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 /// XXREVERSE - The PPC VSX reverse instruction
93 /// VECSHL - The PPC VSX shift left instruction
97 /// XXPERMDI - The PPC XXPERMDI instruction
101 /// The CMPB instruction (takes two operands of i32 or i64).
104 /// Hi/Lo - These represent the high and low 16-bit parts of a global
105 /// address respectively. These nodes have two operands, the first of
106 /// which must be a TargetGlobalAddress, and the second of which must be a
107 /// Constant. Selected naively, these turn into 'lis G+C' and 'li G+C',
108 /// though these are usually folded into other nodes.
111 /// The following two target-specific nodes are used for calls through
112 /// function pointers in the 64-bit SVR4 ABI.
114 /// OPRC, CHAIN = DYNALLOC(CHAIN, NEGSIZE, FRAME_INDEX)
115 /// This instruction is lowered in PPCRegisterInfo::eliminateFrameIndex to
116 /// compute an allocation on the stack.
119 /// This instruction is lowered in PPCRegisterInfo::eliminateFrameIndex to
120 /// compute an offset from native SP to the address of the most recent
124 /// GlobalBaseReg - On Darwin, this node represents the result of the mflr
125 /// at function entry, used for PIC code.
128 /// These nodes represent PPC shifts.
130 /// For scalar types, only the last `n + 1` bits of the shift amounts
131 /// are used, where n is log2(sizeof(element) * 8). See sld/slw, etc.
132 /// for exact behaviors.
134 /// For vector types, only the last n bits are used. See vsld.
137 /// The combination of sra[wd]i and addze used to implemented signed
138 /// integer division by a power of 2. The first operand is the dividend,
139 /// and the second is the constant shift amount (representing the
143 /// CALL - A direct function call.
144 /// CALL_NOP is a call with the special NOP which follows 64-bit
148 /// CHAIN,FLAG = MTCTR(VAL, CHAIN[, INFLAG]) - Directly corresponds to a
149 /// MTCTR instruction.
152 /// CHAIN,FLAG = BCTRL(CHAIN, INFLAG) - Directly corresponds to a
153 /// BCTRL instruction.
156 /// CHAIN,FLAG = BCTRL(CHAIN, ADDR, INFLAG) - The combination of a bctrl
157 /// instruction and the TOC reload required on SVR4 PPC64.
160 /// Return with a flag operand, matched by 'blr'
163 /// R32 = MFOCRF(CRREG, INFLAG) - Represents the MFOCRF instruction.
164 /// This copies the bits corresponding to the specified CRREG into the
165 /// resultant GPR. Bits corresponding to other CR regs are undefined.
168 /// Direct move from a VSX register to a GPR
171 /// Direct move from a GPR to a VSX register (algebraic)
174 /// Direct move from a GPR to a VSX register (zero)
177 /// Extract a subvector from signed integer vector and convert to FP.
178 /// It is primarily used to convert a (widened) illegal integer vector
179 /// type to a legal floating point vector type.
180 /// For example v2i32 -> widened to v4i32 -> v2f64
183 /// Extract a subvector from unsigned integer vector and convert to FP.
184 /// As with SINT_VEC_TO_FP, used for converting illegal types.
187 // FIXME: Remove these once the ANDI glue bug is fixed:
188 /// i1 = ANDIo_1_[EQ|GT]_BIT(i32 or i64 x) - Represents the result of the
189 /// eq or gt bit of CR0 after executing andi. x, 1. This is used to
190 /// implement truncation of i32 or i64 to i1.
191 ANDIo_1_EQ_BIT, ANDIo_1_GT_BIT,
193 // READ_TIME_BASE - A read of the 64-bit time-base register on a 32-bit
194 // target (returns (Lo, Hi)). It takes a chain operand.
197 // EH_SJLJ_SETJMP - SjLj exception handling setjmp.
200 // EH_SJLJ_LONGJMP - SjLj exception handling longjmp.
203 /// RESVEC = VCMP(LHS, RHS, OPC) - Represents one of the altivec VCMP*
204 /// instructions. For lack of better number, we use the opcode number
205 /// encoding for the OPC field to identify the compare. For example, 838
209 /// RESVEC, OUTFLAG = VCMPo(LHS, RHS, OPC) - Represents one of the
210 /// altivec VCMP*o instructions. For lack of better number, we use the
211 /// opcode number encoding for the OPC field to identify the compare. For
212 /// example, 838 is VCMPGTSH.
215 /// CHAIN = COND_BRANCH CHAIN, CRRC, OPC, DESTBB [, INFLAG] - This
216 /// corresponds to the COND_BRANCH pseudo instruction. CRRC is the
217 /// condition register to branch on, OPC is the branch opcode to use (e.g.
218 /// PPC::BLE), DESTBB is the destination block to branch to, and INFLAG is
219 /// an optional input flag argument.
