Bug Summary

File:llvm/lib/Transforms/Scalar/LICM.cpp
Warning:line 1142, column 33
Called C++ object pointer is null

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name LICM.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mframe-pointer=none -fmath-errno -fno-rounding-math -mconstructor-aliases -munwind-tables -target-cpu x86-64 -tune-cpu generic -fno-split-dwarf-inlining -debugger-tuning=gdb -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-12/lib/clang/12.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/build-llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/build-llvm/include -I /build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0/backward -internal-isystem /usr/local/include -internal-isystem /usr/lib/llvm-12/lib/clang/12.0.0/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-comment -std=c++14 -fdeprecated-macro -fdebug-compilation-dir /build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/build-llvm/lib/Transforms/Scalar -fdebug-prefix-map=/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191=. -ferror-limit 19 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -o /tmp/scan-build-2020-11-21-121427-42170-1 -x c++ /build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp

/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp

1//===-- LICM.cpp - Loop Invariant Code Motion Pass ------------------------===//
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 pass performs loop invariant code motion, attempting to remove as much
10// code from the body of a loop as possible. It does this by either hoisting
11// code into the preheader block, or by sinking code to the exit blocks if it is
12// safe. This pass also promotes must-aliased memory locations in the loop to
13// live in registers, thus hoisting and sinking "invariant" loads and stores.
14//
15// This pass uses alias analysis for two purposes:
16//
17// 1. Moving loop invariant loads and calls out of loops. If we can determine
18// that a load or call inside of a loop never aliases anything stored to,
19// we can hoist it or sink it like any other instruction.
20// 2. Scalar Promotion of Memory - If there is a store instruction inside of
21// the loop, we try to move the store to happen AFTER the loop instead of
22// inside of the loop. This can only happen if a few conditions are true:
23// A. The pointer stored through is loop invariant
24// B. There are no stores or loads in the loop which _may_ alias the
25// pointer. There are no calls in the loop which mod/ref the pointer.
26// If these conditions are true, we can promote the loads and stores in the
27// loop of the pointer to use a temporary alloca'd variable. We then use
28// the SSAUpdater to construct the appropriate SSA form for the value.
29//
30//===----------------------------------------------------------------------===//
31
32#include "llvm/Transforms/Scalar/LICM.h"
33#include "llvm/ADT/SetOperations.h"
34#include "llvm/ADT/Statistic.h"
35#include "llvm/Analysis/AliasAnalysis.h"
36#include "llvm/Analysis/AliasSetTracker.h"
37#include "llvm/Analysis/BasicAliasAnalysis.h"
38#include "llvm/Analysis/BlockFrequencyInfo.h"
39#include "llvm/Analysis/CaptureTracking.h"
40#include "llvm/Analysis/ConstantFolding.h"
41#include "llvm/Analysis/GlobalsModRef.h"
42#include "llvm/Analysis/GuardUtils.h"
43#include "llvm/Analysis/LazyBlockFrequencyInfo.h"
44#include "llvm/Analysis/Loads.h"
45#include "llvm/Analysis/LoopInfo.h"
46#include "llvm/Analysis/LoopIterator.h"
47#include "llvm/Analysis/LoopPass.h"
48#include "llvm/Analysis/MemoryBuiltins.h"
49#include "llvm/Analysis/MemorySSA.h"
50#include "llvm/Analysis/MemorySSAUpdater.h"
51#include "llvm/Analysis/MustExecute.h"
52#include "llvm/Analysis/OptimizationRemarkEmitter.h"
53#include "llvm/Analysis/ScalarEvolution.h"
54#include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
55#include "llvm/Analysis/TargetLibraryInfo.h"
56#include "llvm/Analysis/ValueTracking.h"
57#include "llvm/IR/CFG.h"
58#include "llvm/IR/Constants.h"
59#include "llvm/IR/DataLayout.h"
60#include "llvm/IR/DebugInfoMetadata.h"
61#include "llvm/IR/DerivedTypes.h"
62#include "llvm/IR/Dominators.h"
63#include "llvm/IR/Instructions.h"
64#include "llvm/IR/IntrinsicInst.h"
65#include "llvm/IR/LLVMContext.h"
66#include "llvm/IR/Metadata.h"
67#include "llvm/IR/PatternMatch.h"
68#include "llvm/IR/PredIteratorCache.h"
69#include "llvm/InitializePasses.h"
70#include "llvm/Support/CommandLine.h"
71#include "llvm/Support/Debug.h"
72#include "llvm/Support/raw_ostream.h"
73#include "llvm/Transforms/Scalar.h"
74#include "llvm/Transforms/Scalar/LoopPassManager.h"
75#include "llvm/Transforms/Utils/AssumeBundleBuilder.h"
76#include "llvm/Transforms/Utils/BasicBlockUtils.h"
77#include "llvm/Transforms/Utils/Local.h"
78#include "llvm/Transforms/Utils/LoopUtils.h"
79#include "llvm/Transforms/Utils/SSAUpdater.h"
80#include <algorithm>
81#include <utility>
82using namespace llvm;
83
84#define DEBUG_TYPE"licm" "licm"
85
86STATISTIC(NumCreatedBlocks, "Number of blocks created")static llvm::Statistic NumCreatedBlocks = {"licm", "NumCreatedBlocks"
, "Number of blocks created"}
;
87STATISTIC(NumClonedBranches, "Number of branches cloned")static llvm::Statistic NumClonedBranches = {"licm", "NumClonedBranches"
, "Number of branches cloned"}
;
88STATISTIC(NumSunk, "Number of instructions sunk out of loop")static llvm::Statistic NumSunk = {"licm", "NumSunk", "Number of instructions sunk out of loop"
}
;
89STATISTIC(NumHoisted, "Number of instructions hoisted out of loop")static llvm::Statistic NumHoisted = {"licm", "NumHoisted", "Number of instructions hoisted out of loop"
}
;
90STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk")static llvm::Statistic NumMovedLoads = {"licm", "NumMovedLoads"
, "Number of load insts hoisted or sunk"}
;
91STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk")static llvm::Statistic NumMovedCalls = {"licm", "NumMovedCalls"
, "Number of call insts hoisted or sunk"}
;
92STATISTIC(NumPromoted, "Number of memory locations promoted to registers")static llvm::Statistic NumPromoted = {"licm", "NumPromoted", "Number of memory locations promoted to registers"
}
;
93
94/// Memory promotion is enabled by default.
95static cl::opt<bool>
96 DisablePromotion("disable-licm-promotion", cl::Hidden, cl::init(false),
97 cl::desc("Disable memory promotion in LICM pass"));
98
99static cl::opt<bool> ControlFlowHoisting(
100 "licm-control-flow-hoisting", cl::Hidden, cl::init(false),
101 cl::desc("Enable control flow (and PHI) hoisting in LICM"));
102
103static cl::opt<unsigned> HoistSinkColdnessThreshold(
104 "licm-coldness-threshold", cl::Hidden, cl::init(4),
105 cl::desc("Relative coldness Threshold of hoisting/sinking destination "
106 "block for LICM to be considered beneficial"));
107
108static cl::opt<uint32_t> MaxNumUsesTraversed(
109 "licm-max-num-uses-traversed", cl::Hidden, cl::init(8),
110 cl::desc("Max num uses visited for identifying load "
111 "invariance in loop using invariant start (default = 8)"));
112
113// Default value of zero implies we use the regular alias set tracker mechanism
114// instead of the cross product using AA to identify aliasing of the memory
115// location we are interested in.
116static cl::opt<int>
117LICMN2Theshold("licm-n2-threshold", cl::Hidden, cl::init(0),
118 cl::desc("How many instruction to cross product using AA"));
119
120// Experimental option to allow imprecision in LICM in pathological cases, in
121// exchange for faster compile. This is to be removed if MemorySSA starts to
122// address the same issue. This flag applies only when LICM uses MemorySSA
123// instead on AliasSetTracker. LICM calls MemorySSAWalker's
124// getClobberingMemoryAccess, up to the value of the Cap, getting perfect
125// accuracy. Afterwards, LICM will call into MemorySSA's getDefiningAccess,
126// which may not be precise, since optimizeUses is capped. The result is
127// correct, but we may not get as "far up" as possible to get which access is
128// clobbering the one queried.
129cl::opt<unsigned> llvm::SetLicmMssaOptCap(
130 "licm-mssa-optimization-cap", cl::init(100), cl::Hidden,
131 cl::desc("Enable imprecision in LICM in pathological cases, in exchange "
132 "for faster compile. Caps the MemorySSA clobbering calls."));
133
134// Experimentally, memory promotion carries less importance than sinking and
135// hoisting. Limit when we do promotion when using MemorySSA, in order to save
136// compile time.
137cl::opt<unsigned> llvm::SetLicmMssaNoAccForPromotionCap(
138 "licm-mssa-max-acc-promotion", cl::init(250), cl::Hidden,
139 cl::desc("[LICM & MemorySSA] When MSSA in LICM is disabled, this has no "
140 "effect. When MSSA in LICM is enabled, then this is the maximum "
141 "number of accesses allowed to be present in a loop in order to "
142 "enable memory promotion."));
143
144static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI);
145static bool isNotUsedOrFreeInLoop(const Instruction &I, const Loop *CurLoop,
146 const LoopSafetyInfo *SafetyInfo,
147 TargetTransformInfo *TTI, bool &FreeInLoop);
148static void hoist(Instruction &I, const DominatorTree *DT, const Loop *CurLoop,
149 BasicBlock *Dest, ICFLoopSafetyInfo *SafetyInfo,
150 MemorySSAUpdater *MSSAU, ScalarEvolution *SE,
151 OptimizationRemarkEmitter *ORE);
152static bool sink(Instruction &I, LoopInfo *LI, DominatorTree *DT,
153 BlockFrequencyInfo *BFI, const Loop *CurLoop,
154 ICFLoopSafetyInfo *SafetyInfo, MemorySSAUpdater *MSSAU,
155 OptimizationRemarkEmitter *ORE);
156static bool isSafeToExecuteUnconditionally(Instruction &Inst,
157 const DominatorTree *DT,
158 const Loop *CurLoop,
159 const LoopSafetyInfo *SafetyInfo,
160 OptimizationRemarkEmitter *ORE,
161 const Instruction *CtxI = nullptr);
162static bool pointerInvalidatedByLoop(MemoryLocation MemLoc,
163 AliasSetTracker *CurAST, Loop *CurLoop,
164 AAResults *AA);
165static bool pointerInvalidatedByLoopWithMSSA(MemorySSA *MSSA, MemoryUse *MU,
166 Loop *CurLoop,
167 SinkAndHoistLICMFlags &Flags);
168static Instruction *cloneInstructionInExitBlock(
169 Instruction &I, BasicBlock &ExitBlock, PHINode &PN, const LoopInfo *LI,
170 const LoopSafetyInfo *SafetyInfo, MemorySSAUpdater *MSSAU);
171
172static void eraseInstruction(Instruction &I, ICFLoopSafetyInfo &SafetyInfo,
173 AliasSetTracker *AST, MemorySSAUpdater *MSSAU);
174
175static void moveInstructionBefore(Instruction &I, Instruction &Dest,
176 ICFLoopSafetyInfo &SafetyInfo,
177 MemorySSAUpdater *MSSAU, ScalarEvolution *SE);
178
179namespace {
180struct LoopInvariantCodeMotion {
181 bool runOnLoop(Loop *L, AAResults *AA, LoopInfo *LI, DominatorTree *DT,
182 BlockFrequencyInfo *BFI, TargetLibraryInfo *TLI,
183 TargetTransformInfo *TTI, ScalarEvolution *SE, MemorySSA *MSSA,
184 OptimizationRemarkEmitter *ORE);
185
186 LoopInvariantCodeMotion(unsigned LicmMssaOptCap,
187 unsigned LicmMssaNoAccForPromotionCap)
188 : LicmMssaOptCap(LicmMssaOptCap),
189 LicmMssaNoAccForPromotionCap(LicmMssaNoAccForPromotionCap) {}
190
191private:
192 unsigned LicmMssaOptCap;
193 unsigned LicmMssaNoAccForPromotionCap;
194
195 std::unique_ptr<AliasSetTracker>
196 collectAliasInfoForLoop(Loop *L, LoopInfo *LI, AAResults *AA);
197 std::unique_ptr<AliasSetTracker>
198 collectAliasInfoForLoopWithMSSA(Loop *L, AAResults *AA,
199 MemorySSAUpdater *MSSAU);
200};
201
202struct LegacyLICMPass : public LoopPass {
203 static char ID; // Pass identification, replacement for typeid
204 LegacyLICMPass(
205 unsigned LicmMssaOptCap = SetLicmMssaOptCap,
206 unsigned LicmMssaNoAccForPromotionCap = SetLicmMssaNoAccForPromotionCap)
207 : LoopPass(ID), LICM(LicmMssaOptCap, LicmMssaNoAccForPromotionCap) {
208 initializeLegacyLICMPassPass(*PassRegistry::getPassRegistry());
209 }
210
211 bool runOnLoop(Loop *L, LPPassManager &LPM) override {
212 if (skipLoop(L))
213 return false;
214
215 auto *SE = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>();
216 MemorySSA *MSSA = EnableMSSALoopDependency
217 ? (&getAnalysis<MemorySSAWrapperPass>().getMSSA())
218 : nullptr;
219 bool hasProfileData = L->getHeader()->getParent()->hasProfileData();
220 BlockFrequencyInfo *BFI =
221 hasProfileData ? &getAnalysis<LazyBlockFrequencyInfoPass>().getBFI()
222 : nullptr;
223 // For the old PM, we can't use OptimizationRemarkEmitter as an analysis
224 // pass. Function analyses need to be preserved across loop transformations
225 // but ORE cannot be preserved (see comment before the pass definition).
226 OptimizationRemarkEmitter ORE(L->getHeader()->getParent());
227 return LICM.runOnLoop(
228 L, &getAnalysis<AAResultsWrapperPass>().getAAResults(),
229 &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(),
230 &getAnalysis<DominatorTreeWrapperPass>().getDomTree(), BFI,
231 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(
232 *L->getHeader()->getParent()),
233 &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
234 *L->getHeader()->getParent()),
235 SE ? &SE->getSE() : nullptr, MSSA, &ORE);
236 }
237
238 /// This transformation requires natural loop information & requires that
239 /// loop preheaders be inserted into the CFG...
240 ///
241 void getAnalysisUsage(AnalysisUsage &AU) const override {
242 AU.addPreserved<DominatorTreeWrapperPass>();
243 AU.addPreserved<LoopInfoWrapperPass>();
244 AU.addRequired<TargetLibraryInfoWrapperPass>();
245 if (EnableMSSALoopDependency) {
246 AU.addRequired<MemorySSAWrapperPass>();
247 AU.addPreserved<MemorySSAWrapperPass>();
248 }
249 AU.addRequired<TargetTransformInfoWrapperPass>();
250 getLoopAnalysisUsage(AU);
251 LazyBlockFrequencyInfoPass::getLazyBFIAnalysisUsage(AU);
252 AU.addPreserved<LazyBlockFrequencyInfoPass>();
253 AU.addPreserved<LazyBranchProbabilityInfoPass>();
254 }
255
256private:
257 LoopInvariantCodeMotion LICM;
258};
259} // namespace
260
261PreservedAnalyses LICMPass::run(Loop &L, LoopAnalysisManager &AM,
262 LoopStandardAnalysisResults &AR, LPMUpdater &) {
263 // For the new PM, we also can't use OptimizationRemarkEmitter as an analysis
264 // pass. Function analyses need to be preserved across loop transformations
265 // but ORE cannot be preserved (see comment before the pass definition).
266 OptimizationRemarkEmitter ORE(L.getHeader()->getParent());
267
268 LoopInvariantCodeMotion LICM(LicmMssaOptCap, LicmMssaNoAccForPromotionCap);
269 if (!LICM.runOnLoop(&L, &AR.AA, &AR.LI, &AR.DT, AR.BFI, &AR.TLI, &AR.TTI,
270 &AR.SE, AR.MSSA, &ORE))
271 return PreservedAnalyses::all();
272
273 auto PA = getLoopPassPreservedAnalyses();
274
275 PA.preserve<DominatorTreeAnalysis>();
276 PA.preserve<LoopAnalysis>();
277 if (AR.MSSA)
278 PA.preserve<MemorySSAAnalysis>();
279
280 return PA;
281}
282
283char LegacyLICMPass::ID = 0;
284INITIALIZE_PASS_BEGIN(LegacyLICMPass, "licm", "Loop Invariant Code Motion",static void *initializeLegacyLICMPassPassOnce(PassRegistry &
Registry) {
285 false, false)static void *initializeLegacyLICMPassPassOnce(PassRegistry &
Registry) {
286INITIALIZE_PASS_DEPENDENCY(LoopPass)initializeLoopPassPass(Registry);
287INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry);
288INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry);
289INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)initializeMemorySSAWrapperPassPass(Registry);
290INITIALIZE_PASS_DEPENDENCY(LazyBFIPass)initializeLazyBFIPassPass(Registry);
291INITIALIZE_PASS_END(LegacyLICMPass, "licm", "Loop Invariant Code Motion", false,PassInfo *PI = new PassInfo( "Loop Invariant Code Motion", "licm"
, &LegacyLICMPass::ID, PassInfo::NormalCtor_t(callDefaultCtor
<LegacyLICMPass>), false, false); Registry.registerPass
(*PI, true); return PI; } static llvm::once_flag InitializeLegacyLICMPassPassFlag
; void llvm::initializeLegacyLICMPassPass(PassRegistry &Registry
) { llvm::call_once(InitializeLegacyLICMPassPassFlag, initializeLegacyLICMPassPassOnce
, std::ref(Registry)); }
292 false)PassInfo *PI = new PassInfo( "Loop Invariant Code Motion", "licm"
, &LegacyLICMPass::ID, PassInfo::NormalCtor_t(callDefaultCtor
<LegacyLICMPass>), false, false); Registry.registerPass
(*PI, true); return PI; } static llvm::once_flag InitializeLegacyLICMPassPassFlag
; void llvm::initializeLegacyLICMPassPass(PassRegistry &Registry
) { llvm::call_once(InitializeLegacyLICMPassPassFlag, initializeLegacyLICMPassPassOnce
, std::ref(Registry)); }
293
294Pass *llvm::createLICMPass() { return new LegacyLICMPass(); }
295Pass *llvm::createLICMPass(unsigned LicmMssaOptCap,
296 unsigned LicmMssaNoAccForPromotionCap) {
297 return new LegacyLICMPass(LicmMssaOptCap, LicmMssaNoAccForPromotionCap);
298}
299
300/// Hoist expressions out of the specified loop. Note, alias info for inner
301/// loop is not preserved so it is not a good idea to run LICM multiple
302/// times on one loop.
303bool LoopInvariantCodeMotion::runOnLoop(
304 Loop *L, AAResults *AA, LoopInfo *LI, DominatorTree *DT,
305 BlockFrequencyInfo *BFI, TargetLibraryInfo *TLI, TargetTransformInfo *TTI,
306 ScalarEvolution *SE, MemorySSA *MSSA, OptimizationRemarkEmitter *ORE) {
307 bool Changed = false;
308
309 assert(L->isLCSSAForm(*DT) && "Loop is not in LCSSA form.")((L->isLCSSAForm(*DT) && "Loop is not in LCSSA form."
) ? static_cast<void> (0) : __assert_fail ("L->isLCSSAForm(*DT) && \"Loop is not in LCSSA form.\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 309, __PRETTY_FUNCTION__))
;
310
311 // If this loop has metadata indicating that LICM is not to be performed then
312 // just exit.
313 if (hasDisableLICMTransformsHint(L)) {
314 return false;
315 }
316
317 std::unique_ptr<AliasSetTracker> CurAST;
318 std::unique_ptr<MemorySSAUpdater> MSSAU;
319 bool NoOfMemAccTooLarge = false;
320 unsigned LicmMssaOptCounter = 0;
321
322 if (!MSSA) {
323 LLVM_DEBUG(dbgs() << "LICM: Using Alias Set Tracker.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "LICM: Using Alias Set Tracker.\n"
; } } while (false)
;
324 CurAST = collectAliasInfoForLoop(L, LI, AA);
325 } else {
326 LLVM_DEBUG(dbgs() << "LICM: Using MemorySSA.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "LICM: Using MemorySSA.\n"; } } while
(false)
;
327 MSSAU = std::make_unique<MemorySSAUpdater>(MSSA);
328
329 unsigned AccessCapCount = 0;
330 for (auto *BB : L->getBlocks()) {
331 if (auto *Accesses = MSSA->getBlockAccesses(BB)) {
332 for (const auto &MA : *Accesses) {
333 (void)MA;
334 AccessCapCount++;
335 if (AccessCapCount > LicmMssaNoAccForPromotionCap) {
336 NoOfMemAccTooLarge = true;
337 break;
338 }
339 }
340 }
341 if (NoOfMemAccTooLarge)
342 break;
343 }
344 }
345
346 // Get the preheader block to move instructions into...
347 BasicBlock *Preheader = L->getLoopPreheader();
348
349 // Compute loop safety information.
350 ICFLoopSafetyInfo SafetyInfo;
351 SafetyInfo.computeLoopSafetyInfo(L);
352
353 // We want to visit all of the instructions in this loop... that are not parts
354 // of our subloops (they have already had their invariants hoisted out of
355 // their loop, into this loop, so there is no need to process the BODIES of
356 // the subloops).
357 //
358 // Traverse the body of the loop in depth first order on the dominator tree so
359 // that we are guaranteed to see definitions before we see uses. This allows
360 // us to sink instructions in one pass, without iteration. After sinking
361 // instructions, we perform another pass to hoist them out of the loop.
362 SinkAndHoistLICMFlags Flags = {NoOfMemAccTooLarge, LicmMssaOptCounter,
363 LicmMssaOptCap, LicmMssaNoAccForPromotionCap,
364 /*IsSink=*/true};
365 if (L->hasDedicatedExits())
366 Changed |=
367 sinkRegion(DT->getNode(L->getHeader()), AA, LI, DT, BFI, TLI, TTI, L,
368 CurAST.get(), MSSAU.get(), &SafetyInfo, Flags, ORE);
369 Flags.IsSink = false;
370 if (Preheader)
371 Changed |=
372 hoistRegion(DT->getNode(L->getHeader()), AA, LI, DT, BFI, TLI, L,
373 CurAST.get(), MSSAU.get(), SE, &SafetyInfo, Flags, ORE);
374
375 // Now that all loop invariants have been removed from the loop, promote any
376 // memory references to scalars that we can.
377 // Don't sink stores from loops without dedicated block exits. Exits
378 // containing indirect branches are not transformed by loop simplify,
379 // make sure we catch that. An additional load may be generated in the
380 // preheader for SSA updater, so also avoid sinking when no preheader
381 // is available.
382 if (!DisablePromotion && Preheader && L->hasDedicatedExits() &&
383 !NoOfMemAccTooLarge) {
384 // Figure out the loop exits and their insertion points
385 SmallVector<BasicBlock *, 8> ExitBlocks;
386 L->getUniqueExitBlocks(ExitBlocks);
387
388 // We can't insert into a catchswitch.
389 bool HasCatchSwitch = llvm::any_of(ExitBlocks, [](BasicBlock *Exit) {
390 return isa<CatchSwitchInst>(Exit->getTerminator());
391 });
392
393 if (!HasCatchSwitch) {
394 SmallVector<Instruction *, 8> InsertPts;
395 SmallVector<MemoryAccess *, 8> MSSAInsertPts;
396 InsertPts.reserve(ExitBlocks.size());
397 if (MSSAU)
398 MSSAInsertPts.reserve(ExitBlocks.size());
399 for (BasicBlock *ExitBlock : ExitBlocks) {
400 InsertPts.push_back(&*ExitBlock->getFirstInsertionPt());
401 if (MSSAU)
402 MSSAInsertPts.push_back(nullptr);
403 }
404
405 PredIteratorCache PIC;
406
407 bool Promoted = false;
408
409 // Build an AST using MSSA.
410 if (!CurAST.get())
411 CurAST = collectAliasInfoForLoopWithMSSA(L, AA, MSSAU.get());
412
413 // Loop over all of the alias sets in the tracker object.
414 for (AliasSet &AS : *CurAST) {
415 // We can promote this alias set if it has a store, if it is a "Must"
416 // alias set, if the pointer is loop invariant, and if we are not
417 // eliminating any volatile loads or stores.
418 if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() ||
419 !L->isLoopInvariant(AS.begin()->getValue()))
420 continue;
421
422 assert(((!AS.empty() && "Must alias set should have at least one pointer element in it!"
) ? static_cast<void> (0) : __assert_fail ("!AS.empty() && \"Must alias set should have at least one pointer element in it!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 424, __PRETTY_FUNCTION__))
423 !AS.empty() &&((!AS.empty() && "Must alias set should have at least one pointer element in it!"
) ? static_cast<void> (0) : __assert_fail ("!AS.empty() && \"Must alias set should have at least one pointer element in it!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 424, __PRETTY_FUNCTION__))
424 "Must alias set should have at least one pointer element in it!")((!AS.empty() && "Must alias set should have at least one pointer element in it!"
