Revert r354244 "[DAGCombiner] Eliminate dead stores to stack."
[llvm-complete.git] / lib / CodeGen / PeepholeOptimizer.cpp
blobf5277490ae61f974fba728d0c5c5b55a633990ba
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:
11 // - Optimize Extensions
13 // Optimization of sign / zero extension instructions. It may be extended to
14 // handle other instructions with similar properties.
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.
21 // - Optimize Comparisons
23 // Optimization of comparison instructions. For instance, in this code:
25 // sub r1, 1
26 // cmp r1, 0
27 // bz L1
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.
33 // Another instance, in this code:
35 // sub r1, r3 | sub r1, imm
36 // cmp r3, r1 or cmp r1, r3 | cmp r1, imm
37 // bge L1
39 // If the branch instruction can use flag from "sub", then we can replace
40 // "sub" with "subs" and eliminate the "cmp" instruction.
42 // - Optimize Loads:
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.
48 // - Optimize Copies and Bitcast (more generally, target specific copies):
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
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 //===----------------------------------------------------------------------===//
68 #include "llvm/ADT/DenseMap.h"
69 #include "llvm/ADT/Optional.h"
70 #include "llvm/ADT/SmallPtrSet.h"
71 #include "llvm/ADT/SmallSet.h"
72 #include "llvm/ADT/SmallVector.h"
73 #include "llvm/ADT/Statistic.h"
74 #include "llvm/CodeGen/MachineBasicBlock.h"
75 #include "llvm/CodeGen/MachineDominators.h"
76 #include "llvm/CodeGen/MachineFunction.h"
77 #include "llvm/CodeGen/MachineFunctionPass.h"
78 #include "llvm/CodeGen/MachineInstr.h"
79 #include "llvm/CodeGen/MachineInstrBuilder.h"
80 #include "llvm/CodeGen/MachineLoopInfo.h"
81 #include "llvm/CodeGen/MachineOperand.h"
82 #include "llvm/CodeGen/MachineRegisterInfo.h"
83 #include "llvm/CodeGen/TargetInstrInfo.h"
84 #include "llvm/CodeGen/TargetOpcodes.h"
85 #include "llvm/CodeGen/TargetRegisterInfo.h"
86 #include "llvm/CodeGen/TargetSubtargetInfo.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/ErrorHandling.h"
93 #include "llvm/Support/raw_ostream.h"
94 #include <cassert>
95 #include <cstdint>
96 #include <memory>
97 #include <utility>
99 using namespace llvm;
100 using RegSubRegPair = TargetInstrInfo::RegSubRegPair;
101 using RegSubRegPairAndIdx = TargetInstrInfo::RegSubRegPairAndIdx;
103 #define DEBUG_TYPE "peephole-opt"
105 // Optimize Extensions
106 static cl::opt<bool>
107 Aggressive("aggressive-ext-opt", cl::Hidden,
108 cl::desc("Aggressive extension optimization"));
110 static cl::opt<bool>
111 DisablePeephole("disable-peephole", cl::Hidden, cl::init(false),
112 cl::desc("Disable the peephole optimizer"));
114 /// Specifiy whether or not the value tracking looks through
115 /// complex instructions. When this is true, the value tracker
116 /// bails on everything that is not a copy or a bitcast.
117 static cl::opt<bool>
118 DisableAdvCopyOpt("disable-adv-copy-opt", cl::Hidden, cl::init(false),
119 cl::desc("Disable advanced copy optimization"));
121 static cl::opt<bool> DisableNAPhysCopyOpt(
122 "disable-non-allocatable-phys-copy-opt", cl::Hidden, cl::init(false),
123 cl::desc("Disable non-allocatable physical register copy optimization"));
125 // Limit the number of PHI instructions to process
126 // in PeepholeOptimizer::getNextSource.
127 static cl::opt<unsigned> RewritePHILimit(
128 "rewrite-phi-limit", cl::Hidden, cl::init(10),
129 cl::desc("Limit the length of PHI chains to lookup"));
131 // Limit the length of recurrence chain when evaluating the benefit of
132 // commuting operands.
133 static cl::opt<unsigned> MaxRecurrenceChain(
134 "recurrence-chain-limit", cl::Hidden, cl::init(3),
135 cl::desc("Maximum length of recurrence chain when evaluating the benefit "
136 "of commuting operands"));
139 STATISTIC(NumReuse, "Number of extension results reused");
140 STATISTIC(NumCmps, "Number of compares eliminated");
141 STATISTIC(NumImmFold, "Number of move immediate folded");
142 STATISTIC(NumLoadFold, "Number of loads folded");
143 STATISTIC(NumSelects, "Number of selects optimized");
144 STATISTIC(NumUncoalescableCopies, "Number of uncoalescable copies optimized");
145 STATISTIC(NumRewrittenCopies, "Number of copies rewritten");
146 STATISTIC(NumNAPhysCopies, "Number of non-allocatable physical copies removed");
148 namespace {
150 class ValueTrackerResult;
151 class RecurrenceInstr;
153 class PeepholeOptimizer : public MachineFunctionPass {
154 const TargetInstrInfo *TII;
155 const TargetRegisterInfo *TRI;
156 MachineRegisterInfo *MRI;
157 MachineDominatorTree *DT; // Machine dominator tree
158 MachineLoopInfo *MLI;
160 public:
161 static char ID; // Pass identification
163 PeepholeOptimizer() : MachineFunctionPass(ID) {
164 initializePeepholeOptimizerPass(*PassRegistry::getPassRegistry());
167 bool runOnMachineFunction(MachineFunction &MF) override;
169 void getAnalysisUsage(AnalysisUsage &AU) const override {
170 AU.setPreservesCFG();
171 MachineFunctionPass::getAnalysisUsage(AU);
172 AU.addRequired<MachineLoopInfo>();
173 AU.addPreserved<MachineLoopInfo>();
174 if (Aggressive) {
175 AU.addRequired<MachineDominatorTree>();
176 AU.addPreserved<MachineDominatorTree>();
180 /// Track Def -> Use info used for rewriting copies.
181 using RewriteMapTy = SmallDenseMap<RegSubRegPair, ValueTrackerResult>;
183 /// Sequence of instructions that formulate recurrence cycle.
184 using RecurrenceCycle = SmallVector<RecurrenceInstr, 4>;
186 private:
187 bool optimizeCmpInstr(MachineInstr &MI);
188 bool optimizeExtInstr(MachineInstr &MI, MachineBasicBlock &MBB,
189 SmallPtrSetImpl<MachineInstr*> &LocalMIs);
190 bool optimizeSelect(MachineInstr &MI,
191 SmallPtrSetImpl<MachineInstr *> &LocalMIs);
192 bool optimizeCondBranch(MachineInstr &MI);
193 bool optimizeCoalescableCopy(MachineInstr &MI);
194 bool optimizeUncoalescableCopy(MachineInstr &MI,
195 SmallPtrSetImpl<MachineInstr *> &LocalMIs);
196 bool optimizeRecurrence(MachineInstr &PHI);
197 bool findNextSource(RegSubRegPair RegSubReg, RewriteMapTy &RewriteMap);
198 bool isMoveImmediate(MachineInstr &MI,
199 SmallSet<unsigned, 4> &ImmDefRegs,
200 DenseMap<unsigned, MachineInstr*> &ImmDefMIs);
201 bool foldImmediate(MachineInstr &MI, SmallSet<unsigned, 4> &ImmDefRegs,
202 DenseMap<unsigned, MachineInstr*> &ImmDefMIs);
204 /// Finds recurrence cycles, but only ones that formulated around
205 /// a def operand and a use operand that are tied. If there is a use
206 /// operand commutable with the tied use operand, find recurrence cycle
207 /// along that operand as well.
208 bool findTargetRecurrence(unsigned Reg,
209 const SmallSet<unsigned, 2> &TargetReg,
210 RecurrenceCycle &RC);
212 /// If copy instruction \p MI is a virtual register copy, track it in
213 /// the set \p CopySrcRegs and \p CopyMIs. If this virtual register was
214 /// previously seen as a copy, replace the uses of this copy with the
215 /// previously seen copy's destination register.
216 bool foldRedundantCopy(MachineInstr &MI,
217 SmallSet<unsigned, 4> &CopySrcRegs,
218 DenseMap<unsigned, MachineInstr *> &CopyMIs);
220 /// Is the register \p Reg a non-allocatable physical register?
221 bool isNAPhysCopy(unsigned Reg);
223 /// If copy instruction \p MI is a non-allocatable virtual<->physical
224 /// register copy, track it in the \p NAPhysToVirtMIs map. If this
225 /// non-allocatable physical register was previously copied to a virtual
226 /// registered and hasn't been clobbered, the virt->phys copy can be
227 /// deleted.
228 bool foldRedundantNAPhysCopy(MachineInstr &MI,
229 DenseMap<unsigned, MachineInstr *> &NAPhysToVirtMIs);
231 bool isLoadFoldable(MachineInstr &MI,
232 SmallSet<unsigned, 16> &FoldAsLoadDefCandidates);
234 /// Check whether \p MI is understood by the register coalescer
235 /// but may require some rewriting.
236 bool isCoalescableCopy(const MachineInstr &MI) {
237 // SubregToRegs are not interesting, because they are already register
238 // coalescer friendly.
239 return MI.isCopy() || (!DisableAdvCopyOpt &&
240 (MI.isRegSequence() || MI.isInsertSubreg() ||
241 MI.isExtractSubreg()));
244 /// Check whether \p MI is a copy like instruction that is
245 /// not recognized by the register coalescer.
246 bool isUncoalescableCopy(const MachineInstr &MI) {
247 return MI.isBitcast() ||
248 (!DisableAdvCopyOpt &&
249 (MI.isRegSequenceLike() || MI.isInsertSubregLike() ||
250 MI.isExtractSubregLike()));
253 MachineInstr &rewriteSource(MachineInstr &CopyLike,
254 RegSubRegPair Def, RewriteMapTy &RewriteMap);
257 /// Helper class to hold instructions that are inside recurrence cycles.
258 /// The recurrence cycle is formulated around 1) a def operand and its
259 /// tied use operand, or 2) a def operand and a use operand that is commutable
260 /// with another use operand which is tied to the def operand. In the latter
261 /// case, index of the tied use operand and the commutable use operand are
262 /// maintained with CommutePair.
263 class RecurrenceInstr {
264 public:
265 using IndexPair = std::pair<unsigned, unsigned>;
267 RecurrenceInstr(MachineInstr *MI) : MI(MI) {}
268 RecurrenceInstr(MachineInstr *MI, unsigned Idx1, unsigned Idx2)
269 : MI(MI), CommutePair(std::make_pair(Idx1, Idx2)) {}
271 MachineInstr *getMI() const { return MI; }
272 Optional<IndexPair> getCommutePair() const { return CommutePair; }
274 private:
275 MachineInstr *MI;
276 Optional<IndexPair> CommutePair;
279 /// Helper class to hold a reply for ValueTracker queries.
280 /// Contains the returned sources for a given search and the instructions
281 /// where the sources were tracked from.
282 class ValueTrackerResult {
283 private:
284 /// Track all sources found by one ValueTracker query.
285 SmallVector<RegSubRegPair, 2> RegSrcs;
287 /// Instruction using the sources in 'RegSrcs'.
