Bug Summary

File:lib/Transforms/Scalar/JumpThreading.cpp
Warning:line 1439, column 7
Called C++ object pointer is null

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name JumpThreading.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mthread-model posix -mframe-pointer=none -fmath-errno -masm-verbose -mconstructor-aliases -munwind-tables -fuse-init-array -target-cpu x86-64 -dwarf-column-info -debugger-tuning=gdb -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-10/lib/clang/10.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-10~svn374877/build-llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-10~svn374877/build-llvm/include -I /build/llvm-toolchain-snapshot-10~svn374877/include -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0/backward -internal-isystem /usr/local/include -internal-isystem /usr/lib/llvm-10/lib/clang/10.0.0/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-comment -std=c++14 -fdeprecated-macro -fdebug-compilation-dir /build/llvm-toolchain-snapshot-10~svn374877/build-llvm/lib/Transforms/Scalar -fdebug-prefix-map=/build/llvm-toolchain-snapshot-10~svn374877=. -ferror-limit 19 -fmessage-length 0 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -fobjc-runtime=gcc -fdiagnostics-show-option -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -o /tmp/scan-build-2019-10-15-233810-7101-1 -x c++ /build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp

/build/llvm-toolchain-snapshot-10~svn374877/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/Optional.h"
17#include "llvm/ADT/STLExtras.h"
18#include "llvm/ADT/SmallPtrSet.h"
19#include "llvm/ADT/SmallVector.h"
20#include "llvm/ADT/Statistic.h"
21#include "llvm/Analysis/AliasAnalysis.h"
22#include "llvm/Analysis/BlockFrequencyInfo.h"
23#include "llvm/Analysis/BranchProbabilityInfo.h"
24#include "llvm/Analysis/CFG.h"
25#include "llvm/Analysis/ConstantFolding.h"
26#include "llvm/Analysis/DomTreeUpdater.h"
27#include "llvm/Analysis/GlobalsModRef.h"
28#include "llvm/Analysis/GuardUtils.h"
29#include "llvm/Analysis/InstructionSimplify.h"
30#include "llvm/Analysis/LazyValueInfo.h"
31#include "llvm/Analysis/Loads.h"
32#include "llvm/Analysis/LoopInfo.h"
33#include "llvm/Analysis/TargetLibraryInfo.h"
34#include "llvm/Analysis/ValueTracking.h"
35#include "llvm/IR/BasicBlock.h"
36#include "llvm/IR/CFG.h"
37#include "llvm/IR/Constant.h"
38#include "llvm/IR/ConstantRange.h"
39#include "llvm/IR/Constants.h"
40#include "llvm/IR/DataLayout.h"
41#include "llvm/IR/Dominators.h"
42#include "llvm/IR/Function.h"
43#include "llvm/IR/InstrTypes.h"
44#include "llvm/IR/Instruction.h"
45#include "llvm/IR/Instructions.h"
46#include "llvm/IR/IntrinsicInst.h"
47#include "llvm/IR/Intrinsics.h"
48#include "llvm/IR/LLVMContext.h"
49#include "llvm/IR/MDBuilder.h"
50#include "llvm/IR/Metadata.h"
51#include "llvm/IR/Module.h"
52#include "llvm/IR/PassManager.h"
53#include "llvm/IR/PatternMatch.h"
54#include "llvm/IR/Type.h"
55#include "llvm/IR/Use.h"
56#include "llvm/IR/User.h"
57#include "llvm/IR/Value.h"
58#include "llvm/Pass.h"
59#include "llvm/Support/BlockFrequency.h"
60#include "llvm/Support/BranchProbability.h"
61#include "llvm/Support/Casting.h"
62#include "llvm/Support/CommandLine.h"
63#include "llvm/Support/Debug.h"
64#include "llvm/Support/raw_ostream.h"
65#include "llvm/Transforms/Scalar.h"
66#include "llvm/Transforms/Utils/BasicBlockUtils.h"
67#include "llvm/Transforms/Utils/Cloning.h"
68#include "llvm/Transforms/Utils/Local.h"
69#include "llvm/Transforms/Utils/SSAUpdater.h"
70#include "llvm/Transforms/Utils/ValueMapper.h"
71#include <algorithm>
72#include <cassert>
73#include <cstddef>
74#include <cstdint>
75#include <iterator>
76#include <memory>
77#include <utility>
78
79using namespace llvm;
80using namespace jumpthreading;
81
82#define DEBUG_TYPE"jump-threading" "jump-threading"
83
84STATISTIC(NumThreads, "Number of jumps threaded")static llvm::Statistic NumThreads = {"jump-threading", "NumThreads"
, "Number of jumps threaded"}
;
85STATISTIC(NumFolds, "Number of terminators folded")static llvm::Statistic NumFolds = {"jump-threading", "NumFolds"
, "Number of terminators folded"}
;
86STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi")static llvm::Statistic NumDupes = {"jump-threading", "NumDupes"
, "Number of branch blocks duplicated to eliminate phi"}
;
87
88static cl::opt<unsigned>
89BBDuplicateThreshold("jump-threading-threshold",
90 cl::desc("Max block size to duplicate for jump threading"),
91 cl::init(6), cl::Hidden);
92
93static cl::opt<unsigned>
94ImplicationSearchThreshold(
95 "jump-threading-implication-search-threshold",
96 cl::desc("The number of predecessors to search for a stronger "
97 "condition to use to thread over a weaker condition"),
98 cl::init(3), cl::Hidden);
99
100static cl::opt<bool> PrintLVIAfterJumpThreading(
101 "print-lvi-after-jump-threading",
102 cl::desc("Print the LazyValueInfo cache after JumpThreading"), cl::init(false),
103 cl::Hidden);
104
105static cl::opt<bool> ThreadAcrossLoopHeaders(
106 "jump-threading-across-loop-headers",
107 cl::desc("Allow JumpThreading to thread across loop headers, for testing"),
108 cl::init(false), cl::Hidden);
109
110
111namespace {
112
113 /// This pass performs 'jump threading', which looks at blocks that have
114 /// multiple predecessors and multiple successors. If one or more of the
115 /// predecessors of the block can be proven to always jump to one of the
116 /// successors, we forward the edge from the predecessor to the successor by
117 /// duplicating the contents of this block.
118 ///
119 /// An example of when this can occur is code like this:
120 ///
121 /// if () { ...
122 /// X = 4;
123 /// }
124 /// if (X < 3) {
125 ///
126 /// In this case, the unconditional branch at the end of the first if can be
127 /// revectored to the false side of the second if.
128 class JumpThreading : public FunctionPass {
129 JumpThreadingPass Impl;
130
131 public:
132 static char ID; // Pass identification
133
134 JumpThreading(int T = -1) : FunctionPass(ID), Impl(T) {
135 initializeJumpThreadingPass(*PassRegistry::getPassRegistry());
136 }
137
138 bool runOnFunction(Function &F) override;
139
140 void getAnalysisUsage(AnalysisUsage &AU) const override {
141 AU.addRequired<DominatorTreeWrapperPass>();
142 AU.addPreserved<DominatorTreeWrapperPass>();
143 AU.addRequired<AAResultsWrapperPass>();
144 AU.addRequired<LazyValueInfoWrapperPass>();
145 AU.addPreserved<LazyValueInfoWrapperPass>();
146 AU.addPreserved<GlobalsAAWrapperPass>();
147 AU.addRequired<TargetLibraryInfoWrapperPass>();
148 }
149
150 void releaseMemory() override { Impl.releaseMemory(); }
151 };
152
153} // end anonymous namespace
154
155char JumpThreading::ID = 0;
156
157INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",static void *initializeJumpThreadingPassOnce(PassRegistry &
Registry) {
158 "Jump Threading", false, false)static void *initializeJumpThreadingPassOnce(PassRegistry &
Registry) {
159INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry);
160INITIALIZE_PASS_DEPENDENCY(LazyValueInfoWrapperPass)initializeLazyValueInfoWrapperPassPass(Registry);
161INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry);
162INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry);
163INITIALIZE_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)); }
164 "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)); }
165
166// Public interface to the Jump Threading pass
167FunctionPass *llvm::createJumpThreadingPass(int Threshold) {
168 return new JumpThreading(Threshold);
169}
170
171JumpThreadingPass::JumpThreadingPass(int T) {
172 BBDupThreshold = (T == -1) ? BBDuplicateThreshold : unsigned(T);
173}
174
175// Update branch probability information according to conditional
176// branch probability. This is usually made possible for cloned branches
177// in inline instances by the context specific profile in the caller.
178// For instance,
179//
180// [Block PredBB]
181// [Branch PredBr]
182// if (t) {
183// Block A;
184// } else {
185// Block B;
186// }
187//
188// [Block BB]
189// cond = PN([true, %A], [..., %B]); // PHI node
190// [Branch CondBr]
191// if (cond) {
192// ... // P(cond == true) = 1%
193// }
194//
195// Here we know that when block A is taken, cond must be true, which means
196// P(cond == true | A) = 1
197//
198// Given that P(cond == true) = P(cond == true | A) * P(A) +
199// P(cond == true | B) * P(B)
200// we get:
201// P(cond == true ) = P(A) + P(cond == true | B) * P(B)
202//
203// which gives us:
204// P(A) is less than P(cond == true), i.e.
205// P(t == true) <= P(cond == true)
206//
207// In other words, if we know P(cond == true) is unlikely, we know
208// that P(t == true) is also unlikely.
209//
210static void updatePredecessorProfileMetadata(PHINode *PN, BasicBlock *BB) {
211 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
212 if (!CondBr)
213 return;
214
215 BranchProbability BP;
216 uint64_t TrueWeight, FalseWeight;
217 if (!CondBr->extractProfMetadata(TrueWeight, FalseWeight))
218 return;
219
220 // Returns the outgoing edge of the dominating predecessor block
221 // that leads to the PhiNode's incoming block:
222 auto GetPredOutEdge =
223 [](BasicBlock *IncomingBB,
224 BasicBlock *PhiBB) -> std::pair<BasicBlock *, BasicBlock *> {
225 auto *PredBB = IncomingBB;
226 auto *SuccBB = PhiBB;
227 SmallPtrSet<BasicBlock *, 16> Visited;
228 while (true) {
229 BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator());
230 if (PredBr && PredBr->isConditional())
231 return {PredBB, SuccBB};
232 Visited.insert(PredBB);
233 auto *SinglePredBB = PredBB->getSinglePredecessor();
234 if (!SinglePredBB)
235 return {nullptr, nullptr};
236
237 // Stop searching when SinglePredBB has been visited. It means we see
238 // an unreachable loop.
239 if (Visited.count(SinglePredBB))
240 return {nullptr, nullptr};
241
242 SuccBB = PredBB;
243 PredBB = SinglePredBB;
244 }
245 };
246
247 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
248 Value *PhiOpnd = PN->getIncomingValue(i);
249 ConstantInt *CI = dyn_cast<ConstantInt>(PhiOpnd);
250
251 if (!CI || !CI->getType()->isIntegerTy(1))
252 continue;
253
254 BP = (CI->isOne() ? BranchProbability::getBranchProbability(
255 TrueWeight, TrueWeight + FalseWeight)
256 : BranchProbability::getBranchProbability(
257 FalseWeight, TrueWeight + FalseWeight));
258
259 auto PredOutEdge = GetPredOutEdge(PN->getIncomingBlock(i), BB);
260 if (!PredOutEdge.first)
261 return;
262
263 BasicBlock *PredBB = PredOutEdge.first;
264 BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator());
265 if (!PredBr)
266 return;
267
268 uint64_t PredTrueWeight, PredFalseWeight;
269 // FIXME: We currently only set the profile data when it is missing.
270 // With PGO, this can be used to refine even existing profile data with
271 // context information. This needs to be done after more performance
272 // testing.
273 if (PredBr->extractProfMetadata(PredTrueWeight, PredFalseWeight))
274 continue;
275
276 // We can not infer anything useful when BP >= 50%, because BP is the
277 // upper bound probability value.
278 if (BP >= BranchProbability(50, 100))
279 continue;
280
281 SmallVector<uint32_t, 2> Weights;
282 if (PredBr->getSuccessor(0) == PredOutEdge.second) {
283 Weights.push_back(BP.getNumerator());
284 Weights.push_back(BP.getCompl().getNumerator());
285 } else {
286 Weights.push_back(BP.getCompl().getNumerator());
287 Weights.push_back(BP.getNumerator());
288 }
289 PredBr->setMetadata(LLVMContext::MD_prof,
290 MDBuilder(PredBr->getParent()->getContext())
291 .createBranchWeights(Weights));
292 }
293}
294
295/// runOnFunction - Toplevel algorithm.
296bool JumpThreading::runOnFunction(Function &F) {
297 if (skipFunction(F))
298 return false;
299 auto TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
300 // Get DT analysis before LVI. When LVI is initialized it conditionally adds
301 // DT if it's available.
302 auto DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
303 auto LVI = &getAnalysis<LazyValueInfoWrapperPass>().getLVI();
304 auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
305 DomTreeUpdater DTU(*DT, DomTreeUpdater::UpdateStrategy::Lazy);
306 std::unique_ptr<BlockFrequencyInfo> BFI;
307 std::unique_ptr<BranchProbabilityInfo> BPI;
308 bool HasProfileData = F.hasProfileData();
309 if (HasProfileData) {
310 LoopInfo LI{DominatorTree(F)};
311 BPI.reset(new BranchProbabilityInfo(F, LI, TLI));
312 BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
313 }
314
315 bool Changed = Impl.runImpl(F, TLI, LVI, AA, &DTU, HasProfileData,
316 std::move(BFI), std::move(BPI));
317 if (PrintLVIAfterJumpThreading) {
318 dbgs() << "LVI for function '" << F.getName() << "':\n";
319 LVI->printLVI(F, *DT, dbgs());
320 }
321 return Changed;
322}
323
324PreservedAnalyses JumpThreadingPass::run(Function &F,
325 FunctionAnalysisManager &AM) {
326 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
327 // Get DT analysis before LVI. When LVI is initialized it conditionally adds
328 // DT if it's available.
329 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
330 auto &LVI = AM.getResult<LazyValueAnalysis>(F);
331 auto &AA = AM.getResult<AAManager>(F);
332 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Lazy);
333
334 std::unique_ptr<BlockFrequencyInfo> BFI;
335 std::unique_ptr<BranchProbabilityInfo> BPI;
336 if (F.hasProfileData()) {
337 LoopInfo LI{DominatorTree(F)};
338 BPI.reset(new BranchProbabilityInfo(F, LI, &TLI));
339 BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
340 }
341
342 bool Changed = runImpl(F, &TLI, &LVI, &AA, &DTU, HasProfileData,
343 std::move(BFI), std::move(BPI));
344
345 if (!Changed)
346 return PreservedAnalyses::all();
347 PreservedAnalyses PA;
348 PA.preserve<GlobalsAA>();
349 PA.preserve<DominatorTreeAnalysis>();
350 PA.preserve<LazyValueAnalysis>();
351 return PA;
352}
353
354bool JumpThreadingPass::runImpl(Function &F, TargetLibraryInfo *TLI_,
355 LazyValueInfo *LVI_, AliasAnalysis *AA_,
356 DomTreeUpdater *DTU_, bool HasProfileData_,
357 std::unique_ptr<BlockFrequencyInfo> BFI_,
358 std::unique_ptr<BranchProbabilityInfo> BPI_) {
359 LLVM_DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "Jump threading on function '"
<< F.getName() << "'\n"; } } while (false)
;
360 TLI = TLI_;
361 LVI = LVI_;
362 AA = AA_;
363 DTU = DTU_;
364 BFI.reset();
365 BPI.reset();
366 // When profile data is available, we need to update edge weights after
367 // successful jump threading, which requires both BPI and BFI being available.
368 HasProfileData = HasProfileData_;
369 auto *GuardDecl = F.getParent()->getFunction(
370 Intrinsic::getName(Intrinsic::experimental_guard));
371 HasGuards = GuardDecl && !GuardDecl->use_empty();
372 if (HasProfileData) {
373 BPI = std::move(BPI_);
374 BFI = std::move(BFI_);
375 }
376
377 // JumpThreading must not processes blocks unreachable from entry. It's a
378 // waste of compute time and can potentially lead to hangs.
379 SmallPtrSet<BasicBlock *, 16> Unreachable;
380 assert(DTU && "DTU isn't passed into JumpThreading before using it.")((DTU && "DTU isn't passed into JumpThreading before using it."
) ? static_cast<void> (0) : __assert_fail ("DTU && \"DTU isn't passed into JumpThreading before using it.\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 380, __PRETTY_FUNCTION__))
;
381 assert(DTU->hasDomTree() && "JumpThreading relies on DomTree to proceed.")((DTU->hasDomTree() && "JumpThreading relies on DomTree to proceed."
) ? static_cast<void> (0) : __assert_fail ("DTU->hasDomTree() && \"JumpThreading relies on DomTree to proceed.\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 381, __PRETTY_FUNCTION__))
;
382 DominatorTree &DT = DTU->getDomTree();
383 for (auto &BB : F)
384 if (!DT.isReachableFromEntry(&BB))
385 Unreachable.insert(&BB);
386
387 if (!ThreadAcrossLoopHeaders)
388 FindLoopHeaders(F);
389
390 bool EverChanged = false;
391 bool Changed;
392 do {
393 Changed = false;
394 for (auto &BB : F) {
395 if (Unreachable.count(&BB))
396 continue;
397 while (ProcessBlock(&BB)) // Thread all of the branches we can over BB.
398 Changed = true;
399 // Stop processing BB if it's the entry or is now deleted. The following
400 // routines attempt to eliminate BB and locating a suitable replacement
401 // for the entry is non-trivial.
402 if (&BB == &F.getEntryBlock() || DTU->isBBPendingDeletion(&BB))
403 continue;
404
405 if (pred_empty(&BB)) {
406 // When ProcessBlock makes BB unreachable it doesn't bother to fix up
407 // the instructions in it. We must remove BB to prevent invalid IR.
408 LLVM_DEBUG(dbgs() << " JT: Deleting dead block '" << BB.getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " JT: Deleting dead block '"
<< BB.getName() << "' with terminator: " <<
*BB.getTerminator() << '\n'; } } while (false)
409 << "' with terminator: " << *BB.getTerminator()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " JT: Deleting dead block '"
<< BB.getName() << "' with terminator: " <<
*BB.getTerminator() << '\n'; } } while (false)
410 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " JT: Deleting dead block '"
<< BB.getName() << "' with terminator: " <<
*BB.getTerminator() << '\n'; } } while (false)
;
411 LoopHeaders.erase(&BB);
412 LVI->eraseBlock(&BB);
413 DeleteDeadBlock(&BB, DTU);
414 Changed = true;
415 continue;
416 }
417
418 // ProcessBlock doesn't thread BBs with unconditional TIs. However, if BB
419 // is "almost empty", we attempt to merge BB with its sole successor.
420 auto *BI = dyn_cast<BranchInst>(BB.getTerminator());
421 if (BI && BI->isUnconditional() &&
422 // The terminator must be the only non-phi instruction in BB.
423 BB.getFirstNonPHIOrDbg()->isTerminator() &&
424 // Don't alter Loop headers and latches to ensure another pass can
425 // detect and transform nested loops later.
426 !LoopHeaders.count(&BB) && !LoopHeaders.count(BI->getSuccessor(0)) &&
427 TryToSimplifyUncondBranchFromEmptyBlock(&BB, DTU)) {
428 // BB is valid for cleanup here because we passed in DTU. F remains
429 // BB's parent until a DTU->getDomTree() event.
430 LVI->eraseBlock(&BB);
431 Changed = true;
432 }
433 }
434 EverChanged |= Changed;
435 } while (Changed);
436
437 LoopHeaders.clear();
438 // Flush only the Dominator Tree.
439 DTU->getDomTree();
440 LVI->enableDT();
441 return EverChanged;
442}
443
444// Replace uses of Cond with ToVal when safe to do so. If all uses are
445// replaced, we can remove Cond. We cannot blindly replace all uses of Cond
446// because we may incorrectly replace uses when guards/assumes are uses of
447// of `Cond` and we used the guards/assume to reason about the `Cond` value
448// at the end of block. RAUW unconditionally replaces all uses
449// including the guards/assumes themselves and the uses before the
450// guard/assume.
451static void ReplaceFoldableUses(Instruction *Cond, Value *ToVal) {
452 assert(Cond->getType() == ToVal->getType())((Cond->getType() == ToVal->getType()) ? static_cast<
void> (0) : __assert_fail ("Cond->getType() == ToVal->getType()"
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 452, __PRETTY_FUNCTION__))
;
453 auto *BB = Cond->getParent();
454 // We can unconditionally replace all uses in non-local blocks (i.e. uses
455 // strictly dominated by BB), since LVI information is true from the
456 // terminator of BB.
457 replaceNonLocalUsesWith(Cond, ToVal);
458 for (Instruction &I : reverse(*BB)) {
459 // Reached the Cond whose uses we are trying to replace, so there are no
460 // more uses.
461 if (&I == Cond)
462 break;
463 // We only replace uses in instructions that are guaranteed to reach the end
464 // of BB, where we know Cond is ToVal.
465 if (!isGuaranteedToTransferExecutionToSuccessor(&I))
466 break;
467 I.replaceUsesOfWith(Cond, ToVal);
468 }
469 if (Cond->use_empty() && !Cond->mayHaveSideEffects())
470 Cond->eraseFromParent();
471}
472
473/// Return the cost of duplicating a piece of this block from first non-phi
474/// and before StopAt instruction to thread across it. Stop scanning the block
475/// when exceeding the threshold. If duplication is impossible, returns ~0U.
476static unsigned getJumpThreadDuplicationCost(BasicBlock *BB,
477 Instruction *StopAt,
478 unsigned Threshold) {
479 assert(StopAt->getParent() == BB && "Not an instruction from proper BB?")((StopAt->getParent() == BB && "Not an instruction from proper BB?"
) ? static_cast<void> (0) : __assert_fail ("StopAt->getParent() == BB && \"Not an instruction from proper BB?\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 479, __PRETTY_FUNCTION__))
;
480 /// Ignore PHI nodes, these will be flattened when duplication happens.
481 BasicBlock::const_iterator I(BB->getFirstNonPHI());
482
483 // FIXME: THREADING will delete values that are just used to compute the
484 // branch, so they shouldn't count against the duplication cost.
485
486 unsigned Bonus = 0;
487 if (BB->getTerminator() == StopAt) {
488 // Threading through a switch statement is particularly profitable. If this
489 // block ends in a switch, decrease its cost to make it more likely to
490 // happen.
491 if (isa<SwitchInst>(StopAt))
492 Bonus = 6;
493
494 // The same holds for indirect branches, but slightly more so.
495 if (isa<IndirectBrInst>(StopAt))
496 Bonus = 8;
497 }
498
499 // Bump the threshold up so the early exit from the loop doesn't skip the
500 // terminator-based Size adjustment at the end.
501 Threshold += Bonus;
502
503 // Sum up the cost of each instruction until we get to the terminator. Don't
504 // include the terminator because the copy won't include it.
505 unsigned Size = 0;
506 for (; &*I != StopAt; ++I) {
507
508 // Stop scanning the block if we've reached the threshold.
509 if (Size > Threshold)
510 return Size;
511
512 // Debugger intrinsics don't incur code size.
513 if (isa<DbgInfoIntrinsic>(I)) continue;
514
515 // If this is a pointer->pointer bitcast, it is free.
516 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
517 continue;
518
519 // Bail out if this instruction gives back a token type, it is not possible
520 // to duplicate it if it is used outside this BB.
521 if (I->getType()->isTokenTy() && I->isUsedOutsideOfBlock(BB))
522 return ~0U;
523
524 // All other instructions count for at least one unit.
525 ++Size;
526
527 // Calls are more expensive. If they are non-intrinsic calls, we model them
528 // as having cost of 4. If they are a non-vector intrinsic, we model them
529 // as having cost of 2 total, and if they are a vector intrinsic, we model
530 // them as having cost 1.
531 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
532 if (CI->cannotDuplicate() || CI->isConvergent())
533 // Blocks with NoDuplicate are modelled as having infinite cost, so they
534 // are never duplicated.
535 return ~0U;
536 else if (!isa<IntrinsicInst>(CI))
537 Size += 3;
538 else if (!CI->getType()->isVectorTy())
539 Size += 1;
540 }
541 }
542
543 return Size > Bonus ? Size - Bonus : 0;
544}
545
546/// FindLoopHeaders - We do not want jump threading to turn proper loop
547/// structures into irreducible loops. Doing this breaks up the loop nesting
548/// hierarchy and pessimizes later transformations. To prevent this from
549/// happening, we first have to find the loop headers. Here we approximate this
550/// by finding targets of backedges in the CFG.
551///
552/// Note that there definitely are cases when we want to allow threading of
553/// edges across a loop header. For example, threading a jump from outside the
554/// loop (the preheader) to an exit block of the loop is definitely profitable.
555/// It is also almost always profitable to thread backedges from within the loop
556/// to exit blocks, and is often profitable to thread backedges to other blocks
557/// within the loop (forming a nested loop). This simple analysis is not rich
558/// enough to track all of these properties and keep it up-to-date as the CFG
559/// mutates, so we don't allow any of these transformations.
560void JumpThreadingPass::FindLoopHeaders(Function &F) {
561 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
562 FindFunctionBackedges(F, Edges);
563
564 for (const auto &Edge : Edges)
565 LoopHeaders.insert(Edge.second);
566}
567
568/// getKnownConstant - Helper method to determine if we can thread over a
569/// terminator with the given value as its condition, and if so what value to
570/// use for that. What kind of value this is depends on whether we want an
571/// integer or a block address, but an undef is always accepted.
572/// Returns null if Val is null or not an appropriate constant.
573static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
574 if (!Val)
575 return nullptr;
576
577 // Undef is "known" enough.
578 if (UndefValue *U = dyn_cast<UndefValue>(Val))
579 return U;
580
581 if (Preference == WantBlockAddress)
582 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
583
584 return dyn_cast<ConstantInt>(Val);
585}
586
587/// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
588/// if we can infer that the value is a known ConstantInt/BlockAddress or undef
589/// in any of our predecessors. If so, return the known list of value and pred
590/// BB in the result vector.
591///
592/// This returns true if there were any known values.
593bool JumpThreadingPass::ComputeValueKnownInPredecessorsImpl(
594 Value *V, BasicBlock *BB, PredValueInfo &Result,
595 ConstantPreference Preference,
596 DenseSet<std::pair<Value *, BasicBlock *>> &RecursionSet,
597 Instruction *CxtI) {
598 // This method walks up use-def chains recursively. Because of this, we could
599 // get into an infinite loop going around loops in the use-def chain. To
600 // prevent this, keep track of what (value, block) pairs we've already visited
601 // and terminate the search if we loop back to them
602 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
603 return false;
604
605 // If V is a constant, then it is known in all predecessors.
606 if (Constant *KC = getKnownConstant(V, Preference)) {
607 for (BasicBlock *Pred : predecessors(BB))
608 Result.push_back(std::make_pair(KC, Pred));
609
610 return !Result.empty();
611 }
612
613 // If V is a non-instruction value, or an instruction in a different block,
614 // then it can't be derived from a PHI.
615 Instruction *I = dyn_cast<Instruction>(V);
616 if (!I || I->getParent() != BB) {
617
618 // Okay, if this is a live-in value, see if it has a known value at the end
619 // of any of our predecessors.
620 //
621 // FIXME: This should be an edge property, not a block end property.
622 /// TODO: Per PR2563, we could infer value range information about a
623 /// predecessor based on its terminator.
624 //
625 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
626 // "I" is a non-local compare-with-a-constant instruction. This would be
627 // able to handle value inequalities better, for example if the compare is
628 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
629 // Perhaps getConstantOnEdge should be smart enough to do this?
630
631 if (DTU->hasPendingDomTreeUpdates())
632 LVI->disableDT();
633 else
634 LVI->enableDT();
635 for (BasicBlock *P : predecessors(BB)) {
636 // If the value is known by LazyValueInfo to be a constant in a
637 // predecessor, use that information to try to thread this block.
638 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB, CxtI);
639 if (Constant *KC = getKnownConstant(PredCst, Preference))
640 Result.push_back(std::make_pair(KC, P));
641 }
642
643 return !Result.empty();
644 }
645
646 /// If I is a PHI node, then we know the incoming values for any constants.
647 if (PHINode *PN = dyn_cast<PHINode>(I)) {
648 if (DTU->hasPendingDomTreeUpdates())
649 LVI->disableDT();
650 else
651 LVI->enableDT();
652 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
653 Value *InVal = PN->getIncomingValue(i);
654 if (Constant *KC = getKnownConstant(InVal, Preference)) {
655 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
656 } else {
657 Constant *CI = LVI->getConstantOnEdge(InVal,
658 PN->getIncomingBlock(i),
659 BB, CxtI);
660 if (Constant *KC = getKnownConstant(CI, Preference))
661 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
662 }
663 }
664
665 return !Result.empty();
666 }
667
668 // Handle Cast instructions. Only see through Cast when the source operand is
669 // PHI or Cmp to save the compilation time.
670 if (CastInst *CI = dyn_cast<CastInst>(I)) {
671 Value *Source = CI->getOperand(0);
672 if (!isa<PHINode>(Source) && !isa<CmpInst>(Source))
673 return false;
674 ComputeValueKnownInPredecessorsImpl(Source, BB, Result, Preference,
675 RecursionSet, CxtI);
676 if (Result.empty())
677 return false;
678
679 // Convert the known values.
680 for (auto &R : Result)
681 R.first = ConstantExpr::getCast(CI->getOpcode(), R.first, CI->getType());
682
683 return true;
684 }
685
686 // Handle some boolean conditions.
687 if (I->getType()->getPrimitiveSizeInBits() == 1) {
688 assert(Preference == WantInteger && "One-bit non-integer type?")((Preference == WantInteger && "One-bit non-integer type?"
) ? static_cast<void> (0) : __assert_fail ("Preference == WantInteger && \"One-bit non-integer type?\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 688, __PRETTY_FUNCTION__))
;
689 // X | true -> true
690 // X & false -> false
691 if (I->getOpcode() == Instruction::Or ||
692 I->getOpcode() == Instruction::And) {
693 PredValueInfoTy LHSVals, RHSVals;
694
695 ComputeValueKnownInPredecessorsImpl(I->getOperand(0), BB, LHSVals,
696 WantInteger, RecursionSet, CxtI);
697 ComputeValueKnownInPredecessorsImpl(I->getOperand(1), BB, RHSVals,
698 WantInteger, RecursionSet, CxtI);
699
700 if (LHSVals.empty() && RHSVals.empty())
701 return false;
702
703 ConstantInt *InterestingVal;
704 if (I->getOpcode() == Instruction::Or)
705 InterestingVal = ConstantInt::getTrue(I->getContext());
706 else
707 InterestingVal = ConstantInt::getFalse(I->getContext());
708
709 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
710
711 // Scan for the sentinel. If we find an undef, force it to the
712 // interesting value: x|undef -> true and x&undef -> false.
713 for (const auto &LHSVal : LHSVals)
714 if (LHSVal.first == InterestingVal || isa<UndefValue>(LHSVal.first)) {
715 Result.emplace_back(InterestingVal, LHSVal.second);
716 LHSKnownBBs.insert(LHSVal.second);
717 }
718 for (const auto &RHSVal : RHSVals)
719 if (RHSVal.first == InterestingVal || isa<UndefValue>(RHSVal.first)) {
720 // If we already inferred a value for this block on the LHS, don't
721 // re-add it.
722 if (!LHSKnownBBs.count(RHSVal.second))
723 Result.emplace_back(InterestingVal, RHSVal.second);
724 }
725
726 return !Result.empty();
727 }
728
729 // Handle the NOT form of XOR.
730 if (I->getOpcode() == Instruction::Xor &&
731 isa<ConstantInt>(I->getOperand(1)) &&
732 cast<ConstantInt>(I->getOperand(1))->isOne()) {
733 ComputeValueKnownInPredecessorsImpl(I->getOperand(0), BB, Result,
734 WantInteger, RecursionSet, CxtI);
735 if (Result.empty())
736 return false;
737
738 // Invert the known values.
739 for (auto &R : Result)
740 R.first = ConstantExpr::getNot(R.first);
741
742 return true;
743 }
744
745 // Try to simplify some other binary operator values.
746 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
747 assert(Preference != WantBlockAddress((Preference != WantBlockAddress && "A binary operator creating a block address?"
) ? static_cast<void> (0) : __assert_fail ("Preference != WantBlockAddress && \"A binary operator creating a block address?\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 748, __PRETTY_FUNCTION__))
748 && "A binary operator creating a block address?")((Preference != WantBlockAddress && "A binary operator creating a block address?"
) ? static_cast<void> (0) : __assert_fail ("Preference != WantBlockAddress && \"A binary operator creating a block address?\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 748, __PRETTY_FUNCTION__))
;
749 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
750 PredValueInfoTy LHSVals;
751 ComputeValueKnownInPredecessorsImpl(BO->getOperand(0), BB, LHSVals,
752 WantInteger, RecursionSet, CxtI);
753
754 // Try to use constant folding to simplify the binary operator.
755 for (const auto &LHSVal : LHSVals) {
756 Constant *V = LHSVal.first;
757 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
758
759 if (Constant *KC = getKnownConstant(Folded, WantInteger))
760 Result.push_back(std::make_pair(KC, LHSVal.second));
761 }
762 }
763
764 return !Result.empty();
765 }
766
767 // Handle compare with phi operand, where the PHI is defined in this block.
