Bug Summary

File:lib/Transforms/Scalar/JumpThreading.cpp
Warning:line 1421, column 7
Called C++ object pointer is null

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name JumpThreading.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mthread-model posix -fmath-errno -masm-verbose -mconstructor-aliases -munwind-tables -fuse-init-array -target-cpu x86-64 -dwarf-column-info -debugger-tuning=gdb -momit-leaf-frame-pointer -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-9/lib/clang/9.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-9~svn358860/build-llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-9~svn358860/build-llvm/include -I /build/llvm-toolchain-snapshot-9~svn358860/include -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0/backward -internal-isystem /usr/include/clang/9.0.0/include/ -internal-isystem /usr/local/include -internal-isystem /usr/lib/llvm-9/lib/clang/9.0.0/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-comment -std=c++11 -fdeprecated-macro -fdebug-compilation-dir /build/llvm-toolchain-snapshot-9~svn358860/build-llvm/lib/Transforms/Scalar -fdebug-prefix-map=/build/llvm-toolchain-snapshot-9~svn358860=. -ferror-limit 19 -fmessage-length 0 -fvisibility-inlines-hidden -stack-protector 2 -fobjc-runtime=gcc -fdiagnostics-show-option -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -o /tmp/scan-build-2019-04-22-050718-5320-1 -x c++ /build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp -faddrsig
1//===- JumpThreading.cpp - Thread control through conditional blocks ------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file implements the Jump Threading pass.
10//
11//===----------------------------------------------------------------------===//
12
13#include "llvm/Transforms/Scalar/JumpThreading.h"
14#include "llvm/ADT/DenseMap.h"
15#include "llvm/ADT/DenseSet.h"
16#include "llvm/ADT/Optional.h"
17#include "llvm/ADT/STLExtras.h"
18#include "llvm/ADT/SmallPtrSet.h"
19#include "llvm/ADT/SmallVector.h"
20#include "llvm/ADT/Statistic.h"
21#include "llvm/Analysis/AliasAnalysis.h"
22#include "llvm/Analysis/BlockFrequencyInfo.h"
23#include "llvm/Analysis/BranchProbabilityInfo.h"
24#include "llvm/Analysis/CFG.h"
25#include "llvm/Analysis/ConstantFolding.h"
26#include "llvm/Analysis/DomTreeUpdater.h"
27#include "llvm/Analysis/GlobalsModRef.h"
28#include "llvm/Analysis/GuardUtils.h"
29#include "llvm/Analysis/InstructionSimplify.h"
30#include "llvm/Analysis/LazyValueInfo.h"
31#include "llvm/Analysis/Loads.h"
32#include "llvm/Analysis/LoopInfo.h"
33#include "llvm/Analysis/TargetLibraryInfo.h"
34#include "llvm/Analysis/ValueTracking.h"
35#include "llvm/IR/BasicBlock.h"
36#include "llvm/IR/CFG.h"
37#include "llvm/IR/Constant.h"
38#include "llvm/IR/ConstantRange.h"
39#include "llvm/IR/Constants.h"
40#include "llvm/IR/DataLayout.h"
41#include "llvm/IR/Dominators.h"
42#include "llvm/IR/Function.h"
43#include "llvm/IR/InstrTypes.h"
44#include "llvm/IR/Instruction.h"
45#include "llvm/IR/Instructions.h"
46#include "llvm/IR/IntrinsicInst.h"
47#include "llvm/IR/Intrinsics.h"
48#include "llvm/IR/LLVMContext.h"
49#include "llvm/IR/MDBuilder.h"
50#include "llvm/IR/Metadata.h"
51#include "llvm/IR/Module.h"
52#include "llvm/IR/PassManager.h"
53#include "llvm/IR/PatternMatch.h"
54#include "llvm/IR/Type.h"
55#include "llvm/IR/Use.h"
56#include "llvm/IR/User.h"
57#include "llvm/IR/Value.h"
58#include "llvm/Pass.h"
59#include "llvm/Support/BlockFrequency.h"
60#include "llvm/Support/BranchProbability.h"
61#include "llvm/Support/Casting.h"
62#include "llvm/Support/CommandLine.h"
63#include "llvm/Support/Debug.h"
64#include "llvm/Support/raw_ostream.h"
65#include "llvm/Transforms/Scalar.h"
66#include "llvm/Transforms/Utils/BasicBlockUtils.h"
67#include "llvm/Transforms/Utils/Cloning.h"
68#include "llvm/Transforms/Utils/Local.h"
69#include "llvm/Transforms/Utils/SSAUpdater.h"
70#include "llvm/Transforms/Utils/ValueMapper.h"
71#include <algorithm>
72#include <cassert>
73#include <cstddef>
74#include <cstdint>
75#include <iterator>
76#include <memory>
77#include <utility>
78
79using namespace llvm;
80using namespace jumpthreading;
81
82#define DEBUG_TYPE"jump-threading" "jump-threading"
83
84STATISTIC(NumThreads, "Number of jumps threaded")static llvm::Statistic NumThreads = {"jump-threading", "NumThreads"
, "Number of jumps threaded", {0}, {false}}
;
85STATISTIC(NumFolds, "Number of terminators folded")static llvm::Statistic NumFolds = {"jump-threading", "NumFolds"
, "Number of terminators folded", {0}, {false}}
;
86STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi")static llvm::Statistic NumDupes = {"jump-threading", "NumDupes"
, "Number of branch blocks duplicated to eliminate phi", {0},
{false}}
;
87
88static cl::opt<unsigned>
89BBDuplicateThreshold("jump-threading-threshold",
90 cl::desc("Max block size to duplicate for jump threading"),
91 cl::init(6), cl::Hidden);
92
93static cl::opt<unsigned>
94ImplicationSearchThreshold(
95 "jump-threading-implication-search-threshold",
96 cl::desc("The number of predecessors to search for a stronger "
97 "condition to use to thread over a weaker condition"),
98 cl::init(3), cl::Hidden);
99
100static cl::opt<bool> PrintLVIAfterJumpThreading(
101 "print-lvi-after-jump-threading",
102 cl::desc("Print the LazyValueInfo cache after JumpThreading"), cl::init(false),
103 cl::Hidden);
104
105namespace {
106
107 /// This pass performs 'jump threading', which looks at blocks that have
108 /// multiple predecessors and multiple successors. If one or more of the
109 /// predecessors of the block can be proven to always jump to one of the
110 /// successors, we forward the edge from the predecessor to the successor by
111 /// duplicating the contents of this block.
112 ///
113 /// An example of when this can occur is code like this:
114 ///
115 /// if () { ...
116 /// X = 4;
117 /// }
118 /// if (X < 3) {
119 ///
120 /// In this case, the unconditional branch at the end of the first if can be
121 /// revectored to the false side of the second if.
122 class JumpThreading : public FunctionPass {
123 JumpThreadingPass Impl;
124
125 public:
126 static char ID; // Pass identification
127
128 JumpThreading(int T = -1) : FunctionPass(ID), Impl(T) {
129 initializeJumpThreadingPass(*PassRegistry::getPassRegistry());
130 }
131
132 bool runOnFunction(Function &F) override;
133
134 void getAnalysisUsage(AnalysisUsage &AU) const override {
135 AU.addRequired<DominatorTreeWrapperPass>();
136 AU.addPreserved<DominatorTreeWrapperPass>();
137 AU.addRequired<AAResultsWrapperPass>();
138 AU.addRequired<LazyValueInfoWrapperPass>();
139 AU.addPreserved<LazyValueInfoWrapperPass>();
140 AU.addPreserved<GlobalsAAWrapperPass>();
141 AU.addRequired<TargetLibraryInfoWrapperPass>();
142 }
143
144 void releaseMemory() override { Impl.releaseMemory(); }
145 };
146
147} // end anonymous namespace
148
149char JumpThreading::ID = 0;
150
151INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",static void *initializeJumpThreadingPassOnce(PassRegistry &
Registry) {
152 "Jump Threading", false, false)static void *initializeJumpThreadingPassOnce(PassRegistry &
Registry) {
153INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry);
154INITIALIZE_PASS_DEPENDENCY(LazyValueInfoWrapperPass)initializeLazyValueInfoWrapperPassPass(Registry);
155INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry);
156INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry);
157INITIALIZE_PASS_END(JumpThreading, "jump-threading",PassInfo *PI = new PassInfo( "Jump Threading", "jump-threading"
, &JumpThreading::ID, PassInfo::NormalCtor_t(callDefaultCtor
<JumpThreading>), false, false); Registry.registerPass(
*PI, true); return PI; } static llvm::once_flag InitializeJumpThreadingPassFlag
; void llvm::initializeJumpThreadingPass(PassRegistry &Registry
) { llvm::call_once(InitializeJumpThreadingPassFlag, initializeJumpThreadingPassOnce
, std::ref(Registry)); }
158 "Jump Threading", false, false)PassInfo *PI = new PassInfo( "Jump Threading", "jump-threading"
, &JumpThreading::ID, PassInfo::NormalCtor_t(callDefaultCtor
<JumpThreading>), false, false); Registry.registerPass(
*PI, true); return PI; } static llvm::once_flag InitializeJumpThreadingPassFlag
; void llvm::initializeJumpThreadingPass(PassRegistry &Registry
) { llvm::call_once(InitializeJumpThreadingPassFlag, initializeJumpThreadingPassOnce
, std::ref(Registry)); }
159
160// Public interface to the Jump Threading pass
161FunctionPass *llvm::createJumpThreadingPass(int Threshold) {
162 return new JumpThreading(Threshold);
163}
164
165JumpThreadingPass::JumpThreadingPass(int T) {
166 BBDupThreshold = (T == -1) ? BBDuplicateThreshold : unsigned(T);
167}
168
169// Update branch probability information according to conditional
170// branch probability. This is usually made possible for cloned branches
171// in inline instances by the context specific profile in the caller.
172// For instance,
173//
174// [Block PredBB]
175// [Branch PredBr]
176// if (t) {
177// Block A;
178// } else {
179// Block B;
180// }
181//
182// [Block BB]
183// cond = PN([true, %A], [..., %B]); // PHI node
184// [Branch CondBr]
185// if (cond) {
186// ... // P(cond == true) = 1%
187// }
188//
189// Here we know that when block A is taken, cond must be true, which means
190// P(cond == true | A) = 1
191//
192// Given that P(cond == true) = P(cond == true | A) * P(A) +
193// P(cond == true | B) * P(B)
194// we get:
195// P(cond == true ) = P(A) + P(cond == true | B) * P(B)
196//
197// which gives us:
198// P(A) is less than P(cond == true), i.e.
199// P(t == true) <= P(cond == true)
200//
201// In other words, if we know P(cond == true) is unlikely, we know
202// that P(t == true) is also unlikely.
203//
204static void updatePredecessorProfileMetadata(PHINode *PN, BasicBlock *BB) {
205 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
206 if (!CondBr)
207 return;
208
209 BranchProbability BP;
210 uint64_t TrueWeight, FalseWeight;
211 if (!CondBr->extractProfMetadata(TrueWeight, FalseWeight))
212 return;
213
214 // Returns the outgoing edge of the dominating predecessor block
215 // that leads to the PhiNode's incoming block:
216 auto GetPredOutEdge =
217 [](BasicBlock *IncomingBB,
218 BasicBlock *PhiBB) -> std::pair<BasicBlock *, BasicBlock *> {
219 auto *PredBB = IncomingBB;
220 auto *SuccBB = PhiBB;
221 while (true) {
222 BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator());
223 if (PredBr && PredBr->isConditional())
224 return {PredBB, SuccBB};
225 auto *SinglePredBB = PredBB->getSinglePredecessor();
226 if (!SinglePredBB)
227 return {nullptr, nullptr};
228 SuccBB = PredBB;
229 PredBB = SinglePredBB;
230 }
231 };
232
233 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
234 Value *PhiOpnd = PN->getIncomingValue(i);
235 ConstantInt *CI = dyn_cast<ConstantInt>(PhiOpnd);
236
237 if (!CI || !CI->getType()->isIntegerTy(1))
238 continue;
239
240 BP = (CI->isOne() ? BranchProbability::getBranchProbability(
241 TrueWeight, TrueWeight + FalseWeight)
242 : BranchProbability::getBranchProbability(
243 FalseWeight, TrueWeight + FalseWeight));
244
245 auto PredOutEdge = GetPredOutEdge(PN->getIncomingBlock(i), BB);
246 if (!PredOutEdge.first)
247 return;
248
249 BasicBlock *PredBB = PredOutEdge.first;
250 BranchInst *PredBr = cast<BranchInst>(PredBB->getTerminator());
251
252 uint64_t PredTrueWeight, PredFalseWeight;
253 // FIXME: We currently only set the profile data when it is missing.
254 // With PGO, this can be used to refine even existing profile data with
255 // context information. This needs to be done after more performance
256 // testing.
257 if (PredBr->extractProfMetadata(PredTrueWeight, PredFalseWeight))
258 continue;
259
260 // We can not infer anything useful when BP >= 50%, because BP is the
261 // upper bound probability value.
262 if (BP >= BranchProbability(50, 100))
263 continue;
264
265 SmallVector<uint32_t, 2> Weights;
266 if (PredBr->getSuccessor(0) == PredOutEdge.second) {
267 Weights.push_back(BP.getNumerator());
268 Weights.push_back(BP.getCompl().getNumerator());
269 } else {
270 Weights.push_back(BP.getCompl().getNumerator());
271 Weights.push_back(BP.getNumerator());
272 }
273 PredBr->setMetadata(LLVMContext::MD_prof,
274 MDBuilder(PredBr->getParent()->getContext())
275 .createBranchWeights(Weights));
276 }
277}
278
279/// runOnFunction - Toplevel algorithm.
280bool JumpThreading::runOnFunction(Function &F) {
281 if (skipFunction(F))
282 return false;
283 auto TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
284 // Get DT analysis before LVI. When LVI is initialized it conditionally adds
285 // DT if it's available.
286 auto DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
287 auto LVI = &getAnalysis<LazyValueInfoWrapperPass>().getLVI();
288 auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
289 DomTreeUpdater DTU(*DT, DomTreeUpdater::UpdateStrategy::Lazy);
290 std::unique_ptr<BlockFrequencyInfo> BFI;
291 std::unique_ptr<BranchProbabilityInfo> BPI;
292 bool HasProfileData = F.hasProfileData();
293 if (HasProfileData) {
294 LoopInfo LI{DominatorTree(F)};
295 BPI.reset(new BranchProbabilityInfo(F, LI, TLI));
296 BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
297 }
298
299 bool Changed = Impl.runImpl(F, TLI, LVI, AA, &DTU, HasProfileData,
300 std::move(BFI), std::move(BPI));
301 if (PrintLVIAfterJumpThreading) {
302 dbgs() << "LVI for function '" << F.getName() << "':\n";
303 LVI->printLVI(F, *DT, dbgs());
304 }
305 return Changed;
306}
307
308PreservedAnalyses JumpThreadingPass::run(Function &F,
309 FunctionAnalysisManager &AM) {
310 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
311 // Get DT analysis before LVI. When LVI is initialized it conditionally adds
312 // DT if it's available.
313 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
314 auto &LVI = AM.getResult<LazyValueAnalysis>(F);
315 auto &AA = AM.getResult<AAManager>(F);
316 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Lazy);
317
318 std::unique_ptr<BlockFrequencyInfo> BFI;
319 std::unique_ptr<BranchProbabilityInfo> BPI;
320 if (F.hasProfileData()) {
321 LoopInfo LI{DominatorTree(F)};
322 BPI.reset(new BranchProbabilityInfo(F, LI, &TLI));
323 BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
324 }
325
326 bool Changed = runImpl(F, &TLI, &LVI, &AA, &DTU, HasProfileData,
327 std::move(BFI), std::move(BPI));
328
329 if (!Changed)
330 return PreservedAnalyses::all();
331 PreservedAnalyses PA;
332 PA.preserve<GlobalsAA>();
333 PA.preserve<DominatorTreeAnalysis>();
334 PA.preserve<LazyValueAnalysis>();
335 return PA;
336}
337
338bool JumpThreadingPass::runImpl(Function &F, TargetLibraryInfo *TLI_,
339 LazyValueInfo *LVI_, AliasAnalysis *AA_,
340 DomTreeUpdater *DTU_, bool HasProfileData_,
341 std::unique_ptr<BlockFrequencyInfo> BFI_,
342 std::unique_ptr<BranchProbabilityInfo> BPI_) {
343 LLVM_DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "Jump threading on function '"
<< F.getName() << "'\n"; } } while (false)
;
344 TLI = TLI_;
345 LVI = LVI_;
346 AA = AA_;
347 DTU = DTU_;
348 BFI.reset();
349 BPI.reset();
350 // When profile data is available, we need to update edge weights after
351 // successful jump threading, which requires both BPI and BFI being available.
