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 /// SExtVElems, takes an input vector of a smaller type and sign
71 /// extends to an output vector of a larger type.
74 /// Reciprocal estimate instructions (unary FP ops).
77 // VMADDFP, VNMSUBFP - The VMADDFP and VNMSUBFP instructions, taking
78 // three v4f32 operands and producing a v4f32 result.
81 /// VPERM - The PPC VPERM Instruction.
85 /// XXSPLT - The PPC VSX splat instructions
89 /// XXINSERT - The PPC VSX insert instruction
93 /// XXREVERSE - The PPC VSX reverse instruction
97 /// VECSHL - The PPC VSX shift left instruction
101 /// XXPERMDI - The PPC XXPERMDI instruction
105 /// The CMPB instruction (takes two operands of i32 or i64).
108 /// Hi/Lo - These represent the high and low 16-bit parts of a global
109 /// address respectively. These nodes have two operands, the first of
110 /// which must be a TargetGlobalAddress, and the second of which must be a
111 /// Constant. Selected naively, these turn into 'lis G+C' and 'li G+C',
112 /// though these are usually folded into other nodes.
115 /// The following two target-specific nodes are used for calls through
116 /// function pointers in the 64-bit SVR4 ABI.
118 /// OPRC, CHAIN = DYNALLOC(CHAIN, NEGSIZE, FRAME_INDEX)
119 /// This instruction is lowered in PPCRegisterInfo::eliminateFrameIndex to
120 /// compute an allocation on the stack.
123 /// This instruction is lowered in PPCRegisterInfo::eliminateFrameIndex to
124 /// compute an offset from native SP to the address of the most recent
128 /// GlobalBaseReg - On Darwin, this node represents the result of the mflr
129 /// at function entry, used for PIC code.
132 /// These nodes represent PPC shifts.
134 /// For scalar types, only the last `n + 1` bits of the shift amounts
135 /// are used, where n is log2(sizeof(element) * 8). See sld/slw, etc.
136 /// for exact behaviors.
138 /// For vector types, only the last n bits are used. See vsld.
141 /// The combination of sra[wd]i and addze used to implemented signed
142 /// integer division by a power of 2. The first operand is the dividend,
143 /// and the second is the constant shift amount (representing the
147 /// CALL - A direct function call.
148 /// CALL_NOP is a call with the special NOP which follows 64-bit
152 /// CHAIN,FLAG = MTCTR(VAL, CHAIN[, INFLAG]) - Directly corresponds to a
153 /// MTCTR instruction.
156 /// CHAIN,FLAG = BCTRL(CHAIN, INFLAG) - Directly corresponds to a
157 /// BCTRL instruction.
160 /// CHAIN,FLAG = BCTRL(CHAIN, ADDR, INFLAG) - The combination of a bctrl
161 /// instruction and the TOC reload required on SVR4 PPC64.
164 /// Return with a flag operand, matched by 'blr'
167 /// R32 = MFOCRF(CRREG, INFLAG) - Represents the MFOCRF instruction.
168 /// This copies the bits corresponding to the specified CRREG into the
169 /// resultant GPR. Bits corresponding to other CR regs are undefined.
172 /// Direct move from a VSX register to a GPR
175 /// Direct move from a GPR to a VSX register (algebraic)
178 /// Direct move from a GPR to a VSX register (zero)
181 /// Extract a subvector from signed integer vector and convert to FP.
182 /// It is primarily used to convert a (widened) illegal integer vector
183 /// type to a legal floating point vector type.
184 /// For example v2i32 -> widened to v4i32 -> v2f64
187 /// Extract a subvector from unsigned integer vector and convert to FP.
188 /// As with SINT_VEC_TO_FP, used for converting illegal types.
191 // FIXME: Remove these once the ANDI glue bug is fixed:
192 /// i1 = ANDIo_1_[EQ|GT]_BIT(i32 or i64 x) - Represents the result of the
193 /// eq or gt bit of CR0 after executing andi. x, 1. This is used to
194 /// implement truncation of i32 or i64 to i1.
195 ANDIo_1_EQ_BIT, ANDIo_1_GT_BIT,
197 // READ_TIME_BASE - A read of the 64-bit time-base register on a 32-bit
198 // target (returns (Lo, Hi)). It takes a chain operand.
201 // EH_SJLJ_SETJMP - SjLj exception handling setjmp.
204 // EH_SJLJ_LONGJMP - SjLj exception handling longjmp.
207 /// RESVEC = VCMP(LHS, RHS, OPC) - Represents one of the altivec VCMP*
208 /// instructions. For lack of better number, we use the opcode number
209 /// encoding for the OPC field to identify the compare. For example, 838
213 /// RESVEC, OUTFLAG = VCMPo(LHS, RHS, OPC) - Represents one of the
214 /// altivec VCMP*o instructions. For lack of better number, we use the
215 /// opcode number encoding for the OPC field to identify the compare. For
216 /// example, 838 is VCMPGTSH.