768 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
769 assert(Preference == WantInteger && "Compares only produce integers")((Preference == WantInteger && "Compares only produce integers"
) ? static_cast<void> (0) : __assert_fail ("Preference == WantInteger && \"Compares only produce integers\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 769, __PRETTY_FUNCTION__))
;
770 Type *CmpType = Cmp->getType();
771 Value *CmpLHS = Cmp->getOperand(0);
772 Value *CmpRHS = Cmp->getOperand(1);
773 CmpInst::Predicate Pred = Cmp->getPredicate();
774
775 PHINode *PN = dyn_cast<PHINode>(CmpLHS);
776 if (!PN)
777 PN = dyn_cast<PHINode>(CmpRHS);
778 if (PN && PN->getParent() == BB) {
779 const DataLayout &DL = PN->getModule()->getDataLayout();
780 // We can do this simplification if any comparisons fold to true or false.
781 // See if any do.
782 if (DTU->hasPendingDomTreeUpdates())
783 LVI->disableDT();
784 else
785 LVI->enableDT();
786 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
787 BasicBlock *PredBB = PN->getIncomingBlock(i);
788 Value *LHS, *RHS;
789 if (PN == CmpLHS) {
790 LHS = PN->getIncomingValue(i);
791 RHS = CmpRHS->DoPHITranslation(BB, PredBB);
792 } else {
793 LHS = CmpLHS->DoPHITranslation(BB, PredBB);
794 RHS = PN->getIncomingValue(i);
795 }
796 Value *Res = SimplifyCmpInst(Pred, LHS, RHS, {DL});
797 if (!Res) {
798 if (!isa<Constant>(RHS))
799 continue;
800
801 // getPredicateOnEdge call will make no sense if LHS is defined in BB.
802 auto LHSInst = dyn_cast<Instruction>(LHS);
803 if (LHSInst && LHSInst->getParent() == BB)
804 continue;
805
806 LazyValueInfo::Tristate
807 ResT = LVI->getPredicateOnEdge(Pred, LHS,
808 cast<Constant>(RHS), PredBB, BB,
809 CxtI ? CxtI : Cmp);
810 if (ResT == LazyValueInfo::Unknown)
811 continue;
812 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
813 }
814
815 if (Constant *KC = getKnownConstant(Res, WantInteger))
816 Result.push_back(std::make_pair(KC, PredBB));
817 }
818
819 return !Result.empty();
820 }
821
822 // If comparing a live-in value against a constant, see if we know the
823 // live-in value on any predecessors.
824 if (isa<Constant>(CmpRHS) && !CmpType->isVectorTy()) {
825 Constant *CmpConst = cast<Constant>(CmpRHS);
826
827 if (!isa<Instruction>(CmpLHS) ||
828 cast<Instruction>(CmpLHS)->getParent() != BB) {
829 if (DTU->hasPendingDomTreeUpdates())
830 LVI->disableDT();
831 else
832 LVI->enableDT();
833 for (BasicBlock *P : predecessors(BB)) {
834 // If the value is known by LazyValueInfo to be a constant in a
835 // predecessor, use that information to try to thread this block.
836 LazyValueInfo::Tristate Res =
837 LVI->getPredicateOnEdge(Pred, CmpLHS,
838 CmpConst, P, BB, CxtI ? CxtI : Cmp);
839 if (Res == LazyValueInfo::Unknown)
840 continue;
841
842 Constant *ResC = ConstantInt::get(CmpType, Res);
843 Result.push_back(std::make_pair(ResC, P));
844 }
845
846 return !Result.empty();
847 }
848
849 // InstCombine can fold some forms of constant range checks into
850 // (icmp (add (x, C1)), C2). See if we have we have such a thing with
851 // x as a live-in.
852 {
853 using namespace PatternMatch;
854
855 Value *AddLHS;
856 ConstantInt *AddConst;
857 if (isa<ConstantInt>(CmpConst) &&
858 match(CmpLHS, m_Add(m_Value(AddLHS), m_ConstantInt(AddConst)))) {
859 if (!isa<Instruction>(AddLHS) ||
860 cast<Instruction>(AddLHS)->getParent() != BB) {
861 if (DTU->hasPendingDomTreeUpdates())
862 LVI->disableDT();
863 else
864 LVI->enableDT();
865 for (BasicBlock *P : predecessors(BB)) {
866 // If the value is known by LazyValueInfo to be a ConstantRange in
867 // a predecessor, use that information to try to thread this
868 // block.
869 ConstantRange CR = LVI->getConstantRangeOnEdge(
870 AddLHS, P, BB, CxtI ? CxtI : cast<Instruction>(CmpLHS));
871 // Propagate the range through the addition.
872 CR = CR.add(AddConst->getValue());
873
874 // Get the range where the compare returns true.
875 ConstantRange CmpRange = ConstantRange::makeExactICmpRegion(
876 Pred, cast<ConstantInt>(CmpConst)->getValue());
877
878 Constant *ResC;
879 if (CmpRange.contains(CR))
880 ResC = ConstantInt::getTrue(CmpType);
881 else if (CmpRange.inverse().contains(CR))
882 ResC = ConstantInt::getFalse(CmpType);
883 else
884 continue;
885
886 Result.push_back(std::make_pair(ResC, P));
887 }
888
889 return !Result.empty();
890 }
891 }
892 }
893
894 // Try to find a constant value for the LHS of a comparison,
895 // and evaluate it statically if we can.
896 PredValueInfoTy LHSVals;
897 ComputeValueKnownInPredecessorsImpl(I->getOperand(0), BB, LHSVals,
898 WantInteger, RecursionSet, CxtI);
899
900 for (const auto &LHSVal : LHSVals) {
901 Constant *V = LHSVal.first;
902 Constant *Folded = ConstantExpr::getCompare(Pred, V, CmpConst);
903 if (Constant *KC = getKnownConstant(Folded, WantInteger))
904 Result.push_back(std::make_pair(KC, LHSVal.second));
905 }
906
907 return !Result.empty();
908 }
909 }
910
911 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
912 // Handle select instructions where at least one operand is a known constant
913 // and we can figure out the condition value for any predecessor block.
914 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
915 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
916 PredValueInfoTy Conds;
917 if ((TrueVal || FalseVal) &&
918 ComputeValueKnownInPredecessorsImpl(SI->getCondition(), BB, Conds,
919 WantInteger, RecursionSet, CxtI)) {
920 for (auto &C : Conds) {
921 Constant *Cond = C.first;
922
923 // Figure out what value to use for the condition.
924 bool KnownCond;
925 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
926 // A known boolean.
927 KnownCond = CI->isOne();
928 } else {
929 assert(isa<UndefValue>(Cond) && "Unexpected condition value")((isa<UndefValue>(Cond) && "Unexpected condition value"
) ? static_cast<void> (0) : __assert_fail ("isa<UndefValue>(Cond) && \"Unexpected condition value\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 929, __PRETTY_FUNCTION__))
;
930 // Either operand will do, so be sure to pick the one that's a known
931 // constant.
932 // FIXME: Do this more cleverly if both values are known constants?
933 KnownCond = (TrueVal != nullptr);
934 }
935
936 // See if the select has a known constant value for this predecessor.
937 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
938 Result.push_back(std::make_pair(Val, C.second));
939 }
940
941 return !Result.empty();
942 }
943 }
944
945 // If all else fails, see if LVI can figure out a constant value for us.
946 if (DTU->hasPendingDomTreeUpdates())
947 LVI->disableDT();
948 else
949 LVI->enableDT();
950 Constant *CI = LVI->getConstant(V, BB, CxtI);
951 if (Constant *KC = getKnownConstant(CI, Preference)) {
952 for (BasicBlock *Pred : predecessors(BB))
953 Result.push_back(std::make_pair(KC, Pred));
954 }
955
956 return !Result.empty();
957}
958
959/// GetBestDestForBranchOnUndef - If we determine that the specified block ends
960/// in an undefined jump, decide which block is best to revector to.
961///
962/// Since we can pick an arbitrary destination, we pick the successor with the
963/// fewest predecessors. This should reduce the in-degree of the others.
964static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
965 Instruction *BBTerm = BB->getTerminator();
966 unsigned MinSucc = 0;
967 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
968 // Compute the successor with the minimum number of predecessors.
969 unsigned MinNumPreds = pred_size(TestBB);
970 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
971 TestBB = BBTerm->getSuccessor(i);
972 unsigned NumPreds = pred_size(TestBB);
973 if (NumPreds < MinNumPreds) {
974 MinSucc = i;
975 MinNumPreds = NumPreds;
976 }
977 }
978
979 return MinSucc;
980}
981
982static bool hasAddressTakenAndUsed(BasicBlock *BB) {
983 if (!BB->hasAddressTaken()) return false;
984
985 // If the block has its address taken, it may be a tree of dead constants
986 // hanging off of it. These shouldn't keep the block alive.
987 BlockAddress *BA = BlockAddress::get(BB);
988 BA->removeDeadConstantUsers();
989 return !BA->use_empty();
990}
991
992/// ProcessBlock - If there are any predecessors whose control can be threaded
993/// through to a successor, transform them now.
994bool JumpThreadingPass::ProcessBlock(BasicBlock *BB) {
995 // If the block is trivially dead, just return and let the caller nuke it.
996 // This simplifies other transformations.
997 if (DTU->isBBPendingDeletion(BB) ||
1
Assuming the condition is false
998 (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()))
2
Assuming the condition is false
999 return false;
1000
1001 // If this block has a single predecessor, and if that pred has a single
1002 // successor, merge the blocks. This encourages recursive jump threading
1003 // because now the condition in this block can be threaded through
1004 // predecessors of our predecessor block.
1005 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
3
Assuming 'SinglePred' is null
4
Taking false branch
1006 const Instruction *TI = SinglePred->getTerminator();
1007 if (!TI->isExceptionalTerminator() && TI->getNumSuccessors() == 1 &&
1008 SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
1009 // If SinglePred was a loop header, BB becomes one.
1010 if (LoopHeaders.erase(SinglePred))
1011 LoopHeaders.insert(BB);
1012
1013 LVI->eraseBlock(SinglePred);
1014 MergeBasicBlockIntoOnlyPred(BB, DTU);
1015
1016 // Now that BB is merged into SinglePred (i.e. SinglePred Code followed by
1017 // BB code within one basic block `BB`), we need to invalidate the LVI
1018 // information associated with BB, because the LVI information need not be
1019 // true for all of BB after the merge. For example,
1020 // Before the merge, LVI info and code is as follows:
1021 // SinglePred: <LVI info1 for %p val>
1022 // %y = use of %p
1023 // call @exit() // need not transfer execution to successor.
1024 // assume(%p) // from this point on %p is true
1025 // br label %BB
1026 // BB: <LVI info2 for %p val, i.e. %p is true>
1027 // %x = use of %p
1028 // br label exit
1029 //
1030 // Note that this LVI info for blocks BB and SinglPred is correct for %p
1031 // (info2 and info1 respectively). After the merge and the deletion of the
1032 // LVI info1 for SinglePred. We have the following code:
1033 // BB: <LVI info2 for %p val>
1034 // %y = use of %p
1035 // call @exit()
1036 // assume(%p)
1037 // %x = use of %p <-- LVI info2 is correct from here onwards.
1038 // br label exit
1039 // LVI info2 for BB is incorrect at the beginning of BB.
1040
1041 // Invalidate LVI information for BB if the LVI is not provably true for
1042 // all of BB.
1043 if (!isGuaranteedToTransferExecutionToSuccessor(BB))
1044 LVI->eraseBlock(BB);
1045 return true;
1046 }
1047 }
1048
1049 if (TryToUnfoldSelectInCurrBB(BB))
5
Taking false branch
1050 return true;
1051
1052 // Look if we can propagate guards to predecessors.
1053 if (HasGuards && ProcessGuards(BB))
6
Assuming field 'HasGuards' is false
1054 return true;
1055
1056 // What kind of constant we're looking for.
1057 ConstantPreference Preference = WantInteger;
1058
1059 // Look to see if the terminator is a conditional branch, switch or indirect
1060 // branch, if not we can't thread it.
1061 Value *Condition;
1062 Instruction *Terminator = BB->getTerminator();
1063 if (BranchInst *BI
7.1
'BI' is null
7.1
'BI' is null
7.1
'BI' is null
7.1
'BI' is null
= dyn_cast<BranchInst>(Terminator)) {
7
Assuming 'Terminator' is not a 'BranchInst'
8
Taking false branch
1064 // Can't thread an unconditional jump.
1065 if (BI->isUnconditional()) return false;
1066 Condition = BI->getCondition();
1067 } else if (SwitchInst *SI
9.1
'SI' is null
9.1
'SI' is null
9.1
'SI' is null
9.1
'SI' is null
= dyn_cast<SwitchInst>(Terminator)) {
9
Assuming 'Terminator' is not a 'SwitchInst'
10
Taking false branch
1068 Condition = SI->getCondition();
1069 } else if (IndirectBrInst *IB
11.1
'IB' is non-null
11.1
'IB' is non-null
11.1
'IB' is non-null
11.1
'IB' is non-null
= dyn_cast<IndirectBrInst>(Terminator)) {
11
Assuming 'Terminator' is a 'IndirectBrInst'
12
Taking true branch
1070 // Can't thread indirect branch with no successors.
1071 if (IB->getNumSuccessors() == 0) return false;
13
Assuming the condition is false
14
Taking false branch
1072 Condition = IB->getAddress()->stripPointerCasts();
1073 Preference = WantBlockAddress;
1074 } else {
1075 return false; // Must be an invoke or callbr.
1076 }
1077
1078 // Run constant folding to see if we can reduce the condition to a simple
1079 // constant.
1080 if (Instruction *I
15.1
'I' is non-null
15.1
'I' is non-null
15.1
'I' is non-null
15.1
'I' is non-null
= dyn_cast<Instruction>(Condition)) {
15
Assuming 'Condition' is a 'Instruction'
16
Taking true branch
1081 Value *SimpleVal =
1082 ConstantFoldInstruction(I, BB->getModule()->getDataLayout(), TLI);
1083 if (SimpleVal) {
17
Assuming 'SimpleVal' is null
18
Taking false branch
1084 I->replaceAllUsesWith(SimpleVal);
1085 if (isInstructionTriviallyDead(I, TLI))
1086 I->eraseFromParent();
1087 Condition = SimpleVal;
1088 }
1089 }
1090
1091 // If the terminator is branching on an undef, we can pick any of the
1092 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
1093 if (isa<UndefValue>(Condition)) {
19
Assuming 'Condition' is not a 'UndefValue'
20
Taking false branch
1094 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
1095 std::vector<DominatorTree::UpdateType> Updates;
1096
1097 // Fold the branch/switch.
1098 Instruction *BBTerm = BB->getTerminator();
1099 Updates.reserve(BBTerm->getNumSuccessors());
1100 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
1101 if (i == BestSucc) continue;
1102 BasicBlock *Succ = BBTerm->getSuccessor(i);
1103 Succ->removePredecessor(BB, true);
1104 Updates.push_back({DominatorTree::Delete, BB, Succ});
1105 }
1106
1107 LLVM_DEBUG(dbgs() << " In block '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " In block '" <<
BB->getName() << "' folding undef terminator: " <<
*BBTerm << '\n'; } } while (false)
1108 << "' folding undef terminator: " << *BBTerm << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " In block '" <<
BB->getName() << "' folding undef terminator: " <<
*BBTerm << '\n'; } } while (false)
;
1109 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
1110 BBTerm->eraseFromParent();
1111 DTU->applyUpdatesPermissive(Updates);
1112 return true;
1113 }
1114
1115 // If the terminator of this block is branching on a constant, simplify the
1116 // terminator to an unconditional branch. This can occur due to threading in
1117 // other blocks.
1118 if (getKnownConstant(Condition, Preference)) {
21
Taking false branch
1119 LLVM_DEBUG(dbgs() << " In block '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " In block '" <<
BB->getName() << "' folding terminator: " << *
BB->getTerminator() << '\n'; } } while (false)
1120 << "' folding terminator: " << *BB->getTerminator()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " In block '" <<
BB->getName() << "' folding terminator: " << *
BB->getTerminator() << '\n'; } } while (false)
1121 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " In block '" <<
BB->getName() << "' folding terminator: " << *
BB->getTerminator() << '\n'; } } while (false)
;
1122 ++NumFolds;
1123 ConstantFoldTerminator(BB, true, nullptr, DTU);
1124 return true;
1125 }
1126
1127 Instruction *CondInst = dyn_cast<Instruction>(Condition);
22
'Condition' is a 'Instruction'
1128
1129 // All the rest of our checks depend on the condition being an instruction.
1130 if (!CondInst
22.1
'CondInst' is non-null
22.1
'CondInst' is non-null
22.1
'CondInst' is non-null
22.1
'CondInst' is non-null
) {
23
Taking false branch
1131 // FIXME: Unify this with code below.
1132 if (ProcessThreadableEdges(Condition, BB, Preference, Terminator))
1133 return true;
1134 return false;
1135 }
1136
1137 if (CmpInst *CondCmp
24.1
'CondCmp' is null
24.1
'CondCmp' is null
24.1
'CondCmp' is null
24.1
'CondCmp' is null
= dyn_cast<CmpInst>(CondInst)) {
24
Assuming 'CondInst' is not a 'CmpInst'
25
Taking false branch
1138 // If we're branching on a conditional, LVI might be able to determine
1139 // it's value at the branch instruction. We only handle comparisons
1140 // against a constant at this time.
1141 // TODO: This should be extended to handle switches as well.
1142 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1143 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
1144 if (CondBr && CondConst) {
1145 // We should have returned as soon as we turn a conditional branch to
1146 // unconditional. Because its no longer interesting as far as jump
1147 // threading is concerned.
1148 assert(CondBr->isConditional() && "Threading on unconditional terminator")((CondBr->isConditional() && "Threading on unconditional terminator"
) ? static_cast<void> (0) : __assert_fail ("CondBr->isConditional() && \"Threading on unconditional terminator\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 1148, __PRETTY_FUNCTION__))
;
1149
1150 if (DTU->hasPendingDomTreeUpdates())
1151 LVI->disableDT();
1152 else
1153 LVI->enableDT();
1154 LazyValueInfo::Tristate Ret =
1155 LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0),
1156 CondConst, CondBr);
1157 if (Ret != LazyValueInfo::Unknown) {
1158 unsigned ToRemove = Ret == LazyValueInfo::True ? 1 : 0;
1159 unsigned ToKeep = Ret == LazyValueInfo::True ? 0 : 1;
1160 BasicBlock *ToRemoveSucc = CondBr->getSuccessor(ToRemove);
1161 ToRemoveSucc->removePredecessor(BB, true);
1162 BranchInst *UncondBr =
1163 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
1164 UncondBr->setDebugLoc(CondBr->getDebugLoc());
1165 CondBr->eraseFromParent();
1166 if (CondCmp->use_empty())
1167 CondCmp->eraseFromParent();
1168 // We can safely replace *some* uses of the CondInst if it has
1169 // exactly one value as returned by LVI. RAUW is incorrect in the
1170 // presence of guards and assumes, that have the `Cond` as the use. This
1171 // is because we use the guards/assume to reason about the `Cond` value
1172 // at the end of block, but RAUW unconditionally replaces all uses
1173 // including the guards/assumes themselves and the uses before the
1174 // guard/assume.
1175 else if (CondCmp->getParent() == BB) {
1176 auto *CI = Ret == LazyValueInfo::True ?
1177 ConstantInt::getTrue(CondCmp->getType()) :
1178 ConstantInt::getFalse(CondCmp->getType());
1179 ReplaceFoldableUses(CondCmp, CI);
1180 }
1181 DTU->applyUpdatesPermissive(
1182 {{DominatorTree::Delete, BB, ToRemoveSucc}});
1183 return true;
1184 }
1185
1186 // We did not manage to simplify this branch, try to see whether
1187 // CondCmp depends on a known phi-select pattern.
1188 if (TryToUnfoldSelect(CondCmp, BB))
1189 return true;
1190 }
1191 }
1192
1193 if (SwitchInst *SI
26.1
'SI' is null
26.1
'SI' is null
26.1
'SI' is null
26.1
'SI' is null
= dyn_cast<SwitchInst>(BB->getTerminator()))
26
Assuming the object is not a 'SwitchInst'
27
Taking false branch
1194 if (TryToUnfoldSelect(SI, BB))
1195 return true;
1196
1197 // Check for some cases that are worth simplifying. Right now we want to look
1198 // for loads that are used by a switch or by the condition for the branch. If
1199 // we see one, check to see if it's partially redundant. If so, insert a PHI
1200 // which can then be used to thread the values.
1201 Value *SimplifyValue = CondInst;
1202 if (CmpInst *CondCmp
28.1
'CondCmp' is null
28.1
'CondCmp' is null
28.1
'CondCmp' is null
28.1
'CondCmp' is null
= dyn_cast<CmpInst>(SimplifyValue))
28
Assuming 'SimplifyValue' is not a 'CmpInst'
29
Taking false branch
1203 if (isa<Constant>(CondCmp->getOperand(1)))
1204 SimplifyValue = CondCmp->getOperand(0);
1205
1206 // TODO: There are other places where load PRE would be profitable, such as
1207 // more complex comparisons.
1208 if (LoadInst *LoadI
30.1
'LoadI' is non-null
30.1
'LoadI' is non-null
30.1
'LoadI' is non-null
30.1
'LoadI' is non-null
= dyn_cast<LoadInst>(SimplifyValue))
30
Assuming 'SimplifyValue' is a 'LoadInst'
31
Taking true branch
1209 if (SimplifyPartiallyRedundantLoad(LoadI))
32
Calling 'JumpThreadingPass::SimplifyPartiallyRedundantLoad'
1210 return true;
1211
1212 // Before threading, try to propagate profile data backwards:
1213 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
1214 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1215 updatePredecessorProfileMetadata(PN, BB);
1216
1217 // Handle a variety of cases where we are branching on something derived from
1218 // a PHI node in the current block. If we can prove that any predecessors
1219 // compute a predictable value based on a PHI node, thread those predecessors.
1220 if (ProcessThreadableEdges(CondInst, BB, Preference, Terminator))
1221 return true;
1222
1223 // If this is an otherwise-unfoldable branch on a phi node in the current
1224 // block, see if we can simplify.
1225 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
1226 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1227 return ProcessBranchOnPHI(PN);
1228
1229 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
1230 if (CondInst->getOpcode() == Instruction::Xor &&
1231 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1232 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
1233
1234 // Search for a stronger dominating condition that can be used to simplify a
1235 // conditional branch leaving BB.
1236 if (ProcessImpliedCondition(BB))
1237 return true;
1238
1239 return false;
1240}
1241
1242bool JumpThreadingPass::ProcessImpliedCondition(BasicBlock *BB) {
1243 auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
1244 if (!BI || !BI->isConditional())
1245 return false;
1246
1247 Value *Cond = BI->getCondition();
1248 BasicBlock *CurrentBB = BB;
1249 BasicBlock *CurrentPred = BB->getSinglePredecessor();
1250 unsigned Iter = 0;
1251
1252 auto &DL = BB->getModule()->getDataLayout();
1253
1254 while (CurrentPred && Iter++ < ImplicationSearchThreshold) {
1255 auto *PBI = dyn_cast<BranchInst>(CurrentPred->getTerminator());
1256 if (!PBI || !PBI->isConditional())
1257 return false;
1258 if (PBI->getSuccessor(0) != CurrentBB && PBI->getSuccessor(1) != CurrentBB)
1259 return false;
1260
1261 bool CondIsTrue = PBI->getSuccessor(0) == CurrentBB;
1262 Optional<bool> Implication =
1263 isImpliedCondition(PBI->getCondition(), Cond, DL, CondIsTrue);
1264 if (Implication) {
1265 BasicBlock *KeepSucc = BI->getSuccessor(*Implication ? 0 : 1);
1266 BasicBlock *RemoveSucc = BI->getSuccessor(*Implication ? 1 : 0);
1267 RemoveSucc->removePredecessor(BB);
1268 BranchInst *UncondBI = BranchInst::Create(KeepSucc, BI);
1269 UncondBI->setDebugLoc(BI->getDebugLoc());
1270 BI->eraseFromParent();
1271 DTU->applyUpdatesPermissive({{DominatorTree::Delete, BB, RemoveSucc}});
1272 return true;
1273 }
1274 CurrentBB = CurrentPred;
1275 CurrentPred = CurrentBB->getSinglePredecessor();
1276 }
1277
1278 return false;
1279}
1280
1281/// Return true if Op is an instruction defined in the given block.
1282static bool isOpDefinedInBlock(Value *Op, BasicBlock *BB) {
1283 if (Instruction *OpInst = dyn_cast<Instruction>(Op))
45
Assuming 'OpInst' is null
46
Taking false branch
1284 if (OpInst->getParent() == BB)
1285 return true;
1286 return false;
47
Returning zero, which participates in a condition later
1287}
1288
1289/// SimplifyPartiallyRedundantLoad - If LoadI is an obviously partially
1290/// redundant load instruction, eliminate it by replacing it with a PHI node.
1291/// This is an important optimization that encourages jump threading, and needs
1292/// to be run interlaced with other jump threading tasks.
1293bool JumpThreadingPass::SimplifyPartiallyRedundantLoad(LoadInst *LoadI) {
1294 // Don't hack volatile and ordered loads.
1295 if (!LoadI->isUnordered()) return false;
33
Calling 'LoadInst::isUnordered'
38
Returning from 'LoadInst::isUnordered'
39
Taking false branch
1296
1297 // If the load is defined in a block with exactly one predecessor, it can't be
1298 // partially redundant.
1299 BasicBlock *LoadBB = LoadI->getParent();
1300 if (LoadBB->getSinglePredecessor())
40
Assuming the condition is false
41
Taking false branch
1301 return false;
1302
1303 // If the load is defined in an EH pad, it can't be partially redundant,
1304 // because the edges between the invoke and the EH pad cannot have other
1305 // instructions between them.
1306 if (LoadBB->isEHPad())
42
Assuming the condition is false
43
Taking false branch
1307 return false;
1308
1309 Value *LoadedPtr = LoadI->getOperand(0);
1310
1311 // If the loaded operand is defined in the LoadBB and its not a phi,
1312 // it can't be available in predecessors.
1313 if (isOpDefinedInBlock(LoadedPtr, LoadBB) && !isa<PHINode>(LoadedPtr))
44
Calling 'isOpDefinedInBlock'
48
Returning from 'isOpDefinedInBlock'
1314 return false;
1315
1316 // Scan a few instructions up from the load, to see if it is obviously live at
1317 // the entry to its block.
1318 BasicBlock::iterator BBIt(LoadI);
1319 bool IsLoadCSE;
1320 if (Value *AvailableVal = FindAvailableLoadedValue(
49
Assuming 'AvailableVal' is null
50
Taking false branch
1321 LoadI, LoadBB, BBIt, DefMaxInstsToScan, AA, &IsLoadCSE)) {
1322 // If the value of the load is locally available within the block, just use
1323 // it. This frequently occurs for reg2mem'd allocas.
1324
1325 if (IsLoadCSE) {
1326 LoadInst *NLoadI = cast<LoadInst>(AvailableVal);
1327 combineMetadataForCSE(NLoadI, LoadI, false);
1328 };
1329
1330 // If the returned value is the load itself, replace with an undef. This can
1331 // only happen in dead loops.
1332 if (AvailableVal == LoadI)
1333 AvailableVal = UndefValue::get(LoadI->getType());
1334 if (AvailableVal->getType() != LoadI->getType())
1335 AvailableVal = CastInst::CreateBitOrPointerCast(
1336 AvailableVal, LoadI->getType(), "", LoadI);
1337 LoadI->replaceAllUsesWith(AvailableVal);
1338 LoadI->eraseFromParent();
1339 return true;
1340 }
1341
1342 // Otherwise, if we scanned the whole block and got to the top of the block,
1343 // we know the block is locally transparent to the load. If not, something
1344 // might clobber its value.
1345 if (BBIt != LoadBB->begin())
51
Calling 'operator!='
54
Returning from 'operator!='
55
Taking false branch
1346 return false;
1347
1348 // If all of the loads and stores that feed the value have the same AA tags,
1349 // then we can propagate them onto any newly inserted loads.
1350 AAMDNodes AATags;
1351 LoadI->getAAMetadata(AATags);
1352
1353 SmallPtrSet<BasicBlock*, 8> PredsScanned;
1354
1355 using AvailablePredsTy = SmallVector<std::pair<BasicBlock *, Value *>, 8>;
1356
1357 AvailablePredsTy AvailablePreds;
1358 BasicBlock *OneUnavailablePred = nullptr;
56
'OneUnavailablePred' initialized to a null pointer value
1359 SmallVector<LoadInst*, 8> CSELoads;
1360
1361 // If we got here, the loaded value is transparent through to the start of the
1362 // block. Check to see if it is available in any of the predecessor blocks.
1363 for (BasicBlock *PredBB : predecessors(LoadBB)) {
1364 // If we already scanned this predecessor, skip it.
1365 if (!PredsScanned.insert(PredBB).second)
1366 continue;
1367
1368 BBIt = PredBB->end();
1369 unsigned NumScanedInst = 0;
1370 Value *PredAvailable = nullptr;
1371 // NOTE: We don't CSE load that is volatile or anything stronger than
1372 // unordered, that should have been checked when we entered the function.
1373 assert(LoadI->isUnordered() &&((LoadI->isUnordered() && "Attempting to CSE volatile or atomic loads"
) ? static_cast<void> (0) : __assert_fail ("LoadI->isUnordered() && \"Attempting to CSE volatile or atomic loads\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 1374, __PRETTY_FUNCTION__))
1374 "Attempting to CSE volatile or atomic loads")((LoadI->isUnordered() && "Attempting to CSE volatile or atomic loads"
) ? static_cast<void> (0) : __assert_fail ("LoadI->isUnordered() && \"Attempting to CSE volatile or atomic loads\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 1374, __PRETTY_FUNCTION__))
;
1375 // If this is a load on a phi pointer, phi-translate it and search
1376 // for available load/store to the pointer in predecessors.
1377 Value *Ptr = LoadedPtr->DoPHITranslation(LoadBB, PredBB);
1378 PredAvailable = FindAvailablePtrLoadStore(
1379 Ptr, LoadI->getType(), LoadI->isAtomic(), PredBB, BBIt,
1380 DefMaxInstsToScan, AA, &IsLoadCSE, &NumScanedInst);
1381
1382 // If PredBB has a single predecessor, continue scanning through the
1383 // single predecessor.
1384 BasicBlock *SinglePredBB = PredBB;
1385 while (!PredAvailable && SinglePredBB && BBIt == SinglePredBB->begin() &&
1386 NumScanedInst < DefMaxInstsToScan) {
1387 SinglePredBB = SinglePredBB->getSinglePredecessor();
1388 if (SinglePredBB) {
1389 BBIt = SinglePredBB->end();
1390 PredAvailable = FindAvailablePtrLoadStore(
1391 Ptr, LoadI->getType(), LoadI->isAtomic(), SinglePredBB, BBIt,
1392 (DefMaxInstsToScan - NumScanedInst), AA, &IsLoadCSE,
1393 &NumScanedInst);
1394 }
1395 }
1396
1397 if (!PredAvailable) {
1398 OneUnavailablePred = PredBB;
1399 continue;
1400 }
1401
1402 if (IsLoadCSE)
1403 CSELoads.push_back(cast<LoadInst>(PredAvailable));
1404
1405 // If so, this load is partially redundant. Remember this info so that we
1406 // can create a PHI node.
1407 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
1408 }
1409
1410 // If the loaded value isn't available in any predecessor, it isn't partially
1411 // redundant.
1412 if (AvailablePreds.empty()) return false;
57
Calling 'SmallVectorBase::empty'
60
Returning from 'SmallVectorBase::empty'
61
Taking false branch
1413
1414 // Okay, the loaded value is available in at least one (and maybe all!)
1415 // predecessors. If the value is unavailable in more than one unique
1416 // predecessor, we want to insert a merge block for those common predecessors.
1417 // This ensures that we only have to insert one reload, thus not increasing
1418 // code size.
1419 BasicBlock *UnavailablePred = nullptr;
1420
1421 // If the value is unavailable in one of predecessors, we will end up
1422 // inserting a new instruction into them. It is only valid if all the
1423 // instructions before LoadI are guaranteed to pass execution to its
1424 // successor, or if LoadI is safe to speculate.
1425 // TODO: If this logic becomes more complex, and we will perform PRE insertion
1426 // farther than to a predecessor, we need to reuse the code from GVN's PRE.
1427 // It requires domination tree analysis, so for this simple case it is an
1428 // overkill.
1429 if (PredsScanned.size() != AvailablePreds.size() &&
62
Assuming the condition is false
1430 !isSafeToSpeculativelyExecute(LoadI))
1431 for (auto I = LoadBB->begin(); &*I != LoadI; ++I)
1432 if (!isGuaranteedToTransferExecutionToSuccessor(&*I))
1433 return false;
1434
1435 // If there is exactly one predecessor where the value is unavailable, the
1436 // already computed 'OneUnavailablePred' block is it. If it ends in an
1437 // unconditional branch, we know that it isn't a critical edge.
1438 if (PredsScanned.size() == AvailablePreds.size()+1 &&
63
Assuming the condition is true
1439 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
64
Called C++ object pointer is null
1440 UnavailablePred = OneUnavailablePred;
1441 } else if (PredsScanned.size() != AvailablePreds.size()) {
1442 // Otherwise, we had multiple unavailable predecessors or we had a critical
1443 // edge from the one.