288 const MachineInstr *Inst = nullptr;
290 public:
291 ValueTrackerResult() = default;
293 ValueTrackerResult(unsigned Reg, unsigned SubReg) {
294 addSource(Reg, SubReg);
297 bool isValid() const { return getNumSources() > 0; }
299 void setInst(const MachineInstr *I) { Inst = I; }
300 const MachineInstr *getInst() const { return Inst; }
302 void clear() {
303 RegSrcs.clear();
304 Inst = nullptr;
307 void addSource(unsigned SrcReg, unsigned SrcSubReg) {
308 RegSrcs.push_back(RegSubRegPair(SrcReg, SrcSubReg));
311 void setSource(int Idx, unsigned SrcReg, unsigned SrcSubReg) {
312 assert(Idx < getNumSources() && "Reg pair source out of index");
313 RegSrcs[Idx] = RegSubRegPair(SrcReg, SrcSubReg);
316 int getNumSources() const { return RegSrcs.size(); }
318 RegSubRegPair getSrc(int Idx) const {
319 return RegSrcs[Idx];
322 unsigned getSrcReg(int Idx) const {
323 assert(Idx < getNumSources() && "Reg source out of index");
324 return RegSrcs[Idx].Reg;
327 unsigned getSrcSubReg(int Idx) const {
328 assert(Idx < getNumSources() && "SubReg source out of index");
329 return RegSrcs[Idx].SubReg;
332 bool operator==(const ValueTrackerResult &Other) {
333 if (Other.getInst() != getInst())
334 return false;
336 if (Other.getNumSources() != getNumSources())
337 return false;
339 for (int i = 0, e = Other.getNumSources(); i != e; ++i)
340 if (Other.getSrcReg(i) != getSrcReg(i) ||
341 Other.getSrcSubReg(i) != getSrcSubReg(i))
342 return false;
343 return true;
347 /// Helper class to track the possible sources of a value defined by
348 /// a (chain of) copy related instructions.
349 /// Given a definition (instruction and definition index), this class
350 /// follows the use-def chain to find successive suitable sources.
351 /// The given source can be used to rewrite the definition into
352 /// def = COPY src.
354 /// For instance, let us consider the following snippet:
355 /// v0 =
356 /// v2 = INSERT_SUBREG v1, v0, sub0
357 /// def = COPY v2.sub0
359 /// Using a ValueTracker for def = COPY v2.sub0 will give the following
360 /// suitable sources:
361 /// v2.sub0 and v0.
362 /// Then, def can be rewritten into def = COPY v0.
363 class ValueTracker {
364 private:
365 /// The current point into the use-def chain.
366 const MachineInstr *Def = nullptr;
368 /// The index of the definition in Def.
369 unsigned DefIdx = 0;
371 /// The sub register index of the definition.
372 unsigned DefSubReg;
374 /// The register where the value can be found.
375 unsigned Reg;
377 /// MachineRegisterInfo used to perform tracking.
378 const MachineRegisterInfo &MRI;
380 /// Optional TargetInstrInfo used to perform some complex tracking.
381 const TargetInstrInfo *TII;
383 /// Dispatcher to the right underlying implementation of getNextSource.
384 ValueTrackerResult getNextSourceImpl();
386 /// Specialized version of getNextSource for Copy instructions.
387 ValueTrackerResult getNextSourceFromCopy();
389 /// Specialized version of getNextSource for Bitcast instructions.
390 ValueTrackerResult getNextSourceFromBitcast();
392 /// Specialized version of getNextSource for RegSequence instructions.
393 ValueTrackerResult getNextSourceFromRegSequence();
395 /// Specialized version of getNextSource for InsertSubreg instructions.
396 ValueTrackerResult getNextSourceFromInsertSubreg();
398 /// Specialized version of getNextSource for ExtractSubreg instructions.
399 ValueTrackerResult getNextSourceFromExtractSubreg();
401 /// Specialized version of getNextSource for SubregToReg instructions.
402 ValueTrackerResult getNextSourceFromSubregToReg();
404 /// Specialized version of getNextSource for PHI instructions.
405 ValueTrackerResult getNextSourceFromPHI();
407 public:
408 /// Create a ValueTracker instance for the value defined by \p Reg.
409 /// \p DefSubReg represents the sub register index the value tracker will
410 /// track. It does not need to match the sub register index used in the
411 /// definition of \p Reg.
412 /// If \p Reg is a physical register, a value tracker constructed with
413 /// this constructor will not find any alternative source.
414 /// Indeed, when \p Reg is a physical register that constructor does not
415 /// know which definition of \p Reg it should track.
416 /// Use the next constructor to track a physical register.
417 ValueTracker(unsigned Reg, unsigned DefSubReg,
418 const MachineRegisterInfo &MRI,
419 const TargetInstrInfo *TII = nullptr)
420 : DefSubReg(DefSubReg), Reg(Reg), MRI(MRI), TII(TII) {
421 if (!TargetRegisterInfo::isPhysicalRegister(Reg)) {
422 Def = MRI.getVRegDef(Reg);
423 DefIdx = MRI.def_begin(Reg).getOperandNo();
427 /// Following the use-def chain, get the next available source
428 /// for the tracked value.
429 /// \return A ValueTrackerResult containing a set of registers
430 /// and sub registers with tracked values. A ValueTrackerResult with
431 /// an empty set of registers means no source was found.
432 ValueTrackerResult getNextSource();
435 } // end anonymous namespace
437 char PeepholeOptimizer::ID = 0;
439 char &llvm::PeepholeOptimizerID = PeepholeOptimizer::ID;
441 INITIALIZE_PASS_BEGIN(PeepholeOptimizer, DEBUG_TYPE,
442 "Peephole Optimizations", false, false)
443 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
444 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
445 INITIALIZE_PASS_END(PeepholeOptimizer, DEBUG_TYPE,
446 "Peephole Optimizations", false, false)
448 /// If instruction is a copy-like instruction, i.e. it reads a single register
449 /// and writes a single register and it does not modify the source, and if the
450 /// source value is preserved as a sub-register of the result, then replace all
451 /// reachable uses of the source with the subreg of the result.
453 /// Do not generate an EXTRACT that is used only in a debug use, as this changes
454 /// the code. Since this code does not currently share EXTRACTs, just ignore all
455 /// debug uses.
456 bool PeepholeOptimizer::
457 optimizeExtInstr(MachineInstr &MI, MachineBasicBlock &MBB,
458 SmallPtrSetImpl<MachineInstr*> &LocalMIs) {
459 unsigned SrcReg, DstReg, SubIdx;
460 if (!TII->isCoalescableExtInstr(MI, SrcReg, DstReg, SubIdx))
461 return false;
463 if (TargetRegisterInfo::isPhysicalRegister(DstReg) ||
464 TargetRegisterInfo::isPhysicalRegister(SrcReg))
465 return false;
467 if (MRI->hasOneNonDBGUse(SrcReg))
468 // No other uses.
469 return false;
471 // Ensure DstReg can get a register class that actually supports
472 // sub-registers. Don't change the class until we commit.
473 const TargetRegisterClass *DstRC = MRI->getRegClass(DstReg);
474 DstRC = TRI->getSubClassWithSubReg(DstRC, SubIdx);
475 if (!DstRC)
476 return false;
478 // The ext instr may be operating on a sub-register of SrcReg as well.
479 // PPC::EXTSW is a 32 -> 64-bit sign extension, but it reads a 64-bit
480 // register.
481 // If UseSrcSubIdx is Set, SubIdx also applies to SrcReg, and only uses of
482 // SrcReg:SubIdx should be replaced.
483 bool UseSrcSubIdx =
484 TRI->getSubClassWithSubReg(MRI->getRegClass(SrcReg), SubIdx) != nullptr;
486 // The source has other uses. See if we can replace the other uses with use of
487 // the result of the extension.
488 SmallPtrSet<MachineBasicBlock*, 4> ReachedBBs;
489 for (MachineInstr &UI : MRI->use_nodbg_instructions(DstReg))
490 ReachedBBs.insert(UI.getParent());
492 // Uses that are in the same BB of uses of the result of the instruction.
493 SmallVector<MachineOperand*, 8> Uses;
495 // Uses that the result of the instruction can reach.
496 SmallVector<MachineOperand*, 8> ExtendedUses;
498 bool ExtendLife = true;
499 for (MachineOperand &UseMO : MRI->use_nodbg_operands(SrcReg)) {
500 MachineInstr *UseMI = UseMO.getParent();
501 if (UseMI == &MI)
502 continue;
504 if (UseMI->isPHI()) {
505 ExtendLife = false;
506 continue;
509 // Only accept uses of SrcReg:SubIdx.
510 if (UseSrcSubIdx && UseMO.getSubReg() != SubIdx)
511 continue;
513 // It's an error to translate this:
515 // %reg1025 = <sext> %reg1024
516 // ...
517 // %reg1026 = SUBREG_TO_REG 0, %reg1024, 4
519 // into this:
521 // %reg1025 = <sext> %reg1024
522 // ...
523 // %reg1027 = COPY %reg1025:4
524 // %reg1026 = SUBREG_TO_REG 0, %reg1027, 4
526 // The problem here is that SUBREG_TO_REG is there to assert that an
527 // implicit zext occurs. It doesn't insert a zext instruction. If we allow
528 // the COPY here, it will give us the value after the <sext>, not the
529 // original value of %reg1024 before <sext>.
530 if (UseMI->getOpcode() == TargetOpcode::SUBREG_TO_REG)
531 continue;
533 MachineBasicBlock *UseMBB = UseMI->getParent();
534 if (UseMBB == &MBB) {
535 // Local uses that come after the extension.
536 if (!LocalMIs.count(UseMI))
537 Uses.push_back(&UseMO);
538 } else if (ReachedBBs.count(UseMBB)) {
539 // Non-local uses where the result of the extension is used. Always
540 // replace these unless it's a PHI.
541 Uses.push_back(&UseMO);
542 } else if (Aggressive && DT->dominates(&MBB, UseMBB)) {
543 // We may want to extend the live range of the extension result in order
544 // to replace these uses.
545 ExtendedUses.push_back(&UseMO);
546 } else {
547 // Both will be live out of the def MBB anyway. Don't extend live range of
548 // the extension result.
549 ExtendLife = false;
550 break;
554 if (ExtendLife && !ExtendedUses.empty())
555 // Extend the liveness of the extension result.
556 Uses.append(ExtendedUses.begin(), ExtendedUses.end());
558 // Now replace all uses.
559 bool Changed = false;
560 if (!Uses.empty()) {
561 SmallPtrSet<MachineBasicBlock*, 4> PHIBBs;
563 // Look for PHI uses of the extended result, we don't want to extend the
564 // liveness of a PHI input. It breaks all kinds of assumptions down
565 // stream. A PHI use is expected to be the kill of its source values.
566 for (MachineInstr &UI : MRI->use_nodbg_instructions(DstReg))
567 if (UI.isPHI())
568 PHIBBs.insert(UI.getParent());
570 const TargetRegisterClass *RC = MRI->getRegClass(SrcReg);
571 for (unsigned i = 0, e = Uses.size(); i != e; ++i) {
572 MachineOperand *UseMO = Uses[i];
573 MachineInstr *UseMI = UseMO->getParent();
574 MachineBasicBlock *UseMBB = UseMI->getParent();
575 if (PHIBBs.count(UseMBB))
576 continue;
578 // About to add uses of DstReg, clear DstReg's kill flags.
