Bug Summary

File:llvm/lib/Transforms/Scalar/JumpThreading.cpp
Warning:line 1458, 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 -debugger-tuning=gdb -ffunction-sections -fdata-sections -fcoverage-compilation-dir=/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/build-llvm/lib/Transforms/Scalar -resource-dir /usr/lib/llvm-13/lib/clang/13.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/build-llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/build-llvm/include -I /build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/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/c++/6.3.0/backward -internal-isystem /usr/local/include -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../x86_64-linux-gnu/include -internal-isystem /usr/lib/llvm-13/lib/clang/13.0.0/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-13~++20210405022414+5f57793c4fe4/build-llvm/lib/Transforms/Scalar -fdebug-prefix-map=/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4=. -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 -D__GCC_HAVE_DWARF2_CFI_ASM=1 -o /tmp/scan-build-2021-04-05-202135-9119-1 -x c++ /build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp

/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/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/MemoryLocation.h"
35#include "llvm/Analysis/TargetLibraryInfo.h"
36#include "llvm/Analysis/TargetTransformInfo.h"
37#include "llvm/Analysis/ValueTracking.h"
38#include "llvm/IR/BasicBlock.h"
39#include "llvm/IR/CFG.h"
40#include "llvm/IR/Constant.h"
41#include "llvm/IR/ConstantRange.h"
42#include "llvm/IR/Constants.h"
43#include "llvm/IR/DataLayout.h"
44#include "llvm/IR/Dominators.h"
45#include "llvm/IR/Function.h"
46#include "llvm/IR/InstrTypes.h"
47#include "llvm/IR/Instruction.h"
48#include "llvm/IR/Instructions.h"
49#include "llvm/IR/IntrinsicInst.h"
50#include "llvm/IR/Intrinsics.h"
51#include "llvm/IR/LLVMContext.h"
52#include "llvm/IR/MDBuilder.h"
53#include "llvm/IR/Metadata.h"
54#include "llvm/IR/Module.h"
55#include "llvm/IR/PassManager.h"
56#include "llvm/IR/PatternMatch.h"
57#include "llvm/IR/Type.h"
58#include "llvm/IR/Use.h"
59#include "llvm/IR/User.h"
60#include "llvm/IR/Value.h"
61#include "llvm/InitializePasses.h"
62#include "llvm/Pass.h"
63#include "llvm/Support/BlockFrequency.h"
64#include "llvm/Support/BranchProbability.h"
65#include "llvm/Support/Casting.h"
66#include "llvm/Support/CommandLine.h"
67#include "llvm/Support/Debug.h"
68#include "llvm/Support/raw_ostream.h"
69#include "llvm/Transforms/Scalar.h"
70#include "llvm/Transforms/Utils/BasicBlockUtils.h"
71#include "llvm/Transforms/Utils/Cloning.h"
72#include "llvm/Transforms/Utils/Local.h"
73#include "llvm/Transforms/Utils/SSAUpdater.h"
74#include "llvm/Transforms/Utils/ValueMapper.h"
75#include <algorithm>
76#include <cassert>
77#include <cstddef>
78#include <cstdint>
79#include <iterator>
80#include <memory>
81#include <utility>
82
83using namespace llvm;
84using namespace jumpthreading;
85
86#define DEBUG_TYPE"jump-threading" "jump-threading"
87
88STATISTIC(NumThreads, "Number of jumps threaded")static llvm::Statistic NumThreads = {"jump-threading", "NumThreads"
, "Number of jumps threaded"}
;
89STATISTIC(NumFolds, "Number of terminators folded")static llvm::Statistic NumFolds = {"jump-threading", "NumFolds"
, "Number of terminators folded"}
;
90STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi")static llvm::Statistic NumDupes = {"jump-threading", "NumDupes"
, "Number of branch blocks duplicated to eliminate phi"}
;
91
92static cl::opt<unsigned>
93BBDuplicateThreshold("jump-threading-threshold",
94 cl::desc("Max block size to duplicate for jump threading"),
95 cl::init(6), cl::Hidden);
96
97static cl::opt<unsigned>
98ImplicationSearchThreshold(
99 "jump-threading-implication-search-threshold",
100 cl::desc("The number of predecessors to search for a stronger "
101 "condition to use to thread over a weaker condition"),
102 cl::init(3), cl::Hidden);
103
104static cl::opt<bool> PrintLVIAfterJumpThreading(
105 "print-lvi-after-jump-threading",
106 cl::desc("Print the LazyValueInfo cache after JumpThreading"), cl::init(false),
107 cl::Hidden);
108
109static cl::opt<bool> JumpThreadingFreezeSelectCond(
110 "jump-threading-freeze-select-cond",
111 cl::desc("Freeze the condition when unfolding select"), cl::init(false),
112 cl::Hidden);
113
114static cl::opt<bool> ThreadAcrossLoopHeaders(
115 "jump-threading-across-loop-headers",
116 cl::desc("Allow JumpThreading to thread across loop headers, for testing"),
117 cl::init(false), cl::Hidden);
118
119
120namespace {
121
122 /// This pass performs 'jump threading', which looks at blocks that have
123 /// multiple predecessors and multiple successors. If one or more of the
124 /// predecessors of the block can be proven to always jump to one of the
125 /// successors, we forward the edge from the predecessor to the successor by
126 /// duplicating the contents of this block.
127 ///
128 /// An example of when this can occur is code like this:
129 ///
130 /// if () { ...
131 /// X = 4;
132 /// }
133 /// if (X < 3) {
134 ///
135 /// In this case, the unconditional branch at the end of the first if can be
136 /// revectored to the false side of the second if.
137 class JumpThreading : public FunctionPass {
138 JumpThreadingPass Impl;
139
140 public:
141 static char ID; // Pass identification
142
143 JumpThreading(bool InsertFreezeWhenUnfoldingSelect = false, int T = -1)
144 : FunctionPass(ID), Impl(InsertFreezeWhenUnfoldingSelect, T) {
145 initializeJumpThreadingPass(*PassRegistry::getPassRegistry());
146 }
147
148 bool runOnFunction(Function &F) override;
149
150 void getAnalysisUsage(AnalysisUsage &AU) const override {
151 AU.addRequired<DominatorTreeWrapperPass>();
152 AU.addPreserved<DominatorTreeWrapperPass>();
153 AU.addRequired<AAResultsWrapperPass>();
154 AU.addRequired<LazyValueInfoWrapperPass>();
155 AU.addPreserved<LazyValueInfoWrapperPass>();
156 AU.addPreserved<GlobalsAAWrapperPass>();
157 AU.addRequired<TargetLibraryInfoWrapperPass>();
158 AU.addRequired<TargetTransformInfoWrapperPass>();
159 }
160
161 void releaseMemory() override { Impl.releaseMemory(); }
162 };
163
164} // end anonymous namespace
165
166char JumpThreading::ID = 0;
167
168INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",static void *initializeJumpThreadingPassOnce(PassRegistry &
Registry) {
169 "Jump Threading", false, false)static void *initializeJumpThreadingPassOnce(PassRegistry &
Registry) {
170INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry);
171INITIALIZE_PASS_DEPENDENCY(LazyValueInfoWrapperPass)initializeLazyValueInfoWrapperPassPass(Registry);
172INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry);
173INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry);
174INITIALIZE_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)); }
175 "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)); }
176
177// Public interface to the Jump Threading pass
178FunctionPass *llvm::createJumpThreadingPass(bool InsertFr, int Threshold) {
179 return new JumpThreading(InsertFr, Threshold);
180}
181
182JumpThreadingPass::JumpThreadingPass(bool InsertFr, int T) {
183 InsertFreezeWhenUnfoldingSelect = JumpThreadingFreezeSelectCond | InsertFr;
184 DefaultBBDupThreshold = (T == -1) ? BBDuplicateThreshold : unsigned(T);
185}
186
187// Update branch probability information according to conditional
188// branch probability. This is usually made possible for cloned branches
189// in inline instances by the context specific profile in the caller.
190// For instance,
191//
192// [Block PredBB]
193// [Branch PredBr]
194// if (t) {
195// Block A;
196// } else {
197// Block B;
198// }
199//
200// [Block BB]
201// cond = PN([true, %A], [..., %B]); // PHI node
202// [Branch CondBr]
203// if (cond) {
204// ... // P(cond == true) = 1%
205// }
206//
207// Here we know that when block A is taken, cond must be true, which means
208// P(cond == true | A) = 1
209//
210// Given that P(cond == true) = P(cond == true | A) * P(A) +
211// P(cond == true | B) * P(B)
212// we get:
213// P(cond == true ) = P(A) + P(cond == true | B) * P(B)
214//
215// which gives us:
216// P(A) is less than P(cond == true), i.e.
217// P(t == true) <= P(cond == true)
218//
219// In other words, if we know P(cond == true) is unlikely, we know
220// that P(t == true) is also unlikely.
221//
222static void updatePredecessorProfileMetadata(PHINode *PN, BasicBlock *BB) {
223 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
224 if (!CondBr)
225 return;
226
227 uint64_t TrueWeight, FalseWeight;
228 if (!CondBr->extractProfMetadata(TrueWeight, FalseWeight))
229 return;
230
231 if (TrueWeight + FalseWeight == 0)
232 // Zero branch_weights do not give a hint for getting branch probabilities.
233 // Technically it would result in division by zero denominator, which is
234 // TrueWeight + FalseWeight.
235 return;
236
237 // Returns the outgoing edge of the dominating predecessor block
238 // that leads to the PhiNode's incoming block:
239 auto GetPredOutEdge =
240 [](BasicBlock *IncomingBB,
241 BasicBlock *PhiBB) -> std::pair<BasicBlock *, BasicBlock *> {
242 auto *PredBB = IncomingBB;
243 auto *SuccBB = PhiBB;
244 SmallPtrSet<BasicBlock *, 16> Visited;
245 while (true) {
246 BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator());
247 if (PredBr && PredBr->isConditional())
248 return {PredBB, SuccBB};
249 Visited.insert(PredBB);
250 auto *SinglePredBB = PredBB->getSinglePredecessor();
251 if (!SinglePredBB)
252 return {nullptr, nullptr};
253
254 // Stop searching when SinglePredBB has been visited. It means we see
255 // an unreachable loop.
256 if (Visited.count(SinglePredBB))
257 return {nullptr, nullptr};
258
259 SuccBB = PredBB;
260 PredBB = SinglePredBB;
261 }
262 };
263
264 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
265 Value *PhiOpnd = PN->getIncomingValue(i);
266 ConstantInt *CI = dyn_cast<ConstantInt>(PhiOpnd);
267
268 if (!CI || !CI->getType()->isIntegerTy(1))
269 continue;
270
271 BranchProbability BP =
272 (CI->isOne() ? BranchProbability::getBranchProbability(
273 TrueWeight, TrueWeight + FalseWeight)
274 : BranchProbability::getBranchProbability(
275 FalseWeight, TrueWeight + FalseWeight));
276
277 auto PredOutEdge = GetPredOutEdge(PN->getIncomingBlock(i), BB);
278 if (!PredOutEdge.first)
279 return;
280
281 BasicBlock *PredBB = PredOutEdge.first;
282 BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator());
283 if (!PredBr)
284 return;
285
286 uint64_t PredTrueWeight, PredFalseWeight;
287 // FIXME: We currently only set the profile data when it is missing.
288 // With PGO, this can be used to refine even existing profile data with
289 // context information. This needs to be done after more performance
290 // testing.
291 if (PredBr->extractProfMetadata(PredTrueWeight, PredFalseWeight))
292 continue;
293
294 // We can not infer anything useful when BP >= 50%, because BP is the
295 // upper bound probability value.
296 if (BP >= BranchProbability(50, 100))
297 continue;
298
299 SmallVector<uint32_t, 2> Weights;
300 if (PredBr->getSuccessor(0) == PredOutEdge.second) {
301 Weights.push_back(BP.getNumerator());
302 Weights.push_back(BP.getCompl().getNumerator());
303 } else {
304 Weights.push_back(BP.getCompl().getNumerator());
305 Weights.push_back(BP.getNumerator());
306 }
307 PredBr->setMetadata(LLVMContext::MD_prof,
308 MDBuilder(PredBr->getParent()->getContext())
309 .createBranchWeights(Weights));
310 }
311}
312
313/// runOnFunction - Toplevel algorithm.
314bool JumpThreading::runOnFunction(Function &F) {
315 if (skipFunction(F))
316 return false;
317 auto TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
318 // Jump Threading has no sense for the targets with divergent CF
319 if (TTI->hasBranchDivergence())
320 return false;
321 auto TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
322 auto DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
323 auto LVI = &getAnalysis<LazyValueInfoWrapperPass>().getLVI();
324 auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
325 DomTreeUpdater DTU(*DT, DomTreeUpdater::UpdateStrategy::Lazy);
326 std::unique_ptr<BlockFrequencyInfo> BFI;
327 std::unique_ptr<BranchProbabilityInfo> BPI;
328 if (F.hasProfileData()) {
329 LoopInfo LI{DominatorTree(F)};
330 BPI.reset(new BranchProbabilityInfo(F, LI, TLI));
331 BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
332 }
333
334 bool Changed = Impl.runImpl(F, TLI, LVI, AA, &DTU, F.hasProfileData(),
335 std::move(BFI), std::move(BPI));
336 if (PrintLVIAfterJumpThreading) {
337 dbgs() << "LVI for function '" << F.getName() << "':\n";
338 LVI->printLVI(F, DTU.getDomTree(), dbgs());
339 }
340 return Changed;
341}
342
343PreservedAnalyses JumpThreadingPass::run(Function &F,
344 FunctionAnalysisManager &AM) {
345 auto &TTI = AM.getResult<TargetIRAnalysis>(F);
346 // Jump Threading has no sense for the targets with divergent CF
347 if (TTI.hasBranchDivergence())
348 return PreservedAnalyses::all();
349 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
350 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
351 auto &LVI = AM.getResult<LazyValueAnalysis>(F);
352 auto &AA = AM.getResult<AAManager>(F);
353 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Lazy);
354
355 std::unique_ptr<BlockFrequencyInfo> BFI;
356 std::unique_ptr<BranchProbabilityInfo> BPI;
357 if (F.hasProfileData()) {
358 LoopInfo LI{DominatorTree(F)};
359 BPI.reset(new BranchProbabilityInfo(F, LI, &TLI));
360 BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
361 }
362
363 bool Changed = runImpl(F, &TLI, &LVI, &AA, &DTU, F.hasProfileData(),
364 std::move(BFI), std::move(BPI));
365
366 if (PrintLVIAfterJumpThreading) {
367 dbgs() << "LVI for function '" << F.getName() << "':\n";
368 LVI.printLVI(F, DTU.getDomTree(), dbgs());
369 }
370
371 if (!Changed)
372 return PreservedAnalyses::all();
373 PreservedAnalyses PA;
374 PA.preserve<GlobalsAA>();
375 PA.preserve<DominatorTreeAnalysis>();
376 PA.preserve<LazyValueAnalysis>();
377 return PA;
378}
379
380bool JumpThreadingPass::runImpl(Function &F, TargetLibraryInfo *TLI_,
381 LazyValueInfo *LVI_, AliasAnalysis *AA_,
382 DomTreeUpdater *DTU_, bool HasProfileData_,
383 std::unique_ptr<BlockFrequencyInfo> BFI_,
384 std::unique_ptr<BranchProbabilityInfo> BPI_) {
385 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)
;
386 TLI = TLI_;
387 LVI = LVI_;
388 AA = AA_;
389 DTU = DTU_;
390 BFI.reset();
391 BPI.reset();
392 // When profile data is available, we need to update edge weights after
393 // successful jump threading, which requires both BPI and BFI being available.
394 HasProfileData = HasProfileData_;
395 auto *GuardDecl = F.getParent()->getFunction(
396 Intrinsic::getName(Intrinsic::experimental_guard));
397 HasGuards = GuardDecl && !GuardDecl->use_empty();
398 if (HasProfileData) {
399 BPI = std::move(BPI_);
400 BFI = std::move(BFI_);
401 }
402
403 // Reduce the number of instructions duplicated when optimizing strictly for
404 // size.
405 if (BBDuplicateThreshold.getNumOccurrences())
406 BBDupThreshold = BBDuplicateThreshold;
407 else if (F.hasFnAttribute(Attribute::MinSize))
408 BBDupThreshold = 3;
409 else
410 BBDupThreshold = DefaultBBDupThreshold;
411
412 // JumpThreading must not processes blocks unreachable from entry. It's a
413 // waste of compute time and can potentially lead to hangs.
414 SmallPtrSet<BasicBlock *, 16> Unreachable;
415 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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 415, __PRETTY_FUNCTION__))
;
416 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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 416, __PRETTY_FUNCTION__))
;
417 DominatorTree &DT = DTU->getDomTree();
418 for (auto &BB : F)
419 if (!DT.isReachableFromEntry(&BB))
420 Unreachable.insert(&BB);
421
422 if (!ThreadAcrossLoopHeaders)
423 findLoopHeaders(F);
424
425 bool EverChanged = false;
426 bool Changed;
427 do {
428 Changed = false;
429 for (auto &BB : F) {
430 if (Unreachable.count(&BB))
431 continue;
432 while (processBlock(&BB)) // Thread all of the branches we can over BB.
433 Changed = true;
434
435 // Jump threading may have introduced redundant debug values into BB
436 // which should be removed.
437 // Remove redundant pseudo probes as well.
438 if (Changed)
439 RemoveRedundantDbgInstrs(&BB, true);
440
441 // Stop processing BB if it's the entry or is now deleted. The following
442 // routines attempt to eliminate BB and locating a suitable replacement
443 // for the entry is non-trivial.
444 if (&BB == &F.getEntryBlock() || DTU->isBBPendingDeletion(&BB))
445 continue;
446
447 if (pred_empty(&BB)) {
448 // When processBlock makes BB unreachable it doesn't bother to fix up
449 // the instructions in it. We must remove BB to prevent invalid IR.
450 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)
451 << "' 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)
452 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " JT: Deleting dead block '"
<< BB.getName() << "' with terminator: " <<
*BB.getTerminator() << '\n'; } } while (false)
;
453 LoopHeaders.erase(&BB);
454 LVI->eraseBlock(&BB);
455 DeleteDeadBlock(&BB, DTU);
456 Changed = true;
457 continue;
458 }
459
460 // processBlock doesn't thread BBs with unconditional TIs. However, if BB
461 // is "almost empty", we attempt to merge BB with its sole successor.
462 auto *BI = dyn_cast<BranchInst>(BB.getTerminator());
463 if (BI && BI->isUnconditional()) {
464 BasicBlock *Succ = BI->getSuccessor(0);
465 if (
466 // The terminator must be the only non-phi instruction in BB.
467 BB.getFirstNonPHIOrDbg(true)->isTerminator() &&
468 // Don't alter Loop headers and latches to ensure another pass can
469 // detect and transform nested loops later.
470 !LoopHeaders.count(&BB) && !LoopHeaders.count(Succ) &&
471 TryToSimplifyUncondBranchFromEmptyBlock(&BB, DTU)) {
472 RemoveRedundantDbgInstrs(Succ, true);
473 // BB is valid for cleanup here because we passed in DTU. F remains
474 // BB's parent until a DTU->getDomTree() event.
475 LVI->eraseBlock(&BB);
476 Changed = true;
477 }
478 }
479 }
480 EverChanged |= Changed;
481 } while (Changed);
482
483 LoopHeaders.clear();
484 return EverChanged;
485}
486
487// Replace uses of Cond with ToVal when safe to do so. If all uses are
488// replaced, we can remove Cond. We cannot blindly replace all uses of Cond
489// because we may incorrectly replace uses when guards/assumes are uses of
490// of `Cond` and we used the guards/assume to reason about the `Cond` value
491// at the end of block. RAUW unconditionally replaces all uses
492// including the guards/assumes themselves and the uses before the
493// guard/assume.
494static void replaceFoldableUses(Instruction *Cond, Value *ToVal) {
495 assert(Cond->getType() == ToVal->getType())((Cond->getType() == ToVal->getType()) ? static_cast<
void> (0) : __assert_fail ("Cond->getType() == ToVal->getType()"
, "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 495, __PRETTY_FUNCTION__))
;
496 auto *BB = Cond->getParent();
497 // We can unconditionally replace all uses in non-local blocks (i.e. uses
498 // strictly dominated by BB), since LVI information is true from the
499 // terminator of BB.
500 replaceNonLocalUsesWith(Cond, ToVal);
501 for (Instruction &I : reverse(*BB)) {
502 // Reached the Cond whose uses we are trying to replace, so there are no
503 // more uses.
504 if (&I == Cond)
505 break;
506 // We only replace uses in instructions that are guaranteed to reach the end
507 // of BB, where we know Cond is ToVal.
508 if (!isGuaranteedToTransferExecutionToSuccessor(&I))
509 break;
510 I.replaceUsesOfWith(Cond, ToVal);
511 }
512 if (Cond->use_empty() && !Cond->mayHaveSideEffects())
513 Cond->eraseFromParent();
514}
515
516/// Return the cost of duplicating a piece of this block from first non-phi
517/// and before StopAt instruction to thread across it. Stop scanning the block
518/// when exceeding the threshold. If duplication is impossible, returns ~0U.