222 /// CHAIN = BDNZ CHAIN, DESTBB - These are used to create counter-based
226 /// F8RC = FADDRTZ F8RC, F8RC - This is an FADD done with rounding
227 /// towards zero. Used only as part of the long double-to-int
228 /// conversion sequence.
231 /// F8RC = MFFS - This moves the FPSCR (not modeled) into the register.
234 /// TC_RETURN - A tail call return.
236 /// operand #1 callee (register or absolute)
237 /// operand #2 stack adjustment
238 /// operand #3 optional in flag
241 /// ch, gl = CR6[UN]SET ch, inglue - Toggle CR bit 6 for SVR4 vararg calls
245 /// GPRC = address of _GLOBAL_OFFSET_TABLE_. Used by initial-exec TLS
249 /// GPRC = address of _GLOBAL_OFFSET_TABLE_. Used by general dynamic and
250 /// local dynamic TLS on PPC32.
253 /// G8RC = ADDIS_GOT_TPREL_HA %X2, Symbol - Used by the initial-exec
254 /// TLS model, produces an ADDIS8 instruction that adds the GOT
255 /// base to sym\@got\@tprel\@ha.
258 /// G8RC = LD_GOT_TPREL_L Symbol, G8RReg - Used by the initial-exec
259 /// TLS model, produces a LD instruction with base register G8RReg
260 /// and offset sym\@got\@tprel\@l. This completes the addition that
261 /// finds the offset of "sym" relative to the thread pointer.
264 /// G8RC = ADD_TLS G8RReg, Symbol - Used by the initial-exec TLS
265 /// model, produces an ADD instruction that adds the contents of
266 /// G8RReg to the thread pointer. Symbol contains a relocation
267 /// sym\@tls which is to be replaced by the thread pointer and
268 /// identifies to the linker that the instruction is part of a
272 /// G8RC = ADDIS_TLSGD_HA %X2, Symbol - For the general-dynamic TLS
273 /// model, produces an ADDIS8 instruction that adds the GOT base
274 /// register to sym\@got\@tlsgd\@ha.
277 /// %X3 = ADDI_TLSGD_L G8RReg, Symbol - For the general-dynamic TLS
278 /// model, produces an ADDI8 instruction that adds G8RReg to
279 /// sym\@got\@tlsgd\@l and stores the result in X3. Hidden by
280 /// ADDIS_TLSGD_L_ADDR until after register assignment.
283 /// %X3 = GET_TLS_ADDR %X3, Symbol - For the general-dynamic TLS
284 /// model, produces a call to __tls_get_addr(sym\@tlsgd). Hidden by
285 /// ADDIS_TLSGD_L_ADDR until after register assignment.
288 /// G8RC = ADDI_TLSGD_L_ADDR G8RReg, Symbol, Symbol - Op that
289 /// combines ADDI_TLSGD_L and GET_TLS_ADDR until expansion following
290 /// register assignment.
293 /// G8RC = ADDIS_TLSLD_HA %X2, Symbol - For the local-dynamic TLS
294 /// model, produces an ADDIS8 instruction that adds the GOT base
295 /// register to sym\@got\@tlsld\@ha.
298 /// %X3 = ADDI_TLSLD_L G8RReg, Symbol - For the local-dynamic TLS
299 /// model, produces an ADDI8 instruction that adds G8RReg to
300 /// sym\@got\@tlsld\@l and stores the result in X3. Hidden by
301 /// ADDIS_TLSLD_L_ADDR until after register assignment.
304 /// %X3 = GET_TLSLD_ADDR %X3, Symbol - For the local-dynamic TLS
305 /// model, produces a call to __tls_get_addr(sym\@tlsld). Hidden by
306 /// ADDIS_TLSLD_L_ADDR until after register assignment.
309 /// G8RC = ADDI_TLSLD_L_ADDR G8RReg, Symbol, Symbol - Op that
310 /// combines ADDI_TLSLD_L and GET_TLSLD_ADDR until expansion
311 /// following register assignment.
314 /// G8RC = ADDIS_DTPREL_HA %X3, Symbol - For the local-dynamic TLS
315 /// model, produces an ADDIS8 instruction that adds X3 to
319 /// G8RC = ADDI_DTPREL_L G8RReg, Symbol - For the local-dynamic TLS
320 /// model, produces an ADDI8 instruction that adds G8RReg to
321 /// sym\@got\@dtprel\@l.
324 /// VRRC = VADD_SPLAT Elt, EltSize - Temporary node to be expanded
325 /// during instruction selection to optimize a BUILD_VECTOR into
326 /// operations on splats. This is necessary to avoid losing these
327 /// optimizations due to constant folding.
330 /// CHAIN = SC CHAIN, Imm128 - System call. The 7-bit unsigned
331 /// operand identifies the operating system entry point.
334 /// CHAIN = CLRBHRB CHAIN - Clear branch history rolling buffer.
337 /// GPRC, CHAIN = MFBHRBE CHAIN, Entry, Dummy - Move from branch
338 /// history rolling buffer entry.