579 if (!Changed) {
580 MRI->clearKillFlags(DstReg);
581 MRI->constrainRegClass(DstReg, DstRC);
584 unsigned NewVR = MRI->createVirtualRegister(RC);
585 MachineInstr *Copy = BuildMI(*UseMBB, UseMI, UseMI->getDebugLoc(),
586 TII->get(TargetOpcode::COPY), NewVR)
587 .addReg(DstReg, 0, SubIdx);
588 // SubIdx applies to both SrcReg and DstReg when UseSrcSubIdx is set.
589 if (UseSrcSubIdx) {
590 Copy->getOperand(0).setSubReg(SubIdx);
591 Copy->getOperand(0).setIsUndef();
593 UseMO->setReg(NewVR);
594 ++NumReuse;
595 Changed = true;
599 return Changed;
602 /// If the instruction is a compare and the previous instruction it's comparing
603 /// against already sets (or could be modified to set) the same flag as the
604 /// compare, then we can remove the comparison and use the flag from the
605 /// previous instruction.
606 bool PeepholeOptimizer::optimizeCmpInstr(MachineInstr &MI) {
607 // If this instruction is a comparison against zero and isn't comparing a
608 // physical register, we can try to optimize it.
609 unsigned SrcReg, SrcReg2;
610 int CmpMask, CmpValue;
611 if (!TII->analyzeCompare(MI, SrcReg, SrcReg2, CmpMask, CmpValue) ||
612 TargetRegisterInfo::isPhysicalRegister(SrcReg) ||
613 (SrcReg2 != 0 && TargetRegisterInfo::isPhysicalRegister(SrcReg2)))
614 return false;
616 // Attempt to optimize the comparison instruction.
617 if (TII->optimizeCompareInstr(MI, SrcReg, SrcReg2, CmpMask, CmpValue, MRI)) {
618 ++NumCmps;
619 return true;
622 return false;
625 /// Optimize a select instruction.
626 bool PeepholeOptimizer::optimizeSelect(MachineInstr &MI,
627 SmallPtrSetImpl<MachineInstr *> &LocalMIs) {
628 unsigned TrueOp = 0;
629 unsigned FalseOp = 0;
630 bool Optimizable = false;
631 SmallVector<MachineOperand, 4> Cond;
632 if (TII->analyzeSelect(MI, Cond, TrueOp, FalseOp, Optimizable))
633 return false;
634 if (!Optimizable)
635 return false;
636 if (!TII->optimizeSelect(MI, LocalMIs))
637 return false;
638 MI.eraseFromParent();
639 ++NumSelects;
640 return true;
643 /// Check if a simpler conditional branch can be generated.
644 bool PeepholeOptimizer::optimizeCondBranch(MachineInstr &MI) {
645 return TII->optimizeCondBranch(MI);
648 /// Try to find the next source that share the same register file
649 /// for the value defined by \p Reg and \p SubReg.
650 /// When true is returned, the \p RewriteMap can be used by the client to
651 /// retrieve all Def -> Use along the way up to the next source. Any found
652 /// Use that is not itself a key for another entry, is the next source to
653 /// use. During the search for the next source, multiple sources can be found
654 /// given multiple incoming sources of a PHI instruction. In this case, we
655 /// look in each PHI source for the next source; all found next sources must
656 /// share the same register file as \p Reg and \p SubReg. The client should
657 /// then be capable to rewrite all intermediate PHIs to get the next source.
658 /// \return False if no alternative sources are available. True otherwise.
659 bool PeepholeOptimizer::findNextSource(RegSubRegPair RegSubReg,
660 RewriteMapTy &RewriteMap) {
661 // Do not try to find a new source for a physical register.
662 // So far we do not have any motivating example for doing that.
663 // Thus, instead of maintaining untested code, we will revisit that if
664 // that changes at some point.
665 unsigned Reg = RegSubReg.Reg;
666 if (TargetRegisterInfo::isPhysicalRegister(Reg))
667 return false;
668 const TargetRegisterClass *DefRC = MRI->getRegClass(Reg);
670 SmallVector<RegSubRegPair, 4> SrcToLook;
671 RegSubRegPair CurSrcPair = RegSubReg;
672 SrcToLook.push_back(CurSrcPair);
674 unsigned PHICount = 0;
675 do {
676 CurSrcPair = SrcToLook.pop_back_val();
677 // As explained above, do not handle physical registers
678 if (TargetRegisterInfo::isPhysicalRegister(CurSrcPair.Reg))
679 return false;
681 ValueTracker ValTracker(CurSrcPair.Reg, CurSrcPair.SubReg, *MRI, TII);
683 // Follow the chain of copies until we find a more suitable source, a phi
684 // or have to abort.
685 while (true) {
686 ValueTrackerResult Res = ValTracker.getNextSource();
687 // Abort at the end of a chain (without finding a suitable source).
688 if (!Res.isValid())
689 return false;
691 // Insert the Def -> Use entry for the recently found source.
692 ValueTrackerResult CurSrcRes = RewriteMap.lookup(CurSrcPair);
693 if (CurSrcRes.isValid()) {
694 assert(CurSrcRes == Res && "ValueTrackerResult found must match");
695 // An existent entry with multiple sources is a PHI cycle we must avoid.
696 // Otherwise it's an entry with a valid next source we already found.
697 if (CurSrcRes.getNumSources() > 1) {
698 LLVM_DEBUG(dbgs()
699 << "findNextSource: found PHI cycle, aborting...\n");
700 return false;
702 break;
704 RewriteMap.insert(std::make_pair(CurSrcPair, Res));
706 // ValueTrackerResult usually have one source unless it's the result from
707 // a PHI instruction. Add the found PHI edges to be looked up further.
708 unsigned NumSrcs = Res.getNumSources();
709 if (NumSrcs > 1) {
710 PHICount++;
711 if (PHICount >= RewritePHILimit) {
712 LLVM_DEBUG(dbgs() << "findNextSource: PHI limit reached\n");
713 return false;
716 for (unsigned i = 0; i < NumSrcs; ++i)
717 SrcToLook.push_back(Res.getSrc(i));
718 break;
721 CurSrcPair = Res.getSrc(0);
722 // Do not extend the live-ranges of physical registers as they add
723 // constraints to the register allocator. Moreover, if we want to extend
724 // the live-range of a physical register, unlike SSA virtual register,
725 // we will have to check that they aren't redefine before the related use.
726 if (TargetRegisterInfo::isPhysicalRegister(CurSrcPair.Reg))
727 return false;
729 // Keep following the chain if the value isn't any better yet.
730 const TargetRegisterClass *SrcRC = MRI->getRegClass(CurSrcPair.Reg);
731 if (!TRI->shouldRewriteCopySrc(DefRC, RegSubReg.SubReg, SrcRC,
732 CurSrcPair.SubReg))
733 continue;
735 // We currently cannot deal with subreg operands on PHI instructions
736 // (see insertPHI()).
737 if (PHICount > 0 && CurSrcPair.SubReg != 0)
738 continue;
740 // We found a suitable source, and are done with this chain.
741 break;
743 } while (!SrcToLook.empty());
745 // If we did not find a more suitable source, there is nothing to optimize.
746 return CurSrcPair.Reg != Reg;
749 /// Insert a PHI instruction with incoming edges \p SrcRegs that are
750 /// guaranteed to have the same register class. This is necessary whenever we
751 /// successfully traverse a PHI instruction and find suitable sources coming
752 /// from its edges. By inserting a new PHI, we provide a rewritten PHI def
753 /// suitable to be used in a new COPY instruction.
754 static MachineInstr &
755 insertPHI(MachineRegisterInfo &MRI, const TargetInstrInfo &TII,
756 const SmallVectorImpl<RegSubRegPair> &SrcRegs,
757 MachineInstr &OrigPHI) {
758 assert(!SrcRegs.empty() && "No sources to create a PHI instruction?");
760 const TargetRegisterClass *NewRC = MRI.getRegClass(SrcRegs[0].Reg);
761 // NewRC is only correct if no subregisters are involved. findNextSource()
762 // should have rejected those cases already.
763 assert(SrcRegs[0].SubReg == 0 && "should not have subreg operand");
764 unsigned NewVR = MRI.createVirtualRegister(NewRC);
765 MachineBasicBlock *MBB = OrigPHI.getParent();
766 MachineInstrBuilder MIB = BuildMI(*MBB, &OrigPHI, OrigPHI.getDebugLoc(),
767 TII.get(TargetOpcode::PHI), NewVR);
769 unsigned MBBOpIdx = 2;
770 for (const RegSubRegPair &RegPair : SrcRegs) {
771 MIB.addReg(RegPair.Reg, 0, RegPair.SubReg);
772 MIB.addMBB(OrigPHI.getOperand(MBBOpIdx).getMBB());
773 // Since we're extended the lifetime of RegPair.Reg, clear the
774 // kill flags to account for that and make RegPair.Reg reaches
775 // the new PHI.
776 MRI.clearKillFlags(RegPair.Reg);
777 MBBOpIdx += 2;
780 return *MIB;
783 namespace {
785 /// Interface to query instructions amenable to copy rewriting.
786 class Rewriter {
787 protected:
788 MachineInstr &CopyLike;
789 unsigned CurrentSrcIdx = 0; ///< The index of the source being rewritten.
790 public:
791 Rewriter(MachineInstr &CopyLike) : CopyLike(CopyLike) {}
792 virtual ~Rewriter() {}
794 /// Get the next rewritable source (SrcReg, SrcSubReg) and
795 /// the related value that it affects (DstReg, DstSubReg).
796 /// A source is considered rewritable if its register class and the
797 /// register class of the related DstReg may not be register
798 /// coalescer friendly. In other words, given a copy-like instruction
799 /// not all the arguments may be returned at rewritable source, since
800 /// some arguments are none to be register coalescer friendly.
802 /// Each call of this method moves the current source to the next
803 /// rewritable source.
804 /// For instance, let CopyLike be the instruction to rewrite.
805 /// CopyLike has one definition and one source:
806 /// dst.dstSubIdx = CopyLike src.srcSubIdx.
808 /// The first call will give the first rewritable source, i.e.,
809 /// the only source this instruction has:
810 /// (SrcReg, SrcSubReg) = (src, srcSubIdx).
811 /// This source defines the whole definition, i.e.,
812 /// (DstReg, DstSubReg) = (dst, dstSubIdx).
814 /// The second and subsequent calls will return false, as there is only one
815 /// rewritable source.
817 /// \return True if a rewritable source has been found, false otherwise.
818 /// The output arguments are valid if and only if true is returned.
819 virtual bool getNextRewritableSource(RegSubRegPair &Src,
820 RegSubRegPair &Dst) = 0;
822 /// Rewrite the current source with \p NewReg and \p NewSubReg if possible.
823 /// \return True if the rewriting was possible, false otherwise.
824 virtual bool RewriteCurrentSource(unsigned NewReg, unsigned NewSubReg) = 0;
827 /// Rewriter for COPY instructions.
828 class CopyRewriter : public Rewriter {
829 public:
830 CopyRewriter(MachineInstr &MI) : Rewriter(MI) {
831 assert(MI.isCopy() && "Expected copy instruction");
833 virtual ~CopyRewriter() = default;
835 bool getNextRewritableSource(RegSubRegPair &Src,
836 RegSubRegPair &Dst) override {
837 // CurrentSrcIdx > 0 means this function has already been called.
