Bug Summary

File:lib/Transforms/Scalar/JumpThreading.cpp
Warning:line 1397, 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-eagerly-assume -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-7/lib/clang/7.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-7~svn329677/build-llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-7~svn329677/build-llvm/include -I /build/llvm-toolchain-snapshot-7~svn329677/include -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/7.3.0/../../../../include/c++/7.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/7.3.0/../../../../include/x86_64-linux-gnu/c++/7.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/7.3.0/../../../../include/x86_64-linux-gnu/c++/7.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/7.3.0/../../../../include/c++/7.3.0/backward -internal-isystem /usr/include/clang/7.0.0/include/ -internal-isystem /usr/local/include -internal-isystem /usr/lib/llvm-7/lib/clang/7.0.0/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-7~svn329677/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-checker optin.performance.Padding -analyzer-output=html -analyzer-config stable-report-filename=true -o /tmp/scan-build-2018-04-11-031539-24776-1 -x c++ /build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp
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/InstructionSimplify.h"
29#include "llvm/Analysis/LazyValueInfo.h"
30#include "llvm/Analysis/Loads.h"
31#include "llvm/Analysis/LoopInfo.h"
32#include "llvm/Analysis/TargetLibraryInfo.h"
33#include "llvm/Analysis/Utils/Local.h"
34#include "llvm/Analysis/ValueTracking.h"
35#include "llvm/IR/BasicBlock.h"
36#include "llvm/IR/CFG.h"
37#include "llvm/IR/Constant.h"
38#include "llvm/IR/ConstantRange.h"
39#include "llvm/IR/Constants.h"
40#include "llvm/IR/DataLayout.h"
41#include "llvm/IR/Dominators.h"
42#include "llvm/IR/Function.h"
43#include "llvm/IR/InstrTypes.h"
44#include "llvm/IR/Instruction.h"
45#include "llvm/IR/Instructions.h"
46#include "llvm/IR/IntrinsicInst.h"
47#include "llvm/IR/Intrinsics.h"
48#include "llvm/IR/LLVMContext.h"
49#include "llvm/IR/MDBuilder.h"
50#include "llvm/IR/Metadata.h"
51#include "llvm/IR/Module.h"
52#include "llvm/IR/PassManager.h"
53#include "llvm/IR/PatternMatch.h"
54#include "llvm/IR/Type.h"
55#include "llvm/IR/Use.h"
56#include "llvm/IR/User.h"
57#include "llvm/IR/Value.h"
58#include "llvm/Pass.h"
59#include "llvm/Support/BlockFrequency.h"
60#include "llvm/Support/BranchProbability.h"
61#include "llvm/Support/Casting.h"
62#include "llvm/Support/CommandLine.h"
63#include "llvm/Support/Debug.h"
64#include "llvm/Support/raw_ostream.h"
65#include "llvm/Transforms/Scalar.h"
66#include "llvm/Transforms/Utils/BasicBlockUtils.h"
67#include "llvm/Transforms/Utils/Cloning.h"
68#include "llvm/Transforms/Utils/SSAUpdater.h"
69#include "llvm/Transforms/Utils/ValueMapper.h"
70#include <algorithm>
71#include <cassert>
72#include <cstddef>
73#include <cstdint>
74#include <iterator>
75#include <memory>
76#include <utility>
77
78using namespace llvm;
79using namespace jumpthreading;
80
81#define DEBUG_TYPE"jump-threading" "jump-threading"
82
83STATISTIC(NumThreads, "Number of jumps threaded")static llvm::Statistic NumThreads = {"jump-threading", "NumThreads"
, "Number of jumps threaded", {0}, {false}}
;
84STATISTIC(NumFolds, "Number of terminators folded")static llvm::Statistic NumFolds = {"jump-threading", "NumFolds"
, "Number of terminators folded", {0}, {false}}
;
85STATISTIC(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}}
;
86
87static cl::opt<unsigned>
88BBDuplicateThreshold("jump-threading-threshold",
89 cl::desc("Max block size to duplicate for jump threading"),
90 cl::init(6), cl::Hidden);
91
92static cl::opt<unsigned>
93ImplicationSearchThreshold(
94 "jump-threading-implication-search-threshold",
95 cl::desc("The number of predecessors to search for a stronger "
96 "condition to use to thread over a weaker condition"),
97 cl::init(3), cl::Hidden);
98
99static cl::opt<bool> PrintLVIAfterJumpThreading(
100 "print-lvi-after-jump-threading",
101 cl::desc("Print the LazyValueInfo cache after JumpThreading"), cl::init(false),
102 cl::Hidden);
103
104namespace {
105
106 /// This pass performs 'jump threading', which looks at blocks that have
107 /// multiple predecessors and multiple successors. If one or more of the
108 /// predecessors of the block can be proven to always jump to one of the
109 /// successors, we forward the edge from the predecessor to the successor by
110 /// duplicating the contents of this block.
111 ///
112 /// An example of when this can occur is code like this:
113 ///
114 /// if () { ...
115 /// X = 4;
116 /// }
117 /// if (X < 3) {
118 ///
119 /// In this case, the unconditional branch at the end of the first if can be
120 /// revectored to the false side of the second if.
121 class JumpThreading : public FunctionPass {
122 JumpThreadingPass Impl;
123
124 public:
125 static char ID; // Pass identification
126
127 JumpThreading(int T = -1) : FunctionPass(ID), Impl(T) {
128 initializeJumpThreadingPass(*PassRegistry::getPassRegistry());
129 }
130
131 bool runOnFunction(Function &F) override;
132
133 void getAnalysisUsage(AnalysisUsage &AU) const override {
134 AU.addRequired<DominatorTreeWrapperPass>();
135 AU.addPreserved<DominatorTreeWrapperPass>();
136 AU.addRequired<AAResultsWrapperPass>();
137 AU.addRequired<LazyValueInfoWrapperPass>();
138 AU.addPreserved<LazyValueInfoWrapperPass>();
139 AU.addPreserved<GlobalsAAWrapperPass>();
140 AU.addRequired<TargetLibraryInfoWrapperPass>();
141 }
142
143 void releaseMemory() override { Impl.releaseMemory(); }
144 };
145
146} // end anonymous namespace
147
148char JumpThreading::ID = 0;
149
150INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",static void *initializeJumpThreadingPassOnce(PassRegistry &
Registry) {
151 "Jump Threading", false, false)static void *initializeJumpThreadingPassOnce(PassRegistry &
Registry) {
152INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry);
153INITIALIZE_PASS_DEPENDENCY(LazyValueInfoWrapperPass)initializeLazyValueInfoWrapperPassPass(Registry);
154INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry);
155INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry);
156INITIALIZE_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)); }
157 "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)); }
158
159// Public interface to the Jump Threading pass
160FunctionPass *llvm::createJumpThreadingPass(int Threshold) {
161 return new JumpThreading(Threshold);
162}
163
164JumpThreadingPass::JumpThreadingPass(int T) {
165 BBDupThreshold = (T == -1) ? BBDuplicateThreshold : unsigned(T);
166}
167
168// Update branch probability information according to conditional
169// branch probablity. This is usually made possible for cloned branches
170// in inline instances by the context specific profile in the caller.
171// For instance,
172//
173// [Block PredBB]
174// [Branch PredBr]
175// if (t) {
176// Block A;
177// } else {
178// Block B;
179// }
180//
181// [Block BB]
182// cond = PN([true, %A], [..., %B]); // PHI node
183// [Branch CondBr]
184// if (cond) {
185// ... // P(cond == true) = 1%
186// }
187//
188// Here we know that when block A is taken, cond must be true, which means
189// P(cond == true | A) = 1
190//
191// Given that P(cond == true) = P(cond == true | A) * P(A) +
192// P(cond == true | B) * P(B)
193// we get:
194// P(cond == true ) = P(A) + P(cond == true | B) * P(B)
195//
196// which gives us:
197// P(A) is less than P(cond == true), i.e.
198// P(t == true) <= P(cond == true)
199//
200// In other words, if we know P(cond == true) is unlikely, we know
201// that P(t == true) is also unlikely.
202//
203static void updatePredecessorProfileMetadata(PHINode *PN, BasicBlock *BB) {
204 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
205 if (!CondBr)
206 return;
207
208 BranchProbability BP;
209 uint64_t TrueWeight, FalseWeight;
210 if (!CondBr->extractProfMetadata(TrueWeight, FalseWeight))
211 return;
212
213 // Returns the outgoing edge of the dominating predecessor block
214 // that leads to the PhiNode's incoming block:
215 auto GetPredOutEdge =
216 [](BasicBlock *IncomingBB,
217 BasicBlock *PhiBB) -> std::pair<BasicBlock *, BasicBlock *> {
218 auto *PredBB = IncomingBB;
219 auto *SuccBB = PhiBB;
220 while (true) {
221 BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator());
222 if (PredBr && PredBr->isConditional())
223 return {PredBB, SuccBB};
224 auto *SinglePredBB = PredBB->getSinglePredecessor();
225 if (!SinglePredBB)
226 return {nullptr, nullptr};
227 SuccBB = PredBB;
228 PredBB = SinglePredBB;
229 }
230 };
231
232 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
233 Value *PhiOpnd = PN->getIncomingValue(i);
234 ConstantInt *CI = dyn_cast<ConstantInt>(PhiOpnd);
235
236 if (!CI || !CI->getType()->isIntegerTy(1))
237 continue;
238
239 BP = (CI->isOne() ? BranchProbability::getBranchProbability(
240 TrueWeight, TrueWeight + FalseWeight)
241 : BranchProbability::getBranchProbability(
242 FalseWeight, TrueWeight + FalseWeight));
243
244 auto PredOutEdge = GetPredOutEdge(PN->getIncomingBlock(i), BB);
245 if (!PredOutEdge.first)
246 return;
247
248 BasicBlock *PredBB = PredOutEdge.first;
249 BranchInst *PredBr = cast<BranchInst>(PredBB->getTerminator());
250
251 uint64_t PredTrueWeight, PredFalseWeight;
252 // FIXME: We currently only set the profile data when it is missing.
253 // With PGO, this can be used to refine even existing profile data with
254 // context information. This needs to be done after more performance
255 // testing.
256 if (PredBr->extractProfMetadata(PredTrueWeight, PredFalseWeight))
257 continue;
258
259 // We can not infer anything useful when BP >= 50%, because BP is the
260 // upper bound probability value.
261 if (BP >= BranchProbability(50, 100))
262 continue;
263
264 SmallVector<uint32_t, 2> Weights;
265 if (PredBr->getSuccessor(0) == PredOutEdge.second) {
266 Weights.push_back(BP.getNumerator());
267 Weights.push_back(BP.getCompl().getNumerator());
268 } else {
269 Weights.push_back(BP.getCompl().getNumerator());
270 Weights.push_back(BP.getNumerator());
271 }
272 PredBr->setMetadata(LLVMContext::MD_prof,
273 MDBuilder(PredBr->getParent()->getContext())
274 .createBranchWeights(Weights));
275 }
276}
277
278/// runOnFunction - Toplevel algorithm.
279bool JumpThreading::runOnFunction(Function &F) {
280 if (skipFunction(F))
281 return false;
282 auto TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
283 // Get DT analysis before LVI. When LVI is initialized it conditionally adds
284 // DT if it's available.
285 auto DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
286 auto LVI = &getAnalysis<LazyValueInfoWrapperPass>().getLVI();
287 auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
288 DeferredDominance DDT(*DT);
289 std::unique_ptr<BlockFrequencyInfo> BFI;
290 std::unique_ptr<BranchProbabilityInfo> BPI;
291 bool HasProfileData = F.hasProfileData();
292 if (HasProfileData) {
293 LoopInfo LI{DominatorTree(F)};
294 BPI.reset(new BranchProbabilityInfo(F, LI, TLI));
295 BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
296 }
297
298 bool Changed = Impl.runImpl(F, TLI, LVI, AA, &DDT, HasProfileData,
299 std::move(BFI), std::move(BPI));
300 if (PrintLVIAfterJumpThreading) {
301 dbgs() << "LVI for function '" << F.getName() << "':\n";
302 LVI->printLVI(F, *DT, dbgs());
303 }
304 return Changed;
305}
306
307PreservedAnalyses JumpThreadingPass::run(Function &F,
308 FunctionAnalysisManager &AM) {
309 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
310 // Get DT analysis before LVI. When LVI is initialized it conditionally adds
311 // DT if it's available.
312 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
313 auto &LVI = AM.getResult<LazyValueAnalysis>(F);
314 auto &AA = AM.getResult<AAManager>(F);
315 DeferredDominance DDT(DT);
316
317 std::unique_ptr<BlockFrequencyInfo> BFI;
318 std::unique_ptr<BranchProbabilityInfo> BPI;
319 if (F.hasProfileData()) {
320 LoopInfo LI{DominatorTree(F)};
321 BPI.reset(new BranchProbabilityInfo(F, LI, &TLI));
322 BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
323 }
324
325 bool Changed = runImpl(F, &TLI, &LVI, &AA, &DDT, HasProfileData,
326 std::move(BFI), std::move(BPI));
327
328 if (!Changed)
329 return PreservedAnalyses::all();
330 PreservedAnalyses PA;
331 PA.preserve<GlobalsAA>();
332 PA.preserve<DominatorTreeAnalysis>();
333 PA.preserve<LazyValueAnalysis>();
334 return PA;
335}
336
337bool JumpThreadingPass::runImpl(Function &F, TargetLibraryInfo *TLI_,
338 LazyValueInfo *LVI_, AliasAnalysis *AA_,
339 DeferredDominance *DDT_, bool HasProfileData_,
340 std::unique_ptr<BlockFrequencyInfo> BFI_,
341 std::unique_ptr<BranchProbabilityInfo> BPI_) {
342 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)
;
343 TLI = TLI_;
344 LVI = LVI_;
345 AA = AA_;
346 DDT = DDT_;
347 BFI.reset();
348 BPI.reset();
349 // When profile data is available, we need to update edge weights after
350 // successful jump threading, which requires both BPI and BFI being available.
351 HasProfileData = HasProfileData_;
352 auto *GuardDecl = F.getParent()->getFunction(
353 Intrinsic::getName(Intrinsic::experimental_guard));
354 HasGuards = GuardDecl && !GuardDecl->use_empty();
355 if (HasProfileData) {
356 BPI = std::move(BPI_);
357 BFI = std::move(BFI_);
358 }
359
360 // JumpThreading must not processes blocks unreachable from entry. It's a
361 // waste of compute time and can potentially lead to hangs.
