1 //===- JumpThreading.cpp - Thread control through conditional blocks ------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements the Jump Threading pass.
10 //
11 //===----------------------------------------------------------------------===//
12
13 #include "llvm/Transforms/Scalar/JumpThreading.h"
14 #include "llvm/ADT/DenseMap.h"
15 #include "llvm/ADT/DenseSet.h"
16 #include "llvm/ADT/MapVector.h"
17 #include "llvm/ADT/Optional.h"
18 #include "llvm/ADT/STLExtras.h"
19 #include "llvm/ADT/SmallPtrSet.h"
20 #include "llvm/ADT/SmallVector.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/Analysis/AliasAnalysis.h"
23 #include "llvm/Analysis/BlockFrequencyInfo.h"
24 #include "llvm/Analysis/BranchProbabilityInfo.h"
25 #include "llvm/Analysis/CFG.h"
26 #include "llvm/Analysis/ConstantFolding.h"
27 #include "llvm/Analysis/DomTreeUpdater.h"
28 #include "llvm/Analysis/GlobalsModRef.h"
29 #include "llvm/Analysis/GuardUtils.h"
30 #include "llvm/Analysis/InstructionSimplify.h"
31 #include "llvm/Analysis/LazyValueInfo.h"
32 #include "llvm/Analysis/Loads.h"
33 #include "llvm/Analysis/LoopInfo.h"
34 #include "llvm/Analysis/TargetLibraryInfo.h"
35 #include "llvm/Analysis/TargetTransformInfo.h"
36 #include "llvm/Analysis/ValueTracking.h"
37 #include "llvm/IR/BasicBlock.h"
38 #include "llvm/IR/CFG.h"
39 #include "llvm/IR/Constant.h"
40 #include "llvm/IR/ConstantRange.h"
41 #include "llvm/IR/Constants.h"
42 #include "llvm/IR/DataLayout.h"
43 #include "llvm/IR/Dominators.h"
44 #include "llvm/IR/Function.h"
45 #include "llvm/IR/InstrTypes.h"
46 #include "llvm/IR/Instruction.h"
47 #include "llvm/IR/Instructions.h"
48 #include "llvm/IR/IntrinsicInst.h"
49 #include "llvm/IR/Intrinsics.h"
50 #include "llvm/IR/LLVMContext.h"
51 #include "llvm/IR/MDBuilder.h"
52 #include "llvm/IR/Metadata.h"
53 #include "llvm/IR/Module.h"
54 #include "llvm/IR/PassManager.h"
55 #include "llvm/IR/PatternMatch.h"
56 #include "llvm/IR/Type.h"
57 #include "llvm/IR/Use.h"
58 #include "llvm/IR/User.h"
59 #include "llvm/IR/Value.h"
60 #include "llvm/InitializePasses.h"
61 #include "llvm/Pass.h"
62 #include "llvm/Support/BlockFrequency.h"
63 #include "llvm/Support/BranchProbability.h"
64 #include "llvm/Support/Casting.h"
65 #include "llvm/Support/CommandLine.h"
66 #include "llvm/Support/Debug.h"
67 #include "llvm/Support/raw_ostream.h"
68 #include "llvm/Transforms/Scalar.h"
69 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
70 #include "llvm/Transforms/Utils/Cloning.h"
71 #include "llvm/Transforms/Utils/Local.h"
72 #include "llvm/Transforms/Utils/SSAUpdater.h"
73 #include "llvm/Transforms/Utils/ValueMapper.h"
74 #include <algorithm>
75 #include <cassert>
76 #include <cstddef>
77 #include <cstdint>
78 #include <iterator>
79 #include <memory>
80 #include <utility>
81
82 using namespace llvm;
83 using namespace jumpthreading;
84
85 #define DEBUG_TYPE "jump-threading"
86
87 STATISTIC(NumThreads, "Number of jumps threaded");
88 STATISTIC(NumFolds, "Number of terminators folded");
89 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
90
91 static cl::opt<unsigned>
92 BBDuplicateThreshold("jump-threading-threshold",
93 cl::desc("Max block size to duplicate for jump threading"),
94 cl::init(6), cl::Hidden);
95
96 static cl::opt<unsigned>
97 ImplicationSearchThreshold(
98 "jump-threading-implication-search-threshold",
99 cl::desc("The number of predecessors to search for a stronger "
100 "condition to use to thread over a weaker condition"),
101 cl::init(3), cl::Hidden);
102
103 static cl::opt<bool> PrintLVIAfterJumpThreading(
104 "print-lvi-after-jump-threading",
105 cl::desc("Print the LazyValueInfo cache after JumpThreading"), cl::init(false),
106 cl::Hidden);
107
108 static cl::opt<bool> JumpThreadingFreezeSelectCond(
109 "jump-threading-freeze-select-cond",
110 cl::desc("Freeze the condition when unfolding select"), cl::init(false),
111 cl::Hidden);
112
113 static cl::opt<bool> ThreadAcrossLoopHeaders(
114 "jump-threading-across-loop-headers",
115 cl::desc("Allow JumpThreading to thread across loop headers, for testing"),
116 cl::init(false), cl::Hidden);
117
118
119 namespace {
120
121 /// This pass performs 'jump threading', which looks at blocks that have
122 /// multiple predecessors and multiple successors. If one or more of the
123 /// predecessors of the block can be proven to always jump to one of the
124 /// successors, we forward the edge from the predecessor to the successor by
125 /// duplicating the contents of this block.
126 ///
127 /// An example of when this can occur is code like this:
128 ///
129 /// if () { ...
130 /// X = 4;
131 /// }
132 /// if (X < 3) {
133 ///
134 /// In this case, the unconditional branch at the end of the first if can be
135 /// revectored to the false side of the second if.
136 class JumpThreading : public FunctionPass {
137 JumpThreadingPass Impl;
138
139 public:
140 static char ID; // Pass identification
141
JumpThreading(bool InsertFreezeWhenUnfoldingSelect=false,int T=-1)142 JumpThreading(bool InsertFreezeWhenUnfoldingSelect = false, int T = -1)
143 : FunctionPass(ID), Impl(InsertFreezeWhenUnfoldingSelect, T) {
144 initializeJumpThreadingPass(*PassRegistry::getPassRegistry());
145 }
146
147 bool runOnFunction(Function &F) override;
148
getAnalysisUsage(AnalysisUsage & AU) const149 void getAnalysisUsage(AnalysisUsage &AU) const override {
150 AU.addRequired<DominatorTreeWrapperPass>();
151 AU.addPreserved<DominatorTreeWrapperPass>();
152 AU.addRequired<AAResultsWrapperPass>();
153 AU.addRequired<LazyValueInfoWrapperPass>();
154 AU.addPreserved<LazyValueInfoWrapperPass>();
155 AU.addPreserved<GlobalsAAWrapperPass>();
156 AU.addRequired<TargetLibraryInfoWrapperPass>();
157 AU.addRequired<TargetTransformInfoWrapperPass>();
158 }
159
releaseMemory()160 void releaseMemory() override { Impl.releaseMemory(); }
161 };
162
163 } // end anonymous namespace
164
165 char JumpThreading::ID = 0;
166
167 INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",
168 "Jump Threading", false, false)
INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)169 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
170 INITIALIZE_PASS_DEPENDENCY(LazyValueInfoWrapperPass)
171 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
172 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
173 INITIALIZE_PASS_END(JumpThreading, "jump-threading",
174 "Jump Threading", false, false)
175
176 // Public interface to the Jump Threading pass
177 FunctionPass *llvm::createJumpThreadingPass(bool InsertFr, int Threshold) {
178 return new JumpThreading(InsertFr, Threshold);
179 }
180
JumpThreadingPass(bool InsertFr,int T)181 JumpThreadingPass::JumpThreadingPass(bool InsertFr, int T) {
182 InsertFreezeWhenUnfoldingSelect = JumpThreadingFreezeSelectCond | InsertFr;
183 DefaultBBDupThreshold = (T == -1) ? BBDuplicateThreshold : unsigned(T);
184 }
185
186 // Update branch probability information according to conditional
187 // branch probability. This is usually made possible for cloned branches
188 // in inline instances by the context specific profile in the caller.
189 // For instance,
190 //
191 // [Block PredBB]
192 // [Branch PredBr]
193 // if (t) {
194 // Block A;
195 // } else {
196 // Block B;
197 // }
198 //
199 // [Block BB]
200 // cond = PN([true, %A], [..., %B]); // PHI node
201 // [Branch CondBr]
202 // if (cond) {
203 // ... // P(cond == true) = 1%
204 // }
205 //
206 // Here we know that when block A is taken, cond must be true, which means
207 // P(cond == true | A) = 1
208 //
209 // Given that P(cond == true) = P(cond == true | A) * P(A) +
210 // P(cond == true | B) * P(B)
211 // we get:
212 // P(cond == true ) = P(A) + P(cond == true | B) * P(B)
213 //
214 // which gives us:
215 // P(A) is less than P(cond == true), i.e.
216 // P(t == true) <= P(cond == true)
217 //
218 // In other words, if we know P(cond == true) is unlikely, we know
219 // that P(t == true) is also unlikely.
220 //
updatePredecessorProfileMetadata(PHINode * PN,BasicBlock * BB)221 static void updatePredecessorProfileMetadata(PHINode *PN, BasicBlock *BB) {
222 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
223 if (!CondBr)
224 return;
225
226 uint64_t TrueWeight, FalseWeight;
227 if (!CondBr->extractProfMetadata(TrueWeight, FalseWeight))
228 return;
229
230 if (TrueWeight + FalseWeight == 0)
231 // Zero branch_weights do not give a hint for getting branch probabilities.
232 // Technically it would result in division by zero denominator, which is
233 // TrueWeight + FalseWeight.
234 return;
235
236 // Returns the outgoing edge of the dominating predecessor block
237 // that leads to the PhiNode's incoming block:
238 auto GetPredOutEdge =
239 [](BasicBlock *IncomingBB,
240 BasicBlock *PhiBB) -> std::pair<BasicBlock *, BasicBlock *> {
241 auto *PredBB = IncomingBB;
242 auto *SuccBB = PhiBB;
243 SmallPtrSet<BasicBlock *, 16> Visited;
244 while (true) {
245 BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator());
246 if (PredBr && PredBr->isConditional())
247 return {PredBB, SuccBB};
248 Visited.insert(PredBB);
249 auto *SinglePredBB = PredBB->getSinglePredecessor();
250 if (!SinglePredBB)
251 return {nullptr, nullptr};
252
253 // Stop searching when SinglePredBB has been visited. It means we see
254 // an unreachable loop.
255 if (Visited.count(SinglePredBB))
256 return {nullptr, nullptr};
257
258 SuccBB = PredBB;
259 PredBB = SinglePredBB;
260 }
261 };
262
263 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
264 Value *PhiOpnd = PN->getIncomingValue(i);
265 ConstantInt *CI = dyn_cast<ConstantInt>(PhiOpnd);
266
267 if (!CI || !CI->getType()->isIntegerTy(1))
268 continue;
269
270 BranchProbability BP =
271 (CI->isOne() ? BranchProbability::getBranchProbability(
272 TrueWeight, TrueWeight + FalseWeight)
273 : BranchProbability::getBranchProbability(
274 FalseWeight, TrueWeight + FalseWeight));
275
276 auto PredOutEdge = GetPredOutEdge(PN->getIncomingBlock(i), BB);
277 if (!PredOutEdge.first)
278 return;
279
280 BasicBlock *PredBB = PredOutEdge.first;
281 BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator());
282 if (!PredBr)
283 return;
284
285 uint64_t PredTrueWeight, PredFalseWeight;
286 // FIXME: We currently only set the profile data when it is missing.
287 // With PGO, this can be used to refine even existing profile data with
288 // context information. This needs to be done after more performance
289 // testing.
290 if (PredBr->extractProfMetadata(PredTrueWeight, PredFalseWeight))
291 continue;
292
293 // We can not infer anything useful when BP >= 50%, because BP is the
294 // upper bound probability value.
295 if (BP >= BranchProbability(50, 100))
296 continue;
297
298 SmallVector<uint32_t, 2> Weights;
299 if (PredBr->getSuccessor(0) == PredOutEdge.second) {
300 Weights.push_back(BP.getNumerator());
301 Weights.push_back(BP.getCompl().getNumerator());
302 } else {
303 Weights.push_back(BP.getCompl().getNumerator());
304 Weights.push_back(BP.getNumerator());
305 }
306 PredBr->setMetadata(LLVMContext::MD_prof,
307 MDBuilder(PredBr->getParent()->getContext())
308 .createBranchWeights(Weights));
309 }
310 }
311
312 /// runOnFunction - Toplevel algorithm.
runOnFunction(Function & F)313 bool JumpThreading::runOnFunction(Function &F) {
314 if (skipFunction(F))
315 return false;
316 auto TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
317 // Jump Threading has no sense for the targets with divergent CF
318 if (TTI->hasBranchDivergence())
319 return false;
320 auto TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
321 auto DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
322 auto LVI = &getAnalysis<LazyValueInfoWrapperPass>().getLVI();
323 auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
324 DomTreeUpdater DTU(*DT, DomTreeUpdater::UpdateStrategy::Lazy);
325 std::unique_ptr<BlockFrequencyInfo> BFI;
326 std::unique_ptr<BranchProbabilityInfo> BPI;
327 if (F.hasProfileData()) {
328 LoopInfo LI{DominatorTree(F)};
329 BPI.reset(new BranchProbabilityInfo(F, LI, TLI));
330 BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
331 }
332
333 bool Changed = Impl.runImpl(F, TLI, LVI, AA, &DTU, F.hasProfileData(),
334 std::move(BFI), std::move(BPI));
335 if (PrintLVIAfterJumpThreading) {
336 dbgs() << "LVI for function '" << F.getName() << "':\n";
337 LVI->printLVI(F, DTU.getDomTree(), dbgs());
338 }
339 return Changed;
340 }
341
run(Function & F,FunctionAnalysisManager & AM)342 PreservedAnalyses JumpThreadingPass::run(Function &F,
343 FunctionAnalysisManager &AM) {
344 auto &TTI = AM.getResult<TargetIRAnalysis>(F);
345 // Jump Threading has no sense for the targets with divergent CF
346 if (TTI.hasBranchDivergence())
347 return PreservedAnalyses::all();
348 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
349 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
350 auto &LVI = AM.getResult<LazyValueAnalysis>(F);
351 auto &AA = AM.getResult<AAManager>(F);
352 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Lazy);
353
354 std::unique_ptr<BlockFrequencyInfo> BFI;
355 std::unique_ptr<BranchProbabilityInfo> BPI;
356 if (F.hasProfileData()) {
357 LoopInfo LI{DominatorTree(F)};
358 BPI.reset(new BranchProbabilityInfo(F, LI, &TLI));
359 BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
360 }
361
362 bool Changed = runImpl(F, &TLI, &LVI, &AA, &DTU, F.hasProfileData(),
363 std::move(BFI), std::move(BPI));
364
365 if (PrintLVIAfterJumpThreading) {
366 dbgs() << "LVI for function '" << F.getName() << "':\n";
367 LVI.printLVI(F, DTU.getDomTree(), dbgs());
368 }
369
370 if (!Changed)
371 return PreservedAnalyses::all();
372 PreservedAnalyses PA;
373 PA.preserve<GlobalsAA>();
374 PA.preserve<DominatorTreeAnalysis>();
375 PA.preserve<LazyValueAnalysis>();
376 return PA;
377 }
378
runImpl(Function & F,TargetLibraryInfo * TLI_,LazyValueInfo * LVI_,AliasAnalysis * AA_,DomTreeUpdater * DTU_,bool HasProfileData_,std::unique_ptr<BlockFrequencyInfo> BFI_,std::unique_ptr<BranchProbabilityInfo> BPI_)379 bool JumpThreadingPass::runImpl(Function &F, TargetLibraryInfo *TLI_,
380 LazyValueInfo *LVI_, AliasAnalysis *AA_,
381 DomTreeUpdater *DTU_, bool HasProfileData_,
382 std::unique_ptr<BlockFrequencyInfo> BFI_,
383 std::unique_ptr<BranchProbabilityInfo> BPI_) {
384 LLVM_DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
385 TLI = TLI_;
386 LVI = LVI_;
387 AA = AA_;
388 DTU = DTU_;
389 BFI.reset();
390 BPI.reset();
391 // When profile data is available, we need to update edge weights after
392 // successful jump threading, which requires both BPI and BFI being available.
393 HasProfileData = HasProfileData_;
394 auto *GuardDecl = F.getParent()->getFunction(
395 Intrinsic::getName(Intrinsic::experimental_guard));
396 HasGuards = GuardDecl && !GuardDecl->use_empty();
397 if (HasProfileData) {
398 BPI = std::move(BPI_);
399 BFI = std::move(BFI_);
400 }
401
402 // Reduce the number of instructions duplicated when optimizing strictly for
403 // size.
