Bug Summary

File:llvm/lib/Transforms/Scalar/JumpThreading.cpp
Warning:line 1459, 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 -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-14~++20210903100615+fd66b44ec19e/build-llvm/lib/Transforms/Scalar -resource-dir /usr/lib/llvm-14/lib/clang/14.0.0 -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-14~++20210903100615+fd66b44ec19e/build-llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-14~++20210903100615+fd66b44ec19e/llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-14~++20210903100615+fd66b44ec19e/build-llvm/include -I /build/llvm-toolchain-snapshot-14~++20210903100615+fd66b44ec19e/llvm/include -D NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/x86_64-linux-gnu/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10/backward -internal-isystem /usr/lib/llvm-14/lib/clang/14.0.0/include -internal-isystem /usr/local/include -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../x86_64-linux-gnu/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-class-memaccess -Wno-redundant-move -Wno-pessimizing-move -Wno-noexcept-type -Wno-comment -std=c++14 -fdeprecated-macro -fdebug-compilation-dir=/build/llvm-toolchain-snapshot-14~++20210903100615+fd66b44ec19e/build-llvm/lib/Transforms/Scalar -fdebug-prefix-map=/build/llvm-toolchain-snapshot-14~++20210903100615+fd66b44ec19e=. -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-2021-09-04-040900-46481-1 -x c++ /build/llvm-toolchain-snapshot-14~++20210903100615+fd66b44ec19e/llvm/lib/Transforms/Scalar/JumpThreading.cpp

/build/llvm-toolchain-snapshot-14~++20210903100615+fd66b44ec19e/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<DominatorTreeAnalysis>();
375 PA.preserve<LazyValueAnalysis>();
376 return PA;
377}
378
379bool JumpThreadingPass::runImpl(Function &F, TargetLibraryInfo *TLI_,
380 LazyValueInfo *LVI_, AliasAnalysis *AA_,
381 DomTreeUpdater *DTU_, bool HasProfileData_,
382 std::unique_ptr<BlockFrequencyInfo> BFI_,
383 std::unique_ptr<BranchProbabilityInfo> BPI_) {
384 LLVM_DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n")do { } while (false);
385 TLI = TLI_;
386 LVI = LVI_;
387 AA = AA_;
388 DTU = DTU_;
389 BFI.reset();
390 BPI.reset();
391 // When profile data is available, we need to update edge weights after
392 // successful jump threading, which requires both BPI and BFI being available.
393 HasProfileData = HasProfileData_;
394 auto *GuardDecl = F.getParent()->getFunction(
395 Intrinsic::getName(Intrinsic::experimental_guard));
396 HasGuards = GuardDecl && !GuardDecl->use_empty();
397 if (HasProfileData) {
398 BPI = std::move(BPI_);
399 BFI = std::move(BFI_);
400 }
401
402 // Reduce the number of instructions duplicated when optimizing strictly for
403 // size.
404 if (BBDuplicateThreshold.getNumOccurrences())
405 BBDupThreshold = BBDuplicateThreshold;
406 else if (F.hasFnAttribute(Attribute::MinSize))
407 BBDupThreshold = 3;
408 else
409 BBDupThreshold = DefaultBBDupThreshold;
410
411 // JumpThreading must not processes blocks unreachable from entry. It's a
412 // waste of compute time and can potentially lead to hangs.
413 SmallPtrSet<BasicBlock *, 16> Unreachable;
414 assert(DTU && "DTU isn't passed into JumpThreading before using it.")(static_cast<void> (0));
415 assert(DTU->hasDomTree() && "JumpThreading relies on DomTree to proceed.")(static_cast<void> (0));
416 DominatorTree &DT = DTU->getDomTree();
417 for (auto &BB : F)
418 if (!DT.isReachableFromEntry(&BB))
419 Unreachable.insert(&BB);
420
421 if (!ThreadAcrossLoopHeaders)
422 findLoopHeaders(F);
423
424 bool EverChanged = false;
425 bool Changed;
426 do {
427 Changed = false;
428 for (auto &BB : F) {
429 if (Unreachable.count(&BB))
430 continue;
431 while (processBlock(&BB)) // Thread all of the branches we can over BB.
432 Changed = true;
433
434 // Jump threading may have introduced redundant debug values into BB
435 // which should be removed.
436 if (Changed)
437 RemoveRedundantDbgInstrs(&BB);
438
439 // Stop processing BB if it's the entry or is now deleted. The following
440 // routines attempt to eliminate BB and locating a suitable replacement
441 // for the entry is non-trivial.
442 if (&BB == &F.getEntryBlock() || DTU->isBBPendingDeletion(&BB))
443 continue;
444
445 if (pred_empty(&BB)) {
446 // When processBlock makes BB unreachable it doesn't bother to fix up
447 // the instructions in it. We must remove BB to prevent invalid IR.
448 LLVM_DEBUG(dbgs() << " JT: Deleting dead block '" << BB.getName()do { } while (false)
449 << "' with terminator: " << *BB.getTerminator()do { } while (false)
450 << '\n')do { } while (false);
451 LoopHeaders.erase(&BB);
452 LVI->eraseBlock(&BB);
453 DeleteDeadBlock(&BB, DTU);
454 Changed = true;
455 continue;
456 }
457
458 // processBlock doesn't thread BBs with unconditional TIs. However, if BB
459 // is "almost empty", we attempt to merge BB with its sole successor.
460 auto *BI = dyn_cast<BranchInst>(BB.getTerminator());
461 if (BI && BI->isUnconditional()) {
462 BasicBlock *Succ = BI->getSuccessor(0);
463 if (
464 // The terminator must be the only non-phi instruction in BB.
465 BB.getFirstNonPHIOrDbg(true)->isTerminator() &&
466 // Don't alter Loop headers and latches to ensure another pass can
467 // detect and transform nested loops later.
468 !LoopHeaders.count(&BB) && !LoopHeaders.count(Succ) &&
469 TryToSimplifyUncondBranchFromEmptyBlock(&BB, DTU)) {
470 RemoveRedundantDbgInstrs(Succ);
471 // BB is valid for cleanup here because we passed in DTU. F remains
472 // BB's parent until a DTU->getDomTree() event.
473 LVI->eraseBlock(&BB);
474 Changed = true;
475 }
476 }
477 }
478 EverChanged |= Changed;
479 } while (Changed);
480
481 LoopHeaders.clear();
482 return EverChanged;
483}
484
485// Replace uses of Cond with ToVal when safe to do so. If all uses are
486// replaced, we can remove Cond. We cannot blindly replace all uses of Cond
487// because we may incorrectly replace uses when guards/assumes are uses of
488// of `Cond` and we used the guards/assume to reason about the `Cond` value
489// at the end of block. RAUW unconditionally replaces all uses
490// including the guards/assumes themselves and the uses before the
491// guard/assume.
492static void replaceFoldableUses(Instruction *Cond, Value *ToVal) {
493 assert(Cond->getType() == ToVal->getType())(static_cast<void> (0));
494 auto *BB = Cond->getParent();
495 // We can unconditionally replace all uses in non-local blocks (i.e. uses
496 // strictly dominated by BB), since LVI information is true from the
497 // terminator of BB.
498 replaceNonLocalUsesWith(Cond, ToVal);
499 for (Instruction &I : reverse(*BB)) {
500 // Reached the Cond whose uses we are trying to replace, so there are no
501 // more uses.
502 if (&I == Cond)
503 break;
504 // We only replace uses in instructions that are guaranteed to reach the end
505 // of BB, where we know Cond is ToVal.
506 if (!isGuaranteedToTransferExecutionToSuccessor(&I))
507 break;
508 I.replaceUsesOfWith(Cond, ToVal);
509 }
510 if (Cond->use_empty() && !Cond->mayHaveSideEffects())
511 Cond->eraseFromParent();
512}
513
514/// Return the cost of duplicating a piece of this block from first non-phi
515/// and before StopAt instruction to thread across it. Stop scanning the block
516/// when exceeding the threshold. If duplication is impossible, returns ~0U.
517static unsigned getJumpThreadDuplicationCost(BasicBlock *BB,
518 Instruction *StopAt,
519 unsigned Threshold) {
520 assert(StopAt->getParent() == BB && "Not an instruction from proper BB?")(static_cast<void> (0));
521 /// Ignore PHI nodes, these will be flattened when duplication happens.
522 BasicBlock::const_iterator I(BB->getFirstNonPHI());
523
524 // FIXME: THREADING will delete values that are just used to compute the
525 // branch, so they shouldn't count against the duplication cost.
526
527 unsigned Bonus = 0;
528 if (BB->getTerminator() == StopAt) {
529 // Threading through a switch statement is particularly profitable. If this
530 // block ends in a switch, decrease its cost to make it more likely to
531 // happen.
532 if (isa<SwitchInst>(StopAt))
533 Bonus = 6;
534
535 // The same holds for indirect branches, but slightly more so.
536 if (isa<IndirectBrInst>(StopAt))
537 Bonus = 8;
538 }
539
540 // Bump the threshold up so the early exit from the loop doesn't skip the
541 // terminator-based Size adjustment at the end.
542 Threshold += Bonus;
543
544 // Sum up the cost of each instruction until we get to the terminator. Don't
545 // include the terminator because the copy won't include it.
546 unsigned Size = 0;
547 for (; &*I != StopAt; ++I) {
548
549 // Stop scanning the block if we've reached the threshold.
550 if (Size > Threshold)
551 return Size;
552
553 // Debugger intrinsics don't incur code size.
554 if (isa<DbgInfoIntrinsic>(I)) continue;
555
556 // Pseudo-probes don't incur code size.
557 if (isa<PseudoProbeInst>(I))
558 continue;
559
560 // If this is a pointer->pointer bitcast, it is free.
561 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
562 continue;
563
564 // Freeze instruction is free, too.
565 if (isa<FreezeInst>(I))
566 continue;
567
568 // Bail out if this instruction gives back a token type, it is not possible
569 // to duplicate it if it is used outside this BB.
570 if (I->getType()->isTokenTy() && I->isUsedOutsideOfBlock(BB))
571 return ~0U;
572
573 // All other instructions count for at least one unit.
574 ++Size;
575
576 // Calls are more expensive. If they are non-intrinsic calls, we model them
577 // as having cost of 4. If they are a non-vector intrinsic, we model them
578 // as having cost of 2 total, and if they are a vector intrinsic, we model
579 // them as having cost 1.
580 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
581 if (CI->cannotDuplicate() || CI->isConvergent())
582 // Blocks with NoDuplicate are modelled as having infinite cost, so they
583 // are never duplicated.
584 return ~0U;
585 else if (!isa<IntrinsicInst>(CI))
586 Size += 3;
587 else if (!CI->getType()->isVectorTy())
588 Size += 1;
589 }
590 }
591
592 return Size > Bonus ? Size - Bonus : 0;
593}
594
595/// findLoopHeaders - We do not want jump threading to turn proper loop
596/// structures into irreducible loops. Doing this breaks up the loop nesting
597/// hierarchy and pessimizes later transformations. To prevent this from
598/// happening, we first have to find the loop headers. Here we approximate this
599/// by finding targets of backedges in the CFG.
600///
601/// Note that there definitely are cases when we want to allow threading of
602/// edges across a loop header. For example, threading a jump from outside the
603/// loop (the preheader) to an exit block of the loop is definitely profitable.
604/// It is also almost always profitable to thread backedges from within the loop
605/// to exit blocks, and is often profitable to thread backedges to other blocks
606/// within the loop (forming a nested loop). This simple analysis is not rich
607/// enough to track all of these properties and keep it up-to-date as the CFG
608/// mutates, so we don't allow any of these transformations.
609void JumpThreadingPass::findLoopHeaders(Function &F) {
610 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
611 FindFunctionBackedges(F, Edges);
612
613 for (const auto &Edge : Edges)
614 LoopHeaders.insert(Edge.second);
615}
616
617/// getKnownConstant - Helper method to determine if we can thread over a
618/// terminator with the given value as its condition, and if so what value to
619/// use for that. What kind of value this is depends on whether we want an
620/// integer or a block address, but an undef is always accepted.
621/// Returns null if Val is null or not an appropriate constant.
622static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
623 if (!Val)
624 return nullptr;
625
626 // Undef is "known" enough.
627 if (UndefValue *U = dyn_cast<UndefValue>(Val))
628 return U;
629
630 if (Preference == WantBlockAddress)
631 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
632
633 return dyn_cast<ConstantInt>(Val);
634}
635
636/// computeValueKnownInPredecessors - Given a basic block BB and a value V, see
637/// if we can infer that the value is a known ConstantInt/BlockAddress or undef
638/// in any of our predecessors. If so, return the known list of value and pred
639/// BB in the result vector.
640///
641/// This returns true if there were any known values.
642bool JumpThreadingPass::computeValueKnownInPredecessorsImpl(
643 Value *V, BasicBlock *BB, PredValueInfo &Result,
644 ConstantPreference Preference, DenseSet<Value *> &RecursionSet,
645 Instruction *CxtI) {
646 // This method walks up use-def chains recursively. Because of this, we could
647 // get into an infinite loop going around loops in the use-def chain. To
648 // prevent this, keep track of what (value, block) pairs we've already visited
649 // and terminate the search if we loop back to them
650 if (!RecursionSet.insert(V).second)
651 return false;
652
653 // If V is a constant, then it is known in all predecessors.
654 if (Constant *KC = getKnownConstant(V, Preference)) {
655 for (BasicBlock *Pred : predecessors(BB))
656 Result.emplace_back(KC, Pred);
657
658 return !Result.empty();
659 }
660
661 // If V is a non-instruction value, or an instruction in a different block,
662 // then it can't be derived from a PHI.
663 Instruction *I = dyn_cast<Instruction>(V);
664 if (!I || I->getParent() != BB) {
665
666 // Okay, if this is a live-in value, see if it has a known value at the end
667 // of any of our predecessors.
668 //
669 // FIXME: This should be an edge property, not a block end property.
670 /// TODO: Per PR2563, we could infer value range information about a
671 /// predecessor based on its terminator.
672 //
673 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
674 // "I" is a non-local compare-with-a-constant instruction. This would be
675 // able to handle value inequalities better, for example if the compare is
676 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
677 // Perhaps getConstantOnEdge should be smart enough to do this?
678 for (BasicBlock *P : predecessors(BB)) {
679 // If the value is known by LazyValueInfo to be a constant in a
680 // predecessor, use that information to try to thread this block.
681 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB, CxtI);
682 if (Constant *KC = getKnownConstant(PredCst, Preference))
683 Result.emplace_back(KC, P);
684 }
685
686 return !Result.empty();
687 }
688
689 /// If I is a PHI node, then we know the incoming values for any constants.
690 if (PHINode *PN = dyn_cast<PHINode>(I)) {
691 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
692 Value *InVal = PN->getIncomingValue(i);
693 if (Constant *KC = getKnownConstant(InVal, Preference)) {
694 Result.emplace_back(KC, PN->getIncomingBlock(i));
695 } else {
696 Constant *CI = LVI->getConstantOnEdge(InVal,
697 PN->getIncomingBlock(i),
698 BB, CxtI);
699 if (Constant *KC = getKnownConstant(CI, Preference))
700 Result.emplace_back(KC, PN->getIncomingBlock(i));
701 }
702 }
703
704 return !Result.empty();
705 }
706
707 // Handle Cast instructions.
708 if (CastInst *CI = dyn_cast<CastInst>(I)) {
709 Value *Source = CI->getOperand(0);
710 computeValueKnownInPredecessorsImpl(Source, BB, Result, Preference,
711 RecursionSet, CxtI);
712 if (Result.empty())
713 return false;
714
715 // Convert the known values.
716 for (auto &R : Result)
717 R.first = ConstantExpr::getCast(CI->getOpcode(), R.first, CI->getType());
718
719 return true;
720 }
721
722 if (FreezeInst *FI = dyn_cast<FreezeInst>(I)) {
723 Value *Source = FI->getOperand(0);
724 computeValueKnownInPredecessorsImpl(Source, BB, Result, Preference,
725 RecursionSet, CxtI);
726
727 erase_if(Result, [](auto &Pair) {
728 return !isGuaranteedNotToBeUndefOrPoison(Pair.first);
729 });
730
731 return !Result.empty();
732 }
733
734 // Handle some boolean conditions.
735 if (I->getType()->getPrimitiveSizeInBits() == 1) {
736 using namespace PatternMatch;
737
738 assert(Preference == WantInteger && "One-bit non-integer type?")(static_cast<void> (0));
739 // X | true -> true
740 // X & false -> false
741 Value *Op0, *Op1;
742 if (match(I, m_LogicalOr(m_Value(Op0), m_Value(Op1))) ||
743 match(I, m_LogicalAnd(m_Value(Op0), m_Value(Op1)))) {
744 PredValueInfoTy LHSVals, RHSVals;
745
746 computeValueKnownInPredecessorsImpl(Op0, BB, LHSVals, WantInteger,
747 RecursionSet, CxtI);
748 computeValueKnownInPredecessorsImpl(Op1, BB, RHSVals, WantInteger,
749 RecursionSet, CxtI);
750
751 if (LHSVals.empty() && RHSVals.empty())
752 return false;
753
754 ConstantInt *InterestingVal;
755 if (match(I, m_LogicalOr()))
756 InterestingVal = ConstantInt::getTrue(I->getContext());
757 else
758 InterestingVal = ConstantInt::getFalse(I->getContext());
759
760 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
761
762 // Scan for the sentinel. If we find an undef, force it to the
763 // interesting value: x|undef -> true and x&undef -> false.
764 for (const auto &LHSVal : LHSVals)
765 if (LHSVal.first == InterestingVal || isa<UndefValue>(LHSVal.first)) {
766 Result.emplace_back(InterestingVal, LHSVal.second);
767 LHSKnownBBs.insert(LHSVal.second);
768 }
769 for (const auto &RHSVal : RHSVals)
770 if (RHSVal.first == InterestingVal || isa<UndefValue>(RHSVal.first)) {
771 // If we already inferred a value for this block on the LHS, don't
772 // re-add it.
773 if (!LHSKnownBBs.count(RHSVal.second))
774 Result.emplace_back(InterestingVal, RHSVal.second);
775 }
776
777 return !Result.empty();
778 }
779
780 // Handle the NOT form of XOR.
781 if (I->getOpcode() == Instruction::Xor &&
782 isa<ConstantInt>(I->getOperand(1)) &&
783 cast<ConstantInt>(I->getOperand(1))->isOne()) {
784 computeValueKnownInPredecessorsImpl(I->getOperand(0), BB, Result,
785 WantInteger, RecursionSet, CxtI);
786 if (Result.empty())
787 return false;
788
789 // Invert the known values.
790 for (auto &R : Result)
791 R.first = ConstantExpr::getNot(R.first);
792
793 return true;
794 }
795
796 // Try to simplify some other binary operator values.
797 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
798 assert(Preference != WantBlockAddress(static_cast<void> (0))
799 && "A binary operator creating a block address?")(static_cast<void> (0));
800 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
801 PredValueInfoTy LHSVals;
802 computeValueKnownInPredecessorsImpl(BO->getOperand(0), BB, LHSVals,
803 WantInteger, RecursionSet, CxtI);
804
805 // Try to use constant folding to simplify the binary operator.
806 for (const auto &LHSVal : LHSVals) {
807 Constant *V = LHSVal.first;
808 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
809
810 if (Constant *KC = getKnownConstant(Folded, WantInteger))
811 Result.emplace_back(KC, LHSVal.second);
812 }
813 }
814
815 return !Result.empty();
816 }
817
818 // Handle compare with phi operand, where the PHI is defined in this block.
819 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
820 assert(Preference == WantInteger && "Compares only produce integers")(static_cast<void> (0));
821 Type *CmpType = Cmp->getType();
822 Value *CmpLHS = Cmp->getOperand(0);
823 Value *CmpRHS = Cmp->getOperand(1);
824 CmpInst::Predicate Pred = Cmp->getPredicate();
825
826 PHINode *PN = dyn_cast<PHINode>(CmpLHS);
827 if (!PN)
828 PN = dyn_cast<PHINode>(CmpRHS);
829 if (PN && PN->getParent() == BB) {
830 const DataLayout &DL = PN->getModule()->getDataLayout();
831 // We can do this simplification if any comparisons fold to true or false.
832 // See if any do.
833 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
834 BasicBlock *PredBB = PN->getIncomingBlock(i);
835 Value *LHS, *RHS;
836 if (PN == CmpLHS) {
837 LHS = PN->getIncomingValue(i);
838 RHS = CmpRHS->DoPHITranslation(BB, PredBB);
839 } else {
840 LHS = CmpLHS->DoPHITranslation(BB, PredBB);
841 RHS = PN->getIncomingValue(i);
842 }
843 Value *Res = SimplifyCmpInst(Pred, LHS, RHS, {DL});
844 if (!Res) {
845 if (!isa<Constant>(RHS))
846 continue;
847
848 // getPredicateOnEdge call will make no sense if LHS is defined in BB.
849 auto LHSInst = dyn_cast<Instruction>(LHS);
850 if (LHSInst && LHSInst->getParent() == BB)
851 continue;
852
853 LazyValueInfo::Tristate
854 ResT = LVI->getPredicateOnEdge(Pred, LHS,
855 cast<Constant>(RHS), PredBB, BB,
856 CxtI ? CxtI : Cmp);
857 if (ResT == LazyValueInfo::Unknown)
858 continue;
859 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
860 }
861
862 if (Constant *KC = getKnownConstant(Res, WantInteger))
863 Result.emplace_back(KC, PredBB);
864 }
865
866 return !Result.empty();
867 }
868
869 // If comparing a live-in value against a constant, see if we know the
870 // live-in value on any predecessors.
871 if (isa<Constant>(CmpRHS) && !CmpType->isVectorTy()) {
872 Constant *CmpConst = cast<Constant>(CmpRHS);
873
874 if (!isa<Instruction>(CmpLHS) ||
875 cast<Instruction>(CmpLHS)->getParent() != BB) {
876 for (BasicBlock *P : predecessors(BB)) {
877 // If the value is known by LazyValueInfo to be a constant in a
878 // predecessor, use that information to try to thread this block.
879 LazyValueInfo::Tristate Res =
880 LVI->getPredicateOnEdge(Pred, CmpLHS,
881 CmpConst, P, BB, CxtI ? CxtI : Cmp);
882 if (Res == LazyValueInfo::Unknown)
883 continue;
884
885 Constant *ResC = ConstantInt::get(CmpType, Res);
886 Result.emplace_back(ResC, P);
887 }
888
889 return !Result.empty();
890 }
891
892 // InstCombine can fold some forms of constant range checks into
893 // (icmp (add (x, C1)), C2). See if we have we have such a thing with
894 // x as a live-in.
895 {
896 using namespace PatternMatch;
897
898 Value *AddLHS;
899 ConstantInt *AddConst;
900 if (isa<ConstantInt>(CmpConst) &&
901 match(CmpLHS, m_Add(m_Value(AddLHS), m_ConstantInt(AddConst)))) {
902 if (!isa<Instruction>(AddLHS) ||
903 cast<Instruction>(AddLHS)->getParent() != BB) {
904 for (BasicBlock *P : predecessors(BB)) {
905 // If the value is known by LazyValueInfo to be a ConstantRange in
906 // a predecessor, use that information to try to thread this
907 // block.
908 ConstantRange CR = LVI->getConstantRangeOnEdge(
909 AddLHS, P, BB, CxtI ? CxtI : cast<Instruction>(CmpLHS));
910 // Propagate the range through the addition.
911 CR = CR.add(AddConst->getValue());
912
913 // Get the range where the compare returns true.
914 ConstantRange CmpRange = ConstantRange::makeExactICmpRegion(
915 Pred, cast<ConstantInt>(CmpConst)->getValue());
916
917 Constant *ResC;
918 if (CmpRange.contains(CR))
919 ResC = ConstantInt::getTrue(CmpType);
920 else if (CmpRange.inverse().contains(CR))
921 ResC = ConstantInt::getFalse(CmpType);
922 else
923 continue;
924
925 Result.emplace_back(ResC, P);
926 }
927
928 return !Result.empty();
929 }
930 }
931 }
932
933 // Try to find a constant value for the LHS of a comparison,
934 // and evaluate it statically if we can.
935 PredValueInfoTy LHSVals;
936 computeValueKnownInPredecessorsImpl(I->getOperand(0), BB, LHSVals,
937 WantInteger, RecursionSet, CxtI);
938
939 for (const auto &LHSVal : LHSVals) {
940 Constant *V = LHSVal.first;
941 Constant *Folded = ConstantExpr::getCompare(Pred, V, CmpConst);
942 if (Constant *KC = getKnownConstant(Folded, WantInteger))
943 Result.emplace_back(KC, LHSVal.second);
944 }
945
946 return !Result.empty();
947 }
948 }
949
950 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
951 // Handle select instructions where at least one operand is a known constant
952 // and we can figure out the condition value for any predecessor block.
953 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
954 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
955 PredValueInfoTy Conds;
956 if ((TrueVal || FalseVal) &&
957 computeValueKnownInPredecessorsImpl(SI->getCondition(), BB, Conds,
958 WantInteger, RecursionSet, CxtI)) {
959 for (auto &C : Conds) {
960 Constant *Cond = C.first;
961
962 // Figure out what value to use for the condition.
963 bool KnownCond;
964 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
965 // A known boolean.
966 KnownCond = CI->isOne();
967 } else {
968 assert(isa<UndefValue>(Cond) && "Unexpected condition value")(static_cast<void> (0));
969 // Either operand will do, so be sure to pick the one that's a known
970 // constant.
971 // FIXME: Do this more cleverly if both values are known constants?
972 KnownCond = (TrueVal != nullptr);
973 }
974
975 // See if the select has a known constant value for this predecessor.
976 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
977 Result.emplace_back(Val, C.second);
978 }
979
980 return !Result.empty();
981 }
982 }
983
984 // If all else fails, see if LVI can figure out a constant value for us.
985 assert(CxtI->getParent() == BB && "CxtI should be in BB")(static_cast<void> (0));
986 Constant *CI = LVI->getConstant(V, CxtI);
987 if (Constant *KC = getKnownConstant(CI, Preference)) {
988 for (BasicBlock *Pred : predecessors(BB))
989 Result.emplace_back(KC, Pred);
990 }
991
992 return !Result.empty();
993}
994
995/// GetBestDestForBranchOnUndef - If we determine that the specified block ends
996/// in an undefined jump, decide which block is best to revector to.
997///
998/// Since we can pick an arbitrary destination, we pick the successor with the
999/// fewest predecessors. This should reduce the in-degree of the others.
1000static unsigned getBestDestForJumpOnUndef(BasicBlock *BB) {
1001 Instruction *BBTerm = BB->getTerminator();
1002 unsigned MinSucc = 0;
1003 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
1004 // Compute the successor with the minimum number of predecessors.
1005 unsigned MinNumPreds = pred_size(TestBB);
1006 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
1007 TestBB = BBTerm->getSuccessor(i);
1008 unsigned NumPreds = pred_size(TestBB);
1009 if (NumPreds < MinNumPreds) {
1010 MinSucc = i;
1011 MinNumPreds = NumPreds;
1012 }
1013 }
1014
1015 return MinSucc;
1016}
1017
1018static bool hasAddressTakenAndUsed(BasicBlock *BB) {
1019 if (!BB->hasAddressTaken()) return false;
1020
1021 // If the block has its address taken, it may be a tree of dead constants
1022 // hanging off of it. These shouldn't keep the block alive.
1023 BlockAddress *BA = BlockAddress::get(BB);
1024 BA->removeDeadConstantUsers();
1025 return !BA->use_empty();
1026}
1027
1028/// processBlock - If there are any predecessors whose control can be threaded
1029/// through to a successor, transform them now.
1030bool JumpThreadingPass::processBlock(BasicBlock *BB) {
1031 // If the block is trivially dead, just return and let the caller nuke it.
