1 //===-- PPCISelLowering.h - PPC32 DAG Lowering Interface --------*- C++ -*-===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file defines the interfaces that PPC uses to lower LLVM code into a
11 // selection DAG.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #ifndef LLVM_LIB_TARGET_POWERPC_PPCISELLOWERING_H
16 #define LLVM_LIB_TARGET_POWERPC_PPCISELLOWERING_H
17 
18 #include "PPC.h"
19 #include "PPCInstrInfo.h"
20 #include "PPCRegisterInfo.h"
21 #include "llvm/CodeGen/CallingConvLower.h"
22 #include "llvm/CodeGen/SelectionDAG.h"
23 #include "llvm/Target/TargetLowering.h"
24 
25 namespace llvm {
26   namespace PPCISD {
27     enum NodeType : unsigned {
28       // Start the numbering where the builtin ops and target ops leave off.
29       FIRST_NUMBER = ISD::BUILTIN_OP_END,
30 
31       /// FSEL - Traditional three-operand fsel node.
32       ///
33       FSEL,
34 
35       /// FCFID - The FCFID instruction, taking an f64 operand and producing
36       /// and f64 value containing the FP representation of the integer that
37       /// was temporarily in the f64 operand.
38       FCFID,
39 
40       /// Newer FCFID[US] integer-to-floating-point conversion instructions for
41       /// unsigned integers and single-precision outputs.
42       FCFIDU, FCFIDS, FCFIDUS,
43 
44       /// FCTI[D,W]Z - The FCTIDZ and FCTIWZ instructions, taking an f32 or f64
45       /// operand, producing an f64 value containing the integer representation
46       /// of that FP value.
47       FCTIDZ, FCTIWZ,
48 
49       /// Newer FCTI[D,W]UZ floating-point-to-integer conversion instructions for
50       /// unsigned integers.
51       FCTIDUZ, FCTIWUZ,
52 
53       /// Reciprocal estimate instructions (unary FP ops).
54       FRE, FRSQRTE,
55 
56       // VMADDFP, VNMSUBFP - The VMADDFP and VNMSUBFP instructions, taking
57       // three v4f32 operands and producing a v4f32 result.
58       VMADDFP, VNMSUBFP,
59 
60       /// VPERM - The PPC VPERM Instruction.
61       ///
62       VPERM,
63 
64       /// XXSPLT - The PPC VSX splat instructions
65       ///
66       XXSPLT,
67 
68       /// The CMPB instruction (takes two operands of i32 or i64).
69       CMPB,
70 
71       /// Hi/Lo - These represent the high and low 16-bit parts of a global
72       /// address respectively.  These nodes have two operands, the first of
73       /// which must be a TargetGlobalAddress, and the second of which must be a
74       /// Constant.  Selected naively, these turn into 'lis G+C' and 'li G+C',
75       /// though these are usually folded into other nodes.
76       Hi, Lo,
77 
78       /// The following two target-specific nodes are used for calls through
79       /// function pointers in the 64-bit SVR4 ABI.
80 
81       /// OPRC, CHAIN = DYNALLOC(CHAIN, NEGSIZE, FRAME_INDEX)
82       /// This instruction is lowered in PPCRegisterInfo::eliminateFrameIndex to
83       /// compute an allocation on the stack.
84       DYNALLOC,
85 
86       /// This instruction is lowered in PPCRegisterInfo::eliminateFrameIndex to
87       /// compute an offset from native SP to the address  of the most recent
88       /// dynamic alloca.
89       DYNAREAOFFSET,
90 
91       /// GlobalBaseReg - On Darwin, this node represents the result of the mflr
92       /// at function entry, used for PIC code.
93       GlobalBaseReg,
94 
95       /// These nodes represent the 32-bit PPC shifts that operate on 6-bit
96       /// shift amounts.  These nodes are generated by the multi-precision shift
97       /// code.
98       SRL, SRA, SHL,
99 
100       /// The combination of sra[wd]i and addze used to implemented signed
101       /// integer division by a power of 2. The first operand is the dividend,
102       /// and the second is the constant shift amount (representing the
103       /// divisor).
104       SRA_ADDZE,
105 
106       /// CALL - A direct function call.
107       /// CALL_NOP is a call with the special NOP which follows 64-bit
108       /// SVR4 calls.
109       CALL, CALL_NOP,
110 
111       /// CHAIN,FLAG = MTCTR(VAL, CHAIN[, INFLAG]) - Directly corresponds to a
112       /// MTCTR instruction.
113       MTCTR,
114 
115       /// CHAIN,FLAG = BCTRL(CHAIN, INFLAG) - Directly corresponds to a
116       /// BCTRL instruction.
117       BCTRL,
118 
119       /// CHAIN,FLAG = BCTRL(CHAIN, ADDR, INFLAG) - The combination of a bctrl
120       /// instruction and the TOC reload required on SVR4 PPC64.
121       BCTRL_LOAD_TOC,
122 
123       /// Return with a flag operand, matched by 'blr'
124       RET_FLAG,
125 
126       /// R32 = MFOCRF(CRREG, INFLAG) - Represents the MFOCRF instruction.
