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
79	using namespace llvm;
80	using namespace jumpthreading;
81
82	#define DEBUG_TYPE "jump-threading"
83
84	STATISTIC(NumThreads, "Number of jumps threaded");
85	STATISTIC(NumFolds, "Number of terminators folded");
86	STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
87
88	static cl::opt<unsigned>
89	BBDuplicateThreshold("jump-threading-threshold",
90	cl::desc("Max block size to duplicate for jump threading"),
91	cl::init(Val: 6), cl::Hidden);
92
93	static cl::opt<unsigned>
94	ImplicationSearchThreshold(
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
100	static 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
105	static 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
110	JumpThreadingPass::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	//
149	static 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);
235	}
236	}
237
238	PreservedAnalyses 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
282	bool 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	Intrinsic::getName(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(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.
393	static 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.
427	static 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.
526	void 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.
539	static 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.
559	bool JumpThreadingPass::computeValueKnownInPredecessorsImpl(
560	Value *V, BasicBlock *BB, PredValueInfo &Result,
561	ConstantPreference Preference, DenseSet<Value *> &RecursionSet,
562	Instruction *CxtI) {
563	const DataLayout &DL = BB->getModule()->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(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	auto Res = LVI->getPredicateOnEdge(Pred, V: Val, C: Cst, FromBB: P, ToBB: BB, CxtI);
601	if (Res != LazyValueInfo::Unknown)
602	PredCst = ConstantInt::getBool(Context&: V->getContext(), V: Res);
603	}
604	if (Constant *KC = getKnownConstant(Val: PredCst, Preference))
605	Result.emplace_back(Args&: KC, Args&: P);
606	}
607
608	return !Result.empty();
609	}
610
611	/// If I is a PHI node, then we know the incoming values for any constants.
612	if (PHINode *PN = dyn_cast<PHINode>(Val: I)) {
613	for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
614	Value *InVal = PN->getIncomingValue(i);
615	if (Constant *KC = getKnownConstant(Val: InVal, Preference)) {
616	Result.emplace_back(Args&: KC, Args: PN->getIncomingBlock(i));
617	} else {
618	Constant *CI = LVI->getConstantOnEdge(V: InVal,
619	FromBB: PN->getIncomingBlock(i),
620	ToBB: BB, CxtI);
621	if (Constant *KC = getKnownConstant(Val: CI, Preference))
622	Result.emplace_back(Args&: KC, Args: PN->getIncomingBlock(i));
623	}
624	}
625
626	return !Result.empty();
627	}
628
629	// Handle Cast instructions.
630	if (CastInst *CI = dyn_cast<CastInst>(Val: I)) {
631	Value *Source = CI->getOperand(i_nocapture: 0);
632	PredValueInfoTy Vals;
633	computeValueKnownInPredecessorsImpl(V: Source, BB, Result&: Vals, Preference,
634	RecursionSet, CxtI);
635	if (Vals.empty())
636	return false;
637
638	// Convert the known values.
639	for (auto &Val : Vals)
640	if (Constant *Folded = ConstantFoldCastOperand(Opcode: CI->getOpcode(), C: Val.first,
641	DestTy: CI->getType(), DL))
642	Result.emplace_back(Args&: Folded, Args&: Val.second);
643
644	return !Result.empty();
645	}
646
647	if (FreezeInst *FI = dyn_cast<FreezeInst>(Val: I)) {
648	Value *Source = FI->getOperand(i_nocapture: 0);
649	computeValueKnownInPredecessorsImpl(V: Source, BB, Result, Preference,
650	RecursionSet, CxtI);
651
652	erase_if(C&: Result, P: [](auto &Pair) {
653	return !isGuaranteedNotToBeUndefOrPoison(Pair.first);
654	});
655
656	return !Result.empty();
657	}
658
659	// Handle some boolean conditions.
660	if (I->getType()->getPrimitiveSizeInBits() == 1) {
661	using namespace PatternMatch;
662	if (Preference != WantInteger)
663	return false;
664	// X | true -> true
665	// X & false -> false
666	Value *Op0, *Op1;
667	if (match(V: I, P: m_LogicalOr(L: m_Value(V&: Op0), R: m_Value(V&: Op1))) ||
668	match(V: I, P: m_LogicalAnd(L: m_Value(V&: Op0), R: m_Value(V&: Op1)))) {
669	PredValueInfoTy LHSVals, RHSVals;
670
671	computeValueKnownInPredecessorsImpl(V: Op0, BB, Result&: LHSVals, Preference: WantInteger,
672	RecursionSet, CxtI);
673	computeValueKnownInPredecessorsImpl(V: Op1, BB, Result&: RHSVals, Preference: WantInteger,
674	RecursionSet, CxtI);
675
676	if (LHSVals.empty() && RHSVals.empty())
677	return false;
678
679	ConstantInt *InterestingVal;
680	if (match(V: I, P: m_LogicalOr()))
681	InterestingVal = ConstantInt::getTrue(Context&: I->getContext());
682	else
683	InterestingVal = ConstantInt::getFalse(Context&: I->getContext());
684
685	SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
686
687	// Scan for the sentinel. If we find an undef, force it to the
688	// interesting value: x|undef -> true and x&undef -> false.
689	for (const auto &LHSVal : LHSVals)
690	if (LHSVal.first == InterestingVal || isa<UndefValue>(Val: LHSVal.first)) {
691	Result.emplace_back(Args&: InterestingVal, Args: LHSVal.second);
692	LHSKnownBBs.insert(Ptr: LHSVal.second);
693	}
694	for (const auto &RHSVal : RHSVals)
695	if (RHSVal.first == InterestingVal || isa<UndefValue>(Val: RHSVal.first)) {
696	// If we already inferred a value for this block on the LHS, don't
697	// re-add it.
698	if (!LHSKnownBBs.count(Ptr: RHSVal.second))
699	Result.emplace_back(Args&: InterestingVal, Args: RHSVal.second);
700	}
701
702	return !Result.empty();
703	}
704
705	// Handle the NOT form of XOR.
706	if (I->getOpcode() == Instruction::Xor &&
707	isa<ConstantInt>(Val: I->getOperand(i: 1)) &&
708	cast<ConstantInt>(Val: I->getOperand(i: 1))->isOne()) {
709	computeValueKnownInPredecessorsImpl(V: I->getOperand(i: 0), BB, Result,
710	Preference: WantInteger, RecursionSet, CxtI);
711	if (Result.empty())
712	return false;
713
714	// Invert the known values.
715	for (auto &R : Result)
716	R.first = ConstantExpr::getNot(C: R.first);
717
718	return true;
719	}
720
721	// Try to simplify some other binary operator values.
722	} else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Val: I)) {
723	if (Preference != WantInteger)
724	return false;
725	if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: BO->getOperand(i_nocapture: 1))) {
726	PredValueInfoTy LHSVals;
727	computeValueKnownInPredecessorsImpl(V: BO->getOperand(i_nocapture: 0), BB, Result&: LHSVals,
728	Preference: WantInteger, RecursionSet, CxtI);
729
730	// Try to use constant folding to simplify the binary operator.
731	for (const auto &LHSVal : LHSVals) {
732	Constant *V = LHSVal.first;
733	Constant *Folded =
734	ConstantFoldBinaryOpOperands(Opcode: BO->getOpcode(), LHS: V, RHS: CI, DL);
735
736	if (Constant *KC = getKnownConstant(Val: Folded, Preference: WantInteger))
737	Result.emplace_back(Args&: KC, Args: LHSVal.second);
738	}
739	}
740
741	return !Result.empty();
742	}
743
744	// Handle compare with phi operand, where the PHI is defined in this block.
745	if (CmpInst *Cmp = dyn_cast<CmpInst>(Val: I)) {
746	if (Preference != WantInteger)
747	return false;
748	Type *CmpType = Cmp->getType();
749	Value *CmpLHS = Cmp->getOperand(i_nocapture: 0);
750	Value *CmpRHS = Cmp->getOperand(i_nocapture: 1);
751	CmpInst::Predicate Pred = Cmp->getPredicate();
752
753	PHINode *PN = dyn_cast<PHINode>(Val: CmpLHS);
754	if (!PN)
755	PN = dyn_cast<PHINode>(Val: CmpRHS);
756	// Do not perform phi translation across a loop header phi, because this
757	// may result in comparison of values from two different loop iterations.
758	// FIXME: This check is broken if LoopHeaders is not populated.
759	if (PN && PN->getParent() == BB && !LoopHeaders.contains(V: BB)) {
760	const DataLayout &DL = PN->getModule()->getDataLayout();
761	// We can do this simplification if any comparisons fold to true or false.
762	// See if any do.
763	for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
764	BasicBlock *PredBB = PN->getIncomingBlock(i);
765	Value *LHS, *RHS;
766	if (PN == CmpLHS) {
767	LHS = PN->getIncomingValue(i);
768	RHS = CmpRHS->DoPHITranslation(CurBB: BB, PredBB);
769	} else {
770	LHS = CmpLHS->DoPHITranslation(CurBB: BB, PredBB);
771	RHS = PN->getIncomingValue(i);
772	}
773	Value *Res = simplifyCmpInst(Predicate: Pred, LHS, RHS, Q: {DL});
774	if (!Res) {
775	if (!isa<Constant>(Val: RHS))
776	continue;
777
778	// getPredicateOnEdge call will make no sense if LHS is defined in BB.
779	auto LHSInst = dyn_cast<Instruction>(Val: LHS);
780	if (LHSInst && LHSInst->getParent() == BB)
781	continue;
782
783	LazyValueInfo::Tristate
784	ResT = LVI->getPredicateOnEdge(Pred, V: LHS,
785	C: cast<Constant>(Val: RHS), FromBB: PredBB, ToBB: BB,
786	CxtI: CxtI ? CxtI : Cmp);
787	if (ResT == LazyValueInfo::Unknown)
788	continue;
789	Res = ConstantInt::get(Ty: Type::getInt1Ty(C&: LHS->getContext()), V: ResT);
790	}
791
792	if (Constant *KC = getKnownConstant(Val: Res, Preference: WantInteger))
793	Result.emplace_back(Args&: KC, Args&: PredBB);
794	}
795
796	return !Result.empty();
797	}
798
799	// If comparing a live-in value against a constant, see if we know the
800	// live-in value on any predecessors.
801	if (isa<Constant>(Val: CmpRHS) && !CmpType->isVectorTy()) {
802	Constant *CmpConst = cast<Constant>(Val: CmpRHS);
803
804	if (!isa<Instruction>(Val: CmpLHS) ||
805	cast<Instruction>(Val: CmpLHS)->getParent() != BB) {
806	for (BasicBlock *P : predecessors(BB)) {
807	// If the value is known by LazyValueInfo to be a constant in a
808	// predecessor, use that information to try to thread this block.
809	LazyValueInfo::Tristate Res =
810	LVI->getPredicateOnEdge(Pred, V: CmpLHS,
811	C: CmpConst, FromBB: P, ToBB: BB, CxtI: CxtI ? CxtI : Cmp);
812	if (Res == LazyValueInfo::Unknown)
813	continue;
814
815	Constant *ResC = ConstantInt::get(Ty: CmpType, V: Res);
816	Result.emplace_back(Args&: ResC, Args&: P);
817	}
818
819	return !Result.empty();
820	}
821
822	// InstCombine can fold some forms of constant range checks into
823	// (icmp (add (x, C1)), C2). See if we have we have such a thing with
824	// x as a live-in.
825	{
826	using namespace PatternMatch;
827
828	Value *AddLHS;
829	ConstantInt *AddConst;
830	if (isa<ConstantInt>(Val: CmpConst) &&
831	match(V: CmpLHS, P: m_Add(L: m_Value(V&: AddLHS), R: m_ConstantInt(CI&: AddConst)))) {
832	if (!isa<Instruction>(Val: AddLHS) ||
833	cast<Instruction>(Val: AddLHS)->getParent() != BB) {
834	for (BasicBlock *P : predecessors(BB)) {
835	// If the value is known by LazyValueInfo to be a ConstantRange in
836	// a predecessor, use that information to try to thread this
837	// block.
838	ConstantRange CR = LVI->getConstantRangeOnEdge(
839	V: AddLHS, FromBB: P, ToBB: BB, CxtI: CxtI ? CxtI : cast<Instruction>(Val: CmpLHS));
840	// Propagate the range through the addition.
841	CR = CR.add(Other: AddConst->getValue());
842
843	// Get the range where the compare returns true.
844	ConstantRange CmpRange = ConstantRange::makeExactICmpRegion(
845	Pred, Other: cast<ConstantInt>(Val: CmpConst)->getValue());
846
847	Constant *ResC;
848	if (CmpRange.contains(CR))
849	ResC = ConstantInt::getTrue(Ty: CmpType);
850	else if (CmpRange.inverse().contains(CR))
851	ResC = ConstantInt::getFalse(Ty: CmpType);
852	else
853	continue;
854
855	Result.emplace_back(Args&: ResC, Args&: P);
856	}
857
858	return !Result.empty();
859	}
860	}
861	}
862
863	// Try to find a constant value for the LHS of a comparison,
864	// and evaluate it statically if we can.
865	PredValueInfoTy LHSVals;
866	computeValueKnownInPredecessorsImpl(V: I->getOperand(i: 0), BB, Result&: LHSVals,
867	Preference: WantInteger, RecursionSet, CxtI);
868
869	for (const auto &LHSVal : LHSVals) {
870	Constant *V = LHSVal.first;
871	Constant *Folded = ConstantExpr::getCompare(pred: Pred, C1: V, C2: CmpConst);
872	if (Constant *KC = getKnownConstant(Val: Folded, Preference: WantInteger))
873	Result.emplace_back(Args&: KC, Args: LHSVal.second);
874	}
875
876	return !Result.empty();
877	}
878	}
879
880	if (SelectInst *SI = dyn_cast<SelectInst>(Val: I)) {
881	// Handle select instructions where at least one operand is a known constant
882	// and we can figure out the condition value for any predecessor block.
883	Constant *TrueVal = getKnownConstant(Val: SI->getTrueValue(), Preference);
884	Constant *FalseVal = getKnownConstant(Val: SI->getFalseValue(), Preference);
885	PredValueInfoTy Conds;
886	if ((TrueVal || FalseVal) &&
887	computeValueKnownInPredecessorsImpl(V: SI->getCondition(), BB, Result&: Conds,
888	Preference: WantInteger, RecursionSet, CxtI)) {
889	for (auto &C : Conds) {
890	Constant *Cond = C.first;
891
892	// Figure out what value to use for the condition.
893	bool KnownCond;
894	if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: Cond)) {
895	// A known boolean.
896	KnownCond = CI->isOne();
897	} else {
898	assert(isa<UndefValue>(Cond) && "Unexpected condition value");
899	// Either operand will do, so be sure to pick the one that's a known
900	// constant.
901	// FIXME: Do this more cleverly if both values are known constants?
902	KnownCond = (TrueVal != nullptr);
903	}
904
905	// See if the select has a known constant value for this predecessor.
906	if (Constant *Val = KnownCond ? TrueVal : FalseVal)
907	Result.emplace_back(Args&: Val, Args&: C.second);
908	}
909
910	return !Result.empty();
911	}
912	}
913
914	// If all else fails, see if LVI can figure out a constant value for us.
915	assert(CxtI->getParent() == BB && "CxtI should be in BB");
916	Constant *CI = LVI->getConstant(V, CxtI);
917	if (Constant *KC = getKnownConstant(Val: CI, Preference)) {
918	for (BasicBlock *Pred : predecessors(BB))
919	Result.emplace_back(Args&: KC, Args&: Pred);
920	}
921
922	return !Result.empty();
923	}
924
925	/// GetBestDestForBranchOnUndef - If we determine that the specified block ends
926	/// in an undefined jump, decide which block is best to revector to.
927	///
928	/// Since we can pick an arbitrary destination, we pick the successor with the
929	/// fewest predecessors. This should reduce the in-degree of the others.
930	static unsigned getBestDestForJumpOnUndef(BasicBlock *BB) {
931	Instruction *BBTerm = BB->getTerminator();
932	unsigned MinSucc = 0;
933	BasicBlock *TestBB = BBTerm->getSuccessor(Idx: MinSucc);
934	// Compute the successor with the minimum number of predecessors.
