1 | //===- PeepholeOptimizer.cpp - Peephole Optimizations ---------------------===// |
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 | // Perform peephole optimizations on the machine code: |
10 | // |
11 | // - Optimize Extensions |
12 | // |
13 | // Optimization of sign / zero extension instructions. It may be extended to |
14 | // handle other instructions with similar properties. |
15 | // |
16 | // On some targets, some instructions, e.g. X86 sign / zero extension, may |
17 | // leave the source value in the lower part of the result. This optimization |
18 | // will replace some uses of the pre-extension value with uses of the |
19 | // sub-register of the results. |
20 | // |
21 | // - Optimize Comparisons |
22 | // |
23 | // Optimization of comparison instructions. For instance, in this code: |
24 | // |
25 | // sub r1, 1 |
26 | // cmp r1, 0 |
27 | // bz L1 |
28 | // |
29 | // If the "sub" instruction all ready sets (or could be modified to set) the |
30 | // same flag that the "cmp" instruction sets and that "bz" uses, then we can |
31 | // eliminate the "cmp" instruction. |
32 | // |
33 | // Another instance, in this code: |
34 | // |
35 | // sub r1, r3 | sub r1, imm |
36 | // cmp r3, r1 or cmp r1, r3 | cmp r1, imm |
37 | // bge L1 |
38 | // |
39 | // If the branch instruction can use flag from "sub", then we can replace |
40 | // "sub" with "subs" and eliminate the "cmp" instruction. |
41 | // |
42 | // - Optimize Loads: |
43 | // |
44 | // Loads that can be folded into a later instruction. A load is foldable |
45 | // if it loads to virtual registers and the virtual register defined has |
46 | // a single use. |
47 | // |
48 | // - Optimize Copies and Bitcast (more generally, target specific copies): |
49 | // |
50 | // Rewrite copies and bitcasts to avoid cross register bank copies |
51 | // when possible. |
52 | // E.g., Consider the following example, where capital and lower |
53 | // letters denote different register file: |
54 | // b = copy A <-- cross-bank copy |
55 | // C = copy b <-- cross-bank copy |
56 | // => |
57 | // b = copy A <-- cross-bank copy |
58 | // C = copy A <-- same-bank copy |
59 | // |
60 | // E.g., for bitcast: |
61 | // b = bitcast A <-- cross-bank copy |
62 | // C = bitcast b <-- cross-bank copy |
63 | // => |
64 | // b = bitcast A <-- cross-bank copy |
65 | // C = copy A <-- same-bank copy |
66 | //===----------------------------------------------------------------------===// |
67 | |
68 | #include "llvm/ADT/DenseMap.h" |
69 | #include "llvm/ADT/SmallPtrSet.h" |
70 | #include "llvm/ADT/SmallSet.h" |
71 | #include "llvm/ADT/SmallVector.h" |
72 | #include "llvm/ADT/Statistic.h" |
73 | #include "llvm/CodeGen/MachineBasicBlock.h" |
74 | #include "llvm/CodeGen/MachineDominators.h" |
75 | #include "llvm/CodeGen/MachineFunction.h" |
76 | #include "llvm/CodeGen/MachineFunctionPass.h" |
77 | #include "llvm/CodeGen/MachineInstr.h" |
78 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
79 | #include "llvm/CodeGen/MachineLoopInfo.h" |
80 | #include "llvm/CodeGen/MachineOperand.h" |
81 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
82 | #include "llvm/CodeGen/TargetInstrInfo.h" |
83 | #include "llvm/CodeGen/TargetOpcodes.h" |
84 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
85 | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
86 | #include "llvm/InitializePasses.h" |
87 | #include "llvm/MC/LaneBitmask.h" |
88 | #include "llvm/MC/MCInstrDesc.h" |
89 | #include "llvm/Pass.h" |
90 | #include "llvm/Support/CommandLine.h" |
91 | #include "llvm/Support/Debug.h" |
92 | #include "llvm/Support/raw_ostream.h" |
93 | #include <cassert> |
94 | #include <cstdint> |
95 | #include <memory> |
96 | #include <utility> |
97 | |
98 | using namespace llvm; |
99 | using RegSubRegPair = TargetInstrInfo::RegSubRegPair; |
100 | using RegSubRegPairAndIdx = TargetInstrInfo::RegSubRegPairAndIdx; |
101 | |
102 | #define DEBUG_TYPE "peephole-opt" |
103 | |
104 | // Optimize Extensions |
105 | static cl::opt<bool> |
106 | Aggressive("aggressive-ext-opt" , cl::Hidden, |
107 | cl::desc("Aggressive extension optimization" )); |
108 | |
109 | static cl::opt<bool> |
110 | DisablePeephole("disable-peephole" , cl::Hidden, cl::init(Val: false), |
111 | cl::desc("Disable the peephole optimizer" )); |
112 | |
113 | /// Specifiy whether or not the value tracking looks through |
114 | /// complex instructions. When this is true, the value tracker |
115 | /// bails on everything that is not a copy or a bitcast. |
116 | static cl::opt<bool> |
117 | DisableAdvCopyOpt("disable-adv-copy-opt" , cl::Hidden, cl::init(Val: false), |
118 | cl::desc("Disable advanced copy optimization" )); |
119 | |
120 | static cl::opt<bool> DisableNAPhysCopyOpt( |
121 | "disable-non-allocatable-phys-copy-opt" , cl::Hidden, cl::init(Val: false), |
122 | cl::desc("Disable non-allocatable physical register copy optimization" )); |
123 | |
124 | // Limit the number of PHI instructions to process |
125 | // in PeepholeOptimizer::getNextSource. |
126 | static cl::opt<unsigned> RewritePHILimit( |
127 | "rewrite-phi-limit" , cl::Hidden, cl::init(Val: 10), |
128 | cl::desc("Limit the length of PHI chains to lookup" )); |
129 | |
130 | // Limit the length of recurrence chain when evaluating the benefit of |
131 | // commuting operands. |
132 | static cl::opt<unsigned> MaxRecurrenceChain( |
133 | "recurrence-chain-limit" , cl::Hidden, cl::init(Val: 3), |
134 | cl::desc("Maximum length of recurrence chain when evaluating the benefit " |
135 | "of commuting operands" )); |
136 | |
137 | |
138 | STATISTIC(NumReuse, "Number of extension results reused" ); |
139 | STATISTIC(NumCmps, "Number of compares eliminated" ); |
140 | STATISTIC(NumImmFold, "Number of move immediate folded" ); |
141 | STATISTIC(NumLoadFold, "Number of loads folded" ); |
142 | STATISTIC(NumSelects, "Number of selects optimized" ); |
143 | STATISTIC(NumUncoalescableCopies, "Number of uncoalescable copies optimized" ); |
144 | STATISTIC(NumRewrittenCopies, "Number of copies rewritten" ); |
145 | STATISTIC(NumNAPhysCopies, "Number of non-allocatable physical copies removed" ); |
146 | |
147 | namespace { |
148 | |
149 | class ValueTrackerResult; |
150 | class RecurrenceInstr; |
151 | |
152 | class PeepholeOptimizer : public MachineFunctionPass, |
153 | private MachineFunction::Delegate { |
154 | const TargetInstrInfo *TII = nullptr; |
155 | const TargetRegisterInfo *TRI = nullptr; |
156 | MachineRegisterInfo *MRI = nullptr; |
157 | MachineDominatorTree *DT = nullptr; // Machine dominator tree |
158 | MachineLoopInfo *MLI = nullptr; |
159 | |
160 | public: |
161 | static char ID; // Pass identification |
162 | |
163 | PeepholeOptimizer() : MachineFunctionPass(ID) { |
164 | initializePeepholeOptimizerPass(*PassRegistry::getPassRegistry()); |
165 | } |
166 | |
167 | bool runOnMachineFunction(MachineFunction &MF) override; |
168 | |
169 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
170 | AU.setPreservesCFG(); |
171 | MachineFunctionPass::getAnalysisUsage(AU); |
172 | AU.addRequired<MachineLoopInfoWrapperPass>(); |
173 | AU.addPreserved<MachineLoopInfoWrapperPass>(); |
174 | if (Aggressive) { |
175 | AU.addRequired<MachineDominatorTreeWrapperPass>(); |
176 | AU.addPreserved<MachineDominatorTreeWrapperPass>(); |
177 | } |
178 | } |
179 | |
180 | MachineFunctionProperties getRequiredProperties() const override { |
181 | return MachineFunctionProperties() |
182 | .set(MachineFunctionProperties::Property::IsSSA); |
183 | } |
184 | |
185 | /// Track Def -> Use info used for rewriting copies. |
186 | using RewriteMapTy = SmallDenseMap<RegSubRegPair, ValueTrackerResult>; |
187 | |
188 | /// Sequence of instructions that formulate recurrence cycle. |
189 | using RecurrenceCycle = SmallVector<RecurrenceInstr, 4>; |
190 | |
191 | private: |
192 | bool optimizeCmpInstr(MachineInstr &MI); |
193 | bool optimizeExtInstr(MachineInstr &MI, MachineBasicBlock &MBB, |
194 | SmallPtrSetImpl<MachineInstr*> &LocalMIs); |
195 | bool optimizeSelect(MachineInstr &MI, |
196 | SmallPtrSetImpl<MachineInstr *> &LocalMIs); |
197 | bool optimizeCondBranch(MachineInstr &MI); |
198 | bool optimizeCoalescableCopy(MachineInstr &MI); |
199 | bool optimizeUncoalescableCopy(MachineInstr &MI, |
200 | SmallPtrSetImpl<MachineInstr *> &LocalMIs); |
201 | bool optimizeRecurrence(MachineInstr &PHI); |
202 | bool findNextSource(RegSubRegPair RegSubReg, RewriteMapTy &RewriteMap); |
203 | bool isMoveImmediate(MachineInstr &MI, SmallSet<Register, 4> &ImmDefRegs, |
204 | DenseMap<Register, MachineInstr *> &ImmDefMIs); |
205 | bool foldImmediate(MachineInstr &MI, SmallSet<Register, 4> &ImmDefRegs, |
206 | DenseMap<Register, MachineInstr *> &ImmDefMIs, |
207 | bool &Deleted); |
208 | |
209 | /// Finds recurrence cycles, but only ones that formulated around |
210 | /// a def operand and a use operand that are tied. If there is a use |
211 | /// operand commutable with the tied use operand, find recurrence cycle |
212 | /// along that operand as well. |
213 | bool findTargetRecurrence(Register Reg, |
214 | const SmallSet<Register, 2> &TargetReg, |
215 | RecurrenceCycle &RC); |
216 | |
217 | /// If copy instruction \p MI is a virtual register copy or a copy of a |
218 | /// constant physical register to a virtual register, track it in the |
219 | /// set CopySrcMIs. If this virtual register was previously seen as a |
220 | /// copy, replace the uses of this copy with the previously seen copy's |
221 | /// destination register. |
222 | bool foldRedundantCopy(MachineInstr &MI); |
223 | |
224 | /// Is the register \p Reg a non-allocatable physical register? |
225 | bool isNAPhysCopy(Register Reg); |
226 | |
227 | /// If copy instruction \p MI is a non-allocatable virtual<->physical |
228 | /// register copy, track it in the \p NAPhysToVirtMIs map. If this |
229 | /// non-allocatable physical register was previously copied to a virtual |
230 | /// registered and hasn't been clobbered, the virt->phys copy can be |
231 | /// deleted. |
232 | bool foldRedundantNAPhysCopy( |
233 | MachineInstr &MI, DenseMap<Register, MachineInstr *> &NAPhysToVirtMIs); |
234 | |
235 | bool isLoadFoldable(MachineInstr &MI, |
236 | SmallSet<Register, 16> &FoldAsLoadDefCandidates); |
237 | |
238 | /// Check whether \p MI is understood by the register coalescer |
239 | /// but may require some rewriting. |
240 | bool isCoalescableCopy(const MachineInstr &MI) { |
241 | // SubregToRegs are not interesting, because they are already register |
242 | // coalescer friendly. |
243 | return MI.isCopy() || (!DisableAdvCopyOpt && |
244 | (MI.isRegSequence() || MI.isInsertSubreg() || |
245 | MI.isExtractSubreg())); |
246 | } |
247 | |
248 | /// Check whether \p MI is a copy like instruction that is |
249 | /// not recognized by the register coalescer. |
250 | bool isUncoalescableCopy(const MachineInstr &MI) { |
251 | return MI.isBitcast() || |
252 | (!DisableAdvCopyOpt && |
253 | (MI.isRegSequenceLike() || MI.isInsertSubregLike() || |
254 | MI.isExtractSubregLike())); |
255 | } |
256 | |
257 | MachineInstr &rewriteSource(MachineInstr &CopyLike, |
258 | RegSubRegPair Def, RewriteMapTy &RewriteMap); |
259 | |
260 | // Set of copies to virtual registers keyed by source register. Never |
261 | // holds any physreg which requires def tracking. |
262 | DenseMap<RegSubRegPair, MachineInstr *> CopySrcMIs; |
263 | |
264 | // MachineFunction::Delegate implementation. Used to maintain CopySrcMIs. |
265 | void MF_HandleInsertion(MachineInstr &MI) override { |
266 | return; |
267 | } |
268 | |
269 | bool getCopySrc(MachineInstr &MI, RegSubRegPair &SrcPair) { |
270 | if (!MI.isCopy()) |
271 | return false; |
272 | |
