Bug Summary

File:llvm/lib/Transforms/Scalar/JumpThreading.cpp
Warning:line 1449, column 7
Called C++ object pointer is null

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name JumpThreading.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -fhalf-no-semantic-interposition -mframe-pointer=none -fmath-errno -fno-rounding-math -mconstructor-aliases -munwind-tables -target-cpu x86-64 -tune-cpu generic -fno-split-dwarf-inlining -debugger-tuning=gdb -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-12/lib/clang/12.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/build-llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/build-llvm/include -I /build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0/backward -internal-isystem /usr/local/include -internal-isystem /usr/lib/llvm-12/lib/clang/12.0.0/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++14 -fdeprecated-macro -fdebug-compilation-dir /build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/build-llvm/lib/Transforms/Scalar -fdebug-prefix-map=/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5=. -ferror-limit 19 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -o /tmp/scan-build-2021-01-16-002530-32805-1 -x c++ /build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp

/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp

1//===- JumpThreading.cpp - Thread control through conditional blocks ------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file implements the Jump Threading pass.
10//
11//===----------------------------------------------------------------------===//
12
13#include "llvm/Transforms/Scalar/JumpThreading.h"
14#include "llvm/ADT/DenseMap.h"
15#include "llvm/ADT/DenseSet.h"
16#include "llvm/ADT/MapVector.h"
17#include "llvm/ADT/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/DomTreeUpdater.h"
28#include "llvm/Analysis/GlobalsModRef.h"
29#include "llvm/Analysis/GuardUtils.h"
30#include "llvm/Analysis/InstructionSimplify.h"
31#include "llvm/Analysis/LazyValueInfo.h"
32#include "llvm/Analysis/Loads.h"
33#include "llvm/Analysis/LoopInfo.h"
34#include "llvm/Analysis/TargetLibraryInfo.h"
35#include "llvm/Analysis/TargetTransformInfo.h"
36#include "llvm/Analysis/ValueTracking.h"
37#include "llvm/IR/BasicBlock.h"
38#include "llvm/IR/CFG.h"
39#include "llvm/IR/Constant.h"
40#include "llvm/IR/ConstantRange.h"
41#include "llvm/IR/Constants.h"
42#include "llvm/IR/DataLayout.h"
43#include "llvm/IR/Dominators.h"
44#include "llvm/IR/Function.h"
45#include "llvm/IR/InstrTypes.h"
46#include "llvm/IR/Instruction.h"
47#include "llvm/IR/Instructions.h"
48#include "llvm/IR/IntrinsicInst.h"
49#include "llvm/IR/Intrinsics.h"
50#include "llvm/IR/LLVMContext.h"
51#include "llvm/IR/MDBuilder.h"
52#include "llvm/IR/Metadata.h"
53#include "llvm/IR/Module.h"
54#include "llvm/IR/PassManager.h"
55#include "llvm/IR/PatternMatch.h"
56#include "llvm/IR/Type.h"
57#include "llvm/IR/Use.h"
58#include "llvm/IR/User.h"
59#include "llvm/IR/Value.h"
60#include "llvm/InitializePasses.h"
61#include "llvm/Pass.h"
62#include "llvm/Support/BlockFrequency.h"
63#include "llvm/Support/BranchProbability.h"
64#include "llvm/Support/Casting.h"
65#include "llvm/Support/CommandLine.h"
66#include "llvm/Support/Debug.h"
67#include "llvm/Support/raw_ostream.h"
68#include "llvm/Transforms/Scalar.h"
69#include "llvm/Transforms/Utils/BasicBlockUtils.h"
70#include "llvm/Transforms/Utils/Cloning.h"
71#include "llvm/Transforms/Utils/Local.h"
72#include "llvm/Transforms/Utils/SSAUpdater.h"
73#include "llvm/Transforms/Utils/ValueMapper.h"
74#include <algorithm>
75#include <cassert>
76#include <cstddef>
77#include <cstdint>
78#include <iterator>
79#include <memory>
80#include <utility>
81
82using namespace llvm;
83using namespace jumpthreading;
84
85#define DEBUG_TYPE"jump-threading" "jump-threading"
86
87STATISTIC(NumThreads, "Number of jumps threaded")static llvm::Statistic NumThreads = {"jump-threading", "NumThreads"
, "Number of jumps threaded"}
;
88STATISTIC(NumFolds, "Number of terminators folded")static llvm::Statistic NumFolds = {"jump-threading", "NumFolds"
, "Number of terminators folded"}
;
89STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi")static llvm::Statistic NumDupes = {"jump-threading", "NumDupes"
, "Number of branch blocks duplicated to eliminate phi"}
;
90
91static cl::opt<unsigned>
92BBDuplicateThreshold("jump-threading-threshold",
93 cl::desc("Max block size to duplicate for jump threading"),
94 cl::init(6), cl::Hidden);
95
96static cl::opt<unsigned>
97ImplicationSearchThreshold(
98 "jump-threading-implication-search-threshold",
99 cl::desc("The number of predecessors to search for a stronger "
100 "condition to use to thread over a weaker condition"),
101 cl::init(3), cl::Hidden);
102
103static cl::opt<bool> PrintLVIAfterJumpThreading(
104 "print-lvi-after-jump-threading",
105 cl::desc("Print the LazyValueInfo cache after JumpThreading"), cl::init(false),
106 cl::Hidden);
107
108static cl::opt<bool> JumpThreadingFreezeSelectCond(
109 "jump-threading-freeze-select-cond",
110 cl::desc("Freeze the condition when unfolding select"), cl::init(false),
111 cl::Hidden);
112
113static cl::opt<bool> ThreadAcrossLoopHeaders(
114 "jump-threading-across-loop-headers",
115 cl::desc("Allow JumpThreading to thread across loop headers, for testing"),
116 cl::init(false), cl::Hidden);
117
118
119namespace {
120
121 /// This pass performs 'jump threading', which looks at blocks that have
122 /// multiple predecessors and multiple successors. If one or more of the
123 /// predecessors of the block can be proven to always jump to one of the
124 /// successors, we forward the edge from the predecessor to the successor by
125 /// duplicating the contents of this block.
126 ///
127 /// An example of when this can occur is code like this:
128 ///
129 /// if () { ...
130 /// X = 4;
131 /// }
132 /// if (X < 3) {
133 ///
134 /// In this case, the unconditional branch at the end of the first if can be
135 /// revectored to the false side of the second if.
136 class JumpThreading : public FunctionPass {
137 JumpThreadingPass Impl;
138
139 public:
140 static char ID; // Pass identification
141
142 JumpThreading(bool InsertFreezeWhenUnfoldingSelect = false, int T = -1)
143 : FunctionPass(ID), Impl(InsertFreezeWhenUnfoldingSelect, T) {
144 initializeJumpThreadingPass(*PassRegistry::getPassRegistry());
145 }
146
147 bool runOnFunction(Function &F) override;
148
149 void getAnalysisUsage(AnalysisUsage &AU) const override {
150 AU.addRequired<DominatorTreeWrapperPass>();
151 AU.addPreserved<DominatorTreeWrapperPass>();
152 AU.addRequired<AAResultsWrapperPass>();
153 AU.addRequired<LazyValueInfoWrapperPass>();
154 AU.addPreserved<LazyValueInfoWrapperPass>();
155 AU.addPreserved<GlobalsAAWrapperPass>();
156 AU.addRequired<TargetLibraryInfoWrapperPass>();
157 AU.addRequired<TargetTransformInfoWrapperPass>();
158 }
159
160 void releaseMemory() override { Impl.releaseMemory(); }
161 };
162
163} // end anonymous namespace
164
165char JumpThreading::ID = 0;
166
167INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",static void *initializeJumpThreadingPassOnce(PassRegistry &
Registry) {
168 "Jump Threading", false, false)static void *initializeJumpThreadingPassOnce(PassRegistry &
Registry) {
169INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry);
170INITIALIZE_PASS_DEPENDENCY(LazyValueInfoWrapperPass)initializeLazyValueInfoWrapperPassPass(Registry);
171INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry);
172INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry);
173INITIALIZE_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)); }
174 "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)); }
175
176// Public interface to the Jump Threading pass
177FunctionPass *llvm::createJumpThreadingPass(bool InsertFr, int Threshold) {
178 return new JumpThreading(InsertFr, Threshold);
179}
180
181JumpThreadingPass::JumpThreadingPass(bool InsertFr, int T) {
182 InsertFreezeWhenUnfoldingSelect = JumpThreadingFreezeSelectCond | InsertFr;
183 DefaultBBDupThreshold = (T == -1) ? BBDuplicateThreshold : unsigned(T);
184}
185
186// Update branch probability information according to conditional
187// branch probability. This is usually made possible for cloned branches
188// in inline instances by the context specific profile in the caller.
189// For instance,
190//
191// [Block PredBB]
192// [Branch PredBr]
193// if (t) {
194// Block A;
195// } else {
196// Block B;
197// }
198//
199// [Block BB]
200// cond = PN([true, %A], [..., %B]); // PHI node
201// [Branch CondBr]
202// if (cond) {
203// ... // P(cond == true) = 1%
204// }
205//
206// Here we know that when block A is taken, cond must be true, which means
207// P(cond == true | A) = 1
208//
209// Given that P(cond == true) = P(cond == true | A) * P(A) +
210// P(cond == true | B) * P(B)
211// we get:
212// P(cond == true ) = P(A) + P(cond == true | B) * P(B)
213//
214// which gives us:
215// P(A) is less than P(cond == true), i.e.
216// P(t == true) <= P(cond == true)
217//
218// In other words, if we know P(cond == true) is unlikely, we know
219// that P(t == true) is also unlikely.
220//
221static void updatePredecessorProfileMetadata(PHINode *PN, BasicBlock *BB) {
222 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
223 if (!CondBr)
224 return;
225
226 uint64_t TrueWeight, FalseWeight;
227 if (!CondBr->extractProfMetadata(TrueWeight, FalseWeight))
228 return;
229
230 if (TrueWeight + FalseWeight == 0)
231 // Zero branch_weights do not give a hint for getting branch probabilities.
232 // Technically it would result in division by zero denominator, which is
233 // TrueWeight + FalseWeight.
234 return;
235
236 // Returns the outgoing edge of the dominating predecessor block
237 // that leads to the PhiNode's incoming block:
238 auto GetPredOutEdge =
239 [](BasicBlock *IncomingBB,
240 BasicBlock *PhiBB) -> std::pair<BasicBlock *, BasicBlock *> {
241 auto *PredBB = IncomingBB;
242 auto *SuccBB = PhiBB;
243 SmallPtrSet<BasicBlock *, 16> Visited;
244 while (true) {
245 BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator());
246 if (PredBr && PredBr->isConditional())
247 return {PredBB, SuccBB};
248 Visited.insert(PredBB);
249 auto *SinglePredBB = PredBB->getSinglePredecessor();
250 if (!SinglePredBB)
251 return {nullptr, nullptr};
252
253 // Stop searching when SinglePredBB has been visited. It means we see
254 // an unreachable loop.
255 if (Visited.count(SinglePredBB))
256 return {nullptr, nullptr};
257
258 SuccBB = PredBB;
259 PredBB = SinglePredBB;
260 }
261 };
262
263 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
264 Value *PhiOpnd = PN->getIncomingValue(i);
265 ConstantInt *CI = dyn_cast<ConstantInt>(PhiOpnd);
266
267 if (!CI || !CI->getType()->isIntegerTy(1))
268 continue;
269
270 BranchProbability BP =
271 (CI->isOne() ? BranchProbability::getBranchProbability(
272 TrueWeight, TrueWeight + FalseWeight)
273 : BranchProbability::getBranchProbability(
274 FalseWeight, TrueWeight + FalseWeight));
275
276 auto PredOutEdge = GetPredOutEdge(PN->getIncomingBlock(i), BB);
277 if (!PredOutEdge.first)
278 return;
279
280 BasicBlock *PredBB = PredOutEdge.first;
281 BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator());
282 if (!PredBr)
283 return;
284
285 uint64_t PredTrueWeight, PredFalseWeight;
286 // FIXME: We currently only set the profile data when it is missing.
287 // With PGO, this can be used to refine even existing profile data with
288 // context information. This needs to be done after more performance
289 // testing.
290 if (PredBr->extractProfMetadata(PredTrueWeight, PredFalseWeight))
291 continue;
292
293 // We can not infer anything useful when BP >= 50%, because BP is the
294 // upper bound probability value.
295 if (BP >= BranchProbability(50, 100))
296 continue;
297
298 SmallVector<uint32_t, 2> Weights;
299 if (PredBr->getSuccessor(0) == PredOutEdge.second) {
300 Weights.push_back(BP.getNumerator());
301 Weights.push_back(BP.getCompl().getNumerator());
302 } else {
303 Weights.push_back(BP.getCompl().getNumerator());
304 Weights.push_back(BP.getNumerator());
305 }
306 PredBr->setMetadata(LLVMContext::MD_prof,
307 MDBuilder(PredBr->getParent()->getContext())
308 .createBranchWeights(Weights));
309 }
310}
311
312/// runOnFunction - Toplevel algorithm.
313bool JumpThreading::runOnFunction(Function &F) {
314 if (skipFunction(F))
315 return false;
316 auto TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
317 // Jump Threading has no sense for the targets with divergent CF
318 if (TTI->hasBranchDivergence())
319 return false;
320 auto TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
321 auto DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
322 auto LVI = &getAnalysis<LazyValueInfoWrapperPass>().getLVI();
323 auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
324 DomTreeUpdater DTU(*DT, DomTreeUpdater::UpdateStrategy::Lazy);
325 std::unique_ptr<BlockFrequencyInfo> BFI;
326 std::unique_ptr<BranchProbabilityInfo> BPI;
327 if (F.hasProfileData()) {
328 LoopInfo LI{DominatorTree(F)};
329 BPI.reset(new BranchProbabilityInfo(F, LI, TLI));
330 BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
331 }
332
333 bool Changed = Impl.runImpl(F, TLI, LVI, AA, &DTU, F.hasProfileData(),
334 std::move(BFI), std::move(BPI));
335 if (PrintLVIAfterJumpThreading) {
336 dbgs() << "LVI for function '" << F.getName() << "':\n";
337 LVI->printLVI(F, DTU.getDomTree(), dbgs());
338 }
339 return Changed;
340}
341
342PreservedAnalyses JumpThreadingPass::run(Function &F,
343 FunctionAnalysisManager &AM) {
344 auto &TTI = AM.getResult<TargetIRAnalysis>(F);
345 // Jump Threading has no sense for the targets with divergent CF
346 if (TTI.hasBranchDivergence())
347 return PreservedAnalyses::all();
348 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
349 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
350 auto &LVI = AM.getResult<LazyValueAnalysis>(F);
351 auto &AA = AM.getResult<AAManager>(F);
352 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Lazy);
353
354 std::unique_ptr<BlockFrequencyInfo> BFI;
355 std::unique_ptr<BranchProbabilityInfo> BPI;
356 if (F.hasProfileData()) {
357 LoopInfo LI{DominatorTree(F)};
358 BPI.reset(new BranchProbabilityInfo(F, LI, &TLI));
359 BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
360 }
361
362 bool Changed = runImpl(F, &TLI, &LVI, &AA, &DTU, F.hasProfileData(),
363 std::move(BFI), std::move(BPI));
364
365 if (PrintLVIAfterJumpThreading) {
366 dbgs() << "LVI for function '" << F.getName() << "':\n";
367 LVI.printLVI(F, DTU.getDomTree(), dbgs());
368 }
369
370 if (!Changed)
371 return PreservedAnalyses::all();
372 PreservedAnalyses PA;
373 PA.preserve<GlobalsAA>();
374 PA.preserve<DominatorTreeAnalysis>();
375 PA.preserve<LazyValueAnalysis>();
376 return PA;
377}
378
379bool JumpThreadingPass::runImpl(Function &F, TargetLibraryInfo *TLI_,
380 LazyValueInfo *LVI_, AliasAnalysis *AA_,
381 DomTreeUpdater *DTU_, bool HasProfileData_,
382 std::unique_ptr<BlockFrequencyInfo> BFI_,
383 std::unique_ptr<BranchProbabilityInfo> BPI_) {
384 LLVM_DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "Jump threading on function '"
<< F.getName() << "'\n"; } } while (false)
;
385 TLI = TLI_;
386 LVI = LVI_;
387 AA = AA_;
388 DTU = DTU_;
389 BFI.reset();
390 BPI.reset();
391 // When profile data is available, we need to update edge weights after
392 // successful jump threading, which requires both BPI and BFI being available.
393 HasProfileData = HasProfileData_;
394 auto *GuardDecl = F.getParent()->getFunction(
395 Intrinsic::getName(Intrinsic::experimental_guard));
396 HasGuards = GuardDecl && !GuardDecl->use_empty();
397 if (HasProfileData) {
398 BPI = std::move(BPI_);
399 BFI = std::move(BFI_);
400 }
401
402 // Reduce the number of instructions duplicated when optimizing strictly for
403 // size.
404 if (BBDuplicateThreshold.getNumOccurrences())
405 BBDupThreshold = BBDuplicateThreshold;
406 else if (F.hasFnAttribute(Attribute::MinSize))
407 BBDupThreshold = 3;
408 else
409 BBDupThreshold = DefaultBBDupThreshold;
410
411 // JumpThreading must not processes blocks unreachable from entry. It's a
412 // waste of compute time and can potentially lead to hangs.
413 SmallPtrSet<BasicBlock *, 16> Unreachable;
414 assert(DTU && "DTU isn't passed into JumpThreading before using it.")((DTU && "DTU isn't passed into JumpThreading before using it."
) ? static_cast<void> (0) : __assert_fail ("DTU && \"DTU isn't passed into JumpThreading before using it.\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 414, __PRETTY_FUNCTION__))
;
415 assert(DTU->hasDomTree() && "JumpThreading relies on DomTree to proceed.")((DTU->hasDomTree() && "JumpThreading relies on DomTree to proceed."
) ? static_cast<void> (0) : __assert_fail ("DTU->hasDomTree() && \"JumpThreading relies on DomTree to proceed.\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 415, __PRETTY_FUNCTION__))
;
416 DominatorTree &DT = DTU->getDomTree();
417 for (auto &BB : F)
418 if (!DT.isReachableFromEntry(&BB))
419 Unreachable.insert(&BB);
420
421 if (!ThreadAcrossLoopHeaders)
422 findLoopHeaders(F);
423
424 bool EverChanged = false;
425 bool Changed;
426 do {
427 Changed = false;
428 for (auto &BB : F) {
429 if (Unreachable.count(&BB))
430 continue;
431 while (processBlock(&BB)) // Thread all of the branches we can over BB.
432 Changed = true;
433
434 // Jump threading may have introduced redundant debug values into BB
435 // which should be removed.
436 if (Changed)
437 RemoveRedundantDbgInstrs(&BB);
438
439 // Stop processing BB if it's the entry or is now deleted. The following
440 // routines attempt to eliminate BB and locating a suitable replacement
441 // for the entry is non-trivial.
442 if (&BB == &F.getEntryBlock() || DTU->isBBPendingDeletion(&BB))
443 continue;
444
445 if (pred_empty(&BB)) {
446 // When processBlock makes BB unreachable it doesn't bother to fix up
447 // the instructions in it. We must remove BB to prevent invalid IR.
448 LLVM_DEBUG(dbgs() << " JT: Deleting dead block '" << BB.getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " JT: Deleting dead block '"
<< BB.getName() << "' with terminator: " <<
*BB.getTerminator() << '\n'; } } while (false)
449 << "' with terminator: " << *BB.getTerminator()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " JT: Deleting dead block '"
<< BB.getName() << "' with terminator: " <<
*BB.getTerminator() << '\n'; } } while (false)
450 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " JT: Deleting dead block '"
<< BB.getName() << "' with terminator: " <<
*BB.getTerminator() << '\n'; } } while (false)
;
451 LoopHeaders.erase(&BB);
452 LVI->eraseBlock(&BB);
453 DeleteDeadBlock(&BB, DTU);
454 Changed = true;
455 continue;
456 }
457
458 // processBlock doesn't thread BBs with unconditional TIs. However, if BB
459 // is "almost empty", we attempt to merge BB with its sole successor.
460 auto *BI = dyn_cast<BranchInst>(BB.getTerminator());
461 if (BI && BI->isUnconditional()) {
462 BasicBlock *Succ = BI->getSuccessor(0);
463 if (
464 // The terminator must be the only non-phi instruction in BB.
465 BB.getFirstNonPHIOrDbg()->isTerminator() &&
466 // Don't alter Loop headers and latches to ensure another pass can
467 // detect and transform nested loops later.
468 !LoopHeaders.count(&BB) && !LoopHeaders.count(Succ) &&
469 TryToSimplifyUncondBranchFromEmptyBlock(&BB, DTU)) {
470 RemoveRedundantDbgInstrs(Succ);
471 // BB is valid for cleanup here because we passed in DTU. F remains
472 // BB's parent until a DTU->getDomTree() event.
473 LVI->eraseBlock(&BB);
474 Changed = true;
475 }
476 }
477 }
478 EverChanged |= Changed;
479 } while (Changed);
480
481 LoopHeaders.clear();
482 return EverChanged;
483}
484
485// Replace uses of Cond with ToVal when safe to do so. If all uses are
486// replaced, we can remove Cond. We cannot blindly replace all uses of Cond
487// because we may incorrectly replace uses when guards/assumes are uses of
488// of `Cond` and we used the guards/assume to reason about the `Cond` value
489// at the end of block. RAUW unconditionally replaces all uses
490// including the guards/assumes themselves and the uses before the
491// guard/assume.
492static void replaceFoldableUses(Instruction *Cond, Value *ToVal) {
493 assert(Cond->getType() == ToVal->getType())((Cond->getType() == ToVal->getType()) ? static_cast<
void> (0) : __assert_fail ("Cond->getType() == ToVal->getType()"
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 493, __PRETTY_FUNCTION__))
;
494 auto *BB = Cond->getParent();
495 // We can unconditionally replace all uses in non-local blocks (i.e. uses
496 // strictly dominated by BB), since LVI information is true from the
497 // terminator of BB.
498 replaceNonLocalUsesWith(Cond, ToVal);
499 for (Instruction &I : reverse(*BB)) {
500 // Reached the Cond whose uses we are trying to replace, so there are no
501 // more uses.
502 if (&I == Cond)
503 break;
504 // We only replace uses in instructions that are guaranteed to reach the end
505 // of BB, where we know Cond is ToVal.
506 if (!isGuaranteedToTransferExecutionToSuccessor(&I))
507 break;
508 I.replaceUsesOfWith(Cond, ToVal);
509 }
510 if (Cond->use_empty() && !Cond->mayHaveSideEffects())
511 Cond->eraseFromParent();
512}
513
514/// Return the cost of duplicating a piece of this block from first non-phi
515/// and before StopAt instruction to thread across it. Stop scanning the block
516/// when exceeding the threshold. If duplication is impossible, returns ~0U.
517static unsigned getJumpThreadDuplicationCost(BasicBlock *BB,
518 Instruction *StopAt,
519 unsigned Threshold) {
520 assert(StopAt->getParent() == BB && "Not an instruction from proper BB?")((StopAt->getParent() == BB && "Not an instruction from proper BB?"
) ? static_cast<void> (0) : __assert_fail ("StopAt->getParent() == BB && \"Not an instruction from proper BB?\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 520, __PRETTY_FUNCTION__))
;
521 /// Ignore PHI nodes, these will be flattened when duplication happens.
522 BasicBlock::const_iterator I(BB->getFirstNonPHI());
523
524 // FIXME: THREADING will delete values that are just used to compute the
525 // branch, so they shouldn't count against the duplication cost.
526
527 unsigned Bonus = 0;
528 if (BB->getTerminator() == StopAt) {
529 // Threading through a switch statement is particularly profitable. If this
530 // block ends in a switch, decrease its cost to make it more likely to
531 // happen.
532 if (isa<SwitchInst>(StopAt))
533 Bonus = 6;
534
535 // The same holds for indirect branches, but slightly more so.
536 if (isa<IndirectBrInst>(StopAt))
537 Bonus = 8;
538 }
539
540 // Bump the threshold up so the early exit from the loop doesn't skip the
541 // terminator-based Size adjustment at the end.
542 Threshold += Bonus;
543
544 // Sum up the cost of each instruction until we get to the terminator. Don't
545 // include the terminator because the copy won't include it.
546 unsigned Size = 0;
547 for (; &*I != StopAt; ++I) {
548
549 // Stop scanning the block if we've reached the threshold.
550 if (Size > Threshold)
551 return Size;
552
553 // Debugger intrinsics don't incur code size.
554 if (isa<DbgInfoIntrinsic>(I)) continue;
555
556 // Pseudo-probes don't incur code size.
557 if (isa<PseudoProbeInst>(I))
558 continue;
559
560 // If this is a pointer->pointer bitcast, it is free.
561 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
562 continue;
563
564 // Freeze instruction is free, too.
565 if (isa<FreezeInst>(I))
566 continue;
567
568 // Bail out if this instruction gives back a token type, it is not possible
569 // to duplicate it if it is used outside this BB.
570 if (I->getType()->isTokenTy() && I->isUsedOutsideOfBlock(BB))
571 return ~0U;
572
573 // All other instructions count for at least one unit.
574 ++Size;
575
576 // Calls are more expensive. If they are non-intrinsic calls, we model them
577 // as having cost of 4. If they are a non-vector intrinsic, we model them
578 // as having cost of 2 total, and if they are a vector intrinsic, we model
579 // them as having cost 1.
580 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
581 if (CI->cannotDuplicate() || CI->isConvergent())
582 // Blocks with NoDuplicate are modelled as having infinite cost, so they
583 // are never duplicated.
584 return ~0U;
585 else if (!isa<IntrinsicInst>(CI))
586 Size += 3;
587 else if (!CI->getType()->isVectorTy())
588 Size += 1;
589 }
590 }
591
592 return Size > Bonus ? Size - Bonus : 0;
593}
594
595/// findLoopHeaders - We do not want jump threading to turn proper loop
596/// structures into irreducible loops. Doing this breaks up the loop nesting
597/// hierarchy and pessimizes later transformations. To prevent this from
598/// happening, we first have to find the loop headers. Here we approximate this
599/// by finding targets of backedges in the CFG.
600///
601/// Note that there definitely are cases when we want to allow threading of
602/// edges across a loop header. For example, threading a jump from outside the
603/// loop (the preheader) to an exit block of the loop is definitely profitable.
604/// It is also almost always profitable to thread backedges from within the loop
605/// to exit blocks, and is often profitable to thread backedges to other blocks
606/// within the loop (forming a nested loop). This simple analysis is not rich
607/// enough to track all of these properties and keep it up-to-date as the CFG
608/// mutates, so we don't allow any of these transformations.
609void JumpThreadingPass::findLoopHeaders(Function &F) {
610 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
611 FindFunctionBackedges(F, Edges);
612
613 for (const auto &Edge : Edges)
614 LoopHeaders.insert(Edge.second);
615}
616
617/// getKnownConstant - Helper method to determine if we can thread over a
618/// terminator with the given value as its condition, and if so what value to
619/// use for that. What kind of value this is depends on whether we want an
620/// integer or a block address, but an undef is always accepted.
621/// Returns null if Val is null or not an appropriate constant.
622static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
623 if (!Val)
624 return nullptr;
625
626 // Undef is "known" enough.
627 if (UndefValue *U = dyn_cast<UndefValue>(Val))
628 return U;
629
630 if (Preference == WantBlockAddress)
631 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
632
633 return dyn_cast<ConstantInt>(Val);
634}
635
636/// computeValueKnownInPredecessors - Given a basic block BB and a value V, see
637/// if we can infer that the value is a known ConstantInt/BlockAddress or undef
638/// in any of our predecessors. If so, return the known list of value and pred
639/// BB in the result vector.
640///
641/// This returns true if there were any known values.
642bool JumpThreadingPass::computeValueKnownInPredecessorsImpl(
643 Value *V, BasicBlock *BB, PredValueInfo &Result,
644 ConstantPreference Preference, DenseSet<Value *> &RecursionSet,
645 Instruction *CxtI) {
646 // This method walks up use-def chains recursively. Because of this, we could
647 // get into an infinite loop going around loops in the use-def chain. To
648 // prevent this, keep track of what (value, block) pairs we've already visited
649 // and terminate the search if we loop back to them
650 if (!RecursionSet.insert(V).second)
651 return false;
652
653 // If V is a constant, then it is known in all predecessors.
654 if (Constant *KC = getKnownConstant(V, Preference)) {
655 for (BasicBlock *Pred : predecessors(BB))
656 Result.emplace_back(KC, Pred);
657
658 return !Result.empty();
659 }
660
661 // If V is a non-instruction value, or an instruction in a different block,
662 // then it can't be derived from a PHI.
663 Instruction *I = dyn_cast<Instruction>(V);
664 if (!I || I->getParent() != BB) {
665
666 // Okay, if this is a live-in value, see if it has a known value at the end
667 // of any of our predecessors.
668 //
669 // FIXME: This should be an edge property, not a block end property.
670 /// TODO: Per PR2563, we could infer value range information about a
671 /// predecessor based on its terminator.
672 //
673 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
674 // "I" is a non-local compare-with-a-constant instruction. This would be
675 // able to handle value inequalities better, for example if the compare is
676 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
677 // Perhaps getConstantOnEdge should be smart enough to do this?
678 for (BasicBlock *P : predecessors(BB)) {
679 // If the value is known by LazyValueInfo to be a constant in a
680 // predecessor, use that information to try to thread this block.
681 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB, CxtI);
682 if (Constant *KC = getKnownConstant(PredCst, Preference))
683 Result.emplace_back(KC, P);
684 }
685
686 return !Result.empty();
687 }
688
689 /// If I is a PHI node, then we know the incoming values for any constants.
690 if (PHINode *PN = dyn_cast<PHINode>(I)) {
691 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
692 Value *InVal = PN->getIncomingValue(i);
693 if (Constant *KC = getKnownConstant(InVal, Preference)) {
694 Result.emplace_back(KC, PN->getIncomingBlock(i));
695 } else {
696 Constant *CI = LVI->getConstantOnEdge(InVal,
697 PN->getIncomingBlock(i),
698 BB, CxtI);
699 if (Constant *KC = getKnownConstant(CI, Preference))
700 Result.emplace_back(KC, PN->getIncomingBlock(i));
701 }
702 }
703
704 return !Result.empty();
705 }
706
707 // Handle Cast instructions.
708 if (CastInst *CI = dyn_cast<CastInst>(I)) {
709 Value *Source = CI->getOperand(0);
710 computeValueKnownInPredecessorsImpl(Source, BB, Result, Preference,
711 RecursionSet, CxtI);
712 if (Result.empty())
713 return false;
714
715 // Convert the known values.
716 for (auto &R : Result)
717 R.first = ConstantExpr::getCast(CI->getOpcode(), R.first, CI->getType());
718
719 return true;
720 }
721
722 if (FreezeInst *FI = dyn_cast<FreezeInst>(I)) {
723 Value *Source = FI->getOperand(0);
724 computeValueKnownInPredecessorsImpl(Source, BB, Result, Preference,
725 RecursionSet, CxtI);
726
727 erase_if(Result, [](auto &Pair) {
728 return !isGuaranteedNotToBeUndefOrPoison(Pair.first);
729 });
730
731 return !Result.empty();
732 }
733
734 // Handle some boolean conditions.
735 if (I->getType()->getPrimitiveSizeInBits() == 1) {
736 assert(Preference == WantInteger && "One-bit non-integer type?")((Preference == WantInteger && "One-bit non-integer type?"
) ? static_cast<void> (0) : __assert_fail ("Preference == WantInteger && \"One-bit non-integer type?\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 736, __PRETTY_FUNCTION__))
;
737 // X | true -> true
738 // X & false -> false
739 if (I->getOpcode() == Instruction::Or ||
740 I->getOpcode() == Instruction::And) {
741 PredValueInfoTy LHSVals, RHSVals;
742
743 computeValueKnownInPredecessorsImpl(I->getOperand(0), BB, LHSVals,
744 WantInteger, RecursionSet, CxtI);
745 computeValueKnownInPredecessorsImpl(I->getOperand(1), BB, RHSVals,
746 WantInteger, RecursionSet, CxtI);
747
748 if (LHSVals.empty() && RHSVals.empty())
749 return false;
750
751 ConstantInt *InterestingVal;
752 if (I->getOpcode() == Instruction::Or)
753 InterestingVal = ConstantInt::getTrue(I->getContext());
754 else
755 InterestingVal = ConstantInt::getFalse(I->getContext());
756
757 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
758
759 // Scan for the sentinel. If we find an undef, force it to the
760 // interesting value: x|undef -> true and x&undef -> false.
761 for (const auto &LHSVal : LHSVals)
762 if (LHSVal.first == InterestingVal || isa<UndefValue>(LHSVal.first)) {
763 Result.emplace_back(InterestingVal, LHSVal.second);
764 LHSKnownBBs.insert(LHSVal.second);
765 }
766 for (const auto &RHSVal : RHSVals)
767 if (RHSVal.first == InterestingVal || isa<UndefValue>(RHSVal.first)) {
768 // If we already inferred a value for this block on the LHS, don't
769 // re-add it.
770 if (!LHSKnownBBs.count(RHSVal.second))
771 Result.emplace_back(InterestingVal, RHSVal.second);
772 }
773
774 return !Result.empty();
775 }
776
777 // Handle the NOT form of XOR.