1444 SmallVector<BasicBlock*, 8> PredsToSplit;
1445 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
1446
1447 for (const auto &AvailablePred : AvailablePreds)
1448 AvailablePredSet.insert(AvailablePred.first);
1449
1450 // Add all the unavailable predecessors to the PredsToSplit list.
1451 for (BasicBlock *P : predecessors(LoadBB)) {
1452 // If the predecessor is an indirect goto, we can't split the edge.
1453 // Same for CallBr.
1454 if (isa<IndirectBrInst>(P->getTerminator()) ||
1455 isa<CallBrInst>(P->getTerminator()))
1456 return false;
1457
1458 if (!AvailablePredSet.count(P))
1459 PredsToSplit.push_back(P);
1460 }
1461
1462 // Split them out to their own block.
1463 UnavailablePred = SplitBlockPreds(LoadBB, PredsToSplit, "thread-pre-split");
1464 }
1465
1466 // If the value isn't available in all predecessors, then there will be
1467 // exactly one where it isn't available. Insert a load on that edge and add
1468 // it to the AvailablePreds list.
1469 if (UnavailablePred) {
1470 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&((UnavailablePred->getTerminator()->getNumSuccessors() ==
1 && "Can't handle critical edge here!") ? static_cast
<void> (0) : __assert_fail ("UnavailablePred->getTerminator()->getNumSuccessors() == 1 && \"Can't handle critical edge here!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 1471, __PRETTY_FUNCTION__))
1471 "Can't handle critical edge here!")((UnavailablePred->getTerminator()->getNumSuccessors() ==
1 && "Can't handle critical edge here!") ? static_cast
<void> (0) : __assert_fail ("UnavailablePred->getTerminator()->getNumSuccessors() == 1 && \"Can't handle critical edge here!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 1471, __PRETTY_FUNCTION__))
;
1472 LoadInst *NewVal = new LoadInst(
1473 LoadI->getType(), LoadedPtr->DoPHITranslation(LoadBB, UnavailablePred),
1474 LoadI->getName() + ".pr", false, LoadI->getAlignment(),
1475 LoadI->getOrdering(), LoadI->getSyncScopeID(),
1476 UnavailablePred->getTerminator());
1477 NewVal->setDebugLoc(LoadI->getDebugLoc());
1478 if (AATags)
1479 NewVal->setAAMetadata(AATags);
1480
1481 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
1482 }
1483
1484 // Now we know that each predecessor of this block has a value in
1485 // AvailablePreds, sort them for efficient access as we're walking the preds.
1486 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1487
1488 // Create a PHI node at the start of the block for the PRE'd load value.
1489 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
1490 PHINode *PN = PHINode::Create(LoadI->getType(), std::distance(PB, PE), "",
1491 &LoadBB->front());
1492 PN->takeName(LoadI);
1493 PN->setDebugLoc(LoadI->getDebugLoc());
1494
1495 // Insert new entries into the PHI for each predecessor. A single block may
1496 // have multiple entries here.
1497 for (pred_iterator PI = PB; PI != PE; ++PI) {
1498 BasicBlock *P = *PI;
1499 AvailablePredsTy::iterator I =
1500 llvm::lower_bound(AvailablePreds, std::make_pair(P, (Value *)nullptr));
1501
1502 assert(I != AvailablePreds.end() && I->first == P &&((I != AvailablePreds.end() && I->first == P &&
"Didn't find entry for predecessor!") ? static_cast<void>
(0) : __assert_fail ("I != AvailablePreds.end() && I->first == P && \"Didn't find entry for predecessor!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 1503, __PRETTY_FUNCTION__))
1503 "Didn't find entry for predecessor!")((I != AvailablePreds.end() && I->first == P &&
"Didn't find entry for predecessor!") ? static_cast<void>
(0) : __assert_fail ("I != AvailablePreds.end() && I->first == P && \"Didn't find entry for predecessor!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 1503, __PRETTY_FUNCTION__))
;
1504
1505 // If we have an available predecessor but it requires casting, insert the
1506 // cast in the predecessor and use the cast. Note that we have to update the
1507 // AvailablePreds vector as we go so that all of the PHI entries for this
1508 // predecessor use the same bitcast.
1509 Value *&PredV = I->second;
1510 if (PredV->getType() != LoadI->getType())
1511 PredV = CastInst::CreateBitOrPointerCast(PredV, LoadI->getType(), "",
1512 P->getTerminator());
1513
1514 PN->addIncoming(PredV, I->first);
1515 }
1516
1517 for (LoadInst *PredLoadI : CSELoads) {
1518 combineMetadataForCSE(PredLoadI, LoadI, true);
1519 }
1520
1521 LoadI->replaceAllUsesWith(PN);
1522 LoadI->eraseFromParent();
1523
1524 return true;
1525}
1526
1527/// FindMostPopularDest - The specified list contains multiple possible
1528/// threadable destinations. Pick the one that occurs the most frequently in
1529/// the list.
1530static BasicBlock *
1531FindMostPopularDest(BasicBlock *BB,
1532 const SmallVectorImpl<std::pair<BasicBlock *,
1533 BasicBlock *>> &PredToDestList) {
1534 assert(!PredToDestList.empty())((!PredToDestList.empty()) ? static_cast<void> (0) : __assert_fail
("!PredToDestList.empty()", "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 1534, __PRETTY_FUNCTION__))
;
1535
1536 // Determine popularity. If there are multiple possible destinations, we
1537 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1538 // blocks with known and real destinations to threading undef. We'll handle
1539 // them later if interesting.
1540 DenseMap<BasicBlock*, unsigned> DestPopularity;
1541 for (const auto &PredToDest : PredToDestList)
1542 if (PredToDest.second)
1543 DestPopularity[PredToDest.second]++;
1544
1545 if (DestPopularity.empty())
1546 return nullptr;
1547
1548 // Find the most popular dest.
1549 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1550 BasicBlock *MostPopularDest = DPI->first;
1551 unsigned Popularity = DPI->second;
1552 SmallVector<BasicBlock*, 4> SamePopularity;
1553
1554 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1555 // If the popularity of this entry isn't higher than the popularity we've
1556 // seen so far, ignore it.
1557 if (DPI->second < Popularity)
1558 ; // ignore.
1559 else if (DPI->second == Popularity) {
1560 // If it is the same as what we've seen so far, keep track of it.
1561 SamePopularity.push_back(DPI->first);
1562 } else {
1563 // If it is more popular, remember it.
1564 SamePopularity.clear();
1565 MostPopularDest = DPI->first;
1566 Popularity = DPI->second;
1567 }
1568 }
1569
1570 // Okay, now we know the most popular destination. If there is more than one
1571 // destination, we need to determine one. This is arbitrary, but we need
1572 // to make a deterministic decision. Pick the first one that appears in the
1573 // successor list.
1574 if (!SamePopularity.empty()) {
1575 SamePopularity.push_back(MostPopularDest);
1576 Instruction *TI = BB->getTerminator();
1577 for (unsigned i = 0; ; ++i) {
1578 assert(i != TI->getNumSuccessors() && "Didn't find any successor!")((i != TI->getNumSuccessors() && "Didn't find any successor!"
) ? static_cast<void> (0) : __assert_fail ("i != TI->getNumSuccessors() && \"Didn't find any successor!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 1578, __PRETTY_FUNCTION__))
;
1579
1580 if (!is_contained(SamePopularity, TI->getSuccessor(i)))
1581 continue;
1582
1583 MostPopularDest = TI->getSuccessor(i);
1584 break;
1585 }
1586 }
1587
1588 // Okay, we have finally picked the most popular destination.
1589 return MostPopularDest;
1590}
1591
1592bool JumpThreadingPass::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1593 ConstantPreference Preference,
1594 Instruction *CxtI) {
1595 // If threading this would thread across a loop header, don't even try to
1596 // thread the edge.
1597 if (LoopHeaders.count(BB))
1598 return false;
1599
1600 PredValueInfoTy PredValues;
1601 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference, CxtI))
1602 return false;
1603
1604 assert(!PredValues.empty() &&((!PredValues.empty() && "ComputeValueKnownInPredecessors returned true with no values"
) ? static_cast<void> (0) : __assert_fail ("!PredValues.empty() && \"ComputeValueKnownInPredecessors returned true with no values\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 1605, __PRETTY_FUNCTION__))
1605 "ComputeValueKnownInPredecessors returned true with no values")((!PredValues.empty() && "ComputeValueKnownInPredecessors returned true with no values"
) ? static_cast<void> (0) : __assert_fail ("!PredValues.empty() && \"ComputeValueKnownInPredecessors returned true with no values\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 1605, __PRETTY_FUNCTION__))
;
1606
1607 LLVM_DEBUG(dbgs() << "IN BB: " << *BB;do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "IN BB: " << *BB;
for (const auto &PredValue : PredValues) { dbgs() <<
" BB '" << BB->getName() << "': FOUND condition = "
<< *PredValue.first << " for pred '" << PredValue
.second->getName() << "'.\n"; }; } } while (false)
1608 for (const auto &PredValue : PredValues) {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "IN BB: " << *BB;
for (const auto &PredValue : PredValues) { dbgs() <<
" BB '" << BB->getName() << "': FOUND condition = "
<< *PredValue.first << " for pred '" << PredValue
.second->getName() << "'.\n"; }; } } while (false)
1609 dbgs() << " BB '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "IN BB: " << *BB;
for (const auto &PredValue : PredValues) { dbgs() <<
" BB '" << BB->getName() << "': FOUND condition = "
<< *PredValue.first << " for pred '" << PredValue
.second->getName() << "'.\n"; }; } } while (false)
1610 << "': FOUND condition = " << *PredValue.firstdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "IN BB: " << *BB;
for (const auto &PredValue : PredValues) { dbgs() <<
" BB '" << BB->getName() << "': FOUND condition = "
<< *PredValue.first << " for pred '" << PredValue
.second->getName() << "'.\n"; }; } } while (false)
1611 << " for pred '" << PredValue.second->getName() << "'.\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "IN BB: " << *BB;
for (const auto &PredValue : PredValues) { dbgs() <<
" BB '" << BB->getName() << "': FOUND condition = "
<< *PredValue.first << " for pred '" << PredValue
.second->getName() << "'.\n"; }; } } while (false)
1612 })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "IN BB: " << *BB;
for (const auto &PredValue : PredValues) { dbgs() <<
" BB '" << BB->getName() << "': FOUND condition = "
<< *PredValue.first << " for pred '" << PredValue
.second->getName() << "'.\n"; }; } } while (false)
;
1613
1614 // Decide what we want to thread through. Convert our list of known values to
1615 // a list of known destinations for each pred. This also discards duplicate
1616 // predecessors and keeps track of the undefined inputs (which are represented
1617 // as a null dest in the PredToDestList).
1618 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1619 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1620
1621 BasicBlock *OnlyDest = nullptr;
1622 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1623 Constant *OnlyVal = nullptr;
1624 Constant *MultipleVal = (Constant *)(intptr_t)~0ULL;
1625
1626 for (const auto &PredValue : PredValues) {
1627 BasicBlock *Pred = PredValue.second;
1628 if (!SeenPreds.insert(Pred).second)
1629 continue; // Duplicate predecessor entry.
1630
1631 Constant *Val = PredValue.first;
1632
1633 BasicBlock *DestBB;
1634 if (isa<UndefValue>(Val))
1635 DestBB = nullptr;
1636 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
1637 assert(isa<ConstantInt>(Val) && "Expecting a constant integer")((isa<ConstantInt>(Val) && "Expecting a constant integer"
) ? static_cast<void> (0) : __assert_fail ("isa<ConstantInt>(Val) && \"Expecting a constant integer\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 1637, __PRETTY_FUNCTION__))
;
1638 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1639 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1640 assert(isa<ConstantInt>(Val) && "Expecting a constant integer")((isa<ConstantInt>(Val) && "Expecting a constant integer"
) ? static_cast<void> (0) : __assert_fail ("isa<ConstantInt>(Val) && \"Expecting a constant integer\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 1640, __PRETTY_FUNCTION__))
;
1641 DestBB = SI->findCaseValue(cast<ConstantInt>(Val))->getCaseSuccessor();
1642 } else {
1643 assert(isa<IndirectBrInst>(BB->getTerminator())((isa<IndirectBrInst>(BB->getTerminator()) &&
"Unexpected terminator") ? static_cast<void> (0) : __assert_fail
("isa<IndirectBrInst>(BB->getTerminator()) && \"Unexpected terminator\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 1644, __PRETTY_FUNCTION__))
1644 && "Unexpected terminator")((isa<IndirectBrInst>(BB->getTerminator()) &&
"Unexpected terminator") ? static_cast<void> (0) : __assert_fail
("isa<IndirectBrInst>(BB->getTerminator()) && \"Unexpected terminator\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 1644, __PRETTY_FUNCTION__))
;
1645 assert(isa<BlockAddress>(Val) && "Expecting a constant blockaddress")((isa<BlockAddress>(Val) && "Expecting a constant blockaddress"
) ? static_cast<void> (0) : __assert_fail ("isa<BlockAddress>(Val) && \"Expecting a constant blockaddress\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 1645, __PRETTY_FUNCTION__))
;
1646 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1647 }
1648
1649 // If we have exactly one destination, remember it for efficiency below.
1650 if (PredToDestList.empty()) {
1651 OnlyDest = DestBB;
1652 OnlyVal = Val;
1653 } else {
1654 if (OnlyDest != DestBB)
1655 OnlyDest = MultipleDestSentinel;
1656 // It possible we have same destination, but different value, e.g. default
1657 // case in switchinst.
1658 if (Val != OnlyVal)
1659 OnlyVal = MultipleVal;
1660 }
1661
1662 // If the predecessor ends with an indirect goto, we can't change its
1663 // destination. Same for CallBr.
1664 if (isa<IndirectBrInst>(Pred->getTerminator()) ||
1665 isa<CallBrInst>(Pred->getTerminator()))
1666 continue;
1667
1668 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1669 }
1670
1671 // If all edges were unthreadable, we fail.
1672 if (PredToDestList.empty())
1673 return false;
1674
1675 // If all the predecessors go to a single known successor, we want to fold,
1676 // not thread. By doing so, we do not need to duplicate the current block and
1677 // also miss potential opportunities in case we dont/cant duplicate.
1678 if (OnlyDest && OnlyDest != MultipleDestSentinel) {
1679 if (BB->hasNPredecessors(PredToDestList.size())) {
1680 bool SeenFirstBranchToOnlyDest = false;
1681 std::vector <DominatorTree::UpdateType> Updates;
1682 Updates.reserve(BB->getTerminator()->getNumSuccessors() - 1);
1683 for (BasicBlock *SuccBB : successors(BB)) {
1684 if (SuccBB == OnlyDest && !SeenFirstBranchToOnlyDest) {
1685 SeenFirstBranchToOnlyDest = true; // Don't modify the first branch.
1686 } else {
1687 SuccBB->removePredecessor(BB, true); // This is unreachable successor.
1688 Updates.push_back({DominatorTree::Delete, BB, SuccBB});
1689 }
1690 }
1691
1692 // Finally update the terminator.
1693 Instruction *Term = BB->getTerminator();
1694 BranchInst::Create(OnlyDest, Term);
1695 Term->eraseFromParent();
1696 DTU->applyUpdatesPermissive(Updates);
1697
1698 // If the condition is now dead due to the removal of the old terminator,
1699 // erase it.
1700 if (auto *CondInst = dyn_cast<Instruction>(Cond)) {
1701 if (CondInst->use_empty() && !CondInst->mayHaveSideEffects())
1702 CondInst->eraseFromParent();
1703 // We can safely replace *some* uses of the CondInst if it has
1704 // exactly one value as returned by LVI. RAUW is incorrect in the
1705 // presence of guards and assumes, that have the `Cond` as the use. This
1706 // is because we use the guards/assume to reason about the `Cond` value
1707 // at the end of block, but RAUW unconditionally replaces all uses
1708 // including the guards/assumes themselves and the uses before the
1709 // guard/assume.
1710 else if (OnlyVal && OnlyVal != MultipleVal &&
1711 CondInst->getParent() == BB)
1712 ReplaceFoldableUses(CondInst, OnlyVal);
1713 }
1714 return true;
1715 }
1716 }
1717
1718 // Determine which is the most common successor. If we have many inputs and
1719 // this block is a switch, we want to start by threading the batch that goes
1720 // to the most popular destination first. If we only know about one
1721 // threadable destination (the common case) we can avoid this.
1722 BasicBlock *MostPopularDest = OnlyDest;
1723
1724 if (MostPopularDest == MultipleDestSentinel) {
1725 // Remove any loop headers from the Dest list, ThreadEdge conservatively
1726 // won't process them, but we might have other destination that are eligible
1727 // and we still want to process.
1728 erase_if(PredToDestList,
1729 [&](const std::pair<BasicBlock *, BasicBlock *> &PredToDest) {
1730 return LoopHeaders.count(PredToDest.second) != 0;
1731 });
1732
1733 if (PredToDestList.empty())
1734 return false;
1735
1736 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1737 }
1738
1739 // Now that we know what the most popular destination is, factor all
1740 // predecessors that will jump to it into a single predecessor.
1741 SmallVector<BasicBlock*, 16> PredsToFactor;
1742 for (const auto &PredToDest : PredToDestList)
1743 if (PredToDest.second == MostPopularDest) {
1744 BasicBlock *Pred = PredToDest.first;
1745
1746 // This predecessor may be a switch or something else that has multiple
1747 // edges to the block. Factor each of these edges by listing them
1748 // according to # occurrences in PredsToFactor.
1749 for (BasicBlock *Succ : successors(Pred))
1750 if (Succ == BB)
1751 PredsToFactor.push_back(Pred);
1752 }
1753
1754 // If the threadable edges are branching on an undefined value, we get to pick
1755 // the destination that these predecessors should get to.
1756 if (!MostPopularDest)
1757 MostPopularDest = BB->getTerminator()->
1758 getSuccessor(GetBestDestForJumpOnUndef(BB));
1759
1760 // Ok, try to thread it!
1761 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1762}
1763
1764/// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1765/// a PHI node in the current block. See if there are any simplifications we
1766/// can do based on inputs to the phi node.
1767bool JumpThreadingPass::ProcessBranchOnPHI(PHINode *PN) {
1768 BasicBlock *BB = PN->getParent();
1769
1770 // TODO: We could make use of this to do it once for blocks with common PHI
1771 // values.
1772 SmallVector<BasicBlock*, 1> PredBBs;
1773 PredBBs.resize(1);
1774
1775 // If any of the predecessor blocks end in an unconditional branch, we can
1776 // *duplicate* the conditional branch into that block in order to further
1777 // encourage jump threading and to eliminate cases where we have branch on a
1778 // phi of an icmp (branch on icmp is much better).
1779 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1780 BasicBlock *PredBB = PN->getIncomingBlock(i);
1781 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1782 if (PredBr->isUnconditional()) {
1783 PredBBs[0] = PredBB;
1784 // Try to duplicate BB into PredBB.
1785 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1786 return true;
1787 }
1788 }
1789
1790 return false;
1791}
1792
1793/// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1794/// a xor instruction in the current block. See if there are any
1795/// simplifications we can do based on inputs to the xor.
1796bool JumpThreadingPass::ProcessBranchOnXOR(BinaryOperator *BO) {
1797 BasicBlock *BB = BO->getParent();
1798
1799 // If either the LHS or RHS of the xor is a constant, don't do this
1800 // optimization.
1801 if (isa<ConstantInt>(BO->getOperand(0)) ||
1802 isa<ConstantInt>(BO->getOperand(1)))
1803 return false;
1804
1805 // If the first instruction in BB isn't a phi, we won't be able to infer
1806 // anything special about any particular predecessor.
1807 if (!isa<PHINode>(BB->front()))
1808 return false;
1809
1810 // If this BB is a landing pad, we won't be able to split the edge into it.
1811 if (BB->isEHPad())
1812 return false;
1813
1814 // If we have a xor as the branch input to this block, and we know that the
1815 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1816 // the condition into the predecessor and fix that value to true, saving some
1817 // logical ops on that path and encouraging other paths to simplify.
1818 //
1819 // This copies something like this:
1820 //
1821 // BB:
1822 // %X = phi i1 [1], [%X']
1823 // %Y = icmp eq i32 %A, %B
1824 // %Z = xor i1 %X, %Y
1825 // br i1 %Z, ...
1826 //
1827 // Into:
1828 // BB':
1829 // %Y = icmp ne i32 %A, %B
1830 // br i1 %Y, ...
1831
1832 PredValueInfoTy XorOpValues;
1833 bool isLHS = true;
1834 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1835 WantInteger, BO)) {
1836 assert(XorOpValues.empty())((XorOpValues.empty()) ? static_cast<void> (0) : __assert_fail
("XorOpValues.empty()", "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 1836, __PRETTY_FUNCTION__))
;
1837 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1838 WantInteger, BO))
1839 return false;
1840 isLHS = false;
1841 }
1842
1843 assert(!XorOpValues.empty() &&((!XorOpValues.empty() && "ComputeValueKnownInPredecessors returned true with no values"
) ? static_cast<void> (0) : __assert_fail ("!XorOpValues.empty() && \"ComputeValueKnownInPredecessors returned true with no values\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 1844, __PRETTY_FUNCTION__))
1844 "ComputeValueKnownInPredecessors returned true with no values")((!XorOpValues.empty() && "ComputeValueKnownInPredecessors returned true with no values"
) ? static_cast<void> (0) : __assert_fail ("!XorOpValues.empty() && \"ComputeValueKnownInPredecessors returned true with no values\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 1844, __PRETTY_FUNCTION__))
;
1845
1846 // Scan the information to see which is most popular: true or false. The
1847 // predecessors can be of the set true, false, or undef.
1848 unsigned NumTrue = 0, NumFalse = 0;
1849 for (const auto &XorOpValue : XorOpValues) {
1850 if (isa<UndefValue>(XorOpValue.first))
1851 // Ignore undefs for the count.
1852 continue;
1853 if (cast<ConstantInt>(XorOpValue.first)->isZero())
1854 ++NumFalse;
1855 else
1856 ++NumTrue;
1857 }
1858
1859 // Determine which value to split on, true, false, or undef if neither.
1860 ConstantInt *SplitVal = nullptr;
1861 if (NumTrue > NumFalse)
1862 SplitVal = ConstantInt::getTrue(BB->getContext());
1863 else if (NumTrue != 0 || NumFalse != 0)
1864 SplitVal = ConstantInt::getFalse(BB->getContext());
1865
1866 // Collect all of the blocks that this can be folded into so that we can
1867 // factor this once and clone it once.
1868 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1869 for (const auto &XorOpValue : XorOpValues) {
1870 if (XorOpValue.first != SplitVal && !isa<UndefValue>(XorOpValue.first))
1871 continue;
1872
1873 BlocksToFoldInto.push_back(XorOpValue.second);
1874 }
1875
1876 // If we inferred a value for all of the predecessors, then duplication won't
1877 // help us. However, we can just replace the LHS or RHS with the constant.
1878 if (BlocksToFoldInto.size() ==
1879 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1880 if (!SplitVal) {
1881 // If all preds provide undef, just nuke the xor, because it is undef too.
1882 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1883 BO->eraseFromParent();
1884 } else if (SplitVal->isZero()) {
1885 // If all preds provide 0, replace the xor with the other input.
1886 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1887 BO->eraseFromParent();
1888 } else {
1889 // If all preds provide 1, set the computed value to 1.
1890 BO->setOperand(!isLHS, SplitVal);
1891 }
1892
1893 return true;
1894 }
1895
1896 // Try to duplicate BB into PredBB.
1897 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1898}
1899
1900/// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1901/// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1902/// NewPred using the entries from OldPred (suitably mapped).
1903static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1904 BasicBlock *OldPred,
1905 BasicBlock *NewPred,
1906 DenseMap<Instruction*, Value*> &ValueMap) {
1907 for (PHINode &PN : PHIBB->phis()) {
1908 // Ok, we have a PHI node. Figure out what the incoming value was for the
1909 // DestBlock.
1910 Value *IV = PN.getIncomingValueForBlock(OldPred);
1911
1912 // Remap the value if necessary.
1913 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1914 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1915 if (I != ValueMap.end())
1916 IV = I->second;
1917 }
1918
1919 PN.addIncoming(IV, NewPred);
1920 }
1921}
1922
1923/// ThreadEdge - We have decided that it is safe and profitable to factor the
1924/// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1925/// across BB. Transform the IR to reflect this change.
1926bool JumpThreadingPass::ThreadEdge(BasicBlock *BB,
1927 const SmallVectorImpl<BasicBlock *> &PredBBs,
1928 BasicBlock *SuccBB) {
1929 // If threading to the same block as we come from, we would infinite loop.
1930 if (SuccBB == BB) {
1931 LLVM_DEBUG(dbgs() << " Not threading across BB '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Not threading across BB '"
<< BB->getName() << "' - would thread to self!\n"
; } } while (false)
1932 << "' - would thread to self!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Not threading across BB '"
<< BB->getName() << "' - would thread to self!\n"
; } } while (false)
;
1933 return false;
1934 }
1935
1936 // If threading this would thread across a loop header, don't thread the edge.
1937 // See the comments above FindLoopHeaders for justifications and caveats.
1938 if (LoopHeaders.count(BB) || LoopHeaders.count(SuccBB)) {
1939 LLVM_DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB
); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() <<
" Not threading across " << (BBIsHeader ? "loop header BB '"
: "block BB '") << BB->getName() << "' to dest "
<< (SuccIsHeader ? "loop header BB '" : "block BB '") <<
SuccBB->getName() << "' - it might create an irreducible loop!\n"
; }; } } while (false)
1940 bool BBIsHeader = LoopHeaders.count(BB);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB
); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() <<
" Not threading across " << (BBIsHeader ? "loop header BB '"
: "block BB '") << BB->getName() << "' to dest "
<< (SuccIsHeader ? "loop header BB '" : "block BB '") <<
SuccBB->getName() << "' - it might create an irreducible loop!\n"
; }; } } while (false)
1941 bool SuccIsHeader = LoopHeaders.count(SuccBB);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB
); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() <<
" Not threading across " << (BBIsHeader ? "loop header BB '"
: "block BB '") << BB->getName() << "' to dest "
<< (SuccIsHeader ? "loop header BB '" : "block BB '") <<
SuccBB->getName() << "' - it might create an irreducible loop!\n"
; }; } } while (false)
1942 dbgs() << " Not threading across "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB
); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() <<
" Not threading across " << (BBIsHeader ? "loop header BB '"
: "block BB '") << BB->getName() << "' to dest "
<< (SuccIsHeader ? "loop header BB '" : "block BB '") <<
SuccBB->getName() << "' - it might create an irreducible loop!\n"
; }; } } while (false)
1943 << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB
); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() <<
" Not threading across " << (BBIsHeader ? "loop header BB '"
: "block BB '") << BB->getName() << "' to dest "
<< (SuccIsHeader ? "loop header BB '" : "block BB '") <<
SuccBB->getName() << "' - it might create an irreducible loop!\n"
; }; } } while (false)
1944 << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB
); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() <<
" Not threading across " << (BBIsHeader ? "loop header BB '"
: "block BB '") << BB->getName() << "' to dest "
<< (SuccIsHeader ? "loop header BB '" : "block BB '") <<
SuccBB->getName() << "' - it might create an irreducible loop!\n"
; }; } } while (false)
1945 << SuccBB->getName() << "' - it might create an irreducible loop!\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB
); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() <<
" Not threading across " << (BBIsHeader ? "loop header BB '"
: "block BB '") << BB->getName() << "' to dest "
<< (SuccIsHeader ? "loop header BB '" : "block BB '") <<
SuccBB->getName() << "' - it might create an irreducible loop!\n"
; }; } } while (false)
1946 })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { { bool BBIsHeader = LoopHeaders.count(BB
); bool SuccIsHeader = LoopHeaders.count(SuccBB); dbgs() <<
" Not threading across " << (BBIsHeader ? "loop header BB '"
: "block BB '") << BB->getName() << "' to dest "
<< (SuccIsHeader ? "loop header BB '" : "block BB '") <<
SuccBB->getName() << "' - it might create an irreducible loop!\n"
; }; } } while (false)
;
1947 return false;
1948 }
1949
1950 unsigned JumpThreadCost =
1951 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
1952 if (JumpThreadCost > BBDupThreshold) {
1953 LLVM_DEBUG(dbgs() << " Not threading BB '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Not threading BB '" <<
BB->getName() << "' - Cost is too high: " << JumpThreadCost
<< "\n"; } } while (false)
1954 << "' - Cost is too high: " << JumpThreadCost << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Not threading BB '" <<
BB->getName() << "' - Cost is too high: " << JumpThreadCost
<< "\n"; } } while (false)
;
1955 return false;
1956 }
1957
1958 // And finally, do it! Start by factoring the predecessors if needed.
1959 BasicBlock *PredBB;
1960 if (PredBBs.size() == 1)
1961 PredBB = PredBBs[0];
1962 else {
1963 LLVM_DEBUG(dbgs() << " Factoring out " << PredBBs.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Factoring out " <<
PredBBs.size() << " common predecessors.\n"; } } while
(false)
1964 << " common predecessors.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Factoring out " <<
PredBBs.size() << " common predecessors.\n"; } } while
(false)
;
1965 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
1966 }
1967
1968 // And finally, do it!
1969 LLVM_DEBUG(dbgs() << " Threading edge from '" << PredBB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Threading edge from '"
<< PredBB->getName() << "' to '" << SuccBB
->getName() << "' with cost: " << JumpThreadCost
<< ", across block:\n " << *BB << "\n";
} } while (false)
1970 << "' to '" << SuccBB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Threading edge from '"
<< PredBB->getName() << "' to '" << SuccBB
->getName() << "' with cost: " << JumpThreadCost
<< ", across block:\n " << *BB << "\n";
} } while (false)
1971 << "' with cost: " << JumpThreadCostdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Threading edge from '"
<< PredBB->getName() << "' to '" << SuccBB
->getName() << "' with cost: " << JumpThreadCost
<< ", across block:\n " << *BB << "\n";
} } while (false)
1972 << ", across block:\n " << *BB << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Threading edge from '"
<< PredBB->getName() << "' to '" << SuccBB
->getName() << "' with cost: " << JumpThreadCost
<< ", across block:\n " << *BB << "\n";
} } while (false)
;
1973
1974 if (DTU->hasPendingDomTreeUpdates())
1975 LVI->disableDT();
1976 else
1977 LVI->enableDT();
1978 LVI->threadEdge(PredBB, BB, SuccBB);
1979
1980 // We are going to have to map operands from the original BB block to the new
1981 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1982 // account for entry from PredBB.
1983 DenseMap<Instruction*, Value*> ValueMapping;
1984
1985 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1986 BB->getName()+".thread",
1987 BB->getParent(), BB);
1988 NewBB->moveAfter(PredBB);
1989
1990 // Set the block frequency of NewBB.
1991 if (HasProfileData) {
1992 auto NewBBFreq =
1993 BFI->getBlockFreq(PredBB) * BPI->getEdgeProbability(PredBB, BB);
1994 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
1995 }
1996
1997 BasicBlock::iterator BI = BB->begin();
1998 // Clone the phi nodes of BB into NewBB. The resulting phi nodes are trivial,
1999 // since NewBB only has one predecessor, but SSAUpdater might need to rewrite
2000 // the operand of the cloned phi.
2001 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) {
2002 PHINode *NewPN = PHINode::Create(PN->getType(), 1, PN->getName(), NewBB);
2003 NewPN->addIncoming(PN->getIncomingValueForBlock(PredBB), PredBB);
2004 ValueMapping[PN] = NewPN;
2005 }
2006
2007 // Clone the non-phi instructions of BB into NewBB, keeping track of the
2008 // mapping and using it to remap operands in the cloned instructions.
2009 for (; !BI->isTerminator(); ++BI) {
2010 Instruction *New = BI->clone();
2011 New->setName(BI->getName());
2012 NewBB->getInstList().push_back(New);
2013 ValueMapping[&*BI] = New;
2014
2015 // Remap operands to patch up intra-block references.
2016 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2017 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
2018 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
2019 if (I != ValueMapping.end())
2020 New->setOperand(i, I->second);
2021 }
2022 }
2023
2024 // We didn't copy the terminator from BB over to NewBB, because there is now
2025 // an unconditional jump to SuccBB. Insert the unconditional jump.
2026 BranchInst *NewBI = BranchInst::Create(SuccBB, NewBB);
2027 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
2028
2029 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
2030 // PHI nodes for NewBB now.
2031 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
2032
2033 // Update the terminator of PredBB to jump to NewBB instead of BB. This
2034 // eliminates predecessors from BB, which requires us to simplify any PHI
2035 // nodes in BB.
2036 Instruction *PredTerm = PredBB->getTerminator();
2037 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
2038 if (PredTerm->getSuccessor(i) == BB) {
2039 BB->removePredecessor(PredBB, true);
2040 PredTerm->setSuccessor(i, NewBB);
2041 }
2042
2043 // Enqueue required DT updates.
2044 DTU->applyUpdatesPermissive({{DominatorTree::Insert, NewBB, SuccBB},
2045 {DominatorTree::Insert, PredBB, NewBB},
2046 {DominatorTree::Delete, PredBB, BB}});
2047
2048 // If there were values defined in BB that are used outside the block, then we
2049 // now have to update all uses of the value to use either the original value,
2050 // the cloned value, or some PHI derived value. This can require arbitrary
2051 // PHI insertion, of which we are prepared to do, clean these up now.
2052 SSAUpdater SSAUpdate;
2053 SmallVector<Use*, 16> UsesToRename;
2054
2055 for (Instruction &I : *BB) {
2056 // Scan all uses of this instruction to see if their uses are no longer
2057 // dominated by the previous def and if so, record them in UsesToRename.