404 if (BBDuplicateThreshold.getNumOccurrences())
405 BBDupThreshold = BBDuplicateThreshold;
406 else if (F.hasFnAttribute(Attribute::MinSize))
407 BBDupThreshold = 3;
408 else
409 BBDupThreshold = DefaultBBDupThreshold;
410
411 // JumpThreading must not processes blocks unreachable from entry. It's a
412 // waste of compute time and can potentially lead to hangs.
413 SmallPtrSet<BasicBlock *, 16> Unreachable;
414 assert(DTU && "DTU isn't passed into JumpThreading before using it.");
415 assert(DTU->hasDomTree() && "JumpThreading relies on DomTree to proceed.");
416 DominatorTree &DT = DTU->getDomTree();
417 for (auto &BB : F)
418 if (!DT.isReachableFromEntry(&BB))
419 Unreachable.insert(&BB);
420
421 if (!ThreadAcrossLoopHeaders)
422 findLoopHeaders(F);
423
424 bool EverChanged = false;
425 bool Changed;
426 do {
427 Changed = false;
428 for (auto &BB : F) {
429 if (Unreachable.count(&BB))
430 continue;
431 while (processBlock(&BB)) // Thread all of the branches we can over BB.
432 Changed = true;
433
434 // Jump threading may have introduced redundant debug values into BB
435 // which should be removed.
436 if (Changed)
437 RemoveRedundantDbgInstrs(&BB);
438
439 // Stop processing BB if it's the entry or is now deleted. The following
440 // routines attempt to eliminate BB and locating a suitable replacement
441 // for the entry is non-trivial.
442 if (&BB == &F.getEntryBlock() || DTU->isBBPendingDeletion(&BB))
443 continue;
444
445 if (pred_empty(&BB)) {
446 // When processBlock makes BB unreachable it doesn't bother to fix up
447 // the instructions in it. We must remove BB to prevent invalid IR.
448 LLVM_DEBUG(dbgs() << " JT: Deleting dead block '" << BB.getName()
449 << "' with terminator: " << *BB.getTerminator()
450 << '\n');
451 LoopHeaders.erase(&BB);
452 LVI->eraseBlock(&BB);
453 DeleteDeadBlock(&BB, DTU);
454 Changed = true;
455 continue;
456 }
457
458 // processBlock doesn't thread BBs with unconditional TIs. However, if BB
459 // is "almost empty", we attempt to merge BB with its sole successor.
460 auto *BI = dyn_cast<BranchInst>(BB.getTerminator());
461 if (BI && BI->isUnconditional()) {
462 BasicBlock *Succ = BI->getSuccessor(0);
463 if (
464 // The terminator must be the only non-phi instruction in BB.
465 BB.getFirstNonPHIOrDbg()->isTerminator() &&
466 // Don't alter Loop headers and latches to ensure another pass can
467 // detect and transform nested loops later.
468 !LoopHeaders.count(&BB) && !LoopHeaders.count(Succ) &&
469 TryToSimplifyUncondBranchFromEmptyBlock(&BB, DTU)) {
470 RemoveRedundantDbgInstrs(Succ);
471 // BB is valid for cleanup here because we passed in DTU. F remains
472 // BB's parent until a DTU->getDomTree() event.
473 LVI->eraseBlock(&BB);
474 Changed = true;
475 }
476 }
477 }
478 EverChanged |= Changed;
479 } while (Changed);
480
481 LoopHeaders.clear();
482 return EverChanged;
483 }
484
485 // Replace uses of Cond with ToVal when safe to do so. If all uses are
486 // replaced, we can remove Cond. We cannot blindly replace all uses of Cond
487 // because we may incorrectly replace uses when guards/assumes are uses of
488 // of `Cond` and we used the guards/assume to reason about the `Cond` value
489 // at the end of block. RAUW unconditionally replaces all uses
490 // including the guards/assumes themselves and the uses before the
491 // guard/assume.
replaceFoldableUses(Instruction * Cond,Value * ToVal)492 static void replaceFoldableUses(Instruction *Cond, Value *ToVal) {
493 assert(Cond->getType() == ToVal->getType());
494 auto *BB = Cond->getParent();
495 // We can unconditionally replace all uses in non-local blocks (i.e. uses
496 // strictly dominated by BB), since LVI information is true from the
497 // terminator of BB.
498 replaceNonLocalUsesWith(Cond, ToVal);
499 for (Instruction &I : reverse(*BB)) {
500 // Reached the Cond whose uses we are trying to replace, so there are no
501 // more uses.
502 if (&I == Cond)
503 break;
504 // We only replace uses in instructions that are guaranteed to reach the end
505 // of BB, where we know Cond is ToVal.
506 if (!isGuaranteedToTransferExecutionToSuccessor(&I))
507 break;
508 I.replaceUsesOfWith(Cond, ToVal);
509 }
510 if (Cond->use_empty() && !Cond->mayHaveSideEffects())
511 Cond->eraseFromParent();
512 }
513
514 /// Return the cost of duplicating a piece of this block from first non-phi
515 /// and before StopAt instruction to thread across it. Stop scanning the block
516 /// when exceeding the threshold. If duplication is impossible, returns ~0U.
getJumpThreadDuplicationCost(BasicBlock * BB,Instruction * StopAt,unsigned Threshold)517 static unsigned getJumpThreadDuplicationCost(BasicBlock *BB,
518 Instruction *StopAt,
519 unsigned Threshold) {
520 assert(StopAt->getParent() == BB && "Not an instruction from proper BB?");
521 /// Ignore PHI nodes, these will be flattened when duplication happens.
522 BasicBlock::const_iterator I(BB->getFirstNonPHI());
523
524 // FIXME: THREADING will delete values that are just used to compute the
525 // branch, so they shouldn't count against the duplication cost.
526
527 unsigned Bonus = 0;
528 if (BB->getTerminator() == StopAt) {
529 // Threading through a switch statement is particularly profitable. If this
530 // block ends in a switch, decrease its cost to make it more likely to
531 // happen.
532 if (isa<SwitchInst>(StopAt))
533 Bonus = 6;
534
535 // The same holds for indirect branches, but slightly more so.
536 if (isa<IndirectBrInst>(StopAt))
537 Bonus = 8;
538 }
539
540 // Bump the threshold up so the early exit from the loop doesn't skip the
541 // terminator-based Size adjustment at the end.
542 Threshold += Bonus;
543
544 // Sum up the cost of each instruction until we get to the terminator. Don't
545 // include the terminator because the copy won't include it.
546 unsigned Size = 0;
547 for (; &*I != StopAt; ++I) {
548
549 // Stop scanning the block if we've reached the threshold.
550 if (Size > Threshold)
551 return Size;
552
553 // Debugger intrinsics don't incur code size.
554 if (isa<DbgInfoIntrinsic>(I)) continue;
555
556 // Pseudo-probes don't incur code size.
557 if (isa<PseudoProbeInst>(I))
558 continue;
559
560 // If this is a pointer->pointer bitcast, it is free.
561 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
562 continue;
563
564 // Freeze instruction is free, too.
565 if (isa<FreezeInst>(I))
566 continue;
567
568 // Bail out if this instruction gives back a token type, it is not possible
569 // to duplicate it if it is used outside this BB.
570 if (I->getType()->isTokenTy() && I->isUsedOutsideOfBlock(BB))
571 return ~0U;
572
573 // All other instructions count for at least one unit.
574 ++Size;
575
576 // Calls are more expensive. If they are non-intrinsic calls, we model them
577 // as having cost of 4. If they are a non-vector intrinsic, we model them
578 // as having cost of 2 total, and if they are a vector intrinsic, we model
579 // them as having cost 1.
580 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
581 if (CI->cannotDuplicate() || CI->isConvergent())
582 // Blocks with NoDuplicate are modelled as having infinite cost, so they
583 // are never duplicated.
584 return ~0U;
585 else if (!isa<IntrinsicInst>(CI))
586 Size += 3;
587 else if (!CI->getType()->isVectorTy())
588 Size += 1;
589 }
590 }
591
592 return Size > Bonus ? Size - Bonus : 0;
593 }
594
595 /// findLoopHeaders - We do not want jump threading to turn proper loop
596 /// structures into irreducible loops. Doing this breaks up the loop nesting
597 /// hierarchy and pessimizes later transformations. To prevent this from
598 /// happening, we first have to find the loop headers. Here we approximate this
599 /// by finding targets of backedges in the CFG.
600 ///
601 /// Note that there definitely are cases when we want to allow threading of
602 /// edges across a loop header. For example, threading a jump from outside the
603 /// loop (the preheader) to an exit block of the loop is definitely profitable.
604 /// It is also almost always profitable to thread backedges from within the loop
605 /// to exit blocks, and is often profitable to thread backedges to other blocks
606 /// within the loop (forming a nested loop). This simple analysis is not rich
607 /// enough to track all of these properties and keep it up-to-date as the CFG
608 /// mutates, so we don't allow any of these transformations.
findLoopHeaders(Function & F)609 void JumpThreadingPass::findLoopHeaders(Function &F) {
610 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
611 FindFunctionBackedges(F, Edges);
612
613 for (const auto &Edge : Edges)
614 LoopHeaders.insert(Edge.second);
615 }
616
617 /// getKnownConstant - Helper method to determine if we can thread over a
618 /// terminator with the given value as its condition, and if so what value to
619 /// use for that. What kind of value this is depends on whether we want an
620 /// integer or a block address, but an undef is always accepted.
621 /// Returns null if Val is null or not an appropriate constant.
getKnownConstant(Value * Val,ConstantPreference Preference)622 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
623 if (!Val)
624 return nullptr;
625
626 // Undef is "known" enough.
627 if (UndefValue *U = dyn_cast<UndefValue>(Val))
628 return U;
629
630 if (Preference == WantBlockAddress)
631 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
632
633 return dyn_cast<ConstantInt>(Val);
634 }
635
636 /// computeValueKnownInPredecessors - Given a basic block BB and a value V, see
637 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
638 /// in any of our predecessors. If so, return the known list of value and pred
639 /// BB in the result vector.
640 ///
641 /// This returns true if there were any known values.
computeValueKnownInPredecessorsImpl(Value * V,BasicBlock * BB,PredValueInfo & Result,ConstantPreference Preference,DenseSet<Value * > & RecursionSet,Instruction * CxtI)642 bool JumpThreadingPass::computeValueKnownInPredecessorsImpl(
643 Value *V, BasicBlock *BB, PredValueInfo &Result,
644 ConstantPreference Preference, DenseSet<Value *> &RecursionSet,
645 Instruction *CxtI) {
646 // This method walks up use-def chains recursively. Because of this, we could
647 // get into an infinite loop going around loops in the use-def chain. To
648 // prevent this, keep track of what (value, block) pairs we've already visited
649 // and terminate the search if we loop back to them
650 if (!RecursionSet.insert(V).second)
651 return false;
652
653 // If V is a constant, then it is known in all predecessors.
654 if (Constant *KC = getKnownConstant(V, Preference)) {
655 for (BasicBlock *Pred : predecessors(BB))
656 Result.emplace_back(KC, Pred);
657
658 return !Result.empty();
659 }
660
661 // If V is a non-instruction value, or an instruction in a different block,
662 // then it can't be derived from a PHI.
663 Instruction *I = dyn_cast<Instruction>(V);
664 if (!I || I->getParent() != BB) {
665
666 // Okay, if this is a live-in value, see if it has a known value at the end
667 // of any of our predecessors.
668 //
669 // FIXME: This should be an edge property, not a block end property.
670 /// TODO: Per PR2563, we could infer value range information about a
671 /// predecessor based on its terminator.
672 //
673 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
674 // "I" is a non-local compare-with-a-constant instruction. This would be
675 // able to handle value inequalities better, for example if the compare is
676 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
677 // Perhaps getConstantOnEdge should be smart enough to do this?
678 for (BasicBlock *P : predecessors(BB)) {
679 // If the value is known by LazyValueInfo to be a constant in a
680 // predecessor, use that information to try to thread this block.
681 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB, CxtI);
682 if (Constant *KC = getKnownConstant(PredCst, Preference))
683 Result.emplace_back(KC, P);
684 }
685
686 return !Result.empty();
687 }
688
689 /// If I is a PHI node, then we know the incoming values for any constants.
690 if (PHINode *PN = dyn_cast<PHINode>(I)) {
691 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
692 Value *InVal = PN->getIncomingValue(i);
693 if (Constant *KC = getKnownConstant(InVal, Preference)) {
694 Result.emplace_back(KC, PN->getIncomingBlock(i));
695 } else {
696 Constant *CI = LVI->getConstantOnEdge(InVal,
697 PN->getIncomingBlock(i),
698 BB, CxtI);
699 if (Constant *KC = getKnownConstant(CI, Preference))
700 Result.emplace_back(KC, PN->getIncomingBlock(i));
701 }
702 }
703
704 return !Result.empty();
705 }
706
707 // Handle Cast instructions.
708 if (CastInst *CI = dyn_cast<CastInst>(I)) {
709 Value *Source = CI->getOperand(0);
710 computeValueKnownInPredecessorsImpl(Source, BB, Result, Preference,
711 RecursionSet, CxtI);
712 if (Result.empty())
713 return false;
714
715 // Convert the known values.
716 for (auto &R : Result)
717 R.first = ConstantExpr::getCast(CI->getOpcode(), R.first, CI->getType());
718
719 return true;
720 }
721
722 if (FreezeInst *FI = dyn_cast<FreezeInst>(I)) {
723 Value *Source = FI->getOperand(0);
724 computeValueKnownInPredecessorsImpl(Source, BB, Result, Preference,
725 RecursionSet, CxtI);
726
727 erase_if(Result, [](auto &Pair) {
728 return !isGuaranteedNotToBeUndefOrPoison(Pair.first);
729 });
730
731 return !Result.empty();
732 }
733
734 // Handle some boolean conditions.
735 if (I->getType()->getPrimitiveSizeInBits() == 1) {
736 assert(Preference == WantInteger && "One-bit non-integer type?");
737 // X | true -> true
738 // X & false -> false
739 if (I->getOpcode() == Instruction::Or ||
740 I->getOpcode() == Instruction::And) {
741 PredValueInfoTy LHSVals, RHSVals;
742
743 computeValueKnownInPredecessorsImpl(I->getOperand(0), BB, LHSVals,
744 WantInteger, RecursionSet, CxtI);
745 computeValueKnownInPredecessorsImpl(I->getOperand(1), BB, RHSVals,
746 WantInteger, RecursionSet, CxtI);
747
748 if (LHSVals.empty() && RHSVals.empty())
749 return false;
750
751 ConstantInt *InterestingVal;
752 if (I->getOpcode() == Instruction::Or)
753 InterestingVal = ConstantInt::getTrue(I->getContext());
754 else
755 InterestingVal = ConstantInt::getFalse(I->getContext());
756
757 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
758
759 // Scan for the sentinel. If we find an undef, force it to the
760 // interesting value: x|undef -> true and x&undef -> false.
761 for (const auto &LHSVal : LHSVals)
762 if (LHSVal.first == InterestingVal || isa<UndefValue>(LHSVal.first)) {
763 Result.emplace_back(InterestingVal, LHSVal.second);
764 LHSKnownBBs.insert(LHSVal.second);
765 }
766 for (const auto &RHSVal : RHSVals)
767 if (RHSVal.first == InterestingVal || isa<UndefValue>(RHSVal.first)) {
768 // If we already inferred a value for this block on the LHS, don't
769 // re-add it.
770 if (!LHSKnownBBs.count(RHSVal.second))
771 Result.emplace_back(InterestingVal, RHSVal.second);
772 }
773
774 return !Result.empty();
775 }
776
777 // Handle the NOT form of XOR.
778 if (I->getOpcode() == Instruction::Xor &&
779 isa<ConstantInt>(I->getOperand(1)) &&
780 cast<ConstantInt>(I->getOperand(1))->isOne()) {
781 computeValueKnownInPredecessorsImpl(I->getOperand(0), BB, Result,
782 WantInteger, RecursionSet, CxtI);
783 if (Result.empty())
784 return false;
785
786 // Invert the known values.
787 for (auto &R : Result)
788 R.first = ConstantExpr::getNot(R.first);
789
790 return true;
791 }
792
793 // Try to simplify some other binary operator values.
794 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
795 assert(Preference != WantBlockAddress
796 && "A binary operator creating a block address?");
797 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
798 PredValueInfoTy LHSVals;
799 computeValueKnownInPredecessorsImpl(BO->getOperand(0), BB, LHSVals,
800 WantInteger, RecursionSet, CxtI);
801
802 // Try to use constant folding to simplify the binary operator.