362 SmallPtrSet<BasicBlock *, 16> Unreachable;
363 DominatorTree &DT = DDT->flush();
364 for (auto &BB : F)
365 if (!DT.isReachableFromEntry(&BB))
366 Unreachable.insert(&BB);
367
368 FindLoopHeaders(F);
369
370 bool EverChanged = false;
371 bool Changed;
372 do {
373 Changed = false;
374 for (auto &BB : F) {
375 if (Unreachable.count(&BB))
376 continue;
377 while (ProcessBlock(&BB)) // Thread all of the branches we can over BB.
378 Changed = true;
379 // Stop processing BB if it's the entry or is now deleted. The following
380 // routines attempt to eliminate BB and locating a suitable replacement
381 // for the entry is non-trivial.
382 if (&BB == &F.getEntryBlock() || DDT->pendingDeletedBB(&BB))
383 continue;
384
385 if (pred_empty(&BB)) {
386 // When ProcessBlock makes BB unreachable it doesn't bother to fix up
387 // the instructions in it. We must remove BB to prevent invalid IR.
388 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)
389 << "' with terminator: " << *BB.getTerminator() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " JT: Deleting dead block '"
<< BB.getName() << "' with terminator: " <<
*BB.getTerminator() << '\n'; } } while (false)
;
390 LoopHeaders.erase(&BB);
391 LVI->eraseBlock(&BB);
392 DeleteDeadBlock(&BB, DDT);
393 Changed = true;
394 continue;
395 }
396
397 // ProcessBlock doesn't thread BBs with unconditional TIs. However, if BB
398 // is "almost empty", we attempt to merge BB with its sole successor.
399 auto *BI = dyn_cast<BranchInst>(BB.getTerminator());
400 if (BI && BI->isUnconditional() &&
401 // The terminator must be the only non-phi instruction in BB.
402 BB.getFirstNonPHIOrDbg()->isTerminator() &&
403 // Don't alter Loop headers and latches to ensure another pass can
404 // detect and transform nested loops later.
405 !LoopHeaders.count(&BB) && !LoopHeaders.count(BI->getSuccessor(0)) &&
406 TryToSimplifyUncondBranchFromEmptyBlock(&BB, DDT)) {
407 // BB is valid for cleanup here because we passed in DDT. F remains
408 // BB's parent until a DDT->flush() event.
409 LVI->eraseBlock(&BB);
410 Changed = true;
411 }
412 }
413 EverChanged |= Changed;
414 } while (Changed);
415
416 LoopHeaders.clear();
417 DDT->flush();
418 LVI->enableDT();
419 return EverChanged;
420}
421
422// Replace uses of Cond with ToVal when safe to do so. If all uses are
423// replaced, we can remove Cond. We cannot blindly replace all uses of Cond
424// because we may incorrectly replace uses when guards/assumes are uses of
425// of `Cond` and we used the guards/assume to reason about the `Cond` value
426// at the end of block. RAUW unconditionally replaces all uses
427// including the guards/assumes themselves and the uses before the
428// guard/assume.
429static void ReplaceFoldableUses(Instruction *Cond, Value *ToVal) {
430 assert(Cond->getType() == ToVal->getType())(static_cast <bool> (Cond->getType() == ToVal->getType
()) ? void (0) : __assert_fail ("Cond->getType() == ToVal->getType()"
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 430, __extension__ __PRETTY_FUNCTION__))
;
431 auto *BB = Cond->getParent();
432 // We can unconditionally replace all uses in non-local blocks (i.e. uses
433 // strictly dominated by BB), since LVI information is true from the
434 // terminator of BB.
435 replaceNonLocalUsesWith(Cond, ToVal);
436 for (Instruction &I : reverse(*BB)) {
437 // Reached the Cond whose uses we are trying to replace, so there are no
438 // more uses.
439 if (&I == Cond)
440 break;
441 // We only replace uses in instructions that are guaranteed to reach the end
442 // of BB, where we know Cond is ToVal.
443 if (!isGuaranteedToTransferExecutionToSuccessor(&I))
444 break;
445 I.replaceUsesOfWith(Cond, ToVal);
446 }
447 if (Cond->use_empty() && !Cond->mayHaveSideEffects())
448 Cond->eraseFromParent();
449}
450
451/// Return the cost of duplicating a piece of this block from first non-phi
452/// and before StopAt instruction to thread across it. Stop scanning the block
453/// when exceeding the threshold. If duplication is impossible, returns ~0U.
454static unsigned getJumpThreadDuplicationCost(BasicBlock *BB,
455 Instruction *StopAt,
456 unsigned Threshold) {
457 assert(StopAt->getParent() == BB && "Not an instruction from proper BB?")(static_cast <bool> (StopAt->getParent() == BB &&
"Not an instruction from proper BB?") ? void (0) : __assert_fail
("StopAt->getParent() == BB && \"Not an instruction from proper BB?\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 457, __extension__ __PRETTY_FUNCTION__))
;
458 /// Ignore PHI nodes, these will be flattened when duplication happens.
459 BasicBlock::const_iterator I(BB->getFirstNonPHI());
460
461 // FIXME: THREADING will delete values that are just used to compute the
462 // branch, so they shouldn't count against the duplication cost.
463
464 unsigned Bonus = 0;
465 if (BB->getTerminator() == StopAt) {
466 // Threading through a switch statement is particularly profitable. If this
467 // block ends in a switch, decrease its cost to make it more likely to
468 // happen.
469 if (isa<SwitchInst>(StopAt))
470 Bonus = 6;
471
472 // The same holds for indirect branches, but slightly more so.
473 if (isa<IndirectBrInst>(StopAt))
474 Bonus = 8;
475 }
476
477 // Bump the threshold up so the early exit from the loop doesn't skip the
478 // terminator-based Size adjustment at the end.
479 Threshold += Bonus;
480
481 // Sum up the cost of each instruction until we get to the terminator. Don't
482 // include the terminator because the copy won't include it.
483 unsigned Size = 0;
484 for (; &*I != StopAt; ++I) {
485
486 // Stop scanning the block if we've reached the threshold.
487 if (Size > Threshold)
488 return Size;
489
490 // Debugger intrinsics don't incur code size.
491 if (isa<DbgInfoIntrinsic>(I)) continue;
492
493 // If this is a pointer->pointer bitcast, it is free.
494 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
495 continue;
496
497 // Bail out if this instruction gives back a token type, it is not possible
498 // to duplicate it if it is used outside this BB.
499 if (I->getType()->isTokenTy() && I->isUsedOutsideOfBlock(BB))
500 return ~0U;
501
502 // All other instructions count for at least one unit.
503 ++Size;
504
505 // Calls are more expensive. If they are non-intrinsic calls, we model them
506 // as having cost of 4. If they are a non-vector intrinsic, we model them
507 // as having cost of 2 total, and if they are a vector intrinsic, we model
508 // them as having cost 1.
509 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
510 if (CI->cannotDuplicate() || CI->isConvergent())
511 // Blocks with NoDuplicate are modelled as having infinite cost, so they
512 // are never duplicated.
513 return ~0U;
514 else if (!isa<IntrinsicInst>(CI))
515 Size += 3;
516 else if (!CI->getType()->isVectorTy())
517 Size += 1;
518 }
519 }
520
521 return Size > Bonus ? Size - Bonus : 0;
522}
523
524/// FindLoopHeaders - We do not want jump threading to turn proper loop
525/// structures into irreducible loops. Doing this breaks up the loop nesting
526/// hierarchy and pessimizes later transformations. To prevent this from
527/// happening, we first have to find the loop headers. Here we approximate this
528/// by finding targets of backedges in the CFG.
529///
530/// Note that there definitely are cases when we want to allow threading of
531/// edges across a loop header. For example, threading a jump from outside the
532/// loop (the preheader) to an exit block of the loop is definitely profitable.
533/// It is also almost always profitable to thread backedges from within the loop
534/// to exit blocks, and is often profitable to thread backedges to other blocks
535/// within the loop (forming a nested loop). This simple analysis is not rich
536/// enough to track all of these properties and keep it up-to-date as the CFG
537/// mutates, so we don't allow any of these transformations.
538void JumpThreadingPass::FindLoopHeaders(Function &F) {
539 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
540 FindFunctionBackedges(F, Edges);
541
542 for (const auto &Edge : Edges)
543 LoopHeaders.insert(Edge.second);
544}
545
546/// getKnownConstant - Helper method to determine if we can thread over a
547/// terminator with the given value as its condition, and if so what value to
548/// use for that. What kind of value this is depends on whether we want an
549/// integer or a block address, but an undef is always accepted.
550/// Returns null if Val is null or not an appropriate constant.
551static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
552 if (!Val)
553 return nullptr;
554
555 // Undef is "known" enough.
556 if (UndefValue *U = dyn_cast<UndefValue>(Val))
557 return U;
558
559 if (Preference == WantBlockAddress)
560 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
561
562 return dyn_cast<ConstantInt>(Val);
563}
564
565/// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
566/// if we can infer that the value is a known ConstantInt/BlockAddress or undef
567/// in any of our predecessors. If so, return the known list of value and pred
568/// BB in the result vector.
569///
570/// This returns true if there were any known values.
571bool JumpThreadingPass::ComputeValueKnownInPredecessors(
572 Value *V, BasicBlock *BB, PredValueInfo &Result,
573 ConstantPreference Preference, Instruction *CxtI) {
574 // This method walks up use-def chains recursively. Because of this, we could
575 // get into an infinite loop going around loops in the use-def chain. To
576 // prevent this, keep track of what (value, block) pairs we've already visited
577 // and terminate the search if we loop back to them
578 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
579 return false;
580
581 // An RAII help to remove this pair from the recursion set once the recursion
582 // stack pops back out again.
583 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
584
585 // If V is a constant, then it is known in all predecessors.
586 if (Constant *KC = getKnownConstant(V, Preference)) {
587 for (BasicBlock *Pred : predecessors(BB))
588 Result.push_back(std::make_pair(KC, Pred));
589
590 return !Result.empty();
591 }
592
593 // If V is a non-instruction value, or an instruction in a different block,
594 // then it can't be derived from a PHI.
595 Instruction *I = dyn_cast<Instruction>(V);
596 if (!I || I->getParent() != BB) {
597
598 // Okay, if this is a live-in value, see if it has a known value at the end
599 // of any of our predecessors.
600 //
601 // FIXME: This should be an edge property, not a block end property.
602 /// TODO: Per PR2563, we could infer value range information about a
603 /// predecessor based on its terminator.
604 //
605 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
606 // "I" is a non-local compare-with-a-constant instruction. This would be
607 // able to handle value inequalities better, for example if the compare is
608 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
609 // Perhaps getConstantOnEdge should be smart enough to do this?
610
611 if (DDT->pending())
612 LVI->disableDT();
613 else
614 LVI->enableDT();
615 for (BasicBlock *P : predecessors(BB)) {
616 // If the value is known by LazyValueInfo to be a constant in a
617 // predecessor, use that information to try to thread this block.
618 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB, CxtI);
619 if (Constant *KC = getKnownConstant(PredCst, Preference))
620 Result.push_back(std::make_pair(KC, P));
621 }
622
623 return !Result.empty();
624 }
625
626 /// If I is a PHI node, then we know the incoming values for any constants.
627 if (PHINode *PN = dyn_cast<PHINode>(I)) {
628 if (DDT->pending())
629 LVI->disableDT();
630 else
631 LVI->enableDT();
632 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
633 Value *InVal = PN->getIncomingValue(i);
634 if (Constant *KC = getKnownConstant(InVal, Preference)) {
635 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
636 } else {
637 Constant *CI = LVI->getConstantOnEdge(InVal,
638 PN->getIncomingBlock(i),
639 BB, CxtI);
640 if (Constant *KC = getKnownConstant(CI, Preference))
641 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
642 }
643 }
644
645 return !Result.empty();
646 }
647
648 // Handle Cast instructions. Only see through Cast when the source operand is
649 // PHI or Cmp to save the compilation time.
650 if (CastInst *CI = dyn_cast<CastInst>(I)) {
651 Value *Source = CI->getOperand(0);
652 if (!isa<PHINode>(Source) && !isa<CmpInst>(Source))
653 return false;
654 ComputeValueKnownInPredecessors(Source, BB, Result, Preference, CxtI);
655 if (Result.empty())
656 return false;
657
658 // Convert the known values.
659 for (auto &R : Result)
660 R.first = ConstantExpr::getCast(CI->getOpcode(), R.first, CI->getType());
661
662 return true;
663 }
664
665 // Handle some boolean conditions.
666 if (I->getType()->getPrimitiveSizeInBits() == 1) {
667 assert(Preference == WantInteger && "One-bit non-integer type?")(static_cast <bool> (Preference == WantInteger &&
"One-bit non-integer type?") ? void (0) : __assert_fail ("Preference == WantInteger && \"One-bit non-integer type?\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 667, __extension__ __PRETTY_FUNCTION__))
;
668 // X | true -> true
669 // X & false -> false
670 if (I->getOpcode() == Instruction::Or ||
671 I->getOpcode() == Instruction::And) {
672 PredValueInfoTy LHSVals, RHSVals;
673
674 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
675 WantInteger, CxtI);
676 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals,
677 WantInteger, CxtI);
678
679 if (LHSVals.empty() && RHSVals.empty())
680 return false;
681
682 ConstantInt *InterestingVal;
683 if (I->getOpcode() == Instruction::Or)
684 InterestingVal = ConstantInt::getTrue(I->getContext());
685 else
686 InterestingVal = ConstantInt::getFalse(I->getContext());
687
688 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
689
690 // Scan for the sentinel. If we find an undef, force it to the
691 // interesting value: x|undef -> true and x&undef -> false.
692 for (const auto &LHSVal : LHSVals)
693 if (LHSVal.first == InterestingVal || isa<UndefValue>(LHSVal.first)) {
694 Result.emplace_back(InterestingVal, LHSVal.second);
695 LHSKnownBBs.insert(LHSVal.second);
696 }
697 for (const auto &RHSVal : RHSVals)
698 if (RHSVal.first == InterestingVal || isa<UndefValue>(RHSVal.first)) {
699 // If we already inferred a value for this block on the LHS, don't
700 // re-add it.
701 if (!LHSKnownBBs.count(RHSVal.second))
702 Result.emplace_back(InterestingVal, RHSVal.second);
703 }
704
705 return !Result.empty();
706 }
707
708 // Handle the NOT form of XOR.
709 if (I->getOpcode() == Instruction::Xor &&
710 isa<ConstantInt>(I->getOperand(1)) &&
711 cast<ConstantInt>(I->getOperand(1))->isOne()) {
712 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result,
713 WantInteger, CxtI);
714 if (Result.empty())
715 return false;
716
717 // Invert the known values.
718 for (auto &R : Result)
719 R.first = ConstantExpr::getNot(R.first);
720
721 return true;
722 }
723
724 // Try to simplify some other binary operator values.