) ? static_cast<void> (0) : __assert_fail ("!AS.empty() && \"Must alias set should have at least one pointer element in it!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 424, __PRETTY_FUNCTION__))
;
425
426 SmallSetVector<Value *, 8> PointerMustAliases;
427 for (const auto &ASI : AS)
428 PointerMustAliases.insert(ASI.getValue());
429
430 Promoted |= promoteLoopAccessesToScalars(
431 PointerMustAliases, ExitBlocks, InsertPts, MSSAInsertPts, PIC, LI,
432 DT, TLI, L, CurAST.get(), MSSAU.get(), &SafetyInfo, ORE);
433 }
434
435 // Once we have promoted values across the loop body we have to
436 // recursively reform LCSSA as any nested loop may now have values defined
437 // within the loop used in the outer loop.
438 // FIXME: This is really heavy handed. It would be a bit better to use an
439 // SSAUpdater strategy during promotion that was LCSSA aware and reformed
440 // it as it went.
441 if (Promoted)
442 formLCSSARecursively(*L, *DT, LI, SE);
443
444 Changed |= Promoted;
445 }
446 }
447
448 // Check that neither this loop nor its parent have had LCSSA broken. LICM is
449 // specifically moving instructions across the loop boundary and so it is
450 // especially in need of sanity checking here.
451 assert(L->isLCSSAForm(*DT) && "Loop not left in LCSSA form after LICM!")((L->isLCSSAForm(*DT) && "Loop not left in LCSSA form after LICM!"
) ? static_cast<void> (0) : __assert_fail ("L->isLCSSAForm(*DT) && \"Loop not left in LCSSA form after LICM!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 451, __PRETTY_FUNCTION__))
;
452 assert((L->isOutermost() || L->getParentLoop()->isLCSSAForm(*DT)) &&(((L->isOutermost() || L->getParentLoop()->isLCSSAForm
(*DT)) && "Parent loop not left in LCSSA form after LICM!"
) ? static_cast<void> (0) : __assert_fail ("(L->isOutermost() || L->getParentLoop()->isLCSSAForm(*DT)) && \"Parent loop not left in LCSSA form after LICM!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 453, __PRETTY_FUNCTION__))
453 "Parent loop not left in LCSSA form after LICM!")(((L->isOutermost() || L->getParentLoop()->isLCSSAForm
(*DT)) && "Parent loop not left in LCSSA form after LICM!"
) ? static_cast<void> (0) : __assert_fail ("(L->isOutermost() || L->getParentLoop()->isLCSSAForm(*DT)) && \"Parent loop not left in LCSSA form after LICM!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 453, __PRETTY_FUNCTION__))
;
454
455 if (MSSAU.get() && VerifyMemorySSA)
456 MSSAU->getMemorySSA()->verifyMemorySSA();
457
458 if (Changed && SE)
459 SE->forgetLoopDispositions(L);
460 return Changed;
461}
462
463/// Walk the specified region of the CFG (defined by all blocks dominated by
464/// the specified block, and that are in the current loop) in reverse depth
465/// first order w.r.t the DominatorTree. This allows us to visit uses before
466/// definitions, allowing us to sink a loop body in one pass without iteration.
467///
468bool llvm::sinkRegion(DomTreeNode *N, AAResults *AA, LoopInfo *LI,
469 DominatorTree *DT, BlockFrequencyInfo *BFI,
470 TargetLibraryInfo *TLI, TargetTransformInfo *TTI,
471 Loop *CurLoop, AliasSetTracker *CurAST,
472 MemorySSAUpdater *MSSAU, ICFLoopSafetyInfo *SafetyInfo,
473 SinkAndHoistLICMFlags &Flags,
474 OptimizationRemarkEmitter *ORE) {
475
476 // Verify inputs.
477 assert(N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr &&((N != nullptr && AA != nullptr && LI != nullptr
&& DT != nullptr && CurLoop != nullptr &&
SafetyInfo != nullptr && "Unexpected input to sinkRegion."
) ? static_cast<void> (0) : __assert_fail ("N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && \"Unexpected input to sinkRegion.\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 479, __PRETTY_FUNCTION__))
478 CurLoop != nullptr && SafetyInfo != nullptr &&((N != nullptr && AA != nullptr && LI != nullptr
&& DT != nullptr && CurLoop != nullptr &&
SafetyInfo != nullptr && "Unexpected input to sinkRegion."
) ? static_cast<void> (0) : __assert_fail ("N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && \"Unexpected input to sinkRegion.\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 479, __PRETTY_FUNCTION__))
479 "Unexpected input to sinkRegion.")((N != nullptr && AA != nullptr && LI != nullptr
&& DT != nullptr && CurLoop != nullptr &&
SafetyInfo != nullptr && "Unexpected input to sinkRegion."
) ? static_cast<void> (0) : __assert_fail ("N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && \"Unexpected input to sinkRegion.\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 479, __PRETTY_FUNCTION__))
;
480 assert(((CurAST != nullptr) ^ (MSSAU != nullptr)) &&((((CurAST != nullptr) ^ (MSSAU != nullptr)) && "Either AliasSetTracker or MemorySSA should be initialized."
) ? static_cast<void> (0) : __assert_fail ("((CurAST != nullptr) ^ (MSSAU != nullptr)) && \"Either AliasSetTracker or MemorySSA should be initialized.\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 481, __PRETTY_FUNCTION__))
481 "Either AliasSetTracker or MemorySSA should be initialized.")((((CurAST != nullptr) ^ (MSSAU != nullptr)) && "Either AliasSetTracker or MemorySSA should be initialized."
) ? static_cast<void> (0) : __assert_fail ("((CurAST != nullptr) ^ (MSSAU != nullptr)) && \"Either AliasSetTracker or MemorySSA should be initialized.\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 481, __PRETTY_FUNCTION__))
;
482
483 // We want to visit children before parents. We will enque all the parents
484 // before their children in the worklist and process the worklist in reverse
485 // order.
486 SmallVector<DomTreeNode *, 16> Worklist = collectChildrenInLoop(N, CurLoop);
487
488 bool Changed = false;
489 for (DomTreeNode *DTN : reverse(Worklist)) {
490 BasicBlock *BB = DTN->getBlock();
491 // Only need to process the contents of this block if it is not part of a
492 // subloop (which would already have been processed).
493 if (inSubLoop(BB, CurLoop, LI))
494 continue;
495
496 for (BasicBlock::iterator II = BB->end(); II != BB->begin();) {
497 Instruction &I = *--II;
498
499 // If the instruction is dead, we would try to sink it because it isn't
500 // used in the loop, instead, just delete it.
501 if (isInstructionTriviallyDead(&I, TLI)) {
502 LLVM_DEBUG(dbgs() << "LICM deleting dead inst: " << I << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "LICM deleting dead inst: " <<
I << '\n'; } } while (false)
;
503 salvageKnowledge(&I);
504 salvageDebugInfo(I);
505 ++II;
506 eraseInstruction(I, *SafetyInfo, CurAST, MSSAU);
507 Changed = true;
508 continue;
509 }
510
511 // Check to see if we can sink this instruction to the exit blocks
512 // of the loop. We can do this if the all users of the instruction are
513 // outside of the loop. In this case, it doesn't even matter if the
514 // operands of the instruction are loop invariant.
515 //
516 bool FreeInLoop = false;
517 if (!I.mayHaveSideEffects() &&
518 isNotUsedOrFreeInLoop(I, CurLoop, SafetyInfo, TTI, FreeInLoop) &&
519 canSinkOrHoistInst(I, AA, DT, CurLoop, CurAST, MSSAU, true, &Flags,
520 ORE)) {
521 if (sink(I, LI, DT, BFI, CurLoop, SafetyInfo, MSSAU, ORE)) {
522 if (!FreeInLoop) {
523 ++II;
524 salvageDebugInfo(I);
525 eraseInstruction(I, *SafetyInfo, CurAST, MSSAU);
526 }
527 Changed = true;
528 }
529 }
530 }
531 }
532 if (MSSAU && VerifyMemorySSA)
533 MSSAU->getMemorySSA()->verifyMemorySSA();
534 return Changed;
535}
536
537namespace {
538// This is a helper class for hoistRegion to make it able to hoist control flow
539// in order to be able to hoist phis. The way this works is that we initially
540// start hoisting to the loop preheader, and when we see a loop invariant branch
541// we make note of this. When we then come to hoist an instruction that's
542// conditional on such a branch we duplicate the branch and the relevant control
543// flow, then hoist the instruction into the block corresponding to its original
544// block in the duplicated control flow.
545class ControlFlowHoister {
546private:
547 // Information about the loop we are hoisting from
548 LoopInfo *LI;
549 DominatorTree *DT;
550 Loop *CurLoop;
551 MemorySSAUpdater *MSSAU;
552
553 // A map of blocks in the loop to the block their instructions will be hoisted
554 // to.
555 DenseMap<BasicBlock *, BasicBlock *> HoistDestinationMap;
556
557 // The branches that we can hoist, mapped to the block that marks a
558 // convergence point of their control flow.
559 DenseMap<BranchInst *, BasicBlock *> HoistableBranches;
560
561public:
562 ControlFlowHoister(LoopInfo *LI, DominatorTree *DT, Loop *CurLoop,
563 MemorySSAUpdater *MSSAU)
564 : LI(LI), DT(DT), CurLoop(CurLoop), MSSAU(MSSAU) {}
565
566 void registerPossiblyHoistableBranch(BranchInst *BI) {
567 // We can only hoist conditional branches with loop invariant operands.
568 if (!ControlFlowHoisting || !BI->isConditional() ||
569 !CurLoop->hasLoopInvariantOperands(BI))
570 return;
571
572 // The branch destinations need to be in the loop, and we don't gain
573 // anything by duplicating conditional branches with duplicate successors,
574 // as it's essentially the same as an unconditional branch.
575 BasicBlock *TrueDest = BI->getSuccessor(0);
576 BasicBlock *FalseDest = BI->getSuccessor(1);
577 if (!CurLoop->contains(TrueDest) || !CurLoop->contains(FalseDest) ||
578 TrueDest == FalseDest)
579 return;
580
581 // We can hoist BI if one branch destination is the successor of the other,
582 // or both have common successor which we check by seeing if the
583 // intersection of their successors is non-empty.
584 // TODO: This could be expanded to allowing branches where both ends
585 // eventually converge to a single block.
586 SmallPtrSet<BasicBlock *, 4> TrueDestSucc, FalseDestSucc;
587 TrueDestSucc.insert(succ_begin(TrueDest), succ_end(TrueDest));
588 FalseDestSucc.insert(succ_begin(FalseDest), succ_end(FalseDest));
589 BasicBlock *CommonSucc = nullptr;
590 if (TrueDestSucc.count(FalseDest)) {
591 CommonSucc = FalseDest;
592 } else if (FalseDestSucc.count(TrueDest)) {
593 CommonSucc = TrueDest;
594 } else {
595 set_intersect(TrueDestSucc, FalseDestSucc);
596 // If there's one common successor use that.
597 if (TrueDestSucc.size() == 1)
598 CommonSucc = *TrueDestSucc.begin();
599 // If there's more than one pick whichever appears first in the block list
600 // (we can't use the value returned by TrueDestSucc.begin() as it's
601 // unpredicatable which element gets returned).
602 else if (!TrueDestSucc.empty()) {
603 Function *F = TrueDest->getParent();
604 auto IsSucc = [&](BasicBlock &BB) { return TrueDestSucc.count(&BB); };
605 auto It = std::find_if(F->begin(), F->end(), IsSucc);
606 assert(It != F->end() && "Could not find successor in function")((It != F->end() && "Could not find successor in function"
) ? static_cast<void> (0) : __assert_fail ("It != F->end() && \"Could not find successor in function\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 606, __PRETTY_FUNCTION__))
;
607 CommonSucc = &*It;
608 }
609 }
610 // The common successor has to be dominated by the branch, as otherwise
611 // there will be some other path to the successor that will not be
612 // controlled by this branch so any phi we hoist would be controlled by the
613 // wrong condition. This also takes care of avoiding hoisting of loop back
614 // edges.
615 // TODO: In some cases this could be relaxed if the successor is dominated
616 // by another block that's been hoisted and we can guarantee that the
617 // control flow has been replicated exactly.
618 if (CommonSucc && DT->dominates(BI, CommonSucc))
619 HoistableBranches[BI] = CommonSucc;
620 }
621
622 bool canHoistPHI(PHINode *PN) {
623 // The phi must have loop invariant operands.
624 if (!ControlFlowHoisting || !CurLoop->hasLoopInvariantOperands(PN))
625 return false;
626 // We can hoist phis if the block they are in is the target of hoistable
627 // branches which cover all of the predecessors of the block.
628 SmallPtrSet<BasicBlock *, 8> PredecessorBlocks;
629 BasicBlock *BB = PN->getParent();
630 for (BasicBlock *PredBB : predecessors(BB))
631 PredecessorBlocks.insert(PredBB);
632 // If we have less predecessor blocks than predecessors then the phi will
633 // have more than one incoming value for the same block which we can't
634 // handle.
635 // TODO: This could be handled be erasing some of the duplicate incoming
636 // values.
637 if (PredecessorBlocks.size() != pred_size(BB))
638 return false;
639 for (auto &Pair : HoistableBranches) {
640 if (Pair.second == BB) {
641 // Which blocks are predecessors via this branch depends on if the
642 // branch is triangle-like or diamond-like.
643 if (Pair.first->getSuccessor(0) == BB) {
644 PredecessorBlocks.erase(Pair.first->getParent());
645 PredecessorBlocks.erase(Pair.first->getSuccessor(1));
646 } else if (Pair.first->getSuccessor(1) == BB) {
647 PredecessorBlocks.erase(Pair.first->getParent());
648 PredecessorBlocks.erase(Pair.first->getSuccessor(0));
649 } else {
650 PredecessorBlocks.erase(Pair.first->getSuccessor(0));
651 PredecessorBlocks.erase(Pair.first->getSuccessor(1));
652 }
653 }
654 }
655 // PredecessorBlocks will now be empty if for every predecessor of BB we
656 // found a hoistable branch source.
657 return PredecessorBlocks.empty();
658 }
659
660 BasicBlock *getOrCreateHoistedBlock(BasicBlock *BB) {
661 if (!ControlFlowHoisting)
662 return CurLoop->getLoopPreheader();
663 // If BB has already been hoisted, return that
664 if (HoistDestinationMap.count(BB))
665 return HoistDestinationMap[BB];
666
667 // Check if this block is conditional based on a pending branch
668 auto HasBBAsSuccessor =
669 [&](DenseMap<BranchInst *, BasicBlock *>::value_type &Pair) {
670 return BB != Pair.second && (Pair.first->getSuccessor(0) == BB ||
671 Pair.first->getSuccessor(1) == BB);
672 };
673 auto It = std::find_if(HoistableBranches.begin(), HoistableBranches.end(),
674 HasBBAsSuccessor);
675
676 // If not involved in a pending branch, hoist to preheader
677 BasicBlock *InitialPreheader = CurLoop->getLoopPreheader();
678 if (It == HoistableBranches.end()) {
679 LLVM_DEBUG(dbgs() << "LICM using " << InitialPreheader->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "LICM using " << InitialPreheader
->getName() << " as hoist destination for " <<
BB->getName() << "\n"; } } while (false)
680 << " as hoist destination for " << BB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "LICM using " << InitialPreheader
->getName() << " as hoist destination for " <<
BB->getName() << "\n"; } } while (false)
681 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "LICM using " << InitialPreheader
->getName() << " as hoist destination for " <<
BB->getName() << "\n"; } } while (false)
;
682 HoistDestinationMap[BB] = InitialPreheader;
683 return InitialPreheader;
684 }
685 BranchInst *BI = It->first;
686 assert(std::find_if(++It, HoistableBranches.end(), HasBBAsSuccessor) ==((std::find_if(++It, HoistableBranches.end(), HasBBAsSuccessor
) == HoistableBranches.end() && "BB is expected to be the target of at most one branch"
) ? static_cast<void> (0) : __assert_fail ("std::find_if(++It, HoistableBranches.end(), HasBBAsSuccessor) == HoistableBranches.end() && \"BB is expected to be the target of at most one branch\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 688, __PRETTY_FUNCTION__))
687 HoistableBranches.end() &&((std::find_if(++It, HoistableBranches.end(), HasBBAsSuccessor
) == HoistableBranches.end() && "BB is expected to be the target of at most one branch"
) ? static_cast<void> (0) : __assert_fail ("std::find_if(++It, HoistableBranches.end(), HasBBAsSuccessor) == HoistableBranches.end() && \"BB is expected to be the target of at most one branch\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 688, __PRETTY_FUNCTION__))
688 "BB is expected to be the target of at most one branch")((std::find_if(++It, HoistableBranches.end(), HasBBAsSuccessor
) == HoistableBranches.end() && "BB is expected to be the target of at most one branch"
) ? static_cast<void> (0) : __assert_fail ("std::find_if(++It, HoistableBranches.end(), HasBBAsSuccessor) == HoistableBranches.end() && \"BB is expected to be the target of at most one branch\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 688, __PRETTY_FUNCTION__))
;
689
690 LLVMContext &C = BB->getContext();
691 BasicBlock *TrueDest = BI->getSuccessor(0);
692 BasicBlock *FalseDest = BI->getSuccessor(1);
693 BasicBlock *CommonSucc = HoistableBranches[BI];
694 BasicBlock *HoistTarget = getOrCreateHoistedBlock(BI->getParent());
695
696 // Create hoisted versions of blocks that currently don't have them
697 auto CreateHoistedBlock = [&](BasicBlock *Orig) {
698 if (HoistDestinationMap.count(Orig))
699 return HoistDestinationMap[Orig];
700 BasicBlock *New =
701 BasicBlock::Create(C, Orig->getName() + ".licm", Orig->getParent());
702 HoistDestinationMap[Orig] = New;
703 DT->addNewBlock(New, HoistTarget);
704 if (CurLoop->getParentLoop())
705 CurLoop->getParentLoop()->addBasicBlockToLoop(New, *LI);
706 ++NumCreatedBlocks;
707 LLVM_DEBUG(dbgs() << "LICM created " << New->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "LICM created " << New->
getName() << " as hoist destination for " << Orig
->getName() << "\n"; } } while (false)
708 << " as hoist destination for " << Orig->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "LICM created " << New->
getName() << " as hoist destination for " << Orig
->getName() << "\n"; } } while (false)
709 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "LICM created " << New->
getName() << " as hoist destination for " << Orig
->getName() << "\n"; } } while (false)
;
710 return New;
711 };
712 BasicBlock *HoistTrueDest = CreateHoistedBlock(TrueDest);
713 BasicBlock *HoistFalseDest = CreateHoistedBlock(FalseDest);
714 BasicBlock *HoistCommonSucc = CreateHoistedBlock(CommonSucc);
715
716 // Link up these blocks with branches.
717 if (!HoistCommonSucc->getTerminator()) {
718 // The new common successor we've generated will branch to whatever that
719 // hoist target branched to.
720 BasicBlock *TargetSucc = HoistTarget->getSingleSuccessor();
721 assert(TargetSucc && "Expected hoist target to have a single successor")((TargetSucc && "Expected hoist target to have a single successor"
) ? static_cast<void> (0) : __assert_fail ("TargetSucc && \"Expected hoist target to have a single successor\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 721, __PRETTY_FUNCTION__))
;
722 HoistCommonSucc->moveBefore(TargetSucc);
723 BranchInst::Create(TargetSucc, HoistCommonSucc);
724 }
725 if (!HoistTrueDest->getTerminator()) {
726 HoistTrueDest->moveBefore(HoistCommonSucc);
727 BranchInst::Create(HoistCommonSucc, HoistTrueDest);
728 }
729 if (!HoistFalseDest->getTerminator()) {
730 HoistFalseDest->moveBefore(HoistCommonSucc);
731 BranchInst::Create(HoistCommonSucc, HoistFalseDest);
732 }
733
734 // If BI is being cloned to what was originally the preheader then
735 // HoistCommonSucc will now be the new preheader.
736 if (HoistTarget == InitialPreheader) {
737 // Phis in the loop header now need to use the new preheader.
738 InitialPreheader->replaceSuccessorsPhiUsesWith(HoistCommonSucc);
739 if (MSSAU)
740 MSSAU->wireOldPredecessorsToNewImmediatePredecessor(
741 HoistTarget->getSingleSuccessor(), HoistCommonSucc, {HoistTarget});
742 // The new preheader dominates the loop header.
743 DomTreeNode *PreheaderNode = DT->getNode(HoistCommonSucc);
744 DomTreeNode *HeaderNode = DT->getNode(CurLoop->getHeader());
745 DT->changeImmediateDominator(HeaderNode, PreheaderNode);
746 // The preheader hoist destination is now the new preheader, with the
747 // exception of the hoist destination of this branch.
748 for (auto &Pair : HoistDestinationMap)
749 if (Pair.second == InitialPreheader && Pair.first != BI->getParent())
750 Pair.second = HoistCommonSucc;
751 }
752
753 // Now finally clone BI.
754 ReplaceInstWithInst(
755 HoistTarget->getTerminator(),
756 BranchInst::Create(HoistTrueDest, HoistFalseDest, BI->getCondition()));
757 ++NumClonedBranches;
758
759 assert(CurLoop->getLoopPreheader() &&((CurLoop->getLoopPreheader() && "Hoisting blocks should not have destroyed preheader"
) ? static_cast<void> (0) : __assert_fail ("CurLoop->getLoopPreheader() && \"Hoisting blocks should not have destroyed preheader\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 760, __PRETTY_FUNCTION__))
760 "Hoisting blocks should not have destroyed preheader")((CurLoop->getLoopPreheader() && "Hoisting blocks should not have destroyed preheader"
) ? static_cast<void> (0) : __assert_fail ("CurLoop->getLoopPreheader() && \"Hoisting blocks should not have destroyed preheader\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 760, __PRETTY_FUNCTION__))
;
761 return HoistDestinationMap[BB];
762 }
763};
764} // namespace
765
766// Hoisting/sinking instruction out of a loop isn't always beneficial. It's only
767// only worthwhile if the destination block is actually colder than current
768// block.
769static bool worthSinkOrHoistInst(Instruction &I, BasicBlock *DstBlock,
770 OptimizationRemarkEmitter *ORE,
771 BlockFrequencyInfo *BFI) {
772 // Check block frequency only when runtime profile is available
773 // to avoid pathological cases. With static profile, lean towards
774 // hosting because it helps canonicalize the loop for vectorizer.
775 if (!DstBlock->getParent()->hasProfileData())
776 return true;
777
778 if (!HoistSinkColdnessThreshold || !BFI)
779 return true;
780
781 BasicBlock *SrcBlock = I.getParent();
782 if (BFI->getBlockFreq(DstBlock).getFrequency() / HoistSinkColdnessThreshold >
783 BFI->getBlockFreq(SrcBlock).getFrequency()) {
784 ORE->emit([&]() {
785 return OptimizationRemarkMissed(DEBUG_TYPE"licm", "SinkHoistInst", &I)
786 << "failed to sink or hoist instruction because containing block "
787 "has lower frequency than destination block";
788 });
789 return false;
790 }
791
792 return true;
793}
794
795/// Walk the specified region of the CFG (defined by all blocks dominated by
796/// the specified block, and that are in the current loop) in depth first
797/// order w.r.t the DominatorTree. This allows us to visit definitions before
798/// uses, allowing us to hoist a loop body in one pass without iteration.
799///
800bool llvm::hoistRegion(DomTreeNode *N, AAResults *AA, LoopInfo *LI,
801 DominatorTree *DT, BlockFrequencyInfo *BFI,
802 TargetLibraryInfo *TLI, Loop *CurLoop,
803 AliasSetTracker *CurAST, MemorySSAUpdater *MSSAU,
804 ScalarEvolution *SE, ICFLoopSafetyInfo *SafetyInfo,
805 SinkAndHoistLICMFlags &Flags,
806 OptimizationRemarkEmitter *ORE) {
807 // Verify inputs.
808 assert(N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr &&((N != nullptr && AA != nullptr && LI != nullptr
&& DT != nullptr && CurLoop != nullptr &&
SafetyInfo != nullptr && "Unexpected input to hoistRegion."
) ? static_cast<void> (0) : __assert_fail ("N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && \"Unexpected input to hoistRegion.\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 810, __PRETTY_FUNCTION__))
809 CurLoop != nullptr && SafetyInfo != nullptr &&((N != nullptr && AA != nullptr && LI != nullptr
&& DT != nullptr && CurLoop != nullptr &&
SafetyInfo != nullptr && "Unexpected input to hoistRegion."
) ? static_cast<void> (0) : __assert_fail ("N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && \"Unexpected input to hoistRegion.\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 810, __PRETTY_FUNCTION__))
810 "Unexpected input to hoistRegion.")((N != nullptr && AA != nullptr && LI != nullptr
&& DT != nullptr && CurLoop != nullptr &&
SafetyInfo != nullptr && "Unexpected input to hoistRegion."
