1//===- JumpThreading.cpp - Thread control through conditional blocks ------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file implements the Jump Threading pass.
10//
11//===----------------------------------------------------------------------===//
12
13#include "llvm/Transforms/Scalar/JumpThreading.h"
14#include "llvm/ADT/DenseMap.h"
15#include "llvm/ADT/DenseSet.h"
16#include "llvm/ADT/MapVector.h"
17#include "llvm/ADT/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/GlobalsModRef.h"
27#include "llvm/Analysis/GuardUtils.h"
28#include "llvm/Analysis/InstructionSimplify.h"
29#include "llvm/Analysis/LazyValueInfo.h"
30#include "llvm/Analysis/Loads.h"
31#include "llvm/Analysis/LoopInfo.h"
32#include "llvm/Analysis/MemoryLocation.h"
33#include "llvm/Analysis/PostDominators.h"
34#include "llvm/Analysis/TargetLibraryInfo.h"
35#include "llvm/Analysis/TargetTransformInfo.h"
36#include "llvm/Analysis/ValueTracking.h"
37#include "llvm/IR/BasicBlock.h"
38#include "llvm/IR/CFG.h"
39#include "llvm/IR/Constant.h"
40#include "llvm/IR/ConstantRange.h"
41#include "llvm/IR/Constants.h"
42#include "llvm/IR/DataLayout.h"
43#include "llvm/IR/DebugInfo.h"
44#include "llvm/IR/Dominators.h"
45#include "llvm/IR/Function.h"
46#include "llvm/IR/InstrTypes.h"
47#include "llvm/IR/Instruction.h"
48#include "llvm/IR/Instructions.h"
49#include "llvm/IR/IntrinsicInst.h"
50#include "llvm/IR/Intrinsics.h"
51#include "llvm/IR/LLVMContext.h"
52#include "llvm/IR/MDBuilder.h"
53#include "llvm/IR/Metadata.h"
54#include "llvm/IR/Module.h"
55#include "llvm/IR/PassManager.h"
56#include "llvm/IR/PatternMatch.h"
57#include "llvm/IR/ProfDataUtils.h"
58#include "llvm/IR/Type.h"
59#include "llvm/IR/Use.h"
60#include "llvm/IR/Value.h"
61#include "llvm/Support/BlockFrequency.h"
62#include "llvm/Support/BranchProbability.h"
63#include "llvm/Support/Casting.h"
64#include "llvm/Support/CommandLine.h"
65#include "llvm/Support/Debug.h"
66#include "llvm/Support/raw_ostream.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 <cstdint>
75#include <iterator>
76#include <memory>
77#include <utility>
78
79using namespace llvm;
80using namespace jumpthreading;
81
82#define DEBUG_TYPE "jump-threading"
83
84STATISTIC(NumThreads, "Number of jumps threaded");
85STATISTIC(NumFolds, "Number of terminators folded");
86STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
87
88static cl::opt<unsigned>
89BBDuplicateThreshold("jump-threading-threshold",
90 cl::desc("Max block size to duplicate for jump threading"),
91 cl::init(Val: 6), cl::Hidden);
92
93static cl::opt<unsigned>
94ImplicationSearchThreshold(
95 "jump-threading-implication-search-threshold",
96 cl::desc("The number of predecessors to search for a stronger "
97 "condition to use to thread over a weaker condition"),
98 cl::init(Val: 3), cl::Hidden);
99
100static cl::opt<unsigned> PhiDuplicateThreshold(
101 "jump-threading-phi-threshold",
102 cl::desc("Max PHIs in BB to duplicate for jump threading"), cl::init(Val: 76),
103 cl::Hidden);
104
105static cl::opt<bool> ThreadAcrossLoopHeaders(
106 "jump-threading-across-loop-headers",
107 cl::desc("Allow JumpThreading to thread across loop headers, for testing"),
108 cl::init(Val: false), cl::Hidden);
109
110JumpThreadingPass::JumpThreadingPass(int T) {
111 DefaultBBDupThreshold = (T == -1) ? BBDuplicateThreshold : unsigned(T);
112}
113
114// Update branch probability information according to conditional
115// branch probability. This is usually made possible for cloned branches
116// in inline instances by the context specific profile in the caller.
117// For instance,
118//
119// [Block PredBB]
120// [Branch PredBr]
121// if (t) {
122// Block A;
123// } else {
124// Block B;
125// }
126//
127// [Block BB]
128// cond = PN([true, %A], [..., %B]); // PHI node
129// [Branch CondBr]
130// if (cond) {
131// ... // P(cond == true) = 1%
132// }
133//
134// Here we know that when block A is taken, cond must be true, which means
135// P(cond == true | A) = 1
136//
137// Given that P(cond == true) = P(cond == true | A) * P(A) +
138// P(cond == true | B) * P(B)
139// we get:
140// P(cond == true ) = P(A) + P(cond == true | B) * P(B)
141//
142// which gives us:
143// P(A) is less than P(cond == true), i.e.
144// P(t == true) <= P(cond == true)
145//
146// In other words, if we know P(cond == true) is unlikely, we know
147// that P(t == true) is also unlikely.
148//
149static void updatePredecessorProfileMetadata(PHINode *PN, BasicBlock *BB) {
150 BranchInst *CondBr = dyn_cast<BranchInst>(Val: BB->getTerminator());
151 if (!CondBr)
152 return;
153
154 uint64_t TrueWeight, FalseWeight;
155 if (!extractBranchWeights(I: *CondBr, TrueVal&: TrueWeight, FalseVal&: FalseWeight))
156 return;
157
158 if (TrueWeight + FalseWeight == 0)
159 // Zero branch_weights do not give a hint for getting branch probabilities.
160 // Technically it would result in division by zero denominator, which is
161 // TrueWeight + FalseWeight.
162 return;
163
164 // Returns the outgoing edge of the dominating predecessor block
165 // that leads to the PhiNode's incoming block:
166 auto GetPredOutEdge =
167 [](BasicBlock *IncomingBB,
168 BasicBlock *PhiBB) -> std::pair<BasicBlock *, BasicBlock *> {
169 auto *PredBB = IncomingBB;
170 auto *SuccBB = PhiBB;
171 SmallPtrSet<BasicBlock *, 16> Visited;
172 while (true) {
173 BranchInst *PredBr = dyn_cast<BranchInst>(Val: PredBB->getTerminator());
174 if (PredBr && PredBr->isConditional())
175 return {PredBB, SuccBB};
176 Visited.insert(Ptr: PredBB);
177 auto *SinglePredBB = PredBB->getSinglePredecessor();
178 if (!SinglePredBB)
179 return {nullptr, nullptr};
180
181 // Stop searching when SinglePredBB has been visited. It means we see
182 // an unreachable loop.
183 if (Visited.count(Ptr: SinglePredBB))
184 return {nullptr, nullptr};
185
186 SuccBB = PredBB;
187 PredBB = SinglePredBB;
188 }
189 };
190
191 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
192 Value *PhiOpnd = PN->getIncomingValue(i);
193 ConstantInt *CI = dyn_cast<ConstantInt>(Val: PhiOpnd);
194
195 if (!CI || !CI->getType()->isIntegerTy(Bitwidth: 1))
196 continue;
197
198 BranchProbability BP =
199 (CI->isOne() ? BranchProbability::getBranchProbability(
200 Numerator: TrueWeight, Denominator: TrueWeight + FalseWeight)
201 : BranchProbability::getBranchProbability(
202 Numerator: FalseWeight, Denominator: TrueWeight + FalseWeight));
203
204 auto PredOutEdge = GetPredOutEdge(PN->getIncomingBlock(i), BB);
205 if (!PredOutEdge.first)
206 return;
207
208 BasicBlock *PredBB = PredOutEdge.first;
209 BranchInst *PredBr = dyn_cast<BranchInst>(Val: PredBB->getTerminator());
210 if (!PredBr)
211 return;
212
213 uint64_t PredTrueWeight, PredFalseWeight;
214 // FIXME: We currently only set the profile data when it is missing.
215 // With PGO, this can be used to refine even existing profile data with
216 // context information. This needs to be done after more performance
217 // testing.
218 if (extractBranchWeights(I: *PredBr, TrueVal&: PredTrueWeight, FalseVal&: PredFalseWeight))
219 continue;
220
221 // We can not infer anything useful when BP >= 50%, because BP is the
222 // upper bound probability value.
223 if (BP >= BranchProbability(50, 100))
224 continue;
225
226 uint32_t Weights[2];
227 if (PredBr->getSuccessor(i: 0) == PredOutEdge.second) {
228 Weights[0] = BP.getNumerator();
229 Weights[1] = BP.getCompl().getNumerator();
230 } else {
231 Weights[0] = BP.getCompl().getNumerator();
232 Weights[1] = BP.getNumerator();
233 }
234 setBranchWeights(I&: *PredBr, Weights, IsExpected: hasBranchWeightOrigin(I: *PredBr));
235 }
236}
237
238PreservedAnalyses JumpThreadingPass::run(Function &F,
239 FunctionAnalysisManager &AM) {
240 auto &TTI = AM.getResult<TargetIRAnalysis>(IR&: F);
241 // Jump Threading has no sense for the targets with divergent CF
242 if (TTI.hasBranchDivergence(F: &F))
243 return PreservedAnalyses::all();
244 auto &TLI = AM.getResult<TargetLibraryAnalysis>(IR&: F);
245 auto &LVI = AM.getResult<LazyValueAnalysis>(IR&: F);
246 auto &AA = AM.getResult<AAManager>(IR&: F);
247 auto &DT = AM.getResult<DominatorTreeAnalysis>(IR&: F);
248
249 bool Changed =
250 runImpl(F, FAM: &AM, TLI: &TLI, TTI: &TTI, LVI: &LVI, AA: &AA,
251 DTU: std::make_unique<DomTreeUpdater>(
252 args: &DT, args: nullptr, args: DomTreeUpdater::UpdateStrategy::Lazy),
253 BFI: std::nullopt, BPI: std::nullopt);
254
255 if (!Changed)
256 return PreservedAnalyses::all();
257
258
259 getDomTreeUpdater()->flush();
260
261#if defined(EXPENSIVE_CHECKS)
262 assert(getDomTreeUpdater()->getDomTree().verify(
263 DominatorTree::VerificationLevel::Full) &&
264 "DT broken after JumpThreading");
265 assert((!getDomTreeUpdater()->hasPostDomTree() ||
266 getDomTreeUpdater()->getPostDomTree().verify(
267 PostDominatorTree::VerificationLevel::Full)) &&
268 "PDT broken after JumpThreading");
269#else
270 assert(getDomTreeUpdater()->getDomTree().verify(
271 DominatorTree::VerificationLevel::Fast) &&
272 "DT broken after JumpThreading");
273 assert((!getDomTreeUpdater()->hasPostDomTree() ||
274 getDomTreeUpdater()->getPostDomTree().verify(
275 PostDominatorTree::VerificationLevel::Fast)) &&
276 "PDT broken after JumpThreading");
277#endif
278
279 return getPreservedAnalysis();
280}
281
282bool JumpThreadingPass::runImpl(Function &F_, FunctionAnalysisManager *FAM_,
283 TargetLibraryInfo *TLI_,
284 TargetTransformInfo *TTI_, LazyValueInfo *LVI_,
285 AliasAnalysis *AA_,
286 std::unique_ptr<DomTreeUpdater> DTU_,
287 std::optional<BlockFrequencyInfo *> BFI_,
288 std::optional<BranchProbabilityInfo *> BPI_) {
289 LLVM_DEBUG(dbgs() << "Jump threading on function '" << F_.getName() << "'\n");
290 F = &F_;
291 FAM = FAM_;
292 TLI = TLI_;
293 TTI = TTI_;
294 LVI = LVI_;
295 AA = AA_;
296 DTU = std::move(DTU_);
297 BFI = BFI_;
298 BPI = BPI_;
299 auto *GuardDecl = F->getParent()->getFunction(
300 Name: Intrinsic::getName(id: Intrinsic::experimental_guard));
301 HasGuards = GuardDecl && !GuardDecl->use_empty();
302
303 // Reduce the number of instructions duplicated when optimizing strictly for
304 // size.
305 if (BBDuplicateThreshold.getNumOccurrences())
306 BBDupThreshold = BBDuplicateThreshold;
307 else if (F->hasFnAttribute(Kind: Attribute::MinSize))
308 BBDupThreshold = 3;
309 else
310 BBDupThreshold = DefaultBBDupThreshold;
311
312 // JumpThreading must not processes blocks unreachable from entry. It's a
313 // waste of compute time and can potentially lead to hangs.
314 SmallPtrSet<BasicBlock *, 16> Unreachable;
315 assert(DTU && "DTU isn't passed into JumpThreading before using it.");
316 assert(DTU->hasDomTree() && "JumpThreading relies on DomTree to proceed.");
317 DominatorTree &DT = DTU->getDomTree();
318 for (auto &BB : *F)
319 if (!DT.isReachableFromEntry(A: &BB))
320 Unreachable.insert(Ptr: &BB);
321
322 if (!ThreadAcrossLoopHeaders)
323 findLoopHeaders(F&: *F);
324
325 bool EverChanged = false;
326 bool Changed;
327 do {
328 Changed = false;
329 for (auto &BB : *F) {
330 if (Unreachable.count(Ptr: &BB))
331 continue;
332 while (processBlock(BB: &BB)) // Thread all of the branches we can over BB.
333 Changed = ChangedSinceLastAnalysisUpdate = true;
334
335 // Jump threading may have introduced redundant debug values into BB
336 // which should be removed.
337 if (Changed)
338 RemoveRedundantDbgInstrs(BB: &BB);
339
340 // Stop processing BB if it's the entry or is now deleted. The following
341 // routines attempt to eliminate BB and locating a suitable replacement
342 // for the entry is non-trivial.
343 if (&BB == &F->getEntryBlock() || DTU->isBBPendingDeletion(DelBB: &BB))
344 continue;
345
346 if (pred_empty(BB: &BB)) {
347 // When processBlock makes BB unreachable it doesn't bother to fix up
348 // the instructions in it. We must remove BB to prevent invalid IR.
349 LLVM_DEBUG(dbgs() << " JT: Deleting dead block '" << BB.getName()
350 << "' with terminator: " << *BB.getTerminator()
351 << '\n');
352 LoopHeaders.erase(V: &BB);
353 LVI->eraseBlock(BB: &BB);
354 DeleteDeadBlock(BB: &BB, DTU: DTU.get());
355 Changed = ChangedSinceLastAnalysisUpdate = true;
356 continue;
357 }
358
359 // processBlock doesn't thread BBs with unconditional TIs. However, if BB
360 // is "almost empty", we attempt to merge BB with its sole successor.
361 auto *BI = dyn_cast<BranchInst>(Val: BB.getTerminator());
362 if (BI && BI->isUnconditional()) {
363 BasicBlock *Succ = BI->getSuccessor(i: 0);
364 if (
365 // The terminator must be the only non-phi instruction in BB.
366 BB.getFirstNonPHIOrDbg(SkipPseudoOp: true)->isTerminator() &&
367 // Don't alter Loop headers and latches to ensure another pass can
368 // detect and transform nested loops later.
369 !LoopHeaders.count(V: &BB) && !LoopHeaders.count(V: Succ) &&
370 TryToSimplifyUncondBranchFromEmptyBlock(BB: &BB, DTU: DTU.get())) {
371 RemoveRedundantDbgInstrs(BB: Succ);
372 // BB is valid for cleanup here because we passed in DTU. F remains
373 // BB's parent until a DTU->getDomTree() event.
374 LVI->eraseBlock(BB: &BB);
375 Changed = ChangedSinceLastAnalysisUpdate = true;
376 }
377 }
378 }
379 EverChanged |= Changed;
380 } while (Changed);
381
382 LoopHeaders.clear();
383 return EverChanged;
384}
385
386// Replace uses of Cond with ToVal when safe to do so. If all uses are
387// replaced, we can remove Cond. We cannot blindly replace all uses of Cond
388// because we may incorrectly replace uses when guards/assumes are uses of
389// of `Cond` and we used the guards/assume to reason about the `Cond` value
390// at the end of block. RAUW unconditionally replaces all uses
391// including the guards/assumes themselves and the uses before the
392// guard/assume.
393static bool replaceFoldableUses(Instruction *Cond, Value *ToVal,
394 BasicBlock *KnownAtEndOfBB) {
395 bool Changed = false;
396 assert(Cond->getType() == ToVal->getType());
397 // We can unconditionally replace all uses in non-local blocks (i.e. uses
398 // strictly dominated by BB), since LVI information is true from the
399 // terminator of BB.
400 if (Cond->getParent() == KnownAtEndOfBB)
401 Changed |= replaceNonLocalUsesWith(From: Cond, To: ToVal);
402 for (Instruction &I : reverse(C&: *KnownAtEndOfBB)) {
403 // Replace any debug-info record users of Cond with ToVal.
404 for (DbgVariableRecord &DVR : filterDbgVars(R: I.getDbgRecordRange()))
405 DVR.replaceVariableLocationOp(OldValue: Cond, NewValue: ToVal, AllowEmpty: true);
406
407 // Reached the Cond whose uses we are trying to replace, so there are no
408 // more uses.
409 if (&I == Cond)
410 break;
411 // We only replace uses in instructions that are guaranteed to reach the end
412 // of BB, where we know Cond is ToVal.
413 if (!isGuaranteedToTransferExecutionToSuccessor(I: &I))
414 break;
415 Changed |= I.replaceUsesOfWith(From: Cond, To: ToVal);
416 }
417 if (Cond->use_empty() && !Cond->mayHaveSideEffects()) {
418 Cond->eraseFromParent();
419 Changed = true;
420 }
421 return Changed;
422}
423
424/// Return the cost of duplicating a piece of this block from first non-phi
425/// and before StopAt instruction to thread across it. Stop scanning the block
426/// when exceeding the threshold. If duplication is impossible, returns ~0U.
427static unsigned getJumpThreadDuplicationCost(const TargetTransformInfo *TTI,
428 BasicBlock *BB,
429 Instruction *StopAt,
430 unsigned Threshold) {
431 assert(StopAt->getParent() == BB && "Not an instruction from proper BB?");
432
433 // Do not duplicate the BB if it has a lot of PHI nodes.
434 // If a threadable chain is too long then the number of PHI nodes can add up,
435 // leading to a substantial increase in compile time when rewriting the SSA.
436 unsigned PhiCount = 0;
437 Instruction *FirstNonPHI = nullptr;
438 for (Instruction &I : *BB) {
439 if (!isa<PHINode>(Val: &I)) {
440 FirstNonPHI = &I;
441 break;
442 }
443 if (++PhiCount > PhiDuplicateThreshold)
444 return ~0U;
445 }
446
447 /// Ignore PHI nodes, these will be flattened when duplication happens.
448 BasicBlock::const_iterator I(FirstNonPHI);
449
450 // FIXME: THREADING will delete values that are just used to compute the
451 // branch, so they shouldn't count against the duplication cost.