838 if (CurrentSrcIdx > 0)
839 return false;
840 // This is the first call to getNextRewritableSource.
841 // Move the CurrentSrcIdx to remember that we made that call.
842 CurrentSrcIdx = 1;
843 // The rewritable source is the argument.
844 const MachineOperand &MOSrc = CopyLike.getOperand(1);
845 Src = RegSubRegPair(MOSrc.getReg(), MOSrc.getSubReg());
846 // What we track are the alternative sources of the definition.
847 const MachineOperand &MODef = CopyLike.getOperand(0);
848 Dst = RegSubRegPair(MODef.getReg(), MODef.getSubReg());
849 return true;
852 bool RewriteCurrentSource(unsigned NewReg, unsigned NewSubReg) override {
853 if (CurrentSrcIdx != 1)
854 return false;
855 MachineOperand &MOSrc = CopyLike.getOperand(CurrentSrcIdx);
856 MOSrc.setReg(NewReg);
857 MOSrc.setSubReg(NewSubReg);
858 return true;
862 /// Helper class to rewrite uncoalescable copy like instructions
863 /// into new COPY (coalescable friendly) instructions.
864 class UncoalescableRewriter : public Rewriter {
865 unsigned NumDefs; ///< Number of defs in the bitcast.
867 public:
868 UncoalescableRewriter(MachineInstr &MI) : Rewriter(MI) {
869 NumDefs = MI.getDesc().getNumDefs();
872 /// \see See Rewriter::getNextRewritableSource()
873 /// All such sources need to be considered rewritable in order to
874 /// rewrite a uncoalescable copy-like instruction. This method return
875 /// each definition that must be checked if rewritable.
876 bool getNextRewritableSource(RegSubRegPair &Src,
877 RegSubRegPair &Dst) override {
878 // Find the next non-dead definition and continue from there.
879 if (CurrentSrcIdx == NumDefs)
880 return false;
882 while (CopyLike.getOperand(CurrentSrcIdx).isDead()) {
883 ++CurrentSrcIdx;
884 if (CurrentSrcIdx == NumDefs)
885 return false;
888 // What we track are the alternative sources of the definition.
889 Src = RegSubRegPair(0, 0);
890 const MachineOperand &MODef = CopyLike.getOperand(CurrentSrcIdx);
891 Dst = RegSubRegPair(MODef.getReg(), MODef.getSubReg());
893 CurrentSrcIdx++;
894 return true;
897 bool RewriteCurrentSource(unsigned NewReg, unsigned NewSubReg) override {
898 return false;
902 /// Specialized rewriter for INSERT_SUBREG instruction.
903 class InsertSubregRewriter : public Rewriter {
904 public:
905 InsertSubregRewriter(MachineInstr &MI) : Rewriter(MI) {
906 assert(MI.isInsertSubreg() && "Invalid instruction");
909 /// \see See Rewriter::getNextRewritableSource()
910 /// Here CopyLike has the following form:
911 /// dst = INSERT_SUBREG Src1, Src2.src2SubIdx, subIdx.
912 /// Src1 has the same register class has dst, hence, there is
913 /// nothing to rewrite.
914 /// Src2.src2SubIdx, may not be register coalescer friendly.
915 /// Therefore, the first call to this method returns:
916 /// (SrcReg, SrcSubReg) = (Src2, src2SubIdx).
917 /// (DstReg, DstSubReg) = (dst, subIdx).
919 /// Subsequence calls will return false.
920 bool getNextRewritableSource(RegSubRegPair &Src,
921 RegSubRegPair &Dst) override {
922 // If we already get the only source we can rewrite, return false.
923 if (CurrentSrcIdx == 2)
924 return false;
925 // We are looking at v2 = INSERT_SUBREG v0, v1, sub0.
926 CurrentSrcIdx = 2;
927 const MachineOperand &MOInsertedReg = CopyLike.getOperand(2);
928 Src = RegSubRegPair(MOInsertedReg.getReg(), MOInsertedReg.getSubReg());
929 const MachineOperand &MODef = CopyLike.getOperand(0);
931 // We want to track something that is compatible with the
932 // partial definition.
933 if (MODef.getSubReg())
934 // Bail if we have to compose sub-register indices.
935 return false;
936 Dst = RegSubRegPair(MODef.getReg(),
937 (unsigned)CopyLike.getOperand(3).getImm());
938 return true;
941 bool RewriteCurrentSource(unsigned NewReg, unsigned NewSubReg) override {
942 if (CurrentSrcIdx != 2)
943 return false;
944 // We are rewriting the inserted reg.
945 MachineOperand &MO = CopyLike.getOperand(CurrentSrcIdx);
946 MO.setReg(NewReg);
947 MO.setSubReg(NewSubReg);
948 return true;
952 /// Specialized rewriter for EXTRACT_SUBREG instruction.
953 class ExtractSubregRewriter : public Rewriter {
954 const TargetInstrInfo &TII;
956 public:
957 ExtractSubregRewriter(MachineInstr &MI, const TargetInstrInfo &TII)
958 : Rewriter(MI), TII(TII) {
959 assert(MI.isExtractSubreg() && "Invalid instruction");
962 /// \see Rewriter::getNextRewritableSource()
963 /// Here CopyLike has the following form:
964 /// dst.dstSubIdx = EXTRACT_SUBREG Src, subIdx.
965 /// There is only one rewritable source: Src.subIdx,
966 /// which defines dst.dstSubIdx.
967 bool getNextRewritableSource(RegSubRegPair &Src,
968 RegSubRegPair &Dst) override {
969 // If we already get the only source we can rewrite, return false.
970 if (CurrentSrcIdx == 1)
971 return false;
972 // We are looking at v1 = EXTRACT_SUBREG v0, sub0.
973 CurrentSrcIdx = 1;
974 const MachineOperand &MOExtractedReg = CopyLike.getOperand(1);
975 // If we have to compose sub-register indices, bail out.
976 if (MOExtractedReg.getSubReg())
977 return false;
979 Src = RegSubRegPair(MOExtractedReg.getReg(),
980 CopyLike.getOperand(2).getImm());
982 // We want to track something that is compatible with the definition.
983 const MachineOperand &MODef = CopyLike.getOperand(0);
984 Dst = RegSubRegPair(MODef.getReg(), MODef.getSubReg());
985 return true;
988 bool RewriteCurrentSource(unsigned NewReg, unsigned NewSubReg) override {
989 // The only source we can rewrite is the input register.
990 if (CurrentSrcIdx != 1)
991 return false;
993 CopyLike.getOperand(CurrentSrcIdx).setReg(NewReg);
995 // If we find a source that does not require to extract something,
996 // rewrite the operation with a copy.
997 if (!NewSubReg) {
998 // Move the current index to an invalid position.
999 // We do not want another call to this method to be able
1000 // to do any change.
1001 CurrentSrcIdx = -1;
1002 // Rewrite the operation as a COPY.
1003 // Get rid of the sub-register index.
1004 CopyLike.RemoveOperand(2);
1005 // Morph the operation into a COPY.
1006 CopyLike.setDesc(TII.get(TargetOpcode::COPY));
1007 return true;
1009 CopyLike.getOperand(CurrentSrcIdx + 1).setImm(NewSubReg);
1010 return true;
1014 /// Specialized rewriter for REG_SEQUENCE instruction.
1015 class RegSequenceRewriter : public Rewriter {
1016 public:
1017 RegSequenceRewriter(MachineInstr &MI) : Rewriter(MI) {
1018 assert(MI.isRegSequence() && "Invalid instruction");
1021 /// \see Rewriter::getNextRewritableSource()
1022 /// Here CopyLike has the following form:
1023 /// dst = REG_SEQUENCE Src1.src1SubIdx, subIdx1, Src2.src2SubIdx, subIdx2.
1024 /// Each call will return a different source, walking all the available
1025 /// source.
1027 /// The first call returns:
1028 /// (SrcReg, SrcSubReg) = (Src1, src1SubIdx).
1029 /// (DstReg, DstSubReg) = (dst, subIdx1).
1031 /// The second call returns:
1032 /// (SrcReg, SrcSubReg) = (Src2, src2SubIdx).
1033 /// (DstReg, DstSubReg) = (dst, subIdx2).
1035 /// And so on, until all the sources have been traversed, then
1036 /// it returns false.
1037 bool getNextRewritableSource(RegSubRegPair &Src,
1038 RegSubRegPair &Dst) override {
1039 // We are looking at v0 = REG_SEQUENCE v1, sub1, v2, sub2, etc.
1041 // If this is the first call, move to the first argument.
1042 if (CurrentSrcIdx == 0) {
1043 CurrentSrcIdx = 1;
1044 } else {
1045 // Otherwise, move to the next argument and check that it is valid.
1046 CurrentSrcIdx += 2;
1047 if (CurrentSrcIdx >= CopyLike.getNumOperands())
1048 return false;
1050 const MachineOperand &MOInsertedReg = CopyLike.getOperand(CurrentSrcIdx);
1051 Src.Reg = MOInsertedReg.getReg();
1052 // If we have to compose sub-register indices, bail out.
1053 if ((Src.SubReg = MOInsertedReg.getSubReg()))
1054 return false;
1056 // We want to track something that is compatible with the related
1057 // partial definition.
1058 Dst.SubReg = CopyLike.getOperand(CurrentSrcIdx + 1).getImm();
1060 const MachineOperand &MODef = CopyLike.getOperand(0);
1061 Dst.Reg = MODef.getReg();
1062 // If we have to compose sub-registers, bail.
1063 return MODef.getSubReg() == 0;
1066 bool RewriteCurrentSource(unsigned NewReg, unsigned NewSubReg) override {
1067 // We cannot rewrite out of bound operands.
1068 // Moreover, rewritable sources are at odd positions.
1069 if ((CurrentSrcIdx & 1) != 1 || CurrentSrcIdx > CopyLike.getNumOperands())
1070 return false;
1072 MachineOperand &MO = CopyLike.getOperand(CurrentSrcIdx);
1073 MO.setReg(NewReg);
1074 MO.setSubReg(NewSubReg);
1075 return true;
1079 } // end anonymous namespace
1081 /// Get the appropriated Rewriter for \p MI.
1082 /// \return A pointer to a dynamically allocated Rewriter or nullptr if no
1083 /// rewriter works for \p MI.
1084 static Rewriter *getCopyRewriter(MachineInstr &MI, const TargetInstrInfo &TII) {
1085 // Handle uncoalescable copy-like instructions.
1086 if (MI.isBitcast() || MI.isRegSequenceLike() || MI.isInsertSubregLike() ||
1087 MI.isExtractSubregLike())
1088 return new UncoalescableRewriter(MI);
1090 switch (MI.getOpcode()) {
1091 default:
1092 return nullptr;
1093 case TargetOpcode::COPY:
1094 return new CopyRewriter(MI);
1095 case TargetOpcode::INSERT_SUBREG:
1096 return new InsertSubregRewriter(MI);
1097 case TargetOpcode::EXTRACT_SUBREG:
1098 return new ExtractSubregRewriter(MI, TII);
1099 case TargetOpcode::REG_SEQUENCE:
1100 return new RegSequenceRewriter(MI);
1104 /// Given a \p Def.Reg and Def.SubReg pair, use \p RewriteMap to find
1105 /// the new source to use for rewrite. If \p HandleMultipleSources is true and
1106 /// multiple sources for a given \p Def are found along the way, we found a
1107 /// PHI instructions that needs to be rewritten.
