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 {
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       /// The CMPB instruction (takes two operands of i32 or i64).
65       CMPB,
66 
67       /// Hi/Lo - These represent the high and low 16-bit parts of a global
68       /// address respectively.  These nodes have two operands, the first of
69       /// which must be a TargetGlobalAddress, and the second of which must be a
70       /// Constant.  Selected naively, these turn into 'lis G+C' and 'li G+C',
71       /// though these are usually folded into other nodes.
72       Hi, Lo,
73 
74       /// The following two target-specific nodes are used for calls through
75       /// function pointers in the 64-bit SVR4 ABI.
76 
77       /// OPRC, CHAIN = DYNALLOC(CHAIN, NEGSIZE, FRAME_INDEX)
78       /// This instruction is lowered in PPCRegisterInfo::eliminateFrameIndex to
79       /// compute an allocation on the stack.
80       DYNALLOC,
81 
82       /// GlobalBaseReg - On Darwin, this node represents the result of the mflr
83       /// at function entry, used for PIC code.
84       GlobalBaseReg,
85 
86       /// These nodes represent the 32-bit PPC shifts that operate on 6-bit
87       /// shift amounts.  These nodes are generated by the multi-precision shift
88       /// code.
89       SRL, SRA, SHL,
90 
91       /// The combination of sra[wd]i and addze used to implemented signed
92       /// integer division by a power of 2. The first operand is the dividend,
93       /// and the second is the constant shift amount (representing the
94       /// divisor).
95       SRA_ADDZE,
96 
97       /// CALL - A direct function call.
98       /// CALL_NOP is a call with the special NOP which follows 64-bit
99       /// SVR4 calls.
100       CALL, CALL_NOP,
101 
102       /// CHAIN,FLAG = MTCTR(VAL, CHAIN[, INFLAG]) - Directly corresponds to a
103       /// MTCTR instruction.
104       MTCTR,
105 
106       /// CHAIN,FLAG = BCTRL(CHAIN, INFLAG) - Directly corresponds to a
107       /// BCTRL instruction.
108       BCTRL,
109 
110       /// CHAIN,FLAG = BCTRL(CHAIN, ADDR, INFLAG) - The combination of a bctrl
111       /// instruction and the TOC reload required on SVR4 PPC64.
112       BCTRL_LOAD_TOC,
113 
114       /// Return with a flag operand, matched by 'blr'
115       RET_FLAG,
116 
117       /// R32 = MFOCRF(CRREG, INFLAG) - Represents the MFOCRF instruction.
118       /// This copies the bits corresponding to the specified CRREG into the
119       /// resultant GPR.  Bits corresponding to other CR regs are undefined.
120       MFOCRF,
121 
122       /// Direct move from a VSX register to a GPR
123       MFVSR,
124 
125       /// Direct move from a GPR to a VSX register (algebraic)
126       MTVSRA,
127 
128       /// Direct move from a GPR to a VSX register (zero)
129       MTVSRZ,
130 
131       // FIXME: Remove these once the ANDI glue bug is fixed:
132       /// i1 = ANDIo_1_[EQ|GT]_BIT(i32 or i64 x) - Represents the result of the
133       /// eq or gt bit of CR0 after executing andi. x, 1. This is used to
134       /// implement truncation of i32 or i64 to i1.
135       ANDIo_1_EQ_BIT, ANDIo_1_GT_BIT,
136 
137       // READ_TIME_BASE - A read of the 64-bit time-base register on a 32-bit
138       // target (returns (Lo, Hi)). It takes a chain operand.
139       READ_TIME_BASE,
140 
141       // EH_SJLJ_SETJMP - SjLj exception handling setjmp.
142       EH_SJLJ_SETJMP,
143 
144       // EH_SJLJ_LONGJMP - SjLj exception handling longjmp.
145       EH_SJLJ_LONGJMP,
146 
147       /// RESVEC = VCMP(LHS, RHS, OPC) - Represents one of the altivec VCMP*
148       /// instructions.  For lack of better number, we use the opcode number
149       /// encoding for the OPC field to identify the compare.  For example, 838
150       /// is VCMPGTSH.
151       VCMP,
152 
153       /// RESVEC, OUTFLAG = VCMPo(LHS, RHS, OPC) - Represents one of the
154       /// altivec VCMP*o instructions.  For lack of better number, we use the
155       /// opcode number encoding for the OPC field to identify the compare.  For
156       /// example, 838 is VCMPGTSH.
157       VCMPo,
158 
159       /// CHAIN = COND_BRANCH CHAIN, CRRC, OPC, DESTBB [, INFLAG] - This
160       /// corresponds to the COND_BRANCH pseudo instruction.  CRRC is the
161       /// condition register to branch on, OPC is the branch opcode to use (e.g.
