Bug Summary

File:lib/Transforms/Scalar/JumpThreading.cpp
Warning:line 1416, 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 -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~svn345461/build-llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-8~svn345461/build-llvm/include -I /build/llvm-toolchain-snapshot-8~svn345461/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~svn345461/build-llvm/lib/Transforms/Scalar -ferror-limit 19 -fmessage-length 0 -fvisibility-inlines-hidden -fobjc-runtime=gcc -fdiagnostics-show-option -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -o /tmp/scan-build-2018-10-27-211344-32123-1 -x c++ /build/llvm-toolchain-snapshot-8~svn345461/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~svn345461/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~svn345461/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~svn345461/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~svn345461/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::ComputeValueKnownInPredecessors(
578 Value *V, BasicBlock *BB, PredValueInfo &Result,
579 ConstantPreference Preference, Instruction *CxtI) {
580 // This method walks up use-def chains recursively. Because of this, we could
581 // get into an infinite loop going around loops in the use-def chain. To
582 // prevent this, keep track of what (value, block) pairs we've already visited
583 // and terminate the search if we loop back to them
584 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
585 return false;
586
587 // An RAII help to remove this pair from the recursion set once the recursion
588 // stack pops back out again.
589 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
590
591 // If V is a constant, then it is known in all predecessors.
592 if (Constant *KC = getKnownConstant(V, Preference)) {
593 for (BasicBlock *Pred : predecessors(BB))
594 Result.push_back(std::make_pair(KC, Pred));
595
596 return !Result.empty();
597 }
598
599 // If V is a non-instruction value, or an instruction in a different block,
600 // then it can't be derived from a PHI.
601 Instruction *I = dyn_cast<Instruction>(V);
602 if (!I || I->getParent() != BB) {
603
604 // Okay, if this is a live-in value, see if it has a known value at the end
605 // of any of our predecessors.
606 //
607 // FIXME: This should be an edge property, not a block end property.
608 /// TODO: Per PR2563, we could infer value range information about a
609 /// predecessor based on its terminator.
610 //
611 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
612 // "I" is a non-local compare-with-a-constant instruction. This would be
613 // able to handle value inequalities better, for example if the compare is
614 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
615 // Perhaps getConstantOnEdge should be smart enough to do this?
616
617 if (DTU->hasPendingDomTreeUpdates())
618 LVI->disableDT();
619 else
620 LVI->enableDT();
621 for (BasicBlock *P : predecessors(BB)) {
622 // If the value is known by LazyValueInfo to be a constant in a
623 // predecessor, use that information to try to thread this block.
624 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB, CxtI);
625 if (Constant *KC = getKnownConstant(PredCst, Preference))
626 Result.push_back(std::make_pair(KC, P));
627 }
628
629 return !Result.empty();
630 }
631
632 /// If I is a PHI node, then we know the incoming values for any constants.
633 if (PHINode *PN = dyn_cast<PHINode>(I)) {
634 if (DTU->hasPendingDomTreeUpdates())
635 LVI->disableDT();
636 else
637 LVI->enableDT();
638 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
639 Value *InVal = PN->getIncomingValue(i);
640 if (Constant *KC = getKnownConstant(InVal, Preference)) {
641 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
642 } else {
643 Constant *CI = LVI->getConstantOnEdge(InVal,
644 PN->getIncomingBlock(i),
645 BB, CxtI);
646 if (Constant *KC = getKnownConstant(CI, Preference))
647 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
648 }
649 }
650
651 return !Result.empty();
652 }
653
654 // Handle Cast instructions. Only see through Cast when the source operand is
655 // PHI or Cmp to save the compilation time.
656 if (CastInst *CI = dyn_cast<CastInst>(I)) {
657 Value *Source = CI->getOperand(0);
658 if (!isa<PHINode>(Source) && !isa<CmpInst>(Source))
659 return false;
660 ComputeValueKnownInPredecessors(Source, BB, Result, Preference, CxtI);
661 if (Result.empty())
662 return false;
663
664 // Convert the known values.
665 for (auto &R : Result)
666 R.first = ConstantExpr::getCast(CI->getOpcode(), R.first, CI->getType());
667
668 return true;
669 }
670
671 // Handle some boolean conditions.
672 if (I->getType()->getPrimitiveSizeInBits() == 1) {
673 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~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 673, __PRETTY_FUNCTION__))
;
674 // X | true -> true
675 // X & false -> false
676 if (I->getOpcode() == Instruction::Or ||
677 I->getOpcode() == Instruction::And) {
678 PredValueInfoTy LHSVals, RHSVals;
679
680 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
681 WantInteger, CxtI);
682 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals,
683 WantInteger, CxtI);
684
685 if (LHSVals.empty() && RHSVals.empty())
686 return false;
687
688 ConstantInt *InterestingVal;
689 if (I->getOpcode() == Instruction::Or)
690 InterestingVal = ConstantInt::getTrue(I->getContext());
691 else
692 InterestingVal = ConstantInt::getFalse(I->getContext());
693
694 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
695
696 // Scan for the sentinel. If we find an undef, force it to the
697 // interesting value: x|undef -> true and x&undef -> false.
698 for (const auto &LHSVal : LHSVals)
699 if (LHSVal.first == InterestingVal || isa<UndefValue>(LHSVal.first)) {
700 Result.emplace_back(InterestingVal, LHSVal.second);
701 LHSKnownBBs.insert(LHSVal.second);
702 }
703 for (const auto &RHSVal : RHSVals)
704 if (RHSVal.first == InterestingVal || isa<UndefValue>(RHSVal.first)) {
705 // If we already inferred a value for this block on the LHS, don't
706 // re-add it.
707 if (!LHSKnownBBs.count(RHSVal.second))
708 Result.emplace_back(InterestingVal, RHSVal.second);
709 }
710
711 return !Result.empty();
712 }
713
714 // Handle the NOT form of XOR.
715 if (I->getOpcode() == Instruction::Xor &&
716 isa<ConstantInt>(I->getOperand(1)) &&
717 cast<ConstantInt>(I->getOperand(1))->isOne()) {
718 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result,
719 WantInteger, CxtI);
720 if (Result.empty())
721 return false;
722
723 // Invert the known values.
724 for (auto &R : Result)
725 R.first = ConstantExpr::getNot(R.first);
726
727 return true;
728 }
729
730 // Try to simplify some other binary operator values.
731 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
732 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~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 733, __PRETTY_FUNCTION__))
733 && "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~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 733, __PRETTY_FUNCTION__))
;
734 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
735 PredValueInfoTy LHSVals;
736 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals,
737 WantInteger, CxtI);
738
739 // Try to use constant folding to simplify the binary operator.
740 for (const auto &LHSVal : LHSVals) {
741 Constant *V = LHSVal.first;
742 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
743
744 if (Constant *KC = getKnownConstant(Folded, WantInteger))
745 Result.push_back(std::make_pair(KC, LHSVal.second));
746 }
747 }
748
749 return !Result.empty();
750 }
751
752 // Handle compare with phi operand, where the PHI is defined in this block.
753 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
754 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~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 754, __PRETTY_FUNCTION__))
;
755 Type *CmpType = Cmp->getType();
756 Value *CmpLHS = Cmp->getOperand(0);
757 Value *CmpRHS = Cmp->getOperand(1);
758 CmpInst::Predicate Pred = Cmp->getPredicate();
759
760 PHINode *PN = dyn_cast<PHINode>(CmpLHS);
761 if (!PN)
762 PN = dyn_cast<PHINode>(CmpRHS);
763 if (PN && PN->getParent() == BB) {
764 const DataLayout &DL = PN->getModule()->getDataLayout();
765 // We can do this simplification if any comparisons fold to true or false.
766 // See if any do.
767 if (DTU->hasPendingDomTreeUpdates())
768 LVI->disableDT();
769 else
770 LVI->enableDT();
771 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
772 BasicBlock *PredBB = PN->getIncomingBlock(i);
773 Value *LHS, *RHS;
774 if (PN == CmpLHS) {
775 LHS = PN->getIncomingValue(i);
776 RHS = CmpRHS->DoPHITranslation(BB, PredBB);
777 } else {
778 LHS = CmpLHS->DoPHITranslation(BB, PredBB);
779 RHS = PN->getIncomingValue(i);
780 }
781 Value *Res = SimplifyCmpInst(Pred, LHS, RHS, {DL});
782 if (!Res) {
783 if (!isa<Constant>(RHS))
784 continue;
785
786 // getPredicateOnEdge call will make no sense if LHS is defined in BB.
787 auto LHSInst = dyn_cast<Instruction>(LHS);
788 if (LHSInst && LHSInst->getParent() == BB)
789 continue;
790
791 LazyValueInfo::Tristate
792 ResT = LVI->getPredicateOnEdge(Pred, LHS,
793 cast<Constant>(RHS), PredBB, BB,
794 CxtI ? CxtI : Cmp);
795 if (ResT == LazyValueInfo::Unknown)
796 continue;
797 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
798 }
799
800 if (Constant *KC = getKnownConstant(Res, WantInteger))
801 Result.push_back(std::make_pair(KC, PredBB));
802 }
803
804 return !Result.empty();
805 }
806
807 // If comparing a live-in value against a constant, see if we know the
808 // live-in value on any predecessors.
809 if (isa<Constant>(CmpRHS) && !CmpType->isVectorTy()) {
810 Constant *CmpConst = cast<Constant>(CmpRHS);
811
812 if (!isa<Instruction>(CmpLHS) ||
813 cast<Instruction>(CmpLHS)->getParent() != BB) {
814 if (DTU->hasPendingDomTreeUpdates())
815 LVI->disableDT();
816 else
817 LVI->enableDT();
818 for (BasicBlock *P : predecessors(BB)) {
819 // If the value is known by LazyValueInfo to be a constant in a
820 // predecessor, use that information to try to thread this block.
821 LazyValueInfo::Tristate Res =
822 LVI->getPredicateOnEdge(Pred, CmpLHS,
823 CmpConst, P, BB, CxtI ? CxtI : Cmp);
824 if (Res == LazyValueInfo::Unknown)
825 continue;
826
827 Constant *ResC = ConstantInt::get(CmpType, Res);
828 Result.push_back(std::make_pair(ResC, P));
829 }
830
831 return !Result.empty();
832 }
833
834 // InstCombine can fold some forms of constant range checks into
835 // (icmp (add (x, C1)), C2). See if we have we have such a thing with
836 // x as a live-in.
837 {
838 using namespace PatternMatch;
839
840 Value *AddLHS;
841 ConstantInt *AddConst;
842 if (isa<ConstantInt>(CmpConst) &&
843 match(CmpLHS, m_Add(m_Value(AddLHS), m_ConstantInt(AddConst)))) {
844 if (!isa<Instruction>(AddLHS) ||
845 cast<Instruction>(AddLHS)->getParent() != BB) {
846 if (DTU->hasPendingDomTreeUpdates())
847 LVI->disableDT();
848 else
849 LVI->enableDT();
850 for (BasicBlock *P : predecessors(BB)) {
851 // If the value is known by LazyValueInfo to be a ConstantRange in
852 // a predecessor, use that information to try to thread this
853 // block.