341 /// CHAIN = RFEBB CHAIN, State - Return from event-based branch.
344 /// VSRC, CHAIN = XXSWAPD CHAIN, VSRC - Occurs only for little
345 /// endian. Maps to an xxswapd instruction that corrects an lxvd2x
346 /// or stxvd2x instruction. The chain is necessary because the
347 /// sequence replaces a load and needs to provide the same number
351 /// An SDNode for swaps that are not associated with any loads/stores
352 /// and thereby have no chain.
355 /// QVFPERM = This corresponds to the QPX qvfperm instruction.
358 /// QVGPCI = This corresponds to the QPX qvgpci instruction.
361 /// QVALIGNI = This corresponds to the QPX qvaligni instruction.
364 /// QVESPLATI = This corresponds to the QPX qvesplati instruction.
367 /// QBFLT = Access the underlying QPX floating-point boolean
371 /// CHAIN = STBRX CHAIN, GPRC, Ptr, Type - This is a
372 /// byte-swapping store instruction. It byte-swaps the low "Type" bits of
373 /// the GPRC input, then stores it through Ptr. Type can be either i16 or
375 STBRX = ISD::FIRST_TARGET_MEMORY_OPCODE,
377 /// GPRC, CHAIN = LBRX CHAIN, Ptr, Type - This is a
378 /// byte-swapping load instruction. It loads "Type" bits, byte swaps it,
379 /// then puts it in the bottom bits of the GPRC. TYPE can be either i16
383 /// STFIWX - The STFIWX instruction. The first operand is an input token
384 /// chain, then an f64 value to store, then an address to store it to.
387 /// GPRC, CHAIN = LFIWAX CHAIN, Ptr - This is a floating-point
388 /// load which sign-extends from a 32-bit integer value into the
389 /// destination 64-bit register.
392 /// GPRC, CHAIN = LFIWZX CHAIN, Ptr - This is a floating-point
393 /// load which zero-extends from a 32-bit integer value into the
394 /// destination 64-bit register.
397 /// GPRC, CHAIN = LXSIZX, CHAIN, Ptr, ByteWidth - This is a load of an
398 /// integer smaller than 64 bits into a VSR. The integer is zero-extended.
399 /// This can be used for converting loaded integers to floating point.
402 /// STXSIX - The STXSI[bh]X instruction. The first operand is an input
403 /// chain, then an f64 value to store, then an address to store it to,
404 /// followed by a byte-width for the store.
407 /// VSRC, CHAIN = LXVD2X_LE CHAIN, Ptr - Occurs only for little endian.
408 /// Maps directly to an lxvd2x instruction that will be followed by
412 /// CHAIN = STXVD2X CHAIN, VSRC, Ptr - Occurs only for little endian.
413 /// Maps directly to an stxvd2x instruction that will be preceded by
417 /// QBRC, CHAIN = QVLFSb CHAIN, Ptr
418 /// The 4xf32 load used for v4i1 constants.
421 /// GPRC = TOC_ENTRY GA, TOC
422 /// Loads the entry for GA from the TOC, where the TOC base is given by
423 /// the last operand.
427 } // end namespace PPCISD
429 /// Define some predicates that are used for node matching.
432 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
433 /// VPKUHUM instruction.
434 bool isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
437 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
438 /// VPKUWUM instruction.
439 bool isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
442 /// isVPKUDUMShuffleMask - Return true if this is the shuffle mask for a
443 /// VPKUDUM instruction.
444 bool isVPKUDUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
447 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
448 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
449 bool isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
450 unsigned ShuffleKind, SelectionDAG &DAG);
452 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
453 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
454 bool isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
455 unsigned ShuffleKind, SelectionDAG &DAG);
457 /// isVMRGEOShuffleMask - Return true if this is a shuffle mask suitable for
458 /// a VMRGEW or VMRGOW instruction
459 bool isVMRGEOShuffleMask(ShuffleVectorSDNode *N, bool CheckEven,
460 unsigned ShuffleKind, SelectionDAG &DAG);
461 /// isXXSLDWIShuffleMask - Return true if this is a shuffle mask suitable
462 /// for a XXSLDWI instruction.
463 bool isXXSLDWIShuffleMask(ShuffleVectorSDNode *N, unsigned &ShiftElts,
464 bool &Swap, bool IsLE);
466 /// isXXBRHShuffleMask - Return true if this is a shuffle mask suitable
467 /// for a XXBRH instruction.
468 bool isXXBRHShuffleMask(ShuffleVectorSDNode *N);
470 /// isXXBRWShuffleMask - Return true if this is a shuffle mask suitable
471 /// for a XXBRW instruction.
472 bool isXXBRWShuffleMask(ShuffleVectorSDNode *N);
474 /// isXXBRDShuffleMask - Return true if this is a shuffle mask suitable
475 /// for a XXBRD instruction.
476 bool isXXBRDShuffleMask(ShuffleVectorSDNode *N);
478 /// isXXBRQShuffleMask - Return true if this is a shuffle mask suitable
479 /// for a XXBRQ instruction.