935	unsigned MinNumPreds = pred_size(BB: TestBB);
936	for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
937	TestBB = BBTerm->getSuccessor(Idx: i);
938	unsigned NumPreds = pred_size(BB: TestBB);
939	if (NumPreds < MinNumPreds) {
940	MinSucc = i;
941	MinNumPreds = NumPreds;
942	}
943	}
944
945	return MinSucc;
946	}
947
948	static bool hasAddressTakenAndUsed(BasicBlock *BB) {
949	if (!BB->hasAddressTaken()) return false;
950
951	// If the block has its address taken, it may be a tree of dead constants
952	// hanging off of it. These shouldn't keep the block alive.
953	BlockAddress *BA = BlockAddress::get(BB);
954	BA->removeDeadConstantUsers();
955	return !BA->use_empty();
956	}
957
958	/// processBlock - If there are any predecessors whose control can be threaded
959	/// through to a successor, transform them now.
960	bool JumpThreadingPass::processBlock(BasicBlock *BB) {
961	// If the block is trivially dead, just return and let the caller nuke it.
962	// This simplifies other transformations.
963	if (DTU->isBBPendingDeletion(DelBB: BB) ||
964	(pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()))
965	return false;
966
967	// If this block has a single predecessor, and if that pred has a single
968	// successor, merge the blocks. This encourages recursive jump threading
969	// because now the condition in this block can be threaded through
970	// predecessors of our predecessor block.
971	if (maybeMergeBasicBlockIntoOnlyPred(BB))
972	return true;
973
974	if (tryToUnfoldSelectInCurrBB(BB))
975	return true;
976
977	// Look if we can propagate guards to predecessors.
978	if (HasGuards && processGuards(BB))
979	return true;
980
981	// What kind of constant we're looking for.
982	ConstantPreference Preference = WantInteger;
983
984	// Look to see if the terminator is a conditional branch, switch or indirect
985	// branch, if not we can't thread it.
986	Value *Condition;
987	Instruction *Terminator = BB->getTerminator();
988	if (BranchInst *BI = dyn_cast<BranchInst>(Val: Terminator)) {
989	// Can't thread an unconditional jump.
990	if (BI->isUnconditional()) return false;
991	Condition = BI->getCondition();
992	} else if (SwitchInst *SI = dyn_cast<SwitchInst>(Val: Terminator)) {
993	Condition = SI->getCondition();
994	} else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Val: Terminator)) {
995	// Can't thread indirect branch with no successors.
996	if (IB->getNumSuccessors() == 0) return false;
997	Condition = IB->getAddress()->stripPointerCasts();
998	Preference = WantBlockAddress;
999	} else {
1000	return false; // Must be an invoke or callbr.
1001	}
1002
1003	// Keep track if we constant folded the condition in this invocation.
1004	bool ConstantFolded = false;
1005
1006	// Run constant folding to see if we can reduce the condition to a simple
1007	// constant.
1008	if (Instruction *I = dyn_cast<Instruction>(Val: Condition)) {
1009	Value *SimpleVal =
1010	ConstantFoldInstruction(I, DL: BB->getModule()->getDataLayout(), TLI);
1011	if (SimpleVal) {
1012	I->replaceAllUsesWith(V: SimpleVal);
1013	if (isInstructionTriviallyDead(I, TLI))
1014	I->eraseFromParent();
1015	Condition = SimpleVal;
1016	ConstantFolded = true;
1017	}
1018	}
1019
1020	// If the terminator is branching on an undef or freeze undef, we can pick any
1021	// of the successors to branch to. Let getBestDestForJumpOnUndef decide.
1022	auto *FI = dyn_cast<FreezeInst>(Val: Condition);
1023	if (isa<UndefValue>(Val: Condition) ||
1024	(FI && isa<UndefValue>(Val: FI->getOperand(i_nocapture: 0)) && FI->hasOneUse())) {
1025	unsigned BestSucc = getBestDestForJumpOnUndef(BB);
1026	std::vector<DominatorTree::UpdateType> Updates;
1027
1028	// Fold the branch/switch.
1029	Instruction *BBTerm = BB->getTerminator();
1030	Updates.reserve(n: BBTerm->getNumSuccessors());
1031	for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
1032	if (i == BestSucc) continue;
1033	BasicBlock *Succ = BBTerm->getSuccessor(Idx: i);
1034	Succ->removePredecessor(Pred: BB, KeepOneInputPHIs: true);
1035	Updates.push_back(x: {DominatorTree::Delete, BB, Succ});
1036	}
1037
1038	LLVM_DEBUG(dbgs() << " In block '" << BB->getName()
1039	<< "' folding undef terminator: " << *BBTerm << '\n');
1040	BranchInst::Create(IfTrue: BBTerm->getSuccessor(Idx: BestSucc), InsertBefore: BBTerm->getIterator());
1041	++NumFolds;
1042	BBTerm->eraseFromParent();
1043	DTU->applyUpdatesPermissive(Updates);
1044	if (FI)
1045	FI->eraseFromParent();
1046	return true;
1047	}
1048
1049	// If the terminator of this block is branching on a constant, simplify the
1050	// terminator to an unconditional branch. This can occur due to threading in
1051	// other blocks.
1052	if (getKnownConstant(Val: Condition, Preference)) {
1053	LLVM_DEBUG(dbgs() << " In block '" << BB->getName()
1054	<< "' folding terminator: " << *BB->getTerminator()
1055	<< '\n');
1056	++NumFolds;
1057	ConstantFoldTerminator(BB, DeleteDeadConditions: true, TLI: nullptr, DTU: DTU.get());
1058	if (auto *BPI = getBPI())
1059	BPI->eraseBlock(BB);
1060	return true;
1061	}
1062
1063	Instruction *CondInst = dyn_cast<Instruction>(Val: Condition);
1064
1065	// All the rest of our checks depend on the condition being an instruction.
1066	if (!CondInst) {
1067	// FIXME: Unify this with code below.
1068	if (processThreadableEdges(Cond: Condition, BB, Preference, CxtI: Terminator))
1069	return true;
1070	return ConstantFolded;
1071	}
1072
1073	// Some of the following optimization can safely work on the unfrozen cond.
1074	Value *CondWithoutFreeze = CondInst;
1075	if (auto *FI = dyn_cast<FreezeInst>(Val: CondInst))
1076	CondWithoutFreeze = FI->getOperand(i_nocapture: 0);
1077
1078	if (CmpInst *CondCmp = dyn_cast<CmpInst>(Val: CondWithoutFreeze)) {
1079	// If we're branching on a conditional, LVI might be able to determine
1080	// it's value at the branch instruction. We only handle comparisons
1081	// against a constant at this time.
1082	if (Constant *CondConst = dyn_cast<Constant>(Val: CondCmp->getOperand(i_nocapture: 1))) {
1083	LazyValueInfo::Tristate Ret =
1084	LVI->getPredicateAt(Pred: CondCmp->getPredicate(), V: CondCmp->getOperand(i_nocapture: 0),
1085	C: CondConst, CxtI: BB->getTerminator(),
1086	/*UseBlockValue=*/false);
1087	if (Ret != LazyValueInfo::Unknown) {
1088	// We can safely replace *some* uses of the CondInst if it has
1089	// exactly one value as returned by LVI. RAUW is incorrect in the
1090	// presence of guards and assumes, that have the `Cond` as the use. This
1091	// is because we use the guards/assume to reason about the `Cond` value
1092	// at the end of block, but RAUW unconditionally replaces all uses
1093	// including the guards/assumes themselves and the uses before the
1094	// guard/assume.
1095	auto *CI = Ret == LazyValueInfo::True ?
1096	ConstantInt::getTrue(Ty: CondCmp->getType()) :
1097	ConstantInt::getFalse(Ty: CondCmp->getType());
1098	if (replaceFoldableUses(Cond: CondCmp, ToVal: CI, KnownAtEndOfBB: BB))
1099	return true;
1100	}
1101
1102	// We did not manage to simplify this branch, try to see whether
1103	// CondCmp depends on a known phi-select pattern.
1104	if (tryToUnfoldSelect(CondCmp, BB))
1105	return true;
1106	}
1107	}
1108
1109	if (SwitchInst *SI = dyn_cast<SwitchInst>(Val: BB->getTerminator()))
1110	if (tryToUnfoldSelect(SI, BB))
1111	return true;
1112
1113	// Check for some cases that are worth simplifying. Right now we want to look
1114	// for loads that are used by a switch or by the condition for the branch. If
1115	// we see one, check to see if it's partially redundant. If so, insert a PHI
1116	// which can then be used to thread the values.
1117	Value *SimplifyValue = CondWithoutFreeze;
1118
1119	if (CmpInst *CondCmp = dyn_cast<CmpInst>(Val: SimplifyValue))
1120	if (isa<Constant>(Val: CondCmp->getOperand(i_nocapture: 1)))
1121	SimplifyValue = CondCmp->getOperand(i_nocapture: 0);
1122
1123	// TODO: There are other places where load PRE would be profitable, such as
1124	// more complex comparisons.
1125	if (LoadInst *LoadI = dyn_cast<LoadInst>(Val: SimplifyValue))
1126	if (simplifyPartiallyRedundantLoad(LI: LoadI))
1127	return true;
1128
1129	// Before threading, try to propagate profile data backwards:
1130	if (PHINode *PN = dyn_cast<PHINode>(Val: CondInst))
1131	if (PN->getParent() == BB && isa<BranchInst>(Val: BB->getTerminator()))
1132	updatePredecessorProfileMetadata(PN, BB);
1133
1134	// Handle a variety of cases where we are branching on something derived from
1135	// a PHI node in the current block. If we can prove that any predecessors
1136	// compute a predictable value based on a PHI node, thread those predecessors.
1137	if (processThreadableEdges(Cond: CondInst, BB, Preference, CxtI: Terminator))
1138	return true;
1139
1140	// If this is an otherwise-unfoldable branch on a phi node or freeze(phi) in
1141	// the current block, see if we can simplify.
1142	PHINode *PN = dyn_cast<PHINode>(Val: CondWithoutFreeze);
1143	if (PN && PN->getParent() == BB && isa<BranchInst>(Val: BB->getTerminator()))
1144	return processBranchOnPHI(PN);
1145
1146	// If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
1147	if (CondInst->getOpcode() == Instruction::Xor &&
1148	CondInst->getParent() == BB && isa<BranchInst>(Val: BB->getTerminator()))
1149	return processBranchOnXOR(BO: cast<BinaryOperator>(Val: CondInst));
1150
1151	// Search for a stronger dominating condition that can be used to simplify a
1152	// conditional branch leaving BB.
1153	if (processImpliedCondition(BB))
1154	return true;
1155
1156	return false;
1157	}
1158
1159	bool JumpThreadingPass::processImpliedCondition(BasicBlock *BB) {
1160	auto *BI = dyn_cast<BranchInst>(Val: BB->getTerminator());
1161	if (!BI || !BI->isConditional())
1162	return false;
1163
1164	Value *Cond = BI->getCondition();
1165	// Assuming that predecessor's branch was taken, if pred's branch condition
1166	// (V) implies Cond, Cond can be either true, undef, or poison. In this case,
1167	// freeze(Cond) is either true or a nondeterministic value.
1168	// If freeze(Cond) has only one use, we can freely fold freeze(Cond) to true
1169	// without affecting other instructions.
1170	auto *FICond = dyn_cast<FreezeInst>(Val: Cond);
1171	if (FICond && FICond->hasOneUse())
1172	Cond = FICond->getOperand(i_nocapture: 0);
1173	else
1174	FICond = nullptr;
1175
1176	BasicBlock *CurrentBB = BB;
1177	BasicBlock *CurrentPred = BB->getSinglePredecessor();
1178	unsigned Iter = 0;
1179
1180	auto &DL = BB->getModule()->getDataLayout();
1181
1182	while (CurrentPred && Iter++ < ImplicationSearchThreshold) {
1183	auto *PBI = dyn_cast<BranchInst>(Val: CurrentPred->getTerminator());
1184	if (!PBI || !PBI->isConditional())
1185	return false;
1186	if (PBI->getSuccessor(i: 0) != CurrentBB && PBI->getSuccessor(i: 1) != CurrentBB)
1187	return false;
1188
1189	bool CondIsTrue = PBI->getSuccessor(i: 0) == CurrentBB;
1190	std::optional<bool> Implication =
1191	isImpliedCondition(LHS: PBI->getCondition(), RHS: Cond, DL, LHSIsTrue: CondIsTrue);
1192
1193	// If the branch condition of BB (which is Cond) and CurrentPred are
1194	// exactly the same freeze instruction, Cond can be folded into CondIsTrue.
1195	if (!Implication && FICond && isa<FreezeInst>(Val: PBI->getCondition())) {
1196	if (cast<FreezeInst>(Val: PBI->getCondition())->getOperand(i_nocapture: 0) ==
1197	FICond->getOperand(i_nocapture: 0))
1198	Implication = CondIsTrue;
1199	}
1200
1201	if (Implication) {
1202	BasicBlock *KeepSucc = BI->getSuccessor(i: *Implication ? 0 : 1);
1203	BasicBlock *RemoveSucc = BI->getSuccessor(i: *Implication ? 1 : 0);
1204	RemoveSucc->removePredecessor(Pred: BB);
1205	BranchInst *UncondBI = BranchInst::Create(IfTrue: KeepSucc, InsertBefore: BI->getIterator());
1206	UncondBI->setDebugLoc(BI->getDebugLoc());
1207	++NumFolds;
1208	BI->eraseFromParent();
1209	if (FICond)
1210	FICond->eraseFromParent();
1211
1212	DTU->applyUpdatesPermissive(Updates: {{DominatorTree::Delete, BB, RemoveSucc}});
1213	if (auto *BPI = getBPI())
1214	BPI->eraseBlock(BB);
1215	return true;
1216	}
1217	CurrentBB = CurrentPred;
1218	CurrentPred = CurrentBB->getSinglePredecessor();
1219	}
1220
1221	return false;
1222	}
1223
1224	/// Return true if Op is an instruction defined in the given block.
1225	static bool isOpDefinedInBlock(Value *Op, BasicBlock *BB) {
1226	if (Instruction *OpInst = dyn_cast<Instruction>(Val: Op))
1227	if (OpInst->getParent() == BB)
1228	return true;
1229	return false;
1230	}
1231
1232	/// simplifyPartiallyRedundantLoad - If LoadI is an obviously partially
1233	/// redundant load instruction, eliminate it by replacing it with a PHI node.
1234	/// This is an important optimization that encourages jump threading, and needs
1235	/// to be run interlaced with other jump threading tasks.
1236	bool JumpThreadingPass::simplifyPartiallyRedundantLoad(LoadInst *LoadI) {
1237	// Don't hack volatile and ordered loads.
1238	if (!LoadI->isUnordered()) return false;
1239
1240	// If the load is defined in a block with exactly one predecessor, it can't be
1241	// partially redundant.
1242	BasicBlock *LoadBB = LoadI->getParent();
1243	if (LoadBB->getSinglePredecessor())
1244	return false;
1245
1246	// If the load is defined in an EH pad, it can't be partially redundant,
1247	// because the edges between the invoke and the EH pad cannot have other
1248	// instructions between them.
1249	if (LoadBB->isEHPad())
1250	return false;
1251
1252	Value *LoadedPtr = LoadI->getOperand(i_nocapture: 0);
1253
1254	// If the loaded operand is defined in the LoadBB and its not a phi,
1255	// it can't be available in predecessors.
1256	if (isOpDefinedInBlock(Op: LoadedPtr, BB: LoadBB) && !isa<PHINode>(Val: LoadedPtr))
1257	return false;
1258
1259	// Scan a few instructions up from the load, to see if it is obviously live at
1260	// the entry to its block.
1261	BasicBlock::iterator BBIt(LoadI);
1262	bool IsLoadCSE;
1263	BatchAAResults BatchAA(*AA);
1264	// The dominator tree is updated lazily and may not be valid at this point.