854 ConstantRange CR = LVI->getConstantRangeOnEdge(
855 AddLHS, P, BB, CxtI ? CxtI : cast<Instruction>(CmpLHS));
856 // Propagate the range through the addition.
857 CR = CR.add(AddConst->getValue());
858
859 // Get the range where the compare returns true.
860 ConstantRange CmpRange = ConstantRange::makeExactICmpRegion(
861 Pred, cast<ConstantInt>(CmpConst)->getValue());
862
863 Constant *ResC;
864 if (CmpRange.contains(CR))
865 ResC = ConstantInt::getTrue(CmpType);
866 else if (CmpRange.inverse().contains(CR))
867 ResC = ConstantInt::getFalse(CmpType);
868 else
869 continue;
870
871 Result.push_back(std::make_pair(ResC, P));
872 }
873
874 return !Result.empty();
875 }
876 }
877 }
878
879 // Try to find a constant value for the LHS of a comparison,
880 // and evaluate it statically if we can.
881 PredValueInfoTy LHSVals;
882 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
883 WantInteger, CxtI);
884
885 for (const auto &LHSVal : LHSVals) {
886 Constant *V = LHSVal.first;
887 Constant *Folded = ConstantExpr::getCompare(Pred, V, CmpConst);
888 if (Constant *KC = getKnownConstant(Folded, WantInteger))
889 Result.push_back(std::make_pair(KC, LHSVal.second));
890 }
891
892 return !Result.empty();
893 }
894 }
895
896 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
897 // Handle select instructions where at least one operand is a known constant
898 // and we can figure out the condition value for any predecessor block.
899 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
900 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
901 PredValueInfoTy Conds;
902 if ((TrueVal || FalseVal) &&
903 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds,
904 WantInteger, CxtI)) {
905 for (auto &C : Conds) {
906 Constant *Cond = C.first;
907
908 // Figure out what value to use for the condition.
909 bool KnownCond;
910 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
911 // A known boolean.
912 KnownCond = CI->isOne();
913 } else {
914 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~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 914, __PRETTY_FUNCTION__))
;
915 // Either operand will do, so be sure to pick the one that's a known
916 // constant.
917 // FIXME: Do this more cleverly if both values are known constants?
918 KnownCond = (TrueVal != nullptr);
919 }
920
921 // See if the select has a known constant value for this predecessor.
922 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
923 Result.push_back(std::make_pair(Val, C.second));
924 }
925
926 return !Result.empty();
927 }
928 }
929
930 // If all else fails, see if LVI can figure out a constant value for us.
931 if (DTU->hasPendingDomTreeUpdates())
932 LVI->disableDT();
933 else
934 LVI->enableDT();
935 Constant *CI = LVI->getConstant(V, BB, CxtI);
936 if (Constant *KC = getKnownConstant(CI, Preference)) {
937 for (BasicBlock *Pred : predecessors(BB))
938 Result.push_back(std::make_pair(KC, Pred));
939 }
940
941 return !Result.empty();
942}
943
944/// GetBestDestForBranchOnUndef - If we determine that the specified block ends
945/// in an undefined jump, decide which block is best to revector to.
946///
947/// Since we can pick an arbitrary destination, we pick the successor with the
948/// fewest predecessors. This should reduce the in-degree of the others.
949static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
950 Instruction *BBTerm = BB->getTerminator();
951 unsigned MinSucc = 0;
952 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
953 // Compute the successor with the minimum number of predecessors.
954 unsigned MinNumPreds = pred_size(TestBB);
955 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
956 TestBB = BBTerm->getSuccessor(i);
957 unsigned NumPreds = pred_size(TestBB);
958 if (NumPreds < MinNumPreds) {
959 MinSucc = i;
960 MinNumPreds = NumPreds;
961 }
962 }
963
964 return MinSucc;
965}
966
967static bool hasAddressTakenAndUsed(BasicBlock *BB) {
968 if (!BB->hasAddressTaken()) return false;
969
970 // If the block has its address taken, it may be a tree of dead constants
971 // hanging off of it. These shouldn't keep the block alive.
972 BlockAddress *BA = BlockAddress::get(BB);
973 BA->removeDeadConstantUsers();
974 return !BA->use_empty();
975}
976
977/// ProcessBlock - If there are any predecessors whose control can be threaded
978/// through to a successor, transform them now.
979bool JumpThreadingPass::ProcessBlock(BasicBlock *BB) {
980 // If the block is trivially dead, just return and let the caller nuke it.
981 // This simplifies other transformations.
982 if (DTU->isBBPendingDeletion(BB) ||
1
Assuming the condition is false
983 (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()))
2
Assuming the condition is false
984 return false;
985
986 // If this block has a single predecessor, and if that pred has a single
987 // successor, merge the blocks. This encourages recursive jump threading
988 // because now the condition in this block can be threaded through
989 // predecessors of our predecessor block.
990 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
3
Assuming 'SinglePred' is null
4
Taking false branch
991 const Instruction *TI = SinglePred->getTerminator();
992 if (!TI->isExceptionalTerminator() && TI->getNumSuccessors() == 1 &&
993 SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
994 // If SinglePred was a loop header, BB becomes one.
995 if (LoopHeaders.erase(SinglePred))
996 LoopHeaders.insert(BB);
997
998 LVI->eraseBlock(SinglePred);
999 MergeBasicBlockIntoOnlyPred(BB, DTU);
1000
1001 // Now that BB is merged into SinglePred (i.e. SinglePred Code followed by
1002 // BB code within one basic block `BB`), we need to invalidate the LVI
1003 // information associated with BB, because the LVI information need not be
1004 // true for all of BB after the merge. For example,
1005 // Before the merge, LVI info and code is as follows:
1006 // SinglePred: <LVI info1 for %p val>
1007 // %y = use of %p
1008 // call @exit() // need not transfer execution to successor.
1009 // assume(%p) // from this point on %p is true
1010 // br label %BB
1011 // BB: <LVI info2 for %p val, i.e. %p is true>
1012 // %x = use of %p
1013 // br label exit
1014 //
1015 // Note that this LVI info for blocks BB and SinglPred is correct for %p
1016 // (info2 and info1 respectively). After the merge and the deletion of the
1017 // LVI info1 for SinglePred. We have the following code:
1018 // BB: <LVI info2 for %p val>
1019 // %y = use of %p
1020 // call @exit()
1021 // assume(%p)
1022 // %x = use of %p <-- LVI info2 is correct from here onwards.
1023 // br label exit
1024 // LVI info2 for BB is incorrect at the beginning of BB.
1025
1026 // Invalidate LVI information for BB if the LVI is not provably true for
1027 // all of BB.
1028 if (!isGuaranteedToTransferExecutionToSuccessor(BB))
1029 LVI->eraseBlock(BB);
1030 return true;
1031 }
1032 }
1033
1034 if (TryToUnfoldSelectInCurrBB(BB))
5
Taking false branch
1035 return true;
1036
1037 // Look if we can propagate guards to predecessors.
1038 if (HasGuards && ProcessGuards(BB))
6
Assuming the condition is false
1039 return true;
1040
1041 // What kind of constant we're looking for.
1042 ConstantPreference Preference = WantInteger;
1043
1044 // Look to see if the terminator is a conditional branch, switch or indirect
1045 // branch, if not we can't thread it.
1046 Value *Condition;
1047 Instruction *Terminator = BB->getTerminator();
1048 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
7
Taking true branch
1049 // Can't thread an unconditional jump.
1050 if (BI->isUnconditional()) return false;
8
Taking false branch
1051 Condition = BI->getCondition();
1052 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
1053 Condition = SI->getCondition();
1054 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
1055 // Can't thread indirect branch with no successors.
1056 if (IB->getNumSuccessors() == 0) return false;
1057 Condition = IB->getAddress()->stripPointerCasts();
1058 Preference = WantBlockAddress;
1059 } else {
1060 return false; // Must be an invoke.
1061 }
1062
1063 // Run constant folding to see if we can reduce the condition to a simple
1064 // constant.
1065 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
9
Taking true branch
1066 Value *SimpleVal =
1067 ConstantFoldInstruction(I, BB->getModule()->getDataLayout(), TLI);
1068 if (SimpleVal) {
10
Assuming 'SimpleVal' is null
11
Taking false branch
1069 I->replaceAllUsesWith(SimpleVal);
1070 if (isInstructionTriviallyDead(I, TLI))
1071 I->eraseFromParent();
1072 Condition = SimpleVal;
1073 }
1074 }
1075
1076 // If the terminator is branching on an undef, we can pick any of the
1077 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
1078 if (isa<UndefValue>(Condition)) {
12
Taking false branch
1079 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
1080 std::vector<DominatorTree::UpdateType> Updates;
1081
1082 // Fold the branch/switch.
1083 Instruction *BBTerm = BB->getTerminator();
1084 Updates.reserve(BBTerm->getNumSuccessors());
1085 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
1086 if (i == BestSucc) continue;
1087 BasicBlock *Succ = BBTerm->getSuccessor(i);
1088 Succ->removePredecessor(BB, true);
1089 Updates.push_back({DominatorTree::Delete, BB, Succ});
1090 }
1091
1092 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)
1093 << "' 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)
;
1094 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
1095 BBTerm->eraseFromParent();
1096 DTU->applyUpdates(Updates);
1097 return true;
1098 }
1099
1100 // If the terminator of this block is branching on a constant, simplify the
1101 // terminator to an unconditional branch. This can occur due to threading in
1102 // other blocks.
1103 if (getKnownConstant(Condition, Preference)) {
13
Taking false branch
1104 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)
1105 << "' folding terminator: " << *BB->getTerminator()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " In block '" <<
BB->getName() << "' folding terminator: " << *
BB->getTerminator() << '\n'; } } while (false)
1106 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " In block '" <<
BB->getName() << "' folding terminator: " << *
BB->getTerminator() << '\n'; } } while (false)
;
1107 ++NumFolds;
1108 ConstantFoldTerminator(BB, true, nullptr, DTU);
1109 return true;
1110 }
1111
1112 Instruction *CondInst = dyn_cast<Instruction>(Condition);
1113
1114 // All the rest of our checks depend on the condition being an instruction.
1115 if (!CondInst) {
14
Taking false branch
1116 // FIXME: Unify this with code below.
1117 if (ProcessThreadableEdges(Condition, BB, Preference, Terminator))
1118 return true;
1119 return false;
1120 }
1121
1122 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
15
Taking false branch
1123 // If we're branching on a conditional, LVI might be able to determine
1124 // it's value at the branch instruction. We only handle comparisons
1125 // against a constant at this time.
1126 // TODO: This should be extended to handle switches as well.
1127 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1128 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
1129 if (CondBr && CondConst) {
1130 // We should have returned as soon as we turn a conditional branch to
1131 // unconditional. Because its no longer interesting as far as jump
1132 // threading is concerned.