1108 /// TODO: HandleMultipleSources should be removed once we test PHI handling
1109 /// with coalescable copies.
1110 static RegSubRegPair
1111 getNewSource(MachineRegisterInfo *MRI, const TargetInstrInfo *TII,
1112 RegSubRegPair Def,
1113 const PeepholeOptimizer::RewriteMapTy &RewriteMap,
1114 bool HandleMultipleSources = true) {
1115 RegSubRegPair LookupSrc(Def.Reg, Def.SubReg);
1116 while (true) {
1117 ValueTrackerResult Res = RewriteMap.lookup(LookupSrc);
1118 // If there are no entries on the map, LookupSrc is the new source.
1119 if (!Res.isValid())
1120 return LookupSrc;
1122 // There's only one source for this definition, keep searching...
1123 unsigned NumSrcs = Res.getNumSources();
1124 if (NumSrcs == 1) {
1125 LookupSrc.Reg = Res.getSrcReg(0);
1126 LookupSrc.SubReg = Res.getSrcSubReg(0);
1127 continue;
1130 // TODO: Remove once multiple srcs w/ coalescable copies are supported.
1131 if (!HandleMultipleSources)
1132 break;
1134 // Multiple sources, recurse into each source to find a new source
1135 // for it. Then, rewrite the PHI accordingly to its new edges.
1136 SmallVector<RegSubRegPair, 4> NewPHISrcs;
1137 for (unsigned i = 0; i < NumSrcs; ++i) {
1138 RegSubRegPair PHISrc(Res.getSrcReg(i), Res.getSrcSubReg(i));
1139 NewPHISrcs.push_back(
1140 getNewSource(MRI, TII, PHISrc, RewriteMap, HandleMultipleSources));
1143 // Build the new PHI node and return its def register as the new source.
1144 MachineInstr &OrigPHI = const_cast<MachineInstr &>(*Res.getInst());
1145 MachineInstr &NewPHI = insertPHI(*MRI, *TII, NewPHISrcs, OrigPHI);
1146 LLVM_DEBUG(dbgs() << "-- getNewSource\n");
1147 LLVM_DEBUG(dbgs() << " Replacing: " << OrigPHI);
1148 LLVM_DEBUG(dbgs() << " With: " << NewPHI);
1149 const MachineOperand &MODef = NewPHI.getOperand(0);
1150 return RegSubRegPair(MODef.getReg(), MODef.getSubReg());
1153 return RegSubRegPair(0, 0);
1156 /// Optimize generic copy instructions to avoid cross register bank copy.
1157 /// The optimization looks through a chain of copies and tries to find a source
1158 /// that has a compatible register class.
1159 /// Two register classes are considered to be compatible if they share the same
1160 /// register bank.
1161 /// New copies issued by this optimization are register allocator
1162 /// friendly. This optimization does not remove any copy as it may
1163 /// overconstrain the register allocator, but replaces some operands
1164 /// when possible.
1165 /// \pre isCoalescableCopy(*MI) is true.
1166 /// \return True, when \p MI has been rewritten. False otherwise.
1167 bool PeepholeOptimizer::optimizeCoalescableCopy(MachineInstr &MI) {
1168 assert(isCoalescableCopy(MI) && "Invalid argument");
1169 assert(MI.getDesc().getNumDefs() == 1 &&
1170 "Coalescer can understand multiple defs?!");
1171 const MachineOperand &MODef = MI.getOperand(0);
1172 // Do not rewrite physical definitions.
1173 if (TargetRegisterInfo::isPhysicalRegister(MODef.getReg()))
1174 return false;
1176 bool Changed = false;
1177 // Get the right rewriter for the current copy.
1178 std::unique_ptr<Rewriter> CpyRewriter(getCopyRewriter(MI, *TII));
1179 // If none exists, bail out.
1180 if (!CpyRewriter)
1181 return false;
1182 // Rewrite each rewritable source.
1183 RegSubRegPair Src;
1184 RegSubRegPair TrackPair;
1185 while (CpyRewriter->getNextRewritableSource(Src, TrackPair)) {
1186 // Keep track of PHI nodes and its incoming edges when looking for sources.
1187 RewriteMapTy RewriteMap;
1188 // Try to find a more suitable source. If we failed to do so, or get the
1189 // actual source, move to the next source.
1190 if (!findNextSource(TrackPair, RewriteMap))
1191 continue;
1193 // Get the new source to rewrite. TODO: Only enable handling of multiple
1194 // sources (PHIs) once we have a motivating example and testcases for it.
1195 RegSubRegPair NewSrc = getNewSource(MRI, TII, TrackPair, RewriteMap,
1196 /*HandleMultipleSources=*/false);
1197 if (Src.Reg == NewSrc.Reg || NewSrc.Reg == 0)
1198 continue;
1200 // Rewrite source.
1201 if (CpyRewriter->RewriteCurrentSource(NewSrc.Reg, NewSrc.SubReg)) {
1202 // We may have extended the live-range of NewSrc, account for that.
1203 MRI->clearKillFlags(NewSrc.Reg);
1204 Changed = true;
1207 // TODO: We could have a clean-up method to tidy the instruction.
1208 // E.g., v0 = INSERT_SUBREG v1, v1.sub0, sub0
1209 // => v0 = COPY v1
1210 // Currently we haven't seen motivating example for that and we
1211 // want to avoid untested code.
1212 NumRewrittenCopies += Changed;
1213 return Changed;
1216 /// Rewrite the source found through \p Def, by using the \p RewriteMap
1217 /// and create a new COPY instruction. More info about RewriteMap in
1218 /// PeepholeOptimizer::findNextSource. Right now this is only used to handle
1219 /// Uncoalescable copies, since they are copy like instructions that aren't
1220 /// recognized by the register allocator.
1221 MachineInstr &
1222 PeepholeOptimizer::rewriteSource(MachineInstr &CopyLike,
1223 RegSubRegPair Def, RewriteMapTy &RewriteMap) {
1224 assert(!TargetRegisterInfo::isPhysicalRegister(Def.Reg) &&
1225 "We do not rewrite physical registers");
1227 // Find the new source to use in the COPY rewrite.
1228 RegSubRegPair NewSrc = getNewSource(MRI, TII, Def, RewriteMap);
1230 // Insert the COPY.
1231 const TargetRegisterClass *DefRC = MRI->getRegClass(Def.Reg);
1232 unsigned NewVReg = MRI->createVirtualRegister(DefRC);
1234 MachineInstr *NewCopy =
1235 BuildMI(*CopyLike.getParent(), &CopyLike, CopyLike.getDebugLoc(),
1236 TII->get(TargetOpcode::COPY), NewVReg)
1237 .addReg(NewSrc.Reg, 0, NewSrc.SubReg);
1239 if (Def.SubReg) {
1240 NewCopy->getOperand(0).setSubReg(Def.SubReg);
1241 NewCopy->getOperand(0).setIsUndef();
1244 LLVM_DEBUG(dbgs() << "-- RewriteSource\n");
1245 LLVM_DEBUG(dbgs() << " Replacing: " << CopyLike);
1246 LLVM_DEBUG(dbgs() << " With: " << *NewCopy);
1247 MRI->replaceRegWith(Def.Reg, NewVReg);
1248 MRI->clearKillFlags(NewVReg);
1250 // We extended the lifetime of NewSrc.Reg, clear the kill flags to
1251 // account for that.
1252 MRI->clearKillFlags(NewSrc.Reg);
1254 return *NewCopy;
1257 /// Optimize copy-like instructions to create
1258 /// register coalescer friendly instruction.
1259 /// The optimization tries to kill-off the \p MI by looking
1260 /// through a chain of copies to find a source that has a compatible
1261 /// register class.
1262 /// If such a source is found, it replace \p MI by a generic COPY
1263 /// operation.
1264 /// \pre isUncoalescableCopy(*MI) is true.
1265 /// \return True, when \p MI has been optimized. In that case, \p MI has
1266 /// been removed from its parent.
1267 /// All COPY instructions created, are inserted in \p LocalMIs.
1268 bool PeepholeOptimizer::optimizeUncoalescableCopy(
1269 MachineInstr &MI, SmallPtrSetImpl<MachineInstr *> &LocalMIs) {
1270 assert(isUncoalescableCopy(MI) && "Invalid argument");
1271 UncoalescableRewriter CpyRewriter(MI);
1273 // Rewrite each rewritable source by generating new COPYs. This works
1274 // differently from optimizeCoalescableCopy since it first makes sure that all
1275 // definitions can be rewritten.
1276 RewriteMapTy RewriteMap;
1277 RegSubRegPair Src;
1278 RegSubRegPair Def;
1279 SmallVector<RegSubRegPair, 4> RewritePairs;
1280 while (CpyRewriter.getNextRewritableSource(Src, Def)) {
1281 // If a physical register is here, this is probably for a good reason.
1282 // Do not rewrite that.
1283 if (TargetRegisterInfo::isPhysicalRegister(Def.Reg))
1284 return false;
1286 // If we do not know how to rewrite this definition, there is no point
1287 // in trying to kill this instruction.
1288 if (!findNextSource(Def, RewriteMap))
1289 return false;
1291 RewritePairs.push_back(Def);
1294 // The change is possible for all defs, do it.
1295 for (const RegSubRegPair &Def : RewritePairs) {
1296 // Rewrite the "copy" in a way the register coalescer understands.
1297 MachineInstr &NewCopy = rewriteSource(MI, Def, RewriteMap);
1298 LocalMIs.insert(&NewCopy);
1301 // MI is now dead.
1302 MI.eraseFromParent();
1303 ++NumUncoalescableCopies;
1304 return true;
1307 /// Check whether MI is a candidate for folding into a later instruction.
1308 /// We only fold loads to virtual registers and the virtual register defined
1309 /// has a single use.
1310 bool PeepholeOptimizer::isLoadFoldable(
1311 MachineInstr &MI, SmallSet<unsigned, 16> &FoldAsLoadDefCandidates) {
1312 if (!MI.canFoldAsLoad() || !MI.mayLoad())
1313 return false;
1314 const MCInstrDesc &MCID = MI.getDesc();
1315 if (MCID.getNumDefs() != 1)
1316 return false;
1318 unsigned Reg = MI.getOperand(0).getReg();
1319 // To reduce compilation time, we check MRI->hasOneNonDBGUse when inserting
1320 // loads. It should be checked when processing uses of the load, since
1321 // uses can be removed during peephole.