452
453 unsigned Bonus = 0;
454 if (BB->getTerminator() == StopAt) {
455 // Threading through a switch statement is particularly profitable. If this
456 // block ends in a switch, decrease its cost to make it more likely to
457 // happen.
458 if (isa<SwitchInst>(Val: StopAt))
459 Bonus = 6;
460
461 // The same holds for indirect branches, but slightly more so.
462 if (isa<IndirectBrInst>(Val: StopAt))
463 Bonus = 8;
464 }
465
466 // Bump the threshold up so the early exit from the loop doesn't skip the
467 // terminator-based Size adjustment at the end.
468 Threshold += Bonus;
469
470 // Sum up the cost of each instruction until we get to the terminator. Don't
471 // include the terminator because the copy won't include it.
472 unsigned Size = 0;
473 for (; &*I != StopAt; ++I) {
474
475 // Stop scanning the block if we've reached the threshold.
476 if (Size > Threshold)
477 return Size;
478
479 // Bail out if this instruction gives back a token type, it is not possible
480 // to duplicate it if it is used outside this BB.
481 if (I->getType()->isTokenTy() && I->isUsedOutsideOfBlock(BB))
482 return ~0U;
483
484 // Blocks with NoDuplicate are modelled as having infinite cost, so they
485 // are never duplicated.
486 if (const CallInst *CI = dyn_cast<CallInst>(Val&: I))
487 if (CI->cannotDuplicate() || CI->isConvergent())
488 return ~0U;
489
490 if (TTI->getInstructionCost(U: &*I, CostKind: TargetTransformInfo::TCK_SizeAndLatency) ==
491 TargetTransformInfo::TCC_Free)
492 continue;
493
494 // All other instructions count for at least one unit.
495 ++Size;
496
497 // Calls are more expensive. If they are non-intrinsic calls, we model them
498 // as having cost of 4. If they are a non-vector intrinsic, we model them
499 // as having cost of 2 total, and if they are a vector intrinsic, we model
500 // them as having cost 1.
501 if (const CallInst *CI = dyn_cast<CallInst>(Val&: I)) {
502 if (!isa<IntrinsicInst>(Val: CI))
503 Size += 3;
504 else if (!CI->getType()->isVectorTy())
505 Size += 1;
506 }
507 }
508
509 return Size > Bonus ? Size - Bonus : 0;
510}
511
512/// findLoopHeaders - We do not want jump threading to turn proper loop
513/// structures into irreducible loops. Doing this breaks up the loop nesting
514/// hierarchy and pessimizes later transformations. To prevent this from
515/// happening, we first have to find the loop headers. Here we approximate this
516/// by finding targets of backedges in the CFG.
517///
518/// Note that there definitely are cases when we want to allow threading of
519/// edges across a loop header. For example, threading a jump from outside the
520/// loop (the preheader) to an exit block of the loop is definitely profitable.
521/// It is also almost always profitable to thread backedges from within the loop
522/// to exit blocks, and is often profitable to thread backedges to other blocks
523/// within the loop (forming a nested loop). This simple analysis is not rich
524/// enough to track all of these properties and keep it up-to-date as the CFG
525/// mutates, so we don't allow any of these transformations.
526void JumpThreadingPass::findLoopHeaders(Function &F) {
527 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
528 FindFunctionBackedges(F, Result&: Edges);
529
530 for (const auto &Edge : Edges)
531 LoopHeaders.insert(V: Edge.second);
532}
533
534/// getKnownConstant - Helper method to determine if we can thread over a
535/// terminator with the given value as its condition, and if so what value to
536/// use for that. What kind of value this is depends on whether we want an
537/// integer or a block address, but an undef is always accepted.
538/// Returns null if Val is null or not an appropriate constant.
539static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
540 if (!Val)
541 return nullptr;
542
543 // Undef is "known" enough.
544 if (UndefValue *U = dyn_cast<UndefValue>(Val))
545 return U;
546
547 if (Preference == WantBlockAddress)
548 return dyn_cast<BlockAddress>(Val: Val->stripPointerCasts());
549
550 return dyn_cast<ConstantInt>(Val);
551}
552
553/// computeValueKnownInPredecessors - Given a basic block BB and a value V, see
554/// if we can infer that the value is a known ConstantInt/BlockAddress or undef
555/// in any of our predecessors. If so, return the known list of value and pred
556/// BB in the result vector.
557///
558/// This returns true if there were any known values.
559bool JumpThreadingPass::computeValueKnownInPredecessorsImpl(
560 Value *V, BasicBlock *BB, PredValueInfo &Result,
561 ConstantPreference Preference, SmallPtrSet<Value *, 4> &RecursionSet,
562 Instruction *CxtI) {
563 const DataLayout &DL = BB->getDataLayout();
564
565 // This method walks up use-def chains recursively. Because of this, we could
566 // get into an infinite loop going around loops in the use-def chain. To
567 // prevent this, keep track of what (value, block) pairs we've already visited
568 // and terminate the search if we loop back to them
569 if (!RecursionSet.insert(Ptr: V).second)
570 return false;
571
572 // If V is a constant, then it is known in all predecessors.
573 if (Constant *KC = getKnownConstant(Val: V, Preference)) {
574 for (BasicBlock *Pred : predecessors(BB))
575 Result.emplace_back(Args&: KC, Args&: Pred);
576
577 return !Result.empty();
578 }
579
580 // If V is a non-instruction value, or an instruction in a different block,
581 // then it can't be derived from a PHI.
582 Instruction *I = dyn_cast<Instruction>(Val: V);
583 if (!I || I->getParent() != BB) {
584
585 // Okay, if this is a live-in value, see if it has a known value at the any
586 // edge from our predecessors.
587 for (BasicBlock *P : predecessors(BB)) {
588 using namespace PatternMatch;
589 // If the value is known by LazyValueInfo to be a constant in a
590 // predecessor, use that information to try to thread this block.
591 Constant *PredCst = LVI->getConstantOnEdge(V, FromBB: P, ToBB: BB, CxtI);
592 // If I is a non-local compare-with-constant instruction, use more-rich
593 // 'getPredicateOnEdge' method. This would be able to handle value
594 // inequalities better, for example if the compare is "X < 4" and "X < 3"
595 // is known true but "X < 4" itself is not available.
596 CmpInst::Predicate Pred;
597 Value *Val;
598 Constant *Cst;
599 if (!PredCst && match(V, P: m_Cmp(Pred, L: m_Value(V&: Val), R: m_Constant(C&: Cst))))
600 PredCst = LVI->getPredicateOnEdge(Pred, V: Val, C: Cst, FromBB: P, ToBB: BB, CxtI);
601 if (Constant *KC = getKnownConstant(Val: PredCst, Preference))
602 Result.emplace_back(Args&: KC, Args&: P);
603 }
604
605 return !Result.empty();
606 }
607
608 /// If I is a PHI node, then we know the incoming values for any constants.
609 if (PHINode *PN = dyn_cast<PHINode>(Val: I)) {
610 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
611 Value *InVal = PN->getIncomingValue(i);
612 if (Constant *KC = getKnownConstant(Val: InVal, Preference)) {
613 Result.emplace_back(Args&: KC, Args: PN->getIncomingBlock(i));
614 } else {
615 Constant *CI = LVI->getConstantOnEdge(V: InVal,
616 FromBB: PN->getIncomingBlock(i),
617 ToBB: BB, CxtI);
618 if (Constant *KC = getKnownConstant(Val: CI, Preference))
619 Result.emplace_back(Args&: KC, Args: PN->getIncomingBlock(i));
620 }
621 }
622
623 return !Result.empty();
624 }
625
626 // Handle Cast instructions.
627 if (CastInst *CI = dyn_cast<CastInst>(Val: I)) {
628 Value *Source = CI->getOperand(i_nocapture: 0);
629 PredValueInfoTy Vals;
630 computeValueKnownInPredecessorsImpl(V: Source, BB, Result&: Vals, Preference,
631 RecursionSet, CxtI);
632 if (Vals.empty())
633 return false;
634
635 // Convert the known values.
636 for (auto &Val : Vals)
637 if (Constant *Folded = ConstantFoldCastOperand(Opcode: CI->getOpcode(), C: Val.first,
638 DestTy: CI->getType(), DL))
639 Result.emplace_back(Args&: Folded, Args&: Val.second);
640
641 return !Result.empty();
642 }
643
644 if (FreezeInst *FI = dyn_cast<FreezeInst>(Val: I)) {
645 Value *Source = FI->getOperand(i_nocapture: 0);
646 computeValueKnownInPredecessorsImpl(V: Source, BB, Result, Preference,
647 RecursionSet, CxtI);
648
649 erase_if(C&: Result, P: [](auto &Pair) {
650 return !isGuaranteedNotToBeUndefOrPoison(Pair.first);
651 });
652
653 return !Result.empty();
654 }
655
656 // Handle some boolean conditions.
657 if (I->getType()->getPrimitiveSizeInBits() == 1) {
658 using namespace PatternMatch;
659 if (Preference != WantInteger)
660 return false;
661 // X | true -> true
662 // X & false -> false
663 Value *Op0, *Op1;
664 if (match(V: I, P: m_LogicalOr(L: m_Value(V&: Op0), R: m_Value(V&: Op1))) ||
665 match(V: I, P: m_LogicalAnd(L: m_Value(V&: Op0), R: m_Value(V&: Op1)))) {
666 PredValueInfoTy LHSVals, RHSVals;
667
668 computeValueKnownInPredecessorsImpl(V: Op0, BB, Result&: LHSVals, Preference: WantInteger,
669 RecursionSet, CxtI);
670 computeValueKnownInPredecessorsImpl(V: Op1, BB, Result&: RHSVals, Preference: WantInteger,
671 RecursionSet, CxtI);
672
673 if (LHSVals.empty() && RHSVals.empty())
674 return false;
675
676 ConstantInt *InterestingVal;
677 if (match(V: I, P: m_LogicalOr()))
678 InterestingVal = ConstantInt::getTrue(Context&: I->getContext());
679 else
680 InterestingVal = ConstantInt::getFalse(Context&: I->getContext());
681
682 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
683
684 // Scan for the sentinel. If we find an undef, force it to the
685 // interesting value: x|undef -> true and x&undef -> false.
686 for (const auto &LHSVal : LHSVals)
687 if (LHSVal.first == InterestingVal || isa<UndefValue>(Val: LHSVal.first)) {
688 Result.emplace_back(Args&: InterestingVal, Args: LHSVal.second);
689 LHSKnownBBs.insert(Ptr: LHSVal.second);
690 }
691 for (const auto &RHSVal : RHSVals)
692 if (RHSVal.first == InterestingVal || isa<UndefValue>(Val: RHSVal.first)) {
693 // If we already inferred a value for this block on the LHS, don't
694 // re-add it.
695 if (!LHSKnownBBs.count(Ptr: RHSVal.second))
696 Result.emplace_back(Args&: InterestingVal, Args: RHSVal.second);
697 }
698
699 return !Result.empty();
700 }
701
702 // Handle the NOT form of XOR.
703 if (I->getOpcode() == Instruction::Xor &&
704 isa<ConstantInt>(Val: I->getOperand(i: 1)) &&
705 cast<ConstantInt>(Val: I->getOperand(i: 1))->isOne()) {
706 computeValueKnownInPredecessorsImpl(V: I->getOperand(i: 0), BB, Result,
707 Preference: WantInteger, RecursionSet, CxtI);
708 if (Result.empty())
709 return false;
710
711 // Invert the known values.
712 for (auto &R : Result)
713 R.first = ConstantExpr::getNot(C: R.first);
714
715 return true;
716 }
717
718 // Try to simplify some other binary operator values.
719 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Val: I)) {
720 if (Preference != WantInteger)
721 return false;
722 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: BO->getOperand(i_nocapture: 1))) {
723 PredValueInfoTy LHSVals;
724 computeValueKnownInPredecessorsImpl(V: BO->getOperand(i_nocapture: 0), BB, Result&: LHSVals,
725 Preference: WantInteger, RecursionSet, CxtI);
726
727 // Try to use constant folding to simplify the binary operator.
728 for (const auto &LHSVal : LHSVals) {
729 Constant *V = LHSVal.first;
730 Constant *Folded =
731 ConstantFoldBinaryOpOperands(Opcode: BO->getOpcode(), LHS: V, RHS: CI, DL);
732
733 if (Constant *KC = getKnownConstant(Val: Folded, Preference: WantInteger))
734 Result.emplace_back(Args&: KC, Args: LHSVal.second);
735 }
736 }
737
738 return !Result.empty();
739 }
740
741 // Handle compare with phi operand, where the PHI is defined in this block.
742 if (CmpInst *Cmp = dyn_cast<CmpInst>(Val: I)) {
743 if (Preference != WantInteger)
744 return false;
745 Type *CmpType = Cmp->getType();
746 Value *CmpLHS = Cmp->getOperand(i_nocapture: 0);
747 Value *CmpRHS = Cmp->getOperand(i_nocapture: 1);
748 CmpInst::Predicate Pred = Cmp->getPredicate();
749
750 PHINode *PN = dyn_cast<PHINode>(Val: CmpLHS);
751 if (!PN)
752 PN = dyn_cast<PHINode>(Val: CmpRHS);
753 // Do not perform phi translation across a loop header phi, because this
754 // may result in comparison of values from two different loop iterations.
755 // FIXME: This check is broken if LoopHeaders is not populated.
756 if (PN && PN->getParent() == BB && !LoopHeaders.contains(V: BB)) {
757 const DataLayout &DL = PN->getDataLayout();
758 // We can do this simplification if any comparisons fold to true or false.
759 // See if any do.
760 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
761 BasicBlock *PredBB = PN->getIncomingBlock(i);
762 Value *LHS, *RHS;
763 if (PN == CmpLHS) {
764 LHS = PN->getIncomingValue(i);
765 RHS = CmpRHS->DoPHITranslation(CurBB: BB, PredBB);
766 } else {
767 LHS = CmpLHS->DoPHITranslation(CurBB: BB, PredBB);
768 RHS = PN->getIncomingValue(i);
769 }
770 Value *Res = simplifyCmpInst(Predicate: Pred, LHS, RHS, Q: {DL});
771 if (!Res) {
772 if (!isa<Constant>(Val: RHS))
773 continue;
774
775 // getPredicateOnEdge call will make no sense if LHS is defined in BB.
776 auto LHSInst = dyn_cast<Instruction>(Val: LHS);
777 if (LHSInst && LHSInst->getParent() == BB)
778 continue;
779
780 Res = LVI->getPredicateOnEdge(Pred, V: LHS, C: cast<Constant>(Val: RHS), FromBB: PredBB,
781 ToBB: BB, CxtI: CxtI ? CxtI : Cmp);
782 }
783
784 if (Constant *KC = getKnownConstant(Val: Res, Preference: WantInteger))
785 Result.emplace_back(Args&: KC, Args&: PredBB);
786 }
787
788 return !Result.empty();
789 }
790
791 // If comparing a live-in value against a constant, see if we know the
792 // live-in value on any predecessors.
793 if (isa<Constant>(Val: CmpRHS) && !CmpType->isVectorTy()) {
794 Constant *CmpConst = cast<Constant>(Val: CmpRHS);
795
796 if (!isa<Instruction>(Val: CmpLHS) ||
797 cast<Instruction>(Val: CmpLHS)->getParent() != BB) {
798 for (BasicBlock *P : predecessors(BB)) {
799 // If the value is known by LazyValueInfo to be a constant in a
800 // predecessor, use that information to try to thread this block.
801 Constant *Res = LVI->getPredicateOnEdge(Pred, V: CmpLHS, C: CmpConst, FromBB: P, ToBB: BB,
802 CxtI: CxtI ? CxtI : Cmp);
803 if (Constant *KC = getKnownConstant(Val: Res, Preference: WantInteger))
804 Result.emplace_back(Args&: KC, Args&: P);
805 }
806
807 return !Result.empty();
808 }
809
810 // InstCombine can fold some forms of constant range checks into
811 // (icmp (add (x, C1)), C2). See if we have we have such a thing with
812 // x as a live-in.
813 {
814 using namespace PatternMatch;
815
816 Value *AddLHS;
817 ConstantInt *AddConst;
818 if (isa<ConstantInt>(Val: CmpConst) &&
819 match(V: CmpLHS, P: m_Add(L: m_Value(V&: AddLHS), R: m_ConstantInt(CI&: AddConst)))) {
820 if (!isa<Instruction>(Val: AddLHS) ||
821 cast<Instruction>(Val: AddLHS)->getParent() != BB) {
822 for (BasicBlock *P : predecessors(BB)) {
823 // If the value is known by LazyValueInfo to be a ConstantRange in
824 // a predecessor, use that information to try to thread this
825 // block.
826 ConstantRange CR = LVI->getConstantRangeOnEdge(
827 V: AddLHS, FromBB: P, ToBB: BB, CxtI: CxtI ? CxtI : cast<Instruction>(Val: CmpLHS));
828 // Propagate the range through the addition.
829 CR = CR.add(Other: AddConst->getValue());
830
831 // Get the range where the compare returns true.
832 ConstantRange CmpRange = ConstantRange::makeExactICmpRegion(
833 Pred, Other: cast<ConstantInt>(Val: CmpConst)->getValue());
834
835 Constant *ResC;
836 if (CmpRange.contains(CR))
837 ResC = ConstantInt::getTrue(Ty: CmpType);
838 else if (CmpRange.inverse().contains(CR))
839 ResC = ConstantInt::getFalse(Ty: CmpType);
840 else
841 continue;
842
843 Result.emplace_back(Args&: ResC, Args&: P);
844 }
845
846 return !Result.empty();
847 }
848 }
849 }
850
851 // Try to find a constant value for the LHS of a comparison,
852 // and evaluate it statically if we can.
853 PredValueInfoTy LHSVals;
854 computeValueKnownInPredecessorsImpl(V: I->getOperand(i: 0), BB, Result&: LHSVals,
855 Preference: WantInteger, RecursionSet, CxtI);
856
857 for (const auto &LHSVal : LHSVals) {
858 Constant *V = LHSVal.first;
859 Constant *Folded =
860 ConstantFoldCompareInstOperands(Predicate: Pred, LHS: V, RHS: CmpConst, DL);
861 if (Constant *KC = getKnownConstant(Val: Folded, Preference: WantInteger))
862 Result.emplace_back(Args&: KC, Args: LHSVal.second);
863 }
864
865 return !Result.empty();
866 }
867 }
868
869 if (SelectInst *SI = dyn_cast<SelectInst>(Val: I)) {
870 // Handle select instructions where at least one operand is a known constant
871 // and we can figure out the condition value for any predecessor block.
872 Constant *TrueVal = getKnownConstant(Val: SI->getTrueValue(), Preference);
873 Constant *FalseVal = getKnownConstant(Val: SI->getFalseValue(), Preference);
874 PredValueInfoTy Conds;
875 if ((TrueVal || FalseVal) &&
876 computeValueKnownInPredecessorsImpl(V: SI->getCondition(), BB, Result&: Conds,
877 Preference: WantInteger, RecursionSet, CxtI)) {
878 for (auto &C : Conds) {
879 Constant *Cond = C.first;
880
881 // Figure out what value to use for the condition.