778 if (I->getOpcode() == Instruction::Xor &&
779 isa<ConstantInt>(I->getOperand(1)) &&
780 cast<ConstantInt>(I->getOperand(1))->isOne()) {
781 computeValueKnownInPredecessorsImpl(I->getOperand(0), BB, Result,
782 WantInteger, RecursionSet, CxtI);
783 if (Result.empty())
784 return false;
785
786 // Invert the known values.
787 for (auto &R : Result)
788 R.first = ConstantExpr::getNot(R.first);
789
790 return true;
791 }
792
793 // Try to simplify some other binary operator values.
794 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
795 assert(Preference != WantBlockAddress((Preference != WantBlockAddress && "A binary operator creating a block address?"
) ? static_cast<void> (0) : __assert_fail ("Preference != WantBlockAddress && \"A binary operator creating a block address?\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 796, __PRETTY_FUNCTION__))
796 && "A binary operator creating a block address?")((Preference != WantBlockAddress && "A binary operator creating a block address?"
) ? static_cast<void> (0) : __assert_fail ("Preference != WantBlockAddress && \"A binary operator creating a block address?\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 796, __PRETTY_FUNCTION__))
;
797 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
798 PredValueInfoTy LHSVals;
799 computeValueKnownInPredecessorsImpl(BO->getOperand(0), BB, LHSVals,
800 WantInteger, RecursionSet, CxtI);
801
802 // Try to use constant folding to simplify the binary operator.
803 for (const auto &LHSVal : LHSVals) {
804 Constant *V = LHSVal.first;
805 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
806
807 if (Constant *KC = getKnownConstant(Folded, WantInteger))
808 Result.emplace_back(KC, LHSVal.second);
809 }
810 }
811
812 return !Result.empty();
813 }
814
815 // Handle compare with phi operand, where the PHI is defined in this block.
816 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
817 assert(Preference == WantInteger && "Compares only produce integers")((Preference == WantInteger && "Compares only produce integers"
) ? static_cast<void> (0) : __assert_fail ("Preference == WantInteger && \"Compares only produce integers\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 817, __PRETTY_FUNCTION__))
;
818 Type *CmpType = Cmp->getType();
819 Value *CmpLHS = Cmp->getOperand(0);
820 Value *CmpRHS = Cmp->getOperand(1);
821 CmpInst::Predicate Pred = Cmp->getPredicate();
822
823 PHINode *PN = dyn_cast<PHINode>(CmpLHS);
824 if (!PN)
825 PN = dyn_cast<PHINode>(CmpRHS);
826 if (PN && PN->getParent() == BB) {
827 const DataLayout &DL = PN->getModule()->getDataLayout();
828 // We can do this simplification if any comparisons fold to true or false.
829 // See if any do.
830 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
831 BasicBlock *PredBB = PN->getIncomingBlock(i);
832 Value *LHS, *RHS;
833 if (PN == CmpLHS) {
834 LHS = PN->getIncomingValue(i);
835 RHS = CmpRHS->DoPHITranslation(BB, PredBB);
836 } else {
837 LHS = CmpLHS->DoPHITranslation(BB, PredBB);
838 RHS = PN->getIncomingValue(i);
839 }
840 Value *Res = SimplifyCmpInst(Pred, LHS, RHS, {DL});
841 if (!Res) {
842 if (!isa<Constant>(RHS))
843 continue;
844
845 // getPredicateOnEdge call will make no sense if LHS is defined in BB.
846 auto LHSInst = dyn_cast<Instruction>(LHS);
847 if (LHSInst && LHSInst->getParent() == BB)
848 continue;
849
850 LazyValueInfo::Tristate
851 ResT = LVI->getPredicateOnEdge(Pred, LHS,
852 cast<Constant>(RHS), PredBB, BB,
853 CxtI ? CxtI : Cmp);
854 if (ResT == LazyValueInfo::Unknown)
855 continue;
856 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
857 }
858
859 if (Constant *KC = getKnownConstant(Res, WantInteger))
860 Result.emplace_back(KC, PredBB);
861 }
862
863 return !Result.empty();
864 }
865
866 // If comparing a live-in value against a constant, see if we know the
867 // live-in value on any predecessors.
868 if (isa<Constant>(CmpRHS) && !CmpType->isVectorTy()) {
869 Constant *CmpConst = cast<Constant>(CmpRHS);
870
871 if (!isa<Instruction>(CmpLHS) ||
872 cast<Instruction>(CmpLHS)->getParent() != BB) {
873 for (BasicBlock *P : predecessors(BB)) {
874 // If the value is known by LazyValueInfo to be a constant in a
875 // predecessor, use that information to try to thread this block.
876 LazyValueInfo::Tristate Res =
877 LVI->getPredicateOnEdge(Pred, CmpLHS,
878 CmpConst, P, BB, CxtI ? CxtI : Cmp);
879 if (Res == LazyValueInfo::Unknown)
880 continue;
881
882 Constant *ResC = ConstantInt::get(CmpType, Res);
883 Result.emplace_back(ResC, P);
884 }
885
886 return !Result.empty();
887 }
888
889 // InstCombine can fold some forms of constant range checks into
890 // (icmp (add (x, C1)), C2). See if we have we have such a thing with
891 // x as a live-in.
892 {
893 using namespace PatternMatch;
894
895 Value *AddLHS;
896 ConstantInt *AddConst;
897 if (isa<ConstantInt>(CmpConst) &&
898 match(CmpLHS, m_Add(m_Value(AddLHS), m_ConstantInt(AddConst)))) {
899 if (!isa<Instruction>(AddLHS) ||
900 cast<Instruction>(AddLHS)->getParent() != BB) {
901 for (BasicBlock *P : predecessors(BB)) {
902 // If the value is known by LazyValueInfo to be a ConstantRange in
903 // a predecessor, use that information to try to thread this
904 // block.
905 ConstantRange CR = LVI->getConstantRangeOnEdge(
906 AddLHS, P, BB, CxtI ? CxtI : cast<Instruction>(CmpLHS));
907 // Propagate the range through the addition.
908 CR = CR.add(AddConst->getValue());
909
910 // Get the range where the compare returns true.
911 ConstantRange CmpRange = ConstantRange::makeExactICmpRegion(
912 Pred, cast<ConstantInt>(CmpConst)->getValue());
913
914 Constant *ResC;
915 if (CmpRange.contains(CR))
916 ResC = ConstantInt::getTrue(CmpType);
917 else if (CmpRange.inverse().contains(CR))
918 ResC = ConstantInt::getFalse(CmpType);
919 else
920 continue;
921
922 Result.emplace_back(ResC, P);
923 }
924
925 return !Result.empty();
926 }
927 }
928 }
929
930 // Try to find a constant value for the LHS of a comparison,
931 // and evaluate it statically if we can.
932 PredValueInfoTy LHSVals;
933 computeValueKnownInPredecessorsImpl(I->getOperand(0), BB, LHSVals,
934 WantInteger, RecursionSet, CxtI);
935
936 for (const auto &LHSVal : LHSVals) {
937 Constant *V = LHSVal.first;
938 Constant *Folded = ConstantExpr::getCompare(Pred, V, CmpConst);
939 if (Constant *KC = getKnownConstant(Folded, WantInteger))
940 Result.emplace_back(KC, LHSVal.second);
941 }
942
943 return !Result.empty();
944 }
945 }
946
947 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
948 // Handle select instructions where at least one operand is a known constant
949 // and we can figure out the condition value for any predecessor block.
950 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
951 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
952 PredValueInfoTy Conds;
953 if ((TrueVal || FalseVal) &&
954 computeValueKnownInPredecessorsImpl(SI->getCondition(), BB, Conds,
955 WantInteger, RecursionSet, CxtI)) {
956 for (auto &C : Conds) {
957 Constant *Cond = C.first;
958
959 // Figure out what value to use for the condition.
960 bool KnownCond;
961 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
962 // A known boolean.
963 KnownCond = CI->isOne();
964 } else {
965 assert(isa<UndefValue>(Cond) && "Unexpected condition value")((isa<UndefValue>(Cond) && "Unexpected condition value"
) ? static_cast<void> (0) : __assert_fail ("isa<UndefValue>(Cond) && \"Unexpected condition value\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 965, __PRETTY_FUNCTION__))
;
966 // Either operand will do, so be sure to pick the one that's a known
967 // constant.
968 // FIXME: Do this more cleverly if both values are known constants?
969 KnownCond = (TrueVal != nullptr);
970 }
971
972 // See if the select has a known constant value for this predecessor.
973 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
974 Result.emplace_back(Val, C.second);
975 }
976
977 return !Result.empty();
978 }
979 }
980
981 // If all else fails, see if LVI can figure out a constant value for us.
982 assert(CxtI->getParent() == BB && "CxtI should be in BB")((CxtI->getParent() == BB && "CxtI should be in BB"
) ? static_cast<void> (0) : __assert_fail ("CxtI->getParent() == BB && \"CxtI should be in BB\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 982, __PRETTY_FUNCTION__))
;
983 Constant *CI = LVI->getConstant(V, CxtI);
984 if (Constant *KC = getKnownConstant(CI, Preference)) {
985 for (BasicBlock *Pred : predecessors(BB))
986 Result.emplace_back(KC, Pred);
987 }
988
989 return !Result.empty();
990}
991
992/// GetBestDestForBranchOnUndef - If we determine that the specified block ends
993/// in an undefined jump, decide which block is best to revector to.
994///
995/// Since we can pick an arbitrary destination, we pick the successor with the
996/// fewest predecessors. This should reduce the in-degree of the others.
997static unsigned getBestDestForJumpOnUndef(BasicBlock *BB) {
998 Instruction *BBTerm = BB->getTerminator();
999 unsigned MinSucc = 0;
1000 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
1001 // Compute the successor with the minimum number of predecessors.
1002 unsigned MinNumPreds = pred_size(TestBB);
1003 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
1004 TestBB = BBTerm->getSuccessor(i);
1005 unsigned NumPreds = pred_size(TestBB);
1006 if (NumPreds < MinNumPreds) {
1007 MinSucc = i;
1008 MinNumPreds = NumPreds;
1009 }
1010 }
1011
1012 return MinSucc;
1013}
1014
1015static bool hasAddressTakenAndUsed(BasicBlock *BB) {
1016 if (!BB->hasAddressTaken()) return false;
1017
1018 // If the block has its address taken, it may be a tree of dead constants
1019 // hanging off of it. These shouldn't keep the block alive.
1020 BlockAddress *BA = BlockAddress::get(BB);
1021 BA->removeDeadConstantUsers();
1022 return !BA->use_empty();
1023}
1024
1025/// processBlock - If there are any predecessors whose control can be threaded
1026/// through to a successor, transform them now.
1027bool JumpThreadingPass::processBlock(BasicBlock *BB) {
1028 // If the block is trivially dead, just return and let the caller nuke it.
1029 // This simplifies other transformations.
1030 if (DTU->isBBPendingDeletion(BB) ||
1031 (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()))
1032 return false;
1033
1034 // If this block has a single predecessor, and if that pred has a single
1035 // successor, merge the blocks. This encourages recursive jump threading
1036 // because now the condition in this block can be threaded through
1037 // predecessors of our predecessor block.
1038 if (maybeMergeBasicBlockIntoOnlyPred(BB))
1039 return true;
1040
1041 if (tryToUnfoldSelectInCurrBB(BB))
1042 return true;
1043
1044 // Look if we can propagate guards to predecessors.
1045 if (HasGuards && processGuards(BB))
1046 return true;
1047
1048 // What kind of constant we're looking for.
1049 ConstantPreference Preference = WantInteger;
1050
1051 // Look to see if the terminator is a conditional branch, switch or indirect
1052 // branch, if not we can't thread it.
1053 Value *Condition;
1054 Instruction *Terminator = BB->getTerminator();
1055 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
1056 // Can't thread an unconditional jump.
1057 if (BI->isUnconditional()) return false;
1058 Condition = BI->getCondition();
1059 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
1060 Condition = SI->getCondition();
1061 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
1062 // Can't thread indirect branch with no successors.
1063 if (IB->getNumSuccessors() == 0) return false;
1064 Condition = IB->getAddress()->stripPointerCasts();
1065 Preference = WantBlockAddress;
1066 } else {
1067 return false; // Must be an invoke or callbr.
1068 }
1069
1070 // Keep track if we constant folded the condition in this invocation.
1071 bool ConstantFolded = false;
1072
1073 // Run constant folding to see if we can reduce the condition to a simple
1074 // constant.
1075 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
1076 Value *SimpleVal =
1077 ConstantFoldInstruction(I, BB->getModule()->getDataLayout(), TLI);
1078 if (SimpleVal) {
1079 I->replaceAllUsesWith(SimpleVal);
1080 if (isInstructionTriviallyDead(I, TLI))
1081 I->eraseFromParent();
1082 Condition = SimpleVal;
1083 ConstantFolded = true;
1084 }
1085 }
1086
1087 // If the terminator is branching on an undef or freeze undef, we can pick any
1088 // of the successors to branch to. Let getBestDestForJumpOnUndef decide.
1089 auto *FI = dyn_cast<FreezeInst>(Condition);
1090 if (isa<UndefValue>(Condition) ||
1091 (FI && isa<UndefValue>(FI->getOperand(0)) && FI->hasOneUse())) {
1092 unsigned BestSucc = getBestDestForJumpOnUndef(BB);
1093 std::vector<DominatorTree::UpdateType> Updates;
1094
1095 // Fold the branch/switch.
1096 Instruction *BBTerm = BB->getTerminator();
1097 Updates.reserve(BBTerm->getNumSuccessors());
1098 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
1099 if (i == BestSucc) continue;
1100 BasicBlock *Succ = BBTerm->getSuccessor(i);
1101 Succ->removePredecessor(BB, true);
1102 Updates.push_back({DominatorTree::Delete, BB, Succ});
1103 }
1104
1105 LLVM_DEBUG(dbgs() << " In block '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " In block '" <<
BB->getName() << "' folding undef terminator: " <<
*BBTerm << '\n'; } } while (false)
1106 << "' 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)
;
1107 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
1108 BBTerm->eraseFromParent();
1109 DTU->applyUpdatesPermissive(Updates);
1110 if (FI)
1111 FI->eraseFromParent();
1112 return true;
1113 }
1114
1115 // If the terminator of this block is branching on a constant, simplify the
1116 // terminator to an unconditional branch. This can occur due to threading in
1117 // other blocks.
1118 if (getKnownConstant(Condition, Preference)) {
1119 LLVM_DEBUG(dbgs() << " In block '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " In block '" <<
BB->getName() << "' folding terminator: " << *
BB->getTerminator() << '\n'; } } while (false)
1120 << "' folding terminator: " << *BB->getTerminator()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " In block '" <<
BB->getName() << "' folding terminator: " << *
BB->getTerminator() << '\n'; } } while (false)
1121 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " In block '" <<
BB->getName() << "' folding terminator: " << *
BB->getTerminator() << '\n'; } } while (false)
;
1122 ++NumFolds;
1123 ConstantFoldTerminator(BB, true, nullptr, DTU);
1124 if (HasProfileData)
1125 BPI->eraseBlock(BB);
1126 return true;
1127 }
1128
1129 Instruction *CondInst = dyn_cast<Instruction>(Condition);
1130
1131 // All the rest of our checks depend on the condition being an instruction.
1132 if (!CondInst) {
1133 // FIXME: Unify this with code below.
1134 if (processThreadableEdges(Condition, BB, Preference, Terminator))
1135 return true;
1136 return ConstantFolded;
1137 }
1138
1139 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
1140 // If we're branching on a conditional, LVI might be able to determine
1141 // it's value at the branch instruction. We only handle comparisons
1142 // against a constant at this time.
1143 // TODO: This should be extended to handle switches as well.
1144 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1145 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
1146 if (CondBr && CondConst) {
1147 // We should have returned as soon as we turn a conditional branch to
1148 // unconditional. Because its no longer interesting as far as jump
1149 // threading is concerned.
1150 assert(CondBr->isConditional() && "Threading on unconditional terminator")((CondBr->isConditional() && "Threading on unconditional terminator"
) ? static_cast<void> (0) : __assert_fail ("CondBr->isConditional() && \"Threading on unconditional terminator\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1150, __PRETTY_FUNCTION__))
;
1151
1152 LazyValueInfo::Tristate Ret =
1153 LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0),
1154 CondConst, CondBr);
1155 if (Ret != LazyValueInfo::Unknown) {
1156 unsigned ToRemove = Ret == LazyValueInfo::True ? 1 : 0;
1157 unsigned ToKeep = Ret == LazyValueInfo::True ? 0 : 1;
1158 BasicBlock *ToRemoveSucc = CondBr->getSuccessor(ToRemove);
1159 ToRemoveSucc->removePredecessor(BB, true);
1160 BranchInst *UncondBr =
1161 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
1162 UncondBr->setDebugLoc(CondBr->getDebugLoc());
1163 CondBr->eraseFromParent();
1164 if (CondCmp->use_empty())
1165 CondCmp->eraseFromParent();
1166 // We can safely replace *some* uses of the CondInst if it has
1167 // exactly one value as returned by LVI. RAUW is incorrect in the
1168 // presence of guards and assumes, that have the `Cond` as the use. This
1169 // is because we use the guards/assume to reason about the `Cond` value
1170 // at the end of block, but RAUW unconditionally replaces all uses
1171 // including the guards/assumes themselves and the uses before the
1172 // guard/assume.
1173 else if (CondCmp->getParent() == BB) {
1174 auto *CI = Ret == LazyValueInfo::True ?
1175 ConstantInt::getTrue(CondCmp->getType()) :
1176 ConstantInt::getFalse(CondCmp->getType());
1177 replaceFoldableUses(CondCmp, CI);
1178 }
1179 DTU->applyUpdatesPermissive(
1180 {{DominatorTree::Delete, BB, ToRemoveSucc}});
1181 if (HasProfileData)
1182 BPI->eraseBlock(BB);
1183 return true;
1184 }
1185
1186 // We did not manage to simplify this branch, try to see whether
1187 // CondCmp depends on a known phi-select pattern.
1188 if (tryToUnfoldSelect(CondCmp, BB))
1189 return true;
1190 }
1191 }
1192
1193 if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
1194 if (tryToUnfoldSelect(SI, BB))
1195 return true;
1196
1197 // Check for some cases that are worth simplifying. Right now we want to look
1198 // for loads that are used by a switch or by the condition for the branch. If
1199 // we see one, check to see if it's partially redundant. If so, insert a PHI
1200 // which can then be used to thread the values.
1201 Value *SimplifyValue = CondInst;
1202
1203 if (auto *FI = dyn_cast<FreezeInst>(SimplifyValue))
1204 // Look into freeze's operand
1205 SimplifyValue = FI->getOperand(0);
1206
1207 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
1208 if (isa<Constant>(CondCmp->getOperand(1)))
1209 SimplifyValue = CondCmp->getOperand(0);
1210
1211 // TODO: There are other places where load PRE would be profitable, such as
1212 // more complex comparisons.
1213 if (LoadInst *LoadI = dyn_cast<LoadInst>(SimplifyValue))
1214 if (simplifyPartiallyRedundantLoad(LoadI))
1215 return true;
1216
1217 // Before threading, try to propagate profile data backwards:
1218 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
1219 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1220 updatePredecessorProfileMetadata(PN, BB);
1221
1222 // Handle a variety of cases where we are branching on something derived from
1223 // a PHI node in the current block. If we can prove that any predecessors
1224 // compute a predictable value based on a PHI node, thread those predecessors.
1225 if (processThreadableEdges(CondInst, BB, Preference, Terminator))
1226 return true;
1227
1228 // If this is an otherwise-unfoldable branch on a phi node or freeze(phi) in
1229 // the current block, see if we can simplify.
1230 PHINode *PN = dyn_cast<PHINode>(
1231 isa<FreezeInst>(CondInst) ? cast<FreezeInst>(CondInst)->getOperand(0)
1232 : CondInst);
1233
1234 if (PN && PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1235 return processBranchOnPHI(PN);
1236
1237 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
1238 if (CondInst->getOpcode() == Instruction::Xor &&
1239 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1240 return processBranchOnXOR(cast<BinaryOperator>(CondInst));
1241
1242 // Search for a stronger dominating condition that can be used to simplify a
1243 // conditional branch leaving BB.
1244 if (processImpliedCondition(BB))
1245 return true;
1246
1247 return false;
1248}
1249
1250bool JumpThreadingPass::processImpliedCondition(BasicBlock *BB) {
1251 auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
1252 if (!BI || !BI->isConditional())
1253 return false;
1254
1255 Value *Cond = BI->getCondition();
1256 BasicBlock *CurrentBB = BB;
1257 BasicBlock *CurrentPred = BB->getSinglePredecessor();
1258 unsigned Iter = 0;
1259
1260 auto &DL = BB->getModule()->getDataLayout();
1261
1262 while (CurrentPred && Iter++ < ImplicationSearchThreshold) {
1263 auto *PBI = dyn_cast<BranchInst>(CurrentPred->getTerminator());
1264 if (!PBI || !PBI->isConditional())
1265 return false;
1266 if (PBI->getSuccessor(0) != CurrentBB && PBI->getSuccessor(1) != CurrentBB)
1267 return false;
1268
1269 bool CondIsTrue = PBI->getSuccessor(0) == CurrentBB;
1270 Optional<bool> Implication =
1271 isImpliedCondition(PBI->getCondition(), Cond, DL, CondIsTrue);
1272 if (Implication) {
1273 BasicBlock *KeepSucc = BI->getSuccessor(*Implication ? 0 : 1);
1274 BasicBlock *RemoveSucc = BI->getSuccessor(*Implication ? 1 : 0);
1275 RemoveSucc->removePredecessor(BB);
1276 BranchInst *UncondBI = BranchInst::Create(KeepSucc, BI);
1277 UncondBI->setDebugLoc(BI->getDebugLoc());
1278 BI->eraseFromParent();
1279 DTU->applyUpdatesPermissive({{DominatorTree::Delete, BB, RemoveSucc}});
1280 if (HasProfileData)
1281 BPI->eraseBlock(BB);
1282 return true;
1283 }
1284 CurrentBB = CurrentPred;
1285 CurrentPred = CurrentBB->getSinglePredecessor();
1286 }
1287
1288 return false;
1289}
1290
1291/// Return true if Op is an instruction defined in the given block.
1292static bool isOpDefinedInBlock(Value *Op, BasicBlock *BB) {
1293 if (Instruction *OpInst = dyn_cast<Instruction>(Op))
12
Assuming 'OpInst' is null
13
Taking false branch
1294 if (OpInst->getParent() == BB)
1295 return true;
1296 return false;
14
Returning zero, which participates in a condition later
1297}
1298
1299/// simplifyPartiallyRedundantLoad - If LoadI is an obviously partially
1300/// redundant load instruction, eliminate it by replacing it with a PHI node.
1301/// This is an important optimization that encourages jump threading, and needs
1302/// to be run interlaced with other jump threading tasks.
1303bool JumpThreadingPass::simplifyPartiallyRedundantLoad(LoadInst *LoadI) {
1304 // Don't hack volatile and ordered loads.
1305 if (!LoadI->isUnordered()) return false;
1
Calling 'LoadInst::isUnordered'
5
Returning from 'LoadInst::isUnordered'
6
Taking false branch
1306
1307 // If the load is defined in a block with exactly one predecessor, it can't be
1308 // partially redundant.
1309 BasicBlock *LoadBB = LoadI->getParent();
1310 if (LoadBB->getSinglePredecessor())
7
Assuming the condition is false
8
Taking false branch
1311 return false;
1312
1313 // If the load is defined in an EH pad, it can't be partially redundant,
1314 // because the edges between the invoke and the EH pad cannot have other
1315 // instructions between them.
1316 if (LoadBB->isEHPad())
9
Assuming the condition is false
10
Taking false branch
1317 return false;
1318
1319 Value *LoadedPtr = LoadI->getOperand(0);
1320
1321 // If the loaded operand is defined in the LoadBB and its not a phi,
1322 // it can't be available in predecessors.
1323 if (isOpDefinedInBlock(LoadedPtr, LoadBB) && !isa<PHINode>(LoadedPtr))
11
Calling 'isOpDefinedInBlock'
15
Returning from 'isOpDefinedInBlock'
1324 return false;
1325
1326 // Scan a few instructions up from the load, to see if it is obviously live at
1327 // the entry to its block.
1328 BasicBlock::iterator BBIt(LoadI);
1329 bool IsLoadCSE;
1330 if (Value *AvailableVal = FindAvailableLoadedValue(
16
Assuming 'AvailableVal' is null
17
Taking false branch
1331 LoadI, LoadBB, BBIt, DefMaxInstsToScan, AA, &IsLoadCSE)) {
1332 // If the value of the load is locally available within the block, just use
1333 // it. This frequently occurs for reg2mem'd allocas.
1334
1335 if (IsLoadCSE) {
1336 LoadInst *NLoadI = cast<LoadInst>(AvailableVal);
1337 combineMetadataForCSE(NLoadI, LoadI, false);
1338 };
1339
1340 // If the returned value is the load itself, replace with an undef. This can
1341 // only happen in dead loops.
1342 if (AvailableVal == LoadI)
1343 AvailableVal = UndefValue::get(LoadI->getType());
1344 if (AvailableVal->getType() != LoadI->getType())
1345 AvailableVal = CastInst::CreateBitOrPointerCast(
1346 AvailableVal, LoadI->getType(), "", LoadI);
1347 LoadI->replaceAllUsesWith(AvailableVal);
1348 LoadI->eraseFromParent();
1349 return true;
1350 }
1351
1352 // Otherwise, if we scanned the whole block and got to the top of the block,
1353 // we know the block is locally transparent to the load. If not, something
1354 // might clobber its value.
1355 if (BBIt != LoadBB->begin())
18
Calling 'operator!='
21
Returning from 'operator!='
22
Taking false branch
1356 return false;
1357
1358 // If all of the loads and stores that feed the value have the same AA tags,
1359 // then we can propagate them onto any newly inserted loads.
1360 AAMDNodes AATags;
1361 LoadI->getAAMetadata(AATags);
1362
1363 SmallPtrSet<BasicBlock*, 8> PredsScanned;
1364
1365 using AvailablePredsTy = SmallVector<std::pair<BasicBlock *, Value *>, 8>;
1366
1367 AvailablePredsTy AvailablePreds;
1368 BasicBlock *OneUnavailablePred = nullptr;
23
'OneUnavailablePred' initialized to a null pointer value
1369 SmallVector<LoadInst*, 8> CSELoads;
1370
1371 // If we got here, the loaded value is transparent through to the start of the
1372 // block. Check to see if it is available in any of the predecessor blocks.
1373 for (BasicBlock *PredBB : predecessors(LoadBB)) {
1374 // If we already scanned this predecessor, skip it.
1375 if (!PredsScanned.insert(PredBB).second)
1376 continue;
1377
1378 BBIt = PredBB->end();
1379 unsigned NumScanedInst = 0;
1380 Value *PredAvailable = nullptr;
1381 // NOTE: We don't CSE load that is volatile or anything stronger than
1382 // unordered, that should have been checked when we entered the function.
1383 assert(LoadI->isUnordered() &&((LoadI->isUnordered() && "Attempting to CSE volatile or atomic loads"
) ? static_cast<void> (0) : __assert_fail ("LoadI->isUnordered() && \"Attempting to CSE volatile or atomic loads\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1384, __PRETTY_FUNCTION__))
1384 "Attempting to CSE volatile or atomic loads")((LoadI->isUnordered() && "Attempting to CSE volatile or atomic loads"
) ? static_cast<void> (0) : __assert_fail ("LoadI->isUnordered() && \"Attempting to CSE volatile or atomic loads\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1384, __PRETTY_FUNCTION__))
;
1385 // If this is a load on a phi pointer, phi-translate it and search
1386 // for available load/store to the pointer in predecessors.
1387 Value *Ptr = LoadedPtr->DoPHITranslation(LoadBB, PredBB);
1388 PredAvailable = FindAvailablePtrLoadStore(
1389 Ptr, LoadI->getType(), LoadI->isAtomic(), PredBB, BBIt,
1390 DefMaxInstsToScan, AA, &IsLoadCSE, &NumScanedInst);
1391
1392 // If PredBB has a single predecessor, continue scanning through the
1393 // single predecessor.
1394 BasicBlock *SinglePredBB = PredBB;
1395 while (!PredAvailable && SinglePredBB && BBIt == SinglePredBB->begin() &&
1396 NumScanedInst < DefMaxInstsToScan) {
1397 SinglePredBB = SinglePredBB->getSinglePredecessor();
1398 if (SinglePredBB) {
1399 BBIt = SinglePredBB->end();
1400 PredAvailable = FindAvailablePtrLoadStore(
1401 Ptr, LoadI->getType(), LoadI->isAtomic(), SinglePredBB, BBIt,
1402 (DefMaxInstsToScan - NumScanedInst), AA, &IsLoadCSE,
1403 &NumScanedInst);
1404 }
1405 }
1406
1407 if (!PredAvailable) {
1408 OneUnavailablePred = PredBB;
1409 continue;
1410 }
1411
1412 if (IsLoadCSE)
1413 CSELoads.push_back(cast<LoadInst>(PredAvailable));
1414
1415 // If so, this load is partially redundant. Remember this info so that we
1416 // can create a PHI node.
1417 AvailablePreds.emplace_back(PredBB, PredAvailable);
1418 }
1419
1420 // If the loaded value isn't available in any predecessor, it isn't partially
1421 // redundant.
1422 if (AvailablePreds.empty()) return false;
24
Calling 'SmallVectorBase::empty'
27
Returning from 'SmallVectorBase::empty'
28
Taking false branch
1423
1424 // Okay, the loaded value is available in at least one (and maybe all!)
1425 // predecessors. If the value is unavailable in more than one unique
1426 // predecessor, we want to insert a merge block for those common predecessors.
1427 // This ensures that we only have to insert one reload, thus not increasing
1428 // code size.
1429 BasicBlock *UnavailablePred = nullptr;
1430
1431 // If the value is unavailable in one of predecessors, we will end up
1432 // inserting a new instruction into them. It is only valid if all the
1433 // instructions before LoadI are guaranteed to pass execution to its
1434 // successor, or if LoadI is safe to speculate.
1435 // TODO: If this logic becomes more complex, and we will perform PRE insertion
1436 // farther than to a predecessor, we need to reuse the code from GVN's PRE.
1437 // It requires domination tree analysis, so for this simple case it is an
1438 // overkill.
1439 if (PredsScanned.size() != AvailablePreds.size() &&
29
Assuming the condition is false
1440 !isSafeToSpeculativelyExecute(LoadI))
1441 for (auto I = LoadBB->begin(); &*I != LoadI; ++I)
1442 if (!isGuaranteedToTransferExecutionToSuccessor(&*I))
1443 return false;
1444
1445 // If there is exactly one predecessor where the value is unavailable, the
1446 // already computed 'OneUnavailablePred' block is it. If it ends in an
1447 // unconditional branch, we know that it isn't a critical edge.
1448 if (PredsScanned.size() == AvailablePreds.size()+1 &&
30
Assuming the condition is true
1449 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
31
Called C++ object pointer is null
1450 UnavailablePred = OneUnavailablePred;
1451 } else if (PredsScanned.size() != AvailablePreds.size()) {
1452 // Otherwise, we had multiple unavailable predecessors or we had a critical
1453 // edge from the one.
1454 SmallVector<BasicBlock*, 8> PredsToSplit;
1455 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
1456
1457 for (const auto &AvailablePred : AvailablePreds)
1458 AvailablePredSet.insert(AvailablePred.first);
1459
1460 // Add all the unavailable predecessors to the PredsToSplit list.
1461 for (BasicBlock *P : predecessors(LoadBB)) {
1462 // If the predecessor is an indirect goto, we can't split the edge.
1463 // Same for CallBr.
1464 if (isa<IndirectBrInst>(P->getTerminator()) ||
1465 isa<CallBrInst>(P->getTerminator()))
1466 return false;
1467
1468 if (!AvailablePredSet.count(P))
1469 PredsToSplit.push_back(P);
1470 }
1471
1472 // Split them out to their own block.
1473 UnavailablePred = splitBlockPreds(LoadBB, PredsToSplit, "thread-pre-split");
1474 }
1475
1476 // If the value isn't available in all predecessors, then there will be
1477 // exactly one where it isn't available. Insert a load on that edge and add
1478 // it to the AvailablePreds list.
1479 if (UnavailablePred) {
1480 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&((UnavailablePred->getTerminator()->getNumSuccessors() ==
1 && "Can't handle critical edge here!") ? static_cast
<void> (0) : __assert_fail ("UnavailablePred->getTerminator()->getNumSuccessors() == 1 && \"Can't handle critical edge here!\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1481, __PRETTY_FUNCTION__))
1481 "Can't handle critical edge here!")((UnavailablePred->getTerminator()->getNumSuccessors() ==
1 && "Can't handle critical edge here!") ? static_cast
<void> (0) : __assert_fail ("UnavailablePred->getTerminator()->getNumSuccessors() == 1 && \"Can't handle critical edge here!\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1481, __PRETTY_FUNCTION__))
;
1482 LoadInst *NewVal = new LoadInst(
1483 LoadI->getType(), LoadedPtr->DoPHITranslation(LoadBB, UnavailablePred),
1484 LoadI->getName() + ".pr", false, LoadI->getAlign(),
1485 LoadI->getOrdering(), LoadI->getSyncScopeID(),
1486 UnavailablePred->getTerminator());
1487 NewVal->setDebugLoc(LoadI->getDebugLoc());
1488 if (AATags)
1489 NewVal->setAAMetadata(AATags);
1490
1491 AvailablePreds.emplace_back(UnavailablePred, NewVal);
1492 }
1493
1494 // Now we know that each predecessor of this block has a value in
1495 // AvailablePreds, sort them for efficient access as we're walking the preds.
