1 | //===-- PPCInstrInfo.h - PowerPC Instruction Information --------*- C++ -*-===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file contains the PowerPC implementation of the TargetInstrInfo class. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #ifndef LLVM_LIB_TARGET_POWERPC_PPCINSTRINFO_H |
14 | #define LLVM_LIB_TARGET_POWERPC_PPCINSTRINFO_H |
15 | |
16 | #include "MCTargetDesc/PPCMCTargetDesc.h" |
17 | #include "PPC.h" |
18 | #include "PPCRegisterInfo.h" |
19 | #include "llvm/ADT/SmallSet.h" |
20 | #include "llvm/CodeGen/LiveVariables.h" |
21 | #include "llvm/CodeGen/TargetInstrInfo.h" |
22 | |
23 | #define |
24 | #include "PPCGenInstrInfo.inc" |
25 | |
26 | namespace llvm { |
27 | |
28 | // Instructions that have an immediate form might be convertible to that |
29 | // form if the correct input is a result of a load immediate. In order to |
30 | // know whether the transformation is special, we might need to know some |
31 | // of the details of the two forms. |
32 | struct ImmInstrInfo { |
33 | // Is the immediate field in the immediate form signed or unsigned? |
34 | uint64_t SignedImm : 1; |
35 | // Does the immediate need to be a multiple of some value? |
36 | uint64_t ImmMustBeMultipleOf : 5; |
37 | // Is R0/X0 treated specially by the original r+r instruction? |
38 | // If so, in which operand? |
39 | uint64_t ZeroIsSpecialOrig : 3; |
40 | // Is R0/X0 treated specially by the new r+i instruction? |
41 | // If so, in which operand? |
42 | uint64_t ZeroIsSpecialNew : 3; |
43 | // Is the operation commutative? |
44 | uint64_t IsCommutative : 1; |
45 | // The operand number to check for add-immediate def. |
46 | uint64_t OpNoForForwarding : 3; |
47 | // The operand number for the immediate. |
48 | uint64_t ImmOpNo : 3; |
49 | // The opcode of the new instruction. |
50 | uint64_t ImmOpcode : 16; |
51 | // The size of the immediate. |
52 | uint64_t ImmWidth : 5; |
53 | // The immediate should be truncated to N bits. |
54 | uint64_t TruncateImmTo : 5; |
55 | // Is the instruction summing the operand |
56 | uint64_t IsSummingOperands : 1; |
57 | }; |
58 | |
59 | // Information required to convert an instruction to just a materialized |
60 | // immediate. |
61 | struct LoadImmediateInfo { |
62 | unsigned Imm : 16; |
63 | unsigned Is64Bit : 1; |
64 | unsigned SetCR : 1; |
65 | }; |
66 | |
67 | // Index into the OpcodesForSpill array. |
68 | enum SpillOpcodeKey { |
69 | SOK_Int4Spill, |
70 | SOK_Int8Spill, |
71 | SOK_Float8Spill, |
72 | SOK_Float4Spill, |
73 | SOK_CRSpill, |
74 | SOK_CRBitSpill, |
75 | SOK_VRVectorSpill, |
76 | SOK_VSXVectorSpill, |
77 | SOK_VectorFloat8Spill, |
78 | SOK_VectorFloat4Spill, |
79 | SOK_SpillToVSR, |
80 | SOK_PairedVecSpill, |
81 | SOK_AccumulatorSpill, |
82 | SOK_UAccumulatorSpill, |
83 | SOK_WAccumulatorSpill, |
84 | SOK_DMRpSpill, |
85 | SOK_DMRSpill, |
86 | SOK_SPESpill, |
87 | SOK_PairedG8Spill, |
88 | SOK_LastOpcodeSpill // This must be last on the enum. |
89 | }; |
90 | |
91 | // PPC MachineCombiner patterns |
92 | enum PPCMachineCombinerPattern : unsigned { |
93 | // These are patterns matched by the PowerPC to reassociate FMA chains. |
94 | REASSOC_XY_AMM_BMM = MachineCombinerPattern::TARGET_PATTERN_START, |
95 | REASSOC_XMM_AMM_BMM, |
96 | |
97 | // These are patterns matched by the PowerPC to reassociate FMA and FSUB to |
98 | // reduce register pressure. |
99 | REASSOC_XY_BCA, |
100 | REASSOC_XY_BAC, |
101 | |
102 | }; |
103 | |
104 | // Define list of load and store spill opcodes. |
105 | #define NoInstr PPC::INSTRUCTION_LIST_END |
106 | #define Pwr8LoadOpcodes \ |
107 | {PPC::LWZ, \ |
108 | PPC::LD, \ |
109 | PPC::LFD, \ |
110 | PPC::LFS, \ |
111 | PPC::RESTORE_CR, \ |
112 | PPC::RESTORE_CRBIT, \ |
113 | PPC::LVX, \ |
114 | PPC::LXVD2X, \ |
115 | PPC::LXSDX, \ |
116 | PPC::LXSSPX, \ |
117 | PPC::SPILLTOVSR_LD, \ |
118 | NoInstr, \ |
119 | NoInstr, \ |
120 | NoInstr, \ |
121 | NoInstr, \ |
122 | NoInstr, \ |
123 | NoInstr, \ |