725 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
726 assert(Preference != WantBlockAddress(static_cast <bool> (Preference != WantBlockAddress &&
"A binary operator creating a block address?") ? void (0) : __assert_fail
("Preference != WantBlockAddress && \"A binary operator creating a block address?\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 727, __extension__ __PRETTY_FUNCTION__))
727 && "A binary operator creating a block address?")(static_cast <bool> (Preference != WantBlockAddress &&
"A binary operator creating a block address?") ? void (0) : __assert_fail
("Preference != WantBlockAddress && \"A binary operator creating a block address?\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 727, __extension__ __PRETTY_FUNCTION__))
;
728 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
729 PredValueInfoTy LHSVals;
730 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals,
731 WantInteger, CxtI);
732
733 // Try to use constant folding to simplify the binary operator.
734 for (const auto &LHSVal : LHSVals) {
735 Constant *V = LHSVal.first;
736 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
737
738 if (Constant *KC = getKnownConstant(Folded, WantInteger))
739 Result.push_back(std::make_pair(KC, LHSVal.second));
740 }
741 }
742
743 return !Result.empty();
744 }
745
746 // Handle compare with phi operand, where the PHI is defined in this block.
747 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
748 assert(Preference == WantInteger && "Compares only produce integers")(static_cast <bool> (Preference == WantInteger &&
"Compares only produce integers") ? void (0) : __assert_fail
("Preference == WantInteger && \"Compares only produce integers\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 748, __extension__ __PRETTY_FUNCTION__))
;
749 Type *CmpType = Cmp->getType();
750 Value *CmpLHS = Cmp->getOperand(0);
751 Value *CmpRHS = Cmp->getOperand(1);
752 CmpInst::Predicate Pred = Cmp->getPredicate();
753
754 PHINode *PN = dyn_cast<PHINode>(CmpLHS);
755 if (PN && PN->getParent() == BB) {
756 const DataLayout &DL = PN->getModule()->getDataLayout();
757 // We can do this simplification if any comparisons fold to true or false.
758 // See if any do.
759 if (DDT->pending())
760 LVI->disableDT();
761 else
762 LVI->enableDT();
763 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
764 BasicBlock *PredBB = PN->getIncomingBlock(i);
765 Value *LHS = PN->getIncomingValue(i);
766 Value *RHS = CmpRHS->DoPHITranslation(BB, PredBB);
767
768 Value *Res = SimplifyCmpInst(Pred, LHS, RHS, {DL});
769 if (!Res) {
770 if (!isa<Constant>(RHS))
771 continue;
772
773 LazyValueInfo::Tristate
774 ResT = LVI->getPredicateOnEdge(Pred, LHS,
775 cast<Constant>(RHS), PredBB, BB,
776 CxtI ? CxtI : Cmp);
777 if (ResT == LazyValueInfo::Unknown)
778 continue;
779 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
780 }
781
782 if (Constant *KC = getKnownConstant(Res, WantInteger))
783 Result.push_back(std::make_pair(KC, PredBB));
784 }
785
786 return !Result.empty();
787 }
788
789 // If comparing a live-in value against a constant, see if we know the
790 // live-in value on any predecessors.
791 if (isa<Constant>(CmpRHS) && !CmpType->isVectorTy()) {
792 Constant *CmpConst = cast<Constant>(CmpRHS);
793
794 if (!isa<Instruction>(CmpLHS) ||
795 cast<Instruction>(CmpLHS)->getParent() != BB) {
796 if (DDT->pending())
797 LVI->disableDT();
798 else
799 LVI->enableDT();
800 for (BasicBlock *P : predecessors(BB)) {
801 // If the value is known by LazyValueInfo to be a constant in a
802 // predecessor, use that information to try to thread this block.
803 LazyValueInfo::Tristate Res =
804 LVI->getPredicateOnEdge(Pred, CmpLHS,
805 CmpConst, P, BB, CxtI ? CxtI : Cmp);
806 if (Res == LazyValueInfo::Unknown)
807 continue;
808
809 Constant *ResC = ConstantInt::get(CmpType, Res);
810 Result.push_back(std::make_pair(ResC, P));
811 }
812
813 return !Result.empty();
814 }
815
816 // InstCombine can fold some forms of constant range checks into
817 // (icmp (add (x, C1)), C2). See if we have we have such a thing with
818 // x as a live-in.
819 {
820 using namespace PatternMatch;
821
822 Value *AddLHS;
823 ConstantInt *AddConst;
824 if (isa<ConstantInt>(CmpConst) &&
825 match(CmpLHS, m_Add(m_Value(AddLHS), m_ConstantInt(AddConst)))) {
826 if (!isa<Instruction>(AddLHS) ||
827 cast<Instruction>(AddLHS)->getParent() != BB) {
828 if (DDT->pending())
829 LVI->disableDT();
830 else
831 LVI->enableDT();
832 for (BasicBlock *P : predecessors(BB)) {
833 // If the value is known by LazyValueInfo to be a ConstantRange in
834 // a predecessor, use that information to try to thread this
835 // block.
836 ConstantRange CR = LVI->getConstantRangeOnEdge(
837 AddLHS, P, BB, CxtI ? CxtI : cast<Instruction>(CmpLHS));
838 // Propagate the range through the addition.
839 CR = CR.add(AddConst->getValue());
840
841 // Get the range where the compare returns true.
842 ConstantRange CmpRange = ConstantRange::makeExactICmpRegion(
843 Pred, cast<ConstantInt>(CmpConst)->getValue());
844
845 Constant *ResC;
846 if (CmpRange.contains(CR))
847 ResC = ConstantInt::getTrue(CmpType);
848 else if (CmpRange.inverse().contains(CR))
849 ResC = ConstantInt::getFalse(CmpType);
850 else
851 continue;
852
853 Result.push_back(std::make_pair(ResC, P));
854 }
855
856 return !Result.empty();
857 }
858 }
859 }
860
861 // Try to find a constant value for the LHS of a comparison,
862 // and evaluate it statically if we can.
863 PredValueInfoTy LHSVals;
864 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
865 WantInteger, CxtI);
866
867 for (const auto &LHSVal : LHSVals) {
868 Constant *V = LHSVal.first;
869 Constant *Folded = ConstantExpr::getCompare(Pred, V, CmpConst);
870 if (Constant *KC = getKnownConstant(Folded, WantInteger))
871 Result.push_back(std::make_pair(KC, LHSVal.second));
872 }
873
874 return !Result.empty();
875 }
876 }
877
878 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
879 // Handle select instructions where at least one operand is a known constant
880 // and we can figure out the condition value for any predecessor block.
881 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
882 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
883 PredValueInfoTy Conds;
884 if ((TrueVal || FalseVal) &&
885 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds,
886 WantInteger, CxtI)) {
887 for (auto &C : Conds) {
888 Constant *Cond = C.first;
889
890 // Figure out what value to use for the condition.
891 bool KnownCond;
892 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
893 // A known boolean.
894 KnownCond = CI->isOne();
895 } else {
896 assert(isa<UndefValue>(Cond) && "Unexpected condition value")(static_cast <bool> (isa<UndefValue>(Cond) &&
"Unexpected condition value") ? void (0) : __assert_fail ("isa<UndefValue>(Cond) && \"Unexpected condition value\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 896, __extension__ __PRETTY_FUNCTION__))
;
897 // Either operand will do, so be sure to pick the one that's a known
898 // constant.
899 // FIXME: Do this more cleverly if both values are known constants?
900 KnownCond = (TrueVal != nullptr);
901 }
902
903 // See if the select has a known constant value for this predecessor.
904 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
905 Result.push_back(std::make_pair(Val, C.second));
906 }
907
908 return !Result.empty();
909 }
910 }
911
912 // If all else fails, see if LVI can figure out a constant value for us.
913 if (DDT->pending())
914 LVI->disableDT();
915 else
916 LVI->enableDT();
917 Constant *CI = LVI->getConstant(V, BB, CxtI);
918 if (Constant *KC = getKnownConstant(CI, Preference)) {
919 for (BasicBlock *Pred : predecessors(BB))
920 Result.push_back(std::make_pair(KC, Pred));
921 }
922
923 return !Result.empty();
924}
925
926/// GetBestDestForBranchOnUndef - If we determine that the specified block ends
927/// in an undefined jump, decide which block is best to revector to.
928///
929/// Since we can pick an arbitrary destination, we pick the successor with the
930/// fewest predecessors. This should reduce the in-degree of the others.
931static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
932 TerminatorInst *BBTerm = BB->getTerminator();
933 unsigned MinSucc = 0;
934 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
935 // Compute the successor with the minimum number of predecessors.
936 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
937 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
938 TestBB = BBTerm->getSuccessor(i);
939 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
940 if (NumPreds < MinNumPreds) {
941 MinSucc = i;
942 MinNumPreds = NumPreds;
943 }
944 }
945
946 return MinSucc;
947}
948
949static bool hasAddressTakenAndUsed(BasicBlock *BB) {
950 if (!BB->hasAddressTaken()) return false;
951
952 // If the block has its address taken, it may be a tree of dead constants
953 // hanging off of it. These shouldn't keep the block alive.
954 BlockAddress *BA = BlockAddress::get(BB);
955 BA->removeDeadConstantUsers();
956 return !BA->use_empty();
957}
958
959/// ProcessBlock - If there are any predecessors whose control can be threaded
960/// through to a successor, transform them now.
961bool JumpThreadingPass::ProcessBlock(BasicBlock *BB) {
962 // If the block is trivially dead, just return and let the caller nuke it.
963 // This simplifies other transformations.
964 if (DDT->pendingDeletedBB(BB) ||
1
Assuming the condition is false
965 (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()))
2
Assuming the condition is false
966 return false;
967
968 // If this block has a single predecessor, and if that pred has a single
969 // successor, merge the blocks. This encourages recursive jump threading
970 // because now the condition in this block can be threaded through
971 // predecessors of our predecessor block.
972 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
3
Assuming 'SinglePred' is null
4
Taking false branch
973 const TerminatorInst *TI = SinglePred->getTerminator();
974 if (!TI->isExceptional() && TI->getNumSuccessors() == 1 &&
975 SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
976 // If SinglePred was a loop header, BB becomes one.
977 if (LoopHeaders.erase(SinglePred))
978 LoopHeaders.insert(BB);
979
980 LVI->eraseBlock(SinglePred);
981 MergeBasicBlockIntoOnlyPred(BB, nullptr, DDT);
982
983 // Now that BB is merged into SinglePred (i.e. SinglePred Code followed by
984 // BB code within one basic block `BB`), we need to invalidate the LVI
985 // information associated with BB, because the LVI information need not be
986 // true for all of BB after the merge. For example,
987 // Before the merge, LVI info and code is as follows:
988 // SinglePred: <LVI info1 for %p val>
989 // %y = use of %p
990 // call @exit() // need not transfer execution to successor.
991 // assume(%p) // from this point on %p is true
992 // br label %BB
993 // BB: <LVI info2 for %p val, i.e. %p is true>
994 // %x = use of %p
995 // br label exit
996 //
997 // Note that this LVI info for blocks BB and SinglPred is correct for %p
998 // (info2 and info1 respectively). After the merge and the deletion of the
999 // LVI info1 for SinglePred. We have the following code:
1000 // BB: <LVI info2 for %p val>
1001 // %y = use of %p
1002 // call @exit()
1003 // assume(%p)
1004 // %x = use of %p <-- LVI info2 is correct from here onwards.
1005 // br label exit
1006 // LVI info2 for BB is incorrect at the beginning of BB.
1007
1008 // Invalidate LVI information for BB if the LVI is not provably true for
1009 // all of BB.
1010 if (!isGuaranteedToTransferExecutionToSuccessor(BB))
1011 LVI->eraseBlock(BB);
1012 return true;
1013 }
1014 }
1015
1016 if (TryToUnfoldSelectInCurrBB(BB))
5
Taking false branch
1017 return true;
1018
1019 // Look if we can propagate guards to predecessors.
1020 if (HasGuards && ProcessGuards(BB))
6
Assuming the condition is false
1021 return true;
1022
1023 // What kind of constant we're looking for.
1024 ConstantPreference Preference = WantInteger;
1025
1026 // Look to see if the terminator is a conditional branch, switch or indirect
1027 // branch, if not we can't thread it.
1028 Value *Condition;
1029 Instruction *Terminator = BB->getTerminator();
1030 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
7
Assuming 'BI' is non-null
8
Taking true branch
1031 // Can't thread an unconditional jump.
1032 if (BI->isUnconditional()) return false;
9
Taking false branch
1033 Condition = BI->getCondition();
1034 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
1035 Condition = SI->getCondition();
1036 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
1037 // Can't thread indirect branch with no successors.
1038 if (IB->getNumSuccessors() == 0) return false;
1039 Condition = IB->getAddress()->stripPointerCasts();
1040 Preference = WantBlockAddress;
1041 } else {
1042 return false; // Must be an invoke.
1043 }
1044
1045 // Run constant folding to see if we can reduce the condition to a simple
1046 // constant.
1047 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
10
Assuming 'I' is null
11
Taking false branch
1048 Value *SimpleVal =
1049 ConstantFoldInstruction(I, BB->getModule()->getDataLayout(), TLI);
1050 if (SimpleVal) {
1051 I->replaceAllUsesWith(SimpleVal);
1052 if (isInstructionTriviallyDead(I, TLI))
1053 I->eraseFromParent();
1054 Condition = SimpleVal;
1055 }
1056 }
1057
1058 // If the terminator is branching on an undef, we can pick any of the
1059 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
1060 if (isa<UndefValue>(Condition)) {
12
Taking false branch
1061 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
1062 std::vector<DominatorTree::UpdateType> Updates;
1063
1064 // Fold the branch/switch.
1065 TerminatorInst *BBTerm = BB->getTerminator();
1066 Updates.reserve(BBTerm->getNumSuccessors());
1067 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
1068 if (i == BestSucc) continue;
1069 BasicBlock *Succ = BBTerm->getSuccessor(i);
1070 Succ->removePredecessor(BB, true);
1071 Updates.push_back({DominatorTree::Delete, BB, Succ});
1072 }
1073
1074 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)
1075 << "' 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)
;
1076 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
1077 BBTerm->eraseFromParent();
1078 DDT->applyUpdates(Updates);
1079 return true;
1080 }
1081
1082 // If the terminator of this block is branching on a constant, simplify the
1083 // terminator to an unconditional branch. This can occur due to threading in
1084 // other blocks.
1085 if (getKnownConstant(Condition, Preference)) {
13
Taking false branch
1086 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)
1087 << "' folding terminator: " << *BB->getTerminator() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " In block '" <<
BB->getName() << "' folding terminator: " << *
BB->getTerminator() << '\n'; } } while (false)
;
1088 ++NumFolds;
1089 ConstantFoldTerminator(BB, true, nullptr, DDT);
1090 return true;
1091 }
1092
1093 Instruction *CondInst = dyn_cast<Instruction>(Condition);
1094
1095 // All the rest of our checks depend on the condition being an instruction.