1265	BatchAA.disableDominatorTree();
1266	if (Value *AvailableVal = FindAvailableLoadedValue(
1267	Load: LoadI, ScanBB: LoadBB, ScanFrom&: BBIt, MaxInstsToScan: DefMaxInstsToScan, AA: &BatchAA, IsLoadCSE: &IsLoadCSE)) {
1268	// If the value of the load is locally available within the block, just use
1269	// it. This frequently occurs for reg2mem'd allocas.
1270
1271	if (IsLoadCSE) {
1272	LoadInst *NLoadI = cast<LoadInst>(Val: AvailableVal);
1273	combineMetadataForCSE(K: NLoadI, J: LoadI, DoesKMove: false);
1274	LVI->forgetValue(V: NLoadI);
1275	};
1276
1277	// If the returned value is the load itself, replace with poison. This can
1278	// only happen in dead loops.
1279	if (AvailableVal == LoadI)
1280	AvailableVal = PoisonValue::get(T: LoadI->getType());
1281	if (AvailableVal->getType() != LoadI->getType())
1282	AvailableVal = CastInst::CreateBitOrPointerCast(
1283	S: AvailableVal, Ty: LoadI->getType(), Name: "", InsertBefore: LoadI->getIterator());
1284	LoadI->replaceAllUsesWith(V: AvailableVal);
1285	LoadI->eraseFromParent();
1286	return true;
1287	}
1288
1289	// Otherwise, if we scanned the whole block and got to the top of the block,
1290	// we know the block is locally transparent to the load. If not, something
1291	// might clobber its value.
1292	if (BBIt != LoadBB->begin())
1293	return false;
1294
1295	// If all of the loads and stores that feed the value have the same AA tags,
1296	// then we can propagate them onto any newly inserted loads.
1297	AAMDNodes AATags = LoadI->getAAMetadata();
1298
1299	SmallPtrSet<BasicBlock*, 8> PredsScanned;
1300
1301	using AvailablePredsTy = SmallVector<std::pair<BasicBlock *, Value *>, 8>;
1302
1303	AvailablePredsTy AvailablePreds;
1304	BasicBlock *OneUnavailablePred = nullptr;
1305	SmallVector<LoadInst*, 8> CSELoads;
1306
1307	// If we got here, the loaded value is transparent through to the start of the
1308	// block. Check to see if it is available in any of the predecessor blocks.
1309	for (BasicBlock *PredBB : predecessors(BB: LoadBB)) {
1310	// If we already scanned this predecessor, skip it.
1311	if (!PredsScanned.insert(Ptr: PredBB).second)
1312	continue;
1313
1314	BBIt = PredBB->end();
1315	unsigned NumScanedInst = 0;
1316	Value *PredAvailable = nullptr;
1317	// NOTE: We don't CSE load that is volatile or anything stronger than
1318	// unordered, that should have been checked when we entered the function.
1319	assert(LoadI->isUnordered() &&
1320	"Attempting to CSE volatile or atomic loads");
1321	// If this is a load on a phi pointer, phi-translate it and search
1322	// for available load/store to the pointer in predecessors.
1323	Type *AccessTy = LoadI->getType();
1324	const auto &DL = LoadI->getModule()->getDataLayout();
1325	MemoryLocation Loc(LoadedPtr->DoPHITranslation(CurBB: LoadBB, PredBB),
1326	LocationSize::precise(Value: DL.getTypeStoreSize(Ty: AccessTy)),
1327	AATags);
1328	PredAvailable = findAvailablePtrLoadStore(
1329	Loc, AccessTy, AtLeastAtomic: LoadI->isAtomic(), ScanBB: PredBB, ScanFrom&: BBIt, MaxInstsToScan: DefMaxInstsToScan,
1330	AA: &BatchAA, IsLoadCSE: &IsLoadCSE, NumScanedInst: &NumScanedInst);
1331
1332	// If PredBB has a single predecessor, continue scanning through the
1333	// single predecessor.
1334	BasicBlock *SinglePredBB = PredBB;
1335	while (!PredAvailable && SinglePredBB && BBIt == SinglePredBB->begin() &&
1336	NumScanedInst < DefMaxInstsToScan) {
1337	SinglePredBB = SinglePredBB->getSinglePredecessor();
1338	if (SinglePredBB) {
1339	BBIt = SinglePredBB->end();
1340	PredAvailable = findAvailablePtrLoadStore(
1341	Loc, AccessTy, AtLeastAtomic: LoadI->isAtomic(), ScanBB: SinglePredBB, ScanFrom&: BBIt,
1342	MaxInstsToScan: (DefMaxInstsToScan - NumScanedInst), AA: &BatchAA, IsLoadCSE: &IsLoadCSE,
1343	NumScanedInst: &NumScanedInst);
1344	}
1345	}
1346
1347	if (!PredAvailable) {
1348	OneUnavailablePred = PredBB;
1349	continue;
1350	}
1351
1352	if (IsLoadCSE)
1353	CSELoads.push_back(Elt: cast<LoadInst>(Val: PredAvailable));
1354
1355	// If so, this load is partially redundant. Remember this info so that we
1356	// can create a PHI node.
1357	AvailablePreds.emplace_back(Args&: PredBB, Args&: PredAvailable);
1358	}
1359
1360	// If the loaded value isn't available in any predecessor, it isn't partially
1361	// redundant.
1362	if (AvailablePreds.empty()) return false;
1363
1364	// Okay, the loaded value is available in at least one (and maybe all!)
1365	// predecessors. If the value is unavailable in more than one unique
1366	// predecessor, we want to insert a merge block for those common predecessors.
1367	// This ensures that we only have to insert one reload, thus not increasing
1368	// code size.
1369	BasicBlock *UnavailablePred = nullptr;
1370
1371	// If the value is unavailable in one of predecessors, we will end up
1372	// inserting a new instruction into them. It is only valid if all the
1373	// instructions before LoadI are guaranteed to pass execution to its
1374	// successor, or if LoadI is safe to speculate.
1375	// TODO: If this logic becomes more complex, and we will perform PRE insertion
1376	// farther than to a predecessor, we need to reuse the code from GVN's PRE.
1377	// It requires domination tree analysis, so for this simple case it is an
1378	// overkill.
1379	if (PredsScanned.size() != AvailablePreds.size() &&
1380	!isSafeToSpeculativelyExecute(I: LoadI))
1381	for (auto I = LoadBB->begin(); &*I != LoadI; ++I)
1382	if (!isGuaranteedToTransferExecutionToSuccessor(I: &*I))
1383	return false;
1384
1385	// If there is exactly one predecessor where the value is unavailable, the
1386	// already computed 'OneUnavailablePred' block is it. If it ends in an
1387	// unconditional branch, we know that it isn't a critical edge.
1388	if (PredsScanned.size() == AvailablePreds.size()+1 &&
1389	OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
1390	UnavailablePred = OneUnavailablePred;
1391	} else if (PredsScanned.size() != AvailablePreds.size()) {
1392	// Otherwise, we had multiple unavailable predecessors or we had a critical
1393	// edge from the one.
1394	SmallVector<BasicBlock*, 8> PredsToSplit;
1395	SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
1396
1397	for (const auto &AvailablePred : AvailablePreds)
1398	AvailablePredSet.insert(Ptr: AvailablePred.first);
1399
1400	// Add all the unavailable predecessors to the PredsToSplit list.
1401	for (BasicBlock *P : predecessors(BB: LoadBB)) {
1402	// If the predecessor is an indirect goto, we can't split the edge.
1403	if (isa<IndirectBrInst>(Val: P->getTerminator()))
1404	return false;
1405
1406	if (!AvailablePredSet.count(Ptr: P))
1407	PredsToSplit.push_back(Elt: P);
1408	}
1409
1410	// Split them out to their own block.
1411	UnavailablePred = splitBlockPreds(BB: LoadBB, Preds: PredsToSplit, Suffix: "thread-pre-split");
1412	}
1413
1414	// If the value isn't available in all predecessors, then there will be
1415	// exactly one where it isn't available. Insert a load on that edge and add
1416	// it to the AvailablePreds list.
1417	if (UnavailablePred) {
1418	assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
1419	"Can't handle critical edge here!");
1420	LoadInst *NewVal = new LoadInst(
1421	LoadI->getType(), LoadedPtr->DoPHITranslation(CurBB: LoadBB, PredBB: UnavailablePred),
1422	LoadI->getName() + ".pr", false, LoadI->getAlign(),
1423	LoadI->getOrdering(), LoadI->getSyncScopeID(),
1424	UnavailablePred->getTerminator()->getIterator());
1425	NewVal->setDebugLoc(LoadI->getDebugLoc());
1426	if (AATags)
1427	NewVal->setAAMetadata(AATags);
1428
1429	AvailablePreds.emplace_back(Args&: UnavailablePred, Args&: NewVal);
1430	}
1431
1432	// Now we know that each predecessor of this block has a value in
1433	// AvailablePreds, sort them for efficient access as we're walking the preds.
1434	array_pod_sort(Start: AvailablePreds.begin(), End: AvailablePreds.end());
1435
1436	// Create a PHI node at the start of the block for the PRE'd load value.
1437	PHINode *PN = PHINode::Create(Ty: LoadI->getType(), NumReservedValues: pred_size(BB: LoadBB), NameStr: "");
1438	PN->insertBefore(InsertPos: LoadBB->begin());
1439	PN->takeName(V: LoadI);
1440	PN->setDebugLoc(LoadI->getDebugLoc());
1441
1442	// Insert new entries into the PHI for each predecessor. A single block may
1443	// have multiple entries here.
1444	for (BasicBlock *P : predecessors(BB: LoadBB)) {
1445	AvailablePredsTy::iterator I =
1446	llvm::lower_bound(Range&: AvailablePreds, Value: std::make_pair(x&: P, y: (Value *)nullptr));
1447
1448	assert(I != AvailablePreds.end() && I->first == P &&
1449	"Didn't find entry for predecessor!");
1450
1451	// If we have an available predecessor but it requires casting, insert the
1452	// cast in the predecessor and use the cast. Note that we have to update the
1453	// AvailablePreds vector as we go so that all of the PHI entries for this
1454	// predecessor use the same bitcast.
1455	Value *&PredV = I->second;
1456	if (PredV->getType() != LoadI->getType())
1457	PredV = CastInst::CreateBitOrPointerCast(
1458	S: PredV, Ty: LoadI->getType(), Name: "", InsertBefore: P->getTerminator()->getIterator());
1459
1460	PN->addIncoming(V: PredV, BB: I->first);
1461	}
1462
1463	for (LoadInst *PredLoadI : CSELoads) {
1464	combineMetadataForCSE(K: PredLoadI, J: LoadI, DoesKMove: true);
1465	LVI->forgetValue(V: PredLoadI);
1466	}
1467
1468	LoadI->replaceAllUsesWith(V: PN);
1469	LoadI->eraseFromParent();
1470
1471	return true;
1472	}
1473
1474	/// findMostPopularDest - The specified list contains multiple possible
1475	/// threadable destinations. Pick the one that occurs the most frequently in
1476	/// the list.
1477	static BasicBlock *
1478	findMostPopularDest(BasicBlock *BB,
1479	const SmallVectorImpl<std::pair<BasicBlock *,
1480	BasicBlock *>> &PredToDestList) {
1481	assert(!PredToDestList.empty());
1482
1483	// Determine popularity. If there are multiple possible destinations, we
1484	// explicitly choose to ignore 'undef' destinations. We prefer to thread
1485	// blocks with known and real destinations to threading undef. We'll handle
1486	// them later if interesting.
1487	MapVector<BasicBlock *, unsigned> DestPopularity;
1488
1489	// Populate DestPopularity with the successors in the order they appear in the
1490	// successor list. This way, we ensure determinism by iterating it in the
1491	// same order in llvm::max_element below. We map nullptr to 0 so that we can
1492	// return nullptr when PredToDestList contains nullptr only.
1493	DestPopularity[nullptr] = 0;
1494	for (auto *SuccBB : successors(BB))
1495	DestPopularity[SuccBB] = 0;
1496
1497	for (const auto &PredToDest : PredToDestList)
1498	if (PredToDest.second)
1499	DestPopularity[PredToDest.second]++;
1500
1501	// Find the most popular dest.
1502	auto MostPopular = llvm::max_element(Range&: DestPopularity, C: llvm::less_second());
1503
1504	// Okay, we have finally picked the most popular destination.
1505	return MostPopular->first;
1506	}
1507
1508	// Try to evaluate the value of V when the control flows from PredPredBB to
1509	// BB->getSinglePredecessor() and then on to BB.
1510	Constant *JumpThreadingPass::evaluateOnPredecessorEdge(BasicBlock *BB,
1511	BasicBlock *PredPredBB,
1512	Value *V) {
1513	BasicBlock *PredBB = BB->getSinglePredecessor();
1514	assert(PredBB && "Expected a single predecessor");
1515
1516	if (Constant *Cst = dyn_cast<Constant>(Val: V)) {
1517	return Cst;
1518	}
1519
1520	// Consult LVI if V is not an instruction in BB or PredBB.
1521	Instruction *I = dyn_cast<Instruction>(Val: V);
1522	if (!I || (I->getParent() != BB && I->getParent() != PredBB)) {
1523	return LVI->getConstantOnEdge(V, FromBB: PredPredBB, ToBB: PredBB, CxtI: nullptr);
1524	}
1525
1526	// Look into a PHI argument.
1527	if (PHINode *PHI = dyn_cast<PHINode>(Val: V)) {
1528	if (PHI->getParent() == PredBB)
1529	return dyn_cast<Constant>(Val: PHI->getIncomingValueForBlock(BB: PredPredBB));
1530	return nullptr;
1531	}
1532
1533	// If we have a CmpInst, try to fold it for each incoming edge into PredBB.
1534	if (CmpInst *CondCmp = dyn_cast<CmpInst>(Val: V)) {
1535	if (CondCmp->getParent() == BB) {
1536	Constant *Op0 =
1537	evaluateOnPredecessorEdge(BB, PredPredBB, V: CondCmp->getOperand(i_nocapture: 0));
1538	Constant *Op1 =
1539	evaluateOnPredecessorEdge(BB, PredPredBB, V: CondCmp->getOperand(i_nocapture: 1));
1540	if (Op0 && Op1) {
1541	return ConstantExpr::getCompare(pred: CondCmp->getPredicate(), C1: Op0, C2: Op1);
1542	}
1543	}
1544	return nullptr;
1545	}
1546
1547	return nullptr;
1548	}
1549
1550	bool JumpThreadingPass::processThreadableEdges(Value *Cond, BasicBlock *BB,
1551	ConstantPreference Preference,
1552	Instruction *CxtI) {
1553	// If threading this would thread across a loop header, don't even try to
1554	// thread the edge.
1555	if (LoopHeaders.count(V: BB))
1556	return false;
1557
1558	PredValueInfoTy PredValues;
1559	if (!computeValueKnownInPredecessors(V: Cond, BB, Result&: PredValues, Preference,
1560	CxtI)) {
1561	// We don't have known values in predecessors. See if we can thread through
1562	// BB and its sole predecessor.
1563	return maybethreadThroughTwoBasicBlocks(BB, Cond);
1564	}
1565
1566	assert(!PredValues.empty() &&
1567	"computeValueKnownInPredecessors returned true with no values");
1568
1569	LLVM_DEBUG(dbgs() << "IN BB: " << *BB;
1570	for (const auto &PredValue : PredValues) {
1571	dbgs() << " BB '" << BB->getName()
1572	<< "': FOUND condition = " << *PredValue.first
1573	<< " for pred '" << PredValue.second->getName() << "'.\n";
1574	});
1575
1576	// Decide what we want to thread through. Convert our list of known values to
1577	// a list of known destinations for each pred. This also discards duplicate
1578	// predecessors and keeps track of the undefined inputs (which are represented
1579	// as a null dest in the PredToDestList).
1580	SmallPtrSet<BasicBlock*, 16> SeenPreds;
1581	SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1582
1583	BasicBlock *OnlyDest = nullptr;
1584	BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1585	Constant *OnlyVal = nullptr;
1586	Constant *MultipleVal = (Constant *)(intptr_t)~0ULL;
1587
1588	for (const auto &PredValue : PredValues) {
1589	BasicBlock *Pred = PredValue.second;
1590	if (!SeenPreds.insert(Ptr: Pred).second)
1591	continue; // Duplicate predecessor entry.
