanFMulAndFAdd(EVT VT) const override;
773 const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const override;
775 // Should we expand the build vector with shuffles?
777 shouldExpandBuildVectorWithShuffles(EVT VT,
778 unsigned DefinedValues) const override;
780 /// createFastISel - This method returns a target-specific FastISel object,
781 /// or null if the target does not support "fast" instruction selection.
782 FastISel *createFastISel(FunctionLoweringInfo &FuncInfo,
783 const TargetLibraryInfo *LibInfo) const override;
785 /// \brief Returns true if an argument of type Ty needs to be passed in a
786 /// contiguous block of registers in calling convention CallConv.
787 bool functionArgumentNeedsConsecutiveRegisters(
788 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const override {
789 // We support any array type as "consecutive" block in the parameter
790 // save area. The element type defines the alignment requirement and
791 // whether the argument should go in GPRs, FPRs, or VRs if available.
793 // Note that clang uses this capability both to implement the ELFv2
794 // homogeneous float/vector aggregate ABI, and to avoid having to use
795 // "byval" when passing aggregates that might fully fit in registers.
796 return Ty->isArrayTy();
799 /// If a physical register, this returns the register that receives the
800 /// exception address on entry to an EH pad.
802 getExceptionPointerRegister(const Constant *PersonalityFn) const override;
804 /// If a physical register, this returns the register that receives the
805 /// exception typeid on entry to a landing pad.
807 getExceptionSelectorRegister(const Constant *PersonalityFn) const override;
809 /// Override to support customized stack guard loading.
810 bool useLoadStackGuardNode() const override;
811 void insertSSPDeclarations(Module &M) const override;
813 bool isFPImmLegal(const APFloat &Imm, EVT VT) const override;
815 unsigned getJumpTableEncoding() const override;
816 bool isJumpTableRelative() const override;
817 SDValue getPICJumpTableRelocBase(SDValue Table,
818 SelectionDAG &DAG) const override;
819 const MCExpr *getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
821 MCContext &Ctx) const override;
824 struct ReuseLoadInfo {
828 MachinePointerInfo MPI;
829 bool IsDereferenceable = false;
830 bool IsInvariant = false;
831 unsigned Alignment = 0;
833 const MDNode *Ranges = nullptr;
835 ReuseLoadInfo() = default;
837 MachineMemOperand::Flags MMOFlags() const {
838 MachineMemOperand::Flags F = MachineMemOperand::MONone;
839 if (IsDereferenceable)
840 F |= MachineMemOperand::MODereferenceable;
842 F |= MachineMemOperand::MOInvariant;
847 bool canReuseLoadAddress(SDValue Op, EVT MemVT, ReuseLoadInfo &RLI,
849 ISD::LoadExtType ET = ISD::NON_EXTLOAD) const;
850 void spliceIntoChain(SDValue ResChain, SDValue NewResChain,
851 SelectionDAG &DAG) const;
853 void LowerFP_TO_INTForReuse(SDValue Op, ReuseLoadInfo &RLI,
854 SelectionDAG &DAG, const SDLoc &dl) const;
855 SDValue LowerFP_TO_INTDirectMove(SDValue Op, SelectionDAG &DAG,
856 const SDLoc &dl) const;
858 bool directMoveIsProfitable(const SDValue &Op) const;
859 SDValue LowerINT_TO_FPDirectMove(SDValue Op, SelectionDAG &DAG,
860 const SDLoc &dl) const;
862 SDValue getFramePointerFrameIndex(SelectionDAG & DAG) const;
863 SDValue getReturnAddrFrameIndex(SelectionDAG & DAG) const;
866 IsEligibleForTailCallOptimization(SDValue Callee,
867 CallingConv::ID CalleeCC,
869 const SmallVectorImpl<ISD::InputArg> &Ins,
870 SelectionDAG& DAG) const;
873 IsEligibleForTailCallOptimization_64SVR4(
875 CallingConv::ID CalleeCC,
876 ImmutableCallSite *CS,
878 const SmallVectorImpl<ISD::OutputArg> &Outs,
879 const SmallVectorImpl<ISD::InputArg> &Ins,
880 SelectionDAG& DAG) const;
882 SDValue EmitTailCallLoadFPAndRetAddr(SelectionDAG &DAG, int SPDiff,
883 SDValue Chain, SDValue &LROpOut,
885 const SDLoc &dl) const;
887 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
888 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
