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/Optional.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/ADT/SmallVector.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Analysis/AliasAnalysis.h"
22 #include "llvm/Analysis/BlockFrequencyInfo.h"
23 #include "llvm/Analysis/BranchProbabilityInfo.h"
24 #include "llvm/Analysis/CFG.h"
25 #include "llvm/Analysis/ConstantFolding.h"
26 #include "llvm/Analysis/DomTreeUpdater.h"
27 #include "llvm/Analysis/GlobalsModRef.h"
28 #include "llvm/Analysis/GuardUtils.h"
29 #include "llvm/Analysis/InstructionSimplify.h"
30 #include "llvm/Analysis/LazyValueInfo.h"
31 #include "llvm/Analysis/Loads.h"
32 #include "llvm/Analysis/LoopInfo.h"
33 #include "llvm/Analysis/TargetLibraryInfo.h"
34 #include "llvm/Analysis/ValueTracking.h"
35 #include "llvm/IR/BasicBlock.h"
36 #include "llvm/IR/CFG.h"
37 #include "llvm/IR/Constant.h"
38 #include "llvm/IR/ConstantRange.h"
39 #include "llvm/IR/Constants.h"
40 #include "llvm/IR/DataLayout.h"
41 #include "llvm/IR/Dominators.h"
42 #include "llvm/IR/Function.h"
43 #include "llvm/IR/InstrTypes.h"
44 #include "llvm/IR/Instruction.h"
45 #include "llvm/IR/Instructions.h"
46 #include "llvm/IR/IntrinsicInst.h"
47 #include "llvm/IR/Intrinsics.h"
48 #include "llvm/IR/LLVMContext.h"
49 #include "llvm/IR/MDBuilder.h"
50 #include "llvm/IR/Metadata.h"
51 #include "llvm/IR/Module.h"
52 #include "llvm/IR/PassManager.h"
53 #include "llvm/IR/PatternMatch.h"
54 #include "llvm/IR/Type.h"
55 #include "llvm/IR/Use.h"
56 #include "llvm/IR/User.h"
57 #include "llvm/IR/Value.h"
58 #include "llvm/InitializePasses.h"
59 #include "llvm/Pass.h"
60 #include "llvm/Support/BlockFrequency.h"
61 #include "llvm/Support/BranchProbability.h"
62 #include "llvm/Support/Casting.h"
63 #include "llvm/Support/CommandLine.h"
64 #include "llvm/Support/Debug.h"
65 #include "llvm/Support/raw_ostream.h"
66 #include "llvm/Transforms/Scalar.h"
67 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
68 #include "llvm/Transforms/Utils/Cloning.h"
69 #include "llvm/Transforms/Utils/Local.h"
70 #include "llvm/Transforms/Utils/SSAUpdater.h"
71 #include "llvm/Transforms/Utils/ValueMapper.h"
72 #include <algorithm>
73 #include <cassert>
74 #include <cstddef>
75 #include <cstdint>
76 #include <iterator>
77 #include <memory>
78 #include <utility>
79
80 using namespace llvm;
81 using namespace jumpthreading;
82
83 #define DEBUG_TYPE "jump-threading"
84
85 STATISTIC(NumThreads, "Number of jumps threaded");
86 STATISTIC(NumFolds, "Number of terminators folded");
87 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
88
89 static cl::opt<unsigned>
90 BBDuplicateThreshold("jump-threading-threshold",
91 cl::desc("Max block size to duplicate for jump threading"),
92 cl::init(6), cl::Hidden);
93
94 static cl::opt<unsigned>
95 ImplicationSearchThreshold(
96 "jump-threading-implication-search-threshold",
97 cl::desc("The number of predecessors to search for a stronger "
98 "condition to use to thread over a weaker condition"),
99 cl::init(3), cl::Hidden);
100
101 static cl::opt<bool> PrintLVIAfterJumpThreading(
102 "print-lvi-after-jump-threading",
103 cl::desc("Print the LazyValueInfo cache after JumpThreading"), cl::init(false),
104 cl::Hidden);
105
106 static cl::opt<bool> ThreadAcrossLoopHeaders(
107 "jump-threading-across-loop-headers",
108 cl::desc("Allow JumpThreading to thread across loop headers, for testing"),
109 cl::init(false), cl::Hidden);
110
111
112 namespace {
113
114 /// This pass performs 'jump threading', which looks at blocks that have
115 /// multiple predecessors and multiple successors. If one or more of the
116 /// predecessors of the block can be proven to always jump to one of the
117 /// successors, we forward the edge from the predecessor to the successor by
118 /// duplicating the contents of this block.
119 ///
120 /// An example of when this can occur is code like this:
121 ///
122 /// if () { ...
123 /// X = 4;
124 /// }
125 /// if (X < 3) {
126 ///
127 /// In this case, the unconditional branch at the end of the first if can be
128 /// revectored to the false side of the second if.
129 class JumpThreading : public FunctionPass {
130 JumpThreadingPass Impl;
131
132 public:
133 static char ID; // Pass identification
134
JumpThreading(int T=-1)135 JumpThreading(int T = -1) : FunctionPass(ID), Impl(T) {
136 initializeJumpThreadingPass(*PassRegistry::getPassRegistry());
137 }
138
139 bool runOnFunction(Function &F) override;
140
getAnalysisUsage(AnalysisUsage & AU) const141 void getAnalysisUsage(AnalysisUsage &AU) const override {
142 AU.addRequired<DominatorTreeWrapperPass>();
143 AU.addPreserved<DominatorTreeWrapperPass>();
144 AU.addRequired<AAResultsWrapperPass>();
145 AU.addRequired<LazyValueInfoWrapperPass>();
146 AU.addPreserved<LazyValueInfoWrapperPass>();
147 AU.addPreserved<GlobalsAAWrapperPass>();
148 AU.addRequired<TargetLibraryInfoWrapperPass>();
149 }
150
releaseMemory()151 void releaseMemory() override { Impl.releaseMemory(); }
152 };
153
154 } // end anonymous namespace
155
156 char JumpThreading::ID = 0;
157
158 INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",
159 "Jump Threading", false, false)
INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)160 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
161 INITIALIZE_PASS_DEPENDENCY(LazyValueInfoWrapperPass)
162 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
163 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
164 INITIALIZE_PASS_END(JumpThreading, "jump-threading",
165 "Jump Threading", false, false)
166
167 // Public interface to the Jump Threading pass
168 FunctionPass *llvm::createJumpThreadingPass(int Threshold) {
169 return new JumpThreading(Threshold);
170 }
171
JumpThreadingPass(int T)172 JumpThreadingPass::JumpThreadingPass(int T) {
173 BBDupThreshold = (T == -1) ? BBDuplicateThreshold : unsigned(T);
174 }
175
176 // Update branch probability information according to conditional
177 // branch probability. This is usually made possible for cloned branches
178 // in inline instances by the context specific profile in the caller.
179 // For instance,
180 //
181 // [Block PredBB]
182 // [Branch PredBr]
183 // if (t) {
184 // Block A;
185 // } else {
186 // Block B;
187 // }
188 //
189 // [Block BB]
190 // cond = PN([true, %A], [..., %B]); // PHI node
191 // [Branch CondBr]
192 // if (cond) {
193 // ... // P(cond == true) = 1%
194 // }
195 //
196 // Here we know that when block A is taken, cond must be true, which means
197 // P(cond == true | A) = 1
198 //
199 // Given that P(cond == true) = P(cond == true | A) * P(A) +
200 // P(cond == true | B) * P(B)
201 // we get:
202 // P(cond == true ) = P(A) + P(cond == true | B) * P(B)
203 //
204 // which gives us:
205 // P(A) is less than P(cond == true), i.e.
206 // P(t == true) <= P(cond == true)
207 //
208 // In other words, if we know P(cond == true) is unlikely, we know
209 // that P(t == true) is also unlikely.
210 //
updatePredecessorProfileMetadata(PHINode * PN,BasicBlock * BB)211 static void updatePredecessorProfileMetadata(PHINode *PN, BasicBlock *BB) {
212 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
213 if (!CondBr)
214 return;
215
216 BranchProbability BP;
217 uint64_t TrueWeight, FalseWeight;
218 if (!CondBr->extractProfMetadata(TrueWeight, FalseWeight))
219 return;
220
221 // Returns the outgoing edge of the dominating predecessor block
222 // that leads to the PhiNode's incoming block:
223 auto GetPredOutEdge =
224 [](BasicBlock *IncomingBB,
225 BasicBlock *PhiBB) -> std::pair<BasicBlock *, BasicBlock *> {
226 auto *PredBB = IncomingBB;
227 auto *SuccBB = PhiBB;
228 SmallPtrSet<BasicBlock *, 16> Visited;
229 while (true) {
230 BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator());
231 if (PredBr && PredBr->isConditional())
232 return {PredBB, SuccBB};
233 Visited.insert(PredBB);
234 auto *SinglePredBB = PredBB->getSinglePredecessor();
235 if (!SinglePredBB)
236 return {nullptr, nullptr};
237
238 // Stop searching when SinglePredBB has been visited. It means we see
239 // an unreachable loop.
240 if (Visited.count(SinglePredBB))
241 return {nullptr, nullptr};
242
243 SuccBB = PredBB;
244 PredBB = SinglePredBB;
245 }
246 };
247
248 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
249 Value *PhiOpnd = PN->getIncomingValue(i);
250 ConstantInt *CI = dyn_cast<ConstantInt>(PhiOpnd);
251
252 if (!CI || !CI->getType()->isIntegerTy(1))
253 continue;
254
255 BP = (CI->isOne() ? BranchProbability::getBranchProbability(
256 TrueWeight, TrueWeight + FalseWeight)
257 : BranchProbability::getBranchProbability(
258 FalseWeight, TrueWeight + FalseWeight));
259
260 auto PredOutEdge = GetPredOutEdge(PN->getIncomingBlock(i), BB);
261 if (!PredOutEdge.first)
262 return;
263
264 BasicBlock *PredBB = PredOutEdge.first;
265 BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator());
266 if (!PredBr)
267 return;
268
269 uint64_t PredTrueWeight, PredFalseWeight;
270 // FIXME: We currently only set the profile data when it is missing.
271 // With PGO, this can be used to refine even existing profile data with
272 // context information. This needs to be done after more performance
273 // testing.
274 if (PredBr->extractProfMetadata(PredTrueWeight, PredFalseWeight))
275 continue;
276
277 // We can not infer anything useful when BP >= 50%, because BP is the
278 // upper bound probability value.
279 if (BP >= BranchProbability(50, 100))
280 continue;
281
282 SmallVector<uint32_t, 2> Weights;
283 if (PredBr->getSuccessor(0) == PredOutEdge.second) {
284 Weights.push_back(BP.getNumerator());
285 Weights.push_back(BP.getCompl().getNumerator());
286 } else {
287 Weights.push_back(BP.getCompl().getNumerator());
288 Weights.push_back(BP.getNumerator());
289 }
290 PredBr->setMetadata(LLVMContext::MD_prof,
291 MDBuilder(PredBr->getParent()->getContext())
292 .createBranchWeights(Weights));
293 }
294 }
295
296 /// runOnFunction - Toplevel algorithm.
runOnFunction(Function & F)297 bool JumpThreading::runOnFunction(Function &F) {
298 if (skipFunction(F))
299 return false;
300 auto TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
301 // Get DT analysis before LVI. When LVI is initialized it conditionally adds
302 // DT if it's available.
303 auto DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
304 auto LVI = &getAnalysis<LazyValueInfoWrapperPass>().getLVI();
305 auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
306 DomTreeUpdater DTU(*DT, DomTreeUpdater::UpdateStrategy::Lazy);
307 std::unique_ptr<BlockFrequencyInfo> BFI;
308 std::unique_ptr<BranchProbabilityInfo> BPI;
309 if (F.hasProfileData()) {
310 LoopInfo LI{DominatorTree(F)};
311 BPI.reset(new BranchProbabilityInfo(F, LI, TLI));
312 BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
313 }
314
315 bool Changed = Impl.runImpl(F, TLI, LVI, AA, &DTU, F.hasProfileData(),
316 std::move(BFI), std::move(BPI));
317 if (PrintLVIAfterJumpThreading) {
318 dbgs() << "LVI for function '" << F.getName() << "':\n";
319 LVI->printLVI(F, *DT, dbgs());
320 }
321 return Changed;
322 }
323
run(Function & F,FunctionAnalysisManager & AM)324 PreservedAnalyses JumpThreadingPass::run(Function &F,
325 FunctionAnalysisManager &AM) {
326 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
327 // Get DT analysis before LVI. When LVI is initialized it conditionally adds
328 // DT if it's available.
329 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
330 auto &LVI = AM.getResult<LazyValueAnalysis>(F);
331 auto &AA = AM.getResult<AAManager>(F);
332 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Lazy);
333
334 std::unique_ptr<BlockFrequencyInfo> BFI;
335 std::unique_ptr<BranchProbabilityInfo> BPI;
336 if (F.hasProfileData()) {
337 LoopInfo LI{DominatorTree(F)};
338 BPI.reset(new BranchProbabilityInfo(F, LI, &TLI));
339 BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
340 }
341
342 bool Changed = runImpl(F, &TLI, &LVI, &AA, &DTU, F.hasProfileData(),
343 std::move(BFI), std::move(BPI));
344
345 if (!Changed)
346 return PreservedAnalyses::all();
347 PreservedAnalyses PA;
348 PA.preserve<GlobalsAA>();
349 PA.preserve<DominatorTreeAnalysis>();
350 PA.preserve<LazyValueAnalysis>();
351 return PA;
352 }
353
runImpl(Function & F,TargetLibraryInfo * TLI_,LazyValueInfo * LVI_,AliasAnalysis * AA_,DomTreeUpdater * DTU_,bool HasProfileData_,std::unique_ptr<BlockFrequencyInfo> BFI_,std::unique_ptr<BranchProbabilityInfo> BPI_)354 bool JumpThreadingPass::runImpl(Function &F, TargetLibraryInfo *TLI_,
355 LazyValueInfo *LVI_, AliasAnalysis *AA_,
356 DomTreeUpdater *DTU_, bool HasProfileData_,
357 std::unique_ptr<BlockFrequencyInfo> BFI_,
358 std::unique_ptr<BranchProbabilityInfo> BPI_) {
359 LLVM_DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
360 TLI = TLI_;
361 LVI = LVI_;
362 AA = AA_;
363 DTU = DTU_;
364 BFI.reset();
365 BPI.reset();
366 // When profile data is available, we need to update edge weights after
367 // successful jump threading, which requires both BPI and BFI being available.