127       /// This copies the bits corresponding to the specified CRREG into the
128       /// resultant GPR.  Bits corresponding to other CR regs are undefined.
129       MFOCRF,
130 
131       /// Direct move from a VSX register to a GPR
132       MFVSR,
133 
134       /// Direct move from a GPR to a VSX register (algebraic)
135       MTVSRA,
136 
137       /// Direct move from a GPR to a VSX register (zero)
138       MTVSRZ,
139 
140       // FIXME: Remove these once the ANDI glue bug is fixed:
141       /// i1 = ANDIo_1_[EQ|GT]_BIT(i32 or i64 x) - Represents the result of the
142       /// eq or gt bit of CR0 after executing andi. x, 1. This is used to
143       /// implement truncation of i32 or i64 to i1.
144       ANDIo_1_EQ_BIT, ANDIo_1_GT_BIT,
145 
146       // READ_TIME_BASE - A read of the 64-bit time-base register on a 32-bit
147       // target (returns (Lo, Hi)). It takes a chain operand.
148       READ_TIME_BASE,
149 
150       // EH_SJLJ_SETJMP - SjLj exception handling setjmp.
151       EH_SJLJ_SETJMP,
152 
153       // EH_SJLJ_LONGJMP - SjLj exception handling longjmp.
154       EH_SJLJ_LONGJMP,
155 
156       /// RESVEC = VCMP(LHS, RHS, OPC) - Represents one of the altivec VCMP*
157       /// instructions.  For lack of better number, we use the opcode number
158       /// encoding for the OPC field to identify the compare.  For example, 838
159       /// is VCMPGTSH.
160       VCMP,
161 
162       /// RESVEC, OUTFLAG = VCMPo(LHS, RHS, OPC) - Represents one of the
163       /// altivec VCMP*o instructions.  For lack of better number, we use the
164       /// opcode number encoding for the OPC field to identify the compare.  For
165       /// example, 838 is VCMPGTSH.
166       VCMPo,
167 
168       /// CHAIN = COND_BRANCH CHAIN, CRRC, OPC, DESTBB [, INFLAG] - This
169       /// corresponds to the COND_BRANCH pseudo instruction.  CRRC is the
170       /// condition register to branch on, OPC is the branch opcode to use (e.g.
171       /// PPC::BLE), DESTBB is the destination block to branch to, and INFLAG is
172       /// an optional input flag argument.
173       COND_BRANCH,
174 
175       /// CHAIN = BDNZ CHAIN, DESTBB - These are used to create counter-based
176       /// loops.
177       BDNZ, BDZ,
178 
179       /// F8RC = FADDRTZ F8RC, F8RC - This is an FADD done with rounding
180       /// towards zero.  Used only as part of the long double-to-int
181       /// conversion sequence.
182       FADDRTZ,
183 
184       /// F8RC = MFFS - This moves the FPSCR (not modeled) into the register.
185       MFFS,
186 
187       /// TC_RETURN - A tail call return.
188       ///   operand #0 chain
189       ///   operand #1 callee (register or absolute)
190       ///   operand #2 stack adjustment
191       ///   operand #3 optional in flag
192       TC_RETURN,
193 
194       /// ch, gl = CR6[UN]SET ch, inglue - Toggle CR bit 6 for SVR4 vararg calls
195       CR6SET,
196       CR6UNSET,
197 
198       /// GPRC = address of _GLOBAL_OFFSET_TABLE_. Used by initial-exec TLS
199       /// on PPC32.
200       PPC32_GOT,
201 
202       /// GPRC = address of _GLOBAL_OFFSET_TABLE_. Used by general dynamic and
203       /// local dynamic TLS on PPC32.
204       PPC32_PICGOT,
205 
206       /// G8RC = ADDIS_GOT_TPREL_HA %X2, Symbol - Used by the initial-exec
207       /// TLS model, produces an ADDIS8 instruction that adds the GOT
208       /// base to sym\@got\@tprel\@ha.
209       ADDIS_GOT_TPREL_HA,
210 
211       /// G8RC = LD_GOT_TPREL_L Symbol, G8RReg - Used by the initial-exec
212       /// TLS model, produces a LD instruction with base register G8RReg
213       /// and offset sym\@got\@tprel\@l.  This completes the addition that
214       /// finds the offset of "sym" relative to the thread pointer.
215       LD_GOT_TPREL_L,
216 
217       /// G8RC = ADD_TLS G8RReg, Symbol - Used by the initial-exec TLS
218       /// model, produces an ADD instruction that adds the contents of
219       /// G8RReg to the thread pointer.  Symbol contains a relocation
220       /// sym\@tls which is to be replaced by the thread pointer and
221       /// identifies to the linker that the instruction is part of a
222       /// TLS sequence.
223       ADD_TLS,
224 
225       /// G8RC = ADDIS_TLSGD_HA %X2, Symbol - For the general-dynamic TLS
226       /// model, produces an ADDIS8 instruction that adds the GOT base
227       /// register to sym\@got\@tlsgd\@ha.