273 | Register SrcReg = MI.getOperand(i: 1).getReg(); |
274 | unsigned SrcSubReg = MI.getOperand(i: 1).getSubReg(); |
275 | if (!SrcReg.isVirtual() && !MRI->isConstantPhysReg(PhysReg: SrcReg)) |
276 | return false; |
277 | |
278 | SrcPair = RegSubRegPair(SrcReg, SrcSubReg); |
279 | return true; |
280 | } |
281 | |
282 | // If a COPY instruction is to be deleted or changed, we should also remove |
283 | // it from CopySrcMIs. |
284 | void deleteChangedCopy(MachineInstr &MI) { |
285 | RegSubRegPair SrcPair; |
286 | if (!getCopySrc(MI, SrcPair)) |
287 | return; |
288 | |
289 | auto It = CopySrcMIs.find(Val: SrcPair); |
290 | if (It != CopySrcMIs.end() && It->second == &MI) |
291 | CopySrcMIs.erase(I: It); |
292 | } |
293 | |
294 | void MF_HandleRemoval(MachineInstr &MI) override { |
295 | deleteChangedCopy(MI); |
296 | } |
297 | |
298 | void MF_HandleChangeDesc(MachineInstr &MI, const MCInstrDesc &TID) override |
299 | { |
300 | deleteChangedCopy(MI); |
301 | } |
302 | }; |
303 | |
304 | /// Helper class to hold instructions that are inside recurrence cycles. |
305 | /// The recurrence cycle is formulated around 1) a def operand and its |
306 | /// tied use operand, or 2) a def operand and a use operand that is commutable |
307 | /// with another use operand which is tied to the def operand. In the latter |
308 | /// case, index of the tied use operand and the commutable use operand are |
309 | /// maintained with CommutePair. |
310 | class RecurrenceInstr { |
311 | public: |
312 | using IndexPair = std::pair<unsigned, unsigned>; |
313 | |
314 | RecurrenceInstr(MachineInstr *MI) : MI(MI) {} |
315 | RecurrenceInstr(MachineInstr *MI, unsigned Idx1, unsigned Idx2) |
316 | : MI(MI), CommutePair(std::make_pair(x&: Idx1, y&: Idx2)) {} |
317 | |
318 | MachineInstr *getMI() const { return MI; } |
319 | std::optional<IndexPair> getCommutePair() const { return CommutePair; } |
320 | |
321 | private: |
322 | MachineInstr *MI; |
323 | std::optional<IndexPair> CommutePair; |
324 | }; |
325 | |
326 | /// Helper class to hold a reply for ValueTracker queries. |
327 | /// Contains the returned sources for a given search and the instructions |
328 | /// where the sources were tracked from. |
329 | class ValueTrackerResult { |
330 | private: |
331 | /// Track all sources found by one ValueTracker query. |
332 | SmallVector<RegSubRegPair, 2> RegSrcs; |
333 | |
334 | /// Instruction using the sources in 'RegSrcs'. |
335 | const MachineInstr *Inst = nullptr; |
336 | |
337 | public: |
338 | ValueTrackerResult() = default; |
339 | |
340 | ValueTrackerResult(Register Reg, unsigned SubReg) { |
341 | addSource(SrcReg: Reg, SrcSubReg: SubReg); |
342 | } |
343 | |
344 | bool isValid() const { return getNumSources() > 0; } |
345 | |
346 | void setInst(const MachineInstr *I) { Inst = I; } |
347 | const MachineInstr *getInst() const { return Inst; } |
348 | |
349 | void clear() { |
350 | RegSrcs.clear(); |
351 | Inst = nullptr; |
352 | } |
353 | |
354 | void addSource(Register SrcReg, unsigned SrcSubReg) { |
355 | RegSrcs.push_back(Elt: RegSubRegPair(SrcReg, SrcSubReg)); |
356 | } |
357 | |
358 | void setSource(int Idx, Register SrcReg, unsigned SrcSubReg) { |
359 | assert(Idx < getNumSources() && "Reg pair source out of index" ); |
360 | RegSrcs[Idx] = RegSubRegPair(SrcReg, SrcSubReg); |
361 | } |
362 | |
363 | int getNumSources() const { return RegSrcs.size(); } |
364 | |
365 | RegSubRegPair getSrc(int Idx) const { |
366 | return RegSrcs[Idx]; |
367 | } |
368 | |
369 | Register getSrcReg(int Idx) const { |
370 | assert(Idx < getNumSources() && "Reg source out of index" ); |
371 | return RegSrcs[Idx].Reg; |
372 | } |
373 | |
374 | unsigned getSrcSubReg(int Idx) const { |
375 | assert(Idx < getNumSources() && "SubReg source out of index" ); |
376 | return RegSrcs[Idx].SubReg; |
377 | } |
378 | |
379 | bool operator==(const ValueTrackerResult &Other) const { |
380 | if (Other.getInst() != getInst()) |
381 | return false; |
382 | |
383 | if (Other.getNumSources() != getNumSources()) |
384 | return false; |
385 | |
386 | for (int i = 0, e = Other.getNumSources(); i != e; ++i) |
387 | if (Other.getSrcReg(Idx: i) != getSrcReg(Idx: i) || |
388 | Other.getSrcSubReg(Idx: i) != getSrcSubReg(Idx: i)) |
389 | return false; |
390 | return true; |
391 | } |
392 | }; |
393 | |
394 | /// Helper class to track the possible sources of a value defined by |
395 | /// a (chain of) copy related instructions. |
396 | /// Given a definition (instruction and definition index), this class |
397 | /// follows the use-def chain to find successive suitable sources. |
398 | /// The given source can be used to rewrite the definition into |
399 | /// def = COPY src. |
400 | /// |
401 | /// For instance, let us consider the following snippet: |
402 | /// v0 = |
403 | /// v2 = INSERT_SUBREG v1, v0, sub0 |
404 | /// def = COPY v2.sub0 |
405 | /// |
406 | /// Using a ValueTracker for def = COPY v2.sub0 will give the following |
407 | /// suitable sources: |
408 | /// v2.sub0 and v0. |
409 | /// Then, def can be rewritten into def = COPY v0. |
410 | class ValueTracker { |
411 | private: |
412 | /// The current point into the use-def chain. |
413 | const MachineInstr *Def = nullptr; |
414 | |
415 | /// The index of the definition in Def. |
416 | unsigned DefIdx = 0; |
417 | |
418 | /// The sub register index of the definition. |
419 | unsigned DefSubReg; |
420 | |
421 | /// The register where the value can be found. |
422 | Register Reg; |
423 | |
424 | /// MachineRegisterInfo used to perform tracking. |
425 | const MachineRegisterInfo &MRI; |
426 | |
427 | /// Optional TargetInstrInfo used to perform some complex tracking. |
428 | const TargetInstrInfo *TII; |
429 | |
430 | /// Dispatcher to the right underlying implementation of getNextSource. |
431 | ValueTrackerResult getNextSourceImpl(); |
432 | |
433 | /// Specialized version of getNextSource for Copy instructions. |
434 | ValueTrackerResult getNextSourceFromCopy(); |
435 | |
436 | /// Specialized version of getNextSource for Bitcast instructions. |
437 | ValueTrackerResult getNextSourceFromBitcast(); |
438 | |
439 | /// Specialized version of getNextSource for RegSequence instructions. |
440 | ValueTrackerResult getNextSourceFromRegSequence(); |
441 | |
442 | /// Specialized version of getNextSource for InsertSubreg instructions. |
443 | ValueTrackerResult getNextSourceFromInsertSubreg(); |
444 | |
445 | /// Specialized version of getNextSource for ExtractSubreg instructions. |
446 | ValueTrackerResult getNextSourceFromExtractSubreg(); |
447 | |
448 | /// Specialized version of getNextSource for SubregToReg instructions. |
449 | ValueTrackerResult getNextSourceFromSubregToReg(); |
450 | |
451 | /// Specialized version of getNextSource for PHI instructions. |
452 | ValueTrackerResult getNextSourceFromPHI(); |
453 | |
454 | public: |
455 | /// Create a ValueTracker instance for the value defined by \p Reg. |
456 | /// \p DefSubReg represents the sub register index the value tracker will |
457 | /// track. It does not need to match the sub register index used in the |
458 | /// definition of \p Reg. |
459 | /// If \p Reg is a physical register, a value tracker constructed with |
460 | /// this constructor will not find any alternative source. |
461 | /// Indeed, when \p Reg is a physical register that constructor does not |
462 | /// know which definition of \p Reg it should track. |
463 | /// Use the next constructor to track a physical register. |
464 | ValueTracker(Register Reg, unsigned DefSubReg, |
465 | const MachineRegisterInfo &MRI, |
466 | const TargetInstrInfo *TII = nullptr) |
467 | : DefSubReg(DefSubReg), Reg(Reg), MRI(MRI), TII(TII) { |
468 | if (!Reg.isPhysical()) { |
469 | Def = MRI.getVRegDef(Reg); |
470 | DefIdx = MRI.def_begin(RegNo: Reg).getOperandNo(); |
471 | } |
472 | } |
473 | |
474 | /// Following the use-def chain, get the next available source |
475 | /// for the tracked value. |
476 | /// \return A ValueTrackerResult containing a set of registers |
477 | /// and sub registers with tracked values. A ValueTrackerResult with |
478 | /// an empty set of registers means no source was found. |
479 | ValueTrackerResult getNextSource(); |
480 | }; |
481 | |
482 | } // end anonymous namespace |
483 | |
484 | char PeepholeOptimizer::ID = 0; |
485 | |
486 | char &llvm::PeepholeOptimizerID = PeepholeOptimizer::ID; |
487 | |
488 | INITIALIZE_PASS_BEGIN(PeepholeOptimizer, DEBUG_TYPE, |
489 | "Peephole Optimizations" , false, false) |
490 | INITIALIZE_PASS_DEPENDENCY(MachineDominatorTreeWrapperPass) |
491 | INITIALIZE_PASS_DEPENDENCY(MachineLoopInfoWrapperPass) |
492 | INITIALIZE_PASS_END(PeepholeOptimizer, DEBUG_TYPE, |
493 | "Peephole Optimizations" , false, false) |
494 | |
495 | /// If instruction is a copy-like instruction, i.e. it reads a single register |
496 | /// and writes a single register and it does not modify the source, and if the |
497 | /// source value is preserved as a sub-register of the result, then replace all |
498 | /// reachable uses of the source with the subreg of the result. |
499 | /// |
500 | /// Do not generate an EXTRACT that is used only in a debug use, as this changes |
501 | /// the code. Since this code does not currently share EXTRACTs, just ignore all |
502 | /// debug uses. |
503 | bool PeepholeOptimizer:: |
504 | optimizeExtInstr(MachineInstr &MI, MachineBasicBlock &MBB, |
505 | SmallPtrSetImpl<MachineInstr*> &LocalMIs) { |
506 | Register SrcReg, DstReg; |
507 | unsigned SubIdx; |
508 | if (!TII->isCoalescableExtInstr(MI, SrcReg, DstReg, SubIdx)) |
509 | return false; |
510 | |
511 | if (DstReg.isPhysical() || SrcReg.isPhysical()) |
512 | return false; |
513 | |
514 | if (MRI->hasOneNonDBGUse(RegNo: SrcReg)) |
515 | // No other uses. |
516 | return false; |
517 | |
518 | // Ensure DstReg can get a register class that actually supports |
519 | // sub-registers. Don't change the class until we commit. |
520 | const TargetRegisterClass *DstRC = MRI->getRegClass(Reg: DstReg); |
521 | DstRC = TRI->getSubClassWithSubReg(RC: DstRC, Idx: SubIdx); |
522 | if (!DstRC) |
523 | return false; |
524 | |
525 | // The ext instr may be operating on a sub-register of SrcReg as well. |
526 | // PPC::EXTSW is a 32 -> 64-bit sign extension, but it reads a 64-bit |
527 | // register. |
528 | // If UseSrcSubIdx is Set, SubIdx also applies to SrcReg, and only uses of |
529 | // SrcReg:SubIdx should be replaced. |
530 | bool UseSrcSubIdx = |
531 | TRI->getSubClassWithSubReg(RC: MRI->getRegClass(Reg: SrcReg), Idx: SubIdx) != nullptr; |
532 | |
533 | // The source has other uses. See if we can replace the other uses with use of |
534 | // the result of the extension. |
535 | SmallPtrSet<MachineBasicBlock*, 4> ReachedBBs; |
536 | for (MachineInstr &UI : MRI->use_nodbg_instructions(Reg: DstReg)) |
537 | ReachedBBs.insert(Ptr: UI.getParent()); |
538 | |
539 | // Uses that are in the same BB of uses of the result of the instruction. |
540 | SmallVector<MachineOperand*, 8> Uses; |
541 | |
542 | // Uses that the result of the instruction can reach. |
543 | SmallVector<MachineOperand*, 8> ExtendedUses; |
544 | |
545 | bool ExtendLife = true; |
546 | for (MachineOperand &UseMO : MRI->use_nodbg_operands(Reg: SrcReg)) { |
547 | MachineInstr *UseMI = UseMO.getParent(); |
548 | if (UseMI == &MI) |
549 | continue; |
550 | |
551 | if (UseMI->isPHI()) { |
552 | ExtendLife = false; |
553 | continue; |
554 | } |
555 | |
556 | // Only accept uses of SrcReg:SubIdx. |
557 | if (UseSrcSubIdx && UseMO.getSubReg() != SubIdx) |
558 | continue; |
559 | |
560 | // It's an error to translate this: |
561 | // |
562 | // %reg1025 = <sext> %reg1024 |
563 | // ... |
564 | // %reg1026 = SUBREG_TO_REG 0, %reg1024, 4 |
565 | // |
