Bug Summary

File:lib/Transforms/Scalar/JumpThreading.cpp
Warning:line 1415, 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-8/lib/clang/8.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-8~svn350071/build-llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-8~svn350071/build-llvm/include -I /build/llvm-toolchain-snapshot-8~svn350071/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/8.0.0/include/ -internal-isystem /usr/local/include -internal-isystem /usr/lib/llvm-8/lib/clang/8.0.0/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-8~svn350071/build-llvm/lib/Transforms/Scalar -fdebug-prefix-map=/build/llvm-toolchain-snapshot-8~svn350071=. -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-2018-12-27-042839-1215-1 -x c++ /build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp -faddrsig
1//===- JumpThreading.cpp - Thread control through conditional blocks ------===//
2//
3// The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file implements the Jump Threading pass.
11//
12//===----------------------------------------------------------------------===//
13
14#include "llvm/Transforms/Scalar/JumpThreading.h"
15#include "llvm/ADT/DenseMap.h"
16#include "llvm/ADT/DenseSet.h"
17#include "llvm/ADT/Optional.h"
18#include "llvm/ADT/STLExtras.h"
19#include "llvm/ADT/SmallPtrSet.h"
20#include "llvm/ADT/SmallVector.h"
21#include "llvm/ADT/Statistic.h"
22#include "llvm/Analysis/AliasAnalysis.h"
23#include "llvm/Analysis/BlockFrequencyInfo.h"
24#include "llvm/Analysis/BranchProbabilityInfo.h"
25#include "llvm/Analysis/CFG.h"
26#include "llvm/Analysis/ConstantFolding.h"
27#include "llvm/Analysis/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/DomTreeUpdater.h"
42#include "llvm/IR/Dominators.h"
43#include "llvm/IR/Function.h"
44#include "llvm/IR/InstrTypes.h"
45#include "llvm/IR/Instruction.h"
46#include "llvm/IR/Instructions.h"
47#include "llvm/IR/IntrinsicInst.h"
48#include "llvm/IR/Intrinsics.h"
49#include "llvm/IR/LLVMContext.h"
50#include "llvm/IR/MDBuilder.h"
51#include "llvm/IR/Metadata.h"
52#include "llvm/IR/Module.h"
53#include "llvm/IR/PassManager.h"
54#include "llvm/IR/PatternMatch.h"
55#include "llvm/IR/Type.h"
56#include "llvm/IR/Use.h"
57#include "llvm/IR/User.h"
58#include "llvm/IR/Value.h"
59#include "llvm/Pass.h"
60#include "llvm/Support/BlockFrequency.h"
61#include "llvm/Support/BranchProbability.h"
62#include "llvm/Support/Casting.h"
63#include "llvm/Support/CommandLine.h"
64#include "llvm/Support/Debug.h"
65#include "llvm/Support/raw_ostream.h"
66#include "llvm/Transforms/Scalar.h"
67#include "llvm/Transforms/Utils/BasicBlockUtils.h"
68#include "llvm/Transforms/Utils/Cloning.h"
69#include "llvm/Transforms/Utils/Local.h"
70#include "llvm/Transforms/Utils/SSAUpdater.h"
71#include "llvm/Transforms/Utils/ValueMapper.h"
72#include <algorithm>
73#include <cassert>
74#include <cstddef>
75#include <cstdint>
76#include <iterator>
77#include <memory>
78#include <utility>
79
80using namespace llvm;
81using namespace jumpthreading;
82
83#define DEBUG_TYPE"jump-threading" "jump-threading"
84
85STATISTIC(NumThreads, "Number of jumps threaded")static llvm::Statistic NumThreads = {"jump-threading", "NumThreads"
, "Number of jumps threaded", {0}, {false}}
;
86STATISTIC(NumFolds, "Number of terminators folded")static llvm::Statistic NumFolds = {"jump-threading", "NumFolds"
, "Number of terminators folded", {0}, {false}}
;
87STATISTIC(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}}
;
88
89static cl::opt<unsigned>
90BBDuplicateThreshold("jump-threading-threshold",
91 cl::desc("Max block size to duplicate for jump threading"),
92 cl::init(6), cl::Hidden);
93
94static cl::opt<unsigned>
95ImplicationSearchThreshold(
96 "jump-threading-implication-search-threshold",
97 cl::desc("The number of predecessors to search for a stronger "
98 "condition to use to thread over a weaker condition"),
99 cl::init(3), cl::Hidden);
100
101static cl::opt<bool> PrintLVIAfterJumpThreading(
102 "print-lvi-after-jump-threading",
103 cl::desc("Print the LazyValueInfo cache after JumpThreading"), cl::init(false),
104 cl::Hidden);
105
106namespace {
107
108 /// This pass performs 'jump threading', which looks at blocks that have
109 /// multiple predecessors and multiple successors. If one or more of the
110 /// predecessors of the block can be proven to always jump to one of the
111 /// successors, we forward the edge from the predecessor to the successor by
112 /// duplicating the contents of this block.
113 ///
114 /// An example of when this can occur is code like this:
115 ///
116 /// if () { ...
117 /// X = 4;
118 /// }
119 /// if (X < 3) {
120 ///
121 /// In this case, the unconditional branch at the end of the first if can be
122 /// revectored to the false side of the second if.
123 class JumpThreading : public FunctionPass {
124 JumpThreadingPass Impl;
125
126 public:
127 static char ID; // Pass identification
128
129 JumpThreading(int T = -1) : FunctionPass(ID), Impl(T) {
130 initializeJumpThreadingPass(*PassRegistry::getPassRegistry());
131 }
132
133 bool runOnFunction(Function &F) override;
134
135 void getAnalysisUsage(AnalysisUsage &AU) const override {
136 AU.addRequired<DominatorTreeWrapperPass>();
137 AU.addPreserved<DominatorTreeWrapperPass>();
138 AU.addRequired<AAResultsWrapperPass>();
139 AU.addRequired<LazyValueInfoWrapperPass>();
140 AU.addPreserved<LazyValueInfoWrapperPass>();
141 AU.addPreserved<GlobalsAAWrapperPass>();
142 AU.addRequired<TargetLibraryInfoWrapperPass>();
143 }
144
145 void releaseMemory() override { Impl.releaseMemory(); }
146 };
147
148} // end anonymous namespace
149
150char JumpThreading::ID = 0;
151
152INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",static void *initializeJumpThreadingPassOnce(PassRegistry &
Registry) {
153 "Jump Threading", false, false)static void *initializeJumpThreadingPassOnce(PassRegistry &
Registry) {
154INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry);
155INITIALIZE_PASS_DEPENDENCY(LazyValueInfoWrapperPass)initializeLazyValueInfoWrapperPassPass(Registry);
156INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry);
157INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry);
158INITIALIZE_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)); }
159 "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)); }
160
161// Public interface to the Jump Threading pass
162FunctionPass *llvm::createJumpThreadingPass(int Threshold) {
163 return new JumpThreading(Threshold);
164}
165
166JumpThreadingPass::JumpThreadingPass(int T) {
167 BBDupThreshold = (T == -1) ? BBDuplicateThreshold : unsigned(T);
168}
169
170// Update branch probability information according to conditional
171// branch probability. This is usually made possible for cloned branches
172// in inline instances by the context specific profile in the caller.
173// For instance,
174//
175// [Block PredBB]
176// [Branch PredBr]
177// if (t) {
178// Block A;
179// } else {
180// Block B;
181// }
182//
183// [Block BB]
184// cond = PN([true, %A], [..., %B]); // PHI node
185// [Branch CondBr]
186// if (cond) {
187// ... // P(cond == true) = 1%
188// }
189//
190// Here we know that when block A is taken, cond must be true, which means
191// P(cond == true | A) = 1
192//
193// Given that P(cond == true) = P(cond == true | A) * P(A) +
194// P(cond == true | B) * P(B)
195// we get:
196// P(cond == true ) = P(A) + P(cond == true | B) * P(B)
197//
198// which gives us:
199// P(A) is less than P(cond == true), i.e.
200// P(t == true) <= P(cond == true)
201//
202// In other words, if we know P(cond == true) is unlikely, we know
203// that P(t == true) is also unlikely.
204//
205static void updatePredecessorProfileMetadata(PHINode *PN, BasicBlock *BB) {
206 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
207 if (!CondBr)
208 return;
209
210 BranchProbability BP;
211 uint64_t TrueWeight, FalseWeight;
212 if (!CondBr->extractProfMetadata(TrueWeight, FalseWeight))
213 return;
214
215 // Returns the outgoing edge of the dominating predecessor block
216 // that leads to the PhiNode's incoming block:
217 auto GetPredOutEdge =
218 [](BasicBlock *IncomingBB,
219 BasicBlock *PhiBB) -> std::pair<BasicBlock *, BasicBlock *> {
220 auto *PredBB = IncomingBB;
221 auto *SuccBB = PhiBB;
222 while (true) {
223 BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator());
224 if (PredBr && PredBr->isConditional())
225 return {PredBB, SuccBB};
226 auto *SinglePredBB = PredBB->getSinglePredecessor();
227 if (!SinglePredBB)
228 return {nullptr, nullptr};
229 SuccBB = PredBB;
230 PredBB = SinglePredBB;
231 }
232 };
233
234 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
235 Value *PhiOpnd = PN->getIncomingValue(i);
236 ConstantInt *CI = dyn_cast<ConstantInt>(PhiOpnd);
237
238 if (!CI || !CI->getType()->isIntegerTy(1))
239 continue;
240
241 BP = (CI->isOne() ? BranchProbability::getBranchProbability(
242 TrueWeight, TrueWeight + FalseWeight)
243 : BranchProbability::getBranchProbability(
244 FalseWeight, TrueWeight + FalseWeight));
245
246 auto PredOutEdge = GetPredOutEdge(PN->getIncomingBlock(i), BB);
247 if (!PredOutEdge.first)
248 return;
249
250 BasicBlock *PredBB = PredOutEdge.first;
251 BranchInst *PredBr = cast<BranchInst>(PredBB->getTerminator());
252
253 uint64_t PredTrueWeight, PredFalseWeight;
254 // FIXME: We currently only set the profile data when it is missing.
255 // With PGO, this can be used to refine even existing profile data with
256 // context information. This needs to be done after more performance
257 // testing.
258 if (PredBr->extractProfMetadata(PredTrueWeight, PredFalseWeight))
259 continue;
260
261 // We can not infer anything useful when BP >= 50%, because BP is the
262 // upper bound probability value.
263 if (BP >= BranchProbability(50, 100))
264 continue;
265
266 SmallVector<uint32_t, 2> Weights;
267 if (PredBr->getSuccessor(0) == PredOutEdge.second) {
268 Weights.push_back(BP.getNumerator());
269 Weights.push_back(BP.getCompl().getNumerator());
270 } else {
271 Weights.push_back(BP.getCompl().getNumerator());
272 Weights.push_back(BP.getNumerator());
273 }
274 PredBr->setMetadata(LLVMContext::MD_prof,
275 MDBuilder(PredBr->getParent()->getContext())
276 .createBranchWeights(Weights));
277 }
278}
279
280/// runOnFunction - Toplevel algorithm.
281bool JumpThreading::runOnFunction(Function &F) {
282 if (skipFunction(F))
283 return false;
284 auto TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
285 // Get DT analysis before LVI. When LVI is initialized it conditionally adds
286 // DT if it's available.
287 auto DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
288 auto LVI = &getAnalysis<LazyValueInfoWrapperPass>().getLVI();
289 auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
290 DomTreeUpdater DTU(*DT, DomTreeUpdater::UpdateStrategy::Lazy);
291 std::unique_ptr<BlockFrequencyInfo> BFI;
292 std::unique_ptr<BranchProbabilityInfo> BPI;
293 bool HasProfileData = F.hasProfileData();
294 if (HasProfileData) {
295 LoopInfo LI{DominatorTree(F)};
296 BPI.reset(new BranchProbabilityInfo(F, LI, TLI));
297 BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
298 }
299
300 bool Changed = Impl.runImpl(F, TLI, LVI, AA, &DTU, HasProfileData,
301 std::move(BFI), std::move(BPI));
302 if (PrintLVIAfterJumpThreading) {
303 dbgs() << "LVI for function '" << F.getName() << "':\n";
304 LVI->printLVI(F, *DT, dbgs());
305 }
306 return Changed;
307}
308
309PreservedAnalyses JumpThreadingPass::run(Function &F,
310 FunctionAnalysisManager &AM) {
311 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
312 // Get DT analysis before LVI. When LVI is initialized it conditionally adds
313 // DT if it's available.
314 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
315 auto &LVI = AM.getResult<LazyValueAnalysis>(F);
316 auto &AA = AM.getResult<AAManager>(F);
317 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Lazy);
318
319 std::unique_ptr<BlockFrequencyInfo> BFI;
320 std::unique_ptr<BranchProbabilityInfo> BPI;
321 if (F.hasProfileData()) {
322 LoopInfo LI{DominatorTree(F)};
323 BPI.reset(new BranchProbabilityInfo(F, LI, &TLI));
324 BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
325 }
326
327 bool Changed = runImpl(F, &TLI, &LVI, &AA, &DTU, HasProfileData,
328 std::move(BFI), std::move(BPI));
329
330 if (!Changed)
331 return PreservedAnalyses::all();
332 PreservedAnalyses PA;
333 PA.preserve<GlobalsAA>();
334 PA.preserve<DominatorTreeAnalysis>();
335 PA.preserve<LazyValueAnalysis>();
336 return PA;
337}
338
339bool JumpThreadingPass::runImpl(Function &F, TargetLibraryInfo *TLI_,
340 LazyValueInfo *LVI_, AliasAnalysis *AA_,
341 DomTreeUpdater *DTU_, bool HasProfileData_,
342 std::unique_ptr<BlockFrequencyInfo> BFI_,
343 std::unique_ptr<BranchProbabilityInfo> BPI_) {
344 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)
;
345 TLI = TLI_;
346 LVI = LVI_;
347 AA = AA_;
348 DTU = DTU_;
349 BFI.reset();
350 BPI.reset();
351 // When profile data is available, we need to update edge weights after
352 // successful jump threading, which requires both BPI and BFI being available.