124 | PPC::EVLDD, \ |
125 | PPC::RESTORE_QUADWORD} |
126 | |
127 | #define Pwr9LoadOpcodes \ |
128 | {PPC::LWZ, \ |
129 | PPC::LD, \ |
130 | PPC::LFD, \ |
131 | PPC::LFS, \ |
132 | PPC::RESTORE_CR, \ |
133 | PPC::RESTORE_CRBIT, \ |
134 | PPC::LVX, \ |
135 | PPC::LXV, \ |
136 | PPC::DFLOADf64, \ |
137 | PPC::DFLOADf32, \ |
138 | PPC::SPILLTOVSR_LD, \ |
139 | NoInstr, \ |
140 | NoInstr, \ |
141 | NoInstr, \ |
142 | NoInstr, \ |
143 | NoInstr, \ |
144 | NoInstr, \ |
145 | NoInstr, \ |
146 | PPC::RESTORE_QUADWORD} |
147 | |
148 | #define Pwr10LoadOpcodes \ |
149 | {PPC::LWZ, \ |
150 | PPC::LD, \ |
151 | PPC::LFD, \ |
152 | PPC::LFS, \ |
153 | PPC::RESTORE_CR, \ |
154 | PPC::RESTORE_CRBIT, \ |
155 | PPC::LVX, \ |
156 | PPC::LXV, \ |
157 | PPC::DFLOADf64, \ |
158 | PPC::DFLOADf32, \ |
159 | PPC::SPILLTOVSR_LD, \ |
160 | PPC::LXVP, \ |
161 | PPC::RESTORE_ACC, \ |
162 | PPC::RESTORE_UACC, \ |
163 | NoInstr, \ |
164 | NoInstr, \ |
165 | NoInstr, \ |
166 | NoInstr, \ |
167 | PPC::RESTORE_QUADWORD} |
168 | |
169 | #define FutureLoadOpcodes \ |
170 | {PPC::LWZ, \ |
171 | PPC::LD, \ |
172 | PPC::LFD, \ |
173 | PPC::LFS, \ |
174 | PPC::RESTORE_CR, \ |
175 | PPC::RESTORE_CRBIT, \ |
176 | PPC::LVX, \ |
177 | PPC::LXV, \ |
178 | PPC::DFLOADf64, \ |
179 | PPC::DFLOADf32, \ |
180 | PPC::SPILLTOVSR_LD, \ |
181 | PPC::LXVP, \ |
182 | PPC::RESTORE_ACC, \ |
183 | PPC::RESTORE_UACC, \ |
184 | PPC::RESTORE_WACC, \ |
185 | PPC::RESTORE_DMRP, \ |
186 | PPC::RESTORE_DMR, \ |
187 | NoInstr, \ |
188 | PPC::RESTORE_QUADWORD} |
189 | |
190 | #define Pwr8StoreOpcodes \ |
191 | {PPC::STW, \ |
192 | PPC::STD, \ |
193 | PPC::STFD, \ |
194 | PPC::STFS, \ |
195 | PPC::SPILL_CR, \ |
196 | PPC::SPILL_CRBIT, \ |
197 | PPC::STVX, \ |
198 | PPC::STXVD2X, \ |
199 | PPC::STXSDX, \ |
200 | PPC::STXSSPX, \ |
201 | PPC::SPILLTOVSR_ST, \ |
202 | NoInstr, \ |
203 | NoInstr, \ |
204 | NoInstr, \ |
205 | NoInstr, \ |
206 | NoInstr, \ |
207 | NoInstr, \ |
208 | PPC::EVSTDD, \ |
209 | PPC::SPILL_QUADWORD} |
210 | |
211 | #define Pwr9StoreOpcodes \ |
212 | {PPC::STW, \ |
213 | PPC::STD, \ |
214 | PPC::STFD, \ |
215 | PPC::STFS, \ |
216 | PPC::SPILL_CR, \ |
217 | PPC::SPILL_CRBIT, \ |
218 | PPC::STVX, \ |
219 | PPC::STXV, \ |
220 | PPC::DFSTOREf64, \ |
221 | PPC::DFSTOREf32, \ |
222 | PPC::SPILLTOVSR_ST, \ |
223 | NoInstr, \ |
224 | NoInstr, \ |
225 | NoInstr, \ |
226 | NoInstr, \ |
227 | NoInstr, \ |
228 | NoInstr, \ |
229 | NoInstr, \ |
230 | PPC::SPILL_QUADWORD} |
231 | |
232 | #define Pwr10StoreOpcodes \ |
233 | {PPC::STW, \ |
234 | PPC::STD, \ |
235 | PPC::STFD, \ |
236 | PPC::STFS, \ |
237 | PPC::SPILL_CR, \ |
238 | PPC::SPILL_CRBIT, \ |
239 | PPC::STVX, \ |
240 | PPC::STXV, \ |
241 | PPC::DFSTOREf64, \ |
242 | PPC::DFSTOREf32, \ |
243 | PPC::SPILLTOVSR_ST, \ |
244 | PPC::STXVP, \ |
245 | PPC::SPILL_ACC, \ |
246 | PPC::SPILL_UACC, \ |
247 | NoInstr, \ |
248 | NoInstr, \ |
249 | NoInstr, \ |
250 | NoInstr, \ |
251 | PPC::SPILL_QUADWORD} |
252 | |
253 | #define FutureStoreOpcodes \ |
254 | {PPC::STW, \ |
255 | PPC::STD, \ |
256 | PPC::STFD, \ |
257 | PPC::STFS, \ |
258 | PPC::SPILL_CR, \ |
259 | PPC::SPILL_CRBIT, \ |
260 | PPC::STVX, \ |
261 | PPC::STXV, \ |
262 | PPC::DFSTOREf64, \ |
263 | PPC::DFSTOREf32, \ |
264 | PPC::SPILLTOVSR_ST, \ |
265 | PPC::STXVP, \ |
266 | PPC::SPILL_ACC, \ |
267 | PPC::SPILL_UACC, \ |
268 | PPC::SPILL_WACC, \ |
269 | PPC::SPILL_DMRP, \ |
270 | PPC::SPILL_DMR, \ |
271 | NoInstr, \ |
272 | PPC::SPILL_QUADWORD} |
273 | |
274 | // Initialize arrays for load and store spill opcodes on supported subtargets. |
275 | #define StoreOpcodesForSpill \ |
276 | { Pwr8StoreOpcodes, Pwr9StoreOpcodes, Pwr10StoreOpcodes, FutureStoreOpcodes } |
277 | #define LoadOpcodesForSpill \ |
278 | { Pwr8LoadOpcodes, Pwr9LoadOpcodes, Pwr10LoadOpcodes, FutureLoadOpcodes } |
279 | |
280 | class PPCSubtarget; |
281 | class PPCInstrInfo : public PPCGenInstrInfo { |
282 | PPCSubtarget &Subtarget; |
283 | const PPCRegisterInfo RI; |