2058 // Also, skip phi operands from PredBB - we'll remove them anyway.
2059 for (Use &U : I.uses()) {
2060 Instruction *User = cast<Instruction>(U.getUser());
2061 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
2062 if (UserPN->getIncomingBlock(U) == BB)
2063 continue;
2064 } else if (User->getParent() == BB)
2065 continue;
2066
2067 UsesToRename.push_back(&U);
2068 }
2069
2070 // If there are no uses outside the block, we're done with this instruction.
2071 if (UsesToRename.empty())
2072 continue;
2073 LLVM_DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "JT: Renaming non-local uses of: "
<< I << "\n"; } } while (false)
;
2074
2075 // We found a use of I outside of BB. Rename all uses of I that are outside
2076 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
2077 // with the two values we know.
2078 SSAUpdate.Initialize(I.getType(), I.getName());
2079 SSAUpdate.AddAvailableValue(BB, &I);
2080 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[&I]);
2081
2082 while (!UsesToRename.empty())
2083 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
2084 LLVM_DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "\n"; } } while (false)
;
2085 }
2086
2087 // At this point, the IR is fully up to date and consistent. Do a quick scan
2088 // over the new instructions and zap any that are constants or dead. This
2089 // frequently happens because of phi translation.
2090 SimplifyInstructionsInBlock(NewBB, TLI);
2091
2092 // Update the edge weight from BB to SuccBB, which should be less than before.
2093 UpdateBlockFreqAndEdgeWeight(PredBB, BB, NewBB, SuccBB);
2094
2095 // Threaded an edge!
2096 ++NumThreads;
2097 return true;
2098}
2099
2100/// Create a new basic block that will be the predecessor of BB and successor of
2101/// all blocks in Preds. When profile data is available, update the frequency of
2102/// this new block.
2103BasicBlock *JumpThreadingPass::SplitBlockPreds(BasicBlock *BB,
2104 ArrayRef<BasicBlock *> Preds,
2105 const char *Suffix) {
2106 SmallVector<BasicBlock *, 2> NewBBs;
2107
2108 // Collect the frequencies of all predecessors of BB, which will be used to
2109 // update the edge weight of the result of splitting predecessors.
2110 DenseMap<BasicBlock *, BlockFrequency> FreqMap;
2111 if (HasProfileData)
2112 for (auto Pred : Preds)
2113 FreqMap.insert(std::make_pair(
2114 Pred, BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, BB)));
2115
2116 // In the case when BB is a LandingPad block we create 2 new predecessors
2117 // instead of just one.
2118 if (BB->isLandingPad()) {
2119 std::string NewName = std::string(Suffix) + ".split-lp";
2120 SplitLandingPadPredecessors(BB, Preds, Suffix, NewName.c_str(), NewBBs);
2121 } else {
2122 NewBBs.push_back(SplitBlockPredecessors(BB, Preds, Suffix));
2123 }
2124
2125 std::vector<DominatorTree::UpdateType> Updates;
2126 Updates.reserve((2 * Preds.size()) + NewBBs.size());
2127 for (auto NewBB : NewBBs) {
2128 BlockFrequency NewBBFreq(0);
2129 Updates.push_back({DominatorTree::Insert, NewBB, BB});
2130 for (auto Pred : predecessors(NewBB)) {
2131 Updates.push_back({DominatorTree::Delete, Pred, BB});
2132 Updates.push_back({DominatorTree::Insert, Pred, NewBB});
2133 if (HasProfileData) // Update frequencies between Pred -> NewBB.
2134 NewBBFreq += FreqMap.lookup(Pred);
2135 }
2136 if (HasProfileData) // Apply the summed frequency to NewBB.
2137 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
2138 }
2139
2140 DTU->applyUpdatesPermissive(Updates);
2141 return NewBBs[0];
2142}
2143
2144bool JumpThreadingPass::doesBlockHaveProfileData(BasicBlock *BB) {
2145 const Instruction *TI = BB->getTerminator();
2146 assert(TI->getNumSuccessors() > 1 && "not a split")((TI->getNumSuccessors() > 1 && "not a split") ?
static_cast<void> (0) : __assert_fail ("TI->getNumSuccessors() > 1 && \"not a split\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 2146, __PRETTY_FUNCTION__))
;
2147
2148 MDNode *WeightsNode = TI->getMetadata(LLVMContext::MD_prof);
2149 if (!WeightsNode)
2150 return false;
2151
2152 MDString *MDName = cast<MDString>(WeightsNode->getOperand(0));
2153 if (MDName->getString() != "branch_weights")
2154 return false;
2155
2156 // Ensure there are weights for all of the successors. Note that the first
2157 // operand to the metadata node is a name, not a weight.
2158 return WeightsNode->getNumOperands() == TI->getNumSuccessors() + 1;
2159}
2160
2161/// Update the block frequency of BB and branch weight and the metadata on the
2162/// edge BB->SuccBB. This is done by scaling the weight of BB->SuccBB by 1 -
2163/// Freq(PredBB->BB) / Freq(BB->SuccBB).
2164void JumpThreadingPass::UpdateBlockFreqAndEdgeWeight(BasicBlock *PredBB,
2165 BasicBlock *BB,
2166 BasicBlock *NewBB,
2167 BasicBlock *SuccBB) {
2168 if (!HasProfileData)
2169 return;
2170
2171 assert(BFI && BPI && "BFI & BPI should have been created here")((BFI && BPI && "BFI & BPI should have been created here"
) ? static_cast<void> (0) : __assert_fail ("BFI && BPI && \"BFI & BPI should have been created here\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 2171, __PRETTY_FUNCTION__))
;
2172
2173 // As the edge from PredBB to BB is deleted, we have to update the block
2174 // frequency of BB.
2175 auto BBOrigFreq = BFI->getBlockFreq(BB);
2176 auto NewBBFreq = BFI->getBlockFreq(NewBB);
2177 auto BB2SuccBBFreq = BBOrigFreq * BPI->getEdgeProbability(BB, SuccBB);
2178 auto BBNewFreq = BBOrigFreq - NewBBFreq;
2179 BFI->setBlockFreq(BB, BBNewFreq.getFrequency());
2180
2181 // Collect updated outgoing edges' frequencies from BB and use them to update
2182 // edge probabilities.
2183 SmallVector<uint64_t, 4> BBSuccFreq;
2184 for (BasicBlock *Succ : successors(BB)) {
2185 auto SuccFreq = (Succ == SuccBB)
2186 ? BB2SuccBBFreq - NewBBFreq
2187 : BBOrigFreq * BPI->getEdgeProbability(BB, Succ);
2188 BBSuccFreq.push_back(SuccFreq.getFrequency());
2189 }
2190
2191 uint64_t MaxBBSuccFreq =
2192 *std::max_element(BBSuccFreq.begin(), BBSuccFreq.end());
2193
2194 SmallVector<BranchProbability, 4> BBSuccProbs;
2195 if (MaxBBSuccFreq == 0)
2196 BBSuccProbs.assign(BBSuccFreq.size(),
2197 {1, static_cast<unsigned>(BBSuccFreq.size())});
2198 else {
2199 for (uint64_t Freq : BBSuccFreq)
2200 BBSuccProbs.push_back(
2201 BranchProbability::getBranchProbability(Freq, MaxBBSuccFreq));
2202 // Normalize edge probabilities so that they sum up to one.
2203 BranchProbability::normalizeProbabilities(BBSuccProbs.begin(),
2204 BBSuccProbs.end());
2205 }
2206
2207 // Update edge probabilities in BPI.
2208 for (int I = 0, E = BBSuccProbs.size(); I < E; I++)
2209 BPI->setEdgeProbability(BB, I, BBSuccProbs[I]);
2210
2211 // Update the profile metadata as well.
2212 //
2213 // Don't do this if the profile of the transformed blocks was statically
2214 // estimated. (This could occur despite the function having an entry
2215 // frequency in completely cold parts of the CFG.)
2216 //
2217 // In this case we don't want to suggest to subsequent passes that the
2218 // calculated weights are fully consistent. Consider this graph:
2219 //
2220 // check_1
2221 // 50% / |
2222 // eq_1 | 50%
2223 // \ |
2224 // check_2
2225 // 50% / |
2226 // eq_2 | 50%
2227 // \ |
2228 // check_3
2229 // 50% / |
2230 // eq_3 | 50%
2231 // \ |
2232 //
2233 // Assuming the blocks check_* all compare the same value against 1, 2 and 3,
2234 // the overall probabilities are inconsistent; the total probability that the
2235 // value is either 1, 2 or 3 is 150%.
2236 //
2237 // As a consequence if we thread eq_1 -> check_2 to check_3, check_2->check_3
2238 // becomes 0%. This is even worse if the edge whose probability becomes 0% is
2239 // the loop exit edge. Then based solely on static estimation we would assume
2240 // the loop was extremely hot.
2241 //
2242 // FIXME this locally as well so that BPI and BFI are consistent as well. We
2243 // shouldn't make edges extremely likely or unlikely based solely on static
2244 // estimation.
2245 if (BBSuccProbs.size() >= 2 && doesBlockHaveProfileData(BB)) {
2246 SmallVector<uint32_t, 4> Weights;
2247 for (auto Prob : BBSuccProbs)
2248 Weights.push_back(Prob.getNumerator());
2249
2250 auto TI = BB->getTerminator();
2251 TI->setMetadata(
2252 LLVMContext::MD_prof,
2253 MDBuilder(TI->getParent()->getContext()).createBranchWeights(Weights));
2254 }
2255}
2256
2257/// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
2258/// to BB which contains an i1 PHI node and a conditional branch on that PHI.
2259/// If we can duplicate the contents of BB up into PredBB do so now, this
2260/// improves the odds that the branch will be on an analyzable instruction like
2261/// a compare.
2262bool JumpThreadingPass::DuplicateCondBranchOnPHIIntoPred(
2263 BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs) {
2264 assert(!PredBBs.empty() && "Can't handle an empty set")((!PredBBs.empty() && "Can't handle an empty set") ? static_cast
<void> (0) : __assert_fail ("!PredBBs.empty() && \"Can't handle an empty set\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 2264, __PRETTY_FUNCTION__))
;
2265
2266 // If BB is a loop header, then duplicating this block outside the loop would
2267 // cause us to transform this into an irreducible loop, don't do this.
2268 // See the comments above FindLoopHeaders for justifications and caveats.
2269 if (LoopHeaders.count(BB)) {
2270 LLVM_DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Not duplicating loop header '"
<< BB->getName() << "' into predecessor block '"
<< PredBBs[0]->getName() << "' - it might create an irreducible loop!\n"
; } } while (false)
2271 << "' into predecessor block '" << PredBBs[0]->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Not duplicating loop header '"
<< BB->getName() << "' into predecessor block '"
<< PredBBs[0]->getName() << "' - it might create an irreducible loop!\n"
; } } while (false)
2272 << "' - it might create an irreducible loop!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Not duplicating loop header '"
<< BB->getName() << "' into predecessor block '"
<< PredBBs[0]->getName() << "' - it might create an irreducible loop!\n"
; } } while (false)
;
2273 return false;
2274 }
2275
2276 unsigned DuplicationCost =
2277 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
2278 if (DuplicationCost > BBDupThreshold) {
2279 LLVM_DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Not duplicating BB '"
<< BB->getName() << "' - Cost is too high: " <<
DuplicationCost << "\n"; } } while (false)
2280 << "' - Cost is too high: " << DuplicationCost << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Not duplicating BB '"
<< BB->getName() << "' - Cost is too high: " <<
DuplicationCost << "\n"; } } while (false)
;
2281 return false;
2282 }
2283
2284 // And finally, do it! Start by factoring the predecessors if needed.
2285 std::vector<DominatorTree::UpdateType> Updates;
2286 BasicBlock *PredBB;
2287 if (PredBBs.size() == 1)
2288 PredBB = PredBBs[0];
2289 else {
2290 LLVM_DEBUG(dbgs() << " Factoring out " << PredBBs.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Factoring out " <<
PredBBs.size() << " common predecessors.\n"; } } while
(false)
2291 << " common predecessors.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Factoring out " <<
PredBBs.size() << " common predecessors.\n"; } } while
(false)
;
2292 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
2293 }
2294 Updates.push_back({DominatorTree::Delete, PredBB, BB});
2295
2296 // Okay, we decided to do this! Clone all the instructions in BB onto the end
2297 // of PredBB.
2298 LLVM_DEBUG(dbgs() << " Duplicating block '" << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Duplicating block '"
<< BB->getName() << "' into end of '" <<
PredBB->getName() << "' to eliminate branch on phi. Cost: "
<< DuplicationCost << " block is:" << *BB <<
"\n"; } } while (false)
2299 << "' into end of '" << PredBB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Duplicating block '"
<< BB->getName() << "' into end of '" <<
PredBB->getName() << "' to eliminate branch on phi. Cost: "
<< DuplicationCost << " block is:" << *BB <<
"\n"; } } while (false)
2300 << "' to eliminate branch on phi. Cost: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Duplicating block '"
<< BB->getName() << "' into end of '" <<
PredBB->getName() << "' to eliminate branch on phi. Cost: "
<< DuplicationCost << " block is:" << *BB <<
"\n"; } } while (false)
2301 << DuplicationCost << " block is:" << *BB << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << " Duplicating block '"
<< BB->getName() << "' into end of '" <<
PredBB->getName() << "' to eliminate branch on phi. Cost: "
<< DuplicationCost << " block is:" << *BB <<
"\n"; } } while (false)
;
2302
2303 // Unless PredBB ends with an unconditional branch, split the edge so that we
2304 // can just clone the bits from BB into the end of the new PredBB.
2305 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
2306
2307 if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
2308 BasicBlock *OldPredBB = PredBB;
2309 PredBB = SplitEdge(OldPredBB, BB);
2310 Updates.push_back({DominatorTree::Insert, OldPredBB, PredBB});
2311 Updates.push_back({DominatorTree::Insert, PredBB, BB});
2312 Updates.push_back({DominatorTree::Delete, OldPredBB, BB});
2313 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
2314 }
2315
2316 // We are going to have to map operands from the original BB block into the
2317 // PredBB block. Evaluate PHI nodes in BB.
2318 DenseMap<Instruction*, Value*> ValueMapping;
2319
2320 BasicBlock::iterator BI = BB->begin();
2321 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
2322 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
2323 // Clone the non-phi instructions of BB into PredBB, keeping track of the
2324 // mapping and using it to remap operands in the cloned instructions.
2325 for (; BI != BB->end(); ++BI) {
2326 Instruction *New = BI->clone();
2327
2328 // Remap operands to patch up intra-block references.
2329 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2330 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
2331 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
2332 if (I != ValueMapping.end())
2333 New->setOperand(i, I->second);
2334 }
2335
2336 // If this instruction can be simplified after the operands are updated,
2337 // just use the simplified value instead. This frequently happens due to
2338 // phi translation.
2339 if (Value *IV = SimplifyInstruction(
2340 New,
2341 {BB->getModule()->getDataLayout(), TLI, nullptr, nullptr, New})) {
2342 ValueMapping[&*BI] = IV;
2343 if (!New->mayHaveSideEffects()) {
2344 New->deleteValue();
2345 New = nullptr;
2346 }
2347 } else {
2348 ValueMapping[&*BI] = New;
2349 }
2350 if (New) {
2351 // Otherwise, insert the new instruction into the block.
2352 New->setName(BI->getName());
2353 PredBB->getInstList().insert(OldPredBranch->getIterator(), New);
2354 // Update Dominance from simplified New instruction operands.
2355 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2356 if (BasicBlock *SuccBB = dyn_cast<BasicBlock>(New->getOperand(i)))
2357 Updates.push_back({DominatorTree::Insert, PredBB, SuccBB});
2358 }
2359 }
2360
2361 // Check to see if the targets of the branch had PHI nodes. If so, we need to
2362 // add entries to the PHI nodes for branch from PredBB now.
2363 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
2364 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
2365 ValueMapping);
2366 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
2367 ValueMapping);
2368
2369 // If there were values defined in BB that are used outside the block, then we
2370 // now have to update all uses of the value to use either the original value,
2371 // the cloned value, or some PHI derived value. This can require arbitrary
2372 // PHI insertion, of which we are prepared to do, clean these up now.
2373 SSAUpdater SSAUpdate;
2374 SmallVector<Use*, 16> UsesToRename;
2375 for (Instruction &I : *BB) {
2376 // Scan all uses of this instruction to see if it is used outside of its
2377 // block, and if so, record them in UsesToRename.
2378 for (Use &U : I.uses()) {
2379 Instruction *User = cast<Instruction>(U.getUser());
2380 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
2381 if (UserPN->getIncomingBlock(U) == BB)
2382 continue;
2383 } else if (User->getParent() == BB)
2384 continue;
2385
2386 UsesToRename.push_back(&U);
2387 }
2388
2389 // If there are no uses outside the block, we're done with this instruction.
2390 if (UsesToRename.empty())
2391 continue;
2392
2393 LLVM_DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "JT: Renaming non-local uses of: "
<< I << "\n"; } } while (false)
;
2394
2395 // We found a use of I outside of BB. Rename all uses of I that are outside
2396 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
2397 // with the two values we know.
2398 SSAUpdate.Initialize(I.getType(), I.getName());
2399 SSAUpdate.AddAvailableValue(BB, &I);
2400 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[&I]);
2401
2402 while (!UsesToRename.empty())
2403 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
2404 LLVM_DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "\n"; } } while (false)
;
2405 }
2406
2407 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
2408 // that we nuked.
2409 BB->removePredecessor(PredBB, true);
2410
2411 // Remove the unconditional branch at the end of the PredBB block.
2412 OldPredBranch->eraseFromParent();
2413 DTU->applyUpdatesPermissive(Updates);
2414
2415 ++NumDupes;
2416 return true;
2417}
2418
2419// Pred is a predecessor of BB with an unconditional branch to BB. SI is
2420// a Select instruction in Pred. BB has other predecessors and SI is used in
2421// a PHI node in BB. SI has no other use.
2422// A new basic block, NewBB, is created and SI is converted to compare and
2423// conditional branch. SI is erased from parent.
2424void JumpThreadingPass::UnfoldSelectInstr(BasicBlock *Pred, BasicBlock *BB,
2425 SelectInst *SI, PHINode *SIUse,
2426 unsigned Idx) {
2427 // Expand the select.
2428 //
2429 // Pred --
2430 // | v
2431 // | NewBB
2432 // | |
2433 // |-----
2434 // v
2435 // BB
2436 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
2437 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
2438 BB->getParent(), BB);
2439 // Move the unconditional branch to NewBB.
2440 PredTerm->removeFromParent();
2441 NewBB->getInstList().insert(NewBB->end(), PredTerm);
2442 // Create a conditional branch and update PHI nodes.
2443 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
2444 SIUse->setIncomingValue(Idx, SI->getFalseValue());
2445 SIUse->addIncoming(SI->getTrueValue(), NewBB);
2446
2447 // The select is now dead.
2448 SI->eraseFromParent();
2449 DTU->applyUpdatesPermissive({{DominatorTree::Insert, NewBB, BB},
2450 {DominatorTree::Insert, Pred, NewBB}});
2451
2452 // Update any other PHI nodes in BB.
2453 for (BasicBlock::iterator BI = BB->begin();
2454 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
2455 if (Phi != SIUse)
2456 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);
2457}
2458
2459bool JumpThreadingPass::TryToUnfoldSelect(SwitchInst *SI, BasicBlock *BB) {
2460 PHINode *CondPHI = dyn_cast<PHINode>(SI->getCondition());
2461
2462 if (!CondPHI || CondPHI->getParent() != BB)
2463 return false;
2464
2465 for (unsigned I = 0, E = CondPHI->getNumIncomingValues(); I != E; ++I) {
2466 BasicBlock *Pred = CondPHI->getIncomingBlock(I);
2467 SelectInst *PredSI = dyn_cast<SelectInst>(CondPHI->getIncomingValue(I));
2468
2469 // The second and third condition can be potentially relaxed. Currently
2470 // the conditions help to simplify the code and allow us to reuse existing
2471 // code, developed for TryToUnfoldSelect(CmpInst *, BasicBlock *)
2472 if (!PredSI || PredSI->getParent() != Pred || !PredSI->hasOneUse())
2473 continue;
2474
2475 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
2476 if (!PredTerm || !PredTerm->isUnconditional())
2477 continue;
2478
2479 UnfoldSelectInstr(Pred, BB, PredSI, CondPHI, I);
2480 return true;
2481 }
2482 return false;
2483}
2484
2485/// TryToUnfoldSelect - Look for blocks of the form
2486/// bb1:
2487/// %a = select
2488/// br bb2
2489///
2490/// bb2:
2491/// %p = phi [%a, %bb1] ...
2492/// %c = icmp %p
2493/// br i1 %c
2494///
2495/// And expand the select into a branch structure if one of its arms allows %c
2496/// to be folded. This later enables threading from bb1 over bb2.
2497bool JumpThreadingPass::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
2498 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
2499 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
2500 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
2501
2502 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
2503 CondLHS->getParent() != BB)
2504 return false;
2505
2506 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
2507 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
2508 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
2509
2510 // Look if one of the incoming values is a select in the corresponding
2511 // predecessor.
2512 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
2513 continue;
2514
2515 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
2516 if (!PredTerm || !PredTerm->isUnconditional())
2517 continue;
2518
2519 // Now check if one of the select values would allow us to constant fold the
2520 // terminator in BB. We don't do the transform if both sides fold, those
2521 // cases will be threaded in any case.
2522 if (DTU->hasPendingDomTreeUpdates())
2523 LVI->disableDT();
2524 else
2525 LVI->enableDT();
2526 LazyValueInfo::Tristate LHSFolds =
2527 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
2528 CondRHS, Pred, BB, CondCmp);
2529 LazyValueInfo::Tristate RHSFolds =
2530 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
2531 CondRHS, Pred, BB, CondCmp);
2532 if ((LHSFolds != LazyValueInfo::Unknown ||
2533 RHSFolds != LazyValueInfo::Unknown) &&
2534 LHSFolds != RHSFolds) {
2535 UnfoldSelectInstr(Pred, BB, SI, CondLHS, I);
2536 return true;
2537 }
2538 }
2539 return false;
2540}
2541
2542/// TryToUnfoldSelectInCurrBB - Look for PHI/Select or PHI/CMP/Select in the
2543/// same BB in the form
2544/// bb:
2545/// %p = phi [false, %bb1], [true, %bb2], [false, %bb3], [true, %bb4], ...
2546/// %s = select %p, trueval, falseval
2547///
2548/// or
2549///
2550/// bb:
2551/// %p = phi [0, %bb1], [1, %bb2], [0, %bb3], [1, %bb4], ...
2552/// %c = cmp %p, 0
2553/// %s = select %c, trueval, falseval
2554///
2555/// And expand the select into a branch structure. This later enables
2556/// jump-threading over bb in this pass.
2557///
2558/// Using the similar approach of SimplifyCFG::FoldCondBranchOnPHI(), unfold
2559/// select if the associated PHI has at least one constant. If the unfolded
2560/// select is not jump-threaded, it will be folded again in the later
2561/// optimizations.
2562bool JumpThreadingPass::TryToUnfoldSelectInCurrBB(BasicBlock *BB) {
2563 // If threading this would thread across a loop header, don't thread the edge.
2564 // See the comments above FindLoopHeaders for justifications and caveats.
2565 if (LoopHeaders.count(BB))
2566 return false;
2567
2568 for (BasicBlock::iterator BI = BB->begin();
2569 PHINode *PN = dyn_cast<PHINode>(BI); ++BI) {
2570 // Look for a Phi having at least one constant incoming value.
2571 if (llvm::all_of(PN->incoming_values(),
2572 [](Value *V) { return !isa<ConstantInt>(V); }))
2573 continue;
2574
2575 auto isUnfoldCandidate = [BB](SelectInst *SI, Value *V) {
2576 // Check if SI is in BB and use V as condition.
2577 if (SI->getParent() != BB)
2578 return false;
2579 Value *Cond = SI->getCondition();
2580 return (Cond && Cond == V && Cond->getType()->isIntegerTy(1));
2581 };
2582
2583 SelectInst *SI = nullptr;
2584 for (Use &U : PN->uses()) {
2585 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(U.getUser())) {
2586 // Look for a ICmp in BB that compares PN with a constant and is the
2587 // condition of a Select.
2588 if (Cmp->getParent() == BB && Cmp->hasOneUse() &&
2589 isa<ConstantInt>(Cmp->getOperand(1 - U.getOperandNo())))
2590 if (SelectInst *SelectI = dyn_cast<SelectInst>(Cmp->user_back()))
2591 if (isUnfoldCandidate(SelectI, Cmp->use_begin()->get())) {
2592 SI = SelectI;
2593 break;
2594 }
2595 } else if (SelectInst *SelectI = dyn_cast<SelectInst>(U.getUser())) {
2596 // Look for a Select in BB that uses PN as condition.
2597 if (isUnfoldCandidate(SelectI, U.get())) {
2598 SI = SelectI;
2599 break;
2600 }
2601 }
2602 }
2603
2604 if (!SI)
2605 continue;
2606 // Expand the select.
2607 Instruction *Term =
2608 SplitBlockAndInsertIfThen(SI->getCondition(), SI, false);
2609 BasicBlock *SplitBB = SI->getParent();
2610 BasicBlock *NewBB = Term->getParent();
2611 PHINode *NewPN = PHINode::Create(SI->getType(), 2, "", SI);
2612 NewPN->addIncoming(SI->getTrueValue(), Term->getParent());
2613 NewPN->addIncoming(SI->getFalseValue(), BB);
2614 SI->replaceAllUsesWith(NewPN);
2615 SI->eraseFromParent();
2616 // NewBB and SplitBB are newly created blocks which require insertion.
2617 std::vector<DominatorTree::UpdateType> Updates;
2618 Updates.reserve((2 * SplitBB->getTerminator()->getNumSuccessors()) + 3);
2619 Updates.push_back({DominatorTree::Insert, BB, SplitBB});
2620 Updates.push_back({DominatorTree::Insert, BB, NewBB});
2621 Updates.push_back({DominatorTree::Insert, NewBB, SplitBB});
2622 // BB's successors were moved to SplitBB, update DTU accordingly.
2623 for (auto *Succ : successors(SplitBB)) {
2624 Updates.push_back({DominatorTree::Delete, BB, Succ});
2625 Updates.push_back({DominatorTree::Insert, SplitBB, Succ});
2626 }
2627 DTU->applyUpdatesPermissive(Updates);
2628 return true;
2629 }
2630 return false;
2631}
2632
2633/// Try to propagate a guard from the current BB into one of its predecessors
2634/// in case if another branch of execution implies that the condition of this
2635/// guard is always true. Currently we only process the simplest case that
2636/// looks like:
2637///
2638/// Start:
2639/// %cond = ...
2640/// br i1 %cond, label %T1, label %F1
2641/// T1:
2642/// br label %Merge
2643/// F1:
2644/// br label %Merge
2645/// Merge:
2646/// %condGuard = ...
2647/// call void(i1, ...) @llvm.experimental.guard( i1 %condGuard )[ "deopt"() ]
2648///
2649/// And cond either implies condGuard or !condGuard. In this case all the
2650/// instructions before the guard can be duplicated in both branches, and the
2651/// guard is then threaded to one of them.
2652bool JumpThreadingPass::ProcessGuards(BasicBlock *BB) {
2653 using namespace PatternMatch;
2654
2655 // We only want to deal with two predecessors.
2656 BasicBlock *Pred1, *Pred2;
2657 auto PI = pred_begin(BB), PE = pred_end(BB);
2658 if (PI == PE)
2659 return false;
2660 Pred1 = *PI++;
2661 if (PI == PE)
2662 return false;
2663 Pred2 = *PI++;
2664 if (PI != PE)
2665 return false;
2666 if (Pred1 == Pred2)
2667 return false;
2668
2669 // Try to thread one of the guards of the block.
2670 // TODO: Look up deeper than to immediate predecessor?
2671 auto *Parent = Pred1->getSinglePredecessor();
2672 if (!Parent || Parent != Pred2->getSinglePredecessor())
2673 return false;
2674
2675 if (auto *BI = dyn_cast<BranchInst>(Parent->getTerminator()))
2676 for (auto &I : *BB)
2677 if (isGuard(&I) && ThreadGuard(BB, cast<IntrinsicInst>(&I), BI))
2678 return true;
2679
2680 return false;
2681}
2682
2683/// Try to propagate the guard from BB which is the lower block of a diamond
2684/// to one of its branches, in case if diamond's condition implies guard's
2685/// condition.
2686bool JumpThreadingPass::ThreadGuard(BasicBlock *BB, IntrinsicInst *Guard,
2687 BranchInst *BI) {
2688 assert(BI->getNumSuccessors() == 2 && "Wrong number of successors?")((BI->getNumSuccessors() == 2 && "Wrong number of successors?"
) ? static_cast<void> (0) : __assert_fail ("BI->getNumSuccessors() == 2 && \"Wrong number of successors?\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 2688, __PRETTY_FUNCTION__))
;
2689 assert(BI->isConditional() && "Unconditional branch has 2 successors?")((BI->isConditional() && "Unconditional branch has 2 successors?"
) ? static_cast<void> (0) : __assert_fail ("BI->isConditional() && \"Unconditional branch has 2 successors?\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 2689, __PRETTY_FUNCTION__))
;
2690 Value *GuardCond = Guard->getArgOperand(0);
2691 Value *BranchCond = BI->getCondition();
2692 BasicBlock *TrueDest = BI->getSuccessor(0);
2693 BasicBlock *FalseDest = BI->getSuccessor(1);
2694
2695 auto &DL = BB->getModule()->getDataLayout();
2696 bool TrueDestIsSafe = false;
2697 bool FalseDestIsSafe = false;
2698
2699 // True dest is safe if BranchCond => GuardCond.
2700 auto Impl = isImpliedCondition(BranchCond, GuardCond, DL);
2701 if (Impl && *Impl)
2702 TrueDestIsSafe = true;
2703 else {
2704 // False dest is safe if !BranchCond => GuardCond.
2705 Impl = isImpliedCondition(BranchCond, GuardCond, DL, /* LHSIsTrue */ false);
2706 if (Impl && *Impl)
2707 FalseDestIsSafe = true;
2708 }
2709
2710 if (!TrueDestIsSafe && !FalseDestIsSafe)
2711 return false;
2712
2713 BasicBlock *PredUnguardedBlock = TrueDestIsSafe ? TrueDest : FalseDest;
2714 BasicBlock *PredGuardedBlock = FalseDestIsSafe ? TrueDest : FalseDest;
2715
2716 ValueToValueMapTy UnguardedMapping, GuardedMapping;
2717 Instruction *AfterGuard = Guard->getNextNode();
2718 unsigned Cost = getJumpThreadDuplicationCost(BB, AfterGuard, BBDupThreshold);
2719 if (Cost > BBDupThreshold)
2720 return false;
2721 // Duplicate all instructions before the guard and the guard itself to the
2722 // branch where implication is not proved.
2723 BasicBlock *GuardedBlock = DuplicateInstructionsInSplitBetween(
2724 BB, PredGuardedBlock, AfterGuard, GuardedMapping, *DTU);
2725 assert(GuardedBlock && "Could not create the guarded block?")((GuardedBlock && "Could not create the guarded block?"
) ? static_cast<void> (0) : __assert_fail ("GuardedBlock && \"Could not create the guarded block?\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 2725, __PRETTY_FUNCTION__))
;
2726 // Duplicate all instructions before the guard in the unguarded branch.
2727 // Since we have successfully duplicated the guarded block and this block
2728 // has fewer instructions, we expect it to succeed.
2729 BasicBlock *UnguardedBlock = DuplicateInstructionsInSplitBetween(
2730 BB, PredUnguardedBlock, Guard, UnguardedMapping, *DTU);
2731 assert(UnguardedBlock && "Could not create the unguarded block?")((UnguardedBlock && "Could not create the unguarded block?"
) ? static_cast<void> (0) : __assert_fail ("UnguardedBlock && \"Could not create the unguarded block?\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 2731, __PRETTY_FUNCTION__))
;
2732 LLVM_DEBUG(dbgs() << "Moved guard " << *Guard << " to block "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "Moved guard " <<
*Guard << " to block " << GuardedBlock->getName
() << "\n"; } } while (false)
2733 << GuardedBlock->getName() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("jump-threading")) { dbgs() << "Moved guard " <<
*Guard << " to block " << GuardedBlock->getName
() << "\n"; } } while (false)
;
2734 // Some instructions before the guard may still have uses. For them, we need
2735 // to create Phi nodes merging their copies in both guarded and unguarded
2736 // branches. Those instructions that have no uses can be just removed.
2737 SmallVector<Instruction *, 4> ToRemove;
2738 for (auto BI = BB->begin(); &*BI != AfterGuard; ++BI)
2739 if (!isa<PHINode>(&*BI))
2740 ToRemove.push_back(&*BI);
2741
2742 Instruction *InsertionPoint = &*BB->getFirstInsertionPt();
2743 assert(InsertionPoint && "Empty block?")((InsertionPoint && "Empty block?") ? static_cast<
void> (0) : __assert_fail ("InsertionPoint && \"Empty block?\""
, "/build/llvm-toolchain-snapshot-10~svn374877/lib/Transforms/Scalar/JumpThreading.cpp"
, 2743, __PRETTY_FUNCTION__))
;
2744 // Substitute with Phis & remove.