1592
1593	Constant *Val = PredValue.first;
1594
1595	BasicBlock *DestBB;
1596	if (isa<UndefValue>(Val))
1597	DestBB = nullptr;
1598	else if (BranchInst *BI = dyn_cast<BranchInst>(Val: BB->getTerminator())) {
1599	assert(isa<ConstantInt>(Val) && "Expecting a constant integer");
1600	DestBB = BI->getSuccessor(i: cast<ConstantInt>(Val)->isZero());
1601	} else if (SwitchInst *SI = dyn_cast<SwitchInst>(Val: BB->getTerminator())) {
1602	assert(isa<ConstantInt>(Val) && "Expecting a constant integer");
1603	DestBB = SI->findCaseValue(C: cast<ConstantInt>(Val))->getCaseSuccessor();
1604	} else {
1605	assert(isa<IndirectBrInst>(BB->getTerminator())
1606	&& "Unexpected terminator");
1607	assert(isa<BlockAddress>(Val) && "Expecting a constant blockaddress");
1608	DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1609	}
1610
1611	// If we have exactly one destination, remember it for efficiency below.
1612	if (PredToDestList.empty()) {
1613	OnlyDest = DestBB;
1614	OnlyVal = Val;
1615	} else {
1616	if (OnlyDest != DestBB)
1617	OnlyDest = MultipleDestSentinel;
1618	// It possible we have same destination, but different value, e.g. default
1619	// case in switchinst.
1620	if (Val != OnlyVal)
1621	OnlyVal = MultipleVal;
1622	}
1623
1624	// If the predecessor ends with an indirect goto, we can't change its
1625	// destination.
1626	if (isa<IndirectBrInst>(Val: Pred->getTerminator()))
1627	continue;
1628
1629	PredToDestList.emplace_back(Args&: Pred, Args&: DestBB);
1630	}
1631
1632	// If all edges were unthreadable, we fail.
1633	if (PredToDestList.empty())
1634	return false;
1635
1636	// If all the predecessors go to a single known successor, we want to fold,
1637	// not thread. By doing so, we do not need to duplicate the current block and
1638	// also miss potential opportunities in case we dont/cant duplicate.
1639	if (OnlyDest && OnlyDest != MultipleDestSentinel) {
1640	if (BB->hasNPredecessors(N: PredToDestList.size())) {
1641	bool SeenFirstBranchToOnlyDest = false;
1642	std::vector <DominatorTree::UpdateType> Updates;
1643	Updates.reserve(n: BB->getTerminator()->getNumSuccessors() - 1);
1644	for (BasicBlock *SuccBB : successors(BB)) {
1645	if (SuccBB == OnlyDest && !SeenFirstBranchToOnlyDest) {
1646	SeenFirstBranchToOnlyDest = true; // Don't modify the first branch.
1647	} else {
1648	SuccBB->removePredecessor(Pred: BB, KeepOneInputPHIs: true); // This is unreachable successor.
1649	Updates.push_back(x: {DominatorTree::Delete, BB, SuccBB});
1650	}
1651	}
1652
1653	// Finally update the terminator.
1654	Instruction *Term = BB->getTerminator();
1655	BranchInst::Create(IfTrue: OnlyDest, InsertBefore: Term->getIterator());
1656	++NumFolds;
1657	Term->eraseFromParent();
1658	DTU->applyUpdatesPermissive(Updates);
1659	if (auto *BPI = getBPI())
1660	BPI->eraseBlock(BB);
1661
1662	// If the condition is now dead due to the removal of the old terminator,
1663	// erase it.
1664	if (auto *CondInst = dyn_cast<Instruction>(Val: Cond)) {
1665	if (CondInst->use_empty() && !CondInst->mayHaveSideEffects())
1666	CondInst->eraseFromParent();
1667	// We can safely replace *some* uses of the CondInst if it has
1668	// exactly one value as returned by LVI. RAUW is incorrect in the
1669	// presence of guards and assumes, that have the `Cond` as the use. This
1670	// is because we use the guards/assume to reason about the `Cond` value
1671	// at the end of block, but RAUW unconditionally replaces all uses
1672	// including the guards/assumes themselves and the uses before the
1673	// guard/assume.
1674	else if (OnlyVal && OnlyVal != MultipleVal)
1675	replaceFoldableUses(Cond: CondInst, ToVal: OnlyVal, KnownAtEndOfBB: BB);
1676	}
1677	return true;
1678	}
1679	}
1680
1681	// Determine which is the most common successor. If we have many inputs and
1682	// this block is a switch, we want to start by threading the batch that goes
1683	// to the most popular destination first. If we only know about one
1684	// threadable destination (the common case) we can avoid this.
1685	BasicBlock *MostPopularDest = OnlyDest;
1686
1687	if (MostPopularDest == MultipleDestSentinel) {
1688	// Remove any loop headers from the Dest list, threadEdge conservatively
1689	// won't process them, but we might have other destination that are eligible
1690	// and we still want to process.
1691	erase_if(C&: PredToDestList,
1692	P: [&](const std::pair<BasicBlock *, BasicBlock *> &PredToDest) {
1693	return LoopHeaders.contains(V: PredToDest.second);
1694	});
1695
1696	if (PredToDestList.empty())
1697	return false;
1698
1699	MostPopularDest = findMostPopularDest(BB, PredToDestList);
1700	}
1701
1702	// Now that we know what the most popular destination is, factor all
1703	// predecessors that will jump to it into a single predecessor.
1704	SmallVector<BasicBlock*, 16> PredsToFactor;
1705	for (const auto &PredToDest : PredToDestList)
1706	if (PredToDest.second == MostPopularDest) {
1707	BasicBlock *Pred = PredToDest.first;
1708
1709	// This predecessor may be a switch or something else that has multiple
1710	// edges to the block. Factor each of these edges by listing them
1711	// according to # occurrences in PredsToFactor.
1712	for (BasicBlock *Succ : successors(BB: Pred))
1713	if (Succ == BB)
1714	PredsToFactor.push_back(Elt: Pred);
1715	}
1716
1717	// If the threadable edges are branching on an undefined value, we get to pick
1718	// the destination that these predecessors should get to.
1719	if (!MostPopularDest)
1720	MostPopularDest = BB->getTerminator()->
1721	getSuccessor(Idx: getBestDestForJumpOnUndef(BB));
1722
1723	// Ok, try to thread it!
1724	return tryThreadEdge(BB, PredBBs: PredsToFactor, SuccBB: MostPopularDest);
1725	}
1726
1727	/// processBranchOnPHI - We have an otherwise unthreadable conditional branch on
1728	/// a PHI node (or freeze PHI) in the current block. See if there are any
1729	/// simplifications we can do based on inputs to the phi node.
1730	bool JumpThreadingPass::processBranchOnPHI(PHINode *PN) {
1731	BasicBlock *BB = PN->getParent();
1732
1733	// TODO: We could make use of this to do it once for blocks with common PHI
1734	// values.
1735	SmallVector<BasicBlock*, 1> PredBBs;
1736	PredBBs.resize(N: 1);
1737
1738	// If any of the predecessor blocks end in an unconditional branch, we can
1739	// *duplicate* the conditional branch into that block in order to further
1740	// encourage jump threading and to eliminate cases where we have branch on a
1741	// phi of an icmp (branch on icmp is much better).
1742	// This is still beneficial when a frozen phi is used as the branch condition
1743	// because it allows CodeGenPrepare to further canonicalize br(freeze(icmp))
1744	// to br(icmp(freeze ...)).
1745	for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1746	BasicBlock *PredBB = PN->getIncomingBlock(i);
1747	if (BranchInst *PredBr = dyn_cast<BranchInst>(Val: PredBB->getTerminator()))
1748	if (PredBr->isUnconditional()) {
1749	PredBBs[0] = PredBB;
1750	// Try to duplicate BB into PredBB.
1751	if (duplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1752	return true;
1753	}
1754	}
1755
1756	return false;
1757	}
1758
1759	/// processBranchOnXOR - We have an otherwise unthreadable conditional branch on
1760	/// a xor instruction in the current block. See if there are any
1761	/// simplifications we can do based on inputs to the xor.
1762	bool JumpThreadingPass::processBranchOnXOR(BinaryOperator *BO) {
1763	BasicBlock *BB = BO->getParent();
1764
1765	// If either the LHS or RHS of the xor is a constant, don't do this
1766	// optimization.
1767	if (isa<ConstantInt>(Val: BO->getOperand(i_nocapture: 0)) ||
1768	isa<ConstantInt>(Val: BO->getOperand(i_nocapture: 1)))
1769	return false;
1770
1771	// If the first instruction in BB isn't a phi, we won't be able to infer
1772	// anything special about any particular predecessor.
1773	if (!isa<PHINode>(Val: BB->front()))
1774	return false;
1775
1776	// If this BB is a landing pad, we won't be able to split the edge into it.
1777	if (BB->isEHPad())
1778	return false;
1779
1780	// If we have a xor as the branch input to this block, and we know that the
1781	// LHS or RHS of the xor in any predecessor is true/false, then we can clone
1782	// the condition into the predecessor and fix that value to true, saving some
1783	// logical ops on that path and encouraging other paths to simplify.
1784	//
1785	// This copies something like this:
1786	//
1787	// BB:
1788	// %X = phi i1 [1], [%X']
1789	// %Y = icmp eq i32 %A, %B
1790	// %Z = xor i1 %X, %Y
1791	// br i1 %Z, ...
1792	//
1793	// Into:
1794	// BB':
1795	// %Y = icmp ne i32 %A, %B
1796	// br i1 %Y, ...
1797
1798	PredValueInfoTy XorOpValues;
1799	bool isLHS = true;
1800	if (!computeValueKnownInPredecessors(V: BO->getOperand(i_nocapture: 0), BB, Result&: XorOpValues,
1801	Preference: WantInteger, CxtI: BO)) {
1802	assert(XorOpValues.empty());
1803	if (!computeValueKnownInPredecessors(V: BO->getOperand(i_nocapture: 1), BB, Result&: XorOpValues,
1804	Preference: WantInteger, CxtI: BO))
1805	return false;
1806	isLHS = false;
1807	}
1808
1809	assert(!XorOpValues.empty() &&
1810	"computeValueKnownInPredecessors returned true with no values");
1811
1812	// Scan the information to see which is most popular: true or false. The
1813	// predecessors can be of the set true, false, or undef.
1814	unsigned NumTrue = 0, NumFalse = 0;
1815	for (const auto &XorOpValue : XorOpValues) {
1816	if (isa<UndefValue>(Val: XorOpValue.first))
1817	// Ignore undefs for the count.
1818	continue;
1819	if (cast<ConstantInt>(Val: XorOpValue.first)->isZero())
1820	++NumFalse;
1821	else
1822	++NumTrue;
1823	}
1824
1825	// Determine which value to split on, true, false, or undef if neither.
1826	ConstantInt *SplitVal = nullptr;
1827	if (NumTrue > NumFalse)
1828	SplitVal = ConstantInt::getTrue(Context&: BB->getContext());
1829	else if (NumTrue != 0 || NumFalse != 0)
1830	SplitVal = ConstantInt::getFalse(Context&: BB->getContext());
1831
1832	// Collect all of the blocks that this can be folded into so that we can
1833	// factor this once and clone it once.
1834	SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1835	for (const auto &XorOpValue : XorOpValues) {
1836	if (XorOpValue.first != SplitVal && !isa<UndefValue>(Val: XorOpValue.first))
1837	continue;
1838
1839	BlocksToFoldInto.push_back(Elt: XorOpValue.second);
1840	}
1841
1842	// If we inferred a value for all of the predecessors, then duplication won't
1843	// help us. However, we can just replace the LHS or RHS with the constant.
1844	if (BlocksToFoldInto.size() ==
1845	cast<PHINode>(Val&: BB->front()).getNumIncomingValues()) {
1846	if (!SplitVal) {
1847	// If all preds provide undef, just nuke the xor, because it is undef too.
1848	BO->replaceAllUsesWith(V: UndefValue::get(T: BO->getType()));
1849	BO->eraseFromParent();
1850	} else if (SplitVal->isZero() && BO != BO->getOperand(i_nocapture: isLHS)) {
1851	// If all preds provide 0, replace the xor with the other input.
1852	BO->replaceAllUsesWith(V: BO->getOperand(i_nocapture: isLHS));
1853	BO->eraseFromParent();
1854	} else {
1855	// If all preds provide 1, set the computed value to 1.
1856	BO->setOperand(i_nocapture: !isLHS, Val_nocapture: SplitVal);
1857	}
1858
1859	return true;
1860	}
1861
1862	// If any of predecessors end with an indirect goto, we can't change its
1863	// destination.
1864	if (any_of(Range&: BlocksToFoldInto, P: [](BasicBlock *Pred) {
1865	return isa<IndirectBrInst>(Val: Pred->getTerminator());
1866	}))
1867	return false;
1868
1869	// Try to duplicate BB into PredBB.
1870	return duplicateCondBranchOnPHIIntoPred(BB, PredBBs: BlocksToFoldInto);
1871	}
1872
1873	/// addPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1874	/// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1875	/// NewPred using the entries from OldPred (suitably mapped).
1876	static void addPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1877	BasicBlock *OldPred,
1878	BasicBlock *NewPred,
1879	DenseMap<Instruction*, Value*> &ValueMap) {
1880	for (PHINode &PN : PHIBB->phis()) {
1881	// Ok, we have a PHI node. Figure out what the incoming value was for the
1882	// DestBlock.
1883	Value *IV = PN.getIncomingValueForBlock(BB: OldPred);
1884
1885	// Remap the value if necessary.
1886	if (Instruction *Inst = dyn_cast<Instruction>(Val: IV)) {
1887	DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Val: Inst);
1888	if (I != ValueMap.end())
1889	IV = I->second;
1890	}
1891
1892	PN.addIncoming(V: IV, BB: NewPred);
1893	}
1894	}
1895
1896	/// Merge basic block BB into its sole predecessor if possible.
1897	bool JumpThreadingPass::maybeMergeBasicBlockIntoOnlyPred(BasicBlock *BB) {
1898	BasicBlock *SinglePred = BB->getSinglePredecessor();
1899	if (!SinglePred)
1900	return false;
1901
1902	const Instruction *TI = SinglePred->getTerminator();
1903	if (TI->isSpecialTerminator() || TI->getNumSuccessors() != 1 ||
1904	SinglePred == BB || hasAddressTakenAndUsed(BB))
1905	return false;
1906
1907	// If SinglePred was a loop header, BB becomes one.
1908	if (LoopHeaders.erase(V: SinglePred))
1909	LoopHeaders.insert(V: BB);
1910
1911	LVI->eraseBlock(BB: SinglePred);
1912	MergeBasicBlockIntoOnlyPred(BB, DTU: DTU.get());
1913
1914	// Now that BB is merged into SinglePred (i.e. SinglePred code followed by
1915	// BB code within one basic block `BB`), we need to invalidate the LVI
1916	// information associated with BB, because the LVI information need not be
1917	// true for all of BB after the merge. For example,
1918	// Before the merge, LVI info and code is as follows:
1919	// SinglePred: <LVI info1 for %p val>
1920	// %y = use of %p
1921	// call @exit() // need not transfer execution to successor.
1922	// assume(%p) // from this point on %p is true
1923	// br label %BB
1924	// BB: <LVI info2 for %p val, i.e. %p is true>
1925	// %x = use of %p
1926	// br label exit
1927	//
1928	// Note that this LVI info for blocks BB and SinglPred is correct for %p
1929	// (info2 and info1 respectively). After the merge and the deletion of the
1930	// LVI info1 for SinglePred. We have the following code:
1931	// BB: <LVI info2 for %p val>
1932	// %y = use of %p
1933	// call @exit()
1934	// assume(%p)
1935	// %x = use of %p <-- LVI info2 is correct from here onwards.
1936	// br label exit
1937	// LVI info2 for BB is incorrect at the beginning of BB.
1938
1939	// Invalidate LVI information for BB if the LVI is not provably true for
1940	// all of BB.