480 bool isXXBRQShuffleMask(ShuffleVectorSDNode *N);
482 /// isXXPERMDIShuffleMask - Return true if this is a shuffle mask suitable
483 /// for a XXPERMDI instruction.
484 bool isXXPERMDIShuffleMask(ShuffleVectorSDNode *N, unsigned &ShiftElts,
485 bool &Swap, bool IsLE);
487 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the
488 /// shift amount, otherwise return -1.
489 int isVSLDOIShuffleMask(SDNode *N, unsigned ShuffleKind,
492 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
493 /// specifies a splat of a single element that is suitable for input to
494 /// VSPLTB/VSPLTH/VSPLTW.
495 bool isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize);
497 /// isXXINSERTWMask - Return true if this VECTOR_SHUFFLE can be handled by
498 /// the XXINSERTW instruction introduced in ISA 3.0. This is essentially any
499 /// shuffle of v4f32/v4i32 vectors that just inserts one element from one
500 /// vector into the other. This function will also set a couple of
501 /// output parameters for how much the source vector needs to be shifted and
502 /// what byte number needs to be specified for the instruction to put the
503 /// element in the desired location of the target vector.
504 bool isXXINSERTWMask(ShuffleVectorSDNode *N, unsigned &ShiftElts,
505 unsigned &InsertAtByte, bool &Swap, bool IsLE);
507 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
508 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
509 unsigned getVSPLTImmediate(SDNode *N, unsigned EltSize, SelectionDAG &DAG);
511 /// get_VSPLTI_elt - If this is a build_vector of constants which can be
512 /// formed by using a vspltis[bhw] instruction of the specified element
513 /// size, return the constant being splatted. The ByteSize field indicates
514 /// the number of bytes of each element [124] -> [bhw].
515 SDValue get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG);
517 /// If this is a qvaligni shuffle mask, return the shift
518 /// amount, otherwise return -1.
519 int isQVALIGNIShuffleMask(SDNode *N);
521 } // end namespace PPC
523 class PPCTargetLowering : public TargetLowering {
524 const PPCSubtarget &Subtarget;
527 explicit PPCTargetLowering(const PPCTargetMachine &TM,
528 const PPCSubtarget &STI);
530 /// getTargetNodeName() - This method returns the name of a target specific
532 const char *getTargetNodeName(unsigned Opcode) const override;
534 /// getPreferredVectorAction - The code we generate when vector types are
535 /// legalized by promoting the integer element type is often much worse
536 /// than code we generate if we widen the type for applicable vector types.
537 /// The issue with promoting is that the vector is scalaraized, individual
538 /// elements promoted and then the vector is rebuilt. So say we load a pair
539 /// of v4i8's and shuffle them. This will turn into a mess of 8 extending
540 /// loads, moves back into VSR's (or memory ops if we don't have moves) and
541 /// then the VPERM for the shuffle. All in all a very slow sequence.
542 TargetLoweringBase::LegalizeTypeAction getPreferredVectorAction(EVT VT)
544 if (VT.getScalarSizeInBits() % 8 == 0)
545 return TypeWidenVector;
546 return TargetLoweringBase::getPreferredVectorAction(VT);
549 bool useSoftFloat() const override;
551 MVT getScalarShiftAmountTy(const DataLayout &, EVT) const override {
555 bool isCheapToSpeculateCttz() const override {
559 bool isCheapToSpeculateCtlz() const override {
563 bool isCtlzFast() const override {
567 bool hasAndNotCompare(SDValue) const override {
571 bool convertSetCCLogicToBitwiseLogic(EVT VT) const override {
572 return VT.isScalarInteger();
575 bool supportSplitCSR(MachineFunction *MF) const override {
577 MF->getFunction()->getCallingConv() == CallingConv::CXX_FAST_TLS &&
578 MF->getFunction()->hasFnAttribute(Attribute::NoUnwind);
581 void initializeSplitCSR(MachineBasicBlock *Entry) const override;
583 void insertCopiesSplitCSR(
584 MachineBasicBlock *Entry,
585 const SmallVectorImpl<MachineBasicBlock *> &Exits) const override;
587 /// getSetCCResultType - Return the ISD::SETCC ValueType
588 EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context,
589 EVT VT) const override;
591 /// Return true if target always beneficiates from combining into FMA for a
592 /// given value type. This must typically return false on targets where FMA
593 /// takes more cycles to execute than FADD.
594 bool enableAggressiveFMAFusion(EVT VT) const override;
596 /// getPreIndexedAddressParts - returns true by value, base pointer and
597 /// offset pointer and addressing mode by reference if the node's address
598 /// can be legally represented as pre-indexed load / store address.
599 bool getPreIndexedAddressParts(SDNode *N, SDValue &Base,
601 ISD::MemIndexedMode &AM,
602 SelectionDAG &DAG) const override;
604 /// SelectAddressRegReg - Given the specified addressed, check to see if it
605 /// can be represented as an indexed [r+r] operation. Returns false if it
606 /// can be more efficiently represented with [r+imm].