882 bool KnownCond;
883 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: Cond)) {
884 // A known boolean.
885 KnownCond = CI->isOne();
886 } else {
887 assert(isa<UndefValue>(Cond) && "Unexpected condition value");
888 // Either operand will do, so be sure to pick the one that's a known
889 // constant.
890 // FIXME: Do this more cleverly if both values are known constants?
891 KnownCond = (TrueVal != nullptr);
892 }
893
894 // See if the select has a known constant value for this predecessor.
895 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
896 Result.emplace_back(Args&: Val, Args&: C.second);
897 }
898
899 return !Result.empty();
900 }
901 }
902
903 // If all else fails, see if LVI can figure out a constant value for us.
904 assert(CxtI->getParent() == BB && "CxtI should be in BB");
905 Constant *CI = LVI->getConstant(V, CxtI);
906 if (Constant *KC = getKnownConstant(Val: CI, Preference)) {
907 for (BasicBlock *Pred : predecessors(BB))
908 Result.emplace_back(Args&: KC, Args&: Pred);
909 }
910
911 return !Result.empty();
912}
913
914/// GetBestDestForBranchOnUndef - If we determine that the specified block ends
915/// in an undefined jump, decide which block is best to revector to.
916///
917/// Since we can pick an arbitrary destination, we pick the successor with the
918/// fewest predecessors. This should reduce the in-degree of the others.
919static unsigned getBestDestForJumpOnUndef(BasicBlock *BB) {
920 Instruction *BBTerm = BB->getTerminator();
921 unsigned MinSucc = 0;
922 BasicBlock *TestBB = BBTerm->getSuccessor(Idx: MinSucc);
923 // Compute the successor with the minimum number of predecessors.
924 unsigned MinNumPreds = pred_size(BB: TestBB);
925 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
926 TestBB = BBTerm->getSuccessor(Idx: i);
927 unsigned NumPreds = pred_size(BB: TestBB);
928 if (NumPreds < MinNumPreds) {
929 MinSucc = i;
930 MinNumPreds = NumPreds;
931 }
932 }
933
934 return MinSucc;
935}
936
937static bool hasAddressTakenAndUsed(BasicBlock *BB) {
938 if (!BB->hasAddressTaken()) return false;
939
940 // If the block has its address taken, it may be a tree of dead constants
941 // hanging off of it. These shouldn't keep the block alive.
942 BlockAddress *BA = BlockAddress::get(BB);
943 BA->removeDeadConstantUsers();
944 return !BA->use_empty();
945}
946
947/// processBlock - If there are any predecessors whose control can be threaded
948/// through to a successor, transform them now.
949bool JumpThreadingPass::processBlock(BasicBlock *BB) {
950 // If the block is trivially dead, just return and let the caller nuke it.
951 // This simplifies other transformations.
952 if (DTU->isBBPendingDeletion(DelBB: BB) ||
953 (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()))
954 return false;
955
956 // If this block has a single predecessor, and if that pred has a single
957 // successor, merge the blocks. This encourages recursive jump threading
958 // because now the condition in this block can be threaded through
959 // predecessors of our predecessor block.
960 if (maybeMergeBasicBlockIntoOnlyPred(BB))
961 return true;
962
963 if (tryToUnfoldSelectInCurrBB(BB))
964 return true;
965
966 // Look if we can propagate guards to predecessors.
967 if (HasGuards && processGuards(BB))
968 return true;
969
970 // What kind of constant we're looking for.
971 ConstantPreference Preference = WantInteger;
972
973 // Look to see if the terminator is a conditional branch, switch or indirect
974 // branch, if not we can't thread it.
975 Value *Condition;
976 Instruction *Terminator = BB->getTerminator();
977 if (BranchInst *BI = dyn_cast<BranchInst>(Val: Terminator)) {
978 // Can't thread an unconditional jump.
979 if (BI->isUnconditional()) return false;
980 Condition = BI->getCondition();
981 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Val: Terminator)) {
982 Condition = SI->getCondition();
983 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Val: Terminator)) {
984 // Can't thread indirect branch with no successors.
985 if (IB->getNumSuccessors() == 0) return false;
986 Condition = IB->getAddress()->stripPointerCasts();
987 Preference = WantBlockAddress;
988 } else {
989 return false; // Must be an invoke or callbr.
990 }
991
992 // Keep track if we constant folded the condition in this invocation.
993 bool ConstantFolded = false;
994
995 // Run constant folding to see if we can reduce the condition to a simple
996 // constant.
997 if (Instruction *I = dyn_cast<Instruction>(Val: Condition)) {
998 Value *SimpleVal =
999 ConstantFoldInstruction(I, DL: BB->getDataLayout(), TLI);
1000 if (SimpleVal) {
1001 I->replaceAllUsesWith(V: SimpleVal);
1002 if (isInstructionTriviallyDead(I, TLI))
1003 I->eraseFromParent();
1004 Condition = SimpleVal;
1005 ConstantFolded = true;
1006 }
1007 }
1008
1009 // If the terminator is branching on an undef or freeze undef, we can pick any
1010 // of the successors to branch to. Let getBestDestForJumpOnUndef decide.
1011 auto *FI = dyn_cast<FreezeInst>(Val: Condition);
1012 if (isa<UndefValue>(Val: Condition) ||
1013 (FI && isa<UndefValue>(Val: FI->getOperand(i_nocapture: 0)) && FI->hasOneUse())) {
1014 unsigned BestSucc = getBestDestForJumpOnUndef(BB);
1015 std::vector<DominatorTree::UpdateType> Updates;
1016
1017 // Fold the branch/switch.
1018 Instruction *BBTerm = BB->getTerminator();
1019 Updates.reserve(n: BBTerm->getNumSuccessors());
1020 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
1021 if (i == BestSucc) continue;
1022 BasicBlock *Succ = BBTerm->getSuccessor(Idx: i);
1023 Succ->removePredecessor(Pred: BB, KeepOneInputPHIs: true);
1024 Updates.push_back(x: {DominatorTree::Delete, BB, Succ});
1025 }
1026
1027 LLVM_DEBUG(dbgs() << " In block '" << BB->getName()
1028 << "' folding undef terminator: " << *BBTerm << '\n');
1029 Instruction *NewBI = BranchInst::Create(IfTrue: BBTerm->getSuccessor(Idx: BestSucc), InsertBefore: BBTerm->getIterator());
1030 NewBI->setDebugLoc(BBTerm->getDebugLoc());
1031 ++NumFolds;
1032 BBTerm->eraseFromParent();
1033 DTU->applyUpdatesPermissive(Updates);
1034 if (FI)
1035 FI->eraseFromParent();
1036 return true;
1037 }
1038
1039 // If the terminator of this block is branching on a constant, simplify the
1040 // terminator to an unconditional branch. This can occur due to threading in
1041 // other blocks.
1042 if (getKnownConstant(Val: Condition, Preference)) {
1043 LLVM_DEBUG(dbgs() << " In block '" << BB->getName()
1044 << "' folding terminator: " << *BB->getTerminator()
1045 << '\n');
1046 ++NumFolds;
1047 ConstantFoldTerminator(BB, DeleteDeadConditions: true, TLI: nullptr, DTU: DTU.get());
1048 if (auto *BPI = getBPI())
1049 BPI->eraseBlock(BB);
1050 return true;
1051 }
1052
1053 Instruction *CondInst = dyn_cast<Instruction>(Val: Condition);
1054
1055 // All the rest of our checks depend on the condition being an instruction.
1056 if (!CondInst) {
1057 // FIXME: Unify this with code below.
1058 if (processThreadableEdges(Cond: Condition, BB, Preference, CxtI: Terminator))
1059 return true;
1060 return ConstantFolded;
1061 }
1062
1063 // Some of the following optimization can safely work on the unfrozen cond.
1064 Value *CondWithoutFreeze = CondInst;
1065 if (auto *FI = dyn_cast<FreezeInst>(Val: CondInst))
1066 CondWithoutFreeze = FI->getOperand(i_nocapture: 0);
1067
1068 if (CmpInst *CondCmp = dyn_cast<CmpInst>(Val: CondWithoutFreeze)) {
1069 // If we're branching on a conditional, LVI might be able to determine
1070 // it's value at the branch instruction. We only handle comparisons
1071 // against a constant at this time.
1072 if (Constant *CondConst = dyn_cast<Constant>(Val: CondCmp->getOperand(i_nocapture: 1))) {
1073 Constant *Res =
1074 LVI->getPredicateAt(Pred: CondCmp->getPredicate(), V: CondCmp->getOperand(i_nocapture: 0),
1075 C: CondConst, CxtI: BB->getTerminator(),
1076 /*UseBlockValue=*/false);
1077 if (Res) {
1078 // We can safely replace *some* uses of the CondInst if it has
1079 // exactly one value as returned by LVI. RAUW is incorrect in the
1080 // presence of guards and assumes, that have the `Cond` as the use. This
1081 // is because we use the guards/assume to reason about the `Cond` value
1082 // at the end of block, but RAUW unconditionally replaces all uses
1083 // including the guards/assumes themselves and the uses before the
1084 // guard/assume.
1085 if (replaceFoldableUses(Cond: CondCmp, ToVal: Res, KnownAtEndOfBB: BB))
1086 return true;
1087 }
1088
1089 // We did not manage to simplify this branch, try to see whether
1090 // CondCmp depends on a known phi-select pattern.
1091 if (tryToUnfoldSelect(CondCmp, BB))
1092 return true;
1093 }
1094 }
1095
1096 if (SwitchInst *SI = dyn_cast<SwitchInst>(Val: BB->getTerminator()))
1097 if (tryToUnfoldSelect(SI, BB))
1098 return true;
1099
1100 // Check for some cases that are worth simplifying. Right now we want to look
1101 // for loads that are used by a switch or by the condition for the branch. If
1102 // we see one, check to see if it's partially redundant. If so, insert a PHI
1103 // which can then be used to thread the values.
1104 Value *SimplifyValue = CondWithoutFreeze;
1105
1106 if (CmpInst *CondCmp = dyn_cast<CmpInst>(Val: SimplifyValue))
1107 if (isa<Constant>(Val: CondCmp->getOperand(i_nocapture: 1)))
1108 SimplifyValue = CondCmp->getOperand(i_nocapture: 0);
1109
1110 // TODO: There are other places where load PRE would be profitable, such as
1111 // more complex comparisons.
1112 if (LoadInst *LoadI = dyn_cast<LoadInst>(Val: SimplifyValue))
1113 if (simplifyPartiallyRedundantLoad(LI: LoadI))
1114 return true;
1115
1116 // Before threading, try to propagate profile data backwards:
1117 if (PHINode *PN = dyn_cast<PHINode>(Val: CondInst))
1118 if (PN->getParent() == BB && isa<BranchInst>(Val: BB->getTerminator()))
1119 updatePredecessorProfileMetadata(PN, BB);
1120
1121 // Handle a variety of cases where we are branching on something derived from
1122 // a PHI node in the current block. If we can prove that any predecessors
1123 // compute a predictable value based on a PHI node, thread those predecessors.
1124 if (processThreadableEdges(Cond: CondInst, BB, Preference, CxtI: Terminator))
1125 return true;
1126
1127 // If this is an otherwise-unfoldable branch on a phi node or freeze(phi) in
1128 // the current block, see if we can simplify.
1129 PHINode *PN = dyn_cast<PHINode>(Val: CondWithoutFreeze);
1130 if (PN && PN->getParent() == BB && isa<BranchInst>(Val: BB->getTerminator()))
1131 return processBranchOnPHI(PN);
1132
1133 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
1134 if (CondInst->getOpcode() == Instruction::Xor &&
1135 CondInst->getParent() == BB && isa<BranchInst>(Val: BB->getTerminator()))
1136 return processBranchOnXOR(BO: cast<BinaryOperator>(Val: CondInst));
1137
1138 // Search for a stronger dominating condition that can be used to simplify a
1139 // conditional branch leaving BB.
1140 if (processImpliedCondition(BB))
1141 return true;
1142
1143 return false;
1144}
1145
1146bool JumpThreadingPass::processImpliedCondition(BasicBlock *BB) {
1147 auto *BI = dyn_cast<BranchInst>(Val: BB->getTerminator());
1148 if (!BI || !BI->isConditional())
1149 return false;
1150
1151 Value *Cond = BI->getCondition();
1152 // Assuming that predecessor's branch was taken, if pred's branch condition
1153 // (V) implies Cond, Cond can be either true, undef, or poison. In this case,
1154 // freeze(Cond) is either true or a nondeterministic value.
1155 // If freeze(Cond) has only one use, we can freely fold freeze(Cond) to true
1156 // without affecting other instructions.
1157 auto *FICond = dyn_cast<FreezeInst>(Val: Cond);
1158 if (FICond && FICond->hasOneUse())
1159 Cond = FICond->getOperand(i_nocapture: 0);
1160 else
1161 FICond = nullptr;
1162
1163 BasicBlock *CurrentBB = BB;
1164 BasicBlock *CurrentPred = BB->getSinglePredecessor();
1165 unsigned Iter = 0;
1166
1167 auto &DL = BB->getDataLayout();
1168
1169 while (CurrentPred && Iter++ < ImplicationSearchThreshold) {
1170 auto *PBI = dyn_cast<BranchInst>(Val: CurrentPred->getTerminator());
1171 if (!PBI || !PBI->isConditional())
1172 return false;
1173 if (PBI->getSuccessor(i: 0) != CurrentBB && PBI->getSuccessor(i: 1) != CurrentBB)
1174 return false;
1175
1176 bool CondIsTrue = PBI->getSuccessor(i: 0) == CurrentBB;
1177 std::optional<bool> Implication =
1178 isImpliedCondition(LHS: PBI->getCondition(), RHS: Cond, DL, LHSIsTrue: CondIsTrue);
1179
1180 // If the branch condition of BB (which is Cond) and CurrentPred are
1181 // exactly the same freeze instruction, Cond can be folded into CondIsTrue.
1182 if (!Implication && FICond && isa<FreezeInst>(Val: PBI->getCondition())) {
1183 if (cast<FreezeInst>(Val: PBI->getCondition())->getOperand(i_nocapture: 0) ==
1184 FICond->getOperand(i_nocapture: 0))
1185 Implication = CondIsTrue;
1186 }
1187
1188 if (Implication) {
1189 BasicBlock *KeepSucc = BI->getSuccessor(i: *Implication ? 0 : 1);
1190 BasicBlock *RemoveSucc = BI->getSuccessor(i: *Implication ? 1 : 0);
1191 RemoveSucc->removePredecessor(Pred: BB);
1192 BranchInst *UncondBI = BranchInst::Create(IfTrue: KeepSucc, InsertBefore: BI->getIterator());
1193 UncondBI->setDebugLoc(BI->getDebugLoc());
1194 ++NumFolds;
1195 BI->eraseFromParent();
1196 if (FICond)
1197 FICond->eraseFromParent();
1198
1199 DTU->applyUpdatesPermissive(Updates: {{DominatorTree::Delete, BB, RemoveSucc}});
1200 if (auto *BPI = getBPI())
1201 BPI->eraseBlock(BB);
1202 return true;
1203 }
1204 CurrentBB = CurrentPred;
1205 CurrentPred = CurrentBB->getSinglePredecessor();
1206 }
1207
1208 return false;
1209}
1210
1211/// Return true if Op is an instruction defined in the given block.
1212static bool isOpDefinedInBlock(Value *Op, BasicBlock *BB) {
1213 if (Instruction *OpInst = dyn_cast<Instruction>(Val: Op))
1214 if (OpInst->getParent() == BB)
1215 return true;
1216 return false;
1217}
1218
1219/// simplifyPartiallyRedundantLoad - If LoadI is an obviously partially
1220/// redundant load instruction, eliminate it by replacing it with a PHI node.
1221/// This is an important optimization that encourages jump threading, and needs
1222/// to be run interlaced with other jump threading tasks.
1223bool JumpThreadingPass::simplifyPartiallyRedundantLoad(LoadInst *LoadI) {
1224 // Don't hack volatile and ordered loads.
1225 if (!LoadI->isUnordered()) return false;
1226
1227 // If the load is defined in a block with exactly one predecessor, it can't be
1228 // partially redundant.
1229 BasicBlock *LoadBB = LoadI->getParent();
1230 if (LoadBB->getSinglePredecessor())
1231 return false;
1232
1233 // If the load is defined in an EH pad, it can't be partially redundant,
1234 // because the edges between the invoke and the EH pad cannot have other
1235 // instructions between them.
1236 if (LoadBB->isEHPad())
1237 return false;
1238
1239 Value *LoadedPtr = LoadI->getOperand(i_nocapture: 0);
1240
1241 // If the loaded operand is defined in the LoadBB and its not a phi,
1242 // it can't be available in predecessors.
1243 if (isOpDefinedInBlock(Op: LoadedPtr, BB: LoadBB) && !isa<PHINode>(Val: LoadedPtr))
1244 return false;
1245
1246 // Scan a few instructions up from the load, to see if it is obviously live at
1247 // the entry to its block.
1248 BasicBlock::iterator BBIt(LoadI);
1249 bool IsLoadCSE;
1250 BatchAAResults BatchAA(*AA);
1251 // The dominator tree is updated lazily and may not be valid at this point.
1252 BatchAA.disableDominatorTree();
1253 if (Value *AvailableVal = FindAvailableLoadedValue(
1254 Load: LoadI, ScanBB: LoadBB, ScanFrom&: BBIt, MaxInstsToScan: DefMaxInstsToScan, AA: &BatchAA, IsLoadCSE: &IsLoadCSE)) {
1255 // If the value of the load is locally available within the block, just use
1256 // it. This frequently occurs for reg2mem'd allocas.
1257
1258 if (IsLoadCSE) {
1259 LoadInst *NLoadI = cast<LoadInst>(Val: AvailableVal);
1260 combineMetadataForCSE(K: NLoadI, J: LoadI, DoesKMove: false);
1261 LVI->forgetValue(V: NLoadI);
1262 };
1263
1264 // If the returned value is the load itself, replace with poison. This can
1265 // only happen in dead loops.
1266 if (AvailableVal == LoadI)
1267 AvailableVal = PoisonValue::get(T: LoadI->getType());
1268 if (AvailableVal->getType() != LoadI->getType()) {
1269 AvailableVal = CastInst::CreateBitOrPointerCast(
1270 S: AvailableVal, Ty: LoadI->getType(), Name: "", InsertBefore: LoadI->getIterator());
1271 cast<Instruction>(Val: AvailableVal)->setDebugLoc(LoadI->getDebugLoc());
1272 }
1273 LoadI->replaceAllUsesWith(V: AvailableVal);
1274 LoadI->eraseFromParent();
1275 return true;
1276 }
1277
1278 // Otherwise, if we scanned the whole block and got to the top of the block,
1279 // we know the block is locally transparent to the load. If not, something
1280 // might clobber its value.