2745 for (auto *Inst : reverse(ToRemove)) {
2746 if (!Inst->use_empty()) {
2747 PHINode *NewPN = PHINode::Create(Inst->getType(), 2);
2748 NewPN->addIncoming(UnguardedMapping[Inst], UnguardedBlock);
2749 NewPN->addIncoming(GuardedMapping[Inst], GuardedBlock);
2750 NewPN->insertBefore(InsertionPoint);
2751 Inst->replaceAllUsesWith(NewPN);
2752 }
2753 Inst->eraseFromParent();
2754 }
2755 return true;
2756}

/build/llvm-toolchain-snapshot-10~svn374877/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/None.h"
20#include "llvm/ADT/STLExtras.h"
21#include "llvm/ADT/SmallVector.h"
22#include "llvm/ADT/StringRef.h"
23#include "llvm/ADT/Twine.h"
24#include "llvm/ADT/iterator.h"
25#include "llvm/ADT/iterator_range.h"
26#include "llvm/IR/Attributes.h"
27#include "llvm/IR/BasicBlock.h"
28#include "llvm/IR/CallingConv.h"
29#include "llvm/IR/Constant.h"
30#include "llvm/IR/DerivedTypes.h"
31#include "llvm/IR/Function.h"
32#include "llvm/IR/InstrTypes.h"
33#include "llvm/IR/Instruction.h"
34#include "llvm/IR/OperandTraits.h"
35#include "llvm/IR/Type.h"
36#include "llvm/IR/Use.h"
37#include "llvm/IR/User.h"
38#include "llvm/IR/Value.h"
39#include "llvm/Support/AtomicOrdering.h"
40#include "llvm/Support/Casting.h"
41#include "llvm/Support/ErrorHandling.h"
42#include <cassert>
43#include <cstddef>
44#include <cstdint>
45#include <iterator>
46
47namespace llvm {
48
49class APInt;
50class ConstantInt;
51class DataLayout;
52class LLVMContext;
53
54//===----------------------------------------------------------------------===//
55// AllocaInst Class
56//===----------------------------------------------------------------------===//
57
58/// an instruction to allocate memory on the stack
59class AllocaInst : public UnaryInstruction {
60 Type *AllocatedType;
61
62protected:
63 // Note: Instruction needs to be a friend here to call cloneImpl.
64 friend class Instruction;
65
66 AllocaInst *cloneImpl() const;
67
68public:
69 explicit AllocaInst(Type *Ty, unsigned AddrSpace,
70 Value *ArraySize = nullptr,
71 const Twine &Name = "",
72 Instruction *InsertBefore = nullptr);
73 AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize,
74 const Twine &Name, BasicBlock *InsertAtEnd);
75
76 AllocaInst(Type *Ty, unsigned AddrSpace,
77 const Twine &Name, Instruction *InsertBefore = nullptr);
78 AllocaInst(Type *Ty, unsigned AddrSpace,
79 const Twine &Name, BasicBlock *InsertAtEnd);
80
81 AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize, unsigned Align,
82 const Twine &Name = "", Instruction *InsertBefore = nullptr);
83 AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize, unsigned Align,
84 const Twine &Name, BasicBlock *InsertAtEnd);
85
86 /// Return true if there is an allocation size parameter to the allocation
87 /// instruction that is not 1.
88 bool isArrayAllocation() const;
89
90 /// Get the number of elements allocated. For a simple allocation of a single
91 /// element, this will return a constant 1 value.
92 const Value *getArraySize() const { return getOperand(0); }
93 Value *getArraySize() { return getOperand(0); }
94
95 /// Overload to return most specific pointer type.
96 PointerType *getType() const {
97 return cast<PointerType>(Instruction::getType());
98 }
99
100 /// Get allocation size in bits. Returns None if size can't be determined,
101 /// e.g. in case of a VLA.
102 Optional<uint64_t> getAllocationSizeInBits(const DataLayout &DL) const;
103
104 /// Return the type that is being allocated by the instruction.
105 Type *getAllocatedType() const { return AllocatedType; }
106 /// for use only in special circumstances that need to generically
107 /// transform a whole instruction (eg: IR linking and vectorization).
108 void setAllocatedType(Type *Ty) { AllocatedType = Ty; }
109
110 /// Return the alignment of the memory that is being allocated by the
111 /// instruction.
112 unsigned getAlignment() const {
113 if (const auto MA = decodeMaybeAlign(getSubclassDataFromInstruction() & 31))
114 return MA->value();
115 return 0;
116 }
117 void setAlignment(MaybeAlign Align);
118
119 /// Return true if this alloca is in the entry block of the function and is a
120 /// constant size. If so, the code generator will fold it into the
121 /// prolog/epilog code, so it is basically free.
122 bool isStaticAlloca() const;
123
124 /// Return true if this alloca is used as an inalloca argument to a call. Such
125 /// allocas are never considered static even if they are in the entry block.
126 bool isUsedWithInAlloca() const {
127 return getSubclassDataFromInstruction() & 32;
128 }
129
130 /// Specify whether this alloca is used to represent the arguments to a call.
131 void setUsedWithInAlloca(bool V) {
132 setInstructionSubclassData((getSubclassDataFromInstruction() & ~32) |
133 (V ? 32 : 0));
134 }
135
136 /// Return true if this alloca is used as a swifterror argument to a call.
137 bool isSwiftError() const {
138 return getSubclassDataFromInstruction() & 64;
139 }
140
141 /// Specify whether this alloca is used to represent a swifterror.
142 void setSwiftError(bool V) {
143 setInstructionSubclassData((getSubclassDataFromInstruction() & ~64) |
144 (V ? 64 : 0));
145 }
146
147 // Methods for support type inquiry through isa, cast, and dyn_cast:
148 static bool classof(const Instruction *I) {
149 return (I->getOpcode() == Instruction::Alloca);
150 }
151 static bool classof(const Value *V) {
152 return isa<Instruction>(V) && classof(cast<Instruction>(V));
153 }
154
155private:
156 // Shadow Instruction::setInstructionSubclassData with a private forwarding
157 // method so that subclasses cannot accidentally use it.
158 void setInstructionSubclassData(unsigned short D) {
159 Instruction::setInstructionSubclassData(D);
160 }
161};
162
163//===----------------------------------------------------------------------===//
164// LoadInst Class
165//===----------------------------------------------------------------------===//
166
167/// An instruction for reading from memory. This uses the SubclassData field in
168/// Value to store whether or not the load is volatile.
169class LoadInst : public UnaryInstruction {
170 void AssertOK();
171
172protected:
173 // Note: Instruction needs to be a friend here to call cloneImpl.
174 friend class Instruction;
175
176 LoadInst *cloneImpl() const;
177
178public:
179 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr = "",
180 Instruction *InsertBefore = nullptr);
181 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, BasicBlock *InsertAtEnd);
182 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile,
183 Instruction *InsertBefore = nullptr);
184 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile,
185 BasicBlock *InsertAtEnd);
186 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile,
187 unsigned Align, Instruction *InsertBefore = nullptr);
188 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile,
189 unsigned Align, BasicBlock *InsertAtEnd);
190 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile,
191 unsigned Align, AtomicOrdering Order,
192 SyncScope::ID SSID = SyncScope::System,
193 Instruction *InsertBefore = nullptr);
194 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile,
195 unsigned Align, AtomicOrdering Order, SyncScope::ID SSID,
196 BasicBlock *InsertAtEnd);
197
198 // Deprecated [opaque pointer types]
199 explicit LoadInst(Value *Ptr, const Twine &NameStr = "",
200 Instruction *InsertBefore = nullptr)
201 : LoadInst(Ptr->getType()->getPointerElementType(), Ptr, NameStr,
202 InsertBefore) {}
203 LoadInst(Value *Ptr, const Twine &NameStr, BasicBlock *InsertAtEnd)
204 : LoadInst(Ptr->getType()->getPointerElementType(), Ptr, NameStr,
205 InsertAtEnd) {}
206 LoadInst(Value *Ptr, const Twine &NameStr, bool isVolatile,
207 Instruction *InsertBefore = nullptr)
208 : LoadInst(Ptr->getType()->getPointerElementType(), Ptr, NameStr,
209 isVolatile, InsertBefore) {}
210 LoadInst(Value *Ptr, const Twine &NameStr, bool isVolatile,
211 BasicBlock *InsertAtEnd)
212 : LoadInst(Ptr->getType()->getPointerElementType(), Ptr, NameStr,
213 isVolatile, InsertAtEnd) {}
214 LoadInst(Value *Ptr, const Twine &NameStr, bool isVolatile, unsigned Align,
215 Instruction *InsertBefore = nullptr)
216 : LoadInst(Ptr->getType()->getPointerElementType(), Ptr, NameStr,
217 isVolatile, Align, InsertBefore) {}
218 LoadInst(Value *Ptr, const Twine &NameStr, bool isVolatile, unsigned Align,
219 BasicBlock *InsertAtEnd)
220 : LoadInst(Ptr->getType()->getPointerElementType(), Ptr, NameStr,
221 isVolatile, Align, InsertAtEnd) {}
222 LoadInst(Value *Ptr, const Twine &NameStr, bool isVolatile, unsigned Align,
223 AtomicOrdering Order, SyncScope::ID SSID = SyncScope::System,
224 Instruction *InsertBefore = nullptr)
225 : LoadInst(Ptr->getType()->getPointerElementType(), Ptr, NameStr,
226 isVolatile, Align, Order, SSID, InsertBefore) {}
227 LoadInst(Value *Ptr, const Twine &NameStr, bool isVolatile, unsigned Align,
228 AtomicOrdering Order, SyncScope::ID SSID, BasicBlock *InsertAtEnd)
229 : LoadInst(Ptr->getType()->getPointerElementType(), Ptr, NameStr,
230 isVolatile, Align, Order, SSID, InsertAtEnd) {}
231
232 /// Return true if this is a load from a volatile memory location.
233 bool isVolatile() const { return getSubclassDataFromInstruction() & 1; }
234
235 /// Specify whether this is a volatile load or not.
236 void setVolatile(bool V) {
237 setInstructionSubclassData((getSubclassDataFromInstruction() & ~1) |
238 (V ? 1 : 0));
239 }
240
241 /// Return the alignment of the access that is being performed.
242 unsigned getAlignment() const {
243 if (const auto MA =
244 decodeMaybeAlign((getSubclassDataFromInstruction() >> 1) & 31))
245 return MA->value();
246 return 0;
247 }
248
249 void setAlignment(MaybeAlign Align);
250
251 /// Returns the ordering constraint of this load instruction.
252 AtomicOrdering getOrdering() const {
253 return AtomicOrdering((getSubclassDataFromInstruction() >> 7) & 7);
254 }
255
256 /// Sets the ordering constraint of this load instruction. May not be Release
257 /// or AcquireRelease.
258 void setOrdering(AtomicOrdering Ordering) {
259 setInstructionSubclassData((getSubclassDataFromInstruction() & ~(7 << 7)) |
260 ((unsigned)Ordering << 7));
261 }
262
263 /// Returns the synchronization scope ID of this load instruction.
264 SyncScope::ID getSyncScopeID() const {
265 return SSID;
266 }
267
268 /// Sets the synchronization scope ID of this load instruction.
269 void setSyncScopeID(SyncScope::ID SSID) {
270 this->SSID = SSID;
271 }
272
273 /// Sets the ordering constraint and the synchronization scope ID of this load
274 /// instruction.
275 void setAtomic(AtomicOrdering Ordering,
276 SyncScope::ID SSID = SyncScope::System) {
277 setOrdering(Ordering);
278 setSyncScopeID(SSID);
279 }
280
281 bool isSimple() const { return !isAtomic() && !isVolatile(); }
282
283 bool isUnordered() const {
284 return (getOrdering() == AtomicOrdering::NotAtomic ||
34
Assuming the condition is false
37
Returning the value 1, which participates in a condition later
285 getOrdering() == AtomicOrdering::Unordered) &&
35
Assuming the condition is true
286 !isVolatile();
36
Assuming the condition is true
287 }
288
289 Value *getPointerOperand() { return getOperand(0); }
290 const Value *getPointerOperand() const { return getOperand(0); }
291 static unsigned getPointerOperandIndex() { return 0U; }
292 Type *getPointerOperandType() const { return getPointerOperand()->getType(); }
293
294 /// Returns the address space of the pointer operand.
295 unsigned getPointerAddressSpace() const {
296 return getPointerOperandType()->getPointerAddressSpace();
297 }
298
299 // Methods for support type inquiry through isa, cast, and dyn_cast:
300 static bool classof(const Instruction *I) {
301 return I->getOpcode() == Instruction::Load;
302 }
303 static bool classof(const Value *V) {
304 return isa<Instruction>(V) && classof(cast<Instruction>(V));
305 }
306
307private:
308 // Shadow Instruction::setInstructionSubclassData with a private forwarding
309 // method so that subclasses cannot accidentally use it.
310 void setInstructionSubclassData(unsigned short D) {
311 Instruction::setInstructionSubclassData(D);
312 }
313
314 /// The synchronization scope ID of this load instruction. Not quite enough
315 /// room in SubClassData for everything, so synchronization scope ID gets its
316 /// own field.
317 SyncScope::ID SSID;
318};
319
320//===----------------------------------------------------------------------===//
321// StoreInst Class
322//===----------------------------------------------------------------------===//
323
324/// An instruction for storing to memory.
325class StoreInst : public Instruction {
326 void AssertOK();
327
328protected:
329 // Note: Instruction needs to be a friend here to call cloneImpl.
330 friend class Instruction;
331
332 StoreInst *cloneImpl() const;
333
334public:
335 StoreInst(Value *Val, Value *Ptr, Instruction *InsertBefore);
336 StoreInst(Value *Val, Value *Ptr, BasicBlock *InsertAtEnd);
337 StoreInst(Value *Val, Value *Ptr, bool isVolatile = false,
338 Instruction *InsertBefore = nullptr);
339 StoreInst(Value *Val, Value *Ptr, bool isVolatile, BasicBlock *InsertAtEnd);
340 StoreInst(Value *Val, Value *Ptr, bool isVolatile,
341 unsigned Align, Instruction *InsertBefore = nullptr);
342 StoreInst(Value *Val, Value *Ptr, bool isVolatile,
343 unsigned Align, BasicBlock *InsertAtEnd);
344 StoreInst(Value *Val, Value *Ptr, bool isVolatile,
345 unsigned Align, AtomicOrdering Order,
346 SyncScope::ID SSID = SyncScope::System,
347 Instruction *InsertBefore = nullptr);
348 StoreInst(Value *Val, Value *Ptr, bool isVolatile,
349 unsigned Align, AtomicOrdering Order, SyncScope::ID SSID,
350 BasicBlock *InsertAtEnd);
351
352 // allocate space for exactly two operands
353 void *operator new(size_t s) {
354 return User::operator new(s, 2);
355 }
356
357 /// Return true if this is a store to a volatile memory location.
358 bool isVolatile() const { return getSubclassDataFromInstruction() & 1; }
359
360 /// Specify whether this is a volatile store or not.
361 void setVolatile(bool V) {
362 setInstructionSubclassData((getSubclassDataFromInstruction() & ~1) |
363 (V ? 1 : 0));
364 }
365
366 /// Transparently provide more efficient getOperand methods.
367 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
;
368
369 /// Return the alignment of the access that is being performed
370 unsigned getAlignment() const {
371 if (const auto MA =
372 decodeMaybeAlign((getSubclassDataFromInstruction() >> 1) & 31))
373 return MA->value();
374 return 0;
375 }
376
377 void setAlignment(MaybeAlign Align);
378
379 /// Returns the ordering constraint of this store instruction.
380 AtomicOrdering getOrdering() const {
381 return AtomicOrdering((getSubclassDataFromInstruction() >> 7) & 7);
382 }
383
384 /// Sets the ordering constraint of this store instruction. May not be
385 /// Acquire or AcquireRelease.
386 void setOrdering(AtomicOrdering Ordering) {
387 setInstructionSubclassData((getSubclassDataFromInstruction() & ~(7 << 7)) |
388 ((unsigned)Ordering << 7));
389 }
390
391 /// Returns the synchronization scope ID of this store instruction.
392 SyncScope::ID getSyncScopeID() const {
393 return SSID;
394 }
395
396 /// Sets the synchronization scope ID of this store instruction.
397 void setSyncScopeID(SyncScope::ID SSID) {
398 this->SSID = SSID;
399 }
400
401 /// Sets the ordering constraint and the synchronization scope ID of this
402 /// store instruction.
403 void setAtomic(AtomicOrdering Ordering,
404 SyncScope::ID SSID = SyncScope::System) {
405 setOrdering(Ordering);
406 setSyncScopeID(SSID);
407 }
408
409 bool isSimple() const { return !isAtomic() && !isVolatile(); }
410
411 bool isUnordered() const {
412 return (getOrdering() == AtomicOrdering::NotAtomic ||
413 getOrdering() == AtomicOrdering::Unordered) &&
414 !isVolatile();
415 }
416
417 Value *getValueOperand() { return getOperand(0); }
418 const Value *getValueOperand() const { return getOperand(0); }
419
420 Value *getPointerOperand() { return getOperand(1); }
421 const Value *getPointerOperand() const { return getOperand(1); }
422 static unsigned getPointerOperandIndex() { return 1U; }
423 Type *getPointerOperandType() const { return getPointerOperand()->getType(); }
424
425 /// Returns the address space of the pointer operand.
426 unsigned getPointerAddressSpace() const {
427 return getPointerOperandType()->getPointerAddressSpace();
428 }
429
430 // Methods for support type inquiry through isa, cast, and dyn_cast:
431 static bool classof(const Instruction *I) {
432 return I->getOpcode() == Instruction::Store;
433 }
434 static bool classof(const Value *V) {
435 return isa<Instruction>(V) && classof(cast<Instruction>(V));
436 }
437
438private:
439 // Shadow Instruction::setInstructionSubclassData with a private forwarding
440 // method so that subclasses cannot accidentally use it.
441 void setInstructionSubclassData(unsigned short D) {
442 Instruction::setInstructionSubclassData(D);
443 }
444
445 /// The synchronization scope ID of this store instruction. Not quite enough
446 /// room in SubClassData for everything, so synchronization scope ID gets its
447 /// own field.
448 SyncScope::ID SSID;
449};
450
451template <>
452struct OperandTraits<StoreInst> : public FixedNumOperandTraits<StoreInst, 2> {
453};
454
455DEFINE_TRANSPARENT_OPERAND_ACCESSORS(StoreInst, Value)StoreInst::op_iterator StoreInst::op_begin() { return OperandTraits
<StoreInst>::op_begin(this); } StoreInst::const_op_iterator
StoreInst::op_begin() const { return OperandTraits<StoreInst
>::op_begin(const_cast<StoreInst*>(this)); } StoreInst
::op_iterator StoreInst::op_end() { return OperandTraits<StoreInst
>::op_end(this); } StoreInst::const_op_iterator StoreInst::
op_end() const { return OperandTraits<StoreInst>::op_end
(const_cast<StoreInst*>(this)); } Value *StoreInst::getOperand
(unsigned i_nocapture) const { ((i_nocapture < OperandTraits
<StoreInst>::operands(this) && "getOperand() out of range!"
) ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<StoreInst>::operands(this) && \"getOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 455, __PRETTY_FUNCTION__)); return cast_or_null<Value>
( OperandTraits<StoreInst>::op_begin(const_cast<StoreInst
*>(this))[i_nocapture].get()); } void StoreInst::setOperand
(unsigned i_nocapture, Value *Val_nocapture) { ((i_nocapture <
OperandTraits<StoreInst>::operands(this) && "setOperand() out of range!"
) ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<StoreInst>::operands(this) && \"setOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 455, __PRETTY_FUNCTION__)); OperandTraits<StoreInst>::
op_begin(this)[i_nocapture] = Val_nocapture; } unsigned StoreInst
::getNumOperands() const { return OperandTraits<StoreInst>
::operands(this); } template <int Idx_nocapture> Use &
StoreInst::Op() { return this->OpFrom<Idx_nocapture>
(this); } template <int Idx_nocapture> const Use &StoreInst
::Op() const { return this->OpFrom<Idx_nocapture>(this
); }
456
457//===----------------------------------------------------------------------===//
458// FenceInst Class
459//===----------------------------------------------------------------------===//
460
461/// An instruction for ordering other memory operations.
462class FenceInst : public Instruction {
463 void Init(AtomicOrdering Ordering, SyncScope::ID SSID);
464
465protected:
466 // Note: Instruction needs to be a friend here to call cloneImpl.
467 friend class Instruction;
468
469 FenceInst *cloneImpl() const;
470
471public:
472 // Ordering may only be Acquire, Release, AcquireRelease, or
473 // SequentiallyConsistent.
474 FenceInst(LLVMContext &C, AtomicOrdering Ordering,
475 SyncScope::ID SSID = SyncScope::System,
476 Instruction *InsertBefore = nullptr);
477 FenceInst(LLVMContext &C, AtomicOrdering Ordering, SyncScope::ID SSID,
478 BasicBlock *InsertAtEnd);
479
480 // allocate space for exactly zero operands
481 void *operator new(size_t s) {
482 return User::operator new(s, 0);
483 }
484
485 /// Returns the ordering constraint of this fence instruction.
486 AtomicOrdering getOrdering() const {
487 return AtomicOrdering(getSubclassDataFromInstruction() >> 1);
488 }
489
490 /// Sets the ordering constraint of this fence instruction. May only be
491 /// Acquire, Release, AcquireRelease, or SequentiallyConsistent.
492 void setOrdering(AtomicOrdering Ordering) {
493 setInstructionSubclassData((getSubclassDataFromInstruction() & 1) |
494 ((unsigned)Ordering << 1));
495 }
496
497 /// Returns the synchronization scope ID of this fence instruction.
498 SyncScope::ID getSyncScopeID() const {
499 return SSID;
500 }
501
502 /// Sets the synchronization scope ID of this fence instruction.
503 void setSyncScopeID(SyncScope::ID SSID) {
504 this->SSID = SSID;
505 }
506
507 // Methods for support type inquiry through isa, cast, and dyn_cast:
508 static bool classof(const Instruction *I) {
509 return I->getOpcode() == Instruction::Fence;
510 }
511 static bool classof(const Value *V) {
512 return isa<Instruction>(V) && classof(cast<Instruction>(V));
513 }
514
515private:
516 // Shadow Instruction::setInstructionSubclassData with a private forwarding
517 // method so that subclasses cannot accidentally use it.
518 void setInstructionSubclassData(unsigned short D) {
519 Instruction::setInstructionSubclassData(D);
520 }
521
522 /// The synchronization scope ID of this fence instruction. Not quite enough
523 /// room in SubClassData for everything, so synchronization scope ID gets its
524 /// own field.
525 SyncScope::ID SSID;
526};
527
528//===----------------------------------------------------------------------===//
529// AtomicCmpXchgInst Class
530//===----------------------------------------------------------------------===//
531
532/// An instruction that atomically checks whether a
533/// specified value is in a memory location, and, if it is, stores a new value
534/// there. The value returned by this instruction is a pair containing the
535/// original value as first element, and an i1 indicating success (true) or
536/// failure (false) as second element.
537///
538class AtomicCmpXchgInst : public Instruction {
539 void Init(Value *Ptr, Value *Cmp, Value *NewVal,
540 AtomicOrdering SuccessOrdering, AtomicOrdering FailureOrdering,
541 SyncScope::ID SSID);
542
543protected:
544 // Note: Instruction needs to be a friend here to call cloneImpl.
545 friend class Instruction;
546
547 AtomicCmpXchgInst *cloneImpl() const;
548
549public:
550 AtomicCmpXchgInst(Value *Ptr, Value *Cmp, Value *NewVal,
551 AtomicOrdering SuccessOrdering,
552 AtomicOrdering FailureOrdering,
553 SyncScope::ID SSID, Instruction *InsertBefore = nullptr);
554 AtomicCmpXchgInst(Value *Ptr, Value *Cmp, Value *NewVal,
555 AtomicOrdering SuccessOrdering,
556 AtomicOrdering FailureOrdering,
557 SyncScope::ID SSID, BasicBlock *InsertAtEnd);
558
559 // allocate space for exactly three operands
560 void *operator new(size_t s) {
561 return User::operator new(s, 3);
562 }
563
564 /// Return true if this is a cmpxchg from a volatile memory
565 /// location.
566 ///
567 bool isVolatile() const {
568 return getSubclassDataFromInstruction() & 1;
569 }
570
571 /// Specify whether this is a volatile cmpxchg.
572 ///
573 void setVolatile(bool V) {
574 setInstructionSubclassData((getSubclassDataFromInstruction() & ~1) |
575 (unsigned)V);
576 }
577
578 /// Return true if this cmpxchg may spuriously fail.
579 bool isWeak() const {
580 return getSubclassDataFromInstruction() & 0x100;
581 }
582
583 void setWeak(bool IsWeak) {
584 setInstructionSubclassData((getSubclassDataFromInstruction() & ~0x100) |
585 (IsWeak << 8));
586 }
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 /// Returns the success ordering constraint of this cmpxchg instruction.
592 AtomicOrdering getSuccessOrdering() const {
593 return AtomicOrdering((getSubclassDataFromInstruction() >> 2) & 7);
594 }
595
596 /// Sets the success ordering constraint of this cmpxchg instruction.
597 void setSuccessOrdering(AtomicOrdering Ordering) {
598 assert(Ordering != AtomicOrdering::NotAtomic &&((Ordering != AtomicOrdering::NotAtomic && "CmpXchg instructions can only be atomic."
) ? static_cast<void> (0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"CmpXchg instructions can only be atomic.\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 599, __PRETTY_FUNCTION__))
599 "CmpXchg instructions can only be atomic.")((Ordering != AtomicOrdering::NotAtomic && "CmpXchg instructions can only be atomic."
) ? static_cast<void> (0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"CmpXchg instructions can only be atomic.\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 599, __PRETTY_FUNCTION__))
;
600 setInstructionSubclassData((getSubclassDataFromInstruction() & ~0x1c) |
601 ((unsigned)Ordering << 2));
602 }
603
604 /// Returns the failure ordering constraint of this cmpxchg instruction.
605 AtomicOrdering getFailureOrdering() const {
606 return AtomicOrdering((getSubclassDataFromInstruction() >> 5) & 7);
607 }
608
609 /// Sets the failure ordering constraint of this cmpxchg instruction.
610 void setFailureOrdering(AtomicOrdering Ordering) {
611 assert(Ordering != AtomicOrdering::NotAtomic &&((Ordering != AtomicOrdering::NotAtomic && "CmpXchg instructions can only be atomic."
) ? static_cast<void> (0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"CmpXchg instructions can only be atomic.\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 612, __PRETTY_FUNCTION__))
612 "CmpXchg instructions can only be atomic.")((Ordering != AtomicOrdering::NotAtomic && "CmpXchg instructions can only be atomic."
) ? static_cast<void> (0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"CmpXchg instructions can only be atomic.\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 612, __PRETTY_FUNCTION__))
;
613 setInstructionSubclassData((getSubclassDataFromInstruction() & ~0xe0) |
614 ((unsigned)Ordering << 5));
615 }
616
617 /// Returns the synchronization scope ID of this cmpxchg instruction.
618 SyncScope::ID getSyncScopeID() const {
619 return SSID;
620 }
621
622 /// Sets the synchronization scope ID of this cmpxchg instruction.
623 void setSyncScopeID(SyncScope::ID SSID) {
624 this->SSID = SSID;
625 }
626
627 Value *getPointerOperand() { return getOperand(0); }
628 const Value *getPointerOperand() const { return getOperand(0); }
629 static unsigned getPointerOperandIndex() { return 0U; }
630
631 Value *getCompareOperand() { return getOperand(1); }
632 const Value *getCompareOperand() const { return getOperand(1); }
633
634 Value *getNewValOperand() { return getOperand(2); }
635 const Value *getNewValOperand() const { return getOperand(2); }
636
637 /// Returns the address space of the pointer operand.
638 unsigned getPointerAddressSpace() const {
639 return getPointerOperand()->getType()->getPointerAddressSpace();
640 }
641
642 /// Returns the strongest permitted ordering on failure, given the
643 /// desired ordering on success.
644 ///
645 /// If the comparison in a cmpxchg operation fails, there is no atomic store
646 /// so release semantics cannot be provided. So this function drops explicit
647 /// Release requests from the AtomicOrdering. A SequentiallyConsistent
648 /// operation would remain SequentiallyConsistent.
649 static AtomicOrdering
650 getStrongestFailureOrdering(AtomicOrdering SuccessOrdering) {
651 switch (SuccessOrdering) {
652 default:
653 llvm_unreachable("invalid cmpxchg success ordering")::llvm::llvm_unreachable_internal("invalid cmpxchg success ordering"
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 653)
;
654 case AtomicOrdering::Release:
655 case AtomicOrdering::Monotonic:
656 return AtomicOrdering::Monotonic;
657 case AtomicOrdering::AcquireRelease:
658 case AtomicOrdering::Acquire:
659 return AtomicOrdering::Acquire;
660 case AtomicOrdering::SequentiallyConsistent:
661 return AtomicOrdering::SequentiallyConsistent;
662 }
663 }
664
665 // Methods for support type inquiry through isa, cast, and dyn_cast:
666 static bool classof(const Instruction *I) {
667 return I->getOpcode() == Instruction::AtomicCmpXchg;
668 }
669 static bool classof(const Value *V) {
670 return isa<Instruction>(V) && classof(cast<Instruction>(V));
671 }
672
673private:
674 // Shadow Instruction::setInstructionSubclassData with a private forwarding
675 // method so that subclasses cannot accidentally use it.
676 void setInstructionSubclassData(unsigned short D) {
677 Instruction::setInstructionSubclassData(D);
678 }
679
680 /// The synchronization scope ID of this cmpxchg instruction. Not quite
681 /// enough room in SubClassData for everything, so synchronization scope ID
682 /// gets its own field.
683 SyncScope::ID SSID;
684};
685
686template <>
687struct OperandTraits<AtomicCmpXchgInst> :
688 public FixedNumOperandTraits<AtomicCmpXchgInst, 3> {
689};
690
691DEFINE_TRANSPARENT_OPERAND_ACCESSORS(AtomicCmpXchgInst, Value)AtomicCmpXchgInst::op_iterator AtomicCmpXchgInst::op_begin() {
return OperandTraits<AtomicCmpXchgInst>::op_begin(this
); } AtomicCmpXchgInst::const_op_iterator AtomicCmpXchgInst::
op_begin() const { return OperandTraits<AtomicCmpXchgInst>
::op_begin(const_cast<AtomicCmpXchgInst*>(this)); } AtomicCmpXchgInst
::op_iterator AtomicCmpXchgInst::op_end() { return OperandTraits
<AtomicCmpXchgInst>::op_end(this); } AtomicCmpXchgInst::
const_op_iterator AtomicCmpXchgInst::op_end() const { return OperandTraits
<AtomicCmpXchgInst>::op_end(const_cast<AtomicCmpXchgInst
*>(this)); } Value *AtomicCmpXchgInst::getOperand(unsigned
i_nocapture) const { ((i_nocapture < OperandTraits<AtomicCmpXchgInst
>::operands(this) && "getOperand() out of range!")
? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<AtomicCmpXchgInst>::operands(this) && \"getOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 691, __PRETTY_FUNCTION__)); return cast_or_null<Value>
( OperandTraits<AtomicCmpXchgInst>::op_begin(const_cast
<AtomicCmpXchgInst*>(this))[i_nocapture].get()); } void
AtomicCmpXchgInst::setOperand(unsigned i_nocapture, Value *Val_nocapture
) { ((i_nocapture < OperandTraits<AtomicCmpXchgInst>
::operands(this) && "setOperand() out of range!") ? static_cast
<void> (0) : __assert_fail ("i_nocapture < OperandTraits<AtomicCmpXchgInst>::operands(this) && \"setOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 691, __PRETTY_FUNCTION__)); OperandTraits<AtomicCmpXchgInst
>::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned
AtomicCmpXchgInst::getNumOperands() const { return OperandTraits
<AtomicCmpXchgInst>::operands(this); } template <int
Idx_nocapture> Use &AtomicCmpXchgInst::Op() { return this
->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture
> const Use &AtomicCmpXchgInst::Op() const { return this
->OpFrom<Idx_nocapture>(this); }
692
693//===----------------------------------------------------------------------===//
694// AtomicRMWInst Class
695//===----------------------------------------------------------------------===//
696
697/// an instruction that atomically reads a memory location,
698/// combines it with another value, and then stores the result back. Returns
699/// the old value.
700///
701class AtomicRMWInst : public Instruction {
702protected:
703 // Note: Instruction needs to be a friend here to call cloneImpl.
704 friend class Instruction;
705
706 AtomicRMWInst *cloneImpl() const;
707
708public:
709 /// This enumeration lists the possible modifications atomicrmw can make. In
710 /// the descriptions, 'p' is the pointer to the instruction's memory location,
711 /// 'old' is the initial value of *p, and 'v' is the other value passed to the
712 /// instruction. These instructions always return 'old'.