219 /// CHAIN = COND_BRANCH CHAIN, CRRC, OPC, DESTBB [, INFLAG] - This
220 /// corresponds to the COND_BRANCH pseudo instruction. CRRC is the
221 /// condition register to branch on, OPC is the branch opcode to use (e.g.
222 /// PPC::BLE), DESTBB is the destination block to branch to, and INFLAG is
223 /// an optional input flag argument.
226 /// CHAIN = BDNZ CHAIN, DESTBB - These are used to create counter-based
230 /// F8RC = FADDRTZ F8RC, F8RC - This is an FADD done with rounding
231 /// towards zero. Used only as part of the long double-to-int
232 /// conversion sequence.
235 /// F8RC = MFFS - This moves the FPSCR (not modeled) into the register.
238 /// TC_RETURN - A tail call return.
240 /// operand #1 callee (register or absolute)
241 /// operand #2 stack adjustment
242 /// operand #3 optional in flag
245 /// ch, gl = CR6[UN]SET ch, inglue - Toggle CR bit 6 for SVR4 vararg calls
249 /// GPRC = address of _GLOBAL_OFFSET_TABLE_. Used by initial-exec TLS
253 /// GPRC = address of _GLOBAL_OFFSET_TABLE_. Used by general dynamic and
254 /// local dynamic TLS on PPC32.
257 /// G8RC = ADDIS_GOT_TPREL_HA %X2, Symbol - Used by the initial-exec
258 /// TLS model, produces an ADDIS8 instruction that adds the GOT
259 /// base to sym\@got\@tprel\@ha.
262 /// G8RC = LD_GOT_TPREL_L Symbol, G8RReg - Used by the initial-exec
263 /// TLS model, produces a LD instruction with base register G8RReg
264 /// and offset sym\@got\@tprel\@l. This completes the addition that
265 /// finds the offset of "sym" relative to the thread pointer.
268 /// G8RC = ADD_TLS G8RReg, Symbol - Used by the initial-exec TLS
269 /// model, produces an ADD instruction that adds the contents of
270 /// G8RReg to the thread pointer. Symbol contains a relocation
271 /// sym\@tls which is to be replaced by the thread pointer and
272 /// identifies to the linker that the instruction is part of a
276 /// G8RC = ADDIS_TLSGD_HA %X2, Symbol - For the general-dynamic TLS
277 /// model, produces an ADDIS8 instruction that adds the GOT base
278 /// register to sym\@got\@tlsgd\@ha.
281 /// %X3 = ADDI_TLSGD_L G8RReg, Symbol - For the general-dynamic TLS
282 /// model, produces an ADDI8 instruction that adds G8RReg to
283 /// sym\@got\@tlsgd\@l and stores the result in X3. Hidden by
284 /// ADDIS_TLSGD_L_ADDR until after register assignment.
287 /// %X3 = GET_TLS_ADDR %X3, Symbol - For the general-dynamic TLS
288 /// model, produces a call to __tls_get_addr(sym\@tlsgd). Hidden by
289 /// ADDIS_TLSGD_L_ADDR until after register assignment.
292 /// G8RC = ADDI_TLSGD_L_ADDR G8RReg, Symbol, Symbol - Op that
293 /// combines ADDI_TLSGD_L and GET_TLS_ADDR until expansion following
294 /// register assignment.
297 /// G8RC = ADDIS_TLSLD_HA %X2, Symbol - For the local-dynamic TLS
298 /// model, produces an ADDIS8 instruction that adds the GOT base
299 /// register to sym\@got\@tlsld\@ha.
302 /// %X3 = ADDI_TLSLD_L G8RReg, Symbol - For the local-dynamic TLS
303 /// model, produces an ADDI8 instruction that adds G8RReg to
304 /// sym\@got\@tlsld\@l and stores the result in X3. Hidden by
305 /// ADDIS_TLSLD_L_ADDR until after register assignment.
308 /// %X3 = GET_TLSLD_ADDR %X3, Symbol - For the local-dynamic TLS
309 /// model, produces a call to __tls_get_addr(sym\@tlsld). Hidden by
310 /// ADDIS_TLSLD_L_ADDR until after register assignment.
313 /// G8RC = ADDI_TLSLD_L_ADDR G8RReg, Symbol, Symbol - Op that
314 /// combines ADDI_TLSLD_L and GET_TLSLD_ADDR until expansion
315 /// following register assignment.
318 /// G8RC = ADDIS_DTPREL_HA %X3, Symbol - For the local-dynamic TLS
319 /// model, produces an ADDIS8 instruction that adds X3 to
323 /// G8RC = ADDI_DTPREL_L G8RReg, Symbol - For the local-dynamic TLS
324 /// model, produces an ADDI8 instruction that adds G8RReg to
325 /// sym\@got\@dtprel\@l.
328 /// VRRC = VADD_SPLAT Elt, EltSize - Temporary node to be expanded
329 /// during instruction selection to optimize a BUILD_VECTOR into
330 /// operations on splats. This is necessary to avoid losing these
331 /// optimizations due to constant folding.
334 /// CHAIN = SC CHAIN, Imm128 - System call. The 7-bit unsigned
335 /// operand identifies the operating system entry point.