228       ADDIS_TLSGD_HA,
229 
230       /// %X3 = ADDI_TLSGD_L G8RReg, Symbol - For the general-dynamic TLS
231       /// model, produces an ADDI8 instruction that adds G8RReg to
232       /// sym\@got\@tlsgd\@l and stores the result in X3.  Hidden by
233       /// ADDIS_TLSGD_L_ADDR until after register assignment.
234       ADDI_TLSGD_L,
235 
236       /// %X3 = GET_TLS_ADDR %X3, Symbol - For the general-dynamic TLS
237       /// model, produces a call to __tls_get_addr(sym\@tlsgd).  Hidden by
238       /// ADDIS_TLSGD_L_ADDR until after register assignment.
239       GET_TLS_ADDR,
240 
241       /// G8RC = ADDI_TLSGD_L_ADDR G8RReg, Symbol, Symbol - Op that
242       /// combines ADDI_TLSGD_L and GET_TLS_ADDR until expansion following
243       /// register assignment.
244       ADDI_TLSGD_L_ADDR,
245 
246       /// G8RC = ADDIS_TLSLD_HA %X2, Symbol - For the local-dynamic TLS
247       /// model, produces an ADDIS8 instruction that adds the GOT base
248       /// register to sym\@got\@tlsld\@ha.
249       ADDIS_TLSLD_HA,
250 
251       /// %X3 = ADDI_TLSLD_L G8RReg, Symbol - For the local-dynamic TLS
252       /// model, produces an ADDI8 instruction that adds G8RReg to
253       /// sym\@got\@tlsld\@l and stores the result in X3.  Hidden by
254       /// ADDIS_TLSLD_L_ADDR until after register assignment.
255       ADDI_TLSLD_L,
256 
257       /// %X3 = GET_TLSLD_ADDR %X3, Symbol - For the local-dynamic TLS
258       /// model, produces a call to __tls_get_addr(sym\@tlsld).  Hidden by
259       /// ADDIS_TLSLD_L_ADDR until after register assignment.
260       GET_TLSLD_ADDR,
261 
262       /// G8RC = ADDI_TLSLD_L_ADDR G8RReg, Symbol, Symbol - Op that
263       /// combines ADDI_TLSLD_L and GET_TLSLD_ADDR until expansion
264       /// following register assignment.
265       ADDI_TLSLD_L_ADDR,
266 
267       /// G8RC = ADDIS_DTPREL_HA %X3, Symbol - For the local-dynamic TLS
268       /// model, produces an ADDIS8 instruction that adds X3 to
269       /// sym\@dtprel\@ha.
270       ADDIS_DTPREL_HA,
271 
272       /// G8RC = ADDI_DTPREL_L G8RReg, Symbol - For the local-dynamic TLS
273       /// model, produces an ADDI8 instruction that adds G8RReg to
274       /// sym\@got\@dtprel\@l.
275       ADDI_DTPREL_L,
276 
277       /// VRRC = VADD_SPLAT Elt, EltSize - Temporary node to be expanded
278       /// during instruction selection to optimize a BUILD_VECTOR into
279       /// operations on splats.  This is necessary to avoid losing these
280       /// optimizations due to constant folding.
281       VADD_SPLAT,
282 
283       /// CHAIN = SC CHAIN, Imm128 - System call.  The 7-bit unsigned
284       /// operand identifies the operating system entry point.
285       SC,
286 
287       /// CHAIN = CLRBHRB CHAIN - Clear branch history rolling buffer.
288       CLRBHRB,
289 
290       /// GPRC, CHAIN = MFBHRBE CHAIN, Entry, Dummy - Move from branch
291       /// history rolling buffer entry.
292       MFBHRBE,
293 
294       /// CHAIN = RFEBB CHAIN, State - Return from event-based branch.
295       RFEBB,
296 
297       /// VSRC, CHAIN = XXSWAPD CHAIN, VSRC - Occurs only for little
298       /// endian.  Maps to an xxswapd instruction that corrects an lxvd2x
299       /// or stxvd2x instruction.  The chain is necessary because the
300       /// sequence replaces a load and needs to provide the same number
301       /// of outputs.
302       XXSWAPD,
303 
304       /// QVFPERM = This corresponds to the QPX qvfperm instruction.
305       QVFPERM,
306 
307       /// QVGPCI = This corresponds to the QPX qvgpci instruction.
308       QVGPCI,
309 
310       /// QVALIGNI = This corresponds to the QPX qvaligni instruction.
311       QVALIGNI,
312 
313       /// QVESPLATI = This corresponds to the QPX qvesplati instruction.
314       QVESPLATI,
315 
316       /// QBFLT = Access the underlying QPX floating-point boolean
317       /// representation.
318       QBFLT,
319 
320       /// CHAIN = STBRX CHAIN, GPRC, Ptr, Type - This is a
321       /// byte-swapping store instruction.  It byte-swaps the low "Type" bits of
322       /// the GPRC input, then stores it through Ptr.  Type can be either i16 or
323       /// i32.
324       STBRX = ISD::FIRST_TARGET_MEMORY_OPCODE,
325 
326       /// GPRC, CHAIN = LBRX CHAIN, Ptr, Type - This is a
327       /// byte-swapping load instruction.  It loads "Type" bits, byte swaps it,
328       /// then puts it in the bottom bits of the GPRC.  TYPE can be either i16
329       /// or i32.