352 HasProfileData = HasProfileData_;
353 auto *GuardDecl = F.getParent()->getFunction(
354 Intrinsic::getName(Intrinsic::experimental_guard));
355 HasGuards = GuardDecl && !GuardDecl->use_empty();
356 if (HasProfileData) {
357 BPI = std::move(BPI_);
358 BFI = std::move(BFI_);
359 }
360
361 // JumpThreading must not processes blocks unreachable from entry. It's a
362 // waste of compute time and can potentially lead to hangs.
363 SmallPtrSet<BasicBlock *, 16> Unreachable;
364 assert(DTU && "DTU isn't passed into JumpThreading before using it.")((DTU && "DTU isn't passed into JumpThreading before using it."
) ? static_cast<void> (0) : __assert_fail ("DTU && \"DTU isn't passed into JumpThreading before using it.\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 364, __PRETTY_FUNCTION__))
;
365 assert(DTU->hasDomTree() && "JumpThreading relies on DomTree to proceed.")((DTU->hasDomTree() && "JumpThreading relies on DomTree to proceed."
) ? static_cast<void> (0) : __assert_fail ("DTU->hasDomTree() && \"JumpThreading relies on DomTree to proceed.\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 365, __PRETTY_FUNCTION__))
;
366 DominatorTree &DT = DTU->getDomTree();
367 for (auto &BB : F)
368 if (!DT.isReachableFromEntry(&BB))
369 Unreachable.insert(&BB);
370
371 FindLoopHeaders(F);
372
373 bool EverChanged = false;
374 bool Changed;
375 do {
376 Changed = false;
377 for (auto &BB : F) {
378 if (Unreachable.count(&BB))
379 continue;
380 while (ProcessBlock(&BB)) // Thread all of the branches we can over BB.
381 Changed = true;
382 // Stop processing BB if it's the entry or is now deleted. The following
383 // routines attempt to eliminate BB and locating a suitable replacement
384 // for the entry is non-trivial.
385 if (&BB == &F.getEntryBlock() || DTU->isBBPendingDeletion(&BB))
386 continue;
387
388 if (pred_empty(&BB)) {
389 // When ProcessBlock makes BB unreachable it doesn't bother to fix up
390 // the instructions in it. We must remove BB to prevent invalid IR.
391 LLVM_DEBUG(dbgs() << " JT: Deleting dead block '" << BB.getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " JT: Deleting dead block '"
<< BB.getName() << "' with terminator: " <<
*BB.getTerminator() << '\n'; } } while (false)
392 << "' with terminator: " << *BB.getTerminator()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " JT: Deleting dead block '"
<< BB.getName() << "' with terminator: " <<
*BB.getTerminator() << '\n'; } } while (false)
393 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " JT: Deleting dead block '"
<< BB.getName() << "' with terminator: " <<
*BB.getTerminator() << '\n'; } } while (false)
;
394 LoopHeaders.erase(&BB);
395 LVI->eraseBlock(&BB);
396 DeleteDeadBlock(&BB, DTU);
397 Changed = true;
398 continue;
399 }
400
401 // ProcessBlock doesn't thread BBs with unconditional TIs. However, if BB
402 // is "almost empty", we attempt to merge BB with its sole successor.
403 auto *BI = dyn_cast<BranchInst>(BB.getTerminator());
404 if (BI && BI->isUnconditional() &&
405 // The terminator must be the only non-phi instruction in BB.
406 BB.getFirstNonPHIOrDbg()->isTerminator() &&
407 // Don't alter Loop headers and latches to ensure another pass can
408 // detect and transform nested loops later.
409 !LoopHeaders.count(&BB) && !LoopHeaders.count(BI->getSuccessor(0)) &&
410 TryToSimplifyUncondBranchFromEmptyBlock(&BB, DTU)) {
411 // BB is valid for cleanup here because we passed in DTU. F remains
412 // BB's parent until a DTU->getDomTree() event.
413 LVI->eraseBlock(&BB);
414 Changed = true;
415 }
416 }
417 EverChanged |= Changed;
418 } while (Changed);
419
420 LoopHeaders.clear();
421 // Flush only the Dominator Tree.
422 DTU->getDomTree();
423 LVI->enableDT();
424 return EverChanged;
425}
426
427// Replace uses of Cond with ToVal when safe to do so. If all uses are
428// replaced, we can remove Cond. We cannot blindly replace all uses of Cond
429// because we may incorrectly replace uses when guards/assumes are uses of
430// of `Cond` and we used the guards/assume to reason about the `Cond` value
431// at the end of block. RAUW unconditionally replaces all uses
432// including the guards/assumes themselves and the uses before the
433// guard/assume.
434static void ReplaceFoldableUses(Instruction *Cond, Value *ToVal) {
435 assert(Cond->getType() == ToVal->getType())((Cond->getType() == ToVal->getType()) ? static_cast<
void> (0) : __assert_fail ("Cond->getType() == ToVal->getType()"
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 435, __PRETTY_FUNCTION__))
;
436 auto *BB = Cond->getParent();
437 // We can unconditionally replace all uses in non-local blocks (i.e. uses
438 // strictly dominated by BB), since LVI information is true from the
439 // terminator of BB.
440 replaceNonLocalUsesWith(Cond, ToVal);
441 for (Instruction &I : reverse(*BB)) {
442 // Reached the Cond whose uses we are trying to replace, so there are no
443 // more uses.
444 if (&I == Cond)
445 break;
446 // We only replace uses in instructions that are guaranteed to reach the end
447 // of BB, where we know Cond is ToVal.
448 if (!isGuaranteedToTransferExecutionToSuccessor(&I))
449 break;
450 I.replaceUsesOfWith(Cond, ToVal);
451 }
452 if (Cond->use_empty() && !Cond->mayHaveSideEffects())
453 Cond->eraseFromParent();
454}
455
456/// Return the cost of duplicating a piece of this block from first non-phi
457/// and before StopAt instruction to thread across it. Stop scanning the block
458/// when exceeding the threshold. If duplication is impossible, returns ~0U.
459static unsigned getJumpThreadDuplicationCost(BasicBlock *BB,
460 Instruction *StopAt,
461 unsigned Threshold) {
462 assert(StopAt->getParent() == BB && "Not an instruction from proper BB?")((StopAt->getParent() == BB && "Not an instruction from proper BB?"
) ? static_cast<void> (0) : __assert_fail ("StopAt->getParent() == BB && \"Not an instruction from proper BB?\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 462, __PRETTY_FUNCTION__))
;
463 /// Ignore PHI nodes, these will be flattened when duplication happens.
464 BasicBlock::const_iterator I(BB->getFirstNonPHI());
465
466 // FIXME: THREADING will delete values that are just used to compute the
467 // branch, so they shouldn't count against the duplication cost.
468
469 unsigned Bonus = 0;
470 if (BB->getTerminator() == StopAt) {
471 // Threading through a switch statement is particularly profitable. If this
472 // block ends in a switch, decrease its cost to make it more likely to
473 // happen.
474 if (isa<SwitchInst>(StopAt))
475 Bonus = 6;
476
477 // The same holds for indirect branches, but slightly more so.
478 if (isa<IndirectBrInst>(StopAt))
479 Bonus = 8;
480 }
481
482 // Bump the threshold up so the early exit from the loop doesn't skip the
483 // terminator-based Size adjustment at the end.
484 Threshold += Bonus;
485
486 // Sum up the cost of each instruction until we get to the terminator. Don't
487 // include the terminator because the copy won't include it.
488 unsigned Size = 0;
489 for (; &*I != StopAt; ++I) {
490
491 // Stop scanning the block if we've reached the threshold.
492 if (Size > Threshold)
493 return Size;
494
495 // Debugger intrinsics don't incur code size.
496 if (isa<DbgInfoIntrinsic>(I)) continue;
497
498 // If this is a pointer->pointer bitcast, it is free.
499 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
500 continue;
501
502 // Bail out if this instruction gives back a token type, it is not possible
503 // to duplicate it if it is used outside this BB.
504 if (I->getType()->isTokenTy() && I->isUsedOutsideOfBlock(BB))
505 return ~0U;
506
507 // All other instructions count for at least one unit.
508 ++Size;
509
510 // Calls are more expensive. If they are non-intrinsic calls, we model them
511 // as having cost of 4. If they are a non-vector intrinsic, we model them
512 // as having cost of 2 total, and if they are a vector intrinsic, we model
513 // them as having cost 1.
514 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
515 if (CI->cannotDuplicate() || CI->isConvergent())
516 // Blocks with NoDuplicate are modelled as having infinite cost, so they
517 // are never duplicated.
518 return ~0U;
519 else if (!isa<IntrinsicInst>(CI))
520 Size += 3;
521 else if (!CI->getType()->isVectorTy())
522 Size += 1;
523 }
524 }
525
526 return Size > Bonus ? Size - Bonus : 0;
527}
528
529/// FindLoopHeaders - We do not want jump threading to turn proper loop
530/// structures into irreducible loops. Doing this breaks up the loop nesting
531/// hierarchy and pessimizes later transformations. To prevent this from
532/// happening, we first have to find the loop headers. Here we approximate this
533/// by finding targets of backedges in the CFG.
534///
535/// Note that there definitely are cases when we want to allow threading of
536/// edges across a loop header. For example, threading a jump from outside the
537/// loop (the preheader) to an exit block of the loop is definitely profitable.
538/// It is also almost always profitable to thread backedges from within the loop
539/// to exit blocks, and is often profitable to thread backedges to other blocks
540/// within the loop (forming a nested loop). This simple analysis is not rich
541/// enough to track all of these properties and keep it up-to-date as the CFG
542/// mutates, so we don't allow any of these transformations.
543void JumpThreadingPass::FindLoopHeaders(Function &F) {
544 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
545 FindFunctionBackedges(F, Edges);
546
547 for (const auto &Edge : Edges)
548 LoopHeaders.insert(Edge.second);
549}
550
551/// getKnownConstant - Helper method to determine if we can thread over a
552/// terminator with the given value as its condition, and if so what value to
553/// use for that. What kind of value this is depends on whether we want an
554/// integer or a block address, but an undef is always accepted.
555/// Returns null if Val is null or not an appropriate constant.
556static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
557 if (!Val)
558 return nullptr;
559
560 // Undef is "known" enough.
561 if (UndefValue *U = dyn_cast<UndefValue>(Val))
562 return U;
563
564 if (Preference == WantBlockAddress)
565 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
566
567 return dyn_cast<ConstantInt>(Val);
568}
569
570/// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
571/// if we can infer that the value is a known ConstantInt/BlockAddress or undef
572/// in any of our predecessors. If so, return the known list of value and pred
573/// BB in the result vector.
574///
575/// This returns true if there were any known values.
576bool JumpThreadingPass::ComputeValueKnownInPredecessorsImpl(
577 Value *V, BasicBlock *BB, PredValueInfo &Result,
578 ConstantPreference Preference,
579 DenseSet<std::pair<Value *, BasicBlock *>> &RecursionSet,
580 Instruction *CxtI) {
581 // This method walks up use-def chains recursively. Because of this, we could
582 // get into an infinite loop going around loops in the use-def chain. To
583 // prevent this, keep track of what (value, block) pairs we've already visited
584 // and terminate the search if we loop back to them
585 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
586 return false;
587
588 // If V is a constant, then it is known in all predecessors.
589 if (Constant *KC = getKnownConstant(V, Preference)) {
590 for (BasicBlock *Pred : predecessors(BB))
591 Result.push_back(std::make_pair(KC, Pred));
592
593 return !Result.empty();
594 }
595
596 // If V is a non-instruction value, or an instruction in a different block,
597 // then it can't be derived from a PHI.
598 Instruction *I = dyn_cast<Instruction>(V);
599 if (!I || I->getParent() != BB) {
600
601 // Okay, if this is a live-in value, see if it has a known value at the end
602 // of any of our predecessors.
603 //
604 // FIXME: This should be an edge property, not a block end property.
605 /// TODO: Per PR2563, we could infer value range information about a
606 /// predecessor based on its terminator.
607 //
608 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
609 // "I" is a non-local compare-with-a-constant instruction. This would be
610 // able to handle value inequalities better, for example if the compare is
611 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
612 // Perhaps getConstantOnEdge should be smart enough to do this?
613
614 if (DTU->hasPendingDomTreeUpdates())
615 LVI->disableDT();
616 else
617 LVI->enableDT();
618 for (BasicBlock *P : predecessors(BB)) {
619 // If the value is known by LazyValueInfo to be a constant in a
620 // predecessor, use that information to try to thread this block.
621 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB, CxtI);
622 if (Constant *KC = getKnownConstant(PredCst, Preference))
623 Result.push_back(std::make_pair(KC, P));
624 }
625
626 return !Result.empty();
627 }
628
629 /// If I is a PHI node, then we know the incoming values for any constants.
630 if (PHINode *PN = dyn_cast<PHINode>(I)) {
631 if (DTU->hasPendingDomTreeUpdates())
632 LVI->disableDT();
633 else
634 LVI->enableDT();
635 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
636 Value *InVal = PN->getIncomingValue(i);
637 if (Constant *KC = getKnownConstant(InVal, Preference)) {
638 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
639 } else {
640 Constant *CI = LVI->getConstantOnEdge(InVal,
641 PN->getIncomingBlock(i),
642 BB, CxtI);
643 if (Constant *KC = getKnownConstant(CI, Preference))
644 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
645 }
646 }
647
648 return !Result.empty();
649 }
650
651 // Handle Cast instructions. Only see through Cast when the source operand is
652 // PHI or Cmp to save the compilation time.
653 if (CastInst *CI = dyn_cast<CastInst>(I)) {
654 Value *Source = CI->getOperand(0);
655 if (!isa<PHINode>(Source) && !isa<CmpInst>(Source))
656 return false;
657 ComputeValueKnownInPredecessorsImpl(Source, BB, Result, Preference,
658 RecursionSet, CxtI);
659 if (Result.empty())
660 return false;
661
662 // Convert the known values.
663 for (auto &R : Result)
664 R.first = ConstantExpr::getCast(CI->getOpcode(), R.first, CI->getType());
665
666 return true;
667 }
668
669 // Handle some boolean conditions.
670 if (I->getType()->getPrimitiveSizeInBits() == 1) {
671 assert(Preference == WantInteger && "One-bit non-integer type?")((Preference == WantInteger && "One-bit non-integer type?"
) ? static_cast<void> (0) : __assert_fail ("Preference == WantInteger && \"One-bit non-integer type?\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 671, __PRETTY_FUNCTION__))
;
672 // X | true -> true
673 // X & false -> false
674 if (I->getOpcode() == Instruction::Or ||
675 I->getOpcode() == Instruction::And) {
676 PredValueInfoTy LHSVals, RHSVals;
677
678 ComputeValueKnownInPredecessorsImpl(I->getOperand(0), BB, LHSVals,
679 WantInteger, RecursionSet, CxtI);
680 ComputeValueKnownInPredecessorsImpl(I->getOperand(1), BB, RHSVals,
681 WantInteger, RecursionSet, CxtI);
682
683 if (LHSVals.empty() && RHSVals.empty())
684 return false;
685
686 ConstantInt *InterestingVal;
687 if (I->getOpcode() == Instruction::Or)
688 InterestingVal = ConstantInt::getTrue(I->getContext());
689 else
690 InterestingVal = ConstantInt::getFalse(I->getContext());
691
692 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
693
694 // Scan for the sentinel. If we find an undef, force it to the
695 // interesting value: x|undef -> true and x&undef -> false.
696 for (const auto &LHSVal : LHSVals)
697 if (LHSVal.first == InterestingVal || isa<UndefValue>(LHSVal.first)) {
698 Result.emplace_back(InterestingVal, LHSVal.second);
699 LHSKnownBBs.insert(LHSVal.second);
700 }
701 for (const auto &RHSVal : RHSVals)
702 if (RHSVal.first == InterestingVal || isa<UndefValue>(RHSVal.first)) {
703 // If we already inferred a value for this block on the LHS, don't
704 // re-add it.
705 if (!LHSKnownBBs.count(RHSVal.second))
706 Result.emplace_back(InterestingVal, RHSVal.second);
707 }
708
709 return !Result.empty();
710 }
711
712 // Handle the NOT form of XOR.
713 if (I->getOpcode() == Instruction::Xor &&
714 isa<ConstantInt>(I->getOperand(1)) &&
715 cast<ConstantInt>(I->getOperand(1))->isOne()) {
716 ComputeValueKnownInPredecessorsImpl(I->getOperand(0), BB, Result,
717 WantInteger, RecursionSet, CxtI);
718 if (Result.empty())
719 return false;
720
721 // Invert the known values.