338 /// CHAIN = CLRBHRB CHAIN - Clear branch history rolling buffer.
341 /// GPRC, CHAIN = MFBHRBE CHAIN, Entry, Dummy - Move from branch
342 /// history rolling buffer entry.
345 /// CHAIN = RFEBB CHAIN, State - Return from event-based branch.
348 /// VSRC, CHAIN = XXSWAPD CHAIN, VSRC - Occurs only for little
349 /// endian. Maps to an xxswapd instruction that corrects an lxvd2x
350 /// or stxvd2x instruction. The chain is necessary because the
351 /// sequence replaces a load and needs to provide the same number
355 /// An SDNode for swaps that are not associated with any loads/stores
356 /// and thereby have no chain.
359 /// QVFPERM = This corresponds to the QPX qvfperm instruction.
362 /// QVGPCI = This corresponds to the QPX qvgpci instruction.
365 /// QVALIGNI = This corresponds to the QPX qvaligni instruction.
368 /// QVESPLATI = This corresponds to the QPX qvesplati instruction.
371 /// QBFLT = Access the underlying QPX floating-point boolean
375 /// CHAIN = STBRX CHAIN, GPRC, Ptr, Type - This is a
376 /// byte-swapping store instruction. It byte-swaps the low "Type" bits of
377 /// the GPRC input, then stores it through Ptr. Type can be either i16 or
379 STBRX = ISD::FIRST_TARGET_MEMORY_OPCODE,
381 /// GPRC, CHAIN = LBRX CHAIN, Ptr, Type - This is a
382 /// byte-swapping load instruction. It loads "Type" bits, byte swaps it,
383 /// then puts it in the bottom bits of the GPRC. TYPE can be either i16
387 /// STFIWX - The STFIWX instruction. The first operand is an input token
388 /// chain, then an f64 value to store, then an address to store it to.
391 /// GPRC, CHAIN = LFIWAX CHAIN, Ptr - This is a floating-point
392 /// load which sign-extends from a 32-bit integer value into the
393 /// destination 64-bit register.
396 /// GPRC, CHAIN = LFIWZX CHAIN, Ptr - This is a floating-point
397 /// load which zero-extends from a 32-bit integer value into the
398 /// destination 64-bit register.
401 /// GPRC, CHAIN = LXSIZX, CHAIN, Ptr, ByteWidth - This is a load of an
402 /// integer smaller than 64 bits into a VSR. The integer is zero-extended.
403 /// This can be used for converting loaded integers to floating point.
406 /// STXSIX - The STXSI[bh]X instruction. The first operand is an input
407 /// chain, then an f64 value to store, then an address to store it to,
408 /// followed by a byte-width for the store.
411 /// VSRC, CHAIN = LXVD2X_LE CHAIN, Ptr - Occurs only for little endian.
412 /// Maps directly to an lxvd2x instruction that will be followed by
416 /// CHAIN = STXVD2X CHAIN, VSRC, Ptr - Occurs only for little endian.
417 /// Maps directly to an stxvd2x instruction that will be preceded by
421 /// QBRC, CHAIN = QVLFSb CHAIN, Ptr
422 /// The 4xf32 load used for v4i1 constants.
425 /// GPRC = TOC_ENTRY GA, TOC
426 /// Loads the entry for GA from the TOC, where the TOC base is given by
427 /// the last operand.
431 } // end namespace PPCISD
433 /// Define some predicates that are used for node matching.
436 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
437 /// VPKUHUM instruction.
438 bool isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
441 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
442 /// VPKUWUM instruction.
443 bool isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
446 /// isVPKUDUMShuffleMask - Return true if this is the shuffle mask for a
447 /// VPKUDUM instruction.
448 bool isVPKUDUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
451 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
452 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
453 bool isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
454 unsigned ShuffleKind, SelectionDAG &DAG);
456 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
457 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
458 bool isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
459 unsigned ShuffleKind, SelectionDAG &DAG);
461 /// isVMRGEOShuffleMask - Return true if this is a shuffle mask suitable for
462 /// a VMRGEW or VMRGOW instruction
463 bool isVMRGEOShuffleMask(ShuffleVectorSDNode *N, bool CheckEven,
464 unsigned ShuffleKind, SelectionDAG &DAG);
465 /// isXXSLDWIShuffleMask - Return true if this is a shuffle mask suitable
466 /// for a XXSLDWI instruction.
467 bool isXXSLDWIShuffleMask(ShuffleVectorSDNode *N, unsigned &ShiftElts,
468 bool &Swap, bool IsLE);
470 /// isXXBRHShuffleMask - Return true if this is a shuffle mask suitable
471 /// for a XXBRH instruction.
472 bool isXXBRHShuffleMask(ShuffleVectorSDNode *N);
474 /// isXXBRWShuffleMask - Return true if this is a shuffle mask suitable
475 /// for a XXBRW instruction.
476 bool isXXBRWShuffleMask(ShuffleVectorSDNode *N);
478 /// isXXBRDShuffleMask - Return true if this is a shuffle mask suitable
479 /// for a XXBRD instruction.