162       /// PPC::BLE), DESTBB is the destination block to branch to, and INFLAG is
163       /// an optional input flag argument.
164       COND_BRANCH,
165 
166       /// CHAIN = BDNZ CHAIN, DESTBB - These are used to create counter-based
167       /// loops.
168       BDNZ, BDZ,
169 
170       /// F8RC = FADDRTZ F8RC, F8RC - This is an FADD done with rounding
171       /// towards zero.  Used only as part of the long double-to-int
172       /// conversion sequence.
173       FADDRTZ,
174 
175       /// F8RC = MFFS - This moves the FPSCR (not modeled) into the register.
176       MFFS,
177 
178       /// TC_RETURN - A tail call return.
179       ///   operand #0 chain
180       ///   operand #1 callee (register or absolute)
181       ///   operand #2 stack adjustment
182       ///   operand #3 optional in flag
183       TC_RETURN,
184 
185       /// ch, gl = CR6[UN]SET ch, inglue - Toggle CR bit 6 for SVR4 vararg calls
186       CR6SET,
187       CR6UNSET,
188 
189       /// GPRC = address of _GLOBAL_OFFSET_TABLE_. Used by initial-exec TLS
190       /// on PPC32.
191       PPC32_GOT,
192 
193       /// GPRC = address of _GLOBAL_OFFSET_TABLE_. Used by general dynamic and
194       /// local dynamic TLS on PPC32.
195       PPC32_PICGOT,
196 
197       /// G8RC = ADDIS_GOT_TPREL_HA %X2, Symbol - Used by the initial-exec
198       /// TLS model, produces an ADDIS8 instruction that adds the GOT
199       /// base to sym\@got\@tprel\@ha.
200       ADDIS_GOT_TPREL_HA,
201 
202       /// G8RC = LD_GOT_TPREL_L Symbol, G8RReg - Used by the initial-exec
203       /// TLS model, produces a LD instruction with base register G8RReg
204       /// and offset sym\@got\@tprel\@l.  This completes the addition that
205       /// finds the offset of "sym" relative to the thread pointer.
206       LD_GOT_TPREL_L,
207 
208       /// G8RC = ADD_TLS G8RReg, Symbol - Used by the initial-exec TLS
209       /// model, produces an ADD instruction that adds the contents of
210       /// G8RReg to the thread pointer.  Symbol contains a relocation
211       /// sym\@tls which is to be replaced by the thread pointer and
212       /// identifies to the linker that the instruction is part of a
213       /// TLS sequence.
214       ADD_TLS,
215 
216       /// G8RC = ADDIS_TLSGD_HA %X2, Symbol - For the general-dynamic TLS
217       /// model, produces an ADDIS8 instruction that adds the GOT base
218       /// register to sym\@got\@tlsgd\@ha.
219       ADDIS_TLSGD_HA,
220 
221       /// %X3 = ADDI_TLSGD_L G8RReg, Symbol - For the general-dynamic TLS
222       /// model, produces an ADDI8 instruction that adds G8RReg to
223       /// sym\@got\@tlsgd\@l and stores the result in X3.  Hidden by
224       /// ADDIS_TLSGD_L_ADDR until after register assignment.
225       ADDI_TLSGD_L,
226 
227       /// %X3 = GET_TLS_ADDR %X3, Symbol - For the general-dynamic TLS
228       /// model, produces a call to __tls_get_addr(sym\@tlsgd).  Hidden by
229       /// ADDIS_TLSGD_L_ADDR until after register assignment.
230       GET_TLS_ADDR,
231 
232       /// G8RC = ADDI_TLSGD_L_ADDR G8RReg, Symbol, Symbol - Op that
233       /// combines ADDI_TLSGD_L and GET_TLS_ADDR until expansion following
234       /// register assignment.
235       ADDI_TLSGD_L_ADDR,
236 
237       /// G8RC = ADDIS_TLSLD_HA %X2, Symbol - For the local-dynamic TLS
238       /// model, produces an ADDIS8 instruction that adds the GOT base
239       /// register to sym\@got\@tlsld\@ha.
240       ADDIS_TLSLD_HA,
241 
242       /// %X3 = ADDI_TLSLD_L G8RReg, Symbol - For the local-dynamic TLS
243       /// model, produces an ADDI8 instruction that adds G8RReg to
244       /// sym\@got\@tlsld\@l and stores the result in X3.  Hidden by
245       /// ADDIS_TLSLD_L_ADDR until after register assignment.