284 | const unsigned StoreSpillOpcodesArray[4][SOK_LastOpcodeSpill] = |
285 | StoreOpcodesForSpill; |
286 | const unsigned LoadSpillOpcodesArray[4][SOK_LastOpcodeSpill] = |
287 | LoadOpcodesForSpill; |
288 | |
289 | void StoreRegToStackSlot(MachineFunction &MF, unsigned SrcReg, bool isKill, |
290 | int FrameIdx, const TargetRegisterClass *RC, |
291 | SmallVectorImpl<MachineInstr *> &NewMIs) const; |
292 | void LoadRegFromStackSlot(MachineFunction &MF, const DebugLoc &DL, |
293 | unsigned DestReg, int FrameIdx, |
294 | const TargetRegisterClass *RC, |
295 | SmallVectorImpl<MachineInstr *> &NewMIs) const; |
296 | |
297 | // Replace the instruction with single LI if possible. \p DefMI must be LI or |
298 | // LI8. |
299 | bool simplifyToLI(MachineInstr &MI, MachineInstr &DefMI, |
300 | unsigned OpNoForForwarding, MachineInstr **KilledDef) const; |
301 | // If the inst is imm-form and its register operand is produced by a ADDI, put |
302 | // the imm into the inst directly and remove the ADDI if possible. |
303 | bool transformToNewImmFormFedByAdd(MachineInstr &MI, MachineInstr &DefMI, |
304 | unsigned OpNoForForwarding) const; |
305 | // If the inst is x-form and has imm-form and one of its operand is produced |
306 | // by a LI, put the imm into the inst directly and remove the LI if possible. |
307 | bool transformToImmFormFedByLI(MachineInstr &MI, const ImmInstrInfo &III, |
308 | unsigned ConstantOpNo, |
309 | MachineInstr &DefMI) const; |
310 | // If the inst is x-form and has imm-form and one of its operand is produced |
311 | // by an add-immediate, try to transform it when possible. |
312 | bool transformToImmFormFedByAdd(MachineInstr &MI, const ImmInstrInfo &III, |
313 | unsigned ConstantOpNo, MachineInstr &DefMI, |
314 | bool KillDefMI) const; |
315 | // Try to find that, if the instruction 'MI' contains any operand that |
316 | // could be forwarded from some inst that feeds it. If yes, return the |
317 | // Def of that operand. And OpNoForForwarding is the operand index in |
318 | // the 'MI' for that 'Def'. If we see another use of this Def between |
319 | // the Def and the MI, SeenIntermediateUse becomes 'true'. |
320 | MachineInstr *getForwardingDefMI(MachineInstr &MI, |
321 | unsigned &OpNoForForwarding, |
322 | bool &SeenIntermediateUse) const; |
323 | |
324 | // Can the user MI have it's source at index \p OpNoForForwarding |
325 | // forwarded from an add-immediate that feeds it? |
326 | bool isUseMIElgibleForForwarding(MachineInstr &MI, const ImmInstrInfo &III, |
327 | unsigned OpNoForForwarding) const; |
328 | bool isDefMIElgibleForForwarding(MachineInstr &DefMI, |
329 | const ImmInstrInfo &III, |
330 | MachineOperand *&ImmMO, |
331 | MachineOperand *&RegMO) const; |
332 | bool isImmElgibleForForwarding(const MachineOperand &ImmMO, |
333 | const MachineInstr &DefMI, |
334 | const ImmInstrInfo &III, |
335 | int64_t &Imm, |
336 | int64_t BaseImm = 0) const; |
337 | bool isRegElgibleForForwarding(const MachineOperand &RegMO, |
338 | const MachineInstr &DefMI, |
339 | const MachineInstr &MI, bool KillDefMI, |
340 | bool &IsFwdFeederRegKilled, |
341 | bool &SeenIntermediateUse) const; |
342 | unsigned getSpillTarget() const; |
343 | ArrayRef<unsigned> getStoreOpcodesForSpillArray() const; |
344 | ArrayRef<unsigned> getLoadOpcodesForSpillArray() const; |
345 | unsigned getSpillIndex(const TargetRegisterClass *RC) const; |
346 | int16_t getFMAOpIdxInfo(unsigned Opcode) const; |
347 | void reassociateFMA(MachineInstr &Root, unsigned Pattern, |
348 | SmallVectorImpl<MachineInstr *> &InsInstrs, |
349 | SmallVectorImpl<MachineInstr *> &DelInstrs, |
350 | DenseMap<Register, unsigned> &InstrIdxForVirtReg) const; |
351 | Register |
352 | generateLoadForNewConst(unsigned Idx, MachineInstr *MI, Type *Ty, |
353 | SmallVectorImpl<MachineInstr *> &InsInstrs) const; |
354 | virtual void anchor(); |
355 | |
356 | protected: |
357 | /// Commutes the operands in the given instruction. |