480 bool isXXBRDShuffleMask(ShuffleVectorSDNode *N);
482 /// isXXBRQShuffleMask - Return true if this is a shuffle mask suitable
483 /// for a XXBRQ instruction.
484 bool isXXBRQShuffleMask(ShuffleVectorSDNode *N);
486 /// isXXPERMDIShuffleMask - Return true if this is a shuffle mask suitable
487 /// for a XXPERMDI instruction.
488 bool isXXPERMDIShuffleMask(ShuffleVectorSDNode *N, unsigned &ShiftElts,
489 bool &Swap, bool IsLE);
491 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the
492 /// shift amount, otherwise return -1.
493 int isVSLDOIShuffleMask(SDNode *N, unsigned ShuffleKind,
496 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
497 /// specifies a splat of a single element that is suitable for input to
498 /// VSPLTB/VSPLTH/VSPLTW.
499 bool isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize);
501 /// isXXINSERTWMask - Return true if this VECTOR_SHUFFLE can be handled by
502 /// the XXINSERTW instruction introduced in ISA 3.0. This is essentially any
503 /// shuffle of v4f32/v4i32 vectors that just inserts one element from one
504 /// vector into the other. This function will also set a couple of
505 /// output parameters for how much the source vector needs to be shifted and
506 /// what byte number needs to be specified for the instruction to put the
507 /// element in the desired location of the target vector.
508 bool isXXINSERTWMask(ShuffleVectorSDNode *N, unsigned &ShiftElts,
509 unsigned &InsertAtByte, bool &Swap, bool IsLE);
511 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
512 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
513 unsigned getVSPLTImmediate(SDNode *N, unsigned EltSize, SelectionDAG &DAG);
515 /// get_VSPLTI_elt - If this is a build_vector of constants which can be
516 /// formed by using a vspltis[bhw] instruction of the specified element
517 /// size, return the constant being splatted. The ByteSize field indicates
518 /// the number of bytes of each element [124] -> [bhw].
519 SDValue get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG);
521 /// If this is a qvaligni shuffle mask, return the shift
522 /// amount, otherwise return -1.
523 int isQVALIGNIShuffleMask(SDNode *N);
525 } // end namespace PPC
527 class PPCTargetLowering : public TargetLowering {
528 const PPCSubtarget &Subtarget;
531 explicit PPCTargetLowering(const PPCTargetMachine &TM,
532 const PPCSubtarget &STI);
534 /// getTargetNodeName() - This method returns the name of a target specific
536 const char *getTargetNodeName(unsigned Opcode) const override;
538 /// getPreferredVectorAction - The code we generate when vector types are
539 /// legalized by promoting the integer element type is often much worse
540 /// than code we generate if we widen the type for applicable vector types.
541 /// The issue with promoting is that the vector is scalaraized, individual
542 /// elements promoted and then the vector is rebuilt. So say we load a pair
543 /// of v4i8's and shuffle them. This will turn into a mess of 8 extending
544 /// loads, moves back into VSR's (or memory ops if we don't have moves) and
545 /// then the VPERM for the shuffle. All in all a very slow sequence.
546 TargetLoweringBase::LegalizeTypeAction getPreferredVectorAction(EVT VT)
548 if (VT.getScalarSizeInBits() % 8 == 0)
549 return TypeWidenVector;
550 return TargetLoweringBase::getPreferredVectorAction(VT);
553 bool useSoftFloat() const override;
555 MVT getScalarShiftAmountTy(const DataLayout &, EVT) const override {
559 bool isCheapToSpeculateCttz() const override {
563 bool isCheapToSpeculateCtlz() const override {
567 bool isCtlzFast() const override {
571 bool hasAndNotCompare(SDValue) const override {
575 bool convertSetCCLogicToBitwiseLogic(EVT VT) const override {
576 return VT.isScalarInteger();
579 bool supportSplitCSR(MachineFunction *MF) const override {
581 MF->getFunction()->getCallingConv() == CallingConv::CXX_FAST_TLS &&
582 MF->getFunction()->hasFnAttribute(Attribute::NoUnwind);
585 void initializeSplitCSR(MachineBasicBlock *Entry) const override;
587 void insertCopiesSplitCSR(
588 MachineBasicBlock *Entry,
589 const SmallVectorImpl<MachineBasicBlock *> &Exits) const override;
591 /// getSetCCResultType - Return the ISD::SETCC ValueType
592 EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context,
593 EVT VT) const override;
595 /// Return true if target always beneficiates from combining into FMA for a
596 /// given value type. This must typically return false on targets where FMA
597 /// takes more cycles to execute than FADD.
598 bool enableAggressiveFMAFusion(EVT VT) const override;
600 /// getPreIndexedAddressParts - returns true by value, base pointer and
601 /// offset pointer and addressing mode by reference if the node's address
602 /// can be legally represented as pre-indexed load / store address.