1941	if (!isGuaranteedToTransferExecutionToSuccessor(BB))
1942	LVI->eraseBlock(BB);
1943	return true;
1944	}
1945
1946	/// Update the SSA form. NewBB contains instructions that are copied from BB.
1947	/// ValueMapping maps old values in BB to new ones in NewBB.
1948	void JumpThreadingPass::updateSSA(
1949	BasicBlock *BB, BasicBlock *NewBB,
1950	DenseMap<Instruction *, Value *> &ValueMapping) {
1951	// If there were values defined in BB that are used outside the block, then we
1952	// now have to update all uses of the value to use either the original value,
1953	// the cloned value, or some PHI derived value. This can require arbitrary
1954	// PHI insertion, of which we are prepared to do, clean these up now.
1955	SSAUpdater SSAUpdate;
1956	SmallVector<Use *, 16> UsesToRename;
1957	SmallVector<DbgValueInst *, 4> DbgValues;
1958	SmallVector<DbgVariableRecord *, 4> DbgVariableRecords;
1959
1960	for (Instruction &I : *BB) {
1961	// Scan all uses of this instruction to see if it is used outside of its
1962	// block, and if so, record them in UsesToRename.
1963	for (Use &U : I.uses()) {
1964	Instruction *User = cast<Instruction>(Val: U.getUser());
1965	if (PHINode *UserPN = dyn_cast<PHINode>(Val: User)) {
1966	if (UserPN->getIncomingBlock(U) == BB)
1967	continue;
1968	} else if (User->getParent() == BB)
1969	continue;
1970
1971	UsesToRename.push_back(Elt: &U);
1972	}
1973
1974	// Find debug values outside of the block
1975	findDbgValues(DbgValues, V: &I, DbgVariableRecords: &DbgVariableRecords);
1976	llvm::erase_if(C&: DbgValues, P: [&](const DbgValueInst *DbgVal) {
1977	return DbgVal->getParent() == BB;
1978	});
1979	llvm::erase_if(C&: DbgVariableRecords, P: [&](const DbgVariableRecord *DbgVarRec) {
1980	return DbgVarRec->getParent() == BB;
1981	});
1982
1983	// If there are no uses outside the block, we're done with this instruction.
1984	if (UsesToRename.empty() && DbgValues.empty() && DbgVariableRecords.empty())
1985	continue;
1986	LLVM_DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n");
1987
1988	// We found a use of I outside of BB. Rename all uses of I that are outside
1989	// its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1990	// with the two values we know.
1991	SSAUpdate.Initialize(Ty: I.getType(), Name: I.getName());
1992	SSAUpdate.AddAvailableValue(BB, V: &I);
1993	SSAUpdate.AddAvailableValue(BB: NewBB, V: ValueMapping[&I]);
1994
1995	while (!UsesToRename.empty())
1996	SSAUpdate.RewriteUse(U&: *UsesToRename.pop_back_val());
1997	if (!DbgValues.empty() || !DbgVariableRecords.empty()) {
1998	SSAUpdate.UpdateDebugValues(I: &I, DbgValues);
1999	SSAUpdate.UpdateDebugValues(I: &I, DbgValues&: DbgVariableRecords);
2000	DbgValues.clear();
2001	DbgVariableRecords.clear();
2002	}
2003
2004	LLVM_DEBUG(dbgs() << "\n");
2005	}
2006	}
2007
2008	/// Clone instructions in range [BI, BE) to NewBB. For PHI nodes, we only clone
2009	/// arguments that come from PredBB. Return the map from the variables in the
2010	/// source basic block to the variables in the newly created basic block.
2011	DenseMap<Instruction *, Value *>
2012	JumpThreadingPass::cloneInstructions(BasicBlock::iterator BI,
2013	BasicBlock::iterator BE, BasicBlock *NewBB,
2014	BasicBlock *PredBB) {
2015	// We are going to have to map operands from the source basic block to the new
2016	// copy of the block 'NewBB'. If there are PHI nodes in the source basic
2017	// block, evaluate them to account for entry from PredBB.
2018	DenseMap<Instruction *, Value *> ValueMapping;
2019
2020	// Retargets llvm.dbg.value to any renamed variables.
2021	auto RetargetDbgValueIfPossible = [&](Instruction *NewInst) -> bool {
2022	auto DbgInstruction = dyn_cast<DbgValueInst>(Val: NewInst);
2023	if (!DbgInstruction)
2024	return false;
2025
2026	SmallSet<std::pair<Value *, Value *>, 16> OperandsToRemap;
2027	for (auto DbgOperand : DbgInstruction->location_ops()) {
2028	auto DbgOperandInstruction = dyn_cast<Instruction>(Val: DbgOperand);
2029	if (!DbgOperandInstruction)
2030	continue;
2031
2032	auto I = ValueMapping.find(Val: DbgOperandInstruction);
2033	if (I != ValueMapping.end()) {
2034	OperandsToRemap.insert(
2035	V: std::pair<Value *, Value *>(DbgOperand, I->second));
2036	}
2037	}
2038
2039	for (auto &[OldOp, MappedOp] : OperandsToRemap)
2040	DbgInstruction->replaceVariableLocationOp(OldValue: OldOp, NewValue: MappedOp);
2041	return true;
2042	};
2043
2044	// Duplicate implementation of the above dbg.value code, using
2045	// DbgVariableRecords instead.
2046	auto RetargetDbgVariableRecordIfPossible = [&](DbgVariableRecord *DVR) {
2047	SmallSet<std::pair<Value *, Value *>, 16> OperandsToRemap;
2048	for (auto *Op : DVR->location_ops()) {
2049	Instruction *OpInst = dyn_cast<Instruction>(Val: Op);
2050	if (!OpInst)
2051	continue;
2052
2053	auto I = ValueMapping.find(Val: OpInst);
2054	if (I != ValueMapping.end())
2055	OperandsToRemap.insert(V: {OpInst, I->second});
2056	}
2057
2058	for (auto &[OldOp, MappedOp] : OperandsToRemap)
2059	DVR->replaceVariableLocationOp(OldValue: OldOp, NewValue: MappedOp);
2060	};
2061
2062	BasicBlock *RangeBB = BI->getParent();
2063
2064	// Clone the phi nodes of the source basic block into NewBB. The resulting
2065	// phi nodes are trivial since NewBB only has one predecessor, but SSAUpdater
2066	// might need to rewrite the operand of the cloned phi.
2067	for (; PHINode *PN = dyn_cast<PHINode>(Val&: BI); ++BI) {
2068	PHINode *NewPN = PHINode::Create(Ty: PN->getType(), NumReservedValues: 1, NameStr: PN->getName(), InsertAtEnd: NewBB);
2069	NewPN->addIncoming(V: PN->getIncomingValueForBlock(BB: PredBB), BB: PredBB);
2070	ValueMapping[PN] = NewPN;
2071	}
2072
2073	// Clone noalias scope declarations in the threaded block. When threading a
2074	// loop exit, we would otherwise end up with two idential scope declarations
2075	// visible at the same time.
2076	SmallVector<MDNode *> NoAliasScopes;
2077	DenseMap<MDNode *, MDNode *> ClonedScopes;
2078	LLVMContext &Context = PredBB->getContext();
2079	identifyNoAliasScopesToClone(Start: BI, End: BE, NoAliasDeclScopes&: NoAliasScopes);
2080	cloneNoAliasScopes(NoAliasDeclScopes: NoAliasScopes, ClonedScopes, Ext: "thread", Context);
2081
2082	auto CloneAndRemapDbgInfo = [&](Instruction *NewInst, Instruction *From) {
2083	auto DVRRange = NewInst->cloneDebugInfoFrom(From);
2084	for (DbgVariableRecord &DVR : filterDbgVars(R: DVRRange))
2085	RetargetDbgVariableRecordIfPossible(&DVR);
2086	};
2087
2088	// Clone the non-phi instructions of the source basic block into NewBB,
2089	// keeping track of the mapping and using it to remap operands in the cloned
2090	// instructions.
2091	for (; BI != BE; ++BI) {
2092	Instruction *New = BI->clone();
2093	New->setName(BI->getName());
2094	New->insertInto(ParentBB: NewBB, It: NewBB->end());
2095	ValueMapping[&*BI] = New;
2096	adaptNoAliasScopes(I: New, ClonedScopes, Context);
2097
2098	CloneAndRemapDbgInfo(New, &*BI);
2099
2100	if (RetargetDbgValueIfPossible(New))
2101	continue;
2102
2103	// Remap operands to patch up intra-block references.
2104	for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2105	if (Instruction *Inst = dyn_cast<Instruction>(Val: New->getOperand(i))) {
2106	DenseMap<Instruction *, Value *>::iterator I = ValueMapping.find(Val: Inst);
2107	if (I != ValueMapping.end())
2108	New->setOperand(i, Val: I->second);
2109	}
2110	}
2111
2112	// There may be DbgVariableRecords on the terminator, clone directly from
2113	// marker to marker as there isn't an instruction there.
2114	if (BE != RangeBB->end() && BE->hasDbgRecords()) {
2115	// Dump them at the end.
2116	DbgMarker *Marker = RangeBB->getMarker(It: BE);
2117	DbgMarker *EndMarker = NewBB->createMarker(It: NewBB->end());
2118	auto DVRRange = EndMarker->cloneDebugInfoFrom(From: Marker, FromHere: std::nullopt);
2119	for (DbgVariableRecord &DVR : filterDbgVars(R: DVRRange))
2120	RetargetDbgVariableRecordIfPossible(&DVR);
2121	}
2122
2123	return ValueMapping;
2124	}
2125
2126	/// Attempt to thread through two successive basic blocks.
2127	bool JumpThreadingPass::maybethreadThroughTwoBasicBlocks(BasicBlock *BB,
2128	Value *Cond) {
2129	// Consider:
2130	//
2131	// PredBB:
2132	// %var = phi i32* [ null, %bb1 ], [ @a, %bb2 ]
2133	// %tobool = icmp eq i32 %cond, 0
2134	// br i1 %tobool, label %BB, label ...
2135	//
2136	// BB:
2137	// %cmp = icmp eq i32* %var, null
2138	// br i1 %cmp, label ..., label ...
2139	//
2140	// We don't know the value of %var at BB even if we know which incoming edge
2141	// we take to BB. However, once we duplicate PredBB for each of its incoming
2142	// edges (say, PredBB1 and PredBB2), we know the value of %var in each copy of
2143	// PredBB. Then we can thread edges PredBB1->BB and PredBB2->BB through BB.
2144
2145	// Require that BB end with a Branch for simplicity.
2146	BranchInst *CondBr = dyn_cast<BranchInst>(Val: BB->getTerminator());
2147	if (!CondBr)
2148	return false;
2149
2150	// BB must have exactly one predecessor.
2151	BasicBlock *PredBB = BB->getSinglePredecessor();
2152	if (!PredBB)
2153	return false;
2154
2155	// Require that PredBB end with a conditional Branch. If PredBB ends with an
2156	// unconditional branch, we should be merging PredBB and BB instead. For
2157	// simplicity, we don't deal with a switch.
2158	BranchInst *PredBBBranch = dyn_cast<BranchInst>(Val: PredBB->getTerminator());
2159	if (!PredBBBranch || PredBBBranch->isUnconditional())
2160	return false;
2161
2162	// If PredBB has exactly one incoming edge, we don't gain anything by copying
2163	// PredBB.
2164	if (PredBB->getSinglePredecessor())
2165	return false;
2166
2167	// Don't thread through PredBB if it contains a successor edge to itself, in
2168	// which case we would infinite loop. Suppose we are threading an edge from
2169	// PredPredBB through PredBB and BB to SuccBB with PredBB containing a
2170	// successor edge to itself. If we allowed jump threading in this case, we
2171	// could duplicate PredBB and BB as, say, PredBB.thread and BB.thread. Since
2172	// PredBB.thread has a successor edge to PredBB, we would immediately come up
2173	// with another jump threading opportunity from PredBB.thread through PredBB
2174	// and BB to SuccBB. This jump threading would repeatedly occur. That is, we
2175	// would keep peeling one iteration from PredBB.
2176	if (llvm::is_contained(Range: successors(BB: PredBB), Element: PredBB))
2177	return false;
2178
2179	// Don't thread across a loop header.
2180	if (LoopHeaders.count(V: PredBB))
2181	return false;
2182
2183	// Avoid complication with duplicating EH pads.
2184	if (PredBB->isEHPad())
2185	return false;
2186
2187	// Find a predecessor that we can thread. For simplicity, we only consider a
2188	// successor edge out of BB to which we thread exactly one incoming edge into
2189	// PredBB.
2190	unsigned ZeroCount = 0;
2191	unsigned OneCount = 0;
2192	BasicBlock *ZeroPred = nullptr;
2193	BasicBlock *OnePred = nullptr;
2194	for (BasicBlock *P : predecessors(BB: PredBB)) {
2195	// If PredPred ends with IndirectBrInst, we can't handle it.
2196	if (isa<IndirectBrInst>(Val: P->getTerminator()))
2197	continue;
2198	if (ConstantInt *CI = dyn_cast_or_null<ConstantInt>(
2199	Val: evaluateOnPredecessorEdge(BB, PredPredBB: P, V: Cond))) {
2200	if (CI->isZero()) {
2201	ZeroCount++;
2202	ZeroPred = P;
2203	} else if (CI->isOne()) {
2204	OneCount++;
2205	OnePred = P;
2206	}
2207	}
2208	}
2209
2210	// Disregard complicated cases where we have to thread multiple edges.
2211	BasicBlock *PredPredBB;
2212	if (ZeroCount == 1) {
2213	PredPredBB = ZeroPred;
2214	} else if (OneCount == 1) {
2215	PredPredBB = OnePred;
2216	} else {
2217	return false;
2218	}
2219
2220	BasicBlock *SuccBB = CondBr->getSuccessor(i: PredPredBB == ZeroPred);
2221
2222	// If threading to the same block as we come from, we would infinite loop.
2223	if (SuccBB == BB) {
2224	LLVM_DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
2225	<< "' - would thread to self!\n");
2226	return false;
2227	}
2228
2229	// If threading this would thread across a loop header, don't thread the edge.
2230	// See the comments above findLoopHeaders for justifications and caveats.
2231	if (LoopHeaders.count(V: BB) || LoopHeaders.count(V: SuccBB)) {
2232	LLVM_DEBUG({
2233	bool BBIsHeader = LoopHeaders.count(BB);
2234	bool SuccIsHeader = LoopHeaders.count(SuccBB);
2235	dbgs() << " Not threading across "
2236	<< (BBIsHeader ? "loop header BB '" : "block BB '")
2237	<< BB->getName() << "' to dest "
2238	<< (SuccIsHeader ? "loop header BB '" : "block BB '")
2239	<< SuccBB->getName()
2240	<< "' - it might create an irreducible loop!\n";
2241	});
2242	return false;
2243	}
2244
2245	// Compute the cost of duplicating BB and PredBB.
2246	unsigned BBCost = getJumpThreadDuplicationCost(
2247	TTI, BB, StopAt: BB->getTerminator(), Threshold: BBDupThreshold);
2248	unsigned PredBBCost = getJumpThreadDuplicationCost(
2249	TTI, BB: PredBB, StopAt: PredBB->getTerminator(), Threshold: BBDupThreshold);
2250
2251	// Give up if costs are too high. We need to check BBCost and PredBBCost
2252	// individually before checking their sum because getJumpThreadDuplicationCost
2253	// return (unsigned)~0 for those basic blocks that cannot be duplicated.
2254	if (BBCost > BBDupThreshold || PredBBCost > BBDupThreshold ||
2255	BBCost + PredBBCost > BBDupThreshold) {
2256	LLVM_DEBUG(dbgs() << " Not threading BB '" << BB->getName()
2257	<< "' - Cost is too high: " << PredBBCost
2258	<< " for PredBB, " << BBCost << "for BB\n");
2259	return false;
2260	}
2261
2262	// Now we are ready to duplicate PredBB.
2263	threadThroughTwoBasicBlocks(PredPredBB, PredBB, BB, SuccBB);
2264	return true;
2265	}
2266
2267	void JumpThreadingPass::threadThroughTwoBasicBlocks(BasicBlock *PredPredBB,
2268	BasicBlock *PredBB,
2269	BasicBlock *BB,
2270	BasicBlock *SuccBB) {
2271	LLVM_DEBUG(dbgs() << " Threading through '" << PredBB->getName() << "' and '"
2272	<< BB->getName() << "'\n");
2273
2274	// Build BPI/BFI before any changes are made to IR.