566 | // into this: |
567 | // |
568 | // %reg1025 = <sext> %reg1024 |
569 | // ... |
570 | // %reg1027 = COPY %reg1025:4 |
571 | // %reg1026 = SUBREG_TO_REG 0, %reg1027, 4 |
572 | // |
573 | // The problem here is that SUBREG_TO_REG is there to assert that an |
574 | // implicit zext occurs. It doesn't insert a zext instruction. If we allow |
575 | // the COPY here, it will give us the value after the <sext>, not the |
576 | // original value of %reg1024 before <sext>. |
577 | if (UseMI->getOpcode() == TargetOpcode::SUBREG_TO_REG) |
578 | continue; |
579 | |
580 | MachineBasicBlock *UseMBB = UseMI->getParent(); |
581 | if (UseMBB == &MBB) { |
582 | // Local uses that come after the extension. |
583 | if (!LocalMIs.count(Ptr: UseMI)) |
584 | Uses.push_back(Elt: &UseMO); |
585 | } else if (ReachedBBs.count(Ptr: UseMBB)) { |
586 | // Non-local uses where the result of the extension is used. Always |
587 | // replace these unless it's a PHI. |
588 | Uses.push_back(Elt: &UseMO); |
589 | } else if (Aggressive && DT->dominates(A: &MBB, B: UseMBB)) { |
590 | // We may want to extend the live range of the extension result in order |
591 | // to replace these uses. |
592 | ExtendedUses.push_back(Elt: &UseMO); |
593 | } else { |
594 | // Both will be live out of the def MBB anyway. Don't extend live range of |
595 | // the extension result. |
596 | ExtendLife = false; |
597 | break; |
598 | } |
599 | } |
600 | |
601 | if (ExtendLife && !ExtendedUses.empty()) |
602 | // Extend the liveness of the extension result. |
603 | Uses.append(in_start: ExtendedUses.begin(), in_end: ExtendedUses.end()); |
604 | |
605 | // Now replace all uses. |
606 | bool Changed = false; |
607 | if (!Uses.empty()) { |
608 | SmallPtrSet<MachineBasicBlock*, 4> PHIBBs; |
609 | |
610 | // Look for PHI uses of the extended result, we don't want to extend the |
611 | // liveness of a PHI input. It breaks all kinds of assumptions down |
612 | // stream. A PHI use is expected to be the kill of its source values. |
613 | for (MachineInstr &UI : MRI->use_nodbg_instructions(Reg: DstReg)) |
614 | if (UI.isPHI()) |
615 | PHIBBs.insert(Ptr: UI.getParent()); |
616 | |
617 | const TargetRegisterClass *RC = MRI->getRegClass(Reg: SrcReg); |
618 | for (MachineOperand *UseMO : Uses) { |
619 | MachineInstr *UseMI = UseMO->getParent(); |
620 | MachineBasicBlock *UseMBB = UseMI->getParent(); |
621 | if (PHIBBs.count(Ptr: UseMBB)) |
622 | continue; |
623 | |
624 | // About to add uses of DstReg, clear DstReg's kill flags. |
625 | if (!Changed) { |
626 | MRI->clearKillFlags(Reg: DstReg); |
627 | MRI->constrainRegClass(Reg: DstReg, RC: DstRC); |
628 | } |
629 | |
630 | // SubReg defs are illegal in machine SSA phase, |
631 | // we should not generate SubReg defs. |
632 | // |
633 | // For example, for the instructions: |
634 | // |
635 | // %1:g8rc_and_g8rc_nox0 = EXTSW %0:g8rc |
636 | // %3:gprc_and_gprc_nor0 = COPY %0.sub_32:g8rc |
637 | // |
638 | // We should generate: |
639 | // |
640 | // %1:g8rc_and_g8rc_nox0 = EXTSW %0:g8rc |
641 | // %6:gprc_and_gprc_nor0 = COPY %1.sub_32:g8rc_and_g8rc_nox0 |
642 | // %3:gprc_and_gprc_nor0 = COPY %6:gprc_and_gprc_nor0 |
643 | // |
644 | if (UseSrcSubIdx) |
645 | RC = MRI->getRegClass(Reg: UseMI->getOperand(i: 0).getReg()); |
646 | |
647 | Register NewVR = MRI->createVirtualRegister(RegClass: RC); |
648 | BuildMI(BB&: *UseMBB, I: UseMI, MIMD: UseMI->getDebugLoc(), |
649 | MCID: TII->get(Opcode: TargetOpcode::COPY), DestReg: NewVR) |
650 | .addReg(RegNo: DstReg, flags: 0, SubReg: SubIdx); |
651 | if (UseSrcSubIdx) |
652 | UseMO->setSubReg(0); |
653 | |
654 | UseMO->setReg(NewVR); |
655 | ++NumReuse; |
656 | Changed = true; |
657 | } |
658 | } |
659 | |
660 | return Changed; |
661 | } |
662 | |
663 | /// If the instruction is a compare and the previous instruction it's comparing |
664 | /// against already sets (or could be modified to set) the same flag as the |
665 | /// compare, then we can remove the comparison and use the flag from the |
666 | /// previous instruction. |
667 | bool PeepholeOptimizer::optimizeCmpInstr(MachineInstr &MI) { |
668 | // If this instruction is a comparison against zero and isn't comparing a |
669 | // physical register, we can try to optimize it. |
670 | Register SrcReg, SrcReg2; |
671 | int64_t CmpMask, CmpValue; |
672 | if (!TII->analyzeCompare(MI, SrcReg, SrcReg2, Mask&: CmpMask, Value&: CmpValue) || |
673 | SrcReg.isPhysical() || SrcReg2.isPhysical()) |
674 | return false; |
675 | |
676 | // Attempt to optimize the comparison instruction. |
677 | LLVM_DEBUG(dbgs() << "Attempting to optimize compare: " << MI); |
678 | if (TII->optimizeCompareInstr(CmpInstr&: MI, SrcReg, SrcReg2, Mask: CmpMask, Value: CmpValue, MRI)) { |
679 | LLVM_DEBUG(dbgs() << " -> Successfully optimized compare!\n" ); |
680 | ++NumCmps; |
681 | return true; |
682 | } |
683 | |
684 | return false; |
685 | } |
686 | |
687 | /// Optimize a select instruction. |
688 | bool PeepholeOptimizer::optimizeSelect(MachineInstr &MI, |
689 | SmallPtrSetImpl<MachineInstr *> &LocalMIs) { |
690 | unsigned TrueOp = 0; |
691 | unsigned FalseOp = 0; |
692 | bool Optimizable = false; |
693 | SmallVector<MachineOperand, 4> Cond; |
694 | if (TII->analyzeSelect(MI, Cond, TrueOp, FalseOp, Optimizable)) |
695 | return false; |
696 | if (!Optimizable) |
697 | return false; |
698 | if (!TII->optimizeSelect(MI, NewMIs&: LocalMIs)) |
699 | return false; |
700 | LLVM_DEBUG(dbgs() << "Deleting select: " << MI); |
701 | MI.eraseFromParent(); |
702 | ++NumSelects; |
703 | return true; |
704 | } |
705 | |
706 | /// Check if a simpler conditional branch can be generated. |
707 | bool PeepholeOptimizer::optimizeCondBranch(MachineInstr &MI) { |
708 | return TII->optimizeCondBranch(MI); |
709 | } |
710 | |
711 | /// Try to find the next source that share the same register file |
712 | /// for the value defined by \p Reg and \p SubReg. |
713 | /// When true is returned, the \p RewriteMap can be used by the client to |
714 | /// retrieve all Def -> Use along the way up to the next source. Any found |
715 | /// Use that is not itself a key for another entry, is the next source to |
716 | /// use. During the search for the next source, multiple sources can be found |
717 | /// given multiple incoming sources of a PHI instruction. In this case, we |
718 | /// look in each PHI source for the next source; all found next sources must |
719 | /// share the same register file as \p Reg and \p SubReg. The client should |
720 | /// then be capable to rewrite all intermediate PHIs to get the next source. |
721 | /// \return False if no alternative sources are available. True otherwise. |
722 | bool PeepholeOptimizer::findNextSource(RegSubRegPair RegSubReg, |
723 | RewriteMapTy &RewriteMap) { |
724 | // Do not try to find a new source for a physical register. |
725 | // So far we do not have any motivating example for doing that. |
726 | // Thus, instead of maintaining untested code, we will revisit that if |
727 | // that changes at some point. |
728 | Register Reg = RegSubReg.Reg; |
729 | if (Reg.isPhysical()) |
730 | return false; |
731 | const TargetRegisterClass *DefRC = MRI->getRegClass(Reg); |
732 | |
733 | SmallVector<RegSubRegPair, 4> SrcToLook; |
734 | RegSubRegPair CurSrcPair = RegSubReg; |
735 | SrcToLook.push_back(Elt: CurSrcPair); |
736 | |
737 | unsigned PHICount = 0; |
738 | do { |
739 | CurSrcPair = SrcToLook.pop_back_val(); |
740 | // As explained above, do not handle physical registers |
741 | if (CurSrcPair.Reg.isPhysical()) |
742 | return false; |
743 | |
744 | ValueTracker ValTracker(CurSrcPair.Reg, CurSrcPair.SubReg, *MRI, TII); |
745 | |
746 | // Follow the chain of copies until we find a more suitable source, a phi |
747 | // or have to abort. |
748 | while (true) { |
749 | ValueTrackerResult Res = ValTracker.getNextSource(); |
750 | // Abort at the end of a chain (without finding a suitable source). |
751 | if (!Res.isValid()) |
752 | return false; |
753 | |
754 | // Insert the Def -> Use entry for the recently found source. |
755 | ValueTrackerResult CurSrcRes = RewriteMap.lookup(Val: CurSrcPair); |
756 | if (CurSrcRes.isValid()) { |
757 | assert(CurSrcRes == Res && "ValueTrackerResult found must match" ); |
758 | // An existent entry with multiple sources is a PHI cycle we must avoid. |
759 | // Otherwise it's an entry with a valid next source we already found. |
760 | if (CurSrcRes.getNumSources() > 1) { |
761 | LLVM_DEBUG(dbgs() |
762 | << "findNextSource: found PHI cycle, aborting...\n" ); |
763 | return false; |
764 | } |
765 | break; |
766 | } |
767 | RewriteMap.insert(KV: std::make_pair(x&: CurSrcPair, y&: Res)); |
768 | |
769 | // ValueTrackerResult usually have one source unless it's the result from |
770 | // a PHI instruction. Add the found PHI edges to be looked up further. |
771 | unsigned NumSrcs = Res.getNumSources(); |
772 | if (NumSrcs > 1) { |
773 | PHICount++; |
774 | if (PHICount >= RewritePHILimit) { |
775 | LLVM_DEBUG(dbgs() << "findNextSource: PHI limit reached\n" ); |
776 | return false; |
777 | } |
778 | |
779 | for (unsigned i = 0; i < NumSrcs; ++i) |
780 | SrcToLook.push_back(Elt: Res.getSrc(Idx: i)); |
781 | break; |
782 | } |
783 | |
784 | CurSrcPair = Res.getSrc(Idx: 0); |
785 | // Do not extend the live-ranges of physical registers as they add |
786 | // constraints to the register allocator. Moreover, if we want to extend |
787 | // the live-range of a physical register, unlike SSA virtual register, |
788 | // we will have to check that they aren't redefine before the related use. |
789 | if (CurSrcPair.Reg.isPhysical()) |
790 | return false; |
791 | |
792 | // Keep following the chain if the value isn't any better yet. |
793 | const TargetRegisterClass *SrcRC = MRI->getRegClass(Reg: CurSrcPair.Reg); |
794 | if (!TRI->shouldRewriteCopySrc(DefRC, DefSubReg: RegSubReg.SubReg, SrcRC, |
795 | SrcSubReg: CurSrcPair.SubReg)) |
796 | continue; |
797 | |
798 | // We currently cannot deal with subreg operands on PHI instructions |
799 | // (see insertPHI()). |
800 | if (PHICount > 0 && CurSrcPair.SubReg != 0) |
801 | continue; |
802 | |
803 | // We found a suitable source, and are done with this chain. |
804 | break; |
805 | } |
806 | } while (!SrcToLook.empty()); |
807 | |
808 | // If we did not find a more suitable source, there is nothing to optimize. |
809 | return CurSrcPair.Reg != Reg; |
810 | } |
811 | |
812 | /// Insert a PHI instruction with incoming edges \p SrcRegs that are |
813 | /// guaranteed to have the same register class. This is necessary whenever we |
814 | /// successfully traverse a PHI instruction and find suitable sources coming |
815 | /// from its edges. By inserting a new PHI, we provide a rewritten PHI def |
816 | /// suitable to be used in a new COPY instruction. |
817 | static MachineInstr & |
818 | insertPHI(MachineRegisterInfo &MRI, const TargetInstrInfo &TII, |
819 | const SmallVectorImpl<RegSubRegPair> &SrcRegs, |
820 | MachineInstr &OrigPHI) { |
821 | assert(!SrcRegs.empty() && "No sources to create a PHI instruction?" ); |
822 | |
823 | const TargetRegisterClass *NewRC = MRI.getRegClass(Reg: SrcRegs[0].Reg); |
824 | // NewRC is only correct if no subregisters are involved. findNextSource() |
825 | // should have rejected those cases already. |
826 | assert(SrcRegs[0].SubReg == 0 && "should not have subreg operand" ); |
827 | Register NewVR = MRI.createVirtualRegister(RegClass: NewRC); |
828 | MachineBasicBlock *MBB = OrigPHI.getParent(); |
829 | MachineInstrBuilder MIB = BuildMI(BB&: *MBB, I: &OrigPHI, MIMD: OrigPHI.getDebugLoc(), |
830 | MCID: TII.get(Opcode: TargetOpcode::PHI), DestReg: NewVR); |
831 | |
832 | unsigned MBBOpIdx = 2; |
833 | for (const RegSubRegPair &RegPair : SrcRegs) { |