368 HasProfileData = HasProfileData_;
369 auto *GuardDecl = F.getParent()->getFunction(
370 Intrinsic::getName(Intrinsic::experimental_guard));
371 HasGuards = GuardDecl && !GuardDecl->use_empty();
372 if (HasProfileData) {
373 BPI = std::move(BPI_);
374 BFI = std::move(BFI_);
375 }
376
377 // JumpThreading must not processes blocks unreachable from entry. It's a
378 // waste of compute time and can potentially lead to hangs.
379 SmallPtrSet<BasicBlock *, 16> Unreachable;
380 assert(DTU && "DTU isn't passed into JumpThreading before using it.");
381 assert(DTU->hasDomTree() && "JumpThreading relies on DomTree to proceed.");
382 DominatorTree &DT = DTU->getDomTree();
383 for (auto &BB : F)
384 if (!DT.isReachableFromEntry(&BB))
385 Unreachable.insert(&BB);
386
387 if (!ThreadAcrossLoopHeaders)
388 FindLoopHeaders(F);
389
390 bool EverChanged = false;
391 bool Changed;
392 do {
393 Changed = false;
394 for (auto &BB : F) {
395 if (Unreachable.count(&BB))
396 continue;
397 while (ProcessBlock(&BB)) // Thread all of the branches we can over BB.
398 Changed = true;
399 // Stop processing BB if it's the entry or is now deleted. The following
400 // routines attempt to eliminate BB and locating a suitable replacement
401 // for the entry is non-trivial.
402 if (&BB == &F.getEntryBlock() || DTU->isBBPendingDeletion(&BB))
403 continue;
404
405 if (pred_empty(&BB)) {
406 // When ProcessBlock makes BB unreachable it doesn't bother to fix up
407 // the instructions in it. We must remove BB to prevent invalid IR.
408 LLVM_DEBUG(dbgs() << " JT: Deleting dead block '" << BB.getName()
409 << "' with terminator: " << *BB.getTerminator()
410 << '\n');
411 LoopHeaders.erase(&BB);
412 LVI->eraseBlock(&BB);
413 DeleteDeadBlock(&BB, DTU);
414 Changed = true;
415 continue;
416 }
417
418 // ProcessBlock doesn't thread BBs with unconditional TIs. However, if BB
419 // is "almost empty", we attempt to merge BB with its sole successor.
420 auto *BI = dyn_cast<BranchInst>(BB.getTerminator());
421 if (BI && BI->isUnconditional() &&
422 // The terminator must be the only non-phi instruction in BB.
423 BB.getFirstNonPHIOrDbg()->isTerminator() &&
424 // Don't alter Loop headers and latches to ensure another pass can
425 // detect and transform nested loops later.
426 !LoopHeaders.count(&BB) && !LoopHeaders.count(BI->getSuccessor(0)) &&
427 TryToSimplifyUncondBranchFromEmptyBlock(&BB, DTU)) {
428 // BB is valid for cleanup here because we passed in DTU. F remains
429 // BB's parent until a DTU->getDomTree() event.
430 LVI->eraseBlock(&BB);
431 Changed = true;
432 }
433 }
434 EverChanged |= Changed;
435 } while (Changed);
436
437 LoopHeaders.clear();
438 // Flush only the Dominator Tree.
439 DTU->getDomTree();
440 LVI->enableDT();
441 return EverChanged;
442 }
443
444 // Replace uses of Cond with ToVal when safe to do so. If all uses are
445 // replaced, we can remove Cond. We cannot blindly replace all uses of Cond
446 // because we may incorrectly replace uses when guards/assumes are uses of
447 // of `Cond` and we used the guards/assume to reason about the `Cond` value
448 // at the end of block. RAUW unconditionally replaces all uses
449 // including the guards/assumes themselves and the uses before the
450 // guard/assume.
ReplaceFoldableUses(Instruction * Cond,Value * ToVal)451 static void ReplaceFoldableUses(Instruction *Cond, Value *ToVal) {
452 assert(Cond->getType() == ToVal->getType());
453 auto *BB = Cond->getParent();
454 // We can unconditionally replace all uses in non-local blocks (i.e. uses
455 // strictly dominated by BB), since LVI information is true from the
456 // terminator of BB.
457 replaceNonLocalUsesWith(Cond, ToVal);
458 for (Instruction &I : reverse(*BB)) {
459 // Reached the Cond whose uses we are trying to replace, so there are no
460 // more uses.
461 if (&I == Cond)
462 break;
463 // We only replace uses in instructions that are guaranteed to reach the end
464 // of BB, where we know Cond is ToVal.
465 if (!isGuaranteedToTransferExecutionToSuccessor(&I))
466 break;
467 I.replaceUsesOfWith(Cond, ToVal);
468 }
469 if (Cond->use_empty() && !Cond->mayHaveSideEffects())
470 Cond->eraseFromParent();
471 }
472
473 /// Return the cost of duplicating a piece of this block from first non-phi
474 /// and before StopAt instruction to thread across it. Stop scanning the block
475 /// when exceeding the threshold. If duplication is impossible, returns ~0U.
getJumpThreadDuplicationCost(BasicBlock * BB,Instruction * StopAt,unsigned Threshold)476 static unsigned getJumpThreadDuplicationCost(BasicBlock *BB,
477 Instruction *StopAt,
478 unsigned Threshold) {
479 assert(StopAt->getParent() == BB && "Not an instruction from proper BB?");
480 /// Ignore PHI nodes, these will be flattened when duplication happens.
481 BasicBlock::const_iterator I(BB->getFirstNonPHI());
482
483 // FIXME: THREADING will delete values that are just used to compute the
484 // branch, so they shouldn't count against the duplication cost.
485
486 unsigned Bonus = 0;
487 if (BB->getTerminator() == StopAt) {
488 // Threading through a switch statement is particularly profitable. If this
489 // block ends in a switch, decrease its cost to make it more likely to
490 // happen.
491 if (isa<SwitchInst>(StopAt))
492 Bonus = 6;
493
494 // The same holds for indirect branches, but slightly more so.
495 if (isa<IndirectBrInst>(StopAt))
496 Bonus = 8;
497 }
498
499 // Bump the threshold up so the early exit from the loop doesn't skip the
500 // terminator-based Size adjustment at the end.
501 Threshold += Bonus;
502
503 // Sum up the cost of each instruction until we get to the terminator. Don't
504 // include the terminator because the copy won't include it.
505 unsigned Size = 0;
506 for (; &*I != StopAt; ++I) {
507
508 // Stop scanning the block if we've reached the threshold.
509 if (Size > Threshold)
510 return Size;
511
512 // Debugger intrinsics don't incur code size.
513 if (isa<DbgInfoIntrinsic>(I)) continue;
514
515 // If this is a pointer->pointer bitcast, it is free.
516 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
517 continue;
518
519 // Bail out if this instruction gives back a token type, it is not possible
520 // to duplicate it if it is used outside this BB.
521 if (I->getType()->isTokenTy() && I->isUsedOutsideOfBlock(BB))
522 return ~0U;
523
524 // All other instructions count for at least one unit.
525 ++Size;
526
527 // Calls are more expensive. If they are non-intrinsic calls, we model them
528 // as having cost of 4. If they are a non-vector intrinsic, we model them
529 // as having cost of 2 total, and if they are a vector intrinsic, we model
530 // them as having cost 1.
531 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
532 if (CI->cannotDuplicate() || CI->isConvergent())
533 // Blocks with NoDuplicate are modelled as having infinite cost, so they
534 // are never duplicated.
535 return ~0U;
536 else if (!isa<IntrinsicInst>(CI))
537 Size += 3;
538 else if (!CI->getType()->isVectorTy())
539 Size += 1;
540 }
541 }
542
543 return Size > Bonus ? Size - Bonus : 0;
544 }
545
546 /// FindLoopHeaders - We do not want jump threading to turn proper loop
547 /// structures into irreducible loops. Doing this breaks up the loop nesting
548 /// hierarchy and pessimizes later transformations. To prevent this from
549 /// happening, we first have to find the loop headers. Here we approximate this
550 /// by finding targets of backedges in the CFG.
551 ///
552 /// Note that there definitely are cases when we want to allow threading of
553 /// edges across a loop header. For example, threading a jump from outside the
554 /// loop (the preheader) to an exit block of the loop is definitely profitable.
555 /// It is also almost always profitable to thread backedges from within the loop
556 /// to exit blocks, and is often profitable to thread backedges to other blocks
557 /// within the loop (forming a nested loop). This simple analysis is not rich
558 /// enough to track all of these properties and keep it up-to-date as the CFG
559 /// mutates, so we don't allow any of these transformations.
FindLoopHeaders(Function & F)560 void JumpThreadingPass::FindLoopHeaders(Function &F) {
561 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
562 FindFunctionBackedges(F, Edges);
563
564 for (const auto &Edge : Edges)
565 LoopHeaders.insert(Edge.second);
566 }
567
568 /// getKnownConstant - Helper method to determine if we can thread over a
569 /// terminator with the given value as its condition, and if so what value to
570 /// use for that. What kind of value this is depends on whether we want an
571 /// integer or a block address, but an undef is always accepted.
572 /// Returns null if Val is null or not an appropriate constant.
getKnownConstant(Value * Val,ConstantPreference Preference)573 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
574 if (!Val)
575 return nullptr;
576
577 // Undef is "known" enough.
578 if (UndefValue *U = dyn_cast<UndefValue>(Val))
579 return U;
580
581 if (Preference == WantBlockAddress)
582 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
583
584 return dyn_cast<ConstantInt>(Val);
585 }
586
587 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
588 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
589 /// in any of our predecessors. If so, return the known list of value and pred
590 /// BB in the result vector.
591 ///
592 /// This returns true if there were any known values.
ComputeValueKnownInPredecessorsImpl(Value * V,BasicBlock * BB,PredValueInfo & Result,ConstantPreference Preference,DenseSet<std::pair<Value *,BasicBlock * >> & RecursionSet,Instruction * CxtI)593 bool JumpThreadingPass::ComputeValueKnownInPredecessorsImpl(
594 Value *V, BasicBlock *BB, PredValueInfo &Result,
595 ConstantPreference Preference,
596 DenseSet<std::pair<Value *, BasicBlock *>> &RecursionSet,
597 Instruction *CxtI) {
598 // This method walks up use-def chains recursively. Because of this, we could
599 // get into an infinite loop going around loops in the use-def chain. To
600 // prevent this, keep track of what (value, block) pairs we've already visited
601 // and terminate the search if we loop back to them
602 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
603 return false;
604
605 // If V is a constant, then it is known in all predecessors.
606 if (Constant *KC = getKnownConstant(V, Preference)) {
607 for (BasicBlock *Pred : predecessors(BB))
608 Result.push_back(std::make_pair(KC, Pred));
609
610 return !Result.empty();
611 }
612
613 // If V is a non-instruction value, or an instruction in a different block,
614 // then it can't be derived from a PHI.
615 Instruction *I = dyn_cast<Instruction>(V);
616 if (!I || I->getParent() != BB) {
617
618 // Okay, if this is a live-in value, see if it has a known value at the end
619 // of any of our predecessors.
620 //
621 // FIXME: This should be an edge property, not a block end property.
622 /// TODO: Per PR2563, we could infer value range information about a
623 /// predecessor based on its terminator.
624 //
625 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
626 // "I" is a non-local compare-with-a-constant instruction. This would be
627 // able to handle value inequalities better, for example if the compare is
628 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
629 // Perhaps getConstantOnEdge should be smart enough to do this?
630
631 if (DTU->hasPendingDomTreeUpdates())
632 LVI->disableDT();
633 else
634 LVI->enableDT();
635 for (BasicBlock *P : predecessors(BB)) {
636 // If the value is known by LazyValueInfo to be a constant in a
637 // predecessor, use that information to try to thread this block.
638 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB, CxtI);
639 if (Constant *KC = getKnownConstant(PredCst, Preference))
640 Result.push_back(std::make_pair(KC, P));
641 }
642
643 return !Result.empty();
644 }
645
646 /// If I is a PHI node, then we know the incoming values for any constants.
647 if (PHINode *PN = dyn_cast<PHINode>(I)) {
648 if (DTU->hasPendingDomTreeUpdates())
649 LVI->disableDT();
650 else
651 LVI->enableDT();
652 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
653 Value *InVal = PN->getIncomingValue(i);
654 if (Constant *KC = getKnownConstant(InVal, Preference)) {
655 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
656 } else {
657 Constant *CI = LVI->getConstantOnEdge(InVal,
658 PN->getIncomingBlock(i),
659 BB, CxtI);
660 if (Constant *KC = getKnownConstant(CI, Preference))
661 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
662 }
663 }
664
665 return !Result.empty();
666 }
667
668 // Handle Cast instructions. Only see through Cast when the source operand is
669 // PHI or Cmp to save the compilation time.
670 if (CastInst *CI = dyn_cast<CastInst>(I)) {
671 Value *Source = CI->getOperand(0);
672 if (!isa<PHINode>(Source) && !isa<CmpInst>(Source))
673 return false;
674 ComputeValueKnownInPredecessorsImpl(Source, BB, Result, Preference,
675 RecursionSet, CxtI);
676 if (Result.empty())
677 return false;
678
679 // Convert the known values.
680 for (auto &R : Result)
681 R.first = ConstantExpr::getCast(CI->getOpcode(), R.first, CI->getType());
682
683 return true;
684 }
685
686 // Handle some boolean conditions.
687 if (I->getType()->getPrimitiveSizeInBits() == 1) {
688 assert(Preference == WantInteger && "One-bit non-integer type?");
689 // X | true -> true
690 // X & false -> false
691 if (I->getOpcode() == Instruction::Or ||
692 I->getOpcode() == Instruction::And) {
693 PredValueInfoTy LHSVals, RHSVals;
694
695 ComputeValueKnownInPredecessorsImpl(I->getOperand(0), BB, LHSVals,
696 WantInteger, RecursionSet, CxtI);
697 ComputeValueKnownInPredecessorsImpl(I->getOperand(1), BB, RHSVals,
698 WantInteger, RecursionSet, CxtI);
699
700 if (LHSVals.empty() && RHSVals.empty())
701 return false;
702
703 ConstantInt *InterestingVal;
704 if (I->getOpcode() == Instruction::Or)
705 InterestingVal = ConstantInt::getTrue(I->getContext());
706 else
707 InterestingVal = ConstantInt::getFalse(I->getContext());
708
709 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
710
711 // Scan for the sentinel. If we find an undef, force it to the
712 // interesting value: x|undef -> true and x&undef -> false.
713 for (const auto &LHSVal : LHSVals)
714 if (LHSVal.first == InterestingVal || isa<UndefValue>(LHSVal.first)) {
715 Result.emplace_back(InterestingVal, LHSVal.second);
716 LHSKnownBBs.insert(LHSVal.second);
717 }
718 for (const auto &RHSVal : RHSVals)
719 if (RHSVal.first == InterestingVal || isa<UndefValue>(RHSVal.first)) {
720 // If we already inferred a value for this block on the LHS, don't
721 // re-add it.