353 HasProfileData = HasProfileData_;
354 auto *GuardDecl = F.getParent()->getFunction(
355 Intrinsic::getName(Intrinsic::experimental_guard));
356 HasGuards = GuardDecl && !GuardDecl->use_empty();
357 if (HasProfileData) {
358 BPI = std::move(BPI_);
359 BFI = std::move(BFI_);
360 }
361
362 // JumpThreading must not processes blocks unreachable from entry. It's a
363 // waste of compute time and can potentially lead to hangs.
364 SmallPtrSet<BasicBlock *, 16> Unreachable;
365 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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 365, __PRETTY_FUNCTION__))
;
366 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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 366, __PRETTY_FUNCTION__))
;
367 DominatorTree &DT = DTU->getDomTree();
368 for (auto &BB : F)
369 if (!DT.isReachableFromEntry(&BB))
370 Unreachable.insert(&BB);
371
372 FindLoopHeaders(F);
373
374 bool EverChanged = false;
375 bool Changed;
376 do {
377 Changed = false;
378 for (auto &BB : F) {
379 if (Unreachable.count(&BB))
380 continue;
381 while (ProcessBlock(&BB)) // Thread all of the branches we can over BB.
382 Changed = true;
383 // Stop processing BB if it's the entry or is now deleted. The following
384 // routines attempt to eliminate BB and locating a suitable replacement
385 // for the entry is non-trivial.
386 if (&BB == &F.getEntryBlock() || DTU->isBBPendingDeletion(&BB))
387 continue;
388
389 if (pred_empty(&BB)) {
390 // When ProcessBlock makes BB unreachable it doesn't bother to fix up
391 // the instructions in it. We must remove BB to prevent invalid IR.
392 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)
393 << "' 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)
394 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " JT: Deleting dead block '"
<< BB.getName() << "' with terminator: " <<
*BB.getTerminator() << '\n'; } } while (false)
;
395 LoopHeaders.erase(&BB);
396 LVI->eraseBlock(&BB);
397 DeleteDeadBlock(&BB, DTU);
398 Changed = true;
399 continue;
400 }
401
402 // ProcessBlock doesn't thread BBs with unconditional TIs. However, if BB
403 // is "almost empty", we attempt to merge BB with its sole successor.
404 auto *BI = dyn_cast<BranchInst>(BB.getTerminator());
405 if (BI && BI->isUnconditional() &&
406 // The terminator must be the only non-phi instruction in BB.
407 BB.getFirstNonPHIOrDbg()->isTerminator() &&
408 // Don't alter Loop headers and latches to ensure another pass can
409 // detect and transform nested loops later.
410 !LoopHeaders.count(&BB) && !LoopHeaders.count(BI->getSuccessor(0)) &&
411 TryToSimplifyUncondBranchFromEmptyBlock(&BB, DTU)) {
412 // BB is valid for cleanup here because we passed in DTU. F remains
413 // BB's parent until a DTU->getDomTree() event.
414 LVI->eraseBlock(&BB);
415 Changed = true;
416 }
417 }
418 EverChanged |= Changed;
419 } while (Changed);
420
421 LoopHeaders.clear();
422 // Flush only the Dominator Tree.
423 DTU->getDomTree();
424 LVI->enableDT();
425 return EverChanged;
426}
427
428// Replace uses of Cond with ToVal when safe to do so. If all uses are
429// replaced, we can remove Cond. We cannot blindly replace all uses of Cond
430// because we may incorrectly replace uses when guards/assumes are uses of
431// of `Cond` and we used the guards/assume to reason about the `Cond` value
432// at the end of block. RAUW unconditionally replaces all uses
433// including the guards/assumes themselves and the uses before the
434// guard/assume.
435static void ReplaceFoldableUses(Instruction *Cond, Value *ToVal) {
436 assert(Cond->getType() == ToVal->getType())((Cond->getType() == ToVal->getType()) ? static_cast<
void> (0) : __assert_fail ("Cond->getType() == ToVal->getType()"
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 436, __PRETTY_FUNCTION__))
;
437 auto *BB = Cond->getParent();
438 // We can unconditionally replace all uses in non-local blocks (i.e. uses
439 // strictly dominated by BB), since LVI information is true from the
440 // terminator of BB.
441 replaceNonLocalUsesWith(Cond, ToVal);
442 for (Instruction &I : reverse(*BB)) {
443 // Reached the Cond whose uses we are trying to replace, so there are no
444 // more uses.
445 if (&I == Cond)
446 break;
447 // We only replace uses in instructions that are guaranteed to reach the end
448 // of BB, where we know Cond is ToVal.
449 if (!isGuaranteedToTransferExecutionToSuccessor(&I))
450 break;
451 I.replaceUsesOfWith(Cond, ToVal);
452 }
453 if (Cond->use_empty() && !Cond->mayHaveSideEffects())
454 Cond->eraseFromParent();
455}
456
457/// Return the cost of duplicating a piece of this block from first non-phi
458/// and before StopAt instruction to thread across it. Stop scanning the block
459/// when exceeding the threshold. If duplication is impossible, returns ~0U.
460static unsigned getJumpThreadDuplicationCost(BasicBlock *BB,
461 Instruction *StopAt,
462 unsigned Threshold) {
463 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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 463, __PRETTY_FUNCTION__))
;
464 /// Ignore PHI nodes, these will be flattened when duplication happens.
465 BasicBlock::const_iterator I(BB->getFirstNonPHI());
466
467 // FIXME: THREADING will delete values that are just used to compute the
468 // branch, so they shouldn't count against the duplication cost.
469
470 unsigned Bonus = 0;
471 if (BB->getTerminator() == StopAt) {
472 // Threading through a switch statement is particularly profitable. If this
473 // block ends in a switch, decrease its cost to make it more likely to
474 // happen.
475 if (isa<SwitchInst>(StopAt))
476 Bonus = 6;
477
478 // The same holds for indirect branches, but slightly more so.
479 if (isa<IndirectBrInst>(StopAt))
480 Bonus = 8;
481 }
482
483 // Bump the threshold up so the early exit from the loop doesn't skip the
484 // terminator-based Size adjustment at the end.
485 Threshold += Bonus;
486
487 // Sum up the cost of each instruction until we get to the terminator. Don't
488 // include the terminator because the copy won't include it.
489 unsigned Size = 0;
490 for (; &*I != StopAt; ++I) {
491
492 // Stop scanning the block if we've reached the threshold.
493 if (Size > Threshold)
494 return Size;
495
496 // Debugger intrinsics don't incur code size.
497 if (isa<DbgInfoIntrinsic>(I)) continue;
498
499 // If this is a pointer->pointer bitcast, it is free.
500 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
501 continue;
502
503 // Bail out if this instruction gives back a token type, it is not possible
504 // to duplicate it if it is used outside this BB.
505 if (I->getType()->isTokenTy() && I->isUsedOutsideOfBlock(BB))
506 return ~0U;
507
508 // All other instructions count for at least one unit.
509 ++Size;
510
511 // Calls are more expensive. If they are non-intrinsic calls, we model them
512 // as having cost of 4. If they are a non-vector intrinsic, we model them
513 // as having cost of 2 total, and if they are a vector intrinsic, we model
514 // them as having cost 1.
515 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
516 if (CI->cannotDuplicate() || CI->isConvergent())
517 // Blocks with NoDuplicate are modelled as having infinite cost, so they
518 // are never duplicated.
519 return ~0U;
520 else if (!isa<IntrinsicInst>(CI))
521 Size += 3;
522 else if (!CI->getType()->isVectorTy())
523 Size += 1;
524 }
525 }
526
527 return Size > Bonus ? Size - Bonus : 0;
528}
529
530/// FindLoopHeaders - We do not want jump threading to turn proper loop
531/// structures into irreducible loops. Doing this breaks up the loop nesting
532/// hierarchy and pessimizes later transformations. To prevent this from
533/// happening, we first have to find the loop headers. Here we approximate this
534/// by finding targets of backedges in the CFG.
535///
536/// Note that there definitely are cases when we want to allow threading of
537/// edges across a loop header. For example, threading a jump from outside the
538/// loop (the preheader) to an exit block of the loop is definitely profitable.
539/// It is also almost always profitable to thread backedges from within the loop
540/// to exit blocks, and is often profitable to thread backedges to other blocks
541/// within the loop (forming a nested loop). This simple analysis is not rich
542/// enough to track all of these properties and keep it up-to-date as the CFG
543/// mutates, so we don't allow any of these transformations.
544void JumpThreadingPass::FindLoopHeaders(Function &F) {
545 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
546 FindFunctionBackedges(F, Edges);
547
548 for (const auto &Edge : Edges)
549 LoopHeaders.insert(Edge.second);
550}
551
552/// getKnownConstant - Helper method to determine if we can thread over a
553/// terminator with the given value as its condition, and if so what value to
554/// use for that. What kind of value this is depends on whether we want an
555/// integer or a block address, but an undef is always accepted.
556/// Returns null if Val is null or not an appropriate constant.
557static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
558 if (!Val)
559 return nullptr;
560
561 // Undef is "known" enough.
562 if (UndefValue *U = dyn_cast<UndefValue>(Val))
563 return U;
564
565 if (Preference == WantBlockAddress)
566 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
567
568 return dyn_cast<ConstantInt>(Val);
569}
570
571/// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
572/// if we can infer that the value is a known ConstantInt/BlockAddress or undef
573/// in any of our predecessors. If so, return the known list of value and pred
574/// BB in the result vector.
575///
576/// This returns true if there were any known values.
577bool JumpThreadingPass::ComputeValueKnownInPredecessorsImpl(
578 Value *V, BasicBlock *BB, PredValueInfo &Result,
579 ConstantPreference Preference,
580 DenseSet<std::pair<Value *, BasicBlock *>> &RecursionSet,
581 Instruction *CxtI) {
582 // This method walks up use-def chains recursively. Because of this, we could
583 // get into an infinite loop going around loops in the use-def chain. To
584 // prevent this, keep track of what (value, block) pairs we've already visited
585 // and terminate the search if we loop back to them
586 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
587 return false;
588
589 // If V is a constant, then it is known in all predecessors.
590 if (Constant *KC = getKnownConstant(V, Preference)) {
591 for (BasicBlock *Pred : predecessors(BB))
592 Result.push_back(std::make_pair(KC, Pred));
593
594 return !Result.empty();
595 }
596
597 // If V is a non-instruction value, or an instruction in a different block,
598 // then it can't be derived from a PHI.
599 Instruction *I = dyn_cast<Instruction>(V);
600 if (!I || I->getParent() != BB) {
601
602 // Okay, if this is a live-in value, see if it has a known value at the end
603 // of any of our predecessors.
604 //
605 // FIXME: This should be an edge property, not a block end property.
606 /// TODO: Per PR2563, we could infer value range information about a
607 /// predecessor based on its terminator.
608 //
609 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
610 // "I" is a non-local compare-with-a-constant instruction. This would be
611 // able to handle value inequalities better, for example if the compare is
612 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
613 // Perhaps getConstantOnEdge should be smart enough to do this?
614
615 if (DTU->hasPendingDomTreeUpdates())
616 LVI->disableDT();
617 else
618 LVI->enableDT();
619 for (BasicBlock *P : predecessors(BB)) {
620 // If the value is known by LazyValueInfo to be a constant in a
621 // predecessor, use that information to try to thread this block.
622 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB, CxtI);
623 if (Constant *KC = getKnownConstant(PredCst, Preference))
624 Result.push_back(std::make_pair(KC, P));
625 }
626
627 return !Result.empty();
628 }
629
630 /// If I is a PHI node, then we know the incoming values for any constants.
631 if (PHINode *PN = dyn_cast<PHINode>(I)) {
632 if (DTU->hasPendingDomTreeUpdates())
633 LVI->disableDT();
634 else
635 LVI->enableDT();
636 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
637 Value *InVal = PN->getIncomingValue(i);
638 if (Constant *KC = getKnownConstant(InVal, Preference)) {
639 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
640 } else {
641 Constant *CI = LVI->getConstantOnEdge(InVal,
642 PN->getIncomingBlock(i),
643 BB, CxtI);
644 if (Constant *KC = getKnownConstant(CI, Preference))
645 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
646 }
647 }
648
649 return !Result.empty();
650 }
651
652 // Handle Cast instructions. Only see through Cast when the source operand is
653 // PHI or Cmp to save the compilation time.
654 if (CastInst *CI = dyn_cast<CastInst>(I)) {
655 Value *Source = CI->getOperand(0);
656 if (!isa<PHINode>(Source) && !isa<CmpInst>(Source))
657 return false;
658 ComputeValueKnownInPredecessorsImpl(Source, BB, Result, Preference,
659 RecursionSet, CxtI);
660 if (Result.empty())
661 return false;
662
663 // Convert the known values.
664 for (auto &R : Result)
665 R.first = ConstantExpr::getCast(CI->getOpcode(), R.first, CI->getType());
666
667 return true;
668 }
669
670 // Handle some boolean conditions.
671 if (I->getType()->getPrimitiveSizeInBits() == 1) {
672 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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 672, __PRETTY_FUNCTION__))
;
673 // X | true -> true
674 // X & false -> false
675 if (I->getOpcode() == Instruction::Or ||
676 I->getOpcode() == Instruction::And) {
677 PredValueInfoTy LHSVals, RHSVals;
678
679 ComputeValueKnownInPredecessorsImpl(I->getOperand(0), BB, LHSVals,
680 WantInteger, RecursionSet, CxtI);
681 ComputeValueKnownInPredecessorsImpl(I->getOperand(1), BB, RHSVals,
682 WantInteger, RecursionSet, CxtI);
683
684 if (LHSVals.empty() && RHSVals.empty())
685 return false;
686
687 ConstantInt *InterestingVal;
688 if (I->getOpcode() == Instruction::Or)
689 InterestingVal = ConstantInt::getTrue(I->getContext());
690 else
691 InterestingVal = ConstantInt::getFalse(I->getContext());
692
693 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
694
695 // Scan for the sentinel. If we find an undef, force it to the
696 // interesting value: x|undef -> true and x&undef -> false.