519static unsigned getJumpThreadDuplicationCost(BasicBlock *BB,
520 Instruction *StopAt,
521 unsigned Threshold) {
522 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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 522, __PRETTY_FUNCTION__))
;
523 /// Ignore PHI nodes, these will be flattened when duplication happens.
524 BasicBlock::const_iterator I(BB->getFirstNonPHI());
525
526 // FIXME: THREADING will delete values that are just used to compute the
527 // branch, so they shouldn't count against the duplication cost.
528
529 unsigned Bonus = 0;
530 if (BB->getTerminator() == StopAt) {
531 // Threading through a switch statement is particularly profitable. If this
532 // block ends in a switch, decrease its cost to make it more likely to
533 // happen.
534 if (isa<SwitchInst>(StopAt))
535 Bonus = 6;
536
537 // The same holds for indirect branches, but slightly more so.
538 if (isa<IndirectBrInst>(StopAt))
539 Bonus = 8;
540 }
541
542 // Bump the threshold up so the early exit from the loop doesn't skip the
543 // terminator-based Size adjustment at the end.
544 Threshold += Bonus;
545
546 // Sum up the cost of each instruction until we get to the terminator. Don't
547 // include the terminator because the copy won't include it.
548 unsigned Size = 0;
549 for (; &*I != StopAt; ++I) {
550
551 // Stop scanning the block if we've reached the threshold.
552 if (Size > Threshold)
553 return Size;
554
555 // Debugger intrinsics don't incur code size.
556 if (isa<DbgInfoIntrinsic>(I)) continue;
557
558 // Pseudo-probes don't incur code size.
559 if (isa<PseudoProbeInst>(I))
560 continue;
561
562 // If this is a pointer->pointer bitcast, it is free.
563 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
564 continue;
565
566 // Freeze instruction is free, too.
567 if (isa<FreezeInst>(I))
568 continue;
569
570 // Bail out if this instruction gives back a token type, it is not possible
571 // to duplicate it if it is used outside this BB.
572 if (I->getType()->isTokenTy() && I->isUsedOutsideOfBlock(BB))
573 return ~0U;
574
575 // All other instructions count for at least one unit.
576 ++Size;
577
578 // Calls are more expensive. If they are non-intrinsic calls, we model them
579 // as having cost of 4. If they are a non-vector intrinsic, we model them
580 // as having cost of 2 total, and if they are a vector intrinsic, we model
581 // them as having cost 1.
582 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
583 if (CI->cannotDuplicate() || CI->isConvergent())
584 // Blocks with NoDuplicate are modelled as having infinite cost, so they
585 // are never duplicated.
586 return ~0U;
587 else if (!isa<IntrinsicInst>(CI))
588 Size += 3;
589 else if (!CI->getType()->isVectorTy())
590 Size += 1;
591 }
592 }
593
594 return Size > Bonus ? Size - Bonus : 0;
595}
596
597/// findLoopHeaders - We do not want jump threading to turn proper loop
598/// structures into irreducible loops. Doing this breaks up the loop nesting
599/// hierarchy and pessimizes later transformations. To prevent this from
600/// happening, we first have to find the loop headers. Here we approximate this
601/// by finding targets of backedges in the CFG.
602///
603/// Note that there definitely are cases when we want to allow threading of
604/// edges across a loop header. For example, threading a jump from outside the
605/// loop (the preheader) to an exit block of the loop is definitely profitable.
606/// It is also almost always profitable to thread backedges from within the loop
607/// to exit blocks, and is often profitable to thread backedges to other blocks
608/// within the loop (forming a nested loop). This simple analysis is not rich
609/// enough to track all of these properties and keep it up-to-date as the CFG
610/// mutates, so we don't allow any of these transformations.
611void JumpThreadingPass::findLoopHeaders(Function &F) {
612 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
613 FindFunctionBackedges(F, Edges);
614
615 for (const auto &Edge : Edges)
616 LoopHeaders.insert(Edge.second);
617}
618
619/// getKnownConstant - Helper method to determine if we can thread over a
620/// terminator with the given value as its condition, and if so what value to
621/// use for that. What kind of value this is depends on whether we want an
622/// integer or a block address, but an undef is always accepted.
623/// Returns null if Val is null or not an appropriate constant.
624static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
625 if (!Val)
626 return nullptr;
627
628 // Undef is "known" enough.
629 if (UndefValue *U = dyn_cast<UndefValue>(Val))
630 return U;
631
632 if (Preference == WantBlockAddress)
633 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
634
635 return dyn_cast<ConstantInt>(Val);
636}
637
638/// computeValueKnownInPredecessors - Given a basic block BB and a value V, see
639/// if we can infer that the value is a known ConstantInt/BlockAddress or undef
640/// in any of our predecessors. If so, return the known list of value and pred
641/// BB in the result vector.
642///
643/// This returns true if there were any known values.
644bool JumpThreadingPass::computeValueKnownInPredecessorsImpl(
645 Value *V, BasicBlock *BB, PredValueInfo &Result,
646 ConstantPreference Preference, DenseSet<Value *> &RecursionSet,
647 Instruction *CxtI) {
648 // This method walks up use-def chains recursively. Because of this, we could
649 // get into an infinite loop going around loops in the use-def chain. To
650 // prevent this, keep track of what (value, block) pairs we've already visited
651 // and terminate the search if we loop back to them
652 if (!RecursionSet.insert(V).second)
653 return false;
654
655 // If V is a constant, then it is known in all predecessors.
656 if (Constant *KC = getKnownConstant(V, Preference)) {
657 for (BasicBlock *Pred : predecessors(BB))
658 Result.emplace_back(KC, Pred);
659
660 return !Result.empty();
661 }
662
663 // If V is a non-instruction value, or an instruction in a different block,
664 // then it can't be derived from a PHI.
665 Instruction *I = dyn_cast<Instruction>(V);
666 if (!I || I->getParent() != BB) {
667
668 // Okay, if this is a live-in value, see if it has a known value at the end
669 // of any of our predecessors.
670 //
671 // FIXME: This should be an edge property, not a block end property.
672 /// TODO: Per PR2563, we could infer value range information about a
673 /// predecessor based on its terminator.
674 //
675 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
676 // "I" is a non-local compare-with-a-constant instruction. This would be
677 // able to handle value inequalities better, for example if the compare is
678 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
679 // Perhaps getConstantOnEdge should be smart enough to do this?
680 for (BasicBlock *P : predecessors(BB)) {
681 // If the value is known by LazyValueInfo to be a constant in a
682 // predecessor, use that information to try to thread this block.
683 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB, CxtI);
684 if (Constant *KC = getKnownConstant(PredCst, Preference))
685 Result.emplace_back(KC, P);
686 }
687
688 return !Result.empty();
689 }
690
691 /// If I is a PHI node, then we know the incoming values for any constants.
692 if (PHINode *PN = dyn_cast<PHINode>(I)) {
693 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
694 Value *InVal = PN->getIncomingValue(i);
695 if (Constant *KC = getKnownConstant(InVal, Preference)) {
696 Result.emplace_back(KC, PN->getIncomingBlock(i));
697 } else {
698 Constant *CI = LVI->getConstantOnEdge(InVal,
699 PN->getIncomingBlock(i),
700 BB, CxtI);
701 if (Constant *KC = getKnownConstant(CI, Preference))
702 Result.emplace_back(KC, PN->getIncomingBlock(i));
703 }
704 }
705
706 return !Result.empty();
707 }
708
709 // Handle Cast instructions.
710 if (CastInst *CI = dyn_cast<CastInst>(I)) {
711 Value *Source = CI->getOperand(0);
712 computeValueKnownInPredecessorsImpl(Source, BB, Result, Preference,
713 RecursionSet, CxtI);
714 if (Result.empty())
715 return false;
716
717 // Convert the known values.
718 for (auto &R : Result)
719 R.first = ConstantExpr::getCast(CI->getOpcode(), R.first, CI->getType());
720
721 return true;
722 }
723
724 if (FreezeInst *FI = dyn_cast<FreezeInst>(I)) {
725 Value *Source = FI->getOperand(0);
726 computeValueKnownInPredecessorsImpl(Source, BB, Result, Preference,
727 RecursionSet, CxtI);
728
729 erase_if(Result, [](auto &Pair) {
730 return !isGuaranteedNotToBeUndefOrPoison(Pair.first);
731 });
732
733 return !Result.empty();
734 }
735
736 // Handle some boolean conditions.
737 if (I->getType()->getPrimitiveSizeInBits() == 1) {
738 using namespace PatternMatch;
739
740 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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 740, __PRETTY_FUNCTION__))
;
741 // X | true -> true
742 // X & false -> false
743 Value *Op0, *Op1;
744 if (match(I, m_LogicalOr(m_Value(Op0), m_Value(Op1))) ||
745 match(I, m_LogicalAnd(m_Value(Op0), m_Value(Op1)))) {
746 PredValueInfoTy LHSVals, RHSVals;
747
748 computeValueKnownInPredecessorsImpl(Op0, BB, LHSVals, WantInteger,
749 RecursionSet, CxtI);
750 computeValueKnownInPredecessorsImpl(Op1, BB, RHSVals, WantInteger,
751 RecursionSet, CxtI);
752
753 if (LHSVals.empty() && RHSVals.empty())
754 return false;
755
756 ConstantInt *InterestingVal;
757 if (match(I, m_LogicalOr()))
758 InterestingVal = ConstantInt::getTrue(I->getContext());
759 else
760 InterestingVal = ConstantInt::getFalse(I->getContext());
761
762 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
763
764 // Scan for the sentinel. If we find an undef, force it to the
765 // interesting value: x|undef -> true and x&undef -> false.
766 for (const auto &LHSVal : LHSVals)
767 if (LHSVal.first == InterestingVal || isa<UndefValue>(LHSVal.first)) {
768 Result.emplace_back(InterestingVal, LHSVal.second);
769 LHSKnownBBs.insert(LHSVal.second);
770 }
771 for (const auto &RHSVal : RHSVals)
772 if (RHSVal.first == InterestingVal || isa<UndefValue>(RHSVal.first)) {
773 // If we already inferred a value for this block on the LHS, don't
774 // re-add it.
775 if (!LHSKnownBBs.count(RHSVal.second))
776 Result.emplace_back(InterestingVal, RHSVal.second);
777 }
778
779 return !Result.empty();
780 }
781
782 // Handle the NOT form of XOR.
783 if (I->getOpcode() == Instruction::Xor &&
784 isa<ConstantInt>(I->getOperand(1)) &&
785 cast<ConstantInt>(I->getOperand(1))->isOne()) {
786 computeValueKnownInPredecessorsImpl(I->getOperand(0), BB, Result,
787 WantInteger, RecursionSet, CxtI);
788 if (Result.empty())
789 return false;
790
791 // Invert the known values.
792 for (auto &R : Result)
793 R.first = ConstantExpr::getNot(R.first);
794
795 return true;
796 }
797
798 // Try to simplify some other binary operator values.
799 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
800 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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 801, __PRETTY_FUNCTION__))
801 && "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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 801, __PRETTY_FUNCTION__))
;
802 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
803 PredValueInfoTy LHSVals;
804 computeValueKnownInPredecessorsImpl(BO->getOperand(0), BB, LHSVals,
805 WantInteger, RecursionSet, CxtI);
806
807 // Try to use constant folding to simplify the binary operator.
808 for (const auto &LHSVal : LHSVals) {
809 Constant *V = LHSVal.first;
810 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
811
812 if (Constant *KC = getKnownConstant(Folded, WantInteger))
813 Result.emplace_back(KC, LHSVal.second);
814 }
815 }
816
817 return !Result.empty();
818 }
819
820 // Handle compare with phi operand, where the PHI is defined in this block.
821 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
822 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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 822, __PRETTY_FUNCTION__))
;
823 Type *CmpType = Cmp->getType();
824 Value *CmpLHS = Cmp->getOperand(0);
825 Value *CmpRHS = Cmp->getOperand(1);
826 CmpInst::Predicate Pred = Cmp->getPredicate();
827
828 PHINode *PN = dyn_cast<PHINode>(CmpLHS);
829 if (!PN)
830 PN = dyn_cast<PHINode>(CmpRHS);
831 if (PN && PN->getParent() == BB) {
832 const DataLayout &DL = PN->getModule()->getDataLayout();
833 // We can do this simplification if any comparisons fold to true or false.
834 // See if any do.
835 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
836 BasicBlock *PredBB = PN->getIncomingBlock(i);
837 Value *LHS, *RHS;
838 if (PN == CmpLHS) {
839 LHS = PN->getIncomingValue(i);
840 RHS = CmpRHS->DoPHITranslation(BB, PredBB);
841 } else {
842 LHS = CmpLHS->DoPHITranslation(BB, PredBB);
843 RHS = PN->getIncomingValue(i);
844 }
845 Value *Res = SimplifyCmpInst(Pred, LHS, RHS, {DL});
846 if (!Res) {
847 if (!isa<Constant>(RHS))
848 continue;
849
850 // getPredicateOnEdge call will make no sense if LHS is defined in BB.
851 auto LHSInst = dyn_cast<Instruction>(LHS);
852 if (LHSInst && LHSInst->getParent() == BB)
853 continue;
854
855 LazyValueInfo::Tristate
856 ResT = LVI->getPredicateOnEdge(Pred, LHS,
857 cast<Constant>(RHS), PredBB, BB,
858 CxtI ? CxtI : Cmp);
859 if (ResT == LazyValueInfo::Unknown)
860 continue;
861 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
862 }
863
864 if (Constant *KC = getKnownConstant(Res, WantInteger))
865 Result.emplace_back(KC, PredBB);
866 }
867
868 return !Result.empty();
869 }
870
871 // If comparing a live-in value against a constant, see if we know the
872 // live-in value on any predecessors.
873 if (isa<Constant>(CmpRHS) && !CmpType->isVectorTy()) {
874 Constant *CmpConst = cast<Constant>(CmpRHS);
875
876 if (!isa<Instruction>(CmpLHS) ||
877 cast<Instruction>(CmpLHS)->getParent() != BB) {
878 for (BasicBlock *P : predecessors(BB)) {
879 // If the value is known by LazyValueInfo to be a constant in a
880 // predecessor, use that information to try to thread this block.
881 LazyValueInfo::Tristate Res =
882 LVI->getPredicateOnEdge(Pred, CmpLHS,
883 CmpConst, P, BB, CxtI ? CxtI : Cmp);
884 if (Res == LazyValueInfo::Unknown)
885 continue;
886
887 Constant *ResC = ConstantInt::get(CmpType, Res);
888 Result.emplace_back(ResC, P);
889 }
890
891 return !Result.empty();
892 }
893
894 // InstCombine can fold some forms of constant range checks into
895 // (icmp (add (x, C1)), C2). See if we have we have such a thing with
896 // x as a live-in.
897 {
898 using namespace PatternMatch;
899
900 Value *AddLHS;
901 ConstantInt *AddConst;
902 if (isa<ConstantInt>(CmpConst) &&
903 match(CmpLHS, m_Add(m_Value(AddLHS), m_ConstantInt(AddConst)))) {
904 if (!isa<Instruction>(AddLHS) ||
905 cast<Instruction>(AddLHS)->getParent() != BB) {
906 for (BasicBlock *P : predecessors(BB)) {
907 // If the value is known by LazyValueInfo to be a ConstantRange in
908 // a predecessor, use that information to try to thread this
909 // block.
910 ConstantRange CR = LVI->getConstantRangeOnEdge(
911 AddLHS, P, BB, CxtI ? CxtI : cast<Instruction>(CmpLHS));
912 // Propagate the range through the addition.
913 CR = CR.add(AddConst->getValue());
914
915 // Get the range where the compare returns true.
916 ConstantRange CmpRange = ConstantRange::makeExactICmpRegion(
917 Pred, cast<ConstantInt>(CmpConst)->getValue());
918
919 Constant *ResC;
920 if (CmpRange.contains(CR))
921 ResC = ConstantInt::getTrue(CmpType);
922 else if (CmpRange.inverse().contains(CR))
923 ResC = ConstantInt::getFalse(CmpType);
924 else
925 continue;
926
927 Result.emplace_back(ResC, P);
928 }
929
930 return !Result.empty();
931 }
932 }
933 }
934
935 // Try to find a constant value for the LHS of a comparison,
936 // and evaluate it statically if we can.
937 PredValueInfoTy LHSVals;
938 computeValueKnownInPredecessorsImpl(I->getOperand(0), BB, LHSVals,
939 WantInteger, RecursionSet, CxtI);
940
941 for (const auto &LHSVal : LHSVals) {
942 Constant *V = LHSVal.first;
943 Constant *Folded = ConstantExpr::getCompare(Pred, V, CmpConst);
944 if (Constant *KC = getKnownConstant(Folded, WantInteger))
945 Result.emplace_back(KC, LHSVal.second);
946 }
947
948 return !Result.empty();
949 }
950 }
951
952 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
953 // Handle select instructions where at least one operand is a known constant
954 // and we can figure out the condition value for any predecessor block.
955 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
956 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
957 PredValueInfoTy Conds;
958 if ((TrueVal || FalseVal) &&
959 computeValueKnownInPredecessorsImpl(SI->getCondition(), BB, Conds,
960 WantInteger, RecursionSet, CxtI)) {
961 for (auto &C : Conds) {
962 Constant *Cond = C.first;
963
964 // Figure out what value to use for the condition.
965 bool KnownCond;
966 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
967 // A known boolean.
968 KnownCond = CI->isOne();
969 } else {
970 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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 970, __PRETTY_FUNCTION__))
;
971 // Either operand will do, so be sure to pick the one that's a known
972 // constant.
973 // FIXME: Do this more cleverly if both values are known constants?
974 KnownCond = (TrueVal != nullptr);
975 }
976
977 // See if the select has a known constant value for this predecessor.
978 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
979 Result.emplace_back(Val, C.second);
980 }
981
982 return !Result.empty();
983 }
984 }
985
986 // If all else fails, see if LVI can figure out a constant value for us.
987 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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 987, __PRETTY_FUNCTION__))
;
988 Constant *CI = LVI->getConstant(V, CxtI);
989 if (Constant *KC = getKnownConstant(CI, Preference)) {
990 for (BasicBlock *Pred : predecessors(BB))
991 Result.emplace_back(KC, Pred);
992 }
993
994 return !Result.empty();
995}
996
997/// GetBestDestForBranchOnUndef - If we determine that the specified block ends
998/// in an undefined jump, decide which block is best to revector to.
999///
1000/// Since we can pick an arbitrary destination, we pick the successor with the
1001/// fewest predecessors. This should reduce the in-degree of the others.
1002static unsigned getBestDestForJumpOnUndef(BasicBlock *BB) {
1003 Instruction *BBTerm = BB->getTerminator();
1004 unsigned MinSucc = 0;
1005 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
1006 // Compute the successor with the minimum number of predecessors.
1007 unsigned MinNumPreds = pred_size(TestBB);
1008 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
1009 TestBB = BBTerm->getSuccessor(i);
1010 unsigned NumPreds = pred_size(TestBB);
1011 if (NumPreds < MinNumPreds) {
1012 MinSucc = i;
1013 MinNumPreds = NumPreds;
1014 }
1015 }
1016
1017 return MinSucc;
1018}
1019
1020static bool hasAddressTakenAndUsed(BasicBlock *BB) {
1021 if (!BB->hasAddressTaken()) return false;
1022
1023 // If the block has its address taken, it may be a tree of dead constants
1024 // hanging off of it. These shouldn't keep the block alive.
1025 BlockAddress *BA = BlockAddress::get(BB);
1026 BA->removeDeadConstantUsers();
1027 return !BA->use_empty();
1028}
1029
1030/// processBlock - If there are any predecessors whose control can be threaded
1031/// through to a successor, transform them now.
1032bool JumpThreadingPass::processBlock(BasicBlock *BB) {
1033 // If the block is trivially dead, just return and let the caller nuke it.
1034 // This simplifies other transformations.
1035 if (DTU->isBBPendingDeletion(BB) ||
1036 (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()))
1037 return false;
1038
1039 // If this block has a single predecessor, and if that pred has a single
1040 // successor, merge the blocks. This encourages recursive jump threading
1041 // because now the condition in this block can be threaded through
1042 // predecessors of our predecessor block.
1043 if (maybeMergeBasicBlockIntoOnlyPred(BB))
1044 return true;
1045
1046 if (tryToUnfoldSelectInCurrBB(BB))
1047 return true;
1048
1049 // Look if we can propagate guards to predecessors.
1050 if (HasGuards && processGuards(BB))
1051 return true;
1052
1053 // What kind of constant we're looking for.
1054 ConstantPreference Preference = WantInteger;
1055
1056 // Look to see if the terminator is a conditional branch, switch or indirect
1057 // branch, if not we can't thread it.
1058 Value *Condition;
1059 Instruction *Terminator = BB->getTerminator();
1060 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
1061 // Can't thread an unconditional jump.