330       LBRX,
331 
332       /// STFIWX - The STFIWX instruction.  The first operand is an input token
333       /// chain, then an f64 value to store, then an address to store it to.
334       STFIWX,
335 
336       /// GPRC, CHAIN = LFIWAX CHAIN, Ptr - This is a floating-point
337       /// load which sign-extends from a 32-bit integer value into the
338       /// destination 64-bit register.
339       LFIWAX,
340 
341       /// GPRC, CHAIN = LFIWZX CHAIN, Ptr - This is a floating-point
342       /// load which zero-extends from a 32-bit integer value into the
343       /// destination 64-bit register.
344       LFIWZX,
345 
346       /// VSRC, CHAIN = LXVD2X_LE CHAIN, Ptr - Occurs only for little endian.
347       /// Maps directly to an lxvd2x instruction that will be followed by
348       /// an xxswapd.
349       LXVD2X,
350 
351       /// CHAIN = STXVD2X CHAIN, VSRC, Ptr - Occurs only for little endian.
352       /// Maps directly to an stxvd2x instruction that will be preceded by
353       /// an xxswapd.
354       STXVD2X,
355 
356       /// QBRC, CHAIN = QVLFSb CHAIN, Ptr
357       /// The 4xf32 load used for v4i1 constants.
358       QVLFSb,
359 
360       /// GPRC = TOC_ENTRY GA, TOC
361       /// Loads the entry for GA from the TOC, where the TOC base is given by
362       /// the last operand.
363       TOC_ENTRY
364     };
365   }
366 
367   /// Define some predicates that are used for node matching.
368   namespace PPC {
369     /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
370     /// VPKUHUM instruction.
371     bool isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
372                               SelectionDAG &DAG);
373 
374     /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
375     /// VPKUWUM instruction.
376     bool isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
377                               SelectionDAG &DAG);
378 
379     /// isVPKUDUMShuffleMask - Return true if this is the shuffle mask for a
380     /// VPKUDUM instruction.
381     bool isVPKUDUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
382                               SelectionDAG &DAG);
383 
384     /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
385     /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
386     bool isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
387                             unsigned ShuffleKind, SelectionDAG &DAG);
388 
389     /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
390     /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
391     bool isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
392                             unsigned ShuffleKind, SelectionDAG &DAG);
393 
394     /// isVMRGEOShuffleMask - Return true if this is a shuffle mask suitable for
395     /// a VMRGEW or VMRGOW instruction
396     bool isVMRGEOShuffleMask(ShuffleVectorSDNode *N, bool CheckEven,
397                              unsigned ShuffleKind, SelectionDAG &DAG);
398 
399     /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the
400     /// shift amount, otherwise return -1.
401     int isVSLDOIShuffleMask(SDNode *N, unsigned ShuffleKind,
402                             SelectionDAG &DAG);
403 
404     /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
405     /// specifies a splat of a single element that is suitable for input to
406     /// VSPLTB/VSPLTH/VSPLTW.
407     bool isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize);
408 
409     /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
410     /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
411     unsigned getVSPLTImmediate(SDNode *N, unsigned EltSize, SelectionDAG &DAG);
412 
413     /// get_VSPLTI_elt - If this is a build_vector of constants which can be
414     /// formed by using a vspltis[bhw] instruction of the specified element
415     /// size, return the constant being splatted.  The ByteSize field indicates
416     /// the number of bytes of each element [124] -> [bhw].
417     SDValue get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG);
418 
419     /// If this is a qvaligni shuffle mask, return the shift
420     /// amount, otherwise return -1.
421     int isQVALIGNIShuffleMask(SDNode *N);
422   }
423 
424   class PPCTargetLowering : public TargetLowering {
425     const PPCSubtarget &Subtarget;
426 
427   public:
428     explicit PPCTargetLowering(const PPCTargetMachine &TM,
429                                const PPCSubtarget &STI);
430 
431     /// getTargetNodeName() - This method returns the name of a target specific
432     /// DAG node.
433     const char *getTargetNodeName(unsigned Opcode) const override;
434 
435     bool useSoftFloat() const override;
436 
437     MVT getScalarShiftAmountTy(const DataLayout &, EVT) const override {
438       return MVT::i32;
439     }
440 
441     bool isCheapToSpeculateCttz() const override {
442       return true;
443     }
444 
445     bool isCheapToSpeculateCtlz() const override {
446       return true;
447     }
448 
449     bool supportSplitCSR(MachineFunction *MF) const override {
450       return
451         MF->getFunction()->getCallingConv() == CallingConv::CXX_FAST_TLS &&
452         MF->getFunction()->hasFnAttribute(Attribute::NoUnwind);
453     }
454 
455     void initializeSplitCSR(MachineBasicBlock *Entry) const override;
456 
457     void insertCopiesSplitCSR(
458       MachineBasicBlock *Entry,
459       const SmallVectorImpl<MachineBasicBlock *> &Exits) const override;
460 
461     /// getSetCCResultType - Return the ISD::SETCC ValueType
462     EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context,
463                            EVT VT) const override;
464 
465     /// Return true if target always beneficiates from combining into FMA for a
466     /// given value type. This must typically return false on targets where FMA
467     /// takes more cycles to execute than FADD.