697 for (const auto &LHSVal : LHSVals)
698 if (LHSVal.first == InterestingVal || isa<UndefValue>(LHSVal.first)) {
699 Result.emplace_back(InterestingVal, LHSVal.second);
700 LHSKnownBBs.insert(LHSVal.second);
701 }
702 for (const auto &RHSVal : RHSVals)
703 if (RHSVal.first == InterestingVal || isa<UndefValue>(RHSVal.first)) {
704 // If we already inferred a value for this block on the LHS, don't
705 // re-add it.
706 if (!LHSKnownBBs.count(RHSVal.second))
707 Result.emplace_back(InterestingVal, RHSVal.second);
708 }
709
710 return !Result.empty();
711 }
712
713 // Handle the NOT form of XOR.
714 if (I->getOpcode() == Instruction::Xor &&
715 isa<ConstantInt>(I->getOperand(1)) &&
716 cast<ConstantInt>(I->getOperand(1))->isOne()) {
717 ComputeValueKnownInPredecessorsImpl(I->getOperand(0), BB, Result,
718 WantInteger, RecursionSet, CxtI);
719 if (Result.empty())
720 return false;
721
722 // Invert the known values.
723 for (auto &R : Result)
724 R.first = ConstantExpr::getNot(R.first);
725
726 return true;
727 }
728
729 // Try to simplify some other binary operator values.
730 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
731 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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 732, __PRETTY_FUNCTION__))
732 && "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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 732, __PRETTY_FUNCTION__))
;
733 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
734 PredValueInfoTy LHSVals;
735 ComputeValueKnownInPredecessorsImpl(BO->getOperand(0), BB, LHSVals,
736 WantInteger, RecursionSet, CxtI);
737
738 // Try to use constant folding to simplify the binary operator.
739 for (const auto &LHSVal : LHSVals) {
740 Constant *V = LHSVal.first;
741 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
742
743 if (Constant *KC = getKnownConstant(Folded, WantInteger))
744 Result.push_back(std::make_pair(KC, LHSVal.second));
745 }
746 }
747
748 return !Result.empty();
749 }
750
751 // Handle compare with phi operand, where the PHI is defined in this block.
752 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
753 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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 753, __PRETTY_FUNCTION__))
;
754 Type *CmpType = Cmp->getType();
755 Value *CmpLHS = Cmp->getOperand(0);
756 Value *CmpRHS = Cmp->getOperand(1);
757 CmpInst::Predicate Pred = Cmp->getPredicate();
758
759 PHINode *PN = dyn_cast<PHINode>(CmpLHS);
760 if (!PN)
761 PN = dyn_cast<PHINode>(CmpRHS);
762 if (PN && PN->getParent() == BB) {
763 const DataLayout &DL = PN->getModule()->getDataLayout();
764 // We can do this simplification if any comparisons fold to true or false.
765 // See if any do.
766 if (DTU->hasPendingDomTreeUpdates())
767 LVI->disableDT();
768 else
769 LVI->enableDT();
770 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
771 BasicBlock *PredBB = PN->getIncomingBlock(i);
772 Value *LHS, *RHS;
773 if (PN == CmpLHS) {
774 LHS = PN->getIncomingValue(i);
775 RHS = CmpRHS->DoPHITranslation(BB, PredBB);
776 } else {
777 LHS = CmpLHS->DoPHITranslation(BB, PredBB);
778 RHS = PN->getIncomingValue(i);
779 }
780 Value *Res = SimplifyCmpInst(Pred, LHS, RHS, {DL});
781 if (!Res) {
782 if (!isa<Constant>(RHS))
783 continue;
784
785 // getPredicateOnEdge call will make no sense if LHS is defined in BB.
786 auto LHSInst = dyn_cast<Instruction>(LHS);
787 if (LHSInst && LHSInst->getParent() == BB)
788 continue;
789
790 LazyValueInfo::Tristate
791 ResT = LVI->getPredicateOnEdge(Pred, LHS,
792 cast<Constant>(RHS), PredBB, BB,
793 CxtI ? CxtI : Cmp);
794 if (ResT == LazyValueInfo::Unknown)
795 continue;
796 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
797 }
798
799 if (Constant *KC = getKnownConstant(Res, WantInteger))
800 Result.push_back(std::make_pair(KC, PredBB));
801 }
802
803 return !Result.empty();
804 }
805
806 // If comparing a live-in value against a constant, see if we know the
807 // live-in value on any predecessors.
808 if (isa<Constant>(CmpRHS) && !CmpType->isVectorTy()) {
809 Constant *CmpConst = cast<Constant>(CmpRHS);
810
811 if (!isa<Instruction>(CmpLHS) ||
812 cast<Instruction>(CmpLHS)->getParent() != BB) {
813 if (DTU->hasPendingDomTreeUpdates())
814 LVI->disableDT();
815 else
816 LVI->enableDT();
817 for (BasicBlock *P : predecessors(BB)) {
818 // If the value is known by LazyValueInfo to be a constant in a
819 // predecessor, use that information to try to thread this block.
820 LazyValueInfo::Tristate Res =
821 LVI->getPredicateOnEdge(Pred, CmpLHS,
822 CmpConst, P, BB, CxtI ? CxtI : Cmp);
823 if (Res == LazyValueInfo::Unknown)
824 continue;
825
826 Constant *ResC = ConstantInt::get(CmpType, Res);
827 Result.push_back(std::make_pair(ResC, P));
828 }
829
830 return !Result.empty();
831 }
832
833 // InstCombine can fold some forms of constant range checks into
834 // (icmp (add (x, C1)), C2). See if we have we have such a thing with
835 // x as a live-in.
836 {
837 using namespace PatternMatch;
838
839 Value *AddLHS;
840 ConstantInt *AddConst;
841 if (isa<ConstantInt>(CmpConst) &&
842 match(CmpLHS, m_Add(m_Value(AddLHS), m_ConstantInt(AddConst)))) {
843 if (!isa<Instruction>(AddLHS) ||
844 cast<Instruction>(AddLHS)->getParent() != BB) {
845 if (DTU->hasPendingDomTreeUpdates())
846 LVI->disableDT();
847 else
848 LVI->enableDT();
849 for (BasicBlock *P : predecessors(BB)) {
850 // If the value is known by LazyValueInfo to be a ConstantRange in
851 // a predecessor, use that information to try to thread this
852 // block.
853 ConstantRange CR = LVI->getConstantRangeOnEdge(
854 AddLHS, P, BB, CxtI ? CxtI : cast<Instruction>(CmpLHS));
855 // Propagate the range through the addition.
856 CR = CR.add(AddConst->getValue());
857
858 // Get the range where the compare returns true.
859 ConstantRange CmpRange = ConstantRange::makeExactICmpRegion(
860 Pred, cast<ConstantInt>(CmpConst)->getValue());
861
862 Constant *ResC;
863 if (CmpRange.contains(CR))
864 ResC = ConstantInt::getTrue(CmpType);
865 else if (CmpRange.inverse().contains(CR))
866 ResC = ConstantInt::getFalse(CmpType);
867 else
868 continue;
869
870 Result.push_back(std::make_pair(ResC, P));
871 }
872
873 return !Result.empty();
874 }
875 }
876 }
877
878 // Try to find a constant value for the LHS of a comparison,
879 // and evaluate it statically if we can.
880 PredValueInfoTy LHSVals;
881 ComputeValueKnownInPredecessorsImpl(I->getOperand(0), BB, LHSVals,
882 WantInteger, RecursionSet, CxtI);
883
884 for (const auto &LHSVal : LHSVals) {
885 Constant *V = LHSVal.first;
886 Constant *Folded = ConstantExpr::getCompare(Pred, V, CmpConst);
887 if (Constant *KC = getKnownConstant(Folded, WantInteger))
888 Result.push_back(std::make_pair(KC, LHSVal.second));
889 }
890
891 return !Result.empty();
892 }
893 }
894
895 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
896 // Handle select instructions where at least one operand is a known constant
897 // and we can figure out the condition value for any predecessor block.
898 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
899 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
900 PredValueInfoTy Conds;
901 if ((TrueVal || FalseVal) &&
902 ComputeValueKnownInPredecessorsImpl(SI->getCondition(), BB, Conds,
903 WantInteger, RecursionSet, CxtI)) {
904 for (auto &C : Conds) {
905 Constant *Cond = C.first;
906
907 // Figure out what value to use for the condition.
908 bool KnownCond;
909 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
910 // A known boolean.
911 KnownCond = CI->isOne();
912 } else {
913 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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 913, __PRETTY_FUNCTION__))
;
914 // Either operand will do, so be sure to pick the one that's a known
915 // constant.
916 // FIXME: Do this more cleverly if both values are known constants?
917 KnownCond = (TrueVal != nullptr);
918 }
919
920 // See if the select has a known constant value for this predecessor.
921 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
922 Result.push_back(std::make_pair(Val, C.second));
923 }
924
925 return !Result.empty();
926 }
927 }
928
929 // If all else fails, see if LVI can figure out a constant value for us.
930 if (DTU->hasPendingDomTreeUpdates())
931 LVI->disableDT();
932 else
933 LVI->enableDT();
934 Constant *CI = LVI->getConstant(V, BB, CxtI);
935 if (Constant *KC = getKnownConstant(CI, Preference)) {
936 for (BasicBlock *Pred : predecessors(BB))
937 Result.push_back(std::make_pair(KC, Pred));
938 }
939
940 return !Result.empty();
941}
942
943/// GetBestDestForBranchOnUndef - If we determine that the specified block ends
944/// in an undefined jump, decide which block is best to revector to.
945///
946/// Since we can pick an arbitrary destination, we pick the successor with the
947/// fewest predecessors. This should reduce the in-degree of the others.
948static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
949 Instruction *BBTerm = BB->getTerminator();
950 unsigned MinSucc = 0;
951 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
952 // Compute the successor with the minimum number of predecessors.
953 unsigned MinNumPreds = pred_size(TestBB);
954 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
955 TestBB = BBTerm->getSuccessor(i);
956 unsigned NumPreds = pred_size(TestBB);
957 if (NumPreds < MinNumPreds) {
958 MinSucc = i;
959 MinNumPreds = NumPreds;
960 }
961 }
962
963 return MinSucc;
964}
965
966static bool hasAddressTakenAndUsed(BasicBlock *BB) {
967 if (!BB->hasAddressTaken()) return false;
968
969 // If the block has its address taken, it may be a tree of dead constants
970 // hanging off of it. These shouldn't keep the block alive.
971 BlockAddress *BA = BlockAddress::get(BB);
972 BA->removeDeadConstantUsers();
973 return !BA->use_empty();
974}
975
976/// ProcessBlock - If there are any predecessors whose control can be threaded
977/// through to a successor, transform them now.
978bool JumpThreadingPass::ProcessBlock(BasicBlock *BB) {
979 // If the block is trivially dead, just return and let the caller nuke it.
980 // This simplifies other transformations.
981 if (DTU->isBBPendingDeletion(BB) ||
1
Assuming the condition is false
982 (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()))
2
Assuming the condition is false
983 return false;
984
985 // If this block has a single predecessor, and if that pred has a single
986 // successor, merge the blocks. This encourages recursive jump threading
987 // because now the condition in this block can be threaded through
988 // predecessors of our predecessor block.
989 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
3
Assuming 'SinglePred' is null
4
Taking false branch
990 const Instruction *TI = SinglePred->getTerminator();
991 if (!TI->isExceptionalTerminator() && TI->getNumSuccessors() == 1 &&
992 SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
993 // If SinglePred was a loop header, BB becomes one.
994 if (LoopHeaders.erase(SinglePred))
995 LoopHeaders.insert(BB);
996
997 LVI->eraseBlock(SinglePred);
998 MergeBasicBlockIntoOnlyPred(BB, DTU);
999
1000 // Now that BB is merged into SinglePred (i.e. SinglePred Code followed by
1001 // BB code within one basic block `BB`), we need to invalidate the LVI
1002 // information associated with BB, because the LVI information need not be
1003 // true for all of BB after the merge. For example,
1004 // Before the merge, LVI info and code is as follows:
1005 // SinglePred: <LVI info1 for %p val>
1006 // %y = use of %p
1007 // call @exit() // need not transfer execution to successor.
1008 // assume(%p) // from this point on %p is true
1009 // br label %BB
1010 // BB: <LVI info2 for %p val, i.e. %p is true>
1011 // %x = use of %p
1012 // br label exit
1013 //
1014 // Note that this LVI info for blocks BB and SinglPred is correct for %p
1015 // (info2 and info1 respectively). After the merge and the deletion of the
1016 // LVI info1 for SinglePred. We have the following code:
1017 // BB: <LVI info2 for %p val>
1018 // %y = use of %p
1019 // call @exit()
1020 // assume(%p)
1021 // %x = use of %p <-- LVI info2 is correct from here onwards.
1022 // br label exit
1023 // LVI info2 for BB is incorrect at the beginning of BB.
1024
1025 // Invalidate LVI information for BB if the LVI is not provably true for
1026 // all of BB.
1027 if (!isGuaranteedToTransferExecutionToSuccessor(BB))
1028 LVI->eraseBlock(BB);
1029 return true;
1030 }
1031 }
1032
1033 if (TryToUnfoldSelectInCurrBB(BB))
5
Taking false branch
1034 return true;
1035
1036 // Look if we can propagate guards to predecessors.
1037 if (HasGuards && ProcessGuards(BB))
6
Assuming the condition is false
1038 return true;
1039
1040 // What kind of constant we're looking for.