) ? static_cast<void> (0) : __assert_fail ("N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && \"Unexpected input to hoistRegion.\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 810, __PRETTY_FUNCTION__))
;
811 assert(((CurAST != nullptr) ^ (MSSAU != nullptr)) &&((((CurAST != nullptr) ^ (MSSAU != nullptr)) && "Either AliasSetTracker or MemorySSA should be initialized."
) ? static_cast<void> (0) : __assert_fail ("((CurAST != nullptr) ^ (MSSAU != nullptr)) && \"Either AliasSetTracker or MemorySSA should be initialized.\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 812, __PRETTY_FUNCTION__))
812 "Either AliasSetTracker or MemorySSA should be initialized.")((((CurAST != nullptr) ^ (MSSAU != nullptr)) && "Either AliasSetTracker or MemorySSA should be initialized."
) ? static_cast<void> (0) : __assert_fail ("((CurAST != nullptr) ^ (MSSAU != nullptr)) && \"Either AliasSetTracker or MemorySSA should be initialized.\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 812, __PRETTY_FUNCTION__))
;
813
814 ControlFlowHoister CFH(LI, DT, CurLoop, MSSAU);
815
816 // Keep track of instructions that have been hoisted, as they may need to be
817 // re-hoisted if they end up not dominating all of their uses.
818 SmallVector<Instruction *, 16> HoistedInstructions;
819
820 // For PHI hoisting to work we need to hoist blocks before their successors.
821 // We can do this by iterating through the blocks in the loop in reverse
822 // post-order.
823 LoopBlocksRPO Worklist(CurLoop);
824 Worklist.perform(LI);
825 bool Changed = false;
826 for (BasicBlock *BB : Worklist) {
827 // Only need to process the contents of this block if it is not part of a
828 // subloop (which would already have been processed).
829 if (inSubLoop(BB, CurLoop, LI))
830 continue;
831
832 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E;) {
833 Instruction &I = *II++;
834 // Try constant folding this instruction. If all the operands are
835 // constants, it is technically hoistable, but it would be better to
836 // just fold it.
837 if (Constant *C = ConstantFoldInstruction(
838 &I, I.getModule()->getDataLayout(), TLI)) {
839 LLVM_DEBUG(dbgs() << "LICM folding inst: " << I << " --> " << *Cdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "LICM folding inst: " << I <<
" --> " << *C << '\n'; } } while (false)
840 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "LICM folding inst: " << I <<
" --> " << *C << '\n'; } } while (false)
;
841 if (CurAST)
842 CurAST->copyValue(&I, C);
843 // FIXME MSSA: Such replacements may make accesses unoptimized (D51960).
844 I.replaceAllUsesWith(C);
845 if (isInstructionTriviallyDead(&I, TLI))
846 eraseInstruction(I, *SafetyInfo, CurAST, MSSAU);
847 Changed = true;
848 continue;
849 }
850
851 // Try hoisting the instruction out to the preheader. We can only do
852 // this if all of the operands of the instruction are loop invariant and
853 // if it is safe to hoist the instruction. We also check block frequency
854 // to make sure instruction only gets hoisted into colder blocks.
855 // TODO: It may be safe to hoist if we are hoisting to a conditional block
856 // and we have accurately duplicated the control flow from the loop header
857 // to that block.
858 if (CurLoop->hasLoopInvariantOperands(&I) &&
859 canSinkOrHoistInst(I, AA, DT, CurLoop, CurAST, MSSAU, true, &Flags,
860 ORE) &&
861 worthSinkOrHoistInst(I, CurLoop->getLoopPreheader(), ORE, BFI) &&
862 isSafeToExecuteUnconditionally(
863 I, DT, CurLoop, SafetyInfo, ORE,
864 CurLoop->getLoopPreheader()->getTerminator())) {
865 hoist(I, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), SafetyInfo,
866 MSSAU, SE, ORE);
867 HoistedInstructions.push_back(&I);
868 Changed = true;
869 continue;
870 }
871
872 // Attempt to remove floating point division out of the loop by
873 // converting it to a reciprocal multiplication.
874 if (I.getOpcode() == Instruction::FDiv && I.hasAllowReciprocal() &&
875 CurLoop->isLoopInvariant(I.getOperand(1))) {
876 auto Divisor = I.getOperand(1);
877 auto One = llvm::ConstantFP::get(Divisor->getType(), 1.0);
878 auto ReciprocalDivisor = BinaryOperator::CreateFDiv(One, Divisor);
879 ReciprocalDivisor->setFastMathFlags(I.getFastMathFlags());
880 SafetyInfo->insertInstructionTo(ReciprocalDivisor, I.getParent());
881 ReciprocalDivisor->insertBefore(&I);
882
883 auto Product =
884 BinaryOperator::CreateFMul(I.getOperand(0), ReciprocalDivisor);
885 Product->setFastMathFlags(I.getFastMathFlags());
886 SafetyInfo->insertInstructionTo(Product, I.getParent());
887 Product->insertAfter(&I);
888 I.replaceAllUsesWith(Product);
889 eraseInstruction(I, *SafetyInfo, CurAST, MSSAU);
890
891 hoist(*ReciprocalDivisor, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB),
892 SafetyInfo, MSSAU, SE, ORE);
893 HoistedInstructions.push_back(ReciprocalDivisor);
894 Changed = true;
895 continue;
896 }
897
898 auto IsInvariantStart = [&](Instruction &I) {
899 using namespace PatternMatch;
900 return I.use_empty() &&
901 match(&I, m_Intrinsic<Intrinsic::invariant_start>());
902 };
903 auto MustExecuteWithoutWritesBefore = [&](Instruction &I) {
904 return SafetyInfo->isGuaranteedToExecute(I, DT, CurLoop) &&
905 SafetyInfo->doesNotWriteMemoryBefore(I, CurLoop);
906 };
907 if ((IsInvariantStart(I) || isGuard(&I)) &&
908 CurLoop->hasLoopInvariantOperands(&I) &&
909 MustExecuteWithoutWritesBefore(I)) {
910 hoist(I, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), SafetyInfo,
911 MSSAU, SE, ORE);
912 HoistedInstructions.push_back(&I);
913 Changed = true;
914 continue;
915 }
916
917 if (PHINode *PN = dyn_cast<PHINode>(&I)) {
918 if (CFH.canHoistPHI(PN)) {
919 // Redirect incoming blocks first to ensure that we create hoisted
920 // versions of those blocks before we hoist the phi.
921 for (unsigned int i = 0; i < PN->getNumIncomingValues(); ++i)
922 PN->setIncomingBlock(
923 i, CFH.getOrCreateHoistedBlock(PN->getIncomingBlock(i)));
924 hoist(*PN, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), SafetyInfo,
925 MSSAU, SE, ORE);
926 assert(DT->dominates(PN, BB) && "Conditional PHIs not expected")((DT->dominates(PN, BB) && "Conditional PHIs not expected"
) ? static_cast<void> (0) : __assert_fail ("DT->dominates(PN, BB) && \"Conditional PHIs not expected\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 926, __PRETTY_FUNCTION__))
;
927 Changed = true;
928 continue;
929 }
930 }
931
932 // Remember possibly hoistable branches so we can actually hoist them
933 // later if needed.
934 if (BranchInst *BI = dyn_cast<BranchInst>(&I))
935 CFH.registerPossiblyHoistableBranch(BI);
936 }
937 }
938
939 // If we hoisted instructions to a conditional block they may not dominate
940 // their uses that weren't hoisted (such as phis where some operands are not
941 // loop invariant). If so make them unconditional by moving them to their
942 // immediate dominator. We iterate through the instructions in reverse order
943 // which ensures that when we rehoist an instruction we rehoist its operands,
944 // and also keep track of where in the block we are rehoisting to to make sure
945 // that we rehoist instructions before the instructions that use them.
946 Instruction *HoistPoint = nullptr;
947 if (ControlFlowHoisting) {
948 for (Instruction *I : reverse(HoistedInstructions)) {
949 if (!llvm::all_of(I->uses(),
950 [&](Use &U) { return DT->dominates(I, U); })) {
951 BasicBlock *Dominator =
952 DT->getNode(I->getParent())->getIDom()->getBlock();
953 if (!HoistPoint || !DT->dominates(HoistPoint->getParent(), Dominator)) {
954 if (HoistPoint)
955 assert(DT->dominates(Dominator, HoistPoint->getParent()) &&((DT->dominates(Dominator, HoistPoint->getParent()) &&
"New hoist point expected to dominate old hoist point") ? static_cast
<void> (0) : __assert_fail ("DT->dominates(Dominator, HoistPoint->getParent()) && \"New hoist point expected to dominate old hoist point\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 956, __PRETTY_FUNCTION__))
956 "New hoist point expected to dominate old hoist point")((DT->dominates(Dominator, HoistPoint->getParent()) &&
"New hoist point expected to dominate old hoist point") ? static_cast
<void> (0) : __assert_fail ("DT->dominates(Dominator, HoistPoint->getParent()) && \"New hoist point expected to dominate old hoist point\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 956, __PRETTY_FUNCTION__))
;
957 HoistPoint = Dominator->getTerminator();
958 }
959 LLVM_DEBUG(dbgs() << "LICM rehoisting to "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "LICM rehoisting to " << HoistPoint
->getParent()->getName() << ": " << *I <<
"\n"; } } while (false)
960 << HoistPoint->getParent()->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "LICM rehoisting to " << HoistPoint
->getParent()->getName() << ": " << *I <<
"\n"; } } while (false)
961 << ": " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "LICM rehoisting to " << HoistPoint
->getParent()->getName() << ": " << *I <<
"\n"; } } while (false)
;
962 moveInstructionBefore(*I, *HoistPoint, *SafetyInfo, MSSAU, SE);
963 HoistPoint = I;
964 Changed = true;
965 }
966 }
967 }
968 if (MSSAU && VerifyMemorySSA)
969 MSSAU->getMemorySSA()->verifyMemorySSA();
970
971 // Now that we've finished hoisting make sure that LI and DT are still
972 // valid.
973#ifdef EXPENSIVE_CHECKS
974 if (Changed) {
975 assert(DT->verify(DominatorTree::VerificationLevel::Fast) &&((DT->verify(DominatorTree::VerificationLevel::Fast) &&
"Dominator tree verification failed") ? static_cast<void>
(0) : __assert_fail ("DT->verify(DominatorTree::VerificationLevel::Fast) && \"Dominator tree verification failed\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 976, __PRETTY_FUNCTION__))
976 "Dominator tree verification failed")((DT->verify(DominatorTree::VerificationLevel::Fast) &&
"Dominator tree verification failed") ? static_cast<void>
(0) : __assert_fail ("DT->verify(DominatorTree::VerificationLevel::Fast) && \"Dominator tree verification failed\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 976, __PRETTY_FUNCTION__))
;
977 LI->verify(*DT);
978 }
979#endif
980
981 return Changed;
982}
983
984// Return true if LI is invariant within scope of the loop. LI is invariant if
985// CurLoop is dominated by an invariant.start representing the same memory
986// location and size as the memory location LI loads from, and also the
987// invariant.start has no uses.
988static bool isLoadInvariantInLoop(LoadInst *LI, DominatorTree *DT,
989 Loop *CurLoop) {
990 Value *Addr = LI->getOperand(0);
991 const DataLayout &DL = LI->getModule()->getDataLayout();
992 const TypeSize LocSizeInBits = DL.getTypeSizeInBits(LI->getType());
993
994 // It is not currently possible for clang to generate an invariant.start
995 // intrinsic with scalable vector types because we don't support thread local
996 // sizeless types and we don't permit sizeless types in structs or classes.
997 // Furthermore, even if support is added for this in future the intrinsic
998 // itself is defined to have a size of -1 for variable sized objects. This
999 // makes it impossible to verify if the intrinsic envelops our region of
1000 // interest. For example, both <vscale x 32 x i8> and <vscale x 16 x i8>
1001 // types would have a -1 parameter, but the former is clearly double the size
1002 // of the latter.
1003 if (LocSizeInBits.isScalable())
27
Assuming the condition is true
28
Taking true branch
1004 return false;
29
Returning zero, which participates in a condition later
1005
1006 // if the type is i8 addrspace(x)*, we know this is the type of
1007 // llvm.invariant.start operand
1008 auto *PtrInt8Ty = PointerType::get(Type::getInt8Ty(LI->getContext()),
1009 LI->getPointerAddressSpace());
1010 unsigned BitcastsVisited = 0;
1011 // Look through bitcasts until we reach the i8* type (this is invariant.start
1012 // operand type).
1013 while (Addr->getType() != PtrInt8Ty) {
1014 auto *BC = dyn_cast<BitCastInst>(Addr);
1015 // Avoid traversing high number of bitcast uses.
1016 if (++BitcastsVisited > MaxNumUsesTraversed || !BC)
1017 return false;
1018 Addr = BC->getOperand(0);
1019 }
1020
1021 unsigned UsesVisited = 0;
1022 // Traverse all uses of the load operand value, to see if invariant.start is
1023 // one of the uses, and whether it dominates the load instruction.
1024 for (auto *U : Addr->users()) {
1025 // Avoid traversing for Load operand with high number of users.
1026 if (++UsesVisited > MaxNumUsesTraversed)
1027 return false;
1028 IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
1029 // If there are escaping uses of invariant.start instruction, the load maybe
1030 // non-invariant.
1031 if (!II || II->getIntrinsicID() != Intrinsic::invariant_start ||
1032 !II->use_empty())
1033 continue;
1034 ConstantInt *InvariantSize = cast<ConstantInt>(II->getArgOperand(0));
1035 // The intrinsic supports having a -1 argument for variable sized objects
1036 // so we should check for that here.
1037 if (InvariantSize->isNegative())
1038 continue;
1039 uint64_t InvariantSizeInBits = InvariantSize->getSExtValue() * 8;
1040 // Confirm the invariant.start location size contains the load operand size
1041 // in bits. Also, the invariant.start should dominate the load, and we
1042 // should not hoist the load out of a loop that contains this dominating
1043 // invariant.start.
1044 if (LocSizeInBits.getFixedSize() <= InvariantSizeInBits &&
1045 DT->properlyDominates(II->getParent(), CurLoop->getHeader()))
1046 return true;
1047 }
1048
1049 return false;
1050}
1051
1052namespace {
1053/// Return true if-and-only-if we know how to (mechanically) both hoist and
1054/// sink a given instruction out of a loop. Does not address legality
1055/// concerns such as aliasing or speculation safety.
1056bool isHoistableAndSinkableInst(Instruction &I) {
1057 // Only these instructions are hoistable/sinkable.
1058 return (isa<LoadInst>(I) || isa<StoreInst>(I) || isa<CallInst>(I) ||
2
Assuming 'I' is a 'LoadInst'
3
Returning the value 1, which participates in a condition later
1059 isa<FenceInst>(I) || isa<CastInst>(I) || isa<UnaryOperator>(I) ||
1060 isa<BinaryOperator>(I) || isa<SelectInst>(I) ||
1061 isa<GetElementPtrInst>(I) || isa<CmpInst>(I) ||
1062 isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
1063 isa<ShuffleVectorInst>(I) || isa<ExtractValueInst>(I) ||
1064 isa<InsertValueInst>(I) || isa<FreezeInst>(I));
1065}
1066/// Return true if all of the alias sets within this AST are known not to
1067/// contain a Mod, or if MSSA knows thare are no MemoryDefs in the loop.
1068bool isReadOnly(AliasSetTracker *CurAST, const MemorySSAUpdater *MSSAU,
1069 const Loop *L) {
1070 if (CurAST) {
1071 for (AliasSet &AS : *CurAST) {
1072 if (!AS.isForwardingAliasSet() && AS.isMod()) {
1073 return false;
1074 }
1075 }
1076 return true;
1077 } else { /*MSSAU*/
1078 for (auto *BB : L->getBlocks())
1079 if (MSSAU->getMemorySSA()->getBlockDefs(BB))
1080 return false;
1081 return true;
1082 }
1083}
1084
1085/// Return true if I is the only Instruction with a MemoryAccess in L.
1086bool isOnlyMemoryAccess(const Instruction *I, const Loop *L,
1087 const MemorySSAUpdater *MSSAU) {
1088 for (auto *BB : L->getBlocks())
1089 if (auto *Accs = MSSAU->getMemorySSA()->getBlockAccesses(BB)) {
1090 int NotAPhi = 0;
1091 for (const auto &Acc : *Accs) {
1092 if (isa<MemoryPhi>(&Acc))
1093 continue;
1094 const auto *MUD = cast<MemoryUseOrDef>(&Acc);
1095 if (MUD->getMemoryInst() != I || NotAPhi++ == 1)
1096 return false;
1097 }
1098 }
1099 return true;
1100}
1101}
1102
1103bool llvm::canSinkOrHoistInst(Instruction &I, AAResults *AA, DominatorTree *DT,
1104 Loop *CurLoop, AliasSetTracker *CurAST,
1105 MemorySSAUpdater *MSSAU,
1106 bool TargetExecutesOncePerLoop,
1107 SinkAndHoistLICMFlags *Flags,
1108 OptimizationRemarkEmitter *ORE) {
1109 // If we don't understand the instruction, bail early.
1110 if (!isHoistableAndSinkableInst(I))
1
Calling 'isHoistableAndSinkableInst'
4
Returning from 'isHoistableAndSinkableInst'
5
Taking false branch
1111 return false;
1112
1113 MemorySSA *MSSA = MSSAU ? MSSAU->getMemorySSA() : nullptr;
6
Assuming 'MSSAU' is null
7
'?' condition is false
8
'MSSA' initialized to a null pointer value
1114 if (MSSA
8.1
'MSSA' is null
8.1
'MSSA' is null
8.1
'MSSA' is null
)
9
Taking false branch
1115 assert(Flags != nullptr && "Flags cannot be null.")((Flags != nullptr && "Flags cannot be null.") ? static_cast
<void> (0) : __assert_fail ("Flags != nullptr && \"Flags cannot be null.\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1115, __PRETTY_FUNCTION__))
;
1116
1117 // Loads have extra constraints we have to verify before we can hoist them.
1118 if (LoadInst *LI
10.1
'LI' is non-null
10.1
'LI' is non-null
10.1
'LI' is non-null
= dyn_cast<LoadInst>(&I)) {
10
Assuming the object is a 'LoadInst'
11
Taking true branch
1119 if (!LI->isUnordered())
12
Calling 'LoadInst::isUnordered'
17
Returning from 'LoadInst::isUnordered'
18
Taking false branch
1120 return false; // Don't sink/hoist volatile or ordered atomic loads!
1121
1122 // Loads from constant memory are always safe to move, even if they end up
1123 // in the same alias set as something that ends up being modified.
1124 if (AA->pointsToConstantMemory(LI->getOperand(0)))
19
Assuming the condition is false
20
Taking false branch
1125 return true;
1126 if (LI->hasMetadata(LLVMContext::MD_invariant_load))
21
Calling 'Instruction::hasMetadata'
23
Returning from 'Instruction::hasMetadata'
24
Taking false branch
1127 return true;
1128
1129 if (LI->isAtomic() && !TargetExecutesOncePerLoop)
25
Assuming the condition is false
1130 return false; // Don't risk duplicating unordered loads
1131
1132 // This checks for an invariant.start dominating the load.
1133 if (isLoadInvariantInLoop(LI, DT, CurLoop))
26
Calling 'isLoadInvariantInLoop'
30
Returning from 'isLoadInvariantInLoop'
31
Taking false branch
1134 return true;
1135
1136 bool Invalidated;
1137 if (CurAST)
32
Assuming 'CurAST' is null
33
Taking false branch
1138 Invalidated = pointerInvalidatedByLoop(MemoryLocation::get(LI), CurAST,
1139 CurLoop, AA);
1140 else
1141 Invalidated = pointerInvalidatedByLoopWithMSSA(
1142 MSSA, cast<MemoryUse>(MSSA->getMemoryAccess(LI)), CurLoop, *Flags);
34
Called C++ object pointer is null
1143 // Check loop-invariant address because this may also be a sinkable load
1144 // whose address is not necessarily loop-invariant.
1145 if (ORE && Invalidated && CurLoop->isLoopInvariant(LI->getPointerOperand()))
1146 ORE->emit([&]() {
1147 return OptimizationRemarkMissed(
1148 DEBUG_TYPE"licm", "LoadWithLoopInvariantAddressInvalidated", LI)
1149 << "failed to move load with loop-invariant address "
1150 "because the loop may invalidate its value";
1151 });
1152
1153 return !Invalidated;
1154 } else if (CallInst *CI = dyn_cast<CallInst>(&I)) {
1155 // Don't sink or hoist dbg info; it's legal, but not useful.
1156 if (isa<DbgInfoIntrinsic>(I))
1157 return false;
1158
1159 // Don't sink calls which can throw.
1160 if (CI->mayThrow())
1161 return false;
1162
1163 using namespace PatternMatch;
1164 if (match(CI, m_Intrinsic<Intrinsic::assume>()))
1165 // Assumes don't actually alias anything or throw
1166 return true;
1167
1168 if (match(CI, m_Intrinsic<Intrinsic::experimental_widenable_condition>()))
1169 // Widenable conditions don't actually alias anything or throw
1170 return true;
1171
1172 // Handle simple cases by querying alias analysis.
1173 FunctionModRefBehavior Behavior = AA->getModRefBehavior(CI);
1174 if (Behavior == FMRB_DoesNotAccessMemory)
1175 return true;
1176 if (AAResults::onlyReadsMemory(Behavior)) {
1177 // A readonly argmemonly function only reads from memory pointed to by
1178 // it's arguments with arbitrary offsets. If we can prove there are no
1179 // writes to this memory in the loop, we can hoist or sink.
1180 if (AAResults::onlyAccessesArgPointees(Behavior)) {
1181 // TODO: expand to writeable arguments
1182 for (Value *Op : CI->arg_operands())
1183 if (Op->getType()->isPointerTy()) {
1184 bool Invalidated;
1185 if (CurAST)
1186 Invalidated = pointerInvalidatedByLoop(
1187 MemoryLocation(Op, LocationSize::unknown(), AAMDNodes()),
1188 CurAST, CurLoop, AA);
1189 else
1190 Invalidated = pointerInvalidatedByLoopWithMSSA(
1191 MSSA, cast<MemoryUse>(MSSA->getMemoryAccess(CI)), CurLoop,
1192 *Flags);
1193 if (Invalidated)
1194 return false;
1195 }
1196 return true;
1197 }
1198
1199 // If this call only reads from memory and there are no writes to memory
1200 // in the loop, we can hoist or sink the call as appropriate.
1201 if (isReadOnly(CurAST, MSSAU, CurLoop))
1202 return true;
1203 }
1204
1205 // FIXME: This should use mod/ref information to see if we can hoist or
1206 // sink the call.
1207
1208 return false;
1209 } else if (auto *FI = dyn_cast<FenceInst>(&I)) {
1210 // Fences alias (most) everything to provide ordering. For the moment,
1211 // just give up if there are any other memory operations in the loop.
1212 if (CurAST) {
1213 auto Begin = CurAST->begin();
1214 assert(Begin != CurAST->end() && "must contain FI")((Begin != CurAST->end() && "must contain FI") ? static_cast
<void> (0) : __assert_fail ("Begin != CurAST->end() && \"must contain FI\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1214, __PRETTY_FUNCTION__))
;
1215 if (std::next(Begin) != CurAST->end())
1216 // constant memory for instance, TODO: handle better
1217 return false;
1218 auto *UniqueI = Begin->getUniqueInstruction();
1219 if (!UniqueI)
1220 // other memory op, give up
1221 return false;
1222 (void)FI; // suppress unused variable warning
1223 assert(UniqueI == FI && "AS must contain FI")((UniqueI == FI && "AS must contain FI") ? static_cast
<void> (0) : __assert_fail ("UniqueI == FI && \"AS must contain FI\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1223, __PRETTY_FUNCTION__))
;
1224 return true;
1225 } else // MSSAU
1226 return isOnlyMemoryAccess(FI, CurLoop, MSSAU);
1227 } else if (auto *SI = dyn_cast<StoreInst>(&I)) {
1228 if (!SI->isUnordered())
1229 return false; // Don't sink/hoist volatile or ordered atomic store!
1230
1231 // We can only hoist a store that we can prove writes a value which is not
1232 // read or overwritten within the loop. For those cases, we fallback to
1233 // load store promotion instead. TODO: We can extend this to cases where
1234 // there is exactly one write to the location and that write dominates an
1235 // arbitrary number of reads in the loop.
1236 if (CurAST) {
1237 auto &AS = CurAST->getAliasSetFor(MemoryLocation::get(SI));
1238
1239 if (AS.isRef() || !AS.isMustAlias())
1240 // Quick exit test, handled by the full path below as well.