722 for (auto &R : Result)
723 R.first = ConstantExpr::getNot(R.first);
724
725 return true;
726 }
727
728 // Try to simplify some other binary operator values.
729 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
730 assert(Preference != WantBlockAddress((Preference != WantBlockAddress && "A binary operator creating a block address?"
) ? static_cast<void> (0) : __assert_fail ("Preference != WantBlockAddress && \"A binary operator creating a block address?\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 731, __PRETTY_FUNCTION__))
731 && "A binary operator creating a block address?")((Preference != WantBlockAddress && "A binary operator creating a block address?"
) ? static_cast<void> (0) : __assert_fail ("Preference != WantBlockAddress && \"A binary operator creating a block address?\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 731, __PRETTY_FUNCTION__))
;
732 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
733 PredValueInfoTy LHSVals;
734 ComputeValueKnownInPredecessorsImpl(BO->getOperand(0), BB, LHSVals,
735 WantInteger, RecursionSet, CxtI);
736
737 // Try to use constant folding to simplify the binary operator.
738 for (const auto &LHSVal : LHSVals) {
739 Constant *V = LHSVal.first;
740 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
741
742 if (Constant *KC = getKnownConstant(Folded, WantInteger))
743 Result.push_back(std::make_pair(KC, LHSVal.second));
744 }
745 }
746
747 return !Result.empty();
748 }
749
750 // Handle compare with phi operand, where the PHI is defined in this block.
751 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
752 assert(Preference == WantInteger && "Compares only produce integers")((Preference == WantInteger && "Compares only produce integers"
) ? static_cast<void> (0) : __assert_fail ("Preference == WantInteger && \"Compares only produce integers\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 752, __PRETTY_FUNCTION__))
;
753 Type *CmpType = Cmp->getType();
754 Value *CmpLHS = Cmp->getOperand(0);
755 Value *CmpRHS = Cmp->getOperand(1);
756 CmpInst::Predicate Pred = Cmp->getPredicate();
757
758 PHINode *PN = dyn_cast<PHINode>(CmpLHS);
759 if (!PN)
760 PN = dyn_cast<PHINode>(CmpRHS);
761 if (PN && PN->getParent() == BB) {
762 const DataLayout &DL = PN->getModule()->getDataLayout();
763 // We can do this simplification if any comparisons fold to true or false.
764 // See if any do.
765 if (DTU->hasPendingDomTreeUpdates())
766 LVI->disableDT();
767 else
768 LVI->enableDT();
769 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
770 BasicBlock *PredBB = PN->getIncomingBlock(i);
771 Value *LHS, *RHS;
772 if (PN == CmpLHS) {
773 LHS = PN->getIncomingValue(i);
774 RHS = CmpRHS->DoPHITranslation(BB, PredBB);
775 } else {
776 LHS = CmpLHS->DoPHITranslation(BB, PredBB);
777 RHS = PN->getIncomingValue(i);
778 }
779 Value *Res = SimplifyCmpInst(Pred, LHS, RHS, {DL});
780 if (!Res) {
781 if (!isa<Constant>(RHS))
782 continue;
783
784 // getPredicateOnEdge call will make no sense if LHS is defined in BB.
785 auto LHSInst = dyn_cast<Instruction>(LHS);
786 if (LHSInst && LHSInst->getParent() == BB)
787 continue;
788
789 LazyValueInfo::Tristate
790 ResT = LVI->getPredicateOnEdge(Pred, LHS,
791 cast<Constant>(RHS), PredBB, BB,
792 CxtI ? CxtI : Cmp);
793 if (ResT == LazyValueInfo::Unknown)
794 continue;
795 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
796 }
797
798 if (Constant *KC = getKnownConstant(Res, WantInteger))
799 Result.push_back(std::make_pair(KC, PredBB));
800 }
801
802 return !Result.empty();
803 }
804
805 // If comparing a live-in value against a constant, see if we know the
806 // live-in value on any predecessors.
807 if (isa<Constant>(CmpRHS) && !CmpType->isVectorTy()) {
808 Constant *CmpConst = cast<Constant>(CmpRHS);
809
810 if (!isa<Instruction>(CmpLHS) ||
811 cast<Instruction>(CmpLHS)->getParent() != BB) {
812 if (DTU->hasPendingDomTreeUpdates())
813 LVI->disableDT();
814 else
815 LVI->enableDT();
816 for (BasicBlock *P : predecessors(BB)) {
817 // If the value is known by LazyValueInfo to be a constant in a
818 // predecessor, use that information to try to thread this block.
819 LazyValueInfo::Tristate Res =
820 LVI->getPredicateOnEdge(Pred, CmpLHS,
821 CmpConst, P, BB, CxtI ? CxtI : Cmp);
822 if (Res == LazyValueInfo::Unknown)
823 continue;
824
825 Constant *ResC = ConstantInt::get(CmpType, Res);
826 Result.push_back(std::make_pair(ResC, P));
827 }
828
829 return !Result.empty();
830 }
831
832 // InstCombine can fold some forms of constant range checks into
833 // (icmp (add (x, C1)), C2). See if we have we have such a thing with
834 // x as a live-in.
835 {
836 using namespace PatternMatch;
837
838 Value *AddLHS;
839 ConstantInt *AddConst;
840 if (isa<ConstantInt>(CmpConst) &&
841 match(CmpLHS, m_Add(m_Value(AddLHS), m_ConstantInt(AddConst)))) {
842 if (!isa<Instruction>(AddLHS) ||
843 cast<Instruction>(AddLHS)->getParent() != BB) {
844 if (DTU->hasPendingDomTreeUpdates())
845 LVI->disableDT();
846 else
847 LVI->enableDT();
848 for (BasicBlock *P : predecessors(BB)) {
849 // If the value is known by LazyValueInfo to be a ConstantRange in
850 // a predecessor, use that information to try to thread this
851 // block.
852 ConstantRange CR = LVI->getConstantRangeOnEdge(
853 AddLHS, P, BB, CxtI ? CxtI : cast<Instruction>(CmpLHS));
854 // Propagate the range through the addition.
855 CR = CR.add(AddConst->getValue());
856
857 // Get the range where the compare returns true.
858 ConstantRange CmpRange = ConstantRange::makeExactICmpRegion(
859 Pred, cast<ConstantInt>(CmpConst)->getValue());
860
861 Constant *ResC;
862 if (CmpRange.contains(CR))
863 ResC = ConstantInt::getTrue(CmpType);
864 else if (CmpRange.inverse().contains(CR))
865 ResC = ConstantInt::getFalse(CmpType);
866 else
867 continue;
868
869 Result.push_back(std::make_pair(ResC, P));
870 }
871
872 return !Result.empty();
873 }
874 }
875 }
876
877 // Try to find a constant value for the LHS of a comparison,
878 // and evaluate it statically if we can.
879 PredValueInfoTy LHSVals;
880 ComputeValueKnownInPredecessorsImpl(I->getOperand(0), BB, LHSVals,
881 WantInteger, RecursionSet, CxtI);
882
883 for (const auto &LHSVal : LHSVals) {
884 Constant *V = LHSVal.first;
885 Constant *Folded = ConstantExpr::getCompare(Pred, V, CmpConst);
886 if (Constant *KC = getKnownConstant(Folded, WantInteger))
887 Result.push_back(std::make_pair(KC, LHSVal.second));
888 }
889
890 return !Result.empty();
891 }
892 }
893
894 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
895 // Handle select instructions where at least one operand is a known constant
896 // and we can figure out the condition value for any predecessor block.
897 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
898 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
899 PredValueInfoTy Conds;
900 if ((TrueVal || FalseVal) &&
901 ComputeValueKnownInPredecessorsImpl(SI->getCondition(), BB, Conds,
902 WantInteger, RecursionSet, CxtI)) {
903 for (auto &C : Conds) {
904 Constant *Cond = C.first;
905
906 // Figure out what value to use for the condition.
907 bool KnownCond;
908 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
909 // A known boolean.
910 KnownCond = CI->isOne();
911 } else {
912 assert(isa<UndefValue>(Cond) && "Unexpected condition value")((isa<UndefValue>(Cond) && "Unexpected condition value"
) ? static_cast<void> (0) : __assert_fail ("isa<UndefValue>(Cond) && \"Unexpected condition value\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 912, __PRETTY_FUNCTION__))
;
913 // Either operand will do, so be sure to pick the one that's a known
914 // constant.
915 // FIXME: Do this more cleverly if both values are known constants?
916 KnownCond = (TrueVal != nullptr);
917 }
918
919 // See if the select has a known constant value for this predecessor.
920 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
921 Result.push_back(std::make_pair(Val, C.second));
922 }
923
924 return !Result.empty();
925 }
926 }
927
928 // If all else fails, see if LVI can figure out a constant value for us.
929 if (DTU->hasPendingDomTreeUpdates())
930 LVI->disableDT();
931 else
932 LVI->enableDT();
933 Constant *CI = LVI->getConstant(V, BB, CxtI);
934 if (Constant *KC = getKnownConstant(CI, Preference)) {
935 for (BasicBlock *Pred : predecessors(BB))
936 Result.push_back(std::make_pair(KC, Pred));
937 }
938
939 return !Result.empty();
940}
941
942/// GetBestDestForBranchOnUndef - If we determine that the specified block ends
943/// in an undefined jump, decide which block is best to revector to.
944///
945/// Since we can pick an arbitrary destination, we pick the successor with the
946/// fewest predecessors. This should reduce the in-degree of the others.
947static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
948 Instruction *BBTerm = BB->getTerminator();
949 unsigned MinSucc = 0;
950 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
951 // Compute the successor with the minimum number of predecessors.
952 unsigned MinNumPreds = pred_size(TestBB);
953 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
954 TestBB = BBTerm->getSuccessor(i);
955 unsigned NumPreds = pred_size(TestBB);
956 if (NumPreds < MinNumPreds) {
957 MinSucc = i;
958 MinNumPreds = NumPreds;
959 }
960 }
961
962 return MinSucc;
963}
964
965static bool hasAddressTakenAndUsed(BasicBlock *BB) {
966 if (!BB->hasAddressTaken()) return false;
967
968 // If the block has its address taken, it may be a tree of dead constants
969 // hanging off of it. These shouldn't keep the block alive.
970 BlockAddress *BA = BlockAddress::get(BB);
971 BA->removeDeadConstantUsers();
972 return !BA->use_empty();
973}
974
975/// ProcessBlock - If there are any predecessors whose control can be threaded
976/// through to a successor, transform them now.
977bool JumpThreadingPass::ProcessBlock(BasicBlock *BB) {
978 // If the block is trivially dead, just return and let the caller nuke it.
979 // This simplifies other transformations.
980 if (DTU->isBBPendingDeletion(BB) ||
1
Assuming the condition is false
981 (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()))
2
Assuming the condition is false
982 return false;
983
984 // If this block has a single predecessor, and if that pred has a single
985 // successor, merge the blocks. This encourages recursive jump threading
986 // because now the condition in this block can be threaded through
987 // predecessors of our predecessor block.
988 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
3
Assuming 'SinglePred' is null
4
Taking false branch
989 const Instruction *TI = SinglePred->getTerminator();
990 if (!TI->isExceptionalTerminator() && TI->getNumSuccessors() == 1 &&
991 SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
992 // If SinglePred was a loop header, BB becomes one.
993 if (LoopHeaders.erase(SinglePred))
994 LoopHeaders.insert(BB);
995
996 LVI->eraseBlock(SinglePred);
997 MergeBasicBlockIntoOnlyPred(BB, DTU);
998
999 // Now that BB is merged into SinglePred (i.e. SinglePred Code followed by
1000 // BB code within one basic block `BB`), we need to invalidate the LVI
1001 // information associated with BB, because the LVI information need not be
1002 // true for all of BB after the merge. For example,
1003 // Before the merge, LVI info and code is as follows:
1004 // SinglePred: <LVI info1 for %p val>
1005 // %y = use of %p
1006 // call @exit() // need not transfer execution to successor.
1007 // assume(%p) // from this point on %p is true
1008 // br label %BB
1009 // BB: <LVI info2 for %p val, i.e. %p is true>
1010 // %x = use of %p
1011 // br label exit
1012 //
1013 // Note that this LVI info for blocks BB and SinglPred is correct for %p
1014 // (info2 and info1 respectively). After the merge and the deletion of the
1015 // LVI info1 for SinglePred. We have the following code:
1016 // BB: <LVI info2 for %p val>
1017 // %y = use of %p
1018 // call @exit()
1019 // assume(%p)
1020 // %x = use of %p <-- LVI info2 is correct from here onwards.
1021 // br label exit
1022 // LVI info2 for BB is incorrect at the beginning of BB.
1023
1024 // Invalidate LVI information for BB if the LVI is not provably true for
1025 // all of BB.
1026 if (!isGuaranteedToTransferExecutionToSuccessor(BB))
1027 LVI->eraseBlock(BB);
1028 return true;
1029 }
1030 }
1031
1032 if (TryToUnfoldSelectInCurrBB(BB))
5
Taking false branch
1033 return true;
1034
1035 // Look if we can propagate guards to predecessors.
1036 if (HasGuards && ProcessGuards(BB))
6
Assuming the condition is false
1037 return true;
1038
1039 // What kind of constant we're looking for.
1040 ConstantPreference Preference = WantInteger;
1041
1042 // Look to see if the terminator is a conditional branch, switch or indirect
1043 // branch, if not we can't thread it.
1044 Value *Condition;
1045 Instruction *Terminator = BB->getTerminator();
1046 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
7
Taking true branch
1047 // Can't thread an unconditional jump.
1048 if (BI->isUnconditional()) return false;
8
Taking false branch
1049 Condition = BI->getCondition();
1050 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
1051 Condition = SI->getCondition();
1052 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
1053 // Can't thread indirect branch with no successors.
1054 if (IB->getNumSuccessors() == 0) return false;
1055 Condition = IB->getAddress()->stripPointerCasts();
1056 Preference = WantBlockAddress;
1057 } else {
1058 return false; // Must be an invoke or callbr.
1059 }
1060
1061 // Run constant folding to see if we can reduce the condition to a simple
1062 // constant.
1063 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
9
Taking true branch
1064 Value *SimpleVal =
1065 ConstantFoldInstruction(I, BB->getModule()->getDataLayout(), TLI);
1066 if (SimpleVal) {
10
Assuming 'SimpleVal' is null
11
Taking false branch
1067 I->replaceAllUsesWith(SimpleVal);
1068 if (isInstructionTriviallyDead(I, TLI))
1069 I->eraseFromParent();
1070 Condition = SimpleVal;
1071 }
1072 }
1073
1074 // If the terminator is branching on an undef, we can pick any of the
1075 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
1076 if (isa<UndefValue>(Condition)) {
12
Taking false branch
1077 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
1078 std::vector<DominatorTree::UpdateType> Updates;
1079
1080 // Fold the branch/switch.
1081 Instruction *BBTerm = BB->getTerminator();
1082 Updates.reserve(BBTerm->getNumSuccessors());
1083 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
1084 if (i == BestSucc) continue;
1085 BasicBlock *Succ = BBTerm->getSuccessor(i);
1086 Succ->removePredecessor(BB, true);
1087 Updates.push_back({DominatorTree::Delete, BB, Succ});
1088 }
1089
1090 LLVM_DEBUG(dbgs() << " In block '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " In block '" <<
BB->getName() << "' folding undef terminator: " <<
*BBTerm << '\n'; } } while (false)
1091 << "' folding undef terminator: " << *BBTerm << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " In block '" <<
BB->getName() << "' folding undef terminator: " <<
*BBTerm << '\n'; } } while (false)
;
1092 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
1093 BBTerm->eraseFromParent();
1094 DTU->applyUpdatesPermissive(Updates);
1095 return true;
1096 }
1097
1098 // If the terminator of this block is branching on a constant, simplify the
1099 // terminator to an unconditional branch. This can occur due to threading in
1100 // other blocks.
1101 if (getKnownConstant(Condition, Preference)) {
13
Taking false branch
1102 LLVM_DEBUG(dbgs() << " In block '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " In block '" <<
BB->getName() << "' folding terminator: " << *
BB->getTerminator() << '\n'; } } while (false)
1103 << "' folding terminator: " << *BB->getTerminator()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " In block '" <<
BB->getName() << "' folding terminator: " << *
BB->getTerminator() << '\n'; } } while (false)
1104 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " In block '" <<
BB->getName() << "' folding terminator: " << *
BB->getTerminator() << '\n'; } } while (false)
;
1105 ++NumFolds;
1106 ConstantFoldTerminator(BB, true, nullptr, DTU);
1107 return true;
1108 }
1109
1110 Instruction *CondInst = dyn_cast<Instruction>(Condition);
1111
1112 // All the rest of our checks depend on the condition being an instruction.