1096 if (!CondInst) {
14
Assuming 'CondInst' is non-null
15
Taking false branch
1097 // FIXME: Unify this with code below.
1098 if (ProcessThreadableEdges(Condition, BB, Preference, Terminator))
1099 return true;
1100 return false;
1101 }
1102
1103 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
16
Taking false branch
1104 // If we're branching on a conditional, LVI might be able to determine
1105 // it's value at the branch instruction. We only handle comparisons
1106 // against a constant at this time.
1107 // TODO: This should be extended to handle switches as well.
1108 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1109 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
1110 if (CondBr && CondConst) {
1111 // We should have returned as soon as we turn a conditional branch to
1112 // unconditional. Because its no longer interesting as far as jump
1113 // threading is concerned.
1114 assert(CondBr->isConditional() && "Threading on unconditional terminator")(static_cast <bool> (CondBr->isConditional() &&
"Threading on unconditional terminator") ? void (0) : __assert_fail
("CondBr->isConditional() && \"Threading on unconditional terminator\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 1114, __extension__ __PRETTY_FUNCTION__))
;
1115
1116 if (DDT->pending())
1117 LVI->disableDT();
1118 else
1119 LVI->enableDT();
1120 LazyValueInfo::Tristate Ret =
1121 LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0),
1122 CondConst, CondBr);
1123 if (Ret != LazyValueInfo::Unknown) {
1124 unsigned ToRemove = Ret == LazyValueInfo::True ? 1 : 0;
1125 unsigned ToKeep = Ret == LazyValueInfo::True ? 0 : 1;
1126 BasicBlock *ToRemoveSucc = CondBr->getSuccessor(ToRemove);
1127 ToRemoveSucc->removePredecessor(BB, true);
1128 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
1129 CondBr->eraseFromParent();
1130 if (CondCmp->use_empty())
1131 CondCmp->eraseFromParent();
1132 // We can safely replace *some* uses of the CondInst if it has
1133 // exactly one value as returned by LVI. RAUW is incorrect in the
1134 // presence of guards and assumes, that have the `Cond` as the use. This
1135 // is because we use the guards/assume to reason about the `Cond` value
1136 // at the end of block, but RAUW unconditionally replaces all uses
1137 // including the guards/assumes themselves and the uses before the
1138 // guard/assume.
1139 else if (CondCmp->getParent() == BB) {
1140 auto *CI = Ret == LazyValueInfo::True ?
1141 ConstantInt::getTrue(CondCmp->getType()) :
1142 ConstantInt::getFalse(CondCmp->getType());
1143 ReplaceFoldableUses(CondCmp, CI);
1144 }
1145 DDT->deleteEdge(BB, ToRemoveSucc);
1146 return true;
1147 }
1148
1149 // We did not manage to simplify this branch, try to see whether
1150 // CondCmp depends on a known phi-select pattern.
1151 if (TryToUnfoldSelect(CondCmp, BB))
1152 return true;
1153 }
1154 }
1155
1156 // Check for some cases that are worth simplifying. Right now we want to look
1157 // for loads that are used by a switch or by the condition for the branch. If
1158 // we see one, check to see if it's partially redundant. If so, insert a PHI
1159 // which can then be used to thread the values.
1160 Value *SimplifyValue = CondInst;
1161 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
17
Taking false branch
1162 if (isa<Constant>(CondCmp->getOperand(1)))
1163 SimplifyValue = CondCmp->getOperand(0);
1164
1165 // TODO: There are other places where load PRE would be profitable, such as
1166 // more complex comparisons.
1167 if (LoadInst *LoadI = dyn_cast<LoadInst>(SimplifyValue))
18
Assuming 'LoadI' is non-null
19
Taking true branch
1168 if (SimplifyPartiallyRedundantLoad(LoadI))
20
Calling 'JumpThreadingPass::SimplifyPartiallyRedundantLoad'
1169 return true;
1170
1171 // Before threading, try to propagate profile data backwards:
1172 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
1173 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1174 updatePredecessorProfileMetadata(PN, BB);
1175
1176 // Handle a variety of cases where we are branching on something derived from
1177 // a PHI node in the current block. If we can prove that any predecessors
1178 // compute a predictable value based on a PHI node, thread those predecessors.
1179 if (ProcessThreadableEdges(CondInst, BB, Preference, Terminator))
1180 return true;
1181
1182 // If this is an otherwise-unfoldable branch on a phi node in the current
1183 // block, see if we can simplify.
1184 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
1185 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1186 return ProcessBranchOnPHI(PN);
1187
1188 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
1189 if (CondInst->getOpcode() == Instruction::Xor &&
1190 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1191 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
1192
1193 // Search for a stronger dominating condition that can be used to simplify a
1194 // conditional branch leaving BB.
1195 if (ProcessImpliedCondition(BB))
1196 return true;
1197
1198 return false;
1199}
1200
1201bool JumpThreadingPass::ProcessImpliedCondition(BasicBlock *BB) {
1202 auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
1203 if (!BI || !BI->isConditional())
1204 return false;
1205
1206 Value *Cond = BI->getCondition();
1207 BasicBlock *CurrentBB = BB;
1208 BasicBlock *CurrentPred = BB->getSinglePredecessor();
1209 unsigned Iter = 0;
1210
1211 auto &DL = BB->getModule()->getDataLayout();
1212
1213 while (CurrentPred && Iter++ < ImplicationSearchThreshold) {
1214 auto *PBI = dyn_cast<BranchInst>(CurrentPred->getTerminator());
1215 if (!PBI || !PBI->isConditional())
1216 return false;
1217 if (PBI->getSuccessor(0) != CurrentBB && PBI->getSuccessor(1) != CurrentBB)
1218 return false;
1219
1220 bool CondIsTrue = PBI->getSuccessor(0) == CurrentBB;
1221 Optional<bool> Implication =
1222 isImpliedCondition(PBI->getCondition(), Cond, DL, CondIsTrue);
1223 if (Implication) {
1224 BasicBlock *KeepSucc = BI->getSuccessor(*Implication ? 0 : 1);
1225 BasicBlock *RemoveSucc = BI->getSuccessor(*Implication ? 1 : 0);
1226 RemoveSucc->removePredecessor(BB);
1227 BranchInst::Create(KeepSucc, BI);
1228 BI->eraseFromParent();
1229 DDT->deleteEdge(BB, RemoveSucc);
1230 return true;
1231 }
1232 CurrentBB = CurrentPred;
1233 CurrentPred = CurrentBB->getSinglePredecessor();
1234 }
1235
1236 return false;
1237}
1238
1239/// Return true if Op is an instruction defined in the given block.
1240static bool isOpDefinedInBlock(Value *Op, BasicBlock *BB) {
1241 if (Instruction *OpInst = dyn_cast<Instruction>(Op))
1242 if (OpInst->getParent() == BB)
1243 return true;
1244 return false;
1245}
1246
1247/// SimplifyPartiallyRedundantLoad - If LoadI is an obviously partially
1248/// redundant load instruction, eliminate it by replacing it with a PHI node.
1249/// This is an important optimization that encourages jump threading, and needs
1250/// to be run interlaced with other jump threading tasks.
1251bool JumpThreadingPass::SimplifyPartiallyRedundantLoad(LoadInst *LoadI) {
1252 // Don't hack volatile and ordered loads.
1253 if (!LoadI->isUnordered()) return false;
21
Taking false branch
1254
1255 // If the load is defined in a block with exactly one predecessor, it can't be
1256 // partially redundant.
1257 BasicBlock *LoadBB = LoadI->getParent();
1258 if (LoadBB->getSinglePredecessor())
22
Assuming the condition is false
23
Taking false branch
1259 return false;
1260
1261 // If the load is defined in an EH pad, it can't be partially redundant,
1262 // because the edges between the invoke and the EH pad cannot have other
1263 // instructions between them.
1264 if (LoadBB->isEHPad())
24
Assuming the condition is false
25
Taking false branch
1265 return false;
1266
1267 Value *LoadedPtr = LoadI->getOperand(0);
1268
1269 // If the loaded operand is defined in the LoadBB and its not a phi,
1270 // it can't be available in predecessors.
1271 if (isOpDefinedInBlock(LoadedPtr, LoadBB) && !isa<PHINode>(LoadedPtr))
1272 return false;
1273
1274 // Scan a few instructions up from the load, to see if it is obviously live at
1275 // the entry to its block.
1276 BasicBlock::iterator BBIt(LoadI);
1277 bool IsLoadCSE;
1278 if (Value *AvailableVal = FindAvailableLoadedValue(
26
Assuming 'AvailableVal' is null
27
Taking false branch
1279 LoadI, LoadBB, BBIt, DefMaxInstsToScan, AA, &IsLoadCSE)) {
1280 // If the value of the load is locally available within the block, just use
1281 // it. This frequently occurs for reg2mem'd allocas.
1282
1283 if (IsLoadCSE) {
1284 LoadInst *NLoadI = cast<LoadInst>(AvailableVal);
1285 combineMetadataForCSE(NLoadI, LoadI);
1286 };
1287
1288 // If the returned value is the load itself, replace with an undef. This can
1289 // only happen in dead loops.
1290 if (AvailableVal == LoadI)
1291 AvailableVal = UndefValue::get(LoadI->getType());
1292 if (AvailableVal->getType() != LoadI->getType())
1293 AvailableVal = CastInst::CreateBitOrPointerCast(
1294 AvailableVal, LoadI->getType(), "", LoadI);
1295 LoadI->replaceAllUsesWith(AvailableVal);
1296 LoadI->eraseFromParent();
1297 return true;
1298 }
1299
1300 // Otherwise, if we scanned the whole block and got to the top of the block,
1301 // we know the block is locally transparent to the load. If not, something
1302 // might clobber its value.
1303 if (BBIt != LoadBB->begin())
28
Taking false branch
1304 return false;
1305
1306 // If all of the loads and stores that feed the value have the same AA tags,
1307 // then we can propagate them onto any newly inserted loads.
1308 AAMDNodes AATags;
1309 LoadI->getAAMetadata(AATags);
1310
1311 SmallPtrSet<BasicBlock*, 8> PredsScanned;
1312
1313 using AvailablePredsTy = SmallVector<std::pair<BasicBlock *, Value *>, 8>;
1314
1315 AvailablePredsTy AvailablePreds;
1316 BasicBlock *OneUnavailablePred = nullptr;
29
'OneUnavailablePred' initialized to a null pointer value
1317 SmallVector<LoadInst*, 8> CSELoads;
1318
1319 // If we got here, the loaded value is transparent through to the start of the
1320 // block. Check to see if it is available in any of the predecessor blocks.
1321 for (BasicBlock *PredBB : predecessors(LoadBB)) {
1322 // If we already scanned this predecessor, skip it.
1323 if (!PredsScanned.insert(PredBB).second)
1324 continue;
1325
1326 BBIt = PredBB->end();
1327 unsigned NumScanedInst = 0;
1328 Value *PredAvailable = nullptr;
1329 // NOTE: We don't CSE load that is volatile or anything stronger than
1330 // unordered, that should have been checked when we entered the function.
1331 assert(LoadI->isUnordered() &&(static_cast <bool> (LoadI->isUnordered() &&
"Attempting to CSE volatile or atomic loads") ? void (0) : __assert_fail
("LoadI->isUnordered() && \"Attempting to CSE volatile or atomic loads\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 1332, __extension__ __PRETTY_FUNCTION__))
1332 "Attempting to CSE volatile or atomic loads")(static_cast <bool> (LoadI->isUnordered() &&
"Attempting to CSE volatile or atomic loads") ? void (0) : __assert_fail
("LoadI->isUnordered() && \"Attempting to CSE volatile or atomic loads\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 1332, __extension__ __PRETTY_FUNCTION__))
;
1333 // If this is a load on a phi pointer, phi-translate it and search
1334 // for available load/store to the pointer in predecessors.
1335 Value *Ptr = LoadedPtr->DoPHITranslation(LoadBB, PredBB);
1336 PredAvailable = FindAvailablePtrLoadStore(
1337 Ptr, LoadI->getType(), LoadI->isAtomic(), PredBB, BBIt,
1338 DefMaxInstsToScan, AA, &IsLoadCSE, &NumScanedInst);
1339
1340 // If PredBB has a single predecessor, continue scanning through the
1341 // single precessor.
1342 BasicBlock *SinglePredBB = PredBB;
1343 while (!PredAvailable && SinglePredBB && BBIt == SinglePredBB->begin() &&
1344 NumScanedInst < DefMaxInstsToScan) {
1345 SinglePredBB = SinglePredBB->getSinglePredecessor();
1346 if (SinglePredBB) {
1347 BBIt = SinglePredBB->end();
1348 PredAvailable = FindAvailablePtrLoadStore(
1349 Ptr, LoadI->getType(), LoadI->isAtomic(), SinglePredBB, BBIt,
1350 (DefMaxInstsToScan - NumScanedInst), AA, &IsLoadCSE,
1351 &NumScanedInst);
1352 }
1353 }
1354
1355 if (!PredAvailable) {
1356 OneUnavailablePred = PredBB;
1357 continue;
1358 }
1359
1360 if (IsLoadCSE)
1361 CSELoads.push_back(cast<LoadInst>(PredAvailable));
1362
1363 // If so, this load is partially redundant. Remember this info so that we
1364 // can create a PHI node.
1365 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
1366 }
1367
1368 // If the loaded value isn't available in any predecessor, it isn't partially
1369 // redundant.
1370 if (AvailablePreds.empty()) return false;
30
Taking false branch
1371
1372 // Okay, the loaded value is available in at least one (and maybe all!)
1373 // predecessors. If the value is unavailable in more than one unique
1374 // predecessor, we want to insert a merge block for those common predecessors.
1375 // This ensures that we only have to insert one reload, thus not increasing
1376 // code size.
1377 BasicBlock *UnavailablePred = nullptr;
1378
1379 // If the value is unavailable in one of predecessors, we will end up
1380 // inserting a new instruction into them. It is only valid if all the
1381 // instructions before LoadI are guaranteed to pass execution to its
1382 // successor, or if LoadI is safe to speculate.
1383 // TODO: If this logic becomes more complex, and we will perform PRE insertion
1384 // farther than to a predecessor, we need to reuse the code from GVN's PRE.
1385 // It requires domination tree analysis, so for this simple case it is an
1386 // overkill.
1387 if (PredsScanned.size() != AvailablePreds.size() &&
1388 !isSafeToSpeculativelyExecute(LoadI))
1389 for (auto I = LoadBB->begin(); &*I != LoadI; ++I)
1390 if (!isGuaranteedToTransferExecutionToSuccessor(&*I))
1391 return false;
1392
1393 // If there is exactly one predecessor where the value is unavailable, the
1394 // already computed 'OneUnavailablePred' block is it. If it ends in an
1395 // unconditional branch, we know that it isn't a critical edge.