1041 ConstantPreference Preference = WantInteger;
1042
1043 // Look to see if the terminator is a conditional branch, switch or indirect
1044 // branch, if not we can't thread it.
1045 Value *Condition;
1046 Instruction *Terminator = BB->getTerminator();
1047 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
7
Taking true branch
1048 // Can't thread an unconditional jump.
1049 if (BI->isUnconditional()) return false;
8
Taking false branch
1050 Condition = BI->getCondition();
1051 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
1052 Condition = SI->getCondition();
1053 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
1054 // Can't thread indirect branch with no successors.
1055 if (IB->getNumSuccessors() == 0) return false;
1056 Condition = IB->getAddress()->stripPointerCasts();
1057 Preference = WantBlockAddress;
1058 } else {
1059 return false; // Must be an invoke.
1060 }
1061
1062 // Run constant folding to see if we can reduce the condition to a simple
1063 // constant.
1064 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
9
Taking true branch
1065 Value *SimpleVal =
1066 ConstantFoldInstruction(I, BB->getModule()->getDataLayout(), TLI);
1067 if (SimpleVal) {
10
Assuming 'SimpleVal' is null
11
Taking false branch
1068 I->replaceAllUsesWith(SimpleVal);
1069 if (isInstructionTriviallyDead(I, TLI))
1070 I->eraseFromParent();
1071 Condition = SimpleVal;
1072 }
1073 }
1074
1075 // If the terminator is branching on an undef, we can pick any of the
1076 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
1077 if (isa<UndefValue>(Condition)) {
12
Taking false branch
1078 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
1079 std::vector<DominatorTree::UpdateType> Updates;
1080
1081 // Fold the branch/switch.
1082 Instruction *BBTerm = BB->getTerminator();
1083 Updates.reserve(BBTerm->getNumSuccessors());
1084 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
1085 if (i == BestSucc) continue;
1086 BasicBlock *Succ = BBTerm->getSuccessor(i);
1087 Succ->removePredecessor(BB, true);
1088 Updates.push_back({DominatorTree::Delete, BB, Succ});
1089 }
1090
1091 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)
1092 << "' 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)
;
1093 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
1094 BBTerm->eraseFromParent();
1095 DTU->applyUpdates(Updates);
1096 return true;
1097 }
1098
1099 // If the terminator of this block is branching on a constant, simplify the
1100 // terminator to an unconditional branch. This can occur due to threading in
1101 // other blocks.
1102 if (getKnownConstant(Condition, Preference)) {
13
Taking false branch
1103 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)
1104 << "' folding terminator: " << *BB->getTerminator()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " In block '" <<
BB->getName() << "' folding terminator: " << *
BB->getTerminator() << '\n'; } } while (false)
1105 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " In block '" <<
BB->getName() << "' folding terminator: " << *
BB->getTerminator() << '\n'; } } while (false)
;
1106 ++NumFolds;
1107 ConstantFoldTerminator(BB, true, nullptr, DTU);
1108 return true;
1109 }
1110
1111 Instruction *CondInst = dyn_cast<Instruction>(Condition);
1112
1113 // All the rest of our checks depend on the condition being an instruction.
1114 if (!CondInst) {
14
Taking false branch
1115 // FIXME: Unify this with code below.
1116 if (ProcessThreadableEdges(Condition, BB, Preference, Terminator))
1117 return true;
1118 return false;
1119 }
1120
1121 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
15
Taking false branch
1122 // If we're branching on a conditional, LVI might be able to determine
1123 // it's value at the branch instruction. We only handle comparisons
1124 // against a constant at this time.
1125 // TODO: This should be extended to handle switches as well.
1126 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1127 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
1128 if (CondBr && CondConst) {
1129 // We should have returned as soon as we turn a conditional branch to
1130 // unconditional. Because its no longer interesting as far as jump
1131 // threading is concerned.
1132 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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 1132, __PRETTY_FUNCTION__))
;
1133
1134 if (DTU->hasPendingDomTreeUpdates())
1135 LVI->disableDT();
1136 else
1137 LVI->enableDT();
1138 LazyValueInfo::Tristate Ret =
1139 LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0),
1140 CondConst, CondBr);
1141 if (Ret != LazyValueInfo::Unknown) {
1142 unsigned ToRemove = Ret == LazyValueInfo::True ? 1 : 0;
1143 unsigned ToKeep = Ret == LazyValueInfo::True ? 0 : 1;
1144 BasicBlock *ToRemoveSucc = CondBr->getSuccessor(ToRemove);
1145 ToRemoveSucc->removePredecessor(BB, true);
1146 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
1147 CondBr->eraseFromParent();
1148 if (CondCmp->use_empty())
1149 CondCmp->eraseFromParent();
1150 // We can safely replace *some* uses of the CondInst if it has
1151 // exactly one value as returned by LVI. RAUW is incorrect in the
1152 // presence of guards and assumes, that have the `Cond` as the use. This
1153 // is because we use the guards/assume to reason about the `Cond` value
1154 // at the end of block, but RAUW unconditionally replaces all uses
1155 // including the guards/assumes themselves and the uses before the
1156 // guard/assume.
1157 else if (CondCmp->getParent() == BB) {
1158 auto *CI = Ret == LazyValueInfo::True ?
1159 ConstantInt::getTrue(CondCmp->getType()) :
1160 ConstantInt::getFalse(CondCmp->getType());
1161 ReplaceFoldableUses(CondCmp, CI);
1162 }
1163 DTU->deleteEdgeRelaxed(BB, ToRemoveSucc);
1164 return true;
1165 }
1166
1167 // We did not manage to simplify this branch, try to see whether
1168 // CondCmp depends on a known phi-select pattern.
1169 if (TryToUnfoldSelect(CondCmp, BB))
1170 return true;
1171 }
1172 }
1173
1174 // Check for some cases that are worth simplifying. Right now we want to look
1175 // for loads that are used by a switch or by the condition for the branch. If
1176 // we see one, check to see if it's partially redundant. If so, insert a PHI
1177 // which can then be used to thread the values.
1178 Value *SimplifyValue = CondInst;
1179 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
16
Taking false branch
1180 if (isa<Constant>(CondCmp->getOperand(1)))
1181 SimplifyValue = CondCmp->getOperand(0);
1182
1183 // TODO: There are other places where load PRE would be profitable, such as
1184 // more complex comparisons.
1185 if (LoadInst *LoadI = dyn_cast<LoadInst>(SimplifyValue))
17
Taking true branch
1186 if (SimplifyPartiallyRedundantLoad(LoadI))
18
Calling 'JumpThreadingPass::SimplifyPartiallyRedundantLoad'
1187 return true;
1188
1189 // Before threading, try to propagate profile data backwards:
1190 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
1191 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1192 updatePredecessorProfileMetadata(PN, BB);
1193
1194 // Handle a variety of cases where we are branching on something derived from
1195 // a PHI node in the current block. If we can prove that any predecessors
1196 // compute a predictable value based on a PHI node, thread those predecessors.
1197 if (ProcessThreadableEdges(CondInst, BB, Preference, Terminator))
1198 return true;
1199
1200 // If this is an otherwise-unfoldable branch on a phi node in the current
1201 // block, see if we can simplify.
1202 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
1203 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1204 return ProcessBranchOnPHI(PN);
1205
1206 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
1207 if (CondInst->getOpcode() == Instruction::Xor &&
1208 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1209 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
1210
1211 // Search for a stronger dominating condition that can be used to simplify a
1212 // conditional branch leaving BB.
1213 if (ProcessImpliedCondition(BB))
1214 return true;
1215
1216 return false;
1217}
1218
1219bool JumpThreadingPass::ProcessImpliedCondition(BasicBlock *BB) {
1220 auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
1221 if (!BI || !BI->isConditional())
1222 return false;
1223
1224 Value *Cond = BI->getCondition();
1225 BasicBlock *CurrentBB = BB;
1226 BasicBlock *CurrentPred = BB->getSinglePredecessor();
1227 unsigned Iter = 0;
1228
1229 auto &DL = BB->getModule()->getDataLayout();
1230
1231 while (CurrentPred && Iter++ < ImplicationSearchThreshold) {
1232 auto *PBI = dyn_cast<BranchInst>(CurrentPred->getTerminator());
1233 if (!PBI || !PBI->isConditional())
1234 return false;
1235 if (PBI->getSuccessor(0) != CurrentBB && PBI->getSuccessor(1) != CurrentBB)
1236 return false;
1237
1238 bool CondIsTrue = PBI->getSuccessor(0) == CurrentBB;
1239 Optional<bool> Implication =
1240 isImpliedCondition(PBI->getCondition(), Cond, DL, CondIsTrue);
1241 if (Implication) {
1242 BasicBlock *KeepSucc = BI->getSuccessor(*Implication ? 0 : 1);
1243 BasicBlock *RemoveSucc = BI->getSuccessor(*Implication ? 1 : 0);
1244 RemoveSucc->removePredecessor(BB);
1245 BranchInst::Create(KeepSucc, BI);
1246 BI->eraseFromParent();
1247 DTU->deleteEdgeRelaxed(BB, RemoveSucc);
1248 return true;
1249 }
1250 CurrentBB = CurrentPred;
1251 CurrentPred = CurrentBB->getSinglePredecessor();
1252 }
1253
1254 return false;
1255}
1256
1257/// Return true if Op is an instruction defined in the given block.
1258static bool isOpDefinedInBlock(Value *Op, BasicBlock *BB) {
1259 if (Instruction *OpInst = dyn_cast<Instruction>(Op))
1260 if (OpInst->getParent() == BB)
1261 return true;
1262 return false;
1263}
1264
1265/// SimplifyPartiallyRedundantLoad - If LoadI is an obviously partially
1266/// redundant load instruction, eliminate it by replacing it with a PHI node.
1267/// This is an important optimization that encourages jump threading, and needs
1268/// to be run interlaced with other jump threading tasks.
1269bool JumpThreadingPass::SimplifyPartiallyRedundantLoad(LoadInst *LoadI) {
1270 // Don't hack volatile and ordered loads.
1271 if (!LoadI->isUnordered()) return false;
19
Taking false branch
1272
1273 // If the load is defined in a block with exactly one predecessor, it can't be
1274 // partially redundant.
1275 BasicBlock *LoadBB = LoadI->getParent();
1276 if (LoadBB->getSinglePredecessor())
20
Assuming the condition is false
21
Taking false branch
1277 return false;
1278
1279 // If the load is defined in an EH pad, it can't be partially redundant,
1280 // because the edges between the invoke and the EH pad cannot have other
1281 // instructions between them.
1282 if (LoadBB->isEHPad())
22
Assuming the condition is false
23
Taking false branch
1283 return false;
1284
1285 Value *LoadedPtr = LoadI->getOperand(0);
1286
1287 // If the loaded operand is defined in the LoadBB and its not a phi,
1288 // it can't be available in predecessors.
1289 if (isOpDefinedInBlock(LoadedPtr, LoadBB) && !isa<PHINode>(LoadedPtr))
1290 return false;
1291
1292 // Scan a few instructions up from the load, to see if it is obviously live at
1293 // the entry to its block.
1294 BasicBlock::iterator BBIt(LoadI);
1295 bool IsLoadCSE;
1296 if (Value *AvailableVal = FindAvailableLoadedValue(
24
Assuming 'AvailableVal' is null
25
Taking false branch
1297 LoadI, LoadBB, BBIt, DefMaxInstsToScan, AA, &IsLoadCSE)) {
1298 // If the value of the load is locally available within the block, just use
1299 // it. This frequently occurs for reg2mem'd allocas.
1300
1301 if (IsLoadCSE) {
1302 LoadInst *NLoadI = cast<LoadInst>(AvailableVal);
1303 combineMetadataForCSE(NLoadI, LoadI, false);
1304 };
1305
1306 // If the returned value is the load itself, replace with an undef. This can
1307 // only happen in dead loops.
1308 if (AvailableVal == LoadI)
1309 AvailableVal = UndefValue::get(LoadI->getType());
1310 if (AvailableVal->getType() != LoadI->getType())
1311 AvailableVal = CastInst::CreateBitOrPointerCast(
1312 AvailableVal, LoadI->getType(), "", LoadI);
1313 LoadI->replaceAllUsesWith(AvailableVal);
1314 LoadI->eraseFromParent();
1315 return true;
1316 }
1317
1318 // Otherwise, if we scanned the whole block and got to the top of the block,
1319 // we know the block is locally transparent to the load. If not, something
1320 // might clobber its value.
1321 if (BBIt != LoadBB->begin())
26
Taking false branch
1322 return false;
1323
1324 // If all of the loads and stores that feed the value have the same AA tags,
1325 // then we can propagate them onto any newly inserted loads.
1326 AAMDNodes AATags;
1327 LoadI->getAAMetadata(AATags);
1328
1329 SmallPtrSet<BasicBlock*, 8> PredsScanned;
1330
1331 using AvailablePredsTy = SmallVector<std::pair<BasicBlock *, Value *>, 8>;
1332
1333 AvailablePredsTy AvailablePreds;
1334 BasicBlock *OneUnavailablePred = nullptr;
27
'OneUnavailablePred' initialized to a null pointer value
1335 SmallVector<LoadInst*, 8> CSELoads;
1336
1337 // If we got here, the loaded value is transparent through to the start of the
1338 // block. Check to see if it is available in any of the predecessor blocks.
1339 for (BasicBlock *PredBB : predecessors(LoadBB)) {
1340 // If we already scanned this predecessor, skip it.
1341 if (!PredsScanned.insert(PredBB).second)
1342 continue;
1343
1344 BBIt = PredBB->end();
1345 unsigned NumScanedInst = 0;
1346 Value *PredAvailable = nullptr;
1347 // NOTE: We don't CSE load that is volatile or anything stronger than
1348 // unordered, that should have been checked when we entered the function.