246       ADDI_TLSLD_L,
247 
248       /// %X3 = GET_TLSLD_ADDR %X3, Symbol - For the local-dynamic TLS
249       /// model, produces a call to __tls_get_addr(sym\@tlsld).  Hidden by
250       /// ADDIS_TLSLD_L_ADDR until after register assignment.
251       GET_TLSLD_ADDR,
252 
253       /// G8RC = ADDI_TLSLD_L_ADDR G8RReg, Symbol, Symbol - Op that
254       /// combines ADDI_TLSLD_L and GET_TLSLD_ADDR until expansion
255       /// following register assignment.
256       ADDI_TLSLD_L_ADDR,
257 
258       /// G8RC = ADDIS_DTPREL_HA %X3, Symbol - For the local-dynamic TLS
259       /// model, produces an ADDIS8 instruction that adds X3 to
260       /// sym\@dtprel\@ha.
261       ADDIS_DTPREL_HA,
262 
263       /// G8RC = ADDI_DTPREL_L G8RReg, Symbol - For the local-dynamic TLS
264       /// model, produces an ADDI8 instruction that adds G8RReg to
265       /// sym\@got\@dtprel\@l.
266       ADDI_DTPREL_L,
267 
268       /// VRRC = VADD_SPLAT Elt, EltSize - Temporary node to be expanded
269       /// during instruction selection to optimize a BUILD_VECTOR into
270       /// operations on splats.  This is necessary to avoid losing these
271       /// optimizations due to constant folding.
272       VADD_SPLAT,
273 
274       /// CHAIN = SC CHAIN, Imm128 - System call.  The 7-bit unsigned
275       /// operand identifies the operating system entry point.
276       SC,
277 
278       /// VSRC, CHAIN = XXSWAPD CHAIN, VSRC - Occurs only for little
279       /// endian.  Maps to an xxswapd instruction that corrects an lxvd2x
280       /// or stxvd2x instruction.  The chain is necessary because the
281       /// sequence replaces a load and needs to provide the same number
282       /// of outputs.
283       XXSWAPD,
284 
285       /// QVFPERM = This corresponds to the QPX qvfperm instruction.
286       QVFPERM,
287 
288       /// QVGPCI = This corresponds to the QPX qvgpci instruction.
289       QVGPCI,
290 
291       /// QVALIGNI = This corresponds to the QPX qvaligni instruction.
292       QVALIGNI,
293 
294       /// QVESPLATI = This corresponds to the QPX qvesplati instruction.
295       QVESPLATI,
296 
297       /// QBFLT = Access the underlying QPX floating-point boolean
298       /// representation.
299       QBFLT,
300 
301       /// CHAIN = STBRX CHAIN, GPRC, Ptr, Type - This is a
302       /// byte-swapping store instruction.  It byte-swaps the low "Type" bits of
303       /// the GPRC input, then stores it through Ptr.  Type can be either i16 or
304       /// i32.
305       STBRX = ISD::FIRST_TARGET_MEMORY_OPCODE,
306 
307       /// GPRC, CHAIN = LBRX CHAIN, Ptr, Type - This is a
308       /// byte-swapping load instruction.  It loads "Type" bits, byte swaps it,
309       /// then puts it in the bottom bits of the GPRC.  TYPE can be either i16
310       /// or i32.
311       LBRX,
312 
313       /// STFIWX - The STFIWX instruction.  The first operand is an input token
314       /// chain, then an f64 value to store, then an address to store it to.
315       STFIWX,
316 
317       /// GPRC, CHAIN = LFIWAX CHAIN, Ptr - This is a floating-point
318       /// load which sign-extends from a 32-bit integer value into the
319       /// destination 64-bit register.
320       LFIWAX,
321 
322       /// GPRC, CHAIN = LFIWZX CHAIN, Ptr - This is a floating-point
323       /// load which zero-extends from a 32-bit integer value into the
324       /// destination 64-bit register.
325       LFIWZX,
326 
327       /// VSRC, CHAIN = LXVD2X_LE CHAIN, Ptr - Occurs only for little endian.
328       /// Maps directly to an lxvd2x instruction that will be followed by
329       /// an xxswapd.
330       LXVD2X,
331 
332       /// CHAIN = STXVD2X CHAIN, VSRC, Ptr - Occurs only for little endian.
333       /// Maps directly to an stxvd2x instruction that will be preceded by
334       /// an xxswapd.
335       STXVD2X,
336 
337       /// QBRC, CHAIN = QVLFSb CHAIN, Ptr
338       /// The 4xf32 load used for v4i1 constants.
339       QVLFSb,
340 
341       /// GPRC = TOC_ENTRY GA, TOC
342       /// Loads the entry for GA from the TOC, where the TOC base is given by
343       /// the last operand.