1281 if (BBIt != LoadBB->begin())
1282 return false;
1283
1284 // If all of the loads and stores that feed the value have the same AA tags,
1285 // then we can propagate them onto any newly inserted loads.
1286 AAMDNodes AATags = LoadI->getAAMetadata();
1287
1288 SmallPtrSet<BasicBlock*, 8> PredsScanned;
1289
1290 using AvailablePredsTy = SmallVector<std::pair<BasicBlock *, Value *>, 8>;
1291
1292 AvailablePredsTy AvailablePreds;
1293 BasicBlock *OneUnavailablePred = nullptr;
1294 SmallVector<LoadInst*, 8> CSELoads;
1295
1296 // If we got here, the loaded value is transparent through to the start of the
1297 // block. Check to see if it is available in any of the predecessor blocks.
1298 for (BasicBlock *PredBB : predecessors(BB: LoadBB)) {
1299 // If we already scanned this predecessor, skip it.
1300 if (!PredsScanned.insert(Ptr: PredBB).second)
1301 continue;
1302
1303 BBIt = PredBB->end();
1304 unsigned NumScanedInst = 0;
1305 Value *PredAvailable = nullptr;
1306 // NOTE: We don't CSE load that is volatile or anything stronger than
1307 // unordered, that should have been checked when we entered the function.
1308 assert(LoadI->isUnordered() &&
1309 "Attempting to CSE volatile or atomic loads");
1310 // If this is a load on a phi pointer, phi-translate it and search
1311 // for available load/store to the pointer in predecessors.
1312 Type *AccessTy = LoadI->getType();
1313 const auto &DL = LoadI->getDataLayout();
1314 MemoryLocation Loc(LoadedPtr->DoPHITranslation(CurBB: LoadBB, PredBB),
1315 LocationSize::precise(Value: DL.getTypeStoreSize(Ty: AccessTy)),
1316 AATags);
1317 PredAvailable = findAvailablePtrLoadStore(
1318 Loc, AccessTy, AtLeastAtomic: LoadI->isAtomic(), ScanBB: PredBB, ScanFrom&: BBIt, MaxInstsToScan: DefMaxInstsToScan,
1319 AA: &BatchAA, IsLoadCSE: &IsLoadCSE, NumScanedInst: &NumScanedInst);
1320
1321 // If PredBB has a single predecessor, continue scanning through the
1322 // single predecessor.
1323 BasicBlock *SinglePredBB = PredBB;
1324 while (!PredAvailable && SinglePredBB && BBIt == SinglePredBB->begin() &&
1325 NumScanedInst < DefMaxInstsToScan) {
1326 SinglePredBB = SinglePredBB->getSinglePredecessor();
1327 if (SinglePredBB) {
1328 BBIt = SinglePredBB->end();
1329 PredAvailable = findAvailablePtrLoadStore(
1330 Loc, AccessTy, AtLeastAtomic: LoadI->isAtomic(), ScanBB: SinglePredBB, ScanFrom&: BBIt,
1331 MaxInstsToScan: (DefMaxInstsToScan - NumScanedInst), AA: &BatchAA, IsLoadCSE: &IsLoadCSE,
1332 NumScanedInst: &NumScanedInst);
1333 }
1334 }
1335
1336 if (!PredAvailable) {
1337 OneUnavailablePred = PredBB;
1338 continue;
1339 }
1340
1341 if (IsLoadCSE)
1342 CSELoads.push_back(Elt: cast<LoadInst>(Val: PredAvailable));
1343
1344 // If so, this load is partially redundant. Remember this info so that we
1345 // can create a PHI node.
1346 AvailablePreds.emplace_back(Args&: PredBB, Args&: PredAvailable);
1347 }
1348
1349 // If the loaded value isn't available in any predecessor, it isn't partially
1350 // redundant.
1351 if (AvailablePreds.empty()) return false;
1352
1353 // Okay, the loaded value is available in at least one (and maybe all!)
1354 // predecessors. If the value is unavailable in more than one unique
1355 // predecessor, we want to insert a merge block for those common predecessors.
1356 // This ensures that we only have to insert one reload, thus not increasing
1357 // code size.
1358 BasicBlock *UnavailablePred = nullptr;
1359
1360 // If the value is unavailable in one of predecessors, we will end up
1361 // inserting a new instruction into them. It is only valid if all the
1362 // instructions before LoadI are guaranteed to pass execution to its
1363 // successor, or if LoadI is safe to speculate.
1364 // TODO: If this logic becomes more complex, and we will perform PRE insertion
1365 // farther than to a predecessor, we need to reuse the code from GVN's PRE.
1366 // It requires domination tree analysis, so for this simple case it is an
1367 // overkill.
1368 if (PredsScanned.size() != AvailablePreds.size() &&
1369 !isSafeToSpeculativelyExecute(I: LoadI))
1370 for (auto I = LoadBB->begin(); &*I != LoadI; ++I)
1371 if (!isGuaranteedToTransferExecutionToSuccessor(I: &*I))
1372 return false;
1373
1374 // If there is exactly one predecessor where the value is unavailable, the
1375 // already computed 'OneUnavailablePred' block is it. If it ends in an
1376 // unconditional branch, we know that it isn't a critical edge.
1377 if (PredsScanned.size() == AvailablePreds.size()+1 &&
1378 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
1379 UnavailablePred = OneUnavailablePred;
1380 } else if (PredsScanned.size() != AvailablePreds.size()) {
1381 // Otherwise, we had multiple unavailable predecessors or we had a critical
1382 // edge from the one.
1383 SmallVector<BasicBlock*, 8> PredsToSplit;
1384 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
1385
1386 for (const auto &AvailablePred : AvailablePreds)
1387 AvailablePredSet.insert(Ptr: AvailablePred.first);
1388
1389 // Add all the unavailable predecessors to the PredsToSplit list.
1390 for (BasicBlock *P : predecessors(BB: LoadBB)) {
1391 // If the predecessor is an indirect goto, we can't split the edge.
1392 if (isa<IndirectBrInst>(Val: P->getTerminator()))
1393 return false;
1394
1395 if (!AvailablePredSet.count(Ptr: P))
1396 PredsToSplit.push_back(Elt: P);
1397 }
1398
1399 // Split them out to their own block.
1400 UnavailablePred = splitBlockPreds(BB: LoadBB, Preds: PredsToSplit, Suffix: "thread-pre-split");
1401 }
1402
1403 // If the value isn't available in all predecessors, then there will be
1404 // exactly one where it isn't available. Insert a load on that edge and add
1405 // it to the AvailablePreds list.
1406 if (UnavailablePred) {
1407 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
1408 "Can't handle critical edge here!");
1409 LoadInst *NewVal = new LoadInst(
1410 LoadI->getType(), LoadedPtr->DoPHITranslation(CurBB: LoadBB, PredBB: UnavailablePred),
1411 LoadI->getName() + ".pr", false, LoadI->getAlign(),
1412 LoadI->getOrdering(), LoadI->getSyncScopeID(),
1413 UnavailablePred->getTerminator()->getIterator());
1414 NewVal->setDebugLoc(LoadI->getDebugLoc());
1415 if (AATags)
1416 NewVal->setAAMetadata(AATags);
1417
1418 AvailablePreds.emplace_back(Args&: UnavailablePred, Args&: NewVal);
1419 }
1420
1421 // Now we know that each predecessor of this block has a value in
1422 // AvailablePreds, sort them for efficient access as we're walking the preds.
1423 array_pod_sort(Start: AvailablePreds.begin(), End: AvailablePreds.end());
1424
1425 // Create a PHI node at the start of the block for the PRE'd load value.
1426 PHINode *PN = PHINode::Create(Ty: LoadI->getType(), NumReservedValues: pred_size(BB: LoadBB), NameStr: "");
1427 PN->insertBefore(InsertPos: LoadBB->begin());
1428 PN->takeName(V: LoadI);
1429 PN->setDebugLoc(LoadI->getDebugLoc());
1430
1431 // Insert new entries into the PHI for each predecessor. A single block may
1432 // have multiple entries here.
1433 for (BasicBlock *P : predecessors(BB: LoadBB)) {
1434 AvailablePredsTy::iterator I =
1435 llvm::lower_bound(Range&: AvailablePreds, Value: std::make_pair(x&: P, y: (Value *)nullptr));
1436
1437 assert(I != AvailablePreds.end() && I->first == P &&
1438 "Didn't find entry for predecessor!");
1439
1440 // If we have an available predecessor but it requires casting, insert the
1441 // cast in the predecessor and use the cast. Note that we have to update the
1442 // AvailablePreds vector as we go so that all of the PHI entries for this
1443 // predecessor use the same bitcast.
1444 Value *&PredV = I->second;
1445 if (PredV->getType() != LoadI->getType())
1446 PredV = CastInst::CreateBitOrPointerCast(
1447 S: PredV, Ty: LoadI->getType(), Name: "", InsertBefore: P->getTerminator()->getIterator());
1448
1449 PN->addIncoming(V: PredV, BB: I->first);
1450 }
1451
1452 for (LoadInst *PredLoadI : CSELoads) {
1453 combineMetadataForCSE(K: PredLoadI, J: LoadI, DoesKMove: true);
1454 LVI->forgetValue(V: PredLoadI);
1455 }
1456
1457 LoadI->replaceAllUsesWith(V: PN);
1458 LoadI->eraseFromParent();
1459
1460 return true;
1461}
1462
1463/// findMostPopularDest - The specified list contains multiple possible
1464/// threadable destinations. Pick the one that occurs the most frequently in
1465/// the list.
1466static BasicBlock *
1467findMostPopularDest(BasicBlock *BB,
1468 const SmallVectorImpl<std::pair<BasicBlock *,
1469 BasicBlock *>> &PredToDestList) {
1470 assert(!PredToDestList.empty());
1471
1472 // Determine popularity. If there are multiple possible destinations, we
1473 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1474 // blocks with known and real destinations to threading undef. We'll handle
1475 // them later if interesting.
1476 MapVector<BasicBlock *, unsigned> DestPopularity;
1477
1478 // Populate DestPopularity with the successors in the order they appear in the
1479 // successor list. This way, we ensure determinism by iterating it in the
1480 // same order in llvm::max_element below. We map nullptr to 0 so that we can
1481 // return nullptr when PredToDestList contains nullptr only.
1482 DestPopularity[nullptr] = 0;
1483 for (auto *SuccBB : successors(BB))
1484 DestPopularity[SuccBB] = 0;
1485
1486 for (const auto &PredToDest : PredToDestList)
1487 if (PredToDest.second)
1488 DestPopularity[PredToDest.second]++;
1489
1490 // Find the most popular dest.
1491 auto MostPopular = llvm::max_element(Range&: DestPopularity, C: llvm::less_second());
1492
1493 // Okay, we have finally picked the most popular destination.
1494 return MostPopular->first;
1495}
1496
1497// Try to evaluate the value of V when the control flows from PredPredBB to
1498// BB->getSinglePredecessor() and then on to BB.
1499Constant *JumpThreadingPass::evaluateOnPredecessorEdge(BasicBlock *BB,
1500 BasicBlock *PredPredBB,
1501 Value *V,
1502 const DataLayout &DL) {
1503 BasicBlock *PredBB = BB->getSinglePredecessor();
1504 assert(PredBB && "Expected a single predecessor");
1505
1506 if (Constant *Cst = dyn_cast<Constant>(Val: V)) {
1507 return Cst;
1508 }
1509
1510 // Consult LVI if V is not an instruction in BB or PredBB.
1511 Instruction *I = dyn_cast<Instruction>(Val: V);
1512 if (!I || (I->getParent() != BB && I->getParent() != PredBB)) {
1513 return LVI->getConstantOnEdge(V, FromBB: PredPredBB, ToBB: PredBB, CxtI: nullptr);
1514 }
1515
1516 // Look into a PHI argument.
1517 if (PHINode *PHI = dyn_cast<PHINode>(Val: V)) {
1518 if (PHI->getParent() == PredBB)
1519 return dyn_cast<Constant>(Val: PHI->getIncomingValueForBlock(BB: PredPredBB));
1520 return nullptr;
1521 }
1522
1523 // If we have a CmpInst, try to fold it for each incoming edge into PredBB.
1524 if (CmpInst *CondCmp = dyn_cast<CmpInst>(Val: V)) {
1525 if (CondCmp->getParent() == BB) {
1526 Constant *Op0 =
1527 evaluateOnPredecessorEdge(BB, PredPredBB, V: CondCmp->getOperand(i_nocapture: 0), DL);
1528 Constant *Op1 =
1529 evaluateOnPredecessorEdge(BB, PredPredBB, V: CondCmp->getOperand(i_nocapture: 1), DL);
1530 if (Op0 && Op1) {
1531 return ConstantFoldCompareInstOperands(Predicate: CondCmp->getPredicate(), LHS: Op0,
1532 RHS: Op1, DL);
1533 }
1534 }
1535 return nullptr;
1536 }
1537
1538 return nullptr;
1539}
1540
1541bool JumpThreadingPass::processThreadableEdges(Value *Cond, BasicBlock *BB,
1542 ConstantPreference Preference,
1543 Instruction *CxtI) {
1544 // If threading this would thread across a loop header, don't even try to
1545 // thread the edge.
1546 if (LoopHeaders.count(V: BB))
1547 return false;
1548
1549 PredValueInfoTy PredValues;
1550 if (!computeValueKnownInPredecessors(V: Cond, BB, Result&: PredValues, Preference,
1551 CxtI)) {
1552 // We don't have known values in predecessors. See if we can thread through
1553 // BB and its sole predecessor.
1554 return maybethreadThroughTwoBasicBlocks(BB, Cond);
1555 }
1556
1557 assert(!PredValues.empty() &&
1558 "computeValueKnownInPredecessors returned true with no values");
1559
1560 LLVM_DEBUG(dbgs() << "IN BB: " << *BB;
1561 for (const auto &PredValue : PredValues) {
1562 dbgs() << " BB '" << BB->getName()
1563 << "': FOUND condition = " << *PredValue.first
1564 << " for pred '" << PredValue.second->getName() << "'.\n";
1565 });
1566
1567 // Decide what we want to thread through. Convert our list of known values to
1568 // a list of known destinations for each pred. This also discards duplicate
1569 // predecessors and keeps track of the undefined inputs (which are represented
1570 // as a null dest in the PredToDestList).
1571 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1572 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1573
1574 BasicBlock *OnlyDest = nullptr;
1575 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1576 Constant *OnlyVal = nullptr;
1577 Constant *MultipleVal = (Constant *)(intptr_t)~0ULL;
1578
1579 for (const auto &PredValue : PredValues) {
1580 BasicBlock *Pred = PredValue.second;
1581 if (!SeenPreds.insert(Ptr: Pred).second)
1582 continue; // Duplicate predecessor entry.
1583
1584 Constant *Val = PredValue.first;
1585
1586 BasicBlock *DestBB;
1587 if (isa<UndefValue>(Val))
1588 DestBB = nullptr;
1589 else if (BranchInst *BI = dyn_cast<BranchInst>(Val: BB->getTerminator())) {
1590 assert(isa<ConstantInt>(Val) && "Expecting a constant integer");
1591 DestBB = BI->getSuccessor(i: cast<ConstantInt>(Val)->isZero());
1592 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Val: BB->getTerminator())) {
1593 assert(isa<ConstantInt>(Val) && "Expecting a constant integer");
1594 DestBB = SI->findCaseValue(C: cast<ConstantInt>(Val))->getCaseSuccessor();
1595 } else {
1596 assert(isa<IndirectBrInst>(BB->getTerminator())
1597 && "Unexpected terminator");
1598 assert(isa<BlockAddress>(Val) && "Expecting a constant blockaddress");
1599 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1600 }
1601
1602 // If we have exactly one destination, remember it for efficiency below.
1603 if (PredToDestList.empty()) {
1604 OnlyDest = DestBB;
1605 OnlyVal = Val;
1606 } else {
1607 if (OnlyDest != DestBB)
1608 OnlyDest = MultipleDestSentinel;
1609 // It possible we have same destination, but different value, e.g. default
1610 // case in switchinst.
1611 if (Val != OnlyVal)
1612 OnlyVal = MultipleVal;
1613 }
1614
1615 // If the predecessor ends with an indirect goto, we can't change its
1616 // destination.
1617 if (isa<IndirectBrInst>(Val: Pred->getTerminator()))
1618 continue;
1619
1620 PredToDestList.emplace_back(Args&: Pred, Args&: DestBB);
1621 }
1622
1623 // If all edges were unthreadable, we fail.
1624 if (PredToDestList.empty())
1625 return false;
1626
1627 // If all the predecessors go to a single known successor, we want to fold,
1628 // not thread. By doing so, we do not need to duplicate the current block and
1629 // also miss potential opportunities in case we dont/cant duplicate.
1630 if (OnlyDest && OnlyDest != MultipleDestSentinel) {
1631 if (BB->hasNPredecessors(N: PredToDestList.size())) {
1632 bool SeenFirstBranchToOnlyDest = false;
1633 std::vector <DominatorTree::UpdateType> Updates;
1634 Updates.reserve(n: BB->getTerminator()->getNumSuccessors() - 1);
1635 for (BasicBlock *SuccBB : successors(BB)) {
1636 if (SuccBB == OnlyDest && !SeenFirstBranchToOnlyDest) {
1637 SeenFirstBranchToOnlyDest = true; // Don't modify the first branch.
1638 } else {
1639 SuccBB->removePredecessor(Pred: BB, KeepOneInputPHIs: true); // This is unreachable successor.
1640 Updates.push_back(x: {DominatorTree::Delete, BB, SuccBB});
1641 }
1642 }
1643
1644 // Finally update the terminator.
1645 Instruction *Term = BB->getTerminator();
1646 Instruction *NewBI = BranchInst::Create(IfTrue: OnlyDest, InsertBefore: Term->getIterator());
1647 NewBI->setDebugLoc(Term->getDebugLoc());
1648 ++NumFolds;
1649 Term->eraseFromParent();
1650 DTU->applyUpdatesPermissive(Updates);
1651 if (auto *BPI = getBPI())
1652 BPI->eraseBlock(BB);
1653
1654 // If the condition is now dead due to the removal of the old terminator,
1655 // erase it.
1656 if (auto *CondInst = dyn_cast<Instruction>(Val: Cond)) {
1657 if (CondInst->use_empty() && !CondInst->mayHaveSideEffects())
1658 CondInst->eraseFromParent();
1659 // We can safely replace *some* uses of the CondInst if it has
1660 // exactly one value as returned by LVI. RAUW is incorrect in the
1661 // presence of guards and assumes, that have the `Cond` as the use. This
1662 // is because we use the guards/assume to reason about the `Cond` value
1663 // at the end of block, but RAUW unconditionally replaces all uses
1664 // including the guards/assumes themselves and the uses before the
1665 // guard/assume.