603 bool getPreIndexedAddressParts(SDNode *N, SDValue &Base,
605 ISD::MemIndexedMode &AM,
606 SelectionDAG &DAG) const override;
608 /// SelectAddressRegReg - Given the specified addressed, check to see if it
609 /// can be represented as an indexed [r+r] operation. Returns false if it
610 /// can be more efficiently represented with [r+imm].
611 bool SelectAddressRegReg(SDValue N, SDValue &Base, SDValue &Index,
612 SelectionDAG &DAG) const;
614 /// SelectAddressRegImm - Returns true if the address N can be represented
615 /// by a base register plus a signed 16-bit displacement [r+imm], and if it
616 /// is not better represented as reg+reg. If Aligned is true, only accept
617 /// displacements suitable for STD and friends, i.e. multiples of 4.
618 bool SelectAddressRegImm(SDValue N, SDValue &Disp, SDValue &Base,
619 SelectionDAG &DAG, unsigned Alignment) const;
621 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
622 /// represented as an indexed [r+r] operation.
623 bool SelectAddressRegRegOnly(SDValue N, SDValue &Base, SDValue &Index,
624 SelectionDAG &DAG) const;
626 Sched::Preference getSchedulingPreference(SDNode *N) const override;
628 /// LowerOperation - Provide custom lowering hooks for some operations.
630 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
632 /// ReplaceNodeResults - Replace the results of node with an illegal result
633 /// type with new values built out of custom code.
635 void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
636 SelectionDAG &DAG) const override;
638 SDValue expandVSXLoadForLE(SDNode *N, DAGCombinerInfo &DCI) const;
639 SDValue expandVSXStoreForLE(SDNode *N, DAGCombinerInfo &DCI) const;
641 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
643 SDValue BuildSDIVPow2(SDNode *N, const APInt &Divisor, SelectionDAG &DAG,
644 std::vector<SDNode *> *Created) const override;
646 unsigned getRegisterByName(const char* RegName, EVT VT,
647 SelectionDAG &DAG) const override;
649 void computeKnownBitsForTargetNode(const SDValue Op,
651 const APInt &DemandedElts,
652 const SelectionDAG &DAG,
653 unsigned Depth = 0) const override;
655 unsigned getPrefLoopAlignment(MachineLoop *ML) const override;
657 bool shouldInsertFencesForAtomic(const Instruction *I) const override {
661 Instruction *emitLeadingFence(IRBuilder<> &Builder, Instruction *Inst,
662 AtomicOrdering Ord) const override;
663 Instruction *emitTrailingFence(IRBuilder<> &Builder, Instruction *Inst,
664 AtomicOrdering Ord) const override;
667 EmitInstrWithCustomInserter(MachineInstr &MI,
668 MachineBasicBlock *MBB) const override;
669 MachineBasicBlock *EmitAtomicBinary(MachineInstr &MI,
670 MachineBasicBlock *MBB,
673 unsigned CmpOpcode = 0,
674 unsigned CmpPred = 0) const;
675 MachineBasicBlock *EmitPartwordAtomicBinary(MachineInstr &MI,
676 MachineBasicBlock *MBB,
679 unsigned CmpOpcode = 0,
680 unsigned CmpPred = 0) const;
682 MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr &MI,
683 MachineBasicBlock *MBB) const;
685 MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr &MI,
686 MachineBasicBlock *MBB) const;
688 ConstraintType getConstraintType(StringRef Constraint) const override;
690 /// Examine constraint string and operand type and determine a weight value.
691 /// The operand object must already have been set up with the operand type.
692 ConstraintWeight getSingleConstraintMatchWeight(
693 AsmOperandInfo &info, const char *constraint) const override;
695 std::pair<unsigned, const TargetRegisterClass *>
696 getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
697 StringRef Constraint, MVT VT) const override;
699 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
700 /// function arguments in the caller parameter area. This is the actual
701 /// alignment, not its logarithm.
702 unsigned getByValTypeAlignment(Type *Ty,
703 const DataLayout &DL) const override;
705 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
706 /// vector. If it is invalid, don't add anything to Ops.
707 void LowerAsmOperandForConstraint(SDValue Op,
708 std::string &Constraint,
709 std::vector<SDValue> &Ops,
710 SelectionDAG &DAG) const override;
713 getInlineAsmMemConstraint(StringRef ConstraintCode) const override {
714 if (ConstraintCode == "es")
715 return InlineAsm::Constraint_es;
716 else if (ConstraintCode == "o")
717 return InlineAsm::Constraint_o;
718 else if (ConstraintCode == "Q")
719 return InlineAsm::Constraint_Q;
720 else if (ConstraintCode == "Z")
721 return InlineAsm::Constraint_Z;
722 else if (ConstraintCode == "Zy")
723 return InlineAsm::Constraint_Zy;
724 return TargetLowering::getInlineAsmMemConstraint(ConstraintCode);
727 /// isLegalAddressingMode - Return true if the addressing mode represented
728 /// by AM is legal for this target, for a load/store of the specified type.