1349 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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 1350, __PRETTY_FUNCTION__))
1350 "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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 1350, __PRETTY_FUNCTION__))
;
1351 // If this is a load on a phi pointer, phi-translate it and search
1352 // for available load/store to the pointer in predecessors.
1353 Value *Ptr = LoadedPtr->DoPHITranslation(LoadBB, PredBB);
1354 PredAvailable = FindAvailablePtrLoadStore(
1355 Ptr, LoadI->getType(), LoadI->isAtomic(), PredBB, BBIt,
1356 DefMaxInstsToScan, AA, &IsLoadCSE, &NumScanedInst);
1357
1358 // If PredBB has a single predecessor, continue scanning through the
1359 // single predecessor.
1360 BasicBlock *SinglePredBB = PredBB;
1361 while (!PredAvailable && SinglePredBB && BBIt == SinglePredBB->begin() &&
1362 NumScanedInst < DefMaxInstsToScan) {
1363 SinglePredBB = SinglePredBB->getSinglePredecessor();
1364 if (SinglePredBB) {
1365 BBIt = SinglePredBB->end();
1366 PredAvailable = FindAvailablePtrLoadStore(
1367 Ptr, LoadI->getType(), LoadI->isAtomic(), SinglePredBB, BBIt,
1368 (DefMaxInstsToScan - NumScanedInst), AA, &IsLoadCSE,
1369 &NumScanedInst);
1370 }
1371 }
1372
1373 if (!PredAvailable) {
1374 OneUnavailablePred = PredBB;
1375 continue;
1376 }
1377
1378 if (IsLoadCSE)
1379 CSELoads.push_back(cast<LoadInst>(PredAvailable));
1380
1381 // If so, this load is partially redundant. Remember this info so that we
1382 // can create a PHI node.
1383 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
1384 }
1385
1386 // If the loaded value isn't available in any predecessor, it isn't partially
1387 // redundant.
1388 if (AvailablePreds.empty()) return false;
28
Taking false branch
1389
1390 // Okay, the loaded value is available in at least one (and maybe all!)
1391 // predecessors. If the value is unavailable in more than one unique
1392 // predecessor, we want to insert a merge block for those common predecessors.
1393 // This ensures that we only have to insert one reload, thus not increasing
1394 // code size.
1395 BasicBlock *UnavailablePred = nullptr;
1396
1397 // If the value is unavailable in one of predecessors, we will end up
1398 // inserting a new instruction into them. It is only valid if all the
1399 // instructions before LoadI are guaranteed to pass execution to its
1400 // successor, or if LoadI is safe to speculate.
1401 // TODO: If this logic becomes more complex, and we will perform PRE insertion
1402 // farther than to a predecessor, we need to reuse the code from GVN's PRE.
1403 // It requires domination tree analysis, so for this simple case it is an
1404 // overkill.
1405 if (PredsScanned.size() != AvailablePreds.size() &&
29
Assuming the condition is false
1406 !isSafeToSpeculativelyExecute(LoadI))
1407 for (auto I = LoadBB->begin(); &*I != LoadI; ++I)
1408 if (!isGuaranteedToTransferExecutionToSuccessor(&*I))
1409 return false;
1410
1411 // If there is exactly one predecessor where the value is unavailable, the
1412 // already computed 'OneUnavailablePred' block is it. If it ends in an
1413 // unconditional branch, we know that it isn't a critical edge.
1414 if (PredsScanned.size() == AvailablePreds.size()+1 &&
30
Assuming the condition is true
1415 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
31
Called C++ object pointer is null
1416 UnavailablePred = OneUnavailablePred;
1417 } else if (PredsScanned.size() != AvailablePreds.size()) {
1418 // Otherwise, we had multiple unavailable predecessors or we had a critical
1419 // edge from the one.
1420 SmallVector<BasicBlock*, 8> PredsToSplit;
1421 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
1422
1423 for (const auto &AvailablePred : AvailablePreds)
1424 AvailablePredSet.insert(AvailablePred.first);
1425
1426 // Add all the unavailable predecessors to the PredsToSplit list.
1427 for (BasicBlock *P : predecessors(LoadBB)) {
1428 // If the predecessor is an indirect goto, we can't split the edge.
1429 if (isa<IndirectBrInst>(P->getTerminator()))
1430 return false;
1431
1432 if (!AvailablePredSet.count(P))
1433 PredsToSplit.push_back(P);
1434 }
1435
1436 // Split them out to their own block.
1437 UnavailablePred = SplitBlockPreds(LoadBB, PredsToSplit, "thread-pre-split");
1438 }
1439
1440 // If the value isn't available in all predecessors, then there will be
1441 // exactly one where it isn't available. Insert a load on that edge and add
1442 // it to the AvailablePreds list.
1443 if (UnavailablePred) {
1444 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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 1445, __PRETTY_FUNCTION__))
1445 "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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 1445, __PRETTY_FUNCTION__))
;
1446 LoadInst *NewVal =
1447 new LoadInst(LoadedPtr->DoPHITranslation(LoadBB, UnavailablePred),
1448 LoadI->getName() + ".pr", false, LoadI->getAlignment(),
1449 LoadI->getOrdering(), LoadI->getSyncScopeID(),
1450 UnavailablePred->getTerminator());
1451 NewVal->setDebugLoc(LoadI->getDebugLoc());
1452 if (AATags)
1453 NewVal->setAAMetadata(AATags);
1454
1455 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
1456 }
1457
1458 // Now we know that each predecessor of this block has a value in
1459 // AvailablePreds, sort them for efficient access as we're walking the preds.
1460 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1461
1462 // Create a PHI node at the start of the block for the PRE'd load value.
1463 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
1464 PHINode *PN = PHINode::Create(LoadI->getType(), std::distance(PB, PE), "",
1465 &LoadBB->front());
1466 PN->takeName(LoadI);
1467 PN->setDebugLoc(LoadI->getDebugLoc());
1468
1469 // Insert new entries into the PHI for each predecessor. A single block may
1470 // have multiple entries here.
1471 for (pred_iterator PI = PB; PI != PE; ++PI) {
1472 BasicBlock *P = *PI;
1473 AvailablePredsTy::iterator I =
1474 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1475 std::make_pair(P, (Value*)nullptr));
1476
1477 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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 1478, __PRETTY_FUNCTION__))
1478 "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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 1478, __PRETTY_FUNCTION__))
;
1479
1480 // If we have an available predecessor but it requires casting, insert the
1481 // cast in the predecessor and use the cast. Note that we have to update the
1482 // AvailablePreds vector as we go so that all of the PHI entries for this
1483 // predecessor use the same bitcast.
1484 Value *&PredV = I->second;
1485 if (PredV->getType() != LoadI->getType())
1486 PredV = CastInst::CreateBitOrPointerCast(PredV, LoadI->getType(), "",
1487 P->getTerminator());
1488
1489 PN->addIncoming(PredV, I->first);
1490 }
1491
1492 for (LoadInst *PredLoadI : CSELoads) {
1493 combineMetadataForCSE(PredLoadI, LoadI, true);
1494 }
1495
1496 LoadI->replaceAllUsesWith(PN);
1497 LoadI->eraseFromParent();
1498
1499 return true;
1500}
1501
1502/// FindMostPopularDest - The specified list contains multiple possible
1503/// threadable destinations. Pick the one that occurs the most frequently in
1504/// the list.
1505static BasicBlock *
1506FindMostPopularDest(BasicBlock *BB,
1507 const SmallVectorImpl<std::pair<BasicBlock *,
1508 BasicBlock *>> &PredToDestList) {
1509 assert(!PredToDestList.empty())((!PredToDestList.empty()) ? static_cast<void> (0) : __assert_fail
("!PredToDestList.empty()", "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 1509, __PRETTY_FUNCTION__))
;
1510
1511 // Determine popularity. If there are multiple possible destinations, we
1512 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1513 // blocks with known and real destinations to threading undef. We'll handle
1514 // them later if interesting.
1515 DenseMap<BasicBlock*, unsigned> DestPopularity;
1516 for (const auto &PredToDest : PredToDestList)
1517 if (PredToDest.second)
1518 DestPopularity[PredToDest.second]++;
1519
1520 if (DestPopularity.empty())
1521 return nullptr;
1522
1523 // Find the most popular dest.
1524 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1525 BasicBlock *MostPopularDest = DPI->first;
1526 unsigned Popularity = DPI->second;
1527 SmallVector<BasicBlock*, 4> SamePopularity;
1528
1529 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1530 // If the popularity of this entry isn't higher than the popularity we've
1531 // seen so far, ignore it.
1532 if (DPI->second < Popularity)
1533 ; // ignore.
1534 else if (DPI->second == Popularity) {
1535 // If it is the same as what we've seen so far, keep track of it.
1536 SamePopularity.push_back(DPI->first);
1537 } else {
1538 // If it is more popular, remember it.
1539 SamePopularity.clear();
1540 MostPopularDest = DPI->first;
1541 Popularity = DPI->second;
1542 }
1543 }
1544
1545 // Okay, now we know the most popular destination. If there is more than one
1546 // destination, we need to determine one. This is arbitrary, but we need
1547 // to make a deterministic decision. Pick the first one that appears in the
1548 // successor list.
1549 if (!SamePopularity.empty()) {
1550 SamePopularity.push_back(MostPopularDest);
1551 Instruction *TI = BB->getTerminator();
1552 for (unsigned i = 0; ; ++i) {
1553 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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 1553, __PRETTY_FUNCTION__))
;
1554
1555 if (!is_contained(SamePopularity, TI->getSuccessor(i)))
1556 continue;
1557
1558 MostPopularDest = TI->getSuccessor(i);
1559 break;
1560 }
1561 }
1562
1563 // Okay, we have finally picked the most popular destination.
1564 return MostPopularDest;
1565}
1566
1567bool JumpThreadingPass::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1568 ConstantPreference Preference,
1569 Instruction *CxtI) {
1570 // If threading this would thread across a loop header, don't even try to
1571 // thread the edge.
1572 if (LoopHeaders.count(BB))
1573 return false;
1574
1575 PredValueInfoTy PredValues;
1576 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference, CxtI))
1577 return false;
1578
1579 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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 1580, __PRETTY_FUNCTION__))
1580 "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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 1580, __PRETTY_FUNCTION__))
;
1581
1582 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)
1583 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)
1584 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)
1585 << "': 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)
1586 << " 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)
1587 })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)
;
1588
1589 // Decide what we want to thread through. Convert our list of known values to
1590 // a list of known destinations for each pred. This also discards duplicate
1591 // predecessors and keeps track of the undefined inputs (which are represented
1592 // as a null dest in the PredToDestList).
1593 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1594 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1595
1596 BasicBlock *OnlyDest = nullptr;
1597 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1598 Constant *OnlyVal = nullptr;
1599 Constant *MultipleVal = (Constant *)(intptr_t)~0ULL;
1600
1601 unsigned PredWithKnownDest = 0;
1602 for (const auto &PredValue : PredValues) {
1603 BasicBlock *Pred = PredValue.second;
1604 if (!SeenPreds.insert(Pred).second)
1605 continue; // Duplicate predecessor entry.
1606
1607 Constant *Val = PredValue.first;
1608
1609 BasicBlock *DestBB;
1610 if (isa<UndefValue>(Val))
1611 DestBB = nullptr;
1612 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
1613 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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 1613, __PRETTY_FUNCTION__))
;
1614 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1615 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1616 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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 1616, __PRETTY_FUNCTION__))
;
1617 DestBB = SI->findCaseValue(cast<ConstantInt>(Val))->getCaseSuccessor();
1618 } else {
1619 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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 1620, __PRETTY_FUNCTION__))
1620 && "Unexpected terminator")((isa<IndirectBrInst>(BB->getTerminator()) &&
"Unexpected terminator") ? static_cast<void> (0) : __assert_fail
("isa<IndirectBrInst>(BB->getTerminator()) && \"Unexpected terminator\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 1620, __PRETTY_FUNCTION__))
;
1621 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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 1621, __PRETTY_FUNCTION__))
;
1622 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1623 }
1624
1625 // If we have exactly one destination, remember it for efficiency below.
1626 if (PredToDestList.empty()) {
1627 OnlyDest = DestBB;
1628 OnlyVal = Val;
1629 } else {
1630 if (OnlyDest != DestBB)
1631 OnlyDest = MultipleDestSentinel;
1632 // It possible we have same destination, but different value, e.g. default
1633 // case in switchinst.
1634 if (Val != OnlyVal)
1635 OnlyVal = MultipleVal;
1636 }
1637
1638 // We know where this predecessor is going.
1639 ++PredWithKnownDest;
1640
1641 // If the predecessor ends with an indirect goto, we can't change its
1642 // destination.
1643 if (isa<IndirectBrInst>(Pred->getTerminator()))
1644 continue;
1645
1646 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1647 }
1648
1649 // If all edges were unthreadable, we fail.
1650 if (PredToDestList.empty())
1651 return false;
1652
1653 // If all the predecessors go to a single known successor, we want to fold,
1654 // not thread. By doing so, we do not need to duplicate the current block and
1655 // also miss potential opportunities in case we dont/cant duplicate.
1656 if (OnlyDest && OnlyDest != MultipleDestSentinel) {
1657 if (PredWithKnownDest == (size_t)pred_size(BB)) {
1658 bool SeenFirstBranchToOnlyDest = false;
1659 std::vector <DominatorTree::UpdateType> Updates;
1660 Updates.reserve(BB->getTerminator()->getNumSuccessors() - 1);
1661 for (BasicBlock *SuccBB : successors(BB)) {
1662 if (SuccBB == OnlyDest && !SeenFirstBranchToOnlyDest) {
1663 SeenFirstBranchToOnlyDest = true; // Don't modify the first branch.
1664 } else {
1665 SuccBB->removePredecessor(BB, true); // This is unreachable successor.
1666 Updates.push_back({DominatorTree::Delete, BB, SuccBB});
1667 }
1668 }
1669
1670 // Finally update the terminator.
