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 /// The CMPB instruction (takes two operands of i32 or i64).
96 /// Hi/Lo - These represent the high and low 16-bit parts of a global
97 /// address respectively. These nodes have two operands, the first of
98 /// which must be a TargetGlobalAddress, and the second of which must be a
99 /// Constant. Selected naively, these turn into 'lis G+C' and 'li G+C',
100 /// though these are usually folded into other nodes.
103 /// The following two target-specific nodes are used for calls through
104 /// function pointers in the 64-bit SVR4 ABI.
106 /// OPRC, CHAIN = DYNALLOC(CHAIN, NEGSIZE, FRAME_INDEX)
107 /// This instruction is lowered in PPCRegisterInfo::eliminateFrameIndex to
108 /// compute an allocation on the stack.
111 /// This instruction is lowered in PPCRegisterInfo::eliminateFrameIndex to
112 /// compute an offset from native SP to the address of the most recent
116 /// GlobalBaseReg - On Darwin, this node represents the result of the mflr
117 /// at function entry, used for PIC code.
120 /// These nodes represent PPC shifts.
122 /// For scalar types, only the last `n + 1` bits of the shift amounts
123 /// are used, where n is log2(sizeof(element) * 8). See sld/slw, etc.
124 /// for exact behaviors.
126 /// For vector types, only the last n bits are used. See vsld.
129 /// The combination of sra[wd]i and addze used to implemented signed
130 /// integer division by a power of 2. The first operand is the dividend,
131 /// and the second is the constant shift amount (representing the
135 /// CALL - A direct function call.
136 /// CALL_NOP is a call with the special NOP which follows 64-bit
140 /// CHAIN,FLAG = MTCTR(VAL, CHAIN[, INFLAG]) - Directly corresponds to a
141 /// MTCTR instruction.
144 /// CHAIN,FLAG = BCTRL(CHAIN, INFLAG) - Directly corresponds to a
145 /// BCTRL instruction.
148 /// CHAIN,FLAG = BCTRL(CHAIN, ADDR, INFLAG) - The combination of a bctrl
149 /// instruction and the TOC reload required on SVR4 PPC64.
152 /// Return with a flag operand, matched by 'blr'
155 /// R32 = MFOCRF(CRREG, INFLAG) - Represents the MFOCRF instruction.
156 /// This copies the bits corresponding to the specified CRREG into the
157 /// resultant GPR. Bits corresponding to other CR regs are undefined.
160 /// Direct move from a VSX register to a GPR
163 /// Direct move from a GPR to a VSX register (algebraic)
166 /// Direct move from a GPR to a VSX register (zero)
169 /// Extract a subvector from signed integer vector and convert to FP.
170 /// It is primarily used to convert a (widened) illegal integer vector
171 /// type to a legal floating point vector type.
172 /// For example v2i32 -> widened to v4i32 -> v2f64
175 /// Extract a subvector from unsigned integer vector and convert to FP.
176 /// As with SINT_VEC_TO_FP, used for converting illegal types.
179 // FIXME: Remove these once the ANDI glue bug is fixed:
180 /// i1 = ANDIo_1_[EQ|GT]_BIT(i32 or i64 x) - Represents the result of the
181 /// eq or gt bit of CR0 after executing andi. x, 1. This is used to
182 /// implement truncation of i32 or i64 to i1.
183 ANDIo_1_EQ_BIT, ANDIo_1_GT_BIT,
185 // READ_TIME_BASE - A read of the 64-bit time-base register on a 32-bit
186 // target (returns (Lo, Hi)). It takes a chain operand.
189 // EH_SJLJ_SETJMP - SjLj exception handling setjmp.
192 // EH_SJLJ_LONGJMP - SjLj exception handling longjmp.
195 /// RESVEC = VCMP(LHS, RHS, OPC) - Represents one of the altivec VCMP*
196 /// instructions. For lack of better number, we use the opcode number
197 /// encoding for the OPC field to identify the compare. For example, 838
201 /// RESVEC, OUTFLAG = VCMPo(LHS, RHS, OPC) - Represents one of the
202 /// altivec VCMP*o instructions. For lack of better number, we use the
203 /// opcode number encoding for the OPC field to identify the compare. For
204 /// example, 838 is VCMPGTSH.
207 /// CHAIN = COND_BRANCH CHAIN, CRRC, OPC, DESTBB [, INFLAG] - This
208 /// corresponds to the COND_BRANCH pseudo instruction. CRRC is the
209 /// condition register to branch on, OPC is the branch opcode to use (e.g.
210 /// PPC::BLE), DESTBB is the destination block to branch to, and INFLAG is
211 /// an optional input flag argument.