358 | /// The commutable operands are specified by their indices OpIdx1 and OpIdx2. |
359 | /// |
360 | /// Do not call this method for a non-commutable instruction or for |
361 | /// non-commutable pair of operand indices OpIdx1 and OpIdx2. |
362 | /// Even though the instruction is commutable, the method may still |
363 | /// fail to commute the operands, null pointer is returned in such cases. |
364 | /// |
365 | /// For example, we can commute rlwimi instructions, but only if the |
366 | /// rotate amt is zero. We also have to munge the immediates a bit. |
367 | MachineInstr *commuteInstructionImpl(MachineInstr &MI, bool NewMI, |
368 | unsigned OpIdx1, |
369 | unsigned OpIdx2) const override; |
370 | |
371 | public: |
372 | explicit PPCInstrInfo(PPCSubtarget &STI); |
373 | |
374 | bool isLoadFromConstantPool(MachineInstr *I) const; |
375 | const Constant *getConstantFromConstantPool(MachineInstr *I) const; |
376 | |
377 | /// getRegisterInfo - TargetInstrInfo is a superset of MRegister info. As |
378 | /// such, whenever a client has an instance of instruction info, it should |
379 | /// always be able to get register info as well (through this method). |
380 | /// |
381 | const PPCRegisterInfo &getRegisterInfo() const { return RI; } |
382 | |
383 | bool isXFormMemOp(unsigned Opcode) const { |
384 | return get(Opcode).TSFlags & PPCII::XFormMemOp; |
385 | } |
386 | bool isPrefixed(unsigned Opcode) const { |
387 | return get(Opcode).TSFlags & PPCII::Prefixed; |
388 | } |
389 | bool isSExt32To64(unsigned Opcode) const { |
390 | return get(Opcode).TSFlags & PPCII::SExt32To64; |
391 | } |
392 | bool isZExt32To64(unsigned Opcode) const { |
393 | return get(Opcode).TSFlags & PPCII::ZExt32To64; |
394 | } |
395 | bool isMemriOp(unsigned Opcode) const { |
396 | return get(Opcode).TSFlags & PPCII::MemriOp; |
397 | } |
398 | |
399 | static bool isSameClassPhysRegCopy(unsigned Opcode) { |
400 | unsigned CopyOpcodes[] = {PPC::OR, PPC::OR8, PPC::FMR, |
401 | PPC::VOR, PPC::XXLOR, PPC::XXLORf, |
402 | PPC::XSCPSGNDP, PPC::MCRF, PPC::CROR, |
403 | PPC::EVOR, -1U}; |
404 | for (int i = 0; CopyOpcodes[i] != -1U; i++) |
405 | if (Opcode == CopyOpcodes[i]) |
406 | return true; |
407 | return false; |
408 | } |
409 | |
410 | static bool hasPCRelFlag(unsigned TF) { |
411 | return TF == PPCII::MO_PCREL_FLAG || TF == PPCII::MO_GOT_TLSGD_PCREL_FLAG || |
412 | TF == PPCII::MO_GOT_TLSLD_PCREL_FLAG || |
413 | TF == PPCII::MO_GOT_TPREL_PCREL_FLAG || |
414 | TF == PPCII::MO_TPREL_PCREL_FLAG || TF == PPCII::MO_TLS_PCREL_FLAG || |
415 | TF == PPCII::MO_GOT_PCREL_FLAG; |
416 | } |
417 | |
418 | static bool hasGOTFlag(unsigned TF) { |
419 | return TF == PPCII::MO_GOT_FLAG || TF == PPCII::MO_GOT_TLSGD_PCREL_FLAG || |
420 | TF == PPCII::MO_GOT_TLSLD_PCREL_FLAG || |
421 | TF == PPCII::MO_GOT_TPREL_PCREL_FLAG || |
422 | TF == PPCII::MO_GOT_PCREL_FLAG; |
423 | } |
424 | |
425 | static bool hasTLSFlag(unsigned TF) { |
426 | return TF == PPCII::MO_TLSGD_FLAG || TF == PPCII::MO_TPREL_FLAG || |
427 | TF == PPCII::MO_TLSLD_FLAG || TF == PPCII::MO_TLSGDM_FLAG || |
428 | TF == PPCII::MO_GOT_TLSGD_PCREL_FLAG || |
429 | TF == PPCII::MO_GOT_TLSLD_PCREL_FLAG || |
430 | TF == PPCII::MO_GOT_TPREL_PCREL_FLAG || TF == PPCII::MO_TPREL_LO || |
431 | TF == PPCII::MO_TPREL_HA || TF == PPCII::MO_DTPREL_LO || |
432 | TF == PPCII::MO_TLSLD_LO || TF == PPCII::MO_TLS || |
433 | TF == PPCII::MO_TPREL_PCREL_FLAG || TF == PPCII::MO_TLS_PCREL_FLAG; |
434 | } |
435 | |
436 | ScheduleHazardRecognizer * |
437 | CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI, |
438 | const ScheduleDAG *DAG) const override; |
439 | ScheduleHazardRecognizer * |
440 | CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II, |
441 | const ScheduleDAG *DAG) const override; |
442 | |
443 | unsigned getInstrLatency(const InstrItineraryData *ItinData, |
444 | const MachineInstr &MI, |
445 | unsigned *PredCost = nullptr) const override; |
446 | |
447 | std::optional<unsigned> getOperandLatency(const InstrItineraryData *ItinData, |