1671 Instruction *Term = BB->getTerminator();
1672 BranchInst::Create(OnlyDest, Term);
1673 Term->eraseFromParent();
1674 DTU->applyUpdates(Updates);
1675
1676 // If the condition is now dead due to the removal of the old terminator,
1677 // erase it.
1678 if (auto *CondInst = dyn_cast<Instruction>(Cond)) {
1679 if (CondInst->use_empty() && !CondInst->mayHaveSideEffects())
1680 CondInst->eraseFromParent();
1681 // We can safely replace *some* uses of the CondInst if it has
1682 // exactly one value as returned by LVI. RAUW is incorrect in the
1683 // presence of guards and assumes, that have the `Cond` as the use. This
1684 // is because we use the guards/assume to reason about the `Cond` value
1685 // at the end of block, but RAUW unconditionally replaces all uses
1686 // including the guards/assumes themselves and the uses before the
1687 // guard/assume.
1688 else if (OnlyVal && OnlyVal != MultipleVal &&
1689 CondInst->getParent() == BB)
1690 ReplaceFoldableUses(CondInst, OnlyVal);
1691 }
1692 return true;
1693 }
1694 }
1695
1696 // Determine which is the most common successor. If we have many inputs and
1697 // this block is a switch, we want to start by threading the batch that goes
1698 // to the most popular destination first. If we only know about one
1699 // threadable destination (the common case) we can avoid this.
1700 BasicBlock *MostPopularDest = OnlyDest;
1701
1702 if (MostPopularDest == MultipleDestSentinel) {
1703 // Remove any loop headers from the Dest list, ThreadEdge conservatively
1704 // won't process them, but we might have other destination that are eligible
1705 // and we still want to process.
1706 erase_if(PredToDestList,
1707 [&](const std::pair<BasicBlock *, BasicBlock *> &PredToDest) {
1708 return LoopHeaders.count(PredToDest.second) != 0;
1709 });
1710
1711 if (PredToDestList.empty())
1712 return false;
1713
1714 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1715 }
1716
1717 // Now that we know what the most popular destination is, factor all
1718 // predecessors that will jump to it into a single predecessor.
1719 SmallVector<BasicBlock*, 16> PredsToFactor;
1720 for (const auto &PredToDest : PredToDestList)
1721 if (PredToDest.second == MostPopularDest) {
1722 BasicBlock *Pred = PredToDest.first;
1723
1724 // This predecessor may be a switch or something else that has multiple
1725 // edges to the block. Factor each of these edges by listing them
1726 // according to # occurrences in PredsToFactor.
1727 for (BasicBlock *Succ : successors(Pred))
1728 if (Succ == BB)
1729 PredsToFactor.push_back(Pred);
1730 }
1731
1732 // If the threadable edges are branching on an undefined value, we get to pick
1733 // the destination that these predecessors should get to.
1734 if (!MostPopularDest)
1735 MostPopularDest = BB->getTerminator()->
1736 getSuccessor(GetBestDestForJumpOnUndef(BB));
1737
1738 // Ok, try to thread it!
1739 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1740}
1741
1742/// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1743/// a PHI node in the current block. See if there are any simplifications we
1744/// can do based on inputs to the phi node.
1745bool JumpThreadingPass::ProcessBranchOnPHI(PHINode *PN) {
1746 BasicBlock *BB = PN->getParent();
1747
1748 // TODO: We could make use of this to do it once for blocks with common PHI
1749 // values.
1750 SmallVector<BasicBlock*, 1> PredBBs;
1751 PredBBs.resize(1);
1752
1753 // If any of the predecessor blocks end in an unconditional branch, we can
1754 // *duplicate* the conditional branch into that block in order to further
1755 // encourage jump threading and to eliminate cases where we have branch on a
1756 // phi of an icmp (branch on icmp is much better).
1757 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1758 BasicBlock *PredBB = PN->getIncomingBlock(i);
1759 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1760 if (PredBr->isUnconditional()) {
1761 PredBBs[0] = PredBB;
1762 // Try to duplicate BB into PredBB.
1763 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1764 return true;
1765 }
1766 }
1767
1768 return false;
1769}
1770
1771/// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1772/// a xor instruction in the current block. See if there are any
1773/// simplifications we can do based on inputs to the xor.
1774bool JumpThreadingPass::ProcessBranchOnXOR(BinaryOperator *BO) {
1775 BasicBlock *BB = BO->getParent();
1776
1777 // If either the LHS or RHS of the xor is a constant, don't do this
1778 // optimization.
1779 if (isa<ConstantInt>(BO->getOperand(0)) ||
1780 isa<ConstantInt>(BO->getOperand(1)))
1781 return false;
1782
1783 // If the first instruction in BB isn't a phi, we won't be able to infer
1784 // anything special about any particular predecessor.
1785 if (!isa<PHINode>(BB->front()))
1786 return false;
1787
1788 // If this BB is a landing pad, we won't be able to split the edge into it.
1789 if (BB->isEHPad())
1790 return false;
1791
1792 // If we have a xor as the branch input to this block, and we know that the
1793 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1794 // the condition into the predecessor and fix that value to true, saving some
1795 // logical ops on that path and encouraging other paths to simplify.
1796 //
1797 // This copies something like this:
1798 //
1799 // BB:
1800 // %X = phi i1 [1], [%X']
1801 // %Y = icmp eq i32 %A, %B
1802 // %Z = xor i1 %X, %Y
1803 // br i1 %Z, ...
1804 //
1805 // Into:
1806 // BB':
1807 // %Y = icmp ne i32 %A, %B
1808 // br i1 %Y, ...
1809
1810 PredValueInfoTy XorOpValues;
1811 bool isLHS = true;
1812 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1813 WantInteger, BO)) {
1814 assert(XorOpValues.empty())((XorOpValues.empty()) ? static_cast<void> (0) : __assert_fail
("XorOpValues.empty()", "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 1814, __PRETTY_FUNCTION__))
;
1815 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1816 WantInteger, BO))
1817 return false;
1818 isLHS = false;
1819 }
1820
1821 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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 1822, __PRETTY_FUNCTION__))
1822 "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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 1822, __PRETTY_FUNCTION__))
;
1823
1824 // Scan the information to see which is most popular: true or false. The
1825 // predecessors can be of the set true, false, or undef.
1826 unsigned NumTrue = 0, NumFalse = 0;
1827 for (const auto &XorOpValue : XorOpValues) {
1828 if (isa<UndefValue>(XorOpValue.first))
1829 // Ignore undefs for the count.
1830 continue;
1831 if (cast<ConstantInt>(XorOpValue.first)->isZero())
1832 ++NumFalse;
1833 else
1834 ++NumTrue;
1835 }
1836
1837 // Determine which value to split on, true, false, or undef if neither.
1838 ConstantInt *SplitVal = nullptr;
1839 if (NumTrue > NumFalse)
1840 SplitVal = ConstantInt::getTrue(BB->getContext());
1841 else if (NumTrue != 0 || NumFalse != 0)
1842 SplitVal = ConstantInt::getFalse(BB->getContext());
1843
1844 // Collect all of the blocks that this can be folded into so that we can
1845 // factor this once and clone it once.
1846 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1847 for (const auto &XorOpValue : XorOpValues) {
1848 if (XorOpValue.first != SplitVal && !isa<UndefValue>(XorOpValue.first))
1849 continue;
1850
1851 BlocksToFoldInto.push_back(XorOpValue.second);
1852 }
1853
1854 // If we inferred a value for all of the predecessors, then duplication won't
1855 // help us. However, we can just replace the LHS or RHS with the constant.
1856 if (BlocksToFoldInto.size() ==
1857 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1858 if (!SplitVal) {
1859 // If all preds provide undef, just nuke the xor, because it is undef too.
1860 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1861 BO->eraseFromParent();
1862 } else if (SplitVal->isZero()) {
1863 // If all preds provide 0, replace the xor with the other input.
1864 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1865 BO->eraseFromParent();
1866 } else {
1867 // If all preds provide 1, set the computed value to 1.
1868 BO->setOperand(!isLHS, SplitVal);
1869 }
1870
1871 return true;
1872 }
1873
1874 // Try to duplicate BB into PredBB.
1875 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1876}
1877
1878/// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1879/// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1880/// NewPred using the entries from OldPred (suitably mapped).
1881static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1882 BasicBlock *OldPred,
1883 BasicBlock *NewPred,
1884 DenseMap<Instruction*, Value*> &ValueMap) {
1885 for (PHINode &PN : PHIBB->phis()) {
1886 // Ok, we have a PHI node. Figure out what the incoming value was for the
1887 // DestBlock.
1888 Value *IV = PN.getIncomingValueForBlock(OldPred);
1889
1890 // Remap the value if necessary.
1891 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1892 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1893 if (I != ValueMap.end())
1894 IV = I->second;
1895 }
1896
1897 PN.addIncoming(IV, NewPred);
1898 }
1899}
1900
1901/// ThreadEdge - We have decided that it is safe and profitable to factor the
1902/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1903/// across BB. Transform the IR to reflect this change.
1904bool JumpThreadingPass::ThreadEdge(BasicBlock *BB,
1905 const SmallVectorImpl<BasicBlock *> &PredBBs,
1906 BasicBlock *SuccBB) {
1907 // If threading to the same block as we come from, we would infinite loop.
1908 if (SuccBB == BB) {
1909 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)
1910 << "' - 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)
;
1911 return false;
1912 }
1913
1914 // If threading this would thread across a loop header, don't thread the edge.
1915 // See the comments above FindLoopHeaders for justifications and caveats.
1916 if (LoopHeaders.count(BB) || LoopHeaders.count(SuccBB)) {
1917 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)
1918 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)
1919 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)
1920 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)
1921 << (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)
1922 << "' 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)
1923 << 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)
1924 })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 return false;
1926 }
1927
1928 unsigned JumpThreadCost =
1929 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
1930 if (JumpThreadCost > BBDupThreshold) {
1931 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)
1932 << "' - 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)
;
1933 return false;
1934 }
1935
1936 // And finally, do it! Start by factoring the predecessors if needed.
1937 BasicBlock *PredBB;
1938 if (PredBBs.size() == 1)
1939 PredBB = PredBBs[0];
1940 else {
1941 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)
1942 << " common predecessors.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Factoring out " <<
PredBBs.size() << " common predecessors.\n"; } } while
(false)
;
1943 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
1944 }
1945
1946 // And finally, do it!
1947 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)
1948 << "' 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)
1949 << "' 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)
1950 << ", 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)
;
1951
1952 if (DTU->hasPendingDomTreeUpdates())
1953 LVI->disableDT();
1954 else
1955 LVI->enableDT();
1956 LVI->threadEdge(PredBB, BB, SuccBB);
1957
1958 // We are going to have to map operands from the original BB block to the new
1959 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1960 // account for entry from PredBB.
1961 DenseMap<Instruction*, Value*> ValueMapping;
1962
1963 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1964 BB->getName()+".thread",
1965 BB->getParent(), BB);
1966 NewBB->moveAfter(PredBB);
1967
1968 // Set the block frequency of NewBB.
1969 if (HasProfileData) {
1970 auto NewBBFreq =
1971 BFI->getBlockFreq(PredBB) * BPI->getEdgeProbability(PredBB, BB);
1972 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
1973 }
1974
1975 BasicBlock::iterator BI = BB->begin();
1976 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1977 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1978
1979 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1980 // mapping and using it to remap operands in the cloned instructions.
1981 for (; !BI->isTerminator(); ++BI) {
1982 Instruction *New = BI->clone();
1983 New->setName(BI->getName());
1984 NewBB->getInstList().push_back(New);
1985 ValueMapping[&*BI] = New;
1986
1987 // Remap operands to patch up intra-block references.
1988 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1989 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1990 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1991 if (I != ValueMapping.end())
1992 New->setOperand(i, I->second);
1993 }
1994 }
1995
1996 // We didn't copy the terminator from BB over to NewBB, because there is now
1997 // an unconditional jump to SuccBB. Insert the unconditional jump.
1998 BranchInst *NewBI = BranchInst::Create(SuccBB, NewBB);
1999 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
2000
2001 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
2002 // PHI nodes for NewBB now.
2003 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
2004
2005 // Update the terminator of PredBB to jump to NewBB instead of BB. This
2006 // eliminates predecessors from BB, which requires us to simplify any PHI
2007 // nodes in BB.
2008 Instruction *PredTerm = PredBB->getTerminator();
2009 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
2010 if (PredTerm->getSuccessor(i) == BB) {
2011 BB->removePredecessor(PredBB, true);
2012 PredTerm->setSuccessor(i, NewBB);
2013 }
2014
2015 // Enqueue required DT updates.
2016 DTU->applyUpdates({{DominatorTree::Insert, NewBB, SuccBB},
2017 {DominatorTree::Insert, PredBB, NewBB},
2018 {DominatorTree::Delete, PredBB, BB}});
2019
2020 // If there were values defined in BB that are used outside the block, then we
2021 // now have to update all uses of the value to use either the original value,
2022 // the cloned value, or some PHI derived value. This can require arbitrary
2023 // PHI insertion, of which we are prepared to do, clean these up now.
2024 SSAUpdater SSAUpdate;
2025 SmallVector<Use*, 16> UsesToRename;
2026
2027 for (Instruction &I : *BB) {
2028 // Scan all uses of this instruction to see if their uses are no longer
2029 // dominated by the previous def and if so, record them in UsesToRename.
2030 // Also, skip phi operands from PredBB - we'll remove them anyway.
2031 for (Use &U : I.uses()) {
2032 Instruction *User = cast<Instruction>(U.getUser());
2033 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
2034 if (UserPN->getIncomingBlock(U) == BB)
2035 continue;
2036 } else if (User->getParent() == BB)
2037 continue;
2038
2039 UsesToRename.push_back(&U);
2040 }
2041
2042 // If there are no uses outside the block, we're done with this instruction.