1062 if (BI->isUnconditional()) return false;
1063 Condition = BI->getCondition();
1064 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
1065 Condition = SI->getCondition();
1066 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
1067 // Can't thread indirect branch with no successors.
1068 if (IB->getNumSuccessors() == 0) return false;
1069 Condition = IB->getAddress()->stripPointerCasts();
1070 Preference = WantBlockAddress;
1071 } else {
1072 return false; // Must be an invoke or callbr.
1073 }
1074
1075 // Keep track if we constant folded the condition in this invocation.
1076 bool ConstantFolded = false;
1077
1078 // Run constant folding to see if we can reduce the condition to a simple
1079 // constant.
1080 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
1081 Value *SimpleVal =
1082 ConstantFoldInstruction(I, BB->getModule()->getDataLayout(), TLI);
1083 if (SimpleVal) {
1084 I->replaceAllUsesWith(SimpleVal);
1085 if (isInstructionTriviallyDead(I, TLI))
1086 I->eraseFromParent();
1087 Condition = SimpleVal;
1088 ConstantFolded = true;
1089 }
1090 }
1091
1092 // If the terminator is branching on an undef or freeze undef, we can pick any
1093 // of the successors to branch to. Let getBestDestForJumpOnUndef decide.
1094 auto *FI = dyn_cast<FreezeInst>(Condition);
1095 if (isa<UndefValue>(Condition) ||
1096 (FI && isa<UndefValue>(FI->getOperand(0)) && FI->hasOneUse())) {
1097 unsigned BestSucc = getBestDestForJumpOnUndef(BB);
1098 std::vector<DominatorTree::UpdateType> Updates;
1099
1100 // Fold the branch/switch.
1101 Instruction *BBTerm = BB->getTerminator();
1102 Updates.reserve(BBTerm->getNumSuccessors());
1103 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
1104 if (i == BestSucc) continue;
1105 BasicBlock *Succ = BBTerm->getSuccessor(i);
1106 Succ->removePredecessor(BB, true);
1107 Updates.push_back({DominatorTree::Delete, BB, Succ});
1108 }
1109
1110 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)
1111 << "' 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)
;
1112 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
1113 BBTerm->eraseFromParent();
1114 DTU->applyUpdatesPermissive(Updates);
1115 if (FI)
1116 FI->eraseFromParent();
1117 return true;
1118 }
1119
1120 // If the terminator of this block is branching on a constant, simplify the
1121 // terminator to an unconditional branch. This can occur due to threading in
1122 // other blocks.
1123 if (getKnownConstant(Condition, Preference)) {
1124 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)
1125 << "' folding terminator: " << *BB->getTerminator()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " In block '" <<
BB->getName() << "' folding terminator: " << *
BB->getTerminator() << '\n'; } } while (false)
1126 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " In block '" <<
BB->getName() << "' folding terminator: " << *
BB->getTerminator() << '\n'; } } while (false)
;
1127 ++NumFolds;
1128 ConstantFoldTerminator(BB, true, nullptr, DTU);
1129 if (HasProfileData)
1130 BPI->eraseBlock(BB);
1131 return true;
1132 }
1133
1134 Instruction *CondInst = dyn_cast<Instruction>(Condition);
1135
1136 // All the rest of our checks depend on the condition being an instruction.
1137 if (!CondInst) {
1138 // FIXME: Unify this with code below.
1139 if (processThreadableEdges(Condition, BB, Preference, Terminator))
1140 return true;
1141 return ConstantFolded;
1142 }
1143
1144 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
1145 // If we're branching on a conditional, LVI might be able to determine
1146 // it's value at the branch instruction. We only handle comparisons
1147 // against a constant at this time.
1148 // TODO: This should be extended to handle switches as well.
1149 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1150 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
1151 if (CondBr && CondConst) {
1152 // We should have returned as soon as we turn a conditional branch to
1153 // unconditional. Because its no longer interesting as far as jump
1154 // threading is concerned.
1155 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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1155, __PRETTY_FUNCTION__))
;
1156
1157 LazyValueInfo::Tristate Ret =
1158 LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0),
1159 CondConst, CondBr);
1160 if (Ret != LazyValueInfo::Unknown) {
1161 unsigned ToRemove = Ret == LazyValueInfo::True ? 1 : 0;
1162 unsigned ToKeep = Ret == LazyValueInfo::True ? 0 : 1;
1163 BasicBlock *ToRemoveSucc = CondBr->getSuccessor(ToRemove);
1164 ToRemoveSucc->removePredecessor(BB, true);
1165 BranchInst *UncondBr =
1166 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
1167 UncondBr->setDebugLoc(CondBr->getDebugLoc());
1168 CondBr->eraseFromParent();
1169 if (CondCmp->use_empty())
1170 CondCmp->eraseFromParent();
1171 // We can safely replace *some* uses of the CondInst if it has
1172 // exactly one value as returned by LVI. RAUW is incorrect in the
1173 // presence of guards and assumes, that have the `Cond` as the use. This
1174 // is because we use the guards/assume to reason about the `Cond` value
1175 // at the end of block, but RAUW unconditionally replaces all uses
1176 // including the guards/assumes themselves and the uses before the
1177 // guard/assume.
1178 else if (CondCmp->getParent() == BB) {
1179 auto *CI = Ret == LazyValueInfo::True ?
1180 ConstantInt::getTrue(CondCmp->getType()) :
1181 ConstantInt::getFalse(CondCmp->getType());
1182 replaceFoldableUses(CondCmp, CI);
1183 }
1184 DTU->applyUpdatesPermissive(
1185 {{DominatorTree::Delete, BB, ToRemoveSucc}});
1186 if (HasProfileData)
1187 BPI->eraseBlock(BB);
1188 return true;
1189 }
1190
1191 // We did not manage to simplify this branch, try to see whether
1192 // CondCmp depends on a known phi-select pattern.
1193 if (tryToUnfoldSelect(CondCmp, BB))
1194 return true;
1195 }
1196 }
1197
1198 if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
1199 if (tryToUnfoldSelect(SI, BB))
1200 return true;
1201
1202 // Check for some cases that are worth simplifying. Right now we want to look
1203 // for loads that are used by a switch or by the condition for the branch. If
1204 // we see one, check to see if it's partially redundant. If so, insert a PHI
1205 // which can then be used to thread the values.
1206 Value *SimplifyValue = CondInst;
1207
1208 if (auto *FI = dyn_cast<FreezeInst>(SimplifyValue))
1209 // Look into freeze's operand
1210 SimplifyValue = FI->getOperand(0);
1211
1212 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
1213 if (isa<Constant>(CondCmp->getOperand(1)))
1214 SimplifyValue = CondCmp->getOperand(0);
1215
1216 // TODO: There are other places where load PRE would be profitable, such as
1217 // more complex comparisons.
1218 if (LoadInst *LoadI = dyn_cast<LoadInst>(SimplifyValue))
1219 if (simplifyPartiallyRedundantLoad(LoadI))
1220 return true;
1221
1222 // Before threading, try to propagate profile data backwards:
1223 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
1224 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1225 updatePredecessorProfileMetadata(PN, BB);
1226
1227 // Handle a variety of cases where we are branching on something derived from
1228 // a PHI node in the current block. If we can prove that any predecessors
1229 // compute a predictable value based on a PHI node, thread those predecessors.
1230 if (processThreadableEdges(CondInst, BB, Preference, Terminator))
1231 return true;
1232
1233 // If this is an otherwise-unfoldable branch on a phi node or freeze(phi) in
1234 // the current block, see if we can simplify.
1235 PHINode *PN = dyn_cast<PHINode>(
1236 isa<FreezeInst>(CondInst) ? cast<FreezeInst>(CondInst)->getOperand(0)
1237 : CondInst);
1238
1239 if (PN && PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1240 return processBranchOnPHI(PN);
1241
1242 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
1243 if (CondInst->getOpcode() == Instruction::Xor &&
1244 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1245 return processBranchOnXOR(cast<BinaryOperator>(CondInst));
1246
1247 // Search for a stronger dominating condition that can be used to simplify a
1248 // conditional branch leaving BB.
1249 if (processImpliedCondition(BB))
1250 return true;
1251
1252 return false;
1253}
1254
1255bool JumpThreadingPass::processImpliedCondition(BasicBlock *BB) {
1256 auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
1257 if (!BI || !BI->isConditional())
1258 return false;
1259
1260 Value *Cond = BI->getCondition();
1261 BasicBlock *CurrentBB = BB;
1262 BasicBlock *CurrentPred = BB->getSinglePredecessor();
1263 unsigned Iter = 0;
1264
1265 auto &DL = BB->getModule()->getDataLayout();
1266
1267 while (CurrentPred && Iter++ < ImplicationSearchThreshold) {
1268 auto *PBI = dyn_cast<BranchInst>(CurrentPred->getTerminator());
1269 if (!PBI || !PBI->isConditional())
1270 return false;
1271 if (PBI->getSuccessor(0) != CurrentBB && PBI->getSuccessor(1) != CurrentBB)
1272 return false;
1273
1274 bool CondIsTrue = PBI->getSuccessor(0) == CurrentBB;
1275 Optional<bool> Implication =
1276 isImpliedCondition(PBI->getCondition(), Cond, DL, CondIsTrue);
1277 if (Implication) {
1278 BasicBlock *KeepSucc = BI->getSuccessor(*Implication ? 0 : 1);
1279 BasicBlock *RemoveSucc = BI->getSuccessor(*Implication ? 1 : 0);
1280 RemoveSucc->removePredecessor(BB);
1281 BranchInst *UncondBI = BranchInst::Create(KeepSucc, BI);
1282 UncondBI->setDebugLoc(BI->getDebugLoc());
1283 BI->eraseFromParent();
1284 DTU->applyUpdatesPermissive({{DominatorTree::Delete, BB, RemoveSucc}});
1285 if (HasProfileData)
1286 BPI->eraseBlock(BB);
1287 return true;
1288 }
1289 CurrentBB = CurrentPred;
1290 CurrentPred = CurrentBB->getSinglePredecessor();
1291 }
1292
1293 return false;
1294}
1295
1296/// Return true if Op is an instruction defined in the given block.
1297static bool isOpDefinedInBlock(Value *Op, BasicBlock *BB) {
1298 if (Instruction *OpInst = dyn_cast<Instruction>(Op))
13
Assuming 'OpInst' is null
14
Taking false branch
1299 if (OpInst->getParent() == BB)
1300 return true;
1301 return false;
15
Returning zero, which participates in a condition later
1302}
1303
1304/// simplifyPartiallyRedundantLoad - If LoadI is an obviously partially
1305/// redundant load instruction, eliminate it by replacing it with a PHI node.
1306/// This is an important optimization that encourages jump threading, and needs
1307/// to be run interlaced with other jump threading tasks.
1308bool JumpThreadingPass::simplifyPartiallyRedundantLoad(LoadInst *LoadI) {
1309 // Don't hack volatile and ordered loads.
1310 if (!LoadI->isUnordered()) return false;
1
Calling 'LoadInst::isUnordered'
6
Returning from 'LoadInst::isUnordered'
7
Taking false branch
1311
1312 // If the load is defined in a block with exactly one predecessor, it can't be
1313 // partially redundant.
1314 BasicBlock *LoadBB = LoadI->getParent();
1315 if (LoadBB->getSinglePredecessor())
8
Assuming the condition is false
9
Taking false branch
1316 return false;
1317
1318 // If the load is defined in an EH pad, it can't be partially redundant,
1319 // because the edges between the invoke and the EH pad cannot have other
1320 // instructions between them.
1321 if (LoadBB->isEHPad())
10
Assuming the condition is false
11
Taking false branch
1322 return false;
1323
1324 Value *LoadedPtr = LoadI->getOperand(0);
1325
1326 // If the loaded operand is defined in the LoadBB and its not a phi,
1327 // it can't be available in predecessors.
1328 if (isOpDefinedInBlock(LoadedPtr, LoadBB) && !isa<PHINode>(LoadedPtr))
12
Calling 'isOpDefinedInBlock'
16
Returning from 'isOpDefinedInBlock'
1329 return false;
1330
1331 // Scan a few instructions up from the load, to see if it is obviously live at
1332 // the entry to its block.
1333 BasicBlock::iterator BBIt(LoadI);
1334 bool IsLoadCSE;
1335 if (Value *AvailableVal = FindAvailableLoadedValue(
17
Assuming 'AvailableVal' is null
18
Taking false branch
1336 LoadI, LoadBB, BBIt, DefMaxInstsToScan, AA, &IsLoadCSE)) {
1337 // If the value of the load is locally available within the block, just use
1338 // it. This frequently occurs for reg2mem'd allocas.
1339
1340 if (IsLoadCSE) {
1341 LoadInst *NLoadI = cast<LoadInst>(AvailableVal);
1342 combineMetadataForCSE(NLoadI, LoadI, false);
1343 };
1344
1345 // If the returned value is the load itself, replace with an undef. This can
1346 // only happen in dead loops.
1347 if (AvailableVal == LoadI)
1348 AvailableVal = UndefValue::get(LoadI->getType());
1349 if (AvailableVal->getType() != LoadI->getType())
1350 AvailableVal = CastInst::CreateBitOrPointerCast(
1351 AvailableVal, LoadI->getType(), "", LoadI);
1352 LoadI->replaceAllUsesWith(AvailableVal);
1353 LoadI->eraseFromParent();
1354 return true;
1355 }
1356
1357 // Otherwise, if we scanned the whole block and got to the top of the block,
1358 // we know the block is locally transparent to the load. If not, something
1359 // might clobber its value.
1360 if (BBIt != LoadBB->begin())
19
Calling 'operator!='
22
Returning from 'operator!='
23
Taking false branch
1361 return false;
1362
1363 // If all of the loads and stores that feed the value have the same AA tags,
1364 // then we can propagate them onto any newly inserted loads.
1365 AAMDNodes AATags;
1366 LoadI->getAAMetadata(AATags);
1367
1368 SmallPtrSet<BasicBlock*, 8> PredsScanned;
1369
1370 using AvailablePredsTy = SmallVector<std::pair<BasicBlock *, Value *>, 8>;
1371
1372 AvailablePredsTy AvailablePreds;
1373 BasicBlock *OneUnavailablePred = nullptr;
24
'OneUnavailablePred' initialized to a null pointer value
1374 SmallVector<LoadInst*, 8> CSELoads;
1375
1376 // If we got here, the loaded value is transparent through to the start of the
1377 // block. Check to see if it is available in any of the predecessor blocks.
1378 for (BasicBlock *PredBB : predecessors(LoadBB)) {
1379 // If we already scanned this predecessor, skip it.
1380 if (!PredsScanned.insert(PredBB).second)
1381 continue;
1382
1383 BBIt = PredBB->end();
1384 unsigned NumScanedInst = 0;
1385 Value *PredAvailable = nullptr;
1386 // NOTE: We don't CSE load that is volatile or anything stronger than
1387 // unordered, that should have been checked when we entered the function.
1388 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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1389, __PRETTY_FUNCTION__))
1389 "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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1389, __PRETTY_FUNCTION__))
;
1390 // If this is a load on a phi pointer, phi-translate it and search
1391 // for available load/store to the pointer in predecessors.
1392 Type *AccessTy = LoadI->getType();
1393 const auto &DL = LoadI->getModule()->getDataLayout();
1394 MemoryLocation Loc(LoadedPtr->DoPHITranslation(LoadBB, PredBB),
1395 LocationSize::precise(DL.getTypeStoreSize(AccessTy)),
1396 AATags);
1397 PredAvailable = findAvailablePtrLoadStore(Loc, AccessTy, LoadI->isAtomic(),
1398 PredBB, BBIt, DefMaxInstsToScan,
1399 AA, &IsLoadCSE, &NumScanedInst);
1400
1401 // If PredBB has a single predecessor, continue scanning through the
1402 // single predecessor.
1403 BasicBlock *SinglePredBB = PredBB;
1404 while (!PredAvailable && SinglePredBB && BBIt == SinglePredBB->begin() &&
1405 NumScanedInst < DefMaxInstsToScan) {
1406 SinglePredBB = SinglePredBB->getSinglePredecessor();
1407 if (SinglePredBB) {
1408 BBIt = SinglePredBB->end();
1409 PredAvailable = findAvailablePtrLoadStore(
1410 Loc, AccessTy, LoadI->isAtomic(), SinglePredBB, BBIt,
1411 (DefMaxInstsToScan - NumScanedInst), AA, &IsLoadCSE,
1412 &NumScanedInst);
1413 }
1414 }
1415
1416 if (!PredAvailable) {
1417 OneUnavailablePred = PredBB;
1418 continue;
1419 }
1420
1421 if (IsLoadCSE)
1422 CSELoads.push_back(cast<LoadInst>(PredAvailable));
1423
1424 // If so, this load is partially redundant. Remember this info so that we
1425 // can create a PHI node.
1426 AvailablePreds.emplace_back(PredBB, PredAvailable);
1427 }
1428
1429 // If the loaded value isn't available in any predecessor, it isn't partially
1430 // redundant.
1431 if (AvailablePreds.empty()) return false;
25
Calling 'SmallVectorBase::empty'
28
Returning from 'SmallVectorBase::empty'
29
Taking false branch
1432
1433 // Okay, the loaded value is available in at least one (and maybe all!)
1434 // predecessors. If the value is unavailable in more than one unique
1435 // predecessor, we want to insert a merge block for those common predecessors.
1436 // This ensures that we only have to insert one reload, thus not increasing
1437 // code size.
1438 BasicBlock *UnavailablePred = nullptr;
1439
1440 // If the value is unavailable in one of predecessors, we will end up
1441 // inserting a new instruction into them. It is only valid if all the
1442 // instructions before LoadI are guaranteed to pass execution to its
1443 // successor, or if LoadI is safe to speculate.
1444 // TODO: If this logic becomes more complex, and we will perform PRE insertion
1445 // farther than to a predecessor, we need to reuse the code from GVN's PRE.
1446 // It requires domination tree analysis, so for this simple case it is an
1447 // overkill.
1448 if (PredsScanned.size() != AvailablePreds.size() &&
30
Assuming the condition is false
1449 !isSafeToSpeculativelyExecute(LoadI))
1450 for (auto I = LoadBB->begin(); &*I != LoadI; ++I)
1451 if (!isGuaranteedToTransferExecutionToSuccessor(&*I))
1452 return false;
1453
1454 // If there is exactly one predecessor where the value is unavailable, the
1455 // already computed 'OneUnavailablePred' block is it. If it ends in an
1456 // unconditional branch, we know that it isn't a critical edge.
1457 if (PredsScanned.size() == AvailablePreds.size()+1 &&
31
Assuming the condition is true
1458 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
32
Called C++ object pointer is null
1459 UnavailablePred = OneUnavailablePred;
1460 } else if (PredsScanned.size() != AvailablePreds.size()) {
1461 // Otherwise, we had multiple unavailable predecessors or we had a critical
1462 // edge from the one.
1463 SmallVector<BasicBlock*, 8> PredsToSplit;
1464 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
1465
1466 for (const auto &AvailablePred : AvailablePreds)
1467 AvailablePredSet.insert(AvailablePred.first);
1468
1469 // Add all the unavailable predecessors to the PredsToSplit list.
1470 for (BasicBlock *P : predecessors(LoadBB)) {
1471 // If the predecessor is an indirect goto, we can't split the edge.
1472 // Same for CallBr.
1473 if (isa<IndirectBrInst>(P->getTerminator()) ||
1474 isa<CallBrInst>(P->getTerminator()))
1475 return false;
1476
1477 if (!AvailablePredSet.count(P))
1478 PredsToSplit.push_back(P);
1479 }
1480
1481 // Split them out to their own block.
1482 UnavailablePred = splitBlockPreds(LoadBB, PredsToSplit, "thread-pre-split");
1483 }
1484
1485 // If the value isn't available in all predecessors, then there will be
1486 // exactly one where it isn't available. Insert a load on that edge and add
1487 // it to the AvailablePreds list.
1488 if (UnavailablePred) {
1489 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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1490, __PRETTY_FUNCTION__))
1490 "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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1490, __PRETTY_FUNCTION__))
;
1491 LoadInst *NewVal = new LoadInst(
1492 LoadI->getType(), LoadedPtr->DoPHITranslation(LoadBB, UnavailablePred),
1493 LoadI->getName() + ".pr", false, LoadI->getAlign(),
1494 LoadI->getOrdering(), LoadI->getSyncScopeID(),
1495 UnavailablePred->getTerminator());
1496 NewVal->setDebugLoc(LoadI->getDebugLoc());
1497 if (AATags)
1498 NewVal->setAAMetadata(AATags);
1499
1500 AvailablePreds.emplace_back(UnavailablePred, NewVal);
1501 }
1502
1503 // Now we know that each predecessor of this block has a value in
1504 // AvailablePreds, sort them for efficient access as we're walking the preds.
1505 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1506
1507 // Create a PHI node at the start of the block for the PRE'd load value.
1508 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
1509 PHINode *PN = PHINode::Create(LoadI->getType(), std::distance(PB, PE), "",
1510 &LoadBB->front());
1511 PN->takeName(LoadI);
1512 PN->setDebugLoc(LoadI->getDebugLoc());
1513
1514 // Insert new entries into the PHI for each predecessor. A single block may
1515 // have multiple entries here.