713 enum BinOp {
714 /// *p = v
715 Xchg,
716 /// *p = old + v
717 Add,
718 /// *p = old - v
719 Sub,
720 /// *p = old & v
721 And,
722 /// *p = ~(old & v)
723 Nand,
724 /// *p = old | v
725 Or,
726 /// *p = old ^ v
727 Xor,
728 /// *p = old >signed v ? old : v
729 Max,
730 /// *p = old <signed v ? old : v
731 Min,
732 /// *p = old >unsigned v ? old : v
733 UMax,
734 /// *p = old <unsigned v ? old : v
735 UMin,
736
737 /// *p = old + v
738 FAdd,
739
740 /// *p = old - v
741 FSub,
742
743 FIRST_BINOP = Xchg,
744 LAST_BINOP = FSub,
745 BAD_BINOP
746 };
747
748 AtomicRMWInst(BinOp Operation, Value *Ptr, Value *Val,
749 AtomicOrdering Ordering, SyncScope::ID SSID,
750 Instruction *InsertBefore = nullptr);
751 AtomicRMWInst(BinOp Operation, Value *Ptr, Value *Val,
752 AtomicOrdering Ordering, SyncScope::ID SSID,
753 BasicBlock *InsertAtEnd);
754
755 // allocate space for exactly two operands
756 void *operator new(size_t s) {
757 return User::operator new(s, 2);
758 }
759
760 BinOp getOperation() const {
761 return static_cast<BinOp>(getSubclassDataFromInstruction() >> 5);
762 }
763
764 static StringRef getOperationName(BinOp Op);
765
766 static bool isFPOperation(BinOp Op) {
767 switch (Op) {
768 case AtomicRMWInst::FAdd:
769 case AtomicRMWInst::FSub:
770 return true;
771 default:
772 return false;
773 }
774 }
775
776 void setOperation(BinOp Operation) {
777 unsigned short SubclassData = getSubclassDataFromInstruction();
778 setInstructionSubclassData((SubclassData & 31) |
779 (Operation << 5));
780 }
781
782 /// Return true if this is a RMW on a volatile memory location.
783 ///
784 bool isVolatile() const {
785 return getSubclassDataFromInstruction() & 1;
786 }
787
788 /// Specify whether this is a volatile RMW or not.
789 ///
790 void setVolatile(bool V) {
791 setInstructionSubclassData((getSubclassDataFromInstruction() & ~1) |
792 (unsigned)V);
793 }
794
795 /// Transparently provide more efficient getOperand methods.
796 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
;
797
798 /// Returns the ordering constraint of this rmw instruction.
799 AtomicOrdering getOrdering() const {
800 return AtomicOrdering((getSubclassDataFromInstruction() >> 2) & 7);
801 }
802
803 /// Sets the ordering constraint of this rmw instruction.
804 void setOrdering(AtomicOrdering Ordering) {
805 assert(Ordering != AtomicOrdering::NotAtomic &&((Ordering != AtomicOrdering::NotAtomic && "atomicrmw instructions can only be atomic."
) ? static_cast<void> (0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"atomicrmw instructions can only be atomic.\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 806, __PRETTY_FUNCTION__))
806 "atomicrmw instructions can only be atomic.")((Ordering != AtomicOrdering::NotAtomic && "atomicrmw instructions can only be atomic."
) ? static_cast<void> (0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"atomicrmw instructions can only be atomic.\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 806, __PRETTY_FUNCTION__))
;
807 setInstructionSubclassData((getSubclassDataFromInstruction() & ~(7 << 2)) |
808 ((unsigned)Ordering << 2));
809 }
810
811 /// Returns the synchronization scope ID of this rmw instruction.
812 SyncScope::ID getSyncScopeID() const {
813 return SSID;
814 }
815
816 /// Sets the synchronization scope ID of this rmw instruction.
817 void setSyncScopeID(SyncScope::ID SSID) {
818 this->SSID = SSID;
819 }
820
821 Value *getPointerOperand() { return getOperand(0); }
822 const Value *getPointerOperand() const { return getOperand(0); }
823 static unsigned getPointerOperandIndex() { return 0U; }
824
825 Value *getValOperand() { return getOperand(1); }
826 const Value *getValOperand() const { return getOperand(1); }
827
828 /// Returns the address space of the pointer operand.
829 unsigned getPointerAddressSpace() const {
830 return getPointerOperand()->getType()->getPointerAddressSpace();
831 }
832
833 bool isFloatingPointOperation() const {
834 return isFPOperation(getOperation());
835 }
836
837 // Methods for support type inquiry through isa, cast, and dyn_cast:
838 static bool classof(const Instruction *I) {
839 return I->getOpcode() == Instruction::AtomicRMW;
840 }
841 static bool classof(const Value *V) {
842 return isa<Instruction>(V) && classof(cast<Instruction>(V));
843 }
844
845private:
846 void Init(BinOp Operation, Value *Ptr, Value *Val,
847 AtomicOrdering Ordering, SyncScope::ID SSID);
848
849 // Shadow Instruction::setInstructionSubclassData with a private forwarding
850 // method so that subclasses cannot accidentally use it.
851 void setInstructionSubclassData(unsigned short D) {
852 Instruction::setInstructionSubclassData(D);
853 }
854
855 /// The synchronization scope ID of this rmw instruction. Not quite enough
856 /// room in SubClassData for everything, so synchronization scope ID gets its
857 /// own field.
858 SyncScope::ID SSID;
859};
860
861template <>
862struct OperandTraits<AtomicRMWInst>
863 : public FixedNumOperandTraits<AtomicRMWInst,2> {
864};
865
866DEFINE_TRANSPARENT_OPERAND_ACCESSORS(AtomicRMWInst, Value)AtomicRMWInst::op_iterator AtomicRMWInst::op_begin() { return
OperandTraits<AtomicRMWInst>::op_begin(this); } AtomicRMWInst
::const_op_iterator AtomicRMWInst::op_begin() const { return OperandTraits
<AtomicRMWInst>::op_begin(const_cast<AtomicRMWInst*>
(this)); } AtomicRMWInst::op_iterator AtomicRMWInst::op_end()
{ return OperandTraits<AtomicRMWInst>::op_end(this); }
AtomicRMWInst::const_op_iterator AtomicRMWInst::op_end() const
{ return OperandTraits<AtomicRMWInst>::op_end(const_cast
<AtomicRMWInst*>(this)); } Value *AtomicRMWInst::getOperand
(unsigned i_nocapture) const { ((i_nocapture < OperandTraits
<AtomicRMWInst>::operands(this) && "getOperand() out of range!"
) ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<AtomicRMWInst>::operands(this) && \"getOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 866, __PRETTY_FUNCTION__)); return cast_or_null<Value>
( OperandTraits<AtomicRMWInst>::op_begin(const_cast<
AtomicRMWInst*>(this))[i_nocapture].get()); } void AtomicRMWInst
::setOperand(unsigned i_nocapture, Value *Val_nocapture) { ((
i_nocapture < OperandTraits<AtomicRMWInst>::operands
(this) && "setOperand() out of range!") ? static_cast
<void> (0) : __assert_fail ("i_nocapture < OperandTraits<AtomicRMWInst>::operands(this) && \"setOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 866, __PRETTY_FUNCTION__)); OperandTraits<AtomicRMWInst>
::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned AtomicRMWInst
::getNumOperands() const { return OperandTraits<AtomicRMWInst
>::operands(this); } template <int Idx_nocapture> Use
&AtomicRMWInst::Op() { return this->OpFrom<Idx_nocapture
>(this); } template <int Idx_nocapture> const Use &
AtomicRMWInst::Op() const { return this->OpFrom<Idx_nocapture
>(this); }
867
868//===----------------------------------------------------------------------===//
869// GetElementPtrInst Class
870//===----------------------------------------------------------------------===//
871
872// checkGEPType - Simple wrapper function to give a better assertion failure
873// message on bad indexes for a gep instruction.
874//
875inline Type *checkGEPType(Type *Ty) {
876 assert(Ty && "Invalid GetElementPtrInst indices for type!")((Ty && "Invalid GetElementPtrInst indices for type!"
) ? static_cast<void> (0) : __assert_fail ("Ty && \"Invalid GetElementPtrInst indices for type!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 876, __PRETTY_FUNCTION__))
;
877 return Ty;
878}
879
880/// an instruction for type-safe pointer arithmetic to
881/// access elements of arrays and structs
882///
883class GetElementPtrInst : public Instruction {
884 Type *SourceElementType;
885 Type *ResultElementType;
886
887 GetElementPtrInst(const GetElementPtrInst &GEPI);
888
889 /// Constructors - Create a getelementptr instruction with a base pointer an
890 /// list of indices. The first ctor can optionally insert before an existing
891 /// instruction, the second appends the new instruction to the specified
892 /// BasicBlock.
893 inline GetElementPtrInst(Type *PointeeType, Value *Ptr,
894 ArrayRef<Value *> IdxList, unsigned Values,
895 const Twine &NameStr, Instruction *InsertBefore);
896 inline GetElementPtrInst(Type *PointeeType, Value *Ptr,
897 ArrayRef<Value *> IdxList, unsigned Values,
898 const Twine &NameStr, BasicBlock *InsertAtEnd);
899
900 void init(Value *Ptr, ArrayRef<Value *> IdxList, const Twine &NameStr);
901
902protected:
903 // Note: Instruction needs to be a friend here to call cloneImpl.
904 friend class Instruction;
905
906 GetElementPtrInst *cloneImpl() const;
907
908public:
909 static GetElementPtrInst *Create(Type *PointeeType, Value *Ptr,
910 ArrayRef<Value *> IdxList,
911 const Twine &NameStr = "",
912 Instruction *InsertBefore = nullptr) {
913 unsigned Values = 1 + unsigned(IdxList.size());
914 if (!PointeeType)
915 PointeeType =
916 cast<PointerType>(Ptr->getType()->getScalarType())->getElementType();
917 else
918 assert(((PointeeType == cast<PointerType>(Ptr->getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 920, __PRETTY_FUNCTION__))
919 PointeeType ==((PointeeType == cast<PointerType>(Ptr->getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 920, __PRETTY_FUNCTION__))
920 cast<PointerType>(Ptr->getType()->getScalarType())->getElementType())((PointeeType == cast<PointerType>(Ptr->getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 920, __PRETTY_FUNCTION__))
;
921 return new (Values) GetElementPtrInst(PointeeType, Ptr, IdxList, Values,
922 NameStr, InsertBefore);
923 }
924
925 static GetElementPtrInst *Create(Type *PointeeType, Value *Ptr,
926 ArrayRef<Value *> IdxList,
927 const Twine &NameStr,
928 BasicBlock *InsertAtEnd) {
929 unsigned Values = 1 + unsigned(IdxList.size());
930 if (!PointeeType)
931 PointeeType =
932 cast<PointerType>(Ptr->getType()->getScalarType())->getElementType();
933 else
934 assert(((PointeeType == cast<PointerType>(Ptr->getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 936, __PRETTY_FUNCTION__))
935 PointeeType ==((PointeeType == cast<PointerType>(Ptr->getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 936, __PRETTY_FUNCTION__))
936 cast<PointerType>(Ptr->getType()->getScalarType())->getElementType())((PointeeType == cast<PointerType>(Ptr->getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 936, __PRETTY_FUNCTION__))
;
937 return new (Values) GetElementPtrInst(PointeeType, Ptr, IdxList, Values,
938 NameStr, InsertAtEnd);
939 }
940
941 /// Create an "inbounds" getelementptr. See the documentation for the
942 /// "inbounds" flag in LangRef.html for details.
943 static GetElementPtrInst *CreateInBounds(Value *Ptr,
944 ArrayRef<Value *> IdxList,
945 const Twine &NameStr = "",
946 Instruction *InsertBefore = nullptr){
947 return CreateInBounds(nullptr, Ptr, IdxList, NameStr, InsertBefore);
948 }
949
950 static GetElementPtrInst *
951 CreateInBounds(Type *PointeeType, Value *Ptr, ArrayRef<Value *> IdxList,
952 const Twine &NameStr = "",
953 Instruction *InsertBefore = nullptr) {
954 GetElementPtrInst *GEP =
955 Create(PointeeType, Ptr, IdxList, NameStr, InsertBefore);
956 GEP->setIsInBounds(true);
957 return GEP;
958 }
959
960 static GetElementPtrInst *CreateInBounds(Value *Ptr,
961 ArrayRef<Value *> IdxList,
962 const Twine &NameStr,
963 BasicBlock *InsertAtEnd) {
964 return CreateInBounds(nullptr, Ptr, IdxList, NameStr, InsertAtEnd);
965 }
966
967 static GetElementPtrInst *CreateInBounds(Type *PointeeType, Value *Ptr,
968 ArrayRef<Value *> IdxList,
969 const Twine &NameStr,
970 BasicBlock *InsertAtEnd) {
971 GetElementPtrInst *GEP =
972 Create(PointeeType, Ptr, IdxList, NameStr, InsertAtEnd);
973 GEP->setIsInBounds(true);
974 return GEP;
975 }
976
977 /// Transparently provide more efficient getOperand methods.
978 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
;
979
980 Type *getSourceElementType() const { return SourceElementType; }
981
982 void setSourceElementType(Type *Ty) { SourceElementType = Ty; }
983 void setResultElementType(Type *Ty) { ResultElementType = Ty; }
984
985 Type *getResultElementType() const {
986 assert(ResultElementType ==((ResultElementType == cast<PointerType>(getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("ResultElementType == cast<PointerType>(getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 987, __PRETTY_FUNCTION__))
987 cast<PointerType>(getType()->getScalarType())->getElementType())((ResultElementType == cast<PointerType>(getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("ResultElementType == cast<PointerType>(getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 987, __PRETTY_FUNCTION__))
;
988 return ResultElementType;
989 }
990
991 /// Returns the address space of this instruction's pointer type.
992 unsigned getAddressSpace() const {
993 // Note that this is always the same as the pointer operand's address space
994 // and that is cheaper to compute, so cheat here.
995 return getPointerAddressSpace();
996 }
997
998 /// Returns the type of the element that would be loaded with
999 /// a load instruction with the specified parameters.
1000 ///
1001 /// Null is returned if the indices are invalid for the specified
1002 /// pointer type.
1003 ///
1004 static Type *getIndexedType(Type *Ty, ArrayRef<Value *> IdxList);
1005 static Type *getIndexedType(Type *Ty, ArrayRef<Constant *> IdxList);
1006 static Type *getIndexedType(Type *Ty, ArrayRef<uint64_t> IdxList);
1007
1008 inline op_iterator idx_begin() { return op_begin()+1; }
1009 inline const_op_iterator idx_begin() const { return op_begin()+1; }
1010 inline op_iterator idx_end() { return op_end(); }
1011 inline const_op_iterator idx_end() const { return op_end(); }
1012
1013 inline iterator_range<op_iterator> indices() {
1014 return make_range(idx_begin(), idx_end());
1015 }
1016
1017 inline iterator_range<const_op_iterator> indices() const {
1018 return make_range(idx_begin(), idx_end());
1019 }
1020
1021 Value *getPointerOperand() {
1022 return getOperand(0);
1023 }
1024 const Value *getPointerOperand() const {
1025 return getOperand(0);
1026 }
1027 static unsigned getPointerOperandIndex() {
1028 return 0U; // get index for modifying correct operand.
1029 }
1030
1031 /// Method to return the pointer operand as a
1032 /// PointerType.
1033 Type *getPointerOperandType() const {
1034 return getPointerOperand()->getType();
1035 }
1036
1037 /// Returns the address space of the pointer operand.
1038 unsigned getPointerAddressSpace() const {
1039 return getPointerOperandType()->getPointerAddressSpace();
1040 }
1041
1042 /// Returns the pointer type returned by the GEP
1043 /// instruction, which may be a vector of pointers.
1044 static Type *getGEPReturnType(Value *Ptr, ArrayRef<Value *> IdxList) {
1045 return getGEPReturnType(
1046 cast<PointerType>(Ptr->getType()->getScalarType())->getElementType(),
1047 Ptr, IdxList);
1048 }
1049 static Type *getGEPReturnType(Type *ElTy, Value *Ptr,
1050 ArrayRef<Value *> IdxList) {
1051 Type *PtrTy = PointerType::get(checkGEPType(getIndexedType(ElTy, IdxList)),
1052 Ptr->getType()->getPointerAddressSpace());
1053 // Vector GEP
1054 if (Ptr->getType()->isVectorTy()) {
1055 unsigned NumElem = Ptr->getType()->getVectorNumElements();
1056 return VectorType::get(PtrTy, NumElem);
1057 }
1058 for (Value *Index : IdxList)
1059 if (Index->getType()->isVectorTy()) {
1060 unsigned NumElem = Index->getType()->getVectorNumElements();
1061 return VectorType::get(PtrTy, NumElem);
1062 }
1063 // Scalar GEP
1064 return PtrTy;
1065 }
1066
1067 unsigned getNumIndices() const { // Note: always non-negative
1068 return getNumOperands() - 1;
1069 }
1070
1071 bool hasIndices() const {
1072 return getNumOperands() > 1;
1073 }
1074
1075 /// Return true if all of the indices of this GEP are
1076 /// zeros. If so, the result pointer and the first operand have the same
1077 /// value, just potentially different types.
1078 bool hasAllZeroIndices() const;
1079
1080 /// Return true if all of the indices of this GEP are
1081 /// constant integers. If so, the result pointer and the first operand have
1082 /// a constant offset between them.
1083 bool hasAllConstantIndices() const;
1084
1085 /// Set or clear the inbounds flag on this GEP instruction.
1086 /// See LangRef.html for the meaning of inbounds on a getelementptr.
1087 void setIsInBounds(bool b = true);
1088
1089 /// Determine whether the GEP has the inbounds flag.
1090 bool isInBounds() const;
1091
1092 /// Accumulate the constant address offset of this GEP if possible.
1093 ///
1094 /// This routine accepts an APInt into which it will accumulate the constant
1095 /// offset of this GEP if the GEP is in fact constant. If the GEP is not
1096 /// all-constant, it returns false and the value of the offset APInt is
1097 /// undefined (it is *not* preserved!). The APInt passed into this routine
1098 /// must be at least as wide as the IntPtr type for the address space of
1099 /// the base GEP pointer.
1100 bool accumulateConstantOffset(const DataLayout &DL, APInt &Offset) const;
1101
1102 // Methods for support type inquiry through isa, cast, and dyn_cast:
1103 static bool classof(const Instruction *I) {
1104 return (I->getOpcode() == Instruction::GetElementPtr);
1105 }
1106 static bool classof(const Value *V) {
1107 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1108 }
1109};
1110
1111template <>
1112struct OperandTraits<GetElementPtrInst> :
1113 public VariadicOperandTraits<GetElementPtrInst, 1> {
1114};
1115
1116GetElementPtrInst::GetElementPtrInst(Type *PointeeType, Value *Ptr,
1117 ArrayRef<Value *> IdxList, unsigned Values,
1118 const Twine &NameStr,
1119 Instruction *InsertBefore)
1120 : Instruction(getGEPReturnType(PointeeType, Ptr, IdxList), GetElementPtr,
1121 OperandTraits<GetElementPtrInst>::op_end(this) - Values,
1122 Values, InsertBefore),
1123 SourceElementType(PointeeType),
1124 ResultElementType(getIndexedType(PointeeType, IdxList)) {
1125 assert(ResultElementType ==((ResultElementType == cast<PointerType>(getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("ResultElementType == cast<PointerType>(getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 1126, __PRETTY_FUNCTION__))
1126 cast<PointerType>(getType()->getScalarType())->getElementType())((ResultElementType == cast<PointerType>(getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("ResultElementType == cast<PointerType>(getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 1126, __PRETTY_FUNCTION__))
;
1127 init(Ptr, IdxList, NameStr);
1128}
1129
1130GetElementPtrInst::GetElementPtrInst(Type *PointeeType, Value *Ptr,
1131 ArrayRef<Value *> IdxList, unsigned Values,
1132 const Twine &NameStr,
1133 BasicBlock *InsertAtEnd)
1134 : Instruction(getGEPReturnType(PointeeType, Ptr, IdxList), GetElementPtr,
1135 OperandTraits<GetElementPtrInst>::op_end(this) - Values,
1136 Values, InsertAtEnd),
1137 SourceElementType(PointeeType),
1138 ResultElementType(getIndexedType(PointeeType, IdxList)) {
1139 assert(ResultElementType ==((ResultElementType == cast<PointerType>(getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("ResultElementType == cast<PointerType>(getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 1140, __PRETTY_FUNCTION__))
1140 cast<PointerType>(getType()->getScalarType())->getElementType())((ResultElementType == cast<PointerType>(getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("ResultElementType == cast<PointerType>(getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 1140, __PRETTY_FUNCTION__))
;
1141 init(Ptr, IdxList, NameStr);
1142}
1143
1144DEFINE_TRANSPARENT_OPERAND_ACCESSORS(GetElementPtrInst, Value)GetElementPtrInst::op_iterator GetElementPtrInst::op_begin() {
return OperandTraits<GetElementPtrInst>::op_begin(this
); } GetElementPtrInst::const_op_iterator GetElementPtrInst::
op_begin() const { return OperandTraits<GetElementPtrInst>
::op_begin(const_cast<GetElementPtrInst*>(this)); } GetElementPtrInst
::op_iterator GetElementPtrInst::op_end() { return OperandTraits
<GetElementPtrInst>::op_end(this); } GetElementPtrInst::
const_op_iterator GetElementPtrInst::op_end() const { return OperandTraits
<GetElementPtrInst>::op_end(const_cast<GetElementPtrInst
*>(this)); } Value *GetElementPtrInst::getOperand(unsigned
i_nocapture) const { ((i_nocapture < OperandTraits<GetElementPtrInst
>::operands(this) && "getOperand() out of range!")
? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<GetElementPtrInst>::operands(this) && \"getOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 1144, __PRETTY_FUNCTION__)); return cast_or_null<Value>
( OperandTraits<GetElementPtrInst>::op_begin(const_cast
<GetElementPtrInst*>(this))[i_nocapture].get()); } void
GetElementPtrInst::setOperand(unsigned i_nocapture, Value *Val_nocapture
) { ((i_nocapture < OperandTraits<GetElementPtrInst>
::operands(this) && "setOperand() out of range!") ? static_cast
<void> (0) : __assert_fail ("i_nocapture < OperandTraits<GetElementPtrInst>::operands(this) && \"setOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 1144, __PRETTY_FUNCTION__)); OperandTraits<GetElementPtrInst
>::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned
GetElementPtrInst::getNumOperands() const { return OperandTraits
<GetElementPtrInst>::operands(this); } template <int
Idx_nocapture> Use &GetElementPtrInst::Op() { return this
->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture
> const Use &GetElementPtrInst::Op() const { return this
->OpFrom<Idx_nocapture>(this); }
1145
1146//===----------------------------------------------------------------------===//
1147// ICmpInst Class
1148//===----------------------------------------------------------------------===//
1149
1150/// This instruction compares its operands according to the predicate given
1151/// to the constructor. It only operates on integers or pointers. The operands
1152/// must be identical types.
1153/// Represent an integer comparison operator.
1154class ICmpInst: public CmpInst {
1155 void AssertOK() {
1156 assert(isIntPredicate() &&((isIntPredicate() && "Invalid ICmp predicate value")
? static_cast<void> (0) : __assert_fail ("isIntPredicate() && \"Invalid ICmp predicate value\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 1157, __PRETTY_FUNCTION__))
1157 "Invalid ICmp predicate value")((isIntPredicate() && "Invalid ICmp predicate value")
? static_cast<void> (0) : __assert_fail ("isIntPredicate() && \"Invalid ICmp predicate value\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 1157, __PRETTY_FUNCTION__))
;
1158 assert(getOperand(0)->getType() == getOperand(1)->getType() &&((getOperand(0)->getType() == getOperand(1)->getType() &&
"Both operands to ICmp instruction are not of the same type!"
) ? static_cast<void> (0) : __assert_fail ("getOperand(0)->getType() == getOperand(1)->getType() && \"Both operands to ICmp instruction are not of the same type!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 1159, __PRETTY_FUNCTION__))
1159 "Both operands to ICmp instruction are not of the same type!")((getOperand(0)->getType() == getOperand(1)->getType() &&
"Both operands to ICmp instruction are not of the same type!"
) ? static_cast<void> (0) : __assert_fail ("getOperand(0)->getType() == getOperand(1)->getType() && \"Both operands to ICmp instruction are not of the same type!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 1159, __PRETTY_FUNCTION__))
;
1160 // Check that the operands are the right type
1161 assert((getOperand(0)->getType()->isIntOrIntVectorTy() ||(((getOperand(0)->getType()->isIntOrIntVectorTy() || getOperand
(0)->getType()->isPtrOrPtrVectorTy()) && "Invalid operand types for ICmp instruction"
) ? static_cast<void> (0) : __assert_fail ("(getOperand(0)->getType()->isIntOrIntVectorTy() || getOperand(0)->getType()->isPtrOrPtrVectorTy()) && \"Invalid operand types for ICmp instruction\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 1163, __PRETTY_FUNCTION__))
1162 getOperand(0)->getType()->isPtrOrPtrVectorTy()) &&(((getOperand(0)->getType()->isIntOrIntVectorTy() || getOperand
(0)->getType()->isPtrOrPtrVectorTy()) && "Invalid operand types for ICmp instruction"
) ? static_cast<void> (0) : __assert_fail ("(getOperand(0)->getType()->isIntOrIntVectorTy() || getOperand(0)->getType()->isPtrOrPtrVectorTy()) && \"Invalid operand types for ICmp instruction\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 1163, __PRETTY_FUNCTION__))
1163 "Invalid operand types for ICmp instruction")(((getOperand(0)->getType()->isIntOrIntVectorTy() || getOperand
(0)->getType()->isPtrOrPtrVectorTy()) && "Invalid operand types for ICmp instruction"
) ? static_cast<void> (0) : __assert_fail ("(getOperand(0)->getType()->isIntOrIntVectorTy() || getOperand(0)->getType()->isPtrOrPtrVectorTy()) && \"Invalid operand types for ICmp instruction\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 1163, __PRETTY_FUNCTION__))
;
1164 }
1165
1166protected:
1167 // Note: Instruction needs to be a friend here to call cloneImpl.
1168 friend class Instruction;
1169
1170 /// Clone an identical ICmpInst
1171 ICmpInst *cloneImpl() const;
1172
1173public:
1174 /// Constructor with insert-before-instruction semantics.
1175 ICmpInst(
1176 Instruction *InsertBefore, ///< Where to insert
1177 Predicate pred, ///< The predicate to use for the comparison
1178 Value *LHS, ///< The left-hand-side of the expression
1179 Value *RHS, ///< The right-hand-side of the expression
1180 const Twine &NameStr = "" ///< Name of the instruction
1181 ) : CmpInst(makeCmpResultType(LHS->getType()),
1182 Instruction::ICmp, pred, LHS, RHS, NameStr,
1183 InsertBefore) {
1184#ifndef NDEBUG
1185 AssertOK();
1186#endif
1187 }
1188
1189 /// Constructor with insert-at-end semantics.
1190 ICmpInst(
1191 BasicBlock &InsertAtEnd, ///< Block to insert into.
1192 Predicate pred, ///< The predicate to use for the comparison
1193 Value *LHS, ///< The left-hand-side of the expression
1194 Value *RHS, ///< The right-hand-side of the expression
1195 const Twine &NameStr = "" ///< Name of the instruction
1196 ) : CmpInst(makeCmpResultType(LHS->getType()),
1197 Instruction::ICmp, pred, LHS, RHS, NameStr,
1198 &InsertAtEnd) {
1199#ifndef NDEBUG
1200 AssertOK();
1201#endif
1202 }
1203
1204 /// Constructor with no-insertion semantics
1205 ICmpInst(
1206 Predicate pred, ///< The predicate to use for the comparison
1207 Value *LHS, ///< The left-hand-side of the expression
1208 Value *RHS, ///< The right-hand-side of the expression
1209 const Twine &NameStr = "" ///< Name of the instruction
1210 ) : CmpInst(makeCmpResultType(LHS->getType()),
1211 Instruction::ICmp, pred, LHS, RHS, NameStr) {
1212#ifndef NDEBUG
1213 AssertOK();
1214#endif
1215 }
1216
1217 /// For example, EQ->EQ, SLE->SLE, UGT->SGT, etc.
1218 /// @returns the predicate that would be the result if the operand were
1219 /// regarded as signed.
1220 /// Return the signed version of the predicate
1221 Predicate getSignedPredicate() const {
1222 return getSignedPredicate(getPredicate());
1223 }
1224
1225 /// This is a static version that you can use without an instruction.
1226 /// Return the signed version of the predicate.
1227 static Predicate getSignedPredicate(Predicate pred);
1228
1229 /// For example, EQ->EQ, SLE->ULE, UGT->UGT, etc.
1230 /// @returns the predicate that would be the result if the operand were
1231 /// regarded as unsigned.
1232 /// Return the unsigned version of the predicate
1233 Predicate getUnsignedPredicate() const {
1234 return getUnsignedPredicate(getPredicate());
1235 }
1236
1237 /// This is a static version that you can use without an instruction.
1238 /// Return the unsigned version of the predicate.
1239 static Predicate getUnsignedPredicate(Predicate pred);
1240
1241 /// Return true if this predicate is either EQ or NE. This also
1242 /// tests for commutativity.
1243 static bool isEquality(Predicate P) {
1244 return P == ICMP_EQ || P == ICMP_NE;
1245 }
1246
1247 /// Return true if this predicate is either EQ or NE. This also
1248 /// tests for commutativity.
1249 bool isEquality() const {
1250 return isEquality(getPredicate());
1251 }
1252
1253 /// @returns true if the predicate of this ICmpInst is commutative
1254 /// Determine if this relation is commutative.
1255 bool isCommutative() const { return isEquality(); }
1256
1257 /// Return true if the predicate is relational (not EQ or NE).
1258 ///
1259 bool isRelational() const {
1260 return !isEquality();
1261 }
1262
1263 /// Return true if the predicate is relational (not EQ or NE).
1264 ///
1265 static bool isRelational(Predicate P) {
1266 return !isEquality(P);
1267 }
1268
1269 /// Exchange the two operands to this instruction in such a way that it does
1270 /// not modify the semantics of the instruction. The predicate value may be
1271 /// changed to retain the same result if the predicate is order dependent
1272 /// (e.g. ult).
1273 /// Swap operands and adjust predicate.
1274 void swapOperands() {
1275 setPredicate(getSwappedPredicate());
1276 Op<0>().swap(Op<1>());
1277 }
1278
1279 // Methods for support type inquiry through isa, cast, and dyn_cast:
1280 static bool classof(const Instruction *I) {
1281 return I->getOpcode() == Instruction::ICmp;
1282 }
1283 static bool classof(const Value *V) {
1284 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1285 }
1286};
1287
1288//===----------------------------------------------------------------------===//
1289// FCmpInst Class
1290//===----------------------------------------------------------------------===//
1291
1292/// This instruction compares its operands according to the predicate given
1293/// to the constructor. It only operates on floating point values or packed
1294/// vectors of floating point values. The operands must be identical types.
1295/// Represents a floating point comparison operator.
1296class FCmpInst: public CmpInst {
1297 void AssertOK() {
1298 assert(isFPPredicate() && "Invalid FCmp predicate value")((isFPPredicate() && "Invalid FCmp predicate value") ?
static_cast<void> (0) : __assert_fail ("isFPPredicate() && \"Invalid FCmp predicate value\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 1298, __PRETTY_FUNCTION__))
;
1299 assert(getOperand(0)->getType() == getOperand(1)->getType() &&((getOperand(0)->getType() == getOperand(1)->getType() &&
"Both operands to FCmp instruction are not of the same type!"
) ? static_cast<void> (0) : __assert_fail ("getOperand(0)->getType() == getOperand(1)->getType() && \"Both operands to FCmp instruction are not of the same type!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 1300, __PRETTY_FUNCTION__))
1300 "Both operands to FCmp instruction are not of the same type!")((getOperand(0)->getType() == getOperand(1)->getType() &&
"Both operands to FCmp instruction are not of the same type!"
) ? static_cast<void> (0) : __assert_fail ("getOperand(0)->getType() == getOperand(1)->getType() && \"Both operands to FCmp instruction are not of the same type!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 1300, __PRETTY_FUNCTION__))
;
1301 // Check that the operands are the right type
1302 assert(getOperand(0)->getType()->isFPOrFPVectorTy() &&((getOperand(0)->getType()->isFPOrFPVectorTy() &&
"Invalid operand types for FCmp instruction") ? static_cast<
void> (0) : __assert_fail ("getOperand(0)->getType()->isFPOrFPVectorTy() && \"Invalid operand types for FCmp instruction\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 1303, __PRETTY_FUNCTION__))
1303 "Invalid operand types for FCmp instruction")((getOperand(0)->getType()->isFPOrFPVectorTy() &&
"Invalid operand types for FCmp instruction") ? static_cast<
void> (0) : __assert_fail ("getOperand(0)->getType()->isFPOrFPVectorTy() && \"Invalid operand types for FCmp instruction\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 1303, __PRETTY_FUNCTION__))
;
1304 }
1305
1306protected:
1307 // Note: Instruction needs to be a friend here to call cloneImpl.
1308 friend class Instruction;
1309
1310 /// Clone an identical FCmpInst
1311 FCmpInst *cloneImpl() const;
1312
1313public:
1314 /// Constructor with insert-before-instruction semantics.
1315 FCmpInst(
1316 Instruction *InsertBefore, ///< Where to insert
1317 Predicate pred, ///< The predicate to use for the comparison
1318 Value *LHS, ///< The left-hand-side of the expression
1319 Value *RHS, ///< The right-hand-side of the expression
1320 const Twine &NameStr = "" ///< Name of the instruction
1321 ) : CmpInst(makeCmpResultType(LHS->getType()),
1322 Instruction::FCmp, pred, LHS, RHS, NameStr,
1323 InsertBefore) {
1324 AssertOK();
1325 }
1326
1327 /// Constructor with insert-at-end semantics.