448 | const MachineInstr &DefMI, |
449 | unsigned DefIdx, |
450 | const MachineInstr &UseMI, |
451 | unsigned UseIdx) const override; |
452 | std::optional<unsigned> getOperandLatency(const InstrItineraryData *ItinData, |
453 | SDNode *DefNode, unsigned DefIdx, |
454 | SDNode *UseNode, |
455 | unsigned UseIdx) const override { |
456 | return PPCGenInstrInfo::getOperandLatency(ItinData, DefNode, DefIdx, |
457 | UseNode, UseIdx); |
458 | } |
459 | |
460 | bool hasLowDefLatency(const TargetSchedModel &SchedModel, |
461 | const MachineInstr &DefMI, |
462 | unsigned DefIdx) const override { |
463 | // Machine LICM should hoist all instructions in low-register-pressure |
464 | // situations; none are sufficiently free to justify leaving in a loop |
465 | // body. |
466 | return false; |
467 | } |
468 | |
469 | bool useMachineCombiner() const override { |
470 | return true; |
471 | } |
472 | |
473 | /// When getMachineCombinerPatterns() finds patterns, this function generates |
474 | /// the instructions that could replace the original code sequence |
475 | void genAlternativeCodeSequence( |
476 | MachineInstr &Root, unsigned Pattern, |
477 | SmallVectorImpl<MachineInstr *> &InsInstrs, |
478 | SmallVectorImpl<MachineInstr *> &DelInstrs, |
479 | DenseMap<Register, unsigned> &InstrIdxForVirtReg) const override; |
480 | |
481 | /// Return true when there is potentially a faster code sequence for a fma |
482 | /// chain ending in \p Root. All potential patterns are output in the \p |
483 | /// P array. |
484 | bool getFMAPatterns(MachineInstr &Root, SmallVectorImpl<unsigned> &Patterns, |
485 | bool DoRegPressureReduce) const; |
486 | |
487 | CombinerObjective getCombinerObjective(unsigned Pattern) const override; |
488 | |
489 | /// Return true when there is potentially a faster code sequence |
490 | /// for an instruction chain ending in <Root>. All potential patterns are |
491 | /// output in the <Pattern> array. |
492 | bool getMachineCombinerPatterns(MachineInstr &Root, |
493 | SmallVectorImpl<unsigned> &Patterns, |
494 | bool DoRegPressureReduce) const override; |
495 | |
496 | /// On PowerPC, we leverage machine combiner pass to reduce register pressure |
497 | /// when the register pressure is high for one BB. |
498 | /// Return true if register pressure for \p MBB is high and ABI is supported |
499 | /// to reduce register pressure. Otherwise return false. |
500 | bool shouldReduceRegisterPressure( |
501 | const MachineBasicBlock *MBB, |
502 | const RegisterClassInfo *RegClassInfo) const override; |
503 | |
504 | /// Fixup the placeholders we put in genAlternativeCodeSequence() for |
505 | /// MachineCombiner. |
506 | void |
507 | finalizeInsInstrs(MachineInstr &Root, unsigned &Pattern, |
508 | SmallVectorImpl<MachineInstr *> &InsInstrs) const override; |
509 | |
510 | bool isAssociativeAndCommutative(const MachineInstr &Inst, |
511 | bool Invert) const override; |
512 | |
513 | /// On PowerPC, we try to reassociate FMA chain which will increase |
514 | /// instruction size. Set extension resource length limit to 1 for edge case. |
515 | /// Resource Length is calculated by scaled resource usage in getCycles(). |
516 | /// Because of the division in getCycles(), it returns different cycles due to |
517 | /// legacy scaled resource usage. So new resource length may be same with |
518 | /// legacy or 1 bigger than legacy. |
519 | /// We need to execlude the 1 bigger case even the resource length is not |
520 | /// perserved for more FMA chain reassociations on PowerPC. |
521 | int getExtendResourceLenLimit() const override { return 1; } |
522 | |
523 | // PowerPC specific version of setSpecialOperandAttr that copies Flags to MI |
524 | // and clears nuw, nsw, and exact flags. |
525 | using TargetInstrInfo::setSpecialOperandAttr; |
526 | void setSpecialOperandAttr(MachineInstr &MI, uint32_t Flags) const; |
527 | |
528 | bool isCoalescableExtInstr(const MachineInstr &MI, |