1133 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~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 1133, __PRETTY_FUNCTION__))
;
1134
1135 if (DTU->hasPendingDomTreeUpdates())
1136 LVI->disableDT();
1137 else
1138 LVI->enableDT();
1139 LazyValueInfo::Tristate Ret =
1140 LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0),
1141 CondConst, CondBr);
1142 if (Ret != LazyValueInfo::Unknown) {
1143 unsigned ToRemove = Ret == LazyValueInfo::True ? 1 : 0;
1144 unsigned ToKeep = Ret == LazyValueInfo::True ? 0 : 1;
1145 BasicBlock *ToRemoveSucc = CondBr->getSuccessor(ToRemove);
1146 ToRemoveSucc->removePredecessor(BB, true);
1147 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
1148 CondBr->eraseFromParent();
1149 if (CondCmp->use_empty())
1150 CondCmp->eraseFromParent();
1151 // We can safely replace *some* uses of the CondInst if it has
1152 // exactly one value as returned by LVI. RAUW is incorrect in the
1153 // presence of guards and assumes, that have the `Cond` as the use. This
1154 // is because we use the guards/assume to reason about the `Cond` value
1155 // at the end of block, but RAUW unconditionally replaces all uses
1156 // including the guards/assumes themselves and the uses before the
1157 // guard/assume.
1158 else if (CondCmp->getParent() == BB) {
1159 auto *CI = Ret == LazyValueInfo::True ?
1160 ConstantInt::getTrue(CondCmp->getType()) :
1161 ConstantInt::getFalse(CondCmp->getType());
1162 ReplaceFoldableUses(CondCmp, CI);
1163 }
1164 DTU->deleteEdgeRelaxed(BB, ToRemoveSucc);
1165 return true;
1166 }
1167
1168 // We did not manage to simplify this branch, try to see whether
1169 // CondCmp depends on a known phi-select pattern.
1170 if (TryToUnfoldSelect(CondCmp, BB))
1171 return true;
1172 }
1173 }
1174
1175 // Check for some cases that are worth simplifying. Right now we want to look
1176 // for loads that are used by a switch or by the condition for the branch. If
1177 // we see one, check to see if it's partially redundant. If so, insert a PHI
1178 // which can then be used to thread the values.
1179 Value *SimplifyValue = CondInst;
1180 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
16
Taking false branch
1181 if (isa<Constant>(CondCmp->getOperand(1)))
1182 SimplifyValue = CondCmp->getOperand(0);
1183
1184 // TODO: There are other places where load PRE would be profitable, such as
1185 // more complex comparisons.
1186 if (LoadInst *LoadI = dyn_cast<LoadInst>(SimplifyValue))
17
Taking true branch
1187 if (SimplifyPartiallyRedundantLoad(LoadI))
18
Calling 'JumpThreadingPass::SimplifyPartiallyRedundantLoad'
1188 return true;
1189
1190 // Before threading, try to propagate profile data backwards:
1191 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
1192 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1193 updatePredecessorProfileMetadata(PN, BB);
1194
1195 // Handle a variety of cases where we are branching on something derived from
1196 // a PHI node in the current block. If we can prove that any predecessors
1197 // compute a predictable value based on a PHI node, thread those predecessors.
1198 if (ProcessThreadableEdges(CondInst, BB, Preference, Terminator))
1199 return true;
1200
1201 // If this is an otherwise-unfoldable branch on a phi node in the current
1202 // block, see if we can simplify.
1203 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
1204 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1205 return ProcessBranchOnPHI(PN);
1206
1207 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
1208 if (CondInst->getOpcode() == Instruction::Xor &&
1209 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1210 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
1211
1212 // Search for a stronger dominating condition that can be used to simplify a
1213 // conditional branch leaving BB.
1214 if (ProcessImpliedCondition(BB))
1215 return true;
1216
1217 return false;
1218}
1219
1220bool JumpThreadingPass::ProcessImpliedCondition(BasicBlock *BB) {
1221 auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
1222 if (!BI || !BI->isConditional())
1223 return false;
1224
1225 Value *Cond = BI->getCondition();
1226 BasicBlock *CurrentBB = BB;
1227 BasicBlock *CurrentPred = BB->getSinglePredecessor();
1228 unsigned Iter = 0;
1229
1230 auto &DL = BB->getModule()->getDataLayout();
1231
1232 while (CurrentPred && Iter++ < ImplicationSearchThreshold) {
1233 auto *PBI = dyn_cast<BranchInst>(CurrentPred->getTerminator());
1234 if (!PBI || !PBI->isConditional())
1235 return false;
1236 if (PBI->getSuccessor(0) != CurrentBB && PBI->getSuccessor(1) != CurrentBB)
1237 return false;
1238
1239 bool CondIsTrue = PBI->getSuccessor(0) == CurrentBB;
1240 Optional<bool> Implication =
1241 isImpliedCondition(PBI->getCondition(), Cond, DL, CondIsTrue);
1242 if (Implication) {
1243 BasicBlock *KeepSucc = BI->getSuccessor(*Implication ? 0 : 1);
1244 BasicBlock *RemoveSucc = BI->getSuccessor(*Implication ? 1 : 0);
1245 RemoveSucc->removePredecessor(BB);
1246 BranchInst::Create(KeepSucc, BI);
1247 BI->eraseFromParent();
1248 DTU->deleteEdgeRelaxed(BB, RemoveSucc);
1249 return true;
1250 }
1251 CurrentBB = CurrentPred;
1252 CurrentPred = CurrentBB->getSinglePredecessor();
1253 }
1254
1255 return false;
1256}
1257
1258/// Return true if Op is an instruction defined in the given block.
1259static bool isOpDefinedInBlock(Value *Op, BasicBlock *BB) {
1260 if (Instruction *OpInst = dyn_cast<Instruction>(Op))
1261 if (OpInst->getParent() == BB)
1262 return true;
1263 return false;
1264}
1265
1266/// SimplifyPartiallyRedundantLoad - If LoadI is an obviously partially
1267/// redundant load instruction, eliminate it by replacing it with a PHI node.
1268/// This is an important optimization that encourages jump threading, and needs
1269/// to be run interlaced with other jump threading tasks.
1270bool JumpThreadingPass::SimplifyPartiallyRedundantLoad(LoadInst *LoadI) {
1271 // Don't hack volatile and ordered loads.
1272 if (!LoadI->isUnordered()) return false;
19
Taking false branch
1273
1274 // If the load is defined in a block with exactly one predecessor, it can't be
1275 // partially redundant.
1276 BasicBlock *LoadBB = LoadI->getParent();
1277 if (LoadBB->getSinglePredecessor())
20
Assuming the condition is false
21
Taking false branch
1278 return false;
1279
1280 // If the load is defined in an EH pad, it can't be partially redundant,
1281 // because the edges between the invoke and the EH pad cannot have other
1282 // instructions between them.
1283 if (LoadBB->isEHPad())
22
Assuming the condition is false
23
Taking false branch
1284 return false;
1285
1286 Value *LoadedPtr = LoadI->getOperand(0);
1287
1288 // If the loaded operand is defined in the LoadBB and its not a phi,
1289 // it can't be available in predecessors.
1290 if (isOpDefinedInBlock(LoadedPtr, LoadBB) && !isa<PHINode>(LoadedPtr))
1291 return false;
1292
1293 // Scan a few instructions up from the load, to see if it is obviously live at
1294 // the entry to its block.
1295 BasicBlock::iterator BBIt(LoadI);
1296 bool IsLoadCSE;
1297 if (Value *AvailableVal = FindAvailableLoadedValue(
24
Assuming 'AvailableVal' is null
25
Taking false branch
1298 LoadI, LoadBB, BBIt, DefMaxInstsToScan, AA, &IsLoadCSE)) {
1299 // If the value of the load is locally available within the block, just use
1300 // it. This frequently occurs for reg2mem'd allocas.
1301
1302 if (IsLoadCSE) {
1303 LoadInst *NLoadI = cast<LoadInst>(AvailableVal);
1304 combineMetadataForCSE(NLoadI, LoadI, false);
1305 };
1306
1307 // If the returned value is the load itself, replace with an undef. This can
1308 // only happen in dead loops.
1309 if (AvailableVal == LoadI)
1310 AvailableVal = UndefValue::get(LoadI->getType());
1311 if (AvailableVal->getType() != LoadI->getType())
1312 AvailableVal = CastInst::CreateBitOrPointerCast(
1313 AvailableVal, LoadI->getType(), "", LoadI);
1314 LoadI->replaceAllUsesWith(AvailableVal);
1315 LoadI->eraseFromParent();
1316 return true;
1317 }
1318
1319 // Otherwise, if we scanned the whole block and got to the top of the block,
1320 // we know the block is locally transparent to the load. If not, something
1321 // might clobber its value.
1322 if (BBIt != LoadBB->begin())
26
Taking false branch
1323 return false;
1324
1325 // If all of the loads and stores that feed the value have the same AA tags,
1326 // then we can propagate them onto any newly inserted loads.
1327 AAMDNodes AATags;
1328 LoadI->getAAMetadata(AATags);
1329
1330 SmallPtrSet<BasicBlock*, 8> PredsScanned;
1331
1332 using AvailablePredsTy = SmallVector<std::pair<BasicBlock *, Value *>, 8>;
1333
1334 AvailablePredsTy AvailablePreds;
1335 BasicBlock *OneUnavailablePred = nullptr;
27
'OneUnavailablePred' initialized to a null pointer value
1336 SmallVector<LoadInst*, 8> CSELoads;
1337
1338 // If we got here, the loaded value is transparent through to the start of the
1339 // block. Check to see if it is available in any of the predecessor blocks.
1340 for (BasicBlock *PredBB : predecessors(LoadBB)) {
1341 // If we already scanned this predecessor, skip it.
1342 if (!PredsScanned.insert(PredBB).second)
1343 continue;
1344
1345 BBIt = PredBB->end();
1346 unsigned NumScanedInst = 0;
1347 Value *PredAvailable = nullptr;
1348 // NOTE: We don't CSE load that is volatile or anything stronger than
1349 // unordered, that should have been checked when we entered the function.
1350 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~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 1351, __PRETTY_FUNCTION__))
1351 "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~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 1351, __PRETTY_FUNCTION__))
;
1352 // If this is a load on a phi pointer, phi-translate it and search
1353 // for available load/store to the pointer in predecessors.
1354 Value *Ptr = LoadedPtr->DoPHITranslation(LoadBB, PredBB);
1355 PredAvailable = FindAvailablePtrLoadStore(
1356 Ptr, LoadI->getType(), LoadI->isAtomic(), PredBB, BBIt,
1357 DefMaxInstsToScan, AA, &IsLoadCSE, &NumScanedInst);
1358
1359 // If PredBB has a single predecessor, continue scanning through the
1360 // single predecessor.
1361 BasicBlock *SinglePredBB = PredBB;
1362 while (!PredAvailable && SinglePredBB && BBIt == SinglePredBB->begin() &&
1363 NumScanedInst < DefMaxInstsToScan) {
1364 SinglePredBB = SinglePredBB->getSinglePredecessor();
1365 if (SinglePredBB) {
1366 BBIt = SinglePredBB->end();
1367 PredAvailable = FindAvailablePtrLoadStore(
1368 Ptr, LoadI->getType(), LoadI->isAtomic(), SinglePredBB, BBIt,
1369 (DefMaxInstsToScan - NumScanedInst), AA, &IsLoadCSE,
1370 &NumScanedInst);
1371 }
1372 }
1373
1374 if (!PredAvailable) {
1375 OneUnavailablePred = PredBB;
1376 continue;
1377 }
1378
1379 if (IsLoadCSE)
1380 CSELoads.push_back(cast<LoadInst>(PredAvailable));
1381
1382 // If so, this load is partially redundant. Remember this info so that we
1383 // can create a PHI node.