2043 if (UsesToRename.empty())
2044 continue;
2045 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)
;
2046
2047 // We found a use of I outside of BB. Rename all uses of I that are outside
2048 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
2049 // with the two values we know.
2050 SSAUpdate.Initialize(I.getType(), I.getName());
2051 SSAUpdate.AddAvailableValue(BB, &I);
2052 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[&I]);
2053
2054 while (!UsesToRename.empty())
2055 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
2056 LLVM_DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "\n"; } } while (false)
;
2057 }
2058
2059 // At this point, the IR is fully up to date and consistent. Do a quick scan
2060 // over the new instructions and zap any that are constants or dead. This
2061 // frequently happens because of phi translation.
2062 SimplifyInstructionsInBlock(NewBB, TLI);
2063
2064 // Update the edge weight from BB to SuccBB, which should be less than before.
2065 UpdateBlockFreqAndEdgeWeight(PredBB, BB, NewBB, SuccBB);
2066
2067 // Threaded an edge!
2068 ++NumThreads;
2069 return true;
2070}
2071
2072/// Create a new basic block that will be the predecessor of BB and successor of
2073/// all blocks in Preds. When profile data is available, update the frequency of
2074/// this new block.
2075BasicBlock *JumpThreadingPass::SplitBlockPreds(BasicBlock *BB,
2076 ArrayRef<BasicBlock *> Preds,
2077 const char *Suffix) {
2078 SmallVector<BasicBlock *, 2> NewBBs;
2079
2080 // Collect the frequencies of all predecessors of BB, which will be used to
2081 // update the edge weight of the result of splitting predecessors.
2082 DenseMap<BasicBlock *, BlockFrequency> FreqMap;
2083 if (HasProfileData)
2084 for (auto Pred : Preds)
2085 FreqMap.insert(std::make_pair(
2086 Pred, BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, BB)));
2087
2088 // In the case when BB is a LandingPad block we create 2 new predecessors
2089 // instead of just one.
2090 if (BB->isLandingPad()) {
2091 std::string NewName = std::string(Suffix) + ".split-lp";
2092 SplitLandingPadPredecessors(BB, Preds, Suffix, NewName.c_str(), NewBBs);
2093 } else {
2094 NewBBs.push_back(SplitBlockPredecessors(BB, Preds, Suffix));
2095 }
2096
2097 std::vector<DominatorTree::UpdateType> Updates;
2098 Updates.reserve((2 * Preds.size()) + NewBBs.size());
2099 for (auto NewBB : NewBBs) {
2100 BlockFrequency NewBBFreq(0);
2101 Updates.push_back({DominatorTree::Insert, NewBB, BB});
2102 for (auto Pred : predecessors(NewBB)) {
2103 Updates.push_back({DominatorTree::Delete, Pred, BB});
2104 Updates.push_back({DominatorTree::Insert, Pred, NewBB});
2105 if (HasProfileData) // Update frequencies between Pred -> NewBB.
2106 NewBBFreq += FreqMap.lookup(Pred);
2107 }
2108 if (HasProfileData) // Apply the summed frequency to NewBB.
2109 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
2110 }
2111
2112 DTU->applyUpdates(Updates);
2113 return NewBBs[0];
2114}
2115
2116bool JumpThreadingPass::doesBlockHaveProfileData(BasicBlock *BB) {
2117 const Instruction *TI = BB->getTerminator();
2118 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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 2118, __PRETTY_FUNCTION__))
;
2119
2120 MDNode *WeightsNode = TI->getMetadata(LLVMContext::MD_prof);
2121 if (!WeightsNode)
2122 return false;
2123
2124 MDString *MDName = cast<MDString>(WeightsNode->getOperand(0));
2125 if (MDName->getString() != "branch_weights")
2126 return false;
2127
2128 // Ensure there are weights for all of the successors. Note that the first
2129 // operand to the metadata node is a name, not a weight.
2130 return WeightsNode->getNumOperands() == TI->getNumSuccessors() + 1;
2131}
2132
2133/// Update the block frequency of BB and branch weight and the metadata on the
2134/// edge BB->SuccBB. This is done by scaling the weight of BB->SuccBB by 1 -
2135/// Freq(PredBB->BB) / Freq(BB->SuccBB).
2136void JumpThreadingPass::UpdateBlockFreqAndEdgeWeight(BasicBlock *PredBB,
2137 BasicBlock *BB,
2138 BasicBlock *NewBB,
2139 BasicBlock *SuccBB) {
2140 if (!HasProfileData)
2141 return;
2142
2143 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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 2143, __PRETTY_FUNCTION__))
;
2144
2145 // As the edge from PredBB to BB is deleted, we have to update the block
2146 // frequency of BB.
2147 auto BBOrigFreq = BFI->getBlockFreq(BB);
2148 auto NewBBFreq = BFI->getBlockFreq(NewBB);
2149 auto BB2SuccBBFreq = BBOrigFreq * BPI->getEdgeProbability(BB, SuccBB);
2150 auto BBNewFreq = BBOrigFreq - NewBBFreq;
2151 BFI->setBlockFreq(BB, BBNewFreq.getFrequency());
2152
2153 // Collect updated outgoing edges' frequencies from BB and use them to update
2154 // edge probabilities.
2155 SmallVector<uint64_t, 4> BBSuccFreq;
2156 for (BasicBlock *Succ : successors(BB)) {
2157 auto SuccFreq = (Succ == SuccBB)
2158 ? BB2SuccBBFreq - NewBBFreq
2159 : BBOrigFreq * BPI->getEdgeProbability(BB, Succ);
2160 BBSuccFreq.push_back(SuccFreq.getFrequency());
2161 }
2162
2163 uint64_t MaxBBSuccFreq =
2164 *std::max_element(BBSuccFreq.begin(), BBSuccFreq.end());
2165
2166 SmallVector<BranchProbability, 4> BBSuccProbs;
2167 if (MaxBBSuccFreq == 0)
2168 BBSuccProbs.assign(BBSuccFreq.size(),
2169 {1, static_cast<unsigned>(BBSuccFreq.size())});
2170 else {
2171 for (uint64_t Freq : BBSuccFreq)
2172 BBSuccProbs.push_back(
2173 BranchProbability::getBranchProbability(Freq, MaxBBSuccFreq));
2174 // Normalize edge probabilities so that they sum up to one.
2175 BranchProbability::normalizeProbabilities(BBSuccProbs.begin(),
2176 BBSuccProbs.end());
2177 }
2178
2179 // Update edge probabilities in BPI.
2180 for (int I = 0, E = BBSuccProbs.size(); I < E; I++)
2181 BPI->setEdgeProbability(BB, I, BBSuccProbs[I]);
2182
2183 // Update the profile metadata as well.
2184 //
2185 // Don't do this if the profile of the transformed blocks was statically
2186 // estimated. (This could occur despite the function having an entry
2187 // frequency in completely cold parts of the CFG.)
2188 //
2189 // In this case we don't want to suggest to subsequent passes that the
2190 // calculated weights are fully consistent. Consider this graph:
2191 //
2192 // check_1
2193 // 50% / |
2194 // eq_1 | 50%
2195 // \ |
2196 // check_2
2197 // 50% / |
2198 // eq_2 | 50%
2199 // \ |
2200 // check_3
2201 // 50% / |
2202 // eq_3 | 50%
2203 // \ |
2204 //
2205 // Assuming the blocks check_* all compare the same value against 1, 2 and 3,
2206 // the overall probabilities are inconsistent; the total probability that the
2207 // value is either 1, 2 or 3 is 150%.
2208 //
2209 // As a consequence if we thread eq_1 -> check_2 to check_3, check_2->check_3
2210 // becomes 0%. This is even worse if the edge whose probability becomes 0% is
2211 // the loop exit edge. Then based solely on static estimation we would assume
2212 // the loop was extremely hot.
2213 //
2214 // FIXME this locally as well so that BPI and BFI are consistent as well. We
2215 // shouldn't make edges extremely likely or unlikely based solely on static
2216 // estimation.
2217 if (BBSuccProbs.size() >= 2 && doesBlockHaveProfileData(BB)) {
2218 SmallVector<uint32_t, 4> Weights;
2219 for (auto Prob : BBSuccProbs)
2220 Weights.push_back(Prob.getNumerator());
2221
2222 auto TI = BB->getTerminator();
2223 TI->setMetadata(
2224 LLVMContext::MD_prof,
2225 MDBuilder(TI->getParent()->getContext()).createBranchWeights(Weights));
2226 }
2227}
2228
2229/// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
2230/// to BB which contains an i1 PHI node and a conditional branch on that PHI.
2231/// If we can duplicate the contents of BB up into PredBB do so now, this
2232/// improves the odds that the branch will be on an analyzable instruction like
2233/// a compare.
2234bool JumpThreadingPass::DuplicateCondBranchOnPHIIntoPred(
2235 BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs) {
2236 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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 2236, __PRETTY_FUNCTION__))
;
2237
2238 // If BB is a loop header, then duplicating this block outside the loop would
2239 // cause us to transform this into an irreducible loop, don't do this.
2240 // See the comments above FindLoopHeaders for justifications and caveats.
2241 if (LoopHeaders.count(BB)) {
2242 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)
2243 << "' 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)
2244 << "' - 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)
;
2245 return false;
2246 }
2247
2248 unsigned DuplicationCost =
2249 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
2250 if (DuplicationCost > BBDupThreshold) {
2251 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)
2252 << "' - 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)
;
2253 return false;
2254 }
2255
2256 // And finally, do it! Start by factoring the predecessors if needed.
2257 std::vector<DominatorTree::UpdateType> Updates;
2258 BasicBlock *PredBB;
2259 if (PredBBs.size() == 1)
2260 PredBB = PredBBs[0];
2261 else {
2262 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)
2263 << " common predecessors.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Factoring out " <<
PredBBs.size() << " common predecessors.\n"; } } while
(false)
;
2264 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
2265 }
2266 Updates.push_back({DominatorTree::Delete, PredBB, BB});
2267
2268 // Okay, we decided to do this! Clone all the instructions in BB onto the end
2269 // of PredBB.
2270 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)
2271 << "' 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)
2272 << "' 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)
2273 << 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)
;
2274
2275 // Unless PredBB ends with an unconditional branch, split the edge so that we
2276 // can just clone the bits from BB into the end of the new PredBB.
2277 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
2278
2279 if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
2280 BasicBlock *OldPredBB = PredBB;
2281 PredBB = SplitEdge(OldPredBB, BB);
2282 Updates.push_back({DominatorTree::Insert, OldPredBB, PredBB});
2283 Updates.push_back({DominatorTree::Insert, PredBB, BB});
2284 Updates.push_back({DominatorTree::Delete, OldPredBB, BB});
2285 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
2286 }
2287
2288 // We are going to have to map operands from the original BB block into the
2289 // PredBB block. Evaluate PHI nodes in BB.
2290 DenseMap<Instruction*, Value*> ValueMapping;
2291
2292 BasicBlock::iterator BI = BB->begin();
2293 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
2294 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
2295 // Clone the non-phi instructions of BB into PredBB, keeping track of the
2296 // mapping and using it to remap operands in the cloned instructions.
2297 for (; BI != BB->end(); ++BI) {
2298 Instruction *New = BI->clone();
2299
2300 // Remap operands to patch up intra-block references.
2301 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2302 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
2303 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
2304 if (I != ValueMapping.end())
2305 New->setOperand(i, I->second);
2306 }
2307
2308 // If this instruction can be simplified after the operands are updated,
2309 // just use the simplified value instead. This frequently happens due to
2310 // phi translation.
2311 if (Value *IV = SimplifyInstruction(
2312 New,
2313 {BB->getModule()->getDataLayout(), TLI, nullptr, nullptr, New})) {
2314 ValueMapping[&*BI] = IV;
2315 if (!New->mayHaveSideEffects()) {
2316 New->deleteValue();
2317 New = nullptr;
2318 }
2319 } else {
2320 ValueMapping[&*BI] = New;
2321 }
2322 if (New) {
2323 // Otherwise, insert the new instruction into the block.
2324 New->setName(BI->getName());
2325 PredBB->getInstList().insert(OldPredBranch->getIterator(), New);
2326 // Update Dominance from simplified New instruction operands.
2327 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2328 if (BasicBlock *SuccBB = dyn_cast<BasicBlock>(New->getOperand(i)))
2329 Updates.push_back({DominatorTree::Insert, PredBB, SuccBB});
2330 }
2331 }
2332
2333 // Check to see if the targets of the branch had PHI nodes. If so, we need to
2334 // add entries to the PHI nodes for branch from PredBB now.
2335 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
2336 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
2337 ValueMapping);
2338 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
2339 ValueMapping);
2340
2341 // If there were values defined in BB that are used outside the block, then we
2342 // now have to update all uses of the value to use either the original value,
2343 // the cloned value, or some PHI derived value. This can require arbitrary
2344 // PHI insertion, of which we are prepared to do, clean these up now.
2345 SSAUpdater SSAUpdate;
2346 SmallVector<Use*, 16> UsesToRename;
2347 for (Instruction &I : *BB) {
2348 // Scan all uses of this instruction to see if it is used outside of its
2349 // block, and if so, record them in UsesToRename.
2350 for (Use &U : I.uses()) {
2351 Instruction *User = cast<Instruction>(U.getUser());
2352 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
2353 if (UserPN->getIncomingBlock(U) == BB)
2354 continue;
2355 } else if (User->getParent() == BB)
2356 continue;
2357
2358 UsesToRename.push_back(&U);
2359 }
2360
2361 // If there are no uses outside the block, we're done with this instruction.
2362 if (UsesToRename.empty())
2363 continue;
2364
2365 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)
;
2366
2367 // We found a use of I outside of BB. Rename all uses of I that are outside
2368 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
2369 // with the two values we know.