1396 if (PredsScanned.size() == AvailablePreds.size()+1 &&
1397 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
31
Called C++ object pointer is null
1398 UnavailablePred = OneUnavailablePred;
1399 } else if (PredsScanned.size() != AvailablePreds.size()) {
1400 // Otherwise, we had multiple unavailable predecessors or we had a critical
1401 // edge from the one.
1402 SmallVector<BasicBlock*, 8> PredsToSplit;
1403 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
1404
1405 for (const auto &AvailablePred : AvailablePreds)
1406 AvailablePredSet.insert(AvailablePred.first);
1407
1408 // Add all the unavailable predecessors to the PredsToSplit list.
1409 for (BasicBlock *P : predecessors(LoadBB)) {
1410 // If the predecessor is an indirect goto, we can't split the edge.
1411 if (isa<IndirectBrInst>(P->getTerminator()))
1412 return false;
1413
1414 if (!AvailablePredSet.count(P))
1415 PredsToSplit.push_back(P);
1416 }
1417
1418 // Split them out to their own block.
1419 UnavailablePred = SplitBlockPreds(LoadBB, PredsToSplit, "thread-pre-split");
1420 }
1421
1422 // If the value isn't available in all predecessors, then there will be
1423 // exactly one where it isn't available. Insert a load on that edge and add
1424 // it to the AvailablePreds list.
1425 if (UnavailablePred) {
1426 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&(static_cast <bool> (UnavailablePred->getTerminator(
)->getNumSuccessors() == 1 && "Can't handle critical edge here!"
) ? void (0) : __assert_fail ("UnavailablePred->getTerminator()->getNumSuccessors() == 1 && \"Can't handle critical edge here!\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 1427, __extension__ __PRETTY_FUNCTION__))
1427 "Can't handle critical edge here!")(static_cast <bool> (UnavailablePred->getTerminator(
)->getNumSuccessors() == 1 && "Can't handle critical edge here!"
) ? void (0) : __assert_fail ("UnavailablePred->getTerminator()->getNumSuccessors() == 1 && \"Can't handle critical edge here!\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 1427, __extension__ __PRETTY_FUNCTION__))
;
1428 LoadInst *NewVal =
1429 new LoadInst(LoadedPtr->DoPHITranslation(LoadBB, UnavailablePred),
1430 LoadI->getName() + ".pr", false, LoadI->getAlignment(),
1431 LoadI->getOrdering(), LoadI->getSyncScopeID(),
1432 UnavailablePred->getTerminator());
1433 NewVal->setDebugLoc(LoadI->getDebugLoc());
1434 if (AATags)
1435 NewVal->setAAMetadata(AATags);
1436
1437 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
1438 }
1439
1440 // Now we know that each predecessor of this block has a value in
1441 // AvailablePreds, sort them for efficient access as we're walking the preds.
1442 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1443
1444 // Create a PHI node at the start of the block for the PRE'd load value.
1445 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
1446 PHINode *PN = PHINode::Create(LoadI->getType(), std::distance(PB, PE), "",
1447 &LoadBB->front());
1448 PN->takeName(LoadI);
1449 PN->setDebugLoc(LoadI->getDebugLoc());
1450
1451 // Insert new entries into the PHI for each predecessor. A single block may
1452 // have multiple entries here.
1453 for (pred_iterator PI = PB; PI != PE; ++PI) {
1454 BasicBlock *P = *PI;
1455 AvailablePredsTy::iterator I =
1456 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1457 std::make_pair(P, (Value*)nullptr));
1458
1459 assert(I != AvailablePreds.end() && I->first == P &&(static_cast <bool> (I != AvailablePreds.end() &&
I->first == P && "Didn't find entry for predecessor!"
) ? void (0) : __assert_fail ("I != AvailablePreds.end() && I->first == P && \"Didn't find entry for predecessor!\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 1460, __extension__ __PRETTY_FUNCTION__))
1460 "Didn't find entry for predecessor!")(static_cast <bool> (I != AvailablePreds.end() &&
I->first == P && "Didn't find entry for predecessor!"
) ? void (0) : __assert_fail ("I != AvailablePreds.end() && I->first == P && \"Didn't find entry for predecessor!\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 1460, __extension__ __PRETTY_FUNCTION__))
;
1461
1462 // If we have an available predecessor but it requires casting, insert the
1463 // cast in the predecessor and use the cast. Note that we have to update the
1464 // AvailablePreds vector as we go so that all of the PHI entries for this
1465 // predecessor use the same bitcast.
1466 Value *&PredV = I->second;
1467 if (PredV->getType() != LoadI->getType())
1468 PredV = CastInst::CreateBitOrPointerCast(PredV, LoadI->getType(), "",
1469 P->getTerminator());
1470
1471 PN->addIncoming(PredV, I->first);
1472 }
1473
1474 for (LoadInst *PredLoadI : CSELoads) {
1475 combineMetadataForCSE(PredLoadI, LoadI);
1476 }
1477
1478 LoadI->replaceAllUsesWith(PN);
1479 LoadI->eraseFromParent();
1480
1481 return true;
1482}
1483
1484/// FindMostPopularDest - The specified list contains multiple possible
1485/// threadable destinations. Pick the one that occurs the most frequently in
1486/// the list.
1487static BasicBlock *
1488FindMostPopularDest(BasicBlock *BB,
1489 const SmallVectorImpl<std::pair<BasicBlock *,
1490 BasicBlock *>> &PredToDestList) {
1491 assert(!PredToDestList.empty())(static_cast <bool> (!PredToDestList.empty()) ? void (0
) : __assert_fail ("!PredToDestList.empty()", "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 1491, __extension__ __PRETTY_FUNCTION__))
;
1492
1493 // Determine popularity. If there are multiple possible destinations, we
1494 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1495 // blocks with known and real destinations to threading undef. We'll handle
1496 // them later if interesting.
1497 DenseMap<BasicBlock*, unsigned> DestPopularity;
1498 for (const auto &PredToDest : PredToDestList)
1499 if (PredToDest.second)
1500 DestPopularity[PredToDest.second]++;
1501
1502 if (DestPopularity.empty())
1503 return nullptr;
1504
1505 // Find the most popular dest.
1506 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1507 BasicBlock *MostPopularDest = DPI->first;
1508 unsigned Popularity = DPI->second;
1509 SmallVector<BasicBlock*, 4> SamePopularity;
1510
1511 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1512 // If the popularity of this entry isn't higher than the popularity we've
1513 // seen so far, ignore it.
1514 if (DPI->second < Popularity)
1515 ; // ignore.
1516 else if (DPI->second == Popularity) {
1517 // If it is the same as what we've seen so far, keep track of it.
1518 SamePopularity.push_back(DPI->first);
1519 } else {
1520 // If it is more popular, remember it.
1521 SamePopularity.clear();
1522 MostPopularDest = DPI->first;
1523 Popularity = DPI->second;
1524 }
1525 }
1526
1527 // Okay, now we know the most popular destination. If there is more than one
1528 // destination, we need to determine one. This is arbitrary, but we need
1529 // to make a deterministic decision. Pick the first one that appears in the
1530 // successor list.
1531 if (!SamePopularity.empty()) {
1532 SamePopularity.push_back(MostPopularDest);
1533 TerminatorInst *TI = BB->getTerminator();
1534 for (unsigned i = 0; ; ++i) {
1535 assert(i != TI->getNumSuccessors() && "Didn't find any successor!")(static_cast <bool> (i != TI->getNumSuccessors() &&
"Didn't find any successor!") ? void (0) : __assert_fail ("i != TI->getNumSuccessors() && \"Didn't find any successor!\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 1535, __extension__ __PRETTY_FUNCTION__))
;
1536
1537 if (!is_contained(SamePopularity, TI->getSuccessor(i)))
1538 continue;
1539
1540 MostPopularDest = TI->getSuccessor(i);
1541 break;
1542 }
1543 }
1544
1545 // Okay, we have finally picked the most popular destination.
1546 return MostPopularDest;
1547}
1548
1549bool JumpThreadingPass::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1550 ConstantPreference Preference,
1551 Instruction *CxtI) {
1552 // If threading this would thread across a loop header, don't even try to
1553 // thread the edge.
1554 if (LoopHeaders.count(BB))
1555 return false;
1556
1557 PredValueInfoTy PredValues;
1558 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference, CxtI))
1559 return false;
1560
1561 assert(!PredValues.empty() &&(static_cast <bool> (!PredValues.empty() && "ComputeValueKnownInPredecessors returned true with no values"
) ? void (0) : __assert_fail ("!PredValues.empty() && \"ComputeValueKnownInPredecessors returned true with no values\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 1562, __extension__ __PRETTY_FUNCTION__))
1562 "ComputeValueKnownInPredecessors returned true with no values")(static_cast <bool> (!PredValues.empty() && "ComputeValueKnownInPredecessors returned true with no values"
) ? void (0) : __assert_fail ("!PredValues.empty() && \"ComputeValueKnownInPredecessors returned true with no values\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 1562, __extension__ __PRETTY_FUNCTION__))
;
1563
1564 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)
1565 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)
1566 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "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)
1567 << *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)
1568 << " 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)
1569 })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)
;
1570
1571 // Decide what we want to thread through. Convert our list of known values to
1572 // a list of known destinations for each pred. This also discards duplicate
1573 // predecessors and keeps track of the undefined inputs (which are represented
1574 // as a null dest in the PredToDestList).
1575 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1576 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1577
1578 BasicBlock *OnlyDest = nullptr;
1579 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1580 Constant *OnlyVal = nullptr;
1581 Constant *MultipleVal = (Constant *)(intptr_t)~0ULL;
1582
1583 unsigned PredWithKnownDest = 0;
1584 for (const auto &PredValue : PredValues) {
1585 BasicBlock *Pred = PredValue.second;
1586 if (!SeenPreds.insert(Pred).second)
1587 continue; // Duplicate predecessor entry.
1588
1589 Constant *Val = PredValue.first;
1590
1591 BasicBlock *DestBB;
1592 if (isa<UndefValue>(Val))
1593 DestBB = nullptr;
1594 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
1595 assert(isa<ConstantInt>(Val) && "Expecting a constant integer")(static_cast <bool> (isa<ConstantInt>(Val) &&
"Expecting a constant integer") ? void (0) : __assert_fail (
"isa<ConstantInt>(Val) && \"Expecting a constant integer\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 1595, __extension__ __PRETTY_FUNCTION__))
;
1596 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1597 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1598 assert(isa<ConstantInt>(Val) && "Expecting a constant integer")(static_cast <bool> (isa<ConstantInt>(Val) &&
"Expecting a constant integer") ? void (0) : __assert_fail (
"isa<ConstantInt>(Val) && \"Expecting a constant integer\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 1598, __extension__ __PRETTY_FUNCTION__))
;
1599 DestBB = SI->findCaseValue(cast<ConstantInt>(Val))->getCaseSuccessor();
1600 } else {
1601 assert(isa<IndirectBrInst>(BB->getTerminator())(static_cast <bool> (isa<IndirectBrInst>(BB->getTerminator
()) && "Unexpected terminator") ? void (0) : __assert_fail
("isa<IndirectBrInst>(BB->getTerminator()) && \"Unexpected terminator\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 1602, __extension__ __PRETTY_FUNCTION__))
1602 && "Unexpected terminator")(static_cast <bool> (isa<IndirectBrInst>(BB->getTerminator
()) && "Unexpected terminator") ? void (0) : __assert_fail
("isa<IndirectBrInst>(BB->getTerminator()) && \"Unexpected terminator\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 1602, __extension__ __PRETTY_FUNCTION__))
;
1603 assert(isa<BlockAddress>(Val) && "Expecting a constant blockaddress")(static_cast <bool> (isa<BlockAddress>(Val) &&
"Expecting a constant blockaddress") ? void (0) : __assert_fail
("isa<BlockAddress>(Val) && \"Expecting a constant blockaddress\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 1603, __extension__ __PRETTY_FUNCTION__))
;
1604 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1605 }
1606
1607 // If we have exactly one destination, remember it for efficiency below.
1608 if (PredToDestList.empty()) {
1609 OnlyDest = DestBB;
1610 OnlyVal = Val;
1611 } else {
1612 if (OnlyDest != DestBB)
1613 OnlyDest = MultipleDestSentinel;
1614 // It possible we have same destination, but different value, e.g. default
1615 // case in switchinst.
1616 if (Val != OnlyVal)
1617 OnlyVal = MultipleVal;
1618 }
1619
1620 // We know where this predecessor is going.
1621 ++PredWithKnownDest;
1622
1623 // If the predecessor ends with an indirect goto, we can't change its
1624 // destination.
1625 if (isa<IndirectBrInst>(Pred->getTerminator()))
1626 continue;
1627
1628 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1629 }
1630
1631 // If all edges were unthreadable, we fail.
1632 if (PredToDestList.empty())
1633 return false;
1634
1635 // If all the predecessors go to a single known successor, we want to fold,
1636 // not thread. By doing so, we do not need to duplicate the current block and
1637 // also miss potential opportunities in case we dont/cant duplicate.
1638 if (OnlyDest && OnlyDest != MultipleDestSentinel) {
1639 if (PredWithKnownDest ==
1640 (size_t)std::distance(pred_begin(BB), pred_end(BB))) {
1641 bool SeenFirstBranchToOnlyDest = false;
1642 std::vector <DominatorTree::UpdateType> Updates;
1643 Updates.reserve(BB->getTerminator()->getNumSuccessors() - 1);
1644 for (BasicBlock *SuccBB : successors(BB)) {
1645 if (SuccBB == OnlyDest && !SeenFirstBranchToOnlyDest) {
1646 SeenFirstBranchToOnlyDest = true; // Don't modify the first branch.
1647 } else {
1648 SuccBB->removePredecessor(BB, true); // This is unreachable successor.
1649 Updates.push_back({DominatorTree::Delete, BB, SuccBB});
1650 }
1651 }
1652
1653 // Finally update the terminator.
1654 TerminatorInst *Term = BB->getTerminator();
1655 BranchInst::Create(OnlyDest, Term);
1656 Term->eraseFromParent();
1657 DDT->applyUpdates(Updates);
1658
1659 // If the condition is now dead due to the removal of the old terminator,
1660 // erase it.
1661 if (auto *CondInst = dyn_cast<Instruction>(Cond)) {
1662 if (CondInst->use_empty() && !CondInst->mayHaveSideEffects())
1663 CondInst->eraseFromParent();
1664 // We can safely replace *some* uses of the CondInst if it has
1665 // exactly one value as returned by LVI. RAUW is incorrect in the
1666 // presence of guards and assumes, that have the `Cond` as the use. This
1667 // is because we use the guards/assume to reason about the `Cond` value
1668 // at the end of block, but RAUW unconditionally replaces all uses
1669 // including the guards/assumes themselves and the uses before the
1670 // guard/assume.