1032 // This simplifies other transformations.
1033 if (DTU->isBBPendingDeletion(BB) ||
1034 (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()))
1035 return false;
1036
1037 // If this block has a single predecessor, and if that pred has a single
1038 // successor, merge the blocks. This encourages recursive jump threading
1039 // because now the condition in this block can be threaded through
1040 // predecessors of our predecessor block.
1041 if (maybeMergeBasicBlockIntoOnlyPred(BB))
1042 return true;
1043
1044 if (tryToUnfoldSelectInCurrBB(BB))
1045 return true;
1046
1047 // Look if we can propagate guards to predecessors.
1048 if (HasGuards && processGuards(BB))
1049 return true;
1050
1051 // What kind of constant we're looking for.
1052 ConstantPreference Preference = WantInteger;
1053
1054 // Look to see if the terminator is a conditional branch, switch or indirect
1055 // branch, if not we can't thread it.
1056 Value *Condition;
1057 Instruction *Terminator = BB->getTerminator();
1058 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
1059 // Can't thread an unconditional jump.
1060 if (BI->isUnconditional()) return false;
1061 Condition = BI->getCondition();
1062 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
1063 Condition = SI->getCondition();
1064 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
1065 // Can't thread indirect branch with no successors.
1066 if (IB->getNumSuccessors() == 0) return false;
1067 Condition = IB->getAddress()->stripPointerCasts();
1068 Preference = WantBlockAddress;
1069 } else {
1070 return false; // Must be an invoke or callbr.
1071 }
1072
1073 // Keep track if we constant folded the condition in this invocation.
1074 bool ConstantFolded = false;
1075
1076 // Run constant folding to see if we can reduce the condition to a simple
1077 // constant.
1078 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
1079 Value *SimpleVal =
1080 ConstantFoldInstruction(I, BB->getModule()->getDataLayout(), TLI);
1081 if (SimpleVal) {
1082 I->replaceAllUsesWith(SimpleVal);
1083 if (isInstructionTriviallyDead(I, TLI))
1084 I->eraseFromParent();
1085 Condition = SimpleVal;
1086 ConstantFolded = true;
1087 }
1088 }
1089
1090 // If the terminator is branching on an undef or freeze undef, we can pick any
1091 // of the successors to branch to. Let getBestDestForJumpOnUndef decide.
1092 auto *FI = dyn_cast<FreezeInst>(Condition);
1093 if (isa<UndefValue>(Condition) ||
1094 (FI && isa<UndefValue>(FI->getOperand(0)) && FI->hasOneUse())) {
1095 unsigned BestSucc = getBestDestForJumpOnUndef(BB);
1096 std::vector<DominatorTree::UpdateType> Updates;
1097
1098 // Fold the branch/switch.
1099 Instruction *BBTerm = BB->getTerminator();
1100 Updates.reserve(BBTerm->getNumSuccessors());
1101 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
1102 if (i == BestSucc) continue;
1103 BasicBlock *Succ = BBTerm->getSuccessor(i);
1104 Succ->removePredecessor(BB, true);
1105 Updates.push_back({DominatorTree::Delete, BB, Succ});
1106 }
1107
1108 LLVM_DEBUG(dbgs() << " In block '" << BB->getName()do { } while (false)
1109 << "' folding undef terminator: " << *BBTerm << '\n')do { } while (false);
1110 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
1111 ++NumFolds;
1112 BBTerm->eraseFromParent();
1113 DTU->applyUpdatesPermissive(Updates);
1114 if (FI)
1115 FI->eraseFromParent();
1116 return true;
1117 }
1118
1119 // If the terminator of this block is branching on a constant, simplify the
1120 // terminator to an unconditional branch. This can occur due to threading in
1121 // other blocks.
1122 if (getKnownConstant(Condition, Preference)) {
1123 LLVM_DEBUG(dbgs() << " In block '" << BB->getName()do { } while (false)
1124 << "' folding terminator: " << *BB->getTerminator()do { } while (false)
1125 << '\n')do { } while (false);
1126 ++NumFolds;
1127 ConstantFoldTerminator(BB, true, nullptr, DTU);
1128 if (HasProfileData)
1129 BPI->eraseBlock(BB);
1130 return true;
1131 }
1132
1133 Instruction *CondInst = dyn_cast<Instruction>(Condition);
1134
1135 // All the rest of our checks depend on the condition being an instruction.
1136 if (!CondInst) {
1137 // FIXME: Unify this with code below.
1138 if (processThreadableEdges(Condition, BB, Preference, Terminator))
1139 return true;
1140 return ConstantFolded;
1141 }
1142
1143 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
1144 // If we're branching on a conditional, LVI might be able to determine
1145 // it's value at the branch instruction. We only handle comparisons
1146 // against a constant at this time.
1147 // TODO: This should be extended to handle switches as well.
1148 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1149 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
1150 if (CondBr && CondConst) {
1151 // We should have returned as soon as we turn a conditional branch to
1152 // unconditional. Because its no longer interesting as far as jump
1153 // threading is concerned.
1154 assert(CondBr->isConditional() && "Threading on unconditional terminator")(static_cast<void> (0));
1155
1156 LazyValueInfo::Tristate Ret =
1157 LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0),
1158 CondConst, CondBr, /*UseBlockValue=*/false);
1159 if (Ret != LazyValueInfo::Unknown) {
1160 unsigned ToRemove = Ret == LazyValueInfo::True ? 1 : 0;
1161 unsigned ToKeep = Ret == LazyValueInfo::True ? 0 : 1;
1162 BasicBlock *ToRemoveSucc = CondBr->getSuccessor(ToRemove);
1163 ToRemoveSucc->removePredecessor(BB, true);
1164 BranchInst *UncondBr =
1165 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
1166 UncondBr->setDebugLoc(CondBr->getDebugLoc());
1167 ++NumFolds;
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 ++NumFolds;
1284 BI->eraseFromParent();
1285 DTU->applyUpdatesPermissive({{DominatorTree::Delete, BB, RemoveSucc}});
1286 if (HasProfileData)
1287 BPI->eraseBlock(BB);
1288 return true;
1289 }
1290 CurrentBB = CurrentPred;
1291 CurrentPred = CurrentBB->getSinglePredecessor();
1292 }
1293
1294 return false;
1295}
1296
1297/// Return true if Op is an instruction defined in the given block.
1298static bool isOpDefinedInBlock(Value *Op, BasicBlock *BB) {
1299 if (Instruction *OpInst = dyn_cast<Instruction>(Op))
13
Assuming 'OpInst' is null
14
Taking false branch
1300 if (OpInst->getParent() == BB)
1301 return true;
1302 return false;
15
Returning zero, which participates in a condition later
1303}
1304
1305/// simplifyPartiallyRedundantLoad - If LoadI is an obviously partially
1306/// redundant load instruction, eliminate it by replacing it with a PHI node.
1307/// This is an important optimization that encourages jump threading, and needs
1308/// to be run interlaced with other jump threading tasks.
1309bool JumpThreadingPass::simplifyPartiallyRedundantLoad(LoadInst *LoadI) {
1310 // Don't hack volatile and ordered loads.
1311 if (!LoadI->isUnordered()) return false;
1
Calling 'LoadInst::isUnordered'
6
Returning from 'LoadInst::isUnordered'
7
Taking false branch
1312
1313 // If the load is defined in a block with exactly one predecessor, it can't be
1314 // partially redundant.
1315 BasicBlock *LoadBB = LoadI->getParent();
1316 if (LoadBB->getSinglePredecessor())
8
Assuming the condition is false
9
Taking false branch
1317 return false;
1318
1319 // If the load is defined in an EH pad, it can't be partially redundant,
1320 // because the edges between the invoke and the EH pad cannot have other
1321 // instructions between them.
1322 if (LoadBB->isEHPad())
10
Assuming the condition is false
11
Taking false branch
1323 return false;
1324
1325 Value *LoadedPtr = LoadI->getOperand(0);
1326
1327 // If the loaded operand is defined in the LoadBB and its not a phi,
1328 // it can't be available in predecessors.
1329 if (isOpDefinedInBlock(LoadedPtr, LoadBB) && !isa<PHINode>(LoadedPtr))
12
Calling 'isOpDefinedInBlock'
16
Returning from 'isOpDefinedInBlock'
1330 return false;
1331
1332 // Scan a few instructions up from the load, to see if it is obviously live at
1333 // the entry to its block.
1334 BasicBlock::iterator BBIt(LoadI);
1335 bool IsLoadCSE;
1336 if (Value *AvailableVal = FindAvailableLoadedValue(
17
Assuming 'AvailableVal' is null
18
Taking false branch
1337 LoadI, LoadBB, BBIt, DefMaxInstsToScan, AA, &IsLoadCSE)) {
1338 // If the value of the load is locally available within the block, just use
1339 // it. This frequently occurs for reg2mem'd allocas.
1340
1341 if (IsLoadCSE) {
1342 LoadInst *NLoadI = cast<LoadInst>(AvailableVal);
1343 combineMetadataForCSE(NLoadI, LoadI, false);
1344 };
1345
1346 // If the returned value is the load itself, replace with an undef. This can
1347 // only happen in dead loops.
1348 if (AvailableVal == LoadI)
1349 AvailableVal = UndefValue::get(LoadI->getType());
1350 if (AvailableVal->getType() != LoadI->getType())
1351 AvailableVal = CastInst::CreateBitOrPointerCast(
1352 AvailableVal, LoadI->getType(), "", LoadI);
1353 LoadI->replaceAllUsesWith(AvailableVal);
1354 LoadI->eraseFromParent();
1355 return true;
1356 }
1357
1358 // Otherwise, if we scanned the whole block and got to the top of the block,
1359 // we know the block is locally transparent to the load. If not, something
1360 // might clobber its value.
1361 if (BBIt != LoadBB->begin())
19
Calling 'operator!='
22
Returning from 'operator!='
23
Taking false branch
1362 return false;
1363
1364 // If all of the loads and stores that feed the value have the same AA tags,
1365 // then we can propagate them onto any newly inserted loads.
1366 AAMDNodes AATags;
1367 LoadI->getAAMetadata(AATags);
1368
1369 SmallPtrSet<BasicBlock*, 8> PredsScanned;
1370
1371 using AvailablePredsTy = SmallVector<std::pair<BasicBlock *, Value *>, 8>;
1372
1373 AvailablePredsTy AvailablePreds;
1374 BasicBlock *OneUnavailablePred = nullptr;
24
'OneUnavailablePred' initialized to a null pointer value
1375 SmallVector<LoadInst*, 8> CSELoads;
1376
1377 // If we got here, the loaded value is transparent through to the start of the
1378 // block. Check to see if it is available in any of the predecessor blocks.
1379 for (BasicBlock *PredBB : predecessors(LoadBB)) {
1380 // If we already scanned this predecessor, skip it.
1381 if (!PredsScanned.insert(PredBB).second)
1382 continue;
1383
1384 BBIt = PredBB->end();
1385 unsigned NumScanedInst = 0;
1386 Value *PredAvailable = nullptr;
1387 // NOTE: We don't CSE load that is volatile or anything stronger than
1388 // unordered, that should have been checked when we entered the function.
1389 assert(LoadI->isUnordered() &&(static_cast<void> (0))
1390 "Attempting to CSE volatile or atomic loads")(static_cast<void> (0));
1391 // If this is a load on a phi pointer, phi-translate it and search
1392 // for available load/store to the pointer in predecessors.
1393 Type *AccessTy = LoadI->getType();
1394 const auto &DL = LoadI->getModule()->getDataLayout();
1395 MemoryLocation Loc(LoadedPtr->DoPHITranslation(LoadBB, PredBB),
1396 LocationSize::precise(DL.getTypeStoreSize(AccessTy)),
1397 AATags);
1398 PredAvailable = findAvailablePtrLoadStore(Loc, AccessTy, LoadI->isAtomic(),
1399 PredBB, BBIt, DefMaxInstsToScan,
1400 AA, &IsLoadCSE, &NumScanedInst);
1401
1402 // If PredBB has a single predecessor, continue scanning through the
1403 // single predecessor.
1404 BasicBlock *SinglePredBB = PredBB;
1405 while (!PredAvailable && SinglePredBB && BBIt == SinglePredBB->begin() &&
1406 NumScanedInst < DefMaxInstsToScan) {
1407 SinglePredBB = SinglePredBB->getSinglePredecessor();
1408 if (SinglePredBB) {
1409 BBIt = SinglePredBB->end();
1410 PredAvailable = findAvailablePtrLoadStore(
1411 Loc, AccessTy, LoadI->isAtomic(), SinglePredBB, BBIt,
1412 (DefMaxInstsToScan - NumScanedInst), AA, &IsLoadCSE,
1413 &NumScanedInst);
1414 }
1415 }
1416
1417 if (!PredAvailable) {
1418 OneUnavailablePred = PredBB;
1419 continue;
1420 }
1421
1422 if (IsLoadCSE)
1423 CSELoads.push_back(cast<LoadInst>(PredAvailable));
1424
1425 // If so, this load is partially redundant. Remember this info so that we
1426 // can create a PHI node.
1427 AvailablePreds.emplace_back(PredBB, PredAvailable);
1428 }
1429
1430 // If the loaded value isn't available in any predecessor, it isn't partially
1431 // redundant.
1432 if (AvailablePreds.empty()) return false;
25
Calling 'SmallVectorBase::empty'
28
Returning from 'SmallVectorBase::empty'
29
Taking false branch
1433
1434 // Okay, the loaded value is available in at least one (and maybe all!)
1435 // predecessors. If the value is unavailable in more than one unique
1436 // predecessor, we want to insert a merge block for those common predecessors.
1437 // This ensures that we only have to insert one reload, thus not increasing
1438 // code size.
1439 BasicBlock *UnavailablePred = nullptr;
1440
1441 // If the value is unavailable in one of predecessors, we will end up
1442 // inserting a new instruction into them. It is only valid if all the
1443 // instructions before LoadI are guaranteed to pass execution to its
1444 // successor, or if LoadI is safe to speculate.
1445 // TODO: If this logic becomes more complex, and we will perform PRE insertion
1446 // farther than to a predecessor, we need to reuse the code from GVN's PRE.
1447 // It requires domination tree analysis, so for this simple case it is an
1448 // overkill.
1449 if (PredsScanned.size() != AvailablePreds.size() &&
30
Assuming the condition is false
1450 !isSafeToSpeculativelyExecute(LoadI))
1451 for (auto I = LoadBB->begin(); &*I != LoadI; ++I)
1452 if (!isGuaranteedToTransferExecutionToSuccessor(&*I))
1453 return false;
1454
1455 // If there is exactly one predecessor where the value is unavailable, the
1456 // already computed 'OneUnavailablePred' block is it. If it ends in an
1457 // unconditional branch, we know that it isn't a critical edge.
1458 if (PredsScanned.size() == AvailablePreds.size()+1 &&
31
Assuming the condition is true
1459 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
32
Called C++ object pointer is null
1460 UnavailablePred = OneUnavailablePred;
1461 } else if (PredsScanned.size() != AvailablePreds.size()) {
1462 // Otherwise, we had multiple unavailable predecessors or we had a critical
1463 // edge from the one.
1464 SmallVector<BasicBlock*, 8> PredsToSplit;
1465 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
1466
1467 for (const auto &AvailablePred : AvailablePreds)
1468 AvailablePredSet.insert(AvailablePred.first);
1469
1470 // Add all the unavailable predecessors to the PredsToSplit list.
1471 for (BasicBlock *P : predecessors(LoadBB)) {
1472 // If the predecessor is an indirect goto, we can't split the edge.
1473 // Same for CallBr.
1474 if (isa<IndirectBrInst>(P->getTerminator()) ||
1475 isa<CallBrInst>(P->getTerminator()))
1476 return false;
1477
1478 if (!AvailablePredSet.count(P))
1479 PredsToSplit.push_back(P);
1480 }
1481
1482 // Split them out to their own block.
1483 UnavailablePred = splitBlockPreds(LoadBB, PredsToSplit, "thread-pre-split");
1484 }
1485
1486 // If the value isn't available in all predecessors, then there will be
1487 // exactly one where it isn't available. Insert a load on that edge and add
1488 // it to the AvailablePreds list.
1489 if (UnavailablePred) {
1490 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&(static_cast<void> (0))
1491 "Can't handle critical edge here!")(static_cast<void> (0));
1492 LoadInst *NewVal = new LoadInst(
1493 LoadI->getType(), LoadedPtr->DoPHITranslation(LoadBB, UnavailablePred),
1494 LoadI->getName() + ".pr", false, LoadI->getAlign(),
1495 LoadI->getOrdering(), LoadI->getSyncScopeID(),
1496 UnavailablePred->getTerminator());
1497 NewVal->setDebugLoc(LoadI->getDebugLoc());
1498 if (AATags)
1499 NewVal->setAAMetadata(AATags);
1500
1501 AvailablePreds.emplace_back(UnavailablePred, NewVal);
1502 }
1503
1504 // Now we know that each predecessor of this block has a value in
1505 // AvailablePreds, sort them for efficient access as we're walking the preds.
1506 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1507
1508 // Create a PHI node at the start of the block for the PRE'd load value.
1509 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
1510 PHINode *PN = PHINode::Create(LoadI->getType(), std::distance(PB, PE), "",
1511 &LoadBB->front());
1512 PN->takeName(LoadI);
1513 PN->setDebugLoc(LoadI->getDebugLoc());
1514
1515 // Insert new entries into the PHI for each predecessor. A single block may
1516 // have multiple entries here.
1517 for (pred_iterator PI = PB; PI != PE; ++PI) {
1518 BasicBlock *P = *PI;
1519 AvailablePredsTy::iterator I =
1520 llvm::lower_bound(AvailablePreds, std::make_pair(P, (Value *)nullptr));
1521
1522 assert(I != AvailablePreds.end() && I->first == P &&(static_cast<void> (0))
1523 "Didn't find entry for predecessor!")(static_cast<void> (0));
1524
1525 // If we have an available predecessor but it requires casting, insert the
1526 // cast in the predecessor and use the cast. Note that we have to update the
1527 // AvailablePreds vector as we go so that all of the PHI entries for this
1528 // predecessor use the same bitcast.
1529 Value *&PredV = I->second;
1530 if (PredV->getType() != LoadI->getType())
1531 PredV = CastInst::CreateBitOrPointerCast(PredV, LoadI->getType(), "",
1532 P->getTerminator());
1533
1534 PN->addIncoming(PredV, I->first);
1535 }
1536
1537 for (LoadInst *PredLoadI : CSELoads) {
1538 combineMetadataForCSE(PredLoadI, LoadI, true);
1539 }
1540
1541 LoadI->replaceAllUsesWith(PN);
1542 LoadI->eraseFromParent();
1543
1544 return true;
1545}
1546
1547/// findMostPopularDest - The specified list contains multiple possible
1548/// threadable destinations. Pick the one that occurs the most frequently in
1549/// the list.
1550static BasicBlock *
1551findMostPopularDest(BasicBlock *BB,
1552 const SmallVectorImpl<std::pair<BasicBlock *,
1553 BasicBlock *>> &PredToDestList) {
1554 assert(!PredToDestList.empty())(static_cast<void> (0));
1555
1556 // Determine popularity. If there are multiple possible destinations, we
1557 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1558 // blocks with known and real destinations to threading undef. We'll handle
1559 // them later if interesting.
1560 MapVector<BasicBlock *, unsigned> DestPopularity;
1561
1562 // Populate DestPopularity with the successors in the order they appear in the
1563 // successor list. This way, we ensure determinism by iterating it in the
1564 // same order in std::max_element below. We map nullptr to 0 so that we can
1565 // return nullptr when PredToDestList contains nullptr only.
1566 DestPopularity[nullptr] = 0;
1567 for (auto *SuccBB : successors(BB))
1568 DestPopularity[SuccBB] = 0;
1569
1570 for (const auto &PredToDest : PredToDestList)
1571 if (PredToDest.second)
1572 DestPopularity[PredToDest.second]++;
1573
1574 // Find the most popular dest.
1575 using VT = decltype(DestPopularity)::value_type;
1576 auto MostPopular = std::max_element(
1577 DestPopularity.begin(), DestPopularity.end(),
1578 [](const VT &L, const VT &R) { return L.second < R.second; });
1579
1580 // Okay, we have finally picked the most popular destination.
1581 return MostPopular->first;
1582}
1583
1584// Try to evaluate the value of V when the control flows from PredPredBB to
1585// BB->getSinglePredecessor() and then on to BB.
1586Constant *JumpThreadingPass::evaluateOnPredecessorEdge(BasicBlock *BB,
1587 BasicBlock *PredPredBB,
1588 Value *V) {
1589 BasicBlock *PredBB = BB->getSinglePredecessor();
1590 assert(PredBB && "Expected a single predecessor")(static_cast<void> (0));
1591
1592 if (Constant *Cst = dyn_cast<Constant>(V)) {
1593 return Cst;
1594 }
1595
1596 // Consult LVI if V is not an instruction in BB or PredBB.
1597 Instruction *I = dyn_cast<Instruction>(V);
1598 if (!I || (I->getParent() != BB && I->getParent() != PredBB)) {
1599 return LVI->getConstantOnEdge(V, PredPredBB, PredBB, nullptr);
1600 }
1601
1602 // Look into a PHI argument.
1603 if (PHINode *PHI = dyn_cast<PHINode>(V)) {
1604 if (PHI->getParent() == PredBB)
1605 return dyn_cast<Constant>(PHI->getIncomingValueForBlock(PredPredBB));
1606 return nullptr;
1607 }
1608
1609 // If we have a CmpInst, try to fold it for each incoming edge into PredBB.
1610 if (CmpInst *CondCmp = dyn_cast<CmpInst>(V)) {
1611 if (CondCmp->getParent() == BB) {
1612 Constant *Op0 =
1613 evaluateOnPredecessorEdge(BB, PredPredBB, CondCmp->getOperand(0));
1614 Constant *Op1 =
1615 evaluateOnPredecessorEdge(BB, PredPredBB, CondCmp->getOperand(1));
1616 if (Op0 && Op1) {
1617 return ConstantExpr::getCompare(CondCmp->getPredicate(), Op0, Op1);
1618 }
1619 }
1620 return nullptr;
1621 }
1622
1623 return nullptr;
1624}
1625
1626bool JumpThreadingPass::processThreadableEdges(Value *Cond, BasicBlock *BB,
1627 ConstantPreference Preference,
1628 Instruction *CxtI) {
1629 // If threading this would thread across a loop header, don't even try to
1630 // thread the edge.
1631 if (LoopHeaders.count(BB))
1632 return false;
1633
1634 PredValueInfoTy PredValues;
1635 if (!computeValueKnownInPredecessors(Cond, BB, PredValues, Preference,
1636 CxtI)) {
1637 // We don't have known values in predecessors. See if we can thread through
1638 // BB and its sole predecessor.
1639 return maybethreadThroughTwoBasicBlocks(BB, Cond);
1640 }
1641
1642 assert(!PredValues.empty() &&(static_cast<void> (0))
1643 "computeValueKnownInPredecessors returned true with no values")(static_cast<void> (0));
1644
1645 LLVM_DEBUG(dbgs() << "IN BB: " << *BB;do { } while (false)
1646 for (const auto &PredValue : PredValues) {do { } while (false)
1647 dbgs() << " BB '" << BB->getName()do { } while (false)
1648 << "': FOUND condition = " << *PredValue.firstdo { } while (false)
1649 << " for pred '" << PredValue.second->getName() << "'.\n";do { } while (false)
1650 })do { } while (false);
1651
1652 // Decide what we want to thread through. Convert our list of known values to
1653 // a list of known destinations for each pred. This also discards duplicate
1654 // predecessors and keeps track of the undefined inputs (which are represented
1655 // as a null dest in the PredToDestList).
1656 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1657 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1658
1659 BasicBlock *OnlyDest = nullptr;
1660 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1661 Constant *OnlyVal = nullptr;
1662 Constant *MultipleVal = (Constant *)(intptr_t)~0ULL;
1663
1664 for (const auto &PredValue : PredValues) {
1665 BasicBlock *Pred = PredValue.second;
1666 if (!SeenPreds.insert(Pred).second)
1667 continue; // Duplicate predecessor entry.
1668
1669 Constant *Val = PredValue.first;
1670
1671 BasicBlock *DestBB;
1672 if (isa<UndefValue>(Val))
1673 DestBB = nullptr;
1674 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
1675 assert(isa<ConstantInt>(Val) && "Expecting a constant integer")(static_cast<void> (0));
1676 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1677 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1678 assert(isa<ConstantInt>(Val) && "Expecting a constant integer")(static_cast<void> (0));
1679 DestBB = SI->findCaseValue(cast<ConstantInt>(Val))->getCaseSuccessor();
1680 } else {
1681 assert(isa<IndirectBrInst>(BB->getTerminator())(static_cast<void> (0))
1682 && "Unexpected terminator")(static_cast<void> (0));
1683 assert(isa<BlockAddress>(Val) && "Expecting a constant blockaddress")(static_cast<void> (0));
1684 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1685 }
1686
1687 // If we have exactly one destination, remember it for efficiency below.
1688 if (PredToDestList.empty()) {
1689 OnlyDest = DestBB;
1690 OnlyVal = Val;
1691 } else {
1692 if (OnlyDest != DestBB)
1693 OnlyDest = MultipleDestSentinel;
1694 // It possible we have same destination, but different value, e.g. default
1695 // case in switchinst.
1696 if (Val != OnlyVal)
1697 OnlyVal = MultipleVal;
1698 }
1699
1700 // If the predecessor ends with an indirect goto, we can't change its
1701 // destination. Same for CallBr.
1702 if (isa<IndirectBrInst>(Pred->getTerminator()) ||
1703 isa<CallBrInst>(Pred->getTerminator()))
1704 continue;
1705
1706 PredToDestList.emplace_back(Pred, DestBB);
1707 }
1708
1709 // If all edges were unthreadable, we fail.
1710 if (PredToDestList.empty())
1711 return false;
1712
1713 // If all the predecessors go to a single known successor, we want to fold,
1714 // not thread. By doing so, we do not need to duplicate the current block and
1715 // also miss potential opportunities in case we dont/cant duplicate.
1716 if (OnlyDest && OnlyDest != MultipleDestSentinel) {
1717 if (BB->hasNPredecessors(PredToDestList.size())) {
1718 bool SeenFirstBranchToOnlyDest = false;
1719 std::vector <DominatorTree::UpdateType> Updates;
1720 Updates.reserve(BB->getTerminator()->getNumSuccessors() - 1);
1721 for (BasicBlock *SuccBB : successors(BB)) {
1722 if (SuccBB == OnlyDest && !SeenFirstBranchToOnlyDest) {
1723 SeenFirstBranchToOnlyDest = true; // Don't modify the first branch.
1724 } else {
1725 SuccBB->removePredecessor(BB, true); // This is unreachable successor.
1726 Updates.push_back({DominatorTree::Delete, BB, SuccBB});
1727 }
1728 }
1729
1730 // Finally update the terminator.
1731 Instruction *Term = BB->getTerminator();
1732 BranchInst::Create(OnlyDest, Term);
1733 ++NumFolds;
1734 Term->eraseFromParent();
1735 DTU->applyUpdatesPermissive(Updates);
1736 if (HasProfileData)
1737 BPI->eraseBlock(BB);
1738
1739 // If the condition is now dead due to the removal of the old terminator,
1740 // erase it.
1741 if (auto *CondInst = dyn_cast<Instruction>(Cond)) {
1742 if (CondInst->use_empty() && !CondInst->mayHaveSideEffects())
1743 CondInst->eraseFromParent();
1744 // We can safely replace *some* uses of the CondInst if it has
1745 // exactly one value as returned by LVI. RAUW is incorrect in the
1746 // presence of guards and assumes, that have the `Cond` as the use. This
1747 // is because we use the guards/assume to reason about the `Cond` value
1748 // at the end of block, but RAUW unconditionally replaces all uses
1749 // including the guards/assumes themselves and the uses before the
1750 // guard/assume.