803 for (const auto &LHSVal : LHSVals) {
804 Constant *V = LHSVal.first;
805 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
806
807 if (Constant *KC = getKnownConstant(Folded, WantInteger))
808 Result.emplace_back(KC, LHSVal.second);
809 }
810 }
811
812 return !Result.empty();
813 }
814
815 // Handle compare with phi operand, where the PHI is defined in this block.
816 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
817 assert(Preference == WantInteger && "Compares only produce integers");
818 Type *CmpType = Cmp->getType();
819 Value *CmpLHS = Cmp->getOperand(0);
820 Value *CmpRHS = Cmp->getOperand(1);
821 CmpInst::Predicate Pred = Cmp->getPredicate();
822
823 PHINode *PN = dyn_cast<PHINode>(CmpLHS);
824 if (!PN)
825 PN = dyn_cast<PHINode>(CmpRHS);
826 if (PN && PN->getParent() == BB) {
827 const DataLayout &DL = PN->getModule()->getDataLayout();
828 // We can do this simplification if any comparisons fold to true or false.
829 // See if any do.
830 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
831 BasicBlock *PredBB = PN->getIncomingBlock(i);
832 Value *LHS, *RHS;
833 if (PN == CmpLHS) {
834 LHS = PN->getIncomingValue(i);
835 RHS = CmpRHS->DoPHITranslation(BB, PredBB);
836 } else {
837 LHS = CmpLHS->DoPHITranslation(BB, PredBB);
838 RHS = PN->getIncomingValue(i);
839 }
840 Value *Res = SimplifyCmpInst(Pred, LHS, RHS, {DL});
841 if (!Res) {
842 if (!isa<Constant>(RHS))
843 continue;
844
845 // getPredicateOnEdge call will make no sense if LHS is defined in BB.
846 auto LHSInst = dyn_cast<Instruction>(LHS);
847 if (LHSInst && LHSInst->getParent() == BB)
848 continue;
849
850 LazyValueInfo::Tristate
851 ResT = LVI->getPredicateOnEdge(Pred, LHS,
852 cast<Constant>(RHS), PredBB, BB,
853 CxtI ? CxtI : Cmp);
854 if (ResT == LazyValueInfo::Unknown)
855 continue;
856 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
857 }
858
859 if (Constant *KC = getKnownConstant(Res, WantInteger))
860 Result.emplace_back(KC, PredBB);
861 }
862
863 return !Result.empty();
864 }
865
866 // If comparing a live-in value against a constant, see if we know the
867 // live-in value on any predecessors.
868 if (isa<Constant>(CmpRHS) && !CmpType->isVectorTy()) {
869 Constant *CmpConst = cast<Constant>(CmpRHS);
870
871 if (!isa<Instruction>(CmpLHS) ||
872 cast<Instruction>(CmpLHS)->getParent() != BB) {
873 for (BasicBlock *P : predecessors(BB)) {
874 // If the value is known by LazyValueInfo to be a constant in a
875 // predecessor, use that information to try to thread this block.
876 LazyValueInfo::Tristate Res =
877 LVI->getPredicateOnEdge(Pred, CmpLHS,
878 CmpConst, P, BB, CxtI ? CxtI : Cmp);
879 if (Res == LazyValueInfo::Unknown)
880 continue;
881
882 Constant *ResC = ConstantInt::get(CmpType, Res);
883 Result.emplace_back(ResC, P);
884 }
885
886 return !Result.empty();
887 }
888
889 // InstCombine can fold some forms of constant range checks into
890 // (icmp (add (x, C1)), C2). See if we have we have such a thing with
891 // x as a live-in.
892 {
893 using namespace PatternMatch;
894
895 Value *AddLHS;
896 ConstantInt *AddConst;
897 if (isa<ConstantInt>(CmpConst) &&
898 match(CmpLHS, m_Add(m_Value(AddLHS), m_ConstantInt(AddConst)))) {
899 if (!isa<Instruction>(AddLHS) ||
900 cast<Instruction>(AddLHS)->getParent() != BB) {
901 for (BasicBlock *P : predecessors(BB)) {
902 // If the value is known by LazyValueInfo to be a ConstantRange in
903 // a predecessor, use that information to try to thread this
904 // block.
905 ConstantRange CR = LVI->getConstantRangeOnEdge(
906 AddLHS, P, BB, CxtI ? CxtI : cast<Instruction>(CmpLHS));
907 // Propagate the range through the addition.
908 CR = CR.add(AddConst->getValue());
909
910 // Get the range where the compare returns true.
911 ConstantRange CmpRange = ConstantRange::makeExactICmpRegion(
912 Pred, cast<ConstantInt>(CmpConst)->getValue());
913
914 Constant *ResC;
915 if (CmpRange.contains(CR))
916 ResC = ConstantInt::getTrue(CmpType);
917 else if (CmpRange.inverse().contains(CR))
918 ResC = ConstantInt::getFalse(CmpType);
919 else
920 continue;
921
922 Result.emplace_back(ResC, P);
923 }
924
925 return !Result.empty();
926 }
927 }
928 }
929
930 // Try to find a constant value for the LHS of a comparison,
931 // and evaluate it statically if we can.
932 PredValueInfoTy LHSVals;
933 computeValueKnownInPredecessorsImpl(I->getOperand(0), BB, LHSVals,
934 WantInteger, RecursionSet, CxtI);
935
936 for (const auto &LHSVal : LHSVals) {
937 Constant *V = LHSVal.first;
938 Constant *Folded = ConstantExpr::getCompare(Pred, V, CmpConst);
939 if (Constant *KC = getKnownConstant(Folded, WantInteger))
940 Result.emplace_back(KC, LHSVal.second);
941 }
942
943 return !Result.empty();
944 }
945 }
946
947 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
948 // Handle select instructions where at least one operand is a known constant
949 // and we can figure out the condition value for any predecessor block.
950 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
951 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
952 PredValueInfoTy Conds;
953 if ((TrueVal || FalseVal) &&
954 computeValueKnownInPredecessorsImpl(SI->getCondition(), BB, Conds,
955 WantInteger, RecursionSet, CxtI)) {
956 for (auto &C : Conds) {
957 Constant *Cond = C.first;
958
959 // Figure out what value to use for the condition.
960 bool KnownCond;
961 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
962 // A known boolean.
963 KnownCond = CI->isOne();
964 } else {
965 assert(isa<UndefValue>(Cond) && "Unexpected condition value");
966 // Either operand will do, so be sure to pick the one that's a known
967 // constant.
968 // FIXME: Do this more cleverly if both values are known constants?
969 KnownCond = (TrueVal != nullptr);
970 }
971
972 // See if the select has a known constant value for this predecessor.
973 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
974 Result.emplace_back(Val, C.second);
975 }
976
977 return !Result.empty();
978 }
979 }
980
981 // If all else fails, see if LVI can figure out a constant value for us.
982 assert(CxtI->getParent() == BB && "CxtI should be in BB");
983 Constant *CI = LVI->getConstant(V, CxtI);
984 if (Constant *KC = getKnownConstant(CI, Preference)) {
985 for (BasicBlock *Pred : predecessors(BB))
986 Result.emplace_back(KC, Pred);
987 }
988
989 return !Result.empty();
990 }
991
992 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
993 /// in an undefined jump, decide which block is best to revector to.
994 ///
995 /// Since we can pick an arbitrary destination, we pick the successor with the
996 /// fewest predecessors. This should reduce the in-degree of the others.
getBestDestForJumpOnUndef(BasicBlock * BB)997 static unsigned getBestDestForJumpOnUndef(BasicBlock *BB) {
998 Instruction *BBTerm = BB->getTerminator();
999 unsigned MinSucc = 0;
1000 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
1001 // Compute the successor with the minimum number of predecessors.
1002 unsigned MinNumPreds = pred_size(TestBB);
1003 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
1004 TestBB = BBTerm->getSuccessor(i);
1005 unsigned NumPreds = pred_size(TestBB);
1006 if (NumPreds < MinNumPreds) {
1007 MinSucc = i;
1008 MinNumPreds = NumPreds;
1009 }
1010 }
1011
1012 return MinSucc;
1013 }
1014
hasAddressTakenAndUsed(BasicBlock * BB)1015 static bool hasAddressTakenAndUsed(BasicBlock *BB) {
1016 if (!BB->hasAddressTaken()) return false;
1017
1018 // If the block has its address taken, it may be a tree of dead constants
1019 // hanging off of it. These shouldn't keep the block alive.
1020 BlockAddress *BA = BlockAddress::get(BB);
1021 BA->removeDeadConstantUsers();
1022 return !BA->use_empty();
1023 }
1024
1025 /// processBlock - If there are any predecessors whose control can be threaded
1026 /// through to a successor, transform them now.
processBlock(BasicBlock * BB)1027 bool JumpThreadingPass::processBlock(BasicBlock *BB) {
1028 // If the block is trivially dead, just return and let the caller nuke it.
1029 // This simplifies other transformations.
1030 if (DTU->isBBPendingDeletion(BB) ||
1031 (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()))
1032 return false;
1033
1034 // If this block has a single predecessor, and if that pred has a single
1035 // successor, merge the blocks. This encourages recursive jump threading
1036 // because now the condition in this block can be threaded through
1037 // predecessors of our predecessor block.
1038 if (maybeMergeBasicBlockIntoOnlyPred(BB))
1039 return true;
1040
1041 if (tryToUnfoldSelectInCurrBB(BB))
1042 return true;
1043
1044 // Look if we can propagate guards to predecessors.
1045 if (HasGuards && processGuards(BB))
1046 return true;
1047
1048 // What kind of constant we're looking for.
1049 ConstantPreference Preference = WantInteger;
1050
1051 // Look to see if the terminator is a conditional branch, switch or indirect
1052 // branch, if not we can't thread it.
1053 Value *Condition;
1054 Instruction *Terminator = BB->getTerminator();
1055 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
1056 // Can't thread an unconditional jump.
1057 if (BI->isUnconditional()) return false;
1058 Condition = BI->getCondition();
1059 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
1060 Condition = SI->getCondition();
1061 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
1062 // Can't thread indirect branch with no successors.
1063 if (IB->getNumSuccessors() == 0) return false;
1064 Condition = IB->getAddress()->stripPointerCasts();
1065 Preference = WantBlockAddress;
1066 } else {
1067 return false; // Must be an invoke or callbr.
1068 }
1069
1070 // Keep track if we constant folded the condition in this invocation.
1071 bool ConstantFolded = false;
1072
1073 // Run constant folding to see if we can reduce the condition to a simple
1074 // constant.
1075 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
1076 Value *SimpleVal =
1077 ConstantFoldInstruction(I, BB->getModule()->getDataLayout(), TLI);
1078 if (SimpleVal) {
1079 I->replaceAllUsesWith(SimpleVal);
1080 if (isInstructionTriviallyDead(I, TLI))
1081 I->eraseFromParent();
1082 Condition = SimpleVal;
1083 ConstantFolded = true;
1084 }
1085 }
1086
1087 // If the terminator is branching on an undef or freeze undef, we can pick any
1088 // of the successors to branch to. Let getBestDestForJumpOnUndef decide.
1089 auto *FI = dyn_cast<FreezeInst>(Condition);
1090 if (isa<UndefValue>(Condition) ||
1091 (FI && isa<UndefValue>(FI->getOperand(0)) && FI->hasOneUse())) {
1092 unsigned BestSucc = getBestDestForJumpOnUndef(BB);
1093 std::vector<DominatorTree::UpdateType> Updates;
1094
1095 // Fold the branch/switch.
1096 Instruction *BBTerm = BB->getTerminator();
1097 Updates.reserve(BBTerm->getNumSuccessors());
1098 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
1099 if (i == BestSucc) continue;
1100 BasicBlock *Succ = BBTerm->getSuccessor(i);
1101 Succ->removePredecessor(BB, true);
1102 Updates.push_back({DominatorTree::Delete, BB, Succ});
1103 }
1104
1105 LLVM_DEBUG(dbgs() << " In block '" << BB->getName()
1106 << "' folding undef terminator: " << *BBTerm << '\n');
1107 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
1108 BBTerm->eraseFromParent();
1109 DTU->applyUpdatesPermissive(Updates);
1110 if (FI)
1111 FI->eraseFromParent();
1112 return true;
1113 }
1114
1115 // If the terminator of this block is branching on a constant, simplify the
1116 // terminator to an unconditional branch. This can occur due to threading in
1117 // other blocks.
1118 if (getKnownConstant(Condition, Preference)) {
1119 LLVM_DEBUG(dbgs() << " In block '" << BB->getName()
1120 << "' folding terminator: " << *BB->getTerminator()
1121 << '\n');
1122 ++NumFolds;
1123 ConstantFoldTerminator(BB, true, nullptr, DTU);
1124 if (HasProfileData)
1125 BPI->eraseBlock(BB);
1126 return true;
1127 }
1128
1129 Instruction *CondInst = dyn_cast<Instruction>(Condition);
1130
1131 // All the rest of our checks depend on the condition being an instruction.
1132 if (!CondInst) {
1133 // FIXME: Unify this with code below.
1134 if (processThreadableEdges(Condition, BB, Preference, Terminator))
1135 return true;
1136 return ConstantFolded;
1137 }
1138
1139 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
1140 // If we're branching on a conditional, LVI might be able to determine
1141 // it's value at the branch instruction. We only handle comparisons
1142 // against a constant at this time.
1143 // TODO: This should be extended to handle switches as well.
1144 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1145 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
1146 if (CondBr && CondConst) {
1147 // We should have returned as soon as we turn a conditional branch to
1148 // unconditional. Because its no longer interesting as far as jump
1149 // threading is concerned.
1150 assert(CondBr->isConditional() && "Threading on unconditional terminator");
1151
1152 LazyValueInfo::Tristate Ret =
1153 LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0),
1154 CondConst, CondBr);
1155 if (Ret != LazyValueInfo::Unknown) {
1156 unsigned ToRemove = Ret == LazyValueInfo::True ? 1 : 0;
1157 unsigned ToKeep = Ret == LazyValueInfo::True ? 0 : 1;
1158 BasicBlock *ToRemoveSucc = CondBr->getSuccessor(ToRemove);
1159 ToRemoveSucc->removePredecessor(BB, true);
1160 BranchInst *UncondBr =
1161 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
1162 UncondBr->setDebugLoc(CondBr->getDebugLoc());
1163 CondBr->eraseFromParent();
1164 if (CondCmp->use_empty())
1165 CondCmp->eraseFromParent();
1166 // We can safely replace *some* uses of the CondInst if it has
1167 // exactly one value as returned by LVI. RAUW is incorrect in the
1168 // presence of guards and assumes, that have the `Cond` as the use. This
1169 // is because we use the guards/assume to reason about the `Cond` value
1170 // at the end of block, but RAUW unconditionally replaces all uses
1171 // including the guards/assumes themselves and the uses before the
1172 // guard/assume.
1173 else if (CondCmp->getParent() == BB) {
1174 auto *CI = Ret == LazyValueInfo::True ?
1175 ConstantInt::getTrue(CondCmp->getType()) :
1176 ConstantInt::getFalse(CondCmp->getType());
1177 replaceFoldableUses(CondCmp, CI);
1178 }
1179 DTU->applyUpdatesPermissive(
1180 {{DominatorTree::Delete, BB, ToRemoveSucc}});
1181 if (HasProfileData)
1182 BPI->eraseBlock(BB);
1183 return true;
1184 }
1185
1186 // We did not manage to simplify this branch, try to see whether
1187 // CondCmp depends on a known phi-select pattern.
1188 if (tryToUnfoldSelect(CondCmp, BB))
1189 return true;
1190 }
1191 }
1192
1193 if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
1194 if (tryToUnfoldSelect(SI, BB))
1195 return true;
1196
1197 // Check for some cases that are worth simplifying. Right now we want to look
1198 // for loads that are used by a switch or by the condition for the branch. If
1199 // we see one, check to see if it's partially redundant. If so, insert a PHI
1200 // which can then be used to thread the values.
1201 Value *SimplifyValue = CondInst;
1202
1203 if (auto *FI = dyn_cast<FreezeInst>(SimplifyValue))
1204 // Look into freeze's operand
1205 SimplifyValue = FI->getOperand(0);
1206
1207 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
1208 if (isa<Constant>(CondCmp->getOperand(1)))
1209 SimplifyValue = CondCmp->getOperand(0);
1210
1211 // TODO: There are other places where load PRE would be profitable, such as
1212 // more complex comparisons.
1213 if (LoadInst *LoadI = dyn_cast<LoadInst>(SimplifyValue))
1214 if (simplifyPartiallyRedundantLoad(LoadI))
1215 return true;
1216
1217 // Before threading, try to propagate profile data backwards:
1218 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
1219 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1220 updatePredecessorProfileMetadata(PN, BB);
1221
1222 // Handle a variety of cases where we are branching on something derived from
1223 // a PHI node in the current block. If we can prove that any predecessors
1224 // compute a predictable value based on a PHI node, thread those predecessors.