607 bool SelectAddressRegReg(SDValue N, SDValue &Base, SDValue &Index,
608 SelectionDAG &DAG) const;
610 /// SelectAddressRegImm - Returns true if the address N can be represented
611 /// by a base register plus a signed 16-bit displacement [r+imm], and if it
612 /// is not better represented as reg+reg. If Aligned is true, only accept
613 /// displacements suitable for STD and friends, i.e. multiples of 4.
614 bool SelectAddressRegImm(SDValue N, SDValue &Disp, SDValue &Base,
615 SelectionDAG &DAG, bool Aligned) const;
617 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
618 /// represented as an indexed [r+r] operation.
619 bool SelectAddressRegRegOnly(SDValue N, SDValue &Base, SDValue &Index,
620 SelectionDAG &DAG) const;
622 Sched::Preference getSchedulingPreference(SDNode *N) const override;
624 /// LowerOperation - Provide custom lowering hooks for some operations.
626 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
628 /// ReplaceNodeResults - Replace the results of node with an illegal result
629 /// type with new values built out of custom code.
631 void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
632 SelectionDAG &DAG) const override;
634 SDValue expandVSXLoadForLE(SDNode *N, DAGCombinerInfo &DCI) const;
635 SDValue expandVSXStoreForLE(SDNode *N, DAGCombinerInfo &DCI) const;
637 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
639 SDValue BuildSDIVPow2(SDNode *N, const APInt &Divisor, SelectionDAG &DAG,
640 std::vector<SDNode *> *Created) const override;
642 unsigned getRegisterByName(const char* RegName, EVT VT,
643 SelectionDAG &DAG) const override;
645 void computeKnownBitsForTargetNode(const SDValue Op,
647 const APInt &DemandedElts,
648 const SelectionDAG &DAG,
649 unsigned Depth = 0) const override;
651 unsigned getPrefLoopAlignment(MachineLoop *ML) const override;
653 bool shouldInsertFencesForAtomic(const Instruction *I) const override {
657 Instruction *emitLeadingFence(IRBuilder<> &Builder, Instruction *Inst,
658 AtomicOrdering Ord) const override;
659 Instruction *emitTrailingFence(IRBuilder<> &Builder, Instruction *Inst,
660 AtomicOrdering Ord) const override;
663 EmitInstrWithCustomInserter(MachineInstr &MI,
664 MachineBasicBlock *MBB) const override;
665 MachineBasicBlock *EmitAtomicBinary(MachineInstr &MI,
666 MachineBasicBlock *MBB,
669 unsigned CmpOpcode = 0,
670 unsigned CmpPred = 0) const;
671 MachineBasicBlock *EmitPartwordAtomicBinary(MachineInstr &MI,
672 MachineBasicBlock *MBB,
675 unsigned CmpOpcode = 0,
676 unsigned CmpPred = 0) const;
678 MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr &MI,
679 MachineBasicBlock *MBB) const;
681 MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr &MI,
682 MachineBasicBlock *MBB) const;
684 ConstraintType getConstraintType(StringRef Constraint) const override;
686 /// Examine constraint string and operand type and determine a weight value.
687 /// The operand object must already have been set up with the operand type.
688 ConstraintWeight getSingleConstraintMatchWeight(
689 AsmOperandInfo &info, const char *constraint) const override;
691 std::pair<unsigned, const TargetRegisterClass *>
692 getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
693 StringRef Constraint, MVT VT) const override;
695 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
696 /// function arguments in the caller parameter area. This is the actual
697 /// alignment, not its logarithm.
698 unsigned getByValTypeAlignment(Type *Ty,
699 const DataLayout &DL) const override;
701 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
702 /// vector. If it is invalid, don't add anything to Ops.
703 void LowerAsmOperandForConstraint(SDValue Op,
704 std::string &Constraint,
705 std::vector<SDValue> &Ops,
706 SelectionDAG &DAG) const override;
709 getInlineAsmMemConstraint(StringRef ConstraintCode) const override {
710 if (ConstraintCode == "es")
711 return InlineAsm::Constraint_es;
712 else if (ConstraintCode == "o")
713 return InlineAsm::Constraint_o;
714 else if (ConstraintCode == "Q")
715 return InlineAsm::Constraint_Q;
716 else if (ConstraintCode == "Z")
717 return InlineAsm::Constraint_Z;
718 else if (ConstraintCode == "Zy")
719 return InlineAsm::Constraint_Zy;
720 return TargetLowering::getInlineAsmMemConstraint(ConstraintCode);
723 /// isLegalAddressingMode - Return true if the addressing mode represented
724 /// by AM is legal for this target, for a load/store of the specified type.
725 bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM,
726 Type *Ty, unsigned AS) const override;
728 /// isLegalICmpImmediate - Return true if the specified immediate is legal
729 /// icmp immediate, that is the target has icmp instructions which can
730 /// compare a register against the immediate without having to materialize
731 /// the immediate into a register.