1516 for (pred_iterator PI = PB; PI != PE; ++PI) {
1517 BasicBlock *P = *PI;
1518 AvailablePredsTy::iterator I =
1519 llvm::lower_bound(AvailablePreds, std::make_pair(P, (Value *)nullptr));
1520
1521 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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1522, __PRETTY_FUNCTION__))
1522 "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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1522, __PRETTY_FUNCTION__))
;
1523
1524 // If we have an available predecessor but it requires casting, insert the
1525 // cast in the predecessor and use the cast. Note that we have to update the
1526 // AvailablePreds vector as we go so that all of the PHI entries for this
1527 // predecessor use the same bitcast.
1528 Value *&PredV = I->second;
1529 if (PredV->getType() != LoadI->getType())
1530 PredV = CastInst::CreateBitOrPointerCast(PredV, LoadI->getType(), "",
1531 P->getTerminator());
1532
1533 PN->addIncoming(PredV, I->first);
1534 }
1535
1536 for (LoadInst *PredLoadI : CSELoads) {
1537 combineMetadataForCSE(PredLoadI, LoadI, true);
1538 }
1539
1540 LoadI->replaceAllUsesWith(PN);
1541 LoadI->eraseFromParent();
1542
1543 return true;
1544}
1545
1546/// findMostPopularDest - The specified list contains multiple possible
1547/// threadable destinations. Pick the one that occurs the most frequently in
1548/// the list.
1549static BasicBlock *
1550findMostPopularDest(BasicBlock *BB,
1551 const SmallVectorImpl<std::pair<BasicBlock *,
1552 BasicBlock *>> &PredToDestList) {
1553 assert(!PredToDestList.empty())((!PredToDestList.empty()) ? static_cast<void> (0) : __assert_fail
("!PredToDestList.empty()", "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1553, __PRETTY_FUNCTION__))
;
1554
1555 // Determine popularity. If there are multiple possible destinations, we
1556 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1557 // blocks with known and real destinations to threading undef. We'll handle
1558 // them later if interesting.
1559 MapVector<BasicBlock *, unsigned> DestPopularity;
1560
1561 // Populate DestPopularity with the successors in the order they appear in the
1562 // successor list. This way, we ensure determinism by iterating it in the
1563 // same order in std::max_element below. We map nullptr to 0 so that we can
1564 // return nullptr when PredToDestList contains nullptr only.
1565 DestPopularity[nullptr] = 0;
1566 for (auto *SuccBB : successors(BB))
1567 DestPopularity[SuccBB] = 0;
1568
1569 for (const auto &PredToDest : PredToDestList)
1570 if (PredToDest.second)
1571 DestPopularity[PredToDest.second]++;
1572
1573 // Find the most popular dest.
1574 using VT = decltype(DestPopularity)::value_type;
1575 auto MostPopular = std::max_element(
1576 DestPopularity.begin(), DestPopularity.end(),
1577 [](const VT &L, const VT &R) { return L.second < R.second; });
1578
1579 // Okay, we have finally picked the most popular destination.
1580 return MostPopular->first;
1581}
1582
1583// Try to evaluate the value of V when the control flows from PredPredBB to
1584// BB->getSinglePredecessor() and then on to BB.
1585Constant *JumpThreadingPass::evaluateOnPredecessorEdge(BasicBlock *BB,
1586 BasicBlock *PredPredBB,
1587 Value *V) {
1588 BasicBlock *PredBB = BB->getSinglePredecessor();
1589 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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1589, __PRETTY_FUNCTION__))
;
1590
1591 if (Constant *Cst = dyn_cast<Constant>(V)) {
1592 return Cst;
1593 }
1594
1595 // Consult LVI if V is not an instruction in BB or PredBB.
1596 Instruction *I = dyn_cast<Instruction>(V);
1597 if (!I || (I->getParent() != BB && I->getParent() != PredBB)) {
1598 return LVI->getConstantOnEdge(V, PredPredBB, PredBB, nullptr);
1599 }
1600
1601 // Look into a PHI argument.
1602 if (PHINode *PHI = dyn_cast<PHINode>(V)) {
1603 if (PHI->getParent() == PredBB)
1604 return dyn_cast<Constant>(PHI->getIncomingValueForBlock(PredPredBB));
1605 return nullptr;
1606 }
1607
1608 // If we have a CmpInst, try to fold it for each incoming edge into PredBB.
1609 if (CmpInst *CondCmp = dyn_cast<CmpInst>(V)) {
1610 if (CondCmp->getParent() == BB) {
1611 Constant *Op0 =
1612 evaluateOnPredecessorEdge(BB, PredPredBB, CondCmp->getOperand(0));
1613 Constant *Op1 =
1614 evaluateOnPredecessorEdge(BB, PredPredBB, CondCmp->getOperand(1));
1615 if (Op0 && Op1) {
1616 return ConstantExpr::getCompare(CondCmp->getPredicate(), Op0, Op1);
1617 }
1618 }
1619 return nullptr;
1620 }
1621
1622 return nullptr;
1623}
1624
1625bool JumpThreadingPass::processThreadableEdges(Value *Cond, BasicBlock *BB,
1626 ConstantPreference Preference,
1627 Instruction *CxtI) {
1628 // If threading this would thread across a loop header, don't even try to
1629 // thread the edge.
1630 if (LoopHeaders.count(BB))
1631 return false;
1632
1633 PredValueInfoTy PredValues;
1634 if (!computeValueKnownInPredecessors(Cond, BB, PredValues, Preference,
1635 CxtI)) {
1636 // We don't have known values in predecessors. See if we can thread through
1637 // BB and its sole predecessor.
1638 return maybethreadThroughTwoBasicBlocks(BB, Cond);
1639 }
1640
1641 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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1642, __PRETTY_FUNCTION__))
1642 "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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1642, __PRETTY_FUNCTION__))
;
1643
1644 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)
1645 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)
1646 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)
1647 << "': 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)
1648 << " 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)
1649 })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)
;
1650
1651 // Decide what we want to thread through. Convert our list of known values to
1652 // a list of known destinations for each pred. This also discards duplicate
1653 // predecessors and keeps track of the undefined inputs (which are represented
1654 // as a null dest in the PredToDestList).
1655 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1656 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1657
1658 BasicBlock *OnlyDest = nullptr;
1659 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1660 Constant *OnlyVal = nullptr;
1661 Constant *MultipleVal = (Constant *)(intptr_t)~0ULL;
1662
1663 for (const auto &PredValue : PredValues) {
1664 BasicBlock *Pred = PredValue.second;
1665 if (!SeenPreds.insert(Pred).second)
1666 continue; // Duplicate predecessor entry.
1667
1668 Constant *Val = PredValue.first;
1669
1670 BasicBlock *DestBB;
1671 if (isa<UndefValue>(Val))
1672 DestBB = nullptr;
1673 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
1674 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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1674, __PRETTY_FUNCTION__))
;
1675 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1676 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1677 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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1677, __PRETTY_FUNCTION__))
;
1678 DestBB = SI->findCaseValue(cast<ConstantInt>(Val))->getCaseSuccessor();
1679 } else {
1680 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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1681, __PRETTY_FUNCTION__))
1681 && "Unexpected terminator")((isa<IndirectBrInst>(BB->getTerminator()) &&
"Unexpected terminator") ? static_cast<void> (0) : __assert_fail
("isa<IndirectBrInst>(BB->getTerminator()) && \"Unexpected terminator\""
, "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1681, __PRETTY_FUNCTION__))
;
1682 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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1682, __PRETTY_FUNCTION__))
;
1683 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1684 }
1685
1686 // If we have exactly one destination, remember it for efficiency below.
1687 if (PredToDestList.empty()) {
1688 OnlyDest = DestBB;
1689 OnlyVal = Val;
1690 } else {
1691 if (OnlyDest != DestBB)
1692 OnlyDest = MultipleDestSentinel;
1693 // It possible we have same destination, but different value, e.g. default
1694 // case in switchinst.
1695 if (Val != OnlyVal)
1696 OnlyVal = MultipleVal;
1697 }
1698
1699 // If the predecessor ends with an indirect goto, we can't change its
1700 // destination. Same for CallBr.
1701 if (isa<IndirectBrInst>(Pred->getTerminator()) ||
1702 isa<CallBrInst>(Pred->getTerminator()))
1703 continue;
1704
1705 PredToDestList.emplace_back(Pred, DestBB);
1706 }
1707
1708 // If all edges were unthreadable, we fail.
1709 if (PredToDestList.empty())
1710 return false;
1711
1712 // If all the predecessors go to a single known successor, we want to fold,
1713 // not thread. By doing so, we do not need to duplicate the current block and
1714 // also miss potential opportunities in case we dont/cant duplicate.
1715 if (OnlyDest && OnlyDest != MultipleDestSentinel) {
1716 if (BB->hasNPredecessors(PredToDestList.size())) {
1717 bool SeenFirstBranchToOnlyDest = false;
1718 std::vector <DominatorTree::UpdateType> Updates;
1719 Updates.reserve(BB->getTerminator()->getNumSuccessors() - 1);
1720 for (BasicBlock *SuccBB : successors(BB)) {
1721 if (SuccBB == OnlyDest && !SeenFirstBranchToOnlyDest) {
1722 SeenFirstBranchToOnlyDest = true; // Don't modify the first branch.
1723 } else {
1724 SuccBB->removePredecessor(BB, true); // This is unreachable successor.
1725 Updates.push_back({DominatorTree::Delete, BB, SuccBB});
1726 }
1727 }
1728
1729 // Finally update the terminator.
1730 Instruction *Term = BB->getTerminator();
1731 BranchInst::Create(OnlyDest, Term);
1732 Term->eraseFromParent();
1733 DTU->applyUpdatesPermissive(Updates);
1734 if (HasProfileData)
1735 BPI->eraseBlock(BB);
1736
1737 // If the condition is now dead due to the removal of the old terminator,
1738 // erase it.
1739 if (auto *CondInst = dyn_cast<Instruction>(Cond)) {
1740 if (CondInst->use_empty() && !CondInst->mayHaveSideEffects())
1741 CondInst->eraseFromParent();
1742 // We can safely replace *some* uses of the CondInst if it has
1743 // exactly one value as returned by LVI. RAUW is incorrect in the
1744 // presence of guards and assumes, that have the `Cond` as the use. This
1745 // is because we use the guards/assume to reason about the `Cond` value
1746 // at the end of block, but RAUW unconditionally replaces all uses
1747 // including the guards/assumes themselves and the uses before the
1748 // guard/assume.
1749 else if (OnlyVal && OnlyVal != MultipleVal &&
1750 CondInst->getParent() == BB)
1751 replaceFoldableUses(CondInst, OnlyVal);
1752 }
1753 return true;
1754 }
1755 }
1756
1757 // Determine which is the most common successor. If we have many inputs and
1758 // this block is a switch, we want to start by threading the batch that goes
1759 // to the most popular destination first. If we only know about one
1760 // threadable destination (the common case) we can avoid this.
1761 BasicBlock *MostPopularDest = OnlyDest;
1762
1763 if (MostPopularDest == MultipleDestSentinel) {
1764 // Remove any loop headers from the Dest list, threadEdge conservatively
1765 // won't process them, but we might have other destination that are eligible
1766 // and we still want to process.
1767 erase_if(PredToDestList,
1768 [&](const std::pair<BasicBlock *, BasicBlock *> &PredToDest) {
1769 return LoopHeaders.contains(PredToDest.second);
1770 });
1771
1772 if (PredToDestList.empty())
1773 return false;
1774
1775 MostPopularDest = findMostPopularDest(BB, PredToDestList);
1776 }
1777
1778 // Now that we know what the most popular destination is, factor all
1779 // predecessors that will jump to it into a single predecessor.
1780 SmallVector<BasicBlock*, 16> PredsToFactor;
1781 for (const auto &PredToDest : PredToDestList)
1782 if (PredToDest.second == MostPopularDest) {
1783 BasicBlock *Pred = PredToDest.first;
1784
1785 // This predecessor may be a switch or something else that has multiple
1786 // edges to the block. Factor each of these edges by listing them
1787 // according to # occurrences in PredsToFactor.
1788 for (BasicBlock *Succ : successors(Pred))
1789 if (Succ == BB)
1790 PredsToFactor.push_back(Pred);
1791 }
1792
1793 // If the threadable edges are branching on an undefined value, we get to pick
1794 // the destination that these predecessors should get to.
1795 if (!MostPopularDest)
1796 MostPopularDest = BB->getTerminator()->
1797 getSuccessor(getBestDestForJumpOnUndef(BB));
1798
1799 // Ok, try to thread it!
1800 return tryThreadEdge(BB, PredsToFactor, MostPopularDest);
1801}
1802
1803/// processBranchOnPHI - We have an otherwise unthreadable conditional branch on
1804/// a PHI node (or freeze PHI) in the current block. See if there are any
1805/// simplifications we can do based on inputs to the phi node.
1806bool JumpThreadingPass::processBranchOnPHI(PHINode *PN) {
1807 BasicBlock *BB = PN->getParent();
1808
1809 // TODO: We could make use of this to do it once for blocks with common PHI
1810 // values.
1811 SmallVector<BasicBlock*, 1> PredBBs;
1812 PredBBs.resize(1);
1813
1814 // If any of the predecessor blocks end in an unconditional branch, we can
1815 // *duplicate* the conditional branch into that block in order to further
1816 // encourage jump threading and to eliminate cases where we have branch on a
1817 // phi of an icmp (branch on icmp is much better).
1818 // This is still beneficial when a frozen phi is used as the branch condition
1819 // because it allows CodeGenPrepare to further canonicalize br(freeze(icmp))
1820 // to br(icmp(freeze ...)).
1821 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1822 BasicBlock *PredBB = PN->getIncomingBlock(i);
1823 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1824 if (PredBr->isUnconditional()) {
1825 PredBBs[0] = PredBB;
1826 // Try to duplicate BB into PredBB.
1827 if (duplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1828 return true;
1829 }
1830 }
1831
1832 return false;
1833}
1834
1835/// processBranchOnXOR - We have an otherwise unthreadable conditional branch on
1836/// a xor instruction in the current block. See if there are any
1837/// simplifications we can do based on inputs to the xor.
1838bool JumpThreadingPass::processBranchOnXOR(BinaryOperator *BO) {
1839 BasicBlock *BB = BO->getParent();
1840
1841 // If either the LHS or RHS of the xor is a constant, don't do this
1842 // optimization.
1843 if (isa<ConstantInt>(BO->getOperand(0)) ||
1844 isa<ConstantInt>(BO->getOperand(1)))
1845 return false;
1846
1847 // If the first instruction in BB isn't a phi, we won't be able to infer
1848 // anything special about any particular predecessor.
1849 if (!isa<PHINode>(BB->front()))
1850 return false;
1851
1852 // If this BB is a landing pad, we won't be able to split the edge into it.
1853 if (BB->isEHPad())
1854 return false;
1855
1856 // If we have a xor as the branch input to this block, and we know that the
1857 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1858 // the condition into the predecessor and fix that value to true, saving some
1859 // logical ops on that path and encouraging other paths to simplify.
1860 //
1861 // This copies something like this:
1862 //
1863 // BB:
1864 // %X = phi i1 [1], [%X']
1865 // %Y = icmp eq i32 %A, %B
1866 // %Z = xor i1 %X, %Y
1867 // br i1 %Z, ...
1868 //
1869 // Into:
1870 // BB':
1871 // %Y = icmp ne i32 %A, %B
1872 // br i1 %Y, ...
1873
1874 PredValueInfoTy XorOpValues;
1875 bool isLHS = true;
1876 if (!computeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1877 WantInteger, BO)) {
1878 assert(XorOpValues.empty())((XorOpValues.empty()) ? static_cast<void> (0) : __assert_fail
("XorOpValues.empty()", "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1878, __PRETTY_FUNCTION__))
;
1879 if (!computeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1880 WantInteger, BO))
1881 return false;
1882 isLHS = false;
1883 }
1884
1885 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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1886, __PRETTY_FUNCTION__))
1886 "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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 1886, __PRETTY_FUNCTION__))
;
1887
1888 // Scan the information to see which is most popular: true or false. The
1889 // predecessors can be of the set true, false, or undef.
1890 unsigned NumTrue = 0, NumFalse = 0;
1891 for (const auto &XorOpValue : XorOpValues) {
1892 if (isa<UndefValue>(XorOpValue.first))
1893 // Ignore undefs for the count.
1894 continue;
1895 if (cast<ConstantInt>(XorOpValue.first)->isZero())
1896 ++NumFalse;
1897 else
1898 ++NumTrue;
1899 }
1900
1901 // Determine which value to split on, true, false, or undef if neither.
1902 ConstantInt *SplitVal = nullptr;
1903 if (NumTrue > NumFalse)
1904 SplitVal = ConstantInt::getTrue(BB->getContext());
1905 else if (NumTrue != 0 || NumFalse != 0)
1906 SplitVal = ConstantInt::getFalse(BB->getContext());
1907
1908 // Collect all of the blocks that this can be folded into so that we can
1909 // factor this once and clone it once.
1910 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1911 for (const auto &XorOpValue : XorOpValues) {
1912 if (XorOpValue.first != SplitVal && !isa<UndefValue>(XorOpValue.first))
1913 continue;
1914
1915 BlocksToFoldInto.push_back(XorOpValue.second);
1916 }
1917
1918 // If we inferred a value for all of the predecessors, then duplication won't
1919 // help us. However, we can just replace the LHS or RHS with the constant.
1920 if (BlocksToFoldInto.size() ==
1921 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1922 if (!SplitVal) {
1923 // If all preds provide undef, just nuke the xor, because it is undef too.
1924 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1925 BO->eraseFromParent();
1926 } else if (SplitVal->isZero()) {
1927 // If all preds provide 0, replace the xor with the other input.
1928 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1929 BO->eraseFromParent();
1930 } else {
1931 // If all preds provide 1, set the computed value to 1.
1932 BO->setOperand(!isLHS, SplitVal);
1933 }
1934
1935 return true;
1936 }
1937
1938 // If any of predecessors end with an indirect goto, we can't change its
1939 // destination. Same for CallBr.
1940 if (any_of(BlocksToFoldInto, [](BasicBlock *Pred) {
1941 return isa<IndirectBrInst>(Pred->getTerminator()) ||
1942 isa<CallBrInst>(Pred->getTerminator());
1943 }))
1944 return false;
1945
1946 // Try to duplicate BB into PredBB.
1947 return duplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1948}
1949
1950/// addPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1951/// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1952/// NewPred using the entries from OldPred (suitably mapped).
1953static void addPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1954 BasicBlock *OldPred,
1955 BasicBlock *NewPred,
1956 DenseMap<Instruction*, Value*> &ValueMap) {
1957 for (PHINode &PN : PHIBB->phis()) {
1958 // Ok, we have a PHI node. Figure out what the incoming value was for the
1959 // DestBlock.
1960 Value *IV = PN.getIncomingValueForBlock(OldPred);
1961
1962 // Remap the value if necessary.
1963 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1964 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1965 if (I != ValueMap.end())
1966 IV = I->second;
1967 }
1968
1969 PN.addIncoming(IV, NewPred);
1970 }
1971}
1972
1973/// Merge basic block BB into its sole predecessor if possible.
1974bool JumpThreadingPass::maybeMergeBasicBlockIntoOnlyPred(BasicBlock *BB) {
1975 BasicBlock *SinglePred = BB->getSinglePredecessor();
1976 if (!SinglePred)
1977 return false;
1978
1979 const Instruction *TI = SinglePred->getTerminator();
1980 if (TI->isExceptionalTerminator() || TI->getNumSuccessors() != 1 ||
1981 SinglePred == BB || hasAddressTakenAndUsed(BB))
1982 return false;
1983
1984 // If SinglePred was a loop header, BB becomes one.
1985 if (LoopHeaders.erase(SinglePred))
1986 LoopHeaders.insert(BB);
1987
1988 LVI->eraseBlock(SinglePred);
1989 MergeBasicBlockIntoOnlyPred(BB, DTU);
1990
1991 // Now that BB is merged into SinglePred (i.e. SinglePred code followed by
1992 // BB code within one basic block `BB`), we need to invalidate the LVI
1993 // information associated with BB, because the LVI information need not be
1994 // true for all of BB after the merge. For example,
1995 // Before the merge, LVI info and code is as follows:
1996 // SinglePred: <LVI info1 for %p val>
1997 // %y = use of %p
1998 // call @exit() // need not transfer execution to successor.
1999 // assume(%p) // from this point on %p is true
2000 // br label %BB
2001 // BB: <LVI info2 for %p val, i.e. %p is true>
2002 // %x = use of %p
2003 // br label exit
2004 //
2005 // Note that this LVI info for blocks BB and SinglPred is correct for %p
2006 // (info2 and info1 respectively). After the merge and the deletion of the
2007 // LVI info1 for SinglePred. We have the following code:
2008 // BB: <LVI info2 for %p val>
2009 // %y = use of %p
2010 // call @exit()
2011 // assume(%p)
2012 // %x = use of %p <-- LVI info2 is correct from here onwards.