214 /// CHAIN = BDNZ CHAIN, DESTBB - These are used to create counter-based
218 /// F8RC = FADDRTZ F8RC, F8RC - This is an FADD done with rounding
219 /// towards zero. Used only as part of the long double-to-int
220 /// conversion sequence.
223 /// F8RC = MFFS - This moves the FPSCR (not modeled) into the register.
226 /// TC_RETURN - A tail call return.
228 /// operand #1 callee (register or absolute)
229 /// operand #2 stack adjustment
230 /// operand #3 optional in flag
233 /// ch, gl = CR6[UN]SET ch, inglue - Toggle CR bit 6 for SVR4 vararg calls
237 /// GPRC = address of _GLOBAL_OFFSET_TABLE_. Used by initial-exec TLS
241 /// GPRC = address of _GLOBAL_OFFSET_TABLE_. Used by general dynamic and
242 /// local dynamic TLS on PPC32.
245 /// G8RC = ADDIS_GOT_TPREL_HA %X2, Symbol - Used by the initial-exec
246 /// TLS model, produces an ADDIS8 instruction that adds the GOT
247 /// base to sym\@got\@tprel\@ha.
250 /// G8RC = LD_GOT_TPREL_L Symbol, G8RReg - Used by the initial-exec
251 /// TLS model, produces a LD instruction with base register G8RReg
252 /// and offset sym\@got\@tprel\@l. This completes the addition that
253 /// finds the offset of "sym" relative to the thread pointer.
256 /// G8RC = ADD_TLS G8RReg, Symbol - Used by the initial-exec TLS
257 /// model, produces an ADD instruction that adds the contents of
258 /// G8RReg to the thread pointer. Symbol contains a relocation
259 /// sym\@tls which is to be replaced by the thread pointer and
260 /// identifies to the linker that the instruction is part of a
264 /// G8RC = ADDIS_TLSGD_HA %X2, Symbol - For the general-dynamic TLS
265 /// model, produces an ADDIS8 instruction that adds the GOT base
266 /// register to sym\@got\@tlsgd\@ha.
269 /// %X3 = ADDI_TLSGD_L G8RReg, Symbol - For the general-dynamic TLS
270 /// model, produces an ADDI8 instruction that adds G8RReg to
271 /// sym\@got\@tlsgd\@l and stores the result in X3. Hidden by
272 /// ADDIS_TLSGD_L_ADDR until after register assignment.
275 /// %X3 = GET_TLS_ADDR %X3, Symbol - For the general-dynamic TLS
276 /// model, produces a call to __tls_get_addr(sym\@tlsgd). Hidden by
277 /// ADDIS_TLSGD_L_ADDR until after register assignment.
280 /// G8RC = ADDI_TLSGD_L_ADDR G8RReg, Symbol, Symbol - Op that
281 /// combines ADDI_TLSGD_L and GET_TLS_ADDR until expansion following
282 /// register assignment.
285 /// G8RC = ADDIS_TLSLD_HA %X2, Symbol - For the local-dynamic TLS
286 /// model, produces an ADDIS8 instruction that adds the GOT base
287 /// register to sym\@got\@tlsld\@ha.
290 /// %X3 = ADDI_TLSLD_L G8RReg, Symbol - For the local-dynamic TLS
291 /// model, produces an ADDI8 instruction that adds G8RReg to
292 /// sym\@got\@tlsld\@l and stores the result in X3. Hidden by
293 /// ADDIS_TLSLD_L_ADDR until after register assignment.
296 /// %X3 = GET_TLSLD_ADDR %X3, Symbol - For the local-dynamic TLS
297 /// model, produces a call to __tls_get_addr(sym\@tlsld). Hidden by
298 /// ADDIS_TLSLD_L_ADDR until after register assignment.
301 /// G8RC = ADDI_TLSLD_L_ADDR G8RReg, Symbol, Symbol - Op that
302 /// combines ADDI_TLSLD_L and GET_TLSLD_ADDR until expansion
303 /// following register assignment.
306 /// G8RC = ADDIS_DTPREL_HA %X3, Symbol - For the local-dynamic TLS
307 /// model, produces an ADDIS8 instruction that adds X3 to
311 /// G8RC = ADDI_DTPREL_L G8RReg, Symbol - For the local-dynamic TLS
312 /// model, produces an ADDI8 instruction that adds G8RReg to
313 /// sym\@got\@dtprel\@l.
316 /// VRRC = VADD_SPLAT Elt, EltSize - Temporary node to be expanded
317 /// during instruction selection to optimize a BUILD_VECTOR into
318 /// operations on splats. This is necessary to avoid losing these
319 /// optimizations due to constant folding.
322 /// CHAIN = SC CHAIN, Imm128 - System call. The 7-bit unsigned
323 /// operand identifies the operating system entry point.
326 /// CHAIN = CLRBHRB CHAIN - Clear branch history rolling buffer.