529 | Register &SrcReg, Register &DstReg, |
530 | unsigned &SubIdx) const override; |
531 | Register isLoadFromStackSlot(const MachineInstr &MI, |
532 | int &FrameIndex) const override; |
533 | bool isReallyTriviallyReMaterializable(const MachineInstr &MI) const override; |
534 | Register isStoreToStackSlot(const MachineInstr &MI, |
535 | int &FrameIndex) const override; |
536 | |
537 | bool findCommutedOpIndices(const MachineInstr &MI, unsigned &SrcOpIdx1, |
538 | unsigned &SrcOpIdx2) const override; |
539 | |
540 | void insertNoop(MachineBasicBlock &MBB, |
541 | MachineBasicBlock::iterator MI) const override; |
542 | |
543 | |
544 | // Branch analysis. |
545 | bool analyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB, |
546 | MachineBasicBlock *&FBB, |
547 | SmallVectorImpl<MachineOperand> &Cond, |
548 | bool AllowModify) const override; |
549 | unsigned removeBranch(MachineBasicBlock &MBB, |
550 | int *BytesRemoved = nullptr) const override; |
551 | unsigned insertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB, |
552 | MachineBasicBlock *FBB, ArrayRef<MachineOperand> Cond, |
553 | const DebugLoc &DL, |
554 | int *BytesAdded = nullptr) const override; |
555 | |
556 | // Select analysis. |
557 | bool canInsertSelect(const MachineBasicBlock &, ArrayRef<MachineOperand> Cond, |
558 | Register, Register, Register, int &, int &, |
559 | int &) const override; |
560 | void insertSelect(MachineBasicBlock &MBB, MachineBasicBlock::iterator MI, |
561 | const DebugLoc &DL, Register DstReg, |
562 | ArrayRef<MachineOperand> Cond, Register TrueReg, |
563 | Register FalseReg) const override; |
564 | |
565 | void copyPhysReg(MachineBasicBlock &MBB, MachineBasicBlock::iterator I, |
566 | const DebugLoc &DL, Register DestReg, Register SrcReg, |
567 | bool KillSrc, bool RenamableDest = false, |
568 | bool RenamableSrc = false) const override; |
569 | |
570 | void storeRegToStackSlot( |
571 | MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, Register SrcReg, |
572 | bool isKill, int FrameIndex, const TargetRegisterClass *RC, |
573 | const TargetRegisterInfo *TRI, Register VReg, |
574 | MachineInstr::MIFlag Flags = MachineInstr::NoFlags) const override; |
575 | |
576 | // Emits a register spill without updating the register class for vector |
577 | // registers. This ensures that when we spill a vector register the |
578 | // element order in the register is the same as it was in memory. |
579 | void storeRegToStackSlotNoUpd(MachineBasicBlock &MBB, |
580 | MachineBasicBlock::iterator MBBI, |
581 | unsigned SrcReg, bool isKill, int FrameIndex, |
582 | const TargetRegisterClass *RC, |
583 | const TargetRegisterInfo *TRI) const; |
584 | |
585 | void loadRegFromStackSlot( |
586 | MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, |
587 | Register DestReg, int FrameIndex, const TargetRegisterClass *RC, |
588 | const TargetRegisterInfo *TRI, Register VReg, |
589 | MachineInstr::MIFlag Flags = MachineInstr::NoFlags) const override; |
590 | |
591 | // Emits a register reload without updating the register class for vector |
592 | // registers. This ensures that when we reload a vector register the |
593 | // element order in the register is the same as it was in memory. |
594 | void loadRegFromStackSlotNoUpd(MachineBasicBlock &MBB, |
595 | MachineBasicBlock::iterator MBBI, |
596 | unsigned DestReg, int FrameIndex, |
597 | const TargetRegisterClass *RC, |
598 | const TargetRegisterInfo *TRI) const; |
599 | |
600 | unsigned getStoreOpcodeForSpill(const TargetRegisterClass *RC) const; |
601 | |
602 | unsigned getLoadOpcodeForSpill(const TargetRegisterClass *RC) const; |
603 | |
604 | bool |
605 | reverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const override; |
606 | |
607 | bool foldImmediate(MachineInstr &UseMI, MachineInstr &DefMI, Register Reg, |
608 | MachineRegisterInfo *MRI) const override; |
609 | |
610 | bool onlyFoldImmediate(MachineInstr &UseMI, MachineInstr &DefMI, |