468     bool enableAggressiveFMAFusion(EVT VT) const override;
469 
470     /// getPreIndexedAddressParts - returns true by value, base pointer and
471     /// offset pointer and addressing mode by reference if the node's address
472     /// can be legally represented as pre-indexed load / store address.
473     bool getPreIndexedAddressParts(SDNode *N, SDValue &Base,
474                                    SDValue &Offset,
475                                    ISD::MemIndexedMode &AM,
476                                    SelectionDAG &DAG) const override;
477 
478     /// SelectAddressRegReg - Given the specified addressed, check to see if it
479     /// can be represented as an indexed [r+r] operation.  Returns false if it
480     /// can be more efficiently represented with [r+imm].
481     bool SelectAddressRegReg(SDValue N, SDValue &Base, SDValue &Index,
482                              SelectionDAG &DAG) const;
483 
484     /// SelectAddressRegImm - Returns true if the address N can be represented
485     /// by a base register plus a signed 16-bit displacement [r+imm], and if it
486     /// is not better represented as reg+reg.  If Aligned is true, only accept
487     /// displacements suitable for STD and friends, i.e. multiples of 4.
488     bool SelectAddressRegImm(SDValue N, SDValue &Disp, SDValue &Base,
489                              SelectionDAG &DAG, bool Aligned) const;
490 
491     /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
492     /// represented as an indexed [r+r] operation.
493     bool SelectAddressRegRegOnly(SDValue N, SDValue &Base, SDValue &Index,
494                                  SelectionDAG &DAG) const;
495 
496     Sched::Preference getSchedulingPreference(SDNode *N) const override;
497 
498     /// LowerOperation - Provide custom lowering hooks for some operations.
499     ///
500     SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
501 
502     /// ReplaceNodeResults - Replace the results of node with an illegal result
503     /// type with new values built out of custom code.
504     ///
505     void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
506                             SelectionDAG &DAG) const override;
507 
508     SDValue expandVSXLoadForLE(SDNode *N, DAGCombinerInfo &DCI) const;
509     SDValue expandVSXStoreForLE(SDNode *N, DAGCombinerInfo &DCI) const;
510 
511     SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
512 
513     SDValue BuildSDIVPow2(SDNode *N, const APInt &Divisor, SelectionDAG &DAG,
514                           std::vector<SDNode *> *Created) const override;
515 
516     unsigned getRegisterByName(const char* RegName, EVT VT,
517                                SelectionDAG &DAG) const override;
518 
519     void computeKnownBitsForTargetNode(const SDValue Op,
520                                        APInt &KnownZero,
521                                        APInt &KnownOne,
522                                        const SelectionDAG &DAG,
523                                        unsigned Depth = 0) const override;
524 
525     unsigned getPrefLoopAlignment(MachineLoop *ML) const override;
526 
527     bool shouldInsertFencesForAtomic(const Instruction *I) const override {
528       return true;
529     }
530 
531     Instruction* emitLeadingFence(IRBuilder<> &Builder, AtomicOrdering Ord,
532                                   bool IsStore, bool IsLoad) const override;
533     Instruction* emitTrailingFence(IRBuilder<> &Builder, AtomicOrdering Ord,
534                                    bool IsStore, bool IsLoad) const override;
535 
536     MachineBasicBlock *
537       EmitInstrWithCustomInserter(MachineInstr *MI,
538                                   MachineBasicBlock *MBB) const override;
539     MachineBasicBlock *EmitAtomicBinary(MachineInstr *MI,
540                                         MachineBasicBlock *MBB,
541                                         unsigned AtomicSize,
542                                         unsigned BinOpcode) const;
543     MachineBasicBlock *EmitPartwordAtomicBinary(MachineInstr *MI,
544                                                 MachineBasicBlock *MBB,
545                                             bool is8bit, unsigned Opcode) const;
546 
547     MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr *MI,
548                                         MachineBasicBlock *MBB) const;
549 
550     MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr *MI,
551                                          MachineBasicBlock *MBB) const;
552 
553     ConstraintType getConstraintType(StringRef Constraint) const override;
554 
555     /// Examine constraint string and operand type and determine a weight value.
556     /// The operand object must already have been set up with the operand type.
557     ConstraintWeight getSingleConstraintMatchWeight(
558       AsmOperandInfo &info, const char *constraint) const override;
559 
560     std::pair<unsigned, const TargetRegisterClass *>
561     getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
562                                  StringRef Constraint, MVT VT) const override;
563 
564     /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
565     /// function arguments in the caller parameter area.  This is the actual
566     /// alignment, not its logarithm.
567     unsigned getByValTypeAlignment(Type *Ty,
568                                    const DataLayout &DL) const override;
569 
570     /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
571     /// vector.  If it is invalid, don't add anything to Ops.