1751 else if (OnlyVal && OnlyVal != MultipleVal &&
1752 CondInst->getParent() == BB)
1753 replaceFoldableUses(CondInst, OnlyVal);
1754 }
1755 return true;
1756 }
1757 }
1758
1759 // Determine which is the most common successor. If we have many inputs and
1760 // this block is a switch, we want to start by threading the batch that goes
1761 // to the most popular destination first. If we only know about one
1762 // threadable destination (the common case) we can avoid this.
1763 BasicBlock *MostPopularDest = OnlyDest;
1764
1765 if (MostPopularDest == MultipleDestSentinel) {
1766 // Remove any loop headers from the Dest list, threadEdge conservatively
1767 // won't process them, but we might have other destination that are eligible
1768 // and we still want to process.
1769 erase_if(PredToDestList,
1770 [&](const std::pair<BasicBlock *, BasicBlock *> &PredToDest) {
1771 return LoopHeaders.contains(PredToDest.second);
1772 });
1773
1774 if (PredToDestList.empty())
1775 return false;
1776
1777 MostPopularDest = findMostPopularDest(BB, PredToDestList);
1778 }
1779
1780 // Now that we know what the most popular destination is, factor all
1781 // predecessors that will jump to it into a single predecessor.
1782 SmallVector<BasicBlock*, 16> PredsToFactor;
1783 for (const auto &PredToDest : PredToDestList)
1784 if (PredToDest.second == MostPopularDest) {
1785 BasicBlock *Pred = PredToDest.first;
1786
1787 // This predecessor may be a switch or something else that has multiple
1788 // edges to the block. Factor each of these edges by listing them
1789 // according to # occurrences in PredsToFactor.
1790 for (BasicBlock *Succ : successors(Pred))
1791 if (Succ == BB)
1792 PredsToFactor.push_back(Pred);
1793 }
1794
1795 // If the threadable edges are branching on an undefined value, we get to pick
1796 // the destination that these predecessors should get to.
1797 if (!MostPopularDest)
1798 MostPopularDest = BB->getTerminator()->
1799 getSuccessor(getBestDestForJumpOnUndef(BB));
1800
1801 // Ok, try to thread it!
1802 return tryThreadEdge(BB, PredsToFactor, MostPopularDest);
1803}
1804
1805/// processBranchOnPHI - We have an otherwise unthreadable conditional branch on
1806/// a PHI node (or freeze PHI) in the current block. See if there are any
1807/// simplifications we can do based on inputs to the phi node.
1808bool JumpThreadingPass::processBranchOnPHI(PHINode *PN) {
1809 BasicBlock *BB = PN->getParent();
1810
1811 // TODO: We could make use of this to do it once for blocks with common PHI
1812 // values.
1813 SmallVector<BasicBlock*, 1> PredBBs;
1814 PredBBs.resize(1);
1815
1816 // If any of the predecessor blocks end in an unconditional branch, we can
1817 // *duplicate* the conditional branch into that block in order to further
1818 // encourage jump threading and to eliminate cases where we have branch on a
1819 // phi of an icmp (branch on icmp is much better).
1820 // This is still beneficial when a frozen phi is used as the branch condition
1821 // because it allows CodeGenPrepare to further canonicalize br(freeze(icmp))
1822 // to br(icmp(freeze ...)).
1823 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1824 BasicBlock *PredBB = PN->getIncomingBlock(i);
1825 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1826 if (PredBr->isUnconditional()) {
1827 PredBBs[0] = PredBB;
1828 // Try to duplicate BB into PredBB.
1829 if (duplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1830 return true;
1831 }
1832 }
1833
1834 return false;
1835}
1836
1837/// processBranchOnXOR - We have an otherwise unthreadable conditional branch on
1838/// a xor instruction in the current block. See if there are any
1839/// simplifications we can do based on inputs to the xor.
1840bool JumpThreadingPass::processBranchOnXOR(BinaryOperator *BO) {
1841 BasicBlock *BB = BO->getParent();
1842
1843 // If either the LHS or RHS of the xor is a constant, don't do this
1844 // optimization.
1845 if (isa<ConstantInt>(BO->getOperand(0)) ||
1846 isa<ConstantInt>(BO->getOperand(1)))
1847 return false;
1848
1849 // If the first instruction in BB isn't a phi, we won't be able to infer
1850 // anything special about any particular predecessor.
1851 if (!isa<PHINode>(BB->front()))
1852 return false;
1853
1854 // If this BB is a landing pad, we won't be able to split the edge into it.
1855 if (BB->isEHPad())
1856 return false;
1857
1858 // If we have a xor as the branch input to this block, and we know that the
1859 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1860 // the condition into the predecessor and fix that value to true, saving some
1861 // logical ops on that path and encouraging other paths to simplify.
1862 //
1863 // This copies something like this:
1864 //
1865 // BB:
1866 // %X = phi i1 [1], [%X']
1867 // %Y = icmp eq i32 %A, %B
1868 // %Z = xor i1 %X, %Y
1869 // br i1 %Z, ...
1870 //
1871 // Into:
1872 // BB':
1873 // %Y = icmp ne i32 %A, %B
1874 // br i1 %Y, ...
1875
1876 PredValueInfoTy XorOpValues;
1877 bool isLHS = true;
1878 if (!computeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1879 WantInteger, BO)) {
1880 assert(XorOpValues.empty())(static_cast<void> (0));
1881 if (!computeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1882 WantInteger, BO))
1883 return false;
1884 isLHS = false;
1885 }
1886
1887 assert(!XorOpValues.empty() &&(static_cast<void> (0))
1888 "computeValueKnownInPredecessors returned true with no values")(static_cast<void> (0));
1889
1890 // Scan the information to see which is most popular: true or false. The
1891 // predecessors can be of the set true, false, or undef.
1892 unsigned NumTrue = 0, NumFalse = 0;
1893 for (const auto &XorOpValue : XorOpValues) {
1894 if (isa<UndefValue>(XorOpValue.first))
1895 // Ignore undefs for the count.
1896 continue;
1897 if (cast<ConstantInt>(XorOpValue.first)->isZero())
1898 ++NumFalse;
1899 else
1900 ++NumTrue;
1901 }
1902
1903 // Determine which value to split on, true, false, or undef if neither.
1904 ConstantInt *SplitVal = nullptr;
1905 if (NumTrue > NumFalse)
1906 SplitVal = ConstantInt::getTrue(BB->getContext());
1907 else if (NumTrue != 0 || NumFalse != 0)
1908 SplitVal = ConstantInt::getFalse(BB->getContext());
1909
1910 // Collect all of the blocks that this can be folded into so that we can
1911 // factor this once and clone it once.
1912 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1913 for (const auto &XorOpValue : XorOpValues) {
1914 if (XorOpValue.first != SplitVal && !isa<UndefValue>(XorOpValue.first))
1915 continue;
1916
1917 BlocksToFoldInto.push_back(XorOpValue.second);
1918 }
1919
1920 // If we inferred a value for all of the predecessors, then duplication won't
1921 // help us. However, we can just replace the LHS or RHS with the constant.
1922 if (BlocksToFoldInto.size() ==
1923 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1924 if (!SplitVal) {
1925 // If all preds provide undef, just nuke the xor, because it is undef too.
1926 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1927 BO->eraseFromParent();
1928 } else if (SplitVal->isZero()) {
1929 // If all preds provide 0, replace the xor with the other input.
1930 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1931 BO->eraseFromParent();
1932 } else {
1933 // If all preds provide 1, set the computed value to 1.
1934 BO->setOperand(!isLHS, SplitVal);
1935 }
1936
1937 return true;
1938 }
1939
1940 // If any of predecessors end with an indirect goto, we can't change its
1941 // destination. Same for CallBr.
1942 if (any_of(BlocksToFoldInto, [](BasicBlock *Pred) {
1943 return isa<IndirectBrInst>(Pred->getTerminator()) ||
1944 isa<CallBrInst>(Pred->getTerminator());
1945 }))
1946 return false;
1947
1948 // Try to duplicate BB into PredBB.
1949 return duplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1950}
1951
1952/// addPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1953/// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1954/// NewPred using the entries from OldPred (suitably mapped).
1955static void addPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1956 BasicBlock *OldPred,
1957 BasicBlock *NewPred,
1958 DenseMap<Instruction*, Value*> &ValueMap) {
1959 for (PHINode &PN : PHIBB->phis()) {
1960 // Ok, we have a PHI node. Figure out what the incoming value was for the
1961 // DestBlock.
1962 Value *IV = PN.getIncomingValueForBlock(OldPred);
1963
1964 // Remap the value if necessary.
1965 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1966 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1967 if (I != ValueMap.end())
1968 IV = I->second;
1969 }
1970
1971 PN.addIncoming(IV, NewPred);
1972 }
1973}
1974
1975/// Merge basic block BB into its sole predecessor if possible.
1976bool JumpThreadingPass::maybeMergeBasicBlockIntoOnlyPred(BasicBlock *BB) {
1977 BasicBlock *SinglePred = BB->getSinglePredecessor();
1978 if (!SinglePred)
1979 return false;
1980
1981 const Instruction *TI = SinglePred->getTerminator();
1982 if (TI->isExceptionalTerminator() || TI->getNumSuccessors() != 1 ||
1983 SinglePred == BB || hasAddressTakenAndUsed(BB))
1984 return false;
1985
1986 // If SinglePred was a loop header, BB becomes one.
1987 if (LoopHeaders.erase(SinglePred))
1988 LoopHeaders.insert(BB);
1989
1990 LVI->eraseBlock(SinglePred);
1991 MergeBasicBlockIntoOnlyPred(BB, DTU);
1992
1993 // Now that BB is merged into SinglePred (i.e. SinglePred code followed by
1994 // BB code within one basic block `BB`), we need to invalidate the LVI
1995 // information associated with BB, because the LVI information need not be
1996 // true for all of BB after the merge. For example,
1997 // Before the merge, LVI info and code is as follows:
1998 // SinglePred: <LVI info1 for %p val>
1999 // %y = use of %p
2000 // call @exit() // need not transfer execution to successor.
2001 // assume(%p) // from this point on %p is true
2002 // br label %BB
2003 // BB: <LVI info2 for %p val, i.e. %p is true>
2004 // %x = use of %p
2005 // br label exit
2006 //
2007 // Note that this LVI info for blocks BB and SinglPred is correct for %p
2008 // (info2 and info1 respectively). After the merge and the deletion of the
2009 // LVI info1 for SinglePred. We have the following code:
2010 // BB: <LVI info2 for %p val>
2011 // %y = use of %p
2012 // call @exit()
2013 // assume(%p)
2014 // %x = use of %p <-- LVI info2 is correct from here onwards.
2015 // br label exit
2016 // LVI info2 for BB is incorrect at the beginning of BB.
2017
2018 // Invalidate LVI information for BB if the LVI is not provably true for
2019 // all of BB.
2020 if (!isGuaranteedToTransferExecutionToSuccessor(BB))
2021 LVI->eraseBlock(BB);
2022 return true;
2023}
2024
2025/// Update the SSA form. NewBB contains instructions that are copied from BB.
2026/// ValueMapping maps old values in BB to new ones in NewBB.
2027void JumpThreadingPass::updateSSA(
2028 BasicBlock *BB, BasicBlock *NewBB,
2029 DenseMap<Instruction *, Value *> &ValueMapping) {
2030 // If there were values defined in BB that are used outside the block, then we
2031 // now have to update all uses of the value to use either the original value,
2032 // the cloned value, or some PHI derived value. This can require arbitrary
2033 // PHI insertion, of which we are prepared to do, clean these up now.
2034 SSAUpdater SSAUpdate;
2035 SmallVector<Use *, 16> UsesToRename;
2036
2037 for (Instruction &I : *BB) {
2038 // Scan all uses of this instruction to see if it is used outside of its
2039 // block, and if so, record them in UsesToRename.
2040 for (Use &U : I.uses()) {
2041 Instruction *User = cast<Instruction>(U.getUser());
2042 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
2043 if (UserPN->getIncomingBlock(U) == BB)
2044 continue;
2045 } else if (User->getParent() == BB)
2046 continue;
2047
2048 UsesToRename.push_back(&U);
2049 }
2050
2051 // If there are no uses outside the block, we're done with this instruction.
2052 if (UsesToRename.empty())
2053 continue;
2054 LLVM_DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n")do { } while (false);
2055
2056 // We found a use of I outside of BB. Rename all uses of I that are outside
2057 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
2058 // with the two values we know.
2059 SSAUpdate.Initialize(I.getType(), I.getName());
2060 SSAUpdate.AddAvailableValue(BB, &I);
2061 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[&I]);
2062
2063 while (!UsesToRename.empty())
2064 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
2065 LLVM_DEBUG(dbgs() << "\n")do { } while (false);
2066 }
2067}
2068
2069/// Clone instructions in range [BI, BE) to NewBB. For PHI nodes, we only clone
2070/// arguments that come from PredBB. Return the map from the variables in the
2071/// source basic block to the variables in the newly created basic block.
2072DenseMap<Instruction *, Value *>
2073JumpThreadingPass::cloneInstructions(BasicBlock::iterator BI,
2074 BasicBlock::iterator BE, BasicBlock *NewBB,
2075 BasicBlock *PredBB) {
2076 // We are going to have to map operands from the source basic block to the new
2077 // copy of the block 'NewBB'. If there are PHI nodes in the source basic
2078 // block, evaluate them to account for entry from PredBB.
2079 DenseMap<Instruction *, Value *> ValueMapping;
2080
2081 // Clone the phi nodes of the source basic block into NewBB. The resulting
2082 // phi nodes are trivial since NewBB only has one predecessor, but SSAUpdater
2083 // might need to rewrite the operand of the cloned phi.
2084 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) {
2085 PHINode *NewPN = PHINode::Create(PN->getType(), 1, PN->getName(), NewBB);
2086 NewPN->addIncoming(PN->getIncomingValueForBlock(PredBB), PredBB);
2087 ValueMapping[PN] = NewPN;
2088 }
2089
2090 // Clone noalias scope declarations in the threaded block. When threading a
2091 // loop exit, we would otherwise end up with two idential scope declarations
2092 // visible at the same time.
2093 SmallVector<MDNode *> NoAliasScopes;
2094 DenseMap<MDNode *, MDNode *> ClonedScopes;
2095 LLVMContext &Context = PredBB->getContext();
2096 identifyNoAliasScopesToClone(BI, BE, NoAliasScopes);
2097 cloneNoAliasScopes(NoAliasScopes, ClonedScopes, "thread", Context);
2098
2099 // Clone the non-phi instructions of the source basic block into NewBB,
2100 // keeping track of the mapping and using it to remap operands in the cloned
2101 // instructions.
2102 for (; BI != BE; ++BI) {
2103 Instruction *New = BI->clone();
2104 New->setName(BI->getName());
2105 NewBB->getInstList().push_back(New);
2106 ValueMapping[&*BI] = New;
2107 adaptNoAliasScopes(New, ClonedScopes, Context);
2108
2109 // Remap operands to patch up intra-block references.
2110 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2111 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
2112 DenseMap<Instruction *, Value *>::iterator I = ValueMapping.find(Inst);
2113 if (I != ValueMapping.end())
2114 New->setOperand(i, I->second);
2115 }
2116 }
2117
2118 return ValueMapping;
2119}
2120
2121/// Attempt to thread through two successive basic blocks.
2122bool JumpThreadingPass::maybethreadThroughTwoBasicBlocks(BasicBlock *BB,
2123 Value *Cond) {
2124 // Consider:
2125 //
2126 // PredBB:
2127 // %var = phi i32* [ null, %bb1 ], [ @a, %bb2 ]
2128 // %tobool = icmp eq i32 %cond, 0
2129 // br i1 %tobool, label %BB, label ...
2130 //
2131 // BB:
2132 // %cmp = icmp eq i32* %var, null
2133 // br i1 %cmp, label ..., label ...
2134 //
2135 // We don't know the value of %var at BB even if we know which incoming edge
2136 // we take to BB. However, once we duplicate PredBB for each of its incoming
2137 // edges (say, PredBB1 and PredBB2), we know the value of %var in each copy of
2138 // PredBB. Then we can thread edges PredBB1->BB and PredBB2->BB through BB.
2139
2140 // Require that BB end with a Branch for simplicity.
2141 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
2142 if (!CondBr)
2143 return false;
2144
2145 // BB must have exactly one predecessor.
2146 BasicBlock *PredBB = BB->getSinglePredecessor();
2147 if (!PredBB)
2148 return false;
2149
2150 // Require that PredBB end with a conditional Branch. If PredBB ends with an
2151 // unconditional branch, we should be merging PredBB and BB instead. For
2152 // simplicity, we don't deal with a switch.
2153 BranchInst *PredBBBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
2154 if (!PredBBBranch || PredBBBranch->isUnconditional())
2155 return false;
2156
2157 // If PredBB has exactly one incoming edge, we don't gain anything by copying
2158 // PredBB.
2159 if (PredBB->getSinglePredecessor())
2160 return false;
2161
2162 // Don't thread through PredBB if it contains a successor edge to itself, in
2163 // which case we would infinite loop. Suppose we are threading an edge from
2164 // PredPredBB through PredBB and BB to SuccBB with PredBB containing a
2165 // successor edge to itself. If we allowed jump threading in this case, we
2166 // could duplicate PredBB and BB as, say, PredBB.thread and BB.thread. Since
2167 // PredBB.thread has a successor edge to PredBB, we would immediately come up
2168 // with another jump threading opportunity from PredBB.thread through PredBB
2169 // and BB to SuccBB. This jump threading would repeatedly occur. That is, we
2170 // would keep peeling one iteration from PredBB.
2171 if (llvm::is_contained(successors(PredBB), PredBB))
2172 return false;
2173
2174 // Don't thread across a loop header.
2175 if (LoopHeaders.count(PredBB))
2176 return false;
2177
2178 // Avoid complication with duplicating EH pads.
2179 if (PredBB->isEHPad())
2180 return false;
2181
2182 // Find a predecessor that we can thread. For simplicity, we only consider a
2183 // successor edge out of BB to which we thread exactly one incoming edge into
2184 // PredBB.
2185 unsigned ZeroCount = 0;
2186 unsigned OneCount = 0;
2187 BasicBlock *ZeroPred = nullptr;
2188 BasicBlock *OnePred = nullptr;
2189 for (BasicBlock *P : predecessors(PredBB)) {
2190 if (ConstantInt *CI = dyn_cast_or_null<ConstantInt>(
2191 evaluateOnPredecessorEdge(BB, P, Cond))) {
2192 if (CI->isZero()) {
2193 ZeroCount++;
2194 ZeroPred = P;
2195 } else if (CI->isOne()) {
2196 OneCount++;
2197 OnePred = P;
2198 }
2199 }
2200 }
2201
2202 // Disregard complicated cases where we have to thread multiple edges.
2203 BasicBlock *PredPredBB;
2204 if (ZeroCount == 1) {
2205 PredPredBB = ZeroPred;
2206 } else if (OneCount == 1) {
2207 PredPredBB = OnePred;
2208 } else {
2209 return false;
2210 }
2211
2212 BasicBlock *SuccBB = CondBr->getSuccessor(PredPredBB == ZeroPred);
2213
2214 // If threading to the same block as we come from, we would infinite loop.
2215 if (SuccBB == BB) {
2216 LLVM_DEBUG(dbgs() << " Not threading across BB '" << BB->getName()do { } while (false)
2217 << "' - would thread to self!\n")do { } while (false);
2218 return false;
2219 }
2220
2221 // If threading this would thread across a loop header, don't thread the edge.
2222 // See the comments above findLoopHeaders for justifications and caveats.
2223 if (LoopHeaders.count(BB) || LoopHeaders.count(SuccBB)) {
2224 LLVM_DEBUG({do { } while (false)
2225 bool BBIsHeader = LoopHeaders.count(BB);do { } while (false)
2226 bool SuccIsHeader = LoopHeaders.count(SuccBB);do { } while (false)
2227 dbgs() << " Not threading across "do { } while (false)
2228 << (BBIsHeader ? "loop header BB '" : "block BB '")do { } while (false)
2229 << BB->getName() << "' to dest "do { } while (false)
2230 << (SuccIsHeader ? "loop header BB '" : "block BB '")do { } while (false)
2231 << SuccBB->getName()do { } while (false)
2232 << "' - it might create an irreducible loop!\n";do { } while (false)
2233 })do { } while (false);
2234 return false;
2235 }
2236
2237 // Compute the cost of duplicating BB and PredBB.
2238 unsigned BBCost =
2239 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
2240 unsigned PredBBCost = getJumpThreadDuplicationCost(
2241 PredBB, PredBB->getTerminator(), BBDupThreshold);
2242
2243 // Give up if costs are too high. We need to check BBCost and PredBBCost
2244 // individually before checking their sum because getJumpThreadDuplicationCost
2245 // return (unsigned)~0 for those basic blocks that cannot be duplicated.
2246 if (BBCost > BBDupThreshold || PredBBCost > BBDupThreshold ||
2247 BBCost + PredBBCost > BBDupThreshold) {
2248 LLVM_DEBUG(dbgs() << " Not threading BB '" << BB->getName()do { } while (false)
2249 << "' - Cost is too high: " << PredBBCostdo { } while (false)
2250 << " for PredBB, " << BBCost << "for BB\n")do { } while (false);
2251 return false;
2252 }
2253
2254 // Now we are ready to duplicate PredBB.
2255 threadThroughTwoBasicBlocks(PredPredBB, PredBB, BB, SuccBB);
2256 return true;
2257}
2258
2259void JumpThreadingPass::threadThroughTwoBasicBlocks(BasicBlock *PredPredBB,
2260 BasicBlock *PredBB,
2261 BasicBlock *BB,
2262 BasicBlock *SuccBB) {
2263 LLVM_DEBUG(dbgs() << " Threading through '" << PredBB->getName() << "' and '"do { } while (false)
2264 << BB->getName() << "'\n")do { } while (false);
2265
2266 BranchInst *CondBr = cast<BranchInst>(BB->getTerminator());
2267 BranchInst *PredBBBranch = cast<BranchInst>(PredBB->getTerminator());
2268
2269 BasicBlock *NewBB =
2270 BasicBlock::Create(PredBB->getContext(), PredBB->getName() + ".thread",
2271 PredBB->getParent(), PredBB);
2272 NewBB->moveAfter(PredBB);
2273
2274 // Set the block frequency of NewBB.
2275 if (HasProfileData) {
2276 auto NewBBFreq = BFI->getBlockFreq(PredPredBB) *
2277 BPI->getEdgeProbability(PredPredBB, PredBB);
2278 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
2279 }
2280
2281 // We are going to have to map operands from the original BB block to the new
2282 // copy of the block 'NewBB'. If there are PHI nodes in PredBB, evaluate them
2283 // to account for entry from PredPredBB.
2284 DenseMap<Instruction *, Value *> ValueMapping =
2285 cloneInstructions(PredBB->begin(), PredBB->end(), NewBB, PredPredBB);
2286
2287 // Copy the edge probabilities from PredBB to NewBB.
2288 if (HasProfileData)
2289 BPI->copyEdgeProbabilities(PredBB, NewBB);
2290
2291 // Update the terminator of PredPredBB to jump to NewBB instead of PredBB.
2292 // This eliminates predecessors from PredPredBB, which requires us to simplify
2293 // any PHI nodes in PredBB.
2294 Instruction *PredPredTerm = PredPredBB->getTerminator();
2295 for (unsigned i = 0, e = PredPredTerm->getNumSuccessors(); i != e; ++i)
2296 if (PredPredTerm->getSuccessor(i) == PredBB) {
2297 PredBB->removePredecessor(PredPredBB, true);
2298 PredPredTerm->setSuccessor(i, NewBB);
2299 }
2300
2301 addPHINodeEntriesForMappedBlock(PredBBBranch->getSuccessor(0), PredBB, NewBB,
2302 ValueMapping);
2303 addPHINodeEntriesForMappedBlock(PredBBBranch->getSuccessor(1), PredBB, NewBB,
2304 ValueMapping);
2305
2306 DTU->applyUpdatesPermissive(
2307 {{DominatorTree::Insert, NewBB, CondBr->getSuccessor(0)},
2308 {DominatorTree::Insert, NewBB, CondBr->getSuccessor(1)},
2309 {DominatorTree::Insert, PredPredBB, NewBB},
2310 {DominatorTree::Delete, PredPredBB, PredBB}});
2311
2312 updateSSA(PredBB, NewBB, ValueMapping);
2313
2314 // Clean up things like PHI nodes with single operands, dead instructions,
2315 // etc.
2316 SimplifyInstructionsInBlock(NewBB, TLI);
2317 SimplifyInstructionsInBlock(PredBB, TLI);
2318
2319 SmallVector<BasicBlock *, 1> PredsToFactor;
2320 PredsToFactor.push_back(NewBB);
2321 threadEdge(BB, PredsToFactor, SuccBB);
2322}
2323
2324/// tryThreadEdge - Thread an edge if it's safe and profitable to do so.
2325bool JumpThreadingPass::tryThreadEdge(
2326 BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs,
2327 BasicBlock *SuccBB) {
2328 // If threading to the same block as we come from, we would infinite loop.
2329 if (SuccBB == BB) {
2330 LLVM_DEBUG(dbgs() << " Not threading across BB '" << BB->getName()do { } while (false)
2331 << "' - would thread to self!\n")do { } while (false);
2332 return false;
2333 }
2334
2335 // If threading this would thread across a loop header, don't thread the edge.
2336 // See the comments above findLoopHeaders for justifications and caveats.
2337 if (LoopHeaders.count(BB) || LoopHeaders.count(SuccBB)) {
2338 LLVM_DEBUG({do { } while (false)
2339 bool BBIsHeader = LoopHeaders.count(BB);do { } while (false)
2340 bool SuccIsHeader = LoopHeaders.count(SuccBB);do { } while (false)
2341 dbgs() << " Not threading across "do { } while (false)
2342 << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName()do { } while (false)
2343 << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '")do { } while (false)
2344 << SuccBB->getName() << "' - it might create an irreducible loop!\n";do { } while (false)
2345 })do { } while (false);
2346 return false;
2347 }
2348
2349 unsigned JumpThreadCost =
2350 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
2351 if (JumpThreadCost > BBDupThreshold) {
2352 LLVM_DEBUG(dbgs() << " Not threading BB '" << BB->getName()do { } while (false)
2353 << "' - Cost is too high: " << JumpThreadCost << "\n")do { } while (false);
2354 return false;
2355 }
2356
2357 threadEdge(BB, PredBBs, SuccBB);
2358 return true;
2359}
2360
2361/// threadEdge - We have decided that it is safe and profitable to factor the
2362/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
2363/// across BB. Transform the IR to reflect this change.
2364void JumpThreadingPass::threadEdge(BasicBlock *BB,
2365 const SmallVectorImpl<BasicBlock *> &PredBBs,
2366 BasicBlock *SuccBB) {
2367 assert(SuccBB != BB && "Don't create an infinite loop")(static_cast<void> (0));
2368
2369 assert(!LoopHeaders.count(BB) && !LoopHeaders.count(SuccBB) &&(static_cast<void> (0))
2370 "Don't thread across loop headers")(static_cast<void> (0));
2371
2372 // And finally, do it! Start by factoring the predecessors if needed.
2373 BasicBlock *PredBB;
2374 if (PredBBs.size() == 1)
2375 PredBB = PredBBs[0];
2376 else {
2377 LLVM_DEBUG(dbgs() << " Factoring out " << PredBBs.size()do { } while (false)
2378 << " common predecessors.\n")do { } while (false);
2379 PredBB = splitBlockPreds(BB, PredBBs, ".thr_comm");
2380 }
2381
2382 // And finally, do it!