1671 else if (OnlyVal && OnlyVal != MultipleVal &&
1672 CondInst->getParent() == BB)
1673 ReplaceFoldableUses(CondInst, OnlyVal);
1674 }
1675 return true;
1676 }
1677 }
1678
1679 // Determine which is the most common successor. If we have many inputs and
1680 // this block is a switch, we want to start by threading the batch that goes
1681 // to the most popular destination first. If we only know about one
1682 // threadable destination (the common case) we can avoid this.
1683 BasicBlock *MostPopularDest = OnlyDest;
1684
1685 if (MostPopularDest == MultipleDestSentinel) {
1686 // Remove any loop headers from the Dest list, ThreadEdge conservatively
1687 // won't process them, but we might have other destination that are eligible
1688 // and we still want to process.
1689 erase_if(PredToDestList,
1690 [&](const std::pair<BasicBlock *, BasicBlock *> &PredToDest) {
1691 return LoopHeaders.count(PredToDest.second) != 0;
1692 });
1693
1694 if (PredToDestList.empty())
1695 return false;
1696
1697 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1698 }
1699
1700 // Now that we know what the most popular destination is, factor all
1701 // predecessors that will jump to it into a single predecessor.
1702 SmallVector<BasicBlock*, 16> PredsToFactor;
1703 for (const auto &PredToDest : PredToDestList)
1704 if (PredToDest.second == MostPopularDest) {
1705 BasicBlock *Pred = PredToDest.first;
1706
1707 // This predecessor may be a switch or something else that has multiple
1708 // edges to the block. Factor each of these edges by listing them
1709 // according to # occurrences in PredsToFactor.
1710 for (BasicBlock *Succ : successors(Pred))
1711 if (Succ == BB)
1712 PredsToFactor.push_back(Pred);
1713 }
1714
1715 // If the threadable edges are branching on an undefined value, we get to pick
1716 // the destination that these predecessors should get to.
1717 if (!MostPopularDest)
1718 MostPopularDest = BB->getTerminator()->
1719 getSuccessor(GetBestDestForJumpOnUndef(BB));
1720
1721 // Ok, try to thread it!
1722 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1723}
1724
1725/// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1726/// a PHI node in the current block. See if there are any simplifications we
1727/// can do based on inputs to the phi node.
1728bool JumpThreadingPass::ProcessBranchOnPHI(PHINode *PN) {
1729 BasicBlock *BB = PN->getParent();
1730
1731 // TODO: We could make use of this to do it once for blocks with common PHI
1732 // values.
1733 SmallVector<BasicBlock*, 1> PredBBs;
1734 PredBBs.resize(1);
1735
1736 // If any of the predecessor blocks end in an unconditional branch, we can
1737 // *duplicate* the conditional branch into that block in order to further
1738 // encourage jump threading and to eliminate cases where we have branch on a
1739 // phi of an icmp (branch on icmp is much better).
1740 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1741 BasicBlock *PredBB = PN->getIncomingBlock(i);
1742 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1743 if (PredBr->isUnconditional()) {
1744 PredBBs[0] = PredBB;
1745 // Try to duplicate BB into PredBB.
1746 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1747 return true;
1748 }
1749 }
1750
1751 return false;
1752}
1753
1754/// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1755/// a xor instruction in the current block. See if there are any
1756/// simplifications we can do based on inputs to the xor.
1757bool JumpThreadingPass::ProcessBranchOnXOR(BinaryOperator *BO) {
1758 BasicBlock *BB = BO->getParent();
1759
1760 // If either the LHS or RHS of the xor is a constant, don't do this
1761 // optimization.
1762 if (isa<ConstantInt>(BO->getOperand(0)) ||
1763 isa<ConstantInt>(BO->getOperand(1)))
1764 return false;
1765
1766 // If the first instruction in BB isn't a phi, we won't be able to infer
1767 // anything special about any particular predecessor.
1768 if (!isa<PHINode>(BB->front()))
1769 return false;
1770
1771 // If this BB is a landing pad, we won't be able to split the edge into it.
1772 if (BB->isEHPad())
1773 return false;
1774
1775 // If we have a xor as the branch input to this block, and we know that the
1776 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1777 // the condition into the predecessor and fix that value to true, saving some
1778 // logical ops on that path and encouraging other paths to simplify.
1779 //
1780 // This copies something like this:
1781 //
1782 // BB:
1783 // %X = phi i1 [1], [%X']
1784 // %Y = icmp eq i32 %A, %B
1785 // %Z = xor i1 %X, %Y
1786 // br i1 %Z, ...
1787 //
1788 // Into:
1789 // BB':
1790 // %Y = icmp ne i32 %A, %B
1791 // br i1 %Y, ...
1792
1793 PredValueInfoTy XorOpValues;
1794 bool isLHS = true;
1795 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1796 WantInteger, BO)) {
1797 assert(XorOpValues.empty())(static_cast <bool> (XorOpValues.empty()) ? void (0) : __assert_fail
("XorOpValues.empty()", "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 1797, __extension__ __PRETTY_FUNCTION__))
;
1798 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1799 WantInteger, BO))
1800 return false;
1801 isLHS = false;
1802 }
1803
1804 assert(!XorOpValues.empty() &&(static_cast <bool> (!XorOpValues.empty() && "ComputeValueKnownInPredecessors returned true with no values"
) ? void (0) : __assert_fail ("!XorOpValues.empty() && \"ComputeValueKnownInPredecessors returned true with no values\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 1805, __extension__ __PRETTY_FUNCTION__))
1805 "ComputeValueKnownInPredecessors returned true with no values")(static_cast <bool> (!XorOpValues.empty() && "ComputeValueKnownInPredecessors returned true with no values"
) ? void (0) : __assert_fail ("!XorOpValues.empty() && \"ComputeValueKnownInPredecessors returned true with no values\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 1805, __extension__ __PRETTY_FUNCTION__))
;
1806
1807 // Scan the information to see which is most popular: true or false. The
1808 // predecessors can be of the set true, false, or undef.
1809 unsigned NumTrue = 0, NumFalse = 0;
1810 for (const auto &XorOpValue : XorOpValues) {
1811 if (isa<UndefValue>(XorOpValue.first))
1812 // Ignore undefs for the count.
1813 continue;
1814 if (cast<ConstantInt>(XorOpValue.first)->isZero())
1815 ++NumFalse;
1816 else
1817 ++NumTrue;
1818 }
1819
1820 // Determine which value to split on, true, false, or undef if neither.
1821 ConstantInt *SplitVal = nullptr;
1822 if (NumTrue > NumFalse)
1823 SplitVal = ConstantInt::getTrue(BB->getContext());
1824 else if (NumTrue != 0 || NumFalse != 0)
1825 SplitVal = ConstantInt::getFalse(BB->getContext());
1826
1827 // Collect all of the blocks that this can be folded into so that we can
1828 // factor this once and clone it once.
1829 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1830 for (const auto &XorOpValue : XorOpValues) {
1831 if (XorOpValue.first != SplitVal && !isa<UndefValue>(XorOpValue.first))
1832 continue;
1833
1834 BlocksToFoldInto.push_back(XorOpValue.second);
1835 }
1836
1837 // If we inferred a value for all of the predecessors, then duplication won't
1838 // help us. However, we can just replace the LHS or RHS with the constant.
1839 if (BlocksToFoldInto.size() ==
1840 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1841 if (!SplitVal) {
1842 // If all preds provide undef, just nuke the xor, because it is undef too.
1843 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1844 BO->eraseFromParent();
1845 } else if (SplitVal->isZero()) {
1846 // If all preds provide 0, replace the xor with the other input.
1847 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1848 BO->eraseFromParent();
1849 } else {
1850 // If all preds provide 1, set the computed value to 1.
1851 BO->setOperand(!isLHS, SplitVal);
1852 }
1853
1854 return true;
1855 }
1856
1857 // Try to duplicate BB into PredBB.
1858 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1859}
1860
1861/// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1862/// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1863/// NewPred using the entries from OldPred (suitably mapped).
1864static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1865 BasicBlock *OldPred,
1866 BasicBlock *NewPred,
1867 DenseMap<Instruction*, Value*> &ValueMap) {
1868 for (PHINode &PN : PHIBB->phis()) {
1869 // Ok, we have a PHI node. Figure out what the incoming value was for the
1870 // DestBlock.
1871 Value *IV = PN.getIncomingValueForBlock(OldPred);
1872
1873 // Remap the value if necessary.
1874 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1875 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1876 if (I != ValueMap.end())
1877 IV = I->second;
1878 }
1879
1880 PN.addIncoming(IV, NewPred);
1881 }
1882}
1883
1884/// ThreadEdge - We have decided that it is safe and profitable to factor the
1885/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1886/// across BB. Transform the IR to reflect this change.
1887bool JumpThreadingPass::ThreadEdge(BasicBlock *BB,
1888 const SmallVectorImpl<BasicBlock *> &PredBBs,
1889 BasicBlock *SuccBB) {
1890 // If threading to the same block as we come from, we would infinite loop.
1891 if (SuccBB == BB) {
1892 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)
1893 << "' - 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)
;
1894 return false;
1895 }
1896
1897 // If threading this would thread across a loop header, don't thread the edge.
1898 // See the comments above FindLoopHeaders for justifications and caveats.
1899 if (LoopHeaders.count(BB) || LoopHeaders.count(SuccBB)) {
1900 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)
1901 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)
1902 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)
1903 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)
1904 << (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)
1905 << "' 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)
1906 << 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)
1907 })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)
;
1908 return false;
1909 }
1910
1911 unsigned JumpThreadCost =
1912 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
1913 if (JumpThreadCost > BBDupThreshold) {
1914 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)
1915 << "' - 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)
;
1916 return false;
1917 }
1918
1919 // And finally, do it! Start by factoring the predecessors if needed.
1920 BasicBlock *PredBB;
1921 if (PredBBs.size() == 1)
1922 PredBB = PredBBs[0];
1923 else {
1924 DEBUG(dbgs() << " Factoring out " << PredBBs.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Factoring out " <<
PredBBs.size() << " common predecessors.\n"; } } while
(false)
1925 << " common predecessors.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Factoring out " <<
PredBBs.size() << " common predecessors.\n"; } } while
(false)
;
1926 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
1927 }
1928
1929 // And finally, do it!
1930 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"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)
1931 << SuccBB->getName() << "' 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)
1932 << ", across block:\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)
1933 << *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)
;
1934
1935 if (DDT->pending())
1936 LVI->disableDT();
1937 else
1938 LVI->enableDT();
1939 LVI->threadEdge(PredBB, BB, SuccBB);
1940
1941 // We are going to have to map operands from the original BB block to the new
1942 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1943 // account for entry from PredBB.
1944 DenseMap<Instruction*, Value*> ValueMapping;
1945
1946 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1947 BB->getName()+".thread",
1948 BB->getParent(), BB);
1949 NewBB->moveAfter(PredBB);
1950
1951 // Set the block frequency of NewBB.
1952 if (HasProfileData) {
1953 auto NewBBFreq =
1954 BFI->getBlockFreq(PredBB) * BPI->getEdgeProbability(PredBB, BB);
1955 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
1956 }
1957
1958 BasicBlock::iterator BI = BB->begin();
1959 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1960 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1961
1962 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1963 // mapping and using it to remap operands in the cloned instructions.
1964 for (; !isa<TerminatorInst>(BI); ++BI) {
1965 Instruction *New = BI->clone();
1966 New->setName(BI->getName());
1967 NewBB->getInstList().push_back(New);
1968 ValueMapping[&*BI] = New;
1969
1970 // Remap operands to patch up intra-block references.
1971 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1972 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1973 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1974 if (I != ValueMapping.end())
1975 New->setOperand(i, I->second);
1976 }
1977 }
1978
1979 // We didn't copy the terminator from BB over to NewBB, because there is now
1980 // an unconditional jump to SuccBB. Insert the unconditional jump.
1981 BranchInst *NewBI = BranchInst::Create(SuccBB, NewBB);
1982 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
1983
1984 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1985 // PHI nodes for NewBB now.
1986 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1987
1988 // If there were values defined in BB that are used outside the block, then we
1989 // now have to update all uses of the value to use either the original value,
1990 // the cloned value, or some PHI derived value. This can require arbitrary
1991 // PHI insertion, of which we are prepared to do, clean these up now.
1992 SSAUpdater SSAUpdate;
1993 SmallVector<Use*, 16> UsesToRename;
1994 for (Instruction &I : *BB) {
1995 // Scan all uses of this instruction to see if it is used outside of its
1996 // block, and if so, record them in UsesToRename.
1997 for (Use &U : I.uses()) {
1998 Instruction *User = cast<Instruction>(U.getUser());
1999 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
2000 if (UserPN->getIncomingBlock(U) == BB)
2001 continue;
2002 } else if (User->getParent() == BB)
2003 continue;
2004
2005 UsesToRename.push_back(&U);
2006 }
2007
2008 // If there are no uses outside the block, we're done with this instruction.
2009 if (UsesToRename.empty())
2010 continue;
2011
2012 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)
;
2013
2014 // We found a use of I outside of BB. Rename all uses of I that are outside
2015 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
2016 // with the two values we know.
2017 SSAUpdate.Initialize(I.getType(), I.getName());
2018 SSAUpdate.AddAvailableValue(BB, &I);
2019 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[&I]);
2020
2021 while (!UsesToRename.empty())
2022 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
2023 DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "\n"; } } while (false)
;
2024 }
2025
2026 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
2027 // NewBB instead of BB. This eliminates predecessors from BB, which requires
2028 // us to simplify any PHI nodes in BB.
2029 TerminatorInst *PredTerm = PredBB->getTerminator();
2030 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
2031 if (PredTerm->getSuccessor(i) == BB) {
2032 BB->removePredecessor(PredBB, true);
2033 PredTerm->setSuccessor(i, NewBB);
2034 }
2035
2036 DDT->applyUpdates({{DominatorTree::Insert, NewBB, SuccBB},
2037 {DominatorTree::Insert, PredBB, NewBB},
2038 {DominatorTree::Delete, PredBB, BB}});
2039
2040 // At this point, the IR is fully up to date and consistent. Do a quick scan
2041 // over the new instructions and zap any that are constants or dead. This
2042 // frequently happens because of phi translation.
2043 SimplifyInstructionsInBlock(NewBB, TLI);
2044
2045 // Update the edge weight from BB to SuccBB, which should be less than before.
2046 UpdateBlockFreqAndEdgeWeight(PredBB, BB, NewBB, SuccBB);
2047
2048 // Threaded an edge!