834 | MIB.addReg(RegNo: RegPair.Reg, flags: 0, SubReg: RegPair.SubReg); |
835 | MIB.addMBB(MBB: OrigPHI.getOperand(i: MBBOpIdx).getMBB()); |
836 | // Since we're extended the lifetime of RegPair.Reg, clear the |
837 | // kill flags to account for that and make RegPair.Reg reaches |
838 | // the new PHI. |
839 | MRI.clearKillFlags(Reg: RegPair.Reg); |
840 | MBBOpIdx += 2; |
841 | } |
842 | |
843 | return *MIB; |
844 | } |
845 | |
846 | namespace { |
847 | |
848 | /// Interface to query instructions amenable to copy rewriting. |
849 | class Rewriter { |
850 | protected: |
851 | MachineInstr &CopyLike; |
852 | unsigned CurrentSrcIdx = 0; ///< The index of the source being rewritten. |
853 | public: |
854 | Rewriter(MachineInstr &CopyLike) : CopyLike(CopyLike) {} |
855 | virtual ~Rewriter() = default; |
856 | |
857 | /// Get the next rewritable source (SrcReg, SrcSubReg) and |
858 | /// the related value that it affects (DstReg, DstSubReg). |
859 | /// A source is considered rewritable if its register class and the |
860 | /// register class of the related DstReg may not be register |
861 | /// coalescer friendly. In other words, given a copy-like instruction |
862 | /// not all the arguments may be returned at rewritable source, since |
863 | /// some arguments are none to be register coalescer friendly. |
864 | /// |
865 | /// Each call of this method moves the current source to the next |
866 | /// rewritable source. |
867 | /// For instance, let CopyLike be the instruction to rewrite. |
868 | /// CopyLike has one definition and one source: |
869 | /// dst.dstSubIdx = CopyLike src.srcSubIdx. |
870 | /// |
871 | /// The first call will give the first rewritable source, i.e., |
872 | /// the only source this instruction has: |
873 | /// (SrcReg, SrcSubReg) = (src, srcSubIdx). |
874 | /// This source defines the whole definition, i.e., |
875 | /// (DstReg, DstSubReg) = (dst, dstSubIdx). |
876 | /// |
877 | /// The second and subsequent calls will return false, as there is only one |
878 | /// rewritable source. |
879 | /// |
880 | /// \return True if a rewritable source has been found, false otherwise. |
881 | /// The output arguments are valid if and only if true is returned. |
882 | virtual bool getNextRewritableSource(RegSubRegPair &Src, |
883 | RegSubRegPair &Dst) = 0; |
884 | |
885 | /// Rewrite the current source with \p NewReg and \p NewSubReg if possible. |
886 | /// \return True if the rewriting was possible, false otherwise. |
887 | virtual bool RewriteCurrentSource(Register NewReg, unsigned NewSubReg) = 0; |
888 | }; |
889 | |
890 | /// Rewriter for COPY instructions. |
891 | class CopyRewriter : public Rewriter { |
892 | public: |
893 | CopyRewriter(MachineInstr &MI) : Rewriter(MI) { |
894 | assert(MI.isCopy() && "Expected copy instruction" ); |
895 | } |
896 | virtual ~CopyRewriter() = default; |
897 | |
898 | bool getNextRewritableSource(RegSubRegPair &Src, |
899 | RegSubRegPair &Dst) override { |
900 | // CurrentSrcIdx > 0 means this function has already been called. |
901 | if (CurrentSrcIdx > 0) |
902 | return false; |
903 | // This is the first call to getNextRewritableSource. |
904 | // Move the CurrentSrcIdx to remember that we made that call. |
905 | CurrentSrcIdx = 1; |
906 | // The rewritable source is the argument. |
907 | const MachineOperand &MOSrc = CopyLike.getOperand(i: 1); |
908 | Src = RegSubRegPair(MOSrc.getReg(), MOSrc.getSubReg()); |
909 | // What we track are the alternative sources of the definition. |
910 | const MachineOperand &MODef = CopyLike.getOperand(i: 0); |
911 | Dst = RegSubRegPair(MODef.getReg(), MODef.getSubReg()); |
912 | return true; |
913 | } |
914 | |
915 | bool RewriteCurrentSource(Register NewReg, unsigned NewSubReg) override { |
916 | if (CurrentSrcIdx != 1) |
917 | return false; |
918 | MachineOperand &MOSrc = CopyLike.getOperand(i: CurrentSrcIdx); |
919 | MOSrc.setReg(NewReg); |
920 | MOSrc.setSubReg(NewSubReg); |
921 | return true; |
922 | } |
923 | }; |
924 | |
925 | /// Helper class to rewrite uncoalescable copy like instructions |
926 | /// into new COPY (coalescable friendly) instructions. |
927 | class UncoalescableRewriter : public Rewriter { |
928 | unsigned NumDefs; ///< Number of defs in the bitcast. |
929 | |
930 | public: |
931 | UncoalescableRewriter(MachineInstr &MI) : Rewriter(MI) { |
932 | NumDefs = MI.getDesc().getNumDefs(); |
933 | } |
934 | |
935 | /// \see See Rewriter::getNextRewritableSource() |
936 | /// All such sources need to be considered rewritable in order to |
937 | /// rewrite a uncoalescable copy-like instruction. This method return |
938 | /// each definition that must be checked if rewritable. |
939 | bool getNextRewritableSource(RegSubRegPair &Src, |
940 | RegSubRegPair &Dst) override { |
941 | // Find the next non-dead definition and continue from there. |
942 | if (CurrentSrcIdx == NumDefs) |
943 | return false; |
944 | |
945 | while (CopyLike.getOperand(i: CurrentSrcIdx).isDead()) { |
946 | ++CurrentSrcIdx; |
947 | if (CurrentSrcIdx == NumDefs) |
948 | return false; |
949 | } |
950 | |
951 | // What we track are the alternative sources of the definition. |
952 | Src = RegSubRegPair(0, 0); |
953 | const MachineOperand &MODef = CopyLike.getOperand(i: CurrentSrcIdx); |
954 | Dst = RegSubRegPair(MODef.getReg(), MODef.getSubReg()); |
955 | |
956 | CurrentSrcIdx++; |
957 | return true; |
958 | } |
959 | |
960 | bool RewriteCurrentSource(Register NewReg, unsigned NewSubReg) override { |
961 | return false; |
962 | } |
963 | }; |
964 | |
965 | /// Specialized rewriter for INSERT_SUBREG instruction. |
966 | class InsertSubregRewriter : public Rewriter { |
967 | public: |
968 | InsertSubregRewriter(MachineInstr &MI) : Rewriter(MI) { |
969 | assert(MI.isInsertSubreg() && "Invalid instruction" ); |
970 | } |
971 | |
972 | /// \see See Rewriter::getNextRewritableSource() |
973 | /// Here CopyLike has the following form: |
974 | /// dst = INSERT_SUBREG Src1, Src2.src2SubIdx, subIdx. |
975 | /// Src1 has the same register class has dst, hence, there is |
976 | /// nothing to rewrite. |
977 | /// Src2.src2SubIdx, may not be register coalescer friendly. |
978 | /// Therefore, the first call to this method returns: |
979 | /// (SrcReg, SrcSubReg) = (Src2, src2SubIdx). |
980 | /// (DstReg, DstSubReg) = (dst, subIdx). |
981 | /// |
982 | /// Subsequence calls will return false. |
983 | bool getNextRewritableSource(RegSubRegPair &Src, |
984 | RegSubRegPair &Dst) override { |
985 | // If we already get the only source we can rewrite, return false. |
986 | if (CurrentSrcIdx == 2) |
987 | return false; |
988 | // We are looking at v2 = INSERT_SUBREG v0, v1, sub0. |
989 | CurrentSrcIdx = 2; |
990 | const MachineOperand &MOInsertedReg = CopyLike.getOperand(i: 2); |
991 | Src = RegSubRegPair(MOInsertedReg.getReg(), MOInsertedReg.getSubReg()); |
992 | const MachineOperand &MODef = CopyLike.getOperand(i: 0); |
993 | |
994 | // We want to track something that is compatible with the |
995 | // partial definition. |
996 | if (MODef.getSubReg()) |
997 | // Bail if we have to compose sub-register indices. |
998 | return false; |
999 | Dst = RegSubRegPair(MODef.getReg(), |
1000 | (unsigned)CopyLike.getOperand(i: 3).getImm()); |
1001 | return true; |
1002 | } |
1003 | |
1004 | bool RewriteCurrentSource(Register NewReg, unsigned NewSubReg) override { |
1005 | if (CurrentSrcIdx != 2) |
1006 | return false; |
1007 | // We are rewriting the inserted reg. |
1008 | MachineOperand &MO = CopyLike.getOperand(i: CurrentSrcIdx); |
1009 | MO.setReg(NewReg); |
1010 | MO.setSubReg(NewSubReg); |
1011 | return true; |
1012 | } |
1013 | }; |
1014 | |
1015 | /// Specialized rewriter for EXTRACT_SUBREG instruction. |
1016 | class : public Rewriter { |
1017 | const TargetInstrInfo &; |
1018 | |
1019 | public: |
1020 | (MachineInstr &MI, const TargetInstrInfo &TII) |
1021 | : Rewriter(MI), TII(TII) { |
1022 | assert(MI.isExtractSubreg() && "Invalid instruction" ); |
1023 | } |
1024 | |
1025 | /// \see Rewriter::getNextRewritableSource() |
1026 | /// Here CopyLike has the following form: |
1027 | /// dst.dstSubIdx = EXTRACT_SUBREG Src, subIdx. |
1028 | /// There is only one rewritable source: Src.subIdx, |
1029 | /// which defines dst.dstSubIdx. |
1030 | bool (RegSubRegPair &Src, |
1031 | RegSubRegPair &Dst) override { |
1032 | // If we already get the only source we can rewrite, return false. |
1033 | if (CurrentSrcIdx == 1) |
1034 | return false; |
1035 | // We are looking at v1 = EXTRACT_SUBREG v0, sub0. |
1036 | CurrentSrcIdx = 1; |
1037 | const MachineOperand & = CopyLike.getOperand(i: 1); |
1038 | // If we have to compose sub-register indices, bail out. |
1039 | if (MOExtractedReg.getSubReg()) |
1040 | return false; |
1041 | |
1042 | Src = RegSubRegPair(MOExtractedReg.getReg(), |
1043 | CopyLike.getOperand(i: 2).getImm()); |
1044 | |
1045 | // We want to track something that is compatible with the definition. |
1046 | const MachineOperand &MODef = CopyLike.getOperand(i: 0); |
1047 | Dst = RegSubRegPair(MODef.getReg(), MODef.getSubReg()); |
1048 | return true; |
1049 | } |
1050 | |
1051 | bool (Register NewReg, unsigned NewSubReg) override { |
1052 | // The only source we can rewrite is the input register. |
1053 | if (CurrentSrcIdx != 1) |
1054 | return false; |
1055 | |
1056 | CopyLike.getOperand(i: CurrentSrcIdx).setReg(NewReg); |
1057 | |
1058 | // If we find a source that does not require to extract something, |
1059 | // rewrite the operation with a copy. |
1060 | if (!NewSubReg) { |
1061 | // Move the current index to an invalid position. |
1062 | // We do not want another call to this method to be able |
1063 | // to do any change. |
1064 | CurrentSrcIdx = -1; |
1065 | // Rewrite the operation as a COPY. |
1066 | // Get rid of the sub-register index. |
1067 | CopyLike.removeOperand(OpNo: 2); |
1068 | // Morph the operation into a COPY. |
1069 | CopyLike.setDesc(TII.get(Opcode: TargetOpcode::COPY)); |
1070 | return true; |
1071 | } |
1072 | CopyLike.getOperand(i: CurrentSrcIdx + 1).setImm(NewSubReg); |
1073 | return true; |
1074 | } |
1075 | }; |
1076 | |
1077 | /// Specialized rewriter for REG_SEQUENCE instruction. |
1078 | class RegSequenceRewriter : public Rewriter { |
1079 | public: |
1080 | RegSequenceRewriter(MachineInstr &MI) : Rewriter(MI) { |
1081 | assert(MI.isRegSequence() && "Invalid instruction" ); |
1082 | } |
1083 | |
1084 | /// \see Rewriter::getNextRewritableSource() |
1085 | /// Here CopyLike has the following form: |
1086 | /// dst = REG_SEQUENCE Src1.src1SubIdx, subIdx1, Src2.src2SubIdx, subIdx2. |
1087 | /// Each call will return a different source, walking all the available |
1088 | /// source. |
1089 | /// |
1090 | /// The first call returns: |
1091 | /// (SrcReg, SrcSubReg) = (Src1, src1SubIdx). |
1092 | /// (DstReg, DstSubReg) = (dst, subIdx1). |
1093 | /// |
1094 | /// The second call returns: |
1095 | /// (SrcReg, SrcSubReg) = (Src2, src2SubIdx). |
1096 | /// (DstReg, DstSubReg) = (dst, subIdx2). |
1097 | /// |
1098 | /// And so on, until all the sources have been traversed, then |
1099 | /// it returns false. |
1100 | bool getNextRewritableSource(RegSubRegPair &Src, |
1101 | RegSubRegPair &Dst) override { |
1102 | // We are looking at v0 = REG_SEQUENCE v1, sub1, v2, sub2, etc. |
1103 | |
1104 | // If this is the first call, move to the first argument. |
1105 | if (CurrentSrcIdx == 0) { |
1106 | CurrentSrcIdx = 1; |
1107 | } else { |
1108 | // Otherwise, move to the next argument and check that it is valid. |
1109 | CurrentSrcIdx += 2; |
1110 | if (CurrentSrcIdx >= CopyLike.getNumOperands()) |
1111 | return false; |
1112 | } |
1113 | const MachineOperand &MOInsertedReg = CopyLike.getOperand(i: CurrentSrcIdx); |