572     void LowerAsmOperandForConstraint(SDValue Op,
573                                       std::string &Constraint,
574                                       std::vector<SDValue> &Ops,
575                                       SelectionDAG &DAG) const override;
576 
577     unsigned
578     getInlineAsmMemConstraint(StringRef ConstraintCode) const override {
579       if (ConstraintCode == "es")
580         return InlineAsm::Constraint_es;
581       else if (ConstraintCode == "o")
582         return InlineAsm::Constraint_o;
583       else if (ConstraintCode == "Q")
584         return InlineAsm::Constraint_Q;
585       else if (ConstraintCode == "Z")
586         return InlineAsm::Constraint_Z;
587       else if (ConstraintCode == "Zy")
588         return InlineAsm::Constraint_Zy;
589       return TargetLowering::getInlineAsmMemConstraint(ConstraintCode);
590     }
591 
592     /// isLegalAddressingMode - Return true if the addressing mode represented
593     /// by AM is legal for this target, for a load/store of the specified type.
594     bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM,
595                                Type *Ty, unsigned AS) const override;
596 
597     /// isLegalICmpImmediate - Return true if the specified immediate is legal
598     /// icmp immediate, that is the target has icmp instructions which can
599     /// compare a register against the immediate without having to materialize
600     /// the immediate into a register.
601     bool isLegalICmpImmediate(int64_t Imm) const override;
602 
603     /// isLegalAddImmediate - Return true if the specified immediate is legal
604     /// add immediate, that is the target has add instructions which can
605     /// add a register and the immediate without having to materialize
606     /// the immediate into a register.
607     bool isLegalAddImmediate(int64_t Imm) const override;
608 
609     /// isTruncateFree - Return true if it's free to truncate a value of
610     /// type Ty1 to type Ty2. e.g. On PPC it's free to truncate a i64 value in
611     /// register X1 to i32 by referencing its sub-register R1.
612     bool isTruncateFree(Type *Ty1, Type *Ty2) const override;
613     bool isTruncateFree(EVT VT1, EVT VT2) const override;
614 
615     bool isZExtFree(SDValue Val, EVT VT2) const override;
616 
617     bool isFPExtFree(EVT VT) const override;
618 
619     /// \brief Returns true if it is beneficial to convert a load of a constant
620     /// to just the constant itself.
621     bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
622                                            Type *Ty) const override;
623 
624     bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const override;
625 
626     bool getTgtMemIntrinsic(IntrinsicInfo &Info,
627                             const CallInst &I,
628                             unsigned Intrinsic) const override;
629 
630     /// getOptimalMemOpType - Returns the target specific optimal type for load
631     /// and store operations as a result of memset, memcpy, and memmove
632     /// lowering. If DstAlign is zero that means it's safe to destination
633     /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
634     /// means there isn't a need to check it against alignment requirement,
635     /// probably because the source does not need to be loaded. If 'IsMemset' is
636     /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
637     /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
638     /// source is constant so it does not need to be loaded.
639     /// It returns EVT::Other if the type should be determined using generic
640     /// target-independent logic.
641     EVT
642     getOptimalMemOpType(uint64_t Size, unsigned DstAlign, unsigned SrcAlign,
643                         bool IsMemset, bool ZeroMemset, bool MemcpyStrSrc,
644                         MachineFunction &MF) const override;
645 
646     /// Is unaligned memory access allowed for the given type, and is it fast
647     /// relative to software emulation.
648     bool allowsMisalignedMemoryAccesses(EVT VT,
649                                         unsigned AddrSpace,
650                                         unsigned Align = 1,
651                                         bool *Fast = nullptr) const override;
652 
653     /// isFMAFasterThanFMulAndFAdd - Return true if an FMA operation is faster
654     /// than a pair of fmul and fadd instructions. fmuladd intrinsics will be
655     /// expanded to FMAs when this method returns true, otherwise fmuladd is
656     /// expanded to fmul + fadd.
657     bool isFMAFasterThanFMulAndFAdd(EVT VT) const override;
658 
659     const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const override;
660 
661     // Should we expand the build vector with shuffles?
662     bool
663     shouldExpandBuildVectorWithShuffles(EVT VT,
664                                         unsigned DefinedValues) const override;
665 
666     /// createFastISel - This method returns a target-specific FastISel object,
667     /// or null if the target does not support "fast" instruction selection.
668     FastISel *createFastISel(FunctionLoweringInfo &FuncInfo,
669                              const TargetLibraryInfo *LibInfo) const override;
670 
671     /// \brief Returns true if an argument of type Ty needs to be passed in a
672     /// contiguous block of registers in calling convention CallConv.
673     bool functionArgumentNeedsConsecutiveRegisters(
674       Type *Ty, CallingConv::ID CallConv, bool isVarArg) const override {
675       // We support any array type as "consecutive" block in the parameter
676       // save area.  The element type defines the alignment requirement and
677       // whether the argument should go in GPRs, FPRs, or VRs if available.
678       //
679       // Note that clang uses this capability both to implement the ELFv2
680       // homogeneous float/vector aggregate ABI, and to avoid having to use
681       // "byval" when passing aggregates that might fully fit in registers.
682       return Ty->isArrayTy();
683     }
684 
685     /// If a physical register, this returns the register that receives the
686     /// exception address on entry to an EH pad.
687     unsigned
688     getExceptionPointerRegister(const Constant *PersonalityFn) const override;
689 
690     /// If a physical register, this returns the register that receives the
691     /// exception typeid on entry to a landing pad.