611 | Register Reg) const; |
612 | |
613 | // If conversion by predication (only supported by some branch instructions). |
614 | // All of the profitability checks always return true; it is always |
615 | // profitable to use the predicated branches. |
616 | bool isProfitableToIfCvt(MachineBasicBlock &MBB, |
617 | unsigned NumCycles, unsigned , |
618 | BranchProbability Probability) const override { |
619 | return true; |
620 | } |
621 | |
622 | bool isProfitableToIfCvt(MachineBasicBlock &TMBB, |
623 | unsigned NumT, unsigned , |
624 | MachineBasicBlock &FMBB, |
625 | unsigned NumF, unsigned , |
626 | BranchProbability Probability) const override; |
627 | |
628 | bool isProfitableToDupForIfCvt(MachineBasicBlock &MBB, unsigned NumCycles, |
629 | BranchProbability Probability) const override { |
630 | return true; |
631 | } |
632 | |
633 | bool isProfitableToUnpredicate(MachineBasicBlock &TMBB, |
634 | MachineBasicBlock &FMBB) const override { |
635 | return false; |
636 | } |
637 | |
638 | // Predication support. |
639 | bool isPredicated(const MachineInstr &MI) const override; |
640 | |
641 | bool isSchedulingBoundary(const MachineInstr &MI, |
642 | const MachineBasicBlock *MBB, |
643 | const MachineFunction &MF) const override; |
644 | |
645 | bool PredicateInstruction(MachineInstr &MI, |
646 | ArrayRef<MachineOperand> Pred) const override; |
647 | |
648 | bool SubsumesPredicate(ArrayRef<MachineOperand> Pred1, |
649 | ArrayRef<MachineOperand> Pred2) const override; |
650 | |
651 | bool ClobbersPredicate(MachineInstr &MI, std::vector<MachineOperand> &Pred, |
652 | bool SkipDead) const override; |
653 | |
654 | // Comparison optimization. |
655 | |
656 | bool analyzeCompare(const MachineInstr &MI, Register &SrcReg, |
657 | Register &SrcReg2, int64_t &Mask, |
658 | int64_t &Value) const override; |
659 | |
660 | bool optimizeCompareInstr(MachineInstr &CmpInstr, Register SrcReg, |
661 | Register SrcReg2, int64_t Mask, int64_t Value, |
662 | const MachineRegisterInfo *MRI) const override; |
663 | |
664 | |
665 | /// Return true if get the base operand, byte offset of an instruction and |
666 | /// the memory width. Width is the size of memory that is being |
667 | /// loaded/stored (e.g. 1, 2, 4, 8). |
668 | bool getMemOperandWithOffsetWidth(const MachineInstr &LdSt, |
669 | const MachineOperand *&BaseOp, |
670 | int64_t &Offset, LocationSize &Width, |
671 | const TargetRegisterInfo *TRI) const; |
672 | |
673 | bool optimizeCmpPostRA(MachineInstr &MI) const; |
674 | |
675 | /// Get the base operand and byte offset of an instruction that reads/writes |
676 | /// memory. |
677 | bool getMemOperandsWithOffsetWidth( |
678 | const MachineInstr &LdSt, |
679 | SmallVectorImpl<const MachineOperand *> &BaseOps, int64_t &Offset, |
680 | bool &OffsetIsScalable, LocationSize &Width, |
681 | const TargetRegisterInfo *TRI) const override; |
682 | |
683 | /// Returns true if the two given memory operations should be scheduled |
684 | /// adjacent. |
685 | bool shouldClusterMemOps(ArrayRef<const MachineOperand *> BaseOps1, |
686 | int64_t Offset1, bool OffsetIsScalable1, |
687 | ArrayRef<const MachineOperand *> BaseOps2, |
688 | int64_t Offset2, bool OffsetIsScalable2, |
689 | unsigned ClusterSize, |
690 | unsigned NumBytes) const override; |
691 | |
692 | /// Return true if two MIs access different memory addresses and false |
693 | /// otherwise |
694 | bool |
695 | areMemAccessesTriviallyDisjoint(const MachineInstr &MIa, |
696 | const MachineInstr &MIb) const override; |
697 | |
698 | /// GetInstSize - Return the number of bytes of code the specified |
699 | /// instruction may be. This returns the maximum number of bytes. |
700 | /// |
701 | unsigned getInstSizeInBytes(const MachineInstr &MI) const override; |
702 | |
703 | MCInst getNop() const override; |
704 | |
705 | std::pair<unsigned, unsigned> |
706 | decomposeMachineOperandsTargetFlags(unsigned TF) const override; |
707 | |
708 | ArrayRef<std::pair<unsigned, const char *>> |