722 if (!LHSKnownBBs.count(RHSVal.second))
723 Result.emplace_back(InterestingVal, RHSVal.second);
724 }
725
726 return !Result.empty();
727 }
728
729 // Handle the NOT form of XOR.
730 if (I->getOpcode() == Instruction::Xor &&
731 isa<ConstantInt>(I->getOperand(1)) &&
732 cast<ConstantInt>(I->getOperand(1))->isOne()) {
733 ComputeValueKnownInPredecessorsImpl(I->getOperand(0), BB, Result,
734 WantInteger, RecursionSet, CxtI);
735 if (Result.empty())
736 return false;
737
738 // Invert the known values.
739 for (auto &R : Result)
740 R.first = ConstantExpr::getNot(R.first);
741
742 return true;
743 }
744
745 // Try to simplify some other binary operator values.
746 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
747 assert(Preference != WantBlockAddress
748 && "A binary operator creating a block address?");
749 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
750 PredValueInfoTy LHSVals;
751 ComputeValueKnownInPredecessorsImpl(BO->getOperand(0), BB, LHSVals,
752 WantInteger, RecursionSet, CxtI);
753
754 // Try to use constant folding to simplify the binary operator.
755 for (const auto &LHSVal : LHSVals) {
756 Constant *V = LHSVal.first;
757 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
758
759 if (Constant *KC = getKnownConstant(Folded, WantInteger))
760 Result.push_back(std::make_pair(KC, LHSVal.second));
761 }
762 }
763
764 return !Result.empty();
765 }
766
767 // Handle compare with phi operand, where the PHI is defined in this block.
768 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
769 assert(Preference == WantInteger && "Compares only produce integers");
770 Type *CmpType = Cmp->getType();
771 Value *CmpLHS = Cmp->getOperand(0);
772 Value *CmpRHS = Cmp->getOperand(1);
773 CmpInst::Predicate Pred = Cmp->getPredicate();
774
775 PHINode *PN = dyn_cast<PHINode>(CmpLHS);
776 if (!PN)
777 PN = dyn_cast<PHINode>(CmpRHS);
778 if (PN && PN->getParent() == BB) {
779 const DataLayout &DL = PN->getModule()->getDataLayout();
780 // We can do this simplification if any comparisons fold to true or false.
781 // See if any do.
782 if (DTU->hasPendingDomTreeUpdates())
783 LVI->disableDT();
784 else
785 LVI->enableDT();
786 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
787 BasicBlock *PredBB = PN->getIncomingBlock(i);
788 Value *LHS, *RHS;
789 if (PN == CmpLHS) {
790 LHS = PN->getIncomingValue(i);
791 RHS = CmpRHS->DoPHITranslation(BB, PredBB);
792 } else {
793 LHS = CmpLHS->DoPHITranslation(BB, PredBB);
794 RHS = PN->getIncomingValue(i);
795 }
796 Value *Res = SimplifyCmpInst(Pred, LHS, RHS, {DL});
797 if (!Res) {
798 if (!isa<Constant>(RHS))
799 continue;
800
801 // getPredicateOnEdge call will make no sense if LHS is defined in BB.
802 auto LHSInst = dyn_cast<Instruction>(LHS);
803 if (LHSInst && LHSInst->getParent() == BB)
804 continue;
805
806 LazyValueInfo::Tristate
807 ResT = LVI->getPredicateOnEdge(Pred, LHS,
808 cast<Constant>(RHS), PredBB, BB,
809 CxtI ? CxtI : Cmp);
810 if (ResT == LazyValueInfo::Unknown)
811 continue;
812 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
813 }
814
815 if (Constant *KC = getKnownConstant(Res, WantInteger))
816 Result.push_back(std::make_pair(KC, PredBB));
817 }
818
819 return !Result.empty();
820 }
821
822 // If comparing a live-in value against a constant, see if we know the
823 // live-in value on any predecessors.
824 if (isa<Constant>(CmpRHS) && !CmpType->isVectorTy()) {
825 Constant *CmpConst = cast<Constant>(CmpRHS);
826
827 if (!isa<Instruction>(CmpLHS) ||
828 cast<Instruction>(CmpLHS)->getParent() != BB) {
829 if (DTU->hasPendingDomTreeUpdates())
830 LVI->disableDT();
831 else
832 LVI->enableDT();
833 for (BasicBlock *P : predecessors(BB)) {
834 // If the value is known by LazyValueInfo to be a constant in a
835 // predecessor, use that information to try to thread this block.
836 LazyValueInfo::Tristate Res =
837 LVI->getPredicateOnEdge(Pred, CmpLHS,
838 CmpConst, P, BB, CxtI ? CxtI : Cmp);
839 if (Res == LazyValueInfo::Unknown)
840 continue;
841
842 Constant *ResC = ConstantInt::get(CmpType, Res);
843 Result.push_back(std::make_pair(ResC, P));
844 }
845
846 return !Result.empty();
847 }
848
849 // InstCombine can fold some forms of constant range checks into
850 // (icmp (add (x, C1)), C2). See if we have we have such a thing with
851 // x as a live-in.
852 {
853 using namespace PatternMatch;
854
855 Value *AddLHS;
856 ConstantInt *AddConst;
857 if (isa<ConstantInt>(CmpConst) &&
858 match(CmpLHS, m_Add(m_Value(AddLHS), m_ConstantInt(AddConst)))) {
859 if (!isa<Instruction>(AddLHS) ||
860 cast<Instruction>(AddLHS)->getParent() != BB) {
861 if (DTU->hasPendingDomTreeUpdates())
862 LVI->disableDT();
863 else
864 LVI->enableDT();
865 for (BasicBlock *P : predecessors(BB)) {
866 // If the value is known by LazyValueInfo to be a ConstantRange in
867 // a predecessor, use that information to try to thread this
868 // block.
869 ConstantRange CR = LVI->getConstantRangeOnEdge(
870 AddLHS, P, BB, CxtI ? CxtI : cast<Instruction>(CmpLHS));
871 // Propagate the range through the addition.
872 CR = CR.add(AddConst->getValue());
873
874 // Get the range where the compare returns true.
875 ConstantRange CmpRange = ConstantRange::makeExactICmpRegion(
876 Pred, cast<ConstantInt>(CmpConst)->getValue());
877
878 Constant *ResC;
879 if (CmpRange.contains(CR))
880 ResC = ConstantInt::getTrue(CmpType);
881 else if (CmpRange.inverse().contains(CR))
882 ResC = ConstantInt::getFalse(CmpType);
883 else
884 continue;
885
886 Result.push_back(std::make_pair(ResC, P));
887 }
888
889 return !Result.empty();
890 }
891 }
892 }
893
894 // Try to find a constant value for the LHS of a comparison,
895 // and evaluate it statically if we can.
896 PredValueInfoTy LHSVals;
897 ComputeValueKnownInPredecessorsImpl(I->getOperand(0), BB, LHSVals,
898 WantInteger, RecursionSet, CxtI);
899
900 for (const auto &LHSVal : LHSVals) {
901 Constant *V = LHSVal.first;
902 Constant *Folded = ConstantExpr::getCompare(Pred, V, CmpConst);
903 if (Constant *KC = getKnownConstant(Folded, WantInteger))
904 Result.push_back(std::make_pair(KC, LHSVal.second));
905 }
906
907 return !Result.empty();
908 }
909 }
910
911 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
912 // Handle select instructions where at least one operand is a known constant
913 // and we can figure out the condition value for any predecessor block.
914 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
915 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
916 PredValueInfoTy Conds;
917 if ((TrueVal || FalseVal) &&
918 ComputeValueKnownInPredecessorsImpl(SI->getCondition(), BB, Conds,
919 WantInteger, RecursionSet, CxtI)) {
920 for (auto &C : Conds) {
921 Constant *Cond = C.first;
922
923 // Figure out what value to use for the condition.
924 bool KnownCond;
925 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
926 // A known boolean.
927 KnownCond = CI->isOne();
928 } else {
929 assert(isa<UndefValue>(Cond) && "Unexpected condition value");
930 // Either operand will do, so be sure to pick the one that's a known
931 // constant.
932 // FIXME: Do this more cleverly if both values are known constants?
933 KnownCond = (TrueVal != nullptr);
934 }
935
936 // See if the select has a known constant value for this predecessor.
937 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
938 Result.push_back(std::make_pair(Val, C.second));
939 }
940
941 return !Result.empty();
942 }
943 }
944
945 // If all else fails, see if LVI can figure out a constant value for us.
946 if (DTU->hasPendingDomTreeUpdates())
947 LVI->disableDT();
948 else
949 LVI->enableDT();
950 Constant *CI = LVI->getConstant(V, BB, CxtI);
951 if (Constant *KC = getKnownConstant(CI, Preference)) {
952 for (BasicBlock *Pred : predecessors(BB))
953 Result.push_back(std::make_pair(KC, Pred));
954 }
955
956 return !Result.empty();
957 }
958
959 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
960 /// in an undefined jump, decide which block is best to revector to.
961 ///
962 /// Since we can pick an arbitrary destination, we pick the successor with the
963 /// fewest predecessors. This should reduce the in-degree of the others.
GetBestDestForJumpOnUndef(BasicBlock * BB)964 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
965 Instruction *BBTerm = BB->getTerminator();
966 unsigned MinSucc = 0;
967 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
968 // Compute the successor with the minimum number of predecessors.
969 unsigned MinNumPreds = pred_size(TestBB);
970 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
971 TestBB = BBTerm->getSuccessor(i);
972 unsigned NumPreds = pred_size(TestBB);
973 if (NumPreds < MinNumPreds) {
974 MinSucc = i;
975 MinNumPreds = NumPreds;
976 }
977 }
978
979 return MinSucc;
980 }
981
hasAddressTakenAndUsed(BasicBlock * BB)982 static bool hasAddressTakenAndUsed(BasicBlock *BB) {
983 if (!BB->hasAddressTaken()) return false;
984
985 // If the block has its address taken, it may be a tree of dead constants
986 // hanging off of it. These shouldn't keep the block alive.
987 BlockAddress *BA = BlockAddress::get(BB);
988 BA->removeDeadConstantUsers();
989 return !BA->use_empty();
990 }
991
992 /// ProcessBlock - If there are any predecessors whose control can be threaded
993 /// through to a successor, transform them now.
ProcessBlock(BasicBlock * BB)994 bool JumpThreadingPass::ProcessBlock(BasicBlock *BB) {
995 // If the block is trivially dead, just return and let the caller nuke it.
996 // This simplifies other transformations.
997 if (DTU->isBBPendingDeletion(BB) ||
998 (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()))
999 return false;
1000
1001 // If this block has a single predecessor, and if that pred has a single
1002 // successor, merge the blocks. This encourages recursive jump threading
1003 // because now the condition in this block can be threaded through
1004 // predecessors of our predecessor block.
1005 if (MaybeMergeBasicBlockIntoOnlyPred(BB))
1006 return true;
1007
1008 if (TryToUnfoldSelectInCurrBB(BB))
1009 return true;
1010
1011 // Look if we can propagate guards to predecessors.
1012 if (HasGuards && ProcessGuards(BB))
1013 return true;
1014
1015 // What kind of constant we're looking for.
1016 ConstantPreference Preference = WantInteger;
1017
1018 // Look to see if the terminator is a conditional branch, switch or indirect
1019 // branch, if not we can't thread it.
1020 Value *Condition;
1021 Instruction *Terminator = BB->getTerminator();
1022 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
1023 // Can't thread an unconditional jump.
1024 if (BI->isUnconditional()) return false;
1025 Condition = BI->getCondition();
1026 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
1027 Condition = SI->getCondition();
1028 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
1029 // Can't thread indirect branch with no successors.
1030 if (IB->getNumSuccessors() == 0) return false;
1031 Condition = IB->getAddress()->stripPointerCasts();
1032 Preference = WantBlockAddress;
1033 } else {
1034 return false; // Must be an invoke or callbr.
1035 }
1036
1037 // Run constant folding to see if we can reduce the condition to a simple
1038 // constant.
1039 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
1040 Value *SimpleVal =
1041 ConstantFoldInstruction(I, BB->getModule()->getDataLayout(), TLI);
1042 if (SimpleVal) {
1043 I->replaceAllUsesWith(SimpleVal);
1044 if (isInstructionTriviallyDead(I, TLI))
1045 I->eraseFromParent();
1046 Condition = SimpleVal;
1047 }
1048 }
1049
1050 // If the terminator is branching on an undef, we can pick any of the
1051 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
1052 if (isa<UndefValue>(Condition)) {
1053 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
1054 std::vector<DominatorTree::UpdateType> Updates;
1055
1056 // Fold the branch/switch.
1057 Instruction *BBTerm = BB->getTerminator();
1058 Updates.reserve(BBTerm->getNumSuccessors());
1059 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
1060 if (i == BestSucc) continue;
1061 BasicBlock *Succ = BBTerm->getSuccessor(i);
1062 Succ->removePredecessor(BB, true);
1063 Updates.push_back({DominatorTree::Delete, BB, Succ});
1064 }
1065
1066 LLVM_DEBUG(dbgs() << " In block '" << BB->getName()
1067 << "' folding undef terminator: " << *BBTerm << '\n');
1068 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
1069 BBTerm->eraseFromParent();
1070 DTU->applyUpdatesPermissive(Updates);
1071 return true;
1072 }
1073
1074 // If the terminator of this block is branching on a constant, simplify the
1075 // terminator to an unconditional branch. This can occur due to threading in
1076 // other blocks.
1077 if (getKnownConstant(Condition, Preference)) {
1078 LLVM_DEBUG(dbgs() << " In block '" << BB->getName()
1079 << "' folding terminator: " << *BB->getTerminator()
1080 << '\n');
1081 ++NumFolds;
1082 ConstantFoldTerminator(BB, true, nullptr, DTU);
1083 return true;
1084 }
1085
1086 Instruction *CondInst = dyn_cast<Instruction>(Condition);
1087
1088 // All the rest of our checks depend on the condition being an instruction.
1089 if (!CondInst) {
1090 // FIXME: Unify this with code below.
1091 if (ProcessThreadableEdges(Condition, BB, Preference, Terminator))
1092 return true;
1093 return false;
1094 }
1095
1096 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
1097 // If we're branching on a conditional, LVI might be able to determine
1098 // it's value at the branch instruction. We only handle comparisons
1099 // against a constant at this time.
1100 // TODO: This should be extended to handle switches as well.