329 /// GPRC, CHAIN = MFBHRBE CHAIN, Entry, Dummy - Move from branch
330 /// history rolling buffer entry.
333 /// CHAIN = RFEBB CHAIN, State - Return from event-based branch.
336 /// VSRC, CHAIN = XXSWAPD CHAIN, VSRC - Occurs only for little
337 /// endian. Maps to an xxswapd instruction that corrects an lxvd2x
338 /// or stxvd2x instruction. The chain is necessary because the
339 /// sequence replaces a load and needs to provide the same number
343 /// An SDNode for swaps that are not associated with any loads/stores
344 /// and thereby have no chain.
347 /// QVFPERM = This corresponds to the QPX qvfperm instruction.
350 /// QVGPCI = This corresponds to the QPX qvgpci instruction.
353 /// QVALIGNI = This corresponds to the QPX qvaligni instruction.
356 /// QVESPLATI = This corresponds to the QPX qvesplati instruction.
359 /// QBFLT = Access the underlying QPX floating-point boolean
363 /// CHAIN = STBRX CHAIN, GPRC, Ptr, Type - This is a
364 /// byte-swapping store instruction. It byte-swaps the low "Type" bits of
365 /// the GPRC input, then stores it through Ptr. Type can be either i16 or
367 STBRX = ISD::FIRST_TARGET_MEMORY_OPCODE,
369 /// GPRC, CHAIN = LBRX CHAIN, Ptr, Type - This is a
370 /// byte-swapping load instruction. It loads "Type" bits, byte swaps it,
371 /// then puts it in the bottom bits of the GPRC. TYPE can be either i16
375 /// STFIWX - The STFIWX instruction. The first operand is an input token
376 /// chain, then an f64 value to store, then an address to store it to.
379 /// GPRC, CHAIN = LFIWAX CHAIN, Ptr - This is a floating-point
380 /// load which sign-extends from a 32-bit integer value into the
381 /// destination 64-bit register.
384 /// GPRC, CHAIN = LFIWZX CHAIN, Ptr - This is a floating-point
385 /// load which zero-extends from a 32-bit integer value into the
386 /// destination 64-bit register.
389 /// GPRC, CHAIN = LXSIZX, CHAIN, Ptr, ByteWidth - This is a load of an
390 /// integer smaller than 64 bits into a VSR. The integer is zero-extended.
391 /// This can be used for converting loaded integers to floating point.
394 /// STXSIX - The STXSI[bh]X instruction. The first operand is an input
395 /// chain, then an f64 value to store, then an address to store it to,
396 /// followed by a byte-width for the store.
399 /// VSRC, CHAIN = LXVD2X_LE CHAIN, Ptr - Occurs only for little endian.
400 /// Maps directly to an lxvd2x instruction that will be followed by
404 /// CHAIN = STXVD2X CHAIN, VSRC, Ptr - Occurs only for little endian.
405 /// Maps directly to an stxvd2x instruction that will be preceded by
409 /// QBRC, CHAIN = QVLFSb CHAIN, Ptr
410 /// The 4xf32 load used for v4i1 constants.
413 /// GPRC = TOC_ENTRY GA, TOC
414 /// Loads the entry for GA from the TOC, where the TOC base is given by
415 /// the last operand.
419 } // end namespace PPCISD
421 /// Define some predicates that are used for node matching.
424 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
425 /// VPKUHUM instruction.
426 bool isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
429 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
430 /// VPKUWUM instruction.
431 bool isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
434 /// isVPKUDUMShuffleMask - Return true if this is the shuffle mask for a
435 /// VPKUDUM instruction.
436 bool isVPKUDUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
439 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
440 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
441 bool isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
442 unsigned ShuffleKind, SelectionDAG &DAG);
444 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
445 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
446 bool isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
447 unsigned ShuffleKind, SelectionDAG &DAG);
449 /// isVMRGEOShuffleMask - Return true if this is a shuffle mask suitable for
450 /// a VMRGEW or VMRGOW instruction
451 bool isVMRGEOShuffleMask(ShuffleVectorSDNode *N, bool CheckEven,
452 unsigned ShuffleKind, SelectionDAG &DAG);
453 /// isXXSLDWIShuffleMask - Return true if this is a shuffle mask suitable
454 /// for a XXSLDWI instruction.
455 bool isXXSLDWIShuffleMask(ShuffleVectorSDNode *N, unsigned &ShiftElts,
456 bool &Swap, bool IsLE);
458 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the
459 /// shift amount, otherwise return -1.
460 int isVSLDOIShuffleMask(SDNode *N, unsigned ShuffleKind,
463 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
464 /// specifies a splat of a single element that is suitable for input to
465 /// VSPLTB/VSPLTH/VSPLTW.