1496 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1497
1498 // Create a PHI node at the start of the block for the PRE'd load value.
1499 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
1500 PHINode *PN = PHINode::Create(LoadI->getType(), std::distance(PB, PE), "",
1501 &LoadBB->front());
1502 PN->takeName(LoadI);
1503 PN->setDebugLoc(LoadI->getDebugLoc());
1504
1505 // Insert new entries into the PHI for each predecessor. A single block may
1506 // have multiple entries here.
1507 for (pred_iterator PI = PB; PI != PE; ++PI) {
1508 BasicBlock *P = *PI;
1509 AvailablePredsTy::iterator I =
1510 llvm::lower_bound(AvailablePreds, std::make_pair(P, (Value *)nullptr));
1511
1512 assert(I != AvailablePreds.end() && I->first == P &&((I != AvailablePreds.end() && I->first == P &&
"Didn't find entry for predecessor!") ? static_cast<void>
(0) : __assert_fail ("I != AvailablePreds.end() && I->first == P && \"Didn't find entry for predecessor!\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1513, __PRETTY_FUNCTION__))
1513 "Didn't find entry for predecessor!")((I != AvailablePreds.end() && I->first == P &&
"Didn't find entry for predecessor!") ? static_cast<void>
(0) : __assert_fail ("I != AvailablePreds.end() && I->first == P && \"Didn't find entry for predecessor!\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1513, __PRETTY_FUNCTION__))
;
1514
1515 // If we have an available predecessor but it requires casting, insert the
1516 // cast in the predecessor and use the cast. Note that we have to update the
1517 // AvailablePreds vector as we go so that all of the PHI entries for this
1518 // predecessor use the same bitcast.
1519 Value *&PredV = I->second;
1520 if (PredV->getType() != LoadI->getType())
1521 PredV = CastInst::CreateBitOrPointerCast(PredV, LoadI->getType(), "",
1522 P->getTerminator());
1523
1524 PN->addIncoming(PredV, I->first);
1525 }
1526
1527 for (LoadInst *PredLoadI : CSELoads) {
1528 combineMetadataForCSE(PredLoadI, LoadI, true);
1529 }
1530
1531 LoadI->replaceAllUsesWith(PN);
1532 LoadI->eraseFromParent();
1533
1534 return true;
1535}
1536
1537/// findMostPopularDest - The specified list contains multiple possible
1538/// threadable destinations. Pick the one that occurs the most frequently in
1539/// the list.
1540static BasicBlock *
1541findMostPopularDest(BasicBlock *BB,
1542 const SmallVectorImpl<std::pair<BasicBlock *,
1543 BasicBlock *>> &PredToDestList) {
1544 assert(!PredToDestList.empty())((!PredToDestList.empty()) ? static_cast<void> (0) : __assert_fail
("!PredToDestList.empty()", "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1544, __PRETTY_FUNCTION__))
;
1545
1546 // Determine popularity. If there are multiple possible destinations, we
1547 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1548 // blocks with known and real destinations to threading undef. We'll handle
1549 // them later if interesting.
1550 MapVector<BasicBlock *, unsigned> DestPopularity;
1551
1552 // Populate DestPopularity with the successors in the order they appear in the
1553 // successor list. This way, we ensure determinism by iterating it in the
1554 // same order in std::max_element below. We map nullptr to 0 so that we can
1555 // return nullptr when PredToDestList contains nullptr only.
1556 DestPopularity[nullptr] = 0;
1557 for (auto *SuccBB : successors(BB))
1558 DestPopularity[SuccBB] = 0;
1559
1560 for (const auto &PredToDest : PredToDestList)
1561 if (PredToDest.second)
1562 DestPopularity[PredToDest.second]++;
1563
1564 // Find the most popular dest.
1565 using VT = decltype(DestPopularity)::value_type;
1566 auto MostPopular = std::max_element(
1567 DestPopularity.begin(), DestPopularity.end(),
1568 [](const VT &L, const VT &R) { return L.second < R.second; });
1569
1570 // Okay, we have finally picked the most popular destination.
1571 return MostPopular->first;
1572}
1573
1574// Try to evaluate the value of V when the control flows from PredPredBB to
1575// BB->getSinglePredecessor() and then on to BB.
1576Constant *JumpThreadingPass::evaluateOnPredecessorEdge(BasicBlock *BB,
1577 BasicBlock *PredPredBB,
1578 Value *V) {
1579 BasicBlock *PredBB = BB->getSinglePredecessor();
1580 assert(PredBB && "Expected a single predecessor")((PredBB && "Expected a single predecessor") ? static_cast
<void> (0) : __assert_fail ("PredBB && \"Expected a single predecessor\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1580, __PRETTY_FUNCTION__))
;
1581
1582 if (Constant *Cst = dyn_cast<Constant>(V)) {
1583 return Cst;
1584 }
1585
1586 // Consult LVI if V is not an instruction in BB or PredBB.
1587 Instruction *I = dyn_cast<Instruction>(V);
1588 if (!I || (I->getParent() != BB && I->getParent() != PredBB)) {
1589 return LVI->getConstantOnEdge(V, PredPredBB, PredBB, nullptr);
1590 }
1591
1592 // Look into a PHI argument.
1593 if (PHINode *PHI = dyn_cast<PHINode>(V)) {
1594 if (PHI->getParent() == PredBB)
1595 return dyn_cast<Constant>(PHI->getIncomingValueForBlock(PredPredBB));
1596 return nullptr;
1597 }
1598
1599 // If we have a CmpInst, try to fold it for each incoming edge into PredBB.
1600 if (CmpInst *CondCmp = dyn_cast<CmpInst>(V)) {
1601 if (CondCmp->getParent() == BB) {
1602 Constant *Op0 =
1603 evaluateOnPredecessorEdge(BB, PredPredBB, CondCmp->getOperand(0));
1604 Constant *Op1 =
1605 evaluateOnPredecessorEdge(BB, PredPredBB, CondCmp->getOperand(1));
1606 if (Op0 && Op1) {
1607 return ConstantExpr::getCompare(CondCmp->getPredicate(), Op0, Op1);
1608 }
1609 }
1610 return nullptr;
1611 }
1612
1613 return nullptr;
1614}
1615
1616bool JumpThreadingPass::processThreadableEdges(Value *Cond, BasicBlock *BB,
1617 ConstantPreference Preference,
1618 Instruction *CxtI) {
1619 // If threading this would thread across a loop header, don't even try to
1620 // thread the edge.
1621 if (LoopHeaders.count(BB))
1622 return false;
1623
1624 PredValueInfoTy PredValues;
1625 if (!computeValueKnownInPredecessors(Cond, BB, PredValues, Preference,
1626 CxtI)) {
1627 // We don't have known values in predecessors. See if we can thread through
1628 // BB and its sole predecessor.
1629 return maybethreadThroughTwoBasicBlocks(BB, Cond);
1630 }
1631
1632 assert(!PredValues.empty() &&((!PredValues.empty() && "computeValueKnownInPredecessors returned true with no values"
) ? static_cast<void> (0) : __assert_fail ("!PredValues.empty() && \"computeValueKnownInPredecessors returned true with no values\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1633, __PRETTY_FUNCTION__))
1633 "computeValueKnownInPredecessors returned true with no values")((!PredValues.empty() && "computeValueKnownInPredecessors returned true with no values"
) ? static_cast<void> (0) : __assert_fail ("!PredValues.empty() && \"computeValueKnownInPredecessors returned true with no values\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1633, __PRETTY_FUNCTION__))
;
1634
1635 LLVM_DEBUG(dbgs() << "IN BB: " << *BB;do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "IN BB: " << *BB;
for (const auto &PredValue : PredValues) { dbgs() <<
" BB '" << BB->getName() << "': FOUND condition = "
<< *PredValue.first << " for pred '" << PredValue
.second->getName() << "'.\n"; }; } } while (false)
1636 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)
1637 dbgs() << " BB '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "IN BB: " << *BB;
for (const auto &PredValue : PredValues) { dbgs() <<
" BB '" << BB->getName() << "': FOUND condition = "
<< *PredValue.first << " for pred '" << PredValue
.second->getName() << "'.\n"; }; } } while (false)
1638 << "': FOUND condition = " << *PredValue.firstdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "IN BB: " << *BB;
for (const auto &PredValue : PredValues) { dbgs() <<
" BB '" << BB->getName() << "': FOUND condition = "
<< *PredValue.first << " for pred '" << PredValue
.second->getName() << "'.\n"; }; } } while (false)
1639 << " 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)
1640 })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)
;
1641
1642 // Decide what we want to thread through. Convert our list of known values to
1643 // a list of known destinations for each pred. This also discards duplicate
1644 // predecessors and keeps track of the undefined inputs (which are represented
1645 // as a null dest in the PredToDestList).
1646 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1647 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1648
1649 BasicBlock *OnlyDest = nullptr;
1650 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1651 Constant *OnlyVal = nullptr;
1652 Constant *MultipleVal = (Constant *)(intptr_t)~0ULL;
1653
1654 for (const auto &PredValue : PredValues) {
1655 BasicBlock *Pred = PredValue.second;
1656 if (!SeenPreds.insert(Pred).second)
1657 continue; // Duplicate predecessor entry.
1658
1659 Constant *Val = PredValue.first;
1660
1661 BasicBlock *DestBB;
1662 if (isa<UndefValue>(Val))
1663 DestBB = nullptr;
1664 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
1665 assert(isa<ConstantInt>(Val) && "Expecting a constant integer")((isa<ConstantInt>(Val) && "Expecting a constant integer"
) ? static_cast<void> (0) : __assert_fail ("isa<ConstantInt>(Val) && \"Expecting a constant integer\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1665, __PRETTY_FUNCTION__))
;
1666 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1667 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1668 assert(isa<ConstantInt>(Val) && "Expecting a constant integer")((isa<ConstantInt>(Val) && "Expecting a constant integer"
) ? static_cast<void> (0) : __assert_fail ("isa<ConstantInt>(Val) && \"Expecting a constant integer\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1668, __PRETTY_FUNCTION__))
;
1669 DestBB = SI->findCaseValue(cast<ConstantInt>(Val))->getCaseSuccessor();
1670 } else {
1671 assert(isa<IndirectBrInst>(BB->getTerminator())((isa<IndirectBrInst>(BB->getTerminator()) &&
"Unexpected terminator") ? static_cast<void> (0) : __assert_fail
("isa<IndirectBrInst>(BB->getTerminator()) && \"Unexpected terminator\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1672, __PRETTY_FUNCTION__))
1672 && "Unexpected terminator")((isa<IndirectBrInst>(BB->getTerminator()) &&
"Unexpected terminator") ? static_cast<void> (0) : __assert_fail
("isa<IndirectBrInst>(BB->getTerminator()) && \"Unexpected terminator\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1672, __PRETTY_FUNCTION__))
;
1673 assert(isa<BlockAddress>(Val) && "Expecting a constant blockaddress")((isa<BlockAddress>(Val) && "Expecting a constant blockaddress"
) ? static_cast<void> (0) : __assert_fail ("isa<BlockAddress>(Val) && \"Expecting a constant blockaddress\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1673, __PRETTY_FUNCTION__))
;
1674 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1675 }
1676
1677 // If we have exactly one destination, remember it for efficiency below.
1678 if (PredToDestList.empty()) {
1679 OnlyDest = DestBB;
1680 OnlyVal = Val;
1681 } else {
1682 if (OnlyDest != DestBB)
1683 OnlyDest = MultipleDestSentinel;
1684 // It possible we have same destination, but different value, e.g. default
1685 // case in switchinst.
1686 if (Val != OnlyVal)
1687 OnlyVal = MultipleVal;
1688 }
1689
1690 // If the predecessor ends with an indirect goto, we can't change its
1691 // destination. Same for CallBr.
1692 if (isa<IndirectBrInst>(Pred->getTerminator()) ||
1693 isa<CallBrInst>(Pred->getTerminator()))
1694 continue;
1695
1696 PredToDestList.emplace_back(Pred, DestBB);
1697 }
1698
1699 // If all edges were unthreadable, we fail.
1700 if (PredToDestList.empty())
1701 return false;
1702
1703 // If all the predecessors go to a single known successor, we want to fold,
1704 // not thread. By doing so, we do not need to duplicate the current block and
1705 // also miss potential opportunities in case we dont/cant duplicate.
1706 if (OnlyDest && OnlyDest != MultipleDestSentinel) {
1707 if (BB->hasNPredecessors(PredToDestList.size())) {
1708 bool SeenFirstBranchToOnlyDest = false;
1709 std::vector <DominatorTree::UpdateType> Updates;
1710 Updates.reserve(BB->getTerminator()->getNumSuccessors() - 1);
1711 for (BasicBlock *SuccBB : successors(BB)) {
1712 if (SuccBB == OnlyDest && !SeenFirstBranchToOnlyDest) {
1713 SeenFirstBranchToOnlyDest = true; // Don't modify the first branch.
1714 } else {
1715 SuccBB->removePredecessor(BB, true); // This is unreachable successor.
1716 Updates.push_back({DominatorTree::Delete, BB, SuccBB});
1717 }
1718 }
1719
1720 // Finally update the terminator.
1721 Instruction *Term = BB->getTerminator();
1722 BranchInst::Create(OnlyDest, Term);
1723 Term->eraseFromParent();
1724 DTU->applyUpdatesPermissive(Updates);
1725 if (HasProfileData)
1726 BPI->eraseBlock(BB);
1727
1728 // If the condition is now dead due to the removal of the old terminator,
1729 // erase it.
1730 if (auto *CondInst = dyn_cast<Instruction>(Cond)) {
1731 if (CondInst->use_empty() && !CondInst->mayHaveSideEffects())
1732 CondInst->eraseFromParent();
1733 // We can safely replace *some* uses of the CondInst if it has
1734 // exactly one value as returned by LVI. RAUW is incorrect in the
1735 // presence of guards and assumes, that have the `Cond` as the use. This
1736 // is because we use the guards/assume to reason about the `Cond` value
1737 // at the end of block, but RAUW unconditionally replaces all uses
1738 // including the guards/assumes themselves and the uses before the
1739 // guard/assume.
1740 else if (OnlyVal && OnlyVal != MultipleVal &&
1741 CondInst->getParent() == BB)
1742 replaceFoldableUses(CondInst, OnlyVal);
1743 }
1744 return true;
1745 }
1746 }
1747
1748 // Determine which is the most common successor. If we have many inputs and
1749 // this block is a switch, we want to start by threading the batch that goes
1750 // to the most popular destination first. If we only know about one
1751 // threadable destination (the common case) we can avoid this.
1752 BasicBlock *MostPopularDest = OnlyDest;
1753
1754 if (MostPopularDest == MultipleDestSentinel) {
1755 // Remove any loop headers from the Dest list, threadEdge conservatively
1756 // won't process them, but we might have other destination that are eligible
1757 // and we still want to process.
1758 erase_if(PredToDestList,
1759 [&](const std::pair<BasicBlock *, BasicBlock *> &PredToDest) {
1760 return LoopHeaders.contains(PredToDest.second);
1761 });
1762
1763 if (PredToDestList.empty())
1764 return false;
1765
1766 MostPopularDest = findMostPopularDest(BB, PredToDestList);
1767 }
1768
1769 // Now that we know what the most popular destination is, factor all
1770 // predecessors that will jump to it into a single predecessor.
1771 SmallVector<BasicBlock*, 16> PredsToFactor;
1772 for (const auto &PredToDest : PredToDestList)
1773 if (PredToDest.second == MostPopularDest) {
1774 BasicBlock *Pred = PredToDest.first;
1775
1776 // This predecessor may be a switch or something else that has multiple
1777 // edges to the block. Factor each of these edges by listing them
1778 // according to # occurrences in PredsToFactor.
1779 for (BasicBlock *Succ : successors(Pred))
1780 if (Succ == BB)
1781 PredsToFactor.push_back(Pred);
1782 }
1783
1784 // If the threadable edges are branching on an undefined value, we get to pick
1785 // the destination that these predecessors should get to.
1786 if (!MostPopularDest)
1787 MostPopularDest = BB->getTerminator()->
1788 getSuccessor(getBestDestForJumpOnUndef(BB));
1789
1790 // Ok, try to thread it!
1791 return tryThreadEdge(BB, PredsToFactor, MostPopularDest);
1792}
1793
1794/// processBranchOnPHI - We have an otherwise unthreadable conditional branch on
1795/// a PHI node (or freeze PHI) in the current block. See if there are any
1796/// simplifications we can do based on inputs to the phi node.
1797bool JumpThreadingPass::processBranchOnPHI(PHINode *PN) {
1798 BasicBlock *BB = PN->getParent();
1799
1800 // TODO: We could make use of this to do it once for blocks with common PHI
1801 // values.
1802 SmallVector<BasicBlock*, 1> PredBBs;
1803 PredBBs.resize(1);
1804
1805 // If any of the predecessor blocks end in an unconditional branch, we can
1806 // *duplicate* the conditional branch into that block in order to further
1807 // encourage jump threading and to eliminate cases where we have branch on a
1808 // phi of an icmp (branch on icmp is much better).
1809 // This is still beneficial when a frozen phi is used as the branch condition
1810 // because it allows CodeGenPrepare to further canonicalize br(freeze(icmp))
1811 // to br(icmp(freeze ...)).
1812 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1813 BasicBlock *PredBB = PN->getIncomingBlock(i);
1814 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1815 if (PredBr->isUnconditional()) {
1816 PredBBs[0] = PredBB;
1817 // Try to duplicate BB into PredBB.
1818 if (duplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1819 return true;
1820 }
1821 }
1822
1823 return false;
1824}
1825
1826/// processBranchOnXOR - We have an otherwise unthreadable conditional branch on
1827/// a xor instruction in the current block. See if there are any
1828/// simplifications we can do based on inputs to the xor.
1829bool JumpThreadingPass::processBranchOnXOR(BinaryOperator *BO) {
1830 BasicBlock *BB = BO->getParent();
1831
1832 // If either the LHS or RHS of the xor is a constant, don't do this
1833 // optimization.
1834 if (isa<ConstantInt>(BO->getOperand(0)) ||
1835 isa<ConstantInt>(BO->getOperand(1)))
1836 return false;
1837
1838 // If the first instruction in BB isn't a phi, we won't be able to infer
1839 // anything special about any particular predecessor.
1840 if (!isa<PHINode>(BB->front()))
1841 return false;
1842
1843 // If this BB is a landing pad, we won't be able to split the edge into it.
1844 if (BB->isEHPad())
1845 return false;
1846
1847 // If we have a xor as the branch input to this block, and we know that the
1848 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1849 // the condition into the predecessor and fix that value to true, saving some
1850 // logical ops on that path and encouraging other paths to simplify.
1851 //
1852 // This copies something like this:
1853 //
1854 // BB:
1855 // %X = phi i1 [1], [%X']
1856 // %Y = icmp eq i32 %A, %B
1857 // %Z = xor i1 %X, %Y
1858 // br i1 %Z, ...
1859 //
1860 // Into:
1861 // BB':
1862 // %Y = icmp ne i32 %A, %B
1863 // br i1 %Y, ...
1864
1865 PredValueInfoTy XorOpValues;
1866 bool isLHS = true;
1867 if (!computeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1868 WantInteger, BO)) {
1869 assert(XorOpValues.empty())((XorOpValues.empty()) ? static_cast<void> (0) : __assert_fail
("XorOpValues.empty()", "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1869, __PRETTY_FUNCTION__))
;
1870 if (!computeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1871 WantInteger, BO))
1872 return false;
1873 isLHS = false;
1874 }
1875
1876 assert(!XorOpValues.empty() &&((!XorOpValues.empty() && "computeValueKnownInPredecessors returned true with no values"
) ? static_cast<void> (0) : __assert_fail ("!XorOpValues.empty() && \"computeValueKnownInPredecessors returned true with no values\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1877, __PRETTY_FUNCTION__))
1877 "computeValueKnownInPredecessors returned true with no values")((!XorOpValues.empty() && "computeValueKnownInPredecessors returned true with no values"
) ? static_cast<void> (0) : __assert_fail ("!XorOpValues.empty() && \"computeValueKnownInPredecessors returned true with no values\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1877, __PRETTY_FUNCTION__))
;
1878
1879 // Scan the information to see which is most popular: true or false. The
1880 // predecessors can be of the set true, false, or undef.
1881 unsigned NumTrue = 0, NumFalse = 0;
1882 for (const auto &XorOpValue : XorOpValues) {
1883 if (isa<UndefValue>(XorOpValue.first))
1884 // Ignore undefs for the count.
1885 continue;
1886 if (cast<ConstantInt>(XorOpValue.first)->isZero())
1887 ++NumFalse;
1888 else
1889 ++NumTrue;
1890 }
1891
1892 // Determine which value to split on, true, false, or undef if neither.
1893 ConstantInt *SplitVal = nullptr;
1894 if (NumTrue > NumFalse)
1895 SplitVal = ConstantInt::getTrue(BB->getContext());
1896 else if (NumTrue != 0 || NumFalse != 0)
1897 SplitVal = ConstantInt::getFalse(BB->getContext());
1898
1899 // Collect all of the blocks that this can be folded into so that we can
1900 // factor this once and clone it once.
1901 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1902 for (const auto &XorOpValue : XorOpValues) {
1903 if (XorOpValue.first != SplitVal && !isa<UndefValue>(XorOpValue.first))
1904 continue;
1905
1906 BlocksToFoldInto.push_back(XorOpValue.second);
1907 }
1908
1909 // If we inferred a value for all of the predecessors, then duplication won't
1910 // help us. However, we can just replace the LHS or RHS with the constant.
1911 if (BlocksToFoldInto.size() ==
1912 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1913 if (!SplitVal) {
1914 // If all preds provide undef, just nuke the xor, because it is undef too.
1915 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1916 BO->eraseFromParent();
1917 } else if (SplitVal->isZero()) {
1918 // If all preds provide 0, replace the xor with the other input.
1919 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1920 BO->eraseFromParent();
1921 } else {
1922 // If all preds provide 1, set the computed value to 1.
1923 BO->setOperand(!isLHS, SplitVal);
1924 }
1925
1926 return true;
1927 }
1928
1929 // If any of predecessors end with an indirect goto, we can't change its
1930 // destination. Same for CallBr.
1931 if (any_of(BlocksToFoldInto, [](BasicBlock *Pred) {
1932 return isa<IndirectBrInst>(Pred->getTerminator()) ||
1933 isa<CallBrInst>(Pred->getTerminator());
1934 }))
1935 return false;
1936
1937 // Try to duplicate BB into PredBB.
1938 return duplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1939}
1940
1941/// addPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1942/// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1943/// NewPred using the entries from OldPred (suitably mapped).
1944static void addPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1945 BasicBlock *OldPred,
1946 BasicBlock *NewPred,
1947 DenseMap<Instruction*, Value*> &ValueMap) {
1948 for (PHINode &PN : PHIBB->phis()) {
1949 // Ok, we have a PHI node. Figure out what the incoming value was for the
1950 // DestBlock.
1951 Value *IV = PN.getIncomingValueForBlock(OldPred);
1952
1953 // Remap the value if necessary.
1954 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1955 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1956 if (I != ValueMap.end())
1957 IV = I->second;
1958 }
1959
1960 PN.addIncoming(IV, NewPred);
1961 }
1962}
1963
1964/// Merge basic block BB into its sole predecessor if possible.
1965bool JumpThreadingPass::maybeMergeBasicBlockIntoOnlyPred(BasicBlock *BB) {
1966 BasicBlock *SinglePred = BB->getSinglePredecessor();
1967 if (!SinglePred)
1968 return false;
1969
1970 const Instruction *TI = SinglePred->getTerminator();
1971 if (TI->isExceptionalTerminator() || TI->getNumSuccessors() != 1 ||
1972 SinglePred == BB || hasAddressTakenAndUsed(BB))
1973 return false;
1974
1975 // If SinglePred was a loop header, BB becomes one.
1976 if (LoopHeaders.erase(SinglePred))
1977 LoopHeaders.insert(BB);
1978
1979 LVI->eraseBlock(SinglePred);
1980 MergeBasicBlockIntoOnlyPred(BB, DTU);
1981
1982 // Now that BB is merged into SinglePred (i.e. SinglePred code followed by
1983 // BB code within one basic block `BB`), we need to invalidate the LVI
1984 // information associated with BB, because the LVI information need not be
1985 // true for all of BB after the merge. For example,
1986 // Before the merge, LVI info and code is as follows:
1987 // SinglePred: <LVI info1 for %p val>
1988 // %y = use of %p
1989 // call @exit() // need not transfer execution to successor.
1990 // assume(%p) // from this point on %p is true
1991 // br label %BB
1992 // BB: <LVI info2 for %p val, i.e. %p is true>
1993 // %x = use of %p
1994 // br label exit
1995 //
1996 // Note that this LVI info for blocks BB and SinglPred is correct for %p
1997 // (info2 and info1 respectively). After the merge and the deletion of the
1998 // LVI info1 for SinglePred. We have the following code:
1999 // BB: <LVI info2 for %p val>
2000 // %y = use of %p
2001 // call @exit()
2002 // assume(%p)
2003 // %x = use of %p <-- LVI info2 is correct from here onwards.
2004 // br label exit
2005 // LVI info2 for BB is incorrect at the beginning of BB.
2006
2007 // Invalidate LVI information for BB if the LVI is not provably true for
2008 // all of BB.
2009 if (!isGuaranteedToTransferExecutionToSuccessor(BB))
2010 LVI->eraseBlock(BB);
2011 return true;
2012}
2013
2014/// Update the SSA form. NewBB contains instructions that are copied from BB.
2015/// ValueMapping maps old values in BB to new ones in NewBB.
2016void JumpThreadingPass::updateSSA(
2017 BasicBlock *BB, BasicBlock *NewBB,
2018 DenseMap<Instruction *, Value *> &ValueMapping) {
2019 // If there were values defined in BB that are used outside the block, then we
2020 // now have to update all uses of the value to use either the original value,
2021 // the cloned value, or some PHI derived value. This can require arbitrary
2022 // PHI insertion, of which we are prepared to do, clean these up now.
2023 SSAUpdater SSAUpdate;
2024 SmallVector<Use *, 16> UsesToRename;
2025
2026 for (Instruction &I : *BB) {
2027 // Scan all uses of this instruction to see if it is used outside of its
2028 // block, and if so, record them in UsesToRename.
2029 for (Use &U : I.uses()) {
2030 Instruction *User = cast<Instruction>(U.getUser());
2031 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
2032 if (UserPN->getIncomingBlock(U) == BB)
2033 continue;
2034 } else if (User->getParent() == BB)
2035 continue;
2036
2037 UsesToRename.push_back(&U);
2038 }
2039
2040 // If there are no uses outside the block, we're done with this instruction.
2041 if (UsesToRename.empty())
2042 continue;
2043 LLVM_DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "JT: Renaming non-local uses of: "
<< I << "\n"; } } while (false)
;
2044
2045 // We found a use of I outside of BB. Rename all uses of I that are outside
2046 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
2047 // with the two values we know.
2048 SSAUpdate.Initialize(I.getType(), I.getName());
2049 SSAUpdate.AddAvailableValue(BB, &I);
2050 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[&I]);
2051
2052 while (!UsesToRename.empty())
2053 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
2054 LLVM_DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "\n"; } } while (false)
;
2055 }
2056}
2057
2058/// Clone instructions in range [BI, BE) to NewBB. For PHI nodes, we only clone
2059/// arguments that come from PredBB. Return the map from the variables in the
2060/// source basic block to the variables in the newly created basic block.
2061DenseMap<Instruction *, Value *>
2062JumpThreadingPass::cloneInstructions(BasicBlock::iterator BI,
2063 BasicBlock::iterator BE, BasicBlock *NewBB,
2064 BasicBlock *PredBB) {
2065 // We are going to have to map operands from the source basic block to the new
2066 // copy of the block 'NewBB'. If there are PHI nodes in the source basic
2067 // block, evaluate them to account for entry from PredBB.
2068 DenseMap<Instruction *, Value *> ValueMapping;
2069
2070 // Clone the phi nodes of the source basic block into NewBB. The resulting
2071 // phi nodes are trivial since NewBB only has one predecessor, but SSAUpdater
2072 // might need to rewrite the operand of the cloned phi.
2073 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) {
2074 PHINode *NewPN = PHINode::Create(PN->getType(), 1, PN->getName(), NewBB);
2075 NewPN->addIncoming(PN->getIncomingValueForBlock(PredBB), PredBB);
2076 ValueMapping[PN] = NewPN;
2077 }
2078
2079 // Clone the non-phi instructions of the source basic block into NewBB,
2080 // keeping track of the mapping and using it to remap operands in the cloned
2081 // instructions.
2082 for (; BI != BE; ++BI) {
2083 Instruction *New = BI->clone();
2084 New->setName(BI->getName());
2085 NewBB->getInstList().push_back(New);
2086 ValueMapping[&*BI] = New;
2087
2088 // Remap operands to patch up intra-block references.
2089 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2090 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
2091 DenseMap<Instruction *, Value *>::iterator I = ValueMapping.find(Inst);
2092 if (I != ValueMapping.end())
2093 New->setOperand(i, I->second);
2094 }
2095 }
2096
2097 return ValueMapping;
2098}
2099
2100/// Attempt to thread through two successive basic blocks.
2101bool JumpThreadingPass::maybethreadThroughTwoBasicBlocks(BasicBlock *BB,
2102 Value *Cond) {
2103 // Consider:
2104 //
2105 // PredBB:
2106 // %var = phi i32* [ null, %bb1 ], [ @a, %bb2 ]
2107 // %tobool = icmp eq i32 %cond, 0
2108 // br i1 %tobool, label %BB, label ...
2109 //
2110 // BB:
2111 // %cmp = icmp eq i32* %var, null
2112 // br i1 %cmp, label ..., label ...
2113 //
2114 // We don't know the value of %var at BB even if we know which incoming edge
2115 // we take to BB. However, once we duplicate PredBB for each of its incoming
2116 // edges (say, PredBB1 and PredBB2), we know the value of %var in each copy of
2117 // PredBB. Then we can thread edges PredBB1->BB and PredBB2->BB through BB.
2118
2119 // Require that BB end with a Branch for simplicity.
2120 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
2121 if (!CondBr)
2122 return false;
2123
2124 // BB must have exactly one predecessor.
2125 BasicBlock *PredBB = BB->getSinglePredecessor();
2126 if (!PredBB)
2127 return false;
2128
2129 // Require that PredBB end with a conditional Branch. If PredBB ends with an
2130 // unconditional branch, we should be merging PredBB and BB instead. For
2131 // simplicity, we don't deal with a switch.
2132 BranchInst *PredBBBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
2133 if (!PredBBBranch || PredBBBranch->isUnconditional())
2134 return false;
2135
2136 // If PredBB has exactly one incoming edge, we don't gain anything by copying
2137 // PredBB.
2138 if (PredBB->getSinglePredecessor())
2139 return false;
2140
2141 // Don't thread through PredBB if it contains a successor edge to itself, in
2142 // which case we would infinite loop. Suppose we are threading an edge from
2143 // PredPredBB through PredBB and BB to SuccBB with PredBB containing a
2144 // successor edge to itself. If we allowed jump threading in this case, we
2145 // could duplicate PredBB and BB as, say, PredBB.thread and BB.thread. Since
2146 // PredBB.thread has a successor edge to PredBB, we would immediately come up
2147 // with another jump threading opportunity from PredBB.thread through PredBB
2148 // and BB to SuccBB. This jump threading would repeatedly occur. That is, we
2149 // would keep peeling one iteration from PredBB.
2150 if (llvm::is_contained(successors(PredBB), PredBB))
2151 return false;
2152
2153 // Don't thread across a loop header.
2154 if (LoopHeaders.count(PredBB))
2155 return false;
2156
2157 // Avoid complication with duplicating EH pads.
2158 if (PredBB->isEHPad())
2159 return false;
2160
2161 // Find a predecessor that we can thread. For simplicity, we only consider a
2162 // successor edge out of BB to which we thread exactly one incoming edge into
2163 // PredBB.
2164 unsigned ZeroCount = 0;
2165 unsigned OneCount = 0;
2166 BasicBlock *ZeroPred = nullptr;
2167 BasicBlock *OnePred = nullptr;
2168 for (BasicBlock *P : predecessors(PredBB)) {
2169 if (ConstantInt *CI = dyn_cast_or_null<ConstantInt>(
2170 evaluateOnPredecessorEdge(BB, P, Cond))) {
2171 if (CI->isZero()) {
2172 ZeroCount++;
2173 ZeroPred = P;
2174 } else if (CI->isOne()) {
2175 OneCount++;
2176 OnePred = P;
2177 }
2178 }
2179 }
2180
2181 // Disregard complicated cases where we have to thread multiple edges.
2182 BasicBlock *PredPredBB;
2183 if (ZeroCount == 1) {
2184 PredPredBB = ZeroPred;
2185 } else if (OneCount == 1) {
2186 PredPredBB = OnePred;
2187 } else {
2188 return false;
2189 }
2190
2191 BasicBlock *SuccBB = CondBr->getSuccessor(PredPredBB == ZeroPred);
2192
2193 // If threading to the same block as we come from, we would infinite loop.
2194 if (SuccBB == BB) {
2195 LLVM_DEBUG(dbgs() << " Not threading across BB '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Not threading across BB '"
<< BB->getName() << "' - would thread to self!\n"
; } } while (false)
2196 << "' - 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)
;
2197 return false;
2198 }
2199
2200 // If threading this would thread across a loop header, don't thread the edge.