1225 if (processThreadableEdges(CondInst, BB, Preference, Terminator))
1226 return true;
1227
1228 // If this is an otherwise-unfoldable branch on a phi node or freeze(phi) in
1229 // the current block, see if we can simplify.
1230 PHINode *PN = dyn_cast<PHINode>(
1231 isa<FreezeInst>(CondInst) ? cast<FreezeInst>(CondInst)->getOperand(0)
1232 : CondInst);
1233
1234 if (PN && PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1235 return processBranchOnPHI(PN);
1236
1237 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
1238 if (CondInst->getOpcode() == Instruction::Xor &&
1239 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1240 return processBranchOnXOR(cast<BinaryOperator>(CondInst));
1241
1242 // Search for a stronger dominating condition that can be used to simplify a
1243 // conditional branch leaving BB.
1244 if (processImpliedCondition(BB))
1245 return true;
1246
1247 return false;
1248 }
1249
processImpliedCondition(BasicBlock * BB)1250 bool JumpThreadingPass::processImpliedCondition(BasicBlock *BB) {
1251 auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
1252 if (!BI || !BI->isConditional())
1253 return false;
1254
1255 Value *Cond = BI->getCondition();
1256 BasicBlock *CurrentBB = BB;
1257 BasicBlock *CurrentPred = BB->getSinglePredecessor();
1258 unsigned Iter = 0;
1259
1260 auto &DL = BB->getModule()->getDataLayout();
1261
1262 while (CurrentPred && Iter++ < ImplicationSearchThreshold) {
1263 auto *PBI = dyn_cast<BranchInst>(CurrentPred->getTerminator());
1264 if (!PBI || !PBI->isConditional())
1265 return false;
1266 if (PBI->getSuccessor(0) != CurrentBB && PBI->getSuccessor(1) != CurrentBB)
1267 return false;
1268
1269 bool CondIsTrue = PBI->getSuccessor(0) == CurrentBB;
1270 Optional<bool> Implication =
1271 isImpliedCondition(PBI->getCondition(), Cond, DL, CondIsTrue);
1272 if (Implication) {
1273 BasicBlock *KeepSucc = BI->getSuccessor(*Implication ? 0 : 1);
1274 BasicBlock *RemoveSucc = BI->getSuccessor(*Implication ? 1 : 0);
1275 RemoveSucc->removePredecessor(BB);
1276 BranchInst *UncondBI = BranchInst::Create(KeepSucc, BI);
1277 UncondBI->setDebugLoc(BI->getDebugLoc());
1278 BI->eraseFromParent();
1279 DTU->applyUpdatesPermissive({{DominatorTree::Delete, BB, RemoveSucc}});
1280 if (HasProfileData)
1281 BPI->eraseBlock(BB);
1282 return true;
1283 }
1284 CurrentBB = CurrentPred;
1285 CurrentPred = CurrentBB->getSinglePredecessor();
1286 }
1287
1288 return false;
1289 }
1290
1291 /// Return true if Op is an instruction defined in the given block.
isOpDefinedInBlock(Value * Op,BasicBlock * BB)1292 static bool isOpDefinedInBlock(Value *Op, BasicBlock *BB) {
1293 if (Instruction *OpInst = dyn_cast<Instruction>(Op))
1294 if (OpInst->getParent() == BB)
1295 return true;
1296 return false;
1297 }
1298
1299 /// simplifyPartiallyRedundantLoad - If LoadI is an obviously partially
1300 /// redundant load instruction, eliminate it by replacing it with a PHI node.
1301 /// This is an important optimization that encourages jump threading, and needs
1302 /// to be run interlaced with other jump threading tasks.
simplifyPartiallyRedundantLoad(LoadInst * LoadI)1303 bool JumpThreadingPass::simplifyPartiallyRedundantLoad(LoadInst *LoadI) {
1304 // Don't hack volatile and ordered loads.
1305 if (!LoadI->isUnordered()) return false;
1306
1307 // If the load is defined in a block with exactly one predecessor, it can't be
1308 // partially redundant.
1309 BasicBlock *LoadBB = LoadI->getParent();
1310 if (LoadBB->getSinglePredecessor())
1311 return false;
1312
1313 // If the load is defined in an EH pad, it can't be partially redundant,
1314 // because the edges between the invoke and the EH pad cannot have other
1315 // instructions between them.
1316 if (LoadBB->isEHPad())
1317 return false;
1318
1319 Value *LoadedPtr = LoadI->getOperand(0);
1320
1321 // If the loaded operand is defined in the LoadBB and its not a phi,
1322 // it can't be available in predecessors.
1323 if (isOpDefinedInBlock(LoadedPtr, LoadBB) && !isa<PHINode>(LoadedPtr))
1324 return false;
1325
1326 // Scan a few instructions up from the load, to see if it is obviously live at
1327 // the entry to its block.
1328 BasicBlock::iterator BBIt(LoadI);
1329 bool IsLoadCSE;
1330 if (Value *AvailableVal = FindAvailableLoadedValue(
1331 LoadI, LoadBB, BBIt, DefMaxInstsToScan, AA, &IsLoadCSE)) {
1332 // If the value of the load is locally available within the block, just use
1333 // it. This frequently occurs for reg2mem'd allocas.
1334
1335 if (IsLoadCSE) {
1336 LoadInst *NLoadI = cast<LoadInst>(AvailableVal);
1337 combineMetadataForCSE(NLoadI, LoadI, false);
1338 };
1339
1340 // If the returned value is the load itself, replace with an undef. This can
1341 // only happen in dead loops.
1342 if (AvailableVal == LoadI)
1343 AvailableVal = UndefValue::get(LoadI->getType());
1344 if (AvailableVal->getType() != LoadI->getType())
1345 AvailableVal = CastInst::CreateBitOrPointerCast(
1346 AvailableVal, LoadI->getType(), "", LoadI);
1347 LoadI->replaceAllUsesWith(AvailableVal);
1348 LoadI->eraseFromParent();
1349 return true;
1350 }
1351
1352 // Otherwise, if we scanned the whole block and got to the top of the block,
1353 // we know the block is locally transparent to the load. If not, something
1354 // might clobber its value.
1355 if (BBIt != LoadBB->begin())
1356 return false;
1357
1358 // If all of the loads and stores that feed the value have the same AA tags,
1359 // then we can propagate them onto any newly inserted loads.
1360 AAMDNodes AATags;
1361 LoadI->getAAMetadata(AATags);
1362
1363 SmallPtrSet<BasicBlock*, 8> PredsScanned;
1364
1365 using AvailablePredsTy = SmallVector<std::pair<BasicBlock *, Value *>, 8>;
1366
1367 AvailablePredsTy AvailablePreds;
1368 BasicBlock *OneUnavailablePred = nullptr;
1369 SmallVector<LoadInst*, 8> CSELoads;
1370
1371 // If we got here, the loaded value is transparent through to the start of the
1372 // block. Check to see if it is available in any of the predecessor blocks.
1373 for (BasicBlock *PredBB : predecessors(LoadBB)) {
1374 // If we already scanned this predecessor, skip it.
1375 if (!PredsScanned.insert(PredBB).second)
1376 continue;
1377
1378 BBIt = PredBB->end();
1379 unsigned NumScanedInst = 0;
1380 Value *PredAvailable = nullptr;
1381 // NOTE: We don't CSE load that is volatile or anything stronger than
1382 // unordered, that should have been checked when we entered the function.
1383 assert(LoadI->isUnordered() &&
1384 "Attempting to CSE volatile or atomic loads");
1385 // If this is a load on a phi pointer, phi-translate it and search
1386 // for available load/store to the pointer in predecessors.
1387 Value *Ptr = LoadedPtr->DoPHITranslation(LoadBB, PredBB);
1388 PredAvailable = FindAvailablePtrLoadStore(
1389 Ptr, LoadI->getType(), LoadI->isAtomic(), PredBB, BBIt,
1390 DefMaxInstsToScan, AA, &IsLoadCSE, &NumScanedInst);
1391
1392 // If PredBB has a single predecessor, continue scanning through the
1393 // single predecessor.
1394 BasicBlock *SinglePredBB = PredBB;
1395 while (!PredAvailable && SinglePredBB && BBIt == SinglePredBB->begin() &&
1396 NumScanedInst < DefMaxInstsToScan) {
1397 SinglePredBB = SinglePredBB->getSinglePredecessor();
1398 if (SinglePredBB) {
1399 BBIt = SinglePredBB->end();
1400 PredAvailable = FindAvailablePtrLoadStore(
1401 Ptr, LoadI->getType(), LoadI->isAtomic(), SinglePredBB, BBIt,
1402 (DefMaxInstsToScan - NumScanedInst), AA, &IsLoadCSE,
1403 &NumScanedInst);
1404 }
1405 }
1406
1407 if (!PredAvailable) {
1408 OneUnavailablePred = PredBB;
1409 continue;
1410 }
1411
1412 if (IsLoadCSE)
1413 CSELoads.push_back(cast<LoadInst>(PredAvailable));
1414
1415 // If so, this load is partially redundant. Remember this info so that we
1416 // can create a PHI node.
1417 AvailablePreds.emplace_back(PredBB, PredAvailable);
1418 }
1419
1420 // If the loaded value isn't available in any predecessor, it isn't partially
1421 // redundant.
1422 if (AvailablePreds.empty()) return false;
1423
1424 // Okay, the loaded value is available in at least one (and maybe all!)
1425 // predecessors. If the value is unavailable in more than one unique
1426 // predecessor, we want to insert a merge block for those common predecessors.
1427 // This ensures that we only have to insert one reload, thus not increasing
1428 // code size.
1429 BasicBlock *UnavailablePred = nullptr;
1430
1431 // If the value is unavailable in one of predecessors, we will end up
1432 // inserting a new instruction into them. It is only valid if all the
1433 // instructions before LoadI are guaranteed to pass execution to its
1434 // successor, or if LoadI is safe to speculate.
1435 // TODO: If this logic becomes more complex, and we will perform PRE insertion
1436 // farther than to a predecessor, we need to reuse the code from GVN's PRE.
1437 // It requires domination tree analysis, so for this simple case it is an
1438 // overkill.
1439 if (PredsScanned.size() != AvailablePreds.size() &&
1440 !isSafeToSpeculativelyExecute(LoadI))
1441 for (auto I = LoadBB->begin(); &*I != LoadI; ++I)
1442 if (!isGuaranteedToTransferExecutionToSuccessor(&*I))
1443 return false;
1444
1445 // If there is exactly one predecessor where the value is unavailable, the
1446 // already computed 'OneUnavailablePred' block is it. If it ends in an
1447 // unconditional branch, we know that it isn't a critical edge.
1448 if (PredsScanned.size() == AvailablePreds.size()+1 &&
1449 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
1450 UnavailablePred = OneUnavailablePred;
1451 } else if (PredsScanned.size() != AvailablePreds.size()) {
1452 // Otherwise, we had multiple unavailable predecessors or we had a critical
1453 // edge from the one.
1454 SmallVector<BasicBlock*, 8> PredsToSplit;
1455 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
1456
1457 for (const auto &AvailablePred : AvailablePreds)
1458 AvailablePredSet.insert(AvailablePred.first);
1459
1460 // Add all the unavailable predecessors to the PredsToSplit list.
1461 for (BasicBlock *P : predecessors(LoadBB)) {
1462 // If the predecessor is an indirect goto, we can't split the edge.
1463 // Same for CallBr.
1464 if (isa<IndirectBrInst>(P->getTerminator()) ||
1465 isa<CallBrInst>(P->getTerminator()))
1466 return false;
1467
1468 if (!AvailablePredSet.count(P))
1469 PredsToSplit.push_back(P);
1470 }
1471
1472 // Split them out to their own block.
1473 UnavailablePred = splitBlockPreds(LoadBB, PredsToSplit, "thread-pre-split");
1474 }
1475
1476 // If the value isn't available in all predecessors, then there will be
1477 // exactly one where it isn't available. Insert a load on that edge and add
1478 // it to the AvailablePreds list.
1479 if (UnavailablePred) {
1480 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
1481 "Can't handle critical edge here!");
1482 LoadInst *NewVal = new LoadInst(
1483 LoadI->getType(), LoadedPtr->DoPHITranslation(LoadBB, UnavailablePred),
1484 LoadI->getName() + ".pr", false, LoadI->getAlign(),
1485 LoadI->getOrdering(), LoadI->getSyncScopeID(),
1486 UnavailablePred->getTerminator());
1487 NewVal->setDebugLoc(LoadI->getDebugLoc());
1488 if (AATags)
1489 NewVal->setAAMetadata(AATags);
1490
1491 AvailablePreds.emplace_back(UnavailablePred, NewVal);
1492 }
1493
1494 // Now we know that each predecessor of this block has a value in
1495 // AvailablePreds, sort them for efficient access as we're walking the preds.
1496 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1497
1498 // Create a PHI node at the start of the block for the PRE'd load value.
1499 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
1500 PHINode *PN = PHINode::Create(LoadI->getType(), std::distance(PB, PE), "",
1501 &LoadBB->front());
1502 PN->takeName(LoadI);
1503 PN->setDebugLoc(LoadI->getDebugLoc());
1504
1505 // Insert new entries into the PHI for each predecessor. A single block may
1506 // have multiple entries here.
1507 for (pred_iterator PI = PB; PI != PE; ++PI) {
1508 BasicBlock *P = *PI;
1509 AvailablePredsTy::iterator I =
1510 llvm::lower_bound(AvailablePreds, std::make_pair(P, (Value *)nullptr));
1511
1512 assert(I != AvailablePreds.end() && I->first == P &&
1513 "Didn't find entry for predecessor!");
1514
1515 // If we have an available predecessor but it requires casting, insert the
1516 // cast in the predecessor and use the cast. Note that we have to update the
1517 // AvailablePreds vector as we go so that all of the PHI entries for this
1518 // predecessor use the same bitcast.
1519 Value *&PredV = I->second;
1520 if (PredV->getType() != LoadI->getType())
1521 PredV = CastInst::CreateBitOrPointerCast(PredV, LoadI->getType(), "",
1522 P->getTerminator());
1523
1524 PN->addIncoming(PredV, I->first);
1525 }
1526
1527 for (LoadInst *PredLoadI : CSELoads) {
1528 combineMetadataForCSE(PredLoadI, LoadI, true);
1529 }
1530
1531 LoadI->replaceAllUsesWith(PN);
1532 LoadI->eraseFromParent();
1533
1534 return true;
1535 }
1536
1537 /// findMostPopularDest - The specified list contains multiple possible
1538 /// threadable destinations. Pick the one that occurs the most frequently in
1539 /// the list.
1540 static BasicBlock *
findMostPopularDest(BasicBlock * BB,const SmallVectorImpl<std::pair<BasicBlock *,BasicBlock * >> & PredToDestList)1541 findMostPopularDest(BasicBlock *BB,
1542 const SmallVectorImpl<std::pair<BasicBlock *,
1543 BasicBlock *>> &PredToDestList) {
1544 assert(!PredToDestList.empty());
1545
1546 // Determine popularity. If there are multiple possible destinations, we
1547 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1548 // blocks with known and real destinations to threading undef. We'll handle
1549 // them later if interesting.
1550 MapVector<BasicBlock *, unsigned> DestPopularity;
1551
1552 // Populate DestPopularity with the successors in the order they appear in the
1553 // successor list. This way, we ensure determinism by iterating it in the
1554 // same order in std::max_element below. We map nullptr to 0 so that we can
1555 // return nullptr when PredToDestList contains nullptr only.
1556 DestPopularity[nullptr] = 0;
1557 for (auto *SuccBB : successors(BB))
1558 DestPopularity[SuccBB] = 0;
1559
1560 for (const auto &PredToDest : PredToDestList)
1561 if (PredToDest.second)
1562 DestPopularity[PredToDest.second]++;
1563
1564 // Find the most popular dest.
1565 using VT = decltype(DestPopularity)::value_type;
1566 auto MostPopular = std::max_element(
1567 DestPopularity.begin(), DestPopularity.end(),
1568 [](const VT &L, const VT &R) { return L.second < R.second; });
1569
1570 // Okay, we have finally picked the most popular destination.
1571 return MostPopular->first;
1572 }
1573
1574 // Try to evaluate the value of V when the control flows from PredPredBB to
1575 // BB->getSinglePredecessor() and then on to BB.
evaluateOnPredecessorEdge(BasicBlock * BB,BasicBlock * PredPredBB,Value * V)1576 Constant *JumpThreadingPass::evaluateOnPredecessorEdge(BasicBlock *BB,
1577 BasicBlock *PredPredBB,
1578 Value *V) {
1579 BasicBlock *PredBB = BB->getSinglePredecessor();
1580 assert(PredBB && "Expected a single predecessor");
1581
1582 if (Constant *Cst = dyn_cast<Constant>(V)) {
1583 return Cst;
1584 }
1585
1586 // Consult LVI if V is not an instruction in BB or PredBB.