1241 return false;
1242 auto *UniqueI = AS.getUniqueInstruction();
1243 if (!UniqueI)
1244 // other memory op, give up
1245 return false;
1246 assert(UniqueI == SI && "AS must contain SI")((UniqueI == SI && "AS must contain SI") ? static_cast
<void> (0) : __assert_fail ("UniqueI == SI && \"AS must contain SI\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1246, __PRETTY_FUNCTION__))
;
1247 return true;
1248 } else { // MSSAU
1249 if (isOnlyMemoryAccess(SI, CurLoop, MSSAU))
1250 return true;
1251 // If there are more accesses than the Promotion cap, give up, we're not
1252 // walking a list that long.
1253 if (Flags->NoOfMemAccTooLarge)
1254 return false;
1255 // Check store only if there's still "quota" to check clobber.
1256 if (Flags->LicmMssaOptCounter >= Flags->LicmMssaOptCap)
1257 return false;
1258 // If there are interfering Uses (i.e. their defining access is in the
1259 // loop), or ordered loads (stored as Defs!), don't move this store.
1260 // Could do better here, but this is conservatively correct.
1261 // TODO: Cache set of Uses on the first walk in runOnLoop, update when
1262 // moving accesses. Can also extend to dominating uses.
1263 auto *SIMD = MSSA->getMemoryAccess(SI);
1264 for (auto *BB : CurLoop->getBlocks())
1265 if (auto *Accesses = MSSA->getBlockAccesses(BB)) {
1266 for (const auto &MA : *Accesses)
1267 if (const auto *MU = dyn_cast<MemoryUse>(&MA)) {
1268 auto *MD = MU->getDefiningAccess();
1269 if (!MSSA->isLiveOnEntryDef(MD) &&
1270 CurLoop->contains(MD->getBlock()))
1271 return false;
1272 // Disable hoisting past potentially interfering loads. Optimized
1273 // Uses may point to an access outside the loop, as getClobbering
1274 // checks the previous iteration when walking the backedge.
1275 // FIXME: More precise: no Uses that alias SI.
1276 if (!Flags->IsSink && !MSSA->dominates(SIMD, MU))
1277 return false;
1278 } else if (const auto *MD = dyn_cast<MemoryDef>(&MA)) {
1279 if (auto *LI = dyn_cast<LoadInst>(MD->getMemoryInst())) {
1280 (void)LI; // Silence warning.
1281 assert(!LI->isUnordered() && "Expected unordered load")((!LI->isUnordered() && "Expected unordered load")
? static_cast<void> (0) : __assert_fail ("!LI->isUnordered() && \"Expected unordered load\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1281, __PRETTY_FUNCTION__))
;
1282 return false;
1283 }
1284 // Any call, while it may not be clobbering SI, it may be a use.
1285 if (auto *CI = dyn_cast<CallInst>(MD->getMemoryInst())) {
1286 // Check if the call may read from the memory locattion written
1287 // to by SI. Check CI's attributes and arguments; the number of
1288 // such checks performed is limited above by NoOfMemAccTooLarge.
1289 ModRefInfo MRI = AA->getModRefInfo(CI, MemoryLocation::get(SI));
1290 if (isModOrRefSet(MRI))
1291 return false;
1292 }
1293 }
1294 }
1295
1296 auto *Source = MSSA->getSkipSelfWalker()->getClobberingMemoryAccess(SI);
1297 Flags->LicmMssaOptCounter++;
1298 // If there are no clobbering Defs in the loop, store is safe to hoist.
1299 return MSSA->isLiveOnEntryDef(Source) ||
1300 !CurLoop->contains(Source->getBlock());
1301 }
1302 }
1303
1304 assert(!I.mayReadOrWriteMemory() && "unhandled aliasing")((!I.mayReadOrWriteMemory() && "unhandled aliasing") ?
static_cast<void> (0) : __assert_fail ("!I.mayReadOrWriteMemory() && \"unhandled aliasing\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1304, __PRETTY_FUNCTION__))
;
1305
1306 // We've established mechanical ability and aliasing, it's up to the caller
1307 // to check fault safety
1308 return true;
1309}
1310
1311/// Returns true if a PHINode is a trivially replaceable with an
1312/// Instruction.
1313/// This is true when all incoming values are that instruction.
1314/// This pattern occurs most often with LCSSA PHI nodes.
1315///
1316static bool isTriviallyReplaceablePHI(const PHINode &PN, const Instruction &I) {
1317 for (const Value *IncValue : PN.incoming_values())
1318 if (IncValue != &I)
1319 return false;
1320
1321 return true;
1322}
1323
1324/// Return true if the instruction is free in the loop.
1325static bool isFreeInLoop(const Instruction &I, const Loop *CurLoop,
1326 const TargetTransformInfo *TTI) {
1327
1328 if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(&I)) {
1329 if (TTI->getUserCost(GEP, TargetTransformInfo::TCK_SizeAndLatency) !=
1330 TargetTransformInfo::TCC_Free)
1331 return false;
1332 // For a GEP, we cannot simply use getUserCost because currently it
1333 // optimistically assume that a GEP will fold into addressing mode
1334 // regardless of its users.
1335 const BasicBlock *BB = GEP->getParent();
1336 for (const User *U : GEP->users()) {
1337 const Instruction *UI = cast<Instruction>(U);
1338 if (CurLoop->contains(UI) &&
1339 (BB != UI->getParent() ||
1340 (!isa<StoreInst>(UI) && !isa<LoadInst>(UI))))
1341 return false;
1342 }
1343 return true;
1344 } else
1345 return TTI->getUserCost(&I, TargetTransformInfo::TCK_SizeAndLatency) ==
1346 TargetTransformInfo::TCC_Free;
1347}
1348
1349/// Return true if the only users of this instruction are outside of
1350/// the loop. If this is true, we can sink the instruction to the exit
1351/// blocks of the loop.
1352///
1353/// We also return true if the instruction could be folded away in lowering.
1354/// (e.g., a GEP can be folded into a load as an addressing mode in the loop).
1355static bool isNotUsedOrFreeInLoop(const Instruction &I, const Loop *CurLoop,
1356 const LoopSafetyInfo *SafetyInfo,
1357 TargetTransformInfo *TTI, bool &FreeInLoop) {
1358 const auto &BlockColors = SafetyInfo->getBlockColors();
1359 bool IsFree = isFreeInLoop(I, CurLoop, TTI);
1360 for (const User *U : I.users()) {
1361 const Instruction *UI = cast<Instruction>(U);
1362 if (const PHINode *PN = dyn_cast<PHINode>(UI)) {
1363 const BasicBlock *BB = PN->getParent();
1364 // We cannot sink uses in catchswitches.
1365 if (isa<CatchSwitchInst>(BB->getTerminator()))
1366 return false;
1367
1368 // We need to sink a callsite to a unique funclet. Avoid sinking if the
1369 // phi use is too muddled.
1370 if (isa<CallInst>(I))
1371 if (!BlockColors.empty() &&
1372 BlockColors.find(const_cast<BasicBlock *>(BB))->second.size() != 1)
1373 return false;
1374 }
1375
1376 if (CurLoop->contains(UI)) {
1377 if (IsFree) {
1378 FreeInLoop = true;
1379 continue;
1380 }
1381 return false;
1382 }
1383 }
1384 return true;
1385}
1386
1387static Instruction *cloneInstructionInExitBlock(
1388 Instruction &I, BasicBlock &ExitBlock, PHINode &PN, const LoopInfo *LI,
1389 const LoopSafetyInfo *SafetyInfo, MemorySSAUpdater *MSSAU) {
1390 Instruction *New;
1391 if (auto *CI = dyn_cast<CallInst>(&I)) {
1392 const auto &BlockColors = SafetyInfo->getBlockColors();
1393
1394 // Sinking call-sites need to be handled differently from other
1395 // instructions. The cloned call-site needs a funclet bundle operand
1396 // appropriate for its location in the CFG.
1397 SmallVector<OperandBundleDef, 1> OpBundles;
1398 for (unsigned BundleIdx = 0, BundleEnd = CI->getNumOperandBundles();
1399 BundleIdx != BundleEnd; ++BundleIdx) {
1400 OperandBundleUse Bundle = CI->getOperandBundleAt(BundleIdx);
1401 if (Bundle.getTagID() == LLVMContext::OB_funclet)
1402 continue;
1403
1404 OpBundles.emplace_back(Bundle);
1405 }
1406
1407 if (!BlockColors.empty()) {
1408 const ColorVector &CV = BlockColors.find(&ExitBlock)->second;
1409 assert(CV.size() == 1 && "non-unique color for exit block!")((CV.size() == 1 && "non-unique color for exit block!"
) ? static_cast<void> (0) : __assert_fail ("CV.size() == 1 && \"non-unique color for exit block!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1409, __PRETTY_FUNCTION__))
;
1410 BasicBlock *BBColor = CV.front();
1411 Instruction *EHPad = BBColor->getFirstNonPHI();
1412 if (EHPad->isEHPad())
1413 OpBundles.emplace_back("funclet", EHPad);
1414 }
1415
1416 New = CallInst::Create(CI, OpBundles);
1417 } else {
1418 New = I.clone();
1419 }
1420
1421 ExitBlock.getInstList().insert(ExitBlock.getFirstInsertionPt(), New);
1422 if (!I.getName().empty())
1423 New->setName(I.getName() + ".le");
1424
1425 if (MSSAU && MSSAU->getMemorySSA()->getMemoryAccess(&I)) {
1426 // Create a new MemoryAccess and let MemorySSA set its defining access.
1427 MemoryAccess *NewMemAcc = MSSAU->createMemoryAccessInBB(
1428 New, nullptr, New->getParent(), MemorySSA::Beginning);
1429 if (NewMemAcc) {
1430 if (auto *MemDef = dyn_cast<MemoryDef>(NewMemAcc))
1431 MSSAU->insertDef(MemDef, /*RenameUses=*/true);
1432 else {
1433 auto *MemUse = cast<MemoryUse>(NewMemAcc);
1434 MSSAU->insertUse(MemUse, /*RenameUses=*/true);
1435 }
1436 }
1437 }
1438
1439 // Build LCSSA PHI nodes for any in-loop operands. Note that this is
1440 // particularly cheap because we can rip off the PHI node that we're
1441 // replacing for the number and blocks of the predecessors.
1442 // OPT: If this shows up in a profile, we can instead finish sinking all
1443 // invariant instructions, and then walk their operands to re-establish
1444 // LCSSA. That will eliminate creating PHI nodes just to nuke them when
1445 // sinking bottom-up.
1446 for (User::op_iterator OI = New->op_begin(), OE = New->op_end(); OI != OE;
1447 ++OI)
1448 if (Instruction *OInst = dyn_cast<Instruction>(*OI))
1449 if (Loop *OLoop = LI->getLoopFor(OInst->getParent()))
1450 if (!OLoop->contains(&PN)) {
1451 PHINode *OpPN =
1452 PHINode::Create(OInst->getType(), PN.getNumIncomingValues(),
1453 OInst->getName() + ".lcssa", &ExitBlock.front());
1454 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
1455 OpPN->addIncoming(OInst, PN.getIncomingBlock(i));
1456 *OI = OpPN;
1457 }
1458 return New;
1459}
1460
1461static void eraseInstruction(Instruction &I, ICFLoopSafetyInfo &SafetyInfo,
1462 AliasSetTracker *AST, MemorySSAUpdater *MSSAU) {
1463 if (AST)
1464 AST->deleteValue(&I);
1465 if (MSSAU)
1466 MSSAU->removeMemoryAccess(&I);
1467 SafetyInfo.removeInstruction(&I);
1468 I.eraseFromParent();
1469}
1470
1471static void moveInstructionBefore(Instruction &I, Instruction &Dest,
1472 ICFLoopSafetyInfo &SafetyInfo,
1473 MemorySSAUpdater *MSSAU,
1474 ScalarEvolution *SE) {
1475 SafetyInfo.removeInstruction(&I);
1476 SafetyInfo.insertInstructionTo(&I, Dest.getParent());
1477 I.moveBefore(&Dest);
1478 if (MSSAU)
1479 if (MemoryUseOrDef *OldMemAcc = cast_or_null<MemoryUseOrDef>(
1480 MSSAU->getMemorySSA()->getMemoryAccess(&I)))
1481 MSSAU->moveToPlace(OldMemAcc, Dest.getParent(),
1482 MemorySSA::BeforeTerminator);
1483 if (SE)
1484 SE->forgetValue(&I);
1485}
1486
1487static Instruction *sinkThroughTriviallyReplaceablePHI(
1488 PHINode *TPN, Instruction *I, LoopInfo *LI,
1489 SmallDenseMap<BasicBlock *, Instruction *, 32> &SunkCopies,
1490 const LoopSafetyInfo *SafetyInfo, const Loop *CurLoop,
1491 MemorySSAUpdater *MSSAU) {
1492 assert(isTriviallyReplaceablePHI(*TPN, *I) &&((isTriviallyReplaceablePHI(*TPN, *I) && "Expect only trivially replaceable PHI"
) ? static_cast<void> (0) : __assert_fail ("isTriviallyReplaceablePHI(*TPN, *I) && \"Expect only trivially replaceable PHI\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1493, __PRETTY_FUNCTION__))
1493 "Expect only trivially replaceable PHI")((isTriviallyReplaceablePHI(*TPN, *I) && "Expect only trivially replaceable PHI"
) ? static_cast<void> (0) : __assert_fail ("isTriviallyReplaceablePHI(*TPN, *I) && \"Expect only trivially replaceable PHI\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1493, __PRETTY_FUNCTION__))
;
1494 BasicBlock *ExitBlock = TPN->getParent();
1495 Instruction *New;
1496 auto It = SunkCopies.find(ExitBlock);
1497 if (It != SunkCopies.end())
1498 New = It->second;
1499 else
1500 New = SunkCopies[ExitBlock] = cloneInstructionInExitBlock(
1501 *I, *ExitBlock, *TPN, LI, SafetyInfo, MSSAU);
1502 return New;
1503}
1504
1505static bool canSplitPredecessors(PHINode *PN, LoopSafetyInfo *SafetyInfo) {
1506 BasicBlock *BB = PN->getParent();
1507 if (!BB->canSplitPredecessors())
1508 return false;
1509 // It's not impossible to split EHPad blocks, but if BlockColors already exist
1510 // it require updating BlockColors for all offspring blocks accordingly. By
1511 // skipping such corner case, we can make updating BlockColors after splitting
1512 // predecessor fairly simple.
1513 if (!SafetyInfo->getBlockColors().empty() && BB->getFirstNonPHI()->isEHPad())
1514 return false;
1515 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
1516 BasicBlock *BBPred = *PI;
1517 if (isa<IndirectBrInst>(BBPred->getTerminator()) ||
1518 isa<CallBrInst>(BBPred->getTerminator()))
1519 return false;
1520 }
1521 return true;
1522}
1523
1524static void splitPredecessorsOfLoopExit(PHINode *PN, DominatorTree *DT,
1525 LoopInfo *LI, const Loop *CurLoop,
1526 LoopSafetyInfo *SafetyInfo,
1527 MemorySSAUpdater *MSSAU) {
1528#ifndef NDEBUG
1529 SmallVector<BasicBlock *, 32> ExitBlocks;
1530 CurLoop->getUniqueExitBlocks(ExitBlocks);
1531 SmallPtrSet<BasicBlock *, 32> ExitBlockSet(ExitBlocks.begin(),
1532 ExitBlocks.end());
1533#endif
1534 BasicBlock *ExitBB = PN->getParent();
1535 assert(ExitBlockSet.count(ExitBB) && "Expect the PHI is in an exit block.")((ExitBlockSet.count(ExitBB) && "Expect the PHI is in an exit block."
) ? static_cast<void> (0) : __assert_fail ("ExitBlockSet.count(ExitBB) && \"Expect the PHI is in an exit block.\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1535, __PRETTY_FUNCTION__))
;
1536
1537 // Split predecessors of the loop exit to make instructions in the loop are
1538 // exposed to exit blocks through trivially replaceable PHIs while keeping the
1539 // loop in the canonical form where each predecessor of each exit block should
1540 // be contained within the loop. For example, this will convert the loop below
1541 // from
1542 //
1543 // LB1:
1544 // %v1 =
1545 // br %LE, %LB2
1546 // LB2:
1547 // %v2 =
1548 // br %LE, %LB1
1549 // LE:
1550 // %p = phi [%v1, %LB1], [%v2, %LB2] <-- non-trivially replaceable
1551 //
1552 // to
1553 //
1554 // LB1:
1555 // %v1 =
1556 // br %LE.split, %LB2
1557 // LB2:
1558 // %v2 =
1559 // br %LE.split2, %LB1
1560 // LE.split:
1561 // %p1 = phi [%v1, %LB1] <-- trivially replaceable
1562 // br %LE
1563 // LE.split2:
1564 // %p2 = phi [%v2, %LB2] <-- trivially replaceable
1565 // br %LE
1566 // LE:
1567 // %p = phi [%p1, %LE.split], [%p2, %LE.split2]
1568 //
1569 const auto &BlockColors = SafetyInfo->getBlockColors();
1570 SmallSetVector<BasicBlock *, 8> PredBBs(pred_begin(ExitBB), pred_end(ExitBB));
1571 while (!PredBBs.empty()) {
1572 BasicBlock *PredBB = *PredBBs.begin();
1573 assert(CurLoop->contains(PredBB) &&((CurLoop->contains(PredBB) && "Expect all predecessors are in the loop"
) ? static_cast<void> (0) : __assert_fail ("CurLoop->contains(PredBB) && \"Expect all predecessors are in the loop\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1574, __PRETTY_FUNCTION__))
1574 "Expect all predecessors are in the loop")((CurLoop->contains(PredBB) && "Expect all predecessors are in the loop"
) ? static_cast<void> (0) : __assert_fail ("CurLoop->contains(PredBB) && \"Expect all predecessors are in the loop\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1574, __PRETTY_FUNCTION__))
;
1575 if (PN->getBasicBlockIndex(PredBB) >= 0) {
1576 BasicBlock *NewPred = SplitBlockPredecessors(
1577 ExitBB, PredBB, ".split.loop.exit", DT, LI, MSSAU, true);
1578 // Since we do not allow splitting EH-block with BlockColors in
1579 // canSplitPredecessors(), we can simply assign predecessor's color to
1580 // the new block.
1581 if (!BlockColors.empty())
1582 // Grab a reference to the ColorVector to be inserted before getting the
1583 // reference to the vector we are copying because inserting the new
1584 // element in BlockColors might cause the map to be reallocated.
1585 SafetyInfo->copyColors(NewPred, PredBB);
1586 }
1587 PredBBs.remove(PredBB);
1588 }
1589}
1590
1591/// When an instruction is found to only be used outside of the loop, this
1592/// function moves it to the exit blocks and patches up SSA form as needed.
1593/// This method is guaranteed to remove the original instruction from its
1594/// position, and may either delete it or move it to outside of the loop.
1595///
1596static bool sink(Instruction &I, LoopInfo *LI, DominatorTree *DT,
1597 BlockFrequencyInfo *BFI, const Loop *CurLoop,
1598 ICFLoopSafetyInfo *SafetyInfo, MemorySSAUpdater *MSSAU,
1599 OptimizationRemarkEmitter *ORE) {
1600 LLVM_DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "LICM sinking instruction: " <<
I << "\n"; } } while (false)
;
1601 ORE->emit([&]() {
1602 return OptimizationRemark(DEBUG_TYPE"licm", "InstSunk", &I)
1603 << "sinking " << ore::NV("Inst", &I);
1604 });
1605 bool Changed = false;
1606 if (isa<LoadInst>(I))
1607 ++NumMovedLoads;
1608 else if (isa<CallInst>(I))
1609 ++NumMovedCalls;
1610 ++NumSunk;
1611
1612 // Iterate over users to be ready for actual sinking. Replace users via
1613 // unreachable blocks with undef and make all user PHIs trivially replaceable.
1614 SmallPtrSet<Instruction *, 8> VisitedUsers;
1615 for (Value::user_iterator UI = I.user_begin(), UE = I.user_end(); UI != UE;) {
1616 auto *User = cast<Instruction>(*UI);
1617 Use &U = UI.getUse();
1618 ++UI;
1619
1620 if (VisitedUsers.count(User) || CurLoop->contains(User))
1621 continue;
1622
1623 if (!DT->isReachableFromEntry(User->getParent())) {
1624 U = UndefValue::get(I.getType());
1625 Changed = true;
1626 continue;
1627 }
1628
1629 // The user must be a PHI node.
1630 PHINode *PN = cast<PHINode>(User);
1631
1632 // Surprisingly, instructions can be used outside of loops without any
1633 // exits. This can only happen in PHI nodes if the incoming block is
1634 // unreachable.
1635 BasicBlock *BB = PN->getIncomingBlock(U);
1636 if (!DT->isReachableFromEntry(BB)) {
1637 U = UndefValue::get(I.getType());
1638 Changed = true;
1639 continue;
1640 }
1641
1642 VisitedUsers.insert(PN);
1643 if (isTriviallyReplaceablePHI(*PN, I))
1644 continue;
1645
1646 if (!canSplitPredecessors(PN, SafetyInfo))
1647 return Changed;
1648
1649 // Split predecessors of the PHI so that we can make users trivially
1650 // replaceable.
1651 splitPredecessorsOfLoopExit(PN, DT, LI, CurLoop, SafetyInfo, MSSAU);
1652
1653 // Should rebuild the iterators, as they may be invalidated by
1654 // splitPredecessorsOfLoopExit().
1655 UI = I.user_begin();
1656 UE = I.user_end();
1657 }
1658
1659 if (VisitedUsers.empty())
1660 return Changed;
1661
1662#ifndef NDEBUG
1663 SmallVector<BasicBlock *, 32> ExitBlocks;
1664 CurLoop->getUniqueExitBlocks(ExitBlocks);
1665 SmallPtrSet<BasicBlock *, 32> ExitBlockSet(ExitBlocks.begin(),
1666 ExitBlocks.end());
1667#endif
1668
1669 // Clones of this instruction. Don't create more than one per exit block!
1670 SmallDenseMap<BasicBlock *, Instruction *, 32> SunkCopies;
1671
1672 // If this instruction is only used outside of the loop, then all users are
1673 // PHI nodes in exit blocks due to LCSSA form. Just RAUW them with clones of
1674 // the instruction.
1675 // First check if I is worth sinking for all uses. Sink only when it is worth
1676 // across all uses.
1677 SmallSetVector<User*, 8> Users(I.user_begin(), I.user_end());
1678 SmallVector<PHINode *, 8> ExitPNs;
1679 for (auto *UI : Users) {
1680 auto *User = cast<Instruction>(UI);
1681
1682 if (CurLoop->contains(User))
1683 continue;
1684
1685 PHINode *PN = cast<PHINode>(User);
1686 assert(ExitBlockSet.count(PN->getParent()) &&((ExitBlockSet.count(PN->getParent()) && "The LCSSA PHI is not in an exit block!"
) ? static_cast<void> (0) : __assert_fail ("ExitBlockSet.count(PN->getParent()) && \"The LCSSA PHI is not in an exit block!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1687, __PRETTY_FUNCTION__))
1687 "The LCSSA PHI is not in an exit block!")((ExitBlockSet.count(PN->getParent()) && "The LCSSA PHI is not in an exit block!"
) ? static_cast<void> (0) : __assert_fail ("ExitBlockSet.count(PN->getParent()) && \"The LCSSA PHI is not in an exit block!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1687, __PRETTY_FUNCTION__))
;
1688 if (!worthSinkOrHoistInst(I, PN->getParent(), ORE, BFI)) {
1689 return Changed;
1690 }
1691
1692 ExitPNs.push_back(PN);
1693 }
1694
1695 for (auto *PN : ExitPNs) {
1696
1697 // The PHI must be trivially replaceable.
1698 Instruction *New = sinkThroughTriviallyReplaceablePHI(
1699 PN, &I, LI, SunkCopies, SafetyInfo, CurLoop, MSSAU);
1700 PN->replaceAllUsesWith(New);
1701 eraseInstruction(*PN, *SafetyInfo, nullptr, nullptr);
1702 Changed = true;
1703 }
1704 return Changed;
1705}
1706
1707/// When an instruction is found to only use loop invariant operands that
1708/// is safe to hoist, this instruction is called to do the dirty work.
1709///
1710static void hoist(Instruction &I, const DominatorTree *DT, const Loop *CurLoop,
1711 BasicBlock *Dest, ICFLoopSafetyInfo *SafetyInfo,
1712 MemorySSAUpdater *MSSAU, ScalarEvolution *SE,
1713 OptimizationRemarkEmitter *ORE) {
1714 LLVM_DEBUG(dbgs() << "LICM hoisting to " << Dest->getName() << ": " << Ido { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "LICM hoisting to " << Dest
->getName() << ": " << I << "\n"; } } while
(false)
1715 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "LICM hoisting to " << Dest
->getName() << ": " << I << "\n"; } } while
(false)
;
1716 ORE->emit([&]() {
1717 return OptimizationRemark(DEBUG_TYPE"licm", "Hoisted", &I) << "hoisting "
1718 << ore::NV("Inst", &I);
1719 });
1720
1721 // Metadata can be dependent on conditions we are hoisting above.
1722 // Conservatively strip all metadata on the instruction unless we were
1723 // guaranteed to execute I if we entered the loop, in which case the metadata
1724 // is valid in the loop preheader.