1666 else if (OnlyVal && OnlyVal != MultipleVal)
1667 replaceFoldableUses(Cond: CondInst, ToVal: OnlyVal, KnownAtEndOfBB: BB);
1668 }
1669 return true;
1670 }
1671 }
1672
1673 // Determine which is the most common successor. If we have many inputs and
1674 // this block is a switch, we want to start by threading the batch that goes
1675 // to the most popular destination first. If we only know about one
1676 // threadable destination (the common case) we can avoid this.
1677 BasicBlock *MostPopularDest = OnlyDest;
1678
1679 if (MostPopularDest == MultipleDestSentinel) {
1680 // Remove any loop headers from the Dest list, threadEdge conservatively
1681 // won't process them, but we might have other destination that are eligible
1682 // and we still want to process.
1683 erase_if(C&: PredToDestList,
1684 P: [&](const std::pair<BasicBlock *, BasicBlock *> &PredToDest) {
1685 return LoopHeaders.contains(V: PredToDest.second);
1686 });
1687
1688 if (PredToDestList.empty())
1689 return false;
1690
1691 MostPopularDest = findMostPopularDest(BB, PredToDestList);
1692 }
1693
1694 // Now that we know what the most popular destination is, factor all
1695 // predecessors that will jump to it into a single predecessor.
1696 SmallVector<BasicBlock*, 16> PredsToFactor;
1697 for (const auto &PredToDest : PredToDestList)
1698 if (PredToDest.second == MostPopularDest) {
1699 BasicBlock *Pred = PredToDest.first;
1700
1701 // This predecessor may be a switch or something else that has multiple
1702 // edges to the block. Factor each of these edges by listing them
1703 // according to # occurrences in PredsToFactor.
1704 for (BasicBlock *Succ : successors(BB: Pred))
1705 if (Succ == BB)
1706 PredsToFactor.push_back(Elt: Pred);
1707 }
1708
1709 // If the threadable edges are branching on an undefined value, we get to pick
1710 // the destination that these predecessors should get to.
1711 if (!MostPopularDest)
1712 MostPopularDest = BB->getTerminator()->
1713 getSuccessor(Idx: getBestDestForJumpOnUndef(BB));
1714
1715 // Ok, try to thread it!
1716 return tryThreadEdge(BB, PredBBs: PredsToFactor, SuccBB: MostPopularDest);
1717}
1718
1719/// processBranchOnPHI - We have an otherwise unthreadable conditional branch on
1720/// a PHI node (or freeze PHI) in the current block. See if there are any
1721/// simplifications we can do based on inputs to the phi node.
1722bool JumpThreadingPass::processBranchOnPHI(PHINode *PN) {
1723 BasicBlock *BB = PN->getParent();
1724
1725 // TODO: We could make use of this to do it once for blocks with common PHI
1726 // values.
1727 SmallVector<BasicBlock*, 1> PredBBs;
1728 PredBBs.resize(N: 1);
1729
1730 // If any of the predecessor blocks end in an unconditional branch, we can
1731 // *duplicate* the conditional branch into that block in order to further
1732 // encourage jump threading and to eliminate cases where we have branch on a
1733 // phi of an icmp (branch on icmp is much better).
1734 // This is still beneficial when a frozen phi is used as the branch condition
1735 // because it allows CodeGenPrepare to further canonicalize br(freeze(icmp))
1736 // to br(icmp(freeze ...)).
1737 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1738 BasicBlock *PredBB = PN->getIncomingBlock(i);
1739 if (BranchInst *PredBr = dyn_cast<BranchInst>(Val: PredBB->getTerminator()))
1740 if (PredBr->isUnconditional()) {
1741 PredBBs[0] = PredBB;
1742 // Try to duplicate BB into PredBB.
1743 if (duplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1744 return true;
1745 }
1746 }
1747
1748 return false;
1749}
1750
1751/// processBranchOnXOR - We have an otherwise unthreadable conditional branch on
1752/// a xor instruction in the current block. See if there are any
1753/// simplifications we can do based on inputs to the xor.
1754bool JumpThreadingPass::processBranchOnXOR(BinaryOperator *BO) {
1755 BasicBlock *BB = BO->getParent();
1756
1757 // If either the LHS or RHS of the xor is a constant, don't do this
1758 // optimization.
1759 if (isa<ConstantInt>(Val: BO->getOperand(i_nocapture: 0)) ||
1760 isa<ConstantInt>(Val: BO->getOperand(i_nocapture: 1)))
1761 return false;
1762
1763 // If the first instruction in BB isn't a phi, we won't be able to infer
1764 // anything special about any particular predecessor.
1765 if (!isa<PHINode>(Val: BB->front()))
1766 return false;
1767
1768 // If this BB is a landing pad, we won't be able to split the edge into it.
1769 if (BB->isEHPad())
1770 return false;
1771
1772 // If we have a xor as the branch input to this block, and we know that the
1773 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1774 // the condition into the predecessor and fix that value to true, saving some
1775 // logical ops on that path and encouraging other paths to simplify.
1776 //
1777 // This copies something like this:
1778 //
1779 // BB:
1780 // %X = phi i1 [1], [%X']
1781 // %Y = icmp eq i32 %A, %B
1782 // %Z = xor i1 %X, %Y
1783 // br i1 %Z, ...
1784 //
1785 // Into:
1786 // BB':
1787 // %Y = icmp ne i32 %A, %B
1788 // br i1 %Y, ...
1789
1790 PredValueInfoTy XorOpValues;
1791 bool isLHS = true;
1792 if (!computeValueKnownInPredecessors(V: BO->getOperand(i_nocapture: 0), BB, Result&: XorOpValues,
1793 Preference: WantInteger, CxtI: BO)) {
1794 assert(XorOpValues.empty());
1795 if (!computeValueKnownInPredecessors(V: BO->getOperand(i_nocapture: 1), BB, Result&: XorOpValues,
1796 Preference: WantInteger, CxtI: BO))
1797 return false;
1798 isLHS = false;
1799 }
1800
1801 assert(!XorOpValues.empty() &&
1802 "computeValueKnownInPredecessors returned true with no values");
1803
1804 // Scan the information to see which is most popular: true or false. The
1805 // predecessors can be of the set true, false, or undef.
1806 unsigned NumTrue = 0, NumFalse = 0;
1807 for (const auto &XorOpValue : XorOpValues) {
1808 if (isa<UndefValue>(Val: XorOpValue.first))
1809 // Ignore undefs for the count.
1810 continue;
1811 if (cast<ConstantInt>(Val: XorOpValue.first)->isZero())
1812 ++NumFalse;
1813 else
1814 ++NumTrue;
1815 }
1816
1817 // Determine which value to split on, true, false, or undef if neither.
1818 ConstantInt *SplitVal = nullptr;
1819 if (NumTrue > NumFalse)
1820 SplitVal = ConstantInt::getTrue(Context&: BB->getContext());
1821 else if (NumTrue != 0 || NumFalse != 0)
1822 SplitVal = ConstantInt::getFalse(Context&: BB->getContext());
1823
1824 // Collect all of the blocks that this can be folded into so that we can
1825 // factor this once and clone it once.
1826 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1827 for (const auto &XorOpValue : XorOpValues) {
1828 if (XorOpValue.first != SplitVal && !isa<UndefValue>(Val: XorOpValue.first))
1829 continue;
1830
1831 BlocksToFoldInto.push_back(Elt: XorOpValue.second);
1832 }
1833
1834 // If we inferred a value for all of the predecessors, then duplication won't
1835 // help us. However, we can just replace the LHS or RHS with the constant.
1836 if (BlocksToFoldInto.size() ==
1837 cast<PHINode>(Val&: BB->front()).getNumIncomingValues()) {
1838 if (!SplitVal) {
1839 // If all preds provide undef, just nuke the xor, because it is undef too.
1840 BO->replaceAllUsesWith(V: UndefValue::get(T: BO->getType()));
1841 BO->eraseFromParent();
1842 } else if (SplitVal->isZero() && BO != BO->getOperand(i_nocapture: isLHS)) {
1843 // If all preds provide 0, replace the xor with the other input.
1844 BO->replaceAllUsesWith(V: BO->getOperand(i_nocapture: isLHS));
1845 BO->eraseFromParent();
1846 } else {
1847 // If all preds provide 1, set the computed value to 1.
1848 BO->setOperand(i_nocapture: !isLHS, Val_nocapture: SplitVal);
1849 }
1850
1851 return true;
1852 }
1853
1854 // If any of predecessors end with an indirect goto, we can't change its
1855 // destination.
1856 if (any_of(Range&: BlocksToFoldInto, P: [](BasicBlock *Pred) {
1857 return isa<IndirectBrInst>(Val: Pred->getTerminator());
1858 }))
1859 return false;
1860
1861 // Try to duplicate BB into PredBB.
1862 return duplicateCondBranchOnPHIIntoPred(BB, PredBBs: BlocksToFoldInto);
1863}
1864
1865/// addPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1866/// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1867/// NewPred using the entries from OldPred (suitably mapped).
1868static void addPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1869 BasicBlock *OldPred,
1870 BasicBlock *NewPred,
1871 ValueToValueMapTy &ValueMap) {
1872 for (PHINode &PN : PHIBB->phis()) {
1873 // Ok, we have a PHI node. Figure out what the incoming value was for the
1874 // DestBlock.
1875 Value *IV = PN.getIncomingValueForBlock(BB: OldPred);
1876
1877 // Remap the value if necessary.
1878 if (Instruction *Inst = dyn_cast<Instruction>(Val: IV)) {
1879 ValueToValueMapTy::iterator I = ValueMap.find(Val: Inst);
1880 if (I != ValueMap.end())
1881 IV = I->second;
1882 }
1883
1884 PN.addIncoming(V: IV, BB: NewPred);
1885 }
1886}
1887
1888/// Merge basic block BB into its sole predecessor if possible.
1889bool JumpThreadingPass::maybeMergeBasicBlockIntoOnlyPred(BasicBlock *BB) {
1890 BasicBlock *SinglePred = BB->getSinglePredecessor();
1891 if (!SinglePred)
1892 return false;
1893
1894 const Instruction *TI = SinglePred->getTerminator();
1895 if (TI->isSpecialTerminator() || TI->getNumSuccessors() != 1 ||
1896 SinglePred == BB || hasAddressTakenAndUsed(BB))
1897 return false;
1898
1899 // If SinglePred was a loop header, BB becomes one.
1900 if (LoopHeaders.erase(V: SinglePred))
1901 LoopHeaders.insert(V: BB);
1902
1903 LVI->eraseBlock(BB: SinglePred);
1904 MergeBasicBlockIntoOnlyPred(BB, DTU: DTU.get());
1905
1906 // Now that BB is merged into SinglePred (i.e. SinglePred code followed by
1907 // BB code within one basic block `BB`), we need to invalidate the LVI
1908 // information associated with BB, because the LVI information need not be
1909 // true for all of BB after the merge. For example,
1910 // Before the merge, LVI info and code is as follows:
1911 // SinglePred: <LVI info1 for %p val>
1912 // %y = use of %p
1913 // call @exit() // need not transfer execution to successor.
1914 // assume(%p) // from this point on %p is true
1915 // br label %BB
1916 // BB: <LVI info2 for %p val, i.e. %p is true>
1917 // %x = use of %p
1918 // br label exit
1919 //
1920 // Note that this LVI info for blocks BB and SinglPred is correct for %p
1921 // (info2 and info1 respectively). After the merge and the deletion of the
1922 // LVI info1 for SinglePred. We have the following code:
1923 // BB: <LVI info2 for %p val>
1924 // %y = use of %p
1925 // call @exit()
1926 // assume(%p)
1927 // %x = use of %p <-- LVI info2 is correct from here onwards.
1928 // br label exit
1929 // LVI info2 for BB is incorrect at the beginning of BB.
1930
1931 // Invalidate LVI information for BB if the LVI is not provably true for
1932 // all of BB.
1933 if (!isGuaranteedToTransferExecutionToSuccessor(BB))
1934 LVI->eraseBlock(BB);
1935 return true;
1936}
1937
1938/// Update the SSA form. NewBB contains instructions that are copied from BB.
1939/// ValueMapping maps old values in BB to new ones in NewBB.
1940void JumpThreadingPass::updateSSA(BasicBlock *BB, BasicBlock *NewBB,
1941 ValueToValueMapTy &ValueMapping) {
1942 // If there were values defined in BB that are used outside the block, then we
1943 // now have to update all uses of the value to use either the original value,
1944 // the cloned value, or some PHI derived value. This can require arbitrary
1945 // PHI insertion, of which we are prepared to do, clean these up now.
1946 SSAUpdater SSAUpdate;
1947 SmallVector<Use *, 16> UsesToRename;
1948 SmallVector<DbgValueInst *, 4> DbgValues;
1949 SmallVector<DbgVariableRecord *, 4> DbgVariableRecords;
1950
1951 for (Instruction &I : *BB) {
1952 // Scan all uses of this instruction to see if it is used outside of its
1953 // block, and if so, record them in UsesToRename.
1954 for (Use &U : I.uses()) {
1955 Instruction *User = cast<Instruction>(Val: U.getUser());
1956 if (PHINode *UserPN = dyn_cast<PHINode>(Val: User)) {
1957 if (UserPN->getIncomingBlock(U) == BB)
1958 continue;
1959 } else if (User->getParent() == BB)
1960 continue;
1961
1962 UsesToRename.push_back(Elt: &U);
1963 }
1964
1965 // Find debug values outside of the block
1966 findDbgValues(DbgValues, V: &I, DbgVariableRecords: &DbgVariableRecords);
1967 llvm::erase_if(C&: DbgValues, P: [&](const DbgValueInst *DbgVal) {
1968 return DbgVal->getParent() == BB;
1969 });
1970 llvm::erase_if(C&: DbgVariableRecords, P: [&](const DbgVariableRecord *DbgVarRec) {
1971 return DbgVarRec->getParent() == BB;
1972 });
1973
1974 // If there are no uses outside the block, we're done with this instruction.
1975 if (UsesToRename.empty() && DbgValues.empty() && DbgVariableRecords.empty())
1976 continue;
1977 LLVM_DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n");
1978
1979 // We found a use of I outside of BB. Rename all uses of I that are outside
1980 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1981 // with the two values we know.
1982 SSAUpdate.Initialize(Ty: I.getType(), Name: I.getName());
1983 SSAUpdate.AddAvailableValue(BB, V: &I);
1984 SSAUpdate.AddAvailableValue(BB: NewBB, V: ValueMapping[&I]);
1985
1986 while (!UsesToRename.empty())
1987 SSAUpdate.RewriteUse(U&: *UsesToRename.pop_back_val());
1988 if (!DbgValues.empty() || !DbgVariableRecords.empty()) {
1989 SSAUpdate.UpdateDebugValues(I: &I, DbgValues);
1990 SSAUpdate.UpdateDebugValues(I: &I, DbgValues&: DbgVariableRecords);
1991 DbgValues.clear();
1992 DbgVariableRecords.clear();
1993 }
1994
1995 LLVM_DEBUG(dbgs() << "\n");
1996 }
1997}
1998
1999/// Clone instructions in range [BI, BE) to NewBB. For PHI nodes, we only clone
2000/// arguments that come from PredBB. Return the map from the variables in the
2001/// source basic block to the variables in the newly created basic block.
2002
2003void JumpThreadingPass::cloneInstructions(ValueToValueMapTy &ValueMapping,
2004 BasicBlock::iterator BI,
2005 BasicBlock::iterator BE,
2006 BasicBlock *NewBB,
2007 BasicBlock *PredBB) {
2008 // We are going to have to map operands from the source basic block to the new
2009 // copy of the block 'NewBB'. If there are PHI nodes in the source basic
2010 // block, evaluate them to account for entry from PredBB.
2011
2012 // Retargets llvm.dbg.value to any renamed variables.
2013 auto RetargetDbgValueIfPossible = [&](Instruction *NewInst) -> bool {
2014 auto DbgInstruction = dyn_cast<DbgValueInst>(Val: NewInst);
2015 if (!DbgInstruction)
2016 return false;
2017
2018 SmallSet<std::pair<Value *, Value *>, 16> OperandsToRemap;
2019 for (auto DbgOperand : DbgInstruction->location_ops()) {
2020 auto DbgOperandInstruction = dyn_cast<Instruction>(Val: DbgOperand);
2021 if (!DbgOperandInstruction)
2022 continue;
2023
2024 auto I = ValueMapping.find(Val: DbgOperandInstruction);
2025 if (I != ValueMapping.end()) {
2026 OperandsToRemap.insert(
2027 V: std::pair<Value *, Value *>(DbgOperand, I->second));
2028 }
2029 }
2030
2031 for (auto &[OldOp, MappedOp] : OperandsToRemap)
2032 DbgInstruction->replaceVariableLocationOp(OldValue: OldOp, NewValue: MappedOp);
2033 return true;
2034 };
2035
2036 // Duplicate implementation of the above dbg.value code, using
2037 // DbgVariableRecords instead.
2038 auto RetargetDbgVariableRecordIfPossible = [&](DbgVariableRecord *DVR) {
2039 SmallSet<std::pair<Value *, Value *>, 16> OperandsToRemap;
2040 for (auto *Op : DVR->location_ops()) {
2041 Instruction *OpInst = dyn_cast<Instruction>(Val: Op);
2042 if (!OpInst)
2043 continue;
2044
2045 auto I = ValueMapping.find(Val: OpInst);
2046 if (I != ValueMapping.end())
2047 OperandsToRemap.insert(V: {OpInst, I->second});
2048 }
2049
2050 for (auto &[OldOp, MappedOp] : OperandsToRemap)
2051 DVR->replaceVariableLocationOp(OldValue: OldOp, NewValue: MappedOp);
2052 };
2053
2054 BasicBlock *RangeBB = BI->getParent();
2055
2056 // Clone the phi nodes of the source basic block into NewBB. The resulting
2057 // phi nodes are trivial since NewBB only has one predecessor, but SSAUpdater
2058 // might need to rewrite the operand of the cloned phi.
2059 for (; PHINode *PN = dyn_cast<PHINode>(Val&: BI); ++BI) {
2060 PHINode *NewPN = PHINode::Create(Ty: PN->getType(), NumReservedValues: 1, NameStr: PN->getName(), InsertBefore: NewBB);
2061 NewPN->addIncoming(V: PN->getIncomingValueForBlock(BB: PredBB), BB: PredBB);
2062 ValueMapping[PN] = NewPN;
2063 }
2064
2065 // Clone noalias scope declarations in the threaded block. When threading a
2066 // loop exit, we would otherwise end up with two idential scope declarations
2067 // visible at the same time.