729 bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM,
730 Type *Ty, unsigned AS) const override;
732 /// isLegalICmpImmediate - Return true if the specified immediate is legal
733 /// icmp immediate, that is the target has icmp instructions which can
734 /// compare a register against the immediate without having to materialize
735 /// the immediate into a register.
736 bool isLegalICmpImmediate(int64_t Imm) const override;
738 /// isLegalAddImmediate - Return true if the specified immediate is legal
739 /// add immediate, that is the target has add instructions which can
740 /// add a register and the immediate without having to materialize
741 /// the immediate into a register.
742 bool isLegalAddImmediate(int64_t Imm) const override;
744 /// isTruncateFree - Return true if it's free to truncate a value of
745 /// type Ty1 to type Ty2. e.g. On PPC it's free to truncate a i64 value in
746 /// register X1 to i32 by referencing its sub-register R1.
747 bool isTruncateFree(Type *Ty1, Type *Ty2) const override;
748 bool isTruncateFree(EVT VT1, EVT VT2) const override;
750 bool isZExtFree(SDValue Val, EVT VT2) const override;
752 bool isFPExtFree(EVT VT) const override;
754 /// \brief Returns true if it is beneficial to convert a load of a constant
755 /// to just the constant itself.
756 bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
757 Type *Ty) const override;
759 bool convertSelectOfConstantsToMath() const override {
763 bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const override;
765 bool getTgtMemIntrinsic(IntrinsicInfo &Info,
767 unsigned Intrinsic) const override;
769 /// getOptimalMemOpType - Returns the target specific optimal type for load
770 /// and store operations as a result of memset, memcpy, and memmove
771 /// lowering. If DstAlign is zero that means it's safe to destination
772 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
773 /// means there isn't a need to check it against alignment requirement,
774 /// probably because the source does not need to be loaded. If 'IsMemset' is
775 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
776 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
777 /// source is constant so it does not need to be loaded.
778 /// It returns EVT::Other if the type should be determined using generic
779 /// target-independent logic.
781 getOptimalMemOpType(uint64_t Size, unsigned DstAlign, unsigned SrcAlign,
782 bool IsMemset, bool ZeroMemset, bool MemcpyStrSrc,
783 MachineFunction &MF) const override;
785 /// Is unaligned memory access allowed for the given type, and is it fast
786 /// relative to software emulation.
787 bool allowsMisalignedMemoryAccesses(EVT VT,
790 bool *Fast = nullptr) const override;
792 /// isFMAFasterThanFMulAndFAdd - Return true if an FMA operation is faster
793 /// than a pair of fmul and fadd instructions. fmuladd intrinsics will be
794 /// expanded to FMAs when this method returns true, otherwise fmuladd is
795 /// expanded to fmul + fadd.
796 bool isFMAFasterThanFMulAndFAdd(EVT VT) const override;
798 const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const override;
800 // Should we expand the build vector with shuffles?
802 shouldExpandBuildVectorWithShuffles(EVT VT,
803 unsigned DefinedValues) const override;
805 /// createFastISel - This method returns a target-specific FastISel object,
806 /// or null if the target does not support "fast" instruction selection.
807 FastISel *createFastISel(FunctionLoweringInfo &FuncInfo,
808 const TargetLibraryInfo *LibInfo) const override;
810 /// \brief Returns true if an argument of type Ty needs to be passed in a
811 /// contiguous block of registers in calling convention CallConv.
812 bool functionArgumentNeedsConsecutiveRegisters(
813 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const override {
814 // We support any array type as "consecutive" block in the parameter
815 // save area. The element type defines the alignment requirement and
816 // whether the argument should go in GPRs, FPRs, or VRs if available.
818 // Note that clang uses this capability both to implement the ELFv2
819 // homogeneous float/vector aggregate ABI, and to avoid having to use
820 // "byval" when passing aggregates that might fully fit in registers.
821 return Ty->isArrayTy();
824 /// If a physical register, this returns the register that receives the
825 /// exception address on entry to an EH pad.
827 getExceptionPointerRegister(const Constant *PersonalityFn) const override;
829 /// If a physical register, this returns the register that receives the
830 /// exception typeid on entry to a landing pad.
832 getExceptionSelectorRegister(const Constant *PersonalityFn) const override;
834 /// Override to support customized stack guard loading.