2275	bool HasProfile = doesBlockHaveProfileData(BB);
2276	auto *BFI = getOrCreateBFI(Force: HasProfile);
2277	auto *BPI = getOrCreateBPI(Force: BFI != nullptr);
2278
2279	BranchInst *CondBr = cast<BranchInst>(Val: BB->getTerminator());
2280	BranchInst *PredBBBranch = cast<BranchInst>(Val: PredBB->getTerminator());
2281
2282	BasicBlock *NewBB =
2283	BasicBlock::Create(Context&: PredBB->getContext(), Name: PredBB->getName() + ".thread",
2284	Parent: PredBB->getParent(), InsertBefore: PredBB);
2285	NewBB->moveAfter(MovePos: PredBB);
2286
2287	// Set the block frequency of NewBB.
2288	if (BFI) {
2289	assert(BPI && "It's expected BPI to exist along with BFI");
2290	auto NewBBFreq = BFI->getBlockFreq(BB: PredPredBB) *
2291	BPI->getEdgeProbability(Src: PredPredBB, Dst: PredBB);
2292	BFI->setBlockFreq(BB: NewBB, Freq: NewBBFreq);
2293	}
2294
2295	// We are going to have to map operands from the original BB block to the new
2296	// copy of the block 'NewBB'. If there are PHI nodes in PredBB, evaluate them
2297	// to account for entry from PredPredBB.
2298	DenseMap<Instruction *, Value *> ValueMapping =
2299	cloneInstructions(BI: PredBB->begin(), BE: PredBB->end(), NewBB, PredBB: PredPredBB);
2300
2301	// Copy the edge probabilities from PredBB to NewBB.
2302	if (BPI)
2303	BPI->copyEdgeProbabilities(Src: PredBB, Dst: NewBB);
2304
2305	// Update the terminator of PredPredBB to jump to NewBB instead of PredBB.
2306	// This eliminates predecessors from PredPredBB, which requires us to simplify
2307	// any PHI nodes in PredBB.
2308	Instruction *PredPredTerm = PredPredBB->getTerminator();
2309	for (unsigned i = 0, e = PredPredTerm->getNumSuccessors(); i != e; ++i)
2310	if (PredPredTerm->getSuccessor(Idx: i) == PredBB) {
2311	PredBB->removePredecessor(Pred: PredPredBB, KeepOneInputPHIs: true);
2312	PredPredTerm->setSuccessor(Idx: i, BB: NewBB);
2313	}
2314
2315	addPHINodeEntriesForMappedBlock(PHIBB: PredBBBranch->getSuccessor(i: 0), OldPred: PredBB, NewPred: NewBB,
2316	ValueMap&: ValueMapping);
2317	addPHINodeEntriesForMappedBlock(PHIBB: PredBBBranch->getSuccessor(i: 1), OldPred: PredBB, NewPred: NewBB,
2318	ValueMap&: ValueMapping);
2319
2320	DTU->applyUpdatesPermissive(
2321	Updates: {{DominatorTree::Insert, NewBB, CondBr->getSuccessor(i: 0)},
2322	{DominatorTree::Insert, NewBB, CondBr->getSuccessor(i: 1)},
2323	{DominatorTree::Insert, PredPredBB, NewBB},
2324	{DominatorTree::Delete, PredPredBB, PredBB}});
2325
2326	updateSSA(BB: PredBB, NewBB, ValueMapping);
2327
2328	// Clean up things like PHI nodes with single operands, dead instructions,
2329	// etc.
2330	SimplifyInstructionsInBlock(BB: NewBB, TLI);
2331	SimplifyInstructionsInBlock(BB: PredBB, TLI);
2332
2333	SmallVector<BasicBlock *, 1> PredsToFactor;
2334	PredsToFactor.push_back(Elt: NewBB);
2335	threadEdge(BB, PredBBs: PredsToFactor, SuccBB);
2336	}
2337
2338	/// tryThreadEdge - Thread an edge if it's safe and profitable to do so.
2339	bool JumpThreadingPass::tryThreadEdge(
2340	BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs,
2341	BasicBlock *SuccBB) {
2342	// If threading to the same block as we come from, we would infinite loop.
2343	if (SuccBB == BB) {
2344	LLVM_DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
2345	<< "' - would thread to self!\n");
2346	return false;
2347	}
2348
2349	// If threading this would thread across a loop header, don't thread the edge.
2350	// See the comments above findLoopHeaders for justifications and caveats.
2351	if (LoopHeaders.count(V: BB) || LoopHeaders.count(V: SuccBB)) {
2352	LLVM_DEBUG({
2353	bool BBIsHeader = LoopHeaders.count(BB);
2354	bool SuccIsHeader = LoopHeaders.count(SuccBB);
2355	dbgs() << " Not threading across "
2356	<< (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName()
2357	<< "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '")
2358	<< SuccBB->getName() << "' - it might create an irreducible loop!\n";
2359	});
2360	return false;
2361	}
2362
2363	unsigned JumpThreadCost = getJumpThreadDuplicationCost(
2364	TTI, BB, StopAt: BB->getTerminator(), Threshold: BBDupThreshold);
2365	if (JumpThreadCost > BBDupThreshold) {
2366	LLVM_DEBUG(dbgs() << " Not threading BB '" << BB->getName()
2367	<< "' - Cost is too high: " << JumpThreadCost << "\n");
2368	return false;
2369	}
2370
2371	threadEdge(BB, PredBBs, SuccBB);
2372	return true;
2373	}
2374
2375	/// threadEdge - We have decided that it is safe and profitable to factor the
2376	/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
2377	/// across BB. Transform the IR to reflect this change.
2378	void JumpThreadingPass::threadEdge(BasicBlock *BB,
2379	const SmallVectorImpl<BasicBlock *> &PredBBs,
2380	BasicBlock *SuccBB) {
2381	assert(SuccBB != BB && "Don't create an infinite loop");
2382
2383	assert(!LoopHeaders.count(BB) && !LoopHeaders.count(SuccBB) &&
2384	"Don't thread across loop headers");
2385
2386	// Build BPI/BFI before any changes are made to IR.
2387	bool HasProfile = doesBlockHaveProfileData(BB);
2388	auto *BFI = getOrCreateBFI(Force: HasProfile);
2389	auto *BPI = getOrCreateBPI(Force: BFI != nullptr);
2390
2391	// And finally, do it! Start by factoring the predecessors if needed.
2392	BasicBlock *PredBB;
2393	if (PredBBs.size() == 1)
2394	PredBB = PredBBs[0];
2395	else {
2396	LLVM_DEBUG(dbgs() << " Factoring out " << PredBBs.size()
2397	<< " common predecessors.\n");
2398	PredBB = splitBlockPreds(BB, Preds: PredBBs, Suffix: ".thr_comm");
2399	}
2400
2401	// And finally, do it!
2402	LLVM_DEBUG(dbgs() << " Threading edge from '" << PredBB->getName()
2403	<< "' to '" << SuccBB->getName()
2404	<< ", across block:\n " << *BB << "\n");
2405
2406	LVI->threadEdge(PredBB, OldSucc: BB, NewSucc: SuccBB);
2407
2408	BasicBlock *NewBB = BasicBlock::Create(Context&: BB->getContext(),
2409	Name: BB->getName()+".thread",
2410	Parent: BB->getParent(), InsertBefore: BB);
2411	NewBB->moveAfter(MovePos: PredBB);
2412
2413	// Set the block frequency of NewBB.
2414	if (BFI) {
2415	assert(BPI && "It's expected BPI to exist along with BFI");
2416	auto NewBBFreq =
2417	BFI->getBlockFreq(BB: PredBB) * BPI->getEdgeProbability(Src: PredBB, Dst: BB);
2418	BFI->setBlockFreq(BB: NewBB, Freq: NewBBFreq);
2419	}
2420
2421	// Copy all the instructions from BB to NewBB except the terminator.
2422	DenseMap<Instruction *, Value *> ValueMapping =
2423	cloneInstructions(BI: BB->begin(), BE: std::prev(x: BB->end()), NewBB, PredBB);
2424
2425	// We didn't copy the terminator from BB over to NewBB, because there is now
2426	// an unconditional jump to SuccBB. Insert the unconditional jump.
2427	BranchInst *NewBI = BranchInst::Create(IfTrue: SuccBB, InsertAtEnd: NewBB);
2428	NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
2429
2430	// Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
2431	// PHI nodes for NewBB now.
2432	addPHINodeEntriesForMappedBlock(PHIBB: SuccBB, OldPred: BB, NewPred: NewBB, ValueMap&: ValueMapping);
2433
2434	// Update the terminator of PredBB to jump to NewBB instead of BB. This
2435	// eliminates predecessors from BB, which requires us to simplify any PHI
2436	// nodes in BB.
2437	Instruction *PredTerm = PredBB->getTerminator();
2438	for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
2439	if (PredTerm->getSuccessor(Idx: i) == BB) {
2440	BB->removePredecessor(Pred: PredBB, KeepOneInputPHIs: true);
2441	PredTerm->setSuccessor(Idx: i, BB: NewBB);
2442	}
2443
2444	// Enqueue required DT updates.
2445	DTU->applyUpdatesPermissive(Updates: {{DominatorTree::Insert, NewBB, SuccBB},
2446	{DominatorTree::Insert, PredBB, NewBB},
2447	{DominatorTree::Delete, PredBB, BB}});
2448
2449	updateSSA(BB, NewBB, ValueMapping);
2450
2451	// At this point, the IR is fully up to date and consistent. Do a quick scan
2452	// over the new instructions and zap any that are constants or dead. This
2453	// frequently happens because of phi translation.
2454	SimplifyInstructionsInBlock(BB: NewBB, TLI);
2455
2456	// Update the edge weight from BB to SuccBB, which should be less than before.
2457	updateBlockFreqAndEdgeWeight(PredBB, BB, NewBB, SuccBB, BFI, BPI, HasProfile);
2458
2459	// Threaded an edge!
2460	++NumThreads;
2461	}
2462
2463	/// Create a new basic block that will be the predecessor of BB and successor of
2464	/// all blocks in Preds. When profile data is available, update the frequency of
2465	/// this new block.
2466	BasicBlock *JumpThreadingPass::splitBlockPreds(BasicBlock *BB,
2467	ArrayRef<BasicBlock *> Preds,
2468	const char *Suffix) {
2469	SmallVector<BasicBlock *, 2> NewBBs;
2470
2471	// Collect the frequencies of all predecessors of BB, which will be used to
2472	// update the edge weight of the result of splitting predecessors.
2473	DenseMap<BasicBlock *, BlockFrequency> FreqMap;
2474	auto *BFI = getBFI();
2475	if (BFI) {
2476	auto *BPI = getOrCreateBPI(Force: true);
2477	for (auto *Pred : Preds)
2478	FreqMap.insert(KV: std::make_pair(
2479	x&: Pred, y: BFI->getBlockFreq(BB: Pred) * BPI->getEdgeProbability(Src: Pred, Dst: BB)));
2480	}
2481
2482	// In the case when BB is a LandingPad block we create 2 new predecessors
2483	// instead of just one.
2484	if (BB->isLandingPad()) {
2485	std::string NewName = std::string(Suffix) + ".split-lp";
2486	SplitLandingPadPredecessors(OrigBB: BB, Preds, Suffix, Suffix2: NewName.c_str(), NewBBs);
2487	} else {
2488	NewBBs.push_back(Elt: SplitBlockPredecessors(BB, Preds, Suffix));
2489	}
2490
2491	std::vector<DominatorTree::UpdateType> Updates;
2492	Updates.reserve(n: (2 * Preds.size()) + NewBBs.size());
2493	for (auto *NewBB : NewBBs) {
2494	BlockFrequency NewBBFreq(0);
2495	Updates.push_back(x: {DominatorTree::Insert, NewBB, BB});
2496	for (auto *Pred : predecessors(BB: NewBB)) {
2497	Updates.push_back(x: {DominatorTree::Delete, Pred, BB});
2498	Updates.push_back(x: {DominatorTree::Insert, Pred, NewBB});
2499	if (BFI) // Update frequencies between Pred -> NewBB.
2500	NewBBFreq += FreqMap.lookup(Val: Pred);
2501	}
2502	if (BFI) // Apply the summed frequency to NewBB.
2503	BFI->setBlockFreq(BB: NewBB, Freq: NewBBFreq);
2504	}
2505
2506	DTU->applyUpdatesPermissive(Updates);
2507	return NewBBs[0];
2508	}
2509
2510	bool JumpThreadingPass::doesBlockHaveProfileData(BasicBlock *BB) {
2511	const Instruction *TI = BB->getTerminator();
2512	if (!TI || TI->getNumSuccessors() < 2)
2513	return false;
2514
2515	return hasValidBranchWeightMD(I: *TI);
2516	}
2517
2518	/// Update the block frequency of BB and branch weight and the metadata on the
2519	/// edge BB->SuccBB. This is done by scaling the weight of BB->SuccBB by 1 -
2520	/// Freq(PredBB->BB) / Freq(BB->SuccBB).
2521	void JumpThreadingPass::updateBlockFreqAndEdgeWeight(BasicBlock *PredBB,
2522	BasicBlock *BB,
2523	BasicBlock *NewBB,
2524	BasicBlock *SuccBB,
2525	BlockFrequencyInfo *BFI,
2526	BranchProbabilityInfo *BPI,
2527	bool HasProfile) {
2528	assert(((BFI && BPI) || (!BFI && !BFI)) &&
2529	"Both BFI & BPI should either be set or unset");
2530
2531	if (!BFI) {
2532	assert(!HasProfile &&
2533	"It's expected to have BFI/BPI when profile info exists");
2534	return;
2535	}
2536
2537	// As the edge from PredBB to BB is deleted, we have to update the block
2538	// frequency of BB.
2539	auto BBOrigFreq = BFI->getBlockFreq(BB);
2540	auto NewBBFreq = BFI->getBlockFreq(BB: NewBB);
2541	auto BB2SuccBBFreq = BBOrigFreq * BPI->getEdgeProbability(Src: BB, Dst: SuccBB);
2542	auto BBNewFreq = BBOrigFreq - NewBBFreq;
2543	BFI->setBlockFreq(BB, Freq: BBNewFreq);
2544
2545	// Collect updated outgoing edges' frequencies from BB and use them to update
2546	// edge probabilities.
2547	SmallVector<uint64_t, 4> BBSuccFreq;
2548	for (BasicBlock *Succ : successors(BB)) {
2549	auto SuccFreq = (Succ == SuccBB)
2550	? BB2SuccBBFreq - NewBBFreq
2551	: BBOrigFreq * BPI->getEdgeProbability(Src: BB, Dst: Succ);
2552	BBSuccFreq.push_back(Elt: SuccFreq.getFrequency());
2553	}
2554
2555	uint64_t MaxBBSuccFreq = *llvm::max_element(Range&: BBSuccFreq);
2556
2557	SmallVector<BranchProbability, 4> BBSuccProbs;
2558	if (MaxBBSuccFreq == 0)
2559	BBSuccProbs.assign(NumElts: BBSuccFreq.size(),
2560	Elt: {1, static_cast<unsigned>(BBSuccFreq.size())});
2561	else {
2562	for (uint64_t Freq : BBSuccFreq)
2563	BBSuccProbs.push_back(
2564	Elt: BranchProbability::getBranchProbability(Numerator: Freq, Denominator: MaxBBSuccFreq));
2565	// Normalize edge probabilities so that they sum up to one.
2566	BranchProbability::normalizeProbabilities(Begin: BBSuccProbs.begin(),
2567	End: BBSuccProbs.end());
2568	}
2569
2570	// Update edge probabilities in BPI.
2571	BPI->setEdgeProbability(Src: BB, Probs: BBSuccProbs);
2572
2573	// Update the profile metadata as well.
2574	//
2575	// Don't do this if the profile of the transformed blocks was statically
2576	// estimated. (This could occur despite the function having an entry
2577	// frequency in completely cold parts of the CFG.)