2068 SmallVector<MDNode *> NoAliasScopes;
2069 DenseMap<MDNode *, MDNode *> ClonedScopes;
2070 LLVMContext &Context = PredBB->getContext();
2071 identifyNoAliasScopesToClone(Start: BI, End: BE, NoAliasDeclScopes&: NoAliasScopes);
2072 cloneNoAliasScopes(NoAliasDeclScopes: NoAliasScopes, ClonedScopes, Ext: "thread", Context);
2073
2074 auto CloneAndRemapDbgInfo = [&](Instruction *NewInst, Instruction *From) {
2075 auto DVRRange = NewInst->cloneDebugInfoFrom(From);
2076 for (DbgVariableRecord &DVR : filterDbgVars(R: DVRRange))
2077 RetargetDbgVariableRecordIfPossible(&DVR);
2078 };
2079
2080 // Clone the non-phi instructions of the source basic block into NewBB,
2081 // keeping track of the mapping and using it to remap operands in the cloned
2082 // instructions.
2083 for (; BI != BE; ++BI) {
2084 Instruction *New = BI->clone();
2085 New->setName(BI->getName());
2086 New->insertInto(ParentBB: NewBB, It: NewBB->end());
2087 ValueMapping[&*BI] = New;
2088 adaptNoAliasScopes(I: New, ClonedScopes, Context);
2089
2090 CloneAndRemapDbgInfo(New, &*BI);
2091
2092 if (RetargetDbgValueIfPossible(New))
2093 continue;
2094
2095 // Remap operands to patch up intra-block references.
2096 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2097 if (Instruction *Inst = dyn_cast<Instruction>(Val: New->getOperand(i))) {
2098 ValueToValueMapTy::iterator I = ValueMapping.find(Val: Inst);
2099 if (I != ValueMapping.end())
2100 New->setOperand(i, Val: I->second);
2101 }
2102 }
2103
2104 // There may be DbgVariableRecords on the terminator, clone directly from
2105 // marker to marker as there isn't an instruction there.
2106 if (BE != RangeBB->end() && BE->hasDbgRecords()) {
2107 // Dump them at the end.
2108 DbgMarker *Marker = RangeBB->getMarker(It: BE);
2109 DbgMarker *EndMarker = NewBB->createMarker(It: NewBB->end());
2110 auto DVRRange = EndMarker->cloneDebugInfoFrom(From: Marker, FromHere: std::nullopt);
2111 for (DbgVariableRecord &DVR : filterDbgVars(R: DVRRange))
2112 RetargetDbgVariableRecordIfPossible(&DVR);
2113 }
2114
2115 return;
2116}
2117
2118/// Attempt to thread through two successive basic blocks.
2119bool JumpThreadingPass::maybethreadThroughTwoBasicBlocks(BasicBlock *BB,
2120 Value *Cond) {
2121 // Consider:
2122 //
2123 // PredBB:
2124 // %var = phi i32* [ null, %bb1 ], [ @a, %bb2 ]
2125 // %tobool = icmp eq i32 %cond, 0
2126 // br i1 %tobool, label %BB, label ...
2127 //
2128 // BB:
2129 // %cmp = icmp eq i32* %var, null
2130 // br i1 %cmp, label ..., label ...
2131 //
2132 // We don't know the value of %var at BB even if we know which incoming edge
2133 // we take to BB. However, once we duplicate PredBB for each of its incoming
2134 // edges (say, PredBB1 and PredBB2), we know the value of %var in each copy of
2135 // PredBB. Then we can thread edges PredBB1->BB and PredBB2->BB through BB.
2136
2137 // Require that BB end with a Branch for simplicity.
2138 BranchInst *CondBr = dyn_cast<BranchInst>(Val: BB->getTerminator());
2139 if (!CondBr)
2140 return false;
2141
2142 // BB must have exactly one predecessor.
2143 BasicBlock *PredBB = BB->getSinglePredecessor();
2144 if (!PredBB)
2145 return false;
2146
2147 // Require that PredBB end with a conditional Branch. If PredBB ends with an
2148 // unconditional branch, we should be merging PredBB and BB instead. For
2149 // simplicity, we don't deal with a switch.
2150 BranchInst *PredBBBranch = dyn_cast<BranchInst>(Val: PredBB->getTerminator());
2151 if (!PredBBBranch || PredBBBranch->isUnconditional())
2152 return false;
2153
2154 // If PredBB has exactly one incoming edge, we don't gain anything by copying
2155 // PredBB.
2156 if (PredBB->getSinglePredecessor())
2157 return false;
2158
2159 // Don't thread through PredBB if it contains a successor edge to itself, in
2160 // which case we would infinite loop. Suppose we are threading an edge from
2161 // PredPredBB through PredBB and BB to SuccBB with PredBB containing a
2162 // successor edge to itself. If we allowed jump threading in this case, we
2163 // could duplicate PredBB and BB as, say, PredBB.thread and BB.thread. Since
2164 // PredBB.thread has a successor edge to PredBB, we would immediately come up
2165 // with another jump threading opportunity from PredBB.thread through PredBB
2166 // and BB to SuccBB. This jump threading would repeatedly occur. That is, we
2167 // would keep peeling one iteration from PredBB.
2168 if (llvm::is_contained(Range: successors(BB: PredBB), Element: PredBB))
2169 return false;
2170
2171 // Don't thread across a loop header.
2172 if (LoopHeaders.count(V: PredBB))
2173 return false;
2174
2175 // Avoid complication with duplicating EH pads.
2176 if (PredBB->isEHPad())
2177 return false;
2178
2179 // Find a predecessor that we can thread. For simplicity, we only consider a
2180 // successor edge out of BB to which we thread exactly one incoming edge into
2181 // PredBB.
2182 unsigned ZeroCount = 0;
2183 unsigned OneCount = 0;
2184 BasicBlock *ZeroPred = nullptr;
2185 BasicBlock *OnePred = nullptr;
2186 const DataLayout &DL = BB->getDataLayout();
2187 for (BasicBlock *P : predecessors(BB: PredBB)) {
2188 // If PredPred ends with IndirectBrInst, we can't handle it.
2189 if (isa<IndirectBrInst>(Val: P->getTerminator()))
2190 continue;
2191 if (ConstantInt *CI = dyn_cast_or_null<ConstantInt>(
2192 Val: evaluateOnPredecessorEdge(BB, PredPredBB: P, V: Cond, DL))) {
2193 if (CI->isZero()) {
2194 ZeroCount++;
2195 ZeroPred = P;
2196 } else if (CI->isOne()) {
2197 OneCount++;
2198 OnePred = P;
2199 }
2200 }
2201 }
2202
2203 // Disregard complicated cases where we have to thread multiple edges.
2204 BasicBlock *PredPredBB;
2205 if (ZeroCount == 1) {
2206 PredPredBB = ZeroPred;
2207 } else if (OneCount == 1) {
2208 PredPredBB = OnePred;
2209 } else {
2210 return false;
2211 }
2212
2213 BasicBlock *SuccBB = CondBr->getSuccessor(i: PredPredBB == ZeroPred);
2214
2215 // If threading to the same block as we come from, we would infinite loop.
2216 if (SuccBB == BB) {
2217 LLVM_DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
2218 << "' - would thread to self!\n");
2219 return false;
2220 }
2221
2222 // If threading this would thread across a loop header, don't thread the edge.
2223 // See the comments above findLoopHeaders for justifications and caveats.
2224 if (LoopHeaders.count(V: BB) || LoopHeaders.count(V: SuccBB)) {
2225 LLVM_DEBUG({
2226 bool BBIsHeader = LoopHeaders.count(BB);
2227 bool SuccIsHeader = LoopHeaders.count(SuccBB);
2228 dbgs() << " Not threading across "
2229 << (BBIsHeader ? "loop header BB '" : "block BB '")
2230 << BB->getName() << "' to dest "
2231 << (SuccIsHeader ? "loop header BB '" : "block BB '")
2232 << SuccBB->getName()
2233 << "' - it might create an irreducible loop!\n";
2234 });
2235 return false;
2236 }
2237
2238 // Compute the cost of duplicating BB and PredBB.
2239 unsigned BBCost = getJumpThreadDuplicationCost(
2240 TTI, BB, StopAt: BB->getTerminator(), Threshold: BBDupThreshold);
2241 unsigned PredBBCost = getJumpThreadDuplicationCost(
2242 TTI, BB: PredBB, StopAt: PredBB->getTerminator(), Threshold: BBDupThreshold);
2243
2244 // Give up if costs are too high. We need to check BBCost and PredBBCost
2245 // individually before checking their sum because getJumpThreadDuplicationCost
2246 // return (unsigned)~0 for those basic blocks that cannot be duplicated.
2247 if (BBCost > BBDupThreshold || PredBBCost > BBDupThreshold ||
2248 BBCost + PredBBCost > BBDupThreshold) {
2249 LLVM_DEBUG(dbgs() << " Not threading BB '" << BB->getName()
2250 << "' - Cost is too high: " << PredBBCost
2251 << " for PredBB, " << BBCost << "for BB\n");
2252 return false;
2253 }
2254
2255 // Now we are ready to duplicate PredBB.
2256 threadThroughTwoBasicBlocks(PredPredBB, PredBB, BB, SuccBB);
2257 return true;
2258}
2259
2260void JumpThreadingPass::threadThroughTwoBasicBlocks(BasicBlock *PredPredBB,
2261 BasicBlock *PredBB,
2262 BasicBlock *BB,
2263 BasicBlock *SuccBB) {
2264 LLVM_DEBUG(dbgs() << " Threading through '" << PredBB->getName() << "' and '"
2265 << BB->getName() << "'\n");
2266
2267 // Build BPI/BFI before any changes are made to IR.
2268 bool HasProfile = doesBlockHaveProfileData(BB);
2269 auto *BFI = getOrCreateBFI(Force: HasProfile);
2270 auto *BPI = getOrCreateBPI(Force: BFI != nullptr);
2271
2272 BranchInst *CondBr = cast<BranchInst>(Val: BB->getTerminator());
2273 BranchInst *PredBBBranch = cast<BranchInst>(Val: PredBB->getTerminator());
2274
2275 BasicBlock *NewBB =
2276 BasicBlock::Create(Context&: PredBB->getContext(), Name: PredBB->getName() + ".thread",
2277 Parent: PredBB->getParent(), InsertBefore: PredBB);
2278 NewBB->moveAfter(MovePos: PredBB);
2279
2280 // Set the block frequency of NewBB.
2281 if (BFI) {
2282 assert(BPI && "It's expected BPI to exist along with BFI");
2283 auto NewBBFreq = BFI->getBlockFreq(BB: PredPredBB) *
2284 BPI->getEdgeProbability(Src: PredPredBB, Dst: PredBB);
2285 BFI->setBlockFreq(BB: NewBB, Freq: NewBBFreq);
2286 }
2287
2288 // We are going to have to map operands from the original BB block to the new
2289 // copy of the block 'NewBB'. If there are PHI nodes in PredBB, evaluate them
2290 // to account for entry from PredPredBB.
2291 ValueToValueMapTy ValueMapping;
2292 cloneInstructions(ValueMapping, BI: PredBB->begin(), BE: PredBB->end(), NewBB,
2293 PredBB: PredPredBB);
2294
2295 // Copy the edge probabilities from PredBB to NewBB.
2296 if (BPI)
2297 BPI->copyEdgeProbabilities(Src: PredBB, Dst: NewBB);
2298
2299 // Update the terminator of PredPredBB to jump to NewBB instead of PredBB.
2300 // This eliminates predecessors from PredPredBB, which requires us to simplify
2301 // any PHI nodes in PredBB.
2302 Instruction *PredPredTerm = PredPredBB->getTerminator();
2303 for (unsigned i = 0, e = PredPredTerm->getNumSuccessors(); i != e; ++i)
2304 if (PredPredTerm->getSuccessor(Idx: i) == PredBB) {
2305 PredBB->removePredecessor(Pred: PredPredBB, KeepOneInputPHIs: true);
2306 PredPredTerm->setSuccessor(Idx: i, BB: NewBB);
2307 }
2308
2309 addPHINodeEntriesForMappedBlock(PHIBB: PredBBBranch->getSuccessor(i: 0), OldPred: PredBB, NewPred: NewBB,
2310 ValueMap&: ValueMapping);
2311 addPHINodeEntriesForMappedBlock(PHIBB: PredBBBranch->getSuccessor(i: 1), OldPred: PredBB, NewPred: NewBB,
2312 ValueMap&: ValueMapping);
2313
2314 DTU->applyUpdatesPermissive(
2315 Updates: {{DominatorTree::Insert, NewBB, CondBr->getSuccessor(i: 0)},
2316 {DominatorTree::Insert, NewBB, CondBr->getSuccessor(i: 1)},
2317 {DominatorTree::Insert, PredPredBB, NewBB},
2318 {DominatorTree::Delete, PredPredBB, PredBB}});
2319
2320 updateSSA(BB: PredBB, NewBB, ValueMapping);
2321
2322 // Clean up things like PHI nodes with single operands, dead instructions,
2323 // etc.
2324 SimplifyInstructionsInBlock(BB: NewBB, TLI);
2325 SimplifyInstructionsInBlock(BB: PredBB, TLI);
2326
2327 SmallVector<BasicBlock *, 1> PredsToFactor;
2328 PredsToFactor.push_back(Elt: NewBB);
2329 threadEdge(BB, PredBBs: PredsToFactor, SuccBB);
2330}
2331
2332/// tryThreadEdge - Thread an edge if it's safe and profitable to do so.
2333bool JumpThreadingPass::tryThreadEdge(
2334 BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs,
2335 BasicBlock *SuccBB) {
2336 // If threading to the same block as we come from, we would infinite loop.
2337 if (SuccBB == BB) {
2338 LLVM_DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
2339 << "' - would thread to self!\n");
2340 return false;
2341 }
2342
2343 // If threading this would thread across a loop header, don't thread the edge.
2344 // See the comments above findLoopHeaders for justifications and caveats.
2345 if (LoopHeaders.count(V: BB) || LoopHeaders.count(V: SuccBB)) {
2346 LLVM_DEBUG({
2347 bool BBIsHeader = LoopHeaders.count(BB);
2348 bool SuccIsHeader = LoopHeaders.count(SuccBB);
2349 dbgs() << " Not threading across "
2350 << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName()
2351 << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '")
2352 << SuccBB->getName() << "' - it might create an irreducible loop!\n";
2353 });
2354 return false;
2355 }
2356
2357 unsigned JumpThreadCost = getJumpThreadDuplicationCost(
2358 TTI, BB, StopAt: BB->getTerminator(), Threshold: BBDupThreshold);
2359 if (JumpThreadCost > BBDupThreshold) {
2360 LLVM_DEBUG(dbgs() << " Not threading BB '" << BB->getName()
2361 << "' - Cost is too high: " << JumpThreadCost << "\n");
2362 return false;
2363 }
2364
2365 threadEdge(BB, PredBBs, SuccBB);
2366 return true;
2367}
2368
2369/// threadEdge - We have decided that it is safe and profitable to factor the
2370/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
2371/// across BB. Transform the IR to reflect this change.
2372void JumpThreadingPass::threadEdge(BasicBlock *BB,
2373 const SmallVectorImpl<BasicBlock *> &PredBBs,
2374 BasicBlock *SuccBB) {
2375 assert(SuccBB != BB && "Don't create an infinite loop");
2376
2377 assert(!LoopHeaders.count(BB) && !LoopHeaders.count(SuccBB) &&
2378 "Don't thread across loop headers");
2379
2380 // Build BPI/BFI before any changes are made to IR.
2381 bool HasProfile = doesBlockHaveProfileData(BB);
2382 auto *BFI = getOrCreateBFI(Force: HasProfile);
2383 auto *BPI = getOrCreateBPI(Force: BFI != nullptr);
2384
2385 // And finally, do it! Start by factoring the predecessors if needed.
2386 BasicBlock *PredBB;
2387 if (PredBBs.size() == 1)
2388 PredBB = PredBBs[0];
2389 else {
2390 LLVM_DEBUG(dbgs() << " Factoring out " << PredBBs.size()
2391 << " common predecessors.\n");
2392 PredBB = splitBlockPreds(BB, Preds: PredBBs, Suffix: ".thr_comm");
2393 }
2394
2395 // And finally, do it!
2396 LLVM_DEBUG(dbgs() << " Threading edge from '" << PredBB->getName()
2397 << "' to '" << SuccBB->getName()
2398 << ", across block:\n " << *BB << "\n");
2399
2400 LVI->threadEdge(PredBB, OldSucc: BB, NewSucc: SuccBB);
2401
2402 BasicBlock *NewBB = BasicBlock::Create(Context&: BB->getContext(),
2403 Name: BB->getName()+".thread",
2404 Parent: BB->getParent(), InsertBefore: BB);
2405 NewBB->moveAfter(MovePos: PredBB);
2406
2407 // Set the block frequency of NewBB.
2408 if (BFI) {
2409 assert(BPI && "It's expected BPI to exist along with BFI");
2410 auto NewBBFreq =
2411 BFI->getBlockFreq(BB: PredBB) * BPI->getEdgeProbability(Src: PredBB, Dst: BB);
2412 BFI->setBlockFreq(BB: NewBB, Freq: NewBBFreq);
2413 }
2414
2415 // Copy all the instructions from BB to NewBB except the terminator.
2416 ValueToValueMapTy ValueMapping;
2417 cloneInstructions(ValueMapping, BI: BB->begin(), BE: std::prev(x: BB->end()), NewBB,
2418 PredBB);
2419
2420 // We didn't copy the terminator from BB over to NewBB, because there is now
2421 // an unconditional jump to SuccBB. Insert the unconditional jump.
2422 BranchInst *NewBI = BranchInst::Create(IfTrue: SuccBB, InsertBefore: NewBB);
2423 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
2424
2425 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
2426 // PHI nodes for NewBB now.
2427 addPHINodeEntriesForMappedBlock(PHIBB: SuccBB, OldPred: BB, NewPred: NewBB, ValueMap&: ValueMapping);
2428
2429 // Update the terminator of PredBB to jump to NewBB instead of BB. This
2430 // eliminates predecessors from BB, which requires us to simplify any PHI
2431 // nodes in BB.
2432 Instruction *PredTerm = PredBB->getTerminator();
2433 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
2434 if (PredTerm->getSuccessor(Idx: i) == BB) {
2435 BB->removePredecessor(Pred: PredBB, KeepOneInputPHIs: true);
2436 PredTerm->setSuccessor(Idx: i, BB: NewBB);
2437 }
2438
2439 // Enqueue required DT updates.
2440 DTU->applyUpdatesPermissive(Updates: {{DominatorTree::Insert, NewBB, SuccBB},
2441 {DominatorTree::Insert, PredBB, NewBB},
2442 {DominatorTree::Delete, PredBB, BB}});
2443
2444 updateSSA(BB, NewBB, ValueMapping);
2445
2446 // At this point, the IR is fully up to date and consistent. Do a quick scan
2447 // over the new instructions and zap any that are constants or dead. This
2448 // frequently happens because of phi translation.
2449 SimplifyInstructionsInBlock(BB: NewBB, TLI);
2450
2451 // Update the edge weight from BB to SuccBB, which should be less than before.
2452 updateBlockFreqAndEdgeWeight(PredBB, BB, NewBB, SuccBB, BFI, BPI, HasProfile);
2453
2454 // Threaded an edge!