2049 ++NumThreads;
2050 return true;
2051}
2052
2053/// Create a new basic block that will be the predecessor of BB and successor of
2054/// all blocks in Preds. When profile data is available, update the frequency of
2055/// this new block.
2056BasicBlock *JumpThreadingPass::SplitBlockPreds(BasicBlock *BB,
2057 ArrayRef<BasicBlock *> Preds,
2058 const char *Suffix) {
2059 SmallVector<BasicBlock *, 2> NewBBs;
2060
2061 // Collect the frequencies of all predecessors of BB, which will be used to
2062 // update the edge weight of the result of splitting predecessors.
2063 DenseMap<BasicBlock *, BlockFrequency> FreqMap;
2064 if (HasProfileData)
2065 for (auto Pred : Preds)
2066 FreqMap.insert(std::make_pair(
2067 Pred, BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, BB)));
2068
2069 // In the case when BB is a LandingPad block we create 2 new predecessors
2070 // instead of just one.
2071 if (BB->isLandingPad()) {
2072 std::string NewName = std::string(Suffix) + ".split-lp";
2073 SplitLandingPadPredecessors(BB, Preds, Suffix, NewName.c_str(), NewBBs);
2074 } else {
2075 NewBBs.push_back(SplitBlockPredecessors(BB, Preds, Suffix));
2076 }
2077
2078 std::vector<DominatorTree::UpdateType> Updates;
2079 Updates.reserve((2 * Preds.size()) + NewBBs.size());
2080 for (auto NewBB : NewBBs) {
2081 BlockFrequency NewBBFreq(0);
2082 Updates.push_back({DominatorTree::Insert, NewBB, BB});
2083 for (auto Pred : predecessors(NewBB)) {
2084 Updates.push_back({DominatorTree::Delete, Pred, BB});
2085 Updates.push_back({DominatorTree::Insert, Pred, NewBB});
2086 if (HasProfileData) // Update frequencies between Pred -> NewBB.
2087 NewBBFreq += FreqMap.lookup(Pred);
2088 }
2089 if (HasProfileData) // Apply the summed frequency to NewBB.
2090 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
2091 }
2092
2093 DDT->applyUpdates(Updates);
2094 return NewBBs[0];
2095}
2096
2097bool JumpThreadingPass::doesBlockHaveProfileData(BasicBlock *BB) {
2098 const TerminatorInst *TI = BB->getTerminator();
2099 assert(TI->getNumSuccessors() > 1 && "not a split")(static_cast <bool> (TI->getNumSuccessors() > 1 &&
"not a split") ? void (0) : __assert_fail ("TI->getNumSuccessors() > 1 && \"not a split\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 2099, __extension__ __PRETTY_FUNCTION__))
;
2100
2101 MDNode *WeightsNode = TI->getMetadata(LLVMContext::MD_prof);
2102 if (!WeightsNode)
2103 return false;
2104
2105 MDString *MDName = cast<MDString>(WeightsNode->getOperand(0));
2106 if (MDName->getString() != "branch_weights")
2107 return false;
2108
2109 // Ensure there are weights for all of the successors. Note that the first
2110 // operand to the metadata node is a name, not a weight.
2111 return WeightsNode->getNumOperands() == TI->getNumSuccessors() + 1;
2112}
2113
2114/// Update the block frequency of BB and branch weight and the metadata on the
2115/// edge BB->SuccBB. This is done by scaling the weight of BB->SuccBB by 1 -
2116/// Freq(PredBB->BB) / Freq(BB->SuccBB).
2117void JumpThreadingPass::UpdateBlockFreqAndEdgeWeight(BasicBlock *PredBB,
2118 BasicBlock *BB,
2119 BasicBlock *NewBB,
2120 BasicBlock *SuccBB) {
2121 if (!HasProfileData)
2122 return;
2123
2124 assert(BFI && BPI && "BFI & BPI should have been created here")(static_cast <bool> (BFI && BPI && "BFI & BPI should have been created here"
) ? void (0) : __assert_fail ("BFI && BPI && \"BFI & BPI should have been created here\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 2124, __extension__ __PRETTY_FUNCTION__))
;
2125
2126 // As the edge from PredBB to BB is deleted, we have to update the block
2127 // frequency of BB.
2128 auto BBOrigFreq = BFI->getBlockFreq(BB);
2129 auto NewBBFreq = BFI->getBlockFreq(NewBB);
2130 auto BB2SuccBBFreq = BBOrigFreq * BPI->getEdgeProbability(BB, SuccBB);
2131 auto BBNewFreq = BBOrigFreq - NewBBFreq;
2132 BFI->setBlockFreq(BB, BBNewFreq.getFrequency());
2133
2134 // Collect updated outgoing edges' frequencies from BB and use them to update
2135 // edge probabilities.
2136 SmallVector<uint64_t, 4> BBSuccFreq;
2137 for (BasicBlock *Succ : successors(BB)) {
2138 auto SuccFreq = (Succ == SuccBB)
2139 ? BB2SuccBBFreq - NewBBFreq
2140 : BBOrigFreq * BPI->getEdgeProbability(BB, Succ);
2141 BBSuccFreq.push_back(SuccFreq.getFrequency());
2142 }
2143
2144 uint64_t MaxBBSuccFreq =
2145 *std::max_element(BBSuccFreq.begin(), BBSuccFreq.end());
2146
2147 SmallVector<BranchProbability, 4> BBSuccProbs;
2148 if (MaxBBSuccFreq == 0)
2149 BBSuccProbs.assign(BBSuccFreq.size(),
2150 {1, static_cast<unsigned>(BBSuccFreq.size())});
2151 else {
2152 for (uint64_t Freq : BBSuccFreq)
2153 BBSuccProbs.push_back(
2154 BranchProbability::getBranchProbability(Freq, MaxBBSuccFreq));
2155 // Normalize edge probabilities so that they sum up to one.
2156 BranchProbability::normalizeProbabilities(BBSuccProbs.begin(),
2157 BBSuccProbs.end());
2158 }
2159
2160 // Update edge probabilities in BPI.
2161 for (int I = 0, E = BBSuccProbs.size(); I < E; I++)
2162 BPI->setEdgeProbability(BB, I, BBSuccProbs[I]);
2163
2164 // Update the profile metadata as well.
2165 //
2166 // Don't do this if the profile of the transformed blocks was statically
2167 // estimated. (This could occur despite the function having an entry
2168 // frequency in completely cold parts of the CFG.)
2169 //
2170 // In this case we don't want to suggest to subsequent passes that the
2171 // calculated weights are fully consistent. Consider this graph:
2172 //
2173 // check_1
2174 // 50% / |
2175 // eq_1 | 50%
2176 // \ |
2177 // check_2
2178 // 50% / |
2179 // eq_2 | 50%
2180 // \ |
2181 // check_3
2182 // 50% / |
2183 // eq_3 | 50%
2184 // \ |
2185 //
2186 // Assuming the blocks check_* all compare the same value against 1, 2 and 3,
2187 // the overall probabilities are inconsistent; the total probability that the
2188 // value is either 1, 2 or 3 is 150%.
2189 //
2190 // As a consequence if we thread eq_1 -> check_2 to check_3, check_2->check_3
2191 // becomes 0%. This is even worse if the edge whose probability becomes 0% is
2192 // the loop exit edge. Then based solely on static estimation we would assume
2193 // the loop was extremely hot.
2194 //
2195 // FIXME this locally as well so that BPI and BFI are consistent as well. We
2196 // shouldn't make edges extremely likely or unlikely based solely on static
2197 // estimation.
2198 if (BBSuccProbs.size() >= 2 && doesBlockHaveProfileData(BB)) {
2199 SmallVector<uint32_t, 4> Weights;
2200 for (auto Prob : BBSuccProbs)
2201 Weights.push_back(Prob.getNumerator());
2202
2203 auto TI = BB->getTerminator();
2204 TI->setMetadata(
2205 LLVMContext::MD_prof,
2206 MDBuilder(TI->getParent()->getContext()).createBranchWeights(Weights));
2207 }
2208}
2209
2210/// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
2211/// to BB which contains an i1 PHI node and a conditional branch on that PHI.
2212/// If we can duplicate the contents of BB up into PredBB do so now, this
2213/// improves the odds that the branch will be on an analyzable instruction like
2214/// a compare.
2215bool JumpThreadingPass::DuplicateCondBranchOnPHIIntoPred(
2216 BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs) {
2217 assert(!PredBBs.empty() && "Can't handle an empty set")(static_cast <bool> (!PredBBs.empty() && "Can't handle an empty set"
) ? void (0) : __assert_fail ("!PredBBs.empty() && \"Can't handle an empty set\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 2217, __extension__ __PRETTY_FUNCTION__))
;
2218
2219 // If BB is a loop header, then duplicating this block outside the loop would
2220 // cause us to transform this into an irreducible loop, don't do this.
2221 // See the comments above FindLoopHeaders for justifications and caveats.
2222 if (LoopHeaders.count(BB)) {
2223 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)
2224 << "' 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)
2225 << "' - 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)
;
2226 return false;
2227 }
2228
2229 unsigned DuplicationCost =
2230 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
2231 if (DuplicationCost > BBDupThreshold) {
2232 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)
2233 << "' - 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)
;
2234 return false;
2235 }
2236
2237 // And finally, do it! Start by factoring the predecessors if needed.
2238 std::vector<DominatorTree::UpdateType> Updates;
2239 BasicBlock *PredBB;
2240 if (PredBBs.size() == 1)
2241 PredBB = PredBBs[0];
2242 else {
2243 DEBUG(dbgs() << " Factoring out " << PredBBs.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Factoring out " <<
PredBBs.size() << " common predecessors.\n"; } } while
(false)
2244 << " common predecessors.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Factoring out " <<
PredBBs.size() << " common predecessors.\n"; } } while
(false)
;
2245 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
2246 }
2247 Updates.push_back({DominatorTree::Delete, PredBB, BB});
2248
2249 // Okay, we decided to do this! Clone all the instructions in BB onto the end
2250 // of PredBB.
2251 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"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)
2252 << PredBB->getName() << "' 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)
2253 << 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)
;
2254
2255 // Unless PredBB ends with an unconditional branch, split the edge so that we
2256 // can just clone the bits from BB into the end of the new PredBB.
2257 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
2258
2259 if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
2260 BasicBlock *OldPredBB = PredBB;
2261 PredBB = SplitEdge(OldPredBB, BB);
2262 Updates.push_back({DominatorTree::Insert, OldPredBB, PredBB});
2263 Updates.push_back({DominatorTree::Insert, PredBB, BB});
2264 Updates.push_back({DominatorTree::Delete, OldPredBB, BB});
2265 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
2266 }
2267
2268 // We are going to have to map operands from the original BB block into the
2269 // PredBB block. Evaluate PHI nodes in BB.
2270 DenseMap<Instruction*, Value*> ValueMapping;
2271
2272 BasicBlock::iterator BI = BB->begin();
2273 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
2274 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
2275 // Clone the non-phi instructions of BB into PredBB, keeping track of the
2276 // mapping and using it to remap operands in the cloned instructions.
2277 for (; BI != BB->end(); ++BI) {
2278 Instruction *New = BI->clone();
2279
2280 // Remap operands to patch up intra-block references.
2281 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2282 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
2283 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
2284 if (I != ValueMapping.end())
2285 New->setOperand(i, I->second);
2286 }
2287
2288 // If this instruction can be simplified after the operands are updated,
2289 // just use the simplified value instead. This frequently happens due to
2290 // phi translation.
2291 if (Value *IV = SimplifyInstruction(
2292 New,
2293 {BB->getModule()->getDataLayout(), TLI, nullptr, nullptr, New})) {
2294 ValueMapping[&*BI] = IV;
2295 if (!New->mayHaveSideEffects()) {
2296 New->deleteValue();
2297 New = nullptr;
2298 }
2299 } else {
2300 ValueMapping[&*BI] = New;
2301 }
2302 if (New) {
2303 // Otherwise, insert the new instruction into the block.
2304 New->setName(BI->getName());
2305 PredBB->getInstList().insert(OldPredBranch->getIterator(), New);
2306 // Update Dominance from simplified New instruction operands.
2307 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2308 if (BasicBlock *SuccBB = dyn_cast<BasicBlock>(New->getOperand(i)))
2309 Updates.push_back({DominatorTree::Insert, PredBB, SuccBB});
2310 }
2311 }
2312
2313 // Check to see if the targets of the branch had PHI nodes. If so, we need to
2314 // add entries to the PHI nodes for branch from PredBB now.
2315 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
2316 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
2317 ValueMapping);
2318 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
2319 ValueMapping);
2320
2321 // If there were values defined in BB that are used outside the block, then we
2322 // now have to update all uses of the value to use either the original value,
2323 // the cloned value, or some PHI derived value. This can require arbitrary
2324 // PHI insertion, of which we are prepared to do, clean these up now.
2325 SSAUpdater SSAUpdate;
2326 SmallVector<Use*, 16> UsesToRename;
2327 for (Instruction &I : *BB) {
2328 // Scan all uses of this instruction to see if it is used outside of its
2329 // block, and if so, record them in UsesToRename.
2330 for (Use &U : I.uses()) {
2331 Instruction *User = cast<Instruction>(U.getUser());
2332 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
2333 if (UserPN->getIncomingBlock(U) == BB)
2334 continue;
2335 } else if (User->getParent() == BB)
2336 continue;
2337
2338 UsesToRename.push_back(&U);
2339 }
2340
2341 // If there are no uses outside the block, we're done with this instruction.
2342 if (UsesToRename.empty())
2343 continue;
2344
2345 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)
;
2346
2347 // We found a use of I outside of BB. Rename all uses of I that are outside
2348 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
2349 // with the two values we know.
2350 SSAUpdate.Initialize(I.getType(), I.getName());
2351 SSAUpdate.AddAvailableValue(BB, &I);
2352 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[&I]);
2353
2354 while (!UsesToRename.empty())
2355 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
2356 DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "\n"; } } while (false)
;
2357 }
2358
2359 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
2360 // that we nuked.
2361 BB->removePredecessor(PredBB, true);
2362
2363 // Remove the unconditional branch at the end of the PredBB block.
2364 OldPredBranch->eraseFromParent();
2365 DDT->applyUpdates(Updates);
2366
2367 ++NumDupes;
2368 return true;
2369}
2370
2371/// TryToUnfoldSelect - Look for blocks of the form
2372/// bb1:
2373/// %a = select
2374/// br bb2
2375///
2376/// bb2:
2377/// %p = phi [%a, %bb1] ...
2378/// %c = icmp %p
2379/// br i1 %c
2380///
2381/// And expand the select into a branch structure if one of its arms allows %c
2382/// to be folded. This later enables threading from bb1 over bb2.