1328 FCmpInst(
1329 BasicBlock &InsertAtEnd, ///< Block to insert into.
1330 Predicate pred, ///< The predicate to use for the comparison
1331 Value *LHS, ///< The left-hand-side of the expression
1332 Value *RHS, ///< The right-hand-side of the expression
1333 const Twine &NameStr = "" ///< Name of the instruction
1334 ) : CmpInst(makeCmpResultType(LHS->getType()),
1335 Instruction::FCmp, pred, LHS, RHS, NameStr,
1336 &InsertAtEnd) {
1337 AssertOK();
1338 }
1339
1340 /// Constructor with no-insertion semantics
1341 FCmpInst(
1342 Predicate Pred, ///< The predicate to use for the comparison
1343 Value *LHS, ///< The left-hand-side of the expression
1344 Value *RHS, ///< The right-hand-side of the expression
1345 const Twine &NameStr = "", ///< Name of the instruction
1346 Instruction *FlagsSource = nullptr
1347 ) : CmpInst(makeCmpResultType(LHS->getType()), Instruction::FCmp, Pred, LHS,
1348 RHS, NameStr, nullptr, FlagsSource) {
1349 AssertOK();
1350 }
1351
1352 /// @returns true if the predicate of this instruction is EQ or NE.
1353 /// Determine if this is an equality predicate.
1354 static bool isEquality(Predicate Pred) {
1355 return Pred == FCMP_OEQ || Pred == FCMP_ONE || Pred == FCMP_UEQ ||
1356 Pred == FCMP_UNE;
1357 }
1358
1359 /// @returns true if the predicate of this instruction is EQ or NE.
1360 /// Determine if this is an equality predicate.
1361 bool isEquality() const { return isEquality(getPredicate()); }
1362
1363 /// @returns true if the predicate of this instruction is commutative.
1364 /// Determine if this is a commutative predicate.
1365 bool isCommutative() const {
1366 return isEquality() ||
1367 getPredicate() == FCMP_FALSE ||
1368 getPredicate() == FCMP_TRUE ||
1369 getPredicate() == FCMP_ORD ||
1370 getPredicate() == FCMP_UNO;
1371 }
1372
1373 /// @returns true if the predicate is relational (not EQ or NE).
1374 /// Determine if this a relational predicate.
1375 bool isRelational() const { return !isEquality(); }
1376
1377 /// Exchange the two operands to this instruction in such a way that it does
1378 /// not modify the semantics of the instruction. The predicate value may be
1379 /// changed to retain the same result if the predicate is order dependent
1380 /// (e.g. ult).
1381 /// Swap operands and adjust predicate.
1382 void swapOperands() {
1383 setPredicate(getSwappedPredicate());
1384 Op<0>().swap(Op<1>());
1385 }
1386
1387 /// Methods for support type inquiry through isa, cast, and dyn_cast:
1388 static bool classof(const Instruction *I) {
1389 return I->getOpcode() == Instruction::FCmp;
1390 }
1391 static bool classof(const Value *V) {
1392 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1393 }
1394};
1395
1396//===----------------------------------------------------------------------===//
1397/// This class represents a function call, abstracting a target
1398/// machine's calling convention. This class uses low bit of the SubClassData
1399/// field to indicate whether or not this is a tail call. The rest of the bits
1400/// hold the calling convention of the call.
1401///
1402class CallInst : public CallBase {
1403 CallInst(const CallInst &CI);
1404
1405 /// Construct a CallInst given a range of arguments.
1406 /// Construct a CallInst from a range of arguments
1407 inline CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1408 ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr,
1409 Instruction *InsertBefore);
1410
1411 inline CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1412 const Twine &NameStr, Instruction *InsertBefore)
1413 : CallInst(Ty, Func, Args, None, NameStr, InsertBefore) {}
1414
1415 /// Construct a CallInst given a range of arguments.
1416 /// Construct a CallInst from a range of arguments
1417 inline CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1418 ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr,
1419 BasicBlock *InsertAtEnd);
1420
1421 explicit CallInst(FunctionType *Ty, Value *F, const Twine &NameStr,
1422 Instruction *InsertBefore);
1423
1424 CallInst(FunctionType *ty, Value *F, const Twine &NameStr,
1425 BasicBlock *InsertAtEnd);
1426
1427 void init(FunctionType *FTy, Value *Func, ArrayRef<Value *> Args,
1428 ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr);
1429 void init(FunctionType *FTy, Value *Func, const Twine &NameStr);
1430
1431 /// Compute the number of operands to allocate.
1432 static int ComputeNumOperands(int NumArgs, int NumBundleInputs = 0) {
1433 // We need one operand for the called function, plus the input operand
1434 // counts provided.
1435 return 1 + NumArgs + NumBundleInputs;
1436 }
1437
1438protected:
1439 // Note: Instruction needs to be a friend here to call cloneImpl.
1440 friend class Instruction;
1441
1442 CallInst *cloneImpl() const;
1443
1444public:
1445 static CallInst *Create(FunctionType *Ty, Value *F, const Twine &NameStr = "",
1446 Instruction *InsertBefore = nullptr) {
1447 return new (ComputeNumOperands(0)) CallInst(Ty, F, NameStr, InsertBefore);
1448 }
1449
1450 static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1451 const Twine &NameStr,
1452 Instruction *InsertBefore = nullptr) {
1453 return new (ComputeNumOperands(Args.size()))
1454 CallInst(Ty, Func, Args, None, NameStr, InsertBefore);
1455 }
1456
1457 static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1458 ArrayRef<OperandBundleDef> Bundles = None,
1459 const Twine &NameStr = "",
1460 Instruction *InsertBefore = nullptr) {
1461 const int NumOperands =
1462 ComputeNumOperands(Args.size(), CountBundleInputs(Bundles));
1463 const unsigned DescriptorBytes = Bundles.size() * sizeof(BundleOpInfo);
1464
1465 return new (NumOperands, DescriptorBytes)
1466 CallInst(Ty, Func, Args, Bundles, NameStr, InsertBefore);
1467 }
1468
1469 static CallInst *Create(FunctionType *Ty, Value *F, const Twine &NameStr,
1470 BasicBlock *InsertAtEnd) {
1471 return new (ComputeNumOperands(0)) CallInst(Ty, F, NameStr, InsertAtEnd);
1472 }
1473
1474 static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1475 const Twine &NameStr, BasicBlock *InsertAtEnd) {
1476 return new (ComputeNumOperands(Args.size()))
1477 CallInst(Ty, Func, Args, None, NameStr, InsertAtEnd);
1478 }
1479
1480 static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1481 ArrayRef<OperandBundleDef> Bundles,
1482 const Twine &NameStr, BasicBlock *InsertAtEnd) {
1483 const int NumOperands =
1484 ComputeNumOperands(Args.size(), CountBundleInputs(Bundles));
1485 const unsigned DescriptorBytes = Bundles.size() * sizeof(BundleOpInfo);
1486
1487 return new (NumOperands, DescriptorBytes)
1488 CallInst(Ty, Func, Args, Bundles, NameStr, InsertAtEnd);
1489 }
1490
1491 static CallInst *Create(FunctionCallee Func, const Twine &NameStr = "",
1492 Instruction *InsertBefore = nullptr) {
1493 return Create(Func.getFunctionType(), Func.getCallee(), NameStr,
1494 InsertBefore);
1495 }
1496
1497 static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args,
1498 ArrayRef<OperandBundleDef> Bundles = None,
1499 const Twine &NameStr = "",
1500 Instruction *InsertBefore = nullptr) {
1501 return Create(Func.getFunctionType(), Func.getCallee(), Args, Bundles,
1502 NameStr, InsertBefore);
1503 }
1504
1505 static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args,
1506 const Twine &NameStr,
1507 Instruction *InsertBefore = nullptr) {
1508 return Create(Func.getFunctionType(), Func.getCallee(), Args, NameStr,
1509 InsertBefore);
1510 }
1511
1512 static CallInst *Create(FunctionCallee Func, const Twine &NameStr,
1513 BasicBlock *InsertAtEnd) {
1514 return Create(Func.getFunctionType(), Func.getCallee(), NameStr,
1515 InsertAtEnd);
1516 }
1517
1518 static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args,
1519 const Twine &NameStr, BasicBlock *InsertAtEnd) {
1520 return Create(Func.getFunctionType(), Func.getCallee(), Args, NameStr,
1521 InsertAtEnd);
1522 }
1523
1524 static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args,
1525 ArrayRef<OperandBundleDef> Bundles,
1526 const Twine &NameStr, BasicBlock *InsertAtEnd) {
1527 return Create(Func.getFunctionType(), Func.getCallee(), Args, Bundles,
1528 NameStr, InsertAtEnd);
1529 }
1530
1531 // Deprecated [opaque pointer types]
1532 static CallInst *Create(Value *Func, const Twine &NameStr = "",
1533 Instruction *InsertBefore = nullptr) {
1534 return Create(cast<FunctionType>(
1535 cast<PointerType>(Func->getType())->getElementType()),
1536 Func, NameStr, InsertBefore);
1537 }
1538
1539 // Deprecated [opaque pointer types]
1540 static CallInst *Create(Value *Func, ArrayRef<Value *> Args,
1541 const Twine &NameStr,
1542 Instruction *InsertBefore = nullptr) {
1543 return Create(cast<FunctionType>(
1544 cast<PointerType>(Func->getType())->getElementType()),
1545 Func, Args, NameStr, InsertBefore);
1546 }
1547
1548 // Deprecated [opaque pointer types]
1549 static CallInst *Create(Value *Func, ArrayRef<Value *> Args,
1550 ArrayRef<OperandBundleDef> Bundles = None,
1551 const Twine &NameStr = "",
1552 Instruction *InsertBefore = nullptr) {
1553 return Create(cast<FunctionType>(
1554 cast<PointerType>(Func->getType())->getElementType()),
1555 Func, Args, Bundles, NameStr, InsertBefore);
1556 }
1557
1558 // Deprecated [opaque pointer types]
1559 static CallInst *Create(Value *Func, const Twine &NameStr,
1560 BasicBlock *InsertAtEnd) {
1561 return Create(cast<FunctionType>(
1562 cast<PointerType>(Func->getType())->getElementType()),
1563 Func, NameStr, InsertAtEnd);
1564 }
1565
1566 // Deprecated [opaque pointer types]
1567 static CallInst *Create(Value *Func, ArrayRef<Value *> Args,
1568 const Twine &NameStr, BasicBlock *InsertAtEnd) {
1569 return Create(cast<FunctionType>(
1570 cast<PointerType>(Func->getType())->getElementType()),
1571 Func, Args, NameStr, InsertAtEnd);
1572 }
1573
1574 // Deprecated [opaque pointer types]
1575 static CallInst *Create(Value *Func, ArrayRef<Value *> Args,
1576 ArrayRef<OperandBundleDef> Bundles,
1577 const Twine &NameStr, BasicBlock *InsertAtEnd) {
1578 return Create(cast<FunctionType>(
1579 cast<PointerType>(Func->getType())->getElementType()),
1580 Func, Args, Bundles, NameStr, InsertAtEnd);
1581 }
1582
1583 /// Create a clone of \p CI with a different set of operand bundles and
1584 /// insert it before \p InsertPt.
1585 ///
1586 /// The returned call instruction is identical \p CI in every way except that
1587 /// the operand bundles for the new instruction are set to the operand bundles
1588 /// in \p Bundles.
1589 static CallInst *Create(CallInst *CI, ArrayRef<OperandBundleDef> Bundles,
1590 Instruction *InsertPt = nullptr);
1591
1592 /// Generate the IR for a call to malloc:
1593 /// 1. Compute the malloc call's argument as the specified type's size,
1594 /// possibly multiplied by the array size if the array size is not
1595 /// constant 1.
1596 /// 2. Call malloc with that argument.
1597 /// 3. Bitcast the result of the malloc call to the specified type.
1598 static Instruction *CreateMalloc(Instruction *InsertBefore, Type *IntPtrTy,
1599 Type *AllocTy, Value *AllocSize,
1600 Value *ArraySize = nullptr,
1601 Function *MallocF = nullptr,
1602 const Twine &Name = "");
1603 static Instruction *CreateMalloc(BasicBlock *InsertAtEnd, Type *IntPtrTy,
1604 Type *AllocTy, Value *AllocSize,
1605 Value *ArraySize = nullptr,
1606 Function *MallocF = nullptr,
1607 const Twine &Name = "");
1608 static Instruction *CreateMalloc(Instruction *InsertBefore, Type *IntPtrTy,
1609 Type *AllocTy, Value *AllocSize,
1610 Value *ArraySize = nullptr,
1611 ArrayRef<OperandBundleDef> Bundles = None,
1612 Function *MallocF = nullptr,
1613 const Twine &Name = "");
1614 static Instruction *CreateMalloc(BasicBlock *InsertAtEnd, Type *IntPtrTy,
1615 Type *AllocTy, Value *AllocSize,
1616 Value *ArraySize = nullptr,
1617 ArrayRef<OperandBundleDef> Bundles = None,
1618 Function *MallocF = nullptr,
1619 const Twine &Name = "");
1620 /// Generate the IR for a call to the builtin free function.
1621 static Instruction *CreateFree(Value *Source, Instruction *InsertBefore);
1622 static Instruction *CreateFree(Value *Source, BasicBlock *InsertAtEnd);
1623 static Instruction *CreateFree(Value *Source,
1624 ArrayRef<OperandBundleDef> Bundles,
1625 Instruction *InsertBefore);
1626 static Instruction *CreateFree(Value *Source,
1627 ArrayRef<OperandBundleDef> Bundles,
1628 BasicBlock *InsertAtEnd);
1629
1630 // Note that 'musttail' implies 'tail'.
1631 enum TailCallKind {
1632 TCK_None = 0,
1633 TCK_Tail = 1,
1634 TCK_MustTail = 2,
1635 TCK_NoTail = 3
1636 };
1637 TailCallKind getTailCallKind() const {
1638 return TailCallKind(getSubclassDataFromInstruction() & 3);
1639 }
1640
1641 bool isTailCall() const {
1642 unsigned Kind = getSubclassDataFromInstruction() & 3;
1643 return Kind == TCK_Tail || Kind == TCK_MustTail;
1644 }
1645
1646 bool isMustTailCall() const {
1647 return (getSubclassDataFromInstruction() & 3) == TCK_MustTail;
1648 }
1649
1650 bool isNoTailCall() const {
1651 return (getSubclassDataFromInstruction() & 3) == TCK_NoTail;
1652 }
1653
1654 void setTailCall(bool isTC = true) {
1655 setInstructionSubclassData((getSubclassDataFromInstruction() & ~3) |
1656 unsigned(isTC ? TCK_Tail : TCK_None));
1657 }
1658
1659 void setTailCallKind(TailCallKind TCK) {
1660 setInstructionSubclassData((getSubclassDataFromInstruction() & ~3) |
1661 unsigned(TCK));
1662 }
1663
1664 /// Return true if the call can return twice
1665 bool canReturnTwice() const { return hasFnAttr(Attribute::ReturnsTwice); }
1666 void setCanReturnTwice() {
1667 addAttribute(AttributeList::FunctionIndex, Attribute::ReturnsTwice);
1668 }
1669
1670 // Methods for support type inquiry through isa, cast, and dyn_cast:
1671 static bool classof(const Instruction *I) {
1672 return I->getOpcode() == Instruction::Call;
1673 }
1674 static bool classof(const Value *V) {
1675 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1676 }
1677
1678 /// Updates profile metadata by scaling it by \p S / \p T.
1679 void updateProfWeight(uint64_t S, uint64_t T);
1680
1681private:
1682 // Shadow Instruction::setInstructionSubclassData with a private forwarding
1683 // method so that subclasses cannot accidentally use it.
1684 void setInstructionSubclassData(unsigned short D) {
1685 Instruction::setInstructionSubclassData(D);
1686 }
1687};
1688
1689CallInst::CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1690 ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr,
1691 BasicBlock *InsertAtEnd)
1692 : CallBase(Ty->getReturnType(), Instruction::Call,
1693 OperandTraits<CallBase>::op_end(this) -
1694 (Args.size() + CountBundleInputs(Bundles) + 1),
1695 unsigned(Args.size() + CountBundleInputs(Bundles) + 1),
1696 InsertAtEnd) {
1697 init(Ty, Func, Args, Bundles, NameStr);
1698}
1699
1700CallInst::CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1701 ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr,
1702 Instruction *InsertBefore)
1703 : CallBase(Ty->getReturnType(), Instruction::Call,
1704 OperandTraits<CallBase>::op_end(this) -
1705 (Args.size() + CountBundleInputs(Bundles) + 1),
1706 unsigned(Args.size() + CountBundleInputs(Bundles) + 1),
1707 InsertBefore) {
1708 init(Ty, Func, Args, Bundles, NameStr);
1709}
1710
1711//===----------------------------------------------------------------------===//
1712// SelectInst Class
1713//===----------------------------------------------------------------------===//
1714
1715/// This class represents the LLVM 'select' instruction.
1716///
1717class SelectInst : public Instruction {
1718 SelectInst(Value *C, Value *S1, Value *S2, const Twine &NameStr,
1719 Instruction *InsertBefore)
1720 : Instruction(S1->getType(), Instruction::Select,
1721 &Op<0>(), 3, InsertBefore) {
1722 init(C, S1, S2);
1723 setName(NameStr);
1724 }
1725
1726 SelectInst(Value *C, Value *S1, Value *S2, const Twine &NameStr,
1727 BasicBlock *InsertAtEnd)
1728 : Instruction(S1->getType(), Instruction::Select,
1729 &Op<0>(), 3, InsertAtEnd) {
1730 init(C, S1, S2);
1731 setName(NameStr);
1732 }
1733
1734 void init(Value *C, Value *S1, Value *S2) {
1735 assert(!areInvalidOperands(C, S1, S2) && "Invalid operands for select")((!areInvalidOperands(C, S1, S2) && "Invalid operands for select"
) ? static_cast<void> (0) : __assert_fail ("!areInvalidOperands(C, S1, S2) && \"Invalid operands for select\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 1735, __PRETTY_FUNCTION__))
;
1736 Op<0>() = C;
1737 Op<1>() = S1;
1738 Op<2>() = S2;
1739 }
1740
1741protected:
1742 // Note: Instruction needs to be a friend here to call cloneImpl.
1743 friend class Instruction;
1744
1745 SelectInst *cloneImpl() const;
1746
1747public:
1748 static SelectInst *Create(Value *C, Value *S1, Value *S2,
1749 const Twine &NameStr = "",
1750 Instruction *InsertBefore = nullptr,
1751 Instruction *MDFrom = nullptr) {
1752 SelectInst *Sel = new(3) SelectInst(C, S1, S2, NameStr, InsertBefore);
1753 if (MDFrom)
1754 Sel->copyMetadata(*MDFrom);
1755 return Sel;
1756 }
1757
1758 static SelectInst *Create(Value *C, Value *S1, Value *S2,
1759 const Twine &NameStr,
1760 BasicBlock *InsertAtEnd) {
1761 return new(3) SelectInst(C, S1, S2, NameStr, InsertAtEnd);
1762 }
1763
1764 const Value *getCondition() const { return Op<0>(); }
1765 const Value *getTrueValue() const { return Op<1>(); }
1766 const Value *getFalseValue() const { return Op<2>(); }
1767 Value *getCondition() { return Op<0>(); }
1768 Value *getTrueValue() { return Op<1>(); }
1769 Value *getFalseValue() { return Op<2>(); }
1770
1771 void setCondition(Value *V) { Op<0>() = V; }
1772 void setTrueValue(Value *V) { Op<1>() = V; }
1773 void setFalseValue(Value *V) { Op<2>() = V; }
1774
1775 /// Swap the true and false values of the select instruction.
1776 /// This doesn't swap prof metadata.
1777 void swapValues() { Op<1>().swap(Op<2>()); }
1778
1779 /// Return a string if the specified operands are invalid
1780 /// for a select operation, otherwise return null.
1781 static const char *areInvalidOperands(Value *Cond, Value *True, Value *False);
1782
1783 /// Transparently provide more efficient getOperand methods.
1784 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
;
1785
1786 OtherOps getOpcode() const {
1787 return static_cast<OtherOps>(Instruction::getOpcode());
1788 }
1789
1790 // Methods for support type inquiry through isa, cast, and dyn_cast:
1791 static bool classof(const Instruction *I) {
1792 return I->getOpcode() == Instruction::Select;
1793 }
1794 static bool classof(const Value *V) {
1795 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1796 }
1797};
1798
1799template <>
1800struct OperandTraits<SelectInst> : public FixedNumOperandTraits<SelectInst, 3> {
1801};
1802
1803DEFINE_TRANSPARENT_OPERAND_ACCESSORS(SelectInst, Value)SelectInst::op_iterator SelectInst::op_begin() { return OperandTraits
<SelectInst>::op_begin(this); } SelectInst::const_op_iterator
SelectInst::op_begin() const { return OperandTraits<SelectInst
>::op_begin(const_cast<SelectInst*>(this)); } SelectInst
::op_iterator SelectInst::op_end() { return OperandTraits<
SelectInst>::op_end(this); } SelectInst::const_op_iterator
SelectInst::op_end() const { return OperandTraits<SelectInst
>::op_end(const_cast<SelectInst*>(this)); } Value *SelectInst
::getOperand(unsigned i_nocapture) const { ((i_nocapture <
OperandTraits<SelectInst>::operands(this) && "getOperand() out of range!"
) ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<SelectInst>::operands(this) && \"getOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 1803, __PRETTY_FUNCTION__)); return cast_or_null<Value>
( OperandTraits<SelectInst>::op_begin(const_cast<SelectInst
*>(this))[i_nocapture].get()); } void SelectInst::setOperand
(unsigned i_nocapture, Value *Val_nocapture) { ((i_nocapture <
OperandTraits<SelectInst>::operands(this) && "setOperand() out of range!"
) ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<SelectInst>::operands(this) && \"setOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 1803, __PRETTY_FUNCTION__)); OperandTraits<SelectInst>
::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned SelectInst
::getNumOperands() const { return OperandTraits<SelectInst
>::operands(this); } template <int Idx_nocapture> Use
&SelectInst::Op() { return this->OpFrom<Idx_nocapture
>(this); } template <int Idx_nocapture> const Use &
SelectInst::Op() const { return this->OpFrom<Idx_nocapture
>(this); }
1804
1805//===----------------------------------------------------------------------===//
1806// VAArgInst Class
1807//===----------------------------------------------------------------------===//
1808
1809/// This class represents the va_arg llvm instruction, which returns
1810/// an argument of the specified type given a va_list and increments that list
1811///
1812class VAArgInst : public UnaryInstruction {
1813protected:
1814 // Note: Instruction needs to be a friend here to call cloneImpl.
1815 friend class Instruction;
1816
1817 VAArgInst *cloneImpl() const;
1818
1819public:
1820 VAArgInst(Value *List, Type *Ty, const Twine &NameStr = "",
1821 Instruction *InsertBefore = nullptr)
1822 : UnaryInstruction(Ty, VAArg, List, InsertBefore) {
1823 setName(NameStr);
1824 }
1825
1826 VAArgInst(Value *List, Type *Ty, const Twine &NameStr,
1827 BasicBlock *InsertAtEnd)
1828 : UnaryInstruction(Ty, VAArg, List, InsertAtEnd) {
1829 setName(NameStr);
1830 }
1831
1832 Value *getPointerOperand() { return getOperand(0); }
1833 const Value *getPointerOperand() const { return getOperand(0); }
1834 static unsigned getPointerOperandIndex() { return 0U; }
1835
1836 // Methods for support type inquiry through isa, cast, and dyn_cast:
1837 static bool classof(const Instruction *I) {
1838 return I->getOpcode() == VAArg;
1839 }
1840 static bool classof(const Value *V) {
1841 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1842 }
1843};
1844
1845//===----------------------------------------------------------------------===//
1846// ExtractElementInst Class
1847//===----------------------------------------------------------------------===//
1848
1849/// This instruction extracts a single (scalar)
1850/// element from a VectorType value
1851///
1852class ExtractElementInst : public Instruction {
1853 ExtractElementInst(Value *Vec, Value *Idx, const Twine &NameStr = "",
1854 Instruction *InsertBefore = nullptr);
1855 ExtractElementInst(Value *Vec, Value *Idx, const Twine &NameStr,
1856 BasicBlock *InsertAtEnd);
1857
1858protected:
1859 // Note: Instruction needs to be a friend here to call cloneImpl.
1860 friend class Instruction;
1861
1862 ExtractElementInst *cloneImpl() const;
1863
1864public:
1865 static ExtractElementInst *Create(Value *Vec, Value *Idx,
1866 const Twine &NameStr = "",
1867 Instruction *InsertBefore = nullptr) {
1868 return new(2) ExtractElementInst(Vec, Idx, NameStr, InsertBefore);
1869 }
1870
1871 static ExtractElementInst *Create(Value *Vec, Value *Idx,
1872 const Twine &NameStr,
1873 BasicBlock *InsertAtEnd) {
1874 return new(2) ExtractElementInst(Vec, Idx, NameStr, InsertAtEnd);
1875 }
1876
1877 /// Return true if an extractelement instruction can be
1878 /// formed with the specified operands.
1879 static bool isValidOperands(const Value *Vec, const Value *Idx);
1880
1881 Value *getVectorOperand() { return Op<0>(); }
1882 Value *getIndexOperand() { return Op<1>(); }
1883 const Value *getVectorOperand() const { return Op<0>(); }
1884 const Value *getIndexOperand() const { return Op<1>(); }
1885
1886 VectorType *getVectorOperandType() const {
1887 return cast<VectorType>(getVectorOperand()->getType());
1888 }
1889
1890 /// Transparently provide more efficient getOperand methods.
1891 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
;
1892
1893 // Methods for support type inquiry through isa, cast, and dyn_cast:
1894 static bool classof(const Instruction *I) {
1895 return I->getOpcode() == Instruction::ExtractElement;
1896 }
1897 static bool classof(const Value *V) {
1898 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1899 }
1900};
1901
1902template <>
1903struct OperandTraits<ExtractElementInst> :
1904 public FixedNumOperandTraits<ExtractElementInst, 2> {
1905};
1906
1907DEFINE_TRANSPARENT_OPERAND_ACCESSORS(ExtractElementInst, Value)ExtractElementInst::op_iterator ExtractElementInst::op_begin(
) { return OperandTraits<ExtractElementInst>::op_begin(
this); } ExtractElementInst::const_op_iterator ExtractElementInst
::op_begin() const { return OperandTraits<ExtractElementInst
>::op_begin(const_cast<ExtractElementInst*>(this)); }
ExtractElementInst::op_iterator ExtractElementInst::op_end()
{ return OperandTraits<ExtractElementInst>::op_end(this
); } ExtractElementInst::const_op_iterator ExtractElementInst
::op_end() const { return OperandTraits<ExtractElementInst
>::op_end(const_cast<ExtractElementInst*>(this)); } Value
*ExtractElementInst::getOperand(unsigned i_nocapture) const {
((i_nocapture < OperandTraits<ExtractElementInst>::
operands(this) && "getOperand() out of range!") ? static_cast
<void> (0) : __assert_fail ("i_nocapture < OperandTraits<ExtractElementInst>::operands(this) && \"getOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 1907, __PRETTY_FUNCTION__)); return cast_or_null<Value>
( OperandTraits<ExtractElementInst>::op_begin(const_cast
<ExtractElementInst*>(this))[i_nocapture].get()); } void
ExtractElementInst::setOperand(unsigned i_nocapture, Value *
Val_nocapture) { ((i_nocapture < OperandTraits<ExtractElementInst
>::operands(this) && "setOperand() out of range!")
? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<ExtractElementInst>::operands(this) && \"setOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 1907, __PRETTY_FUNCTION__)); OperandTraits<ExtractElementInst
>::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned
ExtractElementInst::getNumOperands() const { return OperandTraits
<ExtractElementInst>::operands(this); } template <int
Idx_nocapture> Use &ExtractElementInst::Op() { return
this->OpFrom<Idx_nocapture>(this); } template <int
Idx_nocapture> const Use &ExtractElementInst::Op() const
{ return this->OpFrom<Idx_nocapture>(this); }
1908
1909//===----------------------------------------------------------------------===//
1910// InsertElementInst Class
1911//===----------------------------------------------------------------------===//
1912
1913/// This instruction inserts a single (scalar)
1914/// element into a VectorType value
1915///
1916class InsertElementInst : public Instruction {
1917 InsertElementInst(Value *Vec, Value *NewElt, Value *Idx,
1918 const Twine &NameStr = "",
1919 Instruction *InsertBefore = nullptr);
1920 InsertElementInst(Value *Vec, Value *NewElt, Value *Idx, const Twine &NameStr,
1921 BasicBlock *InsertAtEnd);
1922
1923protected:
1924 // Note: Instruction needs to be a friend here to call cloneImpl.
1925 friend class Instruction;
1926
1927 InsertElementInst *cloneImpl() const;
1928
1929public:
1930 static InsertElementInst *Create(Value *Vec, Value *NewElt, Value *Idx,
1931 const Twine &NameStr = "",
1932 Instruction *InsertBefore = nullptr) {
1933 return new(3) InsertElementInst(Vec, NewElt, Idx, NameStr, InsertBefore);
1934 }
1935
1936 static InsertElementInst *Create(Value *Vec, Value *NewElt, Value *Idx,
1937 const Twine &NameStr,
1938 BasicBlock *InsertAtEnd) {
1939 return new(3) InsertElementInst(Vec, NewElt, Idx, NameStr, InsertAtEnd);
1940 }
1941
1942 /// Return true if an insertelement instruction can be
1943 /// formed with the specified operands.
1944 static bool isValidOperands(const Value *Vec, const Value *NewElt,
1945 const Value *Idx);
1946
1947 /// Overload to return most specific vector type.
1948 ///
1949 VectorType *getType() const {
1950 return cast<VectorType>(Instruction::getType());
1951 }
1952
1953 /// Transparently provide more efficient getOperand methods.
1954 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
;
1955
1956 // Methods for support type inquiry through isa, cast, and dyn_cast:
1957 static bool classof(const Instruction *I) {
1958 return I->getOpcode() == Instruction::InsertElement;
1959 }
1960 static bool classof(const Value *V) {
1961 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1962 }
1963};
1964
1965template <>
1966struct OperandTraits<InsertElementInst> :
1967 public FixedNumOperandTraits<InsertElementInst, 3> {
1968};
1969
1970DEFINE_TRANSPARENT_OPERAND_ACCESSORS(InsertElementInst, Value)InsertElementInst::op_iterator InsertElementInst::op_begin() {
return OperandTraits<InsertElementInst>::op_begin(this
); } InsertElementInst::const_op_iterator InsertElementInst::
op_begin() const { return OperandTraits<InsertElementInst>
::op_begin(const_cast<InsertElementInst*>(this)); } InsertElementInst
::op_iterator InsertElementInst::op_end() { return OperandTraits
<InsertElementInst>::op_end(this); } InsertElementInst::
const_op_iterator InsertElementInst::op_end() const { return OperandTraits
<InsertElementInst>::op_end(const_cast<InsertElementInst
*>(this)); } Value *InsertElementInst::getOperand(unsigned
i_nocapture) const { ((i_nocapture < OperandTraits<InsertElementInst
>::operands(this) && "getOperand() out of range!")
? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<InsertElementInst>::operands(this) && \"getOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 1970, __PRETTY_FUNCTION__)); return cast_or_null<Value>
( OperandTraits<InsertElementInst>::op_begin(const_cast
<InsertElementInst*>(this))[i_nocapture].get()); } void
InsertElementInst::setOperand(unsigned i_nocapture, Value *Val_nocapture
) { ((i_nocapture < OperandTraits<InsertElementInst>
::operands(this) && "setOperand() out of range!") ? static_cast
<void> (0) : __assert_fail ("i_nocapture < OperandTraits<InsertElementInst>::operands(this) && \"setOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 1970, __PRETTY_FUNCTION__)); OperandTraits<InsertElementInst
>::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned
InsertElementInst::getNumOperands() const { return OperandTraits
<InsertElementInst>::operands(this); } template <int
Idx_nocapture> Use &InsertElementInst::Op() { return this
->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture
> const Use &InsertElementInst::Op() const { return this
->OpFrom<Idx_nocapture>(this); }
1971
1972//===----------------------------------------------------------------------===//
1973// ShuffleVectorInst Class
1974//===----------------------------------------------------------------------===//
1975
1976/// This instruction constructs a fixed permutation of two
1977/// input vectors.
1978///
1979class ShuffleVectorInst : public Instruction {
1980protected:
1981 // Note: Instruction needs to be a friend here to call cloneImpl.
1982 friend class Instruction;
1983
1984 ShuffleVectorInst *cloneImpl() const;
1985
1986public:
1987 ShuffleVectorInst(Value *V1, Value *V2, Value *Mask,
1988 const Twine &NameStr = "",
1989 Instruction *InsertBefor = nullptr);
1990 ShuffleVectorInst(Value *V1, Value *V2, Value *Mask,
1991 const Twine &NameStr, BasicBlock *InsertAtEnd);
1992
1993 // allocate space for exactly three operands
1994 void *operator new(size_t s) {
1995 return User::operator new(s, 3);
1996 }
1997
1998 /// Swap the first 2 operands and adjust the mask to preserve the semantics
1999 /// of the instruction.
2000 void commute();
2001
2002 /// Return true if a shufflevector instruction can be
2003 /// formed with the specified operands.
2004 static bool isValidOperands(const Value *V1, const Value *V2,
2005 const Value *Mask);
2006
2007 /// Overload to return most specific vector type.
2008 ///
2009 VectorType *getType() const {
2010 return cast<VectorType>(Instruction::getType());
2011 }
2012
2013 /// Transparently provide more efficient getOperand methods.
2014 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
;
2015
2016 Constant *getMask() const {
2017 return cast<Constant>(getOperand(2));
2018 }
2019
2020 /// Return the shuffle mask value for the specified element of the mask.