344       TOC_ENTRY
345     };
346   }
347 
348   /// Define some predicates that are used for node matching.
349   namespace PPC {
350     /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
351     /// VPKUHUM instruction.
352     bool isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
353                               SelectionDAG &DAG);
354 
355     /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
356     /// VPKUWUM instruction.
357     bool isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
358                               SelectionDAG &DAG);
359 
360     /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
361     /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
362     bool isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
363                             unsigned ShuffleKind, SelectionDAG &DAG);
364 
365     /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
366     /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
367     bool isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
368                             unsigned ShuffleKind, SelectionDAG &DAG);
369 
370     /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the
371     /// shift amount, otherwise return -1.
372     int isVSLDOIShuffleMask(SDNode *N, unsigned ShuffleKind,
373                             SelectionDAG &DAG);
374 
375     /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
376     /// specifies a splat of a single element that is suitable for input to
377     /// VSPLTB/VSPLTH/VSPLTW.
378     bool isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize);
379 
380     /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
381     /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
382     unsigned getVSPLTImmediate(SDNode *N, unsigned EltSize, SelectionDAG &DAG);
383 
384     /// get_VSPLTI_elt - If this is a build_vector of constants which can be
385     /// formed by using a vspltis[bhw] instruction of the specified element
386     /// size, return the constant being splatted.  The ByteSize field indicates
387     /// the number of bytes of each element [124] -> [bhw].
388     SDValue get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG);
389 
390     /// If this is a qvaligni shuffle mask, return the shift
391     /// amount, otherwise return -1.
392     int isQVALIGNIShuffleMask(SDNode *N);
393   }
394 
395   class PPCTargetLowering : public TargetLowering {
396     const PPCSubtarget &Subtarget;
397 
398   public:
399     explicit PPCTargetLowering(const PPCTargetMachine &TM,
400                                const PPCSubtarget &STI);
401 
402     /// getTargetNodeName() - This method returns the name of a target specific
403     /// DAG node.
404     const char *getTargetNodeName(unsigned Opcode) const override;
405 
getScalarShiftAmountTy(EVT LHSTy)406     MVT getScalarShiftAmountTy(EVT LHSTy) const override { return MVT::i32; }
407 
isCheapToSpeculateCttz()408     bool isCheapToSpeculateCttz() const override {
409       return true;
410     }
411 
isCheapToSpeculateCtlz()412     bool isCheapToSpeculateCtlz() const override {
413       return true;
414     }
415 
416     /// getSetCCResultType - Return the ISD::SETCC ValueType
417     EVT getSetCCResultType(LLVMContext &Context, EVT VT) const override;
418 
419     /// Return true if target always beneficiates from combining into FMA for a
420     /// given value type. This must typically return false on targets where FMA
421     /// takes more cycles to execute than FADD.
422     bool enableAggressiveFMAFusion(EVT VT) const override;
423 
424     /// getPreIndexedAddressParts - returns true by value, base pointer and
425     /// offset pointer and addressing mode by reference if the node's address
426     /// can be legally represented as pre-indexed load / store address.
427     bool getPreIndexedAddressParts(SDNode *N, SDValue &Base,
428                                    SDValue &Offset,
429                                    ISD::MemIndexedMode &AM,
430                                    SelectionDAG &DAG) const override;
431 
432     /// SelectAddressRegReg - Given the specified addressed, check to see if it
433     /// can be represented as an indexed [r+r] operation.  Returns false if it
434     /// can be more efficiently represented with [r+imm].
435     bool SelectAddressRegReg(SDValue N, SDValue &Base, SDValue &Index,
436                              SelectionDAG &DAG) const;
437 
438     /// SelectAddressRegImm - Returns true if the address N can be represented
439     /// by a base register plus a signed 16-bit displacement [r+imm], and if it
440     /// is not better represented as reg+reg.  If Aligned is true, only accept
441     /// displacements suitable for STD and friends, i.e. multiples of 4.
442     bool SelectAddressRegImm(SDValue N, SDValue &Disp, SDValue &Base,
443                              SelectionDAG &DAG, bool Aligned) const;
444 
445     /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
446     /// represented as an indexed [r+r] operation.
447     bool SelectAddressRegRegOnly(SDValue N, SDValue &Base, SDValue &Index,
448                                  SelectionDAG &DAG) const;
449 
450     Sched::Preference getSchedulingPreference(SDNode *N) const override;
451 
452     /// LowerOperation - Provide custom lowering hooks for some operations.
453     ///
454     SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
455 
456     /// ReplaceNodeResults - Replace the results of node with an illegal result
457     /// type with new values built out of custom code.