1384 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
1385 }
1386
1387 // If the loaded value isn't available in any predecessor, it isn't partially
1388 // redundant.
1389 if (AvailablePreds.empty()) return false;
28
Taking false branch
1390
1391 // Okay, the loaded value is available in at least one (and maybe all!)
1392 // predecessors. If the value is unavailable in more than one unique
1393 // predecessor, we want to insert a merge block for those common predecessors.
1394 // This ensures that we only have to insert one reload, thus not increasing
1395 // code size.
1396 BasicBlock *UnavailablePred = nullptr;
1397
1398 // If the value is unavailable in one of predecessors, we will end up
1399 // inserting a new instruction into them. It is only valid if all the
1400 // instructions before LoadI are guaranteed to pass execution to its
1401 // successor, or if LoadI is safe to speculate.
1402 // TODO: If this logic becomes more complex, and we will perform PRE insertion
1403 // farther than to a predecessor, we need to reuse the code from GVN's PRE.
1404 // It requires domination tree analysis, so for this simple case it is an
1405 // overkill.
1406 if (PredsScanned.size() != AvailablePreds.size() &&
29
Assuming the condition is false
1407 !isSafeToSpeculativelyExecute(LoadI))
1408 for (auto I = LoadBB->begin(); &*I != LoadI; ++I)
1409 if (!isGuaranteedToTransferExecutionToSuccessor(&*I))
1410 return false;
1411
1412 // If there is exactly one predecessor where the value is unavailable, the
1413 // already computed 'OneUnavailablePred' block is it. If it ends in an
1414 // unconditional branch, we know that it isn't a critical edge.
1415 if (PredsScanned.size() == AvailablePreds.size()+1 &&
30
Assuming the condition is true
1416 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
31
Called C++ object pointer is null
1417 UnavailablePred = OneUnavailablePred;
1418 } else if (PredsScanned.size() != AvailablePreds.size()) {
1419 // Otherwise, we had multiple unavailable predecessors or we had a critical
1420 // edge from the one.
1421 SmallVector<BasicBlock*, 8> PredsToSplit;
1422 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
1423
1424 for (const auto &AvailablePred : AvailablePreds)
1425 AvailablePredSet.insert(AvailablePred.first);
1426
1427 // Add all the unavailable predecessors to the PredsToSplit list.
1428 for (BasicBlock *P : predecessors(LoadBB)) {
1429 // If the predecessor is an indirect goto, we can't split the edge.
1430 if (isa<IndirectBrInst>(P->getTerminator()))
1431 return false;
1432
1433 if (!AvailablePredSet.count(P))
1434 PredsToSplit.push_back(P);
1435 }
1436
1437 // Split them out to their own block.
1438 UnavailablePred = SplitBlockPreds(LoadBB, PredsToSplit, "thread-pre-split");
1439 }
1440
1441 // If the value isn't available in all predecessors, then there will be
1442 // exactly one where it isn't available. Insert a load on that edge and add
1443 // it to the AvailablePreds list.
1444 if (UnavailablePred) {
1445 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~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 1446, __PRETTY_FUNCTION__))
1446 "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~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 1446, __PRETTY_FUNCTION__))
;
1447 LoadInst *NewVal =
1448 new LoadInst(LoadedPtr->DoPHITranslation(LoadBB, UnavailablePred),
1449 LoadI->getName() + ".pr", false, LoadI->getAlignment(),
1450 LoadI->getOrdering(), LoadI->getSyncScopeID(),
1451 UnavailablePred->getTerminator());
1452 NewVal->setDebugLoc(LoadI->getDebugLoc());
1453 if (AATags)
1454 NewVal->setAAMetadata(AATags);
1455
1456 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
1457 }
1458
1459 // Now we know that each predecessor of this block has a value in
1460 // AvailablePreds, sort them for efficient access as we're walking the preds.
1461 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1462
1463 // Create a PHI node at the start of the block for the PRE'd load value.
1464 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
1465 PHINode *PN = PHINode::Create(LoadI->getType(), std::distance(PB, PE), "",
1466 &LoadBB->front());
1467 PN->takeName(LoadI);
1468 PN->setDebugLoc(LoadI->getDebugLoc());
1469
1470 // Insert new entries into the PHI for each predecessor. A single block may
1471 // have multiple entries here.
1472 for (pred_iterator PI = PB; PI != PE; ++PI) {
1473 BasicBlock *P = *PI;
1474 AvailablePredsTy::iterator I =
1475 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1476 std::make_pair(P, (Value*)nullptr));
1477
1478 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~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 1479, __PRETTY_FUNCTION__))
1479 "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~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 1479, __PRETTY_FUNCTION__))
;
1480
1481 // If we have an available predecessor but it requires casting, insert the
1482 // cast in the predecessor and use the cast. Note that we have to update the
1483 // AvailablePreds vector as we go so that all of the PHI entries for this
1484 // predecessor use the same bitcast.
1485 Value *&PredV = I->second;
1486 if (PredV->getType() != LoadI->getType())
1487 PredV = CastInst::CreateBitOrPointerCast(PredV, LoadI->getType(), "",
1488 P->getTerminator());
1489
1490 PN->addIncoming(PredV, I->first);
1491 }
1492
1493 for (LoadInst *PredLoadI : CSELoads) {
1494 combineMetadataForCSE(PredLoadI, LoadI, true);
1495 }
1496
1497 LoadI->replaceAllUsesWith(PN);
1498 LoadI->eraseFromParent();
1499
1500 return true;
1501}
1502
1503/// FindMostPopularDest - The specified list contains multiple possible
1504/// threadable destinations. Pick the one that occurs the most frequently in
1505/// the list.
1506static BasicBlock *
1507FindMostPopularDest(BasicBlock *BB,
1508 const SmallVectorImpl<std::pair<BasicBlock *,
1509 BasicBlock *>> &PredToDestList) {
1510 assert(!PredToDestList.empty())((!PredToDestList.empty()) ? static_cast<void> (0) : __assert_fail
("!PredToDestList.empty()", "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 1510, __PRETTY_FUNCTION__))
;
1511
1512 // Determine popularity. If there are multiple possible destinations, we
1513 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1514 // blocks with known and real destinations to threading undef. We'll handle
1515 // them later if interesting.
1516 DenseMap<BasicBlock*, unsigned> DestPopularity;
1517 for (const auto &PredToDest : PredToDestList)
1518 if (PredToDest.second)
1519 DestPopularity[PredToDest.second]++;
1520
1521 if (DestPopularity.empty())
1522 return nullptr;
1523
1524 // Find the most popular dest.
1525 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1526 BasicBlock *MostPopularDest = DPI->first;
1527 unsigned Popularity = DPI->second;
1528 SmallVector<BasicBlock*, 4> SamePopularity;
1529
1530 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1531 // If the popularity of this entry isn't higher than the popularity we've
1532 // seen so far, ignore it.
1533 if (DPI->second < Popularity)
1534 ; // ignore.
1535 else if (DPI->second == Popularity) {
1536 // If it is the same as what we've seen so far, keep track of it.
1537 SamePopularity.push_back(DPI->first);
1538 } else {
1539 // If it is more popular, remember it.
1540 SamePopularity.clear();
1541 MostPopularDest = DPI->first;
1542 Popularity = DPI->second;
1543 }
1544 }
1545
1546 // Okay, now we know the most popular destination. If there is more than one
1547 // destination, we need to determine one. This is arbitrary, but we need
1548 // to make a deterministic decision. Pick the first one that appears in the
1549 // successor list.
1550 if (!SamePopularity.empty()) {
1551 SamePopularity.push_back(MostPopularDest);
1552 Instruction *TI = BB->getTerminator();
1553 for (unsigned i = 0; ; ++i) {
1554 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~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 1554, __PRETTY_FUNCTION__))
;
1555
1556 if (!is_contained(SamePopularity, TI->getSuccessor(i)))
1557 continue;
1558
1559 MostPopularDest = TI->getSuccessor(i);
1560 break;
1561 }
1562 }
1563
1564 // Okay, we have finally picked the most popular destination.
1565 return MostPopularDest;
1566}
1567
1568bool JumpThreadingPass::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1569 ConstantPreference Preference,
1570 Instruction *CxtI) {
1571 // If threading this would thread across a loop header, don't even try to
1572 // thread the edge.
1573 if (LoopHeaders.count(BB))
1574 return false;
1575
1576 PredValueInfoTy PredValues;
1577 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference, CxtI))
1578 return false;
1579
1580 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~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 1581, __PRETTY_FUNCTION__))
1581 "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~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 1581, __PRETTY_FUNCTION__))
;
1582
1583 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)
1584 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)
1585 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)
1586 << "': 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)
1587 << " 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)
1588 })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)
;
1589
1590 // Decide what we want to thread through. Convert our list of known values to
1591 // a list of known destinations for each pred. This also discards duplicate
1592 // predecessors and keeps track of the undefined inputs (which are represented
1593 // as a null dest in the PredToDestList).
1594 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1595 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1596
1597 BasicBlock *OnlyDest = nullptr;
1598 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1599 Constant *OnlyVal = nullptr;
1600 Constant *MultipleVal = (Constant *)(intptr_t)~0ULL;
1601
1602 unsigned PredWithKnownDest = 0;
1603 for (const auto &PredValue : PredValues) {
1604 BasicBlock *Pred = PredValue.second;
1605 if (!SeenPreds.insert(Pred).second)
1606 continue; // Duplicate predecessor entry.
1607
1608 Constant *Val = PredValue.first;
1609
1610 BasicBlock *DestBB;
1611 if (isa<UndefValue>(Val))
1612 DestBB = nullptr;
1613 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
1614 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~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 1614, __PRETTY_FUNCTION__))
;
1615 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1616 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1617 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~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 1617, __PRETTY_FUNCTION__))
;
1618 DestBB = SI->findCaseValue(cast<ConstantInt>(Val))->getCaseSuccessor();
1619 } else {
1620 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~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 1621, __PRETTY_FUNCTION__))
1621 && "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~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 1621, __PRETTY_FUNCTION__))
;
1622 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~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 1622, __PRETTY_FUNCTION__))
;
1623 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1624 }
1625
1626 // If we have exactly one destination, remember it for efficiency below.
1627 if (PredToDestList.empty()) {
1628 OnlyDest = DestBB;
1629 OnlyVal = Val;
1630 } else {
1631 if (OnlyDest != DestBB)
1632 OnlyDest = MultipleDestSentinel;
1633 // It possible we have same destination, but different value, e.g. default
1634 // case in switchinst.
1635 if (Val != OnlyVal)
1636 OnlyVal = MultipleVal;
1637 }
1638
1639 // We know where this predecessor is going.