2455 ++NumThreads;
2456}
2457
2458/// Create a new basic block that will be the predecessor of BB and successor of
2459/// all blocks in Preds. When profile data is available, update the frequency of
2460/// this new block.
2461BasicBlock *JumpThreadingPass::splitBlockPreds(BasicBlock *BB,
2462 ArrayRef<BasicBlock *> Preds,
2463 const char *Suffix) {
2464 SmallVector<BasicBlock *, 2> NewBBs;
2465
2466 // Collect the frequencies of all predecessors of BB, which will be used to
2467 // update the edge weight of the result of splitting predecessors.
2468 DenseMap<BasicBlock *, BlockFrequency> FreqMap;
2469 auto *BFI = getBFI();
2470 if (BFI) {
2471 auto *BPI = getOrCreateBPI(Force: true);
2472 for (auto *Pred : Preds)
2473 FreqMap.insert(KV: std::make_pair(
2474 x&: Pred, y: BFI->getBlockFreq(BB: Pred) * BPI->getEdgeProbability(Src: Pred, Dst: BB)));
2475 }
2476
2477 // In the case when BB is a LandingPad block we create 2 new predecessors
2478 // instead of just one.
2479 if (BB->isLandingPad()) {
2480 std::string NewName = std::string(Suffix) + ".split-lp";
2481 SplitLandingPadPredecessors(OrigBB: BB, Preds, Suffix, Suffix2: NewName.c_str(), NewBBs);
2482 } else {
2483 NewBBs.push_back(Elt: SplitBlockPredecessors(BB, Preds, Suffix));
2484 }
2485
2486 std::vector<DominatorTree::UpdateType> Updates;
2487 Updates.reserve(n: (2 * Preds.size()) + NewBBs.size());
2488 for (auto *NewBB : NewBBs) {
2489 BlockFrequency NewBBFreq(0);
2490 Updates.push_back(x: {DominatorTree::Insert, NewBB, BB});
2491 for (auto *Pred : predecessors(BB: NewBB)) {
2492 Updates.push_back(x: {DominatorTree::Delete, Pred, BB});
2493 Updates.push_back(x: {DominatorTree::Insert, Pred, NewBB});
2494 if (BFI) // Update frequencies between Pred -> NewBB.
2495 NewBBFreq += FreqMap.lookup(Val: Pred);
2496 }
2497 if (BFI) // Apply the summed frequency to NewBB.
2498 BFI->setBlockFreq(BB: NewBB, Freq: NewBBFreq);
2499 }
2500
2501 DTU->applyUpdatesPermissive(Updates);
2502 return NewBBs[0];
2503}
2504
2505bool JumpThreadingPass::doesBlockHaveProfileData(BasicBlock *BB) {
2506 const Instruction *TI = BB->getTerminator();
2507 if (!TI || TI->getNumSuccessors() < 2)
2508 return false;
2509
2510 return hasValidBranchWeightMD(I: *TI);
2511}
2512
2513/// Update the block frequency of BB and branch weight and the metadata on the
2514/// edge BB->SuccBB. This is done by scaling the weight of BB->SuccBB by 1 -
2515/// Freq(PredBB->BB) / Freq(BB->SuccBB).
2516void JumpThreadingPass::updateBlockFreqAndEdgeWeight(BasicBlock *PredBB,
2517 BasicBlock *BB,
2518 BasicBlock *NewBB,
2519 BasicBlock *SuccBB,
2520 BlockFrequencyInfo *BFI,
2521 BranchProbabilityInfo *BPI,
2522 bool HasProfile) {
2523 assert(((BFI && BPI) || (!BFI && !BFI)) &&
2524 "Both BFI & BPI should either be set or unset");
2525
2526 if (!BFI) {
2527 assert(!HasProfile &&
2528 "It's expected to have BFI/BPI when profile info exists");
2529 return;
2530 }
2531
2532 // As the edge from PredBB to BB is deleted, we have to update the block
2533 // frequency of BB.
2534 auto BBOrigFreq = BFI->getBlockFreq(BB);
2535 auto NewBBFreq = BFI->getBlockFreq(BB: NewBB);
2536 auto BB2SuccBBFreq = BBOrigFreq * BPI->getEdgeProbability(Src: BB, Dst: SuccBB);
2537 auto BBNewFreq = BBOrigFreq - NewBBFreq;
2538 BFI->setBlockFreq(BB, Freq: BBNewFreq);
2539
2540 // Collect updated outgoing edges' frequencies from BB and use them to update
2541 // edge probabilities.
2542 SmallVector<uint64_t, 4> BBSuccFreq;
2543 for (BasicBlock *Succ : successors(BB)) {
2544 auto SuccFreq = (Succ == SuccBB)
2545 ? BB2SuccBBFreq - NewBBFreq
2546 : BBOrigFreq * BPI->getEdgeProbability(Src: BB, Dst: Succ);
2547 BBSuccFreq.push_back(Elt: SuccFreq.getFrequency());
2548 }
2549
2550 uint64_t MaxBBSuccFreq = *llvm::max_element(Range&: BBSuccFreq);
2551
2552 SmallVector<BranchProbability, 4> BBSuccProbs;
2553 if (MaxBBSuccFreq == 0)
2554 BBSuccProbs.assign(NumElts: BBSuccFreq.size(),
2555 Elt: {1, static_cast<unsigned>(BBSuccFreq.size())});
2556 else {
2557 for (uint64_t Freq : BBSuccFreq)
2558 BBSuccProbs.push_back(
2559 Elt: BranchProbability::getBranchProbability(Numerator: Freq, Denominator: MaxBBSuccFreq));
2560 // Normalize edge probabilities so that they sum up to one.
2561 BranchProbability::normalizeProbabilities(Begin: BBSuccProbs.begin(),
2562 End: BBSuccProbs.end());
2563 }
2564
2565 // Update edge probabilities in BPI.
2566 BPI->setEdgeProbability(Src: BB, Probs: BBSuccProbs);
2567
2568 // Update the profile metadata as well.
2569 //
2570 // Don't do this if the profile of the transformed blocks was statically
2571 // estimated. (This could occur despite the function having an entry
2572 // frequency in completely cold parts of the CFG.)
2573 //
2574 // In this case we don't want to suggest to subsequent passes that the
2575 // calculated weights are fully consistent. Consider this graph:
2576 //
2577 // check_1
2578 // 50% / |
2579 // eq_1 | 50%
2580 // \ |
2581 // check_2
2582 // 50% / |
2583 // eq_2 | 50%
2584 // \ |
2585 // check_3
2586 // 50% / |
2587 // eq_3 | 50%
2588 // \ |
2589 //
2590 // Assuming the blocks check_* all compare the same value against 1, 2 and 3,
2591 // the overall probabilities are inconsistent; the total probability that the
2592 // value is either 1, 2 or 3 is 150%.
2593 //
2594 // As a consequence if we thread eq_1 -> check_2 to check_3, check_2->check_3
2595 // becomes 0%. This is even worse if the edge whose probability becomes 0% is
2596 // the loop exit edge. Then based solely on static estimation we would assume
2597 // the loop was extremely hot.
2598 //
2599 // FIXME this locally as well so that BPI and BFI are consistent as well. We
2600 // shouldn't make edges extremely likely or unlikely based solely on static
2601 // estimation.
2602 if (BBSuccProbs.size() >= 2 && HasProfile) {
2603 SmallVector<uint32_t, 4> Weights;
2604 for (auto Prob : BBSuccProbs)
2605 Weights.push_back(Elt: Prob.getNumerator());
2606
2607 auto TI = BB->getTerminator();
2608 setBranchWeights(I&: *TI, Weights, IsExpected: hasBranchWeightOrigin(I: *TI));
2609 }
2610}
2611
2612/// duplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
2613/// to BB which contains an i1 PHI node and a conditional branch on that PHI.
2614/// If we can duplicate the contents of BB up into PredBB do so now, this
2615/// improves the odds that the branch will be on an analyzable instruction like
2616/// a compare.
2617bool JumpThreadingPass::duplicateCondBranchOnPHIIntoPred(
2618 BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs) {
2619 assert(!PredBBs.empty() && "Can't handle an empty set");
2620
2621 // If BB is a loop header, then duplicating this block outside the loop would
2622 // cause us to transform this into an irreducible loop, don't do this.
2623 // See the comments above findLoopHeaders for justifications and caveats.
2624 if (LoopHeaders.count(V: BB)) {
2625 LLVM_DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
2626 << "' into predecessor block '" << PredBBs[0]->getName()
2627 << "' - it might create an irreducible loop!\n");
2628 return false;
2629 }
2630
2631 unsigned DuplicationCost = getJumpThreadDuplicationCost(
2632 TTI, BB, StopAt: BB->getTerminator(), Threshold: BBDupThreshold);
2633 if (DuplicationCost > BBDupThreshold) {
2634 LLVM_DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
2635 << "' - Cost is too high: " << DuplicationCost << "\n");
2636 return false;
2637 }
2638
2639 // And finally, do it! Start by factoring the predecessors if needed.
2640 std::vector<DominatorTree::UpdateType> Updates;
2641 BasicBlock *PredBB;
2642 if (PredBBs.size() == 1)
2643 PredBB = PredBBs[0];
2644 else {
2645 LLVM_DEBUG(dbgs() << " Factoring out " << PredBBs.size()
2646 << " common predecessors.\n");
2647 PredBB = splitBlockPreds(BB, Preds: PredBBs, Suffix: ".thr_comm");
2648 }
2649 Updates.push_back(x: {DominatorTree::Delete, PredBB, BB});
2650
2651 // Okay, we decided to do this! Clone all the instructions in BB onto the end
2652 // of PredBB.
2653 LLVM_DEBUG(dbgs() << " Duplicating block '" << BB->getName()
2654 << "' into end of '" << PredBB->getName()
2655 << "' to eliminate branch on phi. Cost: "
2656 << DuplicationCost << " block is:" << *BB << "\n");
2657
2658 // Unless PredBB ends with an unconditional branch, split the edge so that we
2659 // can just clone the bits from BB into the end of the new PredBB.
2660 BranchInst *OldPredBranch = dyn_cast<BranchInst>(Val: PredBB->getTerminator());
2661
2662 if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
2663 BasicBlock *OldPredBB = PredBB;
2664 PredBB = SplitEdge(From: OldPredBB, To: BB);
2665 Updates.push_back(x: {DominatorTree::Insert, OldPredBB, PredBB});
2666 Updates.push_back(x: {DominatorTree::Insert, PredBB, BB});
2667 Updates.push_back(x: {DominatorTree::Delete, OldPredBB, BB});
2668 OldPredBranch = cast<BranchInst>(Val: PredBB->getTerminator());
2669 }
2670
2671 // We are going to have to map operands from the original BB block into the
2672 // PredBB block. Evaluate PHI nodes in BB.
2673 ValueToValueMapTy ValueMapping;
2674
2675 BasicBlock::iterator BI = BB->begin();
2676 for (; PHINode *PN = dyn_cast<PHINode>(Val&: BI); ++BI)
2677 ValueMapping[PN] = PN->getIncomingValueForBlock(BB: PredBB);
2678 // Clone the non-phi instructions of BB into PredBB, keeping track of the
2679 // mapping and using it to remap operands in the cloned instructions.
2680 for (; BI != BB->end(); ++BI) {
2681 Instruction *New = BI->clone();
2682 New->insertInto(ParentBB: PredBB, It: OldPredBranch->getIterator());
2683
2684 // Remap operands to patch up intra-block references.
2685 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2686 if (Instruction *Inst = dyn_cast<Instruction>(Val: New->getOperand(i))) {
2687 ValueToValueMapTy::iterator I = ValueMapping.find(Val: Inst);
2688 if (I != ValueMapping.end())
2689 New->setOperand(i, Val: I->second);
2690 }
2691
2692 // Remap debug variable operands.
2693 remapDebugVariable(Mapping&: ValueMapping, Inst: New);
2694
2695 // If this instruction can be simplified after the operands are updated,
2696 // just use the simplified value instead. This frequently happens due to
2697 // phi translation.
2698 if (Value *IV = simplifyInstruction(
2699 I: New,
2700 Q: {BB->getDataLayout(), TLI, nullptr, nullptr, New})) {
2701 ValueMapping[&*BI] = IV;
2702 if (!New->mayHaveSideEffects()) {
2703 New->eraseFromParent();
2704 New = nullptr;
2705 // Clone debug-info on the elided instruction to the destination
2706 // position.
2707 OldPredBranch->cloneDebugInfoFrom(From: &*BI, FromHere: std::nullopt, InsertAtHead: true);
2708 }
2709 } else {
2710 ValueMapping[&*BI] = New;
2711 }
2712 if (New) {
2713 // Otherwise, insert the new instruction into the block.
2714 New->setName(BI->getName());
2715 // Clone across any debug-info attached to the old instruction.
2716 New->cloneDebugInfoFrom(From: &*BI);
2717 // Update Dominance from simplified New instruction operands.
2718 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2719 if (BasicBlock *SuccBB = dyn_cast<BasicBlock>(Val: New->getOperand(i)))
2720 Updates.push_back(x: {DominatorTree::Insert, PredBB, SuccBB});
2721 }
2722 }
2723
2724 // Check to see if the targets of the branch had PHI nodes. If so, we need to
2725 // add entries to the PHI nodes for branch from PredBB now.
2726 BranchInst *BBBranch = cast<BranchInst>(Val: BB->getTerminator());
2727 addPHINodeEntriesForMappedBlock(PHIBB: BBBranch->getSuccessor(i: 0), OldPred: BB, NewPred: PredBB,
2728 ValueMap&: ValueMapping);
2729 addPHINodeEntriesForMappedBlock(PHIBB: BBBranch->getSuccessor(i: 1), OldPred: BB, NewPred: PredBB,
2730 ValueMap&: ValueMapping);
2731
2732 updateSSA(BB, NewBB: PredBB, ValueMapping);
2733
2734 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
2735 // that we nuked.
2736 BB->removePredecessor(Pred: PredBB, KeepOneInputPHIs: true);
2737
2738 // Remove the unconditional branch at the end of the PredBB block.
2739 OldPredBranch->eraseFromParent();
2740 if (auto *BPI = getBPI())
2741 BPI->copyEdgeProbabilities(Src: BB, Dst: PredBB);
2742 DTU->applyUpdatesPermissive(Updates);
2743
2744 ++NumDupes;
2745 return true;
2746}
2747
2748// Pred is a predecessor of BB with an unconditional branch to BB. SI is
2749// a Select instruction in Pred. BB has other predecessors and SI is used in
2750// a PHI node in BB. SI has no other use.
2751// A new basic block, NewBB, is created and SI is converted to compare and
2752// conditional branch. SI is erased from parent.
2753void JumpThreadingPass::unfoldSelectInstr(BasicBlock *Pred, BasicBlock *BB,
2754 SelectInst *SI, PHINode *SIUse,
2755 unsigned Idx) {
2756 // Expand the select.
2757 //
2758 // Pred --
2759 // | v
2760 // | NewBB
2761 // | |
2762 // |-----
2763 // v
2764 // BB
2765 BranchInst *PredTerm = cast<BranchInst>(Val: Pred->getTerminator());
2766 BasicBlock *NewBB = BasicBlock::Create(Context&: BB->getContext(), Name: "select.unfold",
2767 Parent: BB->getParent(), InsertBefore: BB);
2768 // Move the unconditional branch to NewBB.
2769 PredTerm->removeFromParent();
2770 PredTerm->insertInto(ParentBB: NewBB, It: NewBB->end());
2771 // Create a conditional branch and update PHI nodes.
2772 auto *BI = BranchInst::Create(IfTrue: NewBB, IfFalse: BB, Cond: SI->getCondition(), InsertBefore: Pred);
2773 BI->applyMergedLocation(LocA: PredTerm->getDebugLoc(), LocB: SI->getDebugLoc());
2774 BI->copyMetadata(SrcInst: *SI, WL: {LLVMContext::MD_prof});
2775 SIUse->setIncomingValue(i: Idx, V: SI->getFalseValue());
2776 SIUse->addIncoming(V: SI->getTrueValue(), BB: NewBB);
2777
2778 uint64_t TrueWeight = 1;
2779 uint64_t FalseWeight = 1;
2780 // Copy probabilities from 'SI' to created conditional branch in 'Pred'.
2781 if (extractBranchWeights(I: *SI, TrueVal&: TrueWeight, FalseVal&: FalseWeight) &&
2782 (TrueWeight + FalseWeight) != 0) {
2783 SmallVector<BranchProbability, 2> BP;
2784 BP.emplace_back(Args: BranchProbability::getBranchProbability(
2785 Numerator: TrueWeight, Denominator: TrueWeight + FalseWeight));
2786 BP.emplace_back(Args: BranchProbability::getBranchProbability(
2787 Numerator: FalseWeight, Denominator: TrueWeight + FalseWeight));
2788 // Update BPI if exists.
2789 if (auto *BPI = getBPI())
2790 BPI->setEdgeProbability(Src: Pred, Probs: BP);
2791 }
2792 // Set the block frequency of NewBB.
2793 if (auto *BFI = getBFI()) {
2794 if ((TrueWeight + FalseWeight) == 0) {
2795 TrueWeight = 1;
2796 FalseWeight = 1;
2797 }
2798 BranchProbability PredToNewBBProb = BranchProbability::getBranchProbability(
2799 Numerator: TrueWeight, Denominator: TrueWeight + FalseWeight);
2800 auto NewBBFreq = BFI->getBlockFreq(BB: Pred) * PredToNewBBProb;
2801 BFI->setBlockFreq(BB: NewBB, Freq: NewBBFreq);
2802 }
2803
2804 // The select is now dead.
2805 SI->eraseFromParent();
2806 DTU->applyUpdatesPermissive(Updates: {{DominatorTree::Insert, NewBB, BB},
2807 {DominatorTree::Insert, Pred, NewBB}});
2808
2809 // Update any other PHI nodes in BB.
2810 for (BasicBlock::iterator BI = BB->begin();
2811 PHINode *Phi = dyn_cast<PHINode>(Val&: BI); ++BI)
2812 if (Phi != SIUse)
2813 Phi->addIncoming(V: Phi->getIncomingValueForBlock(BB: Pred), BB: NewBB);
2814}
2815
2816bool JumpThreadingPass::tryToUnfoldSelect(SwitchInst *SI, BasicBlock *BB) {
2817 PHINode *CondPHI = dyn_cast<PHINode>(Val: SI->getCondition());
2818
2819 if (!CondPHI || CondPHI->getParent() != BB)
2820 return false;
2821
2822 for (unsigned I = 0, E = CondPHI->getNumIncomingValues(); I != E; ++I) {
2823 BasicBlock *Pred = CondPHI->getIncomingBlock(i: I);
2824 SelectInst *PredSI = dyn_cast<SelectInst>(Val: CondPHI->getIncomingValue(i: I));
2825
2826 // The second and third condition can be potentially relaxed. Currently
2827 // the conditions help to simplify the code and allow us to reuse existing
2828 // code, developed for tryToUnfoldSelect(CmpInst *, BasicBlock *)
2829 if (!PredSI || PredSI->getParent() != Pred || !PredSI->hasOneUse())
2830 continue;
2831
2832 BranchInst *PredTerm = dyn_cast<BranchInst>(Val: Pred->getTerminator());
2833 if (!PredTerm || !PredTerm->isUnconditional())
2834 continue;
2835
2836 unfoldSelectInstr(Pred, BB, SI: PredSI, SIUse: CondPHI, Idx: I);
2837 return true;
2838 }
2839 return false;
2840}
2841
2842/// tryToUnfoldSelect - Look for blocks of the form
2843/// bb1:
2844/// %a = select
2845/// br bb2
2846///
2847/// bb2:
2848/// %p = phi [%a, %bb1] ...