458     ///
459     void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue>&Results,
460                             SelectionDAG &DAG) const override;
461 
462     SDValue expandVSXLoadForLE(SDNode *N, DAGCombinerInfo &DCI) const;
463     SDValue expandVSXStoreForLE(SDNode *N, DAGCombinerInfo &DCI) const;
464 
465     SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
466 
467     SDValue BuildSDIVPow2(SDNode *N, const APInt &Divisor, SelectionDAG &DAG,
468                           std::vector<SDNode *> *Created) const override;
469 
470     unsigned getRegisterByName(const char* RegName, EVT VT) const override;
471 
472     void computeKnownBitsForTargetNode(const SDValue Op,
473                                        APInt &KnownZero,
474                                        APInt &KnownOne,
475                                        const SelectionDAG &DAG,
476                                        unsigned Depth = 0) const override;
477 
478     unsigned getPrefLoopAlignment(MachineLoop *ML) const override;
479 
480     Instruction* emitLeadingFence(IRBuilder<> &Builder, AtomicOrdering Ord,
481                                   bool IsStore, bool IsLoad) const override;
482     Instruction* emitTrailingFence(IRBuilder<> &Builder, AtomicOrdering Ord,
483                                    bool IsStore, bool IsLoad) const override;
484 
485     MachineBasicBlock *
486       EmitInstrWithCustomInserter(MachineInstr *MI,
487                                   MachineBasicBlock *MBB) const override;
488     MachineBasicBlock *EmitAtomicBinary(MachineInstr *MI,
489                                         MachineBasicBlock *MBB,
490                                         unsigned AtomicSize,
491                                         unsigned BinOpcode) const;
492     MachineBasicBlock *EmitPartwordAtomicBinary(MachineInstr *MI,
493                                                 MachineBasicBlock *MBB,
494                                             bool is8bit, unsigned Opcode) const;
495 
496     MachineBasicBlock *emitEHSjLjSetJmp(MachineInstr *MI,
497                                         MachineBasicBlock *MBB) const;
498 
499     MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr *MI,
500                                          MachineBasicBlock *MBB) const;
501 
502     ConstraintType
503     getConstraintType(const std::string &Constraint) const override;
504 
505     /// Examine constraint string and operand type and determine a weight value.
506     /// The operand object must already have been set up with the operand type.
507     ConstraintWeight getSingleConstraintMatchWeight(
508       AsmOperandInfo &info, const char *constraint) const override;
509 
510     std::pair<unsigned, const TargetRegisterClass *>
511     getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
512                                  const std::string &Constraint,
513                                  MVT VT) const override;
514 
515     /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
516     /// function arguments in the caller parameter area.  This is the actual
517     /// alignment, not its logarithm.
518     unsigned getByValTypeAlignment(Type *Ty) const override;
519 
520     /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
521     /// vector.  If it is invalid, don't add anything to Ops.
522     void LowerAsmOperandForConstraint(SDValue Op,
523                                       std::string &Constraint,
524                                       std::vector<SDValue> &Ops,
525                                       SelectionDAG &DAG) const override;
526 
getInlineAsmMemConstraint(const std::string & ConstraintCode)527     unsigned getInlineAsmMemConstraint(
528         const std::string &ConstraintCode) const override {
529       if (ConstraintCode == "es")
530         return InlineAsm::Constraint_es;
531       else if (ConstraintCode == "o")
532         return InlineAsm::Constraint_o;
533       else if (ConstraintCode == "Q")
534         return InlineAsm::Constraint_Q;
535       else if (ConstraintCode == "Z")
536         return InlineAsm::Constraint_Z;
537       else if (ConstraintCode == "Zy")
538         return InlineAsm::Constraint_Zy;
539       return TargetLowering::getInlineAsmMemConstraint(ConstraintCode);
540     }
541 
542     /// isLegalAddressingMode - Return true if the addressing mode represented
543     /// by AM is legal for this target, for a load/store of the specified type.
544     bool isLegalAddressingMode(const AddrMode &AM, Type *Ty) const override;
545 
546     /// isLegalICmpImmediate - Return true if the specified immediate is legal
547     /// icmp immediate, that is the target has icmp instructions which can
548     /// compare a register against the immediate without having to materialize
549     /// the immediate into a register.
550     bool isLegalICmpImmediate(int64_t Imm) const override;
551 
552     /// isLegalAddImmediate - Return true if the specified immediate is legal
553     /// add immediate, that is the target has add instructions which can
554     /// add a register and the immediate without having to materialize
555     /// the immediate into a register.
556     bool isLegalAddImmediate(int64_t Imm) const override;
557 
558     /// isTruncateFree - Return true if it's free to truncate a value of
559     /// type Ty1 to type Ty2. e.g. On PPC it's free to truncate a i64 value in
560     /// register X1 to i32 by referencing its sub-register R1.