1640 ++PredWithKnownDest;
1641
1642 // If the predecessor ends with an indirect goto, we can't change its
1643 // destination.
1644 if (isa<IndirectBrInst>(Pred->getTerminator()))
1645 continue;
1646
1647 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1648 }
1649
1650 // If all edges were unthreadable, we fail.
1651 if (PredToDestList.empty())
1652 return false;
1653
1654 // If all the predecessors go to a single known successor, we want to fold,
1655 // not thread. By doing so, we do not need to duplicate the current block and
1656 // also miss potential opportunities in case we dont/cant duplicate.
1657 if (OnlyDest && OnlyDest != MultipleDestSentinel) {
1658 if (PredWithKnownDest == (size_t)pred_size(BB)) {
1659 bool SeenFirstBranchToOnlyDest = false;
1660 std::vector <DominatorTree::UpdateType> Updates;
1661 Updates.reserve(BB->getTerminator()->getNumSuccessors() - 1);
1662 for (BasicBlock *SuccBB : successors(BB)) {
1663 if (SuccBB == OnlyDest && !SeenFirstBranchToOnlyDest) {
1664 SeenFirstBranchToOnlyDest = true; // Don't modify the first branch.
1665 } else {
1666 SuccBB->removePredecessor(BB, true); // This is unreachable successor.
1667 Updates.push_back({DominatorTree::Delete, BB, SuccBB});
1668 }
1669 }
1670
1671 // Finally update the terminator.
1672 Instruction *Term = BB->getTerminator();
1673 BranchInst::Create(OnlyDest, Term);
1674 Term->eraseFromParent();
1675 DTU->applyUpdates(Updates);
1676
1677 // If the condition is now dead due to the removal of the old terminator,
1678 // erase it.
1679 if (auto *CondInst = dyn_cast<Instruction>(Cond)) {
1680 if (CondInst->use_empty() && !CondInst->mayHaveSideEffects())
1681 CondInst->eraseFromParent();
1682 // We can safely replace *some* uses of the CondInst if it has
1683 // exactly one value as returned by LVI. RAUW is incorrect in the
1684 // presence of guards and assumes, that have the `Cond` as the use. This
1685 // is because we use the guards/assume to reason about the `Cond` value
1686 // at the end of block, but RAUW unconditionally replaces all uses
1687 // including the guards/assumes themselves and the uses before the
1688 // guard/assume.
1689 else if (OnlyVal && OnlyVal != MultipleVal &&
1690 CondInst->getParent() == BB)
1691 ReplaceFoldableUses(CondInst, OnlyVal);
1692 }
1693 return true;
1694 }
1695 }
1696
1697 // Determine which is the most common successor. If we have many inputs and
1698 // this block is a switch, we want to start by threading the batch that goes
1699 // to the most popular destination first. If we only know about one
1700 // threadable destination (the common case) we can avoid this.
1701 BasicBlock *MostPopularDest = OnlyDest;
1702
1703 if (MostPopularDest == MultipleDestSentinel) {
1704 // Remove any loop headers from the Dest list, ThreadEdge conservatively
1705 // won't process them, but we might have other destination that are eligible
1706 // and we still want to process.
1707 erase_if(PredToDestList,
1708 [&](const std::pair<BasicBlock *, BasicBlock *> &PredToDest) {
1709 return LoopHeaders.count(PredToDest.second) != 0;
1710 });
1711
1712 if (PredToDestList.empty())
1713 return false;
1714
1715 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1716 }
1717
1718 // Now that we know what the most popular destination is, factor all
1719 // predecessors that will jump to it into a single predecessor.
1720 SmallVector<BasicBlock*, 16> PredsToFactor;
1721 for (const auto &PredToDest : PredToDestList)
1722 if (PredToDest.second == MostPopularDest) {
1723 BasicBlock *Pred = PredToDest.first;
1724
1725 // This predecessor may be a switch or something else that has multiple
1726 // edges to the block. Factor each of these edges by listing them
1727 // according to # occurrences in PredsToFactor.
1728 for (BasicBlock *Succ : successors(Pred))
1729 if (Succ == BB)
1730 PredsToFactor.push_back(Pred);
1731 }
1732
1733 // If the threadable edges are branching on an undefined value, we get to pick
1734 // the destination that these predecessors should get to.
1735 if (!MostPopularDest)
1736 MostPopularDest = BB->getTerminator()->
1737 getSuccessor(GetBestDestForJumpOnUndef(BB));
1738
1739 // Ok, try to thread it!
1740 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1741}
1742
1743/// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1744/// a PHI node in the current block. See if there are any simplifications we
1745/// can do based on inputs to the phi node.
1746bool JumpThreadingPass::ProcessBranchOnPHI(PHINode *PN) {
1747 BasicBlock *BB = PN->getParent();
1748
1749 // TODO: We could make use of this to do it once for blocks with common PHI
1750 // values.
1751 SmallVector<BasicBlock*, 1> PredBBs;
1752 PredBBs.resize(1);
1753
1754 // If any of the predecessor blocks end in an unconditional branch, we can
1755 // *duplicate* the conditional branch into that block in order to further
1756 // encourage jump threading and to eliminate cases where we have branch on a
1757 // phi of an icmp (branch on icmp is much better).
1758 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1759 BasicBlock *PredBB = PN->getIncomingBlock(i);
1760 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1761 if (PredBr->isUnconditional()) {
1762 PredBBs[0] = PredBB;
1763 // Try to duplicate BB into PredBB.
1764 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1765 return true;
1766 }
1767 }
1768
1769 return false;
1770}
1771
1772/// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1773/// a xor instruction in the current block. See if there are any
1774/// simplifications we can do based on inputs to the xor.
1775bool JumpThreadingPass::ProcessBranchOnXOR(BinaryOperator *BO) {
1776 BasicBlock *BB = BO->getParent();
1777
1778 // If either the LHS or RHS of the xor is a constant, don't do this
1779 // optimization.
1780 if (isa<ConstantInt>(BO->getOperand(0)) ||
1781 isa<ConstantInt>(BO->getOperand(1)))
1782 return false;
1783
1784 // If the first instruction in BB isn't a phi, we won't be able to infer
1785 // anything special about any particular predecessor.
1786 if (!isa<PHINode>(BB->front()))
1787 return false;
1788
1789 // If this BB is a landing pad, we won't be able to split the edge into it.
1790 if (BB->isEHPad())
1791 return false;
1792
1793 // If we have a xor as the branch input to this block, and we know that the
1794 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1795 // the condition into the predecessor and fix that value to true, saving some
1796 // logical ops on that path and encouraging other paths to simplify.
1797 //
1798 // This copies something like this:
1799 //
1800 // BB:
1801 // %X = phi i1 [1], [%X']
1802 // %Y = icmp eq i32 %A, %B
1803 // %Z = xor i1 %X, %Y
1804 // br i1 %Z, ...
1805 //
1806 // Into:
1807 // BB':
1808 // %Y = icmp ne i32 %A, %B
1809 // br i1 %Y, ...
1810
1811 PredValueInfoTy XorOpValues;
1812 bool isLHS = true;
1813 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1814 WantInteger, BO)) {
1815 assert(XorOpValues.empty())((XorOpValues.empty()) ? static_cast<void> (0) : __assert_fail
("XorOpValues.empty()", "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 1815, __PRETTY_FUNCTION__))
;
1816 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1817 WantInteger, BO))
1818 return false;
1819 isLHS = false;
1820 }
1821
1822 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~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 1823, __PRETTY_FUNCTION__))
1823 "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~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 1823, __PRETTY_FUNCTION__))
;
1824
1825 // Scan the information to see which is most popular: true or false. The
1826 // predecessors can be of the set true, false, or undef.
1827 unsigned NumTrue = 0, NumFalse = 0;
1828 for (const auto &XorOpValue : XorOpValues) {
1829 if (isa<UndefValue>(XorOpValue.first))
1830 // Ignore undefs for the count.
1831 continue;
1832 if (cast<ConstantInt>(XorOpValue.first)->isZero())
1833 ++NumFalse;
1834 else
1835 ++NumTrue;
1836 }
1837
1838 // Determine which value to split on, true, false, or undef if neither.
1839 ConstantInt *SplitVal = nullptr;
1840 if (NumTrue > NumFalse)
1841 SplitVal = ConstantInt::getTrue(BB->getContext());
1842 else if (NumTrue != 0 || NumFalse != 0)
1843 SplitVal = ConstantInt::getFalse(BB->getContext());
1844
1845 // Collect all of the blocks that this can be folded into so that we can
1846 // factor this once and clone it once.
1847 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1848 for (const auto &XorOpValue : XorOpValues) {
1849 if (XorOpValue.first != SplitVal && !isa<UndefValue>(XorOpValue.first))
1850 continue;
1851
1852 BlocksToFoldInto.push_back(XorOpValue.second);
1853 }
1854
1855 // If we inferred a value for all of the predecessors, then duplication won't
1856 // help us. However, we can just replace the LHS or RHS with the constant.
1857 if (BlocksToFoldInto.size() ==
1858 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1859 if (!SplitVal) {
1860 // If all preds provide undef, just nuke the xor, because it is undef too.
1861 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1862 BO->eraseFromParent();
1863 } else if (SplitVal->isZero()) {
1864 // If all preds provide 0, replace the xor with the other input.
1865 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1866 BO->eraseFromParent();
1867 } else {
1868 // If all preds provide 1, set the computed value to 1.
1869 BO->setOperand(!isLHS, SplitVal);
1870 }
1871
1872 return true;
1873 }
1874
1875 // Try to duplicate BB into PredBB.
1876 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1877}
1878
1879/// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1880/// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1881/// NewPred using the entries from OldPred (suitably mapped).
1882static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1883 BasicBlock *OldPred,
1884 BasicBlock *NewPred,
1885 DenseMap<Instruction*, Value*> &ValueMap) {
1886 for (PHINode &PN : PHIBB->phis()) {
1887 // Ok, we have a PHI node. Figure out what the incoming value was for the
1888 // DestBlock.
1889 Value *IV = PN.getIncomingValueForBlock(OldPred);
1890
1891 // Remap the value if necessary.
1892 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1893 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1894 if (I != ValueMap.end())
1895 IV = I->second;
1896 }
1897
1898 PN.addIncoming(IV, NewPred);
1899 }
1900}
1901
1902/// ThreadEdge - We have decided that it is safe and profitable to factor the
1903/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1904/// across BB. Transform the IR to reflect this change.
1905bool JumpThreadingPass::ThreadEdge(BasicBlock *BB,
1906 const SmallVectorImpl<BasicBlock *> &PredBBs,
1907 BasicBlock *SuccBB) {
1908 // If threading to the same block as we come from, we would infinite loop.
1909 if (SuccBB == BB) {
1910 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)
1911 << "' - 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)
;
1912 return false;
1913 }
1914
1915 // If threading this would thread across a loop header, don't thread the edge.
1916 // See the comments above FindLoopHeaders for justifications and caveats.