889 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
890 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
891 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
892 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
893 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
894 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
895 SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
896 SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
897 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
898 SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const;
899 SDValue LowerVACOPY(SDValue Op, SelectionDAG &DAG) const;
900 SDValue LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG) const;
901 SDValue LowerGET_DYNAMIC_AREA_OFFSET(SDValue Op, SelectionDAG &DAG) const;
902 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
903 SDValue LowerEH_DWARF_CFA(SDValue Op, SelectionDAG &DAG) const;
904 SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const;
905 SDValue LowerSTORE(SDValue Op, SelectionDAG &DAG) const;
906 SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const;
907 SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
908 SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
909 const SDLoc &dl) const;
910 SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
911 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
912 SDValue LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const;
913 SDValue LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const;
914 SDValue LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const;
915 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
916 SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
917 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
918 SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
919 SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const;
920 SDValue LowerINTRINSIC_VOID(SDValue Op, SelectionDAG &DAG) const;
921 SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const;
922 SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const;
923 SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) const;
925 SDValue LowerVectorLoad(SDValue Op, SelectionDAG &DAG) const;
926 SDValue LowerVectorStore(SDValue Op, SelectionDAG &DAG) const;
928 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
929 CallingConv::ID CallConv, bool isVarArg,
930 const SmallVectorImpl<ISD::InputArg> &Ins,
931 const SDLoc &dl, SelectionDAG &DAG,
932 SmallVectorImpl<SDValue> &InVals) const;
933 SDValue FinishCall(CallingConv::ID CallConv, const SDLoc &dl,
934 bool isTailCall, bool isVarArg, bool isPatchPoint,
935 bool hasNest, SelectionDAG &DAG,
936 SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
937 SDValue InFlag, SDValue Chain, SDValue CallSeqStart,
938 SDValue &Callee, int SPDiff, unsigned NumBytes,
939 const SmallVectorImpl<ISD::InputArg> &Ins,
940 SmallVectorImpl<SDValue> &InVals,
941 ImmutableCallSite *CS) const;
944 LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
945 const SmallVectorImpl<ISD::InputArg> &Ins,
946 const SDLoc &dl, SelectionDAG &DAG,
947 SmallVectorImpl<SDValue> &InVals) const override;
949 SDValue LowerCall(TargetLowering::CallLoweringInfo &CLI,
950 SmallVectorImpl<SDValue> &InVals) const override;
952 bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
954 const SmallVectorImpl<ISD::OutputArg> &Outs,
955 LLVMContext &Context) const override;
957 SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
958 const SmallVectorImpl<ISD::OutputArg> &Outs,
959 const SmallVectorImpl<SDValue> &OutVals,
960 const SDLoc &dl, SelectionDAG &DAG) const override;
962 SDValue extendArgForPPC64(ISD::ArgFlagsTy Flags, EVT ObjectVT,
963 SelectionDAG &DAG, SDValue ArgVal,
964 const SDLoc &dl) const;
966 SDValue LowerFormalArguments_Darwin(
967 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
968 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
969 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const;