1114 | Src.Reg = MOInsertedReg.getReg(); |
1115 | // If we have to compose sub-register indices, bail out. |
1116 | if ((Src.SubReg = MOInsertedReg.getSubReg())) |
1117 | return false; |
1118 | |
1119 | // We want to track something that is compatible with the related |
1120 | // partial definition. |
1121 | Dst.SubReg = CopyLike.getOperand(i: CurrentSrcIdx + 1).getImm(); |
1122 | |
1123 | const MachineOperand &MODef = CopyLike.getOperand(i: 0); |
1124 | Dst.Reg = MODef.getReg(); |
1125 | // If we have to compose sub-registers, bail. |
1126 | return MODef.getSubReg() == 0; |
1127 | } |
1128 | |
1129 | bool RewriteCurrentSource(Register NewReg, unsigned NewSubReg) override { |
1130 | // We cannot rewrite out of bound operands. |
1131 | // Moreover, rewritable sources are at odd positions. |
1132 | if ((CurrentSrcIdx & 1) != 1 || CurrentSrcIdx > CopyLike.getNumOperands()) |
1133 | return false; |
1134 | |
1135 | MachineOperand &MO = CopyLike.getOperand(i: CurrentSrcIdx); |
1136 | MO.setReg(NewReg); |
1137 | MO.setSubReg(NewSubReg); |
1138 | return true; |
1139 | } |
1140 | }; |
1141 | |
1142 | } // end anonymous namespace |
1143 | |
1144 | /// Get the appropriated Rewriter for \p MI. |
1145 | /// \return A pointer to a dynamically allocated Rewriter or nullptr if no |
1146 | /// rewriter works for \p MI. |
1147 | static Rewriter *getCopyRewriter(MachineInstr &MI, const TargetInstrInfo &TII) { |
1148 | // Handle uncoalescable copy-like instructions. |
1149 | if (MI.isBitcast() || MI.isRegSequenceLike() || MI.isInsertSubregLike() || |
1150 | MI.isExtractSubregLike()) |
1151 | return new UncoalescableRewriter(MI); |
1152 | |
1153 | switch (MI.getOpcode()) { |
1154 | default: |
1155 | return nullptr; |
1156 | case TargetOpcode::COPY: |
1157 | return new CopyRewriter(MI); |
1158 | case TargetOpcode::INSERT_SUBREG: |
1159 | return new InsertSubregRewriter(MI); |
1160 | case TargetOpcode::EXTRACT_SUBREG: |
1161 | return new ExtractSubregRewriter(MI, TII); |
1162 | case TargetOpcode::REG_SEQUENCE: |
1163 | return new RegSequenceRewriter(MI); |
1164 | } |
1165 | } |
1166 | |
1167 | /// Given a \p Def.Reg and Def.SubReg pair, use \p RewriteMap to find |
1168 | /// the new source to use for rewrite. If \p HandleMultipleSources is true and |
1169 | /// multiple sources for a given \p Def are found along the way, we found a |
1170 | /// PHI instructions that needs to be rewritten. |
1171 | /// TODO: HandleMultipleSources should be removed once we test PHI handling |
1172 | /// with coalescable copies. |
1173 | static RegSubRegPair |
1174 | getNewSource(MachineRegisterInfo *MRI, const TargetInstrInfo *TII, |
1175 | RegSubRegPair Def, |
1176 | const PeepholeOptimizer::RewriteMapTy &RewriteMap, |
1177 | bool HandleMultipleSources = true) { |
1178 | RegSubRegPair LookupSrc(Def.Reg, Def.SubReg); |
1179 | while (true) { |
1180 | ValueTrackerResult Res = RewriteMap.lookup(Val: LookupSrc); |
1181 | // If there are no entries on the map, LookupSrc is the new source. |
1182 | if (!Res.isValid()) |
1183 | return LookupSrc; |
1184 | |
1185 | // There's only one source for this definition, keep searching... |
1186 | unsigned NumSrcs = Res.getNumSources(); |
1187 | if (NumSrcs == 1) { |
1188 | LookupSrc.Reg = Res.getSrcReg(Idx: 0); |
1189 | LookupSrc.SubReg = Res.getSrcSubReg(Idx: 0); |
1190 | continue; |
1191 | } |
1192 | |
1193 | // TODO: Remove once multiple srcs w/ coalescable copies are supported. |
1194 | if (!HandleMultipleSources) |
1195 | break; |
1196 | |
1197 | // Multiple sources, recurse into each source to find a new source |
1198 | // for it. Then, rewrite the PHI accordingly to its new edges. |
1199 | SmallVector<RegSubRegPair, 4> NewPHISrcs; |
1200 | for (unsigned i = 0; i < NumSrcs; ++i) { |
1201 | RegSubRegPair PHISrc(Res.getSrcReg(Idx: i), Res.getSrcSubReg(Idx: i)); |
1202 | NewPHISrcs.push_back( |
1203 | Elt: getNewSource(MRI, TII, Def: PHISrc, RewriteMap, HandleMultipleSources)); |
1204 | } |
1205 | |
1206 | // Build the new PHI node and return its def register as the new source. |
1207 | MachineInstr &OrigPHI = const_cast<MachineInstr &>(*Res.getInst()); |
1208 | MachineInstr &NewPHI = insertPHI(MRI&: *MRI, TII: *TII, SrcRegs: NewPHISrcs, OrigPHI); |
1209 | LLVM_DEBUG(dbgs() << "-- getNewSource\n" ); |
1210 | LLVM_DEBUG(dbgs() << " Replacing: " << OrigPHI); |
1211 | LLVM_DEBUG(dbgs() << " With: " << NewPHI); |
1212 | const MachineOperand &MODef = NewPHI.getOperand(i: 0); |
1213 | return RegSubRegPair(MODef.getReg(), MODef.getSubReg()); |
1214 | } |
1215 | |
1216 | return RegSubRegPair(0, 0); |
1217 | } |
1218 | |
1219 | /// Optimize generic copy instructions to avoid cross register bank copy. |
1220 | /// The optimization looks through a chain of copies and tries to find a source |
1221 | /// that has a compatible register class. |
1222 | /// Two register classes are considered to be compatible if they share the same |
1223 | /// register bank. |
1224 | /// New copies issued by this optimization are register allocator |
1225 | /// friendly. This optimization does not remove any copy as it may |
1226 | /// overconstrain the register allocator, but replaces some operands |
1227 | /// when possible. |
1228 | /// \pre isCoalescableCopy(*MI) is true. |
1229 | /// \return True, when \p MI has been rewritten. False otherwise. |
1230 | bool PeepholeOptimizer::optimizeCoalescableCopy(MachineInstr &MI) { |
1231 | assert(isCoalescableCopy(MI) && "Invalid argument" ); |
1232 | assert(MI.getDesc().getNumDefs() == 1 && |
1233 | "Coalescer can understand multiple defs?!" ); |
1234 | const MachineOperand &MODef = MI.getOperand(i: 0); |
1235 | // Do not rewrite physical definitions. |
1236 | if (MODef.getReg().isPhysical()) |
1237 | return false; |
1238 | |
1239 | bool Changed = false; |
1240 | // Get the right rewriter for the current copy. |
1241 | std::unique_ptr<Rewriter> CpyRewriter(getCopyRewriter(MI, TII: *TII)); |
1242 | // If none exists, bail out. |
1243 | if (!CpyRewriter) |
1244 | return false; |
1245 | // Rewrite each rewritable source. |
1246 | RegSubRegPair Src; |
1247 | RegSubRegPair TrackPair; |
1248 | while (CpyRewriter->getNextRewritableSource(Src, Dst&: TrackPair)) { |
1249 | // Keep track of PHI nodes and its incoming edges when looking for sources. |
1250 | RewriteMapTy RewriteMap; |
1251 | // Try to find a more suitable source. If we failed to do so, or get the |
1252 | // actual source, move to the next source. |
1253 | if (!findNextSource(RegSubReg: TrackPair, RewriteMap)) |
1254 | continue; |
1255 | |
1256 | // Get the new source to rewrite. TODO: Only enable handling of multiple |
1257 | // sources (PHIs) once we have a motivating example and testcases for it. |
1258 | RegSubRegPair NewSrc = getNewSource(MRI, TII, Def: TrackPair, RewriteMap, |
1259 | /*HandleMultipleSources=*/false); |
1260 | if (Src.Reg == NewSrc.Reg || NewSrc.Reg == 0) |
1261 | continue; |
1262 | |
1263 | // Rewrite source. |
1264 | if (CpyRewriter->RewriteCurrentSource(NewReg: NewSrc.Reg, NewSubReg: NewSrc.SubReg)) { |
1265 | // We may have extended the live-range of NewSrc, account for that. |
1266 | MRI->clearKillFlags(Reg: NewSrc.Reg); |
1267 | Changed = true; |
1268 | } |
1269 | } |
1270 | // TODO: We could have a clean-up method to tidy the instruction. |
1271 | // E.g., v0 = INSERT_SUBREG v1, v1.sub0, sub0 |
1272 | // => v0 = COPY v1 |
1273 | // Currently we haven't seen motivating example for that and we |
1274 | // want to avoid untested code. |
1275 | NumRewrittenCopies += Changed; |
1276 | return Changed; |
1277 | } |
1278 | |
1279 | /// Rewrite the source found through \p Def, by using the \p RewriteMap |
1280 | /// and create a new COPY instruction. More info about RewriteMap in |
1281 | /// PeepholeOptimizer::findNextSource. Right now this is only used to handle |
1282 | /// Uncoalescable copies, since they are copy like instructions that aren't |
1283 | /// recognized by the register allocator. |
1284 | MachineInstr & |
1285 | PeepholeOptimizer::rewriteSource(MachineInstr &CopyLike, |
1286 | RegSubRegPair Def, RewriteMapTy &RewriteMap) { |
1287 | assert(!Def.Reg.isPhysical() && "We do not rewrite physical registers" ); |
1288 | |
1289 | // Find the new source to use in the COPY rewrite. |
1290 | RegSubRegPair NewSrc = getNewSource(MRI, TII, Def, RewriteMap); |
1291 | |
1292 | // Insert the COPY. |
1293 | const TargetRegisterClass *DefRC = MRI->getRegClass(Reg: Def.Reg); |
1294 | Register NewVReg = MRI->createVirtualRegister(RegClass: DefRC); |
1295 | |
1296 | MachineInstr *NewCopy = |
1297 | BuildMI(BB&: *CopyLike.getParent(), I: &CopyLike, MIMD: CopyLike.getDebugLoc(), |
1298 | MCID: TII->get(Opcode: TargetOpcode::COPY), DestReg: NewVReg) |
1299 | .addReg(RegNo: NewSrc.Reg, flags: 0, SubReg: NewSrc.SubReg); |
1300 | |
1301 | if (Def.SubReg) { |
1302 | NewCopy->getOperand(i: 0).setSubReg(Def.SubReg); |
1303 | NewCopy->getOperand(i: 0).setIsUndef(); |
1304 | } |
1305 | |
1306 | LLVM_DEBUG(dbgs() << "-- RewriteSource\n" ); |
1307 | LLVM_DEBUG(dbgs() << " Replacing: " << CopyLike); |
1308 | LLVM_DEBUG(dbgs() << " With: " << *NewCopy); |
1309 | MRI->replaceRegWith(FromReg: Def.Reg, ToReg: NewVReg); |
1310 | MRI->clearKillFlags(Reg: NewVReg); |
1311 | |
1312 | // We extended the lifetime of NewSrc.Reg, clear the kill flags to |
1313 | // account for that. |
1314 | MRI->clearKillFlags(Reg: NewSrc.Reg); |
1315 | |
1316 | return *NewCopy; |
1317 | } |
1318 | |
1319 | /// Optimize copy-like instructions to create |
1320 | /// register coalescer friendly instruction. |
1321 | /// The optimization tries to kill-off the \p MI by looking |
1322 | /// through a chain of copies to find a source that has a compatible |
1323 | /// register class. |
1324 | /// If such a source is found, it replace \p MI by a generic COPY |
1325 | /// operation. |
1326 | /// \pre isUncoalescableCopy(*MI) is true. |
1327 | /// \return True, when \p MI has been optimized. In that case, \p MI has |
1328 | /// been removed from its parent. |
1329 | /// All COPY instructions created, are inserted in \p LocalMIs. |
1330 | bool PeepholeOptimizer::optimizeUncoalescableCopy( |
1331 | MachineInstr &MI, SmallPtrSetImpl<MachineInstr *> &LocalMIs) { |
1332 | assert(isUncoalescableCopy(MI) && "Invalid argument" ); |
1333 | UncoalescableRewriter CpyRewriter(MI); |
1334 | |
1335 | // Rewrite each rewritable source by generating new COPYs. This works |
1336 | // differently from optimizeCoalescableCopy since it first makes sure that all |
1337 | // definitions can be rewritten. |
1338 | RewriteMapTy RewriteMap; |
1339 | RegSubRegPair Src; |
1340 | RegSubRegPair Def; |
1341 | SmallVector<RegSubRegPair, 4> RewritePairs; |
1342 | while (CpyRewriter.getNextRewritableSource(Src, Dst&: Def)) { |
1343 | // If a physical register is here, this is probably for a good reason. |
1344 | // Do not rewrite that. |
1345 | if (Def.Reg.isPhysical()) |
1346 | return false; |
1347 | |
1348 | // If we do not know how to rewrite this definition, there is no point |
1349 | // in trying to kill this instruction. |
1350 | if (!findNextSource(RegSubReg: Def, RewriteMap)) |
1351 | return false; |
1352 | |
1353 | RewritePairs.push_back(Elt: Def); |
1354 | } |
1355 | |
1356 | // The change is possible for all defs, do it. |
1357 | for (const RegSubRegPair &Def : RewritePairs) { |
1358 | // Rewrite the "copy" in a way the register coalescer understands. |
1359 | MachineInstr &NewCopy = rewriteSource(CopyLike&: MI, Def, RewriteMap); |
1360 | LocalMIs.insert(Ptr: &NewCopy); |
1361 | } |
1362 | |
1363 | // MI is now dead. |
1364 | LLVM_DEBUG(dbgs() << "Deleting uncoalescable copy: " << MI); |
1365 | MI.eraseFromParent(); |
1366 | ++NumUncoalescableCopies; |