1917 if (LoopHeaders.count(BB) || LoopHeaders.count(SuccBB)) {
1918 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)
1919 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)
1920 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)
1921 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)
1922 << (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)
1923 << "' 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)
1924 << 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)
1925 })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB
); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() <<
" Not threading across " << (BBIsHeader ? "loop header BB '"
: "block BB '") << BB->getName() << "' to dest "
<< (SuccIsHeader ? "loop header BB '" : "block BB '") <<
SuccBB->getName() << "' - it might create an irreducible loop!\n"
; }; } } while (false)
;
1926 return false;
1927 }
1928
1929 unsigned JumpThreadCost =
1930 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
1931 if (JumpThreadCost > BBDupThreshold) {
1932 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)
1933 << "' - 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)
;
1934 return false;
1935 }
1936
1937 // And finally, do it! Start by factoring the predecessors if needed.
1938 BasicBlock *PredBB;
1939 if (PredBBs.size() == 1)
1940 PredBB = PredBBs[0];
1941 else {
1942 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)
1943 << " common predecessors.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Factoring out " <<
PredBBs.size() << " common predecessors.\n"; } } while
(false)
;
1944 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
1945 }
1946
1947 // And finally, do it!
1948 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)
1949 << "' 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)
1950 << "' 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)
1951 << ", 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)
;
1952
1953 if (DTU->hasPendingDomTreeUpdates())
1954 LVI->disableDT();
1955 else
1956 LVI->enableDT();
1957 LVI->threadEdge(PredBB, BB, SuccBB);
1958
1959 // We are going to have to map operands from the original BB block to the new
1960 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1961 // account for entry from PredBB.
1962 DenseMap<Instruction*, Value*> ValueMapping;
1963
1964 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1965 BB->getName()+".thread",
1966 BB->getParent(), BB);
1967 NewBB->moveAfter(PredBB);
1968
1969 // Set the block frequency of NewBB.
1970 if (HasProfileData) {
1971 auto NewBBFreq =
1972 BFI->getBlockFreq(PredBB) * BPI->getEdgeProbability(PredBB, BB);
1973 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
1974 }
1975
1976 BasicBlock::iterator BI = BB->begin();
1977 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1978 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1979
1980 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1981 // mapping and using it to remap operands in the cloned instructions.
1982 for (; !BI->isTerminator(); ++BI) {
1983 Instruction *New = BI->clone();
1984 New->setName(BI->getName());
1985 NewBB->getInstList().push_back(New);
1986 ValueMapping[&*BI] = New;
1987
1988 // Remap operands to patch up intra-block references.
1989 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1990 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1991 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1992 if (I != ValueMapping.end())
1993 New->setOperand(i, I->second);
1994 }
1995 }
1996
1997 // We didn't copy the terminator from BB over to NewBB, because there is now
1998 // an unconditional jump to SuccBB. Insert the unconditional jump.
1999 BranchInst *NewBI = BranchInst::Create(SuccBB, NewBB);
2000 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
2001
2002 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
2003 // PHI nodes for NewBB now.
2004 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
2005
2006 // Update the terminator of PredBB to jump to NewBB instead of BB. This
2007 // eliminates predecessors from BB, which requires us to simplify any PHI
2008 // nodes in BB.
2009 Instruction *PredTerm = PredBB->getTerminator();
2010 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
2011 if (PredTerm->getSuccessor(i) == BB) {
2012 BB->removePredecessor(PredBB, true);
2013 PredTerm->setSuccessor(i, NewBB);
2014 }
2015
2016 // Enqueue required DT updates.
2017 DTU->applyUpdates({{DominatorTree::Insert, NewBB, SuccBB},
2018 {DominatorTree::Insert, PredBB, NewBB},
2019 {DominatorTree::Delete, PredBB, BB}});
2020
2021 // If there were values defined in BB that are used outside the block, then we
2022 // now have to update all uses of the value to use either the original value,
2023 // the cloned value, or some PHI derived value. This can require arbitrary
2024 // PHI insertion, of which we are prepared to do, clean these up now.
2025 SSAUpdater SSAUpdate;
2026 SmallVector<Use*, 16> UsesToRename;
2027
2028 for (Instruction &I : *BB) {
2029 // Scan all uses of this instruction to see if their uses are no longer
2030 // dominated by the previous def and if so, record them in UsesToRename.
2031 // Also, skip phi operands from PredBB - we'll remove them anyway.
2032 for (Use &U : I.uses()) {
2033 Instruction *User = cast<Instruction>(U.getUser());
2034 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
2035 if (UserPN->getIncomingBlock(U) == BB)
2036 continue;
2037 } else if (User->getParent() == BB)
2038 continue;
2039
2040 UsesToRename.push_back(&U);
2041 }
2042
2043 // If there are no uses outside the block, we're done with this instruction.
2044 if (UsesToRename.empty())
2045 continue;
2046 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)
;
2047
2048 // We found a use of I outside of BB. Rename all uses of I that are outside
2049 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
2050 // with the two values we know.
2051 SSAUpdate.Initialize(I.getType(), I.getName());
2052 SSAUpdate.AddAvailableValue(BB, &I);
2053 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[&I]);
2054
2055 while (!UsesToRename.empty())
2056 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
2057 LLVM_DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "\n"; } } while (false)
;
2058 }
2059
2060 // At this point, the IR is fully up to date and consistent. Do a quick scan
2061 // over the new instructions and zap any that are constants or dead. This
2062 // frequently happens because of phi translation.
2063 SimplifyInstructionsInBlock(NewBB, TLI);
2064
2065 // Update the edge weight from BB to SuccBB, which should be less than before.
2066 UpdateBlockFreqAndEdgeWeight(PredBB, BB, NewBB, SuccBB);
2067
2068 // Threaded an edge!
2069 ++NumThreads;
2070 return true;
2071}
2072
2073/// Create a new basic block that will be the predecessor of BB and successor of
2074/// all blocks in Preds. When profile data is available, update the frequency of
2075/// this new block.
2076BasicBlock *JumpThreadingPass::SplitBlockPreds(BasicBlock *BB,
2077 ArrayRef<BasicBlock *> Preds,
2078 const char *Suffix) {
2079 SmallVector<BasicBlock *, 2> NewBBs;
2080
2081 // Collect the frequencies of all predecessors of BB, which will be used to
2082 // update the edge weight of the result of splitting predecessors.
2083 DenseMap<BasicBlock *, BlockFrequency> FreqMap;
2084 if (HasProfileData)
2085 for (auto Pred : Preds)
2086 FreqMap.insert(std::make_pair(
2087 Pred, BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, BB)));
2088
2089 // In the case when BB is a LandingPad block we create 2 new predecessors
2090 // instead of just one.
2091 if (BB->isLandingPad()) {
2092 std::string NewName = std::string(Suffix) + ".split-lp";
2093 SplitLandingPadPredecessors(BB, Preds, Suffix, NewName.c_str(), NewBBs);
2094 } else {
2095 NewBBs.push_back(SplitBlockPredecessors(BB, Preds, Suffix));
2096 }
2097
2098 std::vector<DominatorTree::UpdateType> Updates;
2099 Updates.reserve((2 * Preds.size()) + NewBBs.size());
2100 for (auto NewBB : NewBBs) {
2101 BlockFrequency NewBBFreq(0);
2102 Updates.push_back({DominatorTree::Insert, NewBB, BB});
2103 for (auto Pred : predecessors(NewBB)) {
2104 Updates.push_back({DominatorTree::Delete, Pred, BB});
2105 Updates.push_back({DominatorTree::Insert, Pred, NewBB});
2106 if (HasProfileData) // Update frequencies between Pred -> NewBB.
2107 NewBBFreq += FreqMap.lookup(Pred);
2108 }
2109 if (HasProfileData) // Apply the summed frequency to NewBB.
2110 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
2111 }
2112
2113 DTU->applyUpdates(Updates);
2114 return NewBBs[0];
2115}
2116
2117bool JumpThreadingPass::doesBlockHaveProfileData(BasicBlock *BB) {
2118 const Instruction *TI = BB->getTerminator();
2119 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~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 2119, __PRETTY_FUNCTION__))
;
2120
2121 MDNode *WeightsNode = TI->getMetadata(LLVMContext::MD_prof);
2122 if (!WeightsNode)
2123 return false;
2124
2125 MDString *MDName = cast<MDString>(WeightsNode->getOperand(0));
2126 if (MDName->getString() != "branch_weights")
2127 return false;
2128
2129 // Ensure there are weights for all of the successors. Note that the first
2130 // operand to the metadata node is a name, not a weight.
2131 return WeightsNode->getNumOperands() == TI->getNumSuccessors() + 1;
2132}
2133
2134/// Update the block frequency of BB and branch weight and the metadata on the
2135/// edge BB->SuccBB. This is done by scaling the weight of BB->SuccBB by 1 -
2136/// Freq(PredBB->BB) / Freq(BB->SuccBB).
2137void JumpThreadingPass::UpdateBlockFreqAndEdgeWeight(BasicBlock *PredBB,
2138 BasicBlock *BB,
2139 BasicBlock *NewBB,
2140 BasicBlock *SuccBB) {
2141 if (!HasProfileData)
2142 return;
2143
2144 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~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 2144, __PRETTY_FUNCTION__))
;
2145
2146 // As the edge from PredBB to BB is deleted, we have to update the block
2147 // frequency of BB.
2148 auto BBOrigFreq = BFI->getBlockFreq(BB);
2149 auto NewBBFreq = BFI->getBlockFreq(NewBB);
2150 auto BB2SuccBBFreq = BBOrigFreq * BPI->getEdgeProbability(BB, SuccBB);
2151 auto BBNewFreq = BBOrigFreq - NewBBFreq;
2152 BFI->setBlockFreq(BB, BBNewFreq.getFrequency());
2153
2154 // Collect updated outgoing edges' frequencies from BB and use them to update
2155 // edge probabilities.
2156 SmallVector<uint64_t, 4> BBSuccFreq;
2157 for (BasicBlock *Succ : successors(BB)) {
2158 auto SuccFreq = (Succ == SuccBB)
2159 ? BB2SuccBBFreq - NewBBFreq
2160 : BBOrigFreq * BPI->getEdgeProbability(BB, Succ);
2161 BBSuccFreq.push_back(SuccFreq.getFrequency());
2162 }
2163
2164 uint64_t MaxBBSuccFreq =
2165 *std::max_element(BBSuccFreq.begin(), BBSuccFreq.end());
2166
2167 SmallVector<BranchProbability, 4> BBSuccProbs;
2168 if (MaxBBSuccFreq == 0)
2169 BBSuccProbs.assign(BBSuccFreq.size(),
2170 {1, static_cast<unsigned>(BBSuccFreq.size())});
2171 else {
2172 for (uint64_t Freq : BBSuccFreq)
2173 BBSuccProbs.push_back(
2174 BranchProbability::getBranchProbability(Freq, MaxBBSuccFreq));
2175 // Normalize edge probabilities so that they sum up to one.
2176 BranchProbability::normalizeProbabilities(BBSuccProbs.begin(),
2177 BBSuccProbs.end());
2178 }
2179
2180 // Update edge probabilities in BPI.