2578	//
2579	// In this case we don't want to suggest to subsequent passes that the
2580	// calculated weights are fully consistent. Consider this graph:
2581	//
2582	// check_1
2583	// 50% / |
2584	// eq_1 | 50%
2585	// \ |
2586	// check_2
2587	// 50% / |
2588	// eq_2 | 50%
2589	// \ |
2590	// check_3
2591	// 50% / |
2592	// eq_3 | 50%
2593	// \ |
2594	//
2595	// Assuming the blocks check_* all compare the same value against 1, 2 and 3,
2596	// the overall probabilities are inconsistent; the total probability that the
2597	// value is either 1, 2 or 3 is 150%.
2598	//
2599	// As a consequence if we thread eq_1 -> check_2 to check_3, check_2->check_3
2600	// becomes 0%. This is even worse if the edge whose probability becomes 0% is
2601	// the loop exit edge. Then based solely on static estimation we would assume
2602	// the loop was extremely hot.
2603	//
2604	// FIXME this locally as well so that BPI and BFI are consistent as well. We
2605	// shouldn't make edges extremely likely or unlikely based solely on static
2606	// estimation.
2607	if (BBSuccProbs.size() >= 2 && HasProfile) {
2608	SmallVector<uint32_t, 4> Weights;
2609	for (auto Prob : BBSuccProbs)
2610	Weights.push_back(Elt: Prob.getNumerator());
2611
2612	auto TI = BB->getTerminator();
2613	setBranchWeights(I&: *TI, Weights);
2614	}
2615	}
2616
2617	/// duplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
2618	/// to BB which contains an i1 PHI node and a conditional branch on that PHI.
2619	/// If we can duplicate the contents of BB up into PredBB do so now, this
2620	/// improves the odds that the branch will be on an analyzable instruction like
2621	/// a compare.
2622	bool JumpThreadingPass::duplicateCondBranchOnPHIIntoPred(
2623	BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs) {
2624	assert(!PredBBs.empty() && "Can't handle an empty set");
2625
2626	// If BB is a loop header, then duplicating this block outside the loop would
2627	// cause us to transform this into an irreducible loop, don't do this.
2628	// See the comments above findLoopHeaders for justifications and caveats.
2629	if (LoopHeaders.count(V: BB)) {
2630	LLVM_DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
2631	<< "' into predecessor block '" << PredBBs[0]->getName()
2632	<< "' - it might create an irreducible loop!\n");
2633	return false;
2634	}
2635
2636	unsigned DuplicationCost = getJumpThreadDuplicationCost(
2637	TTI, BB, StopAt: BB->getTerminator(), Threshold: BBDupThreshold);
2638	if (DuplicationCost > BBDupThreshold) {
2639	LLVM_DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
2640	<< "' - Cost is too high: " << DuplicationCost << "\n");
2641	return false;
2642	}
2643
2644	// And finally, do it! Start by factoring the predecessors if needed.
2645	std::vector<DominatorTree::UpdateType> Updates;
2646	BasicBlock *PredBB;
2647	if (PredBBs.size() == 1)
2648	PredBB = PredBBs[0];
2649	else {
2650	LLVM_DEBUG(dbgs() << " Factoring out " << PredBBs.size()
2651	<< " common predecessors.\n");
2652	PredBB = splitBlockPreds(BB, Preds: PredBBs, Suffix: ".thr_comm");
2653	}
2654	Updates.push_back(x: {DominatorTree::Delete, PredBB, BB});
2655
2656	// Okay, we decided to do this! Clone all the instructions in BB onto the end
2657	// of PredBB.
2658	LLVM_DEBUG(dbgs() << " Duplicating block '" << BB->getName()
2659	<< "' into end of '" << PredBB->getName()
2660	<< "' to eliminate branch on phi. Cost: "
2661	<< DuplicationCost << " block is:" << *BB << "\n");
2662
2663	// Unless PredBB ends with an unconditional branch, split the edge so that we
2664	// can just clone the bits from BB into the end of the new PredBB.
2665	BranchInst *OldPredBranch = dyn_cast<BranchInst>(Val: PredBB->getTerminator());
2666
2667	if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
2668	BasicBlock *OldPredBB = PredBB;
2669	PredBB = SplitEdge(From: OldPredBB, To: BB);
2670	Updates.push_back(x: {DominatorTree::Insert, OldPredBB, PredBB});
2671	Updates.push_back(x: {DominatorTree::Insert, PredBB, BB});
2672	Updates.push_back(x: {DominatorTree::Delete, OldPredBB, BB});
2673	OldPredBranch = cast<BranchInst>(Val: PredBB->getTerminator());
2674	}
2675
2676	// We are going to have to map operands from the original BB block into the
2677	// PredBB block. Evaluate PHI nodes in BB.
2678	DenseMap<Instruction*, Value*> ValueMapping;
2679
2680	BasicBlock::iterator BI = BB->begin();
2681	for (; PHINode *PN = dyn_cast<PHINode>(Val&: BI); ++BI)
2682	ValueMapping[PN] = PN->getIncomingValueForBlock(BB: PredBB);
2683	// Clone the non-phi instructions of BB into PredBB, keeping track of the
2684	// mapping and using it to remap operands in the cloned instructions.
2685	for (; BI != BB->end(); ++BI) {
2686	Instruction *New = BI->clone();
2687	New->insertInto(ParentBB: PredBB, It: OldPredBranch->getIterator());
2688
2689	// Remap operands to patch up intra-block references.
2690	for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2691	if (Instruction *Inst = dyn_cast<Instruction>(Val: New->getOperand(i))) {
2692	DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Val: Inst);
2693	if (I != ValueMapping.end())
2694	New->setOperand(i, Val: I->second);
2695	}
2696
2697	// If this instruction can be simplified after the operands are updated,
2698	// just use the simplified value instead. This frequently happens due to
2699	// phi translation.
2700	if (Value *IV = simplifyInstruction(
2701	I: New,
2702	Q: {BB->getModule()->getDataLayout(), TLI, nullptr, nullptr, New})) {
2703	ValueMapping[&*BI] = IV;
2704	if (!New->mayHaveSideEffects()) {
2705	New->eraseFromParent();
2706	New = nullptr;
2707	// Clone debug-info on the elided instruction to the destination
2708	// position.
2709	OldPredBranch->cloneDebugInfoFrom(From: &*BI, FromHere: std::nullopt, InsertAtHead: true);
2710	}
2711	} else {
2712	ValueMapping[&*BI] = New;
2713	}
2714	if (New) {
2715	// Otherwise, insert the new instruction into the block.
2716	New->setName(BI->getName());
2717	// Clone across any debug-info attached to the old instruction.
2718	New->cloneDebugInfoFrom(From: &*BI);
2719	// Update Dominance from simplified New instruction operands.
2720	for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2721	if (BasicBlock *SuccBB = dyn_cast<BasicBlock>(Val: New->getOperand(i)))
2722	Updates.push_back(x: {DominatorTree::Insert, PredBB, SuccBB});
2723	}
2724	}
2725
2726	// Check to see if the targets of the branch had PHI nodes. If so, we need to
2727	// add entries to the PHI nodes for branch from PredBB now.
2728	BranchInst *BBBranch = cast<BranchInst>(Val: BB->getTerminator());
2729	addPHINodeEntriesForMappedBlock(PHIBB: BBBranch->getSuccessor(i: 0), OldPred: BB, NewPred: PredBB,
2730	ValueMap&: ValueMapping);
2731	addPHINodeEntriesForMappedBlock(PHIBB: BBBranch->getSuccessor(i: 1), OldPred: BB, NewPred: PredBB,
2732	ValueMap&: ValueMapping);
2733
2734	updateSSA(BB, NewBB: PredBB, ValueMapping);
2735
2736	// PredBB no longer jumps to BB, remove entries in the PHI node for the edge
2737	// that we nuked.
2738	BB->removePredecessor(Pred: PredBB, KeepOneInputPHIs: true);
2739
2740	// Remove the unconditional branch at the end of the PredBB block.
2741	OldPredBranch->eraseFromParent();
2742	if (auto *BPI = getBPI())
2743	BPI->copyEdgeProbabilities(Src: BB, Dst: PredBB);
2744	DTU->applyUpdatesPermissive(Updates);
2745
2746	++NumDupes;
2747	return true;
2748	}
2749
2750	// Pred is a predecessor of BB with an unconditional branch to BB. SI is
2751	// a Select instruction in Pred. BB has other predecessors and SI is used in
2752	// a PHI node in BB. SI has no other use.
2753	// A new basic block, NewBB, is created and SI is converted to compare and
2754	// conditional branch. SI is erased from parent.
2755	void JumpThreadingPass::unfoldSelectInstr(BasicBlock *Pred, BasicBlock *BB,
2756	SelectInst *SI, PHINode *SIUse,
2757	unsigned Idx) {
2758	// Expand the select.
2759	//
2760	// Pred --
2761	// | v
2762	// | NewBB
2763	// | |
2764	// |-----
2765	// v
2766	// BB
2767	BranchInst *PredTerm = cast<BranchInst>(Val: Pred->getTerminator());
2768	BasicBlock *NewBB = BasicBlock::Create(Context&: BB->getContext(), Name: "select.unfold",
2769	Parent: BB->getParent(), InsertBefore: BB);
2770	// Move the unconditional branch to NewBB.
2771	PredTerm->removeFromParent();
2772	PredTerm->insertInto(ParentBB: NewBB, It: NewBB->end());
2773	// Create a conditional branch and update PHI nodes.
2774	auto *BI = BranchInst::Create(IfTrue: NewBB, IfFalse: BB, Cond: SI->getCondition(), InsertAtEnd: Pred);
2775	BI->applyMergedLocation(LocA: PredTerm->getDebugLoc(), LocB: SI->getDebugLoc());
2776	BI->copyMetadata(SrcInst: *SI, WL: {LLVMContext::MD_prof});
2777	SIUse->setIncomingValue(i: Idx, V: SI->getFalseValue());
2778	SIUse->addIncoming(V: SI->getTrueValue(), BB: NewBB);
2779
2780	uint64_t TrueWeight = 1;
2781	uint64_t FalseWeight = 1;
2782	// Copy probabilities from 'SI' to created conditional branch in 'Pred'.
2783	if (extractBranchWeights(I: *SI, TrueVal&: TrueWeight, FalseVal&: FalseWeight) &&
2784	(TrueWeight + FalseWeight) != 0) {
2785	SmallVector<BranchProbability, 2> BP;
2786	BP.emplace_back(Args: BranchProbability::getBranchProbability(
2787	Numerator: TrueWeight, Denominator: TrueWeight + FalseWeight));
2788	BP.emplace_back(Args: BranchProbability::getBranchProbability(
2789	Numerator: FalseWeight, Denominator: TrueWeight + FalseWeight));
2790	// Update BPI if exists.
2791	if (auto *BPI = getBPI())
2792	BPI->setEdgeProbability(Src: Pred, Probs: BP);
2793	}
2794	// Set the block frequency of NewBB.
2795	if (auto *BFI = getBFI()) {
2796	if ((TrueWeight + FalseWeight) == 0) {
2797	TrueWeight = 1;
2798	FalseWeight = 1;
2799	}
2800	BranchProbability PredToNewBBProb = BranchProbability::getBranchProbability(
2801	Numerator: TrueWeight, Denominator: TrueWeight + FalseWeight);
2802	auto NewBBFreq = BFI->getBlockFreq(BB: Pred) * PredToNewBBProb;
2803	BFI->setBlockFreq(BB: NewBB, Freq: NewBBFreq);
2804	}
2805
2806	// The select is now dead.
2807	SI->eraseFromParent();
2808	DTU->applyUpdatesPermissive(Updates: {{DominatorTree::Insert, NewBB, BB},
2809	{DominatorTree::Insert, Pred, NewBB}});
2810
2811	// Update any other PHI nodes in BB.
2812	for (BasicBlock::iterator BI = BB->begin();
2813	PHINode *Phi = dyn_cast<PHINode>(Val&: BI); ++BI)
2814	if (Phi != SIUse)
2815	Phi->addIncoming(V: Phi->getIncomingValueForBlock(BB: Pred), BB: NewBB);
2816	}
2817
2818	bool JumpThreadingPass::tryToUnfoldSelect(SwitchInst *SI, BasicBlock *BB) {
2819	PHINode *CondPHI = dyn_cast<PHINode>(Val: SI->getCondition());
2820
2821	if (!CondPHI || CondPHI->getParent() != BB)
2822	return false;
2823
2824	for (unsigned I = 0, E = CondPHI->getNumIncomingValues(); I != E; ++I) {
2825	BasicBlock *Pred = CondPHI->getIncomingBlock(i: I);
2826	SelectInst *PredSI = dyn_cast<SelectInst>(Val: CondPHI->getIncomingValue(i: I));
2827
2828	// The second and third condition can be potentially relaxed. Currently
2829	// the conditions help to simplify the code and allow us to reuse existing
2830	// code, developed for tryToUnfoldSelect(CmpInst *, BasicBlock *)
2831	if (!PredSI || PredSI->getParent() != Pred || !PredSI->hasOneUse())
2832	continue;
2833
2834	BranchInst *PredTerm = dyn_cast<BranchInst>(Val: Pred->getTerminator());
2835	if (!PredTerm || !PredTerm->isUnconditional())
2836	continue;
2837
2838	unfoldSelectInstr(Pred, BB, SI: PredSI, SIUse: CondPHI, Idx: I);
2839	return true;
2840	}
2841	return false;
2842	}
2843
2844	/// tryToUnfoldSelect - Look for blocks of the form
2845	/// bb1:
2846	/// %a = select
2847	/// br bb2
2848	///
2849	/// bb2:
2850	/// %p = phi [%a, %bb1] ...
2851	/// %c = icmp %p
2852	/// br i1 %c
2853	///
2854	/// And expand the select into a branch structure if one of its arms allows %c
2855	/// to be folded. This later enables threading from bb1 over bb2.
2856	bool JumpThreadingPass::tryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
2857	BranchInst *CondBr = dyn_cast<BranchInst>(Val: BB->getTerminator());
2858	PHINode *CondLHS = dyn_cast<PHINode>(Val: CondCmp->getOperand(i_nocapture: 0));
2859	Constant *CondRHS = cast<Constant>(Val: CondCmp->getOperand(i_nocapture: 1));
2860
2861	if (!CondBr || !CondBr->isConditional() || !CondLHS ||
2862	CondLHS->getParent() != BB)
2863	return false;
2864
2865	for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
2866	BasicBlock *Pred = CondLHS->getIncomingBlock(i: I);
2867	SelectInst *SI = dyn_cast<SelectInst>(Val: CondLHS->getIncomingValue(i: I));
2868
2869	// Look if one of the incoming values is a select in the corresponding
2870	// predecessor.
2871	if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
2872	continue;
2873
2874	BranchInst *PredTerm = dyn_cast<BranchInst>(Val: Pred->getTerminator());
2875	if (!PredTerm || !PredTerm->isUnconditional())
2876	continue;
2877
2878	// Now check if one of the select values would allow us to constant fold the
2879	// terminator in BB. We don't do the transform if both sides fold, those
2880	// cases will be threaded in any case.
2881	LazyValueInfo::Tristate LHSFolds =
2882	LVI->getPredicateOnEdge(Pred: CondCmp->getPredicate(), V: SI->getOperand(i_nocapture: 1),
2883	C: CondRHS, FromBB: Pred, ToBB: BB, CxtI: CondCmp);
2884	LazyValueInfo::Tristate RHSFolds =
2885	LVI->getPredicateOnEdge(Pred: CondCmp->getPredicate(), V: SI->getOperand(i_nocapture: 2),
2886	C: CondRHS, FromBB: Pred, ToBB: BB, CxtI: CondCmp);
2887	if ((LHSFolds != LazyValueInfo::Unknown ||
2888	RHSFolds != LazyValueInfo::Unknown) &&
2889	LHSFolds != RHSFolds) {
2890	unfoldSelectInstr(Pred, BB, SI, SIUse: CondLHS, Idx: I);
2891	return true;
2892	}
2893	}
2894	return false;
2895	}
2896
2897	/// tryToUnfoldSelectInCurrBB - Look for PHI/Select or PHI/CMP/Select in the
2898	/// same BB in the form
2899	/// bb:
2900	/// %p = phi [false, %bb1], [true, %bb2], [false, %bb3], [true, %bb4], ...