1367 | return true; |
1368 | } |
1369 | |
1370 | /// Check whether MI is a candidate for folding into a later instruction. |
1371 | /// We only fold loads to virtual registers and the virtual register defined |
1372 | /// has a single user. |
1373 | bool PeepholeOptimizer::isLoadFoldable( |
1374 | MachineInstr &MI, SmallSet<Register, 16> &FoldAsLoadDefCandidates) { |
1375 | if (!MI.canFoldAsLoad() || !MI.mayLoad()) |
1376 | return false; |
1377 | const MCInstrDesc &MCID = MI.getDesc(); |
1378 | if (MCID.getNumDefs() != 1) |
1379 | return false; |
1380 | |
1381 | Register Reg = MI.getOperand(i: 0).getReg(); |
1382 | // To reduce compilation time, we check MRI->hasOneNonDBGUser when inserting |
1383 | // loads. It should be checked when processing uses of the load, since |
1384 | // uses can be removed during peephole. |
1385 | if (Reg.isVirtual() && !MI.getOperand(i: 0).getSubReg() && |
1386 | MRI->hasOneNonDBGUser(RegNo: Reg)) { |
1387 | FoldAsLoadDefCandidates.insert(V: Reg); |
1388 | return true; |
1389 | } |
1390 | return false; |
1391 | } |
1392 | |
1393 | bool PeepholeOptimizer::isMoveImmediate( |
1394 | MachineInstr &MI, SmallSet<Register, 4> &ImmDefRegs, |
1395 | DenseMap<Register, MachineInstr *> &ImmDefMIs) { |
1396 | const MCInstrDesc &MCID = MI.getDesc(); |
1397 | if (MCID.getNumDefs() != 1 || !MI.getOperand(i: 0).isReg()) |
1398 | return false; |
1399 | Register Reg = MI.getOperand(i: 0).getReg(); |
1400 | if (!Reg.isVirtual()) |
1401 | return false; |
1402 | |
1403 | int64_t ImmVal; |
1404 | if (!MI.isMoveImmediate() && !TII->getConstValDefinedInReg(MI, Reg, ImmVal)) |
1405 | return false; |
1406 | |
1407 | ImmDefMIs.insert(KV: std::make_pair(x&: Reg, y: &MI)); |
1408 | ImmDefRegs.insert(V: Reg); |
1409 | return true; |
1410 | } |
1411 | |
1412 | /// Try folding register operands that are defined by move immediate |
1413 | /// instructions, i.e. a trivial constant folding optimization, if |
1414 | /// and only if the def and use are in the same BB. |
1415 | bool PeepholeOptimizer::foldImmediate( |
1416 | MachineInstr &MI, SmallSet<Register, 4> &ImmDefRegs, |
1417 | DenseMap<Register, MachineInstr *> &ImmDefMIs, bool &Deleted) { |
1418 | Deleted = false; |
1419 | for (unsigned i = 0, e = MI.getDesc().getNumOperands(); i != e; ++i) { |
1420 | MachineOperand &MO = MI.getOperand(i); |
1421 | if (!MO.isReg() || MO.isDef()) |
1422 | continue; |
1423 | Register Reg = MO.getReg(); |
1424 | if (!Reg.isVirtual()) |
1425 | continue; |
1426 | if (ImmDefRegs.count(V: Reg) == 0) |
1427 | continue; |
1428 | DenseMap<Register, MachineInstr *>::iterator II = ImmDefMIs.find(Val: Reg); |
1429 | assert(II != ImmDefMIs.end() && "couldn't find immediate definition" ); |
1430 | if (TII->foldImmediate(UseMI&: MI, DefMI&: *II->second, Reg, MRI)) { |
1431 | ++NumImmFold; |
1432 | // foldImmediate can delete ImmDefMI if MI was its only user. If ImmDefMI |
1433 | // is not deleted, and we happened to get a same MI, we can delete MI and |
1434 | // replace its users. |
1435 | if (MRI->getVRegDef(Reg) && |
1436 | MI.isIdenticalTo(Other: *II->second, Check: MachineInstr::IgnoreVRegDefs)) { |
1437 | Register DstReg = MI.getOperand(i: 0).getReg(); |
1438 | if (DstReg.isVirtual() && |
1439 | MRI->getRegClass(Reg: DstReg) == MRI->getRegClass(Reg)) { |
1440 | MRI->replaceRegWith(FromReg: DstReg, ToReg: Reg); |
1441 | MI.eraseFromParent(); |
1442 | Deleted = true; |
1443 | } |
1444 | } |
1445 | return true; |
1446 | } |
1447 | } |
1448 | return false; |
1449 | } |
1450 | |
1451 | // FIXME: This is very simple and misses some cases which should be handled when |
1452 | // motivating examples are found. |
1453 | // |
1454 | // The copy rewriting logic should look at uses as well as defs and be able to |
1455 | // eliminate copies across blocks. |
1456 | // |
1457 | // Later copies that are subregister extracts will also not be eliminated since |
1458 | // only the first copy is considered. |
1459 | // |
1460 | // e.g. |
1461 | // %1 = COPY %0 |
1462 | // %2 = COPY %0:sub1 |
1463 | // |
1464 | // Should replace %2 uses with %1:sub1 |
1465 | bool PeepholeOptimizer::foldRedundantCopy(MachineInstr &MI) { |
1466 | assert(MI.isCopy() && "expected a COPY machine instruction" ); |
1467 | |
1468 | RegSubRegPair SrcPair; |
1469 | if (!getCopySrc(MI, SrcPair)) |
1470 | return false; |
1471 | |
1472 | Register DstReg = MI.getOperand(i: 0).getReg(); |
1473 | if (!DstReg.isVirtual()) |
1474 | return false; |
1475 | |
1476 | if (CopySrcMIs.insert(KV: std::make_pair(x&: SrcPair, y: &MI)).second) { |
1477 | // First copy of this reg seen. |
1478 | return false; |
1479 | } |
1480 | |
1481 | MachineInstr *PrevCopy = CopySrcMIs.find(Val: SrcPair)->second; |
1482 | |
1483 | assert(SrcPair.SubReg == PrevCopy->getOperand(1).getSubReg() && |
1484 | "Unexpected mismatching subreg!" ); |
1485 | |
1486 | Register PrevDstReg = PrevCopy->getOperand(i: 0).getReg(); |
1487 | |
1488 | // Only replace if the copy register class is the same. |
1489 | // |
1490 | // TODO: If we have multiple copies to different register classes, we may want |
1491 | // to track multiple copies of the same source register. |
1492 | if (MRI->getRegClass(Reg: DstReg) != MRI->getRegClass(Reg: PrevDstReg)) |
1493 | return false; |
1494 | |
1495 | MRI->replaceRegWith(FromReg: DstReg, ToReg: PrevDstReg); |
1496 | |
1497 | // Lifetime of the previous copy has been extended. |
1498 | MRI->clearKillFlags(Reg: PrevDstReg); |
1499 | return true; |
1500 | } |
1501 | |
1502 | bool PeepholeOptimizer::isNAPhysCopy(Register Reg) { |
1503 | return Reg.isPhysical() && !MRI->isAllocatable(PhysReg: Reg); |
1504 | } |
1505 | |
1506 | bool PeepholeOptimizer::foldRedundantNAPhysCopy( |
1507 | MachineInstr &MI, DenseMap<Register, MachineInstr *> &NAPhysToVirtMIs) { |
1508 | assert(MI.isCopy() && "expected a COPY machine instruction" ); |
1509 | |
1510 | if (DisableNAPhysCopyOpt) |
1511 | return false; |
1512 | |
1513 | Register DstReg = MI.getOperand(i: 0).getReg(); |
1514 | Register SrcReg = MI.getOperand(i: 1).getReg(); |
1515 | if (isNAPhysCopy(Reg: SrcReg) && DstReg.isVirtual()) { |
1516 | // %vreg = COPY $physreg |
1517 | // Avoid using a datastructure which can track multiple live non-allocatable |
1518 | // phys->virt copies since LLVM doesn't seem to do this. |
1519 | NAPhysToVirtMIs.insert(KV: {SrcReg, &MI}); |
1520 | return false; |
1521 | } |
1522 | |
1523 | if (!(SrcReg.isVirtual() && isNAPhysCopy(Reg: DstReg))) |
1524 | return false; |
1525 | |
1526 | // $physreg = COPY %vreg |
1527 | auto PrevCopy = NAPhysToVirtMIs.find(Val: DstReg); |
1528 | if (PrevCopy == NAPhysToVirtMIs.end()) { |
1529 | // We can't remove the copy: there was an intervening clobber of the |
1530 | // non-allocatable physical register after the copy to virtual. |
1531 | LLVM_DEBUG(dbgs() << "NAPhysCopy: intervening clobber forbids erasing " |
1532 | << MI); |
1533 | return false; |
1534 | } |
1535 | |
1536 | Register PrevDstReg = PrevCopy->second->getOperand(i: 0).getReg(); |
1537 | if (PrevDstReg == SrcReg) { |
1538 | // Remove the virt->phys copy: we saw the virtual register definition, and |
1539 | // the non-allocatable physical register's state hasn't changed since then. |
1540 | LLVM_DEBUG(dbgs() << "NAPhysCopy: erasing " << MI); |
1541 | ++NumNAPhysCopies; |
1542 | return true; |
1543 | } |
1544 | |
1545 | // Potential missed optimization opportunity: we saw a different virtual |
1546 | // register get a copy of the non-allocatable physical register, and we only |
1547 | // track one such copy. Avoid getting confused by this new non-allocatable |
1548 | // physical register definition, and remove it from the tracked copies. |
1549 | LLVM_DEBUG(dbgs() << "NAPhysCopy: missed opportunity " << MI); |
1550 | NAPhysToVirtMIs.erase(I: PrevCopy); |
1551 | return false; |
1552 | } |
1553 | |
1554 | /// \bried Returns true if \p MO is a virtual register operand. |
1555 | static bool isVirtualRegisterOperand(MachineOperand &MO) { |
1556 | return MO.isReg() && MO.getReg().isVirtual(); |
1557 | } |
1558 | |
1559 | bool PeepholeOptimizer::findTargetRecurrence( |
1560 | Register Reg, const SmallSet<Register, 2> &TargetRegs, |
1561 | RecurrenceCycle &RC) { |
1562 | // Recurrence found if Reg is in TargetRegs. |
1563 | if (TargetRegs.count(V: Reg)) |
1564 | return true; |
1565 | |
1566 | // TODO: Curerntly, we only allow the last instruction of the recurrence |
1567 | // cycle (the instruction that feeds the PHI instruction) to have more than |
1568 | // one uses to guarantee that commuting operands does not tie registers |
1569 | // with overlapping live range. Once we have actual live range info of |
1570 | // each register, this constraint can be relaxed. |
1571 | if (!MRI->hasOneNonDBGUse(RegNo: Reg)) |
1572 | return false; |
1573 | |
1574 | // Give up if the reccurrence chain length is longer than the limit. |
1575 | if (RC.size() >= MaxRecurrenceChain) |
1576 | return false; |
1577 | |
1578 | MachineInstr &MI = *(MRI->use_instr_nodbg_begin(RegNo: Reg)); |
1579 | unsigned Idx = MI.findRegisterUseOperandIdx(Reg, /*TRI=*/nullptr); |
1580 | |
1581 | // Only interested in recurrences whose instructions have only one def, which |
1582 | // is a virtual register. |
1583 | if (MI.getDesc().getNumDefs() != 1) |
1584 | return false; |
1585 | |
1586 | MachineOperand &DefOp = MI.getOperand(i: 0); |
1587 | if (!isVirtualRegisterOperand(MO&: DefOp)) |
1588 | return false; |
1589 | |
1590 | // Check if def operand of MI is tied to any use operand. We are only |
1591 | // interested in the case that all the instructions in the recurrence chain |
1592 | // have there def operand tied with one of the use operand. |
1593 | unsigned TiedUseIdx; |
1594 | if (!MI.isRegTiedToUseOperand(DefOpIdx: 0, UseOpIdx: &TiedUseIdx)) |
1595 | return false; |
1596 | |
1597 | if (Idx == TiedUseIdx) { |
1598 | RC.push_back(Elt: RecurrenceInstr(&MI)); |
1599 | return findTargetRecurrence(Reg: DefOp.getReg(), TargetRegs, RC); |
1600 | } else { |
1601 | // If Idx is not TiedUseIdx, check if Idx is commutable with TiedUseIdx. |
1602 | unsigned CommIdx = TargetInstrInfo::CommuteAnyOperandIndex; |
1603 | if (TII->findCommutedOpIndices(MI, SrcOpIdx1&: Idx, SrcOpIdx2&: CommIdx) && CommIdx == TiedUseIdx) { |
1604 | RC.push_back(Elt: RecurrenceInstr(&MI, Idx, CommIdx)); |
1605 | return findTargetRecurrence(Reg: DefOp.getReg(), TargetRegs, RC); |
1606 | } |
1607 | } |
1608 | |
1609 | return false; |
1610 | } |
1611 | |
1612 | /// Phi instructions will eventually be lowered to copy instructions. |
1613 | /// If phi is in a loop header, a recurrence may formulated around the source |
1614 | /// and destination of the phi. For such case commuting operands of the |
1615 | /// instructions in the recurrence may enable coalescing of the copy instruction |
1616 | /// generated from the phi. For example, if there is a recurrence of |
1617 | /// |
1618 | /// LoopHeader: |
1619 | /// %1 = phi(%0, %100) |
1620 | /// LoopLatch: |
1621 | /// %0<def, tied1> = ADD %2<def, tied0>, %1 |
1622 | /// |
1623 | /// , the fact that %0 and %2 are in the same tied operands set makes |
1624 | /// the coalescing of copy instruction generated from the phi in |
1625 | /// LoopHeader(i.e. %1 = COPY %0) impossible, because %1 and |
1626 | /// %2 have overlapping live range. This introduces additional move |
1627 | /// instruction to the final assembly. However, if we commute %2 and |
1628 | /// %1 of ADD instruction, the redundant move instruction can be |
1629 | /// avoided. |
1630 | bool PeepholeOptimizer::optimizeRecurrence(MachineInstr &PHI) { |
1631 | SmallSet<Register, 2> TargetRegs; |
1632 | for (unsigned Idx = 1; Idx < PHI.getNumOperands(); Idx += 2) { |