1725 if (I.hasMetadataOtherThanDebugLoc() &&
1726 // The check on hasMetadataOtherThanDebugLoc is to prevent us from burning
1727 // time in isGuaranteedToExecute if we don't actually have anything to
1728 // drop. It is a compile time optimization, not required for correctness.
1729 !SafetyInfo->isGuaranteedToExecute(I, DT, CurLoop))
1730 I.dropUnknownNonDebugMetadata();
1731
1732 if (isa<PHINode>(I))
1733 // Move the new node to the end of the phi list in the destination block.
1734 moveInstructionBefore(I, *Dest->getFirstNonPHI(), *SafetyInfo, MSSAU, SE);
1735 else
1736 // Move the new node to the destination block, before its terminator.
1737 moveInstructionBefore(I, *Dest->getTerminator(), *SafetyInfo, MSSAU, SE);
1738
1739 I.updateLocationAfterHoist();
1740
1741 if (isa<LoadInst>(I))
1742 ++NumMovedLoads;
1743 else if (isa<CallInst>(I))
1744 ++NumMovedCalls;
1745 ++NumHoisted;
1746}
1747
1748/// Only sink or hoist an instruction if it is not a trapping instruction,
1749/// or if the instruction is known not to trap when moved to the preheader.
1750/// or if it is a trapping instruction and is guaranteed to execute.
1751static bool isSafeToExecuteUnconditionally(Instruction &Inst,
1752 const DominatorTree *DT,
1753 const Loop *CurLoop,
1754 const LoopSafetyInfo *SafetyInfo,
1755 OptimizationRemarkEmitter *ORE,
1756 const Instruction *CtxI) {
1757 if (isSafeToSpeculativelyExecute(&Inst, CtxI, DT))
1758 return true;
1759
1760 bool GuaranteedToExecute =
1761 SafetyInfo->isGuaranteedToExecute(Inst, DT, CurLoop);
1762
1763 if (!GuaranteedToExecute) {
1764 auto *LI = dyn_cast<LoadInst>(&Inst);
1765 if (LI && CurLoop->isLoopInvariant(LI->getPointerOperand()))
1766 ORE->emit([&]() {
1767 return OptimizationRemarkMissed(
1768 DEBUG_TYPE"licm", "LoadWithLoopInvariantAddressCondExecuted", LI)
1769 << "failed to hoist load with loop-invariant address "
1770 "because load is conditionally executed";
1771 });
1772 }
1773
1774 return GuaranteedToExecute;
1775}
1776
1777namespace {
1778class LoopPromoter : public LoadAndStorePromoter {
1779 Value *SomePtr; // Designated pointer to store to.
1780 const SmallSetVector<Value *, 8> &PointerMustAliases;
1781 SmallVectorImpl<BasicBlock *> &LoopExitBlocks;
1782 SmallVectorImpl<Instruction *> &LoopInsertPts;
1783 SmallVectorImpl<MemoryAccess *> &MSSAInsertPts;
1784 PredIteratorCache &PredCache;
1785 AliasSetTracker *AST;
1786 MemorySSAUpdater *MSSAU;
1787 LoopInfo &LI;
1788 DebugLoc DL;
1789 int Alignment;
1790 bool UnorderedAtomic;
1791 AAMDNodes AATags;
1792 ICFLoopSafetyInfo &SafetyInfo;
1793
1794 Value *maybeInsertLCSSAPHI(Value *V, BasicBlock *BB) const {
1795 if (Instruction *I = dyn_cast<Instruction>(V))
1796 if (Loop *L = LI.getLoopFor(I->getParent()))
1797 if (!L->contains(BB)) {
1798 // We need to create an LCSSA PHI node for the incoming value and
1799 // store that.
1800 PHINode *PN = PHINode::Create(I->getType(), PredCache.size(BB),
1801 I->getName() + ".lcssa", &BB->front());
1802 for (BasicBlock *Pred : PredCache.get(BB))
1803 PN->addIncoming(I, Pred);
1804 return PN;
1805 }
1806 return V;
1807 }
1808
1809public:
1810 LoopPromoter(Value *SP, ArrayRef<const Instruction *> Insts, SSAUpdater &S,
1811 const SmallSetVector<Value *, 8> &PMA,
1812 SmallVectorImpl<BasicBlock *> &LEB,
1813 SmallVectorImpl<Instruction *> &LIP,
1814 SmallVectorImpl<MemoryAccess *> &MSSAIP, PredIteratorCache &PIC,
1815 AliasSetTracker *ast, MemorySSAUpdater *MSSAU, LoopInfo &li,
1816 DebugLoc dl, int alignment, bool UnorderedAtomic,
1817 const AAMDNodes &AATags, ICFLoopSafetyInfo &SafetyInfo)
1818 : LoadAndStorePromoter(Insts, S), SomePtr(SP), PointerMustAliases(PMA),
1819 LoopExitBlocks(LEB), LoopInsertPts(LIP), MSSAInsertPts(MSSAIP),
1820 PredCache(PIC), AST(ast), MSSAU(MSSAU), LI(li), DL(std::move(dl)),
1821 Alignment(alignment), UnorderedAtomic(UnorderedAtomic), AATags(AATags),
1822 SafetyInfo(SafetyInfo) {}
1823
1824 bool isInstInList(Instruction *I,
1825 const SmallVectorImpl<Instruction *> &) const override {
1826 Value *Ptr;
1827 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1828 Ptr = LI->getOperand(0);
1829 else
1830 Ptr = cast<StoreInst>(I)->getPointerOperand();
1831 return PointerMustAliases.count(Ptr);
1832 }
1833
1834 void doExtraRewritesBeforeFinalDeletion() override {
1835 // Insert stores after in the loop exit blocks. Each exit block gets a
1836 // store of the live-out values that feed them. Since we've already told
1837 // the SSA updater about the defs in the loop and the preheader
1838 // definition, it is all set and we can start using it.
1839 for (unsigned i = 0, e = LoopExitBlocks.size(); i != e; ++i) {
1840 BasicBlock *ExitBlock = LoopExitBlocks[i];
1841 Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock);
1842 LiveInValue = maybeInsertLCSSAPHI(LiveInValue, ExitBlock);
1843 Value *Ptr = maybeInsertLCSSAPHI(SomePtr, ExitBlock);
1844 Instruction *InsertPos = LoopInsertPts[i];
1845 StoreInst *NewSI = new StoreInst(LiveInValue, Ptr, InsertPos);
1846 if (UnorderedAtomic)
1847 NewSI->setOrdering(AtomicOrdering::Unordered);
1848 NewSI->setAlignment(Align(Alignment));
1849 NewSI->setDebugLoc(DL);
1850 if (AATags)
1851 NewSI->setAAMetadata(AATags);
1852
1853 if (MSSAU) {
1854 MemoryAccess *MSSAInsertPoint = MSSAInsertPts[i];
1855 MemoryAccess *NewMemAcc;
1856 if (!MSSAInsertPoint) {
1857 NewMemAcc = MSSAU->createMemoryAccessInBB(
1858 NewSI, nullptr, NewSI->getParent(), MemorySSA::Beginning);
1859 } else {
1860 NewMemAcc =
1861 MSSAU->createMemoryAccessAfter(NewSI, nullptr, MSSAInsertPoint);
1862 }
1863 MSSAInsertPts[i] = NewMemAcc;
1864 MSSAU->insertDef(cast<MemoryDef>(NewMemAcc), true);
1865 // FIXME: true for safety, false may still be correct.
1866 }
1867 }
1868 }
1869
1870 void replaceLoadWithValue(LoadInst *LI, Value *V) const override {
1871 // Update alias analysis.
1872 if (AST)
1873 AST->copyValue(LI, V);
1874 }
1875 void instructionDeleted(Instruction *I) const override {
1876 SafetyInfo.removeInstruction(I);
1877 if (AST)
1878 AST->deleteValue(I);
1879 if (MSSAU)
1880 MSSAU->removeMemoryAccess(I);
1881 }
1882};
1883
1884
1885/// Return true iff we can prove that a caller of this function can not inspect
1886/// the contents of the provided object in a well defined program.
1887bool isKnownNonEscaping(Value *Object, const TargetLibraryInfo *TLI) {
1888 if (isa<AllocaInst>(Object))
1889 // Since the alloca goes out of scope, we know the caller can't retain a
1890 // reference to it and be well defined. Thus, we don't need to check for
1891 // capture.
1892 return true;
1893
1894 // For all other objects we need to know that the caller can't possibly
1895 // have gotten a reference to the object. There are two components of
1896 // that:
1897 // 1) Object can't be escaped by this function. This is what
1898 // PointerMayBeCaptured checks.
1899 // 2) Object can't have been captured at definition site. For this, we
1900 // need to know the return value is noalias. At the moment, we use a
1901 // weaker condition and handle only AllocLikeFunctions (which are
1902 // known to be noalias). TODO
1903 return isAllocLikeFn(Object, TLI) &&
1904 !PointerMayBeCaptured(Object, true, true);
1905}
1906
1907} // namespace
1908
1909/// Try to promote memory values to scalars by sinking stores out of the
1910/// loop and moving loads to before the loop. We do this by looping over
1911/// the stores in the loop, looking for stores to Must pointers which are
1912/// loop invariant.
1913///
1914bool llvm::promoteLoopAccessesToScalars(
1915 const SmallSetVector<Value *, 8> &PointerMustAliases,
1916 SmallVectorImpl<BasicBlock *> &ExitBlocks,
1917 SmallVectorImpl<Instruction *> &InsertPts,
1918 SmallVectorImpl<MemoryAccess *> &MSSAInsertPts, PredIteratorCache &PIC,
1919 LoopInfo *LI, DominatorTree *DT, const TargetLibraryInfo *TLI,
1920 Loop *CurLoop, AliasSetTracker *CurAST, MemorySSAUpdater *MSSAU,
1921 ICFLoopSafetyInfo *SafetyInfo, OptimizationRemarkEmitter *ORE) {
1922 // Verify inputs.
1923 assert(LI != nullptr && DT != nullptr && CurLoop != nullptr &&((LI != nullptr && DT != nullptr && CurLoop !=
nullptr && SafetyInfo != nullptr && "Unexpected Input to promoteLoopAccessesToScalars"
) ? static_cast<void> (0) : __assert_fail ("LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && \"Unexpected Input to promoteLoopAccessesToScalars\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1925, __PRETTY_FUNCTION__))
1924 SafetyInfo != nullptr &&((LI != nullptr && DT != nullptr && CurLoop !=
nullptr && SafetyInfo != nullptr && "Unexpected Input to promoteLoopAccessesToScalars"
) ? static_cast<void> (0) : __assert_fail ("LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && \"Unexpected Input to promoteLoopAccessesToScalars\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1925, __PRETTY_FUNCTION__))
1925 "Unexpected Input to promoteLoopAccessesToScalars")((LI != nullptr && DT != nullptr && CurLoop !=
nullptr && SafetyInfo != nullptr && "Unexpected Input to promoteLoopAccessesToScalars"
) ? static_cast<void> (0) : __assert_fail ("LI != nullptr && DT != nullptr && CurLoop != nullptr && SafetyInfo != nullptr && \"Unexpected Input to promoteLoopAccessesToScalars\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1925, __PRETTY_FUNCTION__))
;
1926
1927 Value *SomePtr = *PointerMustAliases.begin();
1928 BasicBlock *Preheader = CurLoop->getLoopPreheader();
1929
1930 // It is not safe to promote a load/store from the loop if the load/store is
1931 // conditional. For example, turning:
1932 //
1933 // for () { if (c) *P += 1; }
1934 //
1935 // into:
1936 //
1937 // tmp = *P; for () { if (c) tmp +=1; } *P = tmp;
1938 //
1939 // is not safe, because *P may only be valid to access if 'c' is true.
1940 //
1941 // The safety property divides into two parts:
1942 // p1) The memory may not be dereferenceable on entry to the loop. In this
1943 // case, we can't insert the required load in the preheader.
1944 // p2) The memory model does not allow us to insert a store along any dynamic
1945 // path which did not originally have one.
1946 //
1947 // If at least one store is guaranteed to execute, both properties are
1948 // satisfied, and promotion is legal.
1949 //
1950 // This, however, is not a necessary condition. Even if no store/load is
1951 // guaranteed to execute, we can still establish these properties.
1952 // We can establish (p1) by proving that hoisting the load into the preheader
1953 // is safe (i.e. proving dereferenceability on all paths through the loop). We
1954 // can use any access within the alias set to prove dereferenceability,
1955 // since they're all must alias.
1956 //
1957 // There are two ways establish (p2):
1958 // a) Prove the location is thread-local. In this case the memory model
1959 // requirement does not apply, and stores are safe to insert.
1960 // b) Prove a store dominates every exit block. In this case, if an exit
1961 // blocks is reached, the original dynamic path would have taken us through
1962 // the store, so inserting a store into the exit block is safe. Note that this
1963 // is different from the store being guaranteed to execute. For instance,
1964 // if an exception is thrown on the first iteration of the loop, the original
1965 // store is never executed, but the exit blocks are not executed either.
1966
1967 bool DereferenceableInPH = false;
1968 bool SafeToInsertStore = false;
1969
1970 SmallVector<Instruction *, 64> LoopUses;
1971
1972 // We start with an alignment of one and try to find instructions that allow
1973 // us to prove better alignment.
1974 Align Alignment;
1975 // Keep track of which types of access we see
1976 bool SawUnorderedAtomic = false;
1977 bool SawNotAtomic = false;
1978 AAMDNodes AATags;
1979
1980 const DataLayout &MDL = Preheader->getModule()->getDataLayout();
1981
1982 bool IsKnownThreadLocalObject = false;
1983 if (SafetyInfo->anyBlockMayThrow()) {
1984 // If a loop can throw, we have to insert a store along each unwind edge.
1985 // That said, we can't actually make the unwind edge explicit. Therefore,
1986 // we have to prove that the store is dead along the unwind edge. We do
1987 // this by proving that the caller can't have a reference to the object
1988 // after return and thus can't possibly load from the object.
1989 Value *Object = getUnderlyingObject(SomePtr);
1990 if (!isKnownNonEscaping(Object, TLI))
1991 return false;
1992 // Subtlety: Alloca's aren't visible to callers, but *are* potentially
1993 // visible to other threads if captured and used during their lifetimes.
1994 IsKnownThreadLocalObject = !isa<AllocaInst>(Object);
1995 }
1996
1997 // Check that all of the pointers in the alias set have the same type. We
1998 // cannot (yet) promote a memory location that is loaded and stored in
1999 // different sizes. While we are at it, collect alignment and AA info.
2000 for (Value *ASIV : PointerMustAliases) {
2001 // Check that all of the pointers in the alias set have the same type. We
2002 // cannot (yet) promote a memory location that is loaded and stored in
2003 // different sizes.
2004 if (SomePtr->getType() != ASIV->getType())
2005 return false;
2006
2007 for (User *U : ASIV->users()) {
2008 // Ignore instructions that are outside the loop.
2009 Instruction *UI = dyn_cast<Instruction>(U);
2010 if (!UI || !CurLoop->contains(UI))
2011 continue;
2012
2013 // If there is an non-load/store instruction in the loop, we can't promote
2014 // it.
2015 if (LoadInst *Load = dyn_cast<LoadInst>(UI)) {
2016 if (!Load->isUnordered())
2017 return false;
2018
2019 SawUnorderedAtomic |= Load->isAtomic();
2020 SawNotAtomic |= !Load->isAtomic();
2021
2022 Align InstAlignment = Load->getAlign();
2023
2024 // Note that proving a load safe to speculate requires proving
2025 // sufficient alignment at the target location. Proving it guaranteed
2026 // to execute does as well. Thus we can increase our guaranteed
2027 // alignment as well.
2028 if (!DereferenceableInPH || (InstAlignment > Alignment))
2029 if (isSafeToExecuteUnconditionally(*Load, DT, CurLoop, SafetyInfo,
2030 ORE, Preheader->getTerminator())) {
2031 DereferenceableInPH = true;
2032 Alignment = std::max(Alignment, InstAlignment);
2033 }
2034 } else if (const StoreInst *Store = dyn_cast<StoreInst>(UI)) {
2035 // Stores *of* the pointer are not interesting, only stores *to* the
2036 // pointer.
2037 if (UI->getOperand(1) != ASIV)
2038 continue;
2039 if (!Store->isUnordered())
2040 return false;
2041
2042 SawUnorderedAtomic |= Store->isAtomic();
2043 SawNotAtomic |= !Store->isAtomic();
2044
2045 // If the store is guaranteed to execute, both properties are satisfied.
2046 // We may want to check if a store is guaranteed to execute even if we
2047 // already know that promotion is safe, since it may have higher
2048 // alignment than any other guaranteed stores, in which case we can
2049 // raise the alignment on the promoted store.
2050 Align InstAlignment = Store->getAlign();
2051
2052 if (!DereferenceableInPH || !SafeToInsertStore ||
2053 (InstAlignment > Alignment)) {
2054 if (SafetyInfo->isGuaranteedToExecute(*UI, DT, CurLoop)) {
2055 DereferenceableInPH = true;
2056 SafeToInsertStore = true;
2057 Alignment = std::max(Alignment, InstAlignment);
2058 }
2059 }
2060
2061 // If a store dominates all exit blocks, it is safe to sink.
2062 // As explained above, if an exit block was executed, a dominating
2063 // store must have been executed at least once, so we are not
2064 // introducing stores on paths that did not have them.
2065 // Note that this only looks at explicit exit blocks. If we ever
2066 // start sinking stores into unwind edges (see above), this will break.
2067 if (!SafeToInsertStore)
2068 SafeToInsertStore = llvm::all_of(ExitBlocks, [&](BasicBlock *Exit) {
2069 return DT->dominates(Store->getParent(), Exit);
2070 });
2071
2072 // If the store is not guaranteed to execute, we may still get
2073 // deref info through it.
2074 if (!DereferenceableInPH) {
2075 DereferenceableInPH = isDereferenceableAndAlignedPointer(
2076 Store->getPointerOperand(), Store->getValueOperand()->getType(),
2077 Store->getAlign(), MDL, Preheader->getTerminator(), DT);
2078 }
2079 } else
2080 return false; // Not a load or store.
2081
2082 // Merge the AA tags.
2083 if (LoopUses.empty()) {
2084 // On the first load/store, just take its AA tags.
2085 UI->getAAMetadata(AATags);
2086 } else if (AATags) {
2087 UI->getAAMetadata(AATags, /* Merge = */ true);
2088 }
2089
2090 LoopUses.push_back(UI);
2091 }
2092 }
2093
2094 // If we found both an unordered atomic instruction and a non-atomic memory
2095 // access, bail. We can't blindly promote non-atomic to atomic since we
2096 // might not be able to lower the result. We can't downgrade since that
2097 // would violate memory model. Also, align 0 is an error for atomics.
2098 if (SawUnorderedAtomic && SawNotAtomic)
2099 return false;
2100
2101 // If we're inserting an atomic load in the preheader, we must be able to
2102 // lower it. We're only guaranteed to be able to lower naturally aligned
2103 // atomics.
2104 auto *SomePtrElemType = SomePtr->getType()->getPointerElementType();
2105 if (SawUnorderedAtomic &&
2106 Alignment < MDL.getTypeStoreSize(SomePtrElemType))
2107 return false;
2108
2109 // If we couldn't prove we can hoist the load, bail.
2110 if (!DereferenceableInPH)
2111 return false;
2112
2113 // We know we can hoist the load, but don't have a guaranteed store.
2114 // Check whether the location is thread-local. If it is, then we can insert
2115 // stores along paths which originally didn't have them without violating the
2116 // memory model.
2117 if (!SafeToInsertStore) {
2118 if (IsKnownThreadLocalObject)
2119 SafeToInsertStore = true;
2120 else {
2121 Value *Object = getUnderlyingObject(SomePtr);
2122 SafeToInsertStore =
2123 (isAllocLikeFn(Object, TLI) || isa<AllocaInst>(Object)) &&
2124 !PointerMayBeCaptured(Object, true, true);
2125 }
2126 }
2127
2128 // If we've still failed to prove we can sink the store, give up.
2129 if (!SafeToInsertStore)
2130 return false;
2131
2132 // Otherwise, this is safe to promote, lets do it!
2133 LLVM_DEBUG(dbgs() << "LICM: Promoting value stored to in loop: " << *SomePtrdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "LICM: Promoting value stored to in loop: "
<< *SomePtr << '\n'; } } while (false)
2134 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "LICM: Promoting value stored to in loop: "
<< *SomePtr << '\n'; } } while (false)
;
2135 ORE->emit([&]() {
2136 return OptimizationRemark(DEBUG_TYPE"licm", "PromoteLoopAccessesToScalar",
2137 LoopUses[0])
2138 << "Moving accesses to memory location out of the loop";
2139 });
2140 ++NumPromoted;
2141
2142 // Look at all the loop uses, and try to merge their locations.
2143 std::vector<const DILocation *> LoopUsesLocs;
2144 for (auto U : LoopUses)
2145 LoopUsesLocs.push_back(U->getDebugLoc().get());
2146 auto DL = DebugLoc(DILocation::getMergedLocations(LoopUsesLocs));
2147
2148 // We use the SSAUpdater interface to insert phi nodes as required.
2149 SmallVector<PHINode *, 16> NewPHIs;
2150 SSAUpdater SSA(&NewPHIs);
2151 LoopPromoter Promoter(SomePtr, LoopUses, SSA, PointerMustAliases, ExitBlocks,
2152 InsertPts, MSSAInsertPts, PIC, CurAST, MSSAU, *LI, DL,
2153 Alignment.value(), SawUnorderedAtomic, AATags,
2154 *SafetyInfo);
2155
2156 // Set up the preheader to have a definition of the value. It is the live-out
2157 // value from the preheader that uses in the loop will use.
2158 LoadInst *PreheaderLoad = new LoadInst(
2159 SomePtr->getType()->getPointerElementType(), SomePtr,
2160 SomePtr->getName() + ".promoted", Preheader->getTerminator());
2161 if (SawUnorderedAtomic)
2162 PreheaderLoad->setOrdering(AtomicOrdering::Unordered);
2163 PreheaderLoad->setAlignment(Alignment);
2164 PreheaderLoad->setDebugLoc(DebugLoc());
2165 if (AATags)
2166 PreheaderLoad->setAAMetadata(AATags);
2167 SSA.AddAvailableValue(Preheader, PreheaderLoad);
2168
2169 if (MSSAU) {
2170 MemoryAccess *PreheaderLoadMemoryAccess = MSSAU->createMemoryAccessInBB(
2171 PreheaderLoad, nullptr, PreheaderLoad->getParent(), MemorySSA::End);
2172 MemoryUse *NewMemUse = cast<MemoryUse>(PreheaderLoadMemoryAccess);
2173 MSSAU->insertUse(NewMemUse, /*RenameUses=*/true);
2174 }
2175
2176 if (MSSAU && VerifyMemorySSA)
2177 MSSAU->getMemorySSA()->verifyMemorySSA();
2178 // Rewrite all the loads in the loop and remember all the definitions from
2179 // stores in the loop.
2180 Promoter.run(LoopUses);
2181
2182 if (MSSAU && VerifyMemorySSA)
2183 MSSAU->getMemorySSA()->verifyMemorySSA();
2184 // If the SSAUpdater didn't use the load in the preheader, just zap it now.
2185 if (PreheaderLoad->use_empty())
2186 eraseInstruction(*PreheaderLoad, *SafetyInfo, CurAST, MSSAU);
2187
2188 return true;
2189}
2190
2191/// Returns an owning pointer to an alias set which incorporates aliasing info
2192/// from L and all subloops of L.
2193std::unique_ptr<AliasSetTracker>
2194LoopInvariantCodeMotion::collectAliasInfoForLoop(Loop *L, LoopInfo *LI,
2195 AAResults *AA) {
2196 auto CurAST = std::make_unique<AliasSetTracker>(*AA);
2197
2198 // Add everything from all the sub loops.
2199 for (Loop *InnerL : L->getSubLoops())
2200 for (BasicBlock *BB : InnerL->blocks())
2201 CurAST->add(*BB);
2202
2203 // And merge in this loop (without anything from inner loops).
2204 for (BasicBlock *BB : L->blocks())
2205 if (LI->getLoopFor(BB) == L)
2206 CurAST->add(*BB);
2207
2208 return CurAST;
2209}
2210
2211std::unique_ptr<AliasSetTracker>
2212LoopInvariantCodeMotion::collectAliasInfoForLoopWithMSSA(
2213 Loop *L, AAResults *AA, MemorySSAUpdater *MSSAU) {
2214 auto *MSSA = MSSAU->getMemorySSA();
2215 auto CurAST = std::make_unique<AliasSetTracker>(*AA, MSSA, L);
2216 CurAST->addAllInstructionsInLoopUsingMSSA();
2217 return CurAST;
2218}
2219
2220static bool pointerInvalidatedByLoop(MemoryLocation MemLoc,
2221 AliasSetTracker *CurAST, Loop *CurLoop,
2222 AAResults *AA) {
2223 // First check to see if any of the basic blocks in CurLoop invalidate *V.