2383 LLVM_DEBUG(dbgs() << " Threading edge from '" << PredBB->getName()do { } while (false)
2384 << "' to '" << SuccBB->getName()do { } while (false)
2385 << ", across block:\n " << *BB << "\n")do { } while (false);
2386
2387 LVI->threadEdge(PredBB, BB, SuccBB);
2388
2389 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
2390 BB->getName()+".thread",
2391 BB->getParent(), BB);
2392 NewBB->moveAfter(PredBB);
2393
2394 // Set the block frequency of NewBB.
2395 if (HasProfileData) {
2396 auto NewBBFreq =
2397 BFI->getBlockFreq(PredBB) * BPI->getEdgeProbability(PredBB, BB);
2398 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
2399 }
2400
2401 // Copy all the instructions from BB to NewBB except the terminator.
2402 DenseMap<Instruction *, Value *> ValueMapping =
2403 cloneInstructions(BB->begin(), std::prev(BB->end()), NewBB, PredBB);
2404
2405 // We didn't copy the terminator from BB over to NewBB, because there is now
2406 // an unconditional jump to SuccBB. Insert the unconditional jump.
2407 BranchInst *NewBI = BranchInst::Create(SuccBB, NewBB);
2408 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
2409
2410 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
2411 // PHI nodes for NewBB now.
2412 addPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
2413
2414 // Update the terminator of PredBB to jump to NewBB instead of BB. This
2415 // eliminates predecessors from BB, which requires us to simplify any PHI
2416 // nodes in BB.
2417 Instruction *PredTerm = PredBB->getTerminator();
2418 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
2419 if (PredTerm->getSuccessor(i) == BB) {
2420 BB->removePredecessor(PredBB, true);
2421 PredTerm->setSuccessor(i, NewBB);
2422 }
2423
2424 // Enqueue required DT updates.
2425 DTU->applyUpdatesPermissive({{DominatorTree::Insert, NewBB, SuccBB},
2426 {DominatorTree::Insert, PredBB, NewBB},
2427 {DominatorTree::Delete, PredBB, BB}});
2428
2429 updateSSA(BB, NewBB, ValueMapping);
2430
2431 // At this point, the IR is fully up to date and consistent. Do a quick scan
2432 // over the new instructions and zap any that are constants or dead. This
2433 // frequently happens because of phi translation.
2434 SimplifyInstructionsInBlock(NewBB, TLI);
2435
2436 // Update the edge weight from BB to SuccBB, which should be less than before.
2437 updateBlockFreqAndEdgeWeight(PredBB, BB, NewBB, SuccBB);
2438
2439 // Threaded an edge!
2440 ++NumThreads;
2441}
2442
2443/// Create a new basic block that will be the predecessor of BB and successor of
2444/// all blocks in Preds. When profile data is available, update the frequency of
2445/// this new block.
2446BasicBlock *JumpThreadingPass::splitBlockPreds(BasicBlock *BB,
2447 ArrayRef<BasicBlock *> Preds,
2448 const char *Suffix) {
2449 SmallVector<BasicBlock *, 2> NewBBs;
2450
2451 // Collect the frequencies of all predecessors of BB, which will be used to
2452 // update the edge weight of the result of splitting predecessors.
2453 DenseMap<BasicBlock *, BlockFrequency> FreqMap;
2454 if (HasProfileData)
2455 for (auto Pred : Preds)
2456 FreqMap.insert(std::make_pair(
2457 Pred, BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, BB)));
2458
2459 // In the case when BB is a LandingPad block we create 2 new predecessors
2460 // instead of just one.
2461 if (BB->isLandingPad()) {
2462 std::string NewName = std::string(Suffix) + ".split-lp";
2463 SplitLandingPadPredecessors(BB, Preds, Suffix, NewName.c_str(), NewBBs);
2464 } else {
2465 NewBBs.push_back(SplitBlockPredecessors(BB, Preds, Suffix));
2466 }
2467
2468 std::vector<DominatorTree::UpdateType> Updates;
2469 Updates.reserve((2 * Preds.size()) + NewBBs.size());
2470 for (auto NewBB : NewBBs) {
2471 BlockFrequency NewBBFreq(0);
2472 Updates.push_back({DominatorTree::Insert, NewBB, BB});
2473 for (auto Pred : predecessors(NewBB)) {
2474 Updates.push_back({DominatorTree::Delete, Pred, BB});
2475 Updates.push_back({DominatorTree::Insert, Pred, NewBB});
2476 if (HasProfileData) // Update frequencies between Pred -> NewBB.
2477 NewBBFreq += FreqMap.lookup(Pred);
2478 }
2479 if (HasProfileData) // Apply the summed frequency to NewBB.
2480 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
2481 }
2482
2483 DTU->applyUpdatesPermissive(Updates);
2484 return NewBBs[0];
2485}
2486
2487bool JumpThreadingPass::doesBlockHaveProfileData(BasicBlock *BB) {
2488 const Instruction *TI = BB->getTerminator();
2489 assert(TI->getNumSuccessors() > 1 && "not a split")(static_cast<void> (0));
2490
2491 MDNode *WeightsNode = TI->getMetadata(LLVMContext::MD_prof);
2492 if (!WeightsNode)
2493 return false;
2494
2495 MDString *MDName = cast<MDString>(WeightsNode->getOperand(0));
2496 if (MDName->getString() != "branch_weights")
2497 return false;
2498
2499 // Ensure there are weights for all of the successors. Note that the first
2500 // operand to the metadata node is a name, not a weight.
2501 return WeightsNode->getNumOperands() == TI->getNumSuccessors() + 1;
2502}
2503
2504/// Update the block frequency of BB and branch weight and the metadata on the
2505/// edge BB->SuccBB. This is done by scaling the weight of BB->SuccBB by 1 -
2506/// Freq(PredBB->BB) / Freq(BB->SuccBB).
2507void JumpThreadingPass::updateBlockFreqAndEdgeWeight(BasicBlock *PredBB,
2508 BasicBlock *BB,
2509 BasicBlock *NewBB,
2510 BasicBlock *SuccBB) {
2511 if (!HasProfileData)
2512 return;
2513
2514 assert(BFI && BPI && "BFI & BPI should have been created here")(static_cast<void> (0));
2515
2516 // As the edge from PredBB to BB is deleted, we have to update the block
2517 // frequency of BB.
2518 auto BBOrigFreq = BFI->getBlockFreq(BB);
2519 auto NewBBFreq = BFI->getBlockFreq(NewBB);
2520 auto BB2SuccBBFreq = BBOrigFreq * BPI->getEdgeProbability(BB, SuccBB);
2521 auto BBNewFreq = BBOrigFreq - NewBBFreq;
2522 BFI->setBlockFreq(BB, BBNewFreq.getFrequency());
2523
2524 // Collect updated outgoing edges' frequencies from BB and use them to update
2525 // edge probabilities.
2526 SmallVector<uint64_t, 4> BBSuccFreq;
2527 for (BasicBlock *Succ : successors(BB)) {
2528 auto SuccFreq = (Succ == SuccBB)
2529 ? BB2SuccBBFreq - NewBBFreq
2530 : BBOrigFreq * BPI->getEdgeProbability(BB, Succ);
2531 BBSuccFreq.push_back(SuccFreq.getFrequency());
2532 }
2533
2534 uint64_t MaxBBSuccFreq =
2535 *std::max_element(BBSuccFreq.begin(), BBSuccFreq.end());
2536
2537 SmallVector<BranchProbability, 4> BBSuccProbs;
2538 if (MaxBBSuccFreq == 0)
2539 BBSuccProbs.assign(BBSuccFreq.size(),
2540 {1, static_cast<unsigned>(BBSuccFreq.size())});
2541 else {
2542 for (uint64_t Freq : BBSuccFreq)
2543 BBSuccProbs.push_back(
2544 BranchProbability::getBranchProbability(Freq, MaxBBSuccFreq));
2545 // Normalize edge probabilities so that they sum up to one.
2546 BranchProbability::normalizeProbabilities(BBSuccProbs.begin(),
2547 BBSuccProbs.end());
2548 }
2549
2550 // Update edge probabilities in BPI.
2551 BPI->setEdgeProbability(BB, BBSuccProbs);
2552
2553 // Update the profile metadata as well.
2554 //
2555 // Don't do this if the profile of the transformed blocks was statically
2556 // estimated. (This could occur despite the function having an entry
2557 // frequency in completely cold parts of the CFG.)
2558 //
2559 // In this case we don't want to suggest to subsequent passes that the
2560 // calculated weights are fully consistent. Consider this graph:
2561 //
2562 // check_1
2563 // 50% / |
2564 // eq_1 | 50%
2565 // \ |
2566 // check_2
2567 // 50% / |
2568 // eq_2 | 50%
2569 // \ |
2570 // check_3
2571 // 50% / |
2572 // eq_3 | 50%
2573 // \ |
2574 //
2575 // Assuming the blocks check_* all compare the same value against 1, 2 and 3,
2576 // the overall probabilities are inconsistent; the total probability that the
2577 // value is either 1, 2 or 3 is 150%.
2578 //
2579 // As a consequence if we thread eq_1 -> check_2 to check_3, check_2->check_3
2580 // becomes 0%. This is even worse if the edge whose probability becomes 0% is
2581 // the loop exit edge. Then based solely on static estimation we would assume
2582 // the loop was extremely hot.
2583 //
2584 // FIXME this locally as well so that BPI and BFI are consistent as well. We
2585 // shouldn't make edges extremely likely or unlikely based solely on static
2586 // estimation.
2587 if (BBSuccProbs.size() >= 2 && doesBlockHaveProfileData(BB)) {
2588 SmallVector<uint32_t, 4> Weights;
2589 for (auto Prob : BBSuccProbs)
2590 Weights.push_back(Prob.getNumerator());
2591
2592 auto TI = BB->getTerminator();
2593 TI->setMetadata(
2594 LLVMContext::MD_prof,
2595 MDBuilder(TI->getParent()->getContext()).createBranchWeights(Weights));
2596 }
2597}
2598
2599/// duplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
2600/// to BB which contains an i1 PHI node and a conditional branch on that PHI.
2601/// If we can duplicate the contents of BB up into PredBB do so now, this
2602/// improves the odds that the branch will be on an analyzable instruction like
2603/// a compare.
2604bool JumpThreadingPass::duplicateCondBranchOnPHIIntoPred(
2605 BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs) {
2606 assert(!PredBBs.empty() && "Can't handle an empty set")(static_cast<void> (0));
2607
2608 // If BB is a loop header, then duplicating this block outside the loop would
2609 // cause us to transform this into an irreducible loop, don't do this.
2610 // See the comments above findLoopHeaders for justifications and caveats.
2611 if (LoopHeaders.count(BB)) {
2612 LLVM_DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()do { } while (false)
2613 << "' into predecessor block '" << PredBBs[0]->getName()do { } while (false)
2614 << "' - it might create an irreducible loop!\n")do { } while (false);
2615 return false;
2616 }
2617
2618 unsigned DuplicationCost =
2619 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
2620 if (DuplicationCost > BBDupThreshold) {
2621 LLVM_DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()do { } while (false)
2622 << "' - Cost is too high: " << DuplicationCost << "\n")do { } while (false);
2623 return false;
2624 }
2625
2626 // And finally, do it! Start by factoring the predecessors if needed.
2627 std::vector<DominatorTree::UpdateType> Updates;
2628 BasicBlock *PredBB;
2629 if (PredBBs.size() == 1)
2630 PredBB = PredBBs[0];
2631 else {
2632 LLVM_DEBUG(dbgs() << " Factoring out " << PredBBs.size()do { } while (false)
2633 << " common predecessors.\n")do { } while (false);
2634 PredBB = splitBlockPreds(BB, PredBBs, ".thr_comm");
2635 }
2636 Updates.push_back({DominatorTree::Delete, PredBB, BB});
2637
2638 // Okay, we decided to do this! Clone all the instructions in BB onto the end
2639 // of PredBB.
2640 LLVM_DEBUG(dbgs() << " Duplicating block '" << BB->getName()do { } while (false)
2641 << "' into end of '" << PredBB->getName()do { } while (false)
2642 << "' to eliminate branch on phi. Cost: "do { } while (false)
2643 << DuplicationCost << " block is:" << *BB << "\n")do { } while (false);
2644
2645 // Unless PredBB ends with an unconditional branch, split the edge so that we
2646 // can just clone the bits from BB into the end of the new PredBB.
2647 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
2648
2649 if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
2650 BasicBlock *OldPredBB = PredBB;
2651 PredBB = SplitEdge(OldPredBB, BB);
2652 Updates.push_back({DominatorTree::Insert, OldPredBB, PredBB});
2653 Updates.push_back({DominatorTree::Insert, PredBB, BB});
2654 Updates.push_back({DominatorTree::Delete, OldPredBB, BB});
2655 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
2656 }
2657
2658 // We are going to have to map operands from the original BB block into the
2659 // PredBB block. Evaluate PHI nodes in BB.
2660 DenseMap<Instruction*, Value*> ValueMapping;
2661
2662 BasicBlock::iterator BI = BB->begin();
2663 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
2664 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
2665 // Clone the non-phi instructions of BB into PredBB, keeping track of the
2666 // mapping and using it to remap operands in the cloned instructions.
2667 for (; BI != BB->end(); ++BI) {
2668 Instruction *New = BI->clone();
2669
2670 // Remap operands to patch up intra-block references.
2671 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2672 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
2673 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
2674 if (I != ValueMapping.end())
2675 New->setOperand(i, I->second);
2676 }
2677
2678 // If this instruction can be simplified after the operands are updated,
2679 // just use the simplified value instead. This frequently happens due to
2680 // phi translation.
2681 if (Value *IV = SimplifyInstruction(
2682 New,
2683 {BB->getModule()->getDataLayout(), TLI, nullptr, nullptr, New})) {
2684 ValueMapping[&*BI] = IV;
2685 if (!New->mayHaveSideEffects()) {
2686 New->deleteValue();
2687 New = nullptr;
2688 }
2689 } else {
2690 ValueMapping[&*BI] = New;
2691 }
2692 if (New) {
2693 // Otherwise, insert the new instruction into the block.
2694 New->setName(BI->getName());
2695 PredBB->getInstList().insert(OldPredBranch->getIterator(), New);
2696 // Update Dominance from simplified New instruction operands.
2697 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2698 if (BasicBlock *SuccBB = dyn_cast<BasicBlock>(New->getOperand(i)))
2699 Updates.push_back({DominatorTree::Insert, PredBB, SuccBB});
2700 }
2701 }
2702
2703 // Check to see if the targets of the branch had PHI nodes. If so, we need to
2704 // add entries to the PHI nodes for branch from PredBB now.
2705 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
2706 addPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
2707 ValueMapping);
2708 addPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
2709 ValueMapping);
2710
2711 updateSSA(BB, PredBB, ValueMapping);
2712
2713 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
2714 // that we nuked.
2715 BB->removePredecessor(PredBB, true);
2716
2717 // Remove the unconditional branch at the end of the PredBB block.
2718 OldPredBranch->eraseFromParent();
2719 if (HasProfileData)
2720 BPI->copyEdgeProbabilities(BB, PredBB);
2721 DTU->applyUpdatesPermissive(Updates);
2722
2723 ++NumDupes;
2724 return true;
2725}
2726
2727// Pred is a predecessor of BB with an unconditional branch to BB. SI is
2728// a Select instruction in Pred. BB has other predecessors and SI is used in
2729// a PHI node in BB. SI has no other use.
2730// A new basic block, NewBB, is created and SI is converted to compare and
2731// conditional branch. SI is erased from parent.
2732void JumpThreadingPass::unfoldSelectInstr(BasicBlock *Pred, BasicBlock *BB,
2733 SelectInst *SI, PHINode *SIUse,
2734 unsigned Idx) {
2735 // Expand the select.
2736 //
2737 // Pred --
2738 // | v
2739 // | NewBB
2740 // | |
2741 // |-----
2742 // v
2743 // BB
2744 BranchInst *PredTerm = cast<BranchInst>(Pred->getTerminator());
2745 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
2746 BB->getParent(), BB);
2747 // Move the unconditional branch to NewBB.
2748 PredTerm->removeFromParent();
2749 NewBB->getInstList().insert(NewBB->end(), PredTerm);
2750 // Create a conditional branch and update PHI nodes.
2751 auto *BI = BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
2752 BI->applyMergedLocation(PredTerm->getDebugLoc(), SI->getDebugLoc());
2753 SIUse->setIncomingValue(Idx, SI->getFalseValue());
2754 SIUse->addIncoming(SI->getTrueValue(), NewBB);
2755
2756 // The select is now dead.
2757 SI->eraseFromParent();
2758 DTU->applyUpdatesPermissive({{DominatorTree::Insert, NewBB, BB},
2759 {DominatorTree::Insert, Pred, NewBB}});
2760
2761 // Update any other PHI nodes in BB.
2762 for (BasicBlock::iterator BI = BB->begin();
2763 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
2764 if (Phi != SIUse)
2765 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);
2766}
2767
2768bool JumpThreadingPass::tryToUnfoldSelect(SwitchInst *SI, BasicBlock *BB) {
2769 PHINode *CondPHI = dyn_cast<PHINode>(SI->getCondition());
2770
2771 if (!CondPHI || CondPHI->getParent() != BB)
2772 return false;
2773
2774 for (unsigned I = 0, E = CondPHI->getNumIncomingValues(); I != E; ++I) {
2775 BasicBlock *Pred = CondPHI->getIncomingBlock(I);
2776 SelectInst *PredSI = dyn_cast<SelectInst>(CondPHI->getIncomingValue(I));
2777
2778 // The second and third condition can be potentially relaxed. Currently
2779 // the conditions help to simplify the code and allow us to reuse existing
2780 // code, developed for tryToUnfoldSelect(CmpInst *, BasicBlock *)
2781 if (!PredSI || PredSI->getParent() != Pred || !PredSI->hasOneUse())
2782 continue;
2783
2784 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
2785 if (!PredTerm || !PredTerm->isUnconditional())
2786 continue;
2787
2788 unfoldSelectInstr(Pred, BB, PredSI, CondPHI, I);
2789 return true;
2790 }
2791 return false;
2792}
2793
2794/// tryToUnfoldSelect - Look for blocks of the form
2795/// bb1:
2796/// %a = select
2797/// br bb2
2798///
2799/// bb2:
2800/// %p = phi [%a, %bb1] ...
2801/// %c = icmp %p
2802/// br i1 %c
2803///
2804/// And expand the select into a branch structure if one of its arms allows %c
2805/// to be folded. This later enables threading from bb1 over bb2.
2806bool JumpThreadingPass::tryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
2807 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
2808 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
2809 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
2810
2811 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
2812 CondLHS->getParent() != BB)
2813 return false;
2814
2815 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
2816 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
2817 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
2818
2819 // Look if one of the incoming values is a select in the corresponding
2820 // predecessor.
2821 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
2822 continue;
2823
2824 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
2825 if (!PredTerm || !PredTerm->isUnconditional())
2826 continue;
2827
2828 // Now check if one of the select values would allow us to constant fold the
2829 // terminator in BB. We don't do the transform if both sides fold, those
2830 // cases will be threaded in any case.
2831 LazyValueInfo::Tristate LHSFolds =
2832 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
2833 CondRHS, Pred, BB, CondCmp);
2834 LazyValueInfo::Tristate RHSFolds =
2835 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
2836 CondRHS, Pred, BB, CondCmp);
2837 if ((LHSFolds != LazyValueInfo::Unknown ||
2838 RHSFolds != LazyValueInfo::Unknown) &&
2839 LHSFolds != RHSFolds) {
2840 unfoldSelectInstr(Pred, BB, SI, CondLHS, I);
2841 return true;
2842 }
2843 }
2844 return false;
2845}
2846
2847/// tryToUnfoldSelectInCurrBB - Look for PHI/Select or PHI/CMP/Select in the
2848/// same BB in the form
2849/// bb:
2850/// %p = phi [false, %bb1], [true, %bb2], [false, %bb3], [true, %bb4], ...
2851/// %s = select %p, trueval, falseval
2852///
2853/// or
2854///
2855/// bb:
2856/// %p = phi [0, %bb1], [1, %bb2], [0, %bb3], [1, %bb4], ...
2857/// %c = cmp %p, 0
2858/// %s = select %c, trueval, falseval
2859///
2860/// And expand the select into a branch structure. This later enables
2861/// jump-threading over bb in this pass.
2862///
2863/// Using the similar approach of SimplifyCFG::FoldCondBranchOnPHI(), unfold
2864/// select if the associated PHI has at least one constant. If the unfolded
2865/// select is not jump-threaded, it will be folded again in the later
2866/// optimizations.
2867bool JumpThreadingPass::tryToUnfoldSelectInCurrBB(BasicBlock *BB) {
2868 // This transform would reduce the quality of msan diagnostics.
2869 // Disable this transform under MemorySanitizer.
2870 if (BB->getParent()->hasFnAttribute(Attribute::SanitizeMemory))
2871 return false;
2872
2873 // If threading this would thread across a loop header, don't thread the edge.
2874 // See the comments above findLoopHeaders for justifications and caveats.
2875 if (LoopHeaders.count(BB))
2876 return false;
2877
2878 for (BasicBlock::iterator BI = BB->begin();
2879 PHINode *PN = dyn_cast<PHINode>(BI); ++BI) {
2880 // Look for a Phi having at least one constant incoming value.
2881 if (llvm::all_of(PN->incoming_values(),
2882 [](Value *V) { return !isa<ConstantInt>(V); }))
2883 continue;
2884
2885 auto isUnfoldCandidate = [BB](SelectInst *SI, Value *V) {
2886 using namespace PatternMatch;
2887
2888 // Check if SI is in BB and use V as condition.
2889 if (SI->getParent() != BB)
2890 return false;
2891 Value *Cond = SI->getCondition();
2892 bool IsAndOr = match(SI, m_CombineOr(m_LogicalAnd(), m_LogicalOr()));
2893 return Cond && Cond == V && Cond->getType()->isIntegerTy(1) && !IsAndOr;
2894 };
2895
2896 SelectInst *SI = nullptr;
2897 for (Use &U : PN->uses()) {
2898 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(U.getUser())) {
2899 // Look for a ICmp in BB that compares PN with a constant and is the
2900 // condition of a Select.
2901 if (Cmp->getParent() == BB && Cmp->hasOneUse() &&
2902 isa<ConstantInt>(Cmp->getOperand(1 - U.getOperandNo())))
2903 if (SelectInst *SelectI = dyn_cast<SelectInst>(Cmp->user_back()))
2904 if (isUnfoldCandidate(SelectI, Cmp->use_begin()->get())) {
2905 SI = SelectI;
2906 break;
2907 }
2908 } else if (SelectInst *SelectI = dyn_cast<SelectInst>(U.getUser())) {
2909 // Look for a Select in BB that uses PN as condition.
2910 if (isUnfoldCandidate(SelectI, U.get())) {
2911 SI = SelectI;
2912 break;
2913 }
2914 }
2915 }
2916
2917 if (!SI)
2918 continue;
2919 // Expand the select.
2920 Value *Cond = SI->getCondition();
2921 if (InsertFreezeWhenUnfoldingSelect &&
2922 !isGuaranteedNotToBeUndefOrPoison(Cond, nullptr, SI,
2923 &DTU->getDomTree()))
2924 Cond = new FreezeInst(Cond, "cond.fr", SI);
2925 Instruction *Term = SplitBlockAndInsertIfThen(Cond, SI, false);
2926 BasicBlock *SplitBB = SI->getParent();
2927 BasicBlock *NewBB = Term->getParent();
2928 PHINode *NewPN = PHINode::Create(SI->getType(), 2, "", SI);
2929 NewPN->addIncoming(SI->getTrueValue(), Term->getParent());
2930 NewPN->addIncoming(SI->getFalseValue(), BB);
2931 SI->replaceAllUsesWith(NewPN);
2932 SI->eraseFromParent();
2933 // NewBB and SplitBB are newly created blocks which require insertion.
2934 std::vector<DominatorTree::UpdateType> Updates;
2935 Updates.reserve((2 * SplitBB->getTerminator()->getNumSuccessors()) + 3);
2936 Updates.push_back({DominatorTree::Insert, BB, SplitBB});
2937 Updates.push_back({DominatorTree::Insert, BB, NewBB});
2938 Updates.push_back({DominatorTree::Insert, NewBB, SplitBB});
2939 // BB's successors were moved to SplitBB, update DTU accordingly.
2940 for (auto *Succ : successors(SplitBB)) {
2941 Updates.push_back({DominatorTree::Delete, BB, Succ});
2942 Updates.push_back({DominatorTree::Insert, SplitBB, Succ});
2943 }
2944 DTU->applyUpdatesPermissive(Updates);
2945 return true;
2946 }
2947 return false;
2948}
2949
2950/// Try to propagate a guard from the current BB into one of its predecessors
2951/// in case if another branch of execution implies that the condition of this
2952/// guard is always true. Currently we only process the simplest case that
2953/// looks like:
2954///
2955/// Start:
2956/// %cond = ...
2957/// br i1 %cond, label %T1, label %F1
2958/// T1:
2959/// br label %Merge
2960/// F1:
2961/// br label %Merge
2962/// Merge:
2963/// %condGuard = ...
2964/// call void(i1, ...) @llvm.experimental.guard( i1 %condGuard )[ "deopt"() ]
2965///
2966/// And cond either implies condGuard or !condGuard. In this case all the
2967/// instructions before the guard can be duplicated in both branches, and the
2968/// guard is then threaded to one of them.
2969bool JumpThreadingPass::processGuards(BasicBlock *BB) {
2970 using namespace PatternMatch;
2971
2972 // We only want to deal with two predecessors.
2973 BasicBlock *Pred1, *Pred2;
2974 auto PI = pred_begin(BB), PE = pred_end(BB);
2975 if (PI == PE)
2976 return false;
2977 Pred1 = *PI++;
2978 if (PI == PE)
2979 return false;
2980 Pred2 = *PI++;
2981 if (PI != PE)
2982 return false;
2983 if (Pred1 == Pred2)
2984 return false;
2985
2986 // Try to thread one of the guards of the block.
2987 // TODO: Look up deeper than to immediate predecessor?
2988 auto *Parent = Pred1->getSinglePredecessor();
2989 if (!Parent || Parent != Pred2->getSinglePredecessor())
2990 return false;
2991
2992 if (auto *BI = dyn_cast<BranchInst>(Parent->getTerminator()))
2993 for (auto &I : *BB)
2994 if (isGuard(&I) && threadGuard(BB, cast<IntrinsicInst>(&I), BI))
2995 return true;
2996
2997 return false;
2998}
2999
3000/// Try to propagate the guard from BB which is the lower block of a diamond
3001/// to one of its branches, in case if diamond's condition implies guard's
3002/// condition.
3003bool JumpThreadingPass::threadGuard(BasicBlock *BB, IntrinsicInst *Guard,
3004 BranchInst *BI) {
3005 assert(BI->getNumSuccessors() == 2 && "Wrong number of successors?")(static_cast<void> (0));
3006 assert(BI->isConditional() && "Unconditional branch has 2 successors?")(static_cast<void> (0));
3007 Value *GuardCond = Guard->getArgOperand(0);
3008 Value *BranchCond = BI->getCondition();
3009 BasicBlock *TrueDest = BI->getSuccessor(0);
3010 BasicBlock *FalseDest = BI->getSuccessor(1);
3011
3012 auto &DL = BB->getModule()->getDataLayout();
3013 bool TrueDestIsSafe = false;
3014 bool FalseDestIsSafe = false;
3015
3016 // True dest is safe if BranchCond => GuardCond.
3017 auto Impl = isImpliedCondition(BranchCond, GuardCond, DL);
3018 if (Impl && *Impl)
3019 TrueDestIsSafe = true;
3020 else {
3021 // False dest is safe if !BranchCond => GuardCond.