2013 // br label exit
2014 // LVI info2 for BB is incorrect at the beginning of BB.
2015
2016 // Invalidate LVI information for BB if the LVI is not provably true for
2017 // all of BB.
2018 if (!isGuaranteedToTransferExecutionToSuccessor(BB))
2019 LVI->eraseBlock(BB);
2020 return true;
2021}
2022
2023/// Update the SSA form. NewBB contains instructions that are copied from BB.
2024/// ValueMapping maps old values in BB to new ones in NewBB.
2025void JumpThreadingPass::updateSSA(
2026 BasicBlock *BB, BasicBlock *NewBB,
2027 DenseMap<Instruction *, Value *> &ValueMapping) {
2028 // If there were values defined in BB that are used outside the block, then we
2029 // now have to update all uses of the value to use either the original value,
2030 // the cloned value, or some PHI derived value. This can require arbitrary
2031 // PHI insertion, of which we are prepared to do, clean these up now.
2032 SSAUpdater SSAUpdate;
2033 SmallVector<Use *, 16> UsesToRename;
2034
2035 for (Instruction &I : *BB) {
2036 // Scan all uses of this instruction to see if it is used outside of its
2037 // block, and if so, record them in UsesToRename.
2038 for (Use &U : I.uses()) {
2039 Instruction *User = cast<Instruction>(U.getUser());
2040 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
2041 if (UserPN->getIncomingBlock(U) == BB)
2042 continue;
2043 } else if (User->getParent() == BB)
2044 continue;
2045
2046 UsesToRename.push_back(&U);
2047 }
2048
2049 // If there are no uses outside the block, we're done with this instruction.
2050 if (UsesToRename.empty())
2051 continue;
2052 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)
;
2053
2054 // We found a use of I outside of BB. Rename all uses of I that are outside
2055 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
2056 // with the two values we know.
2057 SSAUpdate.Initialize(I.getType(), I.getName());
2058 SSAUpdate.AddAvailableValue(BB, &I);
2059 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[&I]);
2060
2061 while (!UsesToRename.empty())
2062 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
2063 LLVM_DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "\n"; } } while (false)
;
2064 }
2065}
2066
2067/// Clone instructions in range [BI, BE) to NewBB. For PHI nodes, we only clone
2068/// arguments that come from PredBB. Return the map from the variables in the
2069/// source basic block to the variables in the newly created basic block.
2070DenseMap<Instruction *, Value *>
2071JumpThreadingPass::cloneInstructions(BasicBlock::iterator BI,
2072 BasicBlock::iterator BE, BasicBlock *NewBB,
2073 BasicBlock *PredBB) {
2074 // We are going to have to map operands from the source basic block to the new
2075 // copy of the block 'NewBB'. If there are PHI nodes in the source basic
2076 // block, evaluate them to account for entry from PredBB.
2077 DenseMap<Instruction *, Value *> ValueMapping;
2078
2079 // Clone the phi nodes of the source basic block into NewBB. The resulting
2080 // phi nodes are trivial since NewBB only has one predecessor, but SSAUpdater
2081 // might need to rewrite the operand of the cloned phi.
2082 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) {
2083 PHINode *NewPN = PHINode::Create(PN->getType(), 1, PN->getName(), NewBB);
2084 NewPN->addIncoming(PN->getIncomingValueForBlock(PredBB), PredBB);
2085 ValueMapping[PN] = NewPN;
2086 }
2087
2088 // Clone noalias scope declarations in the threaded block. When threading a
2089 // loop exit, we would otherwise end up with two idential scope declarations
2090 // visible at the same time.
2091 SmallVector<MDNode *> NoAliasScopes;
2092 DenseMap<MDNode *, MDNode *> ClonedScopes;
2093 LLVMContext &Context = PredBB->getContext();
2094 identifyNoAliasScopesToClone(BI, BE, NoAliasScopes);
2095 cloneNoAliasScopes(NoAliasScopes, ClonedScopes, "thread", Context);
2096
2097 // Clone the non-phi instructions of the source basic block into NewBB,
2098 // keeping track of the mapping and using it to remap operands in the cloned
2099 // instructions.
2100 for (; BI != BE; ++BI) {
2101 Instruction *New = BI->clone();
2102 New->setName(BI->getName());
2103 NewBB->getInstList().push_back(New);
2104 ValueMapping[&*BI] = New;
2105 adaptNoAliasScopes(New, ClonedScopes, Context);
2106
2107 // Remap operands to patch up intra-block references.
2108 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2109 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
2110 DenseMap<Instruction *, Value *>::iterator I = ValueMapping.find(Inst);
2111 if (I != ValueMapping.end())
2112 New->setOperand(i, I->second);
2113 }
2114 }
2115
2116 return ValueMapping;
2117}
2118
2119/// Attempt to thread through two successive basic blocks.
2120bool JumpThreadingPass::maybethreadThroughTwoBasicBlocks(BasicBlock *BB,
2121 Value *Cond) {
2122 // Consider:
2123 //
2124 // PredBB:
2125 // %var = phi i32* [ null, %bb1 ], [ @a, %bb2 ]
2126 // %tobool = icmp eq i32 %cond, 0
2127 // br i1 %tobool, label %BB, label ...
2128 //
2129 // BB:
2130 // %cmp = icmp eq i32* %var, null
2131 // br i1 %cmp, label ..., label ...
2132 //
2133 // We don't know the value of %var at BB even if we know which incoming edge
2134 // we take to BB. However, once we duplicate PredBB for each of its incoming
2135 // edges (say, PredBB1 and PredBB2), we know the value of %var in each copy of
2136 // PredBB. Then we can thread edges PredBB1->BB and PredBB2->BB through BB.
2137
2138 // Require that BB end with a Branch for simplicity.
2139 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
2140 if (!CondBr)
2141 return false;
2142
2143 // BB must have exactly one predecessor.
2144 BasicBlock *PredBB = BB->getSinglePredecessor();
2145 if (!PredBB)
2146 return false;
2147
2148 // Require that PredBB end with a conditional Branch. If PredBB ends with an
2149 // unconditional branch, we should be merging PredBB and BB instead. For
2150 // simplicity, we don't deal with a switch.
2151 BranchInst *PredBBBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
2152 if (!PredBBBranch || PredBBBranch->isUnconditional())
2153 return false;
2154
2155 // If PredBB has exactly one incoming edge, we don't gain anything by copying
2156 // PredBB.
2157 if (PredBB->getSinglePredecessor())
2158 return false;
2159
2160 // Don't thread through PredBB if it contains a successor edge to itself, in
2161 // which case we would infinite loop. Suppose we are threading an edge from
2162 // PredPredBB through PredBB and BB to SuccBB with PredBB containing a
2163 // successor edge to itself. If we allowed jump threading in this case, we
2164 // could duplicate PredBB and BB as, say, PredBB.thread and BB.thread. Since
2165 // PredBB.thread has a successor edge to PredBB, we would immediately come up
2166 // with another jump threading opportunity from PredBB.thread through PredBB
2167 // and BB to SuccBB. This jump threading would repeatedly occur. That is, we
2168 // would keep peeling one iteration from PredBB.
2169 if (llvm::is_contained(successors(PredBB), PredBB))
2170 return false;
2171
2172 // Don't thread across a loop header.
2173 if (LoopHeaders.count(PredBB))
2174 return false;
2175
2176 // Avoid complication with duplicating EH pads.
2177 if (PredBB->isEHPad())
2178 return false;
2179
2180 // Find a predecessor that we can thread. For simplicity, we only consider a
2181 // successor edge out of BB to which we thread exactly one incoming edge into
2182 // PredBB.
2183 unsigned ZeroCount = 0;
2184 unsigned OneCount = 0;
2185 BasicBlock *ZeroPred = nullptr;
2186 BasicBlock *OnePred = nullptr;
2187 for (BasicBlock *P : predecessors(PredBB)) {
2188 if (ConstantInt *CI = dyn_cast_or_null<ConstantInt>(
2189 evaluateOnPredecessorEdge(BB, P, Cond))) {
2190 if (CI->isZero()) {
2191 ZeroCount++;
2192 ZeroPred = P;
2193 } else if (CI->isOne()) {
2194 OneCount++;
2195 OnePred = P;
2196 }
2197 }
2198 }
2199
2200 // Disregard complicated cases where we have to thread multiple edges.
2201 BasicBlock *PredPredBB;
2202 if (ZeroCount == 1) {
2203 PredPredBB = ZeroPred;
2204 } else if (OneCount == 1) {
2205 PredPredBB = OnePred;
2206 } else {
2207 return false;
2208 }
2209
2210 BasicBlock *SuccBB = CondBr->getSuccessor(PredPredBB == ZeroPred);
2211
2212 // If threading to the same block as we come from, we would infinite loop.
2213 if (SuccBB == BB) {
2214 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)
2215 << "' - 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)
;
2216 return false;
2217 }
2218
2219 // If threading this would thread across a loop header, don't thread the edge.
2220 // See the comments above findLoopHeaders for justifications and caveats.
2221 if (LoopHeaders.count(BB) || LoopHeaders.count(SuccBB)) {
2222 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)
2223 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)
2224 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)
2225 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)
2226 << (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)
2227 << 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)
2228 << (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)
2229 << 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)
2230 << "' - 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)
2231 })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)
;
2232 return false;
2233 }
2234
2235 // Compute the cost of duplicating BB and PredBB.
2236 unsigned BBCost =
2237 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
2238 unsigned PredBBCost = getJumpThreadDuplicationCost(
2239 PredBB, PredBB->getTerminator(), BBDupThreshold);
2240
2241 // Give up if costs are too high. We need to check BBCost and PredBBCost
2242 // individually before checking their sum because getJumpThreadDuplicationCost
2243 // return (unsigned)~0 for those basic blocks that cannot be duplicated.
2244 if (BBCost > BBDupThreshold || PredBBCost > BBDupThreshold ||
2245 BBCost + PredBBCost > BBDupThreshold) {
2246 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)
2247 << "' - 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)
2248 << " 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)
;
2249 return false;
2250 }
2251
2252 // Now we are ready to duplicate PredBB.
2253 threadThroughTwoBasicBlocks(PredPredBB, PredBB, BB, SuccBB);
2254 return true;
2255}
2256
2257void JumpThreadingPass::threadThroughTwoBasicBlocks(BasicBlock *PredPredBB,
2258 BasicBlock *PredBB,
2259 BasicBlock *BB,
2260 BasicBlock *SuccBB) {
2261 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)
2262 << BB->getName() << "'\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Threading through '"
<< PredBB->getName() << "' and '" << BB
->getName() << "'\n"; } } while (false)
;
2263
2264 BranchInst *CondBr = cast<BranchInst>(BB->getTerminator());
2265 BranchInst *PredBBBranch = cast<BranchInst>(PredBB->getTerminator());
2266
2267 BasicBlock *NewBB =
2268 BasicBlock::Create(PredBB->getContext(), PredBB->getName() + ".thread",
2269 PredBB->getParent(), PredBB);
2270 NewBB->moveAfter(PredBB);
2271
2272 // Set the block frequency of NewBB.
2273 if (HasProfileData) {
2274 auto NewBBFreq = BFI->getBlockFreq(PredPredBB) *
2275 BPI->getEdgeProbability(PredPredBB, PredBB);
2276 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
2277 }
2278
2279 // We are going to have to map operands from the original BB block to the new
2280 // copy of the block 'NewBB'. If there are PHI nodes in PredBB, evaluate them
2281 // to account for entry from PredPredBB.
2282 DenseMap<Instruction *, Value *> ValueMapping =
2283 cloneInstructions(PredBB->begin(), PredBB->end(), NewBB, PredPredBB);
2284
2285 // Copy the edge probabilities from PredBB to NewBB.
2286 if (HasProfileData)
2287 BPI->copyEdgeProbabilities(PredBB, NewBB);
2288
2289 // Update the terminator of PredPredBB to jump to NewBB instead of PredBB.
2290 // This eliminates predecessors from PredPredBB, which requires us to simplify
2291 // any PHI nodes in PredBB.
2292 Instruction *PredPredTerm = PredPredBB->getTerminator();
2293 for (unsigned i = 0, e = PredPredTerm->getNumSuccessors(); i != e; ++i)
2294 if (PredPredTerm->getSuccessor(i) == PredBB) {
2295 PredBB->removePredecessor(PredPredBB, true);
2296 PredPredTerm->setSuccessor(i, NewBB);
2297 }
2298
2299 addPHINodeEntriesForMappedBlock(PredBBBranch->getSuccessor(0), PredBB, NewBB,
2300 ValueMapping);
2301 addPHINodeEntriesForMappedBlock(PredBBBranch->getSuccessor(1), PredBB, NewBB,
2302 ValueMapping);
2303
2304 DTU->applyUpdatesPermissive(
2305 {{DominatorTree::Insert, NewBB, CondBr->getSuccessor(0)},
2306 {DominatorTree::Insert, NewBB, CondBr->getSuccessor(1)},
2307 {DominatorTree::Insert, PredPredBB, NewBB},
2308 {DominatorTree::Delete, PredPredBB, PredBB}});
2309
2310 updateSSA(PredBB, NewBB, ValueMapping);
2311
2312 // Clean up things like PHI nodes with single operands, dead instructions,
2313 // etc.
2314 SimplifyInstructionsInBlock(NewBB, TLI);
2315 SimplifyInstructionsInBlock(PredBB, TLI);
2316
2317 SmallVector<BasicBlock *, 1> PredsToFactor;
2318 PredsToFactor.push_back(NewBB);
2319 threadEdge(BB, PredsToFactor, SuccBB);
2320}
2321
2322/// tryThreadEdge - Thread an edge if it's safe and profitable to do so.
2323bool JumpThreadingPass::tryThreadEdge(
2324 BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs,
2325 BasicBlock *SuccBB) {
2326 // If threading to the same block as we come from, we would infinite loop.
2327 if (SuccBB == BB) {
2328 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)
2329 << "' - 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)
;
2330 return false;
2331 }
2332
2333 // If threading this would thread across a loop header, don't thread the edge.
2334 // See the comments above findLoopHeaders for justifications and caveats.
2335 if (LoopHeaders.count(BB) || LoopHeaders.count(SuccBB)) {
2336 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)
2337 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)
2338 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)
2339 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)
2340 << (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)
2341 << "' 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)
2342 << 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)
2343 })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)
;
2344 return false;
2345 }
2346
2347 unsigned JumpThreadCost =
2348 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
2349 if (JumpThreadCost > BBDupThreshold) {
2350 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)
2351 << "' - 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)
;
2352 return false;
2353 }
2354
2355 threadEdge(BB, PredBBs, SuccBB);
2356 return true;
2357}
2358
2359/// threadEdge - We have decided that it is safe and profitable to factor the
2360/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
2361/// across BB. Transform the IR to reflect this change.
2362void JumpThreadingPass::threadEdge(BasicBlock *BB,
2363 const SmallVectorImpl<BasicBlock *> &PredBBs,
2364 BasicBlock *SuccBB) {
2365 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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 2365, __PRETTY_FUNCTION__))
;
2366
2367 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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 2368, __PRETTY_FUNCTION__))
2368 "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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 2368, __PRETTY_FUNCTION__))
;
2369
2370 // And finally, do it! Start by factoring the predecessors if needed.
2371 BasicBlock *PredBB;
2372 if (PredBBs.size() == 1)
2373 PredBB = PredBBs[0];
2374 else {
2375 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)
2376 << " common predecessors.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Factoring out " <<
PredBBs.size() << " common predecessors.\n"; } } while
(false)
;
2377 PredBB = splitBlockPreds(BB, PredBBs, ".thr_comm");
2378 }
2379
2380 // And finally, do it!
2381 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)
2382 << "' 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)
2383 << ", 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)
;
2384
2385 LVI->threadEdge(PredBB, BB, SuccBB);
2386
2387 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
2388 BB->getName()+".thread",
2389 BB->getParent(), BB);
2390 NewBB->moveAfter(PredBB);
2391
2392 // Set the block frequency of NewBB.
2393 if (HasProfileData) {
2394 auto NewBBFreq =
2395 BFI->getBlockFreq(PredBB) * BPI->getEdgeProbability(PredBB, BB);
2396 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
2397 }
2398
2399 // Copy all the instructions from BB to NewBB except the terminator.
2400 DenseMap<Instruction *, Value *> ValueMapping =
2401 cloneInstructions(BB->begin(), std::prev(BB->end()), NewBB, PredBB);
2402
2403 // We didn't copy the terminator from BB over to NewBB, because there is now
2404 // an unconditional jump to SuccBB. Insert the unconditional jump.
2405 BranchInst *NewBI = BranchInst::Create(SuccBB, NewBB);
2406 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
2407
2408 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
2409 // PHI nodes for NewBB now.
2410 addPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
2411
2412 // Update the terminator of PredBB to jump to NewBB instead of BB. This
2413 // eliminates predecessors from BB, which requires us to simplify any PHI
2414 // nodes in BB.
2415 Instruction *PredTerm = PredBB->getTerminator();
2416 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
2417 if (PredTerm->getSuccessor(i) == BB) {
2418 BB->removePredecessor(PredBB, true);
2419 PredTerm->setSuccessor(i, NewBB);
2420 }
2421
2422 // Enqueue required DT updates.
2423 DTU->applyUpdatesPermissive({{DominatorTree::Insert, NewBB, SuccBB},
2424 {DominatorTree::Insert, PredBB, NewBB},
2425 {DominatorTree::Delete, PredBB, BB}});
2426
2427 updateSSA(BB, NewBB, ValueMapping);
2428
2429 // At this point, the IR is fully up to date and consistent. Do a quick scan
2430 // over the new instructions and zap any that are constants or dead. This
2431 // frequently happens because of phi translation.
2432 SimplifyInstructionsInBlock(NewBB, TLI);
2433
2434 // Update the edge weight from BB to SuccBB, which should be less than before.
2435 updateBlockFreqAndEdgeWeight(PredBB, BB, NewBB, SuccBB);
2436
2437 // Threaded an edge!
2438 ++NumThreads;
2439}
2440
2441/// Create a new basic block that will be the predecessor of BB and successor of
2442/// all blocks in Preds. When profile data is available, update the frequency of
2443/// this new block.
2444BasicBlock *JumpThreadingPass::splitBlockPreds(BasicBlock *BB,
2445 ArrayRef<BasicBlock *> Preds,
2446 const char *Suffix) {
2447 SmallVector<BasicBlock *, 2> NewBBs;
2448
2449 // Collect the frequencies of all predecessors of BB, which will be used to
2450 // update the edge weight of the result of splitting predecessors.
2451 DenseMap<BasicBlock *, BlockFrequency> FreqMap;
2452 if (HasProfileData)
2453 for (auto Pred : Preds)
2454 FreqMap.insert(std::make_pair(
2455 Pred, BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, BB)));
2456
2457 // In the case when BB is a LandingPad block we create 2 new predecessors
2458 // instead of just one.
2459 if (BB->isLandingPad()) {
2460 std::string NewName = std::string(Suffix) + ".split-lp";
2461 SplitLandingPadPredecessors(BB, Preds, Suffix, NewName.c_str(), NewBBs);
2462 } else {
2463 NewBBs.push_back(SplitBlockPredecessors(BB, Preds, Suffix));
2464 }
2465
2466 std::vector<DominatorTree::UpdateType> Updates;
2467 Updates.reserve((2 * Preds.size()) + NewBBs.size());
2468 for (auto NewBB : NewBBs) {
2469 BlockFrequency NewBBFreq(0);
2470 Updates.push_back({DominatorTree::Insert, NewBB, BB});
2471 for (auto Pred : predecessors(NewBB)) {
2472 Updates.push_back({DominatorTree::Delete, Pred, BB});
2473 Updates.push_back({DominatorTree::Insert, Pred, NewBB});
2474 if (HasProfileData) // Update frequencies between Pred -> NewBB.
2475 NewBBFreq += FreqMap.lookup(Pred);
2476 }
2477 if (HasProfileData) // Apply the summed frequency to NewBB.
2478 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
2479 }
2480
2481 DTU->applyUpdatesPermissive(Updates);
2482 return NewBBs[0];
2483}
2484
2485bool JumpThreadingPass::doesBlockHaveProfileData(BasicBlock *BB) {
2486 const Instruction *TI = BB->getTerminator();
2487 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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 2487, __PRETTY_FUNCTION__))
;
2488
2489 MDNode *WeightsNode = TI->getMetadata(LLVMContext::MD_prof);
2490 if (!WeightsNode)
2491 return false;
2492
2493 MDString *MDName = cast<MDString>(WeightsNode->getOperand(0));
2494 if (MDName->getString() != "branch_weights")
2495 return false;
2496
2497 // Ensure there are weights for all of the successors. Note that the first
2498 // operand to the metadata node is a name, not a weight.
2499 return WeightsNode->getNumOperands() == TI->getNumSuccessors() + 1;
2500}
2501
2502/// Update the block frequency of BB and branch weight and the metadata on the
2503/// edge BB->SuccBB. This is done by scaling the weight of BB->SuccBB by 1 -
2504/// Freq(PredBB->BB) / Freq(BB->SuccBB).