2849/// %c = icmp %p
2850/// br i1 %c
2851///
2852/// And expand the select into a branch structure if one of its arms allows %c
2853/// to be folded. This later enables threading from bb1 over bb2.
2854bool JumpThreadingPass::tryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
2855 BranchInst *CondBr = dyn_cast<BranchInst>(Val: BB->getTerminator());
2856 PHINode *CondLHS = dyn_cast<PHINode>(Val: CondCmp->getOperand(i_nocapture: 0));
2857 Constant *CondRHS = cast<Constant>(Val: CondCmp->getOperand(i_nocapture: 1));
2858
2859 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
2860 CondLHS->getParent() != BB)
2861 return false;
2862
2863 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
2864 BasicBlock *Pred = CondLHS->getIncomingBlock(i: I);
2865 SelectInst *SI = dyn_cast<SelectInst>(Val: CondLHS->getIncomingValue(i: I));
2866
2867 // Look if one of the incoming values is a select in the corresponding
2868 // predecessor.
2869 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
2870 continue;
2871
2872 BranchInst *PredTerm = dyn_cast<BranchInst>(Val: Pred->getTerminator());
2873 if (!PredTerm || !PredTerm->isUnconditional())
2874 continue;
2875
2876 // Now check if one of the select values would allow us to constant fold the
2877 // terminator in BB. We don't do the transform if both sides fold, those
2878 // cases will be threaded in any case.
2879 Constant *LHSRes =
2880 LVI->getPredicateOnEdge(Pred: CondCmp->getPredicate(), V: SI->getOperand(i_nocapture: 1),
2881 C: CondRHS, FromBB: Pred, ToBB: BB, CxtI: CondCmp);
2882 Constant *RHSRes =
2883 LVI->getPredicateOnEdge(Pred: CondCmp->getPredicate(), V: SI->getOperand(i_nocapture: 2),
2884 C: CondRHS, FromBB: Pred, ToBB: BB, CxtI: CondCmp);
2885 if ((LHSRes || RHSRes) && LHSRes != RHSRes) {
2886 unfoldSelectInstr(Pred, BB, SI, SIUse: CondLHS, Idx: I);
2887 return true;
2888 }
2889 }
2890 return false;
2891}
2892
2893/// tryToUnfoldSelectInCurrBB - Look for PHI/Select or PHI/CMP/Select in the
2894/// same BB in the form
2895/// bb:
2896/// %p = phi [false, %bb1], [true, %bb2], [false, %bb3], [true, %bb4], ...
2897/// %s = select %p, trueval, falseval
2898///
2899/// or
2900///
2901/// bb:
2902/// %p = phi [0, %bb1], [1, %bb2], [0, %bb3], [1, %bb4], ...
2903/// %c = cmp %p, 0
2904/// %s = select %c, trueval, falseval
2905///
2906/// And expand the select into a branch structure. This later enables
2907/// jump-threading over bb in this pass.
2908///
2909/// Using the similar approach of SimplifyCFG::FoldCondBranchOnPHI(), unfold
2910/// select if the associated PHI has at least one constant. If the unfolded
2911/// select is not jump-threaded, it will be folded again in the later
2912/// optimizations.
2913bool JumpThreadingPass::tryToUnfoldSelectInCurrBB(BasicBlock *BB) {
2914 // This transform would reduce the quality of msan diagnostics.
2915 // Disable this transform under MemorySanitizer.
2916 if (BB->getParent()->hasFnAttribute(Kind: Attribute::SanitizeMemory))
2917 return false;
2918
2919 // If threading this would thread across a loop header, don't thread the edge.
2920 // See the comments above findLoopHeaders for justifications and caveats.
2921 if (LoopHeaders.count(V: BB))
2922 return false;
2923
2924 for (BasicBlock::iterator BI = BB->begin();
2925 PHINode *PN = dyn_cast<PHINode>(Val&: BI); ++BI) {
2926 // Look for a Phi having at least one constant incoming value.
2927 if (llvm::all_of(Range: PN->incoming_values(),
2928 P: [](Value *V) { return !isa<ConstantInt>(Val: V); }))
2929 continue;
2930
2931 auto isUnfoldCandidate = [BB](SelectInst *SI, Value *V) {
2932 using namespace PatternMatch;
2933
2934 // Check if SI is in BB and use V as condition.
2935 if (SI->getParent() != BB)
2936 return false;
2937 Value *Cond = SI->getCondition();
2938 bool IsAndOr = match(V: SI, P: m_CombineOr(L: m_LogicalAnd(), R: m_LogicalOr()));
2939 return Cond && Cond == V && Cond->getType()->isIntegerTy(Bitwidth: 1) && !IsAndOr;
2940 };
2941
2942 SelectInst *SI = nullptr;
2943 for (Use &U : PN->uses()) {
2944 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(Val: U.getUser())) {
2945 // Look for a ICmp in BB that compares PN with a constant and is the
2946 // condition of a Select.
2947 if (Cmp->getParent() == BB && Cmp->hasOneUse() &&
2948 isa<ConstantInt>(Val: Cmp->getOperand(i_nocapture: 1 - U.getOperandNo())))
2949 if (SelectInst *SelectI = dyn_cast<SelectInst>(Val: Cmp->user_back()))
2950 if (isUnfoldCandidate(SelectI, Cmp->use_begin()->get())) {
2951 SI = SelectI;
2952 break;
2953 }
2954 } else if (SelectInst *SelectI = dyn_cast<SelectInst>(Val: U.getUser())) {
2955 // Look for a Select in BB that uses PN as condition.
2956 if (isUnfoldCandidate(SelectI, U.get())) {
2957 SI = SelectI;
2958 break;
2959 }
2960 }
2961 }
2962
2963 if (!SI)
2964 continue;
2965 // Expand the select.
2966 Value *Cond = SI->getCondition();
2967 if (!isGuaranteedNotToBeUndefOrPoison(V: Cond, AC: nullptr, CtxI: SI))
2968 Cond = new FreezeInst(Cond, "cond.fr", SI->getIterator());
2969 MDNode *BranchWeights = getBranchWeightMDNode(I: *SI);
2970 Instruction *Term =
2971 SplitBlockAndInsertIfThen(Cond, SplitBefore: SI, Unreachable: false, BranchWeights);
2972 BasicBlock *SplitBB = SI->getParent();
2973 BasicBlock *NewBB = Term->getParent();
2974 PHINode *NewPN = PHINode::Create(Ty: SI->getType(), NumReservedValues: 2, NameStr: "", InsertBefore: SI->getIterator());
2975 NewPN->addIncoming(V: SI->getTrueValue(), BB: Term->getParent());
2976 NewPN->addIncoming(V: SI->getFalseValue(), BB);
2977 NewPN->setDebugLoc(SI->getDebugLoc());
2978 SI->replaceAllUsesWith(V: NewPN);
2979 SI->eraseFromParent();
2980 // NewBB and SplitBB are newly created blocks which require insertion.
2981 std::vector<DominatorTree::UpdateType> Updates;
2982 Updates.reserve(n: (2 * SplitBB->getTerminator()->getNumSuccessors()) + 3);
2983 Updates.push_back(x: {DominatorTree::Insert, BB, SplitBB});
2984 Updates.push_back(x: {DominatorTree::Insert, BB, NewBB});
2985 Updates.push_back(x: {DominatorTree::Insert, NewBB, SplitBB});
2986 // BB's successors were moved to SplitBB, update DTU accordingly.
2987 for (auto *Succ : successors(BB: SplitBB)) {
2988 Updates.push_back(x: {DominatorTree::Delete, BB, Succ});
2989 Updates.push_back(x: {DominatorTree::Insert, SplitBB, Succ});
2990 }
2991 DTU->applyUpdatesPermissive(Updates);
2992 return true;
2993 }
2994 return false;
2995}
2996
2997/// Try to propagate a guard from the current BB into one of its predecessors
2998/// in case if another branch of execution implies that the condition of this
2999/// guard is always true. Currently we only process the simplest case that
3000/// looks like:
3001///
3002/// Start:
3003/// %cond = ...
3004/// br i1 %cond, label %T1, label %F1
3005/// T1:
3006/// br label %Merge
3007/// F1:
3008/// br label %Merge
3009/// Merge:
3010/// %condGuard = ...
3011/// call void(i1, ...) @llvm.experimental.guard( i1 %condGuard )[ "deopt"() ]
3012///
3013/// And cond either implies condGuard or !condGuard. In this case all the
3014/// instructions before the guard can be duplicated in both branches, and the
3015/// guard is then threaded to one of them.
3016bool JumpThreadingPass::processGuards(BasicBlock *BB) {
3017 using namespace PatternMatch;
3018
3019 // We only want to deal with two predecessors.
3020 BasicBlock *Pred1, *Pred2;
3021 auto PI = pred_begin(BB), PE = pred_end(BB);
3022 if (PI == PE)
3023 return false;
3024 Pred1 = *PI++;
3025 if (PI == PE)
3026 return false;
3027 Pred2 = *PI++;
3028 if (PI != PE)
3029 return false;
3030 if (Pred1 == Pred2)
3031 return false;
3032
3033 // Try to thread one of the guards of the block.
3034 // TODO: Look up deeper than to immediate predecessor?
3035 auto *Parent = Pred1->getSinglePredecessor();
3036 if (!Parent || Parent != Pred2->getSinglePredecessor())
3037 return false;
3038
3039 if (auto *BI = dyn_cast<BranchInst>(Val: Parent->getTerminator()))
3040 for (auto &I : *BB)
3041 if (isGuard(U: &I) && threadGuard(BB, Guard: cast<IntrinsicInst>(Val: &I), BI))
3042 return true;
3043
3044 return false;
3045}
3046
3047/// Try to propagate the guard from BB which is the lower block of a diamond
3048/// to one of its branches, in case if diamond's condition implies guard's
3049/// condition.
3050bool JumpThreadingPass::threadGuard(BasicBlock *BB, IntrinsicInst *Guard,
3051 BranchInst *BI) {
3052 assert(BI->getNumSuccessors() == 2 && "Wrong number of successors?");
3053 assert(BI->isConditional() && "Unconditional branch has 2 successors?");
3054 Value *GuardCond = Guard->getArgOperand(i: 0);
3055 Value *BranchCond = BI->getCondition();
3056 BasicBlock *TrueDest = BI->getSuccessor(i: 0);
3057 BasicBlock *FalseDest = BI->getSuccessor(i: 1);
3058
3059 auto &DL = BB->getDataLayout();
3060 bool TrueDestIsSafe = false;
3061 bool FalseDestIsSafe = false;
3062
3063 // True dest is safe if BranchCond => GuardCond.
3064 auto Impl = isImpliedCondition(LHS: BranchCond, RHS: GuardCond, DL);
3065 if (Impl && *Impl)
3066 TrueDestIsSafe = true;
3067 else {
3068 // False dest is safe if !BranchCond => GuardCond.
3069 Impl = isImpliedCondition(LHS: BranchCond, RHS: GuardCond, DL, /* LHSIsTrue */ false);
3070 if (Impl && *Impl)
3071 FalseDestIsSafe = true;
3072 }
3073
3074 if (!TrueDestIsSafe && !FalseDestIsSafe)
3075 return false;
3076
3077 BasicBlock *PredUnguardedBlock = TrueDestIsSafe ? TrueDest : FalseDest;
3078 BasicBlock *PredGuardedBlock = FalseDestIsSafe ? TrueDest : FalseDest;
3079
3080 ValueToValueMapTy UnguardedMapping, GuardedMapping;
3081 Instruction *AfterGuard = Guard->getNextNode();
3082 unsigned Cost =
3083 getJumpThreadDuplicationCost(TTI, BB, StopAt: AfterGuard, Threshold: BBDupThreshold);
3084 if (Cost > BBDupThreshold)
3085 return false;
3086 // Duplicate all instructions before the guard and the guard itself to the
3087 // branch where implication is not proved.
3088 BasicBlock *GuardedBlock = DuplicateInstructionsInSplitBetween(
3089 BB, PredBB: PredGuardedBlock, StopAt: AfterGuard, ValueMapping&: GuardedMapping, DTU&: *DTU);
3090 assert(GuardedBlock && "Could not create the guarded block?");
3091 // Duplicate all instructions before the guard in the unguarded branch.
3092 // Since we have successfully duplicated the guarded block and this block
3093 // has fewer instructions, we expect it to succeed.
3094 BasicBlock *UnguardedBlock = DuplicateInstructionsInSplitBetween(
3095 BB, PredBB: PredUnguardedBlock, StopAt: Guard, ValueMapping&: UnguardedMapping, DTU&: *DTU);
3096 assert(UnguardedBlock && "Could not create the unguarded block?");
3097 LLVM_DEBUG(dbgs() << "Moved guard " << *Guard << " to block "
3098 << GuardedBlock->getName() << "\n");
3099 // Some instructions before the guard may still have uses. For them, we need
3100 // to create Phi nodes merging their copies in both guarded and unguarded
3101 // branches. Those instructions that have no uses can be just removed.
3102 SmallVector<Instruction *, 4> ToRemove;
3103 for (auto BI = BB->begin(); &*BI != AfterGuard; ++BI)
3104 if (!isa<PHINode>(Val: &*BI))
3105 ToRemove.push_back(Elt: &*BI);
3106
3107 BasicBlock::iterator InsertionPoint = BB->getFirstInsertionPt();
3108 assert(InsertionPoint != BB->end() && "Empty block?");
3109 // Substitute with Phis & remove.
3110 for (auto *Inst : reverse(C&: ToRemove)) {
3111 if (!Inst->use_empty()) {
3112 PHINode *NewPN = PHINode::Create(Ty: Inst->getType(), NumReservedValues: 2);
3113 NewPN->addIncoming(V: UnguardedMapping[Inst], BB: UnguardedBlock);
3114 NewPN->addIncoming(V: GuardedMapping[Inst], BB: GuardedBlock);
3115 NewPN->setDebugLoc(Inst->getDebugLoc());
3116 NewPN->insertBefore(InsertPos: InsertionPoint);
3117 Inst->replaceAllUsesWith(V: NewPN);
3118 }
3119 Inst->dropDbgRecords();
3120 Inst->eraseFromParent();
3121 }
3122 return true;
3123}
3124
3125PreservedAnalyses JumpThreadingPass::getPreservedAnalysis() const {
3126 PreservedAnalyses PA;
3127 PA.preserve<LazyValueAnalysis>();
3128 PA.preserve<DominatorTreeAnalysis>();
3129
3130 // TODO: We would like to preserve BPI/BFI. Enable once all paths update them.
3131 // TODO: Would be nice to verify BPI/BFI consistency as well.
3132 return PA;
3133}
3134
3135template <typename AnalysisT>
3136typename AnalysisT::Result *JumpThreadingPass::runExternalAnalysis() {
3137 assert(FAM && "Can't run external analysis without FunctionAnalysisManager");
3138
3139 // If there were no changes since last call to 'runExternalAnalysis' then all
3140 // analysis is either up to date or explicitly invalidated. Just go ahead and
3141 // run the "external" analysis.
3142 if (!ChangedSinceLastAnalysisUpdate) {
3143 assert(!DTU->hasPendingUpdates() &&
3144 "Lost update of 'ChangedSinceLastAnalysisUpdate'?");
3145 // Run the "external" analysis.
3146 return &FAM->getResult<AnalysisT>(*F);
3147 }
3148 ChangedSinceLastAnalysisUpdate = false;
3149
3150 auto PA = getPreservedAnalysis();
3151 // TODO: This shouldn't be needed once 'getPreservedAnalysis' reports BPI/BFI
3152 // as preserved.
3153 PA.preserve<BranchProbabilityAnalysis>();
3154 PA.preserve<BlockFrequencyAnalysis>();
3155 // Report everything except explicitly preserved as invalid.
3156 FAM->invalidate(IR&: *F, PA);
3157 // Update DT/PDT.
3158 DTU->flush();
3159 // Make sure DT/PDT are valid before running "external" analysis.
3160 assert(DTU->getDomTree().verify(DominatorTree::VerificationLevel::Fast));
3161 assert((!DTU->hasPostDomTree() ||
3162 DTU->getPostDomTree().verify(
3163 PostDominatorTree::VerificationLevel::Fast)));
3164 // Run the "external" analysis.
3165 auto *Result = &FAM->getResult<AnalysisT>(*F);
3166 // Update analysis JumpThreading depends on and not explicitly preserved.
3167 TTI = &FAM->getResult<TargetIRAnalysis>(IR&: *F);
3168 TLI = &FAM->getResult<TargetLibraryAnalysis>(IR&: *F);
3169 AA = &FAM->getResult<AAManager>(IR&: *F);
3170
3171 return Result;
3172}
3173
3174BranchProbabilityInfo *JumpThreadingPass::getBPI() {
3175 if (!BPI) {
3176 assert(FAM && "Can't create BPI without FunctionAnalysisManager");
3177 BPI = FAM->getCachedResult<BranchProbabilityAnalysis>(IR&: *F);
3178 }
3179 return *BPI;
3180}
3181
3182BlockFrequencyInfo *JumpThreadingPass::getBFI() {
3183 if (!BFI) {
3184 assert(FAM && "Can't create BFI without FunctionAnalysisManager");
3185 BFI = FAM->getCachedResult<BlockFrequencyAnalysis>(IR&: *F);
3186 }
3187 return *BFI;
3188}
3189
3190// Important note on validity of BPI/BFI. JumpThreading tries to preserve
3191// BPI/BFI as it goes. Thus if cached instance exists it will be updated.
3192// Otherwise, new instance of BPI/BFI is created (up to date by definition).
3193BranchProbabilityInfo *JumpThreadingPass::getOrCreateBPI(bool Force) {
3194 auto *Res = getBPI();
3195 if (Res)
3196 return Res;
3197
3198 if (Force)
3199 BPI = runExternalAnalysis<BranchProbabilityAnalysis>();
3200
3201 return *BPI;
3202}
3203
3204BlockFrequencyInfo *JumpThreadingPass::getOrCreateBFI(bool Force) {
3205 auto *Res = getBFI();
3206 if (Res)
3207 return Res;
3208
3209 if (Force)
3210 BFI = runExternalAnalysis<BlockFrequencyAnalysis>();
3211
3212 return *BFI;
3213}
3214