1113 if (!CondInst) {
14
Taking false branch
1114 // FIXME: Unify this with code below.
1115 if (ProcessThreadableEdges(Condition, BB, Preference, Terminator))
1116 return true;
1117 return false;
1118 }
1119
1120 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
15
Taking false branch
1121 // If we're branching on a conditional, LVI might be able to determine
1122 // it's value at the branch instruction. We only handle comparisons
1123 // against a constant at this time.
1124 // TODO: This should be extended to handle switches as well.
1125 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1126 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
1127 if (CondBr && CondConst) {
1128 // We should have returned as soon as we turn a conditional branch to
1129 // unconditional. Because its no longer interesting as far as jump
1130 // threading is concerned.
1131 assert(CondBr->isConditional() && "Threading on unconditional terminator")((CondBr->isConditional() && "Threading on unconditional terminator"
) ? static_cast<void> (0) : __assert_fail ("CondBr->isConditional() && \"Threading on unconditional terminator\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 1131, __PRETTY_FUNCTION__))
;
1132
1133 if (DTU->hasPendingDomTreeUpdates())
1134 LVI->disableDT();
1135 else
1136 LVI->enableDT();
1137 LazyValueInfo::Tristate Ret =
1138 LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0),
1139 CondConst, CondBr);
1140 if (Ret != LazyValueInfo::Unknown) {
1141 unsigned ToRemove = Ret == LazyValueInfo::True ? 1 : 0;
1142 unsigned ToKeep = Ret == LazyValueInfo::True ? 0 : 1;
1143 BasicBlock *ToRemoveSucc = CondBr->getSuccessor(ToRemove);
1144 ToRemoveSucc->removePredecessor(BB, true);
1145 BranchInst *UncondBr =
1146 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
1147 UncondBr->setDebugLoc(CondBr->getDebugLoc());
1148 CondBr->eraseFromParent();
1149 if (CondCmp->use_empty())
1150 CondCmp->eraseFromParent();
1151 // We can safely replace *some* uses of the CondInst if it has
1152 // exactly one value as returned by LVI. RAUW is incorrect in the
1153 // presence of guards and assumes, that have the `Cond` as the use. This
1154 // is because we use the guards/assume to reason about the `Cond` value
1155 // at the end of block, but RAUW unconditionally replaces all uses
1156 // including the guards/assumes themselves and the uses before the
1157 // guard/assume.
1158 else if (CondCmp->getParent() == BB) {
1159 auto *CI = Ret == LazyValueInfo::True ?
1160 ConstantInt::getTrue(CondCmp->getType()) :
1161 ConstantInt::getFalse(CondCmp->getType());
1162 ReplaceFoldableUses(CondCmp, CI);
1163 }
1164 DTU->applyUpdatesPermissive(
1165 {{DominatorTree::Delete, BB, ToRemoveSucc}});
1166 return true;
1167 }
1168
1169 // We did not manage to simplify this branch, try to see whether
1170 // CondCmp depends on a known phi-select pattern.
1171 if (TryToUnfoldSelect(CondCmp, BB))
1172 return true;
1173 }
1174 }
1175
1176 if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
16
Taking false branch
1177 TryToUnfoldSelect(SI, BB);
1178
1179 // Check for some cases that are worth simplifying. Right now we want to look
1180 // for loads that are used by a switch or by the condition for the branch. If
1181 // we see one, check to see if it's partially redundant. If so, insert a PHI
1182 // which can then be used to thread the values.
1183 Value *SimplifyValue = CondInst;
1184 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
17
Taking false branch
1185 if (isa<Constant>(CondCmp->getOperand(1)))
1186 SimplifyValue = CondCmp->getOperand(0);
1187
1188 // TODO: There are other places where load PRE would be profitable, such as
1189 // more complex comparisons.
1190 if (LoadInst *LoadI = dyn_cast<LoadInst>(SimplifyValue))
18
Taking true branch
1191 if (SimplifyPartiallyRedundantLoad(LoadI))
19
Calling 'JumpThreadingPass::SimplifyPartiallyRedundantLoad'
1192 return true;
1193
1194 // Before threading, try to propagate profile data backwards:
1195 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
1196 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1197 updatePredecessorProfileMetadata(PN, BB);
1198
1199 // Handle a variety of cases where we are branching on something derived from
1200 // a PHI node in the current block. If we can prove that any predecessors
1201 // compute a predictable value based on a PHI node, thread those predecessors.
1202 if (ProcessThreadableEdges(CondInst, BB, Preference, Terminator))
1203 return true;
1204
1205 // If this is an otherwise-unfoldable branch on a phi node in the current
1206 // block, see if we can simplify.
1207 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
1208 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1209 return ProcessBranchOnPHI(PN);
1210
1211 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
1212 if (CondInst->getOpcode() == Instruction::Xor &&
1213 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1214 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
1215
1216 // Search for a stronger dominating condition that can be used to simplify a
1217 // conditional branch leaving BB.
1218 if (ProcessImpliedCondition(BB))
1219 return true;
1220
1221 return false;
1222}
1223
1224bool JumpThreadingPass::ProcessImpliedCondition(BasicBlock *BB) {
1225 auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
1226 if (!BI || !BI->isConditional())
1227 return false;
1228
1229 Value *Cond = BI->getCondition();
1230 BasicBlock *CurrentBB = BB;
1231 BasicBlock *CurrentPred = BB->getSinglePredecessor();
1232 unsigned Iter = 0;
1233
1234 auto &DL = BB->getModule()->getDataLayout();
1235
1236 while (CurrentPred && Iter++ < ImplicationSearchThreshold) {
1237 auto *PBI = dyn_cast<BranchInst>(CurrentPred->getTerminator());
1238 if (!PBI || !PBI->isConditional())
1239 return false;
1240 if (PBI->getSuccessor(0) != CurrentBB && PBI->getSuccessor(1) != CurrentBB)
1241 return false;
1242
1243 bool CondIsTrue = PBI->getSuccessor(0) == CurrentBB;
1244 Optional<bool> Implication =
1245 isImpliedCondition(PBI->getCondition(), Cond, DL, CondIsTrue);
1246 if (Implication) {
1247 BasicBlock *KeepSucc = BI->getSuccessor(*Implication ? 0 : 1);
1248 BasicBlock *RemoveSucc = BI->getSuccessor(*Implication ? 1 : 0);
1249 RemoveSucc->removePredecessor(BB);
1250 BranchInst *UncondBI = BranchInst::Create(KeepSucc, BI);
1251 UncondBI->setDebugLoc(BI->getDebugLoc());
1252 BI->eraseFromParent();
1253 DTU->applyUpdatesPermissive({{DominatorTree::Delete, BB, RemoveSucc}});
1254 return true;
1255 }
1256 CurrentBB = CurrentPred;
1257 CurrentPred = CurrentBB->getSinglePredecessor();
1258 }
1259
1260 return false;
1261}
1262
1263/// Return true if Op is an instruction defined in the given block.
1264static bool isOpDefinedInBlock(Value *Op, BasicBlock *BB) {
1265 if (Instruction *OpInst = dyn_cast<Instruction>(Op))
1266 if (OpInst->getParent() == BB)
1267 return true;
1268 return false;
1269}
1270
1271/// SimplifyPartiallyRedundantLoad - If LoadI is an obviously partially
1272/// redundant load instruction, eliminate it by replacing it with a PHI node.
1273/// This is an important optimization that encourages jump threading, and needs
1274/// to be run interlaced with other jump threading tasks.
1275bool JumpThreadingPass::SimplifyPartiallyRedundantLoad(LoadInst *LoadI) {
1276 // Don't hack volatile and ordered loads.
1277 if (!LoadI->isUnordered()) return false;
20
Taking false branch
1278
1279 // If the load is defined in a block with exactly one predecessor, it can't be
1280 // partially redundant.
1281 BasicBlock *LoadBB = LoadI->getParent();
1282 if (LoadBB->getSinglePredecessor())
21
Assuming the condition is false
22
Taking false branch
1283 return false;
1284
1285 // If the load is defined in an EH pad, it can't be partially redundant,
1286 // because the edges between the invoke and the EH pad cannot have other
1287 // instructions between them.
1288 if (LoadBB->isEHPad())
23
Assuming the condition is false
24
Taking false branch
1289 return false;
1290
1291 Value *LoadedPtr = LoadI->getOperand(0);
1292
1293 // If the loaded operand is defined in the LoadBB and its not a phi,
1294 // it can't be available in predecessors.
1295 if (isOpDefinedInBlock(LoadedPtr, LoadBB) && !isa<PHINode>(LoadedPtr))
1296 return false;
1297
1298 // Scan a few instructions up from the load, to see if it is obviously live at
1299 // the entry to its block.
1300 BasicBlock::iterator BBIt(LoadI);
1301 bool IsLoadCSE;
1302 if (Value *AvailableVal = FindAvailableLoadedValue(
25
Assuming 'AvailableVal' is null
26
Taking false branch
1303 LoadI, LoadBB, BBIt, DefMaxInstsToScan, AA, &IsLoadCSE)) {
1304 // If the value of the load is locally available within the block, just use
1305 // it. This frequently occurs for reg2mem'd allocas.
1306
1307 if (IsLoadCSE) {
1308 LoadInst *NLoadI = cast<LoadInst>(AvailableVal);
1309 combineMetadataForCSE(NLoadI, LoadI, false);
1310 };
1311
1312 // If the returned value is the load itself, replace with an undef. This can
1313 // only happen in dead loops.
1314 if (AvailableVal == LoadI)
1315 AvailableVal = UndefValue::get(LoadI->getType());
1316 if (AvailableVal->getType() != LoadI->getType())
1317 AvailableVal = CastInst::CreateBitOrPointerCast(
1318 AvailableVal, LoadI->getType(), "", LoadI);
1319 LoadI->replaceAllUsesWith(AvailableVal);
1320 LoadI->eraseFromParent();
1321 return true;
1322 }
1323
1324 // Otherwise, if we scanned the whole block and got to the top of the block,
1325 // we know the block is locally transparent to the load. If not, something
1326 // might clobber its value.
1327 if (BBIt != LoadBB->begin())
27
Taking false branch
1328 return false;
1329
1330 // If all of the loads and stores that feed the value have the same AA tags,
1331 // then we can propagate them onto any newly inserted loads.
1332 AAMDNodes AATags;
1333 LoadI->getAAMetadata(AATags);
1334
1335 SmallPtrSet<BasicBlock*, 8> PredsScanned;
1336
1337 using AvailablePredsTy = SmallVector<std::pair<BasicBlock *, Value *>, 8>;
1338
1339 AvailablePredsTy AvailablePreds;
1340 BasicBlock *OneUnavailablePred = nullptr;
28
'OneUnavailablePred' initialized to a null pointer value
1341 SmallVector<LoadInst*, 8> CSELoads;
1342
1343 // If we got here, the loaded value is transparent through to the start of the
1344 // block. Check to see if it is available in any of the predecessor blocks.
1345 for (BasicBlock *PredBB : predecessors(LoadBB)) {
1346 // If we already scanned this predecessor, skip it.
1347 if (!PredsScanned.insert(PredBB).second)
1348 continue;
1349
1350 BBIt = PredBB->end();
1351 unsigned NumScanedInst = 0;
1352 Value *PredAvailable = nullptr;
1353 // NOTE: We don't CSE load that is volatile or anything stronger than
1354 // unordered, that should have been checked when we entered the function.
1355 assert(LoadI->isUnordered() &&((LoadI->isUnordered() && "Attempting to CSE volatile or atomic loads"
) ? static_cast<void> (0) : __assert_fail ("LoadI->isUnordered() && \"Attempting to CSE volatile or atomic loads\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 1356, __PRETTY_FUNCTION__))
1356 "Attempting to CSE volatile or atomic loads")((LoadI->isUnordered() && "Attempting to CSE volatile or atomic loads"
) ? static_cast<void> (0) : __assert_fail ("LoadI->isUnordered() && \"Attempting to CSE volatile or atomic loads\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 1356, __PRETTY_FUNCTION__))
;
1357 // If this is a load on a phi pointer, phi-translate it and search
1358 // for available load/store to the pointer in predecessors.
1359 Value *Ptr = LoadedPtr->DoPHITranslation(LoadBB, PredBB);
1360 PredAvailable = FindAvailablePtrLoadStore(
1361 Ptr, LoadI->getType(), LoadI->isAtomic(), PredBB, BBIt,
1362 DefMaxInstsToScan, AA, &IsLoadCSE, &NumScanedInst);
1363
1364 // If PredBB has a single predecessor, continue scanning through the
1365 // single predecessor.
1366 BasicBlock *SinglePredBB = PredBB;
1367 while (!PredAvailable && SinglePredBB && BBIt == SinglePredBB->begin() &&
1368 NumScanedInst < DefMaxInstsToScan) {
1369 SinglePredBB = SinglePredBB->getSinglePredecessor();
1370 if (SinglePredBB) {
1371 BBIt = SinglePredBB->end();
1372 PredAvailable = FindAvailablePtrLoadStore(
1373 Ptr, LoadI->getType(), LoadI->isAtomic(), SinglePredBB, BBIt,
1374 (DefMaxInstsToScan - NumScanedInst), AA, &IsLoadCSE,
1375 &NumScanedInst);
1376 }
1377 }
1378
1379 if (!PredAvailable) {
1380 OneUnavailablePred = PredBB;
1381 continue;
1382 }
1383
1384 if (IsLoadCSE)
1385 CSELoads.push_back(cast<LoadInst>(PredAvailable));
1386
1387 // If so, this load is partially redundant. Remember this info so that we
1388 // can create a PHI node.
1389 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
1390 }
1391
1392 // If the loaded value isn't available in any predecessor, it isn't partially
1393 // redundant.
1394 if (AvailablePreds.empty()) return false;
29
Taking false branch
1395
1396 // Okay, the loaded value is available in at least one (and maybe all!)
1397 // predecessors. If the value is unavailable in more than one unique
1398 // predecessor, we want to insert a merge block for those common predecessors.
1399 // This ensures that we only have to insert one reload, thus not increasing
1400 // code size.
1401 BasicBlock *UnavailablePred = nullptr;
1402
1403 // If the value is unavailable in one of predecessors, we will end up
1404 // inserting a new instruction into them. It is only valid if all the
1405 // instructions before LoadI are guaranteed to pass execution to its
1406 // successor, or if LoadI is safe to speculate.
1407 // TODO: If this logic becomes more complex, and we will perform PRE insertion
1408 // farther than to a predecessor, we need to reuse the code from GVN's PRE.
1409 // It requires domination tree analysis, so for this simple case it is an
1410 // overkill.
1411 if (PredsScanned.size() != AvailablePreds.size() &&
30
Assuming the condition is false
1412 !isSafeToSpeculativelyExecute(LoadI))
1413 for (auto I = LoadBB->begin(); &*I != LoadI; ++I)
1414 if (!isGuaranteedToTransferExecutionToSuccessor(&*I))
1415 return false;
1416
1417 // If there is exactly one predecessor where the value is unavailable, the
1418 // already computed 'OneUnavailablePred' block is it. If it ends in an
1419 // unconditional branch, we know that it isn't a critical edge.
1420 if (PredsScanned.size() == AvailablePreds.size()+1 &&
31
Assuming the condition is true
1421 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
32
Called C++ object pointer is null
1422 UnavailablePred = OneUnavailablePred;
1423 } else if (PredsScanned.size() != AvailablePreds.size()) {
1424 // Otherwise, we had multiple unavailable predecessors or we had a critical
1425 // edge from the one.
1426 SmallVector<BasicBlock*, 8> PredsToSplit;
1427 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
1428
1429 for (const auto &AvailablePred : AvailablePreds)
1430 AvailablePredSet.insert(AvailablePred.first);
1431
1432 // Add all the unavailable predecessors to the PredsToSplit list.
1433 for (BasicBlock *P : predecessors(LoadBB)) {
1434 // If the predecessor is an indirect goto, we can't split the edge.
1435 // Same for CallBr.
1436 if (isa<IndirectBrInst>(P->getTerminator()) ||
1437 isa<CallBrInst>(P->getTerminator()))
1438 return false;
1439
1440 if (!AvailablePredSet.count(P))
1441 PredsToSplit.push_back(P);
1442 }
1443
1444 // Split them out to their own block.