3022 Impl = isImpliedCondition(BranchCond, GuardCond, DL, /* LHSIsTrue */ false);
3023 if (Impl && *Impl)
3024 FalseDestIsSafe = true;
3025 }
3026
3027 if (!TrueDestIsSafe && !FalseDestIsSafe)
3028 return false;
3029
3030 BasicBlock *PredUnguardedBlock = TrueDestIsSafe ? TrueDest : FalseDest;
3031 BasicBlock *PredGuardedBlock = FalseDestIsSafe ? TrueDest : FalseDest;
3032
3033 ValueToValueMapTy UnguardedMapping, GuardedMapping;
3034 Instruction *AfterGuard = Guard->getNextNode();
3035 unsigned Cost = getJumpThreadDuplicationCost(BB, AfterGuard, BBDupThreshold);
3036 if (Cost > BBDupThreshold)
3037 return false;
3038 // Duplicate all instructions before the guard and the guard itself to the
3039 // branch where implication is not proved.
3040 BasicBlock *GuardedBlock = DuplicateInstructionsInSplitBetween(
3041 BB, PredGuardedBlock, AfterGuard, GuardedMapping, *DTU);
3042 assert(GuardedBlock && "Could not create the guarded block?")(static_cast<void> (0));
3043 // Duplicate all instructions before the guard in the unguarded branch.
3044 // Since we have successfully duplicated the guarded block and this block
3045 // has fewer instructions, we expect it to succeed.
3046 BasicBlock *UnguardedBlock = DuplicateInstructionsInSplitBetween(
3047 BB, PredUnguardedBlock, Guard, UnguardedMapping, *DTU);
3048 assert(UnguardedBlock && "Could not create the unguarded block?")(static_cast<void> (0));
3049 LLVM_DEBUG(dbgs() << "Moved guard " << *Guard << " to block "do { } while (false)
3050 << GuardedBlock->getName() << "\n")do { } while (false);
3051 // Some instructions before the guard may still have uses. For them, we need
3052 // to create Phi nodes merging their copies in both guarded and unguarded
3053 // branches. Those instructions that have no uses can be just removed.
3054 SmallVector<Instruction *, 4> ToRemove;
3055 for (auto BI = BB->begin(); &*BI != AfterGuard; ++BI)
3056 if (!isa<PHINode>(&*BI))
3057 ToRemove.push_back(&*BI);
3058
3059 Instruction *InsertionPoint = &*BB->getFirstInsertionPt();
3060 assert(InsertionPoint && "Empty block?")(static_cast<void> (0));
3061 // Substitute with Phis & remove.
3062 for (auto *Inst : reverse(ToRemove)) {
3063 if (!Inst->use_empty()) {
3064 PHINode *NewPN = PHINode::Create(Inst->getType(), 2);
3065 NewPN->addIncoming(UnguardedMapping[Inst], UnguardedBlock);
3066 NewPN->addIncoming(GuardedMapping[Inst], GuardedBlock);
3067 NewPN->insertBefore(InsertionPoint);
3068 Inst->replaceAllUsesWith(NewPN);
3069 }
3070 Inst->eraseFromParent();
3071 }
3072 return true;
3073}

/build/llvm-toolchain-snapshot-14~++20210903100615+fd66b44ec19e/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/MapVector.h"
21#include "llvm/ADT/None.h"
22#include "llvm/ADT/STLExtras.h"
23#include "llvm/ADT/SmallVector.h"
24#include "llvm/ADT/StringRef.h"
25#include "llvm/ADT/Twine.h"
26#include "llvm/ADT/iterator.h"
27#include "llvm/ADT/iterator_range.h"
28#include "llvm/IR/Attributes.h"
29#include "llvm/IR/BasicBlock.h"
30#include "llvm/IR/CallingConv.h"
31#include "llvm/IR/CFG.h"
32#include "llvm/IR/Constant.h"
33#include "llvm/IR/DerivedTypes.h"
34#include "llvm/IR/Function.h"
35#include "llvm/IR/InstrTypes.h"
36#include "llvm/IR/Instruction.h"
37#include "llvm/IR/OperandTraits.h"
38#include "llvm/IR/Type.h"
39#include "llvm/IR/Use.h"
40#include "llvm/IR/User.h"
41#include "llvm/IR/Value.h"
42#include "llvm/Support/AtomicOrdering.h"
43#include "llvm/Support/Casting.h"
44#include "llvm/Support/ErrorHandling.h"
45#include <cassert>
46#include <cstddef>
47#include <cstdint>
48#include <iterator>
49
50namespace llvm {
51
52class APInt;
53class ConstantInt;
54class DataLayout;
55class LLVMContext;
56
57//===----------------------------------------------------------------------===//
58// AllocaInst Class
59//===----------------------------------------------------------------------===//
60
61/// an instruction to allocate memory on the stack
62class AllocaInst : public UnaryInstruction {
63 Type *AllocatedType;
64
65 using AlignmentField = AlignmentBitfieldElementT<0>;
66 using UsedWithInAllocaField = BoolBitfieldElementT<AlignmentField::NextBit>;
67 using SwiftErrorField = BoolBitfieldElementT<UsedWithInAllocaField::NextBit>;
68 static_assert(Bitfield::areContiguous<AlignmentField, UsedWithInAllocaField,
69 SwiftErrorField>(),
70 "Bitfields must be contiguous");
71
72protected:
73 // Note: Instruction needs to be a friend here to call cloneImpl.
74 friend class Instruction;
75
76 AllocaInst *cloneImpl() const;
77
78public:
79 explicit AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize,
80 const Twine &Name, Instruction *InsertBefore);
81 AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize,
82 const Twine &Name, BasicBlock *InsertAtEnd);
83
84 AllocaInst(Type *Ty, unsigned AddrSpace, const Twine &Name,
85 Instruction *InsertBefore);
86 AllocaInst(Type *Ty, unsigned AddrSpace,
87 const Twine &Name, BasicBlock *InsertAtEnd);
88
89 AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize, Align Align,
90 const Twine &Name = "", Instruction *InsertBefore = nullptr);
91 AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize, Align Align,
92 const Twine &Name, BasicBlock *InsertAtEnd);
93
94 /// Return true if there is an allocation size parameter to the allocation
95 /// instruction that is not 1.
96 bool isArrayAllocation() const;
97
98 /// Get the number of elements allocated. For a simple allocation of a single
99 /// element, this will return a constant 1 value.
100 const Value *getArraySize() const { return getOperand(0); }
101 Value *getArraySize() { return getOperand(0); }
102
103 /// Overload to return most specific pointer type.
104 PointerType *getType() const {
105 return cast<PointerType>(Instruction::getType());
106 }
107
108 /// Get allocation size in bits. Returns None if size can't be determined,
109 /// e.g. in case of a VLA.
110 Optional<TypeSize> getAllocationSizeInBits(const DataLayout &DL) const;
111
112 /// Return the type that is being allocated by the instruction.
113 Type *getAllocatedType() const { return AllocatedType; }
114 /// for use only in special circumstances that need to generically
115 /// transform a whole instruction (eg: IR linking and vectorization).
116 void setAllocatedType(Type *Ty) { AllocatedType = Ty; }
117
118 /// Return the alignment of the memory that is being allocated by the
119 /// instruction.
120 Align getAlign() const {
121 return Align(1ULL << getSubclassData<AlignmentField>());
122 }
123
124 void setAlignment(Align Align) {
125 setSubclassData<AlignmentField>(Log2(Align));
126 }
127
128 // FIXME: Remove this one transition to Align is over.
129 unsigned getAlignment() const { return getAlign().value(); }
130
131 /// Return true if this alloca is in the entry block of the function and is a
132 /// constant size. If so, the code generator will fold it into the
133 /// prolog/epilog code, so it is basically free.
134 bool isStaticAlloca() const;
135
136 /// Return true if this alloca is used as an inalloca argument to a call. Such
137 /// allocas are never considered static even if they are in the entry block.
138 bool isUsedWithInAlloca() const {
139 return getSubclassData<UsedWithInAllocaField>();
140 }
141
142 /// Specify whether this alloca is used to represent the arguments to a call.
143 void setUsedWithInAlloca(bool V) {
144 setSubclassData<UsedWithInAllocaField>(V);
145 }
146
147 /// Return true if this alloca is used as a swifterror argument to a call.
148 bool isSwiftError() const { return getSubclassData<SwiftErrorField>(); }
149 /// Specify whether this alloca is used to represent a swifterror.
150 void setSwiftError(bool V) { setSubclassData<SwiftErrorField>(V); }
151
152 // Methods for support type inquiry through isa, cast, and dyn_cast:
153 static bool classof(const Instruction *I) {
154 return (I->getOpcode() == Instruction::Alloca);
155 }
156 static bool classof(const Value *V) {
157 return isa<Instruction>(V) && classof(cast<Instruction>(V));
158 }
159
160private:
161 // Shadow Instruction::setInstructionSubclassData with a private forwarding
162 // method so that subclasses cannot accidentally use it.
163 template <typename Bitfield>
164 void setSubclassData(typename Bitfield::Type Value) {
165 Instruction::setSubclassData<Bitfield>(Value);
166 }
167};
168
169//===----------------------------------------------------------------------===//
170// LoadInst Class
171//===----------------------------------------------------------------------===//
172
173/// An instruction for reading from memory. This uses the SubclassData field in
174/// Value to store whether or not the load is volatile.
175class LoadInst : public UnaryInstruction {
176 using VolatileField = BoolBitfieldElementT<0>;
177 using AlignmentField = AlignmentBitfieldElementT<VolatileField::NextBit>;
178 using OrderingField = AtomicOrderingBitfieldElementT<AlignmentField::NextBit>;
179 static_assert(
180 Bitfield::areContiguous<VolatileField, AlignmentField, OrderingField>(),
181 "Bitfields must be contiguous");
182
183 void AssertOK();
184
185protected:
186 // Note: Instruction needs to be a friend here to call cloneImpl.
187 friend class Instruction;
188
189 LoadInst *cloneImpl() const;
190
191public:
192 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr,
193 Instruction *InsertBefore);
194 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, BasicBlock *InsertAtEnd);
195 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile,
196 Instruction *InsertBefore);
197 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile,
198 BasicBlock *InsertAtEnd);
199 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile,
200 Align Align, Instruction *InsertBefore = nullptr);
201 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile,
202 Align Align, BasicBlock *InsertAtEnd);
203 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile,
204 Align Align, AtomicOrdering Order,
205 SyncScope::ID SSID = SyncScope::System,
206 Instruction *InsertBefore = nullptr);
207 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile,
208 Align Align, AtomicOrdering Order, SyncScope::ID SSID,
209 BasicBlock *InsertAtEnd);
210
211 /// Return true if this is a load from a volatile memory location.
212 bool isVolatile() const { return getSubclassData<VolatileField>(); }
213
214 /// Specify whether this is a volatile load or not.
215 void setVolatile(bool V) { setSubclassData<VolatileField>(V); }
216
217 /// Return the alignment of the access that is being performed.
218 /// FIXME: Remove this function once transition to Align is over.
219 /// Use getAlign() instead.
220 unsigned getAlignment() const { return getAlign().value(); }
221
222 /// Return the alignment of the access that is being performed.
223 Align getAlign() const {
224 return Align(1ULL << (getSubclassData<AlignmentField>()));
225 }
226
227 void setAlignment(Align Align) {
228 setSubclassData<AlignmentField>(Log2(Align));
229 }
230
231 /// Returns the ordering constraint of this load instruction.
232 AtomicOrdering getOrdering() const {
233 return getSubclassData<OrderingField>();
234 }
235 /// Sets the ordering constraint of this load instruction. May not be Release
236 /// or AcquireRelease.
237 void setOrdering(AtomicOrdering Ordering) {
238 setSubclassData<OrderingField>(Ordering);
239 }
240
241 /// Returns the synchronization scope ID of this load instruction.
242 SyncScope::ID getSyncScopeID() const {
243 return SSID;
244 }
245
246 /// Sets the synchronization scope ID of this load instruction.
247 void setSyncScopeID(SyncScope::ID SSID) {
248 this->SSID = SSID;
249 }
250
251 /// Sets the ordering constraint and the synchronization scope ID of this load
252 /// instruction.
253 void setAtomic(AtomicOrdering Ordering,
254 SyncScope::ID SSID = SyncScope::System) {
255 setOrdering(Ordering);
256 setSyncScopeID(SSID);
257 }
258
259 bool isSimple() const { return !isAtomic() && !isVolatile(); }
260
261 bool isUnordered() const {
262 return (getOrdering() == AtomicOrdering::NotAtomic ||
2
Assuming the condition is false
5
Returning the value 1, which participates in a condition later
263 getOrdering() == AtomicOrdering::Unordered) &&
3
Assuming the condition is true
264 !isVolatile();
4
Assuming the condition is true
265 }
266
267 Value *getPointerOperand() { return getOperand(0); }
268 const Value *getPointerOperand() const { return getOperand(0); }
269 static unsigned getPointerOperandIndex() { return 0U; }
270 Type *getPointerOperandType() const { return getPointerOperand()->getType(); }
271
272 /// Returns the address space of the pointer operand.
273 unsigned getPointerAddressSpace() const {
274 return getPointerOperandType()->getPointerAddressSpace();
275 }
276
277 // Methods for support type inquiry through isa, cast, and dyn_cast:
278 static bool classof(const Instruction *I) {
279 return I->getOpcode() == Instruction::Load;
280 }
281 static bool classof(const Value *V) {
282 return isa<Instruction>(V) && classof(cast<Instruction>(V));
283 }
284
285private:
286 // Shadow Instruction::setInstructionSubclassData with a private forwarding
287 // method so that subclasses cannot accidentally use it.
288 template <typename Bitfield>
289 void setSubclassData(typename Bitfield::Type Value) {
290 Instruction::setSubclassData<Bitfield>(Value);
291 }
292
293 /// The synchronization scope ID of this load instruction. Not quite enough
294 /// room in SubClassData for everything, so synchronization scope ID gets its
295 /// own field.
296 SyncScope::ID SSID;
297};
298
299//===----------------------------------------------------------------------===//
300// StoreInst Class
301//===----------------------------------------------------------------------===//
302
303/// An instruction for storing to memory.
304class StoreInst : public Instruction {
305 using VolatileField = BoolBitfieldElementT<0>;
306 using AlignmentField = AlignmentBitfieldElementT<VolatileField::NextBit>;
307 using OrderingField = AtomicOrderingBitfieldElementT<AlignmentField::NextBit>;
308 static_assert(
309 Bitfield::areContiguous<VolatileField, AlignmentField, OrderingField>(),
310 "Bitfields must be contiguous");
311
312 void AssertOK();
313
314protected:
315 // Note: Instruction needs to be a friend here to call cloneImpl.
316 friend class Instruction;
317
318 StoreInst *cloneImpl() const;
319
320public:
321 StoreInst(Value *Val, Value *Ptr, Instruction *InsertBefore);
322 StoreInst(Value *Val, Value *Ptr, BasicBlock *InsertAtEnd);
323 StoreInst(Value *Val, Value *Ptr, bool isVolatile, Instruction *InsertBefore);
324 StoreInst(Value *Val, Value *Ptr, bool isVolatile, BasicBlock *InsertAtEnd);
325 StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align,
326 Instruction *InsertBefore = nullptr);
327 StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align,
328 BasicBlock *InsertAtEnd);
329 StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align,
330 AtomicOrdering Order, SyncScope::ID SSID = SyncScope::System,
331 Instruction *InsertBefore = nullptr);
332 StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align,
333 AtomicOrdering Order, SyncScope::ID SSID, BasicBlock *InsertAtEnd);
334
335 // allocate space for exactly two operands
336 void *operator new(size_t S) { return User::operator new(S, 2); }
337 void operator delete(void *Ptr) { User::operator delete(Ptr); }
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 { (static_cast<void> (0));
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) { (static_cast<void> (0)); 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) { return User::operator new(S, 0); }
466 void operator delete(void *Ptr) { User::operator delete(Ptr); }
467
468 /// Returns the ordering constraint of this fence instruction.
469 AtomicOrdering getOrdering() const {
470 return getSubclassData<OrderingField>();
471 }
472
473 /// Sets the ordering constraint of this fence instruction. May only be
474 /// Acquire, Release, AcquireRelease, or SequentiallyConsistent.
475 void setOrdering(AtomicOrdering Ordering) {
476 setSubclassData<OrderingField>(Ordering);
477 }
478
479 /// Returns the synchronization scope ID of this fence instruction.
480 SyncScope::ID getSyncScopeID() const {
481 return SSID;
482 }
483
484 /// Sets the synchronization scope ID of this fence instruction.
485 void setSyncScopeID(SyncScope::ID SSID) {
486 this->SSID = SSID;
487 }
488
489 // Methods for support type inquiry through isa, cast, and dyn_cast:
490 static bool classof(const Instruction *I) {
491 return I->getOpcode() == Instruction::Fence;
492 }
493 static bool classof(const Value *V) {
494 return isa<Instruction>(V) && classof(cast<Instruction>(V));
495 }
496
497private:
498 // Shadow Instruction::setInstructionSubclassData with a private forwarding
499 // method so that subclasses cannot accidentally use it.
500 template <typename Bitfield>
501 void setSubclassData(typename Bitfield::Type Value) {
502 Instruction::setSubclassData<Bitfield>(Value);
503 }
504
505 /// The synchronization scope ID of this fence instruction. Not quite enough
506 /// room in SubClassData for everything, so synchronization scope ID gets its
507 /// own field.
508 SyncScope::ID SSID;
509};
510
511//===----------------------------------------------------------------------===//
512// AtomicCmpXchgInst Class
513//===----------------------------------------------------------------------===//
514
515/// An instruction that atomically checks whether a
516/// specified value is in a memory location, and, if it is, stores a new value
517/// there. The value returned by this instruction is a pair containing the
518/// original value as first element, and an i1 indicating success (true) or
519/// failure (false) as second element.
520///
521class AtomicCmpXchgInst : public Instruction {
522 void Init(Value *Ptr, Value *Cmp, Value *NewVal, Align Align,
523 AtomicOrdering SuccessOrdering, AtomicOrdering FailureOrdering,
524 SyncScope::ID SSID);
525
526 template <unsigned Offset>
527 using AtomicOrderingBitfieldElement =
528 typename Bitfield::Element<AtomicOrdering, Offset, 3,
529 AtomicOrdering::LAST>;
530
531protected:
532 // Note: Instruction needs to be a friend here to call cloneImpl.
533 friend class Instruction;
534
535 AtomicCmpXchgInst *cloneImpl() const;
536
537public:
538 AtomicCmpXchgInst(Value *Ptr, Value *Cmp, Value *NewVal, Align Alignment,
539 AtomicOrdering SuccessOrdering,
540 AtomicOrdering FailureOrdering, SyncScope::ID SSID,
541 Instruction *InsertBefore = nullptr);
542 AtomicCmpXchgInst(Value *Ptr, Value *Cmp, Value *NewVal, Align Alignment,
543 AtomicOrdering SuccessOrdering,
544 AtomicOrdering FailureOrdering, SyncScope::ID SSID,
545 BasicBlock *InsertAtEnd);
546
547 // allocate space for exactly three operands
548 void *operator new(size_t S) { return User::operator new(S, 3); }
549 void operator delete(void *Ptr) { User::operator delete(Ptr); }
550
551 using VolatileField = BoolBitfieldElementT<0>;
552 using WeakField = BoolBitfieldElementT<VolatileField::NextBit>;
553 using SuccessOrderingField =
554 AtomicOrderingBitfieldElementT<WeakField::NextBit>;
555 using FailureOrderingField =
556 AtomicOrderingBitfieldElementT<SuccessOrderingField::NextBit>;
557 using AlignmentField =
558 AlignmentBitfieldElementT<FailureOrderingField::NextBit>;
559 static_assert(
560 Bitfield::areContiguous<VolatileField, WeakField, SuccessOrderingField,
561 FailureOrderingField, AlignmentField>(),
562 "Bitfields must be contiguous");
563
564 /// Return the alignment of the memory that is being allocated by the
565 /// instruction.
566 Align getAlign() const {
567 return Align(1ULL << getSubclassData<AlignmentField>());
568 }
569
570 void setAlignment(Align Align) {
571 setSubclassData<AlignmentField>(Log2(Align));
572 }
573
574 /// Return true if this is a cmpxchg from a volatile memory
575 /// location.
576 ///
577 bool isVolatile() const { return getSubclassData<VolatileField>(); }
578
579 /// Specify whether this is a volatile cmpxchg.
580 ///
581 void setVolatile(bool V) { setSubclassData<VolatileField>(V); }
582
583 /// Return true if this cmpxchg may spuriously fail.
584 bool isWeak() const { return getSubclassData<WeakField>(); }
585
586 void setWeak(bool IsWeak) { setSubclassData<WeakField>(IsWeak); }
587
588 /// Transparently provide more efficient getOperand methods.
589 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
;
590
591 static bool isValidSuccessOrdering(AtomicOrdering Ordering) {
592 return Ordering != AtomicOrdering::NotAtomic &&
593 Ordering != AtomicOrdering::Unordered;
594 }
595
596 static bool isValidFailureOrdering(AtomicOrdering Ordering) {
597 return Ordering != AtomicOrdering::NotAtomic &&
598 Ordering != AtomicOrdering::Unordered &&
599 Ordering != AtomicOrdering::AcquireRelease &&
600 Ordering != AtomicOrdering::Release;
601 }
602
603 /// Returns the success ordering constraint of this cmpxchg instruction.
604 AtomicOrdering getSuccessOrdering() const {
605 return getSubclassData<SuccessOrderingField>();
606 }
607
608 /// Sets the success ordering constraint of this cmpxchg instruction.
609 void setSuccessOrdering(AtomicOrdering Ordering) {
610 assert(isValidSuccessOrdering(Ordering) &&(static_cast<void> (0))
611 "invalid CmpXchg success ordering")(static_cast<void> (0));
612 setSubclassData<SuccessOrderingField>(Ordering);
613 }
614
615 /// Returns the failure ordering constraint of this cmpxchg instruction.
616 AtomicOrdering getFailureOrdering() const {
617 return getSubclassData<FailureOrderingField>();
618 }
619
620 /// Sets the failure ordering constraint of this cmpxchg instruction.
621 void setFailureOrdering(AtomicOrdering Ordering) {
622 assert(isValidFailureOrdering(Ordering) &&(static_cast<void> (0))
623 "invalid CmpXchg failure ordering")(static_cast<void> (0));
624 setSubclassData<FailureOrderingField>(Ordering);
625 }
626
627 /// Returns a single ordering which is at least as strong as both the
628 /// success and failure orderings for this cmpxchg.
629 AtomicOrdering getMergedOrdering() const {
630 if (getFailureOrdering() == AtomicOrdering::SequentiallyConsistent)
631 return AtomicOrdering::SequentiallyConsistent;
632 if (getFailureOrdering() == AtomicOrdering::Acquire) {
633 if (getSuccessOrdering() == AtomicOrdering::Monotonic)
634 return AtomicOrdering::Acquire;
635 if (getSuccessOrdering() == AtomicOrdering::Release)
636 return AtomicOrdering::AcquireRelease;
637 }
638 return getSuccessOrdering();
639 }
640
641 /// Returns the synchronization scope ID of this cmpxchg instruction.
642 SyncScope::ID getSyncScopeID() const {
643 return SSID;
644 }
645
646 /// Sets the synchronization scope ID of this cmpxchg instruction.
647 void setSyncScopeID(SyncScope::ID SSID) {
648 this->SSID = SSID;
649 }
650
651 Value *getPointerOperand() { return getOperand(0); }
652 const Value *getPointerOperand() const { return getOperand(0); }
653 static unsigned getPointerOperandIndex() { return 0U; }
654
655 Value *getCompareOperand() { return getOperand(1); }
656 const Value *getCompareOperand() const { return getOperand(1); }
657
658 Value *getNewValOperand() { return getOperand(2); }
659 const Value *getNewValOperand() const { return getOperand(2); }
660
661 /// Returns the address space of the pointer operand.
662 unsigned getPointerAddressSpace() const {
663 return getPointerOperand()->getType()->getPointerAddressSpace();
664 }
665
666 /// Returns the strongest permitted ordering on failure, given the
667 /// desired ordering on success.
668 ///
669 /// If the comparison in a cmpxchg operation fails, there is no atomic store
670 /// so release semantics cannot be provided. So this function drops explicit
671 /// Release requests from the AtomicOrdering. A SequentiallyConsistent
672 /// operation would remain SequentiallyConsistent.
673 static AtomicOrdering
674 getStrongestFailureOrdering(AtomicOrdering SuccessOrdering) {
675 switch (SuccessOrdering) {
676 default:
677 llvm_unreachable("invalid cmpxchg success ordering")__builtin_unreachable();
678 case AtomicOrdering::Release:
679 case AtomicOrdering::Monotonic:
680 return AtomicOrdering::Monotonic;
681 case AtomicOrdering::AcquireRelease:
682 case AtomicOrdering::Acquire:
683 return AtomicOrdering::Acquire;
684 case AtomicOrdering::SequentiallyConsistent:
685 return AtomicOrdering::SequentiallyConsistent;
686 }
687 }
688
689 // Methods for support type inquiry through isa, cast, and dyn_cast:
690 static bool classof(const Instruction *I) {
691 return I->getOpcode() == Instruction::AtomicCmpXchg;
692 }
693 static bool classof(const Value *V) {
694 return isa<Instruction>(V) && classof(cast<Instruction>(V));
695 }
696
697private:
698 // Shadow Instruction::setInstructionSubclassData with a private forwarding
699 // method so that subclasses cannot accidentally use it.
700 template <typename Bitfield>
701 void setSubclassData(typename Bitfield::Type Value) {
702 Instruction::setSubclassData<Bitfield>(Value);
703 }
704
705 /// The synchronization scope ID of this cmpxchg instruction. Not quite
706 /// enough room in SubClassData for everything, so synchronization scope ID
707 /// gets its own field.
708 SyncScope::ID SSID;
709};
710
711template <>
712struct OperandTraits<AtomicCmpXchgInst> :
713 public FixedNumOperandTraits<AtomicCmpXchgInst, 3> {
714};
715
716DEFINE_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 { (static_cast<void> (0)); 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) { (static_cast<void> (0)); 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
); }
717
718//===----------------------------------------------------------------------===//
719// AtomicRMWInst Class
720//===----------------------------------------------------------------------===//
721
722/// an instruction that atomically reads a memory location,
723/// combines it with another value, and then stores the result back. Returns
724/// the old value.
725///
726class AtomicRMWInst : public Instruction {
727protected:
728 // Note: Instruction needs to be a friend here to call cloneImpl.
729 friend class Instruction;
730
731 AtomicRMWInst *cloneImpl() const;
732
733public:
734 /// This enumeration lists the possible modifications atomicrmw can make. In
735 /// the descriptions, 'p' is the pointer to the instruction's memory location,
736 /// 'old' is the initial value of *p, and 'v' is the other value passed to the
737 /// instruction. These instructions always return 'old'.