1633 | MachineOperand &MO = PHI.getOperand(i: Idx); |
1634 | assert(isVirtualRegisterOperand(MO) && "Invalid PHI instruction" ); |
1635 | TargetRegs.insert(V: MO.getReg()); |
1636 | } |
1637 | |
1638 | bool Changed = false; |
1639 | RecurrenceCycle RC; |
1640 | if (findTargetRecurrence(Reg: PHI.getOperand(i: 0).getReg(), TargetRegs, RC)) { |
1641 | // Commutes operands of instructions in RC if necessary so that the copy to |
1642 | // be generated from PHI can be coalesced. |
1643 | LLVM_DEBUG(dbgs() << "Optimize recurrence chain from " << PHI); |
1644 | for (auto &RI : RC) { |
1645 | LLVM_DEBUG(dbgs() << "\tInst: " << *(RI.getMI())); |
1646 | auto CP = RI.getCommutePair(); |
1647 | if (CP) { |
1648 | Changed = true; |
1649 | TII->commuteInstruction(MI&: *(RI.getMI()), NewMI: false, OpIdx1: (*CP).first, |
1650 | OpIdx2: (*CP).second); |
1651 | LLVM_DEBUG(dbgs() << "\t\tCommuted: " << *(RI.getMI())); |
1652 | } |
1653 | } |
1654 | } |
1655 | |
1656 | return Changed; |
1657 | } |
1658 | |
1659 | bool PeepholeOptimizer::runOnMachineFunction(MachineFunction &MF) { |
1660 | if (skipFunction(F: MF.getFunction())) |
1661 | return false; |
1662 | |
1663 | LLVM_DEBUG(dbgs() << "********** PEEPHOLE OPTIMIZER **********\n" ); |
1664 | LLVM_DEBUG(dbgs() << "********** Function: " << MF.getName() << '\n'); |
1665 | |
1666 | if (DisablePeephole) |
1667 | return false; |
1668 | |
1669 | TII = MF.getSubtarget().getInstrInfo(); |
1670 | TRI = MF.getSubtarget().getRegisterInfo(); |
1671 | MRI = &MF.getRegInfo(); |
1672 | DT = Aggressive ? &getAnalysis<MachineDominatorTreeWrapperPass>().getDomTree() |
1673 | : nullptr; |
1674 | MLI = &getAnalysis<MachineLoopInfoWrapperPass>().getLI(); |
1675 | MF.setDelegate(this); |
1676 | |
1677 | bool Changed = false; |
1678 | |
1679 | for (MachineBasicBlock &MBB : MF) { |
1680 | bool SeenMoveImm = false; |
1681 | |
1682 | // During this forward scan, at some point it needs to answer the question |
1683 | // "given a pointer to an MI in the current BB, is it located before or |
1684 | // after the current instruction". |
1685 | // To perform this, the following set keeps track of the MIs already seen |
1686 | // during the scan, if a MI is not in the set, it is assumed to be located |
1687 | // after. Newly created MIs have to be inserted in the set as well. |
1688 | SmallPtrSet<MachineInstr*, 16> LocalMIs; |
1689 | SmallSet<Register, 4> ImmDefRegs; |
1690 | DenseMap<Register, MachineInstr *> ImmDefMIs; |
1691 | SmallSet<Register, 16> FoldAsLoadDefCandidates; |
1692 | |
1693 | // Track when a non-allocatable physical register is copied to a virtual |
1694 | // register so that useless moves can be removed. |
1695 | // |
1696 | // $physreg is the map index; MI is the last valid `%vreg = COPY $physreg` |
1697 | // without any intervening re-definition of $physreg. |
1698 | DenseMap<Register, MachineInstr *> NAPhysToVirtMIs; |
1699 | |
1700 | CopySrcMIs.clear(); |
1701 | |
1702 | bool = MLI->isLoopHeader(BB: &MBB); |
1703 | |
1704 | for (MachineBasicBlock::iterator MII = MBB.begin(), MIE = MBB.end(); |
1705 | MII != MIE; ) { |
1706 | MachineInstr *MI = &*MII; |
1707 | // We may be erasing MI below, increment MII now. |
1708 | ++MII; |
1709 | LocalMIs.insert(Ptr: MI); |
1710 | |
1711 | // Skip debug instructions. They should not affect this peephole |
1712 | // optimization. |
1713 | if (MI->isDebugInstr()) |
1714 | continue; |
1715 | |
1716 | if (MI->isPosition()) |
1717 | continue; |
1718 | |
1719 | if (IsLoopHeader && MI->isPHI()) { |
1720 | if (optimizeRecurrence(PHI&: *MI)) { |
1721 | Changed = true; |
1722 | continue; |
1723 | } |
1724 | } |
1725 | |
1726 | if (!MI->isCopy()) { |
1727 | for (const MachineOperand &MO : MI->operands()) { |
1728 | // Visit all operands: definitions can be implicit or explicit. |
1729 | if (MO.isReg()) { |
1730 | Register Reg = MO.getReg(); |
1731 | if (MO.isDef() && isNAPhysCopy(Reg)) { |
1732 | const auto &Def = NAPhysToVirtMIs.find(Val: Reg); |
1733 | if (Def != NAPhysToVirtMIs.end()) { |
1734 | // A new definition of the non-allocatable physical register |
1735 | // invalidates previous copies. |
1736 | LLVM_DEBUG(dbgs() |
1737 | << "NAPhysCopy: invalidating because of " << *MI); |
1738 | NAPhysToVirtMIs.erase(I: Def); |
1739 | } |
1740 | } |
1741 | } else if (MO.isRegMask()) { |
1742 | const uint32_t *RegMask = MO.getRegMask(); |
1743 | for (auto &RegMI : NAPhysToVirtMIs) { |
1744 | Register Def = RegMI.first; |
1745 | if (MachineOperand::clobbersPhysReg(RegMask, PhysReg: Def)) { |
1746 | LLVM_DEBUG(dbgs() |
1747 | << "NAPhysCopy: invalidating because of " << *MI); |
1748 | NAPhysToVirtMIs.erase(Val: Def); |
1749 | } |
1750 | } |
1751 | } |
1752 | } |
1753 | } |
1754 | |
1755 | if (MI->isImplicitDef() || MI->isKill()) |
1756 | continue; |
1757 | |
1758 | if (MI->isInlineAsm() || MI->hasUnmodeledSideEffects()) { |
1759 | // Blow away all non-allocatable physical registers knowledge since we |
1760 | // don't know what's correct anymore. |
1761 | // |
1762 | // FIXME: handle explicit asm clobbers. |
1763 | LLVM_DEBUG(dbgs() << "NAPhysCopy: blowing away all info due to " |
1764 | << *MI); |
1765 | NAPhysToVirtMIs.clear(); |
1766 | } |
1767 | |
1768 | if ((isUncoalescableCopy(MI: *MI) && |
1769 | optimizeUncoalescableCopy(MI&: *MI, LocalMIs)) || |
1770 | (MI->isCompare() && optimizeCmpInstr(MI&: *MI)) || |
1771 | (MI->isSelect() && optimizeSelect(MI&: *MI, LocalMIs))) { |
1772 | // MI is deleted. |
1773 | LocalMIs.erase(Ptr: MI); |
1774 | Changed = true; |
1775 | continue; |
1776 | } |
1777 | |
1778 | if (MI->isConditionalBranch() && optimizeCondBranch(MI&: *MI)) { |
1779 | Changed = true; |
1780 | continue; |
1781 | } |
1782 | |
1783 | if (isCoalescableCopy(MI: *MI) && optimizeCoalescableCopy(MI&: *MI)) { |
1784 | // MI is just rewritten. |
1785 | Changed = true; |
1786 | continue; |
1787 | } |
1788 | |
1789 | if (MI->isCopy() && (foldRedundantCopy(MI&: *MI) || |
1790 | foldRedundantNAPhysCopy(MI&: *MI, NAPhysToVirtMIs))) { |
1791 | LocalMIs.erase(Ptr: MI); |
1792 | LLVM_DEBUG(dbgs() << "Deleting redundant copy: " << *MI << "\n" ); |
1793 | MI->eraseFromParent(); |
1794 | Changed = true; |
1795 | continue; |
1796 | } |
1797 | |
1798 | if (isMoveImmediate(MI&: *MI, ImmDefRegs, ImmDefMIs)) { |
1799 | SeenMoveImm = true; |
1800 | } else { |
1801 | Changed |= optimizeExtInstr(MI&: *MI, MBB, LocalMIs); |
1802 | // optimizeExtInstr might have created new instructions after MI |
1803 | // and before the already incremented MII. Adjust MII so that the |
1804 | // next iteration sees the new instructions. |
1805 | MII = MI; |
1806 | ++MII; |
1807 | if (SeenMoveImm) { |
1808 | bool Deleted; |
1809 | Changed |= foldImmediate(MI&: *MI, ImmDefRegs, ImmDefMIs, Deleted); |
1810 | if (Deleted) { |
1811 | LocalMIs.erase(Ptr: MI); |
1812 | continue; |
1813 | } |
1814 | } |
1815 | } |
1816 | |
1817 | // Check whether MI is a load candidate for folding into a later |
1818 | // instruction. If MI is not a candidate, check whether we can fold an |
1819 | // earlier load into MI. |
1820 | if (!isLoadFoldable(MI&: *MI, FoldAsLoadDefCandidates) && |
1821 | !FoldAsLoadDefCandidates.empty()) { |
1822 | |
1823 | // We visit each operand even after successfully folding a previous |
1824 | // one. This allows us to fold multiple loads into a single |
1825 | // instruction. We do assume that optimizeLoadInstr doesn't insert |
1826 | // foldable uses earlier in the argument list. Since we don't restart |
1827 | // iteration, we'd miss such cases. |
1828 | const MCInstrDesc &MIDesc = MI->getDesc(); |
1829 | for (unsigned i = MIDesc.getNumDefs(); i != MI->getNumOperands(); |
1830 | ++i) { |
1831 | const MachineOperand &MOp = MI->getOperand(i); |
1832 | if (!MOp.isReg()) |
1833 | continue; |
1834 | Register FoldAsLoadDefReg = MOp.getReg(); |
1835 | if (FoldAsLoadDefCandidates.count(V: FoldAsLoadDefReg)) { |
1836 | // We need to fold load after optimizeCmpInstr, since |
1837 | // optimizeCmpInstr can enable folding by converting SUB to CMP. |
1838 | // Save FoldAsLoadDefReg because optimizeLoadInstr() resets it and |
1839 | // we need it for markUsesInDebugValueAsUndef(). |
1840 | Register FoldedReg = FoldAsLoadDefReg; |
1841 | MachineInstr *DefMI = nullptr; |
1842 | if (MachineInstr *FoldMI = |
1843 | TII->optimizeLoadInstr(MI&: *MI, MRI, FoldAsLoadDefReg, DefMI)) { |
1844 | // Update LocalMIs since we replaced MI with FoldMI and deleted |
1845 | // DefMI. |
1846 | LLVM_DEBUG(dbgs() << "Replacing: " << *MI); |
1847 | LLVM_DEBUG(dbgs() << " With: " << *FoldMI); |
1848 | LocalMIs.erase(Ptr: MI); |
1849 | LocalMIs.erase(Ptr: DefMI); |
1850 | LocalMIs.insert(Ptr: FoldMI); |
1851 | // Update the call site info. |
1852 | if (MI->shouldUpdateCallSiteInfo()) |
1853 | MI->getMF()->moveCallSiteInfo(Old: MI, New: FoldMI); |
1854 | MI->eraseFromParent(); |
1855 | DefMI->eraseFromParent(); |
1856 | MRI->markUsesInDebugValueAsUndef(Reg: FoldedReg); |
1857 | FoldAsLoadDefCandidates.erase(V: FoldedReg); |
1858 | ++NumLoadFold; |
1859 | |
1860 | // MI is replaced with FoldMI so we can continue trying to fold |
1861 | Changed = true; |
1862 | MI = FoldMI; |
1863 | } |
1864 | } |
1865 | } |
1866 | } |
1867 | |
1868 | // If we run into an instruction we can't fold across, discard |
1869 | // the load candidates. Note: We might be able to fold *into* this |
1870 | // instruction, so this needs to be after the folding logic. |
1871 | if (MI->isLoadFoldBarrier()) { |
1872 | LLVM_DEBUG(dbgs() << "Encountered load fold barrier on " << *MI); |
1873 | FoldAsLoadDefCandidates.clear(); |
1874 | } |
1875 | } |
1876 | } |
1877 | |
1878 | MF.resetDelegate(delegate: this); |
1879 | return Changed; |
1880 | } |
1881 | |
1882 | ValueTrackerResult ValueTracker::getNextSourceFromCopy() { |
1883 | assert(Def->isCopy() && "Invalid definition" ); |
1884 | // Copy instruction are supposed to be: Def = Src. |
1885 | // If someone breaks this assumption, bad things will happen everywhere. |
1886 | // There may be implicit uses preventing the copy to be moved across |
1887 | // some target specific register definitions |
1888 | assert(Def->getNumOperands() - Def->getNumImplicitOperands() == 2 && |
1889 | "Invalid number of operands" ); |
1890 | assert(!Def->hasImplicitDef() && "Only implicit uses are allowed" ); |
1891 | |
1892 | if (Def->getOperand(i: DefIdx).getSubReg() != DefSubReg) |
1893 | // If we look for a different subreg, it means we want a subreg of src. |
1894 | // Bails as we do not support composing subregs yet. |
1895 | return ValueTrackerResult(); |
1896 | // Otherwise, we want the whole source. |
1897 | const MachineOperand &Src = Def->getOperand(i: 1); |
1898 | if (Src.isUndef()) |
1899 | return ValueTrackerResult(); |
1900 | return ValueTrackerResult(Src.getReg(), Src.getSubReg()); |
1901 | } |
1902 | |
1903 | ValueTrackerResult ValueTracker::getNextSourceFromBitcast() { |
1904 | assert(Def->isBitcast() && "Invalid definition" ); |
1905 | |
1906 | // Bail if there are effects that a plain copy will not expose. |
1907 | if (Def->mayRaiseFPException() || Def->hasUnmodeledSideEffects()) |
1908 | return ValueTrackerResult(); |
1909 | |
1910 | // Bitcasts with more than one def are not supported. |
1911 | if (Def->getDesc().getNumDefs() != 1) |
1912 | return ValueTrackerResult(); |
1913 | const MachineOperand DefOp = Def->getOperand(i: DefIdx); |
1914 | if (DefOp.getSubReg() != DefSubReg) |
1915 | // If we look for a different subreg, it means we want a subreg of the src. |
1916 | // Bails as we do not support composing subregs yet. |
1917 | return ValueTrackerResult(); |
1918 | |
1919 | unsigned SrcIdx = Def->getNumOperands(); |
1920 | for (unsigned OpIdx = DefIdx + 1, EndOpIdx = SrcIdx; OpIdx != EndOpIdx; |