2201 // See the comments above findLoopHeaders for justifications and caveats.
2202 if (LoopHeaders.count(BB) || LoopHeaders.count(SuccBB)) {
2203 LLVM_DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB
); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() <<
" Not threading across " << (BBIsHeader ? "loop header BB '"
: "block BB '") << BB->getName() << "' to dest "
<< (SuccIsHeader ? "loop header BB '" : "block BB '") <<
SuccBB->getName() << "' - it might create an irreducible loop!\n"
; }; } } while (false)
2204 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)
2205 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)
2206 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)
2207 << (BBIsHeader ? "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)
2208 << BB->getName() << "' to dest "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)
2209 << (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)
2210 << SuccBB->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)
2211 << "' - 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)
2212 })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)
;
2213 return false;
2214 }
2215
2216 // Compute the cost of duplicating BB and PredBB.
2217 unsigned BBCost =
2218 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
2219 unsigned PredBBCost = getJumpThreadDuplicationCost(
2220 PredBB, PredBB->getTerminator(), BBDupThreshold);
2221
2222 // Give up if costs are too high. We need to check BBCost and PredBBCost
2223 // individually before checking their sum because getJumpThreadDuplicationCost
2224 // return (unsigned)~0 for those basic blocks that cannot be duplicated.
2225 if (BBCost > BBDupThreshold || PredBBCost > BBDupThreshold ||
2226 BBCost + PredBBCost > BBDupThreshold) {
2227 LLVM_DEBUG(dbgs() << " Not threading BB '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Not threading BB '" <<
BB->getName() << "' - Cost is too high: " << PredBBCost
<< " for PredBB, " << BBCost << "for BB\n"
; } } while (false)
2228 << "' - Cost is too high: " << PredBBCostdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Not threading BB '" <<
BB->getName() << "' - Cost is too high: " << PredBBCost
<< " for PredBB, " << BBCost << "for BB\n"
; } } while (false)
2229 << " for PredBB, " << BBCost << "for BB\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Not threading BB '" <<
BB->getName() << "' - Cost is too high: " << PredBBCost
<< " for PredBB, " << BBCost << "for BB\n"
; } } while (false)
;
2230 return false;
2231 }
2232
2233 // Now we are ready to duplicate PredBB.
2234 threadThroughTwoBasicBlocks(PredPredBB, PredBB, BB, SuccBB);
2235 return true;
2236}
2237
2238void JumpThreadingPass::threadThroughTwoBasicBlocks(BasicBlock *PredPredBB,
2239 BasicBlock *PredBB,
2240 BasicBlock *BB,
2241 BasicBlock *SuccBB) {
2242 LLVM_DEBUG(dbgs() << " Threading through '" << PredBB->getName() << "' and '"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Threading through '"
<< PredBB->getName() << "' and '" << BB
->getName() << "'\n"; } } while (false)
2243 << BB->getName() << "'\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Threading through '"
<< PredBB->getName() << "' and '" << BB
->getName() << "'\n"; } } while (false)
;
2244
2245 BranchInst *CondBr = cast<BranchInst>(BB->getTerminator());
2246 BranchInst *PredBBBranch = cast<BranchInst>(PredBB->getTerminator());
2247
2248 BasicBlock *NewBB =
2249 BasicBlock::Create(PredBB->getContext(), PredBB->getName() + ".thread",
2250 PredBB->getParent(), PredBB);
2251 NewBB->moveAfter(PredBB);
2252
2253 // Set the block frequency of NewBB.
2254 if (HasProfileData) {
2255 auto NewBBFreq = BFI->getBlockFreq(PredPredBB) *
2256 BPI->getEdgeProbability(PredPredBB, PredBB);
2257 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
2258 }
2259
2260 // We are going to have to map operands from the original BB block to the new
2261 // copy of the block 'NewBB'. If there are PHI nodes in PredBB, evaluate them
2262 // to account for entry from PredPredBB.
2263 DenseMap<Instruction *, Value *> ValueMapping =
2264 cloneInstructions(PredBB->begin(), PredBB->end(), NewBB, PredPredBB);
2265
2266 // Copy the edge probabilities from PredBB to NewBB.
2267 if (HasProfileData)
2268 BPI->copyEdgeProbabilities(PredBB, NewBB);
2269
2270 // Update the terminator of PredPredBB to jump to NewBB instead of PredBB.
2271 // This eliminates predecessors from PredPredBB, which requires us to simplify
2272 // any PHI nodes in PredBB.
2273 Instruction *PredPredTerm = PredPredBB->getTerminator();
2274 for (unsigned i = 0, e = PredPredTerm->getNumSuccessors(); i != e; ++i)
2275 if (PredPredTerm->getSuccessor(i) == PredBB) {
2276 PredBB->removePredecessor(PredPredBB, true);
2277 PredPredTerm->setSuccessor(i, NewBB);
2278 }
2279
2280 addPHINodeEntriesForMappedBlock(PredBBBranch->getSuccessor(0), PredBB, NewBB,
2281 ValueMapping);
2282 addPHINodeEntriesForMappedBlock(PredBBBranch->getSuccessor(1), PredBB, NewBB,
2283 ValueMapping);
2284
2285 DTU->applyUpdatesPermissive(
2286 {{DominatorTree::Insert, NewBB, CondBr->getSuccessor(0)},
2287 {DominatorTree::Insert, NewBB, CondBr->getSuccessor(1)},
2288 {DominatorTree::Insert, PredPredBB, NewBB},
2289 {DominatorTree::Delete, PredPredBB, PredBB}});
2290
2291 updateSSA(PredBB, NewBB, ValueMapping);
2292
2293 // Clean up things like PHI nodes with single operands, dead instructions,
2294 // etc.
2295 SimplifyInstructionsInBlock(NewBB, TLI);
2296 SimplifyInstructionsInBlock(PredBB, TLI);
2297
2298 SmallVector<BasicBlock *, 1> PredsToFactor;
2299 PredsToFactor.push_back(NewBB);
2300 threadEdge(BB, PredsToFactor, SuccBB);
2301}
2302
2303/// tryThreadEdge - Thread an edge if it's safe and profitable to do so.
2304bool JumpThreadingPass::tryThreadEdge(
2305 BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs,
2306 BasicBlock *SuccBB) {
2307 // If threading to the same block as we come from, we would infinite loop.
2308 if (SuccBB == BB) {
2309 LLVM_DEBUG(dbgs() << " Not threading across BB '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Not threading across BB '"
<< BB->getName() << "' - would thread to self!\n"
; } } while (false)
2310 << "' - 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)
;
2311 return false;
2312 }
2313
2314 // If threading this would thread across a loop header, don't thread the edge.
2315 // See the comments above findLoopHeaders for justifications and caveats.
2316 if (LoopHeaders.count(BB) || LoopHeaders.count(SuccBB)) {
2317 LLVM_DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB
); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() <<
" Not threading across " << (BBIsHeader ? "loop header BB '"
: "block BB '") << BB->getName() << "' to dest "
<< (SuccIsHeader ? "loop header BB '" : "block BB '") <<
SuccBB->getName() << "' - it might create an irreducible loop!\n"
; }; } } while (false)
2318 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)
2319 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)
2320 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)
2321 << (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)
2322 << "' 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)
2323 << 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)
2324 })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)
;
2325 return false;
2326 }
2327
2328 unsigned JumpThreadCost =
2329 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
2330 if (JumpThreadCost > BBDupThreshold) {
2331 LLVM_DEBUG(dbgs() << " Not threading BB '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Not threading BB '" <<
BB->getName() << "' - Cost is too high: " << JumpThreadCost
<< "\n"; } } while (false)
2332 << "' - 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)
;
2333 return false;
2334 }
2335
2336 threadEdge(BB, PredBBs, SuccBB);
2337 return true;
2338}
2339
2340/// threadEdge - We have decided that it is safe and profitable to factor the
2341/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
2342/// across BB. Transform the IR to reflect this change.
2343void JumpThreadingPass::threadEdge(BasicBlock *BB,
2344 const SmallVectorImpl<BasicBlock *> &PredBBs,
2345 BasicBlock *SuccBB) {
2346 assert(SuccBB != BB && "Don't create an infinite loop")((SuccBB != BB && "Don't create an infinite loop") ? static_cast
<void> (0) : __assert_fail ("SuccBB != BB && \"Don't create an infinite loop\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 2346, __PRETTY_FUNCTION__))
;
2347
2348 assert(!LoopHeaders.count(BB) && !LoopHeaders.count(SuccBB) &&((!LoopHeaders.count(BB) && !LoopHeaders.count(SuccBB
) && "Don't thread across loop headers") ? static_cast
<void> (0) : __assert_fail ("!LoopHeaders.count(BB) && !LoopHeaders.count(SuccBB) && \"Don't thread across loop headers\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 2349, __PRETTY_FUNCTION__))
2349 "Don't thread across loop headers")((!LoopHeaders.count(BB) && !LoopHeaders.count(SuccBB
) && "Don't thread across loop headers") ? static_cast
<void> (0) : __assert_fail ("!LoopHeaders.count(BB) && !LoopHeaders.count(SuccBB) && \"Don't thread across loop headers\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 2349, __PRETTY_FUNCTION__))
;
2350
2351 // And finally, do it! Start by factoring the predecessors if needed.
2352 BasicBlock *PredBB;
2353 if (PredBBs.size() == 1)
2354 PredBB = PredBBs[0];
2355 else {
2356 LLVM_DEBUG(dbgs() << " Factoring out " << PredBBs.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Factoring out " <<
PredBBs.size() << " common predecessors.\n"; } } while
(false)
2357 << " common predecessors.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Factoring out " <<
PredBBs.size() << " common predecessors.\n"; } } while
(false)
;
2358 PredBB = splitBlockPreds(BB, PredBBs, ".thr_comm");
2359 }
2360
2361 // And finally, do it!
2362 LLVM_DEBUG(dbgs() << " Threading edge from '" << PredBB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Threading edge from '"
<< PredBB->getName() << "' to '" << SuccBB
->getName() << ", across block:\n " << *BB <<
"\n"; } } while (false)
2363 << "' to '" << SuccBB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Threading edge from '"
<< PredBB->getName() << "' to '" << SuccBB
->getName() << ", across block:\n " << *BB <<
"\n"; } } while (false)
2364 << ", across block:\n " << *BB << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Threading edge from '"
<< PredBB->getName() << "' to '" << SuccBB
->getName() << ", across block:\n " << *BB <<
"\n"; } } while (false)
;
2365
2366 LVI->threadEdge(PredBB, BB, SuccBB);
2367
2368 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
2369 BB->getName()+".thread",
2370 BB->getParent(), BB);
2371 NewBB->moveAfter(PredBB);
2372
2373 // Set the block frequency of NewBB.
2374 if (HasProfileData) {
2375 auto NewBBFreq =
2376 BFI->getBlockFreq(PredBB) * BPI->getEdgeProbability(PredBB, BB);
2377 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
2378 }
2379
2380 // Copy all the instructions from BB to NewBB except the terminator.
2381 DenseMap<Instruction *, Value *> ValueMapping =
2382 cloneInstructions(BB->begin(), std::prev(BB->end()), NewBB, PredBB);
2383
2384 // We didn't copy the terminator from BB over to NewBB, because there is now
2385 // an unconditional jump to SuccBB. Insert the unconditional jump.
2386 BranchInst *NewBI = BranchInst::Create(SuccBB, NewBB);
2387 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
2388
2389 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
2390 // PHI nodes for NewBB now.
2391 addPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
2392
2393 // Update the terminator of PredBB to jump to NewBB instead of BB. This
2394 // eliminates predecessors from BB, which requires us to simplify any PHI
2395 // nodes in BB.
2396 Instruction *PredTerm = PredBB->getTerminator();
2397 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
2398 if (PredTerm->getSuccessor(i) == BB) {
2399 BB->removePredecessor(PredBB, true);
2400 PredTerm->setSuccessor(i, NewBB);
2401 }
2402
2403 // Enqueue required DT updates.
2404 DTU->applyUpdatesPermissive({{DominatorTree::Insert, NewBB, SuccBB},
2405 {DominatorTree::Insert, PredBB, NewBB},
2406 {DominatorTree::Delete, PredBB, BB}});
2407
2408 updateSSA(BB, NewBB, ValueMapping);
2409
2410 // At this point, the IR is fully up to date and consistent. Do a quick scan
2411 // over the new instructions and zap any that are constants or dead. This
2412 // frequently happens because of phi translation.
2413 SimplifyInstructionsInBlock(NewBB, TLI);
2414
2415 // Update the edge weight from BB to SuccBB, which should be less than before.
2416 updateBlockFreqAndEdgeWeight(PredBB, BB, NewBB, SuccBB);
2417
2418 // Threaded an edge!
2419 ++NumThreads;
2420}
2421
2422/// Create a new basic block that will be the predecessor of BB and successor of
2423/// all blocks in Preds. When profile data is available, update the frequency of
2424/// this new block.
2425BasicBlock *JumpThreadingPass::splitBlockPreds(BasicBlock *BB,
2426 ArrayRef<BasicBlock *> Preds,
2427 const char *Suffix) {
2428 SmallVector<BasicBlock *, 2> NewBBs;
2429
2430 // Collect the frequencies of all predecessors of BB, which will be used to
2431 // update the edge weight of the result of splitting predecessors.
2432 DenseMap<BasicBlock *, BlockFrequency> FreqMap;
2433 if (HasProfileData)
2434 for (auto Pred : Preds)
2435 FreqMap.insert(std::make_pair(
2436 Pred, BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, BB)));
2437
2438 // In the case when BB is a LandingPad block we create 2 new predecessors
2439 // instead of just one.
2440 if (BB->isLandingPad()) {
2441 std::string NewName = std::string(Suffix) + ".split-lp";
2442 SplitLandingPadPredecessors(BB, Preds, Suffix, NewName.c_str(), NewBBs);
2443 } else {
2444 NewBBs.push_back(SplitBlockPredecessors(BB, Preds, Suffix));
2445 }
2446
2447 std::vector<DominatorTree::UpdateType> Updates;
2448 Updates.reserve((2 * Preds.size()) + NewBBs.size());
2449 for (auto NewBB : NewBBs) {
2450 BlockFrequency NewBBFreq(0);
2451 Updates.push_back({DominatorTree::Insert, NewBB, BB});
2452 for (auto Pred : predecessors(NewBB)) {
2453 Updates.push_back({DominatorTree::Delete, Pred, BB});
2454 Updates.push_back({DominatorTree::Insert, Pred, NewBB});
2455 if (HasProfileData) // Update frequencies between Pred -> NewBB.
2456 NewBBFreq += FreqMap.lookup(Pred);
2457 }
2458 if (HasProfileData) // Apply the summed frequency to NewBB.
2459 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
2460 }
2461
2462 DTU->applyUpdatesPermissive(Updates);
2463 return NewBBs[0];
2464}
2465
2466bool JumpThreadingPass::doesBlockHaveProfileData(BasicBlock *BB) {
2467 const Instruction *TI = BB->getTerminator();
2468 assert(TI->getNumSuccessors() > 1 && "not a split")((TI->getNumSuccessors() > 1 && "not a split") ?
static_cast<void> (0) : __assert_fail ("TI->getNumSuccessors() > 1 && \"not a split\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 2468, __PRETTY_FUNCTION__))
;
2469
2470 MDNode *WeightsNode = TI->getMetadata(LLVMContext::MD_prof);
2471 if (!WeightsNode)
2472 return false;
2473
2474 MDString *MDName = cast<MDString>(WeightsNode->getOperand(0));
2475 if (MDName->getString() != "branch_weights")
2476 return false;
2477
2478 // Ensure there are weights for all of the successors. Note that the first
2479 // operand to the metadata node is a name, not a weight.
2480 return WeightsNode->getNumOperands() == TI->getNumSuccessors() + 1;
2481}
2482
2483/// Update the block frequency of BB and branch weight and the metadata on the
2484/// edge BB->SuccBB. This is done by scaling the weight of BB->SuccBB by 1 -
2485/// Freq(PredBB->BB) / Freq(BB->SuccBB).
2486void JumpThreadingPass::updateBlockFreqAndEdgeWeight(BasicBlock *PredBB,
2487 BasicBlock *BB,
2488 BasicBlock *NewBB,
2489 BasicBlock *SuccBB) {
2490 if (!HasProfileData)
2491 return;
2492
2493 assert(BFI && BPI && "BFI & BPI should have been created here")((BFI && BPI && "BFI & BPI should have been created here"
) ? static_cast<void> (0) : __assert_fail ("BFI && BPI && \"BFI & BPI should have been created here\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 2493, __PRETTY_FUNCTION__))
;
2494
2495 // As the edge from PredBB to BB is deleted, we have to update the block
2496 // frequency of BB.
2497 auto BBOrigFreq = BFI->getBlockFreq(BB);
2498 auto NewBBFreq = BFI->getBlockFreq(NewBB);
2499 auto BB2SuccBBFreq = BBOrigFreq * BPI->getEdgeProbability(BB, SuccBB);
2500 auto BBNewFreq = BBOrigFreq - NewBBFreq;
2501 BFI->setBlockFreq(BB, BBNewFreq.getFrequency());
2502
2503 // Collect updated outgoing edges' frequencies from BB and use them to update
2504 // edge probabilities.
2505 SmallVector<uint64_t, 4> BBSuccFreq;
2506 for (BasicBlock *Succ : successors(BB)) {
2507 auto SuccFreq = (Succ == SuccBB)
2508 ? BB2SuccBBFreq - NewBBFreq
2509 : BBOrigFreq * BPI->getEdgeProbability(BB, Succ);
2510 BBSuccFreq.push_back(SuccFreq.getFrequency());
2511 }
2512
2513 uint64_t MaxBBSuccFreq =
2514 *std::max_element(BBSuccFreq.begin(), BBSuccFreq.end());
2515
2516 SmallVector<BranchProbability, 4> BBSuccProbs;
2517 if (MaxBBSuccFreq == 0)
2518 BBSuccProbs.assign(BBSuccFreq.size(),
2519 {1, static_cast<unsigned>(BBSuccFreq.size())});
2520 else {
2521 for (uint64_t Freq : BBSuccFreq)
2522 BBSuccProbs.push_back(
2523 BranchProbability::getBranchProbability(Freq, MaxBBSuccFreq));
2524 // Normalize edge probabilities so that they sum up to one.
2525 BranchProbability::normalizeProbabilities(BBSuccProbs.begin(),
2526 BBSuccProbs.end());
2527 }
2528
2529 // Update edge probabilities in BPI.
2530 BPI->setEdgeProbability(BB, BBSuccProbs);
2531
2532 // Update the profile metadata as well.
2533 //
2534 // Don't do this if the profile of the transformed blocks was statically
2535 // estimated. (This could occur despite the function having an entry
2536 // frequency in completely cold parts of the CFG.)
2537 //
2538 // In this case we don't want to suggest to subsequent passes that the
2539 // calculated weights are fully consistent. Consider this graph:
2540 //
2541 // check_1
2542 // 50% / |
2543 // eq_1 | 50%
2544 // \ |
2545 // check_2
2546 // 50% / |
2547 // eq_2 | 50%
2548 // \ |
2549 // check_3
2550 // 50% / |
2551 // eq_3 | 50%
2552 // \ |
2553 //
2554 // Assuming the blocks check_* all compare the same value against 1, 2 and 3,
2555 // the overall probabilities are inconsistent; the total probability that the
2556 // value is either 1, 2 or 3 is 150%.
2557 //
2558 // As a consequence if we thread eq_1 -> check_2 to check_3, check_2->check_3
2559 // becomes 0%. This is even worse if the edge whose probability becomes 0% is
2560 // the loop exit edge. Then based solely on static estimation we would assume
2561 // the loop was extremely hot.
2562 //
2563 // FIXME this locally as well so that BPI and BFI are consistent as well. We
2564 // shouldn't make edges extremely likely or unlikely based solely on static
2565 // estimation.
2566 if (BBSuccProbs.size() >= 2 && doesBlockHaveProfileData(BB)) {
2567 SmallVector<uint32_t, 4> Weights;
2568 for (auto Prob : BBSuccProbs)
2569 Weights.push_back(Prob.getNumerator());
2570
2571 auto TI = BB->getTerminator();
2572 TI->setMetadata(
2573 LLVMContext::MD_prof,
2574 MDBuilder(TI->getParent()->getContext()).createBranchWeights(Weights));
2575 }
2576}
2577
2578/// duplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
2579/// to BB which contains an i1 PHI node and a conditional branch on that PHI.
2580/// If we can duplicate the contents of BB up into PredBB do so now, this
2581/// improves the odds that the branch will be on an analyzable instruction like
2582/// a compare.
2583bool JumpThreadingPass::duplicateCondBranchOnPHIIntoPred(
2584 BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs) {
2585 assert(!PredBBs.empty() && "Can't handle an empty set")((!PredBBs.empty() && "Can't handle an empty set") ? static_cast
<void> (0) : __assert_fail ("!PredBBs.empty() && \"Can't handle an empty set\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 2585, __PRETTY_FUNCTION__))
;
2586
2587 // If BB is a loop header, then duplicating this block outside the loop would
2588 // cause us to transform this into an irreducible loop, don't do this.
2589 // See the comments above findLoopHeaders for justifications and caveats.
2590 if (LoopHeaders.count(BB)) {
2591 LLVM_DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Not duplicating loop header '"
<< BB->getName() << "' into predecessor block '"
<< PredBBs[0]->getName() << "' - it might create an irreducible loop!\n"
; } } while (false)
2592 << "' 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)
2593 << "' - 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)
;
2594 return false;
2595 }
2596
2597 unsigned DuplicationCost =
2598 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
2599 if (DuplicationCost > BBDupThreshold) {
2600 LLVM_DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Not duplicating BB '"
<< BB->getName() << "' - Cost is too high: " <<
DuplicationCost << "\n"; } } while (false)
2601 << "' - 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)
;
2602 return false;
2603 }
2604
2605 // And finally, do it! Start by factoring the predecessors if needed.
2606 std::vector<DominatorTree::UpdateType> Updates;
2607 BasicBlock *PredBB;
2608 if (PredBBs.size() == 1)
2609 PredBB = PredBBs[0];
2610 else {
2611 LLVM_DEBUG(dbgs() << " Factoring out " << PredBBs.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Factoring out " <<
PredBBs.size() << " common predecessors.\n"; } } while
(false)
2612 << " common predecessors.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Factoring out " <<
PredBBs.size() << " common predecessors.\n"; } } while
(false)
;
2613 PredBB = splitBlockPreds(BB, PredBBs, ".thr_comm");
2614 }
2615 Updates.push_back({DominatorTree::Delete, PredBB, BB});
2616
2617 // Okay, we decided to do this! Clone all the instructions in BB onto the end
2618 // of PredBB.
2619 LLVM_DEBUG(dbgs() << " Duplicating block '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Duplicating block '"
<< BB->getName() << "' into end of '" <<
PredBB->getName() << "' to eliminate branch on phi. Cost: "
<< DuplicationCost << " block is:" << *BB <<
"\n"; } } while (false)
2620 << "' into end of '" << PredBB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Duplicating block '"
<< BB->getName() << "' into end of '" <<
PredBB->getName() << "' to eliminate branch on phi. Cost: "
<< DuplicationCost << " block is:" << *BB <<
"\n"; } } while (false)
2621 << "' 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)
2622 << 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)
;
2623
2624 // Unless PredBB ends with an unconditional branch, split the edge so that we
2625 // can just clone the bits from BB into the end of the new PredBB.
2626 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
2627
2628 if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
2629 BasicBlock *OldPredBB = PredBB;
2630 PredBB = SplitEdge(OldPredBB, BB);
2631 Updates.push_back({DominatorTree::Insert, OldPredBB, PredBB});
2632 Updates.push_back({DominatorTree::Insert, PredBB, BB});
2633 Updates.push_back({DominatorTree::Delete, OldPredBB, BB});
2634 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
2635 }
2636
2637 // We are going to have to map operands from the original BB block into the
2638 // PredBB block. Evaluate PHI nodes in BB.
2639 DenseMap<Instruction*, Value*> ValueMapping;
2640
2641 BasicBlock::iterator BI = BB->begin();
2642 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
2643 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
2644 // Clone the non-phi instructions of BB into PredBB, keeping track of the
2645 // mapping and using it to remap operands in the cloned instructions.
2646 for (; BI != BB->end(); ++BI) {
2647 Instruction *New = BI->clone();
2648
2649 // Remap operands to patch up intra-block references.
2650 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2651 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
2652 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
2653 if (I != ValueMapping.end())
2654 New->setOperand(i, I->second);
2655 }
2656
2657 // If this instruction can be simplified after the operands are updated,
2658 // just use the simplified value instead. This frequently happens due to
2659 // phi translation.
2660 if (Value *IV = SimplifyInstruction(
2661 New,
2662 {BB->getModule()->getDataLayout(), TLI, nullptr, nullptr, New})) {
2663 ValueMapping[&*BI] = IV;
2664 if (!New->mayHaveSideEffects()) {
2665 New->deleteValue();
2666 New = nullptr;
2667 }
2668 } else {
2669 ValueMapping[&*BI] = New;
2670 }
2671 if (New) {
2672 // Otherwise, insert the new instruction into the block.
2673 New->setName(BI->getName());
2674 PredBB->getInstList().insert(OldPredBranch->getIterator(), New);
2675 // Update Dominance from simplified New instruction operands.
2676 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2677 if (BasicBlock *SuccBB = dyn_cast<BasicBlock>(New->getOperand(i)))
2678 Updates.push_back({DominatorTree::Insert, PredBB, SuccBB});
2679 }
2680 }
2681
2682 // Check to see if the targets of the branch had PHI nodes. If so, we need to
2683 // add entries to the PHI nodes for branch from PredBB now.
2684 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
2685 addPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
2686 ValueMapping);
2687 addPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
2688 ValueMapping);
2689
2690 updateSSA(BB, PredBB, ValueMapping);
2691
2692 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
2693 // that we nuked.
2694 BB->removePredecessor(PredBB, true);
2695
2696 // Remove the unconditional branch at the end of the PredBB block.
2697 OldPredBranch->eraseFromParent();
2698 if (HasProfileData)
2699 BPI->copyEdgeProbabilities(BB, PredBB);
2700 DTU->applyUpdatesPermissive(Updates);
2701
2702 ++NumDupes;
2703 return true;
2704}
2705
2706// Pred is a predecessor of BB with an unconditional branch to BB. SI is
2707// a Select instruction in Pred. BB has other predecessors and SI is used in
2708// a PHI node in BB. SI has no other use.
2709// A new basic block, NewBB, is created and SI is converted to compare and
2710// conditional branch. SI is erased from parent.
2711void JumpThreadingPass::unfoldSelectInstr(BasicBlock *Pred, BasicBlock *BB,
2712 SelectInst *SI, PHINode *SIUse,
2713 unsigned Idx) {
2714 // Expand the select.
2715 //
2716 // Pred --
2717 // | v
2718 // | NewBB
2719 // | |
2720 // |-----
2721 // v
2722 // BB
2723 BranchInst *PredTerm = cast<BranchInst>(Pred->getTerminator());
2724 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
2725 BB->getParent(), BB);
2726 // Move the unconditional branch to NewBB.
2727 PredTerm->removeFromParent();
2728 NewBB->getInstList().insert(NewBB->end(), PredTerm);
2729 // Create a conditional branch and update PHI nodes.
2730 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
2731 SIUse->setIncomingValue(Idx, SI->getFalseValue());
2732 SIUse->addIncoming(SI->getTrueValue(), NewBB);
2733
2734 // The select is now dead.
2735 SI->eraseFromParent();
2736 DTU->applyUpdatesPermissive({{DominatorTree::Insert, NewBB, BB},
2737 {DominatorTree::Insert, Pred, NewBB}});
2738
2739 // Update any other PHI nodes in BB.
2740 for (BasicBlock::iterator BI = BB->begin();
2741 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
2742 if (Phi != SIUse)
2743 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);
2744}
2745
2746bool JumpThreadingPass::tryToUnfoldSelect(SwitchInst *SI, BasicBlock *BB) {
2747 PHINode *CondPHI = dyn_cast<PHINode>(SI->getCondition());
2748
2749 if (!CondPHI || CondPHI->getParent() != BB)
2750 return false;
2751
2752 for (unsigned I = 0, E = CondPHI->getNumIncomingValues(); I != E; ++I) {
2753 BasicBlock *Pred = CondPHI->getIncomingBlock(I);
2754 SelectInst *PredSI = dyn_cast<SelectInst>(CondPHI->getIncomingValue(I));
2755
2756 // The second and third condition can be potentially relaxed. Currently
2757 // the conditions help to simplify the code and allow us to reuse existing
2758 // code, developed for tryToUnfoldSelect(CmpInst *, BasicBlock *)
2759 if (!PredSI || PredSI->getParent() != Pred || !PredSI->hasOneUse())
2760 continue;
2761
2762 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
2763 if (!PredTerm || !PredTerm->isUnconditional())
2764 continue;
2765
2766 unfoldSelectInstr(Pred, BB, PredSI, CondPHI, I);
2767 return true;
2768 }
2769 return false;
2770}
2771
2772/// tryToUnfoldSelect - Look for blocks of the form
2773/// bb1:
2774/// %a = select
2775/// br bb2
2776///
2777/// bb2:
2778/// %p = phi [%a, %bb1] ...
2779/// %c = icmp %p
2780/// br i1 %c
2781///
2782/// And expand the select into a branch structure if one of its arms allows %c
2783/// to be folded. This later enables threading from bb1 over bb2.
2784bool JumpThreadingPass::tryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
2785 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
2786 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
2787 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
2788
2789 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
2790 CondLHS->getParent() != BB)
2791 return false;
2792
2793 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
2794 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
2795 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
2796
2797 // Look if one of the incoming values is a select in the corresponding
2798 // predecessor.
2799 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
2800 continue;
2801
2802 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
2803 if (!PredTerm || !PredTerm->isUnconditional())
2804 continue;
2805
2806 // Now check if one of the select values would allow us to constant fold the
2807 // terminator in BB. We don't do the transform if both sides fold, those
2808 // cases will be threaded in any case.
2809 LazyValueInfo::Tristate LHSFolds =
2810 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
2811 CondRHS, Pred, BB, CondCmp);
2812 LazyValueInfo::Tristate RHSFolds =
2813 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
2814 CondRHS, Pred, BB, CondCmp);
2815 if ((LHSFolds != LazyValueInfo::Unknown ||
2816 RHSFolds != LazyValueInfo::Unknown) &&
2817 LHSFolds != RHSFolds) {
2818 unfoldSelectInstr(Pred, BB, SI, CondLHS, I);
2819 return true;
2820 }
2821 }
2822 return false;
2823}
2824
2825/// tryToUnfoldSelectInCurrBB - Look for PHI/Select or PHI/CMP/Select in the
2826/// same BB in the form
2827/// bb:
2828/// %p = phi [false, %bb1], [true, %bb2], [false, %bb3], [true, %bb4], ...
2829/// %s = select %p, trueval, falseval
2830///
2831/// or
2832///
2833/// bb:
2834/// %p = phi [0, %bb1], [1, %bb2], [0, %bb3], [1, %bb4], ...
2835/// %c = cmp %p, 0
2836/// %s = select %c, trueval, falseval
2837///
2838/// And expand the select into a branch structure. This later enables
2839/// jump-threading over bb in this pass.
2840///
2841/// Using the similar approach of SimplifyCFG::FoldCondBranchOnPHI(), unfold
2842/// select if the associated PHI has at least one constant. If the unfolded
2843/// select is not jump-threaded, it will be folded again in the later
2844/// optimizations.
2845bool JumpThreadingPass::tryToUnfoldSelectInCurrBB(BasicBlock *BB) {
2846 // This transform would reduce the quality of msan diagnostics.
2847 // Disable this transform under MemorySanitizer.
2848 if (BB->getParent()->hasFnAttribute(Attribute::SanitizeMemory))
2849 return false;
2850
2851 // If threading this would thread across a loop header, don't thread the edge.
2852 // See the comments above findLoopHeaders for justifications and caveats.
2853 if (LoopHeaders.count(BB))
2854 return false;
2855
2856 for (BasicBlock::iterator BI = BB->begin();
2857 PHINode *PN = dyn_cast<PHINode>(BI); ++BI) {
2858 // Look for a Phi having at least one constant incoming value.
2859 if (llvm::all_of(PN->incoming_values(),
2860 [](Value *V) { return !isa<ConstantInt>(V); }))
2861 continue;
2862
2863 auto isUnfoldCandidate = [BB](SelectInst *SI, Value *V) {
2864 // Check if SI is in BB and use V as condition.
2865 if (SI->getParent() != BB)
2866 return false;
2867 Value *Cond = SI->getCondition();
2868 return (Cond && Cond == V && Cond->getType()->isIntegerTy(1));
2869 };
2870
2871 SelectInst *SI = nullptr;
2872 for (Use &U : PN->uses()) {
2873 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(U.getUser())) {
2874 // Look for a ICmp in BB that compares PN with a constant and is the
2875 // condition of a Select.
2876 if (Cmp->getParent() == BB && Cmp->hasOneUse() &&
2877 isa<ConstantInt>(Cmp->getOperand(1 - U.getOperandNo())))
2878 if (SelectInst *SelectI = dyn_cast<SelectInst>(Cmp->user_back()))
2879 if (isUnfoldCandidate(SelectI, Cmp->use_begin()->get())) {
2880 SI = SelectI;
2881 break;
2882 }
2883 } else if (SelectInst *SelectI = dyn_cast<SelectInst>(U.getUser())) {
2884 // Look for a Select in BB that uses PN as condition.