835 bool useLoadStackGuardNode() const override;
836 void insertSSPDeclarations(Module &M) const override;
838 bool isFPImmLegal(const APFloat &Imm, EVT VT) const override;
840 unsigned getJumpTableEncoding() const override;
841 bool isJumpTableRelative() const override;
842 SDValue getPICJumpTableRelocBase(SDValue Table,
843 SelectionDAG &DAG) const override;
844 const MCExpr *getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
846 MCContext &Ctx) const override;
849 struct ReuseLoadInfo {
853 MachinePointerInfo MPI;
854 bool IsDereferenceable = false;
855 bool IsInvariant = false;
856 unsigned Alignment = 0;
858 const MDNode *Ranges = nullptr;
860 ReuseLoadInfo() = default;
862 MachineMemOperand::Flags MMOFlags() const {
863 MachineMemOperand::Flags F = MachineMemOperand::MONone;
864 if (IsDereferenceable)
865 F |= MachineMemOperand::MODereferenceable;
867 F |= MachineMemOperand::MOInvariant;
872 bool canReuseLoadAddress(SDValue Op, EVT MemVT, ReuseLoadInfo &RLI,
874 ISD::LoadExtType ET = ISD::NON_EXTLOAD) const;
875 void spliceIntoChain(SDValue ResChain, SDValue NewResChain,
876 SelectionDAG &DAG) const;
878 void LowerFP_TO_INTForReuse(SDValue Op, ReuseLoadInfo &RLI,
879 SelectionDAG &DAG, const SDLoc &dl) const;
880 SDValue LowerFP_TO_INTDirectMove(SDValue Op, SelectionDAG &DAG,
881 const SDLoc &dl) const;
883 bool directMoveIsProfitable(const SDValue &Op) const;
884 SDValue LowerINT_TO_FPDirectMove(SDValue Op, SelectionDAG &DAG,
885 const SDLoc &dl) const;
887 SDValue getFramePointerFrameIndex(SelectionDAG & DAG) const;
888 SDValue getReturnAddrFrameIndex(SelectionDAG & DAG) const;
891 IsEligibleForTailCallOptimization(SDValue Callee,
892 CallingConv::ID CalleeCC,
894 const SmallVectorImpl<ISD::InputArg> &Ins,
895 SelectionDAG& DAG) const;
898 IsEligibleForTailCallOptimization_64SVR4(
900 CallingConv::ID CalleeCC,
901 ImmutableCallSite *CS,
903 const SmallVectorImpl<ISD::OutputArg> &Outs,
904 const SmallVectorImpl<ISD::InputArg> &Ins,
905 SelectionDAG& DAG) const;
907 SDValue EmitTailCallLoadFPAndRetAddr(SelectionDAG &DAG, int SPDiff,
908 SDValue Chain, SDValue &LROpOut,
910 const SDLoc &dl) const;
912 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
913 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
914 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
915 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
916 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
917 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
918 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
919 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
920 SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
921 SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
922 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
923 SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const;
924 SDValue LowerVACOPY(SDValue Op, SelectionDAG &DAG) const;
925 SDValue LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG) const;
926 SDValue LowerGET_DYNAMIC_AREA_OFFSET(SDValue Op, SelectionDAG &DAG) const;
927 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
928 SDValue LowerEH_DWARF_CFA(SDValue Op, SelectionDAG &DAG) const;
929 SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const;
930 SDValue LowerSTORE(SDValue Op, SelectionDAG &DAG) const;
931 SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const;
932 SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
933 SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
934 const SDLoc &dl) const;
935 SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
936 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
937 SDValue LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const;
938 SDValue LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const;
939 SDValue LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const;
940 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
941 SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
942 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
943 SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
944 SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const;
945 SDValue LowerINTRINSIC_VOID(SDValue Op, SelectionDAG &DAG) const;
946 SDValue LowerREM(SDValue Op, SelectionDAG &DAG) const;
947 SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const;
948 SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const;
949 SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) const;
951 SDValue LowerVectorLoad(SDValue Op, SelectionDAG &DAG) const;
952 SDValue LowerVectorStore(SDValue Op, SelectionDAG &DAG) const;
954 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
955 CallingConv::ID CallConv, bool isVarArg,
956 const SmallVectorImpl<ISD::InputArg> &Ins,
957 const SDLoc &dl, SelectionDAG &DAG,
958 SmallVectorImpl<SDValue> &InVals) const;
959 SDValue FinishCall(CallingConv::ID CallConv, const SDLoc &dl,
960 bool isTailCall, bool isVarArg, bool isPatchPoint,
961 bool hasNest, SelectionDAG &DAG,
962 SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
963 SDValue InFlag, SDValue Chain, SDValue CallSeqStart,
964 SDValue &Callee, int SPDiff, unsigned NumBytes,
965 const SmallVectorImpl<ISD::InputArg> &Ins,
966 SmallVectorImpl<SDValue> &InVals,