1445 UnavailablePred = SplitBlockPreds(LoadBB, PredsToSplit, "thread-pre-split");
1446 }
1447
1448 // If the value isn't available in all predecessors, then there will be
1449 // exactly one where it isn't available. Insert a load on that edge and add
1450 // it to the AvailablePreds list.
1451 if (UnavailablePred) {
1452 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&((UnavailablePred->getTerminator()->getNumSuccessors() ==
1 && "Can't handle critical edge here!") ? static_cast
<void> (0) : __assert_fail ("UnavailablePred->getTerminator()->getNumSuccessors() == 1 && \"Can't handle critical edge here!\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 1453, __PRETTY_FUNCTION__))
1453 "Can't handle critical edge here!")((UnavailablePred->getTerminator()->getNumSuccessors() ==
1 && "Can't handle critical edge here!") ? static_cast
<void> (0) : __assert_fail ("UnavailablePred->getTerminator()->getNumSuccessors() == 1 && \"Can't handle critical edge here!\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 1453, __PRETTY_FUNCTION__))
;
1454 LoadInst *NewVal = new LoadInst(
1455 LoadI->getType(), LoadedPtr->DoPHITranslation(LoadBB, UnavailablePred),
1456 LoadI->getName() + ".pr", false, LoadI->getAlignment(),
1457 LoadI->getOrdering(), LoadI->getSyncScopeID(),
1458 UnavailablePred->getTerminator());
1459 NewVal->setDebugLoc(LoadI->getDebugLoc());
1460 if (AATags)
1461 NewVal->setAAMetadata(AATags);
1462
1463 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
1464 }
1465
1466 // Now we know that each predecessor of this block has a value in
1467 // AvailablePreds, sort them for efficient access as we're walking the preds.
1468 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1469
1470 // Create a PHI node at the start of the block for the PRE'd load value.
1471 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
1472 PHINode *PN = PHINode::Create(LoadI->getType(), std::distance(PB, PE), "",
1473 &LoadBB->front());
1474 PN->takeName(LoadI);
1475 PN->setDebugLoc(LoadI->getDebugLoc());
1476
1477 // Insert new entries into the PHI for each predecessor. A single block may
1478 // have multiple entries here.
1479 for (pred_iterator PI = PB; PI != PE; ++PI) {
1480 BasicBlock *P = *PI;
1481 AvailablePredsTy::iterator I =
1482 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1483 std::make_pair(P, (Value*)nullptr));
1484
1485 assert(I != AvailablePreds.end() && I->first == P &&((I != AvailablePreds.end() && I->first == P &&
"Didn't find entry for predecessor!") ? static_cast<void>
(0) : __assert_fail ("I != AvailablePreds.end() && I->first == P && \"Didn't find entry for predecessor!\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 1486, __PRETTY_FUNCTION__))
1486 "Didn't find entry for predecessor!")((I != AvailablePreds.end() && I->first == P &&
"Didn't find entry for predecessor!") ? static_cast<void>
(0) : __assert_fail ("I != AvailablePreds.end() && I->first == P && \"Didn't find entry for predecessor!\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 1486, __PRETTY_FUNCTION__))
;
1487
1488 // If we have an available predecessor but it requires casting, insert the
1489 // cast in the predecessor and use the cast. Note that we have to update the
1490 // AvailablePreds vector as we go so that all of the PHI entries for this
1491 // predecessor use the same bitcast.
1492 Value *&PredV = I->second;
1493 if (PredV->getType() != LoadI->getType())
1494 PredV = CastInst::CreateBitOrPointerCast(PredV, LoadI->getType(), "",
1495 P->getTerminator());
1496
1497 PN->addIncoming(PredV, I->first);
1498 }
1499
1500 for (LoadInst *PredLoadI : CSELoads) {
1501 combineMetadataForCSE(PredLoadI, LoadI, true);
1502 }
1503
1504 LoadI->replaceAllUsesWith(PN);
1505 LoadI->eraseFromParent();
1506
1507 return true;
1508}
1509
1510/// FindMostPopularDest - The specified list contains multiple possible
1511/// threadable destinations. Pick the one that occurs the most frequently in
1512/// the list.
1513static BasicBlock *
1514FindMostPopularDest(BasicBlock *BB,
1515 const SmallVectorImpl<std::pair<BasicBlock *,
1516 BasicBlock *>> &PredToDestList) {
1517 assert(!PredToDestList.empty())((!PredToDestList.empty()) ? static_cast<void> (0) : __assert_fail
("!PredToDestList.empty()", "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 1517, __PRETTY_FUNCTION__))
;
1518
1519 // Determine popularity. If there are multiple possible destinations, we
1520 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1521 // blocks with known and real destinations to threading undef. We'll handle
1522 // them later if interesting.
1523 DenseMap<BasicBlock*, unsigned> DestPopularity;
1524 for (const auto &PredToDest : PredToDestList)
1525 if (PredToDest.second)
1526 DestPopularity[PredToDest.second]++;
1527
1528 if (DestPopularity.empty())
1529 return nullptr;
1530
1531 // Find the most popular dest.
1532 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1533 BasicBlock *MostPopularDest = DPI->first;
1534 unsigned Popularity = DPI->second;
1535 SmallVector<BasicBlock*, 4> SamePopularity;
1536
1537 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1538 // If the popularity of this entry isn't higher than the popularity we've
1539 // seen so far, ignore it.
1540 if (DPI->second < Popularity)
1541 ; // ignore.
1542 else if (DPI->second == Popularity) {
1543 // If it is the same as what we've seen so far, keep track of it.
1544 SamePopularity.push_back(DPI->first);
1545 } else {
1546 // If it is more popular, remember it.
1547 SamePopularity.clear();
1548 MostPopularDest = DPI->first;
1549 Popularity = DPI->second;
1550 }
1551 }
1552
1553 // Okay, now we know the most popular destination. If there is more than one
1554 // destination, we need to determine one. This is arbitrary, but we need
1555 // to make a deterministic decision. Pick the first one that appears in the
1556 // successor list.
1557 if (!SamePopularity.empty()) {
1558 SamePopularity.push_back(MostPopularDest);
1559 Instruction *TI = BB->getTerminator();
1560 for (unsigned i = 0; ; ++i) {
1561 assert(i != TI->getNumSuccessors() && "Didn't find any successor!")((i != TI->getNumSuccessors() && "Didn't find any successor!"
) ? static_cast<void> (0) : __assert_fail ("i != TI->getNumSuccessors() && \"Didn't find any successor!\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 1561, __PRETTY_FUNCTION__))
;
1562
1563 if (!is_contained(SamePopularity, TI->getSuccessor(i)))
1564 continue;
1565
1566 MostPopularDest = TI->getSuccessor(i);
1567 break;
1568 }
1569 }
1570
1571 // Okay, we have finally picked the most popular destination.
1572 return MostPopularDest;
1573}
1574
1575bool JumpThreadingPass::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1576 ConstantPreference Preference,
1577 Instruction *CxtI) {
1578 // If threading this would thread across a loop header, don't even try to
1579 // thread the edge.
1580 if (LoopHeaders.count(BB))
1581 return false;
1582
1583 PredValueInfoTy PredValues;
1584 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference, CxtI))
1585 return false;
1586
1587 assert(!PredValues.empty() &&((!PredValues.empty() && "ComputeValueKnownInPredecessors returned true with no values"
) ? static_cast<void> (0) : __assert_fail ("!PredValues.empty() && \"ComputeValueKnownInPredecessors returned true with no values\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 1588, __PRETTY_FUNCTION__))
1588 "ComputeValueKnownInPredecessors returned true with no values")((!PredValues.empty() && "ComputeValueKnownInPredecessors returned true with no values"
) ? static_cast<void> (0) : __assert_fail ("!PredValues.empty() && \"ComputeValueKnownInPredecessors returned true with no values\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 1588, __PRETTY_FUNCTION__))
;
1589
1590 LLVM_DEBUG(dbgs() << "IN BB: " << *BB;do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "IN BB: " << *BB;
for (const auto &PredValue : PredValues) { dbgs() <<
" BB '" << BB->getName() << "': FOUND condition = "
<< *PredValue.first << " for pred '" << PredValue
.second->getName() << "'.\n"; }; } } while (false)
1591 for (const auto &PredValue : PredValues) {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "IN BB: " << *BB;
for (const auto &PredValue : PredValues) { dbgs() <<
" BB '" << BB->getName() << "': FOUND condition = "
<< *PredValue.first << " for pred '" << PredValue
.second->getName() << "'.\n"; }; } } while (false)
1592 dbgs() << " BB '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "IN BB: " << *BB;
for (const auto &PredValue : PredValues) { dbgs() <<
" BB '" << BB->getName() << "': FOUND condition = "
<< *PredValue.first << " for pred '" << PredValue
.second->getName() << "'.\n"; }; } } while (false)
1593 << "': FOUND condition = " << *PredValue.firstdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "IN BB: " << *BB;
for (const auto &PredValue : PredValues) { dbgs() <<
" BB '" << BB->getName() << "': FOUND condition = "
<< *PredValue.first << " for pred '" << PredValue
.second->getName() << "'.\n"; }; } } while (false)
1594 << " for pred '" << PredValue.second->getName() << "'.\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "IN BB: " << *BB;
for (const auto &PredValue : PredValues) { dbgs() <<
" BB '" << BB->getName() << "': FOUND condition = "
<< *PredValue.first << " for pred '" << PredValue
.second->getName() << "'.\n"; }; } } while (false)
1595 })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "IN BB: " << *BB;
for (const auto &PredValue : PredValues) { dbgs() <<
" BB '" << BB->getName() << "': FOUND condition = "
<< *PredValue.first << " for pred '" << PredValue
.second->getName() << "'.\n"; }; } } while (false)
;
1596
1597 // Decide what we want to thread through. Convert our list of known values to
1598 // a list of known destinations for each pred. This also discards duplicate
1599 // predecessors and keeps track of the undefined inputs (which are represented
1600 // as a null dest in the PredToDestList).
1601 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1602 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1603
1604 BasicBlock *OnlyDest = nullptr;
1605 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1606 Constant *OnlyVal = nullptr;
1607 Constant *MultipleVal = (Constant *)(intptr_t)~0ULL;
1608
1609 for (const auto &PredValue : PredValues) {
1610 BasicBlock *Pred = PredValue.second;
1611 if (!SeenPreds.insert(Pred).second)
1612 continue; // Duplicate predecessor entry.
1613
1614 Constant *Val = PredValue.first;
1615
1616 BasicBlock *DestBB;
1617 if (isa<UndefValue>(Val))
1618 DestBB = nullptr;
1619 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
1620 assert(isa<ConstantInt>(Val) && "Expecting a constant integer")((isa<ConstantInt>(Val) && "Expecting a constant integer"
) ? static_cast<void> (0) : __assert_fail ("isa<ConstantInt>(Val) && \"Expecting a constant integer\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 1620, __PRETTY_FUNCTION__))
;
1621 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1622 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1623 assert(isa<ConstantInt>(Val) && "Expecting a constant integer")((isa<ConstantInt>(Val) && "Expecting a constant integer"
) ? static_cast<void> (0) : __assert_fail ("isa<ConstantInt>(Val) && \"Expecting a constant integer\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 1623, __PRETTY_FUNCTION__))
;
1624 DestBB = SI->findCaseValue(cast<ConstantInt>(Val))->getCaseSuccessor();
1625 } else {
1626 assert(isa<IndirectBrInst>(BB->getTerminator())((isa<IndirectBrInst>(BB->getTerminator()) &&
"Unexpected terminator") ? static_cast<void> (0) : __assert_fail
("isa<IndirectBrInst>(BB->getTerminator()) && \"Unexpected terminator\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 1627, __PRETTY_FUNCTION__))
1627 && "Unexpected terminator")((isa<IndirectBrInst>(BB->getTerminator()) &&
"Unexpected terminator") ? static_cast<void> (0) : __assert_fail
("isa<IndirectBrInst>(BB->getTerminator()) && \"Unexpected terminator\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 1627, __PRETTY_FUNCTION__))
;
1628 assert(isa<BlockAddress>(Val) && "Expecting a constant blockaddress")((isa<BlockAddress>(Val) && "Expecting a constant blockaddress"
) ? static_cast<void> (0) : __assert_fail ("isa<BlockAddress>(Val) && \"Expecting a constant blockaddress\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 1628, __PRETTY_FUNCTION__))
;
1629 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1630 }
1631
1632 // If we have exactly one destination, remember it for efficiency below.
1633 if (PredToDestList.empty()) {
1634 OnlyDest = DestBB;
1635 OnlyVal = Val;
1636 } else {
1637 if (OnlyDest != DestBB)
1638 OnlyDest = MultipleDestSentinel;
1639 // It possible we have same destination, but different value, e.g. default
1640 // case in switchinst.
1641 if (Val != OnlyVal)
1642 OnlyVal = MultipleVal;
1643 }
1644
1645 // If the predecessor ends with an indirect goto, we can't change its
1646 // destination. Same for CallBr.
1647 if (isa<IndirectBrInst>(Pred->getTerminator()) ||
1648 isa<CallBrInst>(Pred->getTerminator()))
1649 continue;
1650
1651 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1652 }
1653
1654 // If all edges were unthreadable, we fail.
1655 if (PredToDestList.empty())
1656 return false;
1657
1658 // If all the predecessors go to a single known successor, we want to fold,
1659 // not thread. By doing so, we do not need to duplicate the current block and
1660 // also miss potential opportunities in case we dont/cant duplicate.
1661 if (OnlyDest && OnlyDest != MultipleDestSentinel) {
1662 if (BB->hasNPredecessors(PredToDestList.size())) {
1663 bool SeenFirstBranchToOnlyDest = false;
1664 std::vector <DominatorTree::UpdateType> Updates;
1665 Updates.reserve(BB->getTerminator()->getNumSuccessors() - 1);
1666 for (BasicBlock *SuccBB : successors(BB)) {
1667 if (SuccBB == OnlyDest && !SeenFirstBranchToOnlyDest) {
1668 SeenFirstBranchToOnlyDest = true; // Don't modify the first branch.
1669 } else {
1670 SuccBB->removePredecessor(BB, true); // This is unreachable successor.
1671 Updates.push_back({DominatorTree::Delete, BB, SuccBB});
1672 }
1673 }
1674
1675 // Finally update the terminator.
1676 Instruction *Term = BB->getTerminator();
1677 BranchInst::Create(OnlyDest, Term);
1678 Term->eraseFromParent();
1679 DTU->applyUpdatesPermissive(Updates);
1680
1681 // If the condition is now dead due to the removal of the old terminator,
1682 // erase it.
1683 if (auto *CondInst = dyn_cast<Instruction>(Cond)) {
1684 if (CondInst->use_empty() && !CondInst->mayHaveSideEffects())
1685 CondInst->eraseFromParent();
1686 // We can safely replace *some* uses of the CondInst if it has
1687 // exactly one value as returned by LVI. RAUW is incorrect in the
1688 // presence of guards and assumes, that have the `Cond` as the use. This
1689 // is because we use the guards/assume to reason about the `Cond` value
1690 // at the end of block, but RAUW unconditionally replaces all uses
1691 // including the guards/assumes themselves and the uses before the
1692 // guard/assume.
1693 else if (OnlyVal && OnlyVal != MultipleVal &&
1694 CondInst->getParent() == BB)
1695 ReplaceFoldableUses(CondInst, OnlyVal);
1696 }
1697 return true;
1698 }
1699 }
1700
1701 // Determine which is the most common successor. If we have many inputs and
1702 // this block is a switch, we want to start by threading the batch that goes
1703 // to the most popular destination first. If we only know about one
1704 // threadable destination (the common case) we can avoid this.
1705 BasicBlock *MostPopularDest = OnlyDest;
1706
1707 if (MostPopularDest == MultipleDestSentinel) {
1708 // Remove any loop headers from the Dest list, ThreadEdge conservatively
1709 // won't process them, but we might have other destination that are eligible
1710 // and we still want to process.
1711 erase_if(PredToDestList,
1712 [&](const std::pair<BasicBlock *, BasicBlock *> &PredToDest) {
1713 return LoopHeaders.count(PredToDest.second) != 0;
1714 });
1715
1716 if (PredToDestList.empty())
1717 return false;
1718
1719 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1720 }
1721
1722 // Now that we know what the most popular destination is, factor all
1723 // predecessors that will jump to it into a single predecessor.
1724 SmallVector<BasicBlock*, 16> PredsToFactor;
1725 for (const auto &PredToDest : PredToDestList)
1726 if (PredToDest.second == MostPopularDest) {
1727 BasicBlock *Pred = PredToDest.first;
1728
1729 // This predecessor may be a switch or something else that has multiple
1730 // edges to the block. Factor each of these edges by listing them
1731 // according to # occurrences in PredsToFactor.