2885 if (isUnfoldCandidate(SelectI, U.get())) {
2886 SI = SelectI;
2887 break;
2888 }
2889 }
2890 }
2891
2892 if (!SI)
2893 continue;
2894 // Expand the select.
2895 Value *Cond = SI->getCondition();
2896 if (InsertFreezeWhenUnfoldingSelect &&
2897 !isGuaranteedNotToBeUndefOrPoison(Cond, nullptr, SI,
2898 &DTU->getDomTree()))
2899 Cond = new FreezeInst(Cond, "cond.fr", SI);
2900 Instruction *Term = SplitBlockAndInsertIfThen(Cond, SI, false);
2901 BasicBlock *SplitBB = SI->getParent();
2902 BasicBlock *NewBB = Term->getParent();
2903 PHINode *NewPN = PHINode::Create(SI->getType(), 2, "", SI);
2904 NewPN->addIncoming(SI->getTrueValue(), Term->getParent());
2905 NewPN->addIncoming(SI->getFalseValue(), BB);
2906 SI->replaceAllUsesWith(NewPN);
2907 SI->eraseFromParent();
2908 // NewBB and SplitBB are newly created blocks which require insertion.
2909 std::vector<DominatorTree::UpdateType> Updates;
2910 Updates.reserve((2 * SplitBB->getTerminator()->getNumSuccessors()) + 3);
2911 Updates.push_back({DominatorTree::Insert, BB, SplitBB});
2912 Updates.push_back({DominatorTree::Insert, BB, NewBB});
2913 Updates.push_back({DominatorTree::Insert, NewBB, SplitBB});
2914 // BB's successors were moved to SplitBB, update DTU accordingly.
2915 for (auto *Succ : successors(SplitBB)) {
2916 Updates.push_back({DominatorTree::Delete, BB, Succ});
2917 Updates.push_back({DominatorTree::Insert, SplitBB, Succ});
2918 }
2919 DTU->applyUpdatesPermissive(Updates);
2920 return true;
2921 }
2922 return false;
2923}
2924
2925/// Try to propagate a guard from the current BB into one of its predecessors
2926/// in case if another branch of execution implies that the condition of this
2927/// guard is always true. Currently we only process the simplest case that
2928/// looks like:
2929///
2930/// Start:
2931/// %cond = ...
2932/// br i1 %cond, label %T1, label %F1
2933/// T1:
2934/// br label %Merge
2935/// F1:
2936/// br label %Merge
2937/// Merge:
2938/// %condGuard = ...
2939/// call void(i1, ...) @llvm.experimental.guard( i1 %condGuard )[ "deopt"() ]
2940///
2941/// And cond either implies condGuard or !condGuard. In this case all the
2942/// instructions before the guard can be duplicated in both branches, and the
2943/// guard is then threaded to one of them.
2944bool JumpThreadingPass::processGuards(BasicBlock *BB) {
2945 using namespace PatternMatch;
2946
2947 // We only want to deal with two predecessors.
2948 BasicBlock *Pred1, *Pred2;
2949 auto PI = pred_begin(BB), PE = pred_end(BB);
2950 if (PI == PE)
2951 return false;
2952 Pred1 = *PI++;
2953 if (PI == PE)
2954 return false;
2955 Pred2 = *PI++;
2956 if (PI != PE)
2957 return false;
2958 if (Pred1 == Pred2)
2959 return false;
2960
2961 // Try to thread one of the guards of the block.
2962 // TODO: Look up deeper than to immediate predecessor?
2963 auto *Parent = Pred1->getSinglePredecessor();
2964 if (!Parent || Parent != Pred2->getSinglePredecessor())
2965 return false;
2966
2967 if (auto *BI = dyn_cast<BranchInst>(Parent->getTerminator()))
2968 for (auto &I : *BB)
2969 if (isGuard(&I) && threadGuard(BB, cast<IntrinsicInst>(&I), BI))
2970 return true;
2971
2972 return false;
2973}
2974
2975/// Try to propagate the guard from BB which is the lower block of a diamond
2976/// to one of its branches, in case if diamond's condition implies guard's
2977/// condition.
2978bool JumpThreadingPass::threadGuard(BasicBlock *BB, IntrinsicInst *Guard,
2979 BranchInst *BI) {
2980 assert(BI->getNumSuccessors() == 2 && "Wrong number of successors?")((BI->getNumSuccessors() == 2 && "Wrong number of successors?"
) ? static_cast<void> (0) : __assert_fail ("BI->getNumSuccessors() == 2 && \"Wrong number of successors?\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 2980, __PRETTY_FUNCTION__))
;
2981 assert(BI->isConditional() && "Unconditional branch has 2 successors?")((BI->isConditional() && "Unconditional branch has 2 successors?"
) ? static_cast<void> (0) : __assert_fail ("BI->isConditional() && \"Unconditional branch has 2 successors?\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 2981, __PRETTY_FUNCTION__))
;
2982 Value *GuardCond = Guard->getArgOperand(0);
2983 Value *BranchCond = BI->getCondition();
2984 BasicBlock *TrueDest = BI->getSuccessor(0);
2985 BasicBlock *FalseDest = BI->getSuccessor(1);
2986
2987 auto &DL = BB->getModule()->getDataLayout();
2988 bool TrueDestIsSafe = false;
2989 bool FalseDestIsSafe = false;
2990
2991 // True dest is safe if BranchCond => GuardCond.
2992 auto Impl = isImpliedCondition(BranchCond, GuardCond, DL);
2993 if (Impl && *Impl)
2994 TrueDestIsSafe = true;
2995 else {
2996 // False dest is safe if !BranchCond => GuardCond.
2997 Impl = isImpliedCondition(BranchCond, GuardCond, DL, /* LHSIsTrue */ false);
2998 if (Impl && *Impl)
2999 FalseDestIsSafe = true;
3000 }
3001
3002 if (!TrueDestIsSafe && !FalseDestIsSafe)
3003 return false;
3004
3005 BasicBlock *PredUnguardedBlock = TrueDestIsSafe ? TrueDest : FalseDest;
3006 BasicBlock *PredGuardedBlock = FalseDestIsSafe ? TrueDest : FalseDest;
3007
3008 ValueToValueMapTy UnguardedMapping, GuardedMapping;
3009 Instruction *AfterGuard = Guard->getNextNode();
3010 unsigned Cost = getJumpThreadDuplicationCost(BB, AfterGuard, BBDupThreshold);
3011 if (Cost > BBDupThreshold)
3012 return false;
3013 // Duplicate all instructions before the guard and the guard itself to the
3014 // branch where implication is not proved.
3015 BasicBlock *GuardedBlock = DuplicateInstructionsInSplitBetween(
3016 BB, PredGuardedBlock, AfterGuard, GuardedMapping, *DTU);
3017 assert(GuardedBlock && "Could not create the guarded block?")((GuardedBlock && "Could not create the guarded block?"
) ? static_cast<void> (0) : __assert_fail ("GuardedBlock && \"Could not create the guarded block?\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 3017, __PRETTY_FUNCTION__))
;
3018 // Duplicate all instructions before the guard in the unguarded branch.
3019 // Since we have successfully duplicated the guarded block and this block
3020 // has fewer instructions, we expect it to succeed.
3021 BasicBlock *UnguardedBlock = DuplicateInstructionsInSplitBetween(
3022 BB, PredUnguardedBlock, Guard, UnguardedMapping, *DTU);
3023 assert(UnguardedBlock && "Could not create the unguarded block?")((UnguardedBlock && "Could not create the unguarded block?"
) ? static_cast<void> (0) : __assert_fail ("UnguardedBlock && \"Could not create the unguarded block?\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 3023, __PRETTY_FUNCTION__))
;
3024 LLVM_DEBUG(dbgs() << "Moved guard " << *Guard << " to block "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "Moved guard " <<
*Guard << " to block " << GuardedBlock->getName
() << "\n"; } } while (false)
3025 << GuardedBlock->getName() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "Moved guard " <<
*Guard << " to block " << GuardedBlock->getName
() << "\n"; } } while (false)
;
3026 // Some instructions before the guard may still have uses. For them, we need
3027 // to create Phi nodes merging their copies in both guarded and unguarded
3028 // branches. Those instructions that have no uses can be just removed.
3029 SmallVector<Instruction *, 4> ToRemove;
3030 for (auto BI = BB->begin(); &*BI != AfterGuard; ++BI)
3031 if (!isa<PHINode>(&*BI))
3032 ToRemove.push_back(&*BI);
3033
3034 Instruction *InsertionPoint = &*BB->getFirstInsertionPt();
3035 assert(InsertionPoint && "Empty block?")((InsertionPoint && "Empty block?") ? static_cast<
void> (0) : __assert_fail ("InsertionPoint && \"Empty block?\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 3035, __PRETTY_FUNCTION__))
;
3036 // Substitute with Phis & remove.
3037 for (auto *Inst : reverse(ToRemove)) {
3038 if (!Inst->use_empty()) {
3039 PHINode *NewPN = PHINode::Create(Inst->getType(), 2);
3040 NewPN->addIncoming(UnguardedMapping[Inst], UnguardedBlock);
3041 NewPN->addIncoming(GuardedMapping[Inst], GuardedBlock);
3042 NewPN->insertBefore(InsertionPoint);
3043 Inst->replaceAllUsesWith(NewPN);
3044 }
3045 Inst->eraseFromParent();
3046 }
3047 return true;
3048}

/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h

1//===- llvm/Instructions.h - Instruction subclass definitions ---*- C++ -*-===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file exposes the class definitions of all of the subclasses of the
10// Instruction class. This is meant to be an easy way to get access to all
11// instruction subclasses.
12//
13//===----------------------------------------------------------------------===//
14
15#ifndef LLVM_IR_INSTRUCTIONS_H
16#define LLVM_IR_INSTRUCTIONS_H
17
18#include "llvm/ADT/ArrayRef.h"
19#include "llvm/ADT/Bitfields.h"
20#include "llvm/ADT/None.h"
21#include "llvm/ADT/STLExtras.h"
22#include "llvm/ADT/SmallVector.h"
23#include "llvm/ADT/StringRef.h"
24#include "llvm/ADT/Twine.h"
25#include "llvm/ADT/iterator.h"
26#include "llvm/ADT/iterator_range.h"
27#include "llvm/IR/Attributes.h"
28#include "llvm/IR/BasicBlock.h"
29#include "llvm/IR/CallingConv.h"
30#include "llvm/IR/CFG.h"
31#include "llvm/IR/Constant.h"
32#include "llvm/IR/DerivedTypes.h"
33#include "llvm/IR/Function.h"
34#include "llvm/IR/InstrTypes.h"
35#include "llvm/IR/Instruction.h"
36#include "llvm/IR/OperandTraits.h"
37#include "llvm/IR/Type.h"
38#include "llvm/IR/Use.h"
39#include "llvm/IR/User.h"
40#include "llvm/IR/Value.h"
41#include "llvm/Support/AtomicOrdering.h"
42#include "llvm/Support/Casting.h"
43#include "llvm/Support/ErrorHandling.h"
44#include <cassert>
45#include <cstddef>
46#include <cstdint>
47#include <iterator>
48
49namespace llvm {
50
51class APInt;
52class ConstantInt;
53class DataLayout;
54class LLVMContext;
55
56//===----------------------------------------------------------------------===//
57// AllocaInst Class
58//===----------------------------------------------------------------------===//
59
60/// an instruction to allocate memory on the stack
61class AllocaInst : public UnaryInstruction {
62 Type *AllocatedType;
63
64 using AlignmentField = AlignmentBitfieldElementT<0>;
65 using UsedWithInAllocaField = BoolBitfieldElementT<AlignmentField::NextBit>;
66 using SwiftErrorField = BoolBitfieldElementT<UsedWithInAllocaField::NextBit>;
67 static_assert(Bitfield::areContiguous<AlignmentField, UsedWithInAllocaField,
68 SwiftErrorField>(),
69 "Bitfields must be contiguous");
70
71protected:
72 // Note: Instruction needs to be a friend here to call cloneImpl.
73 friend class Instruction;
74
75 AllocaInst *cloneImpl() const;
76
77public:
78 explicit AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize,
79 const Twine &Name, Instruction *InsertBefore);
80 AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize,
81 const Twine &Name, BasicBlock *InsertAtEnd);
82
83 AllocaInst(Type *Ty, unsigned AddrSpace, const Twine &Name,
84 Instruction *InsertBefore);
85 AllocaInst(Type *Ty, unsigned AddrSpace,
86 const Twine &Name, BasicBlock *InsertAtEnd);
87
88 AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize, Align Align,
89 const Twine &Name = "", Instruction *InsertBefore = nullptr);
90 AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize, Align Align,
91 const Twine &Name, BasicBlock *InsertAtEnd);
92
93 /// Return true if there is an allocation size parameter to the allocation
94 /// instruction that is not 1.
95 bool isArrayAllocation() const;
96
97 /// Get the number of elements allocated. For a simple allocation of a single
98 /// element, this will return a constant 1 value.
99 const Value *getArraySize() const { return getOperand(0); }
100 Value *getArraySize() { return getOperand(0); }
101
102 /// Overload to return most specific pointer type.
103 PointerType *getType() const {
104 return cast<PointerType>(Instruction::getType());
105 }
106
107 /// Get allocation size in bits. Returns None if size can't be determined,
108 /// e.g. in case of a VLA.
109 Optional<TypeSize> getAllocationSizeInBits(const DataLayout &DL) const;
110
111 /// Return the type that is being allocated by the instruction.
112 Type *getAllocatedType() const { return AllocatedType; }
113 /// for use only in special circumstances that need to generically
114 /// transform a whole instruction (eg: IR linking and vectorization).
115 void setAllocatedType(Type *Ty) { AllocatedType = Ty; }
116
117 /// Return the alignment of the memory that is being allocated by the
118 /// instruction.
119 Align getAlign() const {
120 return Align(1ULL << getSubclassData<AlignmentField>());
121 }
122
123 void setAlignment(Align Align) {
124 setSubclassData<AlignmentField>(Log2(Align));
125 }
126
127 // FIXME: Remove this one transition to Align is over.
128 unsigned getAlignment() const { return getAlign().value(); }
129
130 /// Return true if this alloca is in the entry block of the function and is a
131 /// constant size. If so, the code generator will fold it into the
132 /// prolog/epilog code, so it is basically free.
133 bool isStaticAlloca() const;
134
135 /// Return true if this alloca is used as an inalloca argument to a call. Such
136 /// allocas are never considered static even if they are in the entry block.
137 bool isUsedWithInAlloca() const {
138 return getSubclassData<UsedWithInAllocaField>();
139 }
140
141 /// Specify whether this alloca is used to represent the arguments to a call.
142 void setUsedWithInAlloca(bool V) {
143 setSubclassData<UsedWithInAllocaField>(V);
144 }
145
146 /// Return true if this alloca is used as a swifterror argument to a call.
147 bool isSwiftError() const { return getSubclassData<SwiftErrorField>(); }
148 /// Specify whether this alloca is used to represent a swifterror.
149 void setSwiftError(bool V) { setSubclassData<SwiftErrorField>(V); }
150
151 // Methods for support type inquiry through isa, cast, and dyn_cast:
152 static bool classof(const Instruction *I) {
153 return (I->getOpcode() == Instruction::Alloca);
154 }
155 static bool classof(const Value *V) {
156 return isa<Instruction>(V) && classof(cast<Instruction>(V));
157 }
158
159private:
160 // Shadow Instruction::setInstructionSubclassData with a private forwarding
161 // method so that subclasses cannot accidentally use it.
162 template <typename Bitfield>
163 void setSubclassData(typename Bitfield::Type Value) {
164 Instruction::setSubclassData<Bitfield>(Value);
165 }
166};
167
168//===----------------------------------------------------------------------===//
169// LoadInst Class
170//===----------------------------------------------------------------------===//
171
172/// An instruction for reading from memory. This uses the SubclassData field in
173/// Value to store whether or not the load is volatile.
174class LoadInst : public UnaryInstruction {
175 using VolatileField = BoolBitfieldElementT<0>;
176 using AlignmentField = AlignmentBitfieldElementT<VolatileField::NextBit>;
177 using OrderingField = AtomicOrderingBitfieldElementT<AlignmentField::NextBit>;
178 static_assert(
179 Bitfield::areContiguous<VolatileField, AlignmentField, OrderingField>(),
180 "Bitfields must be contiguous");
181
182 void AssertOK();
183
184protected:
185 // Note: Instruction needs to be a friend here to call cloneImpl.
186 friend class Instruction;
187
188 LoadInst *cloneImpl() const;
189
190public:
191 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr,
192 Instruction *InsertBefore);
193 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, BasicBlock *InsertAtEnd);
194 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile,
195 Instruction *InsertBefore);
196 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile,
197 BasicBlock *InsertAtEnd);
198 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile,
199 Align Align, Instruction *InsertBefore = nullptr);
200 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile,
201 Align Align, BasicBlock *InsertAtEnd);
202 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile,
203 Align Align, AtomicOrdering Order,
204 SyncScope::ID SSID = SyncScope::System,
205 Instruction *InsertBefore = nullptr);
206 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile,
207 Align Align, AtomicOrdering Order, SyncScope::ID SSID,
208 BasicBlock *InsertAtEnd);
209
210 /// Return true if this is a load from a volatile memory location.
211 bool isVolatile() const { return getSubclassData<VolatileField>(); }
212
213 /// Specify whether this is a volatile load or not.
214 void setVolatile(bool V) { setSubclassData<VolatileField>(V); }
215
216 /// Return the alignment of the access that is being performed.
217 /// FIXME: Remove this function once transition to Align is over.
218 /// Use getAlign() instead.
219 unsigned getAlignment() const { return getAlign().value(); }
220
221 /// Return the alignment of the access that is being performed.
222 Align getAlign() const {
223 return Align(1ULL << (getSubclassData<AlignmentField>()));
224 }
225
226 void setAlignment(Align Align) {
227 setSubclassData<AlignmentField>(Log2(Align));
228 }
229
230 /// Returns the ordering constraint of this load instruction.
231 AtomicOrdering getOrdering() const {
232 return getSubclassData<OrderingField>();
233 }
234 /// Sets the ordering constraint of this load instruction. May not be Release
235 /// or AcquireRelease.
236 void setOrdering(AtomicOrdering Ordering) {
237 setSubclassData<OrderingField>(Ordering);
238 }
239
240 /// Returns the synchronization scope ID of this load instruction.
241 SyncScope::ID getSyncScopeID() const {
242 return SSID;
243 }
244
245 /// Sets the synchronization scope ID of this load instruction.
246 void setSyncScopeID(SyncScope::ID SSID) {
247 this->SSID = SSID;
248 }
249
250 /// Sets the ordering constraint and the synchronization scope ID of this load
251 /// instruction.
252 void setAtomic(AtomicOrdering Ordering,
253 SyncScope::ID SSID = SyncScope::System) {
254 setOrdering(Ordering);
255 setSyncScopeID(SSID);
256 }
257
258 bool isSimple() const { return !isAtomic() && !isVolatile(); }
259
260 bool isUnordered() const {
261 return (getOrdering() == AtomicOrdering::NotAtomic ||
2
Assuming the condition is true
4
Returning the value 1, which participates in a condition later
262 getOrdering() == AtomicOrdering::Unordered) &&
263 !isVolatile();
3
Assuming the condition is true
264 }
265
266 Value *getPointerOperand() { return getOperand(0); }
267 const Value *getPointerOperand() const { return getOperand(0); }
268 static unsigned getPointerOperandIndex() { return 0U; }
269 Type *getPointerOperandType() const { return getPointerOperand()->getType(); }
270
271 /// Returns the address space of the pointer operand.
272 unsigned getPointerAddressSpace() const {
273 return getPointerOperandType()->getPointerAddressSpace();
274 }
275
276 // Methods for support type inquiry through isa, cast, and dyn_cast:
277 static bool classof(const Instruction *I) {
278 return I->getOpcode() == Instruction::Load;
279 }
280 static bool classof(const Value *V) {
281 return isa<Instruction>(V) && classof(cast<Instruction>(V));
282 }
283
284private:
285 // Shadow Instruction::setInstructionSubclassData with a private forwarding
286 // method so that subclasses cannot accidentally use it.
287 template <typename Bitfield>
288 void setSubclassData(typename Bitfield::Type Value) {
289 Instruction::setSubclassData<Bitfield>(Value);
290 }
291
292 /// The synchronization scope ID of this load instruction. Not quite enough
293 /// room in SubClassData for everything, so synchronization scope ID gets its
294 /// own field.
295 SyncScope::ID SSID;
296};
297
298//===----------------------------------------------------------------------===//
299// StoreInst Class
300//===----------------------------------------------------------------------===//
301
302/// An instruction for storing to memory.
303class StoreInst : public Instruction {
304 using VolatileField = BoolBitfieldElementT<0>;
305 using AlignmentField = AlignmentBitfieldElementT<VolatileField::NextBit>;
306 using OrderingField = AtomicOrderingBitfieldElementT<AlignmentField::NextBit>;
307 static_assert(
308 Bitfield::areContiguous<VolatileField, AlignmentField, OrderingField>(),
309 "Bitfields must be contiguous");
310
311 void AssertOK();
312
313protected:
314 // Note: Instruction needs to be a friend here to call cloneImpl.
315 friend class Instruction;
316
317 StoreInst *cloneImpl() const;
318
319public:
320 StoreInst(Value *Val, Value *Ptr, Instruction *InsertBefore);
321 StoreInst(Value *Val, Value *Ptr, BasicBlock *InsertAtEnd);
322 StoreInst(Value *Val, Value *Ptr, bool isVolatile, Instruction *InsertBefore);
323 StoreInst(Value *Val, Value *Ptr, bool isVolatile, BasicBlock *InsertAtEnd);
324 StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align,
325 Instruction *InsertBefore = nullptr);
326 StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align,
327 BasicBlock *InsertAtEnd);
328 StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align,
329 AtomicOrdering Order, SyncScope::ID SSID = SyncScope::System,
330 Instruction *InsertBefore = nullptr);
331 StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align,
332 AtomicOrdering Order, SyncScope::ID SSID, BasicBlock *InsertAtEnd);
333
334 // allocate space for exactly two operands
335 void *operator new(size_t s) {
336 return User::operator new(s, 2);
337 }
338
339 /// Return true if this is a store to a volatile memory location.
340 bool isVolatile() const { return getSubclassData<VolatileField>(); }
341
342 /// Specify whether this is a volatile store or not.
343 void setVolatile(bool V) { setSubclassData<VolatileField>(V); }
344
345 /// Transparently provide more efficient getOperand methods.
346 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
347
348 /// Return the alignment of the access that is being performed
349 /// FIXME: Remove this function once transition to Align is over.
350 /// Use getAlign() instead.
351 unsigned getAlignment() const { return getAlign().value(); }
352
353 Align getAlign() const {
354 return Align(1ULL << (getSubclassData<AlignmentField>()));
355 }
356
357 void setAlignment(Align Align) {
358 setSubclassData<AlignmentField>(Log2(Align));
359 }
360
361 /// Returns the ordering constraint of this store instruction.
362 AtomicOrdering getOrdering() const {
363 return getSubclassData<OrderingField>();
364 }
365
366 /// Sets the ordering constraint of this store instruction. May not be
367 /// Acquire or AcquireRelease.
368 void setOrdering(AtomicOrdering Ordering) {
369 setSubclassData<OrderingField>(Ordering);
370 }
371
372 /// Returns the synchronization scope ID of this store instruction.
373 SyncScope::ID getSyncScopeID() const {
374 return SSID;
375 }
376
377 /// Sets the synchronization scope ID of this store instruction.
378 void setSyncScopeID(SyncScope::ID SSID) {
379 this->SSID = SSID;
380 }
381
382 /// Sets the ordering constraint and the synchronization scope ID of this
383 /// store instruction.
384 void setAtomic(AtomicOrdering Ordering,
385 SyncScope::ID SSID = SyncScope::System) {
386 setOrdering(Ordering);
387 setSyncScopeID(SSID);
388 }
389
390 bool isSimple() const { return !isAtomic() && !isVolatile(); }
391
392 bool isUnordered() const {
393 return (getOrdering() == AtomicOrdering::NotAtomic ||
394 getOrdering() == AtomicOrdering::Unordered) &&
395 !isVolatile();
396 }
397
398 Value *getValueOperand() { return getOperand(0); }
399 const Value *getValueOperand() const { return getOperand(0); }
400
401 Value *getPointerOperand() { return getOperand(1); }
402 const Value *getPointerOperand() const { return getOperand(1); }
403 static unsigned getPointerOperandIndex() { return 1U; }
404 Type *getPointerOperandType() const { return getPointerOperand()->getType(); }
405
406 /// Returns the address space of the pointer operand.
407 unsigned getPointerAddressSpace() const {
408 return getPointerOperandType()->getPointerAddressSpace();
409 }
410
411 // Methods for support type inquiry through isa, cast, and dyn_cast:
412 static bool classof(const Instruction *I) {
413 return I->getOpcode() == Instruction::Store;
414 }
415 static bool classof(const Value *V) {
416 return isa<Instruction>(V) && classof(cast<Instruction>(V));
417 }
418
419private:
420 // Shadow Instruction::setInstructionSubclassData with a private forwarding
421 // method so that subclasses cannot accidentally use it.
422 template <typename Bitfield>
423 void setSubclassData(typename Bitfield::Type Value) {
424 Instruction::setSubclassData<Bitfield>(Value);
425 }
426
427 /// The synchronization scope ID of this store instruction. Not quite enough
428 /// room in SubClassData for everything, so synchronization scope ID gets its
429 /// own field.
430 SyncScope::ID SSID;
431};
432
433template <>
434struct OperandTraits<StoreInst> : public FixedNumOperandTraits<StoreInst, 2> {
435};
436
437DEFINE_TRANSPARENT_OPERAND_ACCESSORS(StoreInst, Value)StoreInst::op_iterator StoreInst::op_begin() { return OperandTraits
<StoreInst>::op_begin(this); } StoreInst::const_op_iterator
StoreInst::op_begin() const { return OperandTraits<StoreInst
>::op_begin(const_cast<StoreInst*>(this)); } StoreInst
::op_iterator StoreInst::op_end() { return OperandTraits<StoreInst
>::op_end(this); } StoreInst::const_op_iterator StoreInst::
op_end() const { return OperandTraits<StoreInst>::op_end
(const_cast<StoreInst*>(this)); } Value *StoreInst::getOperand
(unsigned i_nocapture) const { ((i_nocapture < OperandTraits
<StoreInst>::operands(this) && "getOperand() out of range!"
) ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<StoreInst>::operands(this) && \"getOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 437, __PRETTY_FUNCTION__)); return cast_or_null<Value>
( OperandTraits<StoreInst>::op_begin(const_cast<StoreInst
*>(this))[i_nocapture].get()); } void StoreInst::setOperand
(unsigned i_nocapture, Value *Val_nocapture) { ((i_nocapture <
OperandTraits<StoreInst>::operands(this) && "setOperand() out of range!"
) ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<StoreInst>::operands(this) && \"setOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 437, __PRETTY_FUNCTION__)); OperandTraits<StoreInst>::
op_begin(this)[i_nocapture] = Val_nocapture; } unsigned StoreInst
::getNumOperands() const { return OperandTraits<StoreInst>
::operands(this); } template <int Idx_nocapture> Use &
StoreInst::Op() { return this->OpFrom<Idx_nocapture>
(this); } template <int Idx_nocapture> const Use &StoreInst
::Op() const { return this->OpFrom<Idx_nocapture>(this
); }
438
439//===----------------------------------------------------------------------===//
440// FenceInst Class
441//===----------------------------------------------------------------------===//
442
443/// An instruction for ordering other memory operations.
444class FenceInst : public Instruction {
445 using OrderingField = AtomicOrderingBitfieldElementT<0>;
446
447 void Init(AtomicOrdering Ordering, SyncScope::ID SSID);
448
449protected:
450 // Note: Instruction needs to be a friend here to call cloneImpl.
451 friend class Instruction;
452
453 FenceInst *cloneImpl() const;
454
455public:
456 // Ordering may only be Acquire, Release, AcquireRelease, or
457 // SequentiallyConsistent.
458 FenceInst(LLVMContext &C, AtomicOrdering Ordering,
459 SyncScope::ID SSID = SyncScope::System,
460 Instruction *InsertBefore = nullptr);
461 FenceInst(LLVMContext &C, AtomicOrdering Ordering, SyncScope::ID SSID,
462 BasicBlock *InsertAtEnd);
463
464 // allocate space for exactly zero operands
465 void *operator new(size_t s) {
466 return User::operator new(s, 0);
467 }
468
469 /// Returns the ordering constraint of this fence instruction.
470 AtomicOrdering getOrdering() const {
471 return getSubclassData<OrderingField>();
472 }
473
474 /// Sets the ordering constraint of this fence instruction. May only be
475 /// Acquire, Release, AcquireRelease, or SequentiallyConsistent.
476 void setOrdering(AtomicOrdering Ordering) {
477 setSubclassData<OrderingField>(Ordering);
478 }
479
480 /// Returns the synchronization scope ID of this fence instruction.
481 SyncScope::ID getSyncScopeID() const {
482 return SSID;
483 }
484
485 /// Sets the synchronization scope ID of this fence instruction.
486 void setSyncScopeID(SyncScope::ID SSID) {
487 this->SSID = SSID;
488 }
489
490 // Methods for support type inquiry through isa, cast, and dyn_cast:
491 static bool classof(const Instruction *I) {
492 return I->getOpcode() == Instruction::Fence;
493 }
494 static bool classof(const Value *V) {
495 return isa<Instruction>(V) && classof(cast<Instruction>(V));
496 }
497
498private:
499 // Shadow Instruction::setInstructionSubclassData with a private forwarding
500 // method so that subclasses cannot accidentally use it.
501 template <typename Bitfield>
502 void setSubclassData(typename Bitfield::Type Value) {
503 Instruction::setSubclassData<Bitfield>(Value);
504 }
505
506 /// The synchronization scope ID of this fence instruction. Not quite enough
507 /// room in SubClassData for everything, so synchronization scope ID gets its
508 /// own field.
509 SyncScope::ID SSID;
510};
511
512//===----------------------------------------------------------------------===//
513// AtomicCmpXchgInst Class
514//===----------------------------------------------------------------------===//
515
516/// An instruction that atomically checks whether a
517/// specified value is in a memory location, and, if it is, stores a new value
518/// there. The value returned by this instruction is a pair containing the
519/// original value as first element, and an i1 indicating success (true) or
520/// failure (false) as second element.
521///
522class AtomicCmpXchgInst : public Instruction {
523 void Init(Value *Ptr, Value *Cmp, Value *NewVal, Align Align,
524 AtomicOrdering SuccessOrdering, AtomicOrdering FailureOrdering,
525 SyncScope::ID SSID);
526
527 template <unsigned Offset>
528 using AtomicOrderingBitfieldElement =
529 typename Bitfield::Element<AtomicOrdering, Offset, 3,
530 AtomicOrdering::LAST>;
531
532protected:
533 // Note: Instruction needs to be a friend here to call cloneImpl.
534 friend class Instruction;
535
536 AtomicCmpXchgInst *cloneImpl() const;
537
538public:
539 AtomicCmpXchgInst(Value *Ptr, Value *Cmp, Value *NewVal, Align Alignment,
540 AtomicOrdering SuccessOrdering,
541 AtomicOrdering FailureOrdering, SyncScope::ID SSID,
542 Instruction *InsertBefore = nullptr);
543 AtomicCmpXchgInst(Value *Ptr, Value *Cmp, Value *NewVal, Align Alignment,
544 AtomicOrdering SuccessOrdering,
545 AtomicOrdering FailureOrdering, SyncScope::ID SSID,
546 BasicBlock *InsertAtEnd);
547
548 // allocate space for exactly three operands
549 void *operator new(size_t s) {
550 return User::operator new(s, 3);
551 }
552
553 using VolatileField = BoolBitfieldElementT<0>;
554 using WeakField = BoolBitfieldElementT<VolatileField::NextBit>;
555 using SuccessOrderingField =
556 AtomicOrderingBitfieldElementT<WeakField::NextBit>;
557 using FailureOrderingField =
558 AtomicOrderingBitfieldElementT<SuccessOrderingField::NextBit>;
559 using AlignmentField =
560 AlignmentBitfieldElementT<FailureOrderingField::NextBit>;
561 static_assert(
562 Bitfield::areContiguous<VolatileField, WeakField, SuccessOrderingField,
563 FailureOrderingField, AlignmentField>(),
564 "Bitfields must be contiguous");
565
566 /// Return the alignment of the memory that is being allocated by the
567 /// instruction.
568 Align getAlign() const {
569 return Align(1ULL << getSubclassData<AlignmentField>());
570 }
571
572 void setAlignment(Align Align) {
573 setSubclassData<AlignmentField>(Log2(Align));
574 }
575
576 /// Return true if this is a cmpxchg from a volatile memory
577 /// location.
578 ///
579 bool isVolatile() const { return getSubclassData<VolatileField>(); }
580
581 /// Specify whether this is a volatile cmpxchg.
582 ///
583 void setVolatile(bool V) { setSubclassData<VolatileField>(V); }
584
585 /// Return true if this cmpxchg may spuriously fail.
586 bool isWeak() const { return getSubclassData<WeakField>(); }
587
588 void setWeak(bool IsWeak) { setSubclassData<WeakField>(IsWeak); }
589
590 /// Transparently provide more efficient getOperand methods.