732 bool isLegalICmpImmediate(int64_t Imm) const override;
734 /// isLegalAddImmediate - Return true if the specified immediate is legal
735 /// add immediate, that is the target has add instructions which can
736 /// add a register and the immediate without having to materialize
737 /// the immediate into a register.
738 bool isLegalAddImmediate(int64_t Imm) const override;
740 /// isTruncateFree - Return true if it's free to truncate a value of
741 /// type Ty1 to type Ty2. e.g. On PPC it's free to truncate a i64 value in
742 /// register X1 to i32 by referencing its sub-register R1.
743 bool isTruncateFree(Type *Ty1, Type *Ty2) const override;
744 bool isTruncateFree(EVT VT1, EVT VT2) const override;
746 bool isZExtFree(SDValue Val, EVT VT2) const override;
748 bool isFPExtFree(EVT VT) const override;
750 /// \brief Returns true if it is beneficial to convert a load of a constant
751 /// to just the constant itself.
752 bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
753 Type *Ty) const override;
755 bool convertSelectOfConstantsToMath() const override {
759 bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const override;
761 bool getTgtMemIntrinsic(IntrinsicInfo &Info,
763 unsigned Intrinsic) const override;
765 /// getOptimalMemOpType - Returns the target specific optimal type for load
766 /// and store operations as a result of memset, memcpy, and memmove
767 /// lowering. If DstAlign is zero that means it's safe to destination
768 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
769 /// means there isn't a need to check it against alignment requirement,
770 /// probably because the source does not need to be loaded. If 'IsMemset' is
771 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
772 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
773 /// source is constant so it does not need to be loaded.
774 /// It returns EVT::Other if the type should be determined using generic
775 /// target-independent logic.
777 getOptimalMemOpType(uint64_t Size, unsigned DstAlign, unsigned SrcAlign,
778 bool IsMemset, bool ZeroMemset, bool MemcpyStrSrc,
779 MachineFunction &MF) const override;
781 /// Is unaligned memory access allowed for the given type, and is it fast
782 /// relative to software emulation.
783 bool allowsMisalignedMemoryAccesses(EVT VT,
786 bool *Fast = nullptr) const override;
788 /// isFMAFasterThanFMulAndFAdd - Return true if an FMA operation is faster
789 /// than a pair of fmul and fadd instructions. fmuladd intrinsics will be
790 /// expanded to FMAs when this method returns true, otherwise fmuladd is
791 /// expanded to fmul + fadd.
792 bool isFMAFasterThanFMulAndFAdd(EVT VT) const override;
794 const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const override;
796 // Should we expand the build vector with shuffles?
798 shouldExpandBuildVectorWithShuffles(EVT VT,
799 unsigned DefinedValues) const override;
801 /// createFastISel - This method returns a target-specific FastISel object,
802 /// or null if the target does not support "fast" instruction selection.
803 FastISel *createFastISel(FunctionLoweringInfo &FuncInfo,
804 const TargetLibraryInfo *LibInfo) const override;
806 /// \brief Returns true if an argument of type Ty needs to be passed in a
807 /// contiguous block of registers in calling convention CallConv.
808 bool functionArgumentNeedsConsecutiveRegisters(
809 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const override {
810 // We support any array type as "consecutive" block in the parameter
811 // save area. The element type defines the alignment requirement and
812 // whether the argument should go in GPRs, FPRs, or VRs if available.
814 // Note that clang uses this capability both to implement the ELFv2
815 // homogeneous float/vector aggregate ABI, and to avoid having to use
816 // "byval" when passing aggregates that might fully fit in registers.
817 return Ty->isArrayTy();
820 /// If a physical register, this returns the register that receives the
821 /// exception address on entry to an EH pad.
823 getExceptionPointerRegister(const Constant *PersonalityFn) const override;
825 /// If a physical register, this returns the register that receives the
826 /// exception typeid on entry to a landing pad.
828 getExceptionSelectorRegister(const Constant *PersonalityFn) const override;
830 /// Override to support customized stack guard loading.