561     bool isTruncateFree(Type *Ty1, Type *Ty2) const override;
562     bool isTruncateFree(EVT VT1, EVT VT2) const override;
563 
564     bool isZExtFree(SDValue Val, EVT VT2) const override;
565 
566     bool isFPExtFree(EVT VT) const override;
567 
568     /// \brief Returns true if it is beneficial to convert a load of a constant
569     /// to just the constant itself.
570     bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
571                                            Type *Ty) const override;
572 
573     bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const override;
574 
575     bool getTgtMemIntrinsic(IntrinsicInfo &Info,
576                             const CallInst &I,
577                             unsigned Intrinsic) const override;
578 
579     /// getOptimalMemOpType - Returns the target specific optimal type for load
580     /// and store operations as a result of memset, memcpy, and memmove
581     /// lowering. If DstAlign is zero that means it's safe to destination
582     /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
583     /// means there isn't a need to check it against alignment requirement,
584     /// probably because the source does not need to be loaded. If 'IsMemset' is
585     /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
586     /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
587     /// source is constant so it does not need to be loaded.
588     /// It returns EVT::Other if the type should be determined using generic
589     /// target-independent logic.
590     EVT
591     getOptimalMemOpType(uint64_t Size, unsigned DstAlign, unsigned SrcAlign,
592                         bool IsMemset, bool ZeroMemset, bool MemcpyStrSrc,
593                         MachineFunction &MF) const override;
594 
595     /// Is unaligned memory access allowed for the given type, and is it fast
596     /// relative to software emulation.
597     bool allowsMisalignedMemoryAccesses(EVT VT,
598                                         unsigned AddrSpace,
599                                         unsigned Align = 1,
600                                         bool *Fast = nullptr) const override;
601 
602     /// isFMAFasterThanFMulAndFAdd - Return true if an FMA operation is faster
603     /// than a pair of fmul and fadd instructions. fmuladd intrinsics will be
604     /// expanded to FMAs when this method returns true, otherwise fmuladd is
605     /// expanded to fmul + fadd.
606     bool isFMAFasterThanFMulAndFAdd(EVT VT) const override;
607 
608     const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const override;
609 
610     // Should we expand the build vector with shuffles?
611     bool
612     shouldExpandBuildVectorWithShuffles(EVT VT,
613                                         unsigned DefinedValues) const override;
614 
615     /// createFastISel - This method returns a target-specific FastISel object,
616     /// or null if the target does not support "fast" instruction selection.
617     FastISel *createFastISel(FunctionLoweringInfo &FuncInfo,
618                              const TargetLibraryInfo *LibInfo) const override;
619 
620     /// \brief Returns true if an argument of type Ty needs to be passed in a
621     /// contiguous block of registers in calling convention CallConv.
functionArgumentNeedsConsecutiveRegisters(Type * Ty,CallingConv::ID CallConv,bool isVarArg)622     bool functionArgumentNeedsConsecutiveRegisters(
623       Type *Ty, CallingConv::ID CallConv, bool isVarArg) const override {
624       // We support any array type as "consecutive" block in the parameter
625       // save area.  The element type defines the alignment requirement and
626       // whether the argument should go in GPRs, FPRs, or VRs if available.
627       //
628       // Note that clang uses this capability both to implement the ELFv2
629       // homogeneous float/vector aggregate ABI, and to avoid having to use
630       // "byval" when passing aggregates that might fully fit in registers.