591 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
592
593 /// Returns the success ordering constraint of this cmpxchg instruction.
594 AtomicOrdering getSuccessOrdering() const {
595 return getSubclassData<SuccessOrderingField>();
596 }
597
598 /// Sets the success ordering constraint of this cmpxchg instruction.
599 void setSuccessOrdering(AtomicOrdering Ordering) {
600 assert(Ordering != AtomicOrdering::NotAtomic &&((Ordering != AtomicOrdering::NotAtomic && "CmpXchg instructions can only be atomic."
) ? static_cast<void> (0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"CmpXchg instructions can only be atomic.\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 601, __PRETTY_FUNCTION__))
601 "CmpXchg instructions can only be atomic.")((Ordering != AtomicOrdering::NotAtomic && "CmpXchg instructions can only be atomic."
) ? static_cast<void> (0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"CmpXchg instructions can only be atomic.\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 601, __PRETTY_FUNCTION__))
;
602 setSubclassData<SuccessOrderingField>(Ordering);
603 }
604
605 /// Returns the failure ordering constraint of this cmpxchg instruction.
606 AtomicOrdering getFailureOrdering() const {
607 return getSubclassData<FailureOrderingField>();
608 }
609
610 /// Sets the failure ordering constraint of this cmpxchg instruction.
611 void setFailureOrdering(AtomicOrdering Ordering) {
612 assert(Ordering != AtomicOrdering::NotAtomic &&((Ordering != AtomicOrdering::NotAtomic && "CmpXchg instructions can only be atomic."
) ? static_cast<void> (0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"CmpXchg instructions can only be atomic.\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 613, __PRETTY_FUNCTION__))
613 "CmpXchg instructions can only be atomic.")((Ordering != AtomicOrdering::NotAtomic && "CmpXchg instructions can only be atomic."
) ? static_cast<void> (0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"CmpXchg instructions can only be atomic.\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 613, __PRETTY_FUNCTION__))
;
614 setSubclassData<FailureOrderingField>(Ordering);
615 }
616
617 /// Returns the synchronization scope ID of this cmpxchg instruction.
618 SyncScope::ID getSyncScopeID() const {
619 return SSID;
620 }
621
622 /// Sets the synchronization scope ID of this cmpxchg instruction.
623 void setSyncScopeID(SyncScope::ID SSID) {
624 this->SSID = SSID;
625 }
626
627 Value *getPointerOperand() { return getOperand(0); }
628 const Value *getPointerOperand() const { return getOperand(0); }
629 static unsigned getPointerOperandIndex() { return 0U; }
630
631 Value *getCompareOperand() { return getOperand(1); }
632 const Value *getCompareOperand() const { return getOperand(1); }
633
634 Value *getNewValOperand() { return getOperand(2); }
635 const Value *getNewValOperand() const { return getOperand(2); }
636
637 /// Returns the address space of the pointer operand.
638 unsigned getPointerAddressSpace() const {
639 return getPointerOperand()->getType()->getPointerAddressSpace();
640 }
641
642 /// Returns the strongest permitted ordering on failure, given the
643 /// desired ordering on success.
644 ///
645 /// If the comparison in a cmpxchg operation fails, there is no atomic store
646 /// so release semantics cannot be provided. So this function drops explicit
647 /// Release requests from the AtomicOrdering. A SequentiallyConsistent
648 /// operation would remain SequentiallyConsistent.
649 static AtomicOrdering
650 getStrongestFailureOrdering(AtomicOrdering SuccessOrdering) {
651 switch (SuccessOrdering) {
652 default:
653 llvm_unreachable("invalid cmpxchg success ordering")::llvm::llvm_unreachable_internal("invalid cmpxchg success ordering"
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 653)
;
654 case AtomicOrdering::Release:
655 case AtomicOrdering::Monotonic:
656 return AtomicOrdering::Monotonic;
657 case AtomicOrdering::AcquireRelease:
658 case AtomicOrdering::Acquire:
659 return AtomicOrdering::Acquire;
660 case AtomicOrdering::SequentiallyConsistent:
661 return AtomicOrdering::SequentiallyConsistent;
662 }
663 }
664
665 // Methods for support type inquiry through isa, cast, and dyn_cast:
666 static bool classof(const Instruction *I) {
667 return I->getOpcode() == Instruction::AtomicCmpXchg;
668 }
669 static bool classof(const Value *V) {
670 return isa<Instruction>(V) && classof(cast<Instruction>(V));
671 }
672
673private:
674 // Shadow Instruction::setInstructionSubclassData with a private forwarding
675 // method so that subclasses cannot accidentally use it.
676 template <typename Bitfield>
677 void setSubclassData(typename Bitfield::Type Value) {
678 Instruction::setSubclassData<Bitfield>(Value);
679 }
680
681 /// The synchronization scope ID of this cmpxchg instruction. Not quite
682 /// enough room in SubClassData for everything, so synchronization scope ID
683 /// gets its own field.
684 SyncScope::ID SSID;
685};
686
687template <>
688struct OperandTraits<AtomicCmpXchgInst> :
689 public FixedNumOperandTraits<AtomicCmpXchgInst, 3> {
690};
691
692DEFINE_TRANSPARENT_OPERAND_ACCESSORS(AtomicCmpXchgInst, Value)AtomicCmpXchgInst::op_iterator AtomicCmpXchgInst::op_begin() {
return OperandTraits<AtomicCmpXchgInst>::op_begin(this
); } AtomicCmpXchgInst::const_op_iterator AtomicCmpXchgInst::
op_begin() const { return OperandTraits<AtomicCmpXchgInst>
::op_begin(const_cast<AtomicCmpXchgInst*>(this)); } AtomicCmpXchgInst
::op_iterator AtomicCmpXchgInst::op_end() { return OperandTraits
<AtomicCmpXchgInst>::op_end(this); } AtomicCmpXchgInst::
const_op_iterator AtomicCmpXchgInst::op_end() const { return OperandTraits
<AtomicCmpXchgInst>::op_end(const_cast<AtomicCmpXchgInst
*>(this)); } Value *AtomicCmpXchgInst::getOperand(unsigned
i_nocapture) const { ((i_nocapture < OperandTraits<AtomicCmpXchgInst
>::operands(this) && "getOperand() out of range!")
? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<AtomicCmpXchgInst>::operands(this) && \"getOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 692, __PRETTY_FUNCTION__)); return cast_or_null<Value>
( OperandTraits<AtomicCmpXchgInst>::op_begin(const_cast
<AtomicCmpXchgInst*>(this))[i_nocapture].get()); } void
AtomicCmpXchgInst::setOperand(unsigned i_nocapture, Value *Val_nocapture
) { ((i_nocapture < OperandTraits<AtomicCmpXchgInst>
::operands(this) && "setOperand() out of range!") ? static_cast
<void> (0) : __assert_fail ("i_nocapture < OperandTraits<AtomicCmpXchgInst>::operands(this) && \"setOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 692, __PRETTY_FUNCTION__)); OperandTraits<AtomicCmpXchgInst
>::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned
AtomicCmpXchgInst::getNumOperands() const { return OperandTraits
<AtomicCmpXchgInst>::operands(this); } template <int
Idx_nocapture> Use &AtomicCmpXchgInst::Op() { return this
->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture
> const Use &AtomicCmpXchgInst::Op() const { return this
->OpFrom<Idx_nocapture>(this); }
693
694//===----------------------------------------------------------------------===//
695// AtomicRMWInst Class
696//===----------------------------------------------------------------------===//
697
698/// an instruction that atomically reads a memory location,
699/// combines it with another value, and then stores the result back. Returns
700/// the old value.
701///
702class AtomicRMWInst : public Instruction {
703protected:
704 // Note: Instruction needs to be a friend here to call cloneImpl.
705 friend class Instruction;
706
707 AtomicRMWInst *cloneImpl() const;
708
709public:
710 /// This enumeration lists the possible modifications atomicrmw can make. In
711 /// the descriptions, 'p' is the pointer to the instruction's memory location,
712 /// 'old' is the initial value of *p, and 'v' is the other value passed to the
713 /// instruction. These instructions always return 'old'.
714 enum BinOp : unsigned {
715 /// *p = v
716 Xchg,
717 /// *p = old + v
718 Add,
719 /// *p = old - v
720 Sub,
721 /// *p = old & v
722 And,
723 /// *p = ~(old & v)
724 Nand,
725 /// *p = old | v
726 Or,
727 /// *p = old ^ v
728 Xor,
729 /// *p = old >signed v ? old : v
730 Max,
731 /// *p = old <signed v ? old : v
732 Min,
733 /// *p = old >unsigned v ? old : v
734 UMax,
735 /// *p = old <unsigned v ? old : v
736 UMin,
737
738 /// *p = old + v
739 FAdd,
740
741 /// *p = old - v
742 FSub,
743
744 FIRST_BINOP = Xchg,
745 LAST_BINOP = FSub,
746 BAD_BINOP
747 };
748
749private:
750 template <unsigned Offset>
751 using AtomicOrderingBitfieldElement =
752 typename Bitfield::Element<AtomicOrdering, Offset, 3,
753 AtomicOrdering::LAST>;
754
755 template <unsigned Offset>
756 using BinOpBitfieldElement =
757 typename Bitfield::Element<BinOp, Offset, 4, BinOp::LAST_BINOP>;
758
759public:
760 AtomicRMWInst(BinOp Operation, Value *Ptr, Value *Val, Align Alignment,
761 AtomicOrdering Ordering, SyncScope::ID SSID,
762 Instruction *InsertBefore = nullptr);
763 AtomicRMWInst(BinOp Operation, Value *Ptr, Value *Val, Align Alignment,
764 AtomicOrdering Ordering, SyncScope::ID SSID,
765 BasicBlock *InsertAtEnd);
766
767 // allocate space for exactly two operands
768 void *operator new(size_t s) {
769 return User::operator new(s, 2);
770 }
771
772 using VolatileField = BoolBitfieldElementT<0>;
773 using AtomicOrderingField =
774 AtomicOrderingBitfieldElementT<VolatileField::NextBit>;
775 using OperationField = BinOpBitfieldElement<AtomicOrderingField::NextBit>;
776 using AlignmentField = AlignmentBitfieldElementT<OperationField::NextBit>;
777 static_assert(Bitfield::areContiguous<VolatileField, AtomicOrderingField,
778 OperationField, AlignmentField>(),
779 "Bitfields must be contiguous");
780
781 BinOp getOperation() const { return getSubclassData<OperationField>(); }
782
783 static StringRef getOperationName(BinOp Op);
784
785 static bool isFPOperation(BinOp Op) {
786 switch (Op) {
787 case AtomicRMWInst::FAdd:
788 case AtomicRMWInst::FSub:
789 return true;
790 default:
791 return false;
792 }
793 }
794
795 void setOperation(BinOp Operation) {
796 setSubclassData<OperationField>(Operation);
797 }
798
799 /// Return the alignment of the memory that is being allocated by the
800 /// instruction.
801 Align getAlign() const {
802 return Align(1ULL << getSubclassData<AlignmentField>());
803 }
804
805 void setAlignment(Align Align) {
806 setSubclassData<AlignmentField>(Log2(Align));
807 }
808
809 /// Return true if this is a RMW on a volatile memory location.
810 ///
811 bool isVolatile() const { return getSubclassData<VolatileField>(); }
812
813 /// Specify whether this is a volatile RMW or not.
814 ///
815 void setVolatile(bool V) { setSubclassData<VolatileField>(V); }
816
817 /// Transparently provide more efficient getOperand methods.
818 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
819
820 /// Returns the ordering constraint of this rmw instruction.
821 AtomicOrdering getOrdering() const {
822 return getSubclassData<AtomicOrderingField>();
823 }
824
825 /// Sets the ordering constraint of this rmw instruction.
826 void setOrdering(AtomicOrdering Ordering) {
827 assert(Ordering != AtomicOrdering::NotAtomic &&((Ordering != AtomicOrdering::NotAtomic && "atomicrmw instructions can only be atomic."
) ? static_cast<void> (0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"atomicrmw instructions can only be atomic.\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 828, __PRETTY_FUNCTION__))
828 "atomicrmw instructions can only be atomic.")((Ordering != AtomicOrdering::NotAtomic && "atomicrmw instructions can only be atomic."
) ? static_cast<void> (0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"atomicrmw instructions can only be atomic.\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 828, __PRETTY_FUNCTION__))
;
829 setSubclassData<AtomicOrderingField>(Ordering);
830 }
831
832 /// Returns the synchronization scope ID of this rmw instruction.
833 SyncScope::ID getSyncScopeID() const {
834 return SSID;
835 }
836
837 /// Sets the synchronization scope ID of this rmw instruction.
838 void setSyncScopeID(SyncScope::ID SSID) {
839 this->SSID = SSID;
840 }
841
842 Value *getPointerOperand() { return getOperand(0); }
843 const Value *getPointerOperand() const { return getOperand(0); }
844 static unsigned getPointerOperandIndex() { return 0U; }
845
846 Value *getValOperand() { return getOperand(1); }
847 const Value *getValOperand() const { return getOperand(1); }
848
849 /// Returns the address space of the pointer operand.
850 unsigned getPointerAddressSpace() const {
851 return getPointerOperand()->getType()->getPointerAddressSpace();
852 }
853
854 bool isFloatingPointOperation() const {
855 return isFPOperation(getOperation());
856 }
857
858 // Methods for support type inquiry through isa, cast, and dyn_cast:
859 static bool classof(const Instruction *I) {
860 return I->getOpcode() == Instruction::AtomicRMW;
861 }
862 static bool classof(const Value *V) {
863 return isa<Instruction>(V) && classof(cast<Instruction>(V));
864 }
865
866private:
867 void Init(BinOp Operation, Value *Ptr, Value *Val, Align Align,
868 AtomicOrdering Ordering, SyncScope::ID SSID);
869
870 // Shadow Instruction::setInstructionSubclassData with a private forwarding
871 // method so that subclasses cannot accidentally use it.
872 template <typename Bitfield>
873 void setSubclassData(typename Bitfield::Type Value) {
874 Instruction::setSubclassData<Bitfield>(Value);
875 }
876
877 /// The synchronization scope ID of this rmw instruction. Not quite enough
878 /// room in SubClassData for everything, so synchronization scope ID gets its
879 /// own field.
880 SyncScope::ID SSID;
881};
882
883template <>
884struct OperandTraits<AtomicRMWInst>
885 : public FixedNumOperandTraits<AtomicRMWInst,2> {
886};
887
888DEFINE_TRANSPARENT_OPERAND_ACCESSORS(AtomicRMWInst, Value)AtomicRMWInst::op_iterator AtomicRMWInst::op_begin() { return
OperandTraits<AtomicRMWInst>::op_begin(this); } AtomicRMWInst
::const_op_iterator AtomicRMWInst::op_begin() const { return OperandTraits
<AtomicRMWInst>::op_begin(const_cast<AtomicRMWInst*>
(this)); } AtomicRMWInst::op_iterator AtomicRMWInst::op_end()
{ return OperandTraits<AtomicRMWInst>::op_end(this); }
AtomicRMWInst::const_op_iterator AtomicRMWInst::op_end() const
{ return OperandTraits<AtomicRMWInst>::op_end(const_cast
<AtomicRMWInst*>(this)); } Value *AtomicRMWInst::getOperand
(unsigned i_nocapture) const { ((i_nocapture < OperandTraits
<AtomicRMWInst>::operands(this) && "getOperand() out of range!"
) ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<AtomicRMWInst>::operands(this) && \"getOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 888, __PRETTY_FUNCTION__)); return cast_or_null<Value>
( OperandTraits<AtomicRMWInst>::op_begin(const_cast<
AtomicRMWInst*>(this))[i_nocapture].get()); } void AtomicRMWInst
::setOperand(unsigned i_nocapture, Value *Val_nocapture) { ((
i_nocapture < OperandTraits<AtomicRMWInst>::operands
(this) && "setOperand() out of range!") ? static_cast
<void> (0) : __assert_fail ("i_nocapture < OperandTraits<AtomicRMWInst>::operands(this) && \"setOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 888, __PRETTY_FUNCTION__)); OperandTraits<AtomicRMWInst>
::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned AtomicRMWInst
::getNumOperands() const { return OperandTraits<AtomicRMWInst
>::operands(this); } template <int Idx_nocapture> Use
&AtomicRMWInst::Op() { return this->OpFrom<Idx_nocapture
>(this); } template <int Idx_nocapture> const Use &
AtomicRMWInst::Op() const { return this->OpFrom<Idx_nocapture
>(this); }
889
890//===----------------------------------------------------------------------===//
891// GetElementPtrInst Class
892//===----------------------------------------------------------------------===//
893
894// checkGEPType - Simple wrapper function to give a better assertion failure
895// message on bad indexes for a gep instruction.
896//
897inline Type *checkGEPType(Type *Ty) {
898 assert(Ty && "Invalid GetElementPtrInst indices for type!")((Ty && "Invalid GetElementPtrInst indices for type!"
) ? static_cast<void> (0) : __assert_fail ("Ty && \"Invalid GetElementPtrInst indices for type!\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 898, __PRETTY_FUNCTION__))
;
899 return Ty;
900}
901
902/// an instruction for type-safe pointer arithmetic to
903/// access elements of arrays and structs
904///
905class GetElementPtrInst : public Instruction {
906 Type *SourceElementType;
907 Type *ResultElementType;
908
909 GetElementPtrInst(const GetElementPtrInst &GEPI);
910
911 /// Constructors - Create a getelementptr instruction with a base pointer an
912 /// list of indices. The first ctor can optionally insert before an existing
913 /// instruction, the second appends the new instruction to the specified
914 /// BasicBlock.
915 inline GetElementPtrInst(Type *PointeeType, Value *Ptr,
916 ArrayRef<Value *> IdxList, unsigned Values,
917 const Twine &NameStr, Instruction *InsertBefore);
918 inline GetElementPtrInst(Type *PointeeType, Value *Ptr,
919 ArrayRef<Value *> IdxList, unsigned Values,
920 const Twine &NameStr, BasicBlock *InsertAtEnd);
921
922 void init(Value *Ptr, ArrayRef<Value *> IdxList, const Twine &NameStr);
923
924protected:
925 // Note: Instruction needs to be a friend here to call cloneImpl.
926 friend class Instruction;
927
928 GetElementPtrInst *cloneImpl() const;
929
930public:
931 static GetElementPtrInst *Create(Type *PointeeType, Value *Ptr,
932 ArrayRef<Value *> IdxList,
933 const Twine &NameStr = "",
934 Instruction *InsertBefore = nullptr) {
935 unsigned Values = 1 + unsigned(IdxList.size());
936 if (!PointeeType)
937 PointeeType =
938 cast<PointerType>(Ptr->getType()->getScalarType())->getElementType();
939 else
940 assert(((PointeeType == cast<PointerType>(Ptr->getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 942, __PRETTY_FUNCTION__))
941 PointeeType ==((PointeeType == cast<PointerType>(Ptr->getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 942, __PRETTY_FUNCTION__))
942 cast<PointerType>(Ptr->getType()->getScalarType())->getElementType())((PointeeType == cast<PointerType>(Ptr->getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 942, __PRETTY_FUNCTION__))
;
943 return new (Values) GetElementPtrInst(PointeeType, Ptr, IdxList, Values,
944 NameStr, InsertBefore);
945 }
946
947 static GetElementPtrInst *Create(Type *PointeeType, Value *Ptr,
948 ArrayRef<Value *> IdxList,
949 const Twine &NameStr,
950 BasicBlock *InsertAtEnd) {
951 unsigned Values = 1 + unsigned(IdxList.size());
952 if (!PointeeType)
953 PointeeType =
954 cast<PointerType>(Ptr->getType()->getScalarType())->getElementType();
955 else
956 assert(((PointeeType == cast<PointerType>(Ptr->getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 958, __PRETTY_FUNCTION__))
957 PointeeType ==((PointeeType == cast<PointerType>(Ptr->getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 958, __PRETTY_FUNCTION__))
958 cast<PointerType>(Ptr->getType()->getScalarType())->getElementType())((PointeeType == cast<PointerType>(Ptr->getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 958, __PRETTY_FUNCTION__))
;
959 return new (Values) GetElementPtrInst(PointeeType, Ptr, IdxList, Values,
960 NameStr, InsertAtEnd);
961 }
962
963 /// Create an "inbounds" getelementptr. See the documentation for the
964 /// "inbounds" flag in LangRef.html for details.
965 static GetElementPtrInst *CreateInBounds(Value *Ptr,
966 ArrayRef<Value *> IdxList,
967 const Twine &NameStr = "",
968 Instruction *InsertBefore = nullptr){
969 return CreateInBounds(nullptr, Ptr, IdxList, NameStr, InsertBefore);
970 }
971
972 static GetElementPtrInst *
973 CreateInBounds(Type *PointeeType, Value *Ptr, ArrayRef<Value *> IdxList,
974 const Twine &NameStr = "",
975 Instruction *InsertBefore = nullptr) {
976 GetElementPtrInst *GEP =
977 Create(PointeeType, Ptr, IdxList, NameStr, InsertBefore);
978 GEP->setIsInBounds(true);
979 return GEP;
980 }
981
982 static GetElementPtrInst *CreateInBounds(Value *Ptr,
983 ArrayRef<Value *> IdxList,
984 const Twine &NameStr,
985 BasicBlock *InsertAtEnd) {
986 return CreateInBounds(nullptr, Ptr, IdxList, NameStr, InsertAtEnd);
987 }
988
989 static GetElementPtrInst *CreateInBounds(Type *PointeeType, Value *Ptr,
990 ArrayRef<Value *> IdxList,
991 const Twine &NameStr,
992 BasicBlock *InsertAtEnd) {
993 GetElementPtrInst *GEP =
994 Create(PointeeType, Ptr, IdxList, NameStr, InsertAtEnd);
995 GEP->setIsInBounds(true);
996 return GEP;
997 }
998
999 /// Transparently provide more efficient getOperand methods.
1000 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
1001
1002 Type *getSourceElementType() const { return SourceElementType; }
1003
1004 void setSourceElementType(Type *Ty) { SourceElementType = Ty; }
1005 void setResultElementType(Type *Ty) { ResultElementType = Ty; }
1006
1007 Type *getResultElementType() const {
1008 assert(ResultElementType ==((ResultElementType == cast<PointerType>(getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("ResultElementType == cast<PointerType>(getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 1009, __PRETTY_FUNCTION__))
1009 cast<PointerType>(getType()->getScalarType())->getElementType())((ResultElementType == cast<PointerType>(getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("ResultElementType == cast<PointerType>(getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 1009, __PRETTY_FUNCTION__))
;
1010 return ResultElementType;
1011 }
1012
1013 /// Returns the address space of this instruction's pointer type.
1014 unsigned getAddressSpace() const {
1015 // Note that this is always the same as the pointer operand's address space
1016 // and that is cheaper to compute, so cheat here.
1017 return getPointerAddressSpace();
1018 }
1019
1020 /// Returns the result type of a getelementptr with the given source
1021 /// element type and indexes.
1022 ///
1023 /// Null is returned if the indices are invalid for the specified
1024 /// source element type.
1025 static Type *getIndexedType(Type *Ty, ArrayRef<Value *> IdxList);
1026 static Type *getIndexedType(Type *Ty, ArrayRef<Constant *> IdxList);
1027 static Type *getIndexedType(Type *Ty, ArrayRef<uint64_t> IdxList);
1028
1029 /// Return the type of the element at the given index of an indexable
1030 /// type. This is equivalent to "getIndexedType(Agg, {Zero, Idx})".
1031 ///
1032 /// Returns null if the type can't be indexed, or the given index is not
1033 /// legal for the given type.
1034 static Type *getTypeAtIndex(Type *Ty, Value *Idx);
1035 static Type *getTypeAtIndex(Type *Ty, uint64_t Idx);
1036
1037 inline op_iterator idx_begin() { return op_begin()+1; }
1038 inline const_op_iterator idx_begin() const { return op_begin()+1; }
1039 inline op_iterator idx_end() { return op_end(); }
1040 inline const_op_iterator idx_end() const { return op_end(); }
1041
1042 inline iterator_range<op_iterator> indices() {
1043 return make_range(idx_begin(), idx_end());
1044 }
1045
1046 inline iterator_range<const_op_iterator> indices() const {
1047 return make_range(idx_begin(), idx_end());
1048 }
1049
1050 Value *getPointerOperand() {
1051 return getOperand(0);
1052 }
1053 const Value *getPointerOperand() const {
1054 return getOperand(0);
1055 }
1056 static unsigned getPointerOperandIndex() {
1057 return 0U; // get index for modifying correct operand.
1058 }
1059
1060 /// Method to return the pointer operand as a
1061 /// PointerType.
1062 Type *getPointerOperandType() const {
1063 return getPointerOperand()->getType();
1064 }
1065
1066 /// Returns the address space of the pointer operand.
1067 unsigned getPointerAddressSpace() const {
1068 return getPointerOperandType()->getPointerAddressSpace();
1069 }
1070
1071 /// Returns the pointer type returned by the GEP
1072 /// instruction, which may be a vector of pointers.
1073 static Type *getGEPReturnType(Type *ElTy, Value *Ptr,
1074 ArrayRef<Value *> IdxList) {
1075 Type *PtrTy = PointerType::get(checkGEPType(getIndexedType(ElTy, IdxList)),
1076 Ptr->getType()->getPointerAddressSpace());
1077 // Vector GEP
1078 if (auto *PtrVTy = dyn_cast<VectorType>(Ptr->getType())) {
1079 ElementCount EltCount = PtrVTy->getElementCount();
1080 return VectorType::get(PtrTy, EltCount);
1081 }
1082 for (Value *Index : IdxList)
1083 if (auto *IndexVTy = dyn_cast<VectorType>(Index->getType())) {
1084 ElementCount EltCount = IndexVTy->getElementCount();
1085 return VectorType::get(PtrTy, EltCount);
1086 }
1087 // Scalar GEP
1088 return PtrTy;
1089 }
1090
1091 unsigned getNumIndices() const { // Note: always non-negative
1092 return getNumOperands() - 1;
1093 }
1094
1095 bool hasIndices() const {
1096 return getNumOperands() > 1;
1097 }
1098
1099 /// Return true if all of the indices of this GEP are
1100 /// zeros. If so, the result pointer and the first operand have the same
1101 /// value, just potentially different types.
1102 bool hasAllZeroIndices() const;
1103
1104 /// Return true if all of the indices of this GEP are
1105 /// constant integers. If so, the result pointer and the first operand have
1106 /// a constant offset between them.
1107 bool hasAllConstantIndices() const;
1108
1109 /// Set or clear the inbounds flag on this GEP instruction.
1110 /// See LangRef.html for the meaning of inbounds on a getelementptr.
1111 void setIsInBounds(bool b = true);
1112
1113 /// Determine whether the GEP has the inbounds flag.
1114 bool isInBounds() const;
1115
1116 /// Accumulate the constant address offset of this GEP if possible.
1117 ///
1118 /// This routine accepts an APInt into which it will accumulate the constant
1119 /// offset of this GEP if the GEP is in fact constant. If the GEP is not
1120 /// all-constant, it returns false and the value of the offset APInt is
1121 /// undefined (it is *not* preserved!). The APInt passed into this routine
1122 /// must be at least as wide as the IntPtr type for the address space of
1123 /// the base GEP pointer.
1124 bool accumulateConstantOffset(const DataLayout &DL, APInt &Offset) const;
1125
1126 // Methods for support type inquiry through isa, cast, and dyn_cast:
1127 static bool classof(const Instruction *I) {
1128 return (I->getOpcode() == Instruction::GetElementPtr);
1129 }
1130 static bool classof(const Value *V) {
1131 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1132 }
1133};
1134
1135template <>
1136struct OperandTraits<GetElementPtrInst> :
1137 public VariadicOperandTraits<GetElementPtrInst, 1> {
1138};
1139
1140GetElementPtrInst::GetElementPtrInst(Type *PointeeType, Value *Ptr,
1141 ArrayRef<Value *> IdxList, unsigned Values,
1142 const Twine &NameStr,
1143 Instruction *InsertBefore)
1144 : Instruction(getGEPReturnType(PointeeType, Ptr, IdxList), GetElementPtr,
1145 OperandTraits<GetElementPtrInst>::op_end(this) - Values,
1146 Values, InsertBefore),
1147 SourceElementType(PointeeType),
1148 ResultElementType(getIndexedType(PointeeType, IdxList)) {
1149 assert(ResultElementType ==((ResultElementType == cast<PointerType>(getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("ResultElementType == cast<PointerType>(getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 1150, __PRETTY_FUNCTION__))
1150 cast<PointerType>(getType()->getScalarType())->getElementType())((ResultElementType == cast<PointerType>(getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("ResultElementType == cast<PointerType>(getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 1150, __PRETTY_FUNCTION__))
;
1151 init(Ptr, IdxList, NameStr);
1152}
1153
1154GetElementPtrInst::GetElementPtrInst(Type *PointeeType, Value *Ptr,
1155 ArrayRef<Value *> IdxList, unsigned Values,
1156 const Twine &NameStr,
1157 BasicBlock *InsertAtEnd)
1158 : Instruction(getGEPReturnType(PointeeType, Ptr, IdxList), GetElementPtr,
1159 OperandTraits<GetElementPtrInst>::op_end(this) - Values,
1160 Values, InsertAtEnd),
1161 SourceElementType(PointeeType),
1162 ResultElementType(getIndexedType(PointeeType, IdxList)) {
1163 assert(ResultElementType ==((ResultElementType == cast<PointerType>(getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("ResultElementType == cast<PointerType>(getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 1164, __PRETTY_FUNCTION__))
1164 cast<PointerType>(getType()->getScalarType())->getElementType())((ResultElementType == cast<PointerType>(getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("ResultElementType == cast<PointerType>(getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 1164, __PRETTY_FUNCTION__))
;
1165 init(Ptr, IdxList, NameStr);
1166}
1167
1168DEFINE_TRANSPARENT_OPERAND_ACCESSORS(GetElementPtrInst, Value)GetElementPtrInst::op_iterator GetElementPtrInst::op_begin() {
return OperandTraits<GetElementPtrInst>::op_begin(this
); } GetElementPtrInst::const_op_iterator GetElementPtrInst::
op_begin() const { return OperandTraits<GetElementPtrInst>
::op_begin(const_cast<GetElementPtrInst*>(this)); } GetElementPtrInst
::op_iterator GetElementPtrInst::op_end() { return OperandTraits
<GetElementPtrInst>::op_end(this); } GetElementPtrInst::
const_op_iterator GetElementPtrInst::op_end() const { return OperandTraits
<GetElementPtrInst>::op_end(const_cast<GetElementPtrInst
*>(this)); } Value *GetElementPtrInst::getOperand(unsigned
i_nocapture) const { ((i_nocapture < OperandTraits<GetElementPtrInst
>::operands(this) && "getOperand() out of range!")
? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<GetElementPtrInst>::operands(this) && \"getOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 1168, __PRETTY_FUNCTION__)); return cast_or_null<Value>
( OperandTraits<GetElementPtrInst>::op_begin(const_cast
<GetElementPtrInst*>(this))[i_nocapture].get()); } void
GetElementPtrInst::setOperand(unsigned i_nocapture, Value *Val_nocapture
) { ((i_nocapture < OperandTraits<GetElementPtrInst>
::operands(this) && "setOperand() out of range!") ? static_cast
<void> (0) : __assert_fail ("i_nocapture < OperandTraits<GetElementPtrInst>::operands(this) && \"setOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 1168, __PRETTY_FUNCTION__)); OperandTraits<GetElementPtrInst
>::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned
GetElementPtrInst::getNumOperands() const { return OperandTraits
<GetElementPtrInst>::operands(this); } template <int
Idx_nocapture> Use &GetElementPtrInst::Op() { return this
->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture
> const Use &GetElementPtrInst::Op() const { return this
->OpFrom<Idx_nocapture>(this); }
1169
1170//===----------------------------------------------------------------------===//
1171// ICmpInst Class
1172//===----------------------------------------------------------------------===//
1173
1174/// This instruction compares its operands according to the predicate given
1175/// to the constructor. It only operates on integers or pointers. The operands
1176/// must be identical types.
1177/// Represent an integer comparison operator.
1178class ICmpInst: public CmpInst {
1179 void AssertOK() {
1180 assert(isIntPredicate() &&((isIntPredicate() && "Invalid ICmp predicate value")
? static_cast<void> (0) : __assert_fail ("isIntPredicate() && \"Invalid ICmp predicate value\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 1181, __PRETTY_FUNCTION__))
1181 "Invalid ICmp predicate value")((isIntPredicate() && "Invalid ICmp predicate value")
? static_cast<void> (0) : __assert_fail ("isIntPredicate() && \"Invalid ICmp predicate value\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 1181, __PRETTY_FUNCTION__))
;
1182 assert(getOperand(0)->getType() == getOperand(1)->getType() &&((getOperand(0)->getType() == getOperand(1)->getType() &&
"Both operands to ICmp instruction are not of the same type!"
) ? static_cast<void> (0) : __assert_fail ("getOperand(0)->getType() == getOperand(1)->getType() && \"Both operands to ICmp instruction are not of the same type!\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 1183, __PRETTY_FUNCTION__))
1183 "Both operands to ICmp instruction are not of the same type!")((getOperand(0)->getType() == getOperand(1)->getType() &&
"Both operands to ICmp instruction are not of the same type!"