2224 bool isInvalidatedAccordingToAST = CurAST->getAliasSetFor(MemLoc).isMod();
2225
2226 if (!isInvalidatedAccordingToAST || !LICMN2Theshold)
2227 return isInvalidatedAccordingToAST;
2228
2229 // Check with a diagnostic analysis if we can refine the information above.
2230 // This is to identify the limitations of using the AST.
2231 // The alias set mechanism used by LICM has a major weakness in that it
2232 // combines all things which may alias into a single set *before* asking
2233 // modref questions. As a result, a single readonly call within a loop will
2234 // collapse all loads and stores into a single alias set and report
2235 // invalidation if the loop contains any store. For example, readonly calls
2236 // with deopt states have this form and create a general alias set with all
2237 // loads and stores. In order to get any LICM in loops containing possible
2238 // deopt states we need a more precise invalidation of checking the mod ref
2239 // info of each instruction within the loop and LI. This has a complexity of
2240 // O(N^2), so currently, it is used only as a diagnostic tool since the
2241 // default value of LICMN2Threshold is zero.
2242
2243 // Don't look at nested loops.
2244 if (CurLoop->begin() != CurLoop->end())
2245 return true;
2246
2247 int N = 0;
2248 for (BasicBlock *BB : CurLoop->getBlocks())
2249 for (Instruction &I : *BB) {
2250 if (N >= LICMN2Theshold) {
2251 LLVM_DEBUG(dbgs() << "Alasing N2 threshold exhausted for "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "Alasing N2 threshold exhausted for "
<< *(MemLoc.Ptr) << "\n"; } } while (false)
2252 << *(MemLoc.Ptr) << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "Alasing N2 threshold exhausted for "
<< *(MemLoc.Ptr) << "\n"; } } while (false)
;
2253 return true;
2254 }
2255 N++;
2256 auto Res = AA->getModRefInfo(&I, MemLoc);
2257 if (isModSet(Res)) {
2258 LLVM_DEBUG(dbgs() << "Aliasing failed on " << I << " for "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "Aliasing failed on " << I <<
" for " << *(MemLoc.Ptr) << "\n"; } } while (false
)
2259 << *(MemLoc.Ptr) << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "Aliasing failed on " << I <<
" for " << *(MemLoc.Ptr) << "\n"; } } while (false
)
;
2260 return true;
2261 }
2262 }
2263 LLVM_DEBUG(dbgs() << "Aliasing okay for " << *(MemLoc.Ptr) << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "Aliasing okay for " << *(MemLoc
.Ptr) << "\n"; } } while (false)
;
2264 return false;
2265}
2266
2267static bool pointerInvalidatedByLoopWithMSSA(MemorySSA *MSSA, MemoryUse *MU,
2268 Loop *CurLoop,
2269 SinkAndHoistLICMFlags &Flags) {
2270 // For hoisting, use the walker to determine safety
2271 if (!Flags.IsSink) {
2272 MemoryAccess *Source;
2273 // See declaration of SetLicmMssaOptCap for usage details.
2274 if (Flags.LicmMssaOptCounter >= Flags.LicmMssaOptCap)
2275 Source = MU->getDefiningAccess();
2276 else {
2277 Source = MSSA->getSkipSelfWalker()->getClobberingMemoryAccess(MU);
2278 Flags.LicmMssaOptCounter++;
2279 }
2280 return !MSSA->isLiveOnEntryDef(Source) &&
2281 CurLoop->contains(Source->getBlock());
2282 }
2283
2284 // For sinking, we'd need to check all Defs below this use. The getClobbering
2285 // call will look on the backedge of the loop, but will check aliasing with
2286 // the instructions on the previous iteration.
2287 // For example:
2288 // for (i ... )
2289 // load a[i] ( Use (LoE)
2290 // store a[i] ( 1 = Def (2), with 2 = Phi for the loop.
2291 // i++;
2292 // The load sees no clobbering inside the loop, as the backedge alias check
2293 // does phi translation, and will check aliasing against store a[i-1].
2294 // However sinking the load outside the loop, below the store is incorrect.
2295
2296 // For now, only sink if there are no Defs in the loop, and the existing ones
2297 // precede the use and are in the same block.
2298 // FIXME: Increase precision: Safe to sink if Use post dominates the Def;
2299 // needs PostDominatorTreeAnalysis.
2300 // FIXME: More precise: no Defs that alias this Use.
2301 if (Flags.NoOfMemAccTooLarge)
2302 return true;
2303 for (auto *BB : CurLoop->getBlocks())
2304 if (auto *Accesses = MSSA->getBlockDefs(BB))
2305 for (const auto &MA : *Accesses)
2306 if (const auto *MD = dyn_cast<MemoryDef>(&MA))
2307 if (MU->getBlock() != MD->getBlock() ||
2308 !MSSA->locallyDominates(MD, MU))
2309 return true;
2310 return false;
2311}
2312
2313/// Little predicate that returns true if the specified basic block is in
2314/// a subloop of the current one, not the current one itself.
2315///
2316static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI) {
2317 assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop")((CurLoop->contains(BB) && "Only valid if BB is IN the loop"
) ? static_cast<void> (0) : __assert_fail ("CurLoop->contains(BB) && \"Only valid if BB is IN the loop\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/lib/Transforms/Scalar/LICM.cpp"
, 2317, __PRETTY_FUNCTION__))
;
2318 return LI->getLoopFor(BB) != CurLoop;
2319}

/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h

1//===- llvm/Instructions.h - Instruction subclass definitions ---*- C++ -*-===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file exposes the class definitions of all of the subclasses of the
10// Instruction class. This is meant to be an easy way to get access to all
11// instruction subclasses.
12//
13//===----------------------------------------------------------------------===//
14
15#ifndef LLVM_IR_INSTRUCTIONS_H
16#define LLVM_IR_INSTRUCTIONS_H
17
18#include "llvm/ADT/ArrayRef.h"
19#include "llvm/ADT/Bitfields.h"
20#include "llvm/ADT/None.h"
21#include "llvm/ADT/STLExtras.h"
22#include "llvm/ADT/SmallVector.h"
23#include "llvm/ADT/StringRef.h"
24#include "llvm/ADT/Twine.h"
25#include "llvm/ADT/iterator.h"
26#include "llvm/ADT/iterator_range.h"
27#include "llvm/IR/Attributes.h"
28#include "llvm/IR/BasicBlock.h"
29#include "llvm/IR/CallingConv.h"
30#include "llvm/IR/CFG.h"
31#include "llvm/IR/Constant.h"
32#include "llvm/IR/DerivedTypes.h"
33#include "llvm/IR/Function.h"
34#include "llvm/IR/InstrTypes.h"
35#include "llvm/IR/Instruction.h"
36#include "llvm/IR/OperandTraits.h"
37#include "llvm/IR/Type.h"
38#include "llvm/IR/Use.h"
39#include "llvm/IR/User.h"
40#include "llvm/IR/Value.h"
41#include "llvm/Support/AtomicOrdering.h"
42#include "llvm/Support/Casting.h"
43#include "llvm/Support/ErrorHandling.h"
44#include <cassert>
45#include <cstddef>
46#include <cstdint>
47#include <iterator>
48
49namespace llvm {
50
51class APInt;
52class ConstantInt;
53class DataLayout;
54class LLVMContext;
55
56//===----------------------------------------------------------------------===//
57// AllocaInst Class
58//===----------------------------------------------------------------------===//
59
60/// an instruction to allocate memory on the stack
61class AllocaInst : public UnaryInstruction {
62 Type *AllocatedType;
63
64 using AlignmentField = AlignmentBitfieldElementT<0>;
65 using UsedWithInAllocaField = BoolBitfieldElementT<AlignmentField::NextBit>;
66 using SwiftErrorField = BoolBitfieldElementT<UsedWithInAllocaField::NextBit>;
67 static_assert(Bitfield::areContiguous<AlignmentField, UsedWithInAllocaField,
68 SwiftErrorField>(),
69 "Bitfields must be contiguous");
70
71protected:
72 // Note: Instruction needs to be a friend here to call cloneImpl.
73 friend class Instruction;
74
75 AllocaInst *cloneImpl() const;
76
77public:
78 explicit AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize,
79 const Twine &Name, Instruction *InsertBefore);
80 AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize,
81 const Twine &Name, BasicBlock *InsertAtEnd);
82
83 AllocaInst(Type *Ty, unsigned AddrSpace, const Twine &Name,
84 Instruction *InsertBefore);
85 AllocaInst(Type *Ty, unsigned AddrSpace,
86 const Twine &Name, BasicBlock *InsertAtEnd);
87
88 AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize, Align Align,
89 const Twine &Name = "", Instruction *InsertBefore = nullptr);
90 AllocaInst(Type *Ty, unsigned AddrSpace, Value *ArraySize, Align Align,
91 const Twine &Name, BasicBlock *InsertAtEnd);
92
93 /// Return true if there is an allocation size parameter to the allocation
94 /// instruction that is not 1.
95 bool isArrayAllocation() const;
96
97 /// Get the number of elements allocated. For a simple allocation of a single
98 /// element, this will return a constant 1 value.
99 const Value *getArraySize() const { return getOperand(0); }
100 Value *getArraySize() { return getOperand(0); }
101
102 /// Overload to return most specific pointer type.
103 PointerType *getType() const {
104 return cast<PointerType>(Instruction::getType());
105 }
106
107 /// Get allocation size in bits. Returns None if size can't be determined,
108 /// e.g. in case of a VLA.
109 Optional<uint64_t> getAllocationSizeInBits(const DataLayout &DL) const;
110
111 /// Return the type that is being allocated by the instruction.
112 Type *getAllocatedType() const { return AllocatedType; }
113 /// for use only in special circumstances that need to generically
114 /// transform a whole instruction (eg: IR linking and vectorization).
115 void setAllocatedType(Type *Ty) { AllocatedType = Ty; }
116
117 /// Return the alignment of the memory that is being allocated by the
118 /// instruction.
119 Align getAlign() const {
120 return Align(1ULL << getSubclassData<AlignmentField>());
121 }
122
123 void setAlignment(Align Align) {
124 setSubclassData<AlignmentField>(Log2(Align));
125 }
126
127 // FIXME: Remove this one transition to Align is over.
128 unsigned getAlignment() const { return getAlign().value(); }
129
130 /// Return true if this alloca is in the entry block of the function and is a
131 /// constant size. If so, the code generator will fold it into the
132 /// prolog/epilog code, so it is basically free.
133 bool isStaticAlloca() const;
134
135 /// Return true if this alloca is used as an inalloca argument to a call. Such
136 /// allocas are never considered static even if they are in the entry block.
137 bool isUsedWithInAlloca() const {
138 return getSubclassData<UsedWithInAllocaField>();
139 }
140
141 /// Specify whether this alloca is used to represent the arguments to a call.
142 void setUsedWithInAlloca(bool V) {
143 setSubclassData<UsedWithInAllocaField>(V);
144 }
145
146 /// Return true if this alloca is used as a swifterror argument to a call.
147 bool isSwiftError() const { return getSubclassData<SwiftErrorField>(); }
148 /// Specify whether this alloca is used to represent a swifterror.
149 void setSwiftError(bool V) { setSubclassData<SwiftErrorField>(V); }
150
151 // Methods for support type inquiry through isa, cast, and dyn_cast:
152 static bool classof(const Instruction *I) {
153 return (I->getOpcode() == Instruction::Alloca);
154 }
155 static bool classof(const Value *V) {
156 return isa<Instruction>(V) && classof(cast<Instruction>(V));
157 }
158
159private:
160 // Shadow Instruction::setInstructionSubclassData with a private forwarding
161 // method so that subclasses cannot accidentally use it.
162 template <typename Bitfield>
163 void setSubclassData(typename Bitfield::Type Value) {
164 Instruction::setSubclassData<Bitfield>(Value);
165 }
166};
167
168//===----------------------------------------------------------------------===//
169// LoadInst Class
170//===----------------------------------------------------------------------===//
171
172/// An instruction for reading from memory. This uses the SubclassData field in
173/// Value to store whether or not the load is volatile.
174class LoadInst : public UnaryInstruction {
175 using VolatileField = BoolBitfieldElementT<0>;
176 using AlignmentField = AlignmentBitfieldElementT<VolatileField::NextBit>;
177 using OrderingField = AtomicOrderingBitfieldElementT<AlignmentField::NextBit>;
178 static_assert(
179 Bitfield::areContiguous<VolatileField, AlignmentField, OrderingField>(),
180 "Bitfields must be contiguous");
181
182 void AssertOK();
183
184protected:
185 // Note: Instruction needs to be a friend here to call cloneImpl.
186 friend class Instruction;
187
188 LoadInst *cloneImpl() const;
189
190public:
191 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr,
192 Instruction *InsertBefore);
193 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, BasicBlock *InsertAtEnd);
194 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile,
195 Instruction *InsertBefore);
196 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile,
197 BasicBlock *InsertAtEnd);
198 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile,
199 Align Align, Instruction *InsertBefore = nullptr);
200 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile,
201 Align Align, BasicBlock *InsertAtEnd);
202 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile,
203 Align Align, AtomicOrdering Order,
204 SyncScope::ID SSID = SyncScope::System,
205 Instruction *InsertBefore = nullptr);
206 LoadInst(Type *Ty, Value *Ptr, const Twine &NameStr, bool isVolatile,
207 Align Align, AtomicOrdering Order, SyncScope::ID SSID,
208 BasicBlock *InsertAtEnd);
209
210 /// Return true if this is a load from a volatile memory location.
211 bool isVolatile() const { return getSubclassData<VolatileField>(); }
212
213 /// Specify whether this is a volatile load or not.
214 void setVolatile(bool V) { setSubclassData<VolatileField>(V); }
215
216 /// Return the alignment of the access that is being performed.
217 /// FIXME: Remove this function once transition to Align is over.
218 /// Use getAlign() instead.
219 unsigned getAlignment() const { return getAlign().value(); }
220
221 /// Return the alignment of the access that is being performed.
222 Align getAlign() const {
223 return Align(1ULL << (getSubclassData<AlignmentField>()));
224 }
225
226 void setAlignment(Align Align) {
227 setSubclassData<AlignmentField>(Log2(Align));
228 }
229
230 /// Returns the ordering constraint of this load instruction.
231 AtomicOrdering getOrdering() const {
232 return getSubclassData<OrderingField>();
233 }
234 /// Sets the ordering constraint of this load instruction. May not be Release
235 /// or AcquireRelease.
236 void setOrdering(AtomicOrdering Ordering) {
237 setSubclassData<OrderingField>(Ordering);
238 }
239
240 /// Returns the synchronization scope ID of this load instruction.
241 SyncScope::ID getSyncScopeID() const {
242 return SSID;
243 }
244
245 /// Sets the synchronization scope ID of this load instruction.
246 void setSyncScopeID(SyncScope::ID SSID) {
247 this->SSID = SSID;
248 }
249
250 /// Sets the ordering constraint and the synchronization scope ID of this load
251 /// instruction.
252 void setAtomic(AtomicOrdering Ordering,
253 SyncScope::ID SSID = SyncScope::System) {
254 setOrdering(Ordering);
255 setSyncScopeID(SSID);
256 }
257
258 bool isSimple() const { return !isAtomic() && !isVolatile(); }
259
260 bool isUnordered() const {
261 return (getOrdering() == AtomicOrdering::NotAtomic ||
13
Assuming the condition is false
16
Returning the value 1, which participates in a condition later
262 getOrdering() == AtomicOrdering::Unordered) &&
14
Assuming the condition is true
263 !isVolatile();
15
Assuming the condition is true
264 }
265
266 Value *getPointerOperand() { return getOperand(0); }
267 const Value *getPointerOperand() const { return getOperand(0); }
268 static unsigned getPointerOperandIndex() { return 0U; }
269 Type *getPointerOperandType() const { return getPointerOperand()->getType(); }
270
271 /// Returns the address space of the pointer operand.
272 unsigned getPointerAddressSpace() const {
273 return getPointerOperandType()->getPointerAddressSpace();
274 }
275
276 // Methods for support type inquiry through isa, cast, and dyn_cast:
277 static bool classof(const Instruction *I) {
278 return I->getOpcode() == Instruction::Load;
279 }
280 static bool classof(const Value *V) {
281 return isa<Instruction>(V) && classof(cast<Instruction>(V));
282 }
283
284private:
285 // Shadow Instruction::setInstructionSubclassData with a private forwarding
286 // method so that subclasses cannot accidentally use it.
287 template <typename Bitfield>
288 void setSubclassData(typename Bitfield::Type Value) {
289 Instruction::setSubclassData<Bitfield>(Value);
290 }
291
292 /// The synchronization scope ID of this load instruction. Not quite enough
293 /// room in SubClassData for everything, so synchronization scope ID gets its
294 /// own field.
295 SyncScope::ID SSID;
296};
297
298//===----------------------------------------------------------------------===//
299// StoreInst Class
300//===----------------------------------------------------------------------===//
301
302/// An instruction for storing to memory.
303class StoreInst : public Instruction {
304 using VolatileField = BoolBitfieldElementT<0>;
305 using AlignmentField = AlignmentBitfieldElementT<VolatileField::NextBit>;
306 using OrderingField = AtomicOrderingBitfieldElementT<AlignmentField::NextBit>;
307 static_assert(
308 Bitfield::areContiguous<VolatileField, AlignmentField, OrderingField>(),
309 "Bitfields must be contiguous");
310
311 void AssertOK();
312
313protected:
314 // Note: Instruction needs to be a friend here to call cloneImpl.
315 friend class Instruction;
316
317 StoreInst *cloneImpl() const;
318
319public:
320 StoreInst(Value *Val, Value *Ptr, Instruction *InsertBefore);
321 StoreInst(Value *Val, Value *Ptr, BasicBlock *InsertAtEnd);
322 StoreInst(Value *Val, Value *Ptr, bool isVolatile, Instruction *InsertBefore);
323 StoreInst(Value *Val, Value *Ptr, bool isVolatile, BasicBlock *InsertAtEnd);
324 StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align,
325 Instruction *InsertBefore = nullptr);
326 StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align,
327 BasicBlock *InsertAtEnd);
328 StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align,
329 AtomicOrdering Order, SyncScope::ID SSID = SyncScope::System,
330 Instruction *InsertBefore = nullptr);
331 StoreInst(Value *Val, Value *Ptr, bool isVolatile, Align Align,
332 AtomicOrdering Order, SyncScope::ID SSID, BasicBlock *InsertAtEnd);
333
334 // allocate space for exactly two operands
335 void *operator new(size_t s) {
336 return User::operator new(s, 2);
337 }
338
339 /// Return true if this is a store to a volatile memory location.
340 bool isVolatile() const { return getSubclassData<VolatileField>(); }
341
342 /// Specify whether this is a volatile store or not.
343 void setVolatile(bool V) { setSubclassData<VolatileField>(V); }
344
345 /// Transparently provide more efficient getOperand methods.
346 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
347
348 /// Return the alignment of the access that is being performed
349 /// FIXME: Remove this function once transition to Align is over.
350 /// Use getAlign() instead.
351 unsigned getAlignment() const { return getAlign().value(); }
352
353 Align getAlign() const {
354 return Align(1ULL << (getSubclassData<AlignmentField>()));
355 }
356
357 void setAlignment(Align Align) {
358 setSubclassData<AlignmentField>(Log2(Align));
359 }
360
361 /// Returns the ordering constraint of this store instruction.
362 AtomicOrdering getOrdering() const {
363 return getSubclassData<OrderingField>();
364 }
365
366 /// Sets the ordering constraint of this store instruction. May not be
367 /// Acquire or AcquireRelease.
368 void setOrdering(AtomicOrdering Ordering) {
369 setSubclassData<OrderingField>(Ordering);
370 }
371
372 /// Returns the synchronization scope ID of this store instruction.
373 SyncScope::ID getSyncScopeID() const {
374 return SSID;
375 }
376
377 /// Sets the synchronization scope ID of this store instruction.
378 void setSyncScopeID(SyncScope::ID SSID) {
379 this->SSID = SSID;
380 }
381
382 /// Sets the ordering constraint and the synchronization scope ID of this
383 /// store instruction.
384 void setAtomic(AtomicOrdering Ordering,
385 SyncScope::ID SSID = SyncScope::System) {
386 setOrdering(Ordering);
387 setSyncScopeID(SSID);
388 }
389
390 bool isSimple() const { return !isAtomic() && !isVolatile(); }
391
392 bool isUnordered() const {
393 return (getOrdering() == AtomicOrdering::NotAtomic ||
394 getOrdering() == AtomicOrdering::Unordered) &&
395 !isVolatile();
396 }
397
398 Value *getValueOperand() { return getOperand(0); }
399 const Value *getValueOperand() const { return getOperand(0); }
400
401 Value *getPointerOperand() { return getOperand(1); }
402 const Value *getPointerOperand() const { return getOperand(1); }
403 static unsigned getPointerOperandIndex() { return 1U; }
404 Type *getPointerOperandType() const { return getPointerOperand()->getType(); }
405
406 /// Returns the address space of the pointer operand.
407 unsigned getPointerAddressSpace() const {
408 return getPointerOperandType()->getPointerAddressSpace();
409 }
410
411 // Methods for support type inquiry through isa, cast, and dyn_cast:
412 static bool classof(const Instruction *I) {
413 return I->getOpcode() == Instruction::Store;
414 }
415 static bool classof(const Value *V) {
416 return isa<Instruction>(V) && classof(cast<Instruction>(V));
417 }
418
419private:
420 // Shadow Instruction::setInstructionSubclassData with a private forwarding
421 // method so that subclasses cannot accidentally use it.
422 template <typename Bitfield>
423 void setSubclassData(typename Bitfield::Type Value) {
424 Instruction::setSubclassData<Bitfield>(Value);
425 }
426
427 /// The synchronization scope ID of this store instruction. Not quite enough
428 /// room in SubClassData for everything, so synchronization scope ID gets its
429 /// own field.
430 SyncScope::ID SSID;
431};
432
433template <>
434struct OperandTraits<StoreInst> : public FixedNumOperandTraits<StoreInst, 2> {
435};
436
437DEFINE_TRANSPARENT_OPERAND_ACCESSORS(StoreInst, Value)StoreInst::op_iterator StoreInst::op_begin() { return OperandTraits
<StoreInst>::op_begin(this); } StoreInst::const_op_iterator
StoreInst::op_begin() const { return OperandTraits<StoreInst
>::op_begin(const_cast<StoreInst*>(this)); } StoreInst
::op_iterator StoreInst::op_end() { return OperandTraits<StoreInst
>::op_end(this); } StoreInst::const_op_iterator StoreInst::
op_end() const { return OperandTraits<StoreInst>::op_end
(const_cast<StoreInst*>(this)); } Value *StoreInst::getOperand
(unsigned i_nocapture) const { ((i_nocapture < OperandTraits
<StoreInst>::operands(this) && "getOperand() out of range!"
) ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<StoreInst>::operands(this) && \"getOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 437, __PRETTY_FUNCTION__)); return cast_or_null<Value>
( OperandTraits<StoreInst>::op_begin(const_cast<StoreInst
*>(this))[i_nocapture].get()); } void StoreInst::setOperand
(unsigned i_nocapture, Value *Val_nocapture) { ((i_nocapture <
OperandTraits<StoreInst>::operands(this) && "setOperand() out of range!"
) ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<StoreInst>::operands(this) && \"setOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 437, __PRETTY_FUNCTION__)); OperandTraits<StoreInst>::
op_begin(this)[i_nocapture] = Val_nocapture; } unsigned StoreInst
::getNumOperands() const { return OperandTraits<StoreInst>
::operands(this); } template <int Idx_nocapture> Use &
StoreInst::Op() { return this->OpFrom<Idx_nocapture>
(this); } template <int Idx_nocapture> const Use &StoreInst
::Op() const { return this->OpFrom<Idx_nocapture>(this
); }
438
439//===----------------------------------------------------------------------===//
440// FenceInst Class
441//===----------------------------------------------------------------------===//
442
443/// An instruction for ordering other memory operations.
444class FenceInst : public Instruction {
445 using OrderingField = AtomicOrderingBitfieldElementT<0>;
446
447 void Init(AtomicOrdering Ordering, SyncScope::ID SSID);
448
449protected:
450 // Note: Instruction needs to be a friend here to call cloneImpl.
451 friend class Instruction;
452
453 FenceInst *cloneImpl() const;
454
455public:
456 // Ordering may only be Acquire, Release, AcquireRelease, or
457 // SequentiallyConsistent.
458 FenceInst(LLVMContext &C, AtomicOrdering Ordering,
459 SyncScope::ID SSID = SyncScope::System,
460 Instruction *InsertBefore = nullptr);
461 FenceInst(LLVMContext &C, AtomicOrdering Ordering, SyncScope::ID SSID,
462 BasicBlock *InsertAtEnd);
463
464 // allocate space for exactly zero operands
465 void *operator new(size_t s) {
466 return User::operator new(s, 0);
467 }
468
469 /// Returns the ordering constraint of this fence instruction.