1101 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1102 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
1103 if (CondBr && CondConst) {
1104 // We should have returned as soon as we turn a conditional branch to
1105 // unconditional. Because its no longer interesting as far as jump
1106 // threading is concerned.
1107 assert(CondBr->isConditional() && "Threading on unconditional terminator");
1108
1109 if (DTU->hasPendingDomTreeUpdates())
1110 LVI->disableDT();
1111 else
1112 LVI->enableDT();
1113 LazyValueInfo::Tristate Ret =
1114 LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0),
1115 CondConst, CondBr);
1116 if (Ret != LazyValueInfo::Unknown) {
1117 unsigned ToRemove = Ret == LazyValueInfo::True ? 1 : 0;
1118 unsigned ToKeep = Ret == LazyValueInfo::True ? 0 : 1;
1119 BasicBlock *ToRemoveSucc = CondBr->getSuccessor(ToRemove);
1120 ToRemoveSucc->removePredecessor(BB, true);
1121 BranchInst *UncondBr =
1122 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
1123 UncondBr->setDebugLoc(CondBr->getDebugLoc());
1124 CondBr->eraseFromParent();
1125 if (CondCmp->use_empty())
1126 CondCmp->eraseFromParent();
1127 // We can safely replace *some* uses of the CondInst if it has
1128 // exactly one value as returned by LVI. RAUW is incorrect in the
1129 // presence of guards and assumes, that have the `Cond` as the use. This
1130 // is because we use the guards/assume to reason about the `Cond` value
1131 // at the end of block, but RAUW unconditionally replaces all uses
1132 // including the guards/assumes themselves and the uses before the
1133 // guard/assume.
1134 else if (CondCmp->getParent() == BB) {
1135 auto *CI = Ret == LazyValueInfo::True ?
1136 ConstantInt::getTrue(CondCmp->getType()) :
1137 ConstantInt::getFalse(CondCmp->getType());
1138 ReplaceFoldableUses(CondCmp, CI);
1139 }
1140 DTU->applyUpdatesPermissive(
1141 {{DominatorTree::Delete, BB, ToRemoveSucc}});
1142 return true;
1143 }
1144
1145 // We did not manage to simplify this branch, try to see whether
1146 // CondCmp depends on a known phi-select pattern.
1147 if (TryToUnfoldSelect(CondCmp, BB))
1148 return true;
1149 }
1150 }
1151
1152 if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
1153 if (TryToUnfoldSelect(SI, BB))
1154 return true;
1155
1156 // Check for some cases that are worth simplifying. Right now we want to look
1157 // for loads that are used by a switch or by the condition for the branch. If
1158 // we see one, check to see if it's partially redundant. If so, insert a PHI
1159 // which can then be used to thread the values.
1160 Value *SimplifyValue = CondInst;
1161 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
1162 if (isa<Constant>(CondCmp->getOperand(1)))
1163 SimplifyValue = CondCmp->getOperand(0);
1164
1165 // TODO: There are other places where load PRE would be profitable, such as
1166 // more complex comparisons.
1167 if (LoadInst *LoadI = dyn_cast<LoadInst>(SimplifyValue))
1168 if (SimplifyPartiallyRedundantLoad(LoadI))
1169 return true;
1170
1171 // Before threading, try to propagate profile data backwards:
1172 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
1173 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1174 updatePredecessorProfileMetadata(PN, BB);
1175
1176 // Handle a variety of cases where we are branching on something derived from
1177 // a PHI node in the current block. If we can prove that any predecessors
1178 // compute a predictable value based on a PHI node, thread those predecessors.
1179 if (ProcessThreadableEdges(CondInst, BB, Preference, Terminator))
1180 return true;
1181
1182 // If this is an otherwise-unfoldable branch on a phi node in the current
1183 // block, see if we can simplify.
1184 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
1185 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1186 return ProcessBranchOnPHI(PN);
1187
1188 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
1189 if (CondInst->getOpcode() == Instruction::Xor &&
1190 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1191 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
1192
1193 // Search for a stronger dominating condition that can be used to simplify a
1194 // conditional branch leaving BB.
1195 if (ProcessImpliedCondition(BB))
1196 return true;
1197
1198 return false;
1199 }
1200
ProcessImpliedCondition(BasicBlock * BB)1201 bool JumpThreadingPass::ProcessImpliedCondition(BasicBlock *BB) {
1202 auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
1203 if (!BI || !BI->isConditional())
1204 return false;
1205
1206 Value *Cond = BI->getCondition();
1207 BasicBlock *CurrentBB = BB;
1208 BasicBlock *CurrentPred = BB->getSinglePredecessor();
1209 unsigned Iter = 0;
1210
1211 auto &DL = BB->getModule()->getDataLayout();
1212
1213 while (CurrentPred && Iter++ < ImplicationSearchThreshold) {
1214 auto *PBI = dyn_cast<BranchInst>(CurrentPred->getTerminator());
1215 if (!PBI || !PBI->isConditional())
1216 return false;
1217 if (PBI->getSuccessor(0) != CurrentBB && PBI->getSuccessor(1) != CurrentBB)
1218 return false;
1219
1220 bool CondIsTrue = PBI->getSuccessor(0) == CurrentBB;
1221 Optional<bool> Implication =
1222 isImpliedCondition(PBI->getCondition(), Cond, DL, CondIsTrue);
1223 if (Implication) {
1224 BasicBlock *KeepSucc = BI->getSuccessor(*Implication ? 0 : 1);
1225 BasicBlock *RemoveSucc = BI->getSuccessor(*Implication ? 1 : 0);
1226 RemoveSucc->removePredecessor(BB);
1227 BranchInst *UncondBI = BranchInst::Create(KeepSucc, BI);
1228 UncondBI->setDebugLoc(BI->getDebugLoc());
1229 BI->eraseFromParent();
1230 DTU->applyUpdatesPermissive({{DominatorTree::Delete, BB, RemoveSucc}});
1231 return true;
1232 }
1233 CurrentBB = CurrentPred;
1234 CurrentPred = CurrentBB->getSinglePredecessor();
1235 }
1236
1237 return false;
1238 }
1239
1240 /// Return true if Op is an instruction defined in the given block.
isOpDefinedInBlock(Value * Op,BasicBlock * BB)1241 static bool isOpDefinedInBlock(Value *Op, BasicBlock *BB) {
1242 if (Instruction *OpInst = dyn_cast<Instruction>(Op))
1243 if (OpInst->getParent() == BB)
1244 return true;
1245 return false;
1246 }
1247
1248 /// SimplifyPartiallyRedundantLoad - If LoadI is an obviously partially
1249 /// redundant load instruction, eliminate it by replacing it with a PHI node.
1250 /// This is an important optimization that encourages jump threading, and needs
1251 /// to be run interlaced with other jump threading tasks.
SimplifyPartiallyRedundantLoad(LoadInst * LoadI)1252 bool JumpThreadingPass::SimplifyPartiallyRedundantLoad(LoadInst *LoadI) {
1253 // Don't hack volatile and ordered loads.
1254 if (!LoadI->isUnordered()) return false;
1255
1256 // If the load is defined in a block with exactly one predecessor, it can't be
1257 // partially redundant.
1258 BasicBlock *LoadBB = LoadI->getParent();
1259 if (LoadBB->getSinglePredecessor())
1260 return false;
1261
1262 // If the load is defined in an EH pad, it can't be partially redundant,
1263 // because the edges between the invoke and the EH pad cannot have other
1264 // instructions between them.
1265 if (LoadBB->isEHPad())
1266 return false;
1267
1268 Value *LoadedPtr = LoadI->getOperand(0);
1269
1270 // If the loaded operand is defined in the LoadBB and its not a phi,
1271 // it can't be available in predecessors.
1272 if (isOpDefinedInBlock(LoadedPtr, LoadBB) && !isa<PHINode>(LoadedPtr))
1273 return false;
1274
1275 // Scan a few instructions up from the load, to see if it is obviously live at
1276 // the entry to its block.
1277 BasicBlock::iterator BBIt(LoadI);
1278 bool IsLoadCSE;
1279 if (Value *AvailableVal = FindAvailableLoadedValue(
1280 LoadI, LoadBB, BBIt, DefMaxInstsToScan, AA, &IsLoadCSE)) {
1281 // If the value of the load is locally available within the block, just use
1282 // it. This frequently occurs for reg2mem'd allocas.
1283
1284 if (IsLoadCSE) {
1285 LoadInst *NLoadI = cast<LoadInst>(AvailableVal);
1286 combineMetadataForCSE(NLoadI, LoadI, false);
1287 };
1288
1289 // If the returned value is the load itself, replace with an undef. This can
1290 // only happen in dead loops.
1291 if (AvailableVal == LoadI)
1292 AvailableVal = UndefValue::get(LoadI->getType());
1293 if (AvailableVal->getType() != LoadI->getType())
1294 AvailableVal = CastInst::CreateBitOrPointerCast(
1295 AvailableVal, LoadI->getType(), "", LoadI);
1296 LoadI->replaceAllUsesWith(AvailableVal);
1297 LoadI->eraseFromParent();
1298 return true;
1299 }
1300
1301 // Otherwise, if we scanned the whole block and got to the top of the block,
1302 // we know the block is locally transparent to the load. If not, something
1303 // might clobber its value.
1304 if (BBIt != LoadBB->begin())
1305 return false;
1306
1307 // If all of the loads and stores that feed the value have the same AA tags,
1308 // then we can propagate them onto any newly inserted loads.
1309 AAMDNodes AATags;
1310 LoadI->getAAMetadata(AATags);
1311
1312 SmallPtrSet<BasicBlock*, 8> PredsScanned;
1313
1314 using AvailablePredsTy = SmallVector<std::pair<BasicBlock *, Value *>, 8>;
1315
1316 AvailablePredsTy AvailablePreds;
1317 BasicBlock *OneUnavailablePred = nullptr;
1318 SmallVector<LoadInst*, 8> CSELoads;
1319
1320 // If we got here, the loaded value is transparent through to the start of the
1321 // block. Check to see if it is available in any of the predecessor blocks.
1322 for (BasicBlock *PredBB : predecessors(LoadBB)) {
1323 // If we already scanned this predecessor, skip it.
1324 if (!PredsScanned.insert(PredBB).second)
1325 continue;
1326
1327 BBIt = PredBB->end();
1328 unsigned NumScanedInst = 0;
1329 Value *PredAvailable = nullptr;
1330 // NOTE: We don't CSE load that is volatile or anything stronger than
1331 // unordered, that should have been checked when we entered the function.
1332 assert(LoadI->isUnordered() &&
1333 "Attempting to CSE volatile or atomic loads");
1334 // If this is a load on a phi pointer, phi-translate it and search
1335 // for available load/store to the pointer in predecessors.
1336 Value *Ptr = LoadedPtr->DoPHITranslation(LoadBB, PredBB);
1337 PredAvailable = FindAvailablePtrLoadStore(
1338 Ptr, LoadI->getType(), LoadI->isAtomic(), PredBB, BBIt,
1339 DefMaxInstsToScan, AA, &IsLoadCSE, &NumScanedInst);
1340
1341 // If PredBB has a single predecessor, continue scanning through the
1342 // single predecessor.
1343 BasicBlock *SinglePredBB = PredBB;
1344 while (!PredAvailable && SinglePredBB && BBIt == SinglePredBB->begin() &&
1345 NumScanedInst < DefMaxInstsToScan) {
1346 SinglePredBB = SinglePredBB->getSinglePredecessor();
1347 if (SinglePredBB) {
1348 BBIt = SinglePredBB->end();
1349 PredAvailable = FindAvailablePtrLoadStore(
1350 Ptr, LoadI->getType(), LoadI->isAtomic(), SinglePredBB, BBIt,
1351 (DefMaxInstsToScan - NumScanedInst), AA, &IsLoadCSE,
1352 &NumScanedInst);
1353 }
1354 }
1355
1356 if (!PredAvailable) {
1357 OneUnavailablePred = PredBB;
1358 continue;
1359 }
1360
1361 if (IsLoadCSE)
1362 CSELoads.push_back(cast<LoadInst>(PredAvailable));
1363
1364 // If so, this load is partially redundant. Remember this info so that we
1365 // can create a PHI node.
1366 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
1367 }
1368
1369 // If the loaded value isn't available in any predecessor, it isn't partially
1370 // redundant.
1371 if (AvailablePreds.empty()) return false;
1372
1373 // Okay, the loaded value is available in at least one (and maybe all!)
1374 // predecessors. If the value is unavailable in more than one unique
1375 // predecessor, we want to insert a merge block for those common predecessors.
1376 // This ensures that we only have to insert one reload, thus not increasing
1377 // code size.
1378 BasicBlock *UnavailablePred = nullptr;
1379
1380 // If the value is unavailable in one of predecessors, we will end up
1381 // inserting a new instruction into them. It is only valid if all the
1382 // instructions before LoadI are guaranteed to pass execution to its
1383 // successor, or if LoadI is safe to speculate.
1384 // TODO: If this logic becomes more complex, and we will perform PRE insertion
1385 // farther than to a predecessor, we need to reuse the code from GVN's PRE.
1386 // It requires domination tree analysis, so for this simple case it is an
1387 // overkill.
1388 if (PredsScanned.size() != AvailablePreds.size() &&
1389 !isSafeToSpeculativelyExecute(LoadI))
1390 for (auto I = LoadBB->begin(); &*I != LoadI; ++I)
1391 if (!isGuaranteedToTransferExecutionToSuccessor(&*I))
1392 return false;
1393
1394 // If there is exactly one predecessor where the value is unavailable, the
1395 // already computed 'OneUnavailablePred' block is it. If it ends in an
1396 // unconditional branch, we know that it isn't a critical edge.
1397 if (PredsScanned.size() == AvailablePreds.size()+1 &&
1398 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
1399 UnavailablePred = OneUnavailablePred;
1400 } else if (PredsScanned.size() != AvailablePreds.size()) {
1401 // Otherwise, we had multiple unavailable predecessors or we had a critical
1402 // edge from the one.
1403 SmallVector<BasicBlock*, 8> PredsToSplit;
1404 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
1405
1406 for (const auto &AvailablePred : AvailablePreds)
1407 AvailablePredSet.insert(AvailablePred.first);
1408
1409 // Add all the unavailable predecessors to the PredsToSplit list.
1410 for (BasicBlock *P : predecessors(LoadBB)) {
1411 // If the predecessor is an indirect goto, we can't split the edge.
1412 // Same for CallBr.
1413 if (isa<IndirectBrInst>(P->getTerminator()) ||
1414 isa<CallBrInst>(P->getTerminator()))
1415 return false;
1416
1417 if (!AvailablePredSet.count(P))
1418 PredsToSplit.push_back(P);
1419 }
1420
1421 // Split them out to their own block.