2181 for (int I = 0, E = BBSuccProbs.size(); I < E; I++)
2182 BPI->setEdgeProbability(BB, I, BBSuccProbs[I]);
2183
2184 // Update the profile metadata as well.
2185 //
2186 // Don't do this if the profile of the transformed blocks was statically
2187 // estimated. (This could occur despite the function having an entry
2188 // frequency in completely cold parts of the CFG.)
2189 //
2190 // In this case we don't want to suggest to subsequent passes that the
2191 // calculated weights are fully consistent. Consider this graph:
2192 //
2193 // check_1
2194 // 50% / |
2195 // eq_1 | 50%
2196 // \ |
2197 // check_2
2198 // 50% / |
2199 // eq_2 | 50%
2200 // \ |
2201 // check_3
2202 // 50% / |
2203 // eq_3 | 50%
2204 // \ |
2205 //
2206 // Assuming the blocks check_* all compare the same value against 1, 2 and 3,
2207 // the overall probabilities are inconsistent; the total probability that the
2208 // value is either 1, 2 or 3 is 150%.
2209 //
2210 // As a consequence if we thread eq_1 -> check_2 to check_3, check_2->check_3
2211 // becomes 0%. This is even worse if the edge whose probability becomes 0% is
2212 // the loop exit edge. Then based solely on static estimation we would assume
2213 // the loop was extremely hot.
2214 //
2215 // FIXME this locally as well so that BPI and BFI are consistent as well. We
2216 // shouldn't make edges extremely likely or unlikely based solely on static
2217 // estimation.
2218 if (BBSuccProbs.size() >= 2 && doesBlockHaveProfileData(BB)) {
2219 SmallVector<uint32_t, 4> Weights;
2220 for (auto Prob : BBSuccProbs)
2221 Weights.push_back(Prob.getNumerator());
2222
2223 auto TI = BB->getTerminator();
2224 TI->setMetadata(
2225 LLVMContext::MD_prof,
2226 MDBuilder(TI->getParent()->getContext()).createBranchWeights(Weights));
2227 }
2228}
2229
2230/// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
2231/// to BB which contains an i1 PHI node and a conditional branch on that PHI.
2232/// If we can duplicate the contents of BB up into PredBB do so now, this
2233/// improves the odds that the branch will be on an analyzable instruction like
2234/// a compare.
2235bool JumpThreadingPass::DuplicateCondBranchOnPHIIntoPred(
2236 BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs) {
2237 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~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 2237, __PRETTY_FUNCTION__))
;
2238
2239 // If BB is a loop header, then duplicating this block outside the loop would
2240 // cause us to transform this into an irreducible loop, don't do this.
2241 // See the comments above FindLoopHeaders for justifications and caveats.
2242 if (LoopHeaders.count(BB)) {
2243 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)
2244 << "' 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)
2245 << "' - 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)
;
2246 return false;
2247 }
2248
2249 unsigned DuplicationCost =
2250 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
2251 if (DuplicationCost > BBDupThreshold) {
2252 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)
2253 << "' - 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)
;
2254 return false;
2255 }
2256
2257 // And finally, do it! Start by factoring the predecessors if needed.
2258 std::vector<DominatorTree::UpdateType> Updates;
2259 BasicBlock *PredBB;
2260 if (PredBBs.size() == 1)
2261 PredBB = PredBBs[0];
2262 else {
2263 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)
2264 << " common predecessors.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Factoring out " <<
PredBBs.size() << " common predecessors.\n"; } } while
(false)
;
2265 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
2266 }
2267 Updates.push_back({DominatorTree::Delete, PredBB, BB});
2268
2269 // Okay, we decided to do this! Clone all the instructions in BB onto the end
2270 // of PredBB.
2271 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)
2272 << "' 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)
2273 << "' 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)
2274 << 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)
;
2275
2276 // Unless PredBB ends with an unconditional branch, split the edge so that we
2277 // can just clone the bits from BB into the end of the new PredBB.
2278 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
2279
2280 if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
2281 BasicBlock *OldPredBB = PredBB;
2282 PredBB = SplitEdge(OldPredBB, BB);
2283 Updates.push_back({DominatorTree::Insert, OldPredBB, PredBB});
2284 Updates.push_back({DominatorTree::Insert, PredBB, BB});
2285 Updates.push_back({DominatorTree::Delete, OldPredBB, BB});
2286 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
2287 }
2288
2289 // We are going to have to map operands from the original BB block into the
2290 // PredBB block. Evaluate PHI nodes in BB.
2291 DenseMap<Instruction*, Value*> ValueMapping;
2292
2293 BasicBlock::iterator BI = BB->begin();
2294 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
2295 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
2296 // Clone the non-phi instructions of BB into PredBB, keeping track of the
2297 // mapping and using it to remap operands in the cloned instructions.
2298 for (; BI != BB->end(); ++BI) {
2299 Instruction *New = BI->clone();
2300
2301 // Remap operands to patch up intra-block references.
2302 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2303 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
2304 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
2305 if (I != ValueMapping.end())
2306 New->setOperand(i, I->second);
2307 }
2308
2309 // If this instruction can be simplified after the operands are updated,
2310 // just use the simplified value instead. This frequently happens due to
2311 // phi translation.
2312 if (Value *IV = SimplifyInstruction(
2313 New,
2314 {BB->getModule()->getDataLayout(), TLI, nullptr, nullptr, New})) {
2315 ValueMapping[&*BI] = IV;
2316 if (!New->mayHaveSideEffects()) {
2317 New->deleteValue();
2318 New = nullptr;
2319 }
2320 } else {
2321 ValueMapping[&*BI] = New;
2322 }
2323 if (New) {
2324 // Otherwise, insert the new instruction into the block.
2325 New->setName(BI->getName());
2326 PredBB->getInstList().insert(OldPredBranch->getIterator(), New);
2327 // Update Dominance from simplified New instruction operands.
2328 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2329 if (BasicBlock *SuccBB = dyn_cast<BasicBlock>(New->getOperand(i)))
2330 Updates.push_back({DominatorTree::Insert, PredBB, SuccBB});
2331 }
2332 }
2333
2334 // Check to see if the targets of the branch had PHI nodes. If so, we need to
2335 // add entries to the PHI nodes for branch from PredBB now.
2336 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
2337 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
2338 ValueMapping);
2339 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
2340 ValueMapping);
2341
2342 // If there were values defined in BB that are used outside the block, then we
2343 // now have to update all uses of the value to use either the original value,
2344 // the cloned value, or some PHI derived value. This can require arbitrary
2345 // PHI insertion, of which we are prepared to do, clean these up now.
2346 SSAUpdater SSAUpdate;
2347 SmallVector<Use*, 16> UsesToRename;
2348 for (Instruction &I : *BB) {
2349 // Scan all uses of this instruction to see if it is used outside of its
2350 // block, and if so, record them in UsesToRename.
2351 for (Use &U : I.uses()) {
2352 Instruction *User = cast<Instruction>(U.getUser());
2353 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
2354 if (UserPN->getIncomingBlock(U) == BB)
2355 continue;
2356 } else if (User->getParent() == BB)
2357 continue;
2358
2359 UsesToRename.push_back(&U);
2360 }
2361
2362 // If there are no uses outside the block, we're done with this instruction.
2363 if (UsesToRename.empty())
2364 continue;
2365
2366 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)
;
2367
2368 // We found a use of I outside of BB. Rename all uses of I that are outside
2369 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
2370 // with the two values we know.
2371 SSAUpdate.Initialize(I.getType(), I.getName());
2372 SSAUpdate.AddAvailableValue(BB, &I);
2373 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[&I]);
2374
2375 while (!UsesToRename.empty())
2376 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
2377 LLVM_DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "\n"; } } while (false)
;
2378 }
2379
2380 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
2381 // that we nuked.
2382 BB->removePredecessor(PredBB, true);
2383
2384 // Remove the unconditional branch at the end of the PredBB block.
2385 OldPredBranch->eraseFromParent();
2386 DTU->applyUpdates(Updates);
2387
2388 ++NumDupes;
2389 return true;
2390}
2391
2392/// TryToUnfoldSelect - Look for blocks of the form
2393/// bb1:
2394/// %a = select
2395/// br bb2
2396///
2397/// bb2:
2398/// %p = phi [%a, %bb1] ...
2399/// %c = icmp %p
2400/// br i1 %c
2401///
2402/// And expand the select into a branch structure if one of its arms allows %c
2403/// to be folded. This later enables threading from bb1 over bb2.
2404bool JumpThreadingPass::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
2405 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
2406 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
2407 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
2408
2409 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
2410 CondLHS->getParent() != BB)
2411 return false;
2412
2413 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
2414 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
2415 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
2416
2417 // Look if one of the incoming values is a select in the corresponding
2418 // predecessor.
2419 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
2420 continue;
2421
2422 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
2423 if (!PredTerm || !PredTerm->isUnconditional())
2424 continue;
2425
2426 // Now check if one of the select values would allow us to constant fold the
2427 // terminator in BB. We don't do the transform if both sides fold, those
2428 // cases will be threaded in any case.
2429 if (DTU->hasPendingDomTreeUpdates())
2430 LVI->disableDT();
2431 else
2432 LVI->enableDT();
2433 LazyValueInfo::Tristate LHSFolds =
2434 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
2435 CondRHS, Pred, BB, CondCmp);
2436 LazyValueInfo::Tristate RHSFolds =
2437 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
2438 CondRHS, Pred, BB, CondCmp);
2439 if ((LHSFolds != LazyValueInfo::Unknown ||
2440 RHSFolds != LazyValueInfo::Unknown) &&
2441 LHSFolds != RHSFolds) {
2442 // Expand the select.
2443 //
2444 // Pred --
2445 // | v
2446 // | NewBB
2447 // | |
2448 // |-----
2449 // v
2450 // BB
2451 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
2452 BB->getParent(), BB);
2453 // Move the unconditional branch to NewBB.
2454 PredTerm->removeFromParent();
2455 NewBB->getInstList().insert(NewBB->end(), PredTerm);
2456 // Create a conditional branch and update PHI nodes.
2457 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
2458 CondLHS->setIncomingValue(I, SI->getFalseValue());
2459 CondLHS->addIncoming(SI->getTrueValue(), NewBB);
2460 // The select is now dead.
2461 SI->eraseFromParent();
2462
2463 DTU->applyUpdates({{DominatorTree::Insert, NewBB, BB},
2464 {DominatorTree::Insert, Pred, NewBB}});
2465 // Update any other PHI nodes in BB.
2466 for (BasicBlock::iterator BI = BB->begin();
2467 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
2468 if (Phi != CondLHS)
2469 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);
2470 return true;
2471 }
2472 }
2473 return false;
2474}
2475
2476/// TryToUnfoldSelectInCurrBB - Look for PHI/Select or PHI/CMP/Select in the
2477/// same BB in the form
2478/// bb:
2479/// %p = phi [false, %bb1], [true, %bb2], [false, %bb3], [true, %bb4], ...
2480/// %s = select %p, trueval, falseval
2481///
2482/// or
2483///
2484/// bb:
2485/// %p = phi [0, %bb1], [1, %bb2], [0, %bb3], [1, %bb4], ...