2383bool JumpThreadingPass::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
2384 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
2385 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
2386 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
2387
2388 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
2389 CondLHS->getParent() != BB)
2390 return false;
2391
2392 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
2393 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
2394 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
2395
2396 // Look if one of the incoming values is a select in the corresponding
2397 // predecessor.
2398 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
2399 continue;
2400
2401 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
2402 if (!PredTerm || !PredTerm->isUnconditional())
2403 continue;
2404
2405 // Now check if one of the select values would allow us to constant fold the
2406 // terminator in BB. We don't do the transform if both sides fold, those
2407 // cases will be threaded in any case.
2408 if (DDT->pending())
2409 LVI->disableDT();
2410 else
2411 LVI->enableDT();
2412 LazyValueInfo::Tristate LHSFolds =
2413 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
2414 CondRHS, Pred, BB, CondCmp);
2415 LazyValueInfo::Tristate RHSFolds =
2416 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
2417 CondRHS, Pred, BB, CondCmp);
2418 if ((LHSFolds != LazyValueInfo::Unknown ||
2419 RHSFolds != LazyValueInfo::Unknown) &&
2420 LHSFolds != RHSFolds) {
2421 // Expand the select.
2422 //
2423 // Pred --
2424 // | v
2425 // | NewBB
2426 // | |
2427 // |-----
2428 // v
2429 // BB
2430 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
2431 BB->getParent(), BB);
2432 // Move the unconditional branch to NewBB.
2433 PredTerm->removeFromParent();
2434 NewBB->getInstList().insert(NewBB->end(), PredTerm);
2435 // Create a conditional branch and update PHI nodes.
2436 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
2437 CondLHS->setIncomingValue(I, SI->getFalseValue());
2438 CondLHS->addIncoming(SI->getTrueValue(), NewBB);
2439 // The select is now dead.
2440 SI->eraseFromParent();
2441
2442 DDT->applyUpdates({{DominatorTree::Insert, NewBB, BB},
2443 {DominatorTree::Insert, Pred, NewBB}});
2444 // Update any other PHI nodes in BB.
2445 for (BasicBlock::iterator BI = BB->begin();
2446 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
2447 if (Phi != CondLHS)
2448 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);
2449 return true;
2450 }
2451 }
2452 return false;
2453}
2454
2455/// TryToUnfoldSelectInCurrBB - Look for PHI/Select or PHI/CMP/Select in the
2456/// same BB in the form
2457/// bb:
2458/// %p = phi [false, %bb1], [true, %bb2], [false, %bb3], [true, %bb4], ...
2459/// %s = select %p, trueval, falseval
2460///
2461/// or
2462///
2463/// bb:
2464/// %p = phi [0, %bb1], [1, %bb2], [0, %bb3], [1, %bb4], ...
2465/// %c = cmp %p, 0
2466/// %s = select %c, trueval, falseval
2467///
2468/// And expand the select into a branch structure. This later enables
2469/// jump-threading over bb in this pass.
2470///
2471/// Using the similar approach of SimplifyCFG::FoldCondBranchOnPHI(), unfold
2472/// select if the associated PHI has at least one constant. If the unfolded
2473/// select is not jump-threaded, it will be folded again in the later
2474/// optimizations.
2475bool JumpThreadingPass::TryToUnfoldSelectInCurrBB(BasicBlock *BB) {
2476 // If threading this would thread across a loop header, don't thread the edge.
2477 // See the comments above FindLoopHeaders for justifications and caveats.
2478 if (LoopHeaders.count(BB))
2479 return false;
2480
2481 for (BasicBlock::iterator BI = BB->begin();
2482 PHINode *PN = dyn_cast<PHINode>(BI); ++BI) {
2483 // Look for a Phi having at least one constant incoming value.
2484 if (llvm::all_of(PN->incoming_values(),
2485 [](Value *V) { return !isa<ConstantInt>(V); }))
2486 continue;
2487
2488 auto isUnfoldCandidate = [BB](SelectInst *SI, Value *V) {
2489 // Check if SI is in BB and use V as condition.
2490 if (SI->getParent() != BB)
2491 return false;
2492 Value *Cond = SI->getCondition();
2493 return (Cond && Cond == V && Cond->getType()->isIntegerTy(1));
2494 };
2495
2496 SelectInst *SI = nullptr;
2497 for (Use &U : PN->uses()) {
2498 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(U.getUser())) {
2499 // Look for a ICmp in BB that compares PN with a constant and is the
2500 // condition of a Select.
2501 if (Cmp->getParent() == BB && Cmp->hasOneUse() &&
2502 isa<ConstantInt>(Cmp->getOperand(1 - U.getOperandNo())))
2503 if (SelectInst *SelectI = dyn_cast<SelectInst>(Cmp->user_back()))
2504 if (isUnfoldCandidate(SelectI, Cmp->use_begin()->get())) {
2505 SI = SelectI;
2506 break;
2507 }
2508 } else if (SelectInst *SelectI = dyn_cast<SelectInst>(U.getUser())) {
2509 // Look for a Select in BB that uses PN as condtion.
2510 if (isUnfoldCandidate(SelectI, U.get())) {
2511 SI = SelectI;
2512 break;
2513 }
2514 }
2515 }
2516
2517 if (!SI)
2518 continue;
2519 // Expand the select.
2520 TerminatorInst *Term =
2521 SplitBlockAndInsertIfThen(SI->getCondition(), SI, false);
2522 BasicBlock *SplitBB = SI->getParent();
2523 BasicBlock *NewBB = Term->getParent();
2524 PHINode *NewPN = PHINode::Create(SI->getType(), 2, "", SI);
2525 NewPN->addIncoming(SI->getTrueValue(), Term->getParent());
2526 NewPN->addIncoming(SI->getFalseValue(), BB);
2527 SI->replaceAllUsesWith(NewPN);
2528 SI->eraseFromParent();
2529 // NewBB and SplitBB are newly created blocks which require insertion.
2530 std::vector<DominatorTree::UpdateType> Updates;
2531 Updates.reserve((2 * SplitBB->getTerminator()->getNumSuccessors()) + 3);
2532 Updates.push_back({DominatorTree::Insert, BB, SplitBB});
2533 Updates.push_back({DominatorTree::Insert, BB, NewBB});
2534 Updates.push_back({DominatorTree::Insert, NewBB, SplitBB});
2535 // BB's successors were moved to SplitBB, update DDT accordingly.
2536 for (auto *Succ : successors(SplitBB)) {
2537 Updates.push_back({DominatorTree::Delete, BB, Succ});
2538 Updates.push_back({DominatorTree::Insert, SplitBB, Succ});
2539 }
2540 DDT->applyUpdates(Updates);
2541 return true;
2542 }
2543 return false;
2544}
2545
2546/// Try to propagate a guard from the current BB into one of its predecessors
2547/// in case if another branch of execution implies that the condition of this
2548/// guard is always true. Currently we only process the simplest case that
2549/// looks like:
2550///
2551/// Start:
2552/// %cond = ...
2553/// br i1 %cond, label %T1, label %F1
2554/// T1:
2555/// br label %Merge
2556/// F1:
2557/// br label %Merge
2558/// Merge:
2559/// %condGuard = ...
2560/// call void(i1, ...) @llvm.experimental.guard( i1 %condGuard )[ "deopt"() ]
2561///
2562/// And cond either implies condGuard or !condGuard. In this case all the
2563/// instructions before the guard can be duplicated in both branches, and the
2564/// guard is then threaded to one of them.
2565bool JumpThreadingPass::ProcessGuards(BasicBlock *BB) {
2566 using namespace PatternMatch;
2567
2568 // We only want to deal with two predecessors.
2569 BasicBlock *Pred1, *Pred2;
2570 auto PI = pred_begin(BB), PE = pred_end(BB);
2571 if (PI == PE)
2572 return false;
2573 Pred1 = *PI++;
2574 if (PI == PE)
2575 return false;
2576 Pred2 = *PI++;
2577 if (PI != PE)
2578 return false;
2579 if (Pred1 == Pred2)
2580 return false;
2581
2582 // Try to thread one of the guards of the block.
2583 // TODO: Look up deeper than to immediate predecessor?
2584 auto *Parent = Pred1->getSinglePredecessor();
2585 if (!Parent || Parent != Pred2->getSinglePredecessor())
2586 return false;
2587
2588 if (auto *BI = dyn_cast<BranchInst>(Parent->getTerminator()))
2589 for (auto &I : *BB)
2590 if (match(&I, m_Intrinsic<Intrinsic::experimental_guard>()))
2591 if (ThreadGuard(BB, cast<IntrinsicInst>(&I), BI))
2592 return true;
2593
2594 return false;
2595}
2596
2597/// Try to propagate the guard from BB which is the lower block of a diamond
2598/// to one of its branches, in case if diamond's condition implies guard's
2599/// condition.
2600bool JumpThreadingPass::ThreadGuard(BasicBlock *BB, IntrinsicInst *Guard,
2601 BranchInst *BI) {
2602 assert(BI->getNumSuccessors() == 2 && "Wrong number of successors?")(static_cast <bool> (BI->getNumSuccessors() == 2 &&
"Wrong number of successors?") ? void (0) : __assert_fail ("BI->getNumSuccessors() == 2 && \"Wrong number of successors?\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 2602, __extension__ __PRETTY_FUNCTION__))
;
2603 assert(BI->isConditional() && "Unconditional branch has 2 successors?")(static_cast <bool> (BI->isConditional() && "Unconditional branch has 2 successors?"
) ? void (0) : __assert_fail ("BI->isConditional() && \"Unconditional branch has 2 successors?\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 2603, __extension__ __PRETTY_FUNCTION__))
;
2604 Value *GuardCond = Guard->getArgOperand(0);
2605 Value *BranchCond = BI->getCondition();
2606 BasicBlock *TrueDest = BI->getSuccessor(0);
2607 BasicBlock *FalseDest = BI->getSuccessor(1);
2608
2609 auto &DL = BB->getModule()->getDataLayout();
2610 bool TrueDestIsSafe = false;
2611 bool FalseDestIsSafe = false;
2612
2613 // True dest is safe if BranchCond => GuardCond.
2614 auto Impl = isImpliedCondition(BranchCond, GuardCond, DL);
2615 if (Impl && *Impl)
2616 TrueDestIsSafe = true;
2617 else {
2618 // False dest is safe if !BranchCond => GuardCond.
2619 Impl = isImpliedCondition(BranchCond, GuardCond, DL, /* LHSIsTrue */ false);
2620 if (Impl && *Impl)
2621 FalseDestIsSafe = true;
2622 }
2623
2624 if (!TrueDestIsSafe && !FalseDestIsSafe)
2625 return false;
2626
2627 BasicBlock *PredUnguardedBlock = TrueDestIsSafe ? TrueDest : FalseDest;
2628 BasicBlock *PredGuardedBlock = FalseDestIsSafe ? TrueDest : FalseDest;
2629
2630 ValueToValueMapTy UnguardedMapping, GuardedMapping;
2631 Instruction *AfterGuard = Guard->getNextNode();
2632 unsigned Cost = getJumpThreadDuplicationCost(BB, AfterGuard, BBDupThreshold);
2633 if (Cost > BBDupThreshold)
2634 return false;
2635 // Duplicate all instructions before the guard and the guard itself to the
2636 // branch where implication is not proved.
2637 BasicBlock *GuardedBlock = DuplicateInstructionsInSplitBetween(
2638 BB, PredGuardedBlock, AfterGuard, GuardedMapping);
2639 assert(GuardedBlock && "Could not create the guarded block?")(static_cast <bool> (GuardedBlock && "Could not create the guarded block?"
) ? void (0) : __assert_fail ("GuardedBlock && \"Could not create the guarded block?\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 2639, __extension__ __PRETTY_FUNCTION__))
;
2640 // Duplicate all instructions before the guard in the unguarded branch.
2641 // Since we have successfully duplicated the guarded block and this block
2642 // has fewer instructions, we expect it to succeed.
2643 BasicBlock *UnguardedBlock = DuplicateInstructionsInSplitBetween(
2644 BB, PredUnguardedBlock, Guard, UnguardedMapping);
2645 assert(UnguardedBlock && "Could not create the unguarded block?")(static_cast <bool> (UnguardedBlock && "Could not create the unguarded block?"
) ? void (0) : __assert_fail ("UnguardedBlock && \"Could not create the unguarded block?\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 2645, __extension__ __PRETTY_FUNCTION__))
;
2646 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)
2647 << GuardedBlock->getName() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "Moved guard " <<
*Guard << " to block " << GuardedBlock->getName
() << "\n"; } } while (false)
;
2648 // DuplicateInstructionsInSplitBetween inserts a new block "BB.split" between
2649 // PredBB and BB. We need to perform two inserts and one delete for each of
2650 // the above calls to update Dominators.
2651 DDT->applyUpdates(
2652 {// Guarded block split.
2653 {DominatorTree::Delete, PredGuardedBlock, BB},
2654 {DominatorTree::Insert, PredGuardedBlock, GuardedBlock},
2655 {DominatorTree::Insert, GuardedBlock, BB},
2656 // Unguarded block split.
2657 {DominatorTree::Delete, PredUnguardedBlock, BB},
2658 {DominatorTree::Insert, PredUnguardedBlock, UnguardedBlock},
2659 {DominatorTree::Insert, UnguardedBlock, BB}});
2660 // Some instructions before the guard may still have uses. For them, we need
2661 // to create Phi nodes merging their copies in both guarded and unguarded
2662 // branches. Those instructions that have no uses can be just removed.
2663 SmallVector<Instruction *, 4> ToRemove;
2664 for (auto BI = BB->begin(); &*BI != AfterGuard; ++BI)
2665 if (!isa<PHINode>(&*BI))
2666 ToRemove.push_back(&*BI);
2667
2668 Instruction *InsertionPoint = &*BB->getFirstInsertionPt();
2669 assert(InsertionPoint && "Empty block?")(static_cast <bool> (InsertionPoint && "Empty block?"
) ? void (0) : __assert_fail ("InsertionPoint && \"Empty block?\""
, "/build/llvm-toolchain-snapshot-7~svn329677/lib/Transforms/Scalar/JumpThreading.cpp"
, 2669, __extension__ __PRETTY_FUNCTION__))
;
2670 // Substitute with Phis & remove.
2671 for (auto *Inst : reverse(ToRemove)) {
2672 if (!Inst->use_empty()) {
2673 PHINode *NewPN = PHINode::Create(Inst->getType(), 2);
2674 NewPN->addIncoming(UnguardedMapping[Inst], UnguardedBlock);
2675 NewPN->addIncoming(GuardedMapping[Inst], GuardedBlock);
2676 NewPN->insertBefore(InsertionPoint);
2677 Inst->replaceAllUsesWith(NewPN);
2678 }
2679 Inst->eraseFromParent();
2680 }
2681 return true;
2682}