470 AtomicOrdering getOrdering() const {
471 return getSubclassData<OrderingField>();
472 }
473
474 /// Sets the ordering constraint of this fence instruction. May only be
475 /// Acquire, Release, AcquireRelease, or SequentiallyConsistent.
476 void setOrdering(AtomicOrdering Ordering) {
477 setSubclassData<OrderingField>(Ordering);
478 }
479
480 /// Returns the synchronization scope ID of this fence instruction.
481 SyncScope::ID getSyncScopeID() const {
482 return SSID;
483 }
484
485 /// Sets the synchronization scope ID of this fence instruction.
486 void setSyncScopeID(SyncScope::ID SSID) {
487 this->SSID = SSID;
488 }
489
490 // Methods for support type inquiry through isa, cast, and dyn_cast:
491 static bool classof(const Instruction *I) {
492 return I->getOpcode() == Instruction::Fence;
493 }
494 static bool classof(const Value *V) {
495 return isa<Instruction>(V) && classof(cast<Instruction>(V));
496 }
497
498private:
499 // Shadow Instruction::setInstructionSubclassData with a private forwarding
500 // method so that subclasses cannot accidentally use it.
501 template <typename Bitfield>
502 void setSubclassData(typename Bitfield::Type Value) {
503 Instruction::setSubclassData<Bitfield>(Value);
504 }
505
506 /// The synchronization scope ID of this fence instruction. Not quite enough
507 /// room in SubClassData for everything, so synchronization scope ID gets its
508 /// own field.
509 SyncScope::ID SSID;
510};
511
512//===----------------------------------------------------------------------===//
513// AtomicCmpXchgInst Class
514//===----------------------------------------------------------------------===//
515
516/// An instruction that atomically checks whether a
517/// specified value is in a memory location, and, if it is, stores a new value
518/// there. The value returned by this instruction is a pair containing the
519/// original value as first element, and an i1 indicating success (true) or
520/// failure (false) as second element.
521///
522class AtomicCmpXchgInst : public Instruction {
523 void Init(Value *Ptr, Value *Cmp, Value *NewVal, Align Align,
524 AtomicOrdering SuccessOrdering, AtomicOrdering FailureOrdering,
525 SyncScope::ID SSID);
526
527 template <unsigned Offset>
528 using AtomicOrderingBitfieldElement =
529 typename Bitfield::Element<AtomicOrdering, Offset, 3,
530 AtomicOrdering::LAST>;
531
532protected:
533 // Note: Instruction needs to be a friend here to call cloneImpl.
534 friend class Instruction;
535
536 AtomicCmpXchgInst *cloneImpl() const;
537
538public:
539 AtomicCmpXchgInst(Value *Ptr, Value *Cmp, Value *NewVal, Align Alignment,
540 AtomicOrdering SuccessOrdering,
541 AtomicOrdering FailureOrdering, SyncScope::ID SSID,
542 Instruction *InsertBefore = nullptr);
543 AtomicCmpXchgInst(Value *Ptr, Value *Cmp, Value *NewVal, Align Alignment,
544 AtomicOrdering SuccessOrdering,
545 AtomicOrdering FailureOrdering, SyncScope::ID SSID,
546 BasicBlock *InsertAtEnd);
547
548 // allocate space for exactly three operands
549 void *operator new(size_t s) {
550 return User::operator new(s, 3);
551 }
552
553 using VolatileField = BoolBitfieldElementT<0>;
554 using WeakField = BoolBitfieldElementT<VolatileField::NextBit>;
555 using SuccessOrderingField =
556 AtomicOrderingBitfieldElementT<WeakField::NextBit>;
557 using FailureOrderingField =
558 AtomicOrderingBitfieldElementT<SuccessOrderingField::NextBit>;
559 using AlignmentField =
560 AlignmentBitfieldElementT<FailureOrderingField::NextBit>;
561 static_assert(
562 Bitfield::areContiguous<VolatileField, WeakField, SuccessOrderingField,
563 FailureOrderingField, AlignmentField>(),
564 "Bitfields must be contiguous");
565
566 /// Return the alignment of the memory that is being allocated by the
567 /// instruction.
568 Align getAlign() const {
569 return Align(1ULL << getSubclassData<AlignmentField>());
570 }
571
572 void setAlignment(Align Align) {
573 setSubclassData<AlignmentField>(Log2(Align));
574 }
575
576 /// Return true if this is a cmpxchg from a volatile memory
577 /// location.
578 ///
579 bool isVolatile() const { return getSubclassData<VolatileField>(); }
580
581 /// Specify whether this is a volatile cmpxchg.
582 ///
583 void setVolatile(bool V) { setSubclassData<VolatileField>(V); }
584
585 /// Return true if this cmpxchg may spuriously fail.
586 bool isWeak() const { return getSubclassData<WeakField>(); }
587
588 void setWeak(bool IsWeak) { setSubclassData<WeakField>(IsWeak); }
589
590 /// Transparently provide more efficient getOperand methods.
591 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
592
593 /// Returns the success ordering constraint of this cmpxchg instruction.
594 AtomicOrdering getSuccessOrdering() const {
595 return getSubclassData<SuccessOrderingField>();
596 }
597
598 /// Sets the success ordering constraint of this cmpxchg instruction.
599 void setSuccessOrdering(AtomicOrdering Ordering) {
600 assert(Ordering != AtomicOrdering::NotAtomic &&((Ordering != AtomicOrdering::NotAtomic && "CmpXchg instructions can only be atomic."
) ? static_cast<void> (0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"CmpXchg instructions can only be atomic.\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 601, __PRETTY_FUNCTION__))
601 "CmpXchg instructions can only be atomic.")((Ordering != AtomicOrdering::NotAtomic && "CmpXchg instructions can only be atomic."
) ? static_cast<void> (0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"CmpXchg instructions can only be atomic.\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 601, __PRETTY_FUNCTION__))
;
602 setSubclassData<SuccessOrderingField>(Ordering);
603 }
604
605 /// Returns the failure ordering constraint of this cmpxchg instruction.
606 AtomicOrdering getFailureOrdering() const {
607 return getSubclassData<FailureOrderingField>();
608 }
609
610 /// Sets the failure ordering constraint of this cmpxchg instruction.
611 void setFailureOrdering(AtomicOrdering Ordering) {
612 assert(Ordering != AtomicOrdering::NotAtomic &&((Ordering != AtomicOrdering::NotAtomic && "CmpXchg instructions can only be atomic."
) ? static_cast<void> (0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"CmpXchg instructions can only be atomic.\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 613, __PRETTY_FUNCTION__))
613 "CmpXchg instructions can only be atomic.")((Ordering != AtomicOrdering::NotAtomic && "CmpXchg instructions can only be atomic."
) ? static_cast<void> (0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"CmpXchg instructions can only be atomic.\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 613, __PRETTY_FUNCTION__))
;
614 setSubclassData<FailureOrderingField>(Ordering);
615 }
616
617 /// Returns the synchronization scope ID of this cmpxchg instruction.
618 SyncScope::ID getSyncScopeID() const {
619 return SSID;
620 }
621
622 /// Sets the synchronization scope ID of this cmpxchg instruction.
623 void setSyncScopeID(SyncScope::ID SSID) {
624 this->SSID = SSID;
625 }
626
627 Value *getPointerOperand() { return getOperand(0); }
628 const Value *getPointerOperand() const { return getOperand(0); }
629 static unsigned getPointerOperandIndex() { return 0U; }
630
631 Value *getCompareOperand() { return getOperand(1); }
632 const Value *getCompareOperand() const { return getOperand(1); }
633
634 Value *getNewValOperand() { return getOperand(2); }
635 const Value *getNewValOperand() const { return getOperand(2); }
636
637 /// Returns the address space of the pointer operand.
638 unsigned getPointerAddressSpace() const {
639 return getPointerOperand()->getType()->getPointerAddressSpace();
640 }
641
642 /// Returns the strongest permitted ordering on failure, given the
643 /// desired ordering on success.
644 ///
645 /// If the comparison in a cmpxchg operation fails, there is no atomic store
646 /// so release semantics cannot be provided. So this function drops explicit
647 /// Release requests from the AtomicOrdering. A SequentiallyConsistent
648 /// operation would remain SequentiallyConsistent.
649 static AtomicOrdering
650 getStrongestFailureOrdering(AtomicOrdering SuccessOrdering) {
651 switch (SuccessOrdering) {
652 default:
653 llvm_unreachable("invalid cmpxchg success ordering")::llvm::llvm_unreachable_internal("invalid cmpxchg success ordering"
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 653)
;
654 case AtomicOrdering::Release:
655 case AtomicOrdering::Monotonic:
656 return AtomicOrdering::Monotonic;
657 case AtomicOrdering::AcquireRelease:
658 case AtomicOrdering::Acquire:
659 return AtomicOrdering::Acquire;
660 case AtomicOrdering::SequentiallyConsistent:
661 return AtomicOrdering::SequentiallyConsistent;
662 }
663 }
664
665 // Methods for support type inquiry through isa, cast, and dyn_cast:
666 static bool classof(const Instruction *I) {
667 return I->getOpcode() == Instruction::AtomicCmpXchg;
668 }
669 static bool classof(const Value *V) {
670 return isa<Instruction>(V) && classof(cast<Instruction>(V));
671 }
672
673private:
674 // Shadow Instruction::setInstructionSubclassData with a private forwarding
675 // method so that subclasses cannot accidentally use it.
676 template <typename Bitfield>
677 void setSubclassData(typename Bitfield::Type Value) {
678 Instruction::setSubclassData<Bitfield>(Value);
679 }
680
681 /// The synchronization scope ID of this cmpxchg instruction. Not quite
682 /// enough room in SubClassData for everything, so synchronization scope ID
683 /// gets its own field.
684 SyncScope::ID SSID;
685};
686
687template <>
688struct OperandTraits<AtomicCmpXchgInst> :
689 public FixedNumOperandTraits<AtomicCmpXchgInst, 3> {
690};
691
692DEFINE_TRANSPARENT_OPERAND_ACCESSORS(AtomicCmpXchgInst, Value)AtomicCmpXchgInst::op_iterator AtomicCmpXchgInst::op_begin() {
return OperandTraits<AtomicCmpXchgInst>::op_begin(this
); } AtomicCmpXchgInst::const_op_iterator AtomicCmpXchgInst::
op_begin() const { return OperandTraits<AtomicCmpXchgInst>
::op_begin(const_cast<AtomicCmpXchgInst*>(this)); } AtomicCmpXchgInst
::op_iterator AtomicCmpXchgInst::op_end() { return OperandTraits
<AtomicCmpXchgInst>::op_end(this); } AtomicCmpXchgInst::
const_op_iterator AtomicCmpXchgInst::op_end() const { return OperandTraits
<AtomicCmpXchgInst>::op_end(const_cast<AtomicCmpXchgInst
*>(this)); } Value *AtomicCmpXchgInst::getOperand(unsigned
i_nocapture) const { ((i_nocapture < OperandTraits<AtomicCmpXchgInst
>::operands(this) && "getOperand() out of range!")
? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<AtomicCmpXchgInst>::operands(this) && \"getOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 692, __PRETTY_FUNCTION__)); return cast_or_null<Value>
( OperandTraits<AtomicCmpXchgInst>::op_begin(const_cast
<AtomicCmpXchgInst*>(this))[i_nocapture].get()); } void
AtomicCmpXchgInst::setOperand(unsigned i_nocapture, Value *Val_nocapture
) { ((i_nocapture < OperandTraits<AtomicCmpXchgInst>
::operands(this) && "setOperand() out of range!") ? static_cast
<void> (0) : __assert_fail ("i_nocapture < OperandTraits<AtomicCmpXchgInst>::operands(this) && \"setOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 692, __PRETTY_FUNCTION__)); OperandTraits<AtomicCmpXchgInst
>::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned
AtomicCmpXchgInst::getNumOperands() const { return OperandTraits
<AtomicCmpXchgInst>::operands(this); } template <int
Idx_nocapture> Use &AtomicCmpXchgInst::Op() { return this
->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture
> const Use &AtomicCmpXchgInst::Op() const { return this
->OpFrom<Idx_nocapture>(this); }
693
694//===----------------------------------------------------------------------===//
695// AtomicRMWInst Class
696//===----------------------------------------------------------------------===//
697
698/// an instruction that atomically reads a memory location,
699/// combines it with another value, and then stores the result back. Returns
700/// the old value.
701///
702class AtomicRMWInst : public Instruction {
703protected:
704 // Note: Instruction needs to be a friend here to call cloneImpl.
705 friend class Instruction;
706
707 AtomicRMWInst *cloneImpl() const;
708
709public:
710 /// This enumeration lists the possible modifications atomicrmw can make. In
711 /// the descriptions, 'p' is the pointer to the instruction's memory location,
712 /// 'old' is the initial value of *p, and 'v' is the other value passed to the
713 /// instruction. These instructions always return 'old'.
714 enum BinOp : unsigned {
715 /// *p = v
716 Xchg,
717 /// *p = old + v
718 Add,
719 /// *p = old - v
720 Sub,
721 /// *p = old & v
722 And,
723 /// *p = ~(old & v)
724 Nand,
725 /// *p = old | v
726 Or,
727 /// *p = old ^ v
728 Xor,
729 /// *p = old >signed v ? old : v
730 Max,
731 /// *p = old <signed v ? old : v
732 Min,
733 /// *p = old >unsigned v ? old : v
734 UMax,
735 /// *p = old <unsigned v ? old : v
736 UMin,
737
738 /// *p = old + v
739 FAdd,
740
741 /// *p = old - v
742 FSub,
743
744 FIRST_BINOP = Xchg,
745 LAST_BINOP = FSub,
746 BAD_BINOP
747 };
748
749private:
750 template <unsigned Offset>
751 using AtomicOrderingBitfieldElement =
752 typename Bitfield::Element<AtomicOrdering, Offset, 3,
753 AtomicOrdering::LAST>;
754
755 template <unsigned Offset>
756 using BinOpBitfieldElement =
757 typename Bitfield::Element<BinOp, Offset, 4, BinOp::LAST_BINOP>;
758
759public:
760 AtomicRMWInst(BinOp Operation, Value *Ptr, Value *Val, Align Alignment,
761 AtomicOrdering Ordering, SyncScope::ID SSID,
762 Instruction *InsertBefore = nullptr);
763 AtomicRMWInst(BinOp Operation, Value *Ptr, Value *Val, Align Alignment,
764 AtomicOrdering Ordering, SyncScope::ID SSID,
765 BasicBlock *InsertAtEnd);
766
767 // allocate space for exactly two operands
768 void *operator new(size_t s) {
769 return User::operator new(s, 2);
770 }
771
772 using VolatileField = BoolBitfieldElementT<0>;
773 using AtomicOrderingField =
774 AtomicOrderingBitfieldElementT<VolatileField::NextBit>;
775 using OperationField = BinOpBitfieldElement<AtomicOrderingField::NextBit>;
776 using AlignmentField = AlignmentBitfieldElementT<OperationField::NextBit>;
777 static_assert(Bitfield::areContiguous<VolatileField, AtomicOrderingField,
778 OperationField, AlignmentField>(),
779 "Bitfields must be contiguous");
780
781 BinOp getOperation() const { return getSubclassData<OperationField>(); }
782
783 static StringRef getOperationName(BinOp Op);
784
785 static bool isFPOperation(BinOp Op) {
786 switch (Op) {
787 case AtomicRMWInst::FAdd:
788 case AtomicRMWInst::FSub:
789 return true;
790 default:
791 return false;
792 }
793 }
794
795 void setOperation(BinOp Operation) {
796 setSubclassData<OperationField>(Operation);
797 }
798
799 /// Return the alignment of the memory that is being allocated by the
800 /// instruction.
801 Align getAlign() const {
802 return Align(1ULL << getSubclassData<AlignmentField>());
803 }
804
805 void setAlignment(Align Align) {
806 setSubclassData<AlignmentField>(Log2(Align));
807 }
808
809 /// Return true if this is a RMW on a volatile memory location.
810 ///
811 bool isVolatile() const { return getSubclassData<VolatileField>(); }
812
813 /// Specify whether this is a volatile RMW or not.
814 ///
815 void setVolatile(bool V) { setSubclassData<VolatileField>(V); }
816
817 /// Transparently provide more efficient getOperand methods.
818 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
819
820 /// Returns the ordering constraint of this rmw instruction.
821 AtomicOrdering getOrdering() const {
822 return getSubclassData<AtomicOrderingField>();
823 }
824
825 /// Sets the ordering constraint of this rmw instruction.
826 void setOrdering(AtomicOrdering Ordering) {
827 assert(Ordering != AtomicOrdering::NotAtomic &&((Ordering != AtomicOrdering::NotAtomic && "atomicrmw instructions can only be atomic."
) ? static_cast<void> (0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"atomicrmw instructions can only be atomic.\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 828, __PRETTY_FUNCTION__))
828 "atomicrmw instructions can only be atomic.")((Ordering != AtomicOrdering::NotAtomic && "atomicrmw instructions can only be atomic."
) ? static_cast<void> (0) : __assert_fail ("Ordering != AtomicOrdering::NotAtomic && \"atomicrmw instructions can only be atomic.\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 828, __PRETTY_FUNCTION__))
;
829 setSubclassData<AtomicOrderingField>(Ordering);
830 }
831
832 /// Returns the synchronization scope ID of this rmw instruction.
833 SyncScope::ID getSyncScopeID() const {
834 return SSID;
835 }
836
837 /// Sets the synchronization scope ID of this rmw instruction.
838 void setSyncScopeID(SyncScope::ID SSID) {
839 this->SSID = SSID;
840 }
841
842 Value *getPointerOperand() { return getOperand(0); }
843 const Value *getPointerOperand() const { return getOperand(0); }
844 static unsigned getPointerOperandIndex() { return 0U; }
845
846 Value *getValOperand() { return getOperand(1); }
847 const Value *getValOperand() const { return getOperand(1); }
848
849 /// Returns the address space of the pointer operand.
850 unsigned getPointerAddressSpace() const {
851 return getPointerOperand()->getType()->getPointerAddressSpace();
852 }
853
854 bool isFloatingPointOperation() const {
855 return isFPOperation(getOperation());
856 }
857
858 // Methods for support type inquiry through isa, cast, and dyn_cast:
859 static bool classof(const Instruction *I) {
860 return I->getOpcode() == Instruction::AtomicRMW;
861 }
862 static bool classof(const Value *V) {
863 return isa<Instruction>(V) && classof(cast<Instruction>(V));
864 }
865
866private:
867 void Init(BinOp Operation, Value *Ptr, Value *Val, Align Align,
868 AtomicOrdering Ordering, SyncScope::ID SSID);
869
870 // Shadow Instruction::setInstructionSubclassData with a private forwarding
871 // method so that subclasses cannot accidentally use it.
872 template <typename Bitfield>
873 void setSubclassData(typename Bitfield::Type Value) {
874 Instruction::setSubclassData<Bitfield>(Value);
875 }
876
877 /// The synchronization scope ID of this rmw instruction. Not quite enough
878 /// room in SubClassData for everything, so synchronization scope ID gets its
879 /// own field.
880 SyncScope::ID SSID;
881};
882
883template <>
884struct OperandTraits<AtomicRMWInst>
885 : public FixedNumOperandTraits<AtomicRMWInst,2> {
886};
887
888DEFINE_TRANSPARENT_OPERAND_ACCESSORS(AtomicRMWInst, Value)AtomicRMWInst::op_iterator AtomicRMWInst::op_begin() { return
OperandTraits<AtomicRMWInst>::op_begin(this); } AtomicRMWInst
::const_op_iterator AtomicRMWInst::op_begin() const { return OperandTraits
<AtomicRMWInst>::op_begin(const_cast<AtomicRMWInst*>
(this)); } AtomicRMWInst::op_iterator AtomicRMWInst::op_end()
{ return OperandTraits<AtomicRMWInst>::op_end(this); }
AtomicRMWInst::const_op_iterator AtomicRMWInst::op_end() const
{ return OperandTraits<AtomicRMWInst>::op_end(const_cast
<AtomicRMWInst*>(this)); } Value *AtomicRMWInst::getOperand
(unsigned i_nocapture) const { ((i_nocapture < OperandTraits
<AtomicRMWInst>::operands(this) && "getOperand() out of range!"
) ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<AtomicRMWInst>::operands(this) && \"getOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 888, __PRETTY_FUNCTION__)); return cast_or_null<Value>
( OperandTraits<AtomicRMWInst>::op_begin(const_cast<
AtomicRMWInst*>(this))[i_nocapture].get()); } void AtomicRMWInst
::setOperand(unsigned i_nocapture, Value *Val_nocapture) { ((
i_nocapture < OperandTraits<AtomicRMWInst>::operands
(this) && "setOperand() out of range!") ? static_cast
<void> (0) : __assert_fail ("i_nocapture < OperandTraits<AtomicRMWInst>::operands(this) && \"setOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 888, __PRETTY_FUNCTION__)); OperandTraits<AtomicRMWInst>
::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned AtomicRMWInst
::getNumOperands() const { return OperandTraits<AtomicRMWInst
>::operands(this); } template <int Idx_nocapture> Use
&AtomicRMWInst::Op() { return this->OpFrom<Idx_nocapture
>(this); } template <int Idx_nocapture> const Use &
AtomicRMWInst::Op() const { return this->OpFrom<Idx_nocapture
>(this); }
889
890//===----------------------------------------------------------------------===//
891// GetElementPtrInst Class
892//===----------------------------------------------------------------------===//
893
894// checkGEPType - Simple wrapper function to give a better assertion failure
895// message on bad indexes for a gep instruction.
896//
897inline Type *checkGEPType(Type *Ty) {
898 assert(Ty && "Invalid GetElementPtrInst indices for type!")((Ty && "Invalid GetElementPtrInst indices for type!"
) ? static_cast<void> (0) : __assert_fail ("Ty && \"Invalid GetElementPtrInst indices for type!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 898, __PRETTY_FUNCTION__))
;
899 return Ty;
900}
901
902/// an instruction for type-safe pointer arithmetic to
903/// access elements of arrays and structs
904///
905class GetElementPtrInst : public Instruction {
906 Type *SourceElementType;
907 Type *ResultElementType;
908
909 GetElementPtrInst(const GetElementPtrInst &GEPI);
910
911 /// Constructors - Create a getelementptr instruction with a base pointer an
912 /// list of indices. The first ctor can optionally insert before an existing
913 /// instruction, the second appends the new instruction to the specified
914 /// BasicBlock.
915 inline GetElementPtrInst(Type *PointeeType, Value *Ptr,
916 ArrayRef<Value *> IdxList, unsigned Values,
917 const Twine &NameStr, Instruction *InsertBefore);
918 inline GetElementPtrInst(Type *PointeeType, Value *Ptr,
919 ArrayRef<Value *> IdxList, unsigned Values,
920 const Twine &NameStr, BasicBlock *InsertAtEnd);
921
922 void init(Value *Ptr, ArrayRef<Value *> IdxList, const Twine &NameStr);
923
924protected:
925 // Note: Instruction needs to be a friend here to call cloneImpl.
926 friend class Instruction;
927
928 GetElementPtrInst *cloneImpl() const;
929
930public:
931 static GetElementPtrInst *Create(Type *PointeeType, Value *Ptr,
932 ArrayRef<Value *> IdxList,
933 const Twine &NameStr = "",
934 Instruction *InsertBefore = nullptr) {
935 unsigned Values = 1 + unsigned(IdxList.size());
936 if (!PointeeType)
937 PointeeType =
938 cast<PointerType>(Ptr->getType()->getScalarType())->getElementType();
939 else
940 assert(((PointeeType == cast<PointerType>(Ptr->getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 942, __PRETTY_FUNCTION__))
941 PointeeType ==((PointeeType == cast<PointerType>(Ptr->getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 942, __PRETTY_FUNCTION__))
942 cast<PointerType>(Ptr->getType()->getScalarType())->getElementType())((PointeeType == cast<PointerType>(Ptr->getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 942, __PRETTY_FUNCTION__))
;
943 return new (Values) GetElementPtrInst(PointeeType, Ptr, IdxList, Values,
944 NameStr, InsertBefore);
945 }
946
947 static GetElementPtrInst *Create(Type *PointeeType, Value *Ptr,
948 ArrayRef<Value *> IdxList,
949 const Twine &NameStr,
950 BasicBlock *InsertAtEnd) {
951 unsigned Values = 1 + unsigned(IdxList.size());
952 if (!PointeeType)
953 PointeeType =
954 cast<PointerType>(Ptr->getType()->getScalarType())->getElementType();
955 else
956 assert(((PointeeType == cast<PointerType>(Ptr->getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 958, __PRETTY_FUNCTION__))
957 PointeeType ==((PointeeType == cast<PointerType>(Ptr->getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 958, __PRETTY_FUNCTION__))
958 cast<PointerType>(Ptr->getType()->getScalarType())->getElementType())((PointeeType == cast<PointerType>(Ptr->getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("PointeeType == cast<PointerType>(Ptr->getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 958, __PRETTY_FUNCTION__))
;
959 return new (Values) GetElementPtrInst(PointeeType, Ptr, IdxList, Values,
960 NameStr, InsertAtEnd);
961 }
962
963 /// Create an "inbounds" getelementptr. See the documentation for the
964 /// "inbounds" flag in LangRef.html for details.