466 bool isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize);
468 /// isXXINSERTWMask - Return true if this VECTOR_SHUFFLE can be handled by
469 /// the XXINSERTW instruction introduced in ISA 3.0. This is essentially any
470 /// shuffle of v4f32/v4i32 vectors that just inserts one element from one
471 /// vector into the other. This function will also set a couple of
472 /// output parameters for how much the source vector needs to be shifted and
473 /// what byte number needs to be specified for the instruction to put the
474 /// element in the desired location of the target vector.
475 bool isXXINSERTWMask(ShuffleVectorSDNode *N, unsigned &ShiftElts,
476 unsigned &InsertAtByte, bool &Swap, bool IsLE);
478 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
479 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
480 unsigned getVSPLTImmediate(SDNode *N, unsigned EltSize, SelectionDAG &DAG);
482 /// get_VSPLTI_elt - If this is a build_vector of constants which can be
483 /// formed by using a vspltis[bhw] instruction of the specified element
484 /// size, return the constant being splatted. The ByteSize field indicates
485 /// the number of bytes of each element [124] -> [bhw].
486 SDValue get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG);
488 /// If this is a qvaligni shuffle mask, return the shift
489 /// amount, otherwise return -1.
490 int isQVALIGNIShuffleMask(SDNode *N);
492 } // end namespace PPC
494 class PPCTargetLowering : public TargetLowering {
495 const PPCSubtarget &Subtarget;
498 explicit PPCTargetLowering(const PPCTargetMachine &TM,
499 const PPCSubtarget &STI);
501 /// getTargetNodeName() - This method returns the name of a target specific
503 const char *getTargetNodeName(unsigned Opcode) const override;
505 /// getPreferredVectorAction - The code we generate when vector types are
506 /// legalized by promoting the integer element type is often much worse
507 /// than code we generate if we widen the type for applicable vector types.
508 /// The issue with promoting is that the vector is scalaraized, individual
509 /// elements promoted and then the vector is rebuilt. So say we load a pair
510 /// of v4i8's and shuffle them. This will turn into a mess of 8 extending
511 /// loads, moves back into VSR's (or memory ops if we don't have moves) and
512 /// then the VPERM for the shuffle. All in all a very slow sequence.
513 TargetLoweringBase::LegalizeTypeAction getPreferredVectorAction(EVT VT)
515 if (VT.getScalarSizeInBits() % 8 == 0)
516 return TypeWidenVector;
517 return TargetLoweringBase::getPreferredVectorAction(VT);
520 bool useSoftFloat() const override;
522 MVT getScalarShiftAmountTy(const DataLayout &, EVT) const override {
526 bool isCheapToSpeculateCttz() const override {
530 bool isCheapToSpeculateCtlz() const override {
534 bool isCtlzFast() const override {
538 bool hasAndNotCompare(SDValue) const override {
542 bool convertSetCCLogicToBitwiseLogic(EVT VT) const override {
543 return VT.isScalarInteger();
546 bool supportSplitCSR(MachineFunction *MF) const override {
548 MF->getFunction()->getCallingConv() == CallingConv::CXX_FAST_TLS &&
549 MF->getFunction()->hasFnAttribute(Attribute::NoUnwind);
552 void initializeSplitCSR(MachineBasicBlock *Entry) const override;
554 void insertCopiesSplitCSR(
555 MachineBasicBlock *Entry,
556 const SmallVectorImpl<MachineBasicBlock *> &Exits) const override;
558 /// getSetCCResultType - Return the ISD::SETCC ValueType
559 EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context,
560 EVT VT) const override;
562 /// Return true if target always beneficiates from combining into FMA for a
563 /// given value type. This must typically return false on targets where FMA
564 /// takes more cycles to execute than FADD.
565 bool enableAggressiveFMAFusion(EVT VT) const override;
567 /// getPreIndexedAddressParts - returns true by value, base pointer and
568 /// offset pointer and addressing mode by reference if the node's address
569 /// can be legally represented as pre-indexed load / store address.
570 bool getPreIndexedAddressParts(SDNode *N, SDValue &Base,
572 ISD::MemIndexedMode &AM,
573 SelectionDAG &DAG) const override;
575 /// SelectAddressRegReg - Given the specified addressed, check to see if it
576 /// can be represented as an indexed [r+r] operation. Returns false if it
577 /// can be more efficiently represented with [r+imm].
578 bool SelectAddressRegReg(SDValue N, SDValue &Base, SDValue &Index,
579 SelectionDAG &DAG) const;
581 /// SelectAddressRegImm - Returns true if the address N can be represented
582 /// by a base register plus a signed 16-bit displacement [r+imm], and if it
583 /// is not better represented as reg+reg. If Aligned is true, only accept
584 /// displacements suitable for STD and friends, i.e. multiples of 4.