1587 Instruction *I = dyn_cast<Instruction>(V);
1588 if (!I || (I->getParent() != BB && I->getParent() != PredBB)) {
1589 return LVI->getConstantOnEdge(V, PredPredBB, PredBB, nullptr);
1590 }
1591
1592 // Look into a PHI argument.
1593 if (PHINode *PHI = dyn_cast<PHINode>(V)) {
1594 if (PHI->getParent() == PredBB)
1595 return dyn_cast<Constant>(PHI->getIncomingValueForBlock(PredPredBB));
1596 return nullptr;
1597 }
1598
1599 // If we have a CmpInst, try to fold it for each incoming edge into PredBB.
1600 if (CmpInst *CondCmp = dyn_cast<CmpInst>(V)) {
1601 if (CondCmp->getParent() == BB) {
1602 Constant *Op0 =
1603 evaluateOnPredecessorEdge(BB, PredPredBB, CondCmp->getOperand(0));
1604 Constant *Op1 =
1605 evaluateOnPredecessorEdge(BB, PredPredBB, CondCmp->getOperand(1));
1606 if (Op0 && Op1) {
1607 return ConstantExpr::getCompare(CondCmp->getPredicate(), Op0, Op1);
1608 }
1609 }
1610 return nullptr;
1611 }
1612
1613 return nullptr;
1614 }
1615
processThreadableEdges(Value * Cond,BasicBlock * BB,ConstantPreference Preference,Instruction * CxtI)1616 bool JumpThreadingPass::processThreadableEdges(Value *Cond, BasicBlock *BB,
1617 ConstantPreference Preference,
1618 Instruction *CxtI) {
1619 // If threading this would thread across a loop header, don't even try to
1620 // thread the edge.
1621 if (LoopHeaders.count(BB))
1622 return false;
1623
1624 PredValueInfoTy PredValues;
1625 if (!computeValueKnownInPredecessors(Cond, BB, PredValues, Preference,
1626 CxtI)) {
1627 // We don't have known values in predecessors. See if we can thread through
1628 // BB and its sole predecessor.
1629 return maybethreadThroughTwoBasicBlocks(BB, Cond);
1630 }
1631
1632 assert(!PredValues.empty() &&
1633 "computeValueKnownInPredecessors returned true with no values");
1634
1635 LLVM_DEBUG(dbgs() << "IN BB: " << *BB;
1636 for (const auto &PredValue : PredValues) {
1637 dbgs() << " BB '" << BB->getName()
1638 << "': FOUND condition = " << *PredValue.first
1639 << " for pred '" << PredValue.second->getName() << "'.\n";
1640 });
1641
1642 // Decide what we want to thread through. Convert our list of known values to
1643 // a list of known destinations for each pred. This also discards duplicate
1644 // predecessors and keeps track of the undefined inputs (which are represented
1645 // as a null dest in the PredToDestList).
1646 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1647 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1648
1649 BasicBlock *OnlyDest = nullptr;
1650 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1651 Constant *OnlyVal = nullptr;
1652 Constant *MultipleVal = (Constant *)(intptr_t)~0ULL;
1653
1654 for (const auto &PredValue : PredValues) {
1655 BasicBlock *Pred = PredValue.second;
1656 if (!SeenPreds.insert(Pred).second)
1657 continue; // Duplicate predecessor entry.
1658
1659 Constant *Val = PredValue.first;
1660
1661 BasicBlock *DestBB;
1662 if (isa<UndefValue>(Val))
1663 DestBB = nullptr;
1664 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
1665 assert(isa<ConstantInt>(Val) && "Expecting a constant integer");
1666 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1667 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1668 assert(isa<ConstantInt>(Val) && "Expecting a constant integer");
1669 DestBB = SI->findCaseValue(cast<ConstantInt>(Val))->getCaseSuccessor();
1670 } else {
1671 assert(isa<IndirectBrInst>(BB->getTerminator())
1672 && "Unexpected terminator");
1673 assert(isa<BlockAddress>(Val) && "Expecting a constant blockaddress");
1674 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1675 }
1676
1677 // If we have exactly one destination, remember it for efficiency below.
1678 if (PredToDestList.empty()) {
1679 OnlyDest = DestBB;
1680 OnlyVal = Val;
1681 } else {
1682 if (OnlyDest != DestBB)
1683 OnlyDest = MultipleDestSentinel;
1684 // It possible we have same destination, but different value, e.g. default
1685 // case in switchinst.
1686 if (Val != OnlyVal)
1687 OnlyVal = MultipleVal;
1688 }
1689
1690 // If the predecessor ends with an indirect goto, we can't change its
1691 // destination. Same for CallBr.
1692 if (isa<IndirectBrInst>(Pred->getTerminator()) ||
1693 isa<CallBrInst>(Pred->getTerminator()))
1694 continue;
1695
1696 PredToDestList.emplace_back(Pred, DestBB);
1697 }
1698
1699 // If all edges were unthreadable, we fail.
1700 if (PredToDestList.empty())
1701 return false;
1702
1703 // If all the predecessors go to a single known successor, we want to fold,
1704 // not thread. By doing so, we do not need to duplicate the current block and
1705 // also miss potential opportunities in case we dont/cant duplicate.
1706 if (OnlyDest && OnlyDest != MultipleDestSentinel) {
1707 if (BB->hasNPredecessors(PredToDestList.size())) {
1708 bool SeenFirstBranchToOnlyDest = false;
1709 std::vector <DominatorTree::UpdateType> Updates;
1710 Updates.reserve(BB->getTerminator()->getNumSuccessors() - 1);
1711 for (BasicBlock *SuccBB : successors(BB)) {
1712 if (SuccBB == OnlyDest && !SeenFirstBranchToOnlyDest) {
1713 SeenFirstBranchToOnlyDest = true; // Don't modify the first branch.
1714 } else {
1715 SuccBB->removePredecessor(BB, true); // This is unreachable successor.
1716 Updates.push_back({DominatorTree::Delete, BB, SuccBB});
1717 }
1718 }
1719
1720 // Finally update the terminator.
1721 Instruction *Term = BB->getTerminator();
1722 BranchInst::Create(OnlyDest, Term);
1723 Term->eraseFromParent();
1724 DTU->applyUpdatesPermissive(Updates);
1725 if (HasProfileData)
1726 BPI->eraseBlock(BB);
1727
1728 // If the condition is now dead due to the removal of the old terminator,
1729 // erase it.
1730 if (auto *CondInst = dyn_cast<Instruction>(Cond)) {
1731 if (CondInst->use_empty() && !CondInst->mayHaveSideEffects())
1732 CondInst->eraseFromParent();
1733 // We can safely replace *some* uses of the CondInst if it has
1734 // exactly one value as returned by LVI. RAUW is incorrect in the
1735 // presence of guards and assumes, that have the `Cond` as the use. This
1736 // is because we use the guards/assume to reason about the `Cond` value
1737 // at the end of block, but RAUW unconditionally replaces all uses
1738 // including the guards/assumes themselves and the uses before the
1739 // guard/assume.
1740 else if (OnlyVal && OnlyVal != MultipleVal &&
1741 CondInst->getParent() == BB)
1742 replaceFoldableUses(CondInst, OnlyVal);
1743 }
1744 return true;
1745 }
1746 }
1747
1748 // Determine which is the most common successor. If we have many inputs and
1749 // this block is a switch, we want to start by threading the batch that goes
1750 // to the most popular destination first. If we only know about one
1751 // threadable destination (the common case) we can avoid this.
1752 BasicBlock *MostPopularDest = OnlyDest;
1753
1754 if (MostPopularDest == MultipleDestSentinel) {
1755 // Remove any loop headers from the Dest list, threadEdge conservatively
1756 // won't process them, but we might have other destination that are eligible
1757 // and we still want to process.
1758 erase_if(PredToDestList,
1759 [&](const std::pair<BasicBlock *, BasicBlock *> &PredToDest) {
1760 return LoopHeaders.contains(PredToDest.second);
1761 });
1762
1763 if (PredToDestList.empty())
1764 return false;
1765
1766 MostPopularDest = findMostPopularDest(BB, PredToDestList);
1767 }
1768
1769 // Now that we know what the most popular destination is, factor all
1770 // predecessors that will jump to it into a single predecessor.
1771 SmallVector<BasicBlock*, 16> PredsToFactor;
1772 for (const auto &PredToDest : PredToDestList)
1773 if (PredToDest.second == MostPopularDest) {
1774 BasicBlock *Pred = PredToDest.first;
1775
1776 // This predecessor may be a switch or something else that has multiple
1777 // edges to the block. Factor each of these edges by listing them
1778 // according to # occurrences in PredsToFactor.
1779 for (BasicBlock *Succ : successors(Pred))
1780 if (Succ == BB)
1781 PredsToFactor.push_back(Pred);
1782 }
1783
1784 // If the threadable edges are branching on an undefined value, we get to pick
1785 // the destination that these predecessors should get to.
1786 if (!MostPopularDest)
1787 MostPopularDest = BB->getTerminator()->
1788 getSuccessor(getBestDestForJumpOnUndef(BB));
1789
1790 // Ok, try to thread it!
1791 return tryThreadEdge(BB, PredsToFactor, MostPopularDest);
1792 }
1793
1794 /// processBranchOnPHI - We have an otherwise unthreadable conditional branch on
1795 /// a PHI node (or freeze PHI) in the current block. See if there are any
1796 /// simplifications we can do based on inputs to the phi node.
processBranchOnPHI(PHINode * PN)1797 bool JumpThreadingPass::processBranchOnPHI(PHINode *PN) {
1798 BasicBlock *BB = PN->getParent();
1799
1800 // TODO: We could make use of this to do it once for blocks with common PHI
1801 // values.
1802 SmallVector<BasicBlock*, 1> PredBBs;
1803 PredBBs.resize(1);
1804
1805 // If any of the predecessor blocks end in an unconditional branch, we can
1806 // *duplicate* the conditional branch into that block in order to further
1807 // encourage jump threading and to eliminate cases where we have branch on a
1808 // phi of an icmp (branch on icmp is much better).
1809 // This is still beneficial when a frozen phi is used as the branch condition
1810 // because it allows CodeGenPrepare to further canonicalize br(freeze(icmp))
1811 // to br(icmp(freeze ...)).
1812 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1813 BasicBlock *PredBB = PN->getIncomingBlock(i);
1814 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1815 if (PredBr->isUnconditional()) {
1816 PredBBs[0] = PredBB;
1817 // Try to duplicate BB into PredBB.
1818 if (duplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1819 return true;
1820 }
1821 }
1822
1823 return false;
1824 }
1825
1826 /// processBranchOnXOR - We have an otherwise unthreadable conditional branch on
1827 /// a xor instruction in the current block. See if there are any
1828 /// simplifications we can do based on inputs to the xor.
processBranchOnXOR(BinaryOperator * BO)1829 bool JumpThreadingPass::processBranchOnXOR(BinaryOperator *BO) {
1830 BasicBlock *BB = BO->getParent();
1831
1832 // If either the LHS or RHS of the xor is a constant, don't do this
1833 // optimization.
1834 if (isa<ConstantInt>(BO->getOperand(0)) ||
1835 isa<ConstantInt>(BO->getOperand(1)))
1836 return false;
1837
1838 // If the first instruction in BB isn't a phi, we won't be able to infer
1839 // anything special about any particular predecessor.
1840 if (!isa<PHINode>(BB->front()))
1841 return false;
1842
1843 // If this BB is a landing pad, we won't be able to split the edge into it.
1844 if (BB->isEHPad())
1845 return false;
1846
1847 // If we have a xor as the branch input to this block, and we know that the
1848 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1849 // the condition into the predecessor and fix that value to true, saving some
1850 // logical ops on that path and encouraging other paths to simplify.
1851 //
1852 // This copies something like this:
1853 //
1854 // BB:
1855 // %X = phi i1 [1], [%X']
1856 // %Y = icmp eq i32 %A, %B
1857 // %Z = xor i1 %X, %Y
1858 // br i1 %Z, ...
1859 //
1860 // Into:
1861 // BB':
1862 // %Y = icmp ne i32 %A, %B
1863 // br i1 %Y, ...
1864
1865 PredValueInfoTy XorOpValues;
1866 bool isLHS = true;
1867 if (!computeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1868 WantInteger, BO)) {
1869 assert(XorOpValues.empty());
1870 if (!computeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1871 WantInteger, BO))
1872 return false;
1873 isLHS = false;
1874 }
1875
1876 assert(!XorOpValues.empty() &&
1877 "computeValueKnownInPredecessors returned true with no values");
1878
1879 // Scan the information to see which is most popular: true or false. The
1880 // predecessors can be of the set true, false, or undef.
1881 unsigned NumTrue = 0, NumFalse = 0;
1882 for (const auto &XorOpValue : XorOpValues) {
1883 if (isa<UndefValue>(XorOpValue.first))
1884 // Ignore undefs for the count.
1885 continue;
1886 if (cast<ConstantInt>(XorOpValue.first)->isZero())
1887 ++NumFalse;
1888 else
1889 ++NumTrue;
1890 }
1891
1892 // Determine which value to split on, true, false, or undef if neither.
1893 ConstantInt *SplitVal = nullptr;
1894 if (NumTrue > NumFalse)
1895 SplitVal = ConstantInt::getTrue(BB->getContext());
1896 else if (NumTrue != 0 || NumFalse != 0)
1897 SplitVal = ConstantInt::getFalse(BB->getContext());
1898
1899 // Collect all of the blocks that this can be folded into so that we can
1900 // factor this once and clone it once.
1901 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1902 for (const auto &XorOpValue : XorOpValues) {
1903 if (XorOpValue.first != SplitVal && !isa<UndefValue>(XorOpValue.first))
1904 continue;
1905
1906 BlocksToFoldInto.push_back(XorOpValue.second);
1907 }
1908
1909 // If we inferred a value for all of the predecessors, then duplication won't
1910 // help us. However, we can just replace the LHS or RHS with the constant.
1911 if (BlocksToFoldInto.size() ==
1912 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1913 if (!SplitVal) {
1914 // If all preds provide undef, just nuke the xor, because it is undef too.
1915 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1916 BO->eraseFromParent();
1917 } else if (SplitVal->isZero()) {
1918 // If all preds provide 0, replace the xor with the other input.
1919 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1920 BO->eraseFromParent();
1921 } else {
1922 // If all preds provide 1, set the computed value to 1.
1923 BO->setOperand(!isLHS, SplitVal);
1924 }
1925
1926 return true;
1927 }
1928
1929 // If any of predecessors end with an indirect goto, we can't change its
1930 // destination. Same for CallBr.
1931 if (any_of(BlocksToFoldInto, [](BasicBlock *Pred) {
1932 return isa<IndirectBrInst>(Pred->getTerminator()) ||
1933 isa<CallBrInst>(Pred->getTerminator());
1934 }))
1935 return false;
1936
1937 // Try to duplicate BB into PredBB.
1938 return duplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1939 }
1940
1941 /// addPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1942 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1943 /// NewPred using the entries from OldPred (suitably mapped).
addPHINodeEntriesForMappedBlock(BasicBlock * PHIBB,BasicBlock * OldPred,BasicBlock * NewPred,DenseMap<Instruction *,Value * > & ValueMap)1944 static void addPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1945 BasicBlock *OldPred,
1946 BasicBlock *NewPred,
1947 DenseMap<Instruction*, Value*> &ValueMap) {
1948 for (PHINode &PN : PHIBB->phis()) {
1949 // Ok, we have a PHI node. Figure out what the incoming value was for the
1950 // DestBlock.
1951 Value *IV = PN.getIncomingValueForBlock(OldPred);
1952
1953 // Remap the value if necessary.
1954 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1955 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1956 if (I != ValueMap.end())
1957 IV = I->second;
1958 }
1959
1960 PN.addIncoming(IV, NewPred);
1961 }
1962 }
1963
1964 /// Merge basic block BB into its sole predecessor if possible.
maybeMergeBasicBlockIntoOnlyPred(BasicBlock * BB)1965 bool JumpThreadingPass::maybeMergeBasicBlockIntoOnlyPred(BasicBlock *BB) {
1966 BasicBlock *SinglePred = BB->getSinglePredecessor();
1967 if (!SinglePred)
1968 return false;
1969
1970 const Instruction *TI = SinglePred->getTerminator();
1971 if (TI->isExceptionalTerminator() || TI->getNumSuccessors() != 1 ||
1972 SinglePred == BB || hasAddressTakenAndUsed(BB))
1973 return false;
1974
1975 // If SinglePred was a loop header, BB becomes one.