1422 UnavailablePred = SplitBlockPreds(LoadBB, PredsToSplit, "thread-pre-split");
1423 }
1424
1425 // If the value isn't available in all predecessors, then there will be
1426 // exactly one where it isn't available. Insert a load on that edge and add
1427 // it to the AvailablePreds list.
1428 if (UnavailablePred) {
1429 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
1430 "Can't handle critical edge here!");
1431 LoadInst *NewVal = new LoadInst(
1432 LoadI->getType(), LoadedPtr->DoPHITranslation(LoadBB, UnavailablePred),
1433 LoadI->getName() + ".pr", false, MaybeAlign(LoadI->getAlignment()),
1434 LoadI->getOrdering(), LoadI->getSyncScopeID(),
1435 UnavailablePred->getTerminator());
1436 NewVal->setDebugLoc(LoadI->getDebugLoc());
1437 if (AATags)
1438 NewVal->setAAMetadata(AATags);
1439
1440 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
1441 }
1442
1443 // Now we know that each predecessor of this block has a value in
1444 // AvailablePreds, sort them for efficient access as we're walking the preds.
1445 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1446
1447 // Create a PHI node at the start of the block for the PRE'd load value.
1448 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
1449 PHINode *PN = PHINode::Create(LoadI->getType(), std::distance(PB, PE), "",
1450 &LoadBB->front());
1451 PN->takeName(LoadI);
1452 PN->setDebugLoc(LoadI->getDebugLoc());
1453
1454 // Insert new entries into the PHI for each predecessor. A single block may
1455 // have multiple entries here.
1456 for (pred_iterator PI = PB; PI != PE; ++PI) {
1457 BasicBlock *P = *PI;
1458 AvailablePredsTy::iterator I =
1459 llvm::lower_bound(AvailablePreds, std::make_pair(P, (Value *)nullptr));
1460
1461 assert(I != AvailablePreds.end() && I->first == P &&
1462 "Didn't find entry for predecessor!");
1463
1464 // If we have an available predecessor but it requires casting, insert the
1465 // cast in the predecessor and use the cast. Note that we have to update the
1466 // AvailablePreds vector as we go so that all of the PHI entries for this
1467 // predecessor use the same bitcast.
1468 Value *&PredV = I->second;
1469 if (PredV->getType() != LoadI->getType())
1470 PredV = CastInst::CreateBitOrPointerCast(PredV, LoadI->getType(), "",
1471 P->getTerminator());
1472
1473 PN->addIncoming(PredV, I->first);
1474 }
1475
1476 for (LoadInst *PredLoadI : CSELoads) {
1477 combineMetadataForCSE(PredLoadI, LoadI, true);
1478 }
1479
1480 LoadI->replaceAllUsesWith(PN);
1481 LoadI->eraseFromParent();
1482
1483 return true;
1484 }
1485
1486 /// FindMostPopularDest - The specified list contains multiple possible
1487 /// threadable destinations. Pick the one that occurs the most frequently in
1488 /// the list.
1489 static BasicBlock *
FindMostPopularDest(BasicBlock * BB,const SmallVectorImpl<std::pair<BasicBlock *,BasicBlock * >> & PredToDestList)1490 FindMostPopularDest(BasicBlock *BB,
1491 const SmallVectorImpl<std::pair<BasicBlock *,
1492 BasicBlock *>> &PredToDestList) {
1493 assert(!PredToDestList.empty());
1494
1495 // Determine popularity. If there are multiple possible destinations, we
1496 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1497 // blocks with known and real destinations to threading undef. We'll handle
1498 // them later if interesting.
1499 DenseMap<BasicBlock*, unsigned> DestPopularity;
1500 for (const auto &PredToDest : PredToDestList)
1501 if (PredToDest.second)
1502 DestPopularity[PredToDest.second]++;
1503
1504 if (DestPopularity.empty())
1505 return nullptr;
1506
1507 // Find the most popular dest.
1508 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1509 BasicBlock *MostPopularDest = DPI->first;
1510 unsigned Popularity = DPI->second;
1511 SmallVector<BasicBlock*, 4> SamePopularity;
1512
1513 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1514 // If the popularity of this entry isn't higher than the popularity we've
1515 // seen so far, ignore it.
1516 if (DPI->second < Popularity)
1517 ; // ignore.
1518 else if (DPI->second == Popularity) {
1519 // If it is the same as what we've seen so far, keep track of it.
1520 SamePopularity.push_back(DPI->first);
1521 } else {
1522 // If it is more popular, remember it.
1523 SamePopularity.clear();
1524 MostPopularDest = DPI->first;
1525 Popularity = DPI->second;
1526 }
1527 }
1528
1529 // Okay, now we know the most popular destination. If there is more than one
1530 // destination, we need to determine one. This is arbitrary, but we need
1531 // to make a deterministic decision. Pick the first one that appears in the
1532 // successor list.
1533 if (!SamePopularity.empty()) {
1534 SamePopularity.push_back(MostPopularDest);
1535 Instruction *TI = BB->getTerminator();
1536 for (unsigned i = 0; ; ++i) {
1537 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1538
1539 if (!is_contained(SamePopularity, TI->getSuccessor(i)))
1540 continue;
1541
1542 MostPopularDest = TI->getSuccessor(i);
1543 break;
1544 }
1545 }
1546
1547 // Okay, we have finally picked the most popular destination.
1548 return MostPopularDest;
1549 }
1550
ProcessThreadableEdges(Value * Cond,BasicBlock * BB,ConstantPreference Preference,Instruction * CxtI)1551 bool JumpThreadingPass::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1552 ConstantPreference Preference,
1553 Instruction *CxtI) {
1554 // If threading this would thread across a loop header, don't even try to
1555 // thread the edge.
1556 if (LoopHeaders.count(BB))
1557 return false;
1558
1559 PredValueInfoTy PredValues;
1560 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference, CxtI))
1561 return false;
1562
1563 assert(!PredValues.empty() &&
1564 "ComputeValueKnownInPredecessors returned true with no values");
1565
1566 LLVM_DEBUG(dbgs() << "IN BB: " << *BB;
1567 for (const auto &PredValue : PredValues) {
1568 dbgs() << " BB '" << BB->getName()
1569 << "': FOUND condition = " << *PredValue.first
1570 << " for pred '" << PredValue.second->getName() << "'.\n";
1571 });
1572
1573 // Decide what we want to thread through. Convert our list of known values to
1574 // a list of known destinations for each pred. This also discards duplicate
1575 // predecessors and keeps track of the undefined inputs (which are represented
1576 // as a null dest in the PredToDestList).
1577 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1578 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1579
1580 BasicBlock *OnlyDest = nullptr;
1581 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1582 Constant *OnlyVal = nullptr;
1583 Constant *MultipleVal = (Constant *)(intptr_t)~0ULL;
1584
1585 for (const auto &PredValue : PredValues) {
1586 BasicBlock *Pred = PredValue.second;
1587 if (!SeenPreds.insert(Pred).second)
1588 continue; // Duplicate predecessor entry.
1589
1590 Constant *Val = PredValue.first;
1591
1592 BasicBlock *DestBB;
1593 if (isa<UndefValue>(Val))
1594 DestBB = nullptr;
1595 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
1596 assert(isa<ConstantInt>(Val) && "Expecting a constant integer");
1597 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1598 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1599 assert(isa<ConstantInt>(Val) && "Expecting a constant integer");
1600 DestBB = SI->findCaseValue(cast<ConstantInt>(Val))->getCaseSuccessor();
1601 } else {
1602 assert(isa<IndirectBrInst>(BB->getTerminator())
1603 && "Unexpected terminator");
1604 assert(isa<BlockAddress>(Val) && "Expecting a constant blockaddress");
1605 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1606 }
1607
1608 // If we have exactly one destination, remember it for efficiency below.
1609 if (PredToDestList.empty()) {
1610 OnlyDest = DestBB;
1611 OnlyVal = Val;
1612 } else {
1613 if (OnlyDest != DestBB)
1614 OnlyDest = MultipleDestSentinel;
1615 // It possible we have same destination, but different value, e.g. default
1616 // case in switchinst.
1617 if (Val != OnlyVal)
1618 OnlyVal = MultipleVal;
1619 }
1620
1621 // If the predecessor ends with an indirect goto, we can't change its
1622 // destination. Same for CallBr.
1623 if (isa<IndirectBrInst>(Pred->getTerminator()) ||
1624 isa<CallBrInst>(Pred->getTerminator()))
1625 continue;
1626
1627 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1628 }
1629
1630 // If all edges were unthreadable, we fail.
1631 if (PredToDestList.empty())
1632 return false;
1633
1634 // If all the predecessors go to a single known successor, we want to fold,
1635 // not thread. By doing so, we do not need to duplicate the current block and
1636 // also miss potential opportunities in case we dont/cant duplicate.
1637 if (OnlyDest && OnlyDest != MultipleDestSentinel) {
1638 if (BB->hasNPredecessors(PredToDestList.size())) {
1639 bool SeenFirstBranchToOnlyDest = false;
1640 std::vector <DominatorTree::UpdateType> Updates;
1641 Updates.reserve(BB->getTerminator()->getNumSuccessors() - 1);
1642 for (BasicBlock *SuccBB : successors(BB)) {
1643 if (SuccBB == OnlyDest && !SeenFirstBranchToOnlyDest) {
1644 SeenFirstBranchToOnlyDest = true; // Don't modify the first branch.
1645 } else {
1646 SuccBB->removePredecessor(BB, true); // This is unreachable successor.
1647 Updates.push_back({DominatorTree::Delete, BB, SuccBB});
1648 }
1649 }
1650
1651 // Finally update the terminator.
1652 Instruction *Term = BB->getTerminator();
1653 BranchInst::Create(OnlyDest, Term);
1654 Term->eraseFromParent();
1655 DTU->applyUpdatesPermissive(Updates);
1656
1657 // If the condition is now dead due to the removal of the old terminator,
1658 // erase it.
1659 if (auto *CondInst = dyn_cast<Instruction>(Cond)) {
1660 if (CondInst->use_empty() && !CondInst->mayHaveSideEffects())
1661 CondInst->eraseFromParent();
1662 // We can safely replace *some* uses of the CondInst if it has
1663 // exactly one value as returned by LVI. RAUW is incorrect in the
1664 // presence of guards and assumes, that have the `Cond` as the use. This
1665 // is because we use the guards/assume to reason about the `Cond` value
1666 // at the end of block, but RAUW unconditionally replaces all uses
1667 // including the guards/assumes themselves and the uses before the
1668 // guard/assume.
1669 else if (OnlyVal && OnlyVal != MultipleVal &&
1670 CondInst->getParent() == BB)
1671 ReplaceFoldableUses(CondInst, OnlyVal);
1672 }
1673 return true;
1674 }
1675 }
1676
1677 // Determine which is the most common successor. If we have many inputs and
1678 // this block is a switch, we want to start by threading the batch that goes
1679 // to the most popular destination first. If we only know about one
1680 // threadable destination (the common case) we can avoid this.
1681 BasicBlock *MostPopularDest = OnlyDest;
1682
1683 if (MostPopularDest == MultipleDestSentinel) {
1684 // Remove any loop headers from the Dest list, ThreadEdge conservatively
1685 // won't process them, but we might have other destination that are eligible
1686 // and we still want to process.
1687 erase_if(PredToDestList,
1688 [&](const std::pair<BasicBlock *, BasicBlock *> &PredToDest) {
1689 return LoopHeaders.count(PredToDest.second) != 0;
1690 });
1691
1692 if (PredToDestList.empty())
1693 return false;
1694
1695 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1696 }
1697
1698 // Now that we know what the most popular destination is, factor all
1699 // predecessors that will jump to it into a single predecessor.
1700 SmallVector<BasicBlock*, 16> PredsToFactor;
1701 for (const auto &PredToDest : PredToDestList)
1702 if (PredToDest.second == MostPopularDest) {
1703 BasicBlock *Pred = PredToDest.first;
1704
1705 // This predecessor may be a switch or something else that has multiple
1706 // edges to the block. Factor each of these edges by listing them
1707 // according to # occurrences in PredsToFactor.
1708 for (BasicBlock *Succ : successors(Pred))
1709 if (Succ == BB)
1710 PredsToFactor.push_back(Pred);
1711 }
1712
1713 // If the threadable edges are branching on an undefined value, we get to pick
1714 // the destination that these predecessors should get to.
1715 if (!MostPopularDest)
1716 MostPopularDest = BB->getTerminator()->
1717 getSuccessor(GetBestDestForJumpOnUndef(BB));
1718
1719 // Ok, try to thread it!
1720 return TryThreadEdge(BB, PredsToFactor, MostPopularDest);
1721 }
1722
1723 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1724 /// a PHI node in the current block. See if there are any simplifications we
1725 /// can do based on inputs to the phi node.
ProcessBranchOnPHI(PHINode * PN)1726 bool JumpThreadingPass::ProcessBranchOnPHI(PHINode *PN) {
1727 BasicBlock *BB = PN->getParent();
1728
1729 // TODO: We could make use of this to do it once for blocks with common PHI
1730 // values.
1731 SmallVector<BasicBlock*, 1> PredBBs;
1732 PredBBs.resize(1);
1733
1734 // If any of the predecessor blocks end in an unconditional branch, we can
1735 // *duplicate* the conditional branch into that block in order to further
1736 // encourage jump threading and to eliminate cases where we have branch on a
1737 // phi of an icmp (branch on icmp is much better).
1738 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1739 BasicBlock *PredBB = PN->getIncomingBlock(i);
1740 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1741 if (PredBr->isUnconditional()) {
1742 PredBBs[0] = PredBB;
1743 // Try to duplicate BB into PredBB.
1744 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1745 return true;
1746 }
1747 }
1748
1749 return false;
1750 }
1751
1752 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1753 /// a xor instruction in the current block. See if there are any
1754 /// simplifications we can do based on inputs to the xor.
ProcessBranchOnXOR(BinaryOperator * BO)1755 bool JumpThreadingPass::ProcessBranchOnXOR(BinaryOperator *BO) {
1756 BasicBlock *BB = BO->getParent();
1757
1758 // If either the LHS or RHS of the xor is a constant, don't do this
1759 // optimization.
1760 if (isa<ConstantInt>(BO->getOperand(0)) ||
1761 isa<ConstantInt>(BO->getOperand(1)))
1762 return false;
1763
1764 // If the first instruction in BB isn't a phi, we won't be able to infer
1765 // anything special about any particular predecessor.
1766 if (!isa<PHINode>(BB->front()))
1767 return false;
1768
1769 // If this BB is a landing pad, we won't be able to split the edge into it.
1770 if (BB->isEHPad())
1771 return false;
1772
1773 // If we have a xor as the branch input to this block, and we know that the
1774 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1775 // the condition into the predecessor and fix that value to true, saving some
1776 // logical ops on that path and encouraging other paths to simplify.