585 bool SelectAddressRegImm(SDValue N, SDValue &Disp, SDValue &Base,
586 SelectionDAG &DAG, bool Aligned) const;
588 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
589 /// represented as an indexed [r+r] operation.
590 bool SelectAddressRegRegOnly(SDValue N, SDValue &Base, SDValue &Index,
591 SelectionDAG &DAG) const;
593 Sched::Preference getSchedulingPreference(SDNode *N) const override;
595 /// LowerOperation - Provide custom lowering hooks for some operations.
597 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
599 /// ReplaceNodeResults - Replace the results of node with an illegal result
600 /// type with new values built out of custom code.
602 void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
603 SelectionDAG &DAG) const override;
605 SDValue expandVSXLoadForLE(SDNode *N, DAGCombinerInfo &DCI) const;
606 SDValue expandVSXStoreForLE(SDNode *N, DAGCombinerInfo &DCI) const;
608 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
610 SDValue BuildSDIVPow2(SDNode *N, const APInt &Divisor, SelectionDAG &DAG,
611 std::vector<SDNode *> *Created) const override;
613 unsigned getRegisterByName(const char* RegName, EVT VT,
614 SelectionDAG &DAG) const override;
616 void computeKnownBitsForTargetNode(const SDValue Op,
618 const APInt &DemandedElts,
619 const SelectionDAG &DAG,
620 unsigned Depth = 0) const override;
622 unsigned getPrefLoopAlignment(MachineLoop *ML) const override;
624 bool shouldInsertFencesForAtomic(const Instruction *I) const override {
628 Instruction *emitLeadingFence(IRBuilder<> &Builder, Instruction *Inst,
629 AtomicOrdering Ord) const override;
630 Instruction *emitTrailingFence(IRBuilder<> &Builder, Instruction *Inst,
631 AtomicOrdering Ord) const override;
634 EmitInstrWithCustomInserter(MachineInstr &MI,
635 MachineBasicBlock *MBB) const override;
636 MachineBasicBlock *EmitAtomicBinary(MachineInstr &MI,
637 MachineBasicBlock *MBB,
640 unsigned CmpOpcode = 0,
641 unsigned CmpPred = 0) const;
642 MachineBasicBlock *EmitPartwordAtomicBinary(MachineInstr &MI,
643 MachineBasicBlock *MBB,
646 unsigned CmpOpcode = 0,
647 unsigned CmpPred = 0) const;
649 MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr &MI,
650 MachineBasicBlock *MBB) const;
652 MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr &MI,
653 MachineBasicBlock *MBB) const;
655 ConstraintType getConstraintType(StringRef Constraint) const override;
657 /// Examine constraint string and operand type and determine a weight value.
658 /// The operand object must already have been set up with the operand type.
659 ConstraintWeight getSingleConstraintMatchWeight(
660 AsmOperandInfo &info, const char *constraint) const override;
662 std::pair<unsigned, const TargetRegisterClass *>
663 getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
664 StringRef Constraint, MVT VT) const override;
666 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
667 /// function arguments in the caller parameter area. This is the actual
668 /// alignment, not its logarithm.
669 unsigned getByValTypeAlignment(Type *Ty,
670 const DataLayout &DL) const override;
672 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
673 /// vector. If it is invalid, don't add anything to Ops.
674 void LowerAsmOperandForConstraint(SDValue Op,
675 std::string &Constraint,
676 std::vector<SDValue> &Ops,
677 SelectionDAG &DAG) const override;
680 getInlineAsmMemConstraint(StringRef ConstraintCode) const override {
681 if (ConstraintCode == "es")
682 return InlineAsm::Constraint_es;
683 else if (ConstraintCode == "o")
684 return InlineAsm::Constraint_o;
685 else if (ConstraintCode == "Q")
686 return InlineAsm::Constraint_Q;
687 else if (ConstraintCode == "Z")
688 return InlineAsm::Constraint_Z;
689 else if (ConstraintCode == "Zy")
690 return InlineAsm::Constraint_Zy;
691 return TargetLowering::getInlineAsmMemConstraint(ConstraintCode);
694 /// isLegalAddressingMode - Return true if the addressing mode represented
695 /// by AM is legal for this target, for a load/store of the specified type.
696 bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM,
697 Type *Ty, unsigned AS) const override;
699 /// isLegalICmpImmediate - Return true if the specified immediate is legal
700 /// icmp immediate, that is the target has icmp instructions which can
701 /// compare a register against the immediate without having to materialize
702 /// the immediate into a register.
703 bool isLegalICmpImmediate(int64_t Imm) const override;
705 /// isLegalAddImmediate - Return true if the specified immediate is legal
706 /// add immediate, that is the target has add instructions which can
707 /// add a register and the immediate without having to materialize
708 /// the immediate into a register.