2021 /// Return -1 if the element is undef.
2022 static int getMaskValue(const Constant *Mask, unsigned Elt);
2023
2024 /// Return the shuffle mask value of this instruction for the given element
2025 /// index. Return -1 if the element is undef.
2026 int getMaskValue(unsigned Elt) const {
2027 return getMaskValue(getMask(), Elt);
2028 }
2029
2030 /// Convert the input shuffle mask operand to a vector of integers. Undefined
2031 /// elements of the mask are returned as -1.
2032 static void getShuffleMask(const Constant *Mask,
2033 SmallVectorImpl<int> &Result);
2034
2035 /// Return the mask for this instruction as a vector of integers. Undefined
2036 /// elements of the mask are returned as -1.
2037 void getShuffleMask(SmallVectorImpl<int> &Result) const {
2038 return getShuffleMask(getMask(), Result);
2039 }
2040
2041 SmallVector<int, 16> getShuffleMask() const {
2042 SmallVector<int, 16> Mask;
2043 getShuffleMask(Mask);
2044 return Mask;
2045 }
2046
2047 /// Return true if this shuffle returns a vector with a different number of
2048 /// elements than its source vectors.
2049 /// Examples: shufflevector <4 x n> A, <4 x n> B, <1,2,3>
2050 /// shufflevector <4 x n> A, <4 x n> B, <1,2,3,4,5>
2051 bool changesLength() const {
2052 unsigned NumSourceElts = Op<0>()->getType()->getVectorNumElements();
2053 unsigned NumMaskElts = getMask()->getType()->getVectorNumElements();
2054 return NumSourceElts != NumMaskElts;
2055 }
2056
2057 /// Return true if this shuffle returns a vector with a greater number of
2058 /// elements than its source vectors.
2059 /// Example: shufflevector <2 x n> A, <2 x n> B, <1,2,3>
2060 bool increasesLength() const {
2061 unsigned NumSourceElts = Op<0>()->getType()->getVectorNumElements();
2062 unsigned NumMaskElts = getMask()->getType()->getVectorNumElements();
2063 return NumSourceElts < NumMaskElts;
2064 }
2065
2066 /// Return true if this shuffle mask chooses elements from exactly one source
2067 /// vector.
2068 /// Example: <7,5,undef,7>
2069 /// This assumes that vector operands are the same length as the mask.
2070 static bool isSingleSourceMask(ArrayRef<int> Mask);
2071 static bool isSingleSourceMask(const Constant *Mask) {
2072 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((Mask->getType()->isVectorTy() && "Shuffle needs vector constant."
) ? static_cast<void> (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 2072, __PRETTY_FUNCTION__))
;
2073 SmallVector<int, 16> MaskAsInts;
2074 getShuffleMask(Mask, MaskAsInts);
2075 return isSingleSourceMask(MaskAsInts);
2076 }
2077
2078 /// Return true if this shuffle chooses elements from exactly one source
2079 /// vector without changing the length of that vector.
2080 /// Example: shufflevector <4 x n> A, <4 x n> B, <3,0,undef,3>
2081 /// TODO: Optionally allow length-changing shuffles.
2082 bool isSingleSource() const {
2083 return !changesLength() && isSingleSourceMask(getMask());
2084 }
2085
2086 /// Return true if this shuffle mask chooses elements from exactly one source
2087 /// vector without lane crossings. A shuffle using this mask is not
2088 /// necessarily a no-op because it may change the number of elements from its
2089 /// input vectors or it may provide demanded bits knowledge via undef lanes.
2090 /// Example: <undef,undef,2,3>
2091 static bool isIdentityMask(ArrayRef<int> Mask);
2092 static bool isIdentityMask(const Constant *Mask) {
2093 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((Mask->getType()->isVectorTy() && "Shuffle needs vector constant."
) ? static_cast<void> (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 2093, __PRETTY_FUNCTION__))
;
2094 SmallVector<int, 16> MaskAsInts;
2095 getShuffleMask(Mask, MaskAsInts);
2096 return isIdentityMask(MaskAsInts);
2097 }
2098
2099 /// Return true if this shuffle chooses elements from exactly one source
2100 /// vector without lane crossings and does not change the number of elements
2101 /// from its input vectors.
2102 /// Example: shufflevector <4 x n> A, <4 x n> B, <4,undef,6,undef>
2103 bool isIdentity() const {
2104 return !changesLength() && isIdentityMask(getShuffleMask());
2105 }
2106
2107 /// Return true if this shuffle lengthens exactly one source vector with
2108 /// undefs in the high elements.
2109 bool isIdentityWithPadding() const;
2110
2111 /// Return true if this shuffle extracts the first N elements of exactly one
2112 /// source vector.
2113 bool isIdentityWithExtract() const;
2114
2115 /// Return true if this shuffle concatenates its 2 source vectors. This
2116 /// returns false if either input is undefined. In that case, the shuffle is
2117 /// is better classified as an identity with padding operation.
2118 bool isConcat() const;
2119
2120 /// Return true if this shuffle mask chooses elements from its source vectors
2121 /// without lane crossings. A shuffle using this mask would be
2122 /// equivalent to a vector select with a constant condition operand.
2123 /// Example: <4,1,6,undef>
2124 /// This returns false if the mask does not choose from both input vectors.
2125 /// In that case, the shuffle is better classified as an identity shuffle.
2126 /// This assumes that vector operands are the same length as the mask
2127 /// (a length-changing shuffle can never be equivalent to a vector select).
2128 static bool isSelectMask(ArrayRef<int> Mask);
2129 static bool isSelectMask(const Constant *Mask) {
2130 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((Mask->getType()->isVectorTy() && "Shuffle needs vector constant."
) ? static_cast<void> (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 2130, __PRETTY_FUNCTION__))
;
2131 SmallVector<int, 16> MaskAsInts;
2132 getShuffleMask(Mask, MaskAsInts);
2133 return isSelectMask(MaskAsInts);
2134 }
2135
2136 /// Return true if this shuffle chooses elements from its source vectors
2137 /// without lane crossings and all operands have the same number of elements.
2138 /// In other words, this shuffle is equivalent to a vector select with a
2139 /// constant condition operand.
2140 /// Example: shufflevector <4 x n> A, <4 x n> B, <undef,1,6,3>
2141 /// This returns false if the mask does not choose from both input vectors.
2142 /// In that case, the shuffle is better classified as an identity shuffle.
2143 /// TODO: Optionally allow length-changing shuffles.
2144 bool isSelect() const {
2145 return !changesLength() && isSelectMask(getMask());
2146 }
2147
2148 /// Return true if this shuffle mask swaps the order of elements from exactly
2149 /// one source vector.
2150 /// Example: <7,6,undef,4>
2151 /// This assumes that vector operands are the same length as the mask.
2152 static bool isReverseMask(ArrayRef<int> Mask);
2153 static bool isReverseMask(const Constant *Mask) {
2154 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((Mask->getType()->isVectorTy() && "Shuffle needs vector constant."
) ? static_cast<void> (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 2154, __PRETTY_FUNCTION__))
;
2155 SmallVector<int, 16> MaskAsInts;
2156 getShuffleMask(Mask, MaskAsInts);
2157 return isReverseMask(MaskAsInts);
2158 }
2159
2160 /// Return true if this shuffle swaps the order of elements from exactly
2161 /// one source vector.
2162 /// Example: shufflevector <4 x n> A, <4 x n> B, <3,undef,1,undef>
2163 /// TODO: Optionally allow length-changing shuffles.
2164 bool isReverse() const {
2165 return !changesLength() && isReverseMask(getMask());
2166 }
2167
2168 /// Return true if this shuffle mask chooses all elements with the same value
2169 /// as the first element of exactly one source vector.
2170 /// Example: <4,undef,undef,4>
2171 /// This assumes that vector operands are the same length as the mask.
2172 static bool isZeroEltSplatMask(ArrayRef<int> Mask);
2173 static bool isZeroEltSplatMask(const Constant *Mask) {
2174 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((Mask->getType()->isVectorTy() && "Shuffle needs vector constant."
) ? static_cast<void> (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 2174, __PRETTY_FUNCTION__))
;
2175 SmallVector<int, 16> MaskAsInts;
2176 getShuffleMask(Mask, MaskAsInts);
2177 return isZeroEltSplatMask(MaskAsInts);
2178 }
2179
2180 /// Return true if all elements of this shuffle are the same value as the
2181 /// first element of exactly one source vector without changing the length
2182 /// of that vector.
2183 /// Example: shufflevector <4 x n> A, <4 x n> B, <undef,0,undef,0>
2184 /// TODO: Optionally allow length-changing shuffles.
2185 /// TODO: Optionally allow splats from other elements.
2186 bool isZeroEltSplat() const {
2187 return !changesLength() && isZeroEltSplatMask(getMask());
2188 }
2189
2190 /// Return true if this shuffle mask is a transpose mask.
2191 /// Transpose vector masks transpose a 2xn matrix. They read corresponding
2192 /// even- or odd-numbered vector elements from two n-dimensional source
2193 /// vectors and write each result into consecutive elements of an
2194 /// n-dimensional destination vector. Two shuffles are necessary to complete
2195 /// the transpose, one for the even elements and another for the odd elements.
2196 /// This description closely follows how the TRN1 and TRN2 AArch64
2197 /// instructions operate.
2198 ///
2199 /// For example, a simple 2x2 matrix can be transposed with:
2200 ///
2201 /// ; Original matrix
2202 /// m0 = < a, b >
2203 /// m1 = < c, d >
2204 ///
2205 /// ; Transposed matrix
2206 /// t0 = < a, c > = shufflevector m0, m1, < 0, 2 >
2207 /// t1 = < b, d > = shufflevector m0, m1, < 1, 3 >
2208 ///
2209 /// For matrices having greater than n columns, the resulting nx2 transposed
2210 /// matrix is stored in two result vectors such that one vector contains
2211 /// interleaved elements from all the even-numbered rows and the other vector
2212 /// contains interleaved elements from all the odd-numbered rows. For example,
2213 /// a 2x4 matrix can be transposed with:
2214 ///
2215 /// ; Original matrix
2216 /// m0 = < a, b, c, d >
2217 /// m1 = < e, f, g, h >
2218 ///
2219 /// ; Transposed matrix
2220 /// t0 = < a, e, c, g > = shufflevector m0, m1 < 0, 4, 2, 6 >
2221 /// t1 = < b, f, d, h > = shufflevector m0, m1 < 1, 5, 3, 7 >
2222 static bool isTransposeMask(ArrayRef<int> Mask);
2223 static bool isTransposeMask(const Constant *Mask) {
2224 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((Mask->getType()->isVectorTy() && "Shuffle needs vector constant."
) ? static_cast<void> (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 2224, __PRETTY_FUNCTION__))
;
2225 SmallVector<int, 16> MaskAsInts;
2226 getShuffleMask(Mask, MaskAsInts);
2227 return isTransposeMask(MaskAsInts);
2228 }
2229
2230 /// Return true if this shuffle transposes the elements of its inputs without
2231 /// changing the length of the vectors. This operation may also be known as a
2232 /// merge or interleave. See the description for isTransposeMask() for the
2233 /// exact specification.
2234 /// Example: shufflevector <4 x n> A, <4 x n> B, <0,4,2,6>
2235 bool isTranspose() const {
2236 return !changesLength() && isTransposeMask(getMask());
2237 }
2238
2239 /// Return true if this shuffle mask is an extract subvector mask.
2240 /// A valid extract subvector mask returns a smaller vector from a single
2241 /// source operand. The base extraction index is returned as well.
2242 static bool isExtractSubvectorMask(ArrayRef<int> Mask, int NumSrcElts,
2243 int &Index);
2244 static bool isExtractSubvectorMask(const Constant *Mask, int NumSrcElts,
2245 int &Index) {
2246 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((Mask->getType()->isVectorTy() && "Shuffle needs vector constant."
) ? static_cast<void> (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 2246, __PRETTY_FUNCTION__))
;
2247 SmallVector<int, 16> MaskAsInts;
2248 getShuffleMask(Mask, MaskAsInts);
2249 return isExtractSubvectorMask(MaskAsInts, NumSrcElts, Index);
2250 }
2251
2252 /// Return true if this shuffle mask is an extract subvector mask.
2253 bool isExtractSubvectorMask(int &Index) const {
2254 int NumSrcElts = Op<0>()->getType()->getVectorNumElements();
2255 return isExtractSubvectorMask(getMask(), NumSrcElts, Index);
2256 }
2257
2258 /// Change values in a shuffle permute mask assuming the two vector operands
2259 /// of length InVecNumElts have swapped position.
2260 static void commuteShuffleMask(MutableArrayRef<int> Mask,
2261 unsigned InVecNumElts) {
2262 for (int &Idx : Mask) {
2263 if (Idx == -1)
2264 continue;
2265 Idx = Idx < (int)InVecNumElts ? Idx + InVecNumElts : Idx - InVecNumElts;
2266 assert(Idx >= 0 && Idx < (int)InVecNumElts * 2 &&((Idx >= 0 && Idx < (int)InVecNumElts * 2 &&
"shufflevector mask index out of range") ? static_cast<void
> (0) : __assert_fail ("Idx >= 0 && Idx < (int)InVecNumElts * 2 && \"shufflevector mask index out of range\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 2267, __PRETTY_FUNCTION__))
2267 "shufflevector mask index out of range")((Idx >= 0 && Idx < (int)InVecNumElts * 2 &&
"shufflevector mask index out of range") ? static_cast<void
> (0) : __assert_fail ("Idx >= 0 && Idx < (int)InVecNumElts * 2 && \"shufflevector mask index out of range\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 2267, __PRETTY_FUNCTION__))
;
2268 }
2269 }
2270
2271 // Methods for support type inquiry through isa, cast, and dyn_cast:
2272 static bool classof(const Instruction *I) {
2273 return I->getOpcode() == Instruction::ShuffleVector;
2274 }
2275 static bool classof(const Value *V) {
2276 return isa<Instruction>(V) && classof(cast<Instruction>(V));
2277 }
2278};
2279
2280template <>
2281struct OperandTraits<ShuffleVectorInst> :
2282 public FixedNumOperandTraits<ShuffleVectorInst, 3> {
2283};
2284
2285DEFINE_TRANSPARENT_OPERAND_ACCESSORS(ShuffleVectorInst, Value)ShuffleVectorInst::op_iterator ShuffleVectorInst::op_begin() {
return OperandTraits<ShuffleVectorInst>::op_begin(this
); } ShuffleVectorInst::const_op_iterator ShuffleVectorInst::
op_begin() const { return OperandTraits<ShuffleVectorInst>
::op_begin(const_cast<ShuffleVectorInst*>(this)); } ShuffleVectorInst
::op_iterator ShuffleVectorInst::op_end() { return OperandTraits
<ShuffleVectorInst>::op_end(this); } ShuffleVectorInst::
const_op_iterator ShuffleVectorInst::op_end() const { return OperandTraits
<ShuffleVectorInst>::op_end(const_cast<ShuffleVectorInst
*>(this)); } Value *ShuffleVectorInst::getOperand(unsigned
i_nocapture) const { ((i_nocapture < OperandTraits<ShuffleVectorInst
>::operands(this) && "getOperand() out of range!")
? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<ShuffleVectorInst>::operands(this) && \"getOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 2285, __PRETTY_FUNCTION__)); return cast_or_null<Value>
( OperandTraits<ShuffleVectorInst>::op_begin(const_cast
<ShuffleVectorInst*>(this))[i_nocapture].get()); } void
ShuffleVectorInst::setOperand(unsigned i_nocapture, Value *Val_nocapture
) { ((i_nocapture < OperandTraits<ShuffleVectorInst>
::operands(this) && "setOperand() out of range!") ? static_cast
<void> (0) : __assert_fail ("i_nocapture < OperandTraits<ShuffleVectorInst>::operands(this) && \"setOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 2285, __PRETTY_FUNCTION__)); OperandTraits<ShuffleVectorInst
>::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned
ShuffleVectorInst::getNumOperands() const { return OperandTraits
<ShuffleVectorInst>::operands(this); } template <int
Idx_nocapture> Use &ShuffleVectorInst::Op() { return this
->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture
> const Use &ShuffleVectorInst::Op() const { return this
->OpFrom<Idx_nocapture>(this); }
2286
2287//===----------------------------------------------------------------------===//
2288// ExtractValueInst Class
2289//===----------------------------------------------------------------------===//
2290
2291/// This instruction extracts a struct member or array
2292/// element value from an aggregate value.
2293///
2294class ExtractValueInst : public UnaryInstruction {
2295 SmallVector<unsigned, 4> Indices;
2296
2297 ExtractValueInst(const ExtractValueInst &EVI);
2298
2299 /// Constructors - Create a extractvalue instruction with a base aggregate
2300 /// value and a list of indices. The first ctor can optionally insert before
2301 /// an existing instruction, the second appends the new instruction to the
2302 /// specified BasicBlock.
2303 inline ExtractValueInst(Value *Agg,
2304 ArrayRef<unsigned> Idxs,
2305 const Twine &NameStr,
2306 Instruction *InsertBefore);
2307 inline ExtractValueInst(Value *Agg,
2308 ArrayRef<unsigned> Idxs,
2309 const Twine &NameStr, BasicBlock *InsertAtEnd);
2310
2311 void init(ArrayRef<unsigned> Idxs, const Twine &NameStr);
2312
2313protected:
2314 // Note: Instruction needs to be a friend here to call cloneImpl.
2315 friend class Instruction;
2316
2317 ExtractValueInst *cloneImpl() const;
2318
2319public:
2320 static ExtractValueInst *Create(Value *Agg,
2321 ArrayRef<unsigned> Idxs,
2322 const Twine &NameStr = "",
2323 Instruction *InsertBefore = nullptr) {
2324 return new
2325 ExtractValueInst(Agg, Idxs, NameStr, InsertBefore);
2326 }
2327
2328 static ExtractValueInst *Create(Value *Agg,
2329 ArrayRef<unsigned> Idxs,
2330 const Twine &NameStr,
2331 BasicBlock *InsertAtEnd) {
2332 return new ExtractValueInst(Agg, Idxs, NameStr, InsertAtEnd);
2333 }
2334
2335 /// Returns the type of the element that would be extracted
2336 /// with an extractvalue instruction with the specified parameters.
2337 ///
2338 /// Null is returned if the indices are invalid for the specified type.
2339 static Type *getIndexedType(Type *Agg, ArrayRef<unsigned> Idxs);
2340
2341 using idx_iterator = const unsigned*;
2342
2343 inline idx_iterator idx_begin() const { return Indices.begin(); }
2344 inline idx_iterator idx_end() const { return Indices.end(); }
2345 inline iterator_range<idx_iterator> indices() const {
2346 return make_range(idx_begin(), idx_end());
2347 }
2348
2349 Value *getAggregateOperand() {
2350 return getOperand(0);
2351 }
2352 const Value *getAggregateOperand() const {
2353 return getOperand(0);
2354 }
2355 static unsigned getAggregateOperandIndex() {
2356 return 0U; // get index for modifying correct operand
2357 }
2358
2359 ArrayRef<unsigned> getIndices() const {
2360 return Indices;
2361 }
2362
2363 unsigned getNumIndices() const {
2364 return (unsigned)Indices.size();
2365 }
2366
2367 bool hasIndices() const {
2368 return true;
2369 }
2370
2371 // Methods for support type inquiry through isa, cast, and dyn_cast:
2372 static bool classof(const Instruction *I) {
2373 return I->getOpcode() == Instruction::ExtractValue;
2374 }
2375 static bool classof(const Value *V) {
2376 return isa<Instruction>(V) && classof(cast<Instruction>(V));
2377 }
2378};
2379
2380ExtractValueInst::ExtractValueInst(Value *Agg,
2381 ArrayRef<unsigned> Idxs,
2382 const Twine &NameStr,
2383 Instruction *InsertBefore)
2384 : UnaryInstruction(checkGEPType(getIndexedType(Agg->getType(), Idxs)),
2385 ExtractValue, Agg, InsertBefore) {
2386 init(Idxs, NameStr);
2387}
2388
2389ExtractValueInst::ExtractValueInst(Value *Agg,
2390 ArrayRef<unsigned> Idxs,
2391 const Twine &NameStr,
2392 BasicBlock *InsertAtEnd)
2393 : UnaryInstruction(checkGEPType(getIndexedType(Agg->getType(), Idxs)),
2394 ExtractValue, Agg, InsertAtEnd) {
2395 init(Idxs, NameStr);
2396}
2397
2398//===----------------------------------------------------------------------===//
2399// InsertValueInst Class
2400//===----------------------------------------------------------------------===//
2401
2402/// This instruction inserts a struct field of array element
2403/// value into an aggregate value.
2404///
2405class InsertValueInst : public Instruction {
2406 SmallVector<unsigned, 4> Indices;
2407
2408 InsertValueInst(const InsertValueInst &IVI);
2409
2410 /// Constructors - Create a insertvalue instruction with a base aggregate
2411 /// value, a value to insert, and a list of indices. The first ctor can
2412 /// optionally insert before an existing instruction, the second appends
2413 /// the new instruction to the specified BasicBlock.
2414 inline InsertValueInst(Value *Agg, Value *Val,
2415 ArrayRef<unsigned> Idxs,
2416 const Twine &NameStr,
2417 Instruction *InsertBefore);
2418 inline InsertValueInst(Value *Agg, Value *Val,
2419 ArrayRef<unsigned> Idxs,
2420 const Twine &NameStr, BasicBlock *InsertAtEnd);
2421
2422 /// Constructors - These two constructors are convenience methods because one
2423 /// and two index insertvalue instructions are so common.
2424 InsertValueInst(Value *Agg, Value *Val, unsigned Idx,
2425 const Twine &NameStr = "",
2426 Instruction *InsertBefore = nullptr);
2427 InsertValueInst(Value *Agg, Value *Val, unsigned Idx, const Twine &NameStr,
2428 BasicBlock *InsertAtEnd);
2429
2430 void init(Value *Agg, Value *Val, ArrayRef<unsigned> Idxs,
2431 const Twine &NameStr);
2432
2433protected:
2434 // Note: Instruction needs to be a friend here to call cloneImpl.
2435 friend class Instruction;
2436
2437 InsertValueInst *cloneImpl() const;
2438
2439public:
2440 // allocate space for exactly two operands
2441 void *operator new(size_t s) {
2442 return User::operator new(s, 2);
2443 }
2444
2445 static InsertValueInst *Create(Value *Agg, Value *Val,
2446 ArrayRef<unsigned> Idxs,
2447 const Twine &NameStr = "",
2448 Instruction *InsertBefore = nullptr) {
2449 return new InsertValueInst(Agg, Val, Idxs, NameStr, InsertBefore);
2450 }
2451
2452 static InsertValueInst *Create(Value *Agg, Value *Val,
2453 ArrayRef<unsigned> Idxs,
2454 const Twine &NameStr,
2455 BasicBlock *InsertAtEnd) {
2456 return new InsertValueInst(Agg, Val, Idxs, NameStr, InsertAtEnd);
2457 }
2458
2459 /// Transparently provide more efficient getOperand methods.
2460 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
;
2461
2462 using idx_iterator = const unsigned*;
2463
2464 inline idx_iterator idx_begin() const { return Indices.begin(); }
2465 inline idx_iterator idx_end() const { return Indices.end(); }
2466 inline iterator_range<idx_iterator> indices() const {
2467 return make_range(idx_begin(), idx_end());
2468 }
2469
2470 Value *getAggregateOperand() {
2471 return getOperand(0);
2472 }
2473 const Value *getAggregateOperand() const {
2474 return getOperand(0);
2475 }
2476 static unsigned getAggregateOperandIndex() {
2477 return 0U; // get index for modifying correct operand
2478 }
2479
2480 Value *getInsertedValueOperand() {
2481 return getOperand(1);
2482 }
2483 const Value *getInsertedValueOperand() const {
2484 return getOperand(1);
2485 }
2486 static unsigned getInsertedValueOperandIndex() {
2487 return 1U; // get index for modifying correct operand
2488 }
2489
2490 ArrayRef<unsigned> getIndices() const {
2491 return Indices;
2492 }
2493
2494 unsigned getNumIndices() const {
2495 return (unsigned)Indices.size();
2496 }
2497
2498 bool hasIndices() const {
2499 return true;
2500 }
2501
2502 // Methods for support type inquiry through isa, cast, and dyn_cast:
2503 static bool classof(const Instruction *I) {
2504 return I->getOpcode() == Instruction::InsertValue;
2505 }
2506 static bool classof(const Value *V) {
2507 return isa<Instruction>(V) && classof(cast<Instruction>(V));
2508 }
2509};
2510
2511template <>
2512struct OperandTraits<InsertValueInst> :
2513 public FixedNumOperandTraits<InsertValueInst, 2> {
2514};
2515
2516InsertValueInst::InsertValueInst(Value *Agg,
2517 Value *Val,
2518 ArrayRef<unsigned> Idxs,
2519 const Twine &NameStr,
2520 Instruction *InsertBefore)
2521 : Instruction(Agg->getType(), InsertValue,
2522 OperandTraits<InsertValueInst>::op_begin(this),
2523 2, InsertBefore) {
2524 init(Agg, Val, Idxs, NameStr);
2525}
2526
2527InsertValueInst::InsertValueInst(Value *Agg,
2528 Value *Val,
2529 ArrayRef<unsigned> Idxs,
2530 const Twine &NameStr,
2531 BasicBlock *InsertAtEnd)
2532 : Instruction(Agg->getType(), InsertValue,
2533 OperandTraits<InsertValueInst>::op_begin(this),
2534 2, InsertAtEnd) {
2535 init(Agg, Val, Idxs, NameStr);
2536}
2537
2538DEFINE_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 { ((i_nocapture <
OperandTraits<InsertValueInst>::operands(this) &&
"getOperand() out of range!") ? static_cast<void> (0) :
__assert_fail ("i_nocapture < OperandTraits<InsertValueInst>::operands(this) && \"getOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 2538, __PRETTY_FUNCTION__)); 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) { ((
i_nocapture < OperandTraits<InsertValueInst>::operands
(this) && "setOperand() out of range!") ? static_cast
<void> (0) : __assert_fail ("i_nocapture < OperandTraits<InsertValueInst>::operands(this) && \"setOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 2538, __PRETTY_FUNCTION__)); 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); }
2539
2540//===----------------------------------------------------------------------===//
2541// PHINode Class
2542//===----------------------------------------------------------------------===//
2543
2544// PHINode - The PHINode class is used to represent the magical mystical PHI
2545// node, that can not exist in nature, but can be synthesized in a computer
2546// scientist's overactive imagination.
2547//
2548class PHINode : public Instruction {
2549 /// The number of operands actually allocated. NumOperands is
2550 /// the number actually in use.
2551 unsigned ReservedSpace;
2552
2553 PHINode(const PHINode &PN);
2554
2555 explicit PHINode(Type *Ty, unsigned NumReservedValues,
2556 const Twine &NameStr = "",
2557 Instruction *InsertBefore = nullptr)
2558 : Instruction(Ty, Instruction::PHI, nullptr, 0, InsertBefore),
2559 ReservedSpace(NumReservedValues) {
2560 setName(NameStr);
2561 allocHungoffUses(ReservedSpace);
2562 }
2563
2564 PHINode(Type *Ty, unsigned NumReservedValues, const Twine &NameStr,
2565 BasicBlock *InsertAtEnd)
2566 : Instruction(Ty, Instruction::PHI, nullptr, 0, InsertAtEnd),
2567 ReservedSpace(NumReservedValues) {
2568 setName(NameStr);
2569 allocHungoffUses(ReservedSpace);
2570 }
2571
2572protected:
2573 // Note: Instruction needs to be a friend here to call cloneImpl.
2574 friend class Instruction;
2575
2576 PHINode *cloneImpl() const;
2577
2578 // allocHungoffUses - this is more complicated than the generic
2579 // User::allocHungoffUses, because we have to allocate Uses for the incoming
2580 // values and pointers to the incoming blocks, all in one allocation.
2581 void allocHungoffUses(unsigned N) {
2582 User::allocHungoffUses(N, /* IsPhi */ true);
2583 }
2584
2585public:
2586 /// Constructors - NumReservedValues is a hint for the number of incoming
2587 /// edges that this phi node will have (use 0 if you really have no idea).
2588 static PHINode *Create(Type *Ty, unsigned NumReservedValues,
2589 const Twine &NameStr = "",
2590 Instruction *InsertBefore = nullptr) {
2591 return new PHINode(Ty, NumReservedValues, NameStr, InsertBefore);
2592 }
2593
2594 static PHINode *Create(Type *Ty, unsigned NumReservedValues,
2595 const Twine &NameStr, BasicBlock *InsertAtEnd) {
2596 return new PHINode(Ty, NumReservedValues, NameStr, InsertAtEnd);
2597 }
2598
2599 /// Provide fast operand accessors
2600 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
;
2601
2602 // Block iterator interface. This provides access to the list of incoming
2603 // basic blocks, which parallels the list of incoming values.
2604
2605 using block_iterator = BasicBlock **;
2606 using const_block_iterator = BasicBlock * const *;
2607
2608 block_iterator block_begin() {
2609 Use::UserRef *ref =
2610 reinterpret_cast<Use::UserRef*>(op_begin() + ReservedSpace);
2611 return reinterpret_cast<block_iterator>(ref + 1);
2612 }
2613
2614 const_block_iterator block_begin() const {
2615 const Use::UserRef *ref =
2616 reinterpret_cast<const Use::UserRef*>(op_begin() + ReservedSpace);
2617 return reinterpret_cast<const_block_iterator>(ref + 1);
2618 }
2619
2620 block_iterator block_end() {
2621 return block_begin() + getNumOperands();
2622 }
2623
2624 const_block_iterator block_end() const {
2625 return block_begin() + getNumOperands();
2626 }
2627
2628 iterator_range<block_iterator> blocks() {
2629 return make_range(block_begin(), block_end());
2630 }
2631
2632 iterator_range<const_block_iterator> blocks() const {
2633 return make_range(block_begin(), block_end());
2634 }
2635
2636 op_range incoming_values() { return operands(); }
2637
2638 const_op_range incoming_values() const { return operands(); }
2639
2640 /// Return the number of incoming edges
2641 ///
2642 unsigned getNumIncomingValues() const { return getNumOperands(); }
2643
2644 /// Return incoming value number x
2645 ///
2646 Value *getIncomingValue(unsigned i) const {
2647 return getOperand(i);
2648 }
2649 void setIncomingValue(unsigned i, Value *V) {
2650 assert(V && "PHI node got a null value!")((V && "PHI node got a null value!") ? static_cast<
void> (0) : __assert_fail ("V && \"PHI node got a null value!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 2650, __PRETTY_FUNCTION__))
;
2651 assert(getType() == V->getType() &&((getType() == V->getType() && "All operands to PHI node must be the same type as the PHI node!"
) ? static_cast<void> (0) : __assert_fail ("getType() == V->getType() && \"All operands to PHI node must be the same type as the PHI node!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 2652, __PRETTY_FUNCTION__))
2652 "All operands to PHI node must be the same type as the PHI node!")((getType() == V->getType() && "All operands to PHI node must be the same type as the PHI node!"
) ? static_cast<void> (0) : __assert_fail ("getType() == V->getType() && \"All operands to PHI node must be the same type as the PHI node!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 2652, __PRETTY_FUNCTION__))
;
2653 setOperand(i, V);
2654 }
2655
2656 static unsigned getOperandNumForIncomingValue(unsigned i) {
2657 return i;
2658 }
2659
2660 static unsigned getIncomingValueNumForOperand(unsigned i) {
2661 return i;
2662 }
2663
2664 /// Return incoming basic block number @p i.
2665 ///
2666 BasicBlock *getIncomingBlock(unsigned i) const {
2667 return block_begin()[i];
2668 }
2669
2670 /// Return incoming basic block corresponding
2671 /// to an operand of the PHI.
2672 ///
2673 BasicBlock *getIncomingBlock(const Use &U) const {
2674 assert(this == U.getUser() && "Iterator doesn't point to PHI's Uses?")((this == U.getUser() && "Iterator doesn't point to PHI's Uses?"
) ? static_cast<void> (0) : __assert_fail ("this == U.getUser() && \"Iterator doesn't point to PHI's Uses?\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 2674, __PRETTY_FUNCTION__))
;
2675 return getIncomingBlock(unsigned(&U - op_begin()));
2676 }
2677
2678 /// Return incoming basic block corresponding
2679 /// to value use iterator.
2680 ///
2681 BasicBlock *getIncomingBlock(Value::const_user_iterator I) const {
2682 return getIncomingBlock(I.getUse());
2683 }
2684
2685 void setIncomingBlock(unsigned i, BasicBlock *BB) {
2686 assert(BB && "PHI node got a null basic block!")((BB && "PHI node got a null basic block!") ? static_cast
<void> (0) : __assert_fail ("BB && \"PHI node got a null basic block!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 2686, __PRETTY_FUNCTION__))
;
2687 block_begin()[i] = BB;
2688 }
2689
2690 /// Replace every incoming basic block \p Old to basic block \p New.
2691 void replaceIncomingBlockWith(const BasicBlock *Old, BasicBlock *New) {
2692 assert(New && Old && "PHI node got a null basic block!")((New && Old && "PHI node got a null basic block!"
) ? static_cast<void> (0) : __assert_fail ("New && Old && \"PHI node got a null basic block!\""
, "/build/llvm-toolchain-snapshot-10~svn374877/include/llvm/IR/Instructions.h"
, 2692, __PRETTY_FUNCTION__))
;
2693 for (unsigned Op = 0, NumOps = getNumOperands(