2505void JumpThreadingPass::updateBlockFreqAndEdgeWeight(BasicBlock *PredBB,
2506 BasicBlock *BB,
2507 BasicBlock *NewBB,
2508 BasicBlock *SuccBB) {
2509 if (!HasProfileData)
2510 return;
2511
2512 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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 2512, __PRETTY_FUNCTION__))
;
2513
2514 // As the edge from PredBB to BB is deleted, we have to update the block
2515 // frequency of BB.
2516 auto BBOrigFreq = BFI->getBlockFreq(BB);
2517 auto NewBBFreq = BFI->getBlockFreq(NewBB);
2518 auto BB2SuccBBFreq = BBOrigFreq * BPI->getEdgeProbability(BB, SuccBB);
2519 auto BBNewFreq = BBOrigFreq - NewBBFreq;
2520 BFI->setBlockFreq(BB, BBNewFreq.getFrequency());
2521
2522 // Collect updated outgoing edges' frequencies from BB and use them to update
2523 // edge probabilities.
2524 SmallVector<uint64_t, 4> BBSuccFreq;
2525 for (BasicBlock *Succ : successors(BB)) {
2526 auto SuccFreq = (Succ == SuccBB)
2527 ? BB2SuccBBFreq - NewBBFreq
2528 : BBOrigFreq * BPI->getEdgeProbability(BB, Succ);
2529 BBSuccFreq.push_back(SuccFreq.getFrequency());
2530 }
2531
2532 uint64_t MaxBBSuccFreq =
2533 *std::max_element(BBSuccFreq.begin(), BBSuccFreq.end());
2534
2535 SmallVector<BranchProbability, 4> BBSuccProbs;
2536 if (MaxBBSuccFreq == 0)
2537 BBSuccProbs.assign(BBSuccFreq.size(),
2538 {1, static_cast<unsigned>(BBSuccFreq.size())});
2539 else {
2540 for (uint64_t Freq : BBSuccFreq)
2541 BBSuccProbs.push_back(
2542 BranchProbability::getBranchProbability(Freq, MaxBBSuccFreq));
2543 // Normalize edge probabilities so that they sum up to one.
2544 BranchProbability::normalizeProbabilities(BBSuccProbs.begin(),
2545 BBSuccProbs.end());
2546 }
2547
2548 // Update edge probabilities in BPI.
2549 BPI->setEdgeProbability(BB, BBSuccProbs);
2550
2551 // Update the profile metadata as well.
2552 //
2553 // Don't do this if the profile of the transformed blocks was statically
2554 // estimated. (This could occur despite the function having an entry
2555 // frequency in completely cold parts of the CFG.)
2556 //
2557 // In this case we don't want to suggest to subsequent passes that the
2558 // calculated weights are fully consistent. Consider this graph:
2559 //
2560 // check_1
2561 // 50% / |
2562 // eq_1 | 50%
2563 // \ |
2564 // check_2
2565 // 50% / |
2566 // eq_2 | 50%
2567 // \ |
2568 // check_3
2569 // 50% / |
2570 // eq_3 | 50%
2571 // \ |
2572 //
2573 // Assuming the blocks check_* all compare the same value against 1, 2 and 3,
2574 // the overall probabilities are inconsistent; the total probability that the
2575 // value is either 1, 2 or 3 is 150%.
2576 //
2577 // As a consequence if we thread eq_1 -> check_2 to check_3, check_2->check_3
2578 // becomes 0%. This is even worse if the edge whose probability becomes 0% is
2579 // the loop exit edge. Then based solely on static estimation we would assume
2580 // the loop was extremely hot.
2581 //
2582 // FIXME this locally as well so that BPI and BFI are consistent as well. We
2583 // shouldn't make edges extremely likely or unlikely based solely on static
2584 // estimation.
2585 if (BBSuccProbs.size() >= 2 && doesBlockHaveProfileData(BB)) {
2586 SmallVector<uint32_t, 4> Weights;
2587 for (auto Prob : BBSuccProbs)
2588 Weights.push_back(Prob.getNumerator());
2589
2590 auto TI = BB->getTerminator();
2591 TI->setMetadata(
2592 LLVMContext::MD_prof,
2593 MDBuilder(TI->getParent()->getContext()).createBranchWeights(Weights));
2594 }
2595}
2596
2597/// duplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
2598/// to BB which contains an i1 PHI node and a conditional branch on that PHI.
2599/// If we can duplicate the contents of BB up into PredBB do so now, this
2600/// improves the odds that the branch will be on an analyzable instruction like
2601/// a compare.
2602bool JumpThreadingPass::duplicateCondBranchOnPHIIntoPred(
2603 BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs) {
2604 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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 2604, __PRETTY_FUNCTION__))
;
2605
2606 // If BB is a loop header, then duplicating this block outside the loop would
2607 // cause us to transform this into an irreducible loop, don't do this.
2608 // See the comments above findLoopHeaders for justifications and caveats.
2609 if (LoopHeaders.count(BB)) {
2610 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)
2611 << "' 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)
2612 << "' - 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)
;
2613 return false;
2614 }
2615
2616 unsigned DuplicationCost =
2617 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
2618 if (DuplicationCost > BBDupThreshold) {
2619 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)
2620 << "' - 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)
;
2621 return false;
2622 }
2623
2624 // And finally, do it! Start by factoring the predecessors if needed.
2625 std::vector<DominatorTree::UpdateType> Updates;
2626 BasicBlock *PredBB;
2627 if (PredBBs.size() == 1)
2628 PredBB = PredBBs[0];
2629 else {
2630 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)
2631 << " common predecessors.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Factoring out " <<
PredBBs.size() << " common predecessors.\n"; } } while
(false)
;
2632 PredBB = splitBlockPreds(BB, PredBBs, ".thr_comm");
2633 }
2634 Updates.push_back({DominatorTree::Delete, PredBB, BB});
2635
2636 // Okay, we decided to do this! Clone all the instructions in BB onto the end
2637 // of PredBB.
2638 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)
2639 << "' 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)
2640 << "' 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)
2641 << 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)
;
2642
2643 // Unless PredBB ends with an unconditional branch, split the edge so that we
2644 // can just clone the bits from BB into the end of the new PredBB.
2645 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
2646
2647 if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
2648 BasicBlock *OldPredBB = PredBB;
2649 PredBB = SplitEdge(OldPredBB, BB);
2650 Updates.push_back({DominatorTree::Insert, OldPredBB, PredBB});
2651 Updates.push_back({DominatorTree::Insert, PredBB, BB});
2652 Updates.push_back({DominatorTree::Delete, OldPredBB, BB});
2653 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
2654 }
2655
2656 // We are going to have to map operands from the original BB block into the
2657 // PredBB block. Evaluate PHI nodes in BB.
2658 DenseMap<Instruction*, Value*> ValueMapping;
2659
2660 BasicBlock::iterator BI = BB->begin();
2661 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
2662 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
2663 // Clone the non-phi instructions of BB into PredBB, keeping track of the
2664 // mapping and using it to remap operands in the cloned instructions.
2665 for (; BI != BB->end(); ++BI) {
2666 Instruction *New = BI->clone();
2667
2668 // Remap operands to patch up intra-block references.
2669 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2670 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
2671 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
2672 if (I != ValueMapping.end())
2673 New->setOperand(i, I->second);
2674 }
2675
2676 // If this instruction can be simplified after the operands are updated,
2677 // just use the simplified value instead. This frequently happens due to
2678 // phi translation.
2679 if (Value *IV = SimplifyInstruction(
2680 New,
2681 {BB->getModule()->getDataLayout(), TLI, nullptr, nullptr, New})) {
2682 ValueMapping[&*BI] = IV;
2683 if (!New->mayHaveSideEffects()) {
2684 New->deleteValue();
2685 New = nullptr;
2686 }
2687 } else {
2688 ValueMapping[&*BI] = New;
2689 }
2690 if (New) {
2691 // Otherwise, insert the new instruction into the block.
2692 New->setName(BI->getName());
2693 PredBB->getInstList().insert(OldPredBranch->getIterator(), New);
2694 // Update Dominance from simplified New instruction operands.
2695 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2696 if (BasicBlock *SuccBB = dyn_cast<BasicBlock>(New->getOperand(i)))
2697 Updates.push_back({DominatorTree::Insert, PredBB, SuccBB});
2698 }
2699 }
2700
2701 // Check to see if the targets of the branch had PHI nodes. If so, we need to
2702 // add entries to the PHI nodes for branch from PredBB now.
2703 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
2704 addPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
2705 ValueMapping);
2706 addPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
2707 ValueMapping);
2708
2709 updateSSA(BB, PredBB, ValueMapping);
2710
2711 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
2712 // that we nuked.
2713 BB->removePredecessor(PredBB, true);
2714
2715 // Remove the unconditional branch at the end of the PredBB block.
2716 OldPredBranch->eraseFromParent();
2717 if (HasProfileData)
2718 BPI->copyEdgeProbabilities(BB, PredBB);
2719 DTU->applyUpdatesPermissive(Updates);
2720
2721 ++NumDupes;
2722 return true;
2723}
2724
2725// Pred is a predecessor of BB with an unconditional branch to BB. SI is
2726// a Select instruction in Pred. BB has other predecessors and SI is used in
2727// a PHI node in BB. SI has no other use.
2728// A new basic block, NewBB, is created and SI is converted to compare and
2729// conditional branch. SI is erased from parent.
2730void JumpThreadingPass::unfoldSelectInstr(BasicBlock *Pred, BasicBlock *BB,
2731 SelectInst *SI, PHINode *SIUse,
2732 unsigned Idx) {
2733 // Expand the select.
2734 //
2735 // Pred --
2736 // | v
2737 // | NewBB
2738 // | |
2739 // |-----
2740 // v
2741 // BB
2742 BranchInst *PredTerm = cast<BranchInst>(Pred->getTerminator());
2743 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
2744 BB->getParent(), BB);
2745 // Move the unconditional branch to NewBB.
2746 PredTerm->removeFromParent();
2747 NewBB->getInstList().insert(NewBB->end(), PredTerm);
2748 // Create a conditional branch and update PHI nodes.
2749 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
2750 SIUse->setIncomingValue(Idx, SI->getFalseValue());
2751 SIUse->addIncoming(SI->getTrueValue(), NewBB);
2752
2753 // The select is now dead.
2754 SI->eraseFromParent();
2755 DTU->applyUpdatesPermissive({{DominatorTree::Insert, NewBB, BB},
2756 {DominatorTree::Insert, Pred, NewBB}});
2757
2758 // Update any other PHI nodes in BB.
2759 for (BasicBlock::iterator BI = BB->begin();
2760 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
2761 if (Phi != SIUse)
2762 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);
2763}
2764
2765bool JumpThreadingPass::tryToUnfoldSelect(SwitchInst *SI, BasicBlock *BB) {
2766 PHINode *CondPHI = dyn_cast<PHINode>(SI->getCondition());
2767
2768 if (!CondPHI || CondPHI->getParent() != BB)
2769 return false;
2770
2771 for (unsigned I = 0, E = CondPHI->getNumIncomingValues(); I != E; ++I) {
2772 BasicBlock *Pred = CondPHI->getIncomingBlock(I);
2773 SelectInst *PredSI = dyn_cast<SelectInst>(CondPHI->getIncomingValue(I));
2774
2775 // The second and third condition can be potentially relaxed. Currently
2776 // the conditions help to simplify the code and allow us to reuse existing
2777 // code, developed for tryToUnfoldSelect(CmpInst *, BasicBlock *)
2778 if (!PredSI || PredSI->getParent() != Pred || !PredSI->hasOneUse())
2779 continue;
2780
2781 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
2782 if (!PredTerm || !PredTerm->isUnconditional())
2783 continue;
2784
2785 unfoldSelectInstr(Pred, BB, PredSI, CondPHI, I);
2786 return true;
2787 }
2788 return false;
2789}
2790
2791/// tryToUnfoldSelect - Look for blocks of the form
2792/// bb1:
2793/// %a = select
2794/// br bb2
2795///
2796/// bb2:
2797/// %p = phi [%a, %bb1] ...
2798/// %c = icmp %p
2799/// br i1 %c
2800///
2801/// And expand the select into a branch structure if one of its arms allows %c
2802/// to be folded. This later enables threading from bb1 over bb2.
2803bool JumpThreadingPass::tryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
2804 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
2805 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
2806 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
2807
2808 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
2809 CondLHS->getParent() != BB)
2810 return false;
2811
2812 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
2813 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
2814 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
2815
2816 // Look if one of the incoming values is a select in the corresponding
2817 // predecessor.
2818 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
2819 continue;
2820
2821 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
2822 if (!PredTerm || !PredTerm->isUnconditional())
2823 continue;
2824
2825 // Now check if one of the select values would allow us to constant fold the
2826 // terminator in BB. We don't do the transform if both sides fold, those
2827 // cases will be threaded in any case.
2828 LazyValueInfo::Tristate LHSFolds =
2829 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
2830 CondRHS, Pred, BB, CondCmp);
2831 LazyValueInfo::Tristate RHSFolds =
2832 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
2833 CondRHS, Pred, BB, CondCmp);
2834 if ((LHSFolds != LazyValueInfo::Unknown ||
2835 RHSFolds != LazyValueInfo::Unknown) &&
2836 LHSFolds != RHSFolds) {
2837 unfoldSelectInstr(Pred, BB, SI, CondLHS, I);
2838 return true;
2839 }
2840 }
2841 return false;
2842}
2843
2844/// tryToUnfoldSelectInCurrBB - Look for PHI/Select or PHI/CMP/Select in the
2845/// same BB in the form
2846/// bb:
2847/// %p = phi [false, %bb1], [true, %bb2], [false, %bb3], [true, %bb4], ...
2848/// %s = select %p, trueval, falseval
2849///
2850/// or
2851///
2852/// bb:
2853/// %p = phi [0, %bb1], [1, %bb2], [0, %bb3], [1, %bb4], ...
2854/// %c = cmp %p, 0
2855/// %s = select %c, trueval, falseval
2856///
2857/// And expand the select into a branch structure. This later enables
2858/// jump-threading over bb in this pass.
2859///
2860/// Using the similar approach of SimplifyCFG::FoldCondBranchOnPHI(), unfold
2861/// select if the associated PHI has at least one constant. If the unfolded
2862/// select is not jump-threaded, it will be folded again in the later
2863/// optimizations.
2864bool JumpThreadingPass::tryToUnfoldSelectInCurrBB(BasicBlock *BB) {
2865 // This transform would reduce the quality of msan diagnostics.
2866 // Disable this transform under MemorySanitizer.
2867 if (BB->getParent()->hasFnAttribute(Attribute::SanitizeMemory))
2868 return false;
2869
2870 // If threading this would thread across a loop header, don't thread the edge.
2871 // See the comments above findLoopHeaders for justifications and caveats.
2872 if (LoopHeaders.count(BB))
2873 return false;
2874
2875 for (BasicBlock::iterator BI = BB->begin();
2876 PHINode *PN = dyn_cast<PHINode>(BI); ++BI) {
2877 // Look for a Phi having at least one constant incoming value.
2878 if (llvm::all_of(PN->incoming_values(),
2879 [](Value *V) { return !isa<ConstantInt>(V); }))
2880 continue;
2881
2882 auto isUnfoldCandidate = [BB](SelectInst *SI, Value *V) {
2883 using namespace PatternMatch;
2884
2885 // Check if SI is in BB and use V as condition.
2886 if (SI->getParent() != BB)
2887 return false;
2888 Value *Cond = SI->getCondition();
2889 bool IsAndOr = match(SI, m_CombineOr(m_LogicalAnd(), m_LogicalOr()));
2890 return Cond && Cond == V && Cond->getType()->isIntegerTy(1) && !IsAndOr;
2891 };
2892
2893 SelectInst *SI = nullptr;
2894 for (Use &U : PN->uses()) {
2895 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(U.getUser())) {
2896 // Look for a ICmp in BB that compares PN with a constant and is the
2897 // condition of a Select.
2898 if (Cmp->getParent() == BB && Cmp->hasOneUse() &&
2899 isa<ConstantInt>(Cmp->getOperand(1 - U.getOperandNo())))
2900 if (SelectInst *SelectI = dyn_cast<SelectInst>(Cmp->user_back()))
2901 if (isUnfoldCandidate(SelectI, Cmp->use_begin()->get())) {
2902 SI = SelectI;
2903 break;
2904 }
2905 } else if (SelectInst *SelectI = dyn_cast<SelectInst>(U.getUser())) {
2906 // Look for a Select in BB that uses PN as condition.
2907 if (isUnfoldCandidate(SelectI, U.get())) {
2908 SI = SelectI;
2909 break;
2910 }
2911 }
2912 }
2913
2914 if (!SI)
2915 continue;
2916 // Expand the select.
2917 Value *Cond = SI->getCondition();
2918 if (InsertFreezeWhenUnfoldingSelect &&
2919 !isGuaranteedNotToBeUndefOrPoison(Cond, nullptr, SI,
2920 &DTU->getDomTree()))
2921 Cond = new FreezeInst(Cond, "cond.fr", SI);
2922 Instruction *Term = SplitBlockAndInsertIfThen(Cond, SI, false);
2923 BasicBlock *SplitBB = SI->getParent();
2924 BasicBlock *NewBB = Term->getParent();
2925 PHINode *NewPN = PHINode::Create(SI->getType(), 2, "", SI);
2926 NewPN->addIncoming(SI->getTrueValue(), Term->getParent());
2927 NewPN->addIncoming(SI->getFalseValue(), BB);
2928 SI->replaceAllUsesWith(NewPN);
2929 SI->eraseFromParent();
2930 // NewBB and SplitBB are newly created blocks which require insertion.
2931 std::vector<DominatorTree::UpdateType> Updates;
2932 Updates.reserve((2 * SplitBB->getTerminator()->getNumSuccessors()) + 3);
2933 Updates.push_back({DominatorTree::Insert, BB, SplitBB});
2934 Updates.push_back({DominatorTree::Insert, BB, NewBB});
2935 Updates.push_back({DominatorTree::Insert, NewBB, SplitBB});
2936 // BB's successors were moved to SplitBB, update DTU accordingly.
2937 for (auto *Succ : successors(SplitBB)) {
2938 Updates.push_back({DominatorTree::Delete, BB, Succ});
2939 Updates.push_back({DominatorTree::Insert, SplitBB, Succ});
2940 }
2941 DTU->applyUpdatesPermissive(Updates);
2942 return true;
2943 }
2944 return false;
2945}
2946
2947/// Try to propagate a guard from the current BB into one of its predecessors
2948/// in case if another branch of execution implies that the condition of this
2949/// guard is always true. Currently we only process the simplest case that
2950/// looks like:
2951///
2952/// Start:
2953/// %cond = ...
2954/// br i1 %cond, label %T1, label %F1
2955/// T1:
2956/// br label %Merge
2957/// F1:
2958/// br label %Merge
2959/// Merge:
2960/// %condGuard = ...
2961/// call void(i1, ...) @llvm.experimental.guard( i1 %condGuard )[ "deopt"() ]
2962///
2963/// And cond either implies condGuard or !condGuard. In this case all the
2964/// instructions before the guard can be duplicated in both branches, and the
2965/// guard is then threaded to one of them.
2966bool JumpThreadingPass::processGuards(BasicBlock *BB) {
2967 using namespace PatternMatch;
2968
2969 // We only want to deal with two predecessors.
2970 BasicBlock *Pred1, *Pred2;
2971 auto PI = pred_begin(BB), PE = pred_end(BB);
2972 if (PI == PE)
2973 return false;
2974 Pred1 = *PI++;
2975 if (PI == PE)
2976 return false;
2977 Pred2 = *PI++;
2978 if (PI != PE)
2979 return false;
2980 if (Pred1 == Pred2)
2981 return false;
2982
2983 // Try to thread one of the guards of the block.
2984 // TODO: Look up deeper than to immediate predecessor?
2985 auto *Parent = Pred1->getSinglePredecessor();
2986 if (!Parent || Parent != Pred2->getSinglePredecessor())
2987 return false;
2988
2989 if (auto *BI = dyn_cast<BranchInst>(Parent->getTerminator()))
2990 for (auto &I : *BB)
2991 if (isGuard(&I) && threadGuard(BB, cast<IntrinsicInst>(&I), BI))
2992 return true;
2993
2994 return false;
2995}
2996
2997/// Try to propagate the guard from BB which is the lower block of a diamond
2998/// to one of its branches, in case if diamond's condition implies guard's
2999/// condition.
3000bool JumpThreadingPass::threadGuard(BasicBlock *BB, IntrinsicInst *Guard,
3001 BranchInst *BI) {
3002 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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 3002, __PRETTY_FUNCTION__))
;
3003 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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 3003, __PRETTY_FUNCTION__))
;
3004 Value *GuardCond = Guard->getArgOperand(0);
3005 Value *BranchCond = BI->getCondition();
3006 BasicBlock *TrueDest = BI->getSuccessor(0);
3007 BasicBlock *FalseDest = BI->getSuccessor(1);
3008
3009 auto &DL = BB->getModule()->getDataLayout();
3010 bool TrueDestIsSafe = false;
3011 bool FalseDestIsSafe = false;
3012
3013 // True dest is safe if BranchCond => GuardCond.