1322 if (!MI.getOperand(0).getSubReg() &&
1323 TargetRegisterInfo::isVirtualRegister(Reg) &&
1324 MRI->hasOneNonDBGUse(Reg)) {
1325 FoldAsLoadDefCandidates.insert(Reg);
1326 return true;
1328 return false;
1331 bool PeepholeOptimizer::isMoveImmediate(
1332 MachineInstr &MI, SmallSet<unsigned, 4> &ImmDefRegs,
1333 DenseMap<unsigned, MachineInstr *> &ImmDefMIs) {
1334 const MCInstrDesc &MCID = MI.getDesc();
1335 if (!MI.isMoveImmediate())
1336 return false;
1337 if (MCID.getNumDefs() != 1)
1338 return false;
1339 unsigned Reg = MI.getOperand(0).getReg();
1340 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
1341 ImmDefMIs.insert(std::make_pair(Reg, &MI));
1342 ImmDefRegs.insert(Reg);
1343 return true;
1346 return false;
1349 /// Try folding register operands that are defined by move immediate
1350 /// instructions, i.e. a trivial constant folding optimization, if
1351 /// and only if the def and use are in the same BB.
1352 bool PeepholeOptimizer::foldImmediate(MachineInstr &MI,
1353 SmallSet<unsigned, 4> &ImmDefRegs,
1354 DenseMap<unsigned, MachineInstr *> &ImmDefMIs) {
1355 for (unsigned i = 0, e = MI.getDesc().getNumOperands(); i != e; ++i) {
1356 MachineOperand &MO = MI.getOperand(i);
1357 if (!MO.isReg() || MO.isDef())
1358 continue;
1359 // Ignore dead implicit defs.
1360 if (MO.isImplicit() && MO.isDead())
1361 continue;
1362 unsigned Reg = MO.getReg();
1363 if (!TargetRegisterInfo::isVirtualRegister(Reg))
1364 continue;
1365 if (ImmDefRegs.count(Reg) == 0)
1366 continue;
1367 DenseMap<unsigned, MachineInstr*>::iterator II = ImmDefMIs.find(Reg);
1368 assert(II != ImmDefMIs.end() && "couldn't find immediate definition");
1369 if (TII->FoldImmediate(MI, *II->second, Reg, MRI)) {
1370 ++NumImmFold;
1371 return true;
1374 return false;
1377 // FIXME: This is very simple and misses some cases which should be handled when
1378 // motivating examples are found.
1380 // The copy rewriting logic should look at uses as well as defs and be able to
1381 // eliminate copies across blocks.
1383 // Later copies that are subregister extracts will also not be eliminated since
1384 // only the first copy is considered.
1386 // e.g.
1387 // %1 = COPY %0
1388 // %2 = COPY %0:sub1
1390 // Should replace %2 uses with %1:sub1
1391 bool PeepholeOptimizer::foldRedundantCopy(MachineInstr &MI,
1392 SmallSet<unsigned, 4> &CopySrcRegs,
1393 DenseMap<unsigned, MachineInstr *> &CopyMIs) {
1394 assert(MI.isCopy() && "expected a COPY machine instruction");
1396 unsigned SrcReg = MI.getOperand(1).getReg();
1397 if (!TargetRegisterInfo::isVirtualRegister(SrcReg))
1398 return false;
1400 unsigned DstReg = MI.getOperand(0).getReg();
1401 if (!TargetRegisterInfo::isVirtualRegister(DstReg))
1402 return false;
1404 if (CopySrcRegs.insert(SrcReg).second) {
1405 // First copy of this reg seen.
1406 CopyMIs.insert(std::make_pair(SrcReg, &MI));
1407 return false;
1410 MachineInstr *PrevCopy = CopyMIs.find(SrcReg)->second;
1412 unsigned SrcSubReg = MI.getOperand(1).getSubReg();
1413 unsigned PrevSrcSubReg = PrevCopy->getOperand(1).getSubReg();
1415 // Can't replace different subregister extracts.
1416 if (SrcSubReg != PrevSrcSubReg)
1417 return false;
1419 unsigned PrevDstReg = PrevCopy->getOperand(0).getReg();
1421 // Only replace if the copy register class is the same.
1423 // TODO: If we have multiple copies to different register classes, we may want
1424 // to track multiple copies of the same source register.
1425 if (MRI->getRegClass(DstReg) != MRI->getRegClass(PrevDstReg))
1426 return false;
1428 MRI->replaceRegWith(DstReg, PrevDstReg);
1430 // Lifetime of the previous copy has been extended.
1431 MRI->clearKillFlags(PrevDstReg);
1432 return true;
1435 bool PeepholeOptimizer::isNAPhysCopy(unsigned Reg) {
1436 return TargetRegisterInfo::isPhysicalRegister(Reg) &&
1437 !MRI->isAllocatable(Reg);
1440 bool PeepholeOptimizer::foldRedundantNAPhysCopy(
1441 MachineInstr &MI, DenseMap<unsigned, MachineInstr *> &NAPhysToVirtMIs) {
1442 assert(MI.isCopy() && "expected a COPY machine instruction");
1444 if (DisableNAPhysCopyOpt)
1445 return false;
1447 unsigned DstReg = MI.getOperand(0).getReg();
1448 unsigned SrcReg = MI.getOperand(1).getReg();
1449 if (isNAPhysCopy(SrcReg) && TargetRegisterInfo::isVirtualRegister(DstReg)) {
1450 // %vreg = COPY %physreg
1451 // Avoid using a datastructure which can track multiple live non-allocatable
1452 // phys->virt copies since LLVM doesn't seem to do this.
1453 NAPhysToVirtMIs.insert({SrcReg, &MI});
1454 return false;
1457 if (!(TargetRegisterInfo::isVirtualRegister(SrcReg) && isNAPhysCopy(DstReg)))
1458 return false;
1460 // %physreg = COPY %vreg
1461 auto PrevCopy = NAPhysToVirtMIs.find(DstReg);
1462 if (PrevCopy == NAPhysToVirtMIs.end()) {
1463 // We can't remove the copy: there was an intervening clobber of the
1464 // non-allocatable physical register after the copy to virtual.
1465 LLVM_DEBUG(dbgs() << "NAPhysCopy: intervening clobber forbids erasing "
1466 << MI);
1467 return false;
1470 unsigned PrevDstReg = PrevCopy->second->getOperand(0).getReg();
1471 if (PrevDstReg == SrcReg) {
1472 // Remove the virt->phys copy: we saw the virtual register definition, and
1473 // the non-allocatable physical register's state hasn't changed since then.
1474 LLVM_DEBUG(dbgs() << "NAPhysCopy: erasing " << MI);
1475 ++NumNAPhysCopies;
1476 return true;
1479 // Potential missed optimization opportunity: we saw a different virtual
1480 // register get a copy of the non-allocatable physical register, and we only
1481 // track one such copy. Avoid getting confused by this new non-allocatable
1482 // physical register definition, and remove it from the tracked copies.
1483 LLVM_DEBUG(dbgs() << "NAPhysCopy: missed opportunity " << MI);
1484 NAPhysToVirtMIs.erase(PrevCopy);
1485 return false;
1488 /// \bried Returns true if \p MO is a virtual register operand.
1489 static bool isVirtualRegisterOperand(MachineOperand &MO) {
1490 if (!MO.isReg())
1491 return false;
1492 return TargetRegisterInfo::isVirtualRegister(MO.getReg());
1495 bool PeepholeOptimizer::findTargetRecurrence(
1496 unsigned Reg, const SmallSet<unsigned, 2> &TargetRegs,
1497 RecurrenceCycle &RC) {
1498 // Recurrence found if Reg is in TargetRegs.
1499 if (TargetRegs.count(Reg))
1500 return true;
1502 // TODO: Curerntly, we only allow the last instruction of the recurrence
1503 // cycle (the instruction that feeds the PHI instruction) to have more than
1504 // one uses to guarantee that commuting operands does not tie registers
1505 // with overlapping live range. Once we have actual live range info of
1506 // each register, this constraint can be relaxed.
1507 if (!MRI->hasOneNonDBGUse(Reg))
1508 return false;
1510 // Give up if the reccurrence chain length is longer than the limit.
1511 if (RC.size() >= MaxRecurrenceChain)
1512 return false;
1514 MachineInstr &MI = *(MRI->use_instr_nodbg_begin(Reg));
1515 unsigned Idx = MI.findRegisterUseOperandIdx(Reg);
1517 // Only interested in recurrences whose instructions have only one def, which
1518 // is a virtual register.
1519 if (MI.getDesc().getNumDefs() != 1)
1520 return false;
1522 MachineOperand &DefOp = MI.getOperand(0);
1523 if (!isVirtualRegisterOperand(DefOp))
1524 return false;
1526 // Check if def operand of MI is tied to any use operand. We are only
1527 // interested in the case that all the instructions in the recurrence chain
1528 // have there def operand tied with one of the use operand.
1529 unsigned TiedUseIdx;
1530 if (!MI.isRegTiedToUseOperand(0, &TiedUseIdx))
1531 return false;
1533 if (Idx == TiedUseIdx) {
1534 RC.push_back(RecurrenceInstr(&MI));
1535 return findTargetRecurrence(DefOp.getReg(), TargetRegs, RC);
1536 } else {
1537 // If Idx is not TiedUseIdx, check if Idx is commutable with TiedUseIdx.
1538 unsigned CommIdx = TargetInstrInfo::CommuteAnyOperandIndex;
1539 if (TII->findCommutedOpIndices(MI, Idx, CommIdx) && CommIdx == TiedUseIdx) {
1540 RC.push_back(RecurrenceInstr(&MI, Idx, CommIdx));
1541 return findTargetRecurrence(DefOp.getReg(), TargetRegs, RC);
1545 return false;
1548 /// Phi instructions will eventually be lowered to copy instructions.
1549 /// If phi is in a loop header, a recurrence may formulated around the source
1550 /// and destination of the phi. For such case commuting operands of the
1551 /// instructions in the recurrence may enable coalescing of the copy instruction
1552 /// generated from the phi. For example, if there is a recurrence of
1554 /// LoopHeader:
1555 /// %1 = phi(%0, %100)
1556 /// LoopLatch:
1557 /// %0<def, tied1> = ADD %2<def, tied0>, %1
1559 /// , the fact that %0 and %2 are in the same tied operands set makes
1560 /// the coalescing of copy instruction generated from the phi in
1561 /// LoopHeader(i.e. %1 = COPY %0) impossible, because %1 and
1562 /// %2 have overlapping live range. This introduces additional move
1563 /// instruction to the final assembly. However, if we commute %2 and
1564 /// %1 of ADD instruction, the redundant move instruction can be
1565 /// avoided.
1566 bool PeepholeOptimizer::optimizeRecurrence(MachineInstr &PHI) {
1567 SmallSet<unsigned, 2> TargetRegs;
1568 for (unsigned Idx = 1; Idx < PHI.getNumOperands(); Idx += 2) {
1569 MachineOperand &MO = PHI.getOperand(Idx);
1570 assert(isVirtualRegisterOperand(MO) && "Invalid PHI instruction");
1571 TargetRegs.insert(MO.getReg());
1574 bool Changed = false;
1575 RecurrenceCycle RC;
1576 if (findTargetRecurrence(PHI.getOperand(0).getReg(), TargetRegs, RC)) {
1577 // Commutes operands of instructions in RC if necessary so that the copy to
1578 // be generated from PHI can be coalesced.