738 enum BinOp : unsigned {
739 /// *p = v
740 Xchg,
741 /// *p = old + v
742 Add,
743 /// *p = old - v
744 Sub,
745 /// *p = old & v
746 And,
747 /// *p = ~(old & v)
748 Nand,
749 /// *p = old | v
750 Or,
751 /// *p = old ^ v
752 Xor,
753 /// *p = old >signed v ? old : v
754 Max,
755 /// *p = old <signed v ? old : v
756 Min,
757 /// *p = old >unsigned v ? old : v
758 UMax,
759 /// *p = old <unsigned v ? old : v
760 UMin,
761
762 /// *p = old + v
763 FAdd,
764
765 /// *p = old - v
766 FSub,
767
768 FIRST_BINOP = Xchg,
769 LAST_BINOP = FSub,
770 BAD_BINOP
771 };
772
773private:
774 template <unsigned Offset>
775 using AtomicOrderingBitfieldElement =
776 typename Bitfield::Element<AtomicOrdering, Offset, 3,
777 AtomicOrdering::LAST>;
778
779 template <unsigned Offset>
780 using BinOpBitfieldElement =
781 typename Bitfield::Element<BinOp, Offset, 4, BinOp::LAST_BINOP>;
782
783public:
784 AtomicRMWInst(BinOp Operation, Value *Ptr, Value *Val, Align Alignment,
785 AtomicOrdering Ordering, SyncScope::ID SSID,
786 Instruction *InsertBefore = nullptr);
787 AtomicRMWInst(BinOp Operation, Value *Ptr, Value *Val, Align Alignment,
788 AtomicOrdering Ordering, SyncScope::ID SSID,
789 BasicBlock *InsertAtEnd);
790
791 // allocate space for exactly two operands
792 void *operator new(size_t S) { return User::operator new(S, 2); }
793 void operator delete(void *Ptr) { User::operator delete(Ptr); }
794
795 using VolatileField = BoolBitfieldElementT<0>;
796 using AtomicOrderingField =
797 AtomicOrderingBitfieldElementT<VolatileField::NextBit>;
798 using OperationField = BinOpBitfieldElement<AtomicOrderingField::NextBit>;
799 using AlignmentField = AlignmentBitfieldElementT<OperationField::NextBit>;
800 static_assert(Bitfield::areContiguous<VolatileField, AtomicOrderingField,
801 OperationField, AlignmentField>(),
802 "Bitfields must be contiguous");
803
804 BinOp getOperation() const { return getSubclassData<OperationField>(); }
805
806 static StringRef getOperationName(BinOp Op);
807
808 static bool isFPOperation(BinOp Op) {
809 switch (Op) {
810 case AtomicRMWInst::FAdd:
811 case AtomicRMWInst::FSub:
812 return true;
813 default:
814 return false;
815 }
816 }
817
818 void setOperation(BinOp Operation) {
819 setSubclassData<OperationField>(Operation);
820 }
821
822 /// Return the alignment of the memory that is being allocated by the
823 /// instruction.
824 Align getAlign() const {
825 return Align(1ULL << getSubclassData<AlignmentField>());
826 }
827
828 void setAlignment(Align Align) {
829 setSubclassData<AlignmentField>(Log2(Align));
830 }
831
832 /// Return true if this is a RMW on a volatile memory location.
833 ///
834 bool isVolatile() const { return getSubclassData<VolatileField>(); }
835
836 /// Specify whether this is a volatile RMW or not.
837 ///
838 void setVolatile(bool V) { setSubclassData<VolatileField>(V); }
839
840 /// Transparently provide more efficient getOperand methods.
841 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
;
842
843 /// Returns the ordering constraint of this rmw instruction.
844 AtomicOrdering getOrdering() const {
845 return getSubclassData<AtomicOrderingField>();
846 }
847
848 /// Sets the ordering constraint of this rmw instruction.
849 void setOrdering(AtomicOrdering Ordering) {
850 assert(Ordering != AtomicOrdering::NotAtomic &&(static_cast<void> (0))
851 "atomicrmw instructions can only be atomic.")(static_cast<void> (0));
852 setSubclassData<AtomicOrderingField>(Ordering);
853 }
854
855 /// Returns the synchronization scope ID of this rmw instruction.
856 SyncScope::ID getSyncScopeID() const {
857 return SSID;
858 }
859
860 /// Sets the synchronization scope ID of this rmw instruction.
861 void setSyncScopeID(SyncScope::ID SSID) {
862 this->SSID = SSID;
863 }
864
865 Value *getPointerOperand() { return getOperand(0); }
866 const Value *getPointerOperand() const { return getOperand(0); }
867 static unsigned getPointerOperandIndex() { return 0U; }
868
869 Value *getValOperand() { return getOperand(1); }
870 const Value *getValOperand() const { return getOperand(1); }
871
872 /// Returns the address space of the pointer operand.
873 unsigned getPointerAddressSpace() const {
874 return getPointerOperand()->getType()->getPointerAddressSpace();
875 }
876
877 bool isFloatingPointOperation() const {
878 return isFPOperation(getOperation());
879 }
880
881 // Methods for support type inquiry through isa, cast, and dyn_cast:
882 static bool classof(const Instruction *I) {
883 return I->getOpcode() == Instruction::AtomicRMW;
884 }
885 static bool classof(const Value *V) {
886 return isa<Instruction>(V) && classof(cast<Instruction>(V));
887 }
888
889private:
890 void Init(BinOp Operation, Value *Ptr, Value *Val, Align Align,
891 AtomicOrdering Ordering, SyncScope::ID SSID);
892
893 // Shadow Instruction::setInstructionSubclassData with a private forwarding
894 // method so that subclasses cannot accidentally use it.
895 template <typename Bitfield>
896 void setSubclassData(typename Bitfield::Type Value) {
897 Instruction::setSubclassData<Bitfield>(Value);
898 }
899
900 /// The synchronization scope ID of this rmw instruction. Not quite enough
901 /// room in SubClassData for everything, so synchronization scope ID gets its
902 /// own field.
903 SyncScope::ID SSID;
904};
905
906template <>
907struct OperandTraits<AtomicRMWInst>
908 : public FixedNumOperandTraits<AtomicRMWInst,2> {
909};
910
911DEFINE_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 { (static_cast<void> (0));
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) { (static_cast<void> (0)); 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); }
912
913//===----------------------------------------------------------------------===//
914// GetElementPtrInst Class
915//===----------------------------------------------------------------------===//
916
917// checkGEPType - Simple wrapper function to give a better assertion failure
918// message on bad indexes for a gep instruction.
919//
920inline Type *checkGEPType(Type *Ty) {
921 assert(Ty && "Invalid GetElementPtrInst indices for type!")(static_cast<void> (0));
922 return Ty;
923}
924
925/// an instruction for type-safe pointer arithmetic to
926/// access elements of arrays and structs
927///
928class GetElementPtrInst : public Instruction {
929 Type *SourceElementType;
930 Type *ResultElementType;
931
932 GetElementPtrInst(const GetElementPtrInst &GEPI);
933
934 /// Constructors - Create a getelementptr instruction with a base pointer an
935 /// list of indices. The first ctor can optionally insert before an existing
936 /// instruction, the second appends the new instruction to the specified
937 /// BasicBlock.
938 inline GetElementPtrInst(Type *PointeeType, Value *Ptr,
939 ArrayRef<Value *> IdxList, unsigned Values,
940 const Twine &NameStr, Instruction *InsertBefore);
941 inline GetElementPtrInst(Type *PointeeType, Value *Ptr,
942 ArrayRef<Value *> IdxList, unsigned Values,
943 const Twine &NameStr, BasicBlock *InsertAtEnd);
944
945 void init(Value *Ptr, ArrayRef<Value *> IdxList, const Twine &NameStr);
946
947protected:
948 // Note: Instruction needs to be a friend here to call cloneImpl.
949 friend class Instruction;
950
951 GetElementPtrInst *cloneImpl() const;
952
953public:
954 static GetElementPtrInst *Create(Type *PointeeType, Value *Ptr,
955 ArrayRef<Value *> IdxList,
956 const Twine &NameStr = "",
957 Instruction *InsertBefore = nullptr) {
958 unsigned Values = 1 + unsigned(IdxList.size());
959 assert(PointeeType && "Must specify element type")(static_cast<void> (0));
960 assert(cast<PointerType>(Ptr->getType()->getScalarType())(static_cast<void> (0))
961 ->isOpaqueOrPointeeTypeMatches(PointeeType))(static_cast<void> (0));
962 return new (Values) GetElementPtrInst(PointeeType, Ptr, IdxList, Values,
963 NameStr, InsertBefore);
964 }
965
966 static GetElementPtrInst *Create(Type *PointeeType, Value *Ptr,
967 ArrayRef<Value *> IdxList,
968 const Twine &NameStr,
969 BasicBlock *InsertAtEnd) {
970 unsigned Values = 1 + unsigned(IdxList.size());
971 assert(PointeeType && "Must specify element type")(static_cast<void> (0));
972 assert(cast<PointerType>(Ptr->getType()->getScalarType())(static_cast<void> (0))
973 ->isOpaqueOrPointeeTypeMatches(PointeeType))(static_cast<void> (0));
974 return new (Values) GetElementPtrInst(PointeeType, Ptr, IdxList, Values,
975 NameStr, InsertAtEnd);
976 }
977
978 LLVM_ATTRIBUTE_DEPRECATED(static GetElementPtrInst *CreateInBounds([[deprecated("Use the version with explicit element type instead"
)]] static GetElementPtrInst *CreateInBounds( Value *Ptr, ArrayRef
<Value *> IdxList, const Twine &NameStr = "", Instruction
*InsertBefore = nullptr)
979 Value *Ptr, ArrayRef<Value *> IdxList, const Twine &NameStr = "",[[deprecated("Use the version with explicit element type instead"
)]] static GetElementPtrInst *CreateInBounds( Value *Ptr, ArrayRef
<Value *> IdxList, const Twine &NameStr = "", Instruction
*InsertBefore = nullptr)
980 Instruction *InsertBefore = nullptr),[[deprecated("Use the version with explicit element type instead"
)]] static GetElementPtrInst *CreateInBounds( Value *Ptr, ArrayRef
<Value *> IdxList, const Twine &NameStr = "", Instruction
*InsertBefore = nullptr)
981 "Use the version with explicit element type instead")[[deprecated("Use the version with explicit element type instead"
)]] static GetElementPtrInst *CreateInBounds( Value *Ptr, ArrayRef
<Value *> IdxList, const Twine &NameStr = "", Instruction
*InsertBefore = nullptr)
{
982 return CreateInBounds(
983 Ptr->getType()->getScalarType()->getPointerElementType(), Ptr, IdxList,
984 NameStr, InsertBefore);
985 }
986
987 /// Create an "inbounds" getelementptr. See the documentation for the
988 /// "inbounds" flag in LangRef.html for details.
989 static GetElementPtrInst *
990 CreateInBounds(Type *PointeeType, Value *Ptr, ArrayRef<Value *> IdxList,
991 const Twine &NameStr = "",
992 Instruction *InsertBefore = nullptr) {
993 GetElementPtrInst *GEP =
994 Create(PointeeType, Ptr, IdxList, NameStr, InsertBefore);
995 GEP->setIsInBounds(true);
996 return GEP;
997 }
998
999 LLVM_ATTRIBUTE_DEPRECATED(static GetElementPtrInst *CreateInBounds([[deprecated("Use the version with explicit element type instead"
)]] static GetElementPtrInst *CreateInBounds( Value *Ptr, ArrayRef
<Value *> IdxList, const Twine &NameStr, BasicBlock
*InsertAtEnd)
1000 Value *Ptr, ArrayRef<Value *> IdxList, const Twine &NameStr,[[deprecated("Use the version with explicit element type instead"
)]] static GetElementPtrInst *CreateInBounds( Value *Ptr, ArrayRef
<Value *> IdxList, const Twine &NameStr, BasicBlock
*InsertAtEnd)
1001 BasicBlock *InsertAtEnd),[[deprecated("Use the version with explicit element type instead"
)]] static GetElementPtrInst *CreateInBounds( Value *Ptr, ArrayRef
<Value *> IdxList, const Twine &NameStr, BasicBlock
*InsertAtEnd)
1002 "Use the version with explicit element type instead")[[deprecated("Use the version with explicit element type instead"
)]] static GetElementPtrInst *CreateInBounds( Value *Ptr, ArrayRef
<Value *> IdxList, const Twine &NameStr, BasicBlock
*InsertAtEnd)
{
1003 return CreateInBounds(
1004 Ptr->getType()->getScalarType()->getPointerElementType(), Ptr, IdxList,
1005 NameStr, InsertAtEnd);
1006 }
1007
1008 static GetElementPtrInst *CreateInBounds(Type *PointeeType, Value *Ptr,
1009 ArrayRef<Value *> IdxList,
1010 const Twine &NameStr,
1011 BasicBlock *InsertAtEnd) {
1012 GetElementPtrInst *GEP =
1013 Create(PointeeType, Ptr, IdxList, NameStr, InsertAtEnd);
1014 GEP->setIsInBounds(true);
1015 return GEP;
1016 }
1017
1018 /// Transparently provide more efficient getOperand methods.
1019 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
;
1020
1021 Type *getSourceElementType() const { return SourceElementType; }
1022
1023 void setSourceElementType(Type *Ty) { SourceElementType = Ty; }
1024 void setResultElementType(Type *Ty) { ResultElementType = Ty; }
1025
1026 Type *getResultElementType() const {
1027 assert(cast<PointerType>(getType()->getScalarType())(static_cast<void> (0))
1028 ->isOpaqueOrPointeeTypeMatches(ResultElementType))(static_cast<void> (0));
1029 return ResultElementType;
1030 }
1031
1032 /// Returns the address space of this instruction's pointer type.
1033 unsigned getAddressSpace() const {
1034 // Note that this is always the same as the pointer operand's address space
1035 // and that is cheaper to compute, so cheat here.
1036 return getPointerAddressSpace();
1037 }
1038
1039 /// Returns the result type of a getelementptr with the given source
1040 /// element type and indexes.
1041 ///
1042 /// Null is returned if the indices are invalid for the specified
1043 /// source element type.
1044 static Type *getIndexedType(Type *Ty, ArrayRef<Value *> IdxList);
1045 static Type *getIndexedType(Type *Ty, ArrayRef<Constant *> IdxList);
1046 static Type *getIndexedType(Type *Ty, ArrayRef<uint64_t> IdxList);
1047
1048 /// Return the type of the element at the given index of an indexable
1049 /// type. This is equivalent to "getIndexedType(Agg, {Zero, Idx})".
1050 ///
1051 /// Returns null if the type can't be indexed, or the given index is not
1052 /// legal for the given type.
1053 static Type *getTypeAtIndex(Type *Ty, Value *Idx);
1054 static Type *getTypeAtIndex(Type *Ty, uint64_t Idx);
1055
1056 inline op_iterator idx_begin() { return op_begin()+1; }
1057 inline const_op_iterator idx_begin() const { return op_begin()+1; }
1058 inline op_iterator idx_end() { return op_end(); }
1059 inline const_op_iterator idx_end() const { return op_end(); }
1060
1061 inline iterator_range<op_iterator> indices() {
1062 return make_range(idx_begin(), idx_end());
1063 }
1064
1065 inline iterator_range<const_op_iterator> indices() const {
1066 return make_range(idx_begin(), idx_end());
1067 }
1068
1069 Value *getPointerOperand() {
1070 return getOperand(0);
1071 }
1072 const Value *getPointerOperand() const {
1073 return getOperand(0);
1074 }
1075 static unsigned getPointerOperandIndex() {
1076 return 0U; // get index for modifying correct operand.
1077 }
1078
1079 /// Method to return the pointer operand as a
1080 /// PointerType.
1081 Type *getPointerOperandType() const {
1082 return getPointerOperand()->getType();
1083 }
1084
1085 /// Returns the address space of the pointer operand.
1086 unsigned getPointerAddressSpace() const {
1087 return getPointerOperandType()->getPointerAddressSpace();
1088 }
1089
1090 /// Returns the pointer type returned by the GEP
1091 /// instruction, which may be a vector of pointers.
1092 static Type *getGEPReturnType(Type *ElTy, Value *Ptr,
1093 ArrayRef<Value *> IdxList) {
1094 PointerType *OrigPtrTy = cast<PointerType>(Ptr->getType()->getScalarType());
1095 unsigned AddrSpace = OrigPtrTy->getAddressSpace();
1096 Type *ResultElemTy = checkGEPType(getIndexedType(ElTy, IdxList));
1097 Type *PtrTy = OrigPtrTy->isOpaque()
1098 ? PointerType::get(OrigPtrTy->getContext(), AddrSpace)
1099 : PointerType::get(ResultElemTy, AddrSpace);
1100 // Vector GEP
1101 if (auto *PtrVTy = dyn_cast<VectorType>(Ptr->getType())) {
1102 ElementCount EltCount = PtrVTy->getElementCount();
1103 return VectorType::get(PtrTy, EltCount);
1104 }
1105 for (Value *Index : IdxList)
1106 if (auto *IndexVTy = dyn_cast<VectorType>(Index->getType())) {
1107 ElementCount EltCount = IndexVTy->getElementCount();
1108 return VectorType::get(PtrTy, EltCount);
1109 }
1110 // Scalar GEP
1111 return PtrTy;
1112 }
1113
1114 unsigned getNumIndices() const { // Note: always non-negative
1115 return getNumOperands() - 1;
1116 }
1117
1118 bool hasIndices() const {
1119 return getNumOperands() > 1;
1120 }
1121
1122 /// Return true if all of the indices of this GEP are
1123 /// zeros. If so, the result pointer and the first operand have the same
1124 /// value, just potentially different types.
1125 bool hasAllZeroIndices() const;
1126
1127 /// Return true if all of the indices of this GEP are
1128 /// constant integers. If so, the result pointer and the first operand have
1129 /// a constant offset between them.
1130 bool hasAllConstantIndices() const;
1131
1132 /// Set or clear the inbounds flag on this GEP instruction.
1133 /// See LangRef.html for the meaning of inbounds on a getelementptr.
1134 void setIsInBounds(bool b = true);
1135
1136 /// Determine whether the GEP has the inbounds flag.
1137 bool isInBounds() const;
1138
1139 /// Accumulate the constant address offset of this GEP if possible.
1140 ///
1141 /// This routine accepts an APInt into which it will accumulate the constant
1142 /// offset of this GEP if the GEP is in fact constant. If the GEP is not
1143 /// all-constant, it returns false and the value of the offset APInt is
1144 /// undefined (it is *not* preserved!). The APInt passed into this routine
1145 /// must be at least as wide as the IntPtr type for the address space of
1146 /// the base GEP pointer.
1147 bool accumulateConstantOffset(const DataLayout &DL, APInt &Offset) const;
1148 bool collectOffset(const DataLayout &DL, unsigned BitWidth,
1149 MapVector<Value *, APInt> &VariableOffsets,
1150 APInt &ConstantOffset) const;
1151 // Methods for support type inquiry through isa, cast, and dyn_cast:
1152 static bool classof(const Instruction *I) {
1153 return (I->getOpcode() == Instruction::GetElementPtr);
1154 }
1155 static bool classof(const Value *V) {
1156 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1157 }
1158};
1159
1160template <>
1161struct OperandTraits<GetElementPtrInst> :
1162 public VariadicOperandTraits<GetElementPtrInst, 1> {
1163};
1164
1165GetElementPtrInst::GetElementPtrInst(Type *PointeeType, Value *Ptr,
1166 ArrayRef<Value *> IdxList, unsigned Values,
1167 const Twine &NameStr,
1168 Instruction *InsertBefore)
1169 : Instruction(getGEPReturnType(PointeeType, Ptr, IdxList), GetElementPtr,
1170 OperandTraits<GetElementPtrInst>::op_end(this) - Values,
1171 Values, InsertBefore),
1172 SourceElementType(PointeeType),
1173 ResultElementType(getIndexedType(PointeeType, IdxList)) {
1174 assert(cast<PointerType>(getType()->getScalarType())(static_cast<void> (0))
1175 ->isOpaqueOrPointeeTypeMatches(ResultElementType))(static_cast<void> (0));
1176 init(Ptr, IdxList, NameStr);
1177}
1178
1179GetElementPtrInst::GetElementPtrInst(Type *PointeeType, Value *Ptr,
1180 ArrayRef<Value *> IdxList, unsigned Values,
1181 const Twine &NameStr,
1182 BasicBlock *InsertAtEnd)
1183 : Instruction(getGEPReturnType(PointeeType, Ptr, IdxList), GetElementPtr,
1184 OperandTraits<GetElementPtrInst>::op_end(this) - Values,
1185 Values, InsertAtEnd),
1186 SourceElementType(PointeeType),
1187 ResultElementType(getIndexedType(PointeeType, IdxList)) {
1188 assert(cast<PointerType>(getType()->getScalarType())(static_cast<void> (0))
1189 ->isOpaqueOrPointeeTypeMatches(ResultElementType))(static_cast<void> (0));
1190 init(Ptr, IdxList, NameStr);
1191}
1192
1193DEFINE_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 { (static_cast<void> (0)); 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) { (static_cast<void> (0)); 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
); }
1194
1195//===----------------------------------------------------------------------===//
1196// ICmpInst Class
1197//===----------------------------------------------------------------------===//
1198
1199/// This instruction compares its operands according to the predicate given
1200/// to the constructor. It only operates on integers or pointers. The operands
1201/// must be identical types.
1202/// Represent an integer comparison operator.
1203class ICmpInst: public CmpInst {
1204 void AssertOK() {
1205 assert(isIntPredicate() &&(static_cast<void> (0))
1206 "Invalid ICmp predicate value")(static_cast<void> (0));
1207 assert(getOperand(0)->getType() == getOperand(1)->getType() &&(static_cast<void> (0))
1208 "Both operands to ICmp instruction are not of the same type!")(static_cast<void> (0));
1209 // Check that the operands are the right type
1210 assert((getOperand(0)->getType()->isIntOrIntVectorTy() ||(static_cast<void> (0))
1211 getOperand(0)->getType()->isPtrOrPtrVectorTy()) &&(static_cast<void> (0))
1212 "Invalid operand types for ICmp instruction")(static_cast<void> (0));
1213 }
1214
1215protected:
1216 // Note: Instruction needs to be a friend here to call cloneImpl.
1217 friend class Instruction;
1218
1219 /// Clone an identical ICmpInst
1220 ICmpInst *cloneImpl() const;
1221
1222public:
1223 /// Constructor with insert-before-instruction semantics.
1224 ICmpInst(
1225 Instruction *InsertBefore, ///< Where to insert
1226 Predicate pred, ///< The predicate to use for the comparison
1227 Value *LHS, ///< The left-hand-side of the expression
1228 Value *RHS, ///< The right-hand-side of the expression
1229 const Twine &NameStr = "" ///< Name of the instruction
1230 ) : CmpInst(makeCmpResultType(LHS->getType()),
1231 Instruction::ICmp, pred, LHS, RHS, NameStr,
1232 InsertBefore) {
1233#ifndef NDEBUG1
1234 AssertOK();
1235#endif
1236 }
1237
1238 /// Constructor with insert-at-end semantics.
1239 ICmpInst(
1240 BasicBlock &InsertAtEnd, ///< Block to insert into.
1241 Predicate pred, ///< The predicate to use for the comparison
1242 Value *LHS, ///< The left-hand-side of the expression
1243 Value *RHS, ///< The right-hand-side of the expression
1244 const Twine &NameStr = "" ///< Name of the instruction
1245 ) : CmpInst(makeCmpResultType(LHS->getType()),
1246 Instruction::ICmp, pred, LHS, RHS, NameStr,
1247 &InsertAtEnd) {
1248#ifndef NDEBUG1
1249 AssertOK();
1250#endif
1251 }
1252
1253 /// Constructor with no-insertion semantics
1254 ICmpInst(
1255 Predicate pred, ///< The predicate to use for the comparison
1256 Value *LHS, ///< The left-hand-side of the expression
1257 Value *RHS, ///< The right-hand-side of the expression
1258 const Twine &NameStr = "" ///< Name of the instruction
1259 ) : CmpInst(makeCmpResultType(LHS->getType()),
1260 Instruction::ICmp, pred, LHS, RHS, NameStr) {
1261#ifndef NDEBUG1
1262 AssertOK();
1263#endif
1264 }
1265
1266 /// For example, EQ->EQ, SLE->SLE, UGT->SGT, etc.
1267 /// @returns the predicate that would be the result if the operand were
1268 /// regarded as signed.
1269 /// Return the signed version of the predicate
1270 Predicate getSignedPredicate() const {
1271 return getSignedPredicate(getPredicate());
1272 }
1273
1274 /// This is a static version that you can use without an instruction.
1275 /// Return the signed version of the predicate.
1276 static Predicate getSignedPredicate(Predicate pred);
1277
1278 /// For example, EQ->EQ, SLE->ULE, UGT->UGT, etc.
1279 /// @returns the predicate that would be the result if the operand were
1280 /// regarded as unsigned.
1281 /// Return the unsigned version of the predicate
1282 Predicate getUnsignedPredicate() const {
1283 return getUnsignedPredicate(getPredicate());
1284 }
1285
1286 /// This is a static version that you can use without an instruction.
1287 /// Return the unsigned version of the predicate.
1288 static Predicate getUnsignedPredicate(Predicate pred);
1289
1290 /// Return true if this predicate is either EQ or NE. This also
1291 /// tests for commutativity.
1292 static bool isEquality(Predicate P) {
1293 return P == ICMP_EQ || P == ICMP_NE;
1294 }
1295
1296 /// Return true if this predicate is either EQ or NE. This also
1297 /// tests for commutativity.
1298 bool isEquality() const {
1299 return isEquality(getPredicate());
1300 }
1301
1302 /// @returns true if the predicate of this ICmpInst is commutative
1303 /// Determine if this relation is commutative.
1304 bool isCommutative() const { return isEquality(); }
1305
1306 /// Return true if the predicate is relational (not EQ or NE).
1307 ///
1308 bool isRelational() const {
1309 return !isEquality();
1310 }
1311
1312 /// Return true if the predicate is relational (not EQ or NE).
1313 ///
1314 static bool isRelational(Predicate P) {
1315 return !isEquality(P);
1316 }
1317
1318 /// Return true if the predicate is SGT or UGT.
1319 ///
1320 static bool isGT(Predicate P) {
1321 return P == ICMP_SGT || P == ICMP_UGT;
1322 }
1323
1324 /// Return true if the predicate is SLT or ULT.
1325 ///
1326 static bool isLT(Predicate P) {
1327 return P == ICMP_SLT || P == ICMP_ULT;
1328 }
1329
1330 /// Return true if the predicate is SGE or UGE.
1331 ///
1332 static bool isGE(Predicate P) {
1333 return P == ICMP_SGE || P == ICMP_UGE;
1334 }
1335
1336 /// Return true if the predicate is SLE or ULE.
1337 ///
1338 static bool isLE(Predicate P) {
1339 return P == ICMP_SLE || P == ICMP_ULE;
1340 }
1341
1342 /// Exchange the two operands to this instruction in such a way that it does
1343 /// not modify the semantics of the instruction. The predicate value may be
1344 /// changed to retain the same result if the predicate is order dependent
1345 /// (e.g. ult).
1346 /// Swap operands and adjust predicate.
1347 void swapOperands() {
1348 setPredicate(getSwappedPredicate());
1349 Op<0>().swap(Op<1>());
1350 }
1351
1352 // Methods for support type inquiry through isa, cast, and dyn_cast:
1353 static bool classof(const Instruction *I) {
1354 return I->getOpcode() == Instruction::ICmp;
1355 }
1356 static bool classof(const Value *V) {
1357 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1358 }
1359};
1360
1361//===----------------------------------------------------------------------===//
1362// FCmpInst Class
1363//===----------------------------------------------------------------------===//
1364
1365/// This instruction compares its operands according to the predicate given
1366/// to the constructor. It only operates on floating point values or packed
1367/// vectors of floating point values. The operands must be identical types.
1368/// Represents a floating point comparison operator.
1369class FCmpInst: public CmpInst {
1370 void AssertOK() {
1371 assert(isFPPredicate() && "Invalid FCmp predicate value")(static_cast<void> (0));
1372 assert(getOperand(0)->getType() == getOperand(1)->getType() &&(static_cast<void> (0))
1373 "Both operands to FCmp instruction are not of the same type!")(static_cast<void> (0));
1374 // Check that the operands are the right type
1375 assert(getOperand(0)->getType()->isFPOrFPVectorTy() &&(static_cast<void> (0))
1376 "Invalid operand types for FCmp instruction")(static_cast<void> (0));
1377 }
1378
1379protected:
1380 // Note: Instruction needs to be a friend here to call cloneImpl.