3014 auto Impl = isImpliedCondition(BranchCond, GuardCond, DL);
3015 if (Impl && *Impl)
3016 TrueDestIsSafe = true;
3017 else {
3018 // False dest is safe if !BranchCond => GuardCond.
3019 Impl = isImpliedCondition(BranchCond, GuardCond, DL, /* LHSIsTrue */ false);
3020 if (Impl && *Impl)
3021 FalseDestIsSafe = true;
3022 }
3023
3024 if (!TrueDestIsSafe && !FalseDestIsSafe)
3025 return false;
3026
3027 BasicBlock *PredUnguardedBlock = TrueDestIsSafe ? TrueDest : FalseDest;
3028 BasicBlock *PredGuardedBlock = FalseDestIsSafe ? TrueDest : FalseDest;
3029
3030 ValueToValueMapTy UnguardedMapping, GuardedMapping;
3031 Instruction *AfterGuard = Guard->getNextNode();
3032 unsigned Cost = getJumpThreadDuplicationCost(BB, AfterGuard, BBDupThreshold);
3033 if (Cost > BBDupThreshold)
3034 return false;
3035 // Duplicate all instructions before the guard and the guard itself to the
3036 // branch where implication is not proved.
3037 BasicBlock *GuardedBlock = DuplicateInstructionsInSplitBetween(
3038 BB, PredGuardedBlock, AfterGuard, GuardedMapping, *DTU);
3039 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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 3039, __PRETTY_FUNCTION__))
;
3040 // Duplicate all instructions before the guard in the unguarded branch.
3041 // Since we have successfully duplicated the guarded block and this block
3042 // has fewer instructions, we expect it to succeed.
3043 BasicBlock *UnguardedBlock = DuplicateInstructionsInSplitBetween(
3044 BB, PredUnguardedBlock, Guard, UnguardedMapping, *DTU);
3045 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-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 3045, __PRETTY_FUNCTION__))
;
3046 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)
3047 << GuardedBlock->getName() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "Moved guard " <<
*Guard << " to block " << GuardedBlock->getName
() << "\n"; } } while (false)
;
3048 // Some instructions before the guard may still have uses. For them, we need
3049 // to create Phi nodes merging their copies in both guarded and unguarded
3050 // branches. Those instructions that have no uses can be just removed.
3051 SmallVector<Instruction *, 4> ToRemove;
3052 for (auto BI = BB->begin(); &*BI != AfterGuard; ++BI)
3053 if (!isa<PHINode>(&*BI))
3054 ToRemove.push_back(&*BI);
3055
3056 Instruction *InsertionPoint = &*BB->getFirstInsertionPt();
3057 assert(InsertionPoint && "Empty block?")((InsertionPoint && "Empty block?") ? static_cast<
void> (0) : __assert_fail ("InsertionPoint && \"Empty block?\""
, "/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/llvm/lib/Transforms/Scalar/JumpThreading.cpp"
, 3057, __PRETTY_FUNCTION__))
;
3058 // Substitute with Phis & remove.
3059 for (auto *Inst : reverse(ToRemove)) {
3060 if (!Inst->use_empty()) {
3061 PHINode *NewPN = PHINode::Create(Inst->getType(), 2);
3062 NewPN->addIncoming(UnguardedMapping[Inst], UnguardedBlock);
3063 NewPN->addIncoming(GuardedMapping[Inst], GuardedBlock);
3064 NewPN->insertBefore(InsertionPoint);
3065 Inst->replaceAllUsesWith(NewPN);
3066 }
3067 Inst->eraseFromParent();
3068 }
3069 return true;
3070}

/build/llvm-toolchain-snapshot-13~++20210405022414+5f57793c4fe4/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 false
5
Returning the value 1, which participates in a condition later
262 getOrdering() == AtomicOrdering::Unordered) &&
3
Assuming the condition is true
263 !isVolatile();
4
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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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-13~++20210405022414+5f57793c4fe4/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 /// Generate the IR for a call to malloc:
1589 /// 1. Compute the malloc call's argument as the specified type's size,
1590 /// possibly multiplied by the array size if the array size is not
1591 /// constant 1.
1592 /// 2. Call malloc with that argument.
1593 /// 3. Bitcast the result of the malloc call to the specified type.
1594 static Instruction *CreateMalloc(Instruction *InsertBefore, Type *IntPtrTy,
1595 Type *AllocTy, Value *AllocSize,
1596 Value *ArraySize = nullptr,
1597 Function *MallocF = nullptr,
1598 const Twine &Name = "");
1599 static Instruction *CreateMalloc(BasicBlock *InsertAtEnd, Type *IntPtrTy,
1600 Type *AllocTy, Value *AllocSize,
1601 Value *ArraySize = nullptr,
1602 Function *MallocF = nullptr,
1603 const Twine &Name = "");
1604 static Instruction *CreateMalloc(Instruction *InsertBefore, Type *IntPtrTy,
1605 Type *AllocTy, Value *AllocSize,
1606 Value *ArraySize = nullptr,
1607 ArrayRef<OperandBundleDef> Bundles = None,
1608 Function *MallocF = nullptr,
1609 const Twine &Name = "");
1610 static Instruction *CreateMalloc(BasicBlock *InsertAtEnd, Type *IntPtrTy,
1611 Type *AllocTy, Value *AllocSize,
1612 Value *ArraySize = nullptr,
1613 ArrayRef<OperandBundleDef> Bundles = None,
1614 Function *MallocF = nullptr,
1615 const Twine &Name = "");
1616 /// Generate the IR for a call to the builtin free function.
1617 static Instruction *CreateFree(Value *Source, Instruction *InsertBefore);
1618 static Instruction *CreateFree(Value *Source, BasicBlock *InsertAtEnd);
1619 static Instruction *CreateFree(Value *Source,
1620 ArrayRef<OperandBundleDef> Bundles,
1621 Instruction *InsertBefore);
1622 static Instruction *CreateFree(Value *Source,
1623 ArrayRef<OperandBundleDef> Bundles,
1624 BasicBlock *InsertAtEnd);
1625
1626 // Note that 'musttail' implies 'tail'.
1627 enum TailCallKind : unsigned {
1628 TCK_None = 0,
1629 TCK_Tail = 1,
1630 TCK_MustTail = 2,
1631 TCK_NoTail = 3,
1632 TCK_LAST = TCK_NoTail
1633 };
1634
1635 using TailCallKindField = Bitfield::Element<TailCallKind, 0, 2, TCK_LAST>;
1636 static_assert(
1637 Bitfield::areContiguous<TailCallKindField, CallBase::CallingConvField>(),
1638 "Bitfields must be contiguous");
1639
1640 TailCallKind getTailCallKind() const {
1641 return getSubclassData<TailCallKindField>();
1642 }
1643
1644 bool isTailCall() const {
1645 TailCallKind Kind = getTailCallKind();
1646 return Kind == TCK_Tail || Kind == TCK_MustTail;
1647 }
1648
1649 bool isMustTailCall() const { return getTailCallKind() == TCK_MustTail; }
1650
1651 bool isNoTailCall() const { return getTailCallKind() == TCK_NoTail; }
1652
1653 void setTailCallKind(TailCallKind TCK) {
1654 setSubclassData<TailCallKindField>(TCK);
1655 }
1656
1657 void setTailCall(bool IsTc = true) {
1658 setTailCallKind(IsTc ? TCK_Tail : TCK_None);
1659 }
1660
1661 /// Return true if the call can return twice
1662 bool canReturnTwice() const { return hasFnAttr(Attribute::ReturnsTwice); }
1663 void setCanReturnTwice() {
1664 addAttribute(AttributeList::FunctionIndex, Attribute::ReturnsTwice);
1665 }
1666
1667 // Methods for support type inquiry through isa, cast, and dyn_cast:
1668 static bool classof(const Instruction *I) {
1669 return I->getOpcode() == Instruction::Call;
1670 }
1671 static bool classof(const Value *V) {
1672 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1673 }
1674
1675 /// Updates profile metadata by scaling it by \p S / \p T.
1676 void updateProfWeight(uint64_t S, uint64_t T);
1677
1678private:
1679 // Shadow Instruction::setInstructionSubclassData with a private forwarding
1680 // method so that subclasses cannot accidentally use it.
1681 template <typename Bitfield>
1682 void setSubclassData(typename Bitfield::Type Value) {
1683 Instruction::setSubclassData<Bitfield>(Value);
1684 }
1685};
1686
1687CallInst::CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1688 ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr,
1689 BasicBlock *InsertAtEnd)
1690 : CallBase(Ty->getReturnType(), Instruction::Call,
1691 OperandTraits<CallBase>::op_end(this) -
1692 (Args.size() + CountBundleInputs(Bundles) + 1),
1693 unsigned(Args.size() + CountBundleInputs(Bundles) + 1),
1694 InsertAtEnd) {
1695 init(Ty, Func, Args, Bundles, NameStr);
1696}
1697
1698CallInst::CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1699 ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr,
1700 Instruction *InsertBefore)
1701 : CallBase(Ty->getReturnType(), Instruction::Call,
1702 OperandTraits<CallBase>::op_end(this) -
1703 (Args.size() + CountBundleInputs(Bundles) + 1),
1704 unsigned(Args.size() + CountBundleInputs(Bundles) + 1),
1705 InsertBefore) {
1706 init(Ty, Func, Args, Bundles, NameStr);
1707}
1708
1709//===----------------------------------------------------------------------===//
1710// SelectInst Class
1711//===----------------------------------------------------------------------===//
1712
1713/// This class represents the LLVM 'select' instruction.
1714///
1715class SelectInst : public Instruction {
1716 SelectInst(Value *C, Value *S1, Value *S2, const Twine &NameStr,
1717 Instruction *InsertBefore)
1718 : Instruction(S1->getType(), Instruction::Select,
1719 &Op<0>(), 3, InsertBefore) {
1720 init(C, S1, S2);
1721 setName(NameStr);
1722 }
1723
1724 SelectInst(Value *C, Value *S1, Value *S2, const Twine &NameStr,
1725 BasicBlock *InsertAtEnd)
1726 : Instruction(S1->getType(), Instruction::Select,
1727 &Op<0>(), 3, InsertAtEnd) {
1728 init(C, S1, S2);
1729 setName(NameStr);
1730 }
1731
1732 void init(Value *C, Value *S1, Value *S2) {
1733 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-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h"
, 1733, __PRETTY_FUNCTION__))
;
1734 Op<0>() = C;
1735 Op<1>() = S1;
1736 Op<2>() = S2;
1737 }
1738
1739protected:
1740 // Note: Instruction needs to be a friend here to call cloneImpl.
1741 friend class Instruction;
1742
1743 SelectInst *cloneImpl() const;
1744
1745public:
1746 static SelectInst *Create(Value *C, Value *S1, Value *S2,
1747 const Twine &NameStr = "",
1748 Instruction *InsertBefore = nullptr,
1749 Instruction *MDFrom = nullptr) {
1750 SelectInst *Sel = new(3) SelectInst(C, S1, S2, NameStr, InsertBefore);
1751 if (MDFrom)
1752 Sel->copyMetadata(*MDFrom);
1753 return Sel;
1754 }
1755
1756 static SelectInst *Create(Value *C, Value *S1, Value *S2,
1757 const Twine &NameStr,
1758 BasicBlock *InsertAtEnd) {
1759 return new(3) SelectInst(C, S1, S2, NameStr, InsertAtEnd);
1760 }
1761
1762 const Value *getCondition() const { return Op<0>(); }
1763 const Value *getTrueValue() const { return Op<1>(); }
1764 const Value *getFalseValue() const { return Op<2>(); }
1765 Value *getCondition() { return Op<0>(); }
1766 Value *getTrueValue() { return Op<1>(); }
1767 Value *getFalseValue() { return Op<2>(); }
1768
1769 void setCondition(Value *V) { Op<0>() = V; }
1770 void setTrueValue(Value *V) { Op<1>() = V; }
1771 void setFalseValue(Value *V) { Op<2>() = V; }
1772
1773 /// Swap the true and false values of the select instruction.
1774 /// This doesn't swap prof metadata.
1775 void swapValues() { Op<1>().swap(Op<2>()); }
1776
1777 /// Return a string if the specified operands are invalid
1778 /// for a select operation, otherwise return null.
1779 static const char *areInvalidOperands(Value *Cond, Value *True, Value *False);
1780
1781 /// Transparently provide more efficient getOperand methods.
1782 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
;
1783
1784 OtherOps getOpcode() const {
1785 return static_cast<OtherOps>(Instruction::getOpcode());
1786 }
1787
1788 // Methods for support type inquiry through isa, cast, and dyn_cast:
1789 static bool classof(const Instruction *I) {
1790 return I->getOpcode() == Instruction::Select;
1791 }
1792 static bool classof(const Value *V) {
1793 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1794 }
1795};
1796
1797template <>
1798struct OperandTraits<SelectInst> : public FixedNumOperandTraits<SelectInst, 3> {
1799};
1800
1801DEFINE_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-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h"
, 1801, __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-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h"
, 1801, __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); }
1802
1803//===----------------------------------------------------------------------===//
1804// VAArgInst Class
1805//===----------------------------------------------------------------------===//
1806
1807/// This class represents the va_arg llvm instruction, which returns
1808/// an argument of the specified type given a va_list and increments that list
1809///
1810class VAArgInst : public UnaryInstruction {
1811protected:
1812 // Note: Instruction needs to be a friend here to call cloneImpl.
1813 friend class Instruction;
1814
1815 VAArgInst *cloneImpl() const;
1816
1817public:
1818 VAArgInst(Value *List, Type *Ty, const Twine &NameStr = "",
1819 Instruction *InsertBefore = nullptr)
1820 : UnaryInstruction(Ty, VAArg, List, InsertBefore) {
1821 setName(NameStr);
1822 }
1823
1824 VAArgInst(Value *List, Type *Ty, const Twine &NameStr,
1825 BasicBlock *InsertAtEnd)
1826 : UnaryInstruction(Ty, VAArg, List, InsertAtEnd) {
1827 setName(NameStr);
1828 }
1829
1830 Value *getPointerOperand() { return getOperand(0); }
1831 const Value *getPointerOperand() const { return getOperand(0); }
1832 static unsigned getPointerOperandIndex() { return 0U; }
1833
1834 // Methods for support type inquiry through isa, cast, and dyn_cast:
1835 static bool classof(const Instruction *I) {
1836 return I->getOpcode() == VAArg;
1837 }
1838 static bool classof(const Value *V) {
1839 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1840 }
1841};
1842
1843//===----------------------------------------------------------------------===//
1844// ExtractElementInst Class
1845//===----------------------------------------------------------------------===//
1846
1847/// This instruction extracts a single (scalar)
1848/// element from a VectorType value
1849///
1850class ExtractElementInst : public Instruction {
1851 ExtractElementInst(Value *Vec, Value *Idx, const Twine &NameStr = "",
1852 Instruction *InsertBefore = nullptr);
1853 ExtractElementInst(Value *Vec, Value *Idx, const Twine &NameStr,
1854 BasicBlock *InsertAtEnd);
1855
1856protected:
1857 // Note: Instruction needs to be a friend here to call cloneImpl.
1858 friend class Instruction;
1859
1860 ExtractElementInst *cloneImpl() const;
1861
1862public:
1863 static ExtractElementInst *Create(Value *Vec, Value *Idx,
1864 const Twine &NameStr = "",
1865 Instruction *InsertBefore = nullptr) {
1866 return new(2) ExtractElementInst(Vec, Idx, NameStr, InsertBefore);
1867 }
1868
1869 static ExtractElementInst *Create(Value *Vec, Value *Idx,
1870 const Twine &NameStr,
1871 BasicBlock *InsertAtEnd) {
1872 return new(2) ExtractElementInst(Vec, Idx, NameStr, InsertAtEnd);
1873 }
1874
1875 /// Return true if an extractelement instruction can be
1876 /// formed with the specified operands.
1877 static bool isValidOperands(const Value *Vec, const Value *Idx);
1878
1879 Value *getVectorOperand() { return Op<0>(); }
1880 Value *getIndexOperand() { return Op<1>(); }
1881 const Value *getVectorOperand() const { return Op<0>(); }
1882 const Value *getIndexOperand() const { return Op<1>(); }
1883
1884 VectorType *getVectorOperandType() const {
1885 return cast<VectorType>(getVectorOperand()->getType());
1886 }
1887
1888 /// Transparently provide more efficient getOperand methods.
1889 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
;
1890
1891 // Methods for support type inquiry through isa, cast, and dyn_cast:
1892 static bool classof(const Instruction *I) {
1893 return I->getOpcode() == Instruction::ExtractElement;
1894 }
1895 static bool classof(const Value *V) {
1896 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1897 }
1898};
1899
1900template <>
1901struct OperandTraits<ExtractElementInst> :
1902 public FixedNumOperandTraits<ExtractElementInst, 2> {
1903};
1904
1905DEFINE_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-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h"
, 1905, __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-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h"
, 1905, __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); }
1906
1907//===----------------------------------------------------------------------===//
1908// InsertElementInst Class
1909//===----------------------------------------------------------------------===//
1910
1911/// This instruction inserts a single (scalar)
1912/// element into a VectorType value
1913///
1914class InsertElementInst : public Instruction {
1915 InsertElementInst(Value *Vec, Value *NewElt, Value *Idx,
1916 const Twine &NameStr = "",
1917 Instruction *InsertBefore = nullptr);
1918 InsertElementInst(Value *Vec, Value *NewElt, Value *Idx, const Twine &NameStr,
1919 BasicBlock *InsertAtEnd);
1920
1921protected:
1922 // Note: Instruction needs to be a friend here to call cloneImpl.
1923 friend class Instruction;
1924
1925 InsertElementInst *cloneImpl() const;
1926
1927public:
1928 static InsertElementInst *Create(Value *Vec, Value *NewElt, Value *Idx,
1929 const Twine &NameStr = "",
1930 Instruction *InsertBefore = nullptr) {
1931 return new(3) InsertElementInst(Vec, NewElt, Idx, NameStr, InsertBefore);
1932 }
1933
1934 static InsertElementInst *Create(Value *Vec, Value *NewElt, Value *Idx,
1935 const Twine &NameStr,
1936 BasicBlock *InsertAtEnd) {
1937 return new(3) InsertElementInst(Vec, NewElt, Idx, NameStr, InsertAtEnd);
1938 }
1939
1940 /// Return true if an insertelement instruction can be
1941 /// formed with the specified operands.
1942 static bool isValidOperands(const Value *Vec, const Value *NewElt,
1943 const Value *Idx);
1944
1945 /// Overload to return most specific vector type.
1946 ///
1947 VectorType *getType() const {
1948 return cast<VectorType>(Instruction::getType());
1949 }
1950
1951 /// Transparently provide more efficient getOperand methods.
1952 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
;
1953
1954 // Methods for support type inquiry through isa, cast, and dyn_cast:
1955 static bool classof(const Instruction *I) {
1956 return I->getOpcode() == Instruction::InsertElement;
1957 }
1958 static bool classof(const Value *V) {
1959 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1960 }
1961};
1962
1963template <>
1964struct OperandTraits<InsertElementInst> :
1965 public FixedNumOperandTraits<InsertElementInst, 3> {
1966};
1967
1968DEFINE_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-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h"
, 1968, __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-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h"
, 1968, __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); }
1969
1970//===----------------------------------------------------------------------===//
1971// ShuffleVectorInst Class
1972//===----------------------------------------------------------------------===//
1973
1974constexpr int UndefMaskElem = -1;
1975
1976/// This instruction constructs a fixed permutation of two
1977/// input vectors.
1978///
1979/// For each element of the result vector, the shuffle mask selects an element
1980/// from one of the input vectors to copy to the result. Non-negative elements
1981/// in the mask represent an index into the concatenated pair of input vectors.
1982/// UndefMaskElem (-1) specifies that the result element is undefined.
1983///
1984/// For scalable vectors, all the elements of the mask must be 0 or -1. This
1985/// requirement may be relaxed in the future.
1986class ShuffleVectorInst : public Instruction {
1987 SmallVector<int, 4> ShuffleMask;
1988 Constant *ShuffleMaskForBitcode;
1989
1990protected:
1991 // Note: Instruction needs to be a friend here to call cloneImpl.
1992 friend class Instruction;
1993
1994 ShuffleVectorInst *cloneImpl() const;
1995
1996public:
1997 ShuffleVectorInst(Value *V1, Value *V2, Value *Mask,
1998 const Twine &NameStr = "",
1999 Instruction *InsertBefor = nullptr);
2000 ShuffleVectorInst(Value *V1, Value *V2, Value *Mask,
2001 const Twine &NameStr, BasicBlock *InsertAtEnd);
2002 ShuffleVectorInst(Value *V1, Value *V2, ArrayRef<int> Mask,
2003 const Twine &NameStr = "",
2004 Instruction *InsertBefor = nullptr);
2005 ShuffleVectorInst(Value *V1, Value *V2, ArrayRef<int> Mask,
2006 const Twine &NameStr, BasicBlock *InsertAtEnd);
2007
2008 void *operator new(size_t s) { return User::operator new(s, 2); }
2009
2010 /// Swap the operands and adjust the mask to preserve the semantics
2011 /// of the instruction.