1579 LLVM_DEBUG(dbgs() << "Optimize recurrence chain from " << PHI);
1580 for (auto &RI : RC) {
1581 LLVM_DEBUG(dbgs() << "\tInst: " << *(RI.getMI()));
1582 auto CP = RI.getCommutePair();
1583 if (CP) {
1584 Changed = true;
1585 TII->commuteInstruction(*(RI.getMI()), false, (*CP).first,
1586 (*CP).second);
1587 LLVM_DEBUG(dbgs() << "\t\tCommuted: " << *(RI.getMI()));
1592 return Changed;
1595 bool PeepholeOptimizer::runOnMachineFunction(MachineFunction &MF) {
1596 if (skipFunction(MF.getFunction()))
1597 return false;
1599 LLVM_DEBUG(dbgs() << "********** PEEPHOLE OPTIMIZER **********\n");
1600 LLVM_DEBUG(dbgs() << "********** Function: " << MF.getName() << '\n');
1602 if (DisablePeephole)
1603 return false;
1605 TII = MF.getSubtarget().getInstrInfo();
1606 TRI = MF.getSubtarget().getRegisterInfo();
1607 MRI = &MF.getRegInfo();
1608 DT = Aggressive ? &getAnalysis<MachineDominatorTree>() : nullptr;
1609 MLI = &getAnalysis<MachineLoopInfo>();
1611 bool Changed = false;
1613 for (MachineBasicBlock &MBB : MF) {
1614 bool SeenMoveImm = false;
1616 // During this forward scan, at some point it needs to answer the question
1617 // "given a pointer to an MI in the current BB, is it located before or
1618 // after the current instruction".
1619 // To perform this, the following set keeps track of the MIs already seen
1620 // during the scan, if a MI is not in the set, it is assumed to be located
1621 // after. Newly created MIs have to be inserted in the set as well.
1622 SmallPtrSet<MachineInstr*, 16> LocalMIs;
1623 SmallSet<unsigned, 4> ImmDefRegs;
1624 DenseMap<unsigned, MachineInstr*> ImmDefMIs;
1625 SmallSet<unsigned, 16> FoldAsLoadDefCandidates;
1627 // Track when a non-allocatable physical register is copied to a virtual
1628 // register so that useless moves can be removed.
1630 // %physreg is the map index; MI is the last valid `%vreg = COPY %physreg`
1631 // without any intervening re-definition of %physreg.
1632 DenseMap<unsigned, MachineInstr *> NAPhysToVirtMIs;
1634 // Set of virtual registers that are copied from.
1635 SmallSet<unsigned, 4> CopySrcRegs;
1636 DenseMap<unsigned, MachineInstr *> CopySrcMIs;
1638 bool IsLoopHeader = MLI->isLoopHeader(&MBB);
1640 for (MachineBasicBlock::iterator MII = MBB.begin(), MIE = MBB.end();
1641 MII != MIE; ) {
1642 MachineInstr *MI = &*MII;
1643 // We may be erasing MI below, increment MII now.
1644 ++MII;
1645 LocalMIs.insert(MI);
1647 // Skip debug instructions. They should not affect this peephole optimization.
1648 if (MI->isDebugInstr())
1649 continue;
1651 if (MI->isPosition())
1652 continue;
1654 if (IsLoopHeader && MI->isPHI()) {
1655 if (optimizeRecurrence(*MI)) {
1656 Changed = true;
1657 continue;
1661 if (!MI->isCopy()) {
1662 for (const MachineOperand &MO : MI->operands()) {
1663 // Visit all operands: definitions can be implicit or explicit.
1664 if (MO.isReg()) {
1665 unsigned Reg = MO.getReg();
1666 if (MO.isDef() && isNAPhysCopy(Reg)) {
1667 const auto &Def = NAPhysToVirtMIs.find(Reg);
1668 if (Def != NAPhysToVirtMIs.end()) {
1669 // A new definition of the non-allocatable physical register
1670 // invalidates previous copies.
1671 LLVM_DEBUG(dbgs()
1672 << "NAPhysCopy: invalidating because of " << *MI);
1673 NAPhysToVirtMIs.erase(Def);
1676 } else if (MO.isRegMask()) {
1677 const uint32_t *RegMask = MO.getRegMask();
1678 for (auto &RegMI : NAPhysToVirtMIs) {
1679 unsigned Def = RegMI.first;
1680 if (MachineOperand::clobbersPhysReg(RegMask, Def)) {
1681 LLVM_DEBUG(dbgs()
1682 << "NAPhysCopy: invalidating because of " << *MI);
1683 NAPhysToVirtMIs.erase(Def);
1690 if (MI->isImplicitDef() || MI->isKill())
1691 continue;
1693 if (MI->isInlineAsm() || MI->hasUnmodeledSideEffects()) {
1694 // Blow away all non-allocatable physical registers knowledge since we
1695 // don't know what's correct anymore.
1697 // FIXME: handle explicit asm clobbers.
1698 LLVM_DEBUG(dbgs() << "NAPhysCopy: blowing away all info due to "
1699 << *MI);
1700 NAPhysToVirtMIs.clear();
1703 if ((isUncoalescableCopy(*MI) &&
1704 optimizeUncoalescableCopy(*MI, LocalMIs)) ||
1705 (MI->isCompare() && optimizeCmpInstr(*MI)) ||
1706 (MI->isSelect() && optimizeSelect(*MI, LocalMIs))) {
1707 // MI is deleted.
1708 LocalMIs.erase(MI);
1709 Changed = true;
1710 continue;
1713 if (MI->isConditionalBranch() && optimizeCondBranch(*MI)) {
1714 Changed = true;
1715 continue;
1718 if (isCoalescableCopy(*MI) && optimizeCoalescableCopy(*MI)) {
1719 // MI is just rewritten.
1720 Changed = true;
1721 continue;
1724 if (MI->isCopy() &&
1725 (foldRedundantCopy(*MI, CopySrcRegs, CopySrcMIs) ||
1726 foldRedundantNAPhysCopy(*MI, NAPhysToVirtMIs))) {
1727 LocalMIs.erase(MI);
1728 MI->eraseFromParent();
1729 Changed = true;
1730 continue;
1733 if (isMoveImmediate(*MI, ImmDefRegs, ImmDefMIs)) {
1734 SeenMoveImm = true;
1735 } else {
1736 Changed |= optimizeExtInstr(*MI, MBB, LocalMIs);
1737 // optimizeExtInstr might have created new instructions after MI
1738 // and before the already incremented MII. Adjust MII so that the
1739 // next iteration sees the new instructions.
1740 MII = MI;
1741 ++MII;
1742 if (SeenMoveImm)
1743 Changed |= foldImmediate(*MI, ImmDefRegs, ImmDefMIs);
1746 // Check whether MI is a load candidate for folding into a later
1747 // instruction. If MI is not a candidate, check whether we can fold an
1748 // earlier load into MI.
1749 if (!isLoadFoldable(*MI, FoldAsLoadDefCandidates) &&
1750 !FoldAsLoadDefCandidates.empty()) {
1752 // We visit each operand even after successfully folding a previous
1753 // one. This allows us to fold multiple loads into a single
1754 // instruction. We do assume that optimizeLoadInstr doesn't insert
1755 // foldable uses earlier in the argument list. Since we don't restart
1756 // iteration, we'd miss such cases.
1757 const MCInstrDesc &MIDesc = MI->getDesc();
1758 for (unsigned i = MIDesc.getNumDefs(); i != MI->getNumOperands();
1759 ++i) {
1760 const MachineOperand &MOp = MI->getOperand(i);
1761 if (!MOp.isReg())
1762 continue;
1763 unsigned FoldAsLoadDefReg = MOp.getReg();
1764 if (FoldAsLoadDefCandidates.count(FoldAsLoadDefReg)) {
1765 // We need to fold load after optimizeCmpInstr, since
1766 // optimizeCmpInstr can enable folding by converting SUB to CMP.
1767 // Save FoldAsLoadDefReg because optimizeLoadInstr() resets it and
1768 // we need it for markUsesInDebugValueAsUndef().
1769 unsigned FoldedReg = FoldAsLoadDefReg;
1770 MachineInstr *DefMI = nullptr;
1771 if (MachineInstr *FoldMI =
1772 TII->optimizeLoadInstr(*MI, MRI, FoldAsLoadDefReg, DefMI)) {
1773 // Update LocalMIs since we replaced MI with FoldMI and deleted
1774 // DefMI.
1775 LLVM_DEBUG(dbgs() << "Replacing: " << *MI);
1776 LLVM_DEBUG(dbgs() << " With: " << *FoldMI);
1777 LocalMIs.erase(MI);
1778 LocalMIs.erase(DefMI);
1779 LocalMIs.insert(FoldMI);
1780 MI->eraseFromParent();
1781 DefMI->eraseFromParent();
1782 MRI->markUsesInDebugValueAsUndef(FoldedReg);
1783 FoldAsLoadDefCandidates.erase(FoldedReg);
1784 ++NumLoadFold;
1786 // MI is replaced with FoldMI so we can continue trying to fold
1787 Changed = true;
1788 MI = FoldMI;
1794 // If we run into an instruction we can't fold across, discard
1795 // the load candidates. Note: We might be able to fold *into* this
1796 // instruction, so this needs to be after the folding logic.
1797 if (MI->isLoadFoldBarrier()) {
1798 LLVM_DEBUG(dbgs() << "Encountered load fold barrier on " << *MI);
1799 FoldAsLoadDefCandidates.clear();
1804 return Changed;
1807 ValueTrackerResult ValueTracker::getNextSourceFromCopy() {
1808 assert(Def->isCopy() && "Invalid definition");
1809 // Copy instruction are supposed to be: Def = Src.
1810 // If someone breaks this assumption, bad things will happen everywhere.
1811 assert(Def->getNumOperands() == 2 && "Invalid number of operands");
1813 if (Def->getOperand(DefIdx).getSubReg() != DefSubReg)
1814 // If we look for a different subreg, it means we want a subreg of src.
1815 // Bails as we do not support composing subregs yet.
1816 return ValueTrackerResult();
1817 // Otherwise, we want the whole source.
1818 const MachineOperand &Src = Def->getOperand(1);
1819 if (Src.isUndef())
1820 return ValueTrackerResult();
1821 return ValueTrackerResult(Src.getReg(), Src.getSubReg());
1824 ValueTrackerResult ValueTracker::getNextSourceFromBitcast() {
1825 assert(Def->isBitcast() && "Invalid definition");
1827 // Bail if there are effects that a plain copy will not expose.
1828 if (Def->hasUnmodeledSideEffects())
1829 return ValueTrackerResult();
1831 // Bitcasts with more than one def are not supported.
1832 if (Def->getDesc().getNumDefs() != 1)
1833 return ValueTrackerResult();
1834 const MachineOperand DefOp = Def->getOperand(DefIdx);
1835 if (DefOp.getSubReg() != DefSubReg)
1836 // If we look for a different subreg, it means we want a subreg of the src.
1837 // Bails as we do not support composing subregs yet.
1838 return ValueTrackerResult();
1840 unsigned SrcIdx = Def->getNumOperands();
1841 for (unsigned OpIdx = DefIdx + 1, EndOpIdx = SrcIdx; OpIdx != EndOpIdx;
1842 ++OpIdx) {
1843 const MachineOperand &MO = Def->getOperand(OpIdx);
1844 if (!MO.isReg() || !MO.getReg())
1845 continue;
1846 // Ignore dead implicit defs.
1847 if (MO.isImplicit() && MO.isDead())
1848 continue;
1849 assert(!MO.isDef() && "We should have skipped all the definitions by now");
1850 if (SrcIdx != EndOpIdx)
1851 // Multiple sources?