1976 if (LoopHeaders.erase(SinglePred))
1977 LoopHeaders.insert(BB);
1978
1979 LVI->eraseBlock(SinglePred);
1980 MergeBasicBlockIntoOnlyPred(BB, DTU);
1981
1982 // Now that BB is merged into SinglePred (i.e. SinglePred code followed by
1983 // BB code within one basic block `BB`), we need to invalidate the LVI
1984 // information associated with BB, because the LVI information need not be
1985 // true for all of BB after the merge. For example,
1986 // Before the merge, LVI info and code is as follows:
1987 // SinglePred: <LVI info1 for %p val>
1988 // %y = use of %p
1989 // call @exit() // need not transfer execution to successor.
1990 // assume(%p) // from this point on %p is true
1991 // br label %BB
1992 // BB: <LVI info2 for %p val, i.e. %p is true>
1993 // %x = use of %p
1994 // br label exit
1995 //
1996 // Note that this LVI info for blocks BB and SinglPred is correct for %p
1997 // (info2 and info1 respectively). After the merge and the deletion of the
1998 // LVI info1 for SinglePred. We have the following code:
1999 // BB: <LVI info2 for %p val>
2000 // %y = use of %p
2001 // call @exit()
2002 // assume(%p)
2003 // %x = use of %p <-- LVI info2 is correct from here onwards.
2004 // br label exit
2005 // LVI info2 for BB is incorrect at the beginning of BB.
2006
2007 // Invalidate LVI information for BB if the LVI is not provably true for
2008 // all of BB.
2009 if (!isGuaranteedToTransferExecutionToSuccessor(BB))
2010 LVI->eraseBlock(BB);
2011 return true;
2012 }
2013
2014 /// Update the SSA form. NewBB contains instructions that are copied from BB.
2015 /// ValueMapping maps old values in BB to new ones in NewBB.
updateSSA(BasicBlock * BB,BasicBlock * NewBB,DenseMap<Instruction *,Value * > & ValueMapping)2016 void JumpThreadingPass::updateSSA(
2017 BasicBlock *BB, BasicBlock *NewBB,
2018 DenseMap<Instruction *, Value *> &ValueMapping) {
2019 // If there were values defined in BB that are used outside the block, then we
2020 // now have to update all uses of the value to use either the original value,
2021 // the cloned value, or some PHI derived value. This can require arbitrary
2022 // PHI insertion, of which we are prepared to do, clean these up now.
2023 SSAUpdater SSAUpdate;
2024 SmallVector<Use *, 16> UsesToRename;
2025
2026 for (Instruction &I : *BB) {
2027 // Scan all uses of this instruction to see if it is used outside of its
2028 // block, and if so, record them in UsesToRename.
2029 for (Use &U : I.uses()) {
2030 Instruction *User = cast<Instruction>(U.getUser());
2031 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
2032 if (UserPN->getIncomingBlock(U) == BB)
2033 continue;
2034 } else if (User->getParent() == BB)
2035 continue;
2036
2037 UsesToRename.push_back(&U);
2038 }
2039
2040 // If there are no uses outside the block, we're done with this instruction.
2041 if (UsesToRename.empty())
2042 continue;
2043 LLVM_DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n");
2044
2045 // We found a use of I outside of BB. Rename all uses of I that are outside
2046 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
2047 // with the two values we know.
2048 SSAUpdate.Initialize(I.getType(), I.getName());
2049 SSAUpdate.AddAvailableValue(BB, &I);
2050 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[&I]);
2051
2052 while (!UsesToRename.empty())
2053 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
2054 LLVM_DEBUG(dbgs() << "\n");
2055 }
2056 }
2057
2058 /// Clone instructions in range [BI, BE) to NewBB. For PHI nodes, we only clone
2059 /// arguments that come from PredBB. Return the map from the variables in the
2060 /// source basic block to the variables in the newly created basic block.
2061 DenseMap<Instruction *, Value *>
cloneInstructions(BasicBlock::iterator BI,BasicBlock::iterator BE,BasicBlock * NewBB,BasicBlock * PredBB)2062 JumpThreadingPass::cloneInstructions(BasicBlock::iterator BI,
2063 BasicBlock::iterator BE, BasicBlock *NewBB,
2064 BasicBlock *PredBB) {
2065 // We are going to have to map operands from the source basic block to the new
2066 // copy of the block 'NewBB'. If there are PHI nodes in the source basic
2067 // block, evaluate them to account for entry from PredBB.
2068 DenseMap<Instruction *, Value *> ValueMapping;
2069
2070 // Clone the phi nodes of the source basic block into NewBB. The resulting
2071 // phi nodes are trivial since NewBB only has one predecessor, but SSAUpdater
2072 // might need to rewrite the operand of the cloned phi.
2073 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) {
2074 PHINode *NewPN = PHINode::Create(PN->getType(), 1, PN->getName(), NewBB);
2075 NewPN->addIncoming(PN->getIncomingValueForBlock(PredBB), PredBB);
2076 ValueMapping[PN] = NewPN;
2077 }
2078
2079 // Clone noalias scope declarations in the threaded block. When threading a
2080 // loop exit, we would otherwise end up with two idential scope declarations
2081 // visible at the same time.
2082 SmallVector<MDNode *> NoAliasScopes;
2083 DenseMap<MDNode *, MDNode *> ClonedScopes;
2084 LLVMContext &Context = PredBB->getContext();
2085 identifyNoAliasScopesToClone(BI, BE, NoAliasScopes);
2086 cloneNoAliasScopes(NoAliasScopes, ClonedScopes, "thread", Context);
2087
2088 // Clone the non-phi instructions of the source basic block into NewBB,
2089 // keeping track of the mapping and using it to remap operands in the cloned
2090 // instructions.
2091 for (; BI != BE; ++BI) {
2092 Instruction *New = BI->clone();
2093 New->setName(BI->getName());
2094 NewBB->getInstList().push_back(New);
2095 ValueMapping[&*BI] = New;
2096 adaptNoAliasScopes(New, ClonedScopes, Context);
2097
2098 // Remap operands to patch up intra-block references.
2099 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2100 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
2101 DenseMap<Instruction *, Value *>::iterator I = ValueMapping.find(Inst);
2102 if (I != ValueMapping.end())
2103 New->setOperand(i, I->second);
2104 }
2105 }
2106
2107 return ValueMapping;
2108 }
2109
2110 /// Attempt to thread through two successive basic blocks.
maybethreadThroughTwoBasicBlocks(BasicBlock * BB,Value * Cond)2111 bool JumpThreadingPass::maybethreadThroughTwoBasicBlocks(BasicBlock *BB,
2112 Value *Cond) {
2113 // Consider:
2114 //
2115 // PredBB:
2116 // %var = phi i32* [ null, %bb1 ], [ @a, %bb2 ]
2117 // %tobool = icmp eq i32 %cond, 0
2118 // br i1 %tobool, label %BB, label ...
2119 //
2120 // BB:
2121 // %cmp = icmp eq i32* %var, null
2122 // br i1 %cmp, label ..., label ...
2123 //
2124 // We don't know the value of %var at BB even if we know which incoming edge
2125 // we take to BB. However, once we duplicate PredBB for each of its incoming
2126 // edges (say, PredBB1 and PredBB2), we know the value of %var in each copy of
2127 // PredBB. Then we can thread edges PredBB1->BB and PredBB2->BB through BB.
2128
2129 // Require that BB end with a Branch for simplicity.
2130 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
2131 if (!CondBr)
2132 return false;
2133
2134 // BB must have exactly one predecessor.
2135 BasicBlock *PredBB = BB->getSinglePredecessor();
2136 if (!PredBB)
2137 return false;
2138
2139 // Require that PredBB end with a conditional Branch. If PredBB ends with an
2140 // unconditional branch, we should be merging PredBB and BB instead. For
2141 // simplicity, we don't deal with a switch.
2142 BranchInst *PredBBBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
2143 if (!PredBBBranch || PredBBBranch->isUnconditional())
2144 return false;
2145
2146 // If PredBB has exactly one incoming edge, we don't gain anything by copying
2147 // PredBB.
2148 if (PredBB->getSinglePredecessor())
2149 return false;
2150
2151 // Don't thread through PredBB if it contains a successor edge to itself, in
2152 // which case we would infinite loop. Suppose we are threading an edge from
2153 // PredPredBB through PredBB and BB to SuccBB with PredBB containing a
2154 // successor edge to itself. If we allowed jump threading in this case, we
2155 // could duplicate PredBB and BB as, say, PredBB.thread and BB.thread. Since
2156 // PredBB.thread has a successor edge to PredBB, we would immediately come up
2157 // with another jump threading opportunity from PredBB.thread through PredBB
2158 // and BB to SuccBB. This jump threading would repeatedly occur. That is, we
2159 // would keep peeling one iteration from PredBB.
2160 if (llvm::is_contained(successors(PredBB), PredBB))
2161 return false;
2162
2163 // Don't thread across a loop header.
2164 if (LoopHeaders.count(PredBB))
2165 return false;
2166
2167 // Avoid complication with duplicating EH pads.
2168 if (PredBB->isEHPad())
2169 return false;
2170
2171 // Find a predecessor that we can thread. For simplicity, we only consider a
2172 // successor edge out of BB to which we thread exactly one incoming edge into
2173 // PredBB.
2174 unsigned ZeroCount = 0;
2175 unsigned OneCount = 0;
2176 BasicBlock *ZeroPred = nullptr;
2177 BasicBlock *OnePred = nullptr;
2178 for (BasicBlock *P : predecessors(PredBB)) {
2179 if (ConstantInt *CI = dyn_cast_or_null<ConstantInt>(
2180 evaluateOnPredecessorEdge(BB, P, Cond))) {
2181 if (CI->isZero()) {
2182 ZeroCount++;
2183 ZeroPred = P;
2184 } else if (CI->isOne()) {
2185 OneCount++;
2186 OnePred = P;
2187 }
2188 }
2189 }
2190
2191 // Disregard complicated cases where we have to thread multiple edges.
2192 BasicBlock *PredPredBB;
2193 if (ZeroCount == 1) {
2194 PredPredBB = ZeroPred;
2195 } else if (OneCount == 1) {
2196 PredPredBB = OnePred;
2197 } else {
2198 return false;
2199 }
2200
2201 BasicBlock *SuccBB = CondBr->getSuccessor(PredPredBB == ZeroPred);
2202
2203 // If threading to the same block as we come from, we would infinite loop.
2204 if (SuccBB == BB) {
2205 LLVM_DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
2206 << "' - would thread to self!\n");
2207 return false;
2208 }
2209
2210 // If threading this would thread across a loop header, don't thread the edge.
2211 // See the comments above findLoopHeaders for justifications and caveats.
2212 if (LoopHeaders.count(BB) || LoopHeaders.count(SuccBB)) {
2213 LLVM_DEBUG({
2214 bool BBIsHeader = LoopHeaders.count(BB);
2215 bool SuccIsHeader = LoopHeaders.count(SuccBB);
2216 dbgs() << " Not threading across "
2217 << (BBIsHeader ? "loop header BB '" : "block BB '")
2218 << BB->getName() << "' to dest "
2219 << (SuccIsHeader ? "loop header BB '" : "block BB '")
2220 << SuccBB->getName()
2221 << "' - it might create an irreducible loop!\n";
2222 });
2223 return false;
2224 }
2225
2226 // Compute the cost of duplicating BB and PredBB.
2227 unsigned BBCost =
2228 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
2229 unsigned PredBBCost = getJumpThreadDuplicationCost(
2230 PredBB, PredBB->getTerminator(), BBDupThreshold);
2231
2232 // Give up if costs are too high. We need to check BBCost and PredBBCost
2233 // individually before checking their sum because getJumpThreadDuplicationCost
2234 // return (unsigned)~0 for those basic blocks that cannot be duplicated.
2235 if (BBCost > BBDupThreshold || PredBBCost > BBDupThreshold ||
2236 BBCost + PredBBCost > BBDupThreshold) {
2237 LLVM_DEBUG(dbgs() << " Not threading BB '" << BB->getName()
2238 << "' - Cost is too high: " << PredBBCost
2239 << " for PredBB, " << BBCost << "for BB\n");
2240 return false;
2241 }
2242
2243 // Now we are ready to duplicate PredBB.
2244 threadThroughTwoBasicBlocks(PredPredBB, PredBB, BB, SuccBB);
2245 return true;
2246 }
2247
threadThroughTwoBasicBlocks(BasicBlock * PredPredBB,BasicBlock * PredBB,BasicBlock * BB,BasicBlock * SuccBB)2248 void JumpThreadingPass::threadThroughTwoBasicBlocks(BasicBlock *PredPredBB,
2249 BasicBlock *PredBB,
2250 BasicBlock *BB,
2251 BasicBlock *SuccBB) {
2252 LLVM_DEBUG(dbgs() << " Threading through '" << PredBB->getName() << "' and '"
2253 << BB->getName() << "'\n");
2254
2255 BranchInst *CondBr = cast<BranchInst>(BB->getTerminator());
2256 BranchInst *PredBBBranch = cast<BranchInst>(PredBB->getTerminator());
2257
2258 BasicBlock *NewBB =
2259 BasicBlock::Create(PredBB->getContext(), PredBB->getName() + ".thread",
2260 PredBB->getParent(), PredBB);
2261 NewBB->moveAfter(PredBB);
2262
2263 // Set the block frequency of NewBB.
2264 if (HasProfileData) {
2265 auto NewBBFreq = BFI->getBlockFreq(PredPredBB) *
2266 BPI->getEdgeProbability(PredPredBB, PredBB);
2267 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
2268 }
2269
2270 // We are going to have to map operands from the original BB block to the new
2271 // copy of the block 'NewBB'. If there are PHI nodes in PredBB, evaluate them
2272 // to account for entry from PredPredBB.
2273 DenseMap<Instruction *, Value *> ValueMapping =
2274 cloneInstructions(PredBB->begin(), PredBB->end(), NewBB, PredPredBB);
2275
2276 // Copy the edge probabilities from PredBB to NewBB.
2277 if (HasProfileData)
2278 BPI->copyEdgeProbabilities(PredBB, NewBB);
2279
2280 // Update the terminator of PredPredBB to jump to NewBB instead of PredBB.
2281 // This eliminates predecessors from PredPredBB, which requires us to simplify
2282 // any PHI nodes in PredBB.
2283 Instruction *PredPredTerm = PredPredBB->getTerminator();
2284 for (unsigned i = 0, e = PredPredTerm->getNumSuccessors(); i != e; ++i)
2285 if (PredPredTerm->getSuccessor(i) == PredBB) {
2286 PredBB->removePredecessor(PredPredBB, true);
2287 PredPredTerm->setSuccessor(i, NewBB);
2288 }
2289
2290 addPHINodeEntriesForMappedBlock(PredBBBranch->getSuccessor(0), PredBB, NewBB,
2291 ValueMapping);
2292 addPHINodeEntriesForMappedBlock(PredBBBranch->getSuccessor(1), PredBB, NewBB,
2293 ValueMapping);
2294
2295 DTU->applyUpdatesPermissive(
2296 {{DominatorTree::Insert, NewBB, CondBr->getSuccessor(0)},
2297 {DominatorTree::Insert, NewBB, CondBr->getSuccessor(1)},
2298 {DominatorTree::Insert, PredPredBB, NewBB},
2299 {DominatorTree::Delete, PredPredBB, PredBB}});
2300
2301 updateSSA(PredBB, NewBB, ValueMapping);
2302
2303 // Clean up things like PHI nodes with single operands, dead instructions,
2304 // etc.
2305 SimplifyInstructionsInBlock(NewBB, TLI);
2306 SimplifyInstructionsInBlock(PredBB, TLI);
2307
2308 SmallVector<BasicBlock *, 1> PredsToFactor;
2309 PredsToFactor.push_back(NewBB);
2310 threadEdge(BB, PredsToFactor, SuccBB);
2311 }
2312
2313 /// tryThreadEdge - Thread an edge if it's safe and profitable to do so.
tryThreadEdge(BasicBlock * BB,const SmallVectorImpl<BasicBlock * > & PredBBs,BasicBlock * SuccBB)2314 bool JumpThreadingPass::tryThreadEdge(
2315 BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs,
2316 BasicBlock *SuccBB) {
2317 // If threading to the same block as we come from, we would infinite loop.
2318 if (SuccBB == BB) {
2319 LLVM_DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
2320 << "' - would thread to self!\n");
2321 return false;
2322 }
2323
2324 // If threading this would thread across a loop header, don't thread the edge.
2325 // See the comments above findLoopHeaders for justifications and caveats.