1777 //
1778 // This copies something like this:
1779 //
1780 // BB:
1781 // %X = phi i1 [1], [%X']
1782 // %Y = icmp eq i32 %A, %B
1783 // %Z = xor i1 %X, %Y
1784 // br i1 %Z, ...
1785 //
1786 // Into:
1787 // BB':
1788 // %Y = icmp ne i32 %A, %B
1789 // br i1 %Y, ...
1790
1791 PredValueInfoTy XorOpValues;
1792 bool isLHS = true;
1793 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1794 WantInteger, BO)) {
1795 assert(XorOpValues.empty());
1796 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1797 WantInteger, BO))
1798 return false;
1799 isLHS = false;
1800 }
1801
1802 assert(!XorOpValues.empty() &&
1803 "ComputeValueKnownInPredecessors returned true with no values");
1804
1805 // Scan the information to see which is most popular: true or false. The
1806 // predecessors can be of the set true, false, or undef.
1807 unsigned NumTrue = 0, NumFalse = 0;
1808 for (const auto &XorOpValue : XorOpValues) {
1809 if (isa<UndefValue>(XorOpValue.first))
1810 // Ignore undefs for the count.
1811 continue;
1812 if (cast<ConstantInt>(XorOpValue.first)->isZero())
1813 ++NumFalse;
1814 else
1815 ++NumTrue;
1816 }
1817
1818 // Determine which value to split on, true, false, or undef if neither.
1819 ConstantInt *SplitVal = nullptr;
1820 if (NumTrue > NumFalse)
1821 SplitVal = ConstantInt::getTrue(BB->getContext());
1822 else if (NumTrue != 0 || NumFalse != 0)
1823 SplitVal = ConstantInt::getFalse(BB->getContext());
1824
1825 // Collect all of the blocks that this can be folded into so that we can
1826 // factor this once and clone it once.
1827 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1828 for (const auto &XorOpValue : XorOpValues) {
1829 if (XorOpValue.first != SplitVal && !isa<UndefValue>(XorOpValue.first))
1830 continue;
1831
1832 BlocksToFoldInto.push_back(XorOpValue.second);
1833 }
1834
1835 // If we inferred a value for all of the predecessors, then duplication won't
1836 // help us. However, we can just replace the LHS or RHS with the constant.
1837 if (BlocksToFoldInto.size() ==
1838 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1839 if (!SplitVal) {
1840 // If all preds provide undef, just nuke the xor, because it is undef too.
1841 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1842 BO->eraseFromParent();
1843 } else if (SplitVal->isZero()) {
1844 // If all preds provide 0, replace the xor with the other input.
1845 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1846 BO->eraseFromParent();
1847 } else {
1848 // If all preds provide 1, set the computed value to 1.
1849 BO->setOperand(!isLHS, SplitVal);
1850 }
1851
1852 return true;
1853 }
1854
1855 // Try to duplicate BB into PredBB.
1856 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1857 }
1858
1859 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1860 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1861 /// NewPred using the entries from OldPred (suitably mapped).
AddPHINodeEntriesForMappedBlock(BasicBlock * PHIBB,BasicBlock * OldPred,BasicBlock * NewPred,DenseMap<Instruction *,Value * > & ValueMap)1862 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1863 BasicBlock *OldPred,
1864 BasicBlock *NewPred,
1865 DenseMap<Instruction*, Value*> &ValueMap) {
1866 for (PHINode &PN : PHIBB->phis()) {
1867 // Ok, we have a PHI node. Figure out what the incoming value was for the
1868 // DestBlock.
1869 Value *IV = PN.getIncomingValueForBlock(OldPred);
1870
1871 // Remap the value if necessary.
1872 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1873 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1874 if (I != ValueMap.end())
1875 IV = I->second;
1876 }
1877
1878 PN.addIncoming(IV, NewPred);
1879 }
1880 }
1881
1882 /// Merge basic block BB into its sole predecessor if possible.
MaybeMergeBasicBlockIntoOnlyPred(BasicBlock * BB)1883 bool JumpThreadingPass::MaybeMergeBasicBlockIntoOnlyPred(BasicBlock *BB) {
1884 BasicBlock *SinglePred = BB->getSinglePredecessor();
1885 if (!SinglePred)
1886 return false;
1887
1888 const Instruction *TI = SinglePred->getTerminator();
1889 if (TI->isExceptionalTerminator() || TI->getNumSuccessors() != 1 ||
1890 SinglePred == BB || hasAddressTakenAndUsed(BB))
1891 return false;
1892
1893 // If SinglePred was a loop header, BB becomes one.
1894 if (LoopHeaders.erase(SinglePred))
1895 LoopHeaders.insert(BB);
1896
1897 LVI->eraseBlock(SinglePred);
1898 MergeBasicBlockIntoOnlyPred(BB, DTU);
1899
1900 // Now that BB is merged into SinglePred (i.e. SinglePred code followed by
1901 // BB code within one basic block `BB`), we need to invalidate the LVI
1902 // information associated with BB, because the LVI information need not be
1903 // true for all of BB after the merge. For example,
1904 // Before the merge, LVI info and code is as follows:
1905 // SinglePred: <LVI info1 for %p val>
1906 // %y = use of %p
1907 // call @exit() // need not transfer execution to successor.
1908 // assume(%p) // from this point on %p is true
1909 // br label %BB
1910 // BB: <LVI info2 for %p val, i.e. %p is true>
1911 // %x = use of %p
1912 // br label exit
1913 //
1914 // Note that this LVI info for blocks BB and SinglPred is correct for %p
1915 // (info2 and info1 respectively). After the merge and the deletion of the
1916 // LVI info1 for SinglePred. We have the following code:
1917 // BB: <LVI info2 for %p val>
1918 // %y = use of %p
1919 // call @exit()
1920 // assume(%p)
1921 // %x = use of %p <-- LVI info2 is correct from here onwards.
1922 // br label exit
1923 // LVI info2 for BB is incorrect at the beginning of BB.
1924
1925 // Invalidate LVI information for BB if the LVI is not provably true for
1926 // all of BB.
1927 if (!isGuaranteedToTransferExecutionToSuccessor(BB))
1928 LVI->eraseBlock(BB);
1929 return true;
1930 }
1931
1932 /// Update the SSA form. NewBB contains instructions that are copied from BB.
1933 /// ValueMapping maps old values in BB to new ones in NewBB.
UpdateSSA(BasicBlock * BB,BasicBlock * NewBB,DenseMap<Instruction *,Value * > & ValueMapping)1934 void JumpThreadingPass::UpdateSSA(
1935 BasicBlock *BB, BasicBlock *NewBB,
1936 DenseMap<Instruction *, Value *> &ValueMapping) {
1937 // If there were values defined in BB that are used outside the block, then we
1938 // now have to update all uses of the value to use either the original value,
1939 // the cloned value, or some PHI derived value. This can require arbitrary
1940 // PHI insertion, of which we are prepared to do, clean these up now.
1941 SSAUpdater SSAUpdate;
1942 SmallVector<Use *, 16> UsesToRename;
1943
1944 for (Instruction &I : *BB) {
1945 // Scan all uses of this instruction to see if it is used outside of its
1946 // block, and if so, record them in UsesToRename.
1947 for (Use &U : I.uses()) {
1948 Instruction *User = cast<Instruction>(U.getUser());
1949 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1950 if (UserPN->getIncomingBlock(U) == BB)
1951 continue;
1952 } else if (User->getParent() == BB)
1953 continue;
1954
1955 UsesToRename.push_back(&U);
1956 }
1957
1958 // If there are no uses outside the block, we're done with this instruction.
1959 if (UsesToRename.empty())
1960 continue;
1961 LLVM_DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n");
1962
1963 // We found a use of I outside of BB. Rename all uses of I that are outside
1964 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1965 // with the two values we know.
1966 SSAUpdate.Initialize(I.getType(), I.getName());
1967 SSAUpdate.AddAvailableValue(BB, &I);
1968 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[&I]);
1969
1970 while (!UsesToRename.empty())
1971 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1972 LLVM_DEBUG(dbgs() << "\n");
1973 }
1974 }
1975
1976 /// Clone instructions in range [BI, BE) to NewBB. For PHI nodes, we only clone
1977 /// arguments that come from PredBB. Return the map from the variables in the
1978 /// source basic block to the variables in the newly created basic block.
1979 DenseMap<Instruction *, Value *>
CloneInstructions(BasicBlock::iterator BI,BasicBlock::iterator BE,BasicBlock * NewBB,BasicBlock * PredBB)1980 JumpThreadingPass::CloneInstructions(BasicBlock::iterator BI,
1981 BasicBlock::iterator BE, BasicBlock *NewBB,
1982 BasicBlock *PredBB) {
1983 // We are going to have to map operands from the source basic block to the new
1984 // copy of the block 'NewBB'. If there are PHI nodes in the source basic
1985 // block, evaluate them to account for entry from PredBB.
1986 DenseMap<Instruction *, Value *> ValueMapping;
1987
1988 // Clone the phi nodes of the source basic block into NewBB. The resulting
1989 // phi nodes are trivial since NewBB only has one predecessor, but SSAUpdater
1990 // might need to rewrite the operand of the cloned phi.
1991 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) {
1992 PHINode *NewPN = PHINode::Create(PN->getType(), 1, PN->getName(), NewBB);
1993 NewPN->addIncoming(PN->getIncomingValueForBlock(PredBB), PredBB);
1994 ValueMapping[PN] = NewPN;
1995 }
1996
1997 // Clone the non-phi instructions of the source basic block into NewBB,
1998 // keeping track of the mapping and using it to remap operands in the cloned
1999 // instructions.
2000 for (; BI != BE; ++BI) {
2001 Instruction *New = BI->clone();
2002 New->setName(BI->getName());
2003 NewBB->getInstList().push_back(New);
2004 ValueMapping[&*BI] = New;
2005
2006 // Remap operands to patch up intra-block references.
2007 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2008 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
2009 DenseMap<Instruction *, Value *>::iterator I = ValueMapping.find(Inst);
2010 if (I != ValueMapping.end())
2011 New->setOperand(i, I->second);
2012 }
2013 }
2014
2015 return ValueMapping;
2016 }
2017
2018 /// TryThreadEdge - Thread an edge if it's safe and profitable to do so.
TryThreadEdge(BasicBlock * BB,const SmallVectorImpl<BasicBlock * > & PredBBs,BasicBlock * SuccBB)2019 bool JumpThreadingPass::TryThreadEdge(
2020 BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs,
2021 BasicBlock *SuccBB) {
2022 // If threading to the same block as we come from, we would infinite loop.
2023 if (SuccBB == BB) {
2024 LLVM_DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
2025 << "' - would thread to self!\n");
2026 return false;
2027 }
2028
2029 // If threading this would thread across a loop header, don't thread the edge.
2030 // See the comments above FindLoopHeaders for justifications and caveats.
2031 if (LoopHeaders.count(BB) || LoopHeaders.count(SuccBB)) {
2032 LLVM_DEBUG({
2033 bool BBIsHeader = LoopHeaders.count(BB);
2034 bool SuccIsHeader = LoopHeaders.count(SuccBB);
2035 dbgs() << " Not threading across "
2036 << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName()
2037 << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '")
2038 << SuccBB->getName() << "' - it might create an irreducible loop!\n";
2039 });
2040 return false;
2041 }
2042
2043 unsigned JumpThreadCost =
2044 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
2045 if (JumpThreadCost > BBDupThreshold) {
2046 LLVM_DEBUG(dbgs() << " Not threading BB '" << BB->getName()
2047 << "' - Cost is too high: " << JumpThreadCost << "\n");
2048 return false;
2049 }
2050
2051 ThreadEdge(BB, PredBBs, SuccBB);
2052 return true;
2053 }
2054
2055 /// ThreadEdge - We have decided that it is safe and profitable to factor the
2056 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
2057 /// across BB. Transform the IR to reflect this change.
ThreadEdge(BasicBlock * BB,const SmallVectorImpl<BasicBlock * > & PredBBs,BasicBlock * SuccBB)2058 void JumpThreadingPass::ThreadEdge(BasicBlock *BB,
2059 const SmallVectorImpl<BasicBlock *> &PredBBs,
2060 BasicBlock *SuccBB) {
2061 assert(SuccBB != BB && "Don't create an infinite loop");
2062
2063 assert(!LoopHeaders.count(BB) && !LoopHeaders.count(SuccBB) &&
2064 "Don't thread across loop headers");
2065
2066 // And finally, do it! Start by factoring the predecessors if needed.
2067 BasicBlock *PredBB;
2068 if (PredBBs.size() == 1)
2069 PredBB = PredBBs[0];
2070 else {
2071 LLVM_DEBUG(dbgs() << " Factoring out " << PredBBs.size()
2072 << " common predecessors.\n");
2073 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
2074 }
2075
2076 // And finally, do it!
2077 LLVM_DEBUG(dbgs() << " Threading edge from '" << PredBB->getName()
2078 << "' to '" << SuccBB->getName()
2079 << ", across block:\n " << *BB << "\n");
2080
2081 if (DTU->hasPendingDomTreeUpdates())
2082 LVI->disableDT();
2083 else
2084 LVI->enableDT();
2085 LVI->threadEdge(PredBB, BB, SuccBB);
2086
2087 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
2088 BB->getName()+".thread",
2089 BB->getParent(), BB);
2090 NewBB->moveAfter(PredBB);
2091
2092 // Set the block frequency of NewBB.
2093 if (HasProfileData) {
2094 auto NewBBFreq =
2095 BFI->getBlockFreq(PredBB) * BPI->getEdgeProbability(PredBB, BB);
2096 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
2097 }
2098
2099 // Copy all the instructions from BB to NewBB except the terminator.
2100 DenseMap<Instruction *, Value *> ValueMapping =
2101 CloneInstructions(BB->begin(), std::prev(BB->end()), NewBB, PredBB);
2102
2103 // We didn't copy the terminator from BB over to NewBB, because there is now
2104 // an unconditional jump to SuccBB. Insert the unconditional jump.
2105 BranchInst *NewBI = BranchInst::Create(SuccBB, NewBB);
2106 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
2107
2108 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
2109 // PHI nodes for NewBB now.
2110 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
2111
2112 // Update the terminator of PredBB to jump to NewBB instead of BB. This
2113 // eliminates predecessors from BB, which requires us to simplify any PHI
2114 // nodes in BB.
2115 Instruction *PredTerm = PredBB->getTerminator();
2116 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
2117 if (PredTerm->getSuccessor(i) == BB) {
2118 BB->removePredecessor(PredBB, true);
2119 PredTerm->setSuccessor(i, NewBB);
2120 }
2121
2122 // Enqueue required DT updates.