2012 void commute();
2013
2014 /// Return true if a shufflevector instruction can be
2015 /// formed with the specified operands.
2016 static bool isValidOperands(const Value *V1, const Value *V2,
2017 const Value *Mask);
2018 static bool isValidOperands(const Value *V1, const Value *V2,
2019 ArrayRef<int> Mask);
2020
2021 /// Overload to return most specific vector type.
2022 ///
2023 VectorType *getType() const {
2024 return cast<VectorType>(Instruction::getType());
2025 }
2026
2027 /// Transparently provide more efficient getOperand methods.
2028 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
;
2029
2030 /// Return the shuffle mask value of this instruction for the given element
2031 /// index. Return UndefMaskElem if the element is undef.
2032 int getMaskValue(unsigned Elt) const { return ShuffleMask[Elt]; }
2033
2034 /// Convert the input shuffle mask operand to a vector of integers. Undefined
2035 /// elements of the mask are returned as UndefMaskElem.
2036 static void getShuffleMask(const Constant *Mask,
2037 SmallVectorImpl<int> &Result);
2038
2039 /// Return the mask for this instruction as a vector of integers. Undefined
2040 /// elements of the mask are returned as UndefMaskElem.
2041 void getShuffleMask(SmallVectorImpl<int> &Result) const {
2042 Result.assign(ShuffleMask.begin(), ShuffleMask.end());
2043 }
2044
2045 /// Return the mask for this instruction, for use in bitcode.
2046 ///
2047 /// TODO: This is temporary until we decide a new bitcode encoding for
2048 /// shufflevector.
2049 Constant *getShuffleMaskForBitcode() const { return ShuffleMaskForBitcode; }
2050
2051 static Constant *convertShuffleMaskForBitcode(ArrayRef<int> Mask,
2052 Type *ResultTy);
2053
2054 void setShuffleMask(ArrayRef<int> Mask);
2055
2056 ArrayRef<int> getShuffleMask() const { return ShuffleMask; }
2057
2058 /// Return true if this shuffle returns a vector with a different number of
2059 /// elements than its source vectors.
2060 /// Examples: shufflevector <4 x n> A, <4 x n> B, <1,2,3>
2061 /// shufflevector <4 x n> A, <4 x n> B, <1,2,3,4,5>
2062 bool changesLength() const {
2063 unsigned NumSourceElts = cast<VectorType>(Op<0>()->getType())
2064 ->getElementCount()
2065 .getKnownMinValue();
2066 unsigned NumMaskElts = ShuffleMask.size();
2067 return NumSourceElts != NumMaskElts;
2068 }
2069
2070 /// Return true if this shuffle returns a vector with a greater number of
2071 /// elements than its source vectors.
2072 /// Example: shufflevector <2 x n> A, <2 x n> B, <1,2,3>
2073 bool increasesLength() const {
2074 unsigned NumSourceElts = cast<VectorType>(Op<0>()->getType())
2075 ->getElementCount()
2076 .getKnownMinValue();
2077 unsigned NumMaskElts = ShuffleMask.size();
2078 return NumSourceElts < NumMaskElts;
2079 }
2080
2081 /// Return true if this shuffle mask chooses elements from exactly one source
2082 /// vector.
2083 /// Example: <7,5,undef,7>
2084 /// This assumes that vector operands are the same length as the mask.
2085 static bool isSingleSourceMask(ArrayRef<int> Mask);
2086 static bool isSingleSourceMask(const Constant *Mask) {
2087 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-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h"
, 2087, __PRETTY_FUNCTION__))
;
2088 SmallVector<int, 16> MaskAsInts;
2089 getShuffleMask(Mask, MaskAsInts);
2090 return isSingleSourceMask(MaskAsInts);
2091 }
2092
2093 /// Return true if this shuffle chooses elements from exactly one source
2094 /// vector without changing the length of that vector.
2095 /// Example: shufflevector <4 x n> A, <4 x n> B, <3,0,undef,3>
2096 /// TODO: Optionally allow length-changing shuffles.
2097 bool isSingleSource() const {
2098 return !changesLength() && isSingleSourceMask(ShuffleMask);
2099 }
2100
2101 /// Return true if this shuffle mask chooses elements from exactly one source
2102 /// vector without lane crossings. A shuffle using this mask is not
2103 /// necessarily a no-op because it may change the number of elements from its
2104 /// input vectors or it may provide demanded bits knowledge via undef lanes.
2105 /// Example: <undef,undef,2,3>
2106 static bool isIdentityMask(ArrayRef<int> Mask);
2107 static bool isIdentityMask(const Constant *Mask) {
2108 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-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h"
, 2108, __PRETTY_FUNCTION__))
;
2109 SmallVector<int, 16> MaskAsInts;
2110 getShuffleMask(Mask, MaskAsInts);
2111 return isIdentityMask(MaskAsInts);
2112 }
2113
2114 /// Return true if this shuffle chooses elements from exactly one source
2115 /// vector without lane crossings and does not change the number of elements
2116 /// from its input vectors.
2117 /// Example: shufflevector <4 x n> A, <4 x n> B, <4,undef,6,undef>
2118 bool isIdentity() const {
2119 return !changesLength() && isIdentityMask(ShuffleMask);
2120 }
2121
2122 /// Return true if this shuffle lengthens exactly one source vector with
2123 /// undefs in the high elements.
2124 bool isIdentityWithPadding() const;
2125
2126 /// Return true if this shuffle extracts the first N elements of exactly one
2127 /// source vector.
2128 bool isIdentityWithExtract() const;
2129
2130 /// Return true if this shuffle concatenates its 2 source vectors. This
2131 /// returns false if either input is undefined. In that case, the shuffle is
2132 /// is better classified as an identity with padding operation.
2133 bool isConcat() const;
2134
2135 /// Return true if this shuffle mask chooses elements from its source vectors
2136 /// without lane crossings. A shuffle using this mask would be
2137 /// equivalent to a vector select with a constant condition operand.
2138 /// Example: <4,1,6,undef>
2139 /// This returns false if the mask does not choose from both input vectors.
2140 /// In that case, the shuffle is better classified as an identity shuffle.
2141 /// This assumes that vector operands are the same length as the mask
2142 /// (a length-changing shuffle can never be equivalent to a vector select).
2143 static bool isSelectMask(ArrayRef<int> Mask);
2144 static bool isSelectMask(const Constant *Mask) {
2145 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-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h"
, 2145, __PRETTY_FUNCTION__))
;
2146 SmallVector<int, 16> MaskAsInts;
2147 getShuffleMask(Mask, MaskAsInts);
2148 return isSelectMask(MaskAsInts);
2149 }
2150
2151 /// Return true if this shuffle chooses elements from its source vectors
2152 /// without lane crossings and all operands have the same number of elements.
2153 /// In other words, this shuffle is equivalent to a vector select with a
2154 /// constant condition operand.
2155 /// Example: shufflevector <4 x n> A, <4 x n> B, <undef,1,6,3>
2156 /// This returns false if the mask does not choose from both input vectors.
2157 /// In that case, the shuffle is better classified as an identity shuffle.
2158 /// TODO: Optionally allow length-changing shuffles.
2159 bool isSelect() const {
2160 return !changesLength() && isSelectMask(ShuffleMask);
2161 }
2162
2163 /// Return true if this shuffle mask swaps the order of elements from exactly
2164 /// one source vector.
2165 /// Example: <7,6,undef,4>
2166 /// This assumes that vector operands are the same length as the mask.
2167 static bool isReverseMask(ArrayRef<int> Mask);
2168 static bool isReverseMask(const Constant *Mask) {
2169 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-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h"
, 2169, __PRETTY_FUNCTION__))
;
2170 SmallVector<int, 16> MaskAsInts;
2171 getShuffleMask(Mask, MaskAsInts);
2172 return isReverseMask(MaskAsInts);
2173 }
2174
2175 /// Return true if this shuffle swaps the order of elements from exactly
2176 /// one source vector.
2177 /// Example: shufflevector <4 x n> A, <4 x n> B, <3,undef,1,undef>
2178 /// TODO: Optionally allow length-changing shuffles.
2179 bool isReverse() const {
2180 return !changesLength() && isReverseMask(ShuffleMask);
2181 }
2182
2183 /// Return true if this shuffle mask chooses all elements with the same value
2184 /// as the first element of exactly one source vector.
2185 /// Example: <4,undef,undef,4>
2186 /// This assumes that vector operands are the same length as the mask.
2187 static bool isZeroEltSplatMask(ArrayRef<int> Mask);
2188 static bool isZeroEltSplatMask(const Constant *Mask) {
2189 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-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h"
, 2189, __PRETTY_FUNCTION__))
;
2190 SmallVector<int, 16> MaskAsInts;
2191 getShuffleMask(Mask, MaskAsInts);
2192 return isZeroEltSplatMask(MaskAsInts);
2193 }
2194
2195 /// Return true if all elements of this shuffle are the same value as the
2196 /// first element of exactly one source vector without changing the length
2197 /// of that vector.
2198 /// Example: shufflevector <4 x n> A, <4 x n> B, <undef,0,undef,0>
2199 /// TODO: Optionally allow length-changing shuffles.
2200 /// TODO: Optionally allow splats from other elements.
2201 bool isZeroEltSplat() const {
2202 return !changesLength() && isZeroEltSplatMask(ShuffleMask);
2203 }
2204
2205 /// Return true if this shuffle mask is a transpose mask.
2206 /// Transpose vector masks transpose a 2xn matrix. They read corresponding
2207 /// even- or odd-numbered vector elements from two n-dimensional source
2208 /// vectors and write each result into consecutive elements of an
2209 /// n-dimensional destination vector. Two shuffles are necessary to complete
2210 /// the transpose, one for the even elements and another for the odd elements.
2211 /// This description closely follows how the TRN1 and TRN2 AArch64
2212 /// instructions operate.
2213 ///
2214 /// For example, a simple 2x2 matrix can be transposed with:
2215 ///
2216 /// ; Original matrix
2217 /// m0 = < a, b >
2218 /// m1 = < c, d >
2219 ///
2220 /// ; Transposed matrix
2221 /// t0 = < a, c > = shufflevector m0, m1, < 0, 2 >
2222 /// t1 = < b, d > = shufflevector m0, m1, < 1, 3 >
2223 ///
2224 /// For matrices having greater than n columns, the resulting nx2 transposed
2225 /// matrix is stored in two result vectors such that one vector contains
2226 /// interleaved elements from all the even-numbered rows and the other vector
2227 /// contains interleaved elements from all the odd-numbered rows. For example,
2228 /// a 2x4 matrix can be transposed with:
2229 ///
2230 /// ; Original matrix
2231 /// m0 = < a, b, c, d >
2232 /// m1 = < e, f, g, h >
2233 ///
2234 /// ; Transposed matrix
2235 /// t0 = < a, e, c, g > = shufflevector m0, m1 < 0, 4, 2, 6 >
2236 /// t1 = < b, f, d, h > = shufflevector m0, m1 < 1, 5, 3, 7 >
2237 static bool isTransposeMask(ArrayRef<int> Mask);
2238 static bool isTransposeMask(const Constant *Mask) {
2239 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-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h"
, 2239, __PRETTY_FUNCTION__))
;
2240 SmallVector<int, 16> MaskAsInts;
2241 getShuffleMask(Mask, MaskAsInts);
2242 return isTransposeMask(MaskAsInts);
2243 }
2244
2245 /// Return true if this shuffle transposes the elements of its inputs without
2246 /// changing the length of the vectors. This operation may also be known as a
2247 /// merge or interleave. See the description for isTransposeMask() for the
2248 /// exact specification.
2249 /// Example: shufflevector <4 x n> A, <4 x n> B, <0,4,2,6>
2250 bool isTranspose() const {
2251 return !changesLength() && isTransposeMask(ShuffleMask);
2252 }
2253
2254 /// Return true if this shuffle mask is an extract subvector mask.
2255 /// A valid extract subvector mask returns a smaller vector from a single
2256 /// source operand. The base extraction index is returned as well.
2257 static bool isExtractSubvectorMask(ArrayRef<int> Mask, int NumSrcElts,
2258 int &Index);
2259 static bool isExtractSubvectorMask(const Constant *Mask, int NumSrcElts,
2260 int &Index) {
2261 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-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h"
, 2261, __PRETTY_FUNCTION__))
;
2262 // Not possible to express a shuffle mask for a scalable vector for this
2263 // case.
2264 if (isa<ScalableVectorType>(Mask->getType()))
2265 return false;
2266 SmallVector<int, 16> MaskAsInts;
2267 getShuffleMask(Mask, MaskAsInts);
2268 return isExtractSubvectorMask(MaskAsInts, NumSrcElts, Index);
2269 }
2270
2271 /// Return true if this shuffle mask is an extract subvector mask.
2272 bool isExtractSubvectorMask(int &Index) const {
2273 // Not possible to express a shuffle mask for a scalable vector for this
2274 // case.
2275 if (isa<ScalableVectorType>(getType()))
2276 return false;
2277
2278 int NumSrcElts =
2279 cast<FixedVectorType>(Op<0>()->getType())->getNumElements();
2280 return isExtractSubvectorMask(ShuffleMask, NumSrcElts, Index);
2281 }
2282
2283 /// Change values in a shuffle permute mask assuming the two vector operands
2284 /// of length InVecNumElts have swapped position.
2285 static void commuteShuffleMask(MutableArrayRef<int> Mask,
2286 unsigned InVecNumElts) {
2287 for (int &Idx : Mask) {
2288 if (Idx == -1)
2289 continue;
2290 Idx = Idx < (int)InVecNumElts ? Idx + InVecNumElts : Idx - InVecNumElts;
2291 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-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h"
, 2292, __PRETTY_FUNCTION__))
2292 "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-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h"
, 2292, __PRETTY_FUNCTION__))
;
2293 }
2294 }
2295
2296 // Methods for support type inquiry through isa, cast, and dyn_cast:
2297 static bool classof(const Instruction *I) {
2298 return I->getOpcode() == Instruction::ShuffleVector;
2299 }
2300 static bool classof(const Value *V) {
2301 return isa<Instruction>(V) && classof(cast<Instruction>(V));
2302 }
2303};
2304
2305template <>
2306struct OperandTraits<ShuffleVectorInst>
2307 : public FixedNumOperandTraits<ShuffleVectorInst, 2> {};
2308
2309DEFINE_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-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h"
, 2309, __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-13~++20210405022414+5f57793c4fe4/llvm/include/llvm/IR/Instructions.h"
, 2309, __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); }
2310
2311//===----------------------------------------------------------------------===//
2312// ExtractValueInst Class
2313//===----------------------------------------------------------------------===//
2314
2315/// This instruction extracts a struct member or array
2316/// element value from an aggregate value.
2317///
2318class ExtractValueInst : public UnaryInstruction {
2319 SmallVector<unsigned, 4> Indices;
2320
2321 ExtractValueInst(const ExtractValueInst &EVI);
2322
2323 /// Constructors - Create a extractvalue instruction with a base aggregate
2324 /// value and a list of indices. The first ctor can optionally insert before
2325 /// an existing instruction, the second appends the new instruction to the
2326 /// specified BasicBlock.
2327 inline ExtractValueInst(Value *Agg,
2328 ArrayRef<unsigned> Idxs,
2329 const Twine &NameStr,
2330 Instruction *InsertBefore);
2331 inline ExtractValueInst(Value *Agg,
2332 ArrayRef<unsigned> Idxs,
2333 const Twine &NameStr, BasicBlock *InsertAtEnd);
2334
2335 void init(ArrayRef<unsigned> Idxs, const Twine &NameStr);
2336
2337protected:
2338 // Note: Instruction needs to be a friend here to call cloneImpl.
2339 friend class Instruction;
2340
2341 ExtractValueInst *cloneImpl() const;
2342
2343public:
2344 static ExtractValueInst *Create(Value *Agg,
2345 ArrayRef<unsigned> Idxs,
2346 const Twine &NameStr = "",
2347 Instruction *InsertBefore = nullptr) {
2348 return new
2349 ExtractValueInst(Agg, Idxs, NameStr, InsertBefore);
2350 }
2351
2352 static ExtractValueInst *Create(Value *Agg,
2353 ArrayRef<unsigned> Idxs,
2354 const Twine &NameStr,
2355 BasicBlock *InsertAtEnd) {
2356 return new ExtractValueInst(Agg, Idxs, NameStr, InsertAtEnd);
2357 }
2358
2359 /// Returns the type of the element that would be extracted
2360 /// with an extractvalue instruction with the specified parameters.
2361 ///
2362 /// Null is returned if the indices are invalid for the specified type.
2363 static Type *getIndexedType(Type *Agg, ArrayRef<unsigned> Idxs);
2364
2365 using idx_iterator = const unsigned*;
2366
2367 inline idx_iterator idx_begin() const { return Indices.begin(); }
2368 inline idx_iterator idx_end() const { return Indices.end(); }
2369 inline iterator_range<idx_iterator> indices() const {
2370 return make_range(idx_begin(), idx_end());
2371 }
2372
2373 Value *getAggregateOperand() {
2374 return getOperand(0);
2375 }
2376 const Value *getAggregateOperand() const {
2377 return getOperand(0);
2378 }
2379 static unsigned getAggregateOperandIndex() {
2380 return 0U; // get index for modifying correct operand
2381 }
2382
2383 ArrayRef<unsigned> getIndices() const {
2384 return Indices;
2385 }
2386
2387 unsigned getNumIndices() const {
2388 return (unsigned)Indices.size();
2389 }
2390
2391 bool hasIndices() const {
2392 return true;
2393 }
2394
2395 // Methods for support type inquiry through isa, cast, and dyn_cast:
2396 static bool classof(const Instruction *I) {
2397 return I->getOpcode() == Instruction::ExtractValue;
2398 }
2399 static bool classof(const Value *V) {
2400 return isa<Instruction>(V) && classof(cast<Instruction>(V));
2401 }
2402};
2403
2404ExtractValueInst::ExtractValueInst(Value *Agg,
2405 ArrayRef<unsigned> Idxs,
2406 const Twine &NameStr,
2407 Instruction *InsertBefore)
2408 : UnaryInstruction(checkGEPType(getIndexedType(Agg->getType(), Idxs)),
2409 ExtractValue, Agg, InsertBefore) {
2410 init(Idxs, NameStr);
2411}
2412
2413ExtractValueInst::ExtractValueInst(Value *Agg,
2414 ArrayRef<unsigned> Idxs,
2415 const Twine &NameStr,
2416 BasicBlock *InsertAtEnd)
2417 : UnaryInstruction(checkGEPType(getIndexedType(Agg->getType(), Idxs)),
2418 ExtractValue, Agg, InsertAtEnd) {
2419 init(Idxs, NameStr);
2420}
2421
2422//===----------------------------------------------------------------------===//
2423// InsertValueInst Class
2424//===----------------------------------------------------------------------===//
2425
2426/// This instruction inserts a struct field of array element
2427/// value into an aggregate value.
2428///
2429class InsertValueInst : public Instruction {
2430 SmallVector<unsigned, 4> Indices;
2431
2432 InsertValueInst(const InsertValueInst &IVI);
2433
2434 /// Constructors - Create a insertvalue instruction with a base aggregate
2435 /// value, a value to insert, and a list of indices. The first ctor can
2436 /// optionally insert before an existing instruction, the second appends
2437 /// the new instruction to the specified BasicBlock.
2438 inline InsertValueInst(Value *Agg, Value *Val,
2439 ArrayRef<unsigned> Idxs,
2440 const Twine &NameStr,
2441 Instruction *InsertBefore);
2442 inline InsertValueInst(Value *Agg, Value *Val,
2443 ArrayRef<unsigned> Idxs,
2444 const Twine &NameStr, BasicBlock *InsertAtEnd);
2445
2446 /// Constructors - These two constructors are convenience methods because one
2447 /// and two index insertvalue instructions are so common.
2448 InsertValueInst(Value *Agg, Value *Val, unsigned Idx,
2449 const Twine &NameStr = "",
2450 Instruction *InsertBefore = nullptr);
2451 InsertValueInst(Value *Agg, Value *Val, unsigned Idx, const Twine &NameStr,
2452 BasicBlock *InsertAtEnd);
2453
2454 void init(Value *Agg, Value *Val, ArrayRef<unsigned> Idxs,
2455 const Twine &NameStr);
2456
2457protected:
2458 // Note: Instruction needs to be a friend here to call cloneImpl.
2459 friend class Instruction;
2460
2461 InsertValueInst *cloneImpl() const;
2462
2463public:
2464 // allocate space for exactly two operands
2465 void *operator new(size_t s) {
2466 return User::operator new(s, 2);
2467 }
2468
2469 static InsertValueInst *Create(Value *Agg, Value *Val,
2470 ArrayRef<unsigned> Idxs,
2471 const Twine &NameStr =