692     unsigned
693     getExceptionSelectorRegister(const Constant *PersonalityFn) const override;
694 
695     /// Override to support customized stack guard loading.
696     bool useLoadStackGuardNode() const override;
697     void insertSSPDeclarations(Module &M) const override;
698 
699   private:
700     struct ReuseLoadInfo {
701       SDValue Ptr;
702       SDValue Chain;
703       SDValue ResChain;
704       MachinePointerInfo MPI;
705       bool IsInvariant;
706       unsigned Alignment;
707       AAMDNodes AAInfo;
708       const MDNode *Ranges;
709 
710       ReuseLoadInfo() : IsInvariant(false), Alignment(0), Ranges(nullptr) {}
711     };
712 
713     bool canReuseLoadAddress(SDValue Op, EVT MemVT, ReuseLoadInfo &RLI,
714                              SelectionDAG &DAG,
715                              ISD::LoadExtType ET = ISD::NON_EXTLOAD) const;
716     void spliceIntoChain(SDValue ResChain, SDValue NewResChain,
717                          SelectionDAG &DAG) const;
718 
719     void LowerFP_TO_INTForReuse(SDValue Op, ReuseLoadInfo &RLI,
720                                 SelectionDAG &DAG, const SDLoc &dl) const;
721     SDValue LowerFP_TO_INTDirectMove(SDValue Op, SelectionDAG &DAG,
722                                      const SDLoc &dl) const;
723     SDValue LowerINT_TO_FPDirectMove(SDValue Op, SelectionDAG &DAG,
724                                      const SDLoc &dl) const;
725 
726     SDValue getFramePointerFrameIndex(SelectionDAG & DAG) const;
727     SDValue getReturnAddrFrameIndex(SelectionDAG & DAG) const;
728 
729     bool
730     IsEligibleForTailCallOptimization(SDValue Callee,
731                                       CallingConv::ID CalleeCC,
732                                       bool isVarArg,
733                                       const SmallVectorImpl<ISD::InputArg> &Ins,
734                                       SelectionDAG& DAG) const;
735 
736     bool
737     IsEligibleForTailCallOptimization_64SVR4(
738                                     SDValue Callee,
739                                     CallingConv::ID CalleeCC,
740                                     ImmutableCallSite *CS,
741                                     bool isVarArg,
742                                     const SmallVectorImpl<ISD::OutputArg> &Outs,
743                                     const SmallVectorImpl<ISD::InputArg> &Ins,
744                                     SelectionDAG& DAG) const;
745 
746     SDValue EmitTailCallLoadFPAndRetAddr(SelectionDAG &DAG, int SPDiff,
747                                          SDValue Chain, SDValue &LROpOut,
748                                          SDValue &FPOpOut, bool isDarwinABI,
749                                          const SDLoc &dl) const;
750 
751     SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
752     SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
753     SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
754     SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
755     SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
756     SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
757     SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
758     SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
759     SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
760     SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
761     SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG,
762                          const PPCSubtarget &Subtarget) const;
763     SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG,
764                        const PPCSubtarget &Subtarget) const;
765     SDValue LowerVACOPY(SDValue Op, SelectionDAG &DAG,
766                         const PPCSubtarget &Subtarget) const;
767     SDValue LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
768                                 const PPCSubtarget &Subtarget) const;
769     SDValue LowerGET_DYNAMIC_AREA_OFFSET(SDValue Op, SelectionDAG &DAG,
770                                          const PPCSubtarget &Subtarget) const;
771     SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG,
772                                       const PPCSubtarget &Subtarget) const;
773     SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const;
774     SDValue LowerSTORE(SDValue Op, SelectionDAG &DAG) const;
775     SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const;
776     SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
777     SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
778                            const SDLoc &dl) const;
779     SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
780     SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
781     SDValue LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const;
782     SDValue LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const;
783     SDValue LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const;
784     SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
785     SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
786     SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
787     SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const;
788     SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const;
789     SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const;
790     SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) const;
791 
792     SDValue LowerVectorLoad(SDValue Op, SelectionDAG &DAG) const;
793     SDValue LowerVectorStore(SDValue Op, SelectionDAG &DAG) const;
794 
795     SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
796                             CallingConv::ID CallConv, bool isVarArg,
797                             const SmallVectorImpl<ISD::InputArg> &Ins,
798                             const SDLoc &dl, SelectionDAG &DAG,
799                             SmallVectorImpl<SDValue> &InVals) const;
800     SDValue FinishCall(CallingConv::ID CallConv, const SDLoc &dl,
801                        bool isTailCall, bool isVarArg, bool IsPatchPoint,
802                        bool hasNest, SelectionDAG &DAG,
803                        SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
804                        SDValue InFlag, SDValue Chain, SDValue CallSeqStart,
805                        SDValue &Callee, int SPDiff, unsigned NumBytes,