) ? static_cast<void> (0) : __assert_fail ("getOperand(0)->getType() == getOperand(1)->getType() && \"Both operands to ICmp instruction are not of the same type!\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 1183, __PRETTY_FUNCTION__))
;
1184 // Check that the operands are the right type
1185 assert((getOperand(0)->getType()->isIntOrIntVectorTy() ||(((getOperand(0)->getType()->isIntOrIntVectorTy() || getOperand
(0)->getType()->isPtrOrPtrVectorTy()) && "Invalid operand types for ICmp instruction"
) ? static_cast<void> (0) : __assert_fail ("(getOperand(0)->getType()->isIntOrIntVectorTy() || getOperand(0)->getType()->isPtrOrPtrVectorTy()) && \"Invalid operand types for ICmp instruction\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 1187, __PRETTY_FUNCTION__))
1186 getOperand(0)->getType()->isPtrOrPtrVectorTy()) &&(((getOperand(0)->getType()->isIntOrIntVectorTy() || getOperand
(0)->getType()->isPtrOrPtrVectorTy()) && "Invalid operand types for ICmp instruction"
) ? static_cast<void> (0) : __assert_fail ("(getOperand(0)->getType()->isIntOrIntVectorTy() || getOperand(0)->getType()->isPtrOrPtrVectorTy()) && \"Invalid operand types for ICmp instruction\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 1187, __PRETTY_FUNCTION__))
1187 "Invalid operand types for ICmp instruction")(((getOperand(0)->getType()->isIntOrIntVectorTy() || getOperand
(0)->getType()->isPtrOrPtrVectorTy()) && "Invalid operand types for ICmp instruction"
) ? static_cast<void> (0) : __assert_fail ("(getOperand(0)->getType()->isIntOrIntVectorTy() || getOperand(0)->getType()->isPtrOrPtrVectorTy()) && \"Invalid operand types for ICmp instruction\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 1187, __PRETTY_FUNCTION__))
;
1188 }
1189
1190protected:
1191 // Note: Instruction needs to be a friend here to call cloneImpl.
1192 friend class Instruction;
1193
1194 /// Clone an identical ICmpInst
1195 ICmpInst *cloneImpl() const;
1196
1197public:
1198 /// Constructor with insert-before-instruction semantics.
1199 ICmpInst(
1200 Instruction *InsertBefore, ///< Where to insert
1201 Predicate pred, ///< The predicate to use for the comparison
1202 Value *LHS, ///< The left-hand-side of the expression
1203 Value *RHS, ///< The right-hand-side of the expression
1204 const Twine &NameStr = "" ///< Name of the instruction
1205 ) : CmpInst(makeCmpResultType(LHS->getType()),
1206 Instruction::ICmp, pred, LHS, RHS, NameStr,
1207 InsertBefore) {
1208#ifndef NDEBUG
1209 AssertOK();
1210#endif
1211 }
1212
1213 /// Constructor with insert-at-end semantics.
1214 ICmpInst(
1215 BasicBlock &InsertAtEnd, ///< Block to insert into.
1216 Predicate pred, ///< The predicate to use for the comparison
1217 Value *LHS, ///< The left-hand-side of the expression
1218 Value *RHS, ///< The right-hand-side of the expression
1219 const Twine &NameStr = "" ///< Name of the instruction
1220 ) : CmpInst(makeCmpResultType(LHS->getType()),
1221 Instruction::ICmp, pred, LHS, RHS, NameStr,
1222 &InsertAtEnd) {
1223#ifndef NDEBUG
1224 AssertOK();
1225#endif
1226 }
1227
1228 /// Constructor with no-insertion semantics
1229 ICmpInst(
1230 Predicate pred, ///< The predicate to use for the comparison
1231 Value *LHS, ///< The left-hand-side of the expression
1232 Value *RHS, ///< The right-hand-side of the expression
1233 const Twine &NameStr = "" ///< Name of the instruction
1234 ) : CmpInst(makeCmpResultType(LHS->getType()),
1235 Instruction::ICmp, pred, LHS, RHS, NameStr) {
1236#ifndef NDEBUG
1237 AssertOK();
1238#endif
1239 }
1240
1241 /// For example, EQ->EQ, SLE->SLE, UGT->SGT, etc.
1242 /// @returns the predicate that would be the result if the operand were
1243 /// regarded as signed.
1244 /// Return the signed version of the predicate
1245 Predicate getSignedPredicate() const {
1246 return getSignedPredicate(getPredicate());
1247 }
1248
1249 /// This is a static version that you can use without an instruction.
1250 /// Return the signed version of the predicate.
1251 static Predicate getSignedPredicate(Predicate pred);
1252
1253 /// For example, EQ->EQ, SLE->ULE, UGT->UGT, etc.
1254 /// @returns the predicate that would be the result if the operand were
1255 /// regarded as unsigned.
1256 /// Return the unsigned version of the predicate
1257 Predicate getUnsignedPredicate() const {
1258 return getUnsignedPredicate(getPredicate());
1259 }
1260
1261 /// This is a static version that you can use without an instruction.
1262 /// Return the unsigned version of the predicate.
1263 static Predicate getUnsignedPredicate(Predicate pred);
1264
1265 /// Return true if this predicate is either EQ or NE. This also
1266 /// tests for commutativity.
1267 static bool isEquality(Predicate P) {
1268 return P == ICMP_EQ || P == ICMP_NE;
1269 }
1270
1271 /// Return true if this predicate is either EQ or NE. This also
1272 /// tests for commutativity.
1273 bool isEquality() const {
1274 return isEquality(getPredicate());
1275 }
1276
1277 /// @returns true if the predicate of this ICmpInst is commutative
1278 /// Determine if this relation is commutative.
1279 bool isCommutative() const { return isEquality(); }
1280
1281 /// Return true if the predicate is relational (not EQ or NE).
1282 ///
1283 bool isRelational() const {
1284 return !isEquality();
1285 }
1286
1287 /// Return true if the predicate is relational (not EQ or NE).
1288 ///
1289 static bool isRelational(Predicate P) {
1290 return !isEquality(P);
1291 }
1292
1293 /// Return true if the predicate is SGT or UGT.
1294 ///
1295 static bool isGT(Predicate P) {
1296 return P == ICMP_SGT || P == ICMP_UGT;
1297 }
1298
1299 /// Return true if the predicate is SLT or ULT.
1300 ///
1301 static bool isLT(Predicate P) {
1302 return P == ICMP_SLT || P == ICMP_ULT;
1303 }
1304
1305 /// Return true if the predicate is SGE or UGE.
1306 ///
1307 static bool isGE(Predicate P) {
1308 return P == ICMP_SGE || P == ICMP_UGE;
1309 }
1310
1311 /// Return true if the predicate is SLE or ULE.
1312 ///
1313 static bool isLE(Predicate P) {
1314 return P == ICMP_SLE || P == ICMP_ULE;
1315 }
1316
1317 /// Exchange the two operands to this instruction in such a way that it does
1318 /// not modify the semantics of the instruction. The predicate value may be
1319 /// changed to retain the same result if the predicate is order dependent
1320 /// (e.g. ult).
1321 /// Swap operands and adjust predicate.
1322 void swapOperands() {
1323 setPredicate(getSwappedPredicate());
1324 Op<0>().swap(Op<1>());
1325 }
1326
1327 // Methods for support type inquiry through isa, cast, and dyn_cast:
1328 static bool classof(const Instruction *I) {
1329 return I->getOpcode() == Instruction::ICmp;
1330 }
1331 static bool classof(const Value *V) {
1332 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1333 }
1334};
1335
1336//===----------------------------------------------------------------------===//
1337// FCmpInst Class
1338//===----------------------------------------------------------------------===//
1339
1340/// This instruction compares its operands according to the predicate given
1341/// to the constructor. It only operates on floating point values or packed
1342/// vectors of floating point values. The operands must be identical types.
1343/// Represents a floating point comparison operator.
1344class FCmpInst: public CmpInst {
1345 void AssertOK() {
1346 assert(isFPPredicate() && "Invalid FCmp predicate value")((isFPPredicate() && "Invalid FCmp predicate value") ?
static_cast<void> (0) : __assert_fail ("isFPPredicate() && \"Invalid FCmp predicate value\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 1346, __PRETTY_FUNCTION__))
;
1347 assert(getOperand(0)->getType() == getOperand(1)->getType() &&((getOperand(0)->getType() == getOperand(1)->getType() &&
"Both operands to FCmp instruction are not of the same type!"
) ? static_cast<void> (0) : __assert_fail ("getOperand(0)->getType() == getOperand(1)->getType() && \"Both operands to FCmp instruction are not of the same type!\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 1348, __PRETTY_FUNCTION__))
1348 "Both operands to FCmp instruction are not of the same type!")((getOperand(0)->getType() == getOperand(1)->getType() &&
"Both operands to FCmp instruction are not of the same type!"
) ? static_cast<void> (0) : __assert_fail ("getOperand(0)->getType() == getOperand(1)->getType() && \"Both operands to FCmp instruction are not of the same type!\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 1348, __PRETTY_FUNCTION__))
;
1349 // Check that the operands are the right type
1350 assert(getOperand(0)->getType()->isFPOrFPVectorTy() &&((getOperand(0)->getType()->isFPOrFPVectorTy() &&
"Invalid operand types for FCmp instruction") ? static_cast<
void> (0) : __assert_fail ("getOperand(0)->getType()->isFPOrFPVectorTy() && \"Invalid operand types for FCmp instruction\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 1351, __PRETTY_FUNCTION__))
1351 "Invalid operand types for FCmp instruction")((getOperand(0)->getType()->isFPOrFPVectorTy() &&
"Invalid operand types for FCmp instruction") ? static_cast<
void> (0) : __assert_fail ("getOperand(0)->getType()->isFPOrFPVectorTy() && \"Invalid operand types for FCmp instruction\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 1351, __PRETTY_FUNCTION__))
;
1352 }
1353
1354protected:
1355 // Note: Instruction needs to be a friend here to call cloneImpl.
1356 friend class Instruction;
1357
1358 /// Clone an identical FCmpInst
1359 FCmpInst *cloneImpl() const;
1360
1361public:
1362 /// Constructor with insert-before-instruction semantics.
1363 FCmpInst(
1364 Instruction *InsertBefore, ///< Where to insert
1365 Predicate pred, ///< The predicate to use for the comparison
1366 Value *LHS, ///< The left-hand-side of the expression
1367 Value *RHS, ///< The right-hand-side of the expression
1368 const Twine &NameStr = "" ///< Name of the instruction
1369 ) : CmpInst(makeCmpResultType(LHS->getType()),
1370 Instruction::FCmp, pred, LHS, RHS, NameStr,
1371 InsertBefore) {
1372 AssertOK();
1373 }
1374
1375 /// Constructor with insert-at-end semantics.
1376 FCmpInst(
1377 BasicBlock &InsertAtEnd, ///< Block to insert into.
1378 Predicate pred, ///< The predicate to use for the comparison
1379 Value *LHS, ///< The left-hand-side of the expression
1380 Value *RHS, ///< The right-hand-side of the expression
1381 const Twine &NameStr = "" ///< Name of the instruction
1382 ) : CmpInst(makeCmpResultType(LHS->getType()),
1383 Instruction::FCmp, pred, LHS, RHS, NameStr,
1384 &InsertAtEnd) {
1385 AssertOK();
1386 }
1387
1388 /// Constructor with no-insertion semantics
1389 FCmpInst(
1390 Predicate Pred, ///< The predicate to use for the comparison
1391 Value *LHS, ///< The left-hand-side of the expression
1392 Value *RHS, ///< The right-hand-side of the expression
1393 const Twine &NameStr = "", ///< Name of the instruction
1394 Instruction *FlagsSource = nullptr
1395 ) : CmpInst(makeCmpResultType(LHS->getType()), Instruction::FCmp, Pred, LHS,
1396 RHS, NameStr, nullptr, FlagsSource) {
1397 AssertOK();
1398 }
1399
1400 /// @returns true if the predicate of this instruction is EQ or NE.
1401 /// Determine if this is an equality predicate.
1402 static bool isEquality(Predicate Pred) {
1403 return Pred == FCMP_OEQ || Pred == FCMP_ONE || Pred == FCMP_UEQ ||
1404 Pred == FCMP_UNE;
1405 }
1406
1407 /// @returns true if the predicate of this instruction is EQ or NE.
1408 /// Determine if this is an equality predicate.
1409 bool isEquality() const { return isEquality(getPredicate()); }
1410
1411 /// @returns true if the predicate of this instruction is commutative.
1412 /// Determine if this is a commutative predicate.
1413 bool isCommutative() const {
1414 return isEquality() ||
1415 getPredicate() == FCMP_FALSE ||
1416 getPredicate() == FCMP_TRUE ||
1417 getPredicate() == FCMP_ORD ||
1418 getPredicate() == FCMP_UNO;
1419 }
1420
1421 /// @returns true if the predicate is relational (not EQ or NE).
1422 /// Determine if this a relational predicate.
1423 bool isRelational() const { return !isEquality(); }
1424
1425 /// Exchange the two operands to this instruction in such a way that it does
1426 /// not modify the semantics of the instruction. The predicate value may be
1427 /// changed to retain the same result if the predicate is order dependent
1428 /// (e.g. ult).
1429 /// Swap operands and adjust predicate.
1430 void swapOperands() {
1431 setPredicate(getSwappedPredicate());
1432 Op<0>().swap(Op<1>());
1433 }
1434
1435 /// Methods for support type inquiry through isa, cast, and dyn_cast:
1436 static bool classof(const Instruction *I) {
1437 return I->getOpcode() == Instruction::FCmp;
1438 }
1439 static bool classof(const Value *V) {
1440 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1441 }
1442};
1443
1444//===----------------------------------------------------------------------===//
1445/// This class represents a function call, abstracting a target
1446/// machine's calling convention. This class uses low bit of the SubClassData
1447/// field to indicate whether or not this is a tail call. The rest of the bits
1448/// hold the calling convention of the call.
1449///
1450class CallInst : public CallBase {
1451 CallInst(const CallInst &CI);
1452
1453 /// Construct a CallInst given a range of arguments.
1454 /// Construct a CallInst from a range of arguments
1455 inline CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1456 ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr,
1457 Instruction *InsertBefore);
1458
1459 inline CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1460 const Twine &NameStr, Instruction *InsertBefore)
1461 : CallInst(Ty, Func, Args, None, NameStr, InsertBefore) {}
1462
1463 /// Construct a CallInst given a range of arguments.
1464 /// Construct a CallInst from a range of arguments
1465 inline CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1466 ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr,
1467 BasicBlock *InsertAtEnd);
1468
1469 explicit CallInst(FunctionType *Ty, Value *F, const Twine &NameStr,
1470 Instruction *InsertBefore);
1471
1472 CallInst(FunctionType *ty, Value *F, const Twine &NameStr,
1473 BasicBlock *InsertAtEnd);
1474
1475 void init(FunctionType *FTy, Value *Func, ArrayRef<Value *> Args,
1476 ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr);
1477 void init(FunctionType *FTy, Value *Func, const Twine &NameStr);
1478
1479 /// Compute the number of operands to allocate.
1480 static int ComputeNumOperands(int NumArgs, int NumBundleInputs = 0) {
1481 // We need one operand for the called function, plus the input operand
1482 // counts provided.
1483 return 1 + NumArgs + NumBundleInputs;
1484 }
1485
1486protected:
1487 // Note: Instruction needs to be a friend here to call cloneImpl.
1488 friend class Instruction;
1489
1490 CallInst *cloneImpl() const;
1491
1492public:
1493 static CallInst *Create(FunctionType *Ty, Value *F, const Twine &NameStr = "",
1494 Instruction *InsertBefore = nullptr) {
1495 return new (ComputeNumOperands(0)) CallInst(Ty, F, NameStr, InsertBefore);
1496 }
1497
1498 static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1499 const Twine &NameStr,
1500 Instruction *InsertBefore = nullptr) {
1501 return new (ComputeNumOperands(Args.size()))
1502 CallInst(Ty, Func, Args, None, NameStr, InsertBefore);
1503 }
1504
1505 static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1506 ArrayRef<OperandBundleDef> Bundles = None,
1507 const Twine &NameStr = "",
1508 Instruction *InsertBefore = nullptr) {
1509 const int NumOperands =
1510 ComputeNumOperands(Args.size(), CountBundleInputs(Bundles));
1511 const unsigned DescriptorBytes = Bundles.size() * sizeof(BundleOpInfo);
1512
1513 return new (NumOperands, DescriptorBytes)
1514 CallInst(Ty, Func, Args, Bundles, NameStr, InsertBefore);
1515 }
1516
1517 static CallInst *Create(FunctionType *Ty, Value *F, const Twine &NameStr,
1518 BasicBlock *InsertAtEnd) {
1519 return new (ComputeNumOperands(0)) CallInst(Ty, F, NameStr, InsertAtEnd);
1520 }
1521
1522 static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1523 const Twine &NameStr, BasicBlock *InsertAtEnd) {
1524 return new (ComputeNumOperands(Args.size()))
1525 CallInst(Ty, Func, Args, None, NameStr, InsertAtEnd);
1526 }
1527
1528 static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1529 ArrayRef<OperandBundleDef> Bundles,
1530 const Twine &NameStr, BasicBlock *InsertAtEnd) {
1531 const int NumOperands =
1532 ComputeNumOperands(Args.size(), CountBundleInputs(Bundles));
1533 const unsigned DescriptorBytes = Bundles.size() * sizeof(BundleOpInfo);
1534
1535 return new (NumOperands, DescriptorBytes)
1536 CallInst(Ty, Func, Args, Bundles, NameStr, InsertAtEnd);
1537 }
1538
1539 static CallInst *Create(FunctionCallee Func, const Twine &NameStr = "",
1540 Instruction *InsertBefore = nullptr) {
1541 return Create(Func.getFunctionType(), Func.getCallee(), NameStr,
1542 InsertBefore);
1543 }
1544
1545 static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args,
1546 ArrayRef<OperandBundleDef> Bundles = None,
1547 const Twine &NameStr = "",
1548 Instruction *InsertBefore = nullptr) {
1549 return Create(Func.getFunctionType(), Func.getCallee(), Args, Bundles,
1550 NameStr, InsertBefore);
1551 }
1552
1553 static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args,
1554 const Twine &NameStr,
1555 Instruction *InsertBefore = nullptr) {
1556 return Create(Func.getFunctionType(), Func.getCallee(), Args, NameStr,
1557 InsertBefore);
1558 }
1559
1560 static CallInst *Create(FunctionCallee Func, const Twine &NameStr,
1561 BasicBlock *InsertAtEnd) {
1562 return Create(Func.getFunctionType(), Func.getCallee(), NameStr,
1563 InsertAtEnd);
1564 }
1565
1566 static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args,
1567 const Twine &NameStr, BasicBlock *InsertAtEnd) {
1568 return Create(Func.getFunctionType(), Func.getCallee(), Args, NameStr,
1569 InsertAtEnd);
1570 }
1571
1572 static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args,
1573 ArrayRef<OperandBundleDef> Bundles,
1574 const Twine &NameStr, BasicBlock *InsertAtEnd) {
1575 return Create(Func.getFunctionType(), Func.getCallee(), Args, Bundles,
1576 NameStr, InsertAtEnd);
1577 }
1578
1579 /// Create a clone of \p CI with a different set of operand bundles and
1580 /// insert it before \p InsertPt.
1581 ///
1582 /// The returned call instruction is identical \p CI in every way except that
1583 /// the operand bundles for the new instruction are set to the operand bundles
1584 /// in \p Bundles.
1585 static CallInst *Create(CallInst *CI, ArrayRef<OperandBundleDef> Bundles,
1586 Instruction *InsertPt = nullptr);
1587
1588 /// Create a clone of \p CI with a different set of operand bundles and
1589 /// insert it before \p InsertPt.
1590 ///
1591 /// The returned call instruction is identical \p CI in every way except that
1592 /// the operand bundle for the new instruction is set to the operand bundle
1593 /// in \p Bundle.
1594 static CallInst *CreateWithReplacedBundle(CallInst *CI,
1595 OperandBundleDef Bundle,
1596 Instruction *InsertPt = nullptr);
1597
1598 /// Generate the IR for a call to malloc:
1599 /// 1. Compute the malloc call's argument as the specified type's size,
1600 /// possibly multiplied by the array size if the array size is not
1601 /// constant 1.
1602 /// 2. Call malloc with that argument.
1603 /// 3. Bitcast the result of the malloc call to the specified type.
1604 static Instruction *CreateMalloc(Instruction *InsertBefore, Type *IntPtrTy,
1605 Type *AllocTy, Value *AllocSize,
1606 Value *ArraySize = nullptr,
1607 Function *MallocF = nullptr,
1608 const Twine &Name = "");
1609 static Instruction *CreateMalloc(BasicBlock *InsertAtEnd, Type *IntPtrTy,
1610 Type *AllocTy, Value *AllocSize,
1611 Value *ArraySize = nullptr,
1612 Function *MallocF = nullptr,
1613 const Twine &Name = "");
1614 static Instruction *CreateMalloc(Instruction *InsertBefore, Type *IntPtrTy,
1615 Type *AllocTy, Value *AllocSize,
1616 Value *ArraySize = nullptr,
1617 ArrayRef<OperandBundleDef> Bundles = None,
1618 Function *MallocF = nullptr,
1619 const Twine &Name = "");
1620 static Instruction *CreateMalloc(BasicBlock *InsertAtEnd, Type *IntPtrTy,
1621 Type *AllocTy, Value *AllocSize,
1622 Value *ArraySize = nullptr,
1623 ArrayRef<OperandBundleDef> Bundles = None,
1624 Function *MallocF = nullptr,
1625 const Twine &Name = "");
1626 /// Generate the IR for a call to the builtin free function.
1627 static Instruction *CreateFree(Value *Source, Instruction *InsertBefore);
1628 static Instruction *CreateFree(Value *Source, BasicBlock *InsertAtEnd);
1629 static Instruction *CreateFree(Value *Source,
1630 ArrayRef<OperandBundleDef> Bundles,
1631 Instruction *InsertBefore);
1632 static Instruction *CreateFree(Value *Source,
1633 ArrayRef<OperandBundleDef> Bundles,
1634 BasicBlock *InsertAtEnd);
1635
1636 // Note that 'musttail' implies 'tail'.
1637 enum TailCallKind : unsigned {
1638 TCK_None = 0,
1639 TCK_Tail = 1,
1640 TCK_MustTail = 2,
1641 TCK_NoTail = 3,
1642 TCK_LAST = TCK_NoTail
1643 };
1644
1645 using TailCallKindField = Bitfield::Element<TailCallKind, 0, 2, TCK_LAST>;
1646 static_assert(
1647 Bitfield::areContiguous<TailCallKindField, CallBase::CallingConvField>(),
1648 "Bitfields must be contiguous");
1649
1650 TailCallKind getTailCallKind() const {
1651 return getSubclassData<TailCallKindField>();
1652 }
1653
1654 bool isTailCall() const {
1655 TailCallKind Kind = getTailCallKind();
1656 return Kind == TCK_Tail || Kind == TCK_MustTail;
1657 }
1658
1659 bool isMustTailCall() const { return getTailCallKind() == TCK_MustTail; }
1660
1661 bool isNoTailCall() const { return getTailCallKind() == TCK_NoTail; }
1662
1663 void setTailCallKind(TailCallKind TCK) {
1664 setSubclassData<TailCallKindField>(TCK);
1665 }
1666
1667 void setTailCall(bool IsTc = true) {
1668 setTailCallKind(IsTc ? TCK_Tail : TCK_None);
1669 }
1670
1671 /// Return true if the call can return twice
1672 bool canReturnTwice() const { return hasFnAttr(Attribute::ReturnsTwice); }
1673 void setCanReturnTwice() {
1674 addAttribute(AttributeList::FunctionIndex, Attribute::ReturnsTwice);
1675 }
1676
1677 // Methods for support type inquiry through isa, cast, and dyn_cast:
1678 static bool classof(const Instruction *I) {
1679 return I->getOpcode() == Instruction::Call;
1680 }
1681 static bool classof(const Value *V) {
1682 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1683 }
1684
1685 /// Updates profile metadata by scaling it by \p S / \p T.
1686 void updateProfWeight(uint64_t S, uint64_t T);
1687
1688private:
1689 // Shadow Instruction::setInstructionSubclassData with a private forwarding
1690 // method so that subclasses cannot accidentally use it.
1691 template <typename Bitfield>
1692 void setSubclassData(typename Bitfield::Type Value) {
1693 Instruction::setSubclassData<Bitfield>(Value);
1694 }
1695};
1696
1697CallInst::CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1698 ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr,
1699 BasicBlock *InsertAtEnd)
1700 : CallBase(Ty->getReturnType(), Instruction::Call,
1701 OperandTraits<CallBase>::op_end(this) -
1702 (Args.size() + CountBundleInputs(Bundles) + 1),
1703 unsigned(Args.size() + CountBundleInputs(Bundles) + 1),
1704 InsertAtEnd) {
1705 init(Ty, Func, Args, Bundles, NameStr);
1706}
1707
1708CallInst::CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1709 ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr,
1710 Instruction *InsertBefore)
1711 : CallBase(Ty->getReturnType(), Instruction::Call,
1712 OperandTraits<CallBase>::op_end(this) -
1713 (Args.size() + CountBundleInputs(Bundles) + 1),
1714 unsigned(Args.size() + CountBundleInputs(Bundles) + 1),
1715 InsertBefore) {
1716 init(Ty, Func, Args, Bundles, NameStr);
1717}
1718
1719//===----------------------------------------------------------------------===//
1720// SelectInst Class
1721//===----------------------------------------------------------------------===//
1722
1723/// This class represents the LLVM 'select' instruction.
1724///
1725class SelectInst : public Instruction {
1726 SelectInst(Value *C, Value *S1, Value *S2, const Twine &NameStr,
1727 Instruction *InsertBefore)
1728 : Instruction(S1->getType(), Instruction::Select,
1729 &Op<0>(), 3, InsertBefore) {
1730 init(C, S1, S2);
1731 setName(NameStr);
1732 }
1733
1734 SelectInst(Value *C, Value *S1, Value *S2, const Twine &NameStr,
1735 BasicBlock *InsertAtEnd)
1736 : Instruction(S1->getType(), Instruction::Select,
1737 &Op<0>(), 3, InsertAtEnd) {
1738 init(C, S1, S2);
1739 setName(NameStr);
1740 }
1741
1742 void init(Value *C, Value *S1, Value *S2) {
1743 assert(!areInvalidOperands(C, S1, S2) && "Invalid operands for select")((!areInvalidOperands(C, S1, S2) && "Invalid operands for select"
) ? static_cast<void> (0) : __assert_fail ("!areInvalidOperands(C, S1, S2) && \"Invalid operands for select\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 1743, __PRETTY_FUNCTION__))
;
1744 Op<0>() = C;
1745 Op<1>() = S1;
1746 Op<2>() = S2;
1747 }
1748
1749protected:
1750 // Note: Instruction needs to be a friend here to call cloneImpl.
1751 friend class Instruction;
1752
1753 SelectInst *cloneImpl() const;
1754
1755public:
1756 static SelectInst *Create(Value *C, Value *S1, Value *S2,
1757 const Twine &NameStr = "",
1758 Instruction *InsertBefore = nullptr,
1759 Instruction *MDFrom = nullptr) {
1760 SelectInst *Sel = new(3) SelectInst(C, S1, S2, NameStr, InsertBefore);
1761 if (MDFrom)
1762 Sel->copyMetadata(*MDFrom);
1763 return Sel;
1764 }
1765
1766 static SelectInst *Create(Value *C, Value *S1, Value *S2,
1767 const Twine &NameStr,
1768 BasicBlock *InsertAtEnd) {
1769 return new(3) SelectInst(C, S1, S2, NameStr, InsertAtEnd);
1770 }
1771
1772 const Value *getCondition() const { return Op<0>(); }
1773 const Value *getTrueValue() const { return Op<1>(); }
1774 const Value *getFalseValue() const { return Op<2>(); }
1775 Value *getCondition() { return Op<0>(); }
1776 Value *getTrueValue() { return Op<1>(); }
1777 Value *getFalseValue() { return Op<2>(); }
1778
1779 void setCondition(Value *V) { Op<0>() = V; }
1780 void setTrueValue(Value *V) { Op<1>() = V; }
1781 void setFalseValue(Value *V) { Op<2>() = V; }
1782
1783 /// Swap the true and false values of the select instruction.
1784 /// This doesn't swap prof metadata.
1785 void swapValues() { Op<1>().swap(Op<2>()); }
1786
1787 /// Return a string if the specified operands are invalid
1788 /// for a select operation, otherwise return null.
1789 static const char *areInvalidOperands(Value *Cond, Value *True, Value *False);
1790
1791 /// Transparently provide more efficient getOperand methods.
1792 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
1793
1794 OtherOps getOpcode() const {
1795 return static_cast<OtherOps>(Instruction::getOpcode());
1796 }
1797
1798 // Methods for support type inquiry through isa, cast, and dyn_cast:
1799 static bool classof(const Instruction *I) {
1800 return I->getOpcode() == Instruction::Select;
1801 }
1802 static bool classof(const Value *V) {
1803 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1804 }
1805};
1806
1807template <>
1808struct OperandTraits<SelectInst> : public FixedNumOperandTraits<SelectInst, 3> {
1809};
1810
1811DEFINE_TRANSPARENT_OPERAND_ACCESSORS(SelectInst, Value)SelectInst::op_iterator SelectInst::op_begin() { return OperandTraits
<SelectInst>::op_begin(this); } SelectInst::const_op_iterator
SelectInst::op_begin() const { return OperandTraits<SelectInst
>::op_begin(const_cast<SelectInst*>(this)); } SelectInst
::op_iterator SelectInst::op_end() { return OperandTraits<
SelectInst>::op_end(this); } SelectInst::const_op_iterator
SelectInst::op_end() const { return OperandTraits<SelectInst
>::op_end(const_cast<SelectInst*>(this)); } Value *SelectInst
::getOperand(unsigned i_nocapture) const { ((i_nocapture <
OperandTraits<SelectInst>::operands(this) && "getOperand() out of range!"
) ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<SelectInst>::operands(this) && \"getOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 1811, __PRETTY_FUNCTION__)); return cast_or_null<Value>
( OperandTraits<SelectInst>::op_begin(const_cast<SelectInst
*>(this))[i_nocapture].get()); } void SelectInst::setOperand
(unsigned i_nocapture, Value *Val_nocapture) { ((i_nocapture <
OperandTraits<SelectInst>::operands(this) && "setOperand() out of range!"
) ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<SelectInst>::operands(this) && \"setOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 1811, __PRETTY_FUNCTION__)); OperandTraits<SelectInst>
::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned SelectInst
::getNumOperands() const { return OperandTraits<SelectInst
>::operands(this); } template <int Idx_nocapture> Use
&SelectInst::Op() { return this->OpFrom<Idx_nocapture
>(this); } template <int Idx_nocapture> const Use &
SelectInst::Op() const { return this->OpFrom<Idx_nocapture
>(this); }
1812
1813//===----------------------------------------------------------------------===//
1814// VAArgInst Class
1815//===----------------------------------------------------------------------===//
1816
1817/// This class represents the va_arg llvm instruction, which returns
1818/// an argument of the specified type given a va_list and increments that list
1819///
1820class VAArgInst : public UnaryInstruction {
1821protected:
1822 // Note: Instruction needs to be a friend here to call cloneImpl.
1823 friend class Instruction;
1824
1825 VAArgInst *cloneImpl() const;
1826
1827public:
1828 VAArgInst(Value *List, Type *Ty, const Twine &NameStr = "",
1829 Instruction *InsertBefore = nullptr)
1830 : UnaryInstruction(Ty, VAArg, List, InsertBefore) {
1831 setName(NameStr);
1832 }
1833
1834 VAArgInst(Value *List, Type *Ty, const Twine &NameStr,
1835 BasicBlock *InsertAtEnd)
1836 : UnaryInstruction(Ty, VAArg, List, InsertAtEnd) {
1837 setName(NameStr);
1838 }
1839
1840 Value *getPointerOperand() { return getOperand(0); }
1841 const Value *getPointerOperand() const { return getOperand(0); }
1842 static unsigned getPointerOperandIndex() { return 0U; }
1843
1844 // Methods for support type inquiry through isa, cast, and dyn_cast:
1845 static bool classof(const Instruction *I) {
1846 return I->getOpcode() == VAArg;
1847 }
1848 static bool classof(const Value *V) {
1849 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1850 }
1851};
1852
1853//===----------------------------------------------------------------------===//
1854// ExtractElementInst Class
1855//===----------------------------------------------------------------------===//
1856
1857/// This instruction extracts a single (scalar)
1858/// element from a VectorType value
1859///
1860class ExtractElementInst : public Instruction {
1861 ExtractElementInst(Value *Vec, Value *Idx, const Twine &NameStr = "",
1862 Instruction *InsertBefore = nullptr);
1863 ExtractElementInst(Value *Vec, Value *Idx, const Twine &NameStr,
1864 BasicBlock *InsertAtEnd);
1865
1866protected:
1867 // Note: Instruction needs to be a friend here to call cloneImpl.
1868 friend class Instruction;
1869
1870 ExtractElementInst *cloneImpl() const;
1871
1872public:
1873 static ExtractElementInst *Create(Value *Vec, Value *Idx,
1874 const Twine &NameStr = "",
1875 Instruction *InsertBefore = nullptr) {
1876 return new(2) ExtractElementInst(Vec, Idx, NameStr, InsertBefore);
1877 }
1878
1879 static ExtractElementInst *Create(Value *Vec, Value *Idx,
1880 const Twine &NameStr,
1881 BasicBlock *InsertAtEnd) {
1882 return new(2) ExtractElementInst(Vec, Idx, NameStr, InsertAtEnd);
1883 }
1884
1885 /// Return true if an extractelement instruction can be
1886 /// formed with the specified operands.