709 bool isLegalAddImmediate(int64_t Imm) const override;
711 /// isTruncateFree - Return true if it's free to truncate a value of
712 /// type Ty1 to type Ty2. e.g. On PPC it's free to truncate a i64 value in
713 /// register X1 to i32 by referencing its sub-register R1.
714 bool isTruncateFree(Type *Ty1, Type *Ty2) const override;
715 bool isTruncateFree(EVT VT1, EVT VT2) const override;
717 bool isZExtFree(SDValue Val, EVT VT2) const override;
719 bool isFPExtFree(EVT VT) const override;
721 /// \brief Returns true if it is beneficial to convert a load of a constant
722 /// to just the constant itself.
723 bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
724 Type *Ty) const override;
726 bool convertSelectOfConstantsToMath() const override {
730 bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const override;
732 bool getTgtMemIntrinsic(IntrinsicInfo &Info,
734 unsigned Intrinsic) const override;
736 /// getOptimalMemOpType - Returns the target specific optimal type for load
737 /// and store operations as a result of memset, memcpy, and memmove
738 /// lowering. If DstAlign is zero that means it's safe to destination
739 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
740 /// means there isn't a need to check it against alignment requirement,
741 /// probably because the source does not need to be loaded. If 'IsMemset' is
742 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
743 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
744 /// source is constant so it does not need to be loaded.
745 /// It returns EVT::Other if the type should be determined using generic
746 /// target-independent logic.
748 getOptimalMemOpType(uint64_t Size, unsigned DstAlign, unsigned SrcAlign,
749 bool IsMemset, bool ZeroMemset, bool MemcpyStrSrc,
750 MachineFunction &MF) const override;
752 /// Is unaligned memory access allowed for the given type, and is it fast
753 /// relative to software emulation.
754 bool allowsMisalignedMemoryAccesses(EVT VT,
757 bool *Fast = nullptr) const override;
759 /// isFMAFasterThanFMulAndFAdd - Return true if an FMA operation is faster
760 /// than a pair of fmul and fadd instructions. fmuladd intrinsics will be
761 /// expanded to FMAs when this method returns true, otherwise fmuladd is
762 /// expanded to fmul + fadd.
763 bool isFMAFasterThanFMulAndFAdd(EVT VT) const override;
765 const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const override;
767 // Should we expand the build vector with shuffles?
769 shouldExpandBuildVectorWithShuffles(EVT VT,
770 unsigned DefinedValues) const override;
772 /// createFastISel - This method returns a target-specific FastISel object,
773 /// or null if the target does not support "fast" instruction selection.
774 FastISel *createFastISel(FunctionLoweringInfo &FuncInfo,
775 const TargetLibraryInfo *LibInfo) const override;
777 /// \brief Returns true if an argument of type Ty needs to be passed in a
778 /// contiguous block of registers in calling convention CallConv.
779 bool functionArgumentNeedsConsecutiveRegisters(
780 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const override {
781 // We support any array type as "consecutive" block in the parameter
782 // save area. The element type defines the alignment requirement and
783 // whether the argument should go in GPRs, FPRs, or VRs if available.
785 // Note that clang uses this capability both to implement the ELFv2
786 // homogeneous float/vector aggregate ABI, and to avoid having to use
787 // "byval" when passing aggregates that might fully fit in registers.
788 return Ty->isArrayTy();
791 /// If a physical register, this returns the register that receives the
792 /// exception address on entry to an EH pad.
794 getExceptionPointerRegister(const Constant *PersonalityFn) const override;
796 /// If a physical register, this returns the register that receives the
797 /// exception typeid on entry to a landing pad.
799 getExceptionSelectorRegister(const Constant *PersonalityFn) const override;
801 /// Override to support customized stack guard loading.