2123 DTU->applyUpdatesPermissive({{DominatorTree::Insert, NewBB, SuccBB},
2124 {DominatorTree::Insert, PredBB, NewBB},
2125 {DominatorTree::Delete, PredBB, BB}});
2126
2127 UpdateSSA(BB, NewBB, ValueMapping);
2128
2129 // At this point, the IR is fully up to date and consistent. Do a quick scan
2130 // over the new instructions and zap any that are constants or dead. This
2131 // frequently happens because of phi translation.
2132 SimplifyInstructionsInBlock(NewBB, TLI);
2133
2134 // Update the edge weight from BB to SuccBB, which should be less than before.
2135 UpdateBlockFreqAndEdgeWeight(PredBB, BB, NewBB, SuccBB);
2136
2137 // Threaded an edge!
2138 ++NumThreads;
2139 }
2140
2141 /// Create a new basic block that will be the predecessor of BB and successor of
2142 /// all blocks in Preds. When profile data is available, update the frequency of
2143 /// this new block.
SplitBlockPreds(BasicBlock * BB,ArrayRef<BasicBlock * > Preds,const char * Suffix)2144 BasicBlock *JumpThreadingPass::SplitBlockPreds(BasicBlock *BB,
2145 ArrayRef<BasicBlock *> Preds,
2146 const char *Suffix) {
2147 SmallVector<BasicBlock *, 2> NewBBs;
2148
2149 // Collect the frequencies of all predecessors of BB, which will be used to
2150 // update the edge weight of the result of splitting predecessors.
2151 DenseMap<BasicBlock *, BlockFrequency> FreqMap;
2152 if (HasProfileData)
2153 for (auto Pred : Preds)
2154 FreqMap.insert(std::make_pair(
2155 Pred, BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, BB)));
2156
2157 // In the case when BB is a LandingPad block we create 2 new predecessors
2158 // instead of just one.
2159 if (BB->isLandingPad()) {
2160 std::string NewName = std::string(Suffix) + ".split-lp";
2161 SplitLandingPadPredecessors(BB, Preds, Suffix, NewName.c_str(), NewBBs);
2162 } else {
2163 NewBBs.push_back(SplitBlockPredecessors(BB, Preds, Suffix));
2164 }
2165
2166 std::vector<DominatorTree::UpdateType> Updates;
2167 Updates.reserve((2 * Preds.size()) + NewBBs.size());
2168 for (auto NewBB : NewBBs) {
2169 BlockFrequency NewBBFreq(0);
2170 Updates.push_back({DominatorTree::Insert, NewBB, BB});
2171 for (auto Pred : predecessors(NewBB)) {
2172 Updates.push_back({DominatorTree::Delete, Pred, BB});
2173 Updates.push_back({DominatorTree::Insert, Pred, NewBB});
2174 if (HasProfileData) // Update frequencies between Pred -> NewBB.
2175 NewBBFreq += FreqMap.lookup(Pred);
2176 }
2177 if (HasProfileData) // Apply the summed frequency to NewBB.
2178 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
2179 }
2180
2181 DTU->applyUpdatesPermissive(Updates);
2182 return NewBBs[0];
2183 }
2184
doesBlockHaveProfileData(BasicBlock * BB)2185 bool JumpThreadingPass::doesBlockHaveProfileData(BasicBlock *BB) {
2186 const Instruction *TI = BB->getTerminator();
2187 assert(TI->getNumSuccessors() > 1 && "not a split");
2188
2189 MDNode *WeightsNode = TI->getMetadata(LLVMContext::MD_prof);
2190 if (!WeightsNode)
2191 return false;
2192
2193 MDString *MDName = cast<MDString>(WeightsNode->getOperand(0));
2194 if (MDName->getString() != "branch_weights")
2195 return false;
2196
2197 // Ensure there are weights for all of the successors. Note that the first
2198 // operand to the metadata node is a name, not a weight.
2199 return WeightsNode->getNumOperands() == TI->getNumSuccessors() + 1;
2200 }
2201
2202 /// Update the block frequency of BB and branch weight and the metadata on the
2203 /// edge BB->SuccBB. This is done by scaling the weight of BB->SuccBB by 1 -
2204 /// Freq(PredBB->BB) / Freq(BB->SuccBB).
UpdateBlockFreqAndEdgeWeight(BasicBlock * PredBB,BasicBlock * BB,BasicBlock * NewBB,BasicBlock * SuccBB)2205 void JumpThreadingPass::UpdateBlockFreqAndEdgeWeight(BasicBlock *PredBB,
2206 BasicBlock *BB,
2207 BasicBlock *NewBB,
2208 BasicBlock *SuccBB) {
2209 if (!HasProfileData)
2210 return;
2211
2212 assert(BFI && BPI && "BFI & BPI should have been created here");
2213
2214 // As the edge from PredBB to BB is deleted, we have to update the block
2215 // frequency of BB.
2216 auto BBOrigFreq = BFI->getBlockFreq(BB);
2217 auto NewBBFreq = BFI->getBlockFreq(NewBB);
2218 auto BB2SuccBBFreq = BBOrigFreq * BPI->getEdgeProbability(BB, SuccBB);
2219 auto BBNewFreq = BBOrigFreq - NewBBFreq;
2220 BFI->setBlockFreq(BB, BBNewFreq.getFrequency());
2221
2222 // Collect updated outgoing edges' frequencies from BB and use them to update
2223 // edge probabilities.
2224 SmallVector<uint64_t, 4> BBSuccFreq;
2225 for (BasicBlock *Succ : successors(BB)) {
2226 auto SuccFreq = (Succ == SuccBB)
2227 ? BB2SuccBBFreq - NewBBFreq
2228 : BBOrigFreq * BPI->getEdgeProbability(BB, Succ);
2229 BBSuccFreq.push_back(SuccFreq.getFrequency());
2230 }
2231
2232 uint64_t MaxBBSuccFreq =
2233 *std::max_element(BBSuccFreq.begin(), BBSuccFreq.end());
2234
2235 SmallVector<BranchProbability, 4> BBSuccProbs;
2236 if (MaxBBSuccFreq == 0)
2237 BBSuccProbs.assign(BBSuccFreq.size(),
2238 {1, static_cast<unsigned>(BBSuccFreq.size())});
2239 else {
2240 for (uint64_t Freq : BBSuccFreq)
2241 BBSuccProbs.push_back(
2242 BranchProbability::getBranchProbability(Freq, MaxBBSuccFreq));
2243 // Normalize edge probabilities so that they sum up to one.
2244 BranchProbability::normalizeProbabilities(BBSuccProbs.begin(),
2245 BBSuccProbs.end());
2246 }
2247
2248 // Update edge probabilities in BPI.
2249 for (int I = 0, E = BBSuccProbs.size(); I < E; I++)
2250 BPI->setEdgeProbability(BB, I, BBSuccProbs[I]);
2251
2252 // Update the profile metadata as well.
2253 //
2254 // Don't do this if the profile of the transformed blocks was statically
2255 // estimated. (This could occur despite the function having an entry
2256 // frequency in completely cold parts of the CFG.)
2257 //
2258 // In this case we don't want to suggest to subsequent passes that the
2259 // calculated weights are fully consistent. Consider this graph:
2260 //
2261 // check_1
2262 // 50% / |
2263 // eq_1 | 50%
2264 // \ |
2265 // check_2
2266 // 50% / |
2267 // eq_2 | 50%
2268 // \ |
2269 // check_3
2270 // 50% / |
2271 // eq_3 | 50%
2272 // \ |
2273 //
2274 // Assuming the blocks check_* all compare the same value against 1, 2 and 3,
2275 // the overall probabilities are inconsistent; the total probability that the
2276 // value is either 1, 2 or 3 is 150%.
2277 //
2278 // As a consequence if we thread eq_1 -> check_2 to check_3, check_2->check_3
2279 // becomes 0%. This is even worse if the edge whose probability becomes 0% is
2280 // the loop exit edge. Then based solely on static estimation we would assume
2281 // the loop was extremely hot.
2282 //
2283 // FIXME this locally as well so that BPI and BFI are consistent as well. We
2284 // shouldn't make edges extremely likely or unlikely based solely on static
2285 // estimation.
2286 if (BBSuccProbs.size() >= 2 && doesBlockHaveProfileData(BB)) {
2287 SmallVector<uint32_t, 4> Weights;
2288 for (auto Prob : BBSuccProbs)
2289 Weights.push_back(Prob.getNumerator());
2290
2291 auto TI = BB->getTerminator();
2292 TI->setMetadata(
2293 LLVMContext::MD_prof,
2294 MDBuilder(TI->getParent()->getContext()).createBranchWeights(Weights));
2295 }
2296 }
2297
2298 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
2299 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
2300 /// If we can duplicate the contents of BB up into PredBB do so now, this
2301 /// improves the odds that the branch will be on an analyzable instruction like
2302 /// a compare.
DuplicateCondBranchOnPHIIntoPred(BasicBlock * BB,const SmallVectorImpl<BasicBlock * > & PredBBs)2303 bool JumpThreadingPass::DuplicateCondBranchOnPHIIntoPred(
2304 BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs) {
2305 assert(!PredBBs.empty() && "Can't handle an empty set");
2306
2307 // If BB is a loop header, then duplicating this block outside the loop would
2308 // cause us to transform this into an irreducible loop, don't do this.
2309 // See the comments above FindLoopHeaders for justifications and caveats.
2310 if (LoopHeaders.count(BB)) {
2311 LLVM_DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
2312 << "' into predecessor block '" << PredBBs[0]->getName()
2313 << "' - it might create an irreducible loop!\n");
2314 return false;
2315 }
2316
2317 unsigned DuplicationCost =
2318 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
2319 if (DuplicationCost > BBDupThreshold) {
2320 LLVM_DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
2321 << "' - Cost is too high: " << DuplicationCost << "\n");
2322 return false;
2323 }
2324
2325 // And finally, do it! Start by factoring the predecessors if needed.
2326 std::vector<DominatorTree::UpdateType> Updates;
2327 BasicBlock *PredBB;
2328 if (PredBBs.size() == 1)
2329 PredBB = PredBBs[0];
2330 else {
2331 LLVM_DEBUG(dbgs() << " Factoring out " << PredBBs.size()
2332 << " common predecessors.\n");
2333 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
2334 }
2335 Updates.push_back({DominatorTree::Delete, PredBB, BB});
2336
2337 // Okay, we decided to do this! Clone all the instructions in BB onto the end
2338 // of PredBB.
2339 LLVM_DEBUG(dbgs() << " Duplicating block '" << BB->getName()
2340 << "' into end of '" << PredBB->getName()
2341 << "' to eliminate branch on phi. Cost: "
2342 << DuplicationCost << " block is:" << *BB << "\n");
2343
2344 // Unless PredBB ends with an unconditional branch, split the edge so that we
2345 // can just clone the bits from BB into the end of the new PredBB.
2346 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
2347
2348 if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
2349 BasicBlock *OldPredBB = PredBB;
2350 PredBB = SplitEdge(OldPredBB, BB);
2351 Updates.push_back({DominatorTree::Insert, OldPredBB, PredBB});
2352 Updates.push_back({DominatorTree::Insert, PredBB, BB});
2353 Updates.push_back({DominatorTree::Delete, OldPredBB, BB});
2354 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
2355 }
2356
2357 // We are going to have to map operands from the original BB block into the
2358 // PredBB block. Evaluate PHI nodes in BB.
2359 DenseMap<Instruction*, Value*> ValueMapping;
2360
2361 BasicBlock::iterator BI = BB->begin();
2362 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
2363 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
2364 // Clone the non-phi instructions of BB into PredBB, keeping track of the
2365 // mapping and using it to remap operands in the cloned instructions.
2366 for (; BI != BB->end(); ++BI) {
2367 Instruction *New = BI->clone();
2368
2369 // Remap operands to patch up intra-block references.
2370 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2371 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
2372 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
2373 if (I != ValueMapping.end())
2374 New->setOperand(i, I->second);
2375 }
2376
2377 // If this instruction can be simplified after the operands are updated,
2378 // just use the simplified value instead. This frequently happens due to
2379 // phi translation.
2380 if (Value *IV = SimplifyInstruction(
2381 New,
2382 {BB->getModule()->getDataLayout(), TLI, nullptr, nullptr, New})) {
2383 ValueMapping[&*BI] = IV;
2384 if (!New->mayHaveSideEffects()) {
2385 New->deleteValue();
2386 New = nullptr;
2387 }
2388 } else {
2389 ValueMapping[&*BI] = New;
2390 }
2391 if (New) {
2392 // Otherwise, insert the new instruction into the block.
2393 New->setName(BI->getName());
2394 PredBB->getInstList().insert(OldPredBranch->getIterator(), New);
2395 // Update Dominance from simplified New instruction operands.
2396 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2397 if (BasicBlock *SuccBB = dyn_cast<BasicBlock>(New->getOperand(i)))
2398 Updates.push_back({DominatorTree::Insert, PredBB, SuccBB});
2399 }
2400 }
2401
2402 // Check to see if the targets of the branch had PHI nodes. If so, we need to
2403 // add entries to the PHI nodes for branch from PredBB now.
2404 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
2405 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
2406 ValueMapping);
2407 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
2408 ValueMapping);
2409
2410 UpdateSSA(BB, PredBB, ValueMapping);
2411
2412 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
2413 // that we nuked.
2414 BB->removePredecessor(PredBB, true);
2415
2416 // Remove the unconditional branch at the end of the PredBB block.
2417 OldPredBranch->eraseFromParent();
2418 DTU->applyUpdatesPermissive(Updates);
2419
2420 ++NumDupes;
2421 return true;
2422 }
2423
2424 // Pred is a predecessor of BB with an unconditional branch to BB. SI is
2425 // a Select instruction in Pred. BB has other predecessors and SI is used in
2426 // a PHI node in BB. SI has no other use.
2427 // A new basic block, NewBB, is created and SI is converted to compare and
2428 // conditional branch. SI is erased from parent.
UnfoldSelectInstr(BasicBlock * Pred,BasicBlock * BB,SelectInst * SI,PHINode * SIUse,unsigned Idx)2429 void JumpThreadingPass::UnfoldSelectInstr(BasicBlock *Pred, BasicBlock *BB,
2430 SelectInst *SI, PHINode *SIUse,
2431 unsigned Idx) {
2432 // Expand the select.
2433 //
2434 // Pred --
2435 // | v
2436 // | NewBB
2437 // | |
2438 // |-----
2439 // v
2440 // BB
2441 BranchInst *PredTerm = cast<BranchInst>(Pred->getTerminator());
2442 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
2443 BB->getParent(), BB);
2444 // Move the unconditional branch to NewBB.