831 bool useLoadStackGuardNode() const override;
832 void insertSSPDeclarations(Module &M) const override;
834 bool isFPImmLegal(const APFloat &Imm, EVT VT) const override;
836 unsigned getJumpTableEncoding() const override;
837 bool isJumpTableRelative() const override;
838 SDValue getPICJumpTableRelocBase(SDValue Table,
839 SelectionDAG &DAG) const override;
840 const MCExpr *getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
842 MCContext &Ctx) const override;
845 struct ReuseLoadInfo {
849 MachinePointerInfo MPI;
850 bool IsDereferenceable = false;
851 bool IsInvariant = false;
852 unsigned Alignment = 0;
854 const MDNode *Ranges = nullptr;
856 ReuseLoadInfo() = default;
858 MachineMemOperand::Flags MMOFlags() const {
859 MachineMemOperand::Flags F = MachineMemOperand::MONone;
860 if (IsDereferenceable)
861 F |= MachineMemOperand::MODereferenceable;
863 F |= MachineMemOperand::MOInvariant;
868 bool canReuseLoadAddress(SDValue Op, EVT MemVT, ReuseLoadInfo &RLI,
870 ISD::LoadExtType ET = ISD::NON_EXTLOAD) const;
871 void spliceIntoChain(SDValue ResChain, SDValue NewResChain,
872 SelectionDAG &DAG) const;
874 void LowerFP_TO_INTForReuse(SDValue Op, ReuseLoadInfo &RLI,
875 SelectionDAG &DAG, const SDLoc &dl) const;
876 SDValue LowerFP_TO_INTDirectMove(SDValue Op, SelectionDAG &DAG,
877 const SDLoc &dl) const;
879 bool directMoveIsProfitable(const SDValue &Op) const;
880 SDValue LowerINT_TO_FPDirectMove(SDValue Op, SelectionDAG &DAG,
881 const SDLoc &dl) const;
883 SDValue getFramePointerFrameIndex(SelectionDAG & DAG) const;
884 SDValue getReturnAddrFrameIndex(SelectionDAG & DAG) const;
887 IsEligibleForTailCallOptimization(SDValue Callee,
888 CallingConv::ID CalleeCC,
890 const SmallVectorImpl<ISD::InputArg> &Ins,
891 SelectionDAG& DAG) const;
894 IsEligibleForTailCallOptimization_64SVR4(
896 CallingConv::ID CalleeCC,
897 ImmutableCallSite *CS,
899 const SmallVectorImpl<ISD::OutputArg> &Outs,
900 const SmallVectorImpl<ISD::InputArg> &Ins,
901 SelectionDAG& DAG) const;
903 SDValue EmitTailCallLoadFPAndRetAddr(SelectionDAG &DAG, int SPDiff,
904 SDValue Chain, SDValue &LROpOut,
906 const SDLoc &dl) const;
908 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
909 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
910 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
911 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
912 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
913 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
914 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
915 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
916 SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
917 SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
918 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
919 SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const;
920 SDValue LowerVACOPY(SDValue Op, SelectionDAG &DAG) const;
921 SDValue LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG) const;
922 SDValue LowerGET_DYNAMIC_AREA_OFFSET(SDValue Op, SelectionDAG &DAG) const;
923 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
924 SDValue LowerEH_DWARF_CFA(SDValue Op, SelectionDAG &DAG) const;
925 SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const;
926 SDValue LowerSTORE(SDValue Op, SelectionDAG &DAG) const;
927 SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const;
928 SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
929 SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
930 const SDLoc &dl) const;
931 SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
932 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
933 SDValue LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const;
934 SDValue LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const;
935 SDValue LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const;
936 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
937 SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
938 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
939 SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
940 SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const;
941 SDValue LowerINTRINSIC_VOID(SDValue Op, SelectionDAG &DAG) const;
942 SDValue LowerREM(SDValue Op, SelectionDAG &DAG) const;
943 SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const;
944 SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const;
945 SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) const;
947 SDValue LowerVectorLoad(SDValue Op, SelectionDAG &DAG) const;
948 SDValue LowerVectorStore(SDValue Op, SelectionDAG &DAG) const;
950 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
951 CallingConv::ID CallConv, bool isVarArg,
952 const SmallVectorImpl<ISD::InputArg> &Ins,
953 const SDLoc &dl, SelectionDAG &DAG,
954 SmallVectorImpl<SDValue> &InVals) const;
955 SDValue FinishCall(CallingConv::ID CallConv, const SDLoc &dl,
956 bool isTailCall, bool isVarArg, bool isPatchPoint,
957 bool hasNest, SelectionDAG &DAG,
958 SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
959 SDValue InFlag, SDValue Chain, SDValue CallSeqStart,