970 SDValue LowerFormalArguments_64SVR4(
971 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
972 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
973 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const;
974 SDValue LowerFormalArguments_32SVR4(
975 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
976 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
977 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const;
979 SDValue createMemcpyOutsideCallSeq(SDValue Arg, SDValue PtrOff,
980 SDValue CallSeqStart,
981 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
982 const SDLoc &dl) const;
984 SDValue LowerCall_Darwin(SDValue Chain, SDValue Callee,
985 CallingConv::ID CallConv, bool isVarArg,
986 bool isTailCall, bool isPatchPoint,
987 const SmallVectorImpl<ISD::OutputArg> &Outs,
988 const SmallVectorImpl<SDValue> &OutVals,
989 const SmallVectorImpl<ISD::InputArg> &Ins,
990 const SDLoc &dl, SelectionDAG &DAG,
991 SmallVectorImpl<SDValue> &InVals,
992 ImmutableCallSite *CS) const;
993 SDValue LowerCall_64SVR4(SDValue Chain, SDValue Callee,
994 CallingConv::ID CallConv, bool isVarArg,
995 bool isTailCall, bool isPatchPoint,
996 const SmallVectorImpl<ISD::OutputArg> &Outs,
997 const SmallVectorImpl<SDValue> &OutVals,
998 const SmallVectorImpl<ISD::InputArg> &Ins,
999 const SDLoc &dl, SelectionDAG &DAG,
1000 SmallVectorImpl<SDValue> &InVals,
1001 ImmutableCallSite *CS) const;
1002 SDValue LowerCall_32SVR4(SDValue Chain, SDValue Callee,
1003 CallingConv::ID CallConv, bool isVarArg,
1004 bool isTailCall, bool isPatchPoint,
1005 const SmallVectorImpl<ISD::OutputArg> &Outs,
1006 const SmallVectorImpl<SDValue> &OutVals,
1007 const SmallVectorImpl<ISD::InputArg> &Ins,
1008 const SDLoc &dl, SelectionDAG &DAG,
1009 SmallVectorImpl<SDValue> &InVals,
1010 ImmutableCallSite *CS) const;
1012 SDValue lowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
1013 SDValue lowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
1015 SDValue DAGCombineExtBoolTrunc(SDNode *N, DAGCombinerInfo &DCI) const;
1016 SDValue DAGCombineBuildVector(SDNode *N, DAGCombinerInfo &DCI) const;
1017 SDValue DAGCombineTruncBoolExt(SDNode *N, DAGCombinerInfo &DCI) const;
1018 SDValue combineFPToIntToFP(SDNode *N, DAGCombinerInfo &DCI) const;
1019 SDValue combineSHL(SDNode *N, DAGCombinerInfo &DCI) const;
1020 SDValue combineSRA(SDNode *N, DAGCombinerInfo &DCI) const;
1021 SDValue combineSRL(SDNode *N, DAGCombinerInfo &DCI) const;
1023 /// ConvertSETCCToSubtract - looks at SETCC that compares ints. It replaces
1024 /// SETCC with integer subtraction when (1) there is a legal way of doing it
1025 /// (2) keeping the result of comparison in GPR has performance benefit.
1026 SDValue ConvertSETCCToSubtract(SDNode *N, DAGCombinerInfo &DCI) const;
1028 SDValue getSqrtEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
1029 int &RefinementSteps, bool &UseOneConstNR,
1030 bool Reciprocal) const override;
1031 SDValue getRecipEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
1032 int &RefinementSteps) const override;
1033 unsigned combineRepeatedFPDivisors() const override;
1035 CCAssignFn *useFastISelCCs(unsigned Flag) const;
1038 combineElementTruncationToVectorTruncation(SDNode *N,
1039 DAGCombinerInfo &DCI) const;
1044 FastISel *createFastISel(FunctionLoweringInfo &FuncInfo,
1045 const TargetLibraryInfo *LibInfo);
1047 } // end namespace PPC
1049 bool CC_PPC32_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
1050 CCValAssign::LocInfo &LocInfo,
1051 ISD::ArgFlagsTy &ArgFlags,
1054 bool CC_PPC32_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
1056 CCValAssign::LocInfo &LocInfo,
1057 ISD::ArgFlagsTy &ArgFlags,
1061 CC_PPC32_SVR4_Custom_SkipLastArgRegsPPCF128(unsigned &ValNo, MVT &ValVT,
1063 CCValAssign::LocInfo &LocInfo,
1064 ISD::ArgFlagsTy &ArgFlags,
1067 bool CC_PPC32_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
1069 CCValAssign::LocInfo &LocInfo,
1070 ISD::ArgFlagsTy &ArgFlags,
1073 } // end namespace llvm
1075 #endif // LLVM_TARGET_POWERPC_PPC32ISELLOWERING_H