2326 if (LoopHeaders.count(BB) || LoopHeaders.count(SuccBB)) {
2327 LLVM_DEBUG({
2328 bool BBIsHeader = LoopHeaders.count(BB);
2329 bool SuccIsHeader = LoopHeaders.count(SuccBB);
2330 dbgs() << " Not threading across "
2331 << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName()
2332 << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '")
2333 << SuccBB->getName() << "' - it might create an irreducible loop!\n";
2334 });
2335 return false;
2336 }
2337
2338 unsigned JumpThreadCost =
2339 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
2340 if (JumpThreadCost > BBDupThreshold) {
2341 LLVM_DEBUG(dbgs() << " Not threading BB '" << BB->getName()
2342 << "' - Cost is too high: " << JumpThreadCost << "\n");
2343 return false;
2344 }
2345
2346 threadEdge(BB, PredBBs, SuccBB);
2347 return true;
2348 }
2349
2350 /// threadEdge - We have decided that it is safe and profitable to factor the
2351 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
2352 /// across BB. Transform the IR to reflect this change.
threadEdge(BasicBlock * BB,const SmallVectorImpl<BasicBlock * > & PredBBs,BasicBlock * SuccBB)2353 void JumpThreadingPass::threadEdge(BasicBlock *BB,
2354 const SmallVectorImpl<BasicBlock *> &PredBBs,
2355 BasicBlock *SuccBB) {
2356 assert(SuccBB != BB && "Don't create an infinite loop");
2357
2358 assert(!LoopHeaders.count(BB) && !LoopHeaders.count(SuccBB) &&
2359 "Don't thread across loop headers");
2360
2361 // And finally, do it! Start by factoring the predecessors if needed.
2362 BasicBlock *PredBB;
2363 if (PredBBs.size() == 1)
2364 PredBB = PredBBs[0];
2365 else {
2366 LLVM_DEBUG(dbgs() << " Factoring out " << PredBBs.size()
2367 << " common predecessors.\n");
2368 PredBB = splitBlockPreds(BB, PredBBs, ".thr_comm");
2369 }
2370
2371 // And finally, do it!
2372 LLVM_DEBUG(dbgs() << " Threading edge from '" << PredBB->getName()
2373 << "' to '" << SuccBB->getName()
2374 << ", across block:\n " << *BB << "\n");
2375
2376 LVI->threadEdge(PredBB, BB, SuccBB);
2377
2378 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
2379 BB->getName()+".thread",
2380 BB->getParent(), BB);
2381 NewBB->moveAfter(PredBB);
2382
2383 // Set the block frequency of NewBB.
2384 if (HasProfileData) {
2385 auto NewBBFreq =
2386 BFI->getBlockFreq(PredBB) * BPI->getEdgeProbability(PredBB, BB);
2387 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
2388 }
2389
2390 // Copy all the instructions from BB to NewBB except the terminator.
2391 DenseMap<Instruction *, Value *> ValueMapping =
2392 cloneInstructions(BB->begin(), std::prev(BB->end()), NewBB, PredBB);
2393
2394 // We didn't copy the terminator from BB over to NewBB, because there is now
2395 // an unconditional jump to SuccBB. Insert the unconditional jump.
2396 BranchInst *NewBI = BranchInst::Create(SuccBB, NewBB);
2397 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
2398
2399 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
2400 // PHI nodes for NewBB now.
2401 addPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
2402
2403 // Update the terminator of PredBB to jump to NewBB instead of BB. This
2404 // eliminates predecessors from BB, which requires us to simplify any PHI
2405 // nodes in BB.
2406 Instruction *PredTerm = PredBB->getTerminator();
2407 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
2408 if (PredTerm->getSuccessor(i) == BB) {
2409 BB->removePredecessor(PredBB, true);
2410 PredTerm->setSuccessor(i, NewBB);
2411 }
2412
2413 // Enqueue required DT updates.
2414 DTU->applyUpdatesPermissive({{DominatorTree::Insert, NewBB, SuccBB},
2415 {DominatorTree::Insert, PredBB, NewBB},
2416 {DominatorTree::Delete, PredBB, BB}});
2417
2418 updateSSA(BB, NewBB, ValueMapping);
2419
2420 // At this point, the IR is fully up to date and consistent. Do a quick scan
2421 // over the new instructions and zap any that are constants or dead. This
2422 // frequently happens because of phi translation.
2423 SimplifyInstructionsInBlock(NewBB, TLI);
2424
2425 // Update the edge weight from BB to SuccBB, which should be less than before.
2426 updateBlockFreqAndEdgeWeight(PredBB, BB, NewBB, SuccBB);
2427
2428 // Threaded an edge!
2429 ++NumThreads;
2430 }
2431
2432 /// Create a new basic block that will be the predecessor of BB and successor of
2433 /// all blocks in Preds. When profile data is available, update the frequency of
2434 /// this new block.
splitBlockPreds(BasicBlock * BB,ArrayRef<BasicBlock * > Preds,const char * Suffix)2435 BasicBlock *JumpThreadingPass::splitBlockPreds(BasicBlock *BB,
2436 ArrayRef<BasicBlock *> Preds,
2437 const char *Suffix) {
2438 SmallVector<BasicBlock *, 2> NewBBs;
2439
2440 // Collect the frequencies of all predecessors of BB, which will be used to
2441 // update the edge weight of the result of splitting predecessors.
2442 DenseMap<BasicBlock *, BlockFrequency> FreqMap;
2443 if (HasProfileData)
2444 for (auto Pred : Preds)
2445 FreqMap.insert(std::make_pair(
2446 Pred, BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, BB)));
2447
2448 // In the case when BB is a LandingPad block we create 2 new predecessors
2449 // instead of just one.
2450 if (BB->isLandingPad()) {
2451 std::string NewName = std::string(Suffix) + ".split-lp";
2452 SplitLandingPadPredecessors(BB, Preds, Suffix, NewName.c_str(), NewBBs);
2453 } else {
2454 NewBBs.push_back(SplitBlockPredecessors(BB, Preds, Suffix));
2455 }
2456
2457 std::vector<DominatorTree::UpdateType> Updates;
2458 Updates.reserve((2 * Preds.size()) + NewBBs.size());
2459 for (auto NewBB : NewBBs) {
2460 BlockFrequency NewBBFreq(0);
2461 Updates.push_back({DominatorTree::Insert, NewBB, BB});
2462 for (auto Pred : predecessors(NewBB)) {
2463 Updates.push_back({DominatorTree::Delete, Pred, BB});
2464 Updates.push_back({DominatorTree::Insert, Pred, NewBB});
2465 if (HasProfileData) // Update frequencies between Pred -> NewBB.
2466 NewBBFreq += FreqMap.lookup(Pred);
2467 }
2468 if (HasProfileData) // Apply the summed frequency to NewBB.
2469 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
2470 }
2471
2472 DTU->applyUpdatesPermissive(Updates);
2473 return NewBBs[0];
2474 }
2475
doesBlockHaveProfileData(BasicBlock * BB)2476 bool JumpThreadingPass::doesBlockHaveProfileData(BasicBlock *BB) {
2477 const Instruction *TI = BB->getTerminator();
2478 assert(TI->getNumSuccessors() > 1 && "not a split");
2479
2480 MDNode *WeightsNode = TI->getMetadata(LLVMContext::MD_prof);
2481 if (!WeightsNode)
2482 return false;
2483
2484 MDString *MDName = cast<MDString>(WeightsNode->getOperand(0));
2485 if (MDName->getString() != "branch_weights")
2486 return false;
2487
2488 // Ensure there are weights for all of the successors. Note that the first
2489 // operand to the metadata node is a name, not a weight.
2490 return WeightsNode->getNumOperands() == TI->getNumSuccessors() + 1;
2491 }
2492
2493 /// Update the block frequency of BB and branch weight and the metadata on the
2494 /// edge BB->SuccBB. This is done by scaling the weight of BB->SuccBB by 1 -
2495 /// Freq(PredBB->BB) / Freq(BB->SuccBB).
updateBlockFreqAndEdgeWeight(BasicBlock * PredBB,BasicBlock * BB,BasicBlock * NewBB,BasicBlock * SuccBB)2496 void JumpThreadingPass::updateBlockFreqAndEdgeWeight(BasicBlock *PredBB,
2497 BasicBlock *BB,
2498 BasicBlock *NewBB,
2499 BasicBlock *SuccBB) {
2500 if (!HasProfileData)
2501 return;
2502
2503 assert(BFI && BPI && "BFI & BPI should have been created here");
2504
2505 // As the edge from PredBB to BB is deleted, we have to update the block
2506 // frequency of BB.
2507 auto BBOrigFreq = BFI->getBlockFreq(BB);
2508 auto NewBBFreq = BFI->getBlockFreq(NewBB);
2509 auto BB2SuccBBFreq = BBOrigFreq * BPI->getEdgeProbability(BB, SuccBB);
2510 auto BBNewFreq = BBOrigFreq - NewBBFreq;
2511 BFI->setBlockFreq(BB, BBNewFreq.getFrequency());
2512
2513 // Collect updated outgoing edges' frequencies from BB and use them to update
2514 // edge probabilities.
2515 SmallVector<uint64_t, 4> BBSuccFreq;
2516 for (BasicBlock *Succ : successors(BB)) {
2517 auto SuccFreq = (Succ == SuccBB)
2518 ? BB2SuccBBFreq - NewBBFreq
2519 : BBOrigFreq * BPI->getEdgeProbability(BB, Succ);
2520 BBSuccFreq.push_back(SuccFreq.getFrequency());
2521 }
2522
2523 uint64_t MaxBBSuccFreq =
2524 *std::max_element(BBSuccFreq.begin(), BBSuccFreq.end());
2525
2526 SmallVector<BranchProbability, 4> BBSuccProbs;
2527 if (MaxBBSuccFreq == 0)
2528 BBSuccProbs.assign(BBSuccFreq.size(),
2529 {1, static_cast<unsigned>(BBSuccFreq.size())});
2530 else {
2531 for (uint64_t Freq : BBSuccFreq)
2532 BBSuccProbs.push_back(
2533 BranchProbability::getBranchProbability(Freq, MaxBBSuccFreq));
2534 // Normalize edge probabilities so that they sum up to one.
2535 BranchProbability::normalizeProbabilities(BBSuccProbs.begin(),
2536 BBSuccProbs.end());
2537 }
2538
2539 // Update edge probabilities in BPI.
2540 BPI->setEdgeProbability(BB, BBSuccProbs);
2541
2542 // Update the profile metadata as well.
2543 //
2544 // Don't do this if the profile of the transformed blocks was statically
2545 // estimated. (This could occur despite the function having an entry
2546 // frequency in completely cold parts of the CFG.)
2547 //
2548 // In this case we don't want to suggest to subsequent passes that the
2549 // calculated weights are fully consistent. Consider this graph:
2550 //
2551 // check_1
2552 // 50% / |
2553 // eq_1 | 50%
2554 // \ |
2555 // check_2
2556 // 50% / |
2557 // eq_2 | 50%
2558 // \ |
2559 // check_3
2560 // 50% / |
2561 // eq_3 | 50%
2562 // \ |
2563 //
2564 // Assuming the blocks check_* all compare the same value against 1, 2 and 3,
2565 // the overall probabilities are inconsistent; the total probability that the
2566 // value is either 1, 2 or 3 is 150%.
2567 //
2568 // As a consequence if we thread eq_1 -> check_2 to check_3, check_2->check_3
2569 // becomes 0%. This is even worse if the edge whose probability becomes 0% is
2570 // the loop exit edge. Then based solely on static estimation we would assume
2571 // the loop was extremely hot.
2572 //
2573 // FIXME this locally as well so that BPI and BFI are consistent as well. We
2574 // shouldn't make edges extremely likely or unlikely based solely on static
2575 // estimation.
2576 if (BBSuccProbs.size() >= 2 && doesBlockHaveProfileData(BB)) {
2577 SmallVector<uint32_t, 4> Weights;
2578 for (auto Prob : BBSuccProbs)
2579 Weights.push_back(Prob.getNumerator());
2580
2581 auto TI = BB->getTerminator();
2582 TI->setMetadata(
2583 LLVMContext::MD_prof,
2584 MDBuilder(TI->getParent()->getContext()).createBranchWeights(Weights));
2585 }
2586 }
2587
2588 /// duplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
2589 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
2590 /// If we can duplicate the contents of BB up into PredBB do so now, this
2591 /// improves the odds that the branch will be on an analyzable instruction like
2592 /// a compare.
duplicateCondBranchOnPHIIntoPred(BasicBlock * BB,const SmallVectorImpl<BasicBlock * > & PredBBs)2593 bool JumpThreadingPass::duplicateCondBranchOnPHIIntoPred(
2594 BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs) {
2595 assert(!PredBBs.empty() && "Can't handle an empty set");
2596
2597 // If BB is a loop header, then duplicating this block outside the loop would
2598 // cause us to transform this into an irreducible loop, don't do this.
2599 // See the comments above findLoopHeaders for justifications and caveats.
2600 if (LoopHeaders.count(BB)) {
2601 LLVM_DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
2602 << "' into predecessor block '" << PredBBs[0]->getName()
2603 << "' - it might create an irreducible loop!\n");
2604 return false;
2605 }
2606
2607 unsigned DuplicationCost =
2608 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
2609 if (DuplicationCost > BBDupThreshold) {
2610 LLVM_DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
2611 << "' - Cost is too high: " << DuplicationCost << "\n");
2612 return false;
2613 }
2614
2615 // And finally, do it! Start by factoring the predecessors if needed.
2616 std::vector<DominatorTree::UpdateType> Updates;
2617 BasicBlock *PredBB;
2618 if (PredBBs.size() == 1)
2619 PredBB = PredBBs[0];
2620 else {
2621 LLVM_DEBUG(dbgs() << " Factoring out " << PredBBs.size()
2622 << " common predecessors.\n");
2623 PredBB = splitBlockPreds(BB, PredBBs, ".thr_comm");
2624 }
2625 Updates.push_back({DominatorTree::Delete, PredBB, BB});
2626
2627 // Okay, we decided to do this! Clone all the instructions in BB onto the end
2628 // of PredBB.
2629 LLVM_DEBUG(dbgs() << " Duplicating block '" << BB->getName()
2630 << "' into end of '" << PredBB->getName()
2631 << "' to eliminate branch on phi. Cost: "
2632 << DuplicationCost << " block is:" << *BB << "\n");
2633
2634 // Unless PredBB ends with an unconditional branch, split the edge so that we
2635 // can just clone the bits from BB into the end of the new PredBB.
2636 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
2637
2638 if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
2639 BasicBlock *OldPredBB = PredBB;
2640 PredBB = SplitEdge(OldPredBB, BB);
2641 Updates.push_back({DominatorTree::Insert, OldPredBB, PredBB});
2642 Updates.push_back({DominatorTree::Insert, PredBB, BB});
2643 Updates.push_back({DominatorTree::Delete, OldPredBB, BB});
2644 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
2645 }
2646
2647 // We are going to have to map operands from the original BB block into the
2648 // PredBB block. Evaluate PHI nodes in BB.
2649 DenseMap<Instruction*, Value*> ValueMapping;
2650
2651 BasicBlock::iterator BI = BB->begin();
2652 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
2653 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
2654 // Clone the non-phi instructions of BB into PredBB, keeping track of the
2655 // mapping and using it to remap operands in the cloned instructions.
2656 for (; BI != BB->end(); ++BI) {
2657 Instruction *New = BI->clone();
2658
2659 // Remap operands to patch up intra-block references.
2660 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2661 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
2662 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
2663 if (I != ValueMapping.end())
2664 New->setOperand(i, I->second);
2665 }
2666
2667 // If this instruction can be simplified after the operands are updated,
2668 // just use the simplified value instead. This frequently happens due to
2669 // phi translation.
2670 if (Value *IV = SimplifyInstruction(
2671 New,
2672 {BB->getModule()->getDataLayout(), TLI, nullptr, nullptr, New})) {
2673 ValueMapping[&*BI] = IV;
2674 if (!New->mayHaveSideEffects()) {
2675 New->deleteValue();
2676 New = nullptr;
2677 }
2678 } else {
2679 ValueMapping[&*BI] = New;
2680 }
2681 if (New) {
2682 // Otherwise, insert the new instruction into the block.
2683 New->setName(BI->getName());
2684 PredBB->getInstList().insert(OldPredBranch->getIterator(), New);
2685 // Update Dominance from simplified New instruction operands.
2686 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2687 if (BasicBlock *SuccBB = dyn_cast<BasicBlock>(New->getOperand(i)))
2688 Updates.push_back({DominatorTree::Insert, PredBB, SuccBB});
2689 }
2690 }
2691
2692 // Check to see if the targets of the branch had PHI nodes. If so, we need to
2693 // add entries to the PHI nodes for branch from PredBB now.