802 bool useLoadStackGuardNode() const override;
803 void insertSSPDeclarations(Module &M) const override;
805 bool isFPImmLegal(const APFloat &Imm, EVT VT) const override;
807 unsigned getJumpTableEncoding() const override;
808 bool isJumpTableRelative() const override;
809 SDValue getPICJumpTableRelocBase(SDValue Table,
810 SelectionDAG &DAG) const override;
811 const MCExpr *getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
813 MCContext &Ctx) const override;
816 struct ReuseLoadInfo {
820 MachinePointerInfo MPI;
821 bool IsDereferenceable = false;
822 bool IsInvariant = false;
823 unsigned Alignment = 0;
825 const MDNode *Ranges = nullptr;
827 ReuseLoadInfo() = default;
829 MachineMemOperand::Flags MMOFlags() const {
830 MachineMemOperand::Flags F = MachineMemOperand::MONone;
831 if (IsDereferenceable)
832 F |= MachineMemOperand::MODereferenceable;
834 F |= MachineMemOperand::MOInvariant;
839 bool canReuseLoadAddress(SDValue Op, EVT MemVT, ReuseLoadInfo &RLI,
841 ISD::LoadExtType ET = ISD::NON_EXTLOAD) const;
842 void spliceIntoChain(SDValue ResChain, SDValue NewResChain,
843 SelectionDAG &DAG) const;
845 void LowerFP_TO_INTForReuse(SDValue Op, ReuseLoadInfo &RLI,
846 SelectionDAG &DAG, const SDLoc &dl) const;
847 SDValue LowerFP_TO_INTDirectMove(SDValue Op, SelectionDAG &DAG,
848 const SDLoc &dl) const;
850 bool directMoveIsProfitable(const SDValue &Op) const;
851 SDValue LowerINT_TO_FPDirectMove(SDValue Op, SelectionDAG &DAG,
852 const SDLoc &dl) const;
854 SDValue getFramePointerFrameIndex(SelectionDAG & DAG) const;
855 SDValue getReturnAddrFrameIndex(SelectionDAG & DAG) const;
858 IsEligibleForTailCallOptimization(SDValue Callee,
859 CallingConv::ID CalleeCC,
861 const SmallVectorImpl<ISD::InputArg> &Ins,
862 SelectionDAG& DAG) const;
865 IsEligibleForTailCallOptimization_64SVR4(
867 CallingConv::ID CalleeCC,
868 ImmutableCallSite *CS,
870 const SmallVectorImpl<ISD::OutputArg> &Outs,
871 const SmallVectorImpl<ISD::InputArg> &Ins,
872 SelectionDAG& DAG) const;
874 SDValue EmitTailCallLoadFPAndRetAddr(SelectionDAG &DAG, int SPDiff,
875 SDValue Chain, SDValue &LROpOut,
877 const SDLoc &dl) const;
879 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
880 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
881 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
882 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
883 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
884 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
885 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
886 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
887 SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
888 SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
889 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
890 SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const;
891 SDValue LowerVACOPY(SDValue Op, SelectionDAG &DAG) const;
892 SDValue LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG) const;
893 SDValue LowerGET_DYNAMIC_AREA_OFFSET(SDValue Op, SelectionDAG &DAG) const;
894 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
895 SDValue LowerEH_DWARF_CFA(SDValue Op, SelectionDAG &DAG) const;
896 SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const;
897 SDValue LowerSTORE(SDValue Op, SelectionDAG &DAG) const;
898 SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const;
899 SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
900 SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
901 const SDLoc &dl) const;
902 SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
903 SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
904 SDValue LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const;
905 SDValue LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const;
906 SDValue LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const;
907 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
908 SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
909 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
910 SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
911 SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const;
912 SDValue LowerINTRINSIC_VOID(SDValue Op, SelectionDAG &DAG) const;
913 SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const;
914 SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const;
915 SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) const;
917 SDValue LowerVectorLoad(SDValue Op, SelectionDAG &DAG) const;
918 SDValue LowerVectorStore(SDValue Op, SelectionDAG &DAG) const;
920 SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
921 CallingConv::ID CallConv, bool isVarArg,
922 const SmallVectorImpl<ISD::InputArg> &Ins,
923 const SDLoc &dl, SelectionDAG &DAG,
924 SmallVectorImpl<SDValue> &InVals) const;
925 SDValue FinishCall(CallingConv::ID CallConv, const SDLoc &dl,
926 bool isTailCall, bool isVarArg, bool isPatchPoint,
927 bool hasNest, SelectionDAG &DAG,
928 SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
929 SDValue InFlag, SDValue Chain, SDValue CallSeqStart,
930 SDValue &Callee, int SPDiff, unsigned NumBytes,
931 const SmallVectorImpl<ISD::InputArg> &Ins,