960 SDValue &Callee, int SPDiff, unsigned NumBytes,
961 const SmallVectorImpl<ISD::InputArg> &Ins,
962 SmallVectorImpl<SDValue> &InVals,
963 ImmutableCallSite *CS) const;
966 LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
967 const SmallVectorImpl<ISD::InputArg> &Ins,
968 const SDLoc &dl, SelectionDAG &DAG,
969 SmallVectorImpl<SDValue> &InVals) const override;
971 SDValue LowerCall(TargetLowering::CallLoweringInfo &CLI,
972 SmallVectorImpl<SDValue> &InVals) const override;
974 bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
976 const SmallVectorImpl<ISD::OutputArg> &Outs,
977 LLVMContext &Context) const override;
979 SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
980 const SmallVectorImpl<ISD::OutputArg> &Outs,
981 const SmallVectorImpl<SDValue> &OutVals,
982 const SDLoc &dl, SelectionDAG &DAG) const override;
984 SDValue extendArgForPPC64(ISD::ArgFlagsTy Flags, EVT ObjectVT,
985 SelectionDAG &DAG, SDValue ArgVal,
986 const SDLoc &dl) const;
988 SDValue LowerFormalArguments_Darwin(
989 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
990 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
991 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const;
992 SDValue LowerFormalArguments_64SVR4(
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_32SVR4(
997 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
998 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
999 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const;
1001 SDValue createMemcpyOutsideCallSeq(SDValue Arg, SDValue PtrOff,
1002 SDValue CallSeqStart,
1003 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1004 const SDLoc &dl) const;
1006 SDValue LowerCall_Darwin(SDValue Chain, SDValue Callee,
1007 CallingConv::ID CallConv, bool isVarArg,
1008 bool isTailCall, bool isPatchPoint,
1009 const SmallVectorImpl<ISD::OutputArg> &Outs,
1010 const SmallVectorImpl<SDValue> &OutVals,
1011 const SmallVectorImpl<ISD::InputArg> &Ins,
1012 const SDLoc &dl, SelectionDAG &DAG,
1013 SmallVectorImpl<SDValue> &InVals,
1014 ImmutableCallSite *CS) const;
1015 SDValue LowerCall_64SVR4(SDValue Chain, SDValue Callee,
1016 CallingConv::ID CallConv, bool isVarArg,
1017 bool isTailCall, bool isPatchPoint,
1018 const SmallVectorImpl<ISD::OutputArg> &Outs,
1019 const SmallVectorImpl<SDValue> &OutVals,
1020 const SmallVectorImpl<ISD::InputArg> &Ins,
1021 const SDLoc &dl, SelectionDAG &DAG,
1022 SmallVectorImpl<SDValue> &InVals,
1023 ImmutableCallSite *CS) const;
1024 SDValue LowerCall_32SVR4(SDValue Chain, SDValue Callee,
1025 CallingConv::ID CallConv, bool isVarArg,
1026 bool isTailCall, bool isPatchPoint,
1027 const SmallVectorImpl<ISD::OutputArg> &Outs,
1028 const SmallVectorImpl<SDValue> &OutVals,
1029 const SmallVectorImpl<ISD::InputArg> &Ins,
1030 const SDLoc &dl, SelectionDAG &DAG,
1031 SmallVectorImpl<SDValue> &InVals,
1032 ImmutableCallSite *CS) const;
1034 SDValue lowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
1035 SDValue lowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
1037 SDValue DAGCombineExtBoolTrunc(SDNode *N, DAGCombinerInfo &DCI) const;
1038 SDValue DAGCombineBuildVector(SDNode *N, DAGCombinerInfo &DCI) const;
1039 SDValue DAGCombineTruncBoolExt(SDNode *N, DAGCombinerInfo &DCI) const;
1040 SDValue combineFPToIntToFP(SDNode *N, DAGCombinerInfo &DCI) const;
1041 SDValue combineSHL(SDNode *N, DAGCombinerInfo &DCI) const;
1042 SDValue combineSRA(SDNode *N, DAGCombinerInfo &DCI) const;
1043 SDValue combineSRL(SDNode *N, DAGCombinerInfo &DCI) const;
1045 /// ConvertSETCCToSubtract - looks at SETCC that compares ints. It replaces
1046 /// SETCC with integer subtraction when (1) there is a legal way of doing it
1047 /// (2) keeping the result of comparison in GPR has performance benefit.
1048 SDValue ConvertSETCCToSubtract(SDNode *N, DAGCombinerInfo &DCI) const;
1050 SDValue getSqrtEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
1051 int &RefinementSteps, bool &UseOneConstNR,
1052 bool Reciprocal) const override;
1053 SDValue getRecipEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
1054 int &RefinementSteps) const override;
1055 unsigned combineRepeatedFPDivisors() const override;
1057 CCAssignFn *useFastISelCCs(unsigned Flag) const;
1060 combineElementTruncationToVectorTruncation(SDNode *N,
1061 DAGCombinerInfo &DCI) const;
1066 FastISel *createFastISel(FunctionLoweringInfo &FuncInfo,
1067 const TargetLibraryInfo *LibInfo);
1069 } // end namespace PPC
1071 bool CC_PPC32_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
1072 CCValAssign::LocInfo &LocInfo,
1073 ISD::ArgFlagsTy &ArgFlags,
1076 bool CC_PPC32_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
1078 CCValAssign::LocInfo &LocInfo,
1079 ISD::ArgFlagsTy &ArgFlags,
1083 CC_PPC32_SVR4_Custom_SkipLastArgRegsPPCF128(unsigned &ValNo, MVT &ValVT,
1085 CCValAssign::LocInfo &LocInfo,
1086 ISD::ArgFlagsTy &ArgFlags,
1089 bool CC_PPC32_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
1091 CCValAssign::LocInfo &LocInfo,
1092 ISD::ArgFlagsTy &ArgFlags,
1095 } // end namespace llvm
1097 #endif // LLVM_TARGET_POWERPC_PPC32ISELLOWERING_H