1732 for (BasicBlock *Succ : successors(Pred))
1733 if (Succ == BB)
1734 PredsToFactor.push_back(Pred);
1735 }
1736
1737 // If the threadable edges are branching on an undefined value, we get to pick
1738 // the destination that these predecessors should get to.
1739 if (!MostPopularDest)
1740 MostPopularDest = BB->getTerminator()->
1741 getSuccessor(GetBestDestForJumpOnUndef(BB));
1742
1743 // Ok, try to thread it!
1744 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1745}
1746
1747/// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1748/// a PHI node in the current block. See if there are any simplifications we
1749/// can do based on inputs to the phi node.
1750bool JumpThreadingPass::ProcessBranchOnPHI(PHINode *PN) {
1751 BasicBlock *BB = PN->getParent();
1752
1753 // TODO: We could make use of this to do it once for blocks with common PHI
1754 // values.
1755 SmallVector<BasicBlock*, 1> PredBBs;
1756 PredBBs.resize(1);
1757
1758 // If any of the predecessor blocks end in an unconditional branch, we can
1759 // *duplicate* the conditional branch into that block in order to further
1760 // encourage jump threading and to eliminate cases where we have branch on a
1761 // phi of an icmp (branch on icmp is much better).
1762 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1763 BasicBlock *PredBB = PN->getIncomingBlock(i);
1764 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1765 if (PredBr->isUnconditional()) {
1766 PredBBs[0] = PredBB;
1767 // Try to duplicate BB into PredBB.
1768 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1769 return true;
1770 }
1771 }
1772
1773 return false;
1774}
1775
1776/// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1777/// a xor instruction in the current block. See if there are any
1778/// simplifications we can do based on inputs to the xor.
1779bool JumpThreadingPass::ProcessBranchOnXOR(BinaryOperator *BO) {
1780 BasicBlock *BB = BO->getParent();
1781
1782 // If either the LHS or RHS of the xor is a constant, don't do this
1783 // optimization.
1784 if (isa<ConstantInt>(BO->getOperand(0)) ||
1785 isa<ConstantInt>(BO->getOperand(1)))
1786 return false;
1787
1788 // If the first instruction in BB isn't a phi, we won't be able to infer
1789 // anything special about any particular predecessor.
1790 if (!isa<PHINode>(BB->front()))
1791 return false;
1792
1793 // If this BB is a landing pad, we won't be able to split the edge into it.
1794 if (BB->isEHPad())
1795 return false;
1796
1797 // If we have a xor as the branch input to this block, and we know that the
1798 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1799 // the condition into the predecessor and fix that value to true, saving some
1800 // logical ops on that path and encouraging other paths to simplify.
1801 //
1802 // This copies something like this:
1803 //
1804 // BB:
1805 // %X = phi i1 [1], [%X']
1806 // %Y = icmp eq i32 %A, %B
1807 // %Z = xor i1 %X, %Y
1808 // br i1 %Z, ...
1809 //
1810 // Into:
1811 // BB':
1812 // %Y = icmp ne i32 %A, %B
1813 // br i1 %Y, ...
1814
1815 PredValueInfoTy XorOpValues;
1816 bool isLHS = true;
1817 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1818 WantInteger, BO)) {
1819 assert(XorOpValues.empty())((XorOpValues.empty()) ? static_cast<void> (0) : __assert_fail
("XorOpValues.empty()", "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 1819, __PRETTY_FUNCTION__))
;
1820 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1821 WantInteger, BO))
1822 return false;
1823 isLHS = false;
1824 }
1825
1826 assert(!XorOpValues.empty() &&((!XorOpValues.empty() && "ComputeValueKnownInPredecessors returned true with no values"
) ? static_cast<void> (0) : __assert_fail ("!XorOpValues.empty() && \"ComputeValueKnownInPredecessors returned true with no values\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 1827, __PRETTY_FUNCTION__))
1827 "ComputeValueKnownInPredecessors returned true with no values")((!XorOpValues.empty() && "ComputeValueKnownInPredecessors returned true with no values"
) ? static_cast<void> (0) : __assert_fail ("!XorOpValues.empty() && \"ComputeValueKnownInPredecessors returned true with no values\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 1827, __PRETTY_FUNCTION__))
;
1828
1829 // Scan the information to see which is most popular: true or false. The
1830 // predecessors can be of the set true, false, or undef.
1831 unsigned NumTrue = 0, NumFalse = 0;
1832 for (const auto &XorOpValue : XorOpValues) {
1833 if (isa<UndefValue>(XorOpValue.first))
1834 // Ignore undefs for the count.
1835 continue;
1836 if (cast<ConstantInt>(XorOpValue.first)->isZero())
1837 ++NumFalse;
1838 else
1839 ++NumTrue;
1840 }
1841
1842 // Determine which value to split on, true, false, or undef if neither.
1843 ConstantInt *SplitVal = nullptr;
1844 if (NumTrue > NumFalse)
1845 SplitVal = ConstantInt::getTrue(BB->getContext());
1846 else if (NumTrue != 0 || NumFalse != 0)
1847 SplitVal = ConstantInt::getFalse(BB->getContext());
1848
1849 // Collect all of the blocks that this can be folded into so that we can
1850 // factor this once and clone it once.
1851 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1852 for (const auto &XorOpValue : XorOpValues) {
1853 if (XorOpValue.first != SplitVal && !isa<UndefValue>(XorOpValue.first))
1854 continue;
1855
1856 BlocksToFoldInto.push_back(XorOpValue.second);
1857 }
1858
1859 // If we inferred a value for all of the predecessors, then duplication won't
1860 // help us. However, we can just replace the LHS or RHS with the constant.
1861 if (BlocksToFoldInto.size() ==
1862 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1863 if (!SplitVal) {
1864 // If all preds provide undef, just nuke the xor, because it is undef too.
1865 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1866 BO->eraseFromParent();
1867 } else if (SplitVal->isZero()) {
1868 // If all preds provide 0, replace the xor with the other input.
1869 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1870 BO->eraseFromParent();
1871 } else {
1872 // If all preds provide 1, set the computed value to 1.
1873 BO->setOperand(!isLHS, SplitVal);
1874 }
1875
1876 return true;
1877 }
1878
1879 // Try to duplicate BB into PredBB.
1880 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1881}
1882
1883/// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1884/// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1885/// NewPred using the entries from OldPred (suitably mapped).
1886static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1887 BasicBlock *OldPred,
1888 BasicBlock *NewPred,
1889 DenseMap<Instruction*, Value*> &ValueMap) {
1890 for (PHINode &PN : PHIBB->phis()) {
1891 // Ok, we have a PHI node. Figure out what the incoming value was for the
1892 // DestBlock.
1893 Value *IV = PN.getIncomingValueForBlock(OldPred);
1894
1895 // Remap the value if necessary.
1896 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1897 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1898 if (I != ValueMap.end())
1899 IV = I->second;
1900 }
1901
1902 PN.addIncoming(IV, NewPred);
1903 }
1904}
1905
1906/// ThreadEdge - We have decided that it is safe and profitable to factor the
1907/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1908/// across BB. Transform the IR to reflect this change.
1909bool JumpThreadingPass::ThreadEdge(BasicBlock *BB,
1910 const SmallVectorImpl<BasicBlock *> &PredBBs,
1911 BasicBlock *SuccBB) {
1912 // If threading to the same block as we come from, we would infinite loop.
1913 if (SuccBB == BB) {
1914 LLVM_DEBUG(dbgs() << " Not threading across BB '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Not threading across BB '"
<< BB->getName() << "' - would thread to self!\n"
; } } while (false)
1915 << "' - would thread to self!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Not threading across BB '"
<< BB->getName() << "' - would thread to self!\n"
; } } while (false)
;
1916 return false;
1917 }
1918
1919 // If threading this would thread across a loop header, don't thread the edge.
1920 // See the comments above FindLoopHeaders for justifications and caveats.
1921 if (LoopHeaders.count(BB) || LoopHeaders.count(SuccBB)) {
1922 LLVM_DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB
); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() <<
" Not threading across " << (BBIsHeader ? "loop header BB '"
: "block BB '") << BB->getName() << "' to dest "
<< (SuccIsHeader ? "loop header BB '" : "block BB '") <<
SuccBB->getName() << "' - it might create an irreducible loop!\n"
; }; } } while (false)
1923 bool BBIsHeader = LoopHeaders.count(BB);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB
); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() <<
" Not threading across " << (BBIsHeader ? "loop header BB '"
: "block BB '") << BB->getName() << "' to dest "
<< (SuccIsHeader ? "loop header BB '" : "block BB '") <<
SuccBB->getName() << "' - it might create an irreducible loop!\n"
; }; } } while (false)
1924 bool SuccIsHeader = LoopHeaders.count(SuccBB);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB
); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() <<
" Not threading across " << (BBIsHeader ? "loop header BB '"
: "block BB '") << BB->getName() << "' to dest "
<< (SuccIsHeader ? "loop header BB '" : "block BB '") <<
SuccBB->getName() << "' - it might create an irreducible loop!\n"
; }; } } while (false)
1925 dbgs() << " Not threading across "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB
); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() <<
" Not threading across " << (BBIsHeader ? "loop header BB '"
: "block BB '") << BB->getName() << "' to dest "
<< (SuccIsHeader ? "loop header BB '" : "block BB '") <<
SuccBB->getName() << "' - it might create an irreducible loop!\n"
; }; } } while (false)
1926 << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB
); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() <<
" Not threading across " << (BBIsHeader ? "loop header BB '"
: "block BB '") << BB->getName() << "' to dest "
<< (SuccIsHeader ? "loop header BB '" : "block BB '") <<
SuccBB->getName() << "' - it might create an irreducible loop!\n"
; }; } } while (false)
1927 << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB
); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() <<
" Not threading across " << (BBIsHeader ? "loop header BB '"
: "block BB '") << BB->getName() << "' to dest "
<< (SuccIsHeader ? "loop header BB '" : "block BB '") <<
SuccBB->getName() << "' - it might create an irreducible loop!\n"
; }; } } while (false)
1928 << SuccBB->getName() << "' - it might create an irreducible loop!\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB
); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() <<
" Not threading across " << (BBIsHeader ? "loop header BB '"
: "block BB '") << BB->getName() << "' to dest "
<< (SuccIsHeader ? "loop header BB '" : "block BB '") <<
SuccBB->getName() << "' - it might create an irreducible loop!\n"
; }; } } while (false)
1929 })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB
); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() <<
" Not threading across " << (BBIsHeader ? "loop header BB '"
: "block BB '") << BB->getName() << "' to dest "
<< (SuccIsHeader ? "loop header BB '" : "block BB '") <<
SuccBB->getName() << "' - it might create an irreducible loop!\n"
; }; } } while (false)
;
1930 return false;
1931 }
1932
1933 unsigned JumpThreadCost =
1934 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
1935 if (JumpThreadCost > BBDupThreshold) {
1936 LLVM_DEBUG(dbgs() << " Not threading BB '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Not threading BB '" <<
BB->getName() << "' - Cost is too high: " << JumpThreadCost
<< "\n"; } } while (false)
1937 << "' - Cost is too high: " << JumpThreadCost << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Not threading BB '" <<
BB->getName() << "' - Cost is too high: " << JumpThreadCost
<< "\n"; } } while (false)
;
1938 return false;
1939 }
1940
1941 // And finally, do it! Start by factoring the predecessors if needed.
1942 BasicBlock *PredBB;
1943 if (PredBBs.size() == 1)
1944 PredBB = PredBBs[0];
1945 else {
1946 LLVM_DEBUG(dbgs() << " Factoring out " << PredBBs.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Factoring out " <<
PredBBs.size() << " common predecessors.\n"; } } while
(false)
1947 << " common predecessors.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Factoring out " <<
PredBBs.size() << " common predecessors.\n"; } } while
(false)
;
1948 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
1949 }
1950
1951 // And finally, do it!
1952 LLVM_DEBUG(dbgs() << " Threading edge from '" << PredBB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Threading edge from '"
<< PredBB->getName() << "' to '" << SuccBB
->getName() << "' with cost: " << JumpThreadCost
<< ", across block:\n " << *BB << "\n";
} } while (false)
1953 << "' to '" << SuccBB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Threading edge from '"
<< PredBB->getName() << "' to '" << SuccBB
->getName() << "' with cost: " << JumpThreadCost
<< ", across block:\n " << *BB << "\n";
} } while (false)
1954 << "' with cost: " << JumpThreadCostdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Threading edge from '"
<< PredBB->getName() << "' to '" << SuccBB
->getName() << "' with cost: " << JumpThreadCost
<< ", across block:\n " << *BB << "\n";
} } while (false)
1955 << ", across block:\n " << *BB << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Threading edge from '"
<< PredBB->getName() << "' to '" << SuccBB
->getName() << "' with cost: " << JumpThreadCost
<< ", across block:\n " << *BB << "\n";
} } while (false)
;
1956
1957 if (DTU->hasPendingDomTreeUpdates())
1958 LVI->disableDT();
1959 else
1960 LVI->enableDT();
1961 LVI->threadEdge(PredBB, BB, SuccBB);
1962
1963 // We are going to have to map operands from the original BB block to the new
1964 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1965 // account for entry from PredBB.
1966 DenseMap<Instruction*, Value*> ValueMapping;
1967
1968 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1969 BB->getName()+".thread",
1970 BB->getParent(), BB);
1971 NewBB->moveAfter(PredBB);
1972
1973 // Set the block frequency of NewBB.
1974 if (HasProfileData) {
1975 auto NewBBFreq =
1976 BFI->getBlockFreq(PredBB) * BPI->getEdgeProbability(PredBB, BB);
1977 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
1978 }
1979
1980 BasicBlock::iterator BI = BB->begin();
1981 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1982 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1983
1984 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1985 // mapping and using it to remap operands in the cloned instructions.
1986 for (; !BI->isTerminator(); ++BI) {
1987 Instruction *New = BI->clone();
1988 New->setName(BI->getName());
1989 NewBB->getInstList().push_back(New);
1990 ValueMapping[&*BI] = New;
1991
1992 // Remap operands to patch up intra-block references.
1993 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1994 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1995 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1996 if (I != ValueMapping.end())
1997 New->setOperand(i, I->second);
1998 }
1999 }
2000
2001 // We didn't copy the terminator from BB over to NewBB, because there is now
2002 // an unconditional jump to SuccBB. Insert the unconditional jump.
2003 BranchInst *NewBI = BranchInst::Create(SuccBB, NewBB);
2004 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
2005
2006 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
2007 // PHI nodes for NewBB now.
2008 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
2009
2010 // Update the terminator of PredBB to jump to NewBB instead of BB. This
2011 // eliminates predecessors from BB, which requires us to simplify any PHI
2012 // nodes in BB.
2013 Instruction *PredTerm = PredBB->getTerminator();
2014 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
2015 if (PredTerm->getSuccessor(i) == BB) {
2016 BB->removePredecessor(PredBB, true);
2017 PredTerm->setSuccessor(i, NewBB);
2018 }
2019
2020 // Enqueue required DT updates.
2021 DTU->applyUpdatesPermissive({{DominatorTree::Insert, NewBB, SuccBB},
2022 {DominatorTree::Insert, PredBB, NewBB},
2023 {DominatorTree::Delete, PredBB, BB}});
2024
2025 // If there were values defined in BB that are used outside the block, then we
2026 // now have to update all uses of the value to use either the original value,
2027 // the cloned value, or some PHI derived value. This can require arbitrary
2028 // PHI insertion, of which we are prepared to do, clean these up now.
2029 SSAUpdater SSAUpdate;
2030 SmallVector<Use*, 16> UsesToRename;
2031
2032 for (Instruction &I : *BB) {
2033 // Scan all uses of this instruction to see if their uses are no longer
2034 // dominated by the previous def and if so, record them in UsesToRename.
2035 // Also, skip phi operands from PredBB - we'll remove them anyway.
2036 for (Use &U : I.uses()) {
2037 Instruction *User = cast<Instruction>(U.getUser());
2038 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
2039 if (UserPN->getIncomingBlock(U) == BB)
2040 continue;
2041 } else if (User->getParent() == BB)
2042 continue;
2043
2044 UsesToRename.push_back(&U);
2045 }
2046
2047 // If there are no uses outside the block, we're done with this instruction.
2048 if (UsesToRename.empty())
2049 continue;
2050 LLVM_DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "JT: Renaming non-local uses of: "
<< I << "\n"; } } while (false)
;
2051
2052 // We found a use of I outside of BB. Rename all uses of I that are outside
2053 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
2054 // with the two values we know.