2694 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
2695 addPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
2696 ValueMapping);
2697 addPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
2698 ValueMapping);
2699
2700 updateSSA(BB, PredBB, ValueMapping);
2701
2702 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
2703 // that we nuked.
2704 BB->removePredecessor(PredBB, true);
2705
2706 // Remove the unconditional branch at the end of the PredBB block.
2707 OldPredBranch->eraseFromParent();
2708 if (HasProfileData)
2709 BPI->copyEdgeProbabilities(BB, PredBB);
2710 DTU->applyUpdatesPermissive(Updates);
2711
2712 ++NumDupes;
2713 return true;
2714 }
2715
2716 // Pred is a predecessor of BB with an unconditional branch to BB. SI is
2717 // a Select instruction in Pred. BB has other predecessors and SI is used in
2718 // a PHI node in BB. SI has no other use.
2719 // A new basic block, NewBB, is created and SI is converted to compare and
2720 // conditional branch. SI is erased from parent.
unfoldSelectInstr(BasicBlock * Pred,BasicBlock * BB,SelectInst * SI,PHINode * SIUse,unsigned Idx)2721 void JumpThreadingPass::unfoldSelectInstr(BasicBlock *Pred, BasicBlock *BB,
2722 SelectInst *SI, PHINode *SIUse,
2723 unsigned Idx) {
2724 // Expand the select.
2725 //
2726 // Pred --
2727 // | v
2728 // | NewBB
2729 // | |
2730 // |-----
2731 // v
2732 // BB
2733 BranchInst *PredTerm = cast<BranchInst>(Pred->getTerminator());
2734 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
2735 BB->getParent(), BB);
2736 // Move the unconditional branch to NewBB.
2737 PredTerm->removeFromParent();
2738 NewBB->getInstList().insert(NewBB->end(), PredTerm);
2739 // Create a conditional branch and update PHI nodes.
2740 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
2741 SIUse->setIncomingValue(Idx, SI->getFalseValue());
2742 SIUse->addIncoming(SI->getTrueValue(), NewBB);
2743
2744 // The select is now dead.
2745 SI->eraseFromParent();
2746 DTU->applyUpdatesPermissive({{DominatorTree::Insert, NewBB, BB},
2747 {DominatorTree::Insert, Pred, NewBB}});
2748
2749 // Update any other PHI nodes in BB.
2750 for (BasicBlock::iterator BI = BB->begin();
2751 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
2752 if (Phi != SIUse)
2753 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);
2754 }
2755
tryToUnfoldSelect(SwitchInst * SI,BasicBlock * BB)2756 bool JumpThreadingPass::tryToUnfoldSelect(SwitchInst *SI, BasicBlock *BB) {
2757 PHINode *CondPHI = dyn_cast<PHINode>(SI->getCondition());
2758
2759 if (!CondPHI || CondPHI->getParent() != BB)
2760 return false;
2761
2762 for (unsigned I = 0, E = CondPHI->getNumIncomingValues(); I != E; ++I) {
2763 BasicBlock *Pred = CondPHI->getIncomingBlock(I);
2764 SelectInst *PredSI = dyn_cast<SelectInst>(CondPHI->getIncomingValue(I));
2765
2766 // The second and third condition can be potentially relaxed. Currently
2767 // the conditions help to simplify the code and allow us to reuse existing
2768 // code, developed for tryToUnfoldSelect(CmpInst *, BasicBlock *)
2769 if (!PredSI || PredSI->getParent() != Pred || !PredSI->hasOneUse())
2770 continue;
2771
2772 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
2773 if (!PredTerm || !PredTerm->isUnconditional())
2774 continue;
2775
2776 unfoldSelectInstr(Pred, BB, PredSI, CondPHI, I);
2777 return true;
2778 }
2779 return false;
2780 }
2781
2782 /// tryToUnfoldSelect - Look for blocks of the form
2783 /// bb1:
2784 /// %a = select
2785 /// br bb2
2786 ///
2787 /// bb2:
2788 /// %p = phi [%a, %bb1] ...
2789 /// %c = icmp %p
2790 /// br i1 %c
2791 ///
2792 /// And expand the select into a branch structure if one of its arms allows %c
2793 /// to be folded. This later enables threading from bb1 over bb2.
tryToUnfoldSelect(CmpInst * CondCmp,BasicBlock * BB)2794 bool JumpThreadingPass::tryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
2795 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
2796 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
2797 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
2798
2799 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
2800 CondLHS->getParent() != BB)
2801 return false;
2802
2803 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
2804 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
2805 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
2806
2807 // Look if one of the incoming values is a select in the corresponding
2808 // predecessor.
2809 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
2810 continue;
2811
2812 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
2813 if (!PredTerm || !PredTerm->isUnconditional())
2814 continue;
2815
2816 // Now check if one of the select values would allow us to constant fold the
2817 // terminator in BB. We don't do the transform if both sides fold, those
2818 // cases will be threaded in any case.
2819 LazyValueInfo::Tristate LHSFolds =
2820 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
2821 CondRHS, Pred, BB, CondCmp);
2822 LazyValueInfo::Tristate RHSFolds =
2823 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
2824 CondRHS, Pred, BB, CondCmp);
2825 if ((LHSFolds != LazyValueInfo::Unknown ||
2826 RHSFolds != LazyValueInfo::Unknown) &&
2827 LHSFolds != RHSFolds) {
2828 unfoldSelectInstr(Pred, BB, SI, CondLHS, I);
2829 return true;
2830 }
2831 }
2832 return false;
2833 }
2834
2835 /// tryToUnfoldSelectInCurrBB - Look for PHI/Select or PHI/CMP/Select in the
2836 /// same BB in the form
2837 /// bb:
2838 /// %p = phi [false, %bb1], [true, %bb2], [false, %bb3], [true, %bb4], ...
2839 /// %s = select %p, trueval, falseval
2840 ///
2841 /// or
2842 ///
2843 /// bb:
2844 /// %p = phi [0, %bb1], [1, %bb2], [0, %bb3], [1, %bb4], ...
2845 /// %c = cmp %p, 0
2846 /// %s = select %c, trueval, falseval
2847 ///
2848 /// And expand the select into a branch structure. This later enables
2849 /// jump-threading over bb in this pass.
2850 ///
2851 /// Using the similar approach of SimplifyCFG::FoldCondBranchOnPHI(), unfold
2852 /// select if the associated PHI has at least one constant. If the unfolded
2853 /// select is not jump-threaded, it will be folded again in the later
2854 /// optimizations.
tryToUnfoldSelectInCurrBB(BasicBlock * BB)2855 bool JumpThreadingPass::tryToUnfoldSelectInCurrBB(BasicBlock *BB) {
2856 // This transform would reduce the quality of msan diagnostics.
2857 // Disable this transform under MemorySanitizer.
2858 if (BB->getParent()->hasFnAttribute(Attribute::SanitizeMemory))
2859 return false;
2860
2861 // If threading this would thread across a loop header, don't thread the edge.
2862 // See the comments above findLoopHeaders for justifications and caveats.
2863 if (LoopHeaders.count(BB))
2864 return false;
2865
2866 for (BasicBlock::iterator BI = BB->begin();
2867 PHINode *PN = dyn_cast<PHINode>(BI); ++BI) {
2868 // Look for a Phi having at least one constant incoming value.
2869 if (llvm::all_of(PN->incoming_values(),
2870 [](Value *V) { return !isa<ConstantInt>(V); }))
2871 continue;
2872
2873 auto isUnfoldCandidate = [BB](SelectInst *SI, Value *V) {
2874 // Check if SI is in BB and use V as condition.
2875 if (SI->getParent() != BB)
2876 return false;
2877 Value *Cond = SI->getCondition();
2878 return (Cond && Cond == V && Cond->getType()->isIntegerTy(1));
2879 };
2880
2881 SelectInst *SI = nullptr;
2882 for (Use &U : PN->uses()) {
2883 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(U.getUser())) {
2884 // Look for a ICmp in BB that compares PN with a constant and is the
2885 // condition of a Select.
2886 if (Cmp->getParent() == BB && Cmp->hasOneUse() &&
2887 isa<ConstantInt>(Cmp->getOperand(1 - U.getOperandNo())))
2888 if (SelectInst *SelectI = dyn_cast<SelectInst>(Cmp->user_back()))
2889 if (isUnfoldCandidate(SelectI, Cmp->use_begin()->get())) {
2890 SI = SelectI;
2891 break;
2892 }
2893 } else if (SelectInst *SelectI = dyn_cast<SelectInst>(U.getUser())) {
2894 // Look for a Select in BB that uses PN as condition.
2895 if (isUnfoldCandidate(SelectI, U.get())) {
2896 SI = SelectI;
2897 break;
2898 }
2899 }
2900 }
2901
2902 if (!SI)
2903 continue;
2904 // Expand the select.
2905 Value *Cond = SI->getCondition();
2906 if (InsertFreezeWhenUnfoldingSelect &&
2907 !isGuaranteedNotToBeUndefOrPoison(Cond, nullptr, SI,
2908 &DTU->getDomTree()))
2909 Cond = new FreezeInst(Cond, "cond.fr", SI);
2910 Instruction *Term = SplitBlockAndInsertIfThen(Cond, SI, false);
2911 BasicBlock *SplitBB = SI->getParent();
2912 BasicBlock *NewBB = Term->getParent();
2913 PHINode *NewPN = PHINode::Create(SI->getType(), 2, "", SI);
2914 NewPN->addIncoming(SI->getTrueValue(), Term->getParent());
2915 NewPN->addIncoming(SI->getFalseValue(), BB);
2916 SI->replaceAllUsesWith(NewPN);
2917 SI->eraseFromParent();
2918 // NewBB and SplitBB are newly created blocks which require insertion.
2919 std::vector<DominatorTree::UpdateType> Updates;
2920 Updates.reserve((2 * SplitBB->getTerminator()->getNumSuccessors()) + 3);
2921 Updates.push_back({DominatorTree::Insert, BB, SplitBB});
2922 Updates.push_back({DominatorTree::Insert, BB, NewBB});
2923 Updates.push_back({DominatorTree::Insert, NewBB, SplitBB});
2924 // BB's successors were moved to SplitBB, update DTU accordingly.
2925 for (auto *Succ : successors(SplitBB)) {
2926 Updates.push_back({DominatorTree::Delete, BB, Succ});
2927 Updates.push_back({DominatorTree::Insert, SplitBB, Succ});
2928 }
2929 DTU->applyUpdatesPermissive(Updates);
2930 return true;
2931 }
2932 return false;
2933 }
2934
2935 /// Try to propagate a guard from the current BB into one of its predecessors
2936 /// in case if another branch of execution implies that the condition of this
2937 /// guard is always true. Currently we only process the simplest case that
2938 /// looks like:
2939 ///
2940 /// Start:
2941 /// %cond = ...
2942 /// br i1 %cond, label %T1, label %F1
2943 /// T1:
2944 /// br label %Merge
2945 /// F1:
2946 /// br label %Merge
2947 /// Merge:
2948 /// %condGuard = ...
2949 /// call void(i1, ...) @llvm.experimental.guard( i1 %condGuard )[ "deopt"() ]
2950 ///
2951 /// And cond either implies condGuard or !condGuard. In this case all the
2952 /// instructions before the guard can be duplicated in both branches, and the
2953 /// guard is then threaded to one of them.
processGuards(BasicBlock * BB)2954 bool JumpThreadingPass::processGuards(BasicBlock *BB) {
2955 using namespace PatternMatch;
2956
2957 // We only want to deal with two predecessors.
2958 BasicBlock *Pred1, *Pred2;
2959 auto PI = pred_begin(BB), PE = pred_end(BB);
2960 if (PI == PE)
2961 return false;
2962 Pred1 = *PI++;
2963 if (PI == PE)
2964 return false;
2965 Pred2 = *PI++;
2966 if (PI != PE)
2967 return false;
2968 if (Pred1 == Pred2)
2969 return false;
2970
2971 // Try to thread one of the guards of the block.
2972 // TODO: Look up deeper than to immediate predecessor?
2973 auto *Parent = Pred1->getSinglePredecessor();
2974 if (!Parent || Parent != Pred2->getSinglePredecessor())
2975 return false;
2976
2977 if (auto *BI = dyn_cast<BranchInst>(Parent->getTerminator()))
2978 for (auto &I : *BB)
2979 if (isGuard(&I) && threadGuard(BB, cast<IntrinsicInst>(&I), BI))
2980 return true;
2981
2982 return false;
2983 }
2984
2985 /// Try to propagate the guard from BB which is the lower block of a diamond
2986 /// to one of its branches, in case if diamond's condition implies guard's
2987 /// condition.
threadGuard(BasicBlock * BB,IntrinsicInst * Guard,BranchInst * BI)2988 bool JumpThreadingPass::threadGuard(BasicBlock *BB, IntrinsicInst *Guard,
2989 BranchInst *BI) {
2990 assert(BI->getNumSuccessors() == 2 && "Wrong number of successors?");
2991 assert(BI->isConditional() && "Unconditional branch has 2 successors?");
2992 Value *GuardCond = Guard->getArgOperand(0);
2993 Value *BranchCond = BI->getCondition();
2994 BasicBlock *TrueDest = BI->getSuccessor(0);
2995 BasicBlock *FalseDest = BI->getSuccessor(1);
2996
2997 auto &DL = BB->getModule()->getDataLayout();
2998 bool TrueDestIsSafe = false;
2999 bool FalseDestIsSafe = false;
3000
3001 // True dest is safe if BranchCond => GuardCond.
3002 auto Impl = isImpliedCondition(BranchCond, GuardCond, DL);
3003 if (Impl && *Impl)
3004 TrueDestIsSafe = true;
3005 else {
3006 // False dest is safe if !BranchCond => GuardCond.
3007 Impl = isImpliedCondition(BranchCond, GuardCond, DL, /* LHSIsTrue */ false);
3008 if (Impl && *Impl)
3009 FalseDestIsSafe = true;
3010 }
3011
3012 if (!TrueDestIsSafe && !FalseDestIsSafe)
3013 return false;
3014
3015 BasicBlock *PredUnguardedBlock = TrueDestIsSafe ? TrueDest : FalseDest;
3016 BasicBlock *PredGuardedBlock = FalseDestIsSafe ? TrueDest : FalseDest;
3017
3018 ValueToValueMapTy UnguardedMapping, GuardedMapping;
3019 Instruction *AfterGuard = Guard->getNextNode();
3020 unsigned Cost = getJumpThreadDuplicationCost(BB, AfterGuard, BBDupThreshold);
3021 if (Cost > BBDupThreshold)
3022 return false;
3023 // Duplicate all instructions before the guard and the guard itself to the
3024 // branch where implication is not proved.
3025 BasicBlock *GuardedBlock = DuplicateInstructionsInSplitBetween(
3026 BB, PredGuardedBlock, AfterGuard, GuardedMapping, *DTU);
3027 assert(GuardedBlock && "Could not create the guarded block?");
3028 // Duplicate all instructions before the guard in the unguarded branch.
3029 // Since we have successfully duplicated the guarded block and this block
3030 // has fewer instructions, we expect it to succeed.
3031 BasicBlock *UnguardedBlock = DuplicateInstructionsInSplitBetween(
3032 BB, PredUnguardedBlock, Guard, UnguardedMapping, *DTU);
3033 assert(UnguardedBlock && "Could not create the unguarded block?");
3034 LLVM_DEBUG(dbgs() << "Moved guard " << *Guard << " to block "
3035 << GuardedBlock->getName() << "\n");
3036 // Some instructions before the guard may still have uses. For them, we need
3037 // to create Phi nodes merging their copies in both guarded and unguarded
3038 // branches. Those instructions that have no uses can be just removed.
3039 SmallVector<Instruction *, 4> ToRemove;
3040 for (auto BI = BB->begin(); &*BI != AfterGuard; ++BI)
3041 if (!isa<PHINode>(&*BI))
3042 ToRemove.push_back(&*BI);
3043
3044 Instruction *InsertionPoint = &*BB->getFirstInsertionPt();
3045 assert(InsertionPoint && "Empty block?");
3046 // Substitute with Phis & remove.
3047 for (auto *Inst : reverse(ToRemove)) {
3048 if (!Inst->use_empty()) {
3049 PHINode *NewPN = PHINode::Create(Inst->getType(), 2);
3050 NewPN->addIncoming(UnguardedMapping[Inst], UnguardedBlock);
3051 NewPN->addIncoming(GuardedMapping[Inst], GuardedBlock);
3052 NewPN->insertBefore(InsertionPoint);
3053 Inst->replaceAllUsesWith(NewPN);
3054 }
3055 Inst->eraseFromParent();
3056 }
3057 return true;
3058 }
3059