932 SmallVectorImpl<SDValue> &InVals,
933 ImmutableCallSite *CS) const;
936 LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
937 const SmallVectorImpl<ISD::InputArg> &Ins,
938 const SDLoc &dl, SelectionDAG &DAG,
939 SmallVectorImpl<SDValue> &InVals) const override;
941 SDValue LowerCall(TargetLowering::CallLoweringInfo &CLI,
942 SmallVectorImpl<SDValue> &InVals) const override;
944 bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
946 const SmallVectorImpl<ISD::OutputArg> &Outs,
947 LLVMContext &Context) const override;
949 SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
950 const SmallVectorImpl<ISD::OutputArg> &Outs,
951 const SmallVectorImpl<SDValue> &OutVals,
952 const SDLoc &dl, SelectionDAG &DAG) const override;
954 SDValue extendArgForPPC64(ISD::ArgFlagsTy Flags, EVT ObjectVT,
955 SelectionDAG &DAG, SDValue ArgVal,
956 const SDLoc &dl) const;
958 SDValue LowerFormalArguments_Darwin(
959 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
960 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
961 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const;
962 SDValue LowerFormalArguments_64SVR4(
963 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
964 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
965 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const;
966 SDValue LowerFormalArguments_32SVR4(
967 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
968 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
969 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const;
971 SDValue createMemcpyOutsideCallSeq(SDValue Arg, SDValue PtrOff,
972 SDValue CallSeqStart,
973 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
974 const SDLoc &dl) const;
976 SDValue LowerCall_Darwin(SDValue Chain, SDValue Callee,
977 CallingConv::ID CallConv, bool isVarArg,
978 bool isTailCall, bool isPatchPoint,
979 const SmallVectorImpl<ISD::OutputArg> &Outs,
980 const SmallVectorImpl<SDValue> &OutVals,
981 const SmallVectorImpl<ISD::InputArg> &Ins,
982 const SDLoc &dl, SelectionDAG &DAG,
983 SmallVectorImpl<SDValue> &InVals,
984 ImmutableCallSite *CS) const;
985 SDValue LowerCall_64SVR4(SDValue Chain, SDValue Callee,
986 CallingConv::ID CallConv, bool isVarArg,
987 bool isTailCall, bool isPatchPoint,
988 const SmallVectorImpl<ISD::OutputArg> &Outs,
989 const SmallVectorImpl<SDValue> &OutVals,
990 const SmallVectorImpl<ISD::InputArg> &Ins,
991 const SDLoc &dl, SelectionDAG &DAG,
992 SmallVectorImpl<SDValue> &InVals,
993 ImmutableCallSite *CS) const;
994 SDValue LowerCall_32SVR4(SDValue Chain, SDValue Callee,
995 CallingConv::ID CallConv, bool isVarArg,
996 bool isTailCall, bool isPatchPoint,
997 const SmallVectorImpl<ISD::OutputArg> &Outs,
998 const SmallVectorImpl<SDValue> &OutVals,
999 const SmallVectorImpl<ISD::InputArg> &Ins,
1000 const SDLoc &dl, SelectionDAG &DAG,
1001 SmallVectorImpl<SDValue> &InVals,
1002 ImmutableCallSite *CS) const;
1004 SDValue lowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
1005 SDValue lowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
1007 SDValue DAGCombineExtBoolTrunc(SDNode *N, DAGCombinerInfo &DCI) const;
1008 SDValue DAGCombineBuildVector(SDNode *N, DAGCombinerInfo &DCI) const;
1009 SDValue DAGCombineTruncBoolExt(SDNode *N, DAGCombinerInfo &DCI) const;
1010 SDValue combineFPToIntToFP(SDNode *N, DAGCombinerInfo &DCI) const;
1011 SDValue combineSHL(SDNode *N, DAGCombinerInfo &DCI) const;
1012 SDValue combineSRA(SDNode *N, DAGCombinerInfo &DCI) const;
1013 SDValue combineSRL(SDNode *N, DAGCombinerInfo &DCI) const;
1015 /// ConvertSETCCToSubtract - looks at SETCC that compares ints. It replaces
1016 /// SETCC with integer subtraction when (1) there is a legal way of doing it
1017 /// (2) keeping the result of comparison in GPR has performance benefit.
1018 SDValue ConvertSETCCToSubtract(SDNode *N, DAGCombinerInfo &DCI) const;
1020 SDValue getSqrtEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
1021 int &RefinementSteps, bool &UseOneConstNR,
1022 bool Reciprocal) const override;
1023 SDValue getRecipEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
1024 int &RefinementSteps) const override;
1025 unsigned combineRepeatedFPDivisors() const override;
1027 CCAssignFn *useFastISelCCs(unsigned Flag) const;
1030 combineElementTruncationToVectorTruncation(SDNode *N,
1031 DAGCombinerInfo &DCI) const;
1036 FastISel *createFastISel(FunctionLoweringInfo &FuncInfo,
1037 const TargetLibraryInfo *LibInfo);
1039 } // end namespace PPC
1041 bool CC_PPC32_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
1042 CCValAssign::LocInfo &LocInfo,
1043 ISD::ArgFlagsTy &ArgFlags,
1046 bool CC_PPC32_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
1048 CCValAssign::LocInfo &LocInfo,
1049 ISD::ArgFlagsTy &ArgFlags,
1053 CC_PPC32_SVR4_Custom_SkipLastArgRegsPPCF128(unsigned &ValNo, MVT &ValVT,
1055 CCValAssign::LocInfo &LocInfo,
1056 ISD::ArgFlagsTy &ArgFlags,
1059 bool CC_PPC32_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
1061 CCValAssign::LocInfo &LocInfo,
1062 ISD::ArgFlagsTy &ArgFlags,
1065 } // end namespace llvm
1067 #endif // LLVM_TARGET_POWERPC_PPC32ISELLOWERING_H