2055 SSAUpdate.Initialize(I.getType(), I.getName());
2056 SSAUpdate.AddAvailableValue(BB, &I);
2057 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[&I]);
2058
2059 while (!UsesToRename.empty())
2060 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
2061 LLVM_DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "\n"; } } while (false)
;
2062 }
2063
2064 // At this point, the IR is fully up to date and consistent. Do a quick scan
2065 // over the new instructions and zap any that are constants or dead. This
2066 // frequently happens because of phi translation.
2067 SimplifyInstructionsInBlock(NewBB, TLI);
2068
2069 // Update the edge weight from BB to SuccBB, which should be less than before.
2070 UpdateBlockFreqAndEdgeWeight(PredBB, BB, NewBB, SuccBB);
2071
2072 // Threaded an edge!
2073 ++NumThreads;
2074 return true;
2075}
2076
2077/// Create a new basic block that will be the predecessor of BB and successor of
2078/// all blocks in Preds. When profile data is available, update the frequency of
2079/// this new block.
2080BasicBlock *JumpThreadingPass::SplitBlockPreds(BasicBlock *BB,
2081 ArrayRef<BasicBlock *> Preds,
2082 const char *Suffix) {
2083 SmallVector<BasicBlock *, 2> NewBBs;
2084
2085 // Collect the frequencies of all predecessors of BB, which will be used to
2086 // update the edge weight of the result of splitting predecessors.
2087 DenseMap<BasicBlock *, BlockFrequency> FreqMap;
2088 if (HasProfileData)
2089 for (auto Pred : Preds)
2090 FreqMap.insert(std::make_pair(
2091 Pred, BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, BB)));
2092
2093 // In the case when BB is a LandingPad block we create 2 new predecessors
2094 // instead of just one.
2095 if (BB->isLandingPad()) {
2096 std::string NewName = std::string(Suffix) + ".split-lp";
2097 SplitLandingPadPredecessors(BB, Preds, Suffix, NewName.c_str(), NewBBs);
2098 } else {
2099 NewBBs.push_back(SplitBlockPredecessors(BB, Preds, Suffix));
2100 }
2101
2102 std::vector<DominatorTree::UpdateType> Updates;
2103 Updates.reserve((2 * Preds.size()) + NewBBs.size());
2104 for (auto NewBB : NewBBs) {
2105 BlockFrequency NewBBFreq(0);
2106 Updates.push_back({DominatorTree::Insert, NewBB, BB});
2107 for (auto Pred : predecessors(NewBB)) {
2108 Updates.push_back({DominatorTree::Delete, Pred, BB});
2109 Updates.push_back({DominatorTree::Insert, Pred, NewBB});
2110 if (HasProfileData) // Update frequencies between Pred -> NewBB.
2111 NewBBFreq += FreqMap.lookup(Pred);
2112 }
2113 if (HasProfileData) // Apply the summed frequency to NewBB.
2114 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
2115 }
2116
2117 DTU->applyUpdatesPermissive(Updates);
2118 return NewBBs[0];
2119}
2120
2121bool JumpThreadingPass::doesBlockHaveProfileData(BasicBlock *BB) {
2122 const Instruction *TI = BB->getTerminator();
2123 assert(TI->getNumSuccessors() > 1 && "not a split")((TI->getNumSuccessors() > 1 && "not a split") ?
static_cast<void> (0) : __assert_fail ("TI->getNumSuccessors() > 1 && \"not a split\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 2123, __PRETTY_FUNCTION__))
;
2124
2125 MDNode *WeightsNode = TI->getMetadata(LLVMContext::MD_prof);
2126 if (!WeightsNode)
2127 return false;
2128
2129 MDString *MDName = cast<MDString>(WeightsNode->getOperand(0));
2130 if (MDName->getString() != "branch_weights")
2131 return false;
2132
2133 // Ensure there are weights for all of the successors. Note that the first
2134 // operand to the metadata node is a name, not a weight.
2135 return WeightsNode->getNumOperands() == TI->getNumSuccessors() + 1;
2136}
2137
2138/// Update the block frequency of BB and branch weight and the metadata on the
2139/// edge BB->SuccBB. This is done by scaling the weight of BB->SuccBB by 1 -
2140/// Freq(PredBB->BB) / Freq(BB->SuccBB).
2141void JumpThreadingPass::UpdateBlockFreqAndEdgeWeight(BasicBlock *PredBB,
2142 BasicBlock *BB,
2143 BasicBlock *NewBB,
2144 BasicBlock *SuccBB) {
2145 if (!HasProfileData)
2146 return;
2147
2148 assert(BFI && BPI && "BFI & BPI should have been created here")((BFI && BPI && "BFI & BPI should have been created here"
) ? static_cast<void> (0) : __assert_fail ("BFI && BPI && \"BFI & BPI should have been created here\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 2148, __PRETTY_FUNCTION__))
;
2149
2150 // As the edge from PredBB to BB is deleted, we have to update the block
2151 // frequency of BB.
2152 auto BBOrigFreq = BFI->getBlockFreq(BB);
2153 auto NewBBFreq = BFI->getBlockFreq(NewBB);
2154 auto BB2SuccBBFreq = BBOrigFreq * BPI->getEdgeProbability(BB, SuccBB);
2155 auto BBNewFreq = BBOrigFreq - NewBBFreq;
2156 BFI->setBlockFreq(BB, BBNewFreq.getFrequency());
2157
2158 // Collect updated outgoing edges' frequencies from BB and use them to update
2159 // edge probabilities.
2160 SmallVector<uint64_t, 4> BBSuccFreq;
2161 for (BasicBlock *Succ : successors(BB)) {
2162 auto SuccFreq = (Succ == SuccBB)
2163 ? BB2SuccBBFreq - NewBBFreq
2164 : BBOrigFreq * BPI->getEdgeProbability(BB, Succ);
2165 BBSuccFreq.push_back(SuccFreq.getFrequency());
2166 }
2167
2168 uint64_t MaxBBSuccFreq =
2169 *std::max_element(BBSuccFreq.begin(), BBSuccFreq.end());
2170
2171 SmallVector<BranchProbability, 4> BBSuccProbs;
2172 if (MaxBBSuccFreq == 0)
2173 BBSuccProbs.assign(BBSuccFreq.size(),
2174 {1, static_cast<unsigned>(BBSuccFreq.size())});
2175 else {
2176 for (uint64_t Freq : BBSuccFreq)
2177 BBSuccProbs.push_back(
2178 BranchProbability::getBranchProbability(Freq, MaxBBSuccFreq));
2179 // Normalize edge probabilities so that they sum up to one.
2180 BranchProbability::normalizeProbabilities(BBSuccProbs.begin(),
2181 BBSuccProbs.end());
2182 }
2183
2184 // Update edge probabilities in BPI.
2185 for (int I = 0, E = BBSuccProbs.size(); I < E; I++)
2186 BPI->setEdgeProbability(BB, I, BBSuccProbs[I]);
2187
2188 // Update the profile metadata as well.
2189 //
2190 // Don't do this if the profile of the transformed blocks was statically
2191 // estimated. (This could occur despite the function having an entry
2192 // frequency in completely cold parts of the CFG.)
2193 //
2194 // In this case we don't want to suggest to subsequent passes that the
2195 // calculated weights are fully consistent. Consider this graph:
2196 //
2197 // check_1
2198 // 50% / |
2199 // eq_1 | 50%
2200 // \ |
2201 // check_2
2202 // 50% / |
2203 // eq_2 | 50%
2204 // \ |
2205 // check_3
2206 // 50% / |
2207 // eq_3 | 50%
2208 // \ |
2209 //
2210 // Assuming the blocks check_* all compare the same value against 1, 2 and 3,
2211 // the overall probabilities are inconsistent; the total probability that the
2212 // value is either 1, 2 or 3 is 150%.
2213 //
2214 // As a consequence if we thread eq_1 -> check_2 to check_3, check_2->check_3
2215 // becomes 0%. This is even worse if the edge whose probability becomes 0% is
2216 // the loop exit edge. Then based solely on static estimation we would assume
2217 // the loop was extremely hot.
2218 //
2219 // FIXME this locally as well so that BPI and BFI are consistent as well. We
2220 // shouldn't make edges extremely likely or unlikely based solely on static
2221 // estimation.
2222 if (BBSuccProbs.size() >= 2 && doesBlockHaveProfileData(BB)) {
2223 SmallVector<uint32_t, 4> Weights;
2224 for (auto Prob : BBSuccProbs)
2225 Weights.push_back(Prob.getNumerator());
2226
2227 auto TI = BB->getTerminator();
2228 TI->setMetadata(
2229 LLVMContext::MD_prof,
2230 MDBuilder(TI->getParent()->getContext()).createBranchWeights(Weights));
2231 }
2232}
2233
2234/// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
2235/// to BB which contains an i1 PHI node and a conditional branch on that PHI.
2236/// If we can duplicate the contents of BB up into PredBB do so now, this
2237/// improves the odds that the branch will be on an analyzable instruction like
2238/// a compare.
2239bool JumpThreadingPass::DuplicateCondBranchOnPHIIntoPred(
2240 BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs) {
2241 assert(!PredBBs.empty() && "Can't handle an empty set")((!PredBBs.empty() && "Can't handle an empty set") ? static_cast
<void> (0) : __assert_fail ("!PredBBs.empty() && \"Can't handle an empty set\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 2241, __PRETTY_FUNCTION__))
;
2242
2243 // If BB is a loop header, then duplicating this block outside the loop would
2244 // cause us to transform this into an irreducible loop, don't do this.
2245 // See the comments above FindLoopHeaders for justifications and caveats.
2246 if (LoopHeaders.count(BB)) {
2247 LLVM_DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Not duplicating loop header '"
<< BB->getName() << "' into predecessor block '"
<< PredBBs[0]->getName() << "' - it might create an irreducible loop!\n"
; } } while (false)
2248 << "' into predecessor block '" << PredBBs[0]->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Not duplicating loop header '"
<< BB->getName() << "' into predecessor block '"
<< PredBBs[0]->getName() << "' - it might create an irreducible loop!\n"
; } } while (false)
2249 << "' - it might create an irreducible loop!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Not duplicating loop header '"
<< BB->getName() << "' into predecessor block '"
<< PredBBs[0]->getName() << "' - it might create an irreducible loop!\n"
; } } while (false)
;
2250 return false;
2251 }
2252
2253 unsigned DuplicationCost =
2254 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
2255 if (DuplicationCost > BBDupThreshold) {
2256 LLVM_DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Not duplicating BB '"
<< BB->getName() << "' - Cost is too high: " <<
DuplicationCost << "\n"; } } while (false)
2257 << "' - Cost is too high: " << DuplicationCost << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Not duplicating BB '"
<< BB->getName() << "' - Cost is too high: " <<
DuplicationCost << "\n"; } } while (false)
;
2258 return false;
2259 }
2260
2261 // And finally, do it! Start by factoring the predecessors if needed.
2262 std::vector<DominatorTree::UpdateType> Updates;
2263 BasicBlock *PredBB;
2264 if (PredBBs.size() == 1)
2265 PredBB = PredBBs[0];
2266 else {
2267 LLVM_DEBUG(dbgs() << " Factoring out " << PredBBs.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Factoring out " <<
PredBBs.size() << " common predecessors.\n"; } } while
(false)
2268 << " common predecessors.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Factoring out " <<
PredBBs.size() << " common predecessors.\n"; } } while
(false)
;
2269 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
2270 }
2271 Updates.push_back({DominatorTree::Delete, PredBB, BB});
2272
2273 // Okay, we decided to do this! Clone all the instructions in BB onto the end
2274 // of PredBB.
2275 LLVM_DEBUG(dbgs() << " Duplicating block '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Duplicating block '"
<< BB->getName() << "' into end of '" <<
PredBB->getName() << "' to eliminate branch on phi. Cost: "
<< DuplicationCost << " block is:" << *BB <<
"\n"; } } while (false)
2276 << "' into end of '" << PredBB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Duplicating block '"
<< BB->getName() << "' into end of '" <<
PredBB->getName() << "' to eliminate branch on phi. Cost: "
<< DuplicationCost << " block is:" << *BB <<
"\n"; } } while (false)
2277 << "' to eliminate branch on phi. Cost: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Duplicating block '"
<< BB->getName() << "' into end of '" <<
PredBB->getName() << "' to eliminate branch on phi. Cost: "
<< DuplicationCost << " block is:" << *BB <<
"\n"; } } while (false)
2278 << DuplicationCost << " block is:" << *BB << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Duplicating block '"
<< BB->getName() << "' into end of '" <<
PredBB->getName() << "' to eliminate branch on phi. Cost: "
<< DuplicationCost << " block is:" << *BB <<
"\n"; } } while (false)
;
2279
2280 // Unless PredBB ends with an unconditional branch, split the edge so that we
2281 // can just clone the bits from BB into the end of the new PredBB.
2282 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
2283
2284 if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
2285 BasicBlock *OldPredBB = PredBB;
2286 PredBB = SplitEdge(OldPredBB, BB);
2287 Updates.push_back({DominatorTree::Insert, OldPredBB, PredBB});
2288 Updates.push_back({DominatorTree::Insert, PredBB, BB});
2289 Updates.push_back({DominatorTree::Delete, OldPredBB, BB});
2290 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
2291 }
2292
2293 // We are going to have to map operands from the original BB block into the
2294 // PredBB block. Evaluate PHI nodes in BB.
2295 DenseMap<Instruction*, Value*> ValueMapping;
2296
2297 BasicBlock::iterator BI = BB->begin();
2298 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
2299 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
2300 // Clone the non-phi instructions of BB into PredBB, keeping track of the
2301 // mapping and using it to remap operands in the cloned instructions.
2302 for (; BI != BB->end(); ++BI) {
2303 Instruction *New = BI->clone();
2304
2305 // Remap operands to patch up intra-block references.
2306 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2307 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
2308 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
2309 if (I != ValueMapping.end())
2310 New->setOperand(i, I->second);
2311 }
2312
2313 // If this instruction can be simplified after the operands are updated,
2314 // just use the simplified value instead. This frequently happens due to
2315 // phi translation.
2316 if (Value *IV = SimplifyInstruction(
2317 New,
2318 {BB->getModule()->getDataLayout(), TLI, nullptr, nullptr, New})) {
2319 ValueMapping[&*BI] = IV;
2320 if (!New->mayHaveSideEffects()) {
2321 New->deleteValue();
2322 New = nullptr;
2323 }
2324 } else {
2325 ValueMapping[&*BI] = New;
2326 }
2327 if (New) {
2328 // Otherwise, insert the new instruction into the block.
2329 New->setName(BI->getName());
2330 PredBB->getInstList().insert(OldPredBranch->getIterator(), New);
2331 // Update Dominance from simplified New instruction operands.
2332 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2333 if (BasicBlock *SuccBB = dyn_cast<BasicBlock>(New->getOperand(i)))
2334 Updates.push_back({DominatorTree::Insert, PredBB, SuccBB});
2335 }
2336 }
2337
2338 // Check to see if the targets of the branch had PHI nodes. If so, we need to
2339 // add entries to the PHI nodes for branch from PredBB now.
2340 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
2341 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
2342 ValueMapping);
2343 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
2344 ValueMapping);
2345
2346 // If there were values defined in BB that are used outside the block, then we
2347 // now have to update all uses of the value to use either the original value,
2348 // the cloned value, or some PHI derived value. This can require arbitrary
2349 // PHI insertion, of which we are prepared to do, clean these up now.
2350 SSAUpdater SSAUpdate;
2351 SmallVector<Use*, 16> UsesToRename;
2352 for (Instruction &I : *BB) {
2353 // Scan all uses of this instruction to see if it is used outside of its
2354 // block, and if so, record them in UsesToRename.
2355 for (Use &U : I.uses()) {
2356 Instruction *User = cast<Instruction>(U.getUser());
2357 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
2358 if (UserPN->getIncomingBlock(U) == BB)
2359 continue;
2360 } else if (User->getParent() == BB)
2361 continue;
2362
2363 UsesToRename.push_back(&U);
2364 }
2365
2366 // If there are no uses outside the block, we're done with this instruction.
2367 if (UsesToRename.empty())
2368 continue;
2369
2370 LLVM_DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "JT: Renaming non-local uses of: "
<< I << "\n"; } } while (false)
;
2371
2372 // We found a use of I outside of BB. Rename all uses of I that are outside
2373 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
2374 // with the two values we know.
2375 SSAUpdate.Initialize(I.getType(), I.getName());
2376 SSAUpdate.AddAvailableValue(BB, &I);
2377 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[&I]);
2378
2379 while (!UsesToRename.empty())
2380 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
2381 LLVM_DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "\n"; } } while (false)
;
2382 }
2383
2384 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
2385 // that we nuked.