1381 friend class Instruction;
1382
1383 /// Clone an identical FCmpInst
1384 FCmpInst *cloneImpl() const;
1385
1386public:
1387 /// Constructor with insert-before-instruction semantics.
1388 FCmpInst(
1389 Instruction *InsertBefore, ///< Where to insert
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 ) : CmpInst(makeCmpResultType(LHS->getType()),
1395 Instruction::FCmp, pred, LHS, RHS, NameStr,
1396 InsertBefore) {
1397 AssertOK();
1398 }
1399
1400 /// Constructor with insert-at-end semantics.
1401 FCmpInst(
1402 BasicBlock &InsertAtEnd, ///< Block to insert into.
1403 Predicate pred, ///< The predicate to use for the comparison
1404 Value *LHS, ///< The left-hand-side of the expression
1405 Value *RHS, ///< The right-hand-side of the expression
1406 const Twine &NameStr = "" ///< Name of the instruction
1407 ) : CmpInst(makeCmpResultType(LHS->getType()),
1408 Instruction::FCmp, pred, LHS, RHS, NameStr,
1409 &InsertAtEnd) {
1410 AssertOK();
1411 }
1412
1413 /// Constructor with no-insertion semantics
1414 FCmpInst(
1415 Predicate Pred, ///< The predicate to use for the comparison
1416 Value *LHS, ///< The left-hand-side of the expression
1417 Value *RHS, ///< The right-hand-side of the expression
1418 const Twine &NameStr = "", ///< Name of the instruction
1419 Instruction *FlagsSource = nullptr
1420 ) : CmpInst(makeCmpResultType(LHS->getType()), Instruction::FCmp, Pred, LHS,
1421 RHS, NameStr, nullptr, FlagsSource) {
1422 AssertOK();
1423 }
1424
1425 /// @returns true if the predicate of this instruction is EQ or NE.
1426 /// Determine if this is an equality predicate.
1427 static bool isEquality(Predicate Pred) {
1428 return Pred == FCMP_OEQ || Pred == FCMP_ONE || Pred == FCMP_UEQ ||
1429 Pred == FCMP_UNE;
1430 }
1431
1432 /// @returns true if the predicate of this instruction is EQ or NE.
1433 /// Determine if this is an equality predicate.
1434 bool isEquality() const { return isEquality(getPredicate()); }
1435
1436 /// @returns true if the predicate of this instruction is commutative.
1437 /// Determine if this is a commutative predicate.
1438 bool isCommutative() const {
1439 return isEquality() ||
1440 getPredicate() == FCMP_FALSE ||
1441 getPredicate() == FCMP_TRUE ||
1442 getPredicate() == FCMP_ORD ||
1443 getPredicate() == FCMP_UNO;
1444 }
1445
1446 /// @returns true if the predicate is relational (not EQ or NE).
1447 /// Determine if this a relational predicate.
1448 bool isRelational() const { return !isEquality(); }
1449
1450 /// Exchange the two operands to this instruction in such a way that it does
1451 /// not modify the semantics of the instruction. The predicate value may be
1452 /// changed to retain the same result if the predicate is order dependent
1453 /// (e.g. ult).
1454 /// Swap operands and adjust predicate.
1455 void swapOperands() {
1456 setPredicate(getSwappedPredicate());
1457 Op<0>().swap(Op<1>());
1458 }
1459
1460 /// Methods for support type inquiry through isa, cast, and dyn_cast:
1461 static bool classof(const Instruction *I) {
1462 return I->getOpcode() == Instruction::FCmp;
1463 }
1464 static bool classof(const Value *V) {
1465 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1466 }
1467};
1468
1469//===----------------------------------------------------------------------===//
1470/// This class represents a function call, abstracting a target
1471/// machine's calling convention. This class uses low bit of the SubClassData
1472/// field to indicate whether or not this is a tail call. The rest of the bits
1473/// hold the calling convention of the call.
1474///
1475class CallInst : public CallBase {
1476 CallInst(const CallInst &CI);
1477
1478 /// Construct a CallInst given a range of arguments.
1479 /// Construct a CallInst from a range of arguments
1480 inline CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1481 ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr,
1482 Instruction *InsertBefore);
1483
1484 inline CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1485 const Twine &NameStr, Instruction *InsertBefore)
1486 : CallInst(Ty, Func, Args, None, NameStr, InsertBefore) {}
1487
1488 /// Construct a CallInst given a range of arguments.
1489 /// Construct a CallInst from a range of arguments
1490 inline CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1491 ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr,
1492 BasicBlock *InsertAtEnd);
1493
1494 explicit CallInst(FunctionType *Ty, Value *F, const Twine &NameStr,
1495 Instruction *InsertBefore);
1496
1497 CallInst(FunctionType *ty, Value *F, const Twine &NameStr,
1498 BasicBlock *InsertAtEnd);
1499
1500 void init(FunctionType *FTy, Value *Func, ArrayRef<Value *> Args,
1501 ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr);
1502 void init(FunctionType *FTy, Value *Func, const Twine &NameStr);
1503
1504 /// Compute the number of operands to allocate.
1505 static int ComputeNumOperands(int NumArgs, int NumBundleInputs = 0) {
1506 // We need one operand for the called function, plus the input operand
1507 // counts provided.
1508 return 1 + NumArgs + NumBundleInputs;
1509 }
1510
1511protected:
1512 // Note: Instruction needs to be a friend here to call cloneImpl.
1513 friend class Instruction;
1514
1515 CallInst *cloneImpl() const;
1516
1517public:
1518 static CallInst *Create(FunctionType *Ty, Value *F, const Twine &NameStr = "",
1519 Instruction *InsertBefore = nullptr) {
1520 return new (ComputeNumOperands(0)) CallInst(Ty, F, NameStr, InsertBefore);
1521 }
1522
1523 static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1524 const Twine &NameStr,
1525 Instruction *InsertBefore = nullptr) {
1526 return new (ComputeNumOperands(Args.size()))
1527 CallInst(Ty, Func, Args, None, NameStr, InsertBefore);
1528 }
1529
1530 static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1531 ArrayRef<OperandBundleDef> Bundles = None,
1532 const Twine &NameStr = "",
1533 Instruction *InsertBefore = nullptr) {
1534 const int NumOperands =
1535 ComputeNumOperands(Args.size(), CountBundleInputs(Bundles));
1536 const unsigned DescriptorBytes = Bundles.size() * sizeof(BundleOpInfo);
1537
1538 return new (NumOperands, DescriptorBytes)
1539 CallInst(Ty, Func, Args, Bundles, NameStr, InsertBefore);
1540 }
1541
1542 static CallInst *Create(FunctionType *Ty, Value *F, const Twine &NameStr,
1543 BasicBlock *InsertAtEnd) {
1544 return new (ComputeNumOperands(0)) CallInst(Ty, F, NameStr, InsertAtEnd);
1545 }
1546
1547 static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1548 const Twine &NameStr, BasicBlock *InsertAtEnd) {
1549 return new (ComputeNumOperands(Args.size()))
1550 CallInst(Ty, Func, Args, None, NameStr, InsertAtEnd);
1551 }
1552
1553 static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1554 ArrayRef<OperandBundleDef> Bundles,
1555 const Twine &NameStr, BasicBlock *InsertAtEnd) {
1556 const int NumOperands =
1557 ComputeNumOperands(Args.size(), CountBundleInputs(Bundles));
1558 const unsigned DescriptorBytes = Bundles.size() * sizeof(BundleOpInfo);
1559
1560 return new (NumOperands, DescriptorBytes)
1561 CallInst(Ty, Func, Args, Bundles, NameStr, InsertAtEnd);
1562 }
1563
1564 static CallInst *Create(FunctionCallee Func, const Twine &NameStr = "",
1565 Instruction *InsertBefore = nullptr) {
1566 return Create(Func.getFunctionType(), Func.getCallee(), NameStr,
1567 InsertBefore);
1568 }
1569
1570 static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args,
1571 ArrayRef<OperandBundleDef> Bundles = None,
1572 const Twine &NameStr = "",
1573 Instruction *InsertBefore = nullptr) {
1574 return Create(Func.getFunctionType(), Func.getCallee(), Args, Bundles,
1575 NameStr, InsertBefore);
1576 }
1577
1578 static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args,
1579 const Twine &NameStr,
1580 Instruction *InsertBefore = nullptr) {
1581 return Create(Func.getFunctionType(), Func.getCallee(), Args, NameStr,
1582 InsertBefore);
1583 }
1584
1585 static CallInst *Create(FunctionCallee Func, const Twine &NameStr,
1586 BasicBlock *InsertAtEnd) {
1587 return Create(Func.getFunctionType(), Func.getCallee(), NameStr,
1588 InsertAtEnd);
1589 }
1590
1591 static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args,
1592 const Twine &NameStr, BasicBlock *InsertAtEnd) {
1593 return Create(Func.getFunctionType(), Func.getCallee(), Args, NameStr,
1594 InsertAtEnd);
1595 }
1596
1597 static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args,
1598 ArrayRef<OperandBundleDef> Bundles,
1599 const Twine &NameStr, BasicBlock *InsertAtEnd) {
1600 return Create(Func.getFunctionType(), Func.getCallee(), Args, Bundles,
1601 NameStr, InsertAtEnd);
1602 }
1603
1604 /// Create a clone of \p CI with a different set of operand bundles and
1605 /// insert it before \p InsertPt.
1606 ///
1607 /// The returned call instruction is identical \p CI in every way except that
1608 /// the operand bundles for the new instruction are set to the operand bundles
1609 /// in \p Bundles.
1610 static CallInst *Create(CallInst *CI, ArrayRef<OperandBundleDef> Bundles,
1611 Instruction *InsertPt = nullptr);
1612
1613 /// Generate the IR for a call to malloc:
1614 /// 1. Compute the malloc call's argument as the specified type's size,
1615 /// possibly multiplied by the array size if the array size is not
1616 /// constant 1.
1617 /// 2. Call malloc with that argument.
1618 /// 3. Bitcast the result of the malloc call to the specified type.
1619 static Instruction *CreateMalloc(Instruction *InsertBefore, Type *IntPtrTy,
1620 Type *AllocTy, Value *AllocSize,
1621 Value *ArraySize = nullptr,
1622 Function *MallocF = nullptr,
1623 const Twine &Name = "");
1624 static Instruction *CreateMalloc(BasicBlock *InsertAtEnd, Type *IntPtrTy,
1625 Type *AllocTy, Value *AllocSize,
1626 Value *ArraySize = nullptr,
1627 Function *MallocF = nullptr,
1628 const Twine &Name = "");
1629 static Instruction *CreateMalloc(Instruction *InsertBefore, Type *IntPtrTy,
1630 Type *AllocTy, Value *AllocSize,
1631 Value *ArraySize = nullptr,
1632 ArrayRef<OperandBundleDef> Bundles = None,
1633 Function *MallocF = nullptr,
1634 const Twine &Name = "");
1635 static Instruction *CreateMalloc(BasicBlock *InsertAtEnd, Type *IntPtrTy,
1636 Type *AllocTy, Value *AllocSize,
1637 Value *ArraySize = nullptr,
1638 ArrayRef<OperandBundleDef> Bundles = None,
1639 Function *MallocF = nullptr,
1640 const Twine &Name = "");
1641 /// Generate the IR for a call to the builtin free function.
1642 static Instruction *CreateFree(Value *Source, Instruction *InsertBefore);
1643 static Instruction *CreateFree(Value *Source, BasicBlock *InsertAtEnd);
1644 static Instruction *CreateFree(Value *Source,
1645 ArrayRef<OperandBundleDef> Bundles,
1646 Instruction *InsertBefore);
1647 static Instruction *CreateFree(Value *Source,
1648 ArrayRef<OperandBundleDef> Bundles,
1649 BasicBlock *InsertAtEnd);
1650
1651 // Note that 'musttail' implies 'tail'.
1652 enum TailCallKind : unsigned {
1653 TCK_None = 0,
1654 TCK_Tail = 1,
1655 TCK_MustTail = 2,
1656 TCK_NoTail = 3,
1657 TCK_LAST = TCK_NoTail
1658 };
1659
1660 using TailCallKindField = Bitfield::Element<TailCallKind, 0, 2, TCK_LAST>;
1661 static_assert(
1662 Bitfield::areContiguous<TailCallKindField, CallBase::CallingConvField>(),
1663 "Bitfields must be contiguous");
1664
1665 TailCallKind getTailCallKind() const {
1666 return getSubclassData<TailCallKindField>();
1667 }
1668
1669 bool isTailCall() const {
1670 TailCallKind Kind = getTailCallKind();
1671 return Kind == TCK_Tail || Kind == TCK_MustTail;
1672 }
1673
1674 bool isMustTailCall() const { return getTailCallKind() == TCK_MustTail; }
1675
1676 bool isNoTailCall() const { return getTailCallKind() == TCK_NoTail; }
1677
1678 void setTailCallKind(TailCallKind TCK) {
1679 setSubclassData<TailCallKindField>(TCK);
1680 }
1681
1682 void setTailCall(bool IsTc = true) {
1683 setTailCallKind(IsTc ? TCK_Tail : TCK_None);
1684 }
1685
1686 /// Return true if the call can return twice
1687 bool canReturnTwice() const { return hasFnAttr(Attribute::ReturnsTwice); }
1688 void setCanReturnTwice() { addFnAttr(Attribute::ReturnsTwice); }
1689
1690 // Methods for support type inquiry through isa, cast, and dyn_cast:
1691 static bool classof(const Instruction *I) {
1692 return I->getOpcode() == Instruction::Call;
1693 }
1694 static bool classof(const Value *V) {
1695 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1696 }
1697
1698 /// Updates profile metadata by scaling it by \p S / \p T.
1699 void updateProfWeight(uint64_t S, uint64_t T);
1700
1701private:
1702 // Shadow Instruction::setInstructionSubclassData with a private forwarding
1703 // method so that subclasses cannot accidentally use it.
1704 template <typename Bitfield>
1705 void setSubclassData(typename Bitfield::Type Value) {
1706 Instruction::setSubclassData<Bitfield>(Value);
1707 }
1708};
1709
1710CallInst::CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1711 ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr,
1712 BasicBlock *InsertAtEnd)
1713 : CallBase(Ty->getReturnType(), Instruction::Call,
1714 OperandTraits<CallBase>::op_end(this) -
1715 (Args.size() + CountBundleInputs(Bundles) + 1),
1716 unsigned(Args.size() + CountBundleInputs(Bundles) + 1),
1717 InsertAtEnd) {
1718 init(Ty, Func, Args, Bundles, NameStr);
1719}
1720
1721CallInst::CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1722 ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr,
1723 Instruction *InsertBefore)
1724 : CallBase(Ty->getReturnType(), Instruction::Call,
1725 OperandTraits<CallBase>::op_end(this) -
1726 (Args.size() + CountBundleInputs(Bundles) + 1),
1727 unsigned(Args.size() + CountBundleInputs(Bundles) + 1),
1728 InsertBefore) {
1729 init(Ty, Func, Args, Bundles, NameStr);
1730}
1731
1732//===----------------------------------------------------------------------===//
1733// SelectInst Class
1734//===----------------------------------------------------------------------===//
1735
1736/// This class represents the LLVM 'select' instruction.
1737///
1738class SelectInst : public Instruction {
1739 SelectInst(Value *C, Value *S1, Value *S2, const Twine &NameStr,
1740 Instruction *InsertBefore)
1741 : Instruction(S1->getType(), Instruction::Select,
1742 &Op<0>(), 3, InsertBefore) {
1743 init(C, S1, S2);
1744 setName(NameStr);
1745 }
1746
1747 SelectInst(Value *C, Value *S1, Value *S2, const Twine &NameStr,
1748 BasicBlock *InsertAtEnd)
1749 : Instruction(S1->getType(), Instruction::Select,
1750 &Op<0>(), 3, InsertAtEnd) {
1751 init(C, S1, S2);
1752 setName(NameStr);
1753 }
1754
1755 void init(Value *C, Value *S1, Value *S2) {
1756 assert(!areInvalidOperands(C, S1, S2) && "Invalid operands for select")(static_cast<void> (0));
1757 Op<0>() = C;
1758 Op<1>() = S1;
1759 Op<2>() = S2;
1760 }
1761
1762protected:
1763 // Note: Instruction needs to be a friend here to call cloneImpl.
1764 friend class Instruction;
1765
1766 SelectInst *cloneImpl() const;
1767
1768public:
1769 static SelectInst *Create(Value *C, Value *S1, Value *S2,
1770 const Twine &NameStr = "",
1771 Instruction *InsertBefore = nullptr,
1772 Instruction *MDFrom = nullptr) {
1773 SelectInst *Sel = new(3) SelectInst(C, S1, S2, NameStr, InsertBefore);
1774 if (MDFrom)
1775 Sel->copyMetadata(*MDFrom);
1776 return Sel;
1777 }
1778
1779 static SelectInst *Create(Value *C, Value *S1, Value *S2,
1780 const Twine &NameStr,
1781 BasicBlock *InsertAtEnd) {
1782 return new(3) SelectInst(C, S1, S2, NameStr, InsertAtEnd);
1783 }
1784
1785 const Value *getCondition() const { return Op<0>(); }
1786 const Value *getTrueValue() const { return Op<1>(); }
1787 const Value *getFalseValue() const { return Op<2>(); }
1788 Value *getCondition() { return Op<0>(); }
1789 Value *getTrueValue() { return Op<1>(); }
1790 Value *getFalseValue() { return Op<2>(); }
1791
1792 void setCondition(Value *V) { Op<0>() = V; }
1793 void setTrueValue(Value *V) { Op<1>() = V; }
1794 void setFalseValue(Value *V) { Op<2>() = V; }
1795
1796 /// Swap the true and false values of the select instruction.
1797 /// This doesn't swap prof metadata.
1798 void swapValues() { Op<1>().swap(Op<2>()); }
1799
1800 /// Return a string if the specified operands are invalid
1801 /// for a select operation, otherwise return null.
1802 static const char *areInvalidOperands(Value *Cond, Value *True, Value *False);
1803
1804 /// Transparently provide more efficient getOperand methods.
1805 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
;
1806
1807 OtherOps getOpcode() const {
1808 return static_cast<OtherOps>(Instruction::getOpcode());
1809 }
1810
1811 // Methods for support type inquiry through isa, cast, and dyn_cast:
1812 static bool classof(const Instruction *I) {
1813 return I->getOpcode() == Instruction::Select;
1814 }
1815 static bool classof(const Value *V) {
1816 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1817 }
1818};
1819
1820template <>
1821struct OperandTraits<SelectInst> : public FixedNumOperandTraits<SelectInst, 3> {
1822};
1823
1824DEFINE_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 { (static_cast<void
> (0)); 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) { (static_cast<void> (0)); 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); }
1825
1826//===----------------------------------------------------------------------===//
1827// VAArgInst Class
1828//===----------------------------------------------------------------------===//
1829
1830/// This class represents the va_arg llvm instruction, which returns
1831/// an argument of the specified type given a va_list and increments that list
1832///
1833class VAArgInst : public UnaryInstruction {
1834protected:
1835 // Note: Instruction needs to be a friend here to call cloneImpl.
1836 friend class Instruction;
1837
1838 VAArgInst *cloneImpl() const;
1839
1840public:
1841 VAArgInst(Value *List, Type *Ty, const Twine &NameStr = "",
1842 Instruction *InsertBefore = nullptr)
1843 : UnaryInstruction(Ty, VAArg, List, InsertBefore) {
1844 setName(NameStr);
1845 }
1846
1847 VAArgInst(Value *List, Type *Ty, const Twine &NameStr,
1848 BasicBlock *InsertAtEnd)
1849 : UnaryInstruction(Ty, VAArg, List, InsertAtEnd) {
1850 setName(NameStr);
1851 }
1852
1853 Value *getPointerOperand() { return getOperand(0); }
1854 const Value *getPointerOperand() const { return getOperand(0); }
1855 static unsigned getPointerOperandIndex() { return 0U; }
1856
1857 // Methods for support type inquiry through isa, cast, and dyn_cast:
1858 static bool classof(const Instruction *I) {
1859 return I->getOpcode() == VAArg;
1860 }
1861 static bool classof(const Value *V) {
1862 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1863 }
1864};
1865
1866//===----------------------------------------------------------------------===//
1867// ExtractElementInst Class
1868//===----------------------------------------------------------------------===//
1869
1870/// This instruction extracts a single (scalar)
1871/// element from a VectorType value
1872///
1873class ExtractElementInst : public Instruction {
1874 ExtractElementInst(Value *Vec, Value *Idx, const Twine &NameStr = "",
1875 Instruction *InsertBefore = nullptr);
1876 ExtractElementInst(Value *Vec, Value *Idx, const Twine &NameStr,
1877 BasicBlock *InsertAtEnd);
1878
1879protected:
1880 // Note: Instruction needs to be a friend here to call cloneImpl.
1881 friend class Instruction;
1882
1883 ExtractElementInst *cloneImpl() const;
1884
1885public:
1886 static ExtractElementInst *Create(Value *Vec, Value *Idx,
1887 const Twine &NameStr = "",
1888 Instruction *InsertBefore = nullptr) {
1889 return new(2) ExtractElementInst(Vec, Idx, NameStr, InsertBefore);
1890 }
1891
1892 static ExtractElementInst *Create(Value *Vec, Value *Idx,
1893 const Twine &NameStr,
1894 BasicBlock *InsertAtEnd) {
1895 return new(2) ExtractElementInst(Vec, Idx, NameStr, InsertAtEnd);
1896 }
1897
1898 /// Return true if an extractelement instruction can be
1899 /// formed with the specified operands.
1900 static bool isValidOperands(const Value *Vec, const Value *Idx);
1901
1902 Value *getVectorOperand() { return Op<0>(); }
1903 Value *getIndexOperand() { return Op<1>(); }
1904 const Value *getVectorOperand() const { return Op<0>(); }
1905 const Value *getIndexOperand() const { return Op<1>(); }
1906
1907 VectorType *getVectorOperandType() const {
1908 return cast<VectorType>(getVectorOperand()->getType());
1909 }
1910
1911 /// Transparently provide more efficient getOperand methods.
1912 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
;
1913
1914 // Methods for support type inquiry through isa, cast, and dyn_cast:
1915 static bool classof(const Instruction *I) {
1916 return I->getOpcode() == Instruction::ExtractElement;
1917 }
1918 static bool classof(const Value *V) {
1919 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1920 }
1921};
1922
1923template <>
1924struct OperandTraits<ExtractElementInst> :
1925 public FixedNumOperandTraits<ExtractElementInst, 2> {
1926};
1927
1928DEFINE_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 {
(static_cast<void> (0)); 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) { (static_cast<void> (0)); 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
); }
1929
1930//===----------------------------------------------------------------------===//
1931// InsertElementInst Class
1932//===----------------------------------------------------------------------===//
1933
1934/// This instruction inserts a single (scalar)
1935/// element into a VectorType value
1936///
1937class InsertElementInst : public Instruction {
1938 InsertElementInst(Value *Vec, Value *NewElt, Value *Idx,
1939 const Twine &NameStr = "",
1940 Instruction *InsertBefore = nullptr);
1941 InsertElementInst(Value *Vec, Value *NewElt, Value *Idx, const Twine &NameStr,
1942 BasicBlock *InsertAtEnd);
1943
1944protected:
1945 // Note: Instruction needs to be a friend here to call cloneImpl.
1946 friend class Instruction;
1947
1948 InsertElementInst *cloneImpl() const;
1949
1950public:
1951 static InsertElementInst *Create(Value *Vec, Value *NewElt, Value *Idx,
1952 const Twine &NameStr = "",
1953 Instruction *InsertBefore = nullptr) {
1954 return new(3) InsertElementInst(Vec, NewElt, Idx, NameStr, InsertBefore);
1955 }
1956
1957 static InsertElementInst *Create(Value *Vec, Value *NewElt, Value *Idx,
1958 const Twine &NameStr,
1959 BasicBlock *InsertAtEnd) {
1960 return new(3) InsertElementInst(Vec, NewElt, Idx, NameStr, InsertAtEnd);
1961 }
1962
1963 /// Return true if an insertelement instruction can be
1964 /// formed with the specified operands.
1965 static bool isValidOperands(const Value *Vec, const Value *NewElt,
1966 const Value *Idx);
1967
1968 /// Overload to return most specific vector type.
1969 ///
1970 VectorType *getType() const {
1971 return cast<VectorType>(Instruction::getType());
1972 }
1973
1974 /// Transparently provide more efficient getOperand methods.
1975 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
;
1976
1977 // Methods for support type inquiry through isa, cast, and dyn_cast:
1978 static bool classof(const Instruction *I) {
1979 return I->getOpcode() == Instruction::InsertElement;
1980 }
1981 static bool classof(const Value *V) {
1982 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1983 }
1984};
1985
1986template <>
1987struct OperandTraits<InsertElementInst> :
1988 public FixedNumOperandTraits<InsertElementInst, 3> {
1989};
1990
1991DEFINE_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 { (static_cast<void> (0)); 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) { (static_cast<void> (0)); 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
); }
1992
1993//===----------------------------------------------------------------------===//
1994// ShuffleVectorInst Class
1995//===----------------------------------------------------------------------===//
1996
1997constexpr int UndefMaskElem = -1;
1998
1999/// This instruction constructs a fixed permutation of two
2000/// input vectors.
2001///
2002/// For each element of the result vector, the shuffle mask selects an element
2003/// from one of the input vectors to copy to the result. Non-negative elements
2004/// in the mask represent an index into the concatenated pair of input vectors.
2005/// UndefMaskElem (-1) specifies that the result element is undefined.
2006///
2007/// For scalable vectors, all the elements of the mask must be 0 or -1. This
2008/// requirement may be relaxed in the future.
2009class ShuffleVectorInst : public Instruction {
2010 SmallVector<int, 4> ShuffleMask;
2011 Constant *ShuffleMaskForBitcode;
2012
2013protected:
2014 // Note: Instruction needs to be a friend here to call cloneImpl.
2015 friend class Instruction;
2016
2017 ShuffleVectorInst *cloneImpl() const;
2018
2019public:
2020 ShuffleVectorInst(Value *V1, Value *V2, Value *Mask,
2021 const Twine &NameStr = "",
2022 Instruction *InsertBefor = nullptr);
2023 ShuffleVectorInst(Value *V1, Value *V2, Value *Mask,
2024 const Twine &NameStr, BasicBlock *InsertAtEnd);
2025 ShuffleVectorInst(Value *V1, Value *V2, ArrayRef<int> Mask,
2026 const Twine &NameStr = "",
2027 Instruction *InsertBefor = nullptr);
2028 ShuffleVectorInst(Value *V1, Value *V2, ArrayRef<int> Mask,
2029 const Twine &NameStr, BasicBlock *InsertAtEnd);
2030
2031 void *operator new(size_t S) { return User::operator new(S, 2); }
2032 void operator delete(void *Ptr) { return User::operator delete(Ptr); }
2033
2034 /// Swap the operands and adjust the mask to preserve the semantics
2035 /// of the instruction.
2036 void commute();
2037
2038 /// Return true if a shufflevector instruction can be
2039 /// formed with the specified operands.
2040 static bool isValidOperands(const Value *V1, const Value *V2,
2041 const Value *Mask);
2042 static bool isValidOperands(const Value *V1, const Value *V2,
2043 ArrayRef<int> Mask);
2044
2045 /// Overload to return most specific vector type.
2046 ///
2047 VectorType *getType() const {
2048 return cast<VectorType>(Instruction::getType());
2049 }
2050
2051 /// Transparently provide more efficient getOperand methods.
2052 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
;
2053
2054 /// Return the shuffle mask value of this instruction for the given element
2055 /// index. Return UndefMaskElem if the element is undef.