2486/// %c = cmp %p, 0
2487/// %s = select %c, trueval, falseval
2488///
2489/// And expand the select into a branch structure. This later enables
2490/// jump-threading over bb in this pass.
2491///
2492/// Using the similar approach of SimplifyCFG::FoldCondBranchOnPHI(), unfold
2493/// select if the associated PHI has at least one constant. If the unfolded
2494/// select is not jump-threaded, it will be folded again in the later
2495/// optimizations.
2496bool JumpThreadingPass::TryToUnfoldSelectInCurrBB(BasicBlock *BB) {
2497 // If threading this would thread across a loop header, don't thread the edge.
2498 // See the comments above FindLoopHeaders for justifications and caveats.
2499 if (LoopHeaders.count(BB))
2500 return false;
2501
2502 for (BasicBlock::iterator BI = BB->begin();
2503 PHINode *PN = dyn_cast<PHINode>(BI); ++BI) {
2504 // Look for a Phi having at least one constant incoming value.
2505 if (llvm::all_of(PN->incoming_values(),
2506 [](Value *V) { return !isa<ConstantInt>(V); }))
2507 continue;
2508
2509 auto isUnfoldCandidate = [BB](SelectInst *SI, Value *V) {
2510 // Check if SI is in BB and use V as condition.
2511 if (SI->getParent() != BB)
2512 return false;
2513 Value *Cond = SI->getCondition();
2514 return (Cond && Cond == V && Cond->getType()->isIntegerTy(1));
2515 };
2516
2517 SelectInst *SI = nullptr;
2518 for (Use &U : PN->uses()) {
2519 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(U.getUser())) {
2520 // Look for a ICmp in BB that compares PN with a constant and is the
2521 // condition of a Select.
2522 if (Cmp->getParent() == BB && Cmp->hasOneUse() &&
2523 isa<ConstantInt>(Cmp->getOperand(1 - U.getOperandNo())))
2524 if (SelectInst *SelectI = dyn_cast<SelectInst>(Cmp->user_back()))
2525 if (isUnfoldCandidate(SelectI, Cmp->use_begin()->get())) {
2526 SI = SelectI;
2527 break;
2528 }
2529 } else if (SelectInst *SelectI = dyn_cast<SelectInst>(U.getUser())) {
2530 // Look for a Select in BB that uses PN as condition.
2531 if (isUnfoldCandidate(SelectI, U.get())) {
2532 SI = SelectI;
2533 break;
2534 }
2535 }
2536 }
2537
2538 if (!SI)
2539 continue;
2540 // Expand the select.
2541 Instruction *Term =
2542 SplitBlockAndInsertIfThen(SI->getCondition(), SI, false);
2543 BasicBlock *SplitBB = SI->getParent();
2544 BasicBlock *NewBB = Term->getParent();
2545 PHINode *NewPN = PHINode::Create(SI->getType(), 2, "", SI);
2546 NewPN->addIncoming(SI->getTrueValue(), Term->getParent());
2547 NewPN->addIncoming(SI->getFalseValue(), BB);
2548 SI->replaceAllUsesWith(NewPN);
2549 SI->eraseFromParent();
2550 // NewBB and SplitBB are newly created blocks which require insertion.
2551 std::vector<DominatorTree::UpdateType> Updates;
2552 Updates.reserve((2 * SplitBB->getTerminator()->getNumSuccessors()) + 3);
2553 Updates.push_back({DominatorTree::Insert, BB, SplitBB});
2554 Updates.push_back({DominatorTree::Insert, BB, NewBB});
2555 Updates.push_back({DominatorTree::Insert, NewBB, SplitBB});
2556 // BB's successors were moved to SplitBB, update DTU accordingly.
2557 for (auto *Succ : successors(SplitBB)) {
2558 Updates.push_back({DominatorTree::Delete, BB, Succ});
2559 Updates.push_back({DominatorTree::Insert, SplitBB, Succ});
2560 }
2561 DTU->applyUpdates(Updates);
2562 return true;
2563 }
2564 return false;
2565}
2566
2567/// Try to propagate a guard from the current BB into one of its predecessors
2568/// in case if another branch of execution implies that the condition of this
2569/// guard is always true. Currently we only process the simplest case that
2570/// looks like:
2571///
2572/// Start:
2573/// %cond = ...
2574/// br i1 %cond, label %T1, label %F1
2575/// T1:
2576/// br label %Merge
2577/// F1:
2578/// br label %Merge
2579/// Merge:
2580/// %condGuard = ...
2581/// call void(i1, ...) @llvm.experimental.guard( i1 %condGuard )[ "deopt"() ]
2582///
2583/// And cond either implies condGuard or !condGuard. In this case all the
2584/// instructions before the guard can be duplicated in both branches, and the
2585/// guard is then threaded to one of them.
2586bool JumpThreadingPass::ProcessGuards(BasicBlock *BB) {
2587 using namespace PatternMatch;
2588
2589 // We only want to deal with two predecessors.
2590 BasicBlock *Pred1, *Pred2;
2591 auto PI = pred_begin(BB), PE = pred_end(BB);
2592 if (PI == PE)
2593 return false;
2594 Pred1 = *PI++;
2595 if (PI == PE)
2596 return false;
2597 Pred2 = *PI++;
2598 if (PI != PE)
2599 return false;
2600 if (Pred1 == Pred2)
2601 return false;
2602
2603 // Try to thread one of the guards of the block.
2604 // TODO: Look up deeper than to immediate predecessor?
2605 auto *Parent = Pred1->getSinglePredecessor();
2606 if (!Parent || Parent != Pred2->getSinglePredecessor())
2607 return false;
2608
2609 if (auto *BI = dyn_cast<BranchInst>(Parent->getTerminator()))
2610 for (auto &I : *BB)
2611 if (isGuard(&I) && ThreadGuard(BB, cast<IntrinsicInst>(&I), BI))
2612 return true;
2613
2614 return false;
2615}
2616
2617/// Try to propagate the guard from BB which is the lower block of a diamond
2618/// to one of its branches, in case if diamond's condition implies guard's
2619/// condition.
2620bool JumpThreadingPass::ThreadGuard(BasicBlock *BB, IntrinsicInst *Guard,
2621 BranchInst *BI) {
2622 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~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 2622, __PRETTY_FUNCTION__))
;
2623 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~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 2623, __PRETTY_FUNCTION__))
;
2624 Value *GuardCond = Guard->getArgOperand(0);
2625 Value *BranchCond = BI->getCondition();
2626 BasicBlock *TrueDest = BI->getSuccessor(0);
2627 BasicBlock *FalseDest = BI->getSuccessor(1);
2628
2629 auto &DL = BB->getModule()->getDataLayout();
2630 bool TrueDestIsSafe = false;
2631 bool FalseDestIsSafe = false;
2632
2633 // True dest is safe if BranchCond => GuardCond.
2634 auto Impl = isImpliedCondition(BranchCond, GuardCond, DL);
2635 if (Impl && *Impl)
2636 TrueDestIsSafe = true;
2637 else {
2638 // False dest is safe if !BranchCond => GuardCond.
2639 Impl = isImpliedCondition(BranchCond, GuardCond, DL, /* LHSIsTrue */ false);
2640 if (Impl && *Impl)
2641 FalseDestIsSafe = true;
2642 }
2643
2644 if (!TrueDestIsSafe && !FalseDestIsSafe)
2645 return false;
2646
2647 BasicBlock *PredUnguardedBlock = TrueDestIsSafe ? TrueDest : FalseDest;
2648 BasicBlock *PredGuardedBlock = FalseDestIsSafe ? TrueDest : FalseDest;
2649
2650 ValueToValueMapTy UnguardedMapping, GuardedMapping;
2651 Instruction *AfterGuard = Guard->getNextNode();
2652 unsigned Cost = getJumpThreadDuplicationCost(BB, AfterGuard, BBDupThreshold);
2653 if (Cost > BBDupThreshold)
2654 return false;
2655 // Duplicate all instructions before the guard and the guard itself to the
2656 // branch where implication is not proved.
2657 BasicBlock *GuardedBlock = DuplicateInstructionsInSplitBetween(
2658 BB, PredGuardedBlock, AfterGuard, GuardedMapping);
2659 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~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 2659, __PRETTY_FUNCTION__))
;
2660 // Duplicate all instructions before the guard in the unguarded branch.
2661 // Since we have successfully duplicated the guarded block and this block
2662 // has fewer instructions, we expect it to succeed.
2663 BasicBlock *UnguardedBlock = DuplicateInstructionsInSplitBetween(
2664 BB, PredUnguardedBlock, Guard, UnguardedMapping);
2665 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~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 2665, __PRETTY_FUNCTION__))
;
2666 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)
2667 << GuardedBlock->getName() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "Moved guard " <<
*Guard << " to block " << GuardedBlock->getName
() << "\n"; } } while (false)
;
2668 // DuplicateInstructionsInSplitBetween inserts a new block "BB.split" between
2669 // PredBB and BB. We need to perform two inserts and one delete for each of
2670 // the above calls to update Dominators.
2671 DTU->applyUpdates(
2672 {// Guarded block split.
2673 {DominatorTree::Delete, PredGuardedBlock, BB},
2674 {DominatorTree::Insert, PredGuardedBlock, GuardedBlock},
2675 {DominatorTree::Insert, GuardedBlock, BB},
2676 // Unguarded block split.
2677 {DominatorTree::Delete, PredUnguardedBlock, BB},
2678 {DominatorTree::Insert, PredUnguardedBlock, UnguardedBlock},
2679 {DominatorTree::Insert, UnguardedBlock, BB}});
2680 // Some instructions before the guard may still have uses. For them, we need
2681 // to create Phi nodes merging their copies in both guarded and unguarded
2682 // branches. Those instructions that have no uses can be just removed.
2683 SmallVector<Instruction *, 4> ToRemove;
2684 for (auto BI = BB->begin(); &*BI != AfterGuard; ++BI)
2685 if (!isa<PHINode>(&*BI))
2686 ToRemove.push_back(&*BI);
2687
2688 Instruction *InsertionPoint = &*BB->getFirstInsertionPt();
2689 assert(InsertionPoint && "Empty block?")((InsertionPoint && "Empty block?") ? static_cast<
void> (0) : __assert_fail ("InsertionPoint && \"Empty block?\""
, "/build/llvm-toolchain-snapshot-8~svn345461/lib/Transforms/Scalar/JumpThreading.cpp"
, 2689, __PRETTY_FUNCTION__))
;
2690 // Substitute with Phis & remove.
2691 for (auto *Inst : reverse(ToRemove)) {
2692 if (!Inst->use_empty()) {
2693 PHINode *NewPN = PHINode::Create(Inst->getType(), 2);
2694 NewPN->addIncoming(UnguardedMapping[Inst], UnguardedBlock);
2695 NewPN->addIncoming(GuardedMapping[Inst], GuardedBlock);
2696 NewPN->insertBefore(InsertionPoint);
2697 Inst->replaceAllUsesWith(NewPN);
2698 }
2699 Inst->eraseFromParent();
2700 }
2701 return true;
2702}