1921 | ++OpIdx) { |
1922 | const MachineOperand &MO = Def->getOperand(i: OpIdx); |
1923 | if (!MO.isReg() || !MO.getReg()) |
1924 | continue; |
1925 | // Ignore dead implicit defs. |
1926 | if (MO.isImplicit() && MO.isDead()) |
1927 | continue; |
1928 | assert(!MO.isDef() && "We should have skipped all the definitions by now" ); |
1929 | if (SrcIdx != EndOpIdx) |
1930 | // Multiple sources? |
1931 | return ValueTrackerResult(); |
1932 | SrcIdx = OpIdx; |
1933 | } |
1934 | |
1935 | // In some rare case, Def has no input, SrcIdx is out of bound, |
1936 | // getOperand(SrcIdx) will fail below. |
1937 | if (SrcIdx >= Def->getNumOperands()) |
1938 | return ValueTrackerResult(); |
1939 | |
1940 | // Stop when any user of the bitcast is a SUBREG_TO_REG, replacing with a COPY |
1941 | // will break the assumed guarantees for the upper bits. |
1942 | for (const MachineInstr &UseMI : MRI.use_nodbg_instructions(Reg: DefOp.getReg())) { |
1943 | if (UseMI.isSubregToReg()) |
1944 | return ValueTrackerResult(); |
1945 | } |
1946 | |
1947 | const MachineOperand &Src = Def->getOperand(i: SrcIdx); |
1948 | if (Src.isUndef()) |
1949 | return ValueTrackerResult(); |
1950 | return ValueTrackerResult(Src.getReg(), Src.getSubReg()); |
1951 | } |
1952 | |
1953 | ValueTrackerResult ValueTracker::getNextSourceFromRegSequence() { |
1954 | assert((Def->isRegSequence() || Def->isRegSequenceLike()) && |
1955 | "Invalid definition" ); |
1956 | |
1957 | if (Def->getOperand(i: DefIdx).getSubReg()) |
1958 | // If we are composing subregs, bail out. |
1959 | // The case we are checking is Def.<subreg> = REG_SEQUENCE. |
1960 | // This should almost never happen as the SSA property is tracked at |
1961 | // the register level (as opposed to the subreg level). |
1962 | // I.e., |
1963 | // Def.sub0 = |
1964 | // Def.sub1 = |
1965 | // is a valid SSA representation for Def.sub0 and Def.sub1, but not for |
1966 | // Def. Thus, it must not be generated. |
1967 | // However, some code could theoretically generates a single |
1968 | // Def.sub0 (i.e, not defining the other subregs) and we would |
1969 | // have this case. |
1970 | // If we can ascertain (or force) that this never happens, we could |
1971 | // turn that into an assertion. |
1972 | return ValueTrackerResult(); |
1973 | |
1974 | if (!TII) |
1975 | // We could handle the REG_SEQUENCE here, but we do not want to |
1976 | // duplicate the code from the generic TII. |
1977 | return ValueTrackerResult(); |
1978 | |
1979 | SmallVector<RegSubRegPairAndIdx, 8> RegSeqInputRegs; |
1980 | if (!TII->getRegSequenceInputs(MI: *Def, DefIdx, InputRegs&: RegSeqInputRegs)) |
1981 | return ValueTrackerResult(); |
1982 | |
1983 | // We are looking at: |
1984 | // Def = REG_SEQUENCE v0, sub0, v1, sub1, ... |
1985 | // Check if one of the operand defines the subreg we are interested in. |
1986 | for (const RegSubRegPairAndIdx &RegSeqInput : RegSeqInputRegs) { |
1987 | if (RegSeqInput.SubIdx == DefSubReg) |
1988 | return ValueTrackerResult(RegSeqInput.Reg, RegSeqInput.SubReg); |
1989 | } |
1990 | |
1991 | // If the subreg we are tracking is super-defined by another subreg, |
1992 | // we could follow this value. However, this would require to compose |
1993 | // the subreg and we do not do that for now. |
1994 | return ValueTrackerResult(); |
1995 | } |
1996 | |
1997 | ValueTrackerResult ValueTracker::getNextSourceFromInsertSubreg() { |
1998 | assert((Def->isInsertSubreg() || Def->isInsertSubregLike()) && |
1999 | "Invalid definition" ); |
2000 | |
2001 | if (Def->getOperand(i: DefIdx).getSubReg()) |
2002 | // If we are composing subreg, bail out. |
2003 | // Same remark as getNextSourceFromRegSequence. |
2004 | // I.e., this may be turned into an assert. |
2005 | return ValueTrackerResult(); |
2006 | |
2007 | if (!TII) |
2008 | // We could handle the REG_SEQUENCE here, but we do not want to |
2009 | // duplicate the code from the generic TII. |
2010 | return ValueTrackerResult(); |
2011 | |
2012 | RegSubRegPair BaseReg; |
2013 | RegSubRegPairAndIdx InsertedReg; |
2014 | if (!TII->getInsertSubregInputs(MI: *Def, DefIdx, BaseReg, InsertedReg)) |
2015 | return ValueTrackerResult(); |
2016 | |
2017 | // We are looking at: |
2018 | // Def = INSERT_SUBREG v0, v1, sub1 |
2019 | // There are two cases: |
2020 | // 1. DefSubReg == sub1, get v1. |
2021 | // 2. DefSubReg != sub1, the value may be available through v0. |
2022 | |
2023 | // #1 Check if the inserted register matches the required sub index. |
2024 | if (InsertedReg.SubIdx == DefSubReg) { |
2025 | return ValueTrackerResult(InsertedReg.Reg, InsertedReg.SubReg); |
2026 | } |
2027 | // #2 Otherwise, if the sub register we are looking for is not partial |
2028 | // defined by the inserted element, we can look through the main |
2029 | // register (v0). |
2030 | const MachineOperand &MODef = Def->getOperand(i: DefIdx); |
2031 | // If the result register (Def) and the base register (v0) do not |
2032 | // have the same register class or if we have to compose |
2033 | // subregisters, bail out. |
2034 | if (MRI.getRegClass(Reg: MODef.getReg()) != MRI.getRegClass(Reg: BaseReg.Reg) || |
2035 | BaseReg.SubReg) |
2036 | return ValueTrackerResult(); |
2037 | |
2038 | // Get the TRI and check if the inserted sub-register overlaps with the |
2039 | // sub-register we are tracking. |
2040 | const TargetRegisterInfo *TRI = MRI.getTargetRegisterInfo(); |
2041 | if (!TRI || |
2042 | !(TRI->getSubRegIndexLaneMask(SubIdx: DefSubReg) & |
2043 | TRI->getSubRegIndexLaneMask(SubIdx: InsertedReg.SubIdx)).none()) |
2044 | return ValueTrackerResult(); |
2045 | // At this point, the value is available in v0 via the same subreg |
2046 | // we used for Def. |
2047 | return ValueTrackerResult(BaseReg.Reg, DefSubReg); |
2048 | } |
2049 | |
2050 | ValueTrackerResult ValueTracker::() { |
2051 | assert((Def->isExtractSubreg() || |
2052 | Def->isExtractSubregLike()) && "Invalid definition" ); |
2053 | // We are looking at: |
2054 | // Def = EXTRACT_SUBREG v0, sub0 |
2055 | |
2056 | // Bail if we have to compose sub registers. |
2057 | // Indeed, if DefSubReg != 0, we would have to compose it with sub0. |
2058 | if (DefSubReg) |
2059 | return ValueTrackerResult(); |
2060 | |
2061 | if (!TII) |
2062 | // We could handle the EXTRACT_SUBREG here, but we do not want to |
2063 | // duplicate the code from the generic TII. |
2064 | return ValueTrackerResult(); |
2065 | |
2066 | RegSubRegPairAndIdx ; |
2067 | if (!TII->getExtractSubregInputs(MI: *Def, DefIdx, InputReg&: ExtractSubregInputReg)) |
2068 | return ValueTrackerResult(); |
2069 | |
2070 | // Bail if we have to compose sub registers. |
2071 | // Likewise, if v0.subreg != 0, we would have to compose v0.subreg with sub0. |
2072 | if (ExtractSubregInputReg.SubReg) |
2073 | return ValueTrackerResult(); |
2074 | // Otherwise, the value is available in the v0.sub0. |
2075 | return ValueTrackerResult(ExtractSubregInputReg.Reg, |
2076 | ExtractSubregInputReg.SubIdx); |
2077 | } |
2078 | |
2079 | ValueTrackerResult ValueTracker::getNextSourceFromSubregToReg() { |
2080 | assert(Def->isSubregToReg() && "Invalid definition" ); |
2081 | // We are looking at: |
2082 | // Def = SUBREG_TO_REG Imm, v0, sub0 |
2083 | |
2084 | // Bail if we have to compose sub registers. |
2085 | // If DefSubReg != sub0, we would have to check that all the bits |
2086 | // we track are included in sub0 and if yes, we would have to |
2087 | // determine the right subreg in v0. |
2088 | if (DefSubReg != Def->getOperand(i: 3).getImm()) |
2089 | return ValueTrackerResult(); |
2090 | // Bail if we have to compose sub registers. |
2091 | // Likewise, if v0.subreg != 0, we would have to compose it with sub0. |
2092 | if (Def->getOperand(i: 2).getSubReg()) |
2093 | return ValueTrackerResult(); |
2094 | |
2095 | return ValueTrackerResult(Def->getOperand(i: 2).getReg(), |
2096 | Def->getOperand(i: 3).getImm()); |
2097 | } |
2098 | |
2099 | /// Explore each PHI incoming operand and return its sources. |
2100 | ValueTrackerResult ValueTracker::getNextSourceFromPHI() { |
2101 | assert(Def->isPHI() && "Invalid definition" ); |
2102 | ValueTrackerResult Res; |
2103 | |
2104 | // If we look for a different subreg, bail as we do not support composing |
2105 | // subregs yet. |
2106 | if (Def->getOperand(i: 0).getSubReg() != DefSubReg) |
2107 | return ValueTrackerResult(); |
2108 | |
2109 | // Return all register sources for PHI instructions. |
2110 | for (unsigned i = 1, e = Def->getNumOperands(); i < e; i += 2) { |
2111 | const MachineOperand &MO = Def->getOperand(i); |
2112 | assert(MO.isReg() && "Invalid PHI instruction" ); |
2113 | // We have no code to deal with undef operands. They shouldn't happen in |
2114 | // normal programs anyway. |
2115 | if (MO.isUndef()) |
2116 | return ValueTrackerResult(); |
2117 | Res.addSource(SrcReg: MO.getReg(), SrcSubReg: MO.getSubReg()); |
2118 | } |
2119 | |
2120 | return Res; |
2121 | } |
2122 | |
2123 | ValueTrackerResult ValueTracker::getNextSourceImpl() { |
2124 | assert(Def && "This method needs a valid definition" ); |
2125 | |
2126 | assert(((Def->getOperand(DefIdx).isDef() && |
2127 | (DefIdx < Def->getDesc().getNumDefs() || |
2128 | Def->getDesc().isVariadic())) || |
2129 | Def->getOperand(DefIdx).isImplicit()) && |
2130 | "Invalid DefIdx" ); |
2131 | if (Def->isCopy()) |
2132 | return getNextSourceFromCopy(); |
2133 | if (Def->isBitcast()) |
2134 | return getNextSourceFromBitcast(); |
2135 | // All the remaining cases involve "complex" instructions. |
2136 | // Bail if we did not ask for the advanced tracking. |
2137 | if (DisableAdvCopyOpt) |
2138 | return ValueTrackerResult(); |
2139 | if (Def->isRegSequence() || Def->isRegSequenceLike()) |
2140 | return getNextSourceFromRegSequence(); |
2141 | if (Def->isInsertSubreg() || Def->isInsertSubregLike()) |
2142 | return getNextSourceFromInsertSubreg(); |
2143 | if (Def->isExtractSubreg() || Def->isExtractSubregLike()) |
2144 | return getNextSourceFromExtractSubreg(); |
2145 | if (Def->isSubregToReg()) |
2146 | return getNextSourceFromSubregToReg(); |
2147 | if (Def->isPHI()) |
2148 | return getNextSourceFromPHI(); |
2149 | return ValueTrackerResult(); |
2150 | } |
2151 | |
2152 | ValueTrackerResult ValueTracker::getNextSource() { |
2153 | // If we reach a point where we cannot move up in the use-def chain, |
2154 | // there is nothing we can get. |
2155 | if (!Def) |
2156 | return ValueTrackerResult(); |
2157 | |
2158 | ValueTrackerResult Res = getNextSourceImpl(); |
2159 | if (Res.isValid()) { |
2160 | // Update definition, definition index, and subregister for the |
2161 | // next call of getNextSource. |
2162 | // Update the current register. |
2163 | bool OneRegSrc = Res.getNumSources() == 1; |
2164 | if (OneRegSrc) |
2165 | Reg = Res.getSrcReg(Idx: 0); |
2166 | // Update the result before moving up in the use-def chain |
2167 | // with the instruction containing the last found sources. |
2168 | Res.setInst(Def); |
2169 | |
2170 | // If we can still move up in the use-def chain, move to the next |
2171 | // definition. |
2172 | if (!Reg.isPhysical() && OneRegSrc) { |
2173 | MachineRegisterInfo::def_iterator DI = MRI.def_begin(RegNo: Reg); |
2174 | if (DI != MRI.def_end()) { |
2175 | Def = DI->getParent(); |
2176 | DefIdx = DI.getOperandNo(); |
2177 | DefSubReg = Res.getSrcSubReg(Idx: 0); |
2178 | } else { |
2179 | Def = nullptr; |
2180 | } |
2181 | return Res; |
2182 | } |
2183 | } |
2184 | // If we end up here, this means we will not be able to find another source |
2185 | // for the next iteration. Make sure any new call to getNextSource bails out |
2186 | // early by cutting the use-def chain. |
2187 | Def = nullptr; |
2188 | return Res; |
2189 | } |
2190 | |