631       return Ty->isArrayTy();
632     }
633 
634   private:
635 
636     struct ReuseLoadInfo {
637       SDValue Ptr;
638       SDValue Chain;
639       SDValue ResChain;
640       MachinePointerInfo MPI;
641       bool IsInvariant;
642       unsigned Alignment;
643       AAMDNodes AAInfo;
644       const MDNode *Ranges;
645 
ReuseLoadInfoReuseLoadInfo646       ReuseLoadInfo() : IsInvariant(false), Alignment(0), Ranges(nullptr) {}
647     };
648 
649     bool canReuseLoadAddress(SDValue Op, EVT MemVT, ReuseLoadInfo &RLI,
650                              SelectionDAG &DAG,
651                              ISD::LoadExtType ET = ISD::NON_EXTLOAD) const;
652     void spliceIntoChain(SDValue ResChain, SDValue NewResChain,
653                          SelectionDAG &DAG) const;
654 
655     void LowerFP_TO_INTForReuse(SDValue Op, ReuseLoadInfo &RLI,
656                                 SelectionDAG &DAG, SDLoc dl) const;
657     SDValue LowerFP_TO_INTDirectMove(SDValue Op, SelectionDAG &DAG,
658                                      SDLoc dl) const;
659     SDValue LowerINT_TO_FPDirectMove(SDValue Op, SelectionDAG &DAG,
660                                      SDLoc dl) const;
661 
662     SDValue getFramePointerFrameIndex(SelectionDAG & DAG) const;
663     SDValue getReturnAddrFrameIndex(SelectionDAG & DAG) const;
664 
665     bool
666     IsEligibleForTailCallOptimization(SDValue Callee,
667                                       CallingConv::ID CalleeCC,
668                                       bool isVarArg,
669                                       const SmallVectorImpl<ISD::InputArg> &Ins,
670                                       SelectionDAG& DAG) const;
671 
672     SDValue EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG,
673                                          int SPDiff,
674                                          SDValue Chain,
675                                          SDValue &LROpOut,
676                                          SDValue &FPOpOut,
677                                          bool isDarwinABI,
678                                          SDLoc dl) const;
679 
680     SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
681     SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
682     SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
683     SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
684     SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
685     SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
686     SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
687     SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
688     SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
689     SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
690     SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG,
691                          const PPCSubtarget &Subtarget) const;
692     SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG,
693                        const PPCSubtarget &Subtarget) const;
694     SDValue LowerVACOPY(SDValue Op, SelectionDAG &DAG,
695                         const PPCSubtarget &Subtarget) const;
696     SDValue LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
697                                 const PPCSubtarget &Subtarget) const;
698     SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG,
699                                       const PPCSubtarget &Subtarget) const;
700     SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const;
701     SDValue LowerSTORE(SDValue Op, SelectionDAG &DAG) const;
702     SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const;
703     SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
704     SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG, SDLoc dl) const;
705     SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
706     SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
707     SDValue LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const;
708     SDValue LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const;
709     SDValue LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const;
710     SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
711     SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
712     SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
713     SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const;
714     SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const;
715     SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const;
716     SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) const;
717 
718     SDValue LowerVectorLoad(SDValue Op, SelectionDAG &DAG) const;
719     SDValue LowerVectorStore(SDValue Op, SelectionDAG &DAG) const;
720 
721     SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
722                             CallingConv::ID CallConv, bool isVarArg,
723                             const SmallVectorImpl<ISD::InputArg> &Ins,
724                             SDLoc dl, SelectionDAG &DAG,
725                             SmallVectorImpl<SDValue> &InVals) const;
726     SDValue FinishCall(CallingConv::ID CallConv, SDLoc dl, bool isTailCall,
727                        bool isVarArg, bool IsPatchPoint,
728                        SelectionDAG &DAG,
729                        SmallVector<std::pair<unsigned, SDValue>, 8>
730                          &RegsToPass,
731                        SDValue InFlag, SDValue Chain, SDValue CallSeqStart,
732                        SDValue &Callee,
733                        int SPDiff, unsigned NumBytes,
734                        const SmallVectorImpl<ISD::InputArg> &Ins,
735                        SmallVectorImpl<SDValue> &InVals,
736                        ImmutableCallSite *CS) const;
737 
738     SDValue
739       LowerFormalArguments(SDValue Chain,
740                            CallingConv::ID CallConv, bool isVarArg,
741                            const SmallVectorImpl<ISD::InputArg> &Ins,
742                            SDLoc dl, SelectionDAG &DAG,
743                            SmallVectorImpl<SDValue> &InVals) const override;