806                        const SmallVectorImpl<ISD::InputArg> &Ins,
807                        SmallVectorImpl<SDValue> &InVals,
808                        ImmutableCallSite *CS) const;
809 
810     SDValue
811     LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
812                          const SmallVectorImpl<ISD::InputArg> &Ins,
813                          const SDLoc &dl, SelectionDAG &DAG,
814                          SmallVectorImpl<SDValue> &InVals) const override;
815 
816     SDValue
817       LowerCall(TargetLowering::CallLoweringInfo &CLI,
818                 SmallVectorImpl<SDValue> &InVals) const override;
819 
820     bool
821       CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
822                    bool isVarArg,
823                    const SmallVectorImpl<ISD::OutputArg> &Outs,
824                    LLVMContext &Context) const override;
825 
826     SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
827                         const SmallVectorImpl<ISD::OutputArg> &Outs,
828                         const SmallVectorImpl<SDValue> &OutVals,
829                         const SDLoc &dl, SelectionDAG &DAG) const override;
830 
831     SDValue extendArgForPPC64(ISD::ArgFlagsTy Flags, EVT ObjectVT,
832                               SelectionDAG &DAG, SDValue ArgVal,
833                               const SDLoc &dl) const;
834 
835     SDValue LowerFormalArguments_Darwin(
836         SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
837         const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
838         SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const;
839     SDValue LowerFormalArguments_64SVR4(
840         SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
841         const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
842         SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const;
843     SDValue LowerFormalArguments_32SVR4(
844         SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
845         const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
846         SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const;
847 
848     SDValue createMemcpyOutsideCallSeq(SDValue Arg, SDValue PtrOff,
849                                        SDValue CallSeqStart,
850                                        ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
851                                        const SDLoc &dl) const;
852 
853     SDValue LowerCall_Darwin(SDValue Chain, SDValue Callee,
854                              CallingConv::ID CallConv, bool isVarArg,
855                              bool isTailCall, bool IsPatchPoint,
856                              const SmallVectorImpl<ISD::OutputArg> &Outs,
857                              const SmallVectorImpl<SDValue> &OutVals,
858                              const SmallVectorImpl<ISD::InputArg> &Ins,
859                              const SDLoc &dl, SelectionDAG &DAG,
860                              SmallVectorImpl<SDValue> &InVals,
861                              ImmutableCallSite *CS) const;
862     SDValue LowerCall_64SVR4(SDValue Chain, SDValue Callee,
863                              CallingConv::ID CallConv, bool isVarArg,
864                              bool isTailCall, bool IsPatchPoint,
865                              const SmallVectorImpl<ISD::OutputArg> &Outs,
866                              const SmallVectorImpl<SDValue> &OutVals,
867                              const SmallVectorImpl<ISD::InputArg> &Ins,
868                              const SDLoc &dl, SelectionDAG &DAG,
869                              SmallVectorImpl<SDValue> &InVals,
870                              ImmutableCallSite *CS) const;
871     SDValue LowerCall_32SVR4(SDValue Chain, SDValue Callee,
872                              CallingConv::ID CallConv, bool isVarArg,
873                              bool isTailCall, bool IsPatchPoint,
874                              const SmallVectorImpl<ISD::OutputArg> &Outs,
875                              const SmallVectorImpl<SDValue> &OutVals,
876                              const SmallVectorImpl<ISD::InputArg> &Ins,
877                              const SDLoc &dl, SelectionDAG &DAG,
878                              SmallVectorImpl<SDValue> &InVals,
879                              ImmutableCallSite *CS) const;
880 
881     SDValue lowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
882     SDValue lowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
883 
884     SDValue DAGCombineExtBoolTrunc(SDNode *N, DAGCombinerInfo &DCI) const;
885     SDValue DAGCombineTruncBoolExt(SDNode *N, DAGCombinerInfo &DCI) const;
886     SDValue combineFPToIntToFP(SDNode *N, DAGCombinerInfo &DCI) const;
887 
888     SDValue getRsqrtEstimate(SDValue Operand, DAGCombinerInfo &DCI,
889                              unsigned &RefinementSteps,
890                              bool &UseOneConstNR) const override;
891     SDValue getRecipEstimate(SDValue Operand, DAGCombinerInfo &DCI,
892                              unsigned &RefinementSteps) const override;
893     unsigned combineRepeatedFPDivisors() const override;
894 
895     CCAssignFn *useFastISelCCs(unsigned Flag) const;
896   };
897 
898   namespace PPC {
899     FastISel *createFastISel(FunctionLoweringInfo &FuncInfo,
900                              const TargetLibraryInfo *LibInfo);
901   }
902 
903   bool CC_PPC32_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
904                                   CCValAssign::LocInfo &LocInfo,
905                                   ISD::ArgFlagsTy &ArgFlags,
906                                   CCState &State);
907 
908   bool CC_PPC32_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
909                                          MVT &LocVT,
910                                          CCValAssign::LocInfo &LocInfo,
911                                          ISD::ArgFlagsTy &ArgFlags,
912                                          CCState &State);
913 
914   bool CC_PPC32_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
915                                            MVT &LocVT,
916                                            CCValAssign::LocInfo &LocInfo,
917                                            ISD::ArgFlagsTy &ArgFlags,
918                                            CCState &State);
919 }
920 
921 #endif   // LLVM_TARGET_POWERPC_PPC32ISELLOWERING_H
922