2445 PredTerm->removeFromParent();
2446 NewBB->getInstList().insert(NewBB->end(), PredTerm);
2447 // Create a conditional branch and update PHI nodes.
2448 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
2449 SIUse->setIncomingValue(Idx, SI->getFalseValue());
2450 SIUse->addIncoming(SI->getTrueValue(), NewBB);
2451
2452 // The select is now dead.
2453 SI->eraseFromParent();
2454 DTU->applyUpdatesPermissive({{DominatorTree::Insert, NewBB, BB},
2455 {DominatorTree::Insert, Pred, NewBB}});
2456
2457 // Update any other PHI nodes in BB.
2458 for (BasicBlock::iterator BI = BB->begin();
2459 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
2460 if (Phi != SIUse)
2461 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);
2462 }
2463
TryToUnfoldSelect(SwitchInst * SI,BasicBlock * BB)2464 bool JumpThreadingPass::TryToUnfoldSelect(SwitchInst *SI, BasicBlock *BB) {
2465 PHINode *CondPHI = dyn_cast<PHINode>(SI->getCondition());
2466
2467 if (!CondPHI || CondPHI->getParent() != BB)
2468 return false;
2469
2470 for (unsigned I = 0, E = CondPHI->getNumIncomingValues(); I != E; ++I) {
2471 BasicBlock *Pred = CondPHI->getIncomingBlock(I);
2472 SelectInst *PredSI = dyn_cast<SelectInst>(CondPHI->getIncomingValue(I));
2473
2474 // The second and third condition can be potentially relaxed. Currently
2475 // the conditions help to simplify the code and allow us to reuse existing
2476 // code, developed for TryToUnfoldSelect(CmpInst *, BasicBlock *)
2477 if (!PredSI || PredSI->getParent() != Pred || !PredSI->hasOneUse())
2478 continue;
2479
2480 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
2481 if (!PredTerm || !PredTerm->isUnconditional())
2482 continue;
2483
2484 UnfoldSelectInstr(Pred, BB, PredSI, CondPHI, I);
2485 return true;
2486 }
2487 return false;
2488 }
2489
2490 /// TryToUnfoldSelect - Look for blocks of the form
2491 /// bb1:
2492 /// %a = select
2493 /// br bb2
2494 ///
2495 /// bb2:
2496 /// %p = phi [%a, %bb1] ...
2497 /// %c = icmp %p
2498 /// br i1 %c
2499 ///
2500 /// And expand the select into a branch structure if one of its arms allows %c
2501 /// to be folded. This later enables threading from bb1 over bb2.
TryToUnfoldSelect(CmpInst * CondCmp,BasicBlock * BB)2502 bool JumpThreadingPass::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
2503 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
2504 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
2505 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
2506
2507 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
2508 CondLHS->getParent() != BB)
2509 return false;
2510
2511 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
2512 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
2513 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
2514
2515 // Look if one of the incoming values is a select in the corresponding
2516 // predecessor.
2517 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
2518 continue;
2519
2520 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
2521 if (!PredTerm || !PredTerm->isUnconditional())
2522 continue;
2523
2524 // Now check if one of the select values would allow us to constant fold the
2525 // terminator in BB. We don't do the transform if both sides fold, those
2526 // cases will be threaded in any case.
2527 if (DTU->hasPendingDomTreeUpdates())
2528 LVI->disableDT();
2529 else
2530 LVI->enableDT();
2531 LazyValueInfo::Tristate LHSFolds =
2532 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
2533 CondRHS, Pred, BB, CondCmp);
2534 LazyValueInfo::Tristate RHSFolds =
2535 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
2536 CondRHS, Pred, BB, CondCmp);
2537 if ((LHSFolds != LazyValueInfo::Unknown ||
2538 RHSFolds != LazyValueInfo::Unknown) &&
2539 LHSFolds != RHSFolds) {
2540 UnfoldSelectInstr(Pred, BB, SI, CondLHS, I);
2541 return true;
2542 }
2543 }
2544 return false;
2545 }
2546
2547 /// TryToUnfoldSelectInCurrBB - Look for PHI/Select or PHI/CMP/Select in the
2548 /// same BB in the form
2549 /// bb:
2550 /// %p = phi [false, %bb1], [true, %bb2], [false, %bb3], [true, %bb4], ...
2551 /// %s = select %p, trueval, falseval
2552 ///
2553 /// or
2554 ///
2555 /// bb:
2556 /// %p = phi [0, %bb1], [1, %bb2], [0, %bb3], [1, %bb4], ...
2557 /// %c = cmp %p, 0
2558 /// %s = select %c, trueval, falseval
2559 ///
2560 /// And expand the select into a branch structure. This later enables
2561 /// jump-threading over bb in this pass.
2562 ///
2563 /// Using the similar approach of SimplifyCFG::FoldCondBranchOnPHI(), unfold
2564 /// select if the associated PHI has at least one constant. If the unfolded
2565 /// select is not jump-threaded, it will be folded again in the later
2566 /// optimizations.
TryToUnfoldSelectInCurrBB(BasicBlock * BB)2567 bool JumpThreadingPass::TryToUnfoldSelectInCurrBB(BasicBlock *BB) {
2568 // If threading this would thread across a loop header, don't thread the edge.
2569 // See the comments above FindLoopHeaders for justifications and caveats.
2570 if (LoopHeaders.count(BB))
2571 return false;
2572
2573 for (BasicBlock::iterator BI = BB->begin();
2574 PHINode *PN = dyn_cast<PHINode>(BI); ++BI) {
2575 // Look for a Phi having at least one constant incoming value.
2576 if (llvm::all_of(PN->incoming_values(),
2577 [](Value *V) { return !isa<ConstantInt>(V); }))
2578 continue;
2579
2580 auto isUnfoldCandidate = [BB](SelectInst *SI, Value *V) {
2581 // Check if SI is in BB and use V as condition.
2582 if (SI->getParent() != BB)
2583 return false;
2584 Value *Cond = SI->getCondition();
2585 return (Cond && Cond == V && Cond->getType()->isIntegerTy(1));
2586 };
2587
2588 SelectInst *SI = nullptr;
2589 for (Use &U : PN->uses()) {
2590 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(U.getUser())) {
2591 // Look for a ICmp in BB that compares PN with a constant and is the
2592 // condition of a Select.
2593 if (Cmp->getParent() == BB && Cmp->hasOneUse() &&
2594 isa<ConstantInt>(Cmp->getOperand(1 - U.getOperandNo())))
2595 if (SelectInst *SelectI = dyn_cast<SelectInst>(Cmp->user_back()))
2596 if (isUnfoldCandidate(SelectI, Cmp->use_begin()->get())) {
2597 SI = SelectI;
2598 break;
2599 }
2600 } else if (SelectInst *SelectI = dyn_cast<SelectInst>(U.getUser())) {
2601 // Look for a Select in BB that uses PN as condition.
2602 if (isUnfoldCandidate(SelectI, U.get())) {
2603 SI = SelectI;
2604 break;
2605 }
2606 }
2607 }
2608
2609 if (!SI)
2610 continue;
2611 // Expand the select.
2612 Instruction *Term =
2613 SplitBlockAndInsertIfThen(SI->getCondition(), SI, false);
2614 BasicBlock *SplitBB = SI->getParent();
2615 BasicBlock *NewBB = Term->getParent();
2616 PHINode *NewPN = PHINode::Create(SI->getType(), 2, "", SI);
2617 NewPN->addIncoming(SI->getTrueValue(), Term->getParent());
2618 NewPN->addIncoming(SI->getFalseValue(), BB);
2619 SI->replaceAllUsesWith(NewPN);
2620 SI->eraseFromParent();
2621 // NewBB and SplitBB are newly created blocks which require insertion.
2622 std::vector<DominatorTree::UpdateType> Updates;
2623 Updates.reserve((2 * SplitBB->getTerminator()->getNumSuccessors()) + 3);
2624 Updates.push_back({DominatorTree::Insert, BB, SplitBB});
2625 Updates.push_back({DominatorTree::Insert, BB, NewBB});
2626 Updates.push_back({DominatorTree::Insert, NewBB, SplitBB});
2627 // BB's successors were moved to SplitBB, update DTU accordingly.
2628 for (auto *Succ : successors(SplitBB)) {
2629 Updates.push_back({DominatorTree::Delete, BB, Succ});
2630 Updates.push_back({DominatorTree::Insert, SplitBB, Succ});
2631 }
2632 DTU->applyUpdatesPermissive(Updates);
2633 return true;
2634 }
2635 return false;
2636 }
2637
2638 /// Try to propagate a guard from the current BB into one of its predecessors
2639 /// in case if another branch of execution implies that the condition of this
2640 /// guard is always true. Currently we only process the simplest case that
2641 /// looks like:
2642 ///
2643 /// Start:
2644 /// %cond = ...
2645 /// br i1 %cond, label %T1, label %F1
2646 /// T1:
2647 /// br label %Merge
2648 /// F1:
2649 /// br label %Merge
2650 /// Merge:
2651 /// %condGuard = ...
2652 /// call void(i1, ...) @llvm.experimental.guard( i1 %condGuard )[ "deopt"() ]
2653 ///
2654 /// And cond either implies condGuard or !condGuard. In this case all the
2655 /// instructions before the guard can be duplicated in both branches, and the
2656 /// guard is then threaded to one of them.
ProcessGuards(BasicBlock * BB)2657 bool JumpThreadingPass::ProcessGuards(BasicBlock *BB) {
2658 using namespace PatternMatch;
2659
2660 // We only want to deal with two predecessors.
2661 BasicBlock *Pred1, *Pred2;
2662 auto PI = pred_begin(BB), PE = pred_end(BB);
2663 if (PI == PE)
2664 return false;
2665 Pred1 = *PI++;
2666 if (PI == PE)
2667 return false;
2668 Pred2 = *PI++;
2669 if (PI != PE)
2670 return false;
2671 if (Pred1 == Pred2)
2672 return false;
2673
2674 // Try to thread one of the guards of the block.
2675 // TODO: Look up deeper than to immediate predecessor?
2676 auto *Parent = Pred1->getSinglePredecessor();
2677 if (!Parent || Parent != Pred2->getSinglePredecessor())
2678 return false;
2679
2680 if (auto *BI = dyn_cast<BranchInst>(Parent->getTerminator()))
2681 for (auto &I : *BB)
2682 if (isGuard(&I) && ThreadGuard(BB, cast<IntrinsicInst>(&I), BI))
2683 return true;
2684
2685 return false;
2686 }
2687
2688 /// Try to propagate the guard from BB which is the lower block of a diamond
2689 /// to one of its branches, in case if diamond's condition implies guard's
2690 /// condition.
ThreadGuard(BasicBlock * BB,IntrinsicInst * Guard,BranchInst * BI)2691 bool JumpThreadingPass::ThreadGuard(BasicBlock *BB, IntrinsicInst *Guard,
2692 BranchInst *BI) {
2693 assert(BI->getNumSuccessors() == 2 && "Wrong number of successors?");
2694 assert(BI->isConditional() && "Unconditional branch has 2 successors?");
2695 Value *GuardCond = Guard->getArgOperand(0);
2696 Value *BranchCond = BI->getCondition();
2697 BasicBlock *TrueDest = BI->getSuccessor(0);
2698 BasicBlock *FalseDest = BI->getSuccessor(1);
2699
2700 auto &DL = BB->getModule()->getDataLayout();
2701 bool TrueDestIsSafe = false;
2702 bool FalseDestIsSafe = false;
2703
2704 // True dest is safe if BranchCond => GuardCond.
2705 auto Impl = isImpliedCondition(BranchCond, GuardCond, DL);
2706 if (Impl && *Impl)
2707 TrueDestIsSafe = true;
2708 else {
2709 // False dest is safe if !BranchCond => GuardCond.
2710 Impl = isImpliedCondition(BranchCond, GuardCond, DL, /* LHSIsTrue */ false);
2711 if (Impl && *Impl)
2712 FalseDestIsSafe = true;
2713 }
2714
2715 if (!TrueDestIsSafe && !FalseDestIsSafe)
2716 return false;
2717
2718 BasicBlock *PredUnguardedBlock = TrueDestIsSafe ? TrueDest : FalseDest;
2719 BasicBlock *PredGuardedBlock = FalseDestIsSafe ? TrueDest : FalseDest;
2720
2721 ValueToValueMapTy UnguardedMapping, GuardedMapping;
2722 Instruction *AfterGuard = Guard->getNextNode();
2723 unsigned Cost = getJumpThreadDuplicationCost(BB, AfterGuard, BBDupThreshold);
2724 if (Cost > BBDupThreshold)
2725 return false;
2726 // Duplicate all instructions before the guard and the guard itself to the
2727 // branch where implication is not proved.
2728 BasicBlock *GuardedBlock = DuplicateInstructionsInSplitBetween(
2729 BB, PredGuardedBlock, AfterGuard, GuardedMapping, *DTU);
2730 assert(GuardedBlock && "Could not create the guarded block?");
2731 // Duplicate all instructions before the guard in the unguarded branch.
2732 // Since we have successfully duplicated the guarded block and this block
2733 // has fewer instructions, we expect it to succeed.
2734 BasicBlock *UnguardedBlock = DuplicateInstructionsInSplitBetween(
2735 BB, PredUnguardedBlock, Guard, UnguardedMapping, *DTU);
2736 assert(UnguardedBlock && "Could not create the unguarded block?");
2737 LLVM_DEBUG(dbgs() << "Moved guard " << *Guard << " to block "
2738 << GuardedBlock->getName() << "\n");
2739 // Some instructions before the guard may still have uses. For them, we need
2740 // to create Phi nodes merging their copies in both guarded and unguarded
2741 // branches. Those instructions that have no uses can be just removed.
2742 SmallVector<Instruction *, 4> ToRemove;
2743 for (auto BI = BB->begin(); &*BI != AfterGuard; ++BI)
2744 if (!isa<PHINode>(&*BI))
2745 ToRemove.push_back(&*BI);
2746
2747 Instruction *InsertionPoint = &*BB->getFirstInsertionPt();
2748 assert(InsertionPoint && "Empty block?");
2749 // Substitute with Phis & remove.
2750 for (auto *Inst : reverse(ToRemove)) {
2751 if (!Inst->use_empty()) {
2752 PHINode *NewPN = PHINode::Create(Inst->getType(), 2);
2753 NewPN->addIncoming(UnguardedMapping[Inst], UnguardedBlock);
2754 NewPN->addIncoming(GuardedMapping[Inst], GuardedBlock);
2755 NewPN->insertBefore(InsertionPoint);
2756 Inst->replaceAllUsesWith(NewPN);
2757 }
2758 Inst->eraseFromParent();
2759 }
2760 return true;
2761 }
2762