967 ImmutableCallSite *CS) const;
970 LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
971 const SmallVectorImpl<ISD::InputArg> &Ins,
972 const SDLoc &dl, SelectionDAG &DAG,
973 SmallVectorImpl<SDValue> &InVals) const override;
975 SDValue LowerCall(TargetLowering::CallLoweringInfo &CLI,
976 SmallVectorImpl<SDValue> &InVals) const override;
978 bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
980 const SmallVectorImpl<ISD::OutputArg> &Outs,
981 LLVMContext &Context) const override;
983 SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
984 const SmallVectorImpl<ISD::OutputArg> &Outs,
985 const SmallVectorImpl<SDValue> &OutVals,
986 const SDLoc &dl, SelectionDAG &DAG) const override;
988 SDValue extendArgForPPC64(ISD::ArgFlagsTy Flags, EVT ObjectVT,
989 SelectionDAG &DAG, SDValue ArgVal,
990 const SDLoc &dl) const;
992 SDValue LowerFormalArguments_Darwin(
993 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
994 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
995 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const;
996 SDValue LowerFormalArguments_64SVR4(
997 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
998 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
999 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const;
1000 SDValue LowerFormalArguments_32SVR4(
1001 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1002 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
1003 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const;
1005 SDValue createMemcpyOutsideCallSeq(SDValue Arg, SDValue PtrOff,
1006 SDValue CallSeqStart,
1007 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1008 const SDLoc &dl) const;
1010 SDValue LowerCall_Darwin(SDValue Chain, SDValue Callee,
1011 CallingConv::ID CallConv, bool isVarArg,
1012 bool isTailCall, bool isPatchPoint,
1013 const SmallVectorImpl<ISD::OutputArg> &Outs,
1014 const SmallVectorImpl<SDValue> &OutVals,
1015 const SmallVectorImpl<ISD::InputArg> &Ins,
1016 const SDLoc &dl, SelectionDAG &DAG,
1017 SmallVectorImpl<SDValue> &InVals,
1018 ImmutableCallSite *CS) const;
1019 SDValue LowerCall_64SVR4(SDValue Chain, SDValue Callee,
1020 CallingConv::ID CallConv, bool isVarArg,
1021 bool isTailCall, bool isPatchPoint,
1022 const SmallVectorImpl<ISD::OutputArg> &Outs,
1023 const SmallVectorImpl<SDValue> &OutVals,
1024 const SmallVectorImpl<ISD::InputArg> &Ins,
1025 const SDLoc &dl, SelectionDAG &DAG,
1026 SmallVectorImpl<SDValue> &InVals,
1027 ImmutableCallSite *CS) const;
1028 SDValue LowerCall_32SVR4(SDValue Chain, SDValue Callee,
1029 CallingConv::ID CallConv, bool isVarArg,
1030 bool isTailCall, bool isPatchPoint,
1031 const SmallVectorImpl<ISD::OutputArg> &Outs,
1032 const SmallVectorImpl<SDValue> &OutVals,
1033 const SmallVectorImpl<ISD::InputArg> &Ins,
1034 const SDLoc &dl, SelectionDAG &DAG,
1035 SmallVectorImpl<SDValue> &InVals,
1036 ImmutableCallSite *CS) const;
1038 SDValue lowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
1039 SDValue lowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
1041 SDValue DAGCombineExtBoolTrunc(SDNode *N, DAGCombinerInfo &DCI) const;
1042 SDValue DAGCombineBuildVector(SDNode *N, DAGCombinerInfo &DCI) const;
1043 SDValue DAGCombineTruncBoolExt(SDNode *N, DAGCombinerInfo &DCI) const;
1044 SDValue combineFPToIntToFP(SDNode *N, DAGCombinerInfo &DCI) const;
1045 SDValue combineSHL(SDNode *N, DAGCombinerInfo &DCI) const;
1046 SDValue combineSRA(SDNode *N, DAGCombinerInfo &DCI) const;
1047 SDValue combineSRL(SDNode *N, DAGCombinerInfo &DCI) const;
1049 /// ConvertSETCCToSubtract - looks at SETCC that compares ints. It replaces
1050 /// SETCC with integer subtraction when (1) there is a legal way of doing it
1051 /// (2) keeping the result of comparison in GPR has performance benefit.
1052 SDValue ConvertSETCCToSubtract(SDNode *N, DAGCombinerInfo &DCI) const;
1054 SDValue getSqrtEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
1055 int &RefinementSteps, bool &UseOneConstNR,
1056 bool Reciprocal) const override;
1057 SDValue getRecipEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
1058 int &RefinementSteps) const override;
1059 unsigned combineRepeatedFPDivisors() const override;
1061 CCAssignFn *useFastISelCCs(unsigned Flag) const;
1064 combineElementTruncationToVectorTruncation(SDNode *N,
1065 DAGCombinerInfo &DCI) const;
1070 FastISel *createFastISel(FunctionLoweringInfo &FuncInfo,
1071 const TargetLibraryInfo *LibInfo);
1073 } // end namespace PPC
1075 bool CC_PPC32_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
1076 CCValAssign::LocInfo &LocInfo,
1077 ISD::ArgFlagsTy &ArgFlags,
1080 bool CC_PPC32_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
1082 CCValAssign::LocInfo &LocInfo,
1083 ISD::ArgFlagsTy &ArgFlags,
1087 CC_PPC32_SVR4_Custom_SkipLastArgRegsPPCF128(unsigned &ValNo, MVT &ValVT,
1089 CCValAssign::LocInfo &LocInfo,
1090 ISD::ArgFlagsTy &ArgFlags,
1093 bool CC_PPC32_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
1095 CCValAssign::LocInfo &LocInfo,
1096 ISD::ArgFlagsTy &ArgFlags,
1099 bool isIntS16Immediate(SDNode *N, int16_t &Imm);
1100 bool isIntS16Immediate(SDValue Op, int16_t &Imm);
1102 } // end namespace llvm
1104 #endif // LLVM_TARGET_POWERPC_PPC32ISELLOWERING_H