1887 static bool isValidOperands(const Value *Vec, const Value *Idx);
1888
1889 Value *getVectorOperand() { return Op<0>(); }
1890 Value *getIndexOperand() { return Op<1>(); }
1891 const Value *getVectorOperand() const { return Op<0>(); }
1892 const Value *getIndexOperand() const { return Op<1>(); }
1893
1894 VectorType *getVectorOperandType() const {
1895 return cast<VectorType>(getVectorOperand()->getType());
1896 }
1897
1898 /// Transparently provide more efficient getOperand methods.
1899 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
1900
1901 // Methods for support type inquiry through isa, cast, and dyn_cast:
1902 static bool classof(const Instruction *I) {
1903 return I->getOpcode() == Instruction::ExtractElement;
1904 }
1905 static bool classof(const Value *V) {
1906 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1907 }
1908};
1909
1910template <>
1911struct OperandTraits<ExtractElementInst> :
1912 public FixedNumOperandTraits<ExtractElementInst, 2> {
1913};
1914
1915DEFINE_TRANSPARENT_OPERAND_ACCESSORS(ExtractElementInst, Value)ExtractElementInst::op_iterator ExtractElementInst::op_begin(
) { return OperandTraits<ExtractElementInst>::op_begin(
this); } ExtractElementInst::const_op_iterator ExtractElementInst
::op_begin() const { return OperandTraits<ExtractElementInst
>::op_begin(const_cast<ExtractElementInst*>(this)); }
ExtractElementInst::op_iterator ExtractElementInst::op_end()
{ return OperandTraits<ExtractElementInst>::op_end(this
); } ExtractElementInst::const_op_iterator ExtractElementInst
::op_end() const { return OperandTraits<ExtractElementInst
>::op_end(const_cast<ExtractElementInst*>(this)); } Value
*ExtractElementInst::getOperand(unsigned i_nocapture) const {
((i_nocapture < OperandTraits<ExtractElementInst>::
operands(this) && "getOperand() out of range!") ? static_cast
<void> (0) : __assert_fail ("i_nocapture < OperandTraits<ExtractElementInst>::operands(this) && \"getOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 1915, __PRETTY_FUNCTION__)); return cast_or_null<Value>
( OperandTraits<ExtractElementInst>::op_begin(const_cast
<ExtractElementInst*>(this))[i_nocapture].get()); } void
ExtractElementInst::setOperand(unsigned i_nocapture, Value *
Val_nocapture) { ((i_nocapture < OperandTraits<ExtractElementInst
>::operands(this) && "setOperand() out of range!")
? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<ExtractElementInst>::operands(this) && \"setOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 1915, __PRETTY_FUNCTION__)); OperandTraits<ExtractElementInst
>::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned
ExtractElementInst::getNumOperands() const { return OperandTraits
<ExtractElementInst>::operands(this); } template <int
Idx_nocapture> Use &ExtractElementInst::Op() { return
this->OpFrom<Idx_nocapture>(this); } template <int
Idx_nocapture> const Use &ExtractElementInst::Op() const
{ return this->OpFrom<Idx_nocapture>(this); }
1916
1917//===----------------------------------------------------------------------===//
1918// InsertElementInst Class
1919//===----------------------------------------------------------------------===//
1920
1921/// This instruction inserts a single (scalar)
1922/// element into a VectorType value
1923///
1924class InsertElementInst : public Instruction {
1925 InsertElementInst(Value *Vec, Value *NewElt, Value *Idx,
1926 const Twine &NameStr = "",
1927 Instruction *InsertBefore = nullptr);
1928 InsertElementInst(Value *Vec, Value *NewElt, Value *Idx, const Twine &NameStr,
1929 BasicBlock *InsertAtEnd);
1930
1931protected:
1932 // Note: Instruction needs to be a friend here to call cloneImpl.
1933 friend class Instruction;
1934
1935 InsertElementInst *cloneImpl() const;
1936
1937public:
1938 static InsertElementInst *Create(Value *Vec, Value *NewElt, Value *Idx,
1939 const Twine &NameStr = "",
1940 Instruction *InsertBefore = nullptr) {
1941 return new(3) InsertElementInst(Vec, NewElt, Idx, NameStr, InsertBefore);
1942 }
1943
1944 static InsertElementInst *Create(Value *Vec, Value *NewElt, Value *Idx,
1945 const Twine &NameStr,
1946 BasicBlock *InsertAtEnd) {
1947 return new(3) InsertElementInst(Vec, NewElt, Idx, NameStr, InsertAtEnd);
1948 }
1949
1950 /// Return true if an insertelement instruction can be
1951 /// formed with the specified operands.
1952 static bool isValidOperands(const Value *Vec, const Value *NewElt,
1953 const Value *Idx);
1954
1955 /// Overload to return most specific vector type.
1956 ///
1957 VectorType *getType() const {
1958 return cast<VectorType>(Instruction::getType());
1959 }
1960
1961 /// Transparently provide more efficient getOperand methods.
1962 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
1963
1964 // Methods for support type inquiry through isa, cast, and dyn_cast:
1965 static bool classof(const Instruction *I) {
1966 return I->getOpcode() == Instruction::InsertElement;
1967 }
1968 static bool classof(const Value *V) {
1969 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1970 }
1971};
1972
1973template <>
1974struct OperandTraits<InsertElementInst> :
1975 public FixedNumOperandTraits<InsertElementInst, 3> {
1976};
1977
1978DEFINE_TRANSPARENT_OPERAND_ACCESSORS(InsertElementInst, Value)InsertElementInst::op_iterator InsertElementInst::op_begin() {
return OperandTraits<InsertElementInst>::op_begin(this
); } InsertElementInst::const_op_iterator InsertElementInst::
op_begin() const { return OperandTraits<InsertElementInst>
::op_begin(const_cast<InsertElementInst*>(this)); } InsertElementInst
::op_iterator InsertElementInst::op_end() { return OperandTraits
<InsertElementInst>::op_end(this); } InsertElementInst::
const_op_iterator InsertElementInst::op_end() const { return OperandTraits
<InsertElementInst>::op_end(const_cast<InsertElementInst
*>(this)); } Value *InsertElementInst::getOperand(unsigned
i_nocapture) const { ((i_nocapture < OperandTraits<InsertElementInst
>::operands(this) && "getOperand() out of range!")
? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<InsertElementInst>::operands(this) && \"getOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 1978, __PRETTY_FUNCTION__)); return cast_or_null<Value>
( OperandTraits<InsertElementInst>::op_begin(const_cast
<InsertElementInst*>(this))[i_nocapture].get()); } void
InsertElementInst::setOperand(unsigned i_nocapture, Value *Val_nocapture
) { ((i_nocapture < OperandTraits<InsertElementInst>
::operands(this) && "setOperand() out of range!") ? static_cast
<void> (0) : __assert_fail ("i_nocapture < OperandTraits<InsertElementInst>::operands(this) && \"setOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 1978, __PRETTY_FUNCTION__)); OperandTraits<InsertElementInst
>::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned
InsertElementInst::getNumOperands() const { return OperandTraits
<InsertElementInst>::operands(this); } template <int
Idx_nocapture> Use &InsertElementInst::Op() { return this
->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture
> const Use &InsertElementInst::Op() const { return this
->OpFrom<Idx_nocapture>(this); }
1979
1980//===----------------------------------------------------------------------===//
1981// ShuffleVectorInst Class
1982//===----------------------------------------------------------------------===//
1983
1984constexpr int UndefMaskElem = -1;
1985
1986/// This instruction constructs a fixed permutation of two
1987/// input vectors.
1988///
1989/// For each element of the result vector, the shuffle mask selects an element
1990/// from one of the input vectors to copy to the result. Non-negative elements
1991/// in the mask represent an index into the concatenated pair of input vectors.
1992/// UndefMaskElem (-1) specifies that the result element is undefined.
1993///
1994/// For scalable vectors, all the elements of the mask must be 0 or -1. This
1995/// requirement may be relaxed in the future.
1996class ShuffleVectorInst : public Instruction {
1997 SmallVector<int, 4> ShuffleMask;
1998 Constant *ShuffleMaskForBitcode;
1999
2000protected:
2001 // Note: Instruction needs to be a friend here to call cloneImpl.
2002 friend class Instruction;
2003
2004 ShuffleVectorInst *cloneImpl() const;
2005
2006public:
2007 ShuffleVectorInst(Value *V1, Value *V2, Value *Mask,
2008 const Twine &NameStr = "",
2009 Instruction *InsertBefor = nullptr);
2010 ShuffleVectorInst(Value *V1, Value *V2, Value *Mask,
2011 const Twine &NameStr, BasicBlock *InsertAtEnd);
2012 ShuffleVectorInst(Value *V1, Value *V2, ArrayRef<int> Mask,
2013 const Twine &NameStr = "",
2014 Instruction *InsertBefor = nullptr);
2015 ShuffleVectorInst(Value *V1, Value *V2, ArrayRef<int> Mask,
2016 const Twine &NameStr, BasicBlock *InsertAtEnd);
2017
2018 void *operator new(size_t s) { return User::operator new(s, 2); }
2019
2020 /// Swap the operands and adjust the mask to preserve the semantics
2021 /// of the instruction.
2022 void commute();
2023
2024 /// Return true if a shufflevector instruction can be
2025 /// formed with the specified operands.
2026 static bool isValidOperands(const Value *V1, const Value *V2,
2027 const Value *Mask);
2028 static bool isValidOperands(const Value *V1, const Value *V2,
2029 ArrayRef<int> Mask);
2030
2031 /// Overload to return most specific vector type.
2032 ///
2033 VectorType *getType() const {
2034 return cast<VectorType>(Instruction::getType());
2035 }
2036
2037 /// Transparently provide more efficient getOperand methods.
2038 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
2039
2040 /// Return the shuffle mask value of this instruction for the given element
2041 /// index. Return UndefMaskElem if the element is undef.
2042 int getMaskValue(unsigned Elt) const { return ShuffleMask[Elt]; }
2043
2044 /// Convert the input shuffle mask operand to a vector of integers. Undefined
2045 /// elements of the mask are returned as UndefMaskElem.
2046 static void getShuffleMask(const Constant *Mask,
2047 SmallVectorImpl<int> &Result);
2048
2049 /// Return the mask for this instruction as a vector of integers. Undefined
2050 /// elements of the mask are returned as UndefMaskElem.
2051 void getShuffleMask(SmallVectorImpl<int> &Result) const {
2052 Result.assign(ShuffleMask.begin(), ShuffleMask.end());
2053 }
2054
2055 /// Return the mask for this instruction, for use in bitcode.
2056 ///
2057 /// TODO: This is temporary until we decide a new bitcode encoding for
2058 /// shufflevector.
2059 Constant *getShuffleMaskForBitcode() const { return ShuffleMaskForBitcode; }
2060
2061 static Constant *convertShuffleMaskForBitcode(ArrayRef<int> Mask,
2062 Type *ResultTy);
2063
2064 void setShuffleMask(ArrayRef<int> Mask);
2065
2066 ArrayRef<int> getShuffleMask() const { return ShuffleMask; }
2067
2068 /// Return true if this shuffle returns a vector with a different number of
2069 /// elements than its source vectors.
2070 /// Examples: shufflevector <4 x n> A, <4 x n> B, <1,2,3>
2071 /// shufflevector <4 x n> A, <4 x n> B, <1,2,3,4,5>
2072 bool changesLength() const {
2073 unsigned NumSourceElts = cast<VectorType>(Op<0>()->getType())
2074 ->getElementCount()
2075 .getKnownMinValue();
2076 unsigned NumMaskElts = ShuffleMask.size();
2077 return NumSourceElts != NumMaskElts;
2078 }
2079
2080 /// Return true if this shuffle returns a vector with a greater number of
2081 /// elements than its source vectors.
2082 /// Example: shufflevector <2 x n> A, <2 x n> B, <1,2,3>
2083 bool increasesLength() const {
2084 unsigned NumSourceElts = cast<VectorType>(Op<0>()->getType())
2085 ->getElementCount()
2086 .getKnownMinValue();
2087 unsigned NumMaskElts = ShuffleMask.size();
2088 return NumSourceElts < NumMaskElts;
2089 }
2090
2091 /// Return true if this shuffle mask chooses elements from exactly one source
2092 /// vector.
2093 /// Example: <7,5,undef,7>
2094 /// This assumes that vector operands are the same length as the mask.
2095 static bool isSingleSourceMask(ArrayRef<int> Mask);
2096 static bool isSingleSourceMask(const Constant *Mask) {
2097 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((Mask->getType()->isVectorTy() && "Shuffle needs vector constant."
) ? static_cast<void> (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 2097, __PRETTY_FUNCTION__))
;
2098 SmallVector<int, 16> MaskAsInts;
2099 getShuffleMask(Mask, MaskAsInts);
2100 return isSingleSourceMask(MaskAsInts);
2101 }
2102
2103 /// Return true if this shuffle chooses elements from exactly one source
2104 /// vector without changing the length of that vector.
2105 /// Example: shufflevector <4 x n> A, <4 x n> B, <3,0,undef,3>
2106 /// TODO: Optionally allow length-changing shuffles.
2107 bool isSingleSource() const {
2108 return !changesLength() && isSingleSourceMask(ShuffleMask);
2109 }
2110
2111 /// Return true if this shuffle mask chooses elements from exactly one source
2112 /// vector without lane crossings. A shuffle using this mask is not
2113 /// necessarily a no-op because it may change the number of elements from its
2114 /// input vectors or it may provide demanded bits knowledge via undef lanes.
2115 /// Example: <undef,undef,2,3>
2116 static bool isIdentityMask(ArrayRef<int> Mask);
2117 static bool isIdentityMask(const Constant *Mask) {
2118 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((Mask->getType()->isVectorTy() && "Shuffle needs vector constant."
) ? static_cast<void> (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 2118, __PRETTY_FUNCTION__))
;
2119 SmallVector<int, 16> MaskAsInts;
2120 getShuffleMask(Mask, MaskAsInts);
2121 return isIdentityMask(MaskAsInts);
2122 }
2123
2124 /// Return true if this shuffle chooses elements from exactly one source
2125 /// vector without lane crossings and does not change the number of elements
2126 /// from its input vectors.
2127 /// Example: shufflevector <4 x n> A, <4 x n> B, <4,undef,6,undef>
2128 bool isIdentity() const {
2129 return !changesLength() && isIdentityMask(ShuffleMask);
2130 }
2131
2132 /// Return true if this shuffle lengthens exactly one source vector with
2133 /// undefs in the high elements.
2134 bool isIdentityWithPadding() const;
2135
2136 /// Return true if this shuffle extracts the first N elements of exactly one
2137 /// source vector.
2138 bool isIdentityWithExtract() const;
2139
2140 /// Return true if this shuffle concatenates its 2 source vectors. This
2141 /// returns false if either input is undefined. In that case, the shuffle is
2142 /// is better classified as an identity with padding operation.
2143 bool isConcat() const;
2144
2145 /// Return true if this shuffle mask chooses elements from its source vectors
2146 /// without lane crossings. A shuffle using this mask would be
2147 /// equivalent to a vector select with a constant condition operand.
2148 /// Example: <4,1,6,undef>
2149 /// This returns false if the mask does not choose from both input vectors.
2150 /// In that case, the shuffle is better classified as an identity shuffle.
2151 /// This assumes that vector operands are the same length as the mask
2152 /// (a length-changing shuffle can never be equivalent to a vector select).
2153 static bool isSelectMask(ArrayRef<int> Mask);
2154 static bool isSelectMask(const Constant *Mask) {
2155 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((Mask->getType()->isVectorTy() && "Shuffle needs vector constant."
) ? static_cast<void> (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 2155, __PRETTY_FUNCTION__))
;
2156 SmallVector<int, 16> MaskAsInts;
2157 getShuffleMask(Mask, MaskAsInts);
2158 return isSelectMask(MaskAsInts);
2159 }
2160
2161 /// Return true if this shuffle chooses elements from its source vectors
2162 /// without lane crossings and all operands have the same number of elements.
2163 /// In other words, this shuffle is equivalent to a vector select with a
2164 /// constant condition operand.
2165 /// Example: shufflevector <4 x n> A, <4 x n> B, <undef,1,6,3>
2166 /// This returns false if the mask does not choose from both input vectors.
2167 /// In that case, the shuffle is better classified as an identity shuffle.
2168 /// TODO: Optionally allow length-changing shuffles.
2169 bool isSelect() const {
2170 return !changesLength() && isSelectMask(ShuffleMask);
2171 }
2172
2173 /// Return true if this shuffle mask swaps the order of elements from exactly
2174 /// one source vector.
2175 /// Example: <7,6,undef,4>
2176 /// This assumes that vector operands are the same length as the mask.
2177 static bool isReverseMask(ArrayRef<int> Mask);
2178 static bool isReverseMask(const Constant *Mask) {
2179 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((Mask->getType()->isVectorTy() && "Shuffle needs vector constant."
) ? static_cast<void> (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 2179, __PRETTY_FUNCTION__))
;
2180 SmallVector<int, 16> MaskAsInts;
2181 getShuffleMask(Mask, MaskAsInts);
2182 return isReverseMask(MaskAsInts);
2183 }
2184
2185 /// Return true if this shuffle swaps the order of elements from exactly
2186 /// one source vector.
2187 /// Example: shufflevector <4 x n> A, <4 x n> B, <3,undef,1,undef>
2188 /// TODO: Optionally allow length-changing shuffles.
2189 bool isReverse() const {
2190 return !changesLength() && isReverseMask(ShuffleMask);
2191 }
2192
2193 /// Return true if this shuffle mask chooses all elements with the same value
2194 /// as the first element of exactly one source vector.
2195 /// Example: <4,undef,undef,4>
2196 /// This assumes that vector operands are the same length as the mask.
2197 static bool isZeroEltSplatMask(ArrayRef<int> Mask);
2198 static bool isZeroEltSplatMask(const Constant *Mask) {
2199 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((Mask->getType()->isVectorTy() && "Shuffle needs vector constant."
) ? static_cast<void> (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 2199, __PRETTY_FUNCTION__))
;
2200 SmallVector<int, 16> MaskAsInts;
2201 getShuffleMask(Mask, MaskAsInts);
2202 return isZeroEltSplatMask(MaskAsInts);
2203 }
2204
2205 /// Return true if all elements of this shuffle are the same value as the
2206 /// first element of exactly one source vector without changing the length
2207 /// of that vector.
2208 /// Example: shufflevector <4 x n> A, <4 x n> B, <undef,0,undef,0>
2209 /// TODO: Optionally allow length-changing shuffles.
2210 /// TODO: Optionally allow splats from other elements.
2211 bool isZeroEltSplat() const {
2212 return !changesLength() && isZeroEltSplatMask(ShuffleMask);
2213 }
2214
2215 /// Return true if this shuffle mask is a transpose mask.
2216 /// Transpose vector masks transpose a 2xn matrix. They read corresponding
2217 /// even- or odd-numbered vector elements from two n-dimensional source
2218 /// vectors and write each result into consecutive elements of an
2219 /// n-dimensional destination vector. Two shuffles are necessary to complete
2220 /// the transpose, one for the even elements and another for the odd elements.
2221 /// This description closely follows how the TRN1 and TRN2 AArch64
2222 /// instructions operate.
2223 ///
2224 /// For example, a simple 2x2 matrix can be transposed with:
2225 ///
2226 /// ; Original matrix
2227 /// m0 = < a, b >
2228 /// m1 = < c, d >
2229 ///
2230 /// ; Transposed matrix
2231 /// t0 = < a, c > = shufflevector m0, m1, < 0, 2 >
2232 /// t1 = < b, d > = shufflevector m0, m1, < 1, 3 >
2233 ///
2234 /// For matrices having greater than n columns, the resulting nx2 transposed
2235 /// matrix is stored in two result vectors such that one vector contains
2236 /// interleaved elements from all the even-numbered rows and the other vector
2237 /// contains interleaved elements from all the odd-numbered rows. For example,
2238 /// a 2x4 matrix can be transposed with:
2239 ///
2240 /// ; Original matrix
2241 /// m0 = < a, b, c, d >
2242 /// m1 = < e, f, g, h >
2243 ///
2244 /// ; Transposed matrix
2245 /// t0 = < a, e, c, g > = shufflevector m0, m1 < 0, 4, 2, 6 >
2246 /// t1 = < b, f, d, h > = shufflevector m0, m1 < 1, 5, 3, 7 >
2247 static bool isTransposeMask(ArrayRef<int> Mask);
2248 static bool isTransposeMask(const Constant *Mask) {
2249 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((Mask->getType()->isVectorTy() && "Shuffle needs vector constant."
) ? static_cast<void> (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 2249, __PRETTY_FUNCTION__))
;
2250 SmallVector<int, 16> MaskAsInts;
2251 getShuffleMask(Mask, MaskAsInts);
2252 return isTransposeMask(MaskAsInts);
2253 }
2254
2255 /// Return true if this shuffle transposes the elements of its inputs without
2256 /// changing the length of the vectors. This operation may also be known as a
2257 /// merge or interleave. See the description for isTransposeMask() for the
2258 /// exact specification.
2259 /// Example: shufflevector <4 x n> A, <4 x n> B, <0,4,2,6>
2260 bool isTranspose() const {
2261 return !changesLength() && isTransposeMask(ShuffleMask);
2262 }
2263
2264 /// Return true if this shuffle mask is an extract subvector mask.
2265 /// A valid extract subvector mask returns a smaller vector from a single
2266 /// source operand. The base extraction index is returned as well.
2267 static bool isExtractSubvectorMask(ArrayRef<int> Mask, int NumSrcElts,
2268 int &Index);
2269 static bool isExtractSubvectorMask(const Constant *Mask, int NumSrcElts,
2270 int &Index) {
2271 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((Mask->getType()->isVectorTy() && "Shuffle needs vector constant."
) ? static_cast<void> (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 2271, __PRETTY_FUNCTION__))
;
2272 // Not possible to express a shuffle mask for a scalable vector for this
2273 // case.
2274 if (isa<ScalableVectorType>(Mask->getType()))
2275 return false;
2276 SmallVector<int, 16> MaskAsInts;
2277 getShuffleMask(Mask, MaskAsInts);
2278 return isExtractSubvectorMask(MaskAsInts, NumSrcElts, Index);
2279 }
2280
2281 /// Return true if this shuffle mask is an extract subvector mask.
2282 bool isExtractSubvectorMask(int &Index) const {
2283 // Not possible to express a shuffle mask for a scalable vector for this
2284 // case.
2285 if (isa<ScalableVectorType>(getType()))
2286 return false;
2287
2288 int NumSrcElts =
2289 cast<FixedVectorType>(Op<0>()->getType())->getNumElements();
2290 return isExtractSubvectorMask(ShuffleMask, NumSrcElts, Index);
2291 }
2292
2293 /// Change values in a shuffle permute mask assuming the two vector operands
2294 /// of length InVecNumElts have swapped position.
2295 static void commuteShuffleMask(MutableArrayRef<int> Mask,
2296 unsigned InVecNumElts) {
2297 for (int &Idx : Mask) {
2298 if (Idx == -1)
2299 continue;
2300 Idx = Idx < (int)InVecNumElts ? Idx + InVecNumElts : Idx - InVecNumElts;
2301 assert(Idx >= 0 && Idx < (int)InVecNumElts * 2 &&((Idx >= 0 && Idx < (int)InVecNumElts * 2 &&
"shufflevector mask index out of range") ? static_cast<void
> (0) : __assert_fail ("Idx >= 0 && Idx < (int)InVecNumElts * 2 && \"shufflevector mask index out of range\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 2302, __PRETTY_FUNCTION__))
2302 "shufflevector mask index out of range")((Idx >= 0 && Idx < (int)InVecNumElts * 2 &&
"shufflevector mask index out of range") ? static_cast<void
> (0) : __assert_fail ("Idx >= 0 && Idx < (int)InVecNumElts * 2 && \"shufflevector mask index out of range\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 2302, __PRETTY_FUNCTION__))
;
2303 }
2304 }
2305
2306 // Methods for support type inquiry through isa, cast, and dyn_cast:
2307 static bool classof(const Instruction *I) {
2308 return I->getOpcode() == Instruction::ShuffleVector;
2309 }
2310 static bool classof(const Value *V) {
2311 return isa<Instruction>(V) && classof(cast<Instruction>(V));
2312 }
2313};
2314
2315template <>
2316struct OperandTraits<ShuffleVectorInst>
2317 : public FixedNumOperandTraits<ShuffleVectorInst, 2> {};
2318
2319DEFINE_TRANSPARENT_OPERAND_ACCESSORS(ShuffleVectorInst, Value)ShuffleVectorInst::op_iterator ShuffleVectorInst::op_begin() {
return OperandTraits<ShuffleVectorInst>::op_begin(this
); } ShuffleVectorInst::const_op_iterator ShuffleVectorInst::
op_begin() const { return OperandTraits<ShuffleVectorInst>
::op_begin(const_cast<ShuffleVectorInst*>(this)); } ShuffleVectorInst
::op_iterator ShuffleVectorInst::op_end() { return OperandTraits
<ShuffleVectorInst>::op_end(this); } ShuffleVectorInst::
const_op_iterator ShuffleVectorInst::op_end() const { return OperandTraits
<ShuffleVectorInst>::op_end(const_cast<ShuffleVectorInst
*>(this)); } Value *ShuffleVectorInst::getOperand(unsigned
i_nocapture) const { ((i_nocapture < OperandTraits<ShuffleVectorInst
>::operands(this) && "getOperand() out of range!")
? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<ShuffleVectorInst>::operands(this) && \"getOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 2319, __PRETTY_FUNCTION__)); return cast_or_null<Value>
( OperandTraits<ShuffleVectorInst>::op_begin(const_cast
<ShuffleVectorInst*>(this))[i_nocapture].get()); } void
ShuffleVectorInst::setOperand(unsigned i_nocapture, Value *Val_nocapture
) { ((i_nocapture < OperandTraits<ShuffleVectorInst>
::operands(this) && "setOperand() out of range!") ? static_cast
<void> (0) : __assert_fail ("i_nocapture < OperandTraits<ShuffleVectorInst>::operands(this) && \"setOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12~++20210115100614+a14c36fe27f5/llvm/include/llvm/IR/Instructions.h"
, 2319, __PRETTY_FUNCTION__)); OperandTraits<ShuffleVectorInst
>::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned
ShuffleVectorInst::getNumOperands() const { return OperandTraits
<ShuffleVectorInst>::operands(this); } template <int
Idx_nocapture> Use &ShuffleVectorInst::Op() { return this
->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture
> const Use &ShuffleVectorInst::Op() const { return this
->OpFrom<Idx_nocapture>(this); }
2320
2321//===----------------------------------------------------------------------===//
2322// ExtractValueInst Class
2323//===----------------------------------------------------------------------===//
2324
2325/// This instruction extracts a struct member or array
2326/// element value from an aggregate value.
2327///
2328class ExtractValueInst : public UnaryInstruction {
2329 SmallVector<unsigned, 4> Indices;
2330
2331 ExtractValueInst(const ExtractValueInst &EVI);
2332
2333 /// Constructors - Create a extractvalue instruction with a base aggregate
2334 /// value and a list of indices. The first ctor can optionally insert before
2335 /// an existing instruction, the second appends the new instruction to the
2336 /// specified BasicBlock.
2337 inline ExtractValueInst(Value *Agg,
2338 ArrayRef<unsigned> Idxs,
2339 const Twine &NameStr,
2340 Instruction *InsertBefore);
2341 inline ExtractValueInst(Value *Agg,
2342 ArrayRef<unsigned> Idxs,
2343 const Twine &NameStr, BasicBlock *InsertAtEnd);
2344
2345 void init(ArrayRef<unsigned> Idxs, const Twine &NameStr);
2346
2347protected:
2348 // Note: Instruction needs to be a friend here to call cloneImpl.
2349 friend class Instruction;
2350
2351 ExtractValueInst *cloneImpl() const;
2352
2353public:
2354 static ExtractValueInst *Create(Value *Agg,
2355 ArrayRef<unsigned> Idxs,
2356 const Twine &NameStr = "",
2357 Instruction *InsertBefore = nullptr) {
2358 return new
2359 ExtractValueInst(Agg, Idxs, NameStr, InsertBefore);
2360 }
2361
2362 static ExtractValueInst *Create(Value *Agg,
2363 ArrayRef<unsigned> Idxs,
2364 const Twine &NameStr,
2365 BasicBlock *InsertAtEnd) {
2366 return new ExtractValueInst(Agg, Idxs, NameStr, InsertAtEnd);
2367 }
2368
2369 /// Returns the type of the element that would be extracted
2370 /// with an extractvalue instruction with the specified parameters.
2371 ///
2372 /// Null is returned if the indices are invalid for the specified type.
2373 static Type *getIndexedType(Type *Agg, ArrayRef<unsigned> Idxs);
2374
2375 using idx_iterator = const unsigned*;
2376
2377 inline idx_iterator idx_begin() const { return Indices.begin(); }
2378 inline idx_iterator idx_end() const { return Indices.end(); }
2379 inline iterator_range<idx_iterator> indices() const {
2380 return make_range(idx_begin(), idx_end());
2381 }
2382
2383 Value *getAggregateOperand() {
2384 return getOperand(0);
2385 }
2386 const Value *getAggregateOperand() const {
2387 return getOperand(0);
2388 }
2389 static unsigned getAggregateOperandIndex() {
2390 return 0U; // get index for modifying correct operand
2391 }
2392
2393 ArrayRef<unsigned> getIndices() const {
2394 return Indices;
2395 }
2396
2397 unsigned getNumIndices() const {
2398 return (unsigned)Indices.size();
2399 }
2400
2401 bool hasIndices() const {
2402 return true;
2403 }
2404
2405 // Methods for support type inquiry through isa, cast, and dyn_cast:
2406 static bool classof(const Instruction *I) {
2407 return I->getOpcode() == Instruction::ExtractValue;
2408 }
2409 static bool classof(const Value *V) {
2410 return isa<Instruction>(V) && classof(cast<Instruction>(V));
2411 }
2412};
2413
2414ExtractValueInst::ExtractValueInst(Value *Agg,
2415 ArrayRef<unsigned> Idxs,
2416 const Twine &NameStr,
2417 Instruction *InsertBefore)
2418 : UnaryInstruction(checkGEPType(getIndexedType(Agg->getType(), Idxs)),
2419 ExtractValue, Agg, InsertBefore) {
2420 init(Idxs, NameStr);
2421}
2422
2423ExtractValueInst::ExtractValueInst(Value *Agg,
2424 ArrayRef<unsigned> Idxs,
2425 const Twine &NameStr,
2426 BasicBlock *InsertAtEnd)
2427 : UnaryInstruction(checkGEPType(getIndexedType(Agg->getType(), Idxs)),
2428 ExtractValue, Agg, InsertAtEnd) {
2429 init(Idxs, NameStr);
2430}
2431
2432//===----------------------------------------------------------------------===//
2433// InsertValueInst Class
2434//===----------------------------------------------------------------------===//
2435
2436/// This instruction inserts a struct field of array element
2437/// value into an aggregate value.
2438///
2439class InsertValueInst : public Instruction {
2440 SmallVector<unsigned, 4> Indices;
2441
2442 InsertValueInst(const InsertValueInst &IVI);
2443
2444 /// Constructors - Create a insertvalue instruction with a base aggregate
2445 /// value, a value to insert, and a list of indices. The first ctor can
2446 /// optionally insert before an existing instruction, the second appends
2447 /// the new instruction to the specified BasicBlock.
2448 inline InsertValueInst(Value *Agg, Value *Val,
2449 ArrayRef<unsigned> Idxs,
2450 const Twine &NameStr,
2451 Instruction *InsertBefore);
2452 inline InsertValueInst(Value *Agg, Value *Val,
2453 ArrayRef<unsigned> Idxs,
2454 const Twine &NameStr, BasicBlock *InsertAtEnd);
2455
2456 /// Constructors - These two constructors are convenience methods because one
2457 /// and two index insertvalue instructions are so common.
2458 InsertValueInst(Value *Agg, Value *Val, unsigned Idx,
2459 const Twine &NameStr = "",
2460 Instruction *InsertBefore = nullptr);
2461 InsertValueInst(Value *Agg, Value *Val, unsigned Idx, const Twine &NameStr,
2462 BasicBlock *InsertAtEnd);
2463
2464 void init(Value *Agg, Value *Val, ArrayRef<unsigned> Idxs,
2465 const Twine &NameStr);
2466
2467protected:
2468 // Note: Instruction needs to be a friend here to call cloneImpl.
2469 friend class Instruction;
2470
2471 InsertValueInst *cloneImpl() const;
2472
2473public:
2474 // allocate space for exactly two operands
2475 void *operator new(size_t s) {
2476 return User::operator new(s, 2);
2477 }
2478
2479 static InsertValueInst *Create(Value *Agg, Value *Val,
2480 ArrayRef<unsigned> Idxs,
2481 const Twine &NameStr = "",
2482 Instruction *InsertBefore = nullptr) {
2483 return new InsertValueInst(Agg, Val, Idxs, NameStr, InsertBefore);
2484 }
2485
2486 static InsertValueInst *Create(Value *Agg, Value *Val,
2487 ArrayRef<unsigned> Idxs,
2488 const Twine &NameStr,
2489 BasicBlock *InsertAtEnd) {
2490 return new InsertValueInst(Agg, Val, Idxs, NameStr, InsertAtEnd);
2491 }
2492
2493 /// Transparently provide more efficient getOperand methods.
2494 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() co