2056 int getMaskValue(unsigned Elt) const { return ShuffleMask[Elt]; }
2057
2058 /// Convert the input shuffle mask operand to a vector of integers. Undefined
2059 /// elements of the mask are returned as UndefMaskElem.
2060 static void getShuffleMask(const Constant *Mask,
2061 SmallVectorImpl<int> &Result);
2062
2063 /// Return the mask for this instruction as a vector of integers. Undefined
2064 /// elements of the mask are returned as UndefMaskElem.
2065 void getShuffleMask(SmallVectorImpl<int> &Result) const {
2066 Result.assign(ShuffleMask.begin(), ShuffleMask.end());
2067 }
2068
2069 /// Return the mask for this instruction, for use in bitcode.
2070 ///
2071 /// TODO: This is temporary until we decide a new bitcode encoding for
2072 /// shufflevector.
2073 Constant *getShuffleMaskForBitcode() const { return ShuffleMaskForBitcode; }
2074
2075 static Constant *convertShuffleMaskForBitcode(ArrayRef<int> Mask,
2076 Type *ResultTy);
2077
2078 void setShuffleMask(ArrayRef<int> Mask);
2079
2080 ArrayRef<int> getShuffleMask() const { return ShuffleMask; }
2081
2082 /// Return true if this shuffle returns a vector with a different number of
2083 /// elements than its source vectors.
2084 /// Examples: shufflevector <4 x n> A, <4 x n> B, <1,2,3>
2085 /// shufflevector <4 x n> A, <4 x n> B, <1,2,3,4,5>
2086 bool changesLength() const {
2087 unsigned NumSourceElts = cast<VectorType>(Op<0>()->getType())
2088 ->getElementCount()
2089 .getKnownMinValue();
2090 unsigned NumMaskElts = ShuffleMask.size();
2091 return NumSourceElts != NumMaskElts;
2092 }
2093
2094 /// Return true if this shuffle returns a vector with a greater number of
2095 /// elements than its source vectors.
2096 /// Example: shufflevector <2 x n> A, <2 x n> B, <1,2,3>
2097 bool increasesLength() const {
2098 unsigned NumSourceElts = cast<VectorType>(Op<0>()->getType())
2099 ->getElementCount()
2100 .getKnownMinValue();
2101 unsigned NumMaskElts = ShuffleMask.size();
2102 return NumSourceElts < NumMaskElts;
2103 }
2104
2105 /// Return true if this shuffle mask chooses elements from exactly one source
2106 /// vector.
2107 /// Example: <7,5,undef,7>
2108 /// This assumes that vector operands are the same length as the mask.
2109 static bool isSingleSourceMask(ArrayRef<int> Mask);
2110 static bool isSingleSourceMask(const Constant *Mask) {
2111 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")(static_cast<void> (0));
2112 SmallVector<int, 16> MaskAsInts;
2113 getShuffleMask(Mask, MaskAsInts);
2114 return isSingleSourceMask(MaskAsInts);
2115 }
2116
2117 /// Return true if this shuffle chooses elements from exactly one source
2118 /// vector without changing the length of that vector.
2119 /// Example: shufflevector <4 x n> A, <4 x n> B, <3,0,undef,3>
2120 /// TODO: Optionally allow length-changing shuffles.
2121 bool isSingleSource() const {
2122 return !changesLength() && isSingleSourceMask(ShuffleMask);
2123 }
2124
2125 /// Return true if this shuffle mask chooses elements from exactly one source
2126 /// vector without lane crossings. A shuffle using this mask is not
2127 /// necessarily a no-op because it may change the number of elements from its
2128 /// input vectors or it may provide demanded bits knowledge via undef lanes.
2129 /// Example: <undef,undef,2,3>
2130 static bool isIdentityMask(ArrayRef<int> Mask);
2131 static bool isIdentityMask(const Constant *Mask) {
2132 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")(static_cast<void> (0));
2133 SmallVector<int, 16> MaskAsInts;
2134 getShuffleMask(Mask, MaskAsInts);
2135 return isIdentityMask(MaskAsInts);
2136 }
2137
2138 /// Return true if this shuffle chooses elements from exactly one source
2139 /// vector without lane crossings and does not change the number of elements
2140 /// from its input vectors.
2141 /// Example: shufflevector <4 x n> A, <4 x n> B, <4,undef,6,undef>
2142 bool isIdentity() const {
2143 return !changesLength() && isIdentityMask(ShuffleMask);
2144 }
2145
2146 /// Return true if this shuffle lengthens exactly one source vector with
2147 /// undefs in the high elements.
2148 bool isIdentityWithPadding() const;
2149
2150 /// Return true if this shuffle extracts the first N elements of exactly one
2151 /// source vector.
2152 bool isIdentityWithExtract() const;
2153
2154 /// Return true if this shuffle concatenates its 2 source vectors. This
2155 /// returns false if either input is undefined. In that case, the shuffle is
2156 /// is better classified as an identity with padding operation.
2157 bool isConcat() const;
2158
2159 /// Return true if this shuffle mask chooses elements from its source vectors
2160 /// without lane crossings. A shuffle using this mask would be
2161 /// equivalent to a vector select with a constant condition operand.
2162 /// Example: <4,1,6,undef>
2163 /// This returns false if the mask does not choose from both input vectors.
2164 /// In that case, the shuffle is better classified as an identity shuffle.
2165 /// This assumes that vector operands are the same length as the mask
2166 /// (a length-changing shuffle can never be equivalent to a vector select).
2167 static bool isSelectMask(ArrayRef<int> Mask);
2168 static bool isSelectMask(const Constant *Mask) {
2169 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")(static_cast<void> (0));
2170 SmallVector<int, 16> MaskAsInts;
2171 getShuffleMask(Mask, MaskAsInts);
2172 return isSelectMask(MaskAsInts);
2173 }
2174
2175 /// Return true if this shuffle chooses elements from its source vectors
2176 /// without lane crossings and all operands have the same number of elements.
2177 /// In other words, this shuffle is equivalent to a vector select with a
2178 /// constant condition operand.
2179 /// Example: shufflevector <4 x n> A, <4 x n> B, <undef,1,6,3>
2180 /// This returns false if the mask does not choose from both input vectors.
2181 /// In that case, the shuffle is better classified as an identity shuffle.
2182 /// TODO: Optionally allow length-changing shuffles.
2183 bool isSelect() const {
2184 return !changesLength() && isSelectMask(ShuffleMask);
2185 }
2186
2187 /// Return true if this shuffle mask swaps the order of elements from exactly
2188 /// one source vector.
2189 /// Example: <7,6,undef,4>
2190 /// This assumes that vector operands are the same length as the mask.
2191 static bool isReverseMask(ArrayRef<int> Mask);
2192 static bool isReverseMask(const Constant *Mask) {
2193 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")(static_cast<void> (0));
2194 SmallVector<int, 16> MaskAsInts;
2195 getShuffleMask(Mask, MaskAsInts);
2196 return isReverseMask(MaskAsInts);
2197 }
2198
2199 /// Return true if this shuffle swaps the order of elements from exactly
2200 /// one source vector.
2201 /// Example: shufflevector <4 x n> A, <4 x n> B, <3,undef,1,undef>
2202 /// TODO: Optionally allow length-changing shuffles.
2203 bool isReverse() const {
2204 return !changesLength() && isReverseMask(ShuffleMask);
2205 }
2206
2207 /// Return true if this shuffle mask chooses all elements with the same value
2208 /// as the first element of exactly one source vector.
2209 /// Example: <4,undef,undef,4>
2210 /// This assumes that vector operands are the same length as the mask.
2211 static bool isZeroEltSplatMask(ArrayRef<int> Mask);
2212 static bool isZeroEltSplatMask(const Constant *Mask) {
2213 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")(static_cast<void> (0));
2214 SmallVector<int, 16> MaskAsInts;
2215 getShuffleMask(Mask, MaskAsInts);
2216 return isZeroEltSplatMask(MaskAsInts);
2217 }
2218
2219 /// Return true if all elements of this shuffle are the same value as the
2220 /// first element of exactly one source vector without changing the length
2221 /// of that vector.
2222 /// Example: shufflevector <4 x n> A, <4 x n> B, <undef,0,undef,0>
2223 /// TODO: Optionally allow length-changing shuffles.
2224 /// TODO: Optionally allow splats from other elements.
2225 bool isZeroEltSplat() const {
2226 return !changesLength() && isZeroEltSplatMask(ShuffleMask);
2227 }
2228
2229 /// Return true if this shuffle mask is a transpose mask.
2230 /// Transpose vector masks transpose a 2xn matrix. They read corresponding
2231 /// even- or odd-numbered vector elements from two n-dimensional source
2232 /// vectors and write each result into consecutive elements of an
2233 /// n-dimensional destination vector. Two shuffles are necessary to complete
2234 /// the transpose, one for the even elements and another for the odd elements.
2235 /// This description closely follows how the TRN1 and TRN2 AArch64
2236 /// instructions operate.
2237 ///
2238 /// For example, a simple 2x2 matrix can be transposed with:
2239 ///
2240 /// ; Original matrix
2241 /// m0 = < a, b >
2242 /// m1 = < c, d >
2243 ///
2244 /// ; Transposed matrix
2245 /// t0 = < a, c > = shufflevector m0, m1, < 0, 2 >
2246 /// t1 = < b, d > = shufflevector m0, m1, < 1, 3 >
2247 ///
2248 /// For matrices having greater than n columns, the resulting nx2 transposed
2249 /// matrix is stored in two result vectors such that one vector contains
2250 /// interleaved elements from all the even-numbered rows and the other vector
2251 /// contains interleaved elements from all the odd-numbered rows. For example,
2252 /// a 2x4 matrix can be transposed with:
2253 ///
2254 /// ; Original matrix
2255 /// m0 = < a, b, c, d >
2256 /// m1 = < e, f, g, h >
2257 ///
2258 /// ; Transposed matrix
2259 /// t0 = < a, e, c, g > = shufflevector m0, m1 < 0, 4, 2, 6 >
2260 /// t1 = < b, f, d, h > = shufflevector m0, m1 < 1, 5, 3, 7 >
2261 static bool isTransposeMask(ArrayRef<int> Mask);
2262 static bool isTransposeMask(const Constant *Mask) {
2263 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")(static_cast<void> (0));
2264 SmallVector<int, 16> MaskAsInts;
2265 getShuffleMask(Mask, MaskAsInts);
2266 return isTransposeMask(MaskAsInts);
2267 }
2268
2269 /// Return true if this shuffle transposes the elements of its inputs without
2270 /// changing the length of the vectors. This operation may also be known as a
2271 /// merge or interleave. See the description for isTransposeMask() for the
2272 /// exact specification.
2273 /// Example: shufflevector <4 x n> A, <4 x n> B, <0,4,2,6>
2274 bool isTranspose() const {
2275 return !changesLength() && isTransposeMask(ShuffleMask);
2276 }
2277
2278 /// Return true if this shuffle mask is an extract subvector mask.
2279 /// A valid extract subvector mask returns a smaller vector from a single
2280 /// source operand. The base extraction index is returned as well.
2281 static bool isExtractSubvectorMask(ArrayRef<int> Mask, int NumSrcElts,
2282 int &Index);
2283 static bool isExtractSubvectorMask(const Constant *Mask, int NumSrcElts,
2284 int &Index) {
2285 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")(static_cast<void> (0));
2286 // Not possible to express a shuffle mask for a scalable vector for this
2287 // case.
2288 if (isa<ScalableVectorType>(Mask->getType()))
2289 return false;
2290 SmallVector<int, 16> MaskAsInts;
2291 getShuffleMask(Mask, MaskAsInts);
2292 return isExtractSubvectorMask(MaskAsInts, NumSrcElts, Index);
2293 }
2294
2295 /// Return true if this shuffle mask is an extract subvector mask.
2296 bool isExtractSubvectorMask(int &Index) const {
2297 // Not possible to express a shuffle mask for a scalable vector for this
2298 // case.
2299 if (isa<ScalableVectorType>(getType()))
2300 return false;
2301
2302 int NumSrcElts =
2303 cast<FixedVectorType>(Op<0>()->getType())->getNumElements();
2304 return isExtractSubvectorMask(ShuffleMask, NumSrcElts, Index);
2305 }
2306
2307 /// Return true if this shuffle mask is an insert subvector mask.
2308 /// A valid insert subvector mask inserts the lowest elements of a second
2309 /// source operand into an in-place first source operand operand.
2310 /// Both the sub vector width and the insertion index is returned.
2311 static bool isInsertSubvectorMask(ArrayRef<int> Mask, int NumSrcElts,
2312 int &NumSubElts, int &Index);
2313 static bool isInsertSubvectorMask(const Constant *Mask, int NumSrcElts,
2314 int &NumSubElts, int &Index) {
2315 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")(static_cast<void> (0));
2316 // Not possible to express a shuffle mask for a scalable vector for this
2317 // case.
2318 if (isa<ScalableVectorType>(Mask->getType()))
2319 return false;
2320 SmallVector<int, 16> MaskAsInts;
2321 getShuffleMask(Mask, MaskAsInts);
2322 return isInsertSubvectorMask(MaskAsInts, NumSrcElts, NumSubElts, Index);
2323 }
2324
2325 /// Return true if this shuffle mask is an insert subvector mask.
2326 bool isInsertSubvectorMask(int &NumSubElts, int &Index) const {
2327 // Not possible to express a shuffle mask for a scalable vector for this
2328 // case.
2329 if (isa<ScalableVectorType>(getType()))
2330 return false;
2331
2332 int NumSrcElts =
2333 cast<FixedVectorType>(Op<0>()->getType())->getNumElements();
2334 return isInsertSubvectorMask(ShuffleMask, NumSrcElts, NumSubElts, Index);
2335 }
2336
2337 /// Change values in a shuffle permute mask assuming the two vector operands
2338 /// of length InVecNumElts have swapped position.
2339 static void commuteShuffleMask(MutableArrayRef<int> Mask,
2340 unsigned InVecNumElts) {
2341 for (int &Idx : Mask) {
2342 if (Idx == -1)
2343 continue;
2344 Idx = Idx < (int)InVecNumElts ? Idx + InVecNumElts : Idx - InVecNumElts;
2345 assert(Idx >= 0 && Idx < (int)InVecNumElts * 2 &&(static_cast<void> (0))
2346 "shufflevector mask index out of range")(static_cast<void> (0));
2347 }
2348 }
2349
2350 // Methods for support type inquiry through isa, cast, and dyn_cast:
2351 static bool classof(const Instruction *I) {
2352 return I->getOpcode() == Instruction::ShuffleVector;
2353 }
2354 static bool classof(const Value *V) {
2355 return isa<Instruction>(V) && classof(cast<Instruction>(V));
2356 }
2357};
2358
2359template <>
2360struct OperandTraits<ShuffleVectorInst>
2361 : public FixedNumOperandTraits<ShuffleVectorInst, 2> {};
2362
2363DEFINE_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 { (static_cast<void> (0)); 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) { (static_cast<void> (0)); 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
); }
2364
2365//===----------------------------------------------------------------------===//
2366// ExtractValueInst Class
2367//===----------------------------------------------------------------------===//
2368
2369/// This instruction extracts a struct member or array
2370/// element value from an aggregate value.
2371///
2372class ExtractValueInst : public UnaryInstruction {
2373 SmallVector<unsigned, 4> Indices;
2374
2375 ExtractValueInst(const ExtractValueInst &EVI);
2376
2377 /// Constructors - Create a extractvalue instruction with a base aggregate
2378 /// value and a list of indices. The first ctor can optionally insert before
2379 /// an existing instruction, the second appends the new instruction to the
2380 /// specified BasicBlock.
2381 inline ExtractValueInst(Value *Agg,
2382 ArrayRef<unsigned> Idxs,
2383 const Twine &NameStr,
2384 Instruction *InsertBefore);
2385 inline ExtractValueInst(Value *Agg,
2386 ArrayRef<unsigned> Idxs,
2387 const Twine &NameStr, BasicBlock *InsertAtEnd);
2388
2389 void init(ArrayRef<unsigned> Idxs, const Twine &NameStr);
2390
2391protected:
2392 // Note: Instruction needs to be a friend here to call cloneImpl.
2393 friend class Instruction;
2394
2395 ExtractValueInst *cloneImpl() const;
2396
2397public:
2398 static ExtractValueInst *Create(Value *Agg,
2399 ArrayRef<unsigned> Idxs,
2400 const Twine &NameStr = "",
2401 Instruction *InsertBefore = nullptr) {
2402 return new
2403 ExtractValueInst(Agg, Idxs, NameStr, InsertBefore);
2404 }
2405
2406 static ExtractValueInst *Create(Value *Agg,
2407 ArrayRef<unsigned> Idxs,
2408 const Twine &NameStr,
2409 BasicBlock *InsertAtEnd) {
2410 return new ExtractValueInst(Agg, Idxs, NameStr, InsertAtEnd);
2411 }
2412
2413 /// Returns the type of the element that would be extracted
2414 /// with an extractvalue instruction with the specified parameters.
2415 ///
2416 /// Null is returned if the indices are invalid for the specified type.
2417 static Type *getIndexedType(Type *Agg, ArrayRef<unsigned> Idxs);
2418
2419 using idx_iterator = const unsigned*;
2420
2421 inline idx_iterator idx_begin() const { return Indices.begin(); }
2422 inline idx_iterator idx_end() const { return Indices.end(); }
2423 inline iterator_range<idx_iterator> indices() const {
2424 return make_range(idx_begin(), idx_end());
2425 }
2426
2427 Value *getAggregateOperand() {
2428 return getOperand(0);
2429 }
2430 const Value *getAggregateOperand() const {
2431 return getOperand(0);
2432 }
2433 static unsigned getAggregateOperandIndex() {
2434 return 0U; // get index for modifying correct operand
2435 }
2436
2437 ArrayRef<unsigned> getIndices() const {
2438 return Indices;
2439 }
2440
2441 unsigned getNumIndices() const {
2442 return (unsigned)Indices.size();
2443 }
2444
2445 bool hasIndices() const {
2446 return true;
2447 }
2448
2449 // Methods for support type inquiry through isa, cast, and dyn_cast:
2450 static bool classof(const Instruction *I) {
2451 return I->getOpcode() == Instruction::ExtractValue;
2452 }
2453 static bool classof(const Value *V) {
2454 return isa<Instruction>(V) && classof(cast<Instruction>(V));
2455 }
2456};
2457
2458ExtractValueInst::ExtractValueInst(Value *Agg,
2459 ArrayRef<unsigned> Idxs,
2460 const Twine &NameStr,
2461 Instruction *InsertBefore)
2462 : UnaryInstruction(checkGEPType(getIndexedType(Agg->getType(), Idxs)),
2463 ExtractValue, Agg, InsertBefore) {
2464 init(Idxs, NameStr);
2465}
2466
2467ExtractValueInst::ExtractValueInst(Value *Agg,
2468 ArrayRef<unsigned> Idxs,
2469 const Twine &NameStr,
2470 BasicBlock *InsertAtEnd)
2471 : UnaryInstruction(checkGEPType(getIndexedType(Agg->getType(), Idxs)),
2472 ExtractValue, Agg, InsertAtEnd) {
2473 init(Idxs, NameStr);
2474}
2475
2476//===----------------------------------------------------------------------===//
2477// InsertValueInst Class
2478//===----------------------------------------------------------------------===//
2479
2480/// This instruction inserts a struct field of array element
2481/// value into an aggregate value.
2482///
2483class InsertValueInst : public Instruction {
2484 SmallVector<unsigned, 4> Indices;
2485
2486 InsertValueInst(const InsertValueInst &IVI);
2487
2488 /// Constructors - Create a insertvalue instruction with a base aggregate
2489 /// value, a value to insert, and a list of indices. The first ctor can
2490 /// optionally insert before an existing instruction, the second appends
2491 /// the new instruction to the specified BasicBlock.
2492 inline InsertValueInst(Value *Agg, Value *Val,
2493 ArrayRef<unsigned> Idxs,
2494 const Twine &NameStr,
2495 Instruction *InsertBefore);
2496 inline InsertValueInst(Value *Agg, Value *Val,
2497 ArrayRef<unsigned> Idxs,
2498 const Twine &NameStr, BasicBlock *InsertAtEnd);
2499
2500 /// Constructors - These two constructors are convenience methods because one
2501 /// and two index insertvalue instructions are so common.
2502 InsertValueInst(Value *Agg, Value *Val, unsigned Idx,
2503 const Twine &NameStr = "",
2504 Instruction *InsertBefore = nullptr);
2505 InsertValueInst(Value *Agg, Value *Val, unsigned Idx, const Twine &NameStr,
2506 BasicBlock *InsertAtEnd);
2507
2508 void init(Value *Agg, Value *Val, ArrayRef<unsigned> Idxs,
2509 const Twine &NameStr);
2510
2511protected:
2512 // Note: Instruction needs to be a friend here to call cloneImpl.
2513 friend class Instruction;
2514
2515 InsertValueInst *cloneImpl() const;
2516
2517public:
2518 // allocate space for exactly two operands
2519 void *operator new(size_t S) { return User::operator new(S, 2); }
2520 void operator delete(void *Ptr) { User::operator delete(Ptr); }
2521
2522 static InsertValueInst *Create(Value *Agg, Value *Val,
2523 ArrayRef<unsigned> Idxs,
2524 const Twine &NameStr = "",
2525 Instruction *InsertBefore = nullptr) {
2526 return new InsertValueInst(Agg, Val, Idxs, NameStr, InsertBefore);
2527 }
2528
2529 static InsertValueInst *Create(Value *Agg, Value *Val,
2530 ArrayRef<unsigned> Idxs,
2531 const Twine &NameStr,
2532 BasicBlock *InsertAtEnd) {
2533 return new InsertValueInst(Agg, Val, Idxs, NameStr, InsertAtEnd);
2534 }
2535
2536 /// Transparently provide more efficient getOperand methods.
2537 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
;
2538
2539 using idx_iterator = const unsigned*;
2540
2541 inline idx_iterator idx_begin() const { return Indices.begin(); }
2542 inline idx_iterator idx_end() const { return Indices.end(); }
2543 inline iterator_range<idx_iterator> indices() const {
2544 return make_range(idx_begin(), idx_end());
2545 }
2546
2547 Value *getAggregateOperand() {
2548 return getOperand(0);
2549 }
2550 const Value *getAggregateOperand() const {
2551 return getOperand(0);
2552 }
2553 static unsigned getAggregateOperandIndex() {
2554 return 0U; // get index for modifying correct operand
2555 }
2556
2557 Value *getInsertedValueOperand() {
2558 return getOperand(1);
2559 }
2560 const Value *getInsertedValueOperand() const {
2561 return getOperand(1);
2562 }
2563 static unsigned getInsertedValueOperandIndex() {
2564 return 1U; // get index for modifying correct operand
2565 }
2566
2567 ArrayRef<unsigned> getIndices() const {
2568 return Indices;
2569 }
2570
2571 unsigned getNumIndices() const {
2572 return (unsigned)Indices.size();
2573 }
2574
2575 bool hasIndices() const {
2576 return true;
2577 }
2578
2579 // Methods for support type inquiry through isa, cast, and dyn_cast:
2580 static bool classof(const Instruction *I) {
2581 return I->getOpcode() == Instruction::InsertValue;
2582 }
2583 static bool classof(const Value *V) {
2584 return isa<Instruction>(V) && classof(cast<Instruction>(V));
2585 }
2586};
2587
2588template <>
2589struct OperandTraits<InsertValueInst> :
2590 public FixedNumOperandTraits<InsertValueInst, 2> {
2591};
2592
2593InsertValueInst::InsertValueInst(Value *Agg,
2594 Value *Val,
2595 ArrayRef<unsigned> Idxs,
2596 const Twine &NameStr,
2597 Instruction *InsertBefore)
2598 : Instruction(Agg->getType(), InsertValue,
2599 OperandTraits<InsertValueInst>::op_begin(this),
2600 2, InsertBefore) {
2601 init(Agg, Val, Idxs, NameStr);
2602}
2603
2604InsertValueInst::InsertValueInst(Value *Agg,
2605 Value *Val,
2606 ArrayRef<unsigned> Idxs,
2607 const Twine &NameStr,
2608 BasicBlock *InsertAtEnd)
2609 : Instruction(Agg->getType(), InsertValue,
2610 OperandTraits<InsertValueInst>::op_begin(this),
2611 2, InsertAtEnd) {
2612 init(Agg, Val, Idxs, NameStr);
2613}
2614
2615DEFINE_TRANSPARENT_OPERAND_ACCESSORS(InsertValueInst, Value)InsertValueInst::op_iterator InsertValueInst::op_begin() { return
OperandTraits<InsertValueInst>::op_begin(this); } InsertValueInst
::const_op_iterator InsertValueInst::op_begin() const { return
OperandTraits<InsertValueInst>::op_begin(const_cast<
InsertValueInst*>(this)); } InsertValueInst::op_iterator InsertValueInst
::op_end() { return OperandTraits<InsertValueInst>::op_end
(this); } InsertValueInst::const_op_iterator InsertValueInst::
op_end() const { return OperandTraits<InsertValueInst>::
op_end(const_cast<InsertValueInst*>(this)); } Value *InsertValueInst
::getOperand(unsigned i_nocapture) const { (static_cast<void
> (0)); return cast_or_null<Value>( OperandTraits<
InsertValueInst>::op_begin(const_cast<InsertValueInst*>
(this))[i_nocapture].get()); } void InsertValueInst::setOperand
(unsigned i_nocapture, Value *Val_nocapture) { (static_cast<
void> (0)); OperandTraits<InsertValueInst>::op_begin
(this)[i_nocapture] = Val_nocapture; } unsigned InsertValueInst
::getNumOperands() const { return OperandTraits<InsertValueInst
>::operands(this); } template <int Idx_nocapture> Use
&InsertValueInst::Op() { return this->OpFrom<Idx_nocapture
>(this); } template <int Idx_nocapture> const Use &
InsertValueInst::Op() const { return this->OpFrom<Idx_nocapture
>(this); }
2616
2617//===----------------------------------------------------------------------===//
2618// PHINode Class
2619//===----------------------------------------------------------------------===//
2620
2621// PHINode - The PHINode class is used to represent the magical mystical PHI
2622// node, that can not exist in nature, but can be synthesized in a computer
2623// scientist's overactive imagination.
2624//
2625class PHINode : public Instruction {
2626 /// The number of operands actually allocated. NumOperands is
2627 /// the number actually in use.
2628 unsigned ReservedSpace;
2629
2630 PHINode(const PHINode &PN);
2631
2632 explicit PHINode(Type *Ty, unsigned NumReservedValues,
2633 const Twine &NameStr = "",
2634 Instruction *InsertBefore = nullptr)
2635 : Instruction(Ty, Instruction::PHI, nullptr, 0, InsertBefore),
2636 ReservedSpace(NumReservedValues) {
2637 assert(!Ty->isTokenTy() && "PHI nodes cannot have token type!")(static_cast<void> (0));
2638 setName(NameStr);
2639 allocHungoffUses(ReservedSpace);
2640 }
2641
2642 PHINode(Type *Ty, unsigned NumReservedValues, const Twine &NameStr,
2643 BasicBlock *InsertAtEnd)
2644 : Instruction(Ty, Instruction::PHI, nullptr, 0, InsertAtEnd),
2645 ReservedSpace(NumReservedValues) {
2646 assert(!Ty->isTokenTy() && "PHI nodes cannot have token type!")(static_cast<void> (0));
2647 setName(NameStr);
2648 allocHungoffUses(ReservedSpace);
2649 }
2650
2651protected:
2652 // Note: Instruction needs to be a friend here to call cloneImpl.
2653 friend class Instruction;
2654
2655 PHINode *cloneImpl() const;
2656
2657 // allocHungoffUses - this is more complicated than the generic