1852 return ValueTrackerResult();
1853 SrcIdx = OpIdx;
1856 // Stop when any user of the bitcast is a SUBREG_TO_REG, replacing with a COPY
1857 // will break the assumed guarantees for the upper bits.
1858 for (const MachineInstr &UseMI : MRI.use_nodbg_instructions(DefOp.getReg())) {
1859 if (UseMI.isSubregToReg())
1860 return ValueTrackerResult();
1863 const MachineOperand &Src = Def->getOperand(SrcIdx);
1864 if (Src.isUndef())
1865 return ValueTrackerResult();
1866 return ValueTrackerResult(Src.getReg(), Src.getSubReg());
1869 ValueTrackerResult ValueTracker::getNextSourceFromRegSequence() {
1870 assert((Def->isRegSequence() || Def->isRegSequenceLike()) &&
1871 "Invalid definition");
1873 if (Def->getOperand(DefIdx).getSubReg())
1874 // If we are composing subregs, bail out.
1875 // The case we are checking is Def.<subreg> = REG_SEQUENCE.
1876 // This should almost never happen as the SSA property is tracked at
1877 // the register level (as opposed to the subreg level).
1878 // I.e.,
1879 // Def.sub0 =
1880 // Def.sub1 =
1881 // is a valid SSA representation for Def.sub0 and Def.sub1, but not for
1882 // Def. Thus, it must not be generated.
1883 // However, some code could theoretically generates a single
1884 // Def.sub0 (i.e, not defining the other subregs) and we would
1885 // have this case.
1886 // If we can ascertain (or force) that this never happens, we could
1887 // turn that into an assertion.
1888 return ValueTrackerResult();
1890 if (!TII)
1891 // We could handle the REG_SEQUENCE here, but we do not want to
1892 // duplicate the code from the generic TII.
1893 return ValueTrackerResult();
1895 SmallVector<RegSubRegPairAndIdx, 8> RegSeqInputRegs;
1896 if (!TII->getRegSequenceInputs(*Def, DefIdx, RegSeqInputRegs))
1897 return ValueTrackerResult();
1899 // We are looking at:
1900 // Def = REG_SEQUENCE v0, sub0, v1, sub1, ...
1901 // Check if one of the operand defines the subreg we are interested in.
1902 for (const RegSubRegPairAndIdx &RegSeqInput : RegSeqInputRegs) {
1903 if (RegSeqInput.SubIdx == DefSubReg) {
1904 if (RegSeqInput.SubReg)
1905 // Bail if we have to compose sub registers.
1906 return ValueTrackerResult();
1908 return ValueTrackerResult(RegSeqInput.Reg, RegSeqInput.SubReg);
1912 // If the subreg we are tracking is super-defined by another subreg,
1913 // we could follow this value. However, this would require to compose
1914 // the subreg and we do not do that for now.
1915 return ValueTrackerResult();
1918 ValueTrackerResult ValueTracker::getNextSourceFromInsertSubreg() {
1919 assert((Def->isInsertSubreg() || Def->isInsertSubregLike()) &&
1920 "Invalid definition");
1922 if (Def->getOperand(DefIdx).getSubReg())
1923 // If we are composing subreg, bail out.
1924 // Same remark as getNextSourceFromRegSequence.
1925 // I.e., this may be turned into an assert.
1926 return ValueTrackerResult();
1928 if (!TII)
1929 // We could handle the REG_SEQUENCE here, but we do not want to
1930 // duplicate the code from the generic TII.
1931 return ValueTrackerResult();
1933 RegSubRegPair BaseReg;
1934 RegSubRegPairAndIdx InsertedReg;
1935 if (!TII->getInsertSubregInputs(*Def, DefIdx, BaseReg, InsertedReg))
1936 return ValueTrackerResult();
1938 // We are looking at:
1939 // Def = INSERT_SUBREG v0, v1, sub1
1940 // There are two cases:
1941 // 1. DefSubReg == sub1, get v1.
1942 // 2. DefSubReg != sub1, the value may be available through v0.
1944 // #1 Check if the inserted register matches the required sub index.
1945 if (InsertedReg.SubIdx == DefSubReg) {
1946 return ValueTrackerResult(InsertedReg.Reg, InsertedReg.SubReg);
1948 // #2 Otherwise, if the sub register we are looking for is not partial
1949 // defined by the inserted element, we can look through the main
1950 // register (v0).
1951 const MachineOperand &MODef = Def->getOperand(DefIdx);
1952 // If the result register (Def) and the base register (v0) do not
1953 // have the same register class or if we have to compose
1954 // subregisters, bail out.
1955 if (MRI.getRegClass(MODef.getReg()) != MRI.getRegClass(BaseReg.Reg) ||
1956 BaseReg.SubReg)
1957 return ValueTrackerResult();
1959 // Get the TRI and check if the inserted sub-register overlaps with the
1960 // sub-register we are tracking.
1961 const TargetRegisterInfo *TRI = MRI.getTargetRegisterInfo();
1962 if (!TRI ||
1963 !(TRI->getSubRegIndexLaneMask(DefSubReg) &
1964 TRI->getSubRegIndexLaneMask(InsertedReg.SubIdx)).none())
1965 return ValueTrackerResult();
1966 // At this point, the value is available in v0 via the same subreg
1967 // we used for Def.
1968 return ValueTrackerResult(BaseReg.Reg, DefSubReg);
1971 ValueTrackerResult ValueTracker::getNextSourceFromExtractSubreg() {
1972 assert((Def->isExtractSubreg() ||
1973 Def->isExtractSubregLike()) && "Invalid definition");
1974 // We are looking at:
1975 // Def = EXTRACT_SUBREG v0, sub0
1977 // Bail if we have to compose sub registers.
1978 // Indeed, if DefSubReg != 0, we would have to compose it with sub0.
1979 if (DefSubReg)
1980 return ValueTrackerResult();
1982 if (!TII)
1983 // We could handle the EXTRACT_SUBREG here, but we do not want to
1984 // duplicate the code from the generic TII.
1985 return ValueTrackerResult();
1987 RegSubRegPairAndIdx ExtractSubregInputReg;
1988 if (!TII->getExtractSubregInputs(*Def, DefIdx, ExtractSubregInputReg))
1989 return ValueTrackerResult();
1991 // Bail if we have to compose sub registers.
1992 // Likewise, if v0.subreg != 0, we would have to compose v0.subreg with sub0.
1993 if (ExtractSubregInputReg.SubReg)
1994 return ValueTrackerResult();
1995 // Otherwise, the value is available in the v0.sub0.
1996 return ValueTrackerResult(ExtractSubregInputReg.Reg,
1997 ExtractSubregInputReg.SubIdx);
2000 ValueTrackerResult ValueTracker::getNextSourceFromSubregToReg() {
2001 assert(Def->isSubregToReg() && "Invalid definition");
2002 // We are looking at:
2003 // Def = SUBREG_TO_REG Imm, v0, sub0
2005 // Bail if we have to compose sub registers.
2006 // If DefSubReg != sub0, we would have to check that all the bits
2007 // we track are included in sub0 and if yes, we would have to
2008 // determine the right subreg in v0.
2009 if (DefSubReg != Def->getOperand(3).getImm())
2010 return ValueTrackerResult();
2011 // Bail if we have to compose sub registers.
2012 // Likewise, if v0.subreg != 0, we would have to compose it with sub0.
2013 if (Def->getOperand(2).getSubReg())
2014 return ValueTrackerResult();
2016 return ValueTrackerResult(Def->getOperand(2).getReg(),
2017 Def->getOperand(3).getImm());
2020 /// Explore each PHI incoming operand and return its sources.
2021 ValueTrackerResult ValueTracker::getNextSourceFromPHI() {
2022 assert(Def->isPHI() && "Invalid definition");
2023 ValueTrackerResult Res;
2025 // If we look for a different subreg, bail as we do not support composing
2026 // subregs yet.
2027 if (Def->getOperand(0).getSubReg() != DefSubReg)
2028 return ValueTrackerResult();
2030 // Return all register sources for PHI instructions.
2031 for (unsigned i = 1, e = Def->getNumOperands(); i < e; i += 2) {
2032 const MachineOperand &MO = Def->getOperand(i);
2033 assert(MO.isReg() && "Invalid PHI instruction");
2034 // We have no code to deal with undef operands. They shouldn't happen in
2035 // normal programs anyway.
2036 if (MO.isUndef())
2037 return ValueTrackerResult();
2038 Res.addSource(MO.getReg(), MO.getSubReg());
2041 return Res;
2044 ValueTrackerResult ValueTracker::getNextSourceImpl() {
2045 assert(Def && "This method needs a valid definition");
2047 assert(((Def->getOperand(DefIdx).isDef() &&
2048 (DefIdx < Def->getDesc().getNumDefs() ||
2049 Def->getDesc().isVariadic())) ||
2050 Def->getOperand(DefIdx).isImplicit()) &&
2051 "Invalid DefIdx");
2052 if (Def->isCopy())
2053 return getNextSourceFromCopy();
2054 if (Def->isBitcast())
2055 return getNextSourceFromBitcast();
2056 // All the remaining cases involve "complex" instructions.
2057 // Bail if we did not ask for the advanced tracking.
2058 if (DisableAdvCopyOpt)
2059 return ValueTrackerResult();
2060 if (Def->isRegSequence() || Def->isRegSequenceLike())
2061 return getNextSourceFromRegSequence();
2062 if (Def->isInsertSubreg() || Def->isInsertSubregLike())
2063 return getNextSourceFromInsertSubreg();
2064 if (Def->isExtractSubreg() || Def->isExtractSubregLike())
2065 return getNextSourceFromExtractSubreg();
2066 if (Def->isSubregToReg())
2067 return getNextSourceFromSubregToReg();
2068 if (Def->isPHI())
2069 return getNextSourceFromPHI();
2070 return ValueTrackerResult();
2073 ValueTrackerResult ValueTracker::getNextSource() {
2074 // If we reach a point where we cannot move up in the use-def chain,
2075 // there is nothing we can get.
2076 if (!Def)
2077 return ValueTrackerResult();
2079 ValueTrackerResult Res = getNextSourceImpl();
2080 if (Res.isValid()) {
2081 // Update definition, definition index, and subregister for the
2082 // next call of getNextSource.
2083 // Update the current register.
2084 bool OneRegSrc = Res.getNumSources() == 1;
2085 if (OneRegSrc)
2086 Reg = Res.getSrcReg(0);
2087 // Update the result before moving up in the use-def chain
2088 // with the instruction containing the last found sources.
2089 Res.setInst(Def);
2091 // If we can still move up in the use-def chain, move to the next
2092 // definition.
2093 if (!TargetRegisterInfo::isPhysicalRegister(Reg) && OneRegSrc) {
2094 MachineRegisterInfo::def_iterator DI = MRI.def_begin(Reg);
2095 if (DI != MRI.def_end()) {
2096 Def = DI->getParent();
2097 DefIdx = DI.getOperandNo();
2098 DefSubReg = Res.getSrcSubReg(0);
2099 } else {
2100 Def = nullptr;
2102 return Res;
2105 // If we end up here, this means we will not be able to find another source
2106 // for the next iteration. Make sure any new call to getNextSource bails out
2107 // early by cutting the use-def chain.
2108 Def = nullptr;
2109 return Res;