2370 SSAUpdate.Initialize(I.getType(), I.getName());
2371 SSAUpdate.AddAvailableValue(BB, &I);
2372 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[&I]);
2373
2374 while (!UsesToRename.empty())
2375 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
2376 LLVM_DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "\n"; } } while (false)
;
2377 }
2378
2379 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
2380 // that we nuked.
2381 BB->removePredecessor(PredBB, true);
2382
2383 // Remove the unconditional branch at the end of the PredBB block.
2384 OldPredBranch->eraseFromParent();
2385 DTU->applyUpdates(Updates);
2386
2387 ++NumDupes;
2388 return true;
2389}
2390
2391/// TryToUnfoldSelect - Look for blocks of the form
2392/// bb1:
2393/// %a = select
2394/// br bb2
2395///
2396/// bb2:
2397/// %p = phi [%a, %bb1] ...
2398/// %c = icmp %p
2399/// br i1 %c
2400///
2401/// And expand the select into a branch structure if one of its arms allows %c
2402/// to be folded. This later enables threading from bb1 over bb2.
2403bool JumpThreadingPass::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
2404 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
2405 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
2406 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
2407
2408 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
2409 CondLHS->getParent() != BB)
2410 return false;
2411
2412 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
2413 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
2414 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
2415
2416 // Look if one of the incoming values is a select in the corresponding
2417 // predecessor.
2418 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
2419 continue;
2420
2421 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
2422 if (!PredTerm || !PredTerm->isUnconditional())
2423 continue;
2424
2425 // Now check if one of the select values would allow us to constant fold the
2426 // terminator in BB. We don't do the transform if both sides fold, those
2427 // cases will be threaded in any case.
2428 if (DTU->hasPendingDomTreeUpdates())
2429 LVI->disableDT();
2430 else
2431 LVI->enableDT();
2432 LazyValueInfo::Tristate LHSFolds =
2433 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
2434 CondRHS, Pred, BB, CondCmp);
2435 LazyValueInfo::Tristate RHSFolds =
2436 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
2437 CondRHS, Pred, BB, CondCmp);
2438 if ((LHSFolds != LazyValueInfo::Unknown ||
2439 RHSFolds != LazyValueInfo::Unknown) &&
2440 LHSFolds != RHSFolds) {
2441 // Expand the select.
2442 //
2443 // Pred --
2444 // | v
2445 // | NewBB
2446 // | |
2447 // |-----
2448 // v
2449 // BB
2450 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
2451 BB->getParent(), BB);
2452 // Move the unconditional branch to NewBB.
2453 PredTerm->removeFromParent();
2454 NewBB->getInstList().insert(NewBB->end(), PredTerm);
2455 // Create a conditional branch and update PHI nodes.
2456 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
2457 CondLHS->setIncomingValue(I, SI->getFalseValue());
2458 CondLHS->addIncoming(SI->getTrueValue(), NewBB);
2459 // The select is now dead.
2460 SI->eraseFromParent();
2461
2462 DTU->applyUpdates({{DominatorTree::Insert, NewBB, BB},
2463 {DominatorTree::Insert, Pred, NewBB}});
2464 // Update any other PHI nodes in BB.
2465 for (BasicBlock::iterator BI = BB->begin();
2466 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
2467 if (Phi != CondLHS)
2468 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);
2469 return true;
2470 }
2471 }
2472 return false;
2473}
2474
2475/// TryToUnfoldSelectInCurrBB - Look for PHI/Select or PHI/CMP/Select in the
2476/// same BB in the form
2477/// bb:
2478/// %p = phi [false, %bb1], [true, %bb2], [false, %bb3], [true, %bb4], ...
2479/// %s = select %p, trueval, falseval
2480///
2481/// or
2482///
2483/// bb:
2484/// %p = phi [0, %bb1], [1, %bb2], [0, %bb3], [1, %bb4], ...
2485/// %c = cmp %p, 0
2486/// %s = select %c, trueval, falseval
2487///
2488/// And expand the select into a branch structure. This later enables
2489/// jump-threading over bb in this pass.
2490///
2491/// Using the similar approach of SimplifyCFG::FoldCondBranchOnPHI(), unfold
2492/// select if the associated PHI has at least one constant. If the unfolded
2493/// select is not jump-threaded, it will be folded again in the later
2494/// optimizations.
2495bool JumpThreadingPass::TryToUnfoldSelectInCurrBB(BasicBlock *BB) {
2496 // If threading this would thread across a loop header, don't thread the edge.
2497 // See the comments above FindLoopHeaders for justifications and caveats.
2498 if (LoopHeaders.count(BB))
2499 return false;
2500
2501 for (BasicBlock::iterator BI = BB->begin();
2502 PHINode *PN = dyn_cast<PHINode>(BI); ++BI) {
2503 // Look for a Phi having at least one constant incoming value.
2504 if (llvm::all_of(PN->incoming_values(),
2505 [](Value *V) { return !isa<ConstantInt>(V); }))
2506 continue;
2507
2508 auto isUnfoldCandidate = [BB](SelectInst *SI, Value *V) {
2509 // Check if SI is in BB and use V as condition.
2510 if (SI->getParent() != BB)
2511 return false;
2512 Value *Cond = SI->getCondition();
2513 return (Cond && Cond == V && Cond->getType()->isIntegerTy(1));
2514 };
2515
2516 SelectInst *SI = nullptr;
2517 for (Use &U : PN->uses()) {
2518 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(U.getUser())) {
2519 // Look for a ICmp in BB that compares PN with a constant and is the
2520 // condition of a Select.
2521 if (Cmp->getParent() == BB && Cmp->hasOneUse() &&
2522 isa<ConstantInt>(Cmp->getOperand(1 - U.getOperandNo())))
2523 if (SelectInst *SelectI = dyn_cast<SelectInst>(Cmp->user_back()))
2524 if (isUnfoldCandidate(SelectI, Cmp->use_begin()->get())) {
2525 SI = SelectI;
2526 break;
2527 }
2528 } else if (SelectInst *SelectI = dyn_cast<SelectInst>(U.getUser())) {
2529 // Look for a Select in BB that uses PN as condition.
2530 if (isUnfoldCandidate(SelectI, U.get())) {
2531 SI = SelectI;
2532 break;
2533 }
2534 }
2535 }
2536
2537 if (!SI)
2538 continue;
2539 // Expand the select.
2540 Instruction *Term =
2541 SplitBlockAndInsertIfThen(SI->getCondition(), SI, false);
2542 BasicBlock *SplitBB = SI->getParent();
2543 BasicBlock *NewBB = Term->getParent();
2544 PHINode *NewPN = PHINode::Create(SI->getType(), 2, "", SI);
2545 NewPN->addIncoming(SI->getTrueValue(), Term->getParent());
2546 NewPN->addIncoming(SI->getFalseValue(), BB);
2547 SI->replaceAllUsesWith(NewPN);
2548 SI->eraseFromParent();
2549 // NewBB and SplitBB are newly created blocks which require insertion.
2550 std::vector<DominatorTree::UpdateType> Updates;
2551 Updates.reserve((2 * SplitBB->getTerminator()->getNumSuccessors()) + 3);
2552 Updates.push_back({DominatorTree::Insert, BB, SplitBB});
2553 Updates.push_back({DominatorTree::Insert, BB, NewBB});
2554 Updates.push_back({DominatorTree::Insert, NewBB, SplitBB});
2555 // BB's successors were moved to SplitBB, update DTU accordingly.
2556 for (auto *Succ : successors(SplitBB)) {
2557 Updates.push_back({DominatorTree::Delete, BB, Succ});
2558 Updates.push_back({DominatorTree::Insert, SplitBB, Succ});
2559 }
2560 DTU->applyUpdates(Updates);
2561 return true;
2562 }
2563 return false;
2564}
2565
2566/// Try to propagate a guard from the current BB into one of its predecessors
2567/// in case if another branch of execution implies that the condition of this
2568/// guard is always true. Currently we only process the simplest case that
2569/// looks like:
2570///
2571/// Start:
2572/// %cond = ...
2573/// br i1 %cond, label %T1, label %F1
2574/// T1:
2575/// br label %Merge
2576/// F1:
2577/// br label %Merge
2578/// Merge:
2579/// %condGuard = ...
2580/// call void(i1, ...) @llvm.experimental.guard( i1 %condGuard )[ "deopt"() ]
2581///
2582/// And cond either implies condGuard or !condGuard. In this case all the
2583/// instructions before the guard can be duplicated in both branches, and the
2584/// guard is then threaded to one of them.
2585bool JumpThreadingPass::ProcessGuards(BasicBlock *BB) {
2586 using namespace PatternMatch;
2587
2588 // We only want to deal with two predecessors.
2589 BasicBlock *Pred1, *Pred2;
2590 auto PI = pred_begin(BB), PE = pred_end(BB);
2591 if (PI == PE)
2592 return false;
2593 Pred1 = *PI++;
2594 if (PI == PE)
2595 return false;
2596 Pred2 = *PI++;
2597 if (PI != PE)
2598 return false;
2599 if (Pred1 == Pred2)
2600 return false;
2601
2602 // Try to thread one of the guards of the block.
2603 // TODO: Look up deeper than to immediate predecessor?
2604 auto *Parent = Pred1->getSinglePredecessor();
2605 if (!Parent || Parent != Pred2->getSinglePredecessor())
2606 return false;
2607
2608 if (auto *BI = dyn_cast<BranchInst>(Parent->getTerminator()))
2609 for (auto &I : *BB)
2610 if (isGuard(&I) && ThreadGuard(BB, cast<IntrinsicInst>(&I), BI))
2611 return true;
2612
2613 return false;
2614}
2615
2616/// Try to propagate the guard from BB which is the lower block of a diamond
2617/// to one of its branches, in case if diamond's condition implies guard's
2618/// condition.
2619bool JumpThreadingPass::ThreadGuard(BasicBlock *BB, IntrinsicInst *Guard,
2620 BranchInst *BI) {
2621 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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 2621, __PRETTY_FUNCTION__))
;
2622 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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 2622, __PRETTY_FUNCTION__))
;
2623 Value *GuardCond = Guard->getArgOperand(0);
2624 Value *BranchCond = BI->getCondition();
2625 BasicBlock *TrueDest = BI->getSuccessor(0);
2626 BasicBlock *FalseDest = BI->getSuccessor(1);
2627
2628 auto &DL = BB->getModule()->getDataLayout();
2629 bool TrueDestIsSafe = false;
2630 bool FalseDestIsSafe = false;
2631
2632 // True dest is safe if BranchCond => GuardCond.
2633 auto Impl = isImpliedCondition(BranchCond, GuardCond, DL);
2634 if (Impl && *Impl)
2635 TrueDestIsSafe = true;
2636 else {
2637 // False dest is safe if !BranchCond => GuardCond.
2638 Impl = isImpliedCondition(BranchCond, GuardCond, DL, /* LHSIsTrue */ false);
2639 if (Impl && *Impl)
2640 FalseDestIsSafe = true;
2641 }
2642
2643 if (!TrueDestIsSafe && !FalseDestIsSafe)
2644 return false;
2645
2646 BasicBlock *PredUnguardedBlock = TrueDestIsSafe ? TrueDest : FalseDest;
2647 BasicBlock *PredGuardedBlock = FalseDestIsSafe ? TrueDest : FalseDest;
2648
2649 ValueToValueMapTy UnguardedMapping, GuardedMapping;
2650 Instruction *AfterGuard = Guard->getNextNode();
2651 unsigned Cost = getJumpThreadDuplicationCost(BB, AfterGuard, BBDupThreshold);
2652 if (Cost > BBDupThreshold)
2653 return false;
2654 // Duplicate all instructions before the guard and the guard itself to the
2655 // branch where implication is not proved.
2656 BasicBlock *GuardedBlock = DuplicateInstructionsInSplitBetween(
2657 BB, PredGuardedBlock, AfterGuard, GuardedMapping, *DTU);
2658 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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 2658, __PRETTY_FUNCTION__))
;
2659 // Duplicate all instructions before the guard in the unguarded branch.
2660 // Since we have successfully duplicated the guarded block and this block
2661 // has fewer instructions, we expect it to succeed.
2662 BasicBlock *UnguardedBlock = DuplicateInstructionsInSplitBetween(
2663 BB, PredUnguardedBlock, Guard, UnguardedMapping, *DTU);
2664 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-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 2664, __PRETTY_FUNCTION__))
;
2665 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)
2666 << GuardedBlock->getName() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "Moved guard " <<
*Guard << " to block " << GuardedBlock->getName
() << "\n"; } } while (false)
;
2667 // Some instructions before the guard may still have uses. For them, we need
2668 // to create Phi nodes merging their copies in both guarded and unguarded
2669 // branches. Those instructions that have no uses can be just removed.
2670 SmallVector<Instruction *, 4> ToRemove;
2671 for (auto BI = BB->begin(); &*BI != AfterGuard; ++BI)
2672 if (!isa<PHINode>(&*BI))
2673 ToRemove.push_back(&*BI);
2674
2675 Instruction *InsertionPoint = &*BB->getFirstInsertionPt();
2676 assert(InsertionPoint && "Empty block?")((InsertionPoint && "Empty block?") ? static_cast<
void> (0) : __assert_fail ("InsertionPoint && \"Empty block?\""
, "/build/llvm-toolchain-snapshot-8~svn350071/lib/Transforms/Scalar/JumpThreading.cpp"
, 2676, __PRETTY_FUNCTION__))
;
2677 // Substitute with Phis & remove.
2678 for (auto *Inst : reverse(ToRemove)) {
2679 if (!Inst->use_empty()) {
2680 PHINode *NewPN = PHINode::Create(Inst->getType(), 2);
2681 NewPN->addIncoming(UnguardedMapping[Inst], UnguardedBlock);
2682 NewPN->addIncoming(GuardedMapping[Inst], GuardedBlock);
2683 NewPN->insertBefore(InsertionPoint);
2684 Inst->replaceAllUsesWith(NewPN);
2685 }
2686 Inst->eraseFromParent();
2687 }
2688 return true;
2689}