744 
745     SDValue
746       LowerCall(TargetLowering::CallLoweringInfo &CLI,
747                 SmallVectorImpl<SDValue> &InVals) const override;
748 
749     bool
750       CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
751                    bool isVarArg,
752                    const SmallVectorImpl<ISD::OutputArg> &Outs,
753                    LLVMContext &Context) const override;
754 
755     SDValue
756       LowerReturn(SDValue Chain,
757                   CallingConv::ID CallConv, bool isVarArg,
758                   const SmallVectorImpl<ISD::OutputArg> &Outs,
759                   const SmallVectorImpl<SDValue> &OutVals,
760                   SDLoc dl, SelectionDAG &DAG) const override;
761 
762     SDValue
763       extendArgForPPC64(ISD::ArgFlagsTy Flags, EVT ObjectVT, SelectionDAG &DAG,
764                         SDValue ArgVal, SDLoc dl) const;
765 
766     SDValue
767       LowerFormalArguments_Darwin(SDValue Chain,
768                                   CallingConv::ID CallConv, bool isVarArg,
769                                   const SmallVectorImpl<ISD::InputArg> &Ins,
770                                   SDLoc dl, SelectionDAG &DAG,
771                                   SmallVectorImpl<SDValue> &InVals) const;
772     SDValue
773       LowerFormalArguments_64SVR4(SDValue Chain,
774                                   CallingConv::ID CallConv, bool isVarArg,
775                                   const SmallVectorImpl<ISD::InputArg> &Ins,
776                                   SDLoc dl, SelectionDAG &DAG,
777                                   SmallVectorImpl<SDValue> &InVals) const;
778     SDValue
779       LowerFormalArguments_32SVR4(SDValue Chain,
780                                   CallingConv::ID CallConv, bool isVarArg,
781                                   const SmallVectorImpl<ISD::InputArg> &Ins,
782                                   SDLoc dl, SelectionDAG &DAG,
783                                   SmallVectorImpl<SDValue> &InVals) const;
784 
785     SDValue
786       createMemcpyOutsideCallSeq(SDValue Arg, SDValue PtrOff,
787                                  SDValue CallSeqStart, ISD::ArgFlagsTy Flags,
788                                  SelectionDAG &DAG, SDLoc dl) const;
789 
790     SDValue
791       LowerCall_Darwin(SDValue Chain, SDValue Callee,
792                        CallingConv::ID CallConv,
793                        bool isVarArg, bool isTailCall, bool IsPatchPoint,
794                        const SmallVectorImpl<ISD::OutputArg> &Outs,
795                        const SmallVectorImpl<SDValue> &OutVals,
796                        const SmallVectorImpl<ISD::InputArg> &Ins,
797                        SDLoc dl, SelectionDAG &DAG,
798                        SmallVectorImpl<SDValue> &InVals,
799                        ImmutableCallSite *CS) const;
800     SDValue
801       LowerCall_64SVR4(SDValue Chain, SDValue Callee,
802                        CallingConv::ID CallConv,
803                        bool isVarArg, bool isTailCall, bool IsPatchPoint,
804                        const SmallVectorImpl<ISD::OutputArg> &Outs,
805                        const SmallVectorImpl<SDValue> &OutVals,
806                        const SmallVectorImpl<ISD::InputArg> &Ins,
807                        SDLoc dl, SelectionDAG &DAG,
808                        SmallVectorImpl<SDValue> &InVals,
809                        ImmutableCallSite *CS) const;
810     SDValue
811     LowerCall_32SVR4(SDValue Chain, SDValue Callee, CallingConv::ID CallConv,
812                      bool isVarArg, bool isTailCall, bool IsPatchPoint,
813                      const SmallVectorImpl<ISD::OutputArg> &Outs,
814                      const SmallVectorImpl<SDValue> &OutVals,
815                      const SmallVectorImpl<ISD::InputArg> &Ins,
816                      SDLoc dl, SelectionDAG &DAG,
817                      SmallVectorImpl<SDValue> &InVals,
818                      ImmutableCallSite *CS) const;
819 
820     SDValue lowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const;
821     SDValue lowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const;
822 
823     SDValue DAGCombineExtBoolTrunc(SDNode *N, DAGCombinerInfo &DCI) const;
824     SDValue DAGCombineTruncBoolExt(SDNode *N, DAGCombinerInfo &DCI) const;
825     SDValue combineFPToIntToFP(SDNode *N, DAGCombinerInfo &DCI) const;
826 
827     SDValue getRsqrtEstimate(SDValue Operand, DAGCombinerInfo &DCI,
828                              unsigned &RefinementSteps,
829                              bool &UseOneConstNR) const override;
830     SDValue getRecipEstimate(SDValue Operand, DAGCombinerInfo &DCI,
831                              unsigned &RefinementSteps) const override;
832     bool combineRepeatedFPDivisors(unsigned NumUsers) const override;
833 
834     CCAssignFn *useFastISelCCs(unsigned Flag) const;
835   };
836 
837   namespace PPC {
838     FastISel *createFastISel(FunctionLoweringInfo &FuncInfo,
839                              const TargetLibraryInfo *LibInfo);
840   }
841 
842   bool CC_PPC32_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
843                                   CCValAssign::LocInfo &LocInfo,
844                                   ISD::ArgFlagsTy &ArgFlags,
845                                   CCState &State);
846 
847   bool CC_PPC32_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
848                                          MVT &LocVT,
849                                          CCValAssign::LocInfo &LocInfo,
850                                          ISD::ArgFlagsTy &ArgFlags,
851                                          CCState &State);
852 
853   bool CC_PPC32_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
854                                            MVT &LocVT,
855                                            CCValAssign::LocInfo &LocInfo,
856                                            ISD::ArgFlagsTy &ArgFlags,
857                                            CCState &State);
858 }
859 
860 #endif   // LLVM_TARGET_POWERPC_PPC32ISELLOWERING_H
861