2386 BB->removePredecessor(PredBB, true);
2387
2388 // Remove the unconditional branch at the end of the PredBB block.
2389 OldPredBranch->eraseFromParent();
2390 DTU->applyUpdatesPermissive(Updates);
2391
2392 ++NumDupes;
2393 return true;
2394}
2395
2396// Pred is a predecessor of BB with an unconditional branch to BB. SI is
2397// a Select instruction in Pred. BB has other predecessors and SI is used in
2398// a PHI node in BB. SI has no other use.
2399// A new basic block, NewBB, is created and SI is converted to compare and
2400// conditional branch. SI is erased from parent.
2401void JumpThreadingPass::UnfoldSelectInstr(BasicBlock *Pred, BasicBlock *BB,
2402 SelectInst *SI, PHINode *SIUse,
2403 unsigned Idx) {
2404 // Expand the select.
2405 //
2406 // Pred --
2407 // | v
2408 // | NewBB
2409 // | |
2410 // |-----
2411 // v
2412 // BB
2413 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
2414 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
2415 BB->getParent(), BB);
2416 // Move the unconditional branch to NewBB.
2417 PredTerm->removeFromParent();
2418 NewBB->getInstList().insert(NewBB->end(), PredTerm);
2419 // Create a conditional branch and update PHI nodes.
2420 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
2421 SIUse->setIncomingValue(Idx, SI->getFalseValue());
2422 SIUse->addIncoming(SI->getTrueValue(), NewBB);
2423
2424 // The select is now dead.
2425 SI->eraseFromParent();
2426 DTU->applyUpdatesPermissive({{DominatorTree::Insert, NewBB, BB},
2427 {DominatorTree::Insert, Pred, NewBB}});
2428
2429 // Update any other PHI nodes in BB.
2430 for (BasicBlock::iterator BI = BB->begin();
2431 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
2432 if (Phi != SIUse)
2433 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);
2434}
2435
2436bool JumpThreadingPass::TryToUnfoldSelect(SwitchInst *SI, BasicBlock *BB) {
2437 PHINode *CondPHI = dyn_cast<PHINode>(SI->getCondition());
2438
2439 if (!CondPHI || CondPHI->getParent() != BB)
2440 return false;
2441
2442 for (unsigned I = 0, E = CondPHI->getNumIncomingValues(); I != E; ++I) {
2443 BasicBlock *Pred = CondPHI->getIncomingBlock(I);
2444 SelectInst *PredSI = dyn_cast<SelectInst>(CondPHI->getIncomingValue(I));
2445
2446 // The second and third condition can be potentially relaxed. Currently
2447 // the conditions help to simplify the code and allow us to reuse existing
2448 // code, developed for TryToUnfoldSelect(CmpInst *, BasicBlock *)
2449 if (!PredSI || PredSI->getParent() != Pred || !PredSI->hasOneUse())
2450 continue;
2451
2452 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
2453 if (!PredTerm || !PredTerm->isUnconditional())
2454 continue;
2455
2456 UnfoldSelectInstr(Pred, BB, PredSI, CondPHI, I);
2457 return true;
2458 }
2459 return false;
2460}
2461
2462/// TryToUnfoldSelect - Look for blocks of the form
2463/// bb1:
2464/// %a = select
2465/// br bb2
2466///
2467/// bb2:
2468/// %p = phi [%a, %bb1] ...
2469/// %c = icmp %p
2470/// br i1 %c
2471///
2472/// And expand the select into a branch structure if one of its arms allows %c
2473/// to be folded. This later enables threading from bb1 over bb2.
2474bool JumpThreadingPass::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
2475 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
2476 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
2477 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
2478
2479 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
2480 CondLHS->getParent() != BB)
2481 return false;
2482
2483 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
2484 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
2485 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
2486
2487 // Look if one of the incoming values is a select in the corresponding
2488 // predecessor.
2489 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
2490 continue;
2491
2492 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
2493 if (!PredTerm || !PredTerm->isUnconditional())
2494 continue;
2495
2496 // Now check if one of the select values would allow us to constant fold the
2497 // terminator in BB. We don't do the transform if both sides fold, those
2498 // cases will be threaded in any case.
2499 if (DTU->hasPendingDomTreeUpdates())
2500 LVI->disableDT();
2501 else
2502 LVI->enableDT();
2503 LazyValueInfo::Tristate LHSFolds =
2504 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
2505 CondRHS, Pred, BB, CondCmp);
2506 LazyValueInfo::Tristate RHSFolds =
2507 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
2508 CondRHS, Pred, BB, CondCmp);
2509 if ((LHSFolds != LazyValueInfo::Unknown ||
2510 RHSFolds != LazyValueInfo::Unknown) &&
2511 LHSFolds != RHSFolds) {
2512 UnfoldSelectInstr(Pred, BB, SI, CondLHS, I);
2513 return true;
2514 }
2515 }
2516 return false;
2517}
2518
2519/// TryToUnfoldSelectInCurrBB - Look for PHI/Select or PHI/CMP/Select in the
2520/// same BB in the form
2521/// bb:
2522/// %p = phi [false, %bb1], [true, %bb2], [false, %bb3], [true, %bb4], ...
2523/// %s = select %p, trueval, falseval
2524///
2525/// or
2526///
2527/// bb:
2528/// %p = phi [0, %bb1], [1, %bb2], [0, %bb3], [1, %bb4], ...
2529/// %c = cmp %p, 0
2530/// %s = select %c, trueval, falseval
2531///
2532/// And expand the select into a branch structure. This later enables
2533/// jump-threading over bb in this pass.
2534///
2535/// Using the similar approach of SimplifyCFG::FoldCondBranchOnPHI(), unfold
2536/// select if the associated PHI has at least one constant. If the unfolded
2537/// select is not jump-threaded, it will be folded again in the later
2538/// optimizations.
2539bool JumpThreadingPass::TryToUnfoldSelectInCurrBB(BasicBlock *BB) {
2540 // If threading this would thread across a loop header, don't thread the edge.
2541 // See the comments above FindLoopHeaders for justifications and caveats.
2542 if (LoopHeaders.count(BB))
2543 return false;
2544
2545 for (BasicBlock::iterator BI = BB->begin();
2546 PHINode *PN = dyn_cast<PHINode>(BI); ++BI) {
2547 // Look for a Phi having at least one constant incoming value.
2548 if (llvm::all_of(PN->incoming_values(),
2549 [](Value *V) { return !isa<ConstantInt>(V); }))
2550 continue;
2551
2552 auto isUnfoldCandidate = [BB](SelectInst *SI, Value *V) {
2553 // Check if SI is in BB and use V as condition.
2554 if (SI->getParent() != BB)
2555 return false;
2556 Value *Cond = SI->getCondition();
2557 return (Cond && Cond == V && Cond->getType()->isIntegerTy(1));
2558 };
2559
2560 SelectInst *SI = nullptr;
2561 for (Use &U : PN->uses()) {
2562 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(U.getUser())) {
2563 // Look for a ICmp in BB that compares PN with a constant and is the
2564 // condition of a Select.
2565 if (Cmp->getParent() == BB && Cmp->hasOneUse() &&
2566 isa<ConstantInt>(Cmp->getOperand(1 - U.getOperandNo())))
2567 if (SelectInst *SelectI = dyn_cast<SelectInst>(Cmp->user_back()))
2568 if (isUnfoldCandidate(SelectI, Cmp->use_begin()->get())) {
2569 SI = SelectI;
2570 break;
2571 }
2572 } else if (SelectInst *SelectI = dyn_cast<SelectInst>(U.getUser())) {
2573 // Look for a Select in BB that uses PN as condition.
2574 if (isUnfoldCandidate(SelectI, U.get())) {
2575 SI = SelectI;
2576 break;
2577 }
2578 }
2579 }
2580
2581 if (!SI)
2582 continue;
2583 // Expand the select.
2584 Instruction *Term =
2585 SplitBlockAndInsertIfThen(SI->getCondition(), SI, false);
2586 BasicBlock *SplitBB = SI->getParent();
2587 BasicBlock *NewBB = Term->getParent();
2588 PHINode *NewPN = PHINode::Create(SI->getType(), 2, "", SI);
2589 NewPN->addIncoming(SI->getTrueValue(), Term->getParent());
2590 NewPN->addIncoming(SI->getFalseValue(), BB);
2591 SI->replaceAllUsesWith(NewPN);
2592 SI->eraseFromParent();
2593 // NewBB and SplitBB are newly created blocks which require insertion.
2594 std::vector<DominatorTree::UpdateType> Updates;
2595 Updates.reserve((2 * SplitBB->getTerminator()->getNumSuccessors()) + 3);
2596 Updates.push_back({DominatorTree::Insert, BB, SplitBB});
2597 Updates.push_back({DominatorTree::Insert, BB, NewBB});
2598 Updates.push_back({DominatorTree::Insert, NewBB, SplitBB});
2599 // BB's successors were moved to SplitBB, update DTU accordingly.
2600 for (auto *Succ : successors(SplitBB)) {
2601 Updates.push_back({DominatorTree::Delete, BB, Succ});
2602 Updates.push_back({DominatorTree::Insert, SplitBB, Succ});
2603 }
2604 DTU->applyUpdatesPermissive(Updates);
2605 return true;
2606 }
2607 return false;
2608}
2609
2610/// Try to propagate a guard from the current BB into one of its predecessors
2611/// in case if another branch of execution implies that the condition of this
2612/// guard is always true. Currently we only process the simplest case that
2613/// looks like:
2614///
2615/// Start:
2616/// %cond = ...
2617/// br i1 %cond, label %T1, label %F1
2618/// T1:
2619/// br label %Merge
2620/// F1:
2621/// br label %Merge
2622/// Merge:
2623/// %condGuard = ...
2624/// call void(i1, ...) @llvm.experimental.guard( i1 %condGuard )[ "deopt"() ]
2625///
2626/// And cond either implies condGuard or !condGuard. In this case all the
2627/// instructions before the guard can be duplicated in both branches, and the
2628/// guard is then threaded to one of them.
2629bool JumpThreadingPass::ProcessGuards(BasicBlock *BB) {
2630 using namespace PatternMatch;
2631
2632 // We only want to deal with two predecessors.
2633 BasicBlock *Pred1, *Pred2;
2634 auto PI = pred_begin(BB), PE = pred_end(BB);
2635 if (PI == PE)
2636 return false;
2637 Pred1 = *PI++;
2638 if (PI == PE)
2639 return false;
2640 Pred2 = *PI++;
2641 if (PI != PE)
2642 return false;
2643 if (Pred1 == Pred2)
2644 return false;
2645
2646 // Try to thread one of the guards of the block.
2647 // TODO: Look up deeper than to immediate predecessor?
2648 auto *Parent = Pred1->getSinglePredecessor();
2649 if (!Parent || Parent != Pred2->getSinglePredecessor())
2650 return false;
2651
2652 if (auto *BI = dyn_cast<BranchInst>(Parent->getTerminator()))
2653 for (auto &I : *BB)
2654 if (isGuard(&I) && ThreadGuard(BB, cast<IntrinsicInst>(&I), BI))
2655 return true;
2656
2657 return false;
2658}
2659
2660/// Try to propagate the guard from BB which is the lower block of a diamond
2661/// to one of its branches, in case if diamond's condition implies guard's
2662/// condition.
2663bool JumpThreadingPass::ThreadGuard(BasicBlock *BB, IntrinsicInst *Guard,
2664 BranchInst *BI) {
2665 assert(BI->getNumSuccessors() == 2 && "Wrong number of successors?")((BI->getNumSuccessors() == 2 && "Wrong number of successors?"
) ? static_cast<void> (0) : __assert_fail ("BI->getNumSuccessors() == 2 && \"Wrong number of successors?\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 2665, __PRETTY_FUNCTION__))
;
2666 assert(BI->isConditional() && "Unconditional branch has 2 successors?")((BI->isConditional() && "Unconditional branch has 2 successors?"
) ? static_cast<void> (0) : __assert_fail ("BI->isConditional() && \"Unconditional branch has 2 successors?\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 2666, __PRETTY_FUNCTION__))
;
2667 Value *GuardCond = Guard->getArgOperand(0);
2668 Value *BranchCond = BI->getCondition();
2669 BasicBlock *TrueDest = BI->getSuccessor(0);
2670 BasicBlock *FalseDest = BI->getSuccessor(1);
2671
2672 auto &DL = BB->getModule()->getDataLayout();
2673 bool TrueDestIsSafe = false;
2674 bool FalseDestIsSafe = false;
2675
2676 // True dest is safe if BranchCond => GuardCond.
2677 auto Impl = isImpliedCondition(BranchCond, GuardCond, DL);
2678 if (Impl && *Impl)
2679 TrueDestIsSafe = true;
2680 else {
2681 // False dest is safe if !BranchCond => GuardCond.
2682 Impl = isImpliedCondition(BranchCond, GuardCond, DL, /* LHSIsTrue */ false);
2683 if (Impl && *Impl)
2684 FalseDestIsSafe = true;
2685 }
2686
2687 if (!TrueDestIsSafe && !FalseDestIsSafe)
2688 return false;
2689
2690 BasicBlock *PredUnguardedBlock = TrueDestIsSafe ? TrueDest : FalseDest;
2691 BasicBlock *PredGuardedBlock = FalseDestIsSafe ? TrueDest : FalseDest;
2692
2693 ValueToValueMapTy UnguardedMapping, GuardedMapping;
2694 Instruction *AfterGuard = Guard->getNextNode();
2695 unsigned Cost = getJumpThreadDuplicationCost(BB, AfterGuard, BBDupThreshold);
2696 if (Cost > BBDupThreshold)
2697 return false;
2698 // Duplicate all instructions before the guard and the guard itself to the
2699 // branch where implication is not proved.
2700 BasicBlock *GuardedBlock = DuplicateInstructionsInSplitBetween(
2701 BB, PredGuardedBlock, AfterGuard, GuardedMapping, *DTU);
2702 assert(GuardedBlock && "Could not create the guarded block?")((GuardedBlock && "Could not create the guarded block?"
) ? static_cast<void> (0) : __assert_fail ("GuardedBlock && \"Could not create the guarded block?\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 2702, __PRETTY_FUNCTION__))
;
2703 // Duplicate all instructions before the guard in the unguarded branch.
2704 // Since we have successfully duplicated the guarded block and this block
2705 // has fewer instructions, we expect it to succeed.
2706 BasicBlock *UnguardedBlock = DuplicateInstructionsInSplitBetween(
2707 BB, PredUnguardedBlock, Guard, UnguardedMapping, *DTU);
2708 assert(UnguardedBlock && "Could not create the unguarded block?")((UnguardedBlock && "Could not create the unguarded block?"
) ? static_cast<void> (0) : __assert_fail ("UnguardedBlock && \"Could not create the unguarded block?\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 2708, __PRETTY_FUNCTION__))
;
2709 LLVM_DEBUG(dbgs() << "Moved guard " << *Guard << " to block "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "Moved guard " <<
*Guard << " to block " << GuardedBlock->getName
() << "\n"; } } while (false)
2710 << GuardedBlock->getName() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "Moved guard " <<
*Guard << " to block " << GuardedBlock->getName
() << "\n"; } } while (false)
;
2711 // Some instructions before the guard may still have uses. For them, we need
2712 // to create Phi nodes merging their copies in both guarded and unguarded
2713 // branches. Those instructions that have no uses can be just removed.
2714 SmallVector<Instruction *, 4> ToRemove;
2715 for (auto BI = BB->begin(); &*BI != AfterGuard; ++BI)
2716 if (!isa<PHINode>(&*BI))
2717 ToRemove.push_back(&*BI);
2718
2719 Instruction *InsertionPoint = &*BB->getFirstInsertionPt();
2720 assert(InsertionPoint && "Empty block?")((InsertionPoint && "Empty block?") ? static_cast<
void> (0) : __assert_fail ("InsertionPoint && \"Empty block?\""
, "/build/llvm-toolchain-snapshot-9~svn358860/lib/Transforms/Scalar/JumpThreading.cpp"
, 2720, __PRETTY_FUNCTION__))
;
2721 // Substitute with Phis & remove.
2722 for (auto *Inst : reverse(ToRemove)) {
2723 if (!Inst->use_empty()) {
2724 PHINode *NewPN = PHINode::Create(Inst->getType(), 2);
2725 NewPN->addIncoming(UnguardedMapping[Inst], UnguardedBlock);
2726 NewPN->addIncoming(GuardedMapping[Inst], GuardedBlock);
2727 NewPN->insertBefore(InsertionPoint);
2728 Inst->replaceAllUsesWith(NewPN);
2729 }
2730 Inst->eraseFromParent();
2731 }
2732 return true;
2733}