2901	/// %s = select %p, trueval, falseval
2902	///
2903	/// or
2904	///
2905	/// bb:
2906	/// %p = phi [0, %bb1], [1, %bb2], [0, %bb3], [1, %bb4], ...
2907	/// %c = cmp %p, 0
2908	/// %s = select %c, trueval, falseval
2909	///
2910	/// And expand the select into a branch structure. This later enables
2911	/// jump-threading over bb in this pass.
2912	///
2913	/// Using the similar approach of SimplifyCFG::FoldCondBranchOnPHI(), unfold
2914	/// select if the associated PHI has at least one constant. If the unfolded
2915	/// select is not jump-threaded, it will be folded again in the later
2916	/// optimizations.
2917	bool JumpThreadingPass::tryToUnfoldSelectInCurrBB(BasicBlock *BB) {
2918	// This transform would reduce the quality of msan diagnostics.
2919	// Disable this transform under MemorySanitizer.
2920	if (BB->getParent()->hasFnAttribute(Attribute::SanitizeMemory))
2921	return false;
2922
2923	// If threading this would thread across a loop header, don't thread the edge.
2924	// See the comments above findLoopHeaders for justifications and caveats.
2925	if (LoopHeaders.count(V: BB))
2926	return false;
2927
2928	for (BasicBlock::iterator BI = BB->begin();
2929	PHINode *PN = dyn_cast<PHINode>(Val&: BI); ++BI) {
2930	// Look for a Phi having at least one constant incoming value.
2931	if (llvm::all_of(Range: PN->incoming_values(),
2932	P: [](Value *V) { return !isa<ConstantInt>(Val: V); }))
2933	continue;
2934
2935	auto isUnfoldCandidate = [BB](SelectInst *SI, Value *V) {
2936	using namespace PatternMatch;
2937
2938	// Check if SI is in BB and use V as condition.
2939	if (SI->getParent() != BB)
2940	return false;
2941	Value *Cond = SI->getCondition();
2942	bool IsAndOr = match(V: SI, P: m_CombineOr(L: m_LogicalAnd(), R: m_LogicalOr()));
2943	return Cond && Cond == V && Cond->getType()->isIntegerTy(Bitwidth: 1) && !IsAndOr;
2944	};
2945
2946	SelectInst *SI = nullptr;
2947	for (Use &U : PN->uses()) {
2948	if (ICmpInst *Cmp = dyn_cast<ICmpInst>(Val: U.getUser())) {
2949	// Look for a ICmp in BB that compares PN with a constant and is the
2950	// condition of a Select.
2951	if (Cmp->getParent() == BB && Cmp->hasOneUse() &&
2952	isa<ConstantInt>(Val: Cmp->getOperand(i_nocapture: 1 - U.getOperandNo())))
2953	if (SelectInst *SelectI = dyn_cast<SelectInst>(Val: Cmp->user_back()))
2954	if (isUnfoldCandidate(SelectI, Cmp->use_begin()->get())) {
2955	SI = SelectI;
2956	break;
2957	}
2958	} else if (SelectInst *SelectI = dyn_cast<SelectInst>(Val: U.getUser())) {
2959	// Look for a Select in BB that uses PN as condition.
2960	if (isUnfoldCandidate(SelectI, U.get())) {
2961	SI = SelectI;
2962	break;
2963	}
2964	}
2965	}
2966
2967	if (!SI)
2968	continue;
2969	// Expand the select.
2970	Value *Cond = SI->getCondition();
2971	if (!isGuaranteedNotToBeUndefOrPoison(V: Cond, AC: nullptr, CtxI: SI))
2972	Cond = new FreezeInst(Cond, "cond.fr", SI->getIterator());
2973	MDNode *BranchWeights = getBranchWeightMDNode(I: *SI);
2974	Instruction *Term =
2975	SplitBlockAndInsertIfThen(Cond, SplitBefore: SI, Unreachable: false, BranchWeights);
2976	BasicBlock *SplitBB = SI->getParent();
2977	BasicBlock *NewBB = Term->getParent();
2978	PHINode *NewPN = PHINode::Create(Ty: SI->getType(), NumReservedValues: 2, NameStr: "", InsertBefore: SI->getIterator());
2979	NewPN->addIncoming(V: SI->getTrueValue(), BB: Term->getParent());
2980	NewPN->addIncoming(V: SI->getFalseValue(), BB);
2981	SI->replaceAllUsesWith(V: NewPN);
2982	SI->eraseFromParent();
2983	// NewBB and SplitBB are newly created blocks which require insertion.
2984	std::vector<DominatorTree::UpdateType> Updates;
2985	Updates.reserve(n: (2 * SplitBB->getTerminator()->getNumSuccessors()) + 3);
2986	Updates.push_back(x: {DominatorTree::Insert, BB, SplitBB});
2987	Updates.push_back(x: {DominatorTree::Insert, BB, NewBB});
2988	Updates.push_back(x: {DominatorTree::Insert, NewBB, SplitBB});
2989	// BB's successors were moved to SplitBB, update DTU accordingly.
2990	for (auto *Succ : successors(BB: SplitBB)) {
2991	Updates.push_back(x: {DominatorTree::Delete, BB, Succ});
2992	Updates.push_back(x: {DominatorTree::Insert, SplitBB, Succ});
2993	}
2994	DTU->applyUpdatesPermissive(Updates);
2995	return true;
2996	}
2997	return false;
2998	}
2999
3000	/// Try to propagate a guard from the current BB into one of its predecessors
3001	/// in case if another branch of execution implies that the condition of this
3002	/// guard is always true. Currently we only process the simplest case that
3003	/// looks like:
3004	///
3005	/// Start:
3006	/// %cond = ...
3007	/// br i1 %cond, label %T1, label %F1
3008	/// T1:
3009	/// br label %Merge
3010	/// F1:
3011	/// br label %Merge
3012	/// Merge:
3013	/// %condGuard = ...
3014	/// call void(i1, ...) @llvm.experimental.guard( i1 %condGuard )[ "deopt"() ]
3015	///
3016	/// And cond either implies condGuard or !condGuard. In this case all the
3017	/// instructions before the guard can be duplicated in both branches, and the
3018	/// guard is then threaded to one of them.
3019	bool JumpThreadingPass::processGuards(BasicBlock *BB) {
3020	using namespace PatternMatch;
3021
3022	// We only want to deal with two predecessors.
3023	BasicBlock *Pred1, *Pred2;
3024	auto PI = pred_begin(BB), PE = pred_end(BB);
3025	if (PI == PE)
3026	return false;
3027	Pred1 = *PI++;
3028	if (PI == PE)
3029	return false;
3030	Pred2 = *PI++;
3031	if (PI != PE)
3032	return false;
3033	if (Pred1 == Pred2)
3034	return false;
3035
3036	// Try to thread one of the guards of the block.
3037	// TODO: Look up deeper than to immediate predecessor?
3038	auto *Parent = Pred1->getSinglePredecessor();
3039	if (!Parent || Parent != Pred2->getSinglePredecessor())
3040	return false;
3041
3042	if (auto *BI = dyn_cast<BranchInst>(Val: Parent->getTerminator()))
3043	for (auto &I : *BB)
3044	if (isGuard(U: &I) && threadGuard(BB, Guard: cast<IntrinsicInst>(Val: &I), BI))
3045	return true;
3046
3047	return false;
3048	}
3049
3050	/// Try to propagate the guard from BB which is the lower block of a diamond
3051	/// to one of its branches, in case if diamond's condition implies guard's
3052	/// condition.
3053	bool JumpThreadingPass::threadGuard(BasicBlock *BB, IntrinsicInst *Guard,
3054	BranchInst *BI) {
3055	assert(BI->getNumSuccessors() == 2 && "Wrong number of successors?");
3056	assert(BI->isConditional() && "Unconditional branch has 2 successors?");
3057	Value *GuardCond = Guard->getArgOperand(i: 0);
3058	Value *BranchCond = BI->getCondition();
3059	BasicBlock *TrueDest = BI->getSuccessor(i: 0);
3060	BasicBlock *FalseDest = BI->getSuccessor(i: 1);
3061
3062	auto &DL = BB->getModule()->getDataLayout();
3063	bool TrueDestIsSafe = false;
3064	bool FalseDestIsSafe = false;
3065
3066	// True dest is safe if BranchCond => GuardCond.
3067	auto Impl = isImpliedCondition(LHS: BranchCond, RHS: GuardCond, DL);
3068	if (Impl && *Impl)
3069	TrueDestIsSafe = true;
3070	else {
3071	// False dest is safe if !BranchCond => GuardCond.
3072	Impl = isImpliedCondition(LHS: BranchCond, RHS: GuardCond, DL, /* LHSIsTrue */ false);
3073	if (Impl && *Impl)
3074	FalseDestIsSafe = true;
3075	}
3076
3077	if (!TrueDestIsSafe && !FalseDestIsSafe)
3078	return false;
3079
3080	BasicBlock *PredUnguardedBlock = TrueDestIsSafe ? TrueDest : FalseDest;
3081	BasicBlock *PredGuardedBlock = FalseDestIsSafe ? TrueDest : FalseDest;
3082
3083	ValueToValueMapTy UnguardedMapping, GuardedMapping;
3084	Instruction *AfterGuard = Guard->getNextNode();
3085	unsigned Cost =
3086	getJumpThreadDuplicationCost(TTI, BB, StopAt: AfterGuard, Threshold: BBDupThreshold);
3087	if (Cost > BBDupThreshold)
3088	return false;
3089	// Duplicate all instructions before the guard and the guard itself to the
3090	// branch where implication is not proved.
3091	BasicBlock *GuardedBlock = DuplicateInstructionsInSplitBetween(
3092	BB, PredBB: PredGuardedBlock, StopAt: AfterGuard, ValueMapping&: GuardedMapping, DTU&: *DTU);
3093	assert(GuardedBlock && "Could not create the guarded block?");
3094	// Duplicate all instructions before the guard in the unguarded branch.
3095	// Since we have successfully duplicated the guarded block and this block
3096	// has fewer instructions, we expect it to succeed.
3097	BasicBlock *UnguardedBlock = DuplicateInstructionsInSplitBetween(
3098	BB, PredBB: PredUnguardedBlock, StopAt: Guard, ValueMapping&: UnguardedMapping, DTU&: *DTU);
3099	assert(UnguardedBlock && "Could not create the unguarded block?");
3100	LLVM_DEBUG(dbgs() << "Moved guard " << *Guard << " to block "
3101	<< GuardedBlock->getName() << "\n");
3102	// Some instructions before the guard may still have uses. For them, we need
3103	// to create Phi nodes merging their copies in both guarded and unguarded
3104	// branches. Those instructions that have no uses can be just removed.
3105	SmallVector<Instruction *, 4> ToRemove;
3106	for (auto BI = BB->begin(); &*BI != AfterGuard; ++BI)
3107	if (!isa<PHINode>(Val: &*BI))
3108	ToRemove.push_back(Elt: &*BI);
3109
3110	BasicBlock::iterator InsertionPoint = BB->getFirstInsertionPt();
3111	assert(InsertionPoint != BB->end() && "Empty block?");
3112	// Substitute with Phis & remove.
3113	for (auto *Inst : reverse(C&: ToRemove)) {
3114	if (!Inst->use_empty()) {
3115	PHINode *NewPN = PHINode::Create(Ty: Inst->getType(), NumReservedValues: 2);
3116	NewPN->addIncoming(V: UnguardedMapping[Inst], BB: UnguardedBlock);
3117	NewPN->addIncoming(V: GuardedMapping[Inst], BB: GuardedBlock);
3118	NewPN->insertBefore(InsertPos: InsertionPoint);
3119	Inst->replaceAllUsesWith(V: NewPN);
3120	}
3121	Inst->dropDbgRecords();
3122	Inst->eraseFromParent();
3123	}
3124	return true;
3125	}
3126
3127	PreservedAnalyses JumpThreadingPass::getPreservedAnalysis() const {
3128	PreservedAnalyses PA;
3129	PA.preserve<LazyValueAnalysis>();
3130	PA.preserve<DominatorTreeAnalysis>();
3131
3132	// TODO: We would like to preserve BPI/BFI. Enable once all paths update them.
3133	// TODO: Would be nice to verify BPI/BFI consistency as well.
3134	return PA;
3135	}
3136
3137	template <typename AnalysisT>
3138	typename AnalysisT::Result *JumpThreadingPass::runExternalAnalysis() {
3139	assert(FAM && "Can't run external analysis without FunctionAnalysisManager");
3140
3141	// If there were no changes since last call to 'runExternalAnalysis' then all
3142	// analysis is either up to date or explicitly invalidated. Just go ahead and
3143	// run the "external" analysis.
3144	if (!ChangedSinceLastAnalysisUpdate) {
3145	assert(!DTU->hasPendingUpdates() &&
3146	"Lost update of 'ChangedSinceLastAnalysisUpdate'?");
3147	// Run the "external" analysis.
3148	return &FAM->getResult<AnalysisT>(*F);
3149	}
3150	ChangedSinceLastAnalysisUpdate = false;
3151
3152	auto PA = getPreservedAnalysis();
3153	// TODO: This shouldn't be needed once 'getPreservedAnalysis' reports BPI/BFI
3154	// as preserved.
3155	PA.preserve<BranchProbabilityAnalysis>();
3156	PA.preserve<BlockFrequencyAnalysis>();
3157	// Report everything except explicitly preserved as invalid.
3158	FAM->invalidate(IR&: *F, PA);
3159	// Update DT/PDT.
3160	DTU->flush();
3161	// Make sure DT/PDT are valid before running "external" analysis.
3162	assert(DTU->getDomTree().verify(DominatorTree::VerificationLevel::Fast));
3163	assert((!DTU->hasPostDomTree() ||
3164	DTU->getPostDomTree().verify(
3165	PostDominatorTree::VerificationLevel::Fast)));
3166	// Run the "external" analysis.
3167	auto *Result = &FAM->getResult<AnalysisT>(*F);
3168	// Update analysis JumpThreading depends on and not explicitly preserved.
3169	TTI = &FAM->getResult<TargetIRAnalysis>(IR&: *F);
3170	TLI = &FAM->getResult<TargetLibraryAnalysis>(IR&: *F);
3171	AA = &FAM->getResult<AAManager>(IR&: *F);
3172
3173	return Result;
3174	}
3175
3176	BranchProbabilityInfo *JumpThreadingPass::getBPI() {
3177	if (!BPI) {
3178	assert(FAM && "Can't create BPI without FunctionAnalysisManager");
3179	BPI = FAM->getCachedResult<BranchProbabilityAnalysis>(IR&: *F);
3180	}
3181	return *BPI;
3182	}
3183
3184	BlockFrequencyInfo *JumpThreadingPass::getBFI() {
3185	if (!BFI) {
3186	assert(FAM && "Can't create BFI without FunctionAnalysisManager");
3187	BFI = FAM->getCachedResult<BlockFrequencyAnalysis>(IR&: *F);
3188	}
3189	return *BFI;
3190	}
3191
3192	// Important note on validity of BPI/BFI. JumpThreading tries to preserve
3193	// BPI/BFI as it goes. Thus if cached instance exists it will be updated.
3194	// Otherwise, new instance of BPI/BFI is created (up to date by definition).
3195	BranchProbabilityInfo *JumpThreadingPass::getOrCreateBPI(bool Force) {
3196	auto *Res = getBPI();
3197	if (Res)
3198	return Res;
3199
3200	if (Force)
3201	BPI = runExternalAnalysis<BranchProbabilityAnalysis>();
3202
3203	return *BPI;
3204	}
3205
3206	BlockFrequencyInfo *JumpThreadingPass::getOrCreateBFI(bool Force) {
3207	auto *Res = getBFI();
3208	if (Res)
3209	return Res;
3210
3211	if (Force)
3212	BFI = runExternalAnalysis<BlockFrequencyAnalysis>();
3213
3214	return *BFI;
3215	}
3216