709 | getSerializableDirectMachineOperandTargetFlags() const override; |
710 | |
711 | // Expand VSX Memory Pseudo instruction to either a VSX or a FP instruction. |
712 | bool expandVSXMemPseudo(MachineInstr &MI) const; |
713 | |
714 | // Lower pseudo instructions after register allocation. |
715 | bool expandPostRAPseudo(MachineInstr &MI) const override; |
716 | |
717 | const TargetRegisterClass *updatedRC(const TargetRegisterClass *RC) const; |
718 | static int getRecordFormOpcode(unsigned Opcode); |
719 | |
720 | bool isTOCSaveMI(const MachineInstr &MI) const; |
721 | |
722 | std::pair<bool, bool> |
723 | isSignOrZeroExtended(const unsigned Reg, const unsigned BinOpDepth, |
724 | const MachineRegisterInfo *MRI) const; |
725 | |
726 | // Return true if the register is sign-extended from 32 to 64 bits. |
727 | bool isSignExtended(const unsigned Reg, |
728 | const MachineRegisterInfo *MRI) const { |
729 | return isSignOrZeroExtended(Reg, BinOpDepth: 0, MRI).first; |
730 | } |
731 | |
732 | // Return true if the register is zero-extended from 32 to 64 bits. |
733 | bool isZeroExtended(const unsigned Reg, |
734 | const MachineRegisterInfo *MRI) const { |
735 | return isSignOrZeroExtended(Reg, BinOpDepth: 0, MRI).second; |
736 | } |
737 | void promoteInstr32To64ForElimEXTSW(const Register &Reg, |
738 | MachineRegisterInfo *MRI, |
739 | unsigned BinOpDepth, |
740 | LiveVariables *LV) const; |
741 | |
742 | bool convertToImmediateForm(MachineInstr &MI, |
743 | SmallSet<Register, 4> &RegsToUpdate, |
744 | MachineInstr **KilledDef = nullptr) const; |
745 | bool foldFrameOffset(MachineInstr &MI) const; |
746 | bool combineRLWINM(MachineInstr &MI, MachineInstr **ToErase = nullptr) const; |
747 | bool isADDIInstrEligibleForFolding(MachineInstr &ADDIMI, int64_t &Imm) const; |
748 | bool isADDInstrEligibleForFolding(MachineInstr &ADDMI) const; |
749 | bool isImmInstrEligibleForFolding(MachineInstr &MI, unsigned &BaseReg, |
750 | unsigned &XFormOpcode, |
751 | int64_t &OffsetOfImmInstr, |
752 | ImmInstrInfo &III) const; |
753 | bool isValidToBeChangedReg(MachineInstr *ADDMI, unsigned Index, |
754 | MachineInstr *&ADDIMI, int64_t &OffsetAddi, |
755 | int64_t OffsetImm) const; |
756 | |
757 | void replaceInstrWithLI(MachineInstr &MI, const LoadImmediateInfo &LII) const; |
758 | void replaceInstrOperandWithImm(MachineInstr &MI, unsigned OpNo, |
759 | int64_t Imm) const; |
760 | |
761 | bool instrHasImmForm(unsigned Opc, bool IsVFReg, ImmInstrInfo &III, |
762 | bool PostRA) const; |
763 | |
764 | // In PostRA phase, try to find instruction defines \p Reg before \p MI. |
765 | // \p SeenIntermediate is set to true if uses between DefMI and \p MI exist. |
766 | MachineInstr *getDefMIPostRA(unsigned Reg, MachineInstr &MI, |
767 | bool &SeenIntermediateUse) const; |
768 | |
769 | // Materialize immediate after RA. |
770 | void materializeImmPostRA(MachineBasicBlock &MBB, |
771 | MachineBasicBlock::iterator MBBI, |
772 | const DebugLoc &DL, Register Reg, |
773 | int64_t Imm) const; |
774 | |
775 | /// Check \p Opcode is BDNZ (Decrement CTR and branch if it is still nonzero). |
776 | bool isBDNZ(unsigned Opcode) const; |
777 | |
778 | /// Find the hardware loop instruction used to set-up the specified loop. |
779 | /// On PPC, we have two instructions used to set-up the hardware loop |
780 | /// (MTCTRloop, MTCTR8loop) with corresponding endloop (BDNZ, BDNZ8) |
781 | /// instructions to indicate the end of a loop. |
782 | MachineInstr * |
783 | findLoopInstr(MachineBasicBlock &, |
784 | SmallPtrSet<MachineBasicBlock *, 8> &Visited) const; |
785 | |
786 | /// Analyze loop L, which must be a single-basic-block loop, and if the |
787 | /// conditions can be understood enough produce a PipelinerLoopInfo object. |
788 | std::unique_ptr<TargetInstrInfo::PipelinerLoopInfo> |
789 | analyzeLoopForPipelining(MachineBasicBlock *LoopBB) const override; |
790 | }; |
791 | |
792 | } |
793 | |
794 | #endif |
795 | |