965 static GetElementPtrInst *CreateInBounds(Value *Ptr,
966 ArrayRef<Value *> IdxList,
967 const Twine &NameStr = "",
968 Instruction *InsertBefore = nullptr){
969 return CreateInBounds(nullptr, Ptr, IdxList, NameStr, InsertBefore);
970 }
971
972 static GetElementPtrInst *
973 CreateInBounds(Type *PointeeType, Value *Ptr, ArrayRef<Value *> IdxList,
974 const Twine &NameStr = "",
975 Instruction *InsertBefore = nullptr) {
976 GetElementPtrInst *GEP =
977 Create(PointeeType, Ptr, IdxList, NameStr, InsertBefore);
978 GEP->setIsInBounds(true);
979 return GEP;
980 }
981
982 static GetElementPtrInst *CreateInBounds(Value *Ptr,
983 ArrayRef<Value *> IdxList,
984 const Twine &NameStr,
985 BasicBlock *InsertAtEnd) {
986 return CreateInBounds(nullptr, Ptr, IdxList, NameStr, InsertAtEnd);
987 }
988
989 static GetElementPtrInst *CreateInBounds(Type *PointeeType, Value *Ptr,
990 ArrayRef<Value *> IdxList,
991 const Twine &NameStr,
992 BasicBlock *InsertAtEnd) {
993 GetElementPtrInst *GEP =
994 Create(PointeeType, Ptr, IdxList, NameStr, InsertAtEnd);
995 GEP->setIsInBounds(true);
996 return GEP;
997 }
998
999 /// Transparently provide more efficient getOperand methods.
1000 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
1001
1002 Type *getSourceElementType() const { return SourceElementType; }
1003
1004 void setSourceElementType(Type *Ty) { SourceElementType = Ty; }
1005 void setResultElementType(Type *Ty) { ResultElementType = Ty; }
1006
1007 Type *getResultElementType() const {
1008 assert(ResultElementType ==((ResultElementType == cast<PointerType>(getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("ResultElementType == cast<PointerType>(getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 1009, __PRETTY_FUNCTION__))
1009 cast<PointerType>(getType()->getScalarType())->getElementType())((ResultElementType == cast<PointerType>(getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("ResultElementType == cast<PointerType>(getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 1009, __PRETTY_FUNCTION__))
;
1010 return ResultElementType;
1011 }
1012
1013 /// Returns the address space of this instruction's pointer type.
1014 unsigned getAddressSpace() const {
1015 // Note that this is always the same as the pointer operand's address space
1016 // and that is cheaper to compute, so cheat here.
1017 return getPointerAddressSpace();
1018 }
1019
1020 /// Returns the result type of a getelementptr with the given source
1021 /// element type and indexes.
1022 ///
1023 /// Null is returned if the indices are invalid for the specified
1024 /// source element type.
1025 static Type *getIndexedType(Type *Ty, ArrayRef<Value *> IdxList);
1026 static Type *getIndexedType(Type *Ty, ArrayRef<Constant *> IdxList);
1027 static Type *getIndexedType(Type *Ty, ArrayRef<uint64_t> IdxList);
1028
1029 /// Return the type of the element at the given index of an indexable
1030 /// type. This is equivalent to "getIndexedType(Agg, {Zero, Idx})".
1031 ///
1032 /// Returns null if the type can't be indexed, or the given index is not
1033 /// legal for the given type.
1034 static Type *getTypeAtIndex(Type *Ty, Value *Idx);
1035 static Type *getTypeAtIndex(Type *Ty, uint64_t Idx);
1036
1037 inline op_iterator idx_begin() { return op_begin()+1; }
1038 inline const_op_iterator idx_begin() const { return op_begin()+1; }
1039 inline op_iterator idx_end() { return op_end(); }
1040 inline const_op_iterator idx_end() const { return op_end(); }
1041
1042 inline iterator_range<op_iterator> indices() {
1043 return make_range(idx_begin(), idx_end());
1044 }
1045
1046 inline iterator_range<const_op_iterator> indices() const {
1047 return make_range(idx_begin(), idx_end());
1048 }
1049
1050 Value *getPointerOperand() {
1051 return getOperand(0);
1052 }
1053 const Value *getPointerOperand() const {
1054 return getOperand(0);
1055 }
1056 static unsigned getPointerOperandIndex() {
1057 return 0U; // get index for modifying correct operand.
1058 }
1059
1060 /// Method to return the pointer operand as a
1061 /// PointerType.
1062 Type *getPointerOperandType() const {
1063 return getPointerOperand()->getType();
1064 }
1065
1066 /// Returns the address space of the pointer operand.
1067 unsigned getPointerAddressSpace() const {
1068 return getPointerOperandType()->getPointerAddressSpace();
1069 }
1070
1071 /// Returns the pointer type returned by the GEP
1072 /// instruction, which may be a vector of pointers.
1073 static Type *getGEPReturnType(Type *ElTy, Value *Ptr,
1074 ArrayRef<Value *> IdxList) {
1075 Type *PtrTy = PointerType::get(checkGEPType(getIndexedType(ElTy, IdxList)),
1076 Ptr->getType()->getPointerAddressSpace());
1077 // Vector GEP
1078 if (auto *PtrVTy = dyn_cast<VectorType>(Ptr->getType())) {
1079 ElementCount EltCount = PtrVTy->getElementCount();
1080 return VectorType::get(PtrTy, EltCount);
1081 }
1082 for (Value *Index : IdxList)
1083 if (auto *IndexVTy = dyn_cast<VectorType>(Index->getType())) {
1084 ElementCount EltCount = IndexVTy->getElementCount();
1085 return VectorType::get(PtrTy, EltCount);
1086 }
1087 // Scalar GEP
1088 return PtrTy;
1089 }
1090
1091 unsigned getNumIndices() const { // Note: always non-negative
1092 return getNumOperands() - 1;
1093 }
1094
1095 bool hasIndices() const {
1096 return getNumOperands() > 1;
1097 }
1098
1099 /// Return true if all of the indices of this GEP are
1100 /// zeros. If so, the result pointer and the first operand have the same
1101 /// value, just potentially different types.
1102 bool hasAllZeroIndices() const;
1103
1104 /// Return true if all of the indices of this GEP are
1105 /// constant integers. If so, the result pointer and the first operand have
1106 /// a constant offset between them.
1107 bool hasAllConstantIndices() const;
1108
1109 /// Set or clear the inbounds flag on this GEP instruction.
1110 /// See LangRef.html for the meaning of inbounds on a getelementptr.
1111 void setIsInBounds(bool b = true);
1112
1113 /// Determine whether the GEP has the inbounds flag.
1114 bool isInBounds() const;
1115
1116 /// Accumulate the constant address offset of this GEP if possible.
1117 ///
1118 /// This routine accepts an APInt into which it will accumulate the constant
1119 /// offset of this GEP if the GEP is in fact constant. If the GEP is not
1120 /// all-constant, it returns false and the value of the offset APInt is
1121 /// undefined (it is *not* preserved!). The APInt passed into this routine
1122 /// must be at least as wide as the IntPtr type for the address space of
1123 /// the base GEP pointer.
1124 bool accumulateConstantOffset(const DataLayout &DL, APInt &Offset) const;
1125
1126 // Methods for support type inquiry through isa, cast, and dyn_cast:
1127 static bool classof(const Instruction *I) {
1128 return (I->getOpcode() == Instruction::GetElementPtr);
1129 }
1130 static bool classof(const Value *V) {
1131 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1132 }
1133};
1134
1135template <>
1136struct OperandTraits<GetElementPtrInst> :
1137 public VariadicOperandTraits<GetElementPtrInst, 1> {
1138};
1139
1140GetElementPtrInst::GetElementPtrInst(Type *PointeeType, Value *Ptr,
1141 ArrayRef<Value *> IdxList, unsigned Values,
1142 const Twine &NameStr,
1143 Instruction *InsertBefore)
1144 : Instruction(getGEPReturnType(PointeeType, Ptr, IdxList), GetElementPtr,
1145 OperandTraits<GetElementPtrInst>::op_end(this) - Values,
1146 Values, InsertBefore),
1147 SourceElementType(PointeeType),
1148 ResultElementType(getIndexedType(PointeeType, IdxList)) {
1149 assert(ResultElementType ==((ResultElementType == cast<PointerType>(getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("ResultElementType == cast<PointerType>(getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 1150, __PRETTY_FUNCTION__))
1150 cast<PointerType>(getType()->getScalarType())->getElementType())((ResultElementType == cast<PointerType>(getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("ResultElementType == cast<PointerType>(getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 1150, __PRETTY_FUNCTION__))
;
1151 init(Ptr, IdxList, NameStr);
1152}
1153
1154GetElementPtrInst::GetElementPtrInst(Type *PointeeType, Value *Ptr,
1155 ArrayRef<Value *> IdxList, unsigned Values,
1156 const Twine &NameStr,
1157 BasicBlock *InsertAtEnd)
1158 : Instruction(getGEPReturnType(PointeeType, Ptr, IdxList), GetElementPtr,
1159 OperandTraits<GetElementPtrInst>::op_end(this) - Values,
1160 Values, InsertAtEnd),
1161 SourceElementType(PointeeType),
1162 ResultElementType(getIndexedType(PointeeType, IdxList)) {
1163 assert(ResultElementType ==((ResultElementType == cast<PointerType>(getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("ResultElementType == cast<PointerType>(getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 1164, __PRETTY_FUNCTION__))
1164 cast<PointerType>(getType()->getScalarType())->getElementType())((ResultElementType == cast<PointerType>(getType()->
getScalarType())->getElementType()) ? static_cast<void>
(0) : __assert_fail ("ResultElementType == cast<PointerType>(getType()->getScalarType())->getElementType()"
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 1164, __PRETTY_FUNCTION__))
;
1165 init(Ptr, IdxList, NameStr);
1166}
1167
1168DEFINE_TRANSPARENT_OPERAND_ACCESSORS(GetElementPtrInst, Value)GetElementPtrInst::op_iterator GetElementPtrInst::op_begin() {
return OperandTraits<GetElementPtrInst>::op_begin(this
); } GetElementPtrInst::const_op_iterator GetElementPtrInst::
op_begin() const { return OperandTraits<GetElementPtrInst>
::op_begin(const_cast<GetElementPtrInst*>(this)); } GetElementPtrInst
::op_iterator GetElementPtrInst::op_end() { return OperandTraits
<GetElementPtrInst>::op_end(this); } GetElementPtrInst::
const_op_iterator GetElementPtrInst::op_end() const { return OperandTraits
<GetElementPtrInst>::op_end(const_cast<GetElementPtrInst
*>(this)); } Value *GetElementPtrInst::getOperand(unsigned
i_nocapture) const { ((i_nocapture < OperandTraits<GetElementPtrInst
>::operands(this) && "getOperand() out of range!")
? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<GetElementPtrInst>::operands(this) && \"getOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 1168, __PRETTY_FUNCTION__)); return cast_or_null<Value>
( OperandTraits<GetElementPtrInst>::op_begin(const_cast
<GetElementPtrInst*>(this))[i_nocapture].get()); } void
GetElementPtrInst::setOperand(unsigned i_nocapture, Value *Val_nocapture
) { ((i_nocapture < OperandTraits<GetElementPtrInst>
::operands(this) && "setOperand() out of range!") ? static_cast
<void> (0) : __assert_fail ("i_nocapture < OperandTraits<GetElementPtrInst>::operands(this) && \"setOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 1168, __PRETTY_FUNCTION__)); OperandTraits<GetElementPtrInst
>::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned
GetElementPtrInst::getNumOperands() const { return OperandTraits
<GetElementPtrInst>::operands(this); } template <int
Idx_nocapture> Use &GetElementPtrInst::Op() { return this
->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture
> const Use &GetElementPtrInst::Op() const { return this
->OpFrom<Idx_nocapture>(this); }
1169
1170//===----------------------------------------------------------------------===//
1171// ICmpInst Class
1172//===----------------------------------------------------------------------===//
1173
1174/// This instruction compares its operands according to the predicate given
1175/// to the constructor. It only operates on integers or pointers. The operands
1176/// must be identical types.
1177/// Represent an integer comparison operator.
1178class ICmpInst: public CmpInst {
1179 void AssertOK() {
1180 assert(isIntPredicate() &&((isIntPredicate() && "Invalid ICmp predicate value")
? static_cast<void> (0) : __assert_fail ("isIntPredicate() && \"Invalid ICmp predicate value\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 1181, __PRETTY_FUNCTION__))
1181 "Invalid ICmp predicate value")((isIntPredicate() && "Invalid ICmp predicate value")
? static_cast<void> (0) : __assert_fail ("isIntPredicate() && \"Invalid ICmp predicate value\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 1181, __PRETTY_FUNCTION__))
;
1182 assert(getOperand(0)->getType() == getOperand(1)->getType() &&((getOperand(0)->getType() == getOperand(1)->getType() &&
"Both operands to ICmp instruction are not of the same type!"
) ? static_cast<void> (0) : __assert_fail ("getOperand(0)->getType() == getOperand(1)->getType() && \"Both operands to ICmp instruction are not of the same type!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 1183, __PRETTY_FUNCTION__))
1183 "Both operands to ICmp instruction are not of the same type!")((getOperand(0)->getType() == getOperand(1)->getType() &&
"Both operands to ICmp instruction are not of the same type!"
) ? static_cast<void> (0) : __assert_fail ("getOperand(0)->getType() == getOperand(1)->getType() && \"Both operands to ICmp instruction are not of the same type!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 1183, __PRETTY_FUNCTION__))
;
1184 // Check that the operands are the right type
1185 assert((getOperand(0)->getType()->isIntOrIntVectorTy() ||(((getOperand(0)->getType()->isIntOrIntVectorTy() || getOperand
(0)->getType()->isPtrOrPtrVectorTy()) && "Invalid operand types for ICmp instruction"
) ? static_cast<void> (0) : __assert_fail ("(getOperand(0)->getType()->isIntOrIntVectorTy() || getOperand(0)->getType()->isPtrOrPtrVectorTy()) && \"Invalid operand types for ICmp instruction\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 1187, __PRETTY_FUNCTION__))
1186 getOperand(0)->getType()->isPtrOrPtrVectorTy()) &&(((getOperand(0)->getType()->isIntOrIntVectorTy() || getOperand
(0)->getType()->isPtrOrPtrVectorTy()) && "Invalid operand types for ICmp instruction"
) ? static_cast<void> (0) : __assert_fail ("(getOperand(0)->getType()->isIntOrIntVectorTy() || getOperand(0)->getType()->isPtrOrPtrVectorTy()) && \"Invalid operand types for ICmp instruction\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 1187, __PRETTY_FUNCTION__))
1187 "Invalid operand types for ICmp instruction")(((getOperand(0)->getType()->isIntOrIntVectorTy() || getOperand
(0)->getType()->isPtrOrPtrVectorTy()) && "Invalid operand types for ICmp instruction"
) ? static_cast<void> (0) : __assert_fail ("(getOperand(0)->getType()->isIntOrIntVectorTy() || getOperand(0)->getType()->isPtrOrPtrVectorTy()) && \"Invalid operand types for ICmp instruction\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 1187, __PRETTY_FUNCTION__))
;
1188 }
1189
1190protected:
1191 // Note: Instruction needs to be a friend here to call cloneImpl.
1192 friend class Instruction;
1193
1194 /// Clone an identical ICmpInst
1195 ICmpInst *cloneImpl() const;
1196
1197public:
1198 /// Constructor with insert-before-instruction semantics.
1199 ICmpInst(
1200 Instruction *InsertBefore, ///< Where to insert
1201 Predicate pred, ///< The predicate to use for the comparison
1202 Value *LHS, ///< The left-hand-side of the expression
1203 Value *RHS, ///< The right-hand-side of the expression
1204 const Twine &NameStr = "" ///< Name of the instruction
1205 ) : CmpInst(makeCmpResultType(LHS->getType()),
1206 Instruction::ICmp, pred, LHS, RHS, NameStr,
1207 InsertBefore) {
1208#ifndef NDEBUG
1209 AssertOK();
1210#endif
1211 }
1212
1213 /// Constructor with insert-at-end semantics.
1214 ICmpInst(
1215 BasicBlock &InsertAtEnd, ///< Block to insert into.
1216 Predicate pred, ///< The predicate to use for the comparison
1217 Value *LHS, ///< The left-hand-side of the expression
1218 Value *RHS, ///< The right-hand-side of the expression
1219 const Twine &NameStr = "" ///< Name of the instruction
1220 ) : CmpInst(makeCmpResultType(LHS->getType()),
1221 Instruction::ICmp, pred, LHS, RHS, NameStr,
1222 &InsertAtEnd) {
1223#ifndef NDEBUG
1224 AssertOK();
1225#endif
1226 }
1227
1228 /// Constructor with no-insertion semantics
1229 ICmpInst(
1230 Predicate pred, ///< The predicate to use for the comparison
1231 Value *LHS, ///< The left-hand-side of the expression
1232 Value *RHS, ///< The right-hand-side of the expression
1233 const Twine &NameStr = "" ///< Name of the instruction
1234 ) : CmpInst(makeCmpResultType(LHS->getType()),
1235 Instruction::ICmp, pred, LHS, RHS, NameStr) {
1236#ifndef NDEBUG
1237 AssertOK();
1238#endif
1239 }
1240
1241 /// For example, EQ->EQ, SLE->SLE, UGT->SGT, etc.
1242 /// @returns the predicate that would be the result if the operand were
1243 /// regarded as signed.
1244 /// Return the signed version of the predicate
1245 Predicate getSignedPredicate() const {
1246 return getSignedPredicate(getPredicate());
1247 }
1248
1249 /// This is a static version that you can use without an instruction.
1250 /// Return the signed version of the predicate.
1251 static Predicate getSignedPredicate(Predicate pred);
1252
1253 /// For example, EQ->EQ, SLE->ULE, UGT->UGT, etc.
1254 /// @returns the predicate that would be the result if the operand were
1255 /// regarded as unsigned.
1256 /// Return the unsigned version of the predicate
1257 Predicate getUnsignedPredicate() const {
1258 return getUnsignedPredicate(getPredicate());
1259 }
1260
1261 /// This is a static version that you can use without an instruction.
1262 /// Return the unsigned version of the predicate.
1263 static Predicate getUnsignedPredicate(Predicate pred);
1264
1265 /// Return true if this predicate is either EQ or NE. This also
1266 /// tests for commutativity.
1267 static bool isEquality(Predicate P) {
1268 return P == ICMP_EQ || P == ICMP_NE;
1269 }
1270
1271 /// Return true if this predicate is either EQ or NE. This also
1272 /// tests for commutativity.
1273 bool isEquality() const {
1274 return isEquality(getPredicate());
1275 }
1276
1277 /// @returns true if the predicate of this ICmpInst is commutative
1278 /// Determine if this relation is commutative.
1279 bool isCommutative() const { return isEquality(); }
1280
1281 /// Return true if the predicate is relational (not EQ or NE).
1282 ///
1283 bool isRelational() const {
1284 return !isEquality();
1285 }
1286
1287 /// Return true if the predicate is relational (not EQ or NE).
1288 ///
1289 static bool isRelational(Predicate P) {
1290 return !isEquality(P);
1291 }
1292
1293 /// Return true if the predicate is SGT or UGT.
1294 ///
1295 static bool isGT(Predicate P) {
1296 return P == ICMP_SGT || P == ICMP_UGT;
1297 }
1298
1299 /// Return true if the predicate is SLT or ULT.
1300 ///
1301 static bool isLT(Predicate P) {
1302 return P == ICMP_SLT || P == ICMP_ULT;
1303 }
1304
1305 /// Return true if the predicate is SGE or UGE.
1306 ///
1307 static bool isGE(Predicate P) {
1308 return P == ICMP_SGE || P == ICMP_UGE;
1309 }
1310
1311 /// Return true if the predicate is SLE or ULE.
1312 ///
1313 static bool isLE(Predicate P) {
1314 return P == ICMP_SLE || P == ICMP_ULE;
1315 }
1316
1317 /// Exchange the two operands to this instruction in such a way that it does
1318 /// not modify the semantics of the instruction. The predicate value may be
1319 /// changed to retain the same result if the predicate is order dependent
1320 /// (e.g. ult).
1321 /// Swap operands and adjust predicate.
1322 void swapOperands() {
1323 setPredicate(getSwappedPredicate());
1324 Op<0>().swap(Op<1>());
1325 }
1326
1327 // Methods for support type inquiry through isa, cast, and dyn_cast:
1328 static bool classof(const Instruction *I) {
1329 return I->getOpcode() == Instruction::ICmp;
1330 }
1331 static bool classof(const Value *V) {
1332 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1333 }
1334};
1335
1336//===----------------------------------------------------------------------===//
1337// FCmpInst Class
1338//===----------------------------------------------------------------------===//
1339
1340/// This instruction compares its operands according to the predicate given
1341/// to the constructor. It only operates on floating point values or packed
1342/// vectors of floating point values. The operands must be identical types.
1343/// Represents a floating point comparison operator.
1344class FCmpInst: public CmpInst {
1345 void AssertOK() {
1346 assert(isFPPredicate() && "Invalid FCmp predicate value")((isFPPredicate() && "Invalid FCmp predicate value") ?
static_cast<void> (0) : __assert_fail ("isFPPredicate() && \"Invalid FCmp predicate value\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 1346, __PRETTY_FUNCTION__))
;
1347 assert(getOperand(0)->getType() == getOperand(1)->getType() &&((getOperand(0)->getType() == getOperand(1)->getType() &&
"Both operands to FCmp instruction are not of the same type!"
) ? static_cast<void> (0) : __assert_fail ("getOperand(0)->getType() == getOperand(1)->getType() && \"Both operands to FCmp instruction are not of the same type!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 1348, __PRETTY_FUNCTION__))
1348 "Both operands to FCmp instruction are not of the same type!")((getOperand(0)->getType() == getOperand(1)->getType() &&
"Both operands to FCmp instruction are not of the same type!"
) ? static_cast<void> (0) : __assert_fail ("getOperand(0)->getType() == getOperand(1)->getType() && \"Both operands to FCmp instruction are not of the same type!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 1348, __PRETTY_FUNCTION__))
;
1349 // Check that the operands are the right type
1350 assert(getOperand(0)->getType()->isFPOrFPVectorTy() &&((getOperand(0)->getType()->isFPOrFPVectorTy() &&
"Invalid operand types for FCmp instruction") ? static_cast<
void> (0) : __assert_fail ("getOperand(0)->getType()->isFPOrFPVectorTy() && \"Invalid operand types for FCmp instruction\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 1351, __PRETTY_FUNCTION__))
1351 "Invalid operand types for FCmp instruction")((getOperand(0)->getType()->isFPOrFPVectorTy() &&
"Invalid operand types for FCmp instruction") ? static_cast<
void> (0) : __assert_fail ("getOperand(0)->getType()->isFPOrFPVectorTy() && \"Invalid operand types for FCmp instruction\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 1351, __PRETTY_FUNCTION__))
;
1352 }
1353
1354protected:
1355 // Note: Instruction needs to be a friend here to call cloneImpl.
1356 friend class Instruction;
1357
1358 /// Clone an identical FCmpInst
1359 FCmpInst *cloneImpl() const;
1360
1361public:
1362 /// Constructor with insert-before-instruction semantics.
1363 FCmpInst(
1364 Instruction *InsertBefore, ///< Where to insert
1365 Predicate pred, ///< The predicate to use for the comparison
1366 Value *LHS, ///< The left-hand-side of the expression
1367 Value *RHS, ///< The right-hand-side of the expression
1368 const Twine &NameStr = "" ///< Name of the instruction
1369 ) : CmpInst(makeCmpResultType(LHS->getType()),
1370 Instruction::FCmp, pred, LHS, RHS, NameStr,
1371 InsertBefore) {
1372 AssertOK();
1373 }
1374
1375 /// Constructor with insert-at-end semantics.
1376 FCmpInst(
1377 BasicBlock &InsertAtEnd, ///< Block to insert into.
1378 Predicate pred, ///< The predicate to use for the comparison
1379 Value *LHS, ///< The left-hand-side of the expression
1380 Value *RHS, ///< The right-hand-side of the expression
1381 const Twine &NameStr = "" ///< Name of the instruction
1382 ) : CmpInst(makeCmpResultType(LHS->getType()),
1383 Instruction::FCmp, pred, LHS, RHS, NameStr,
1384 &InsertAtEnd) {
1385 AssertOK();
1386 }
1387
1388 /// Constructor with no-insertion semantics
1389 FCmpInst(
1390 Predicate Pred, ///< The predicate to use for the comparison
1391 Value *LHS, ///< The left-hand-side of the expression
1392 Value *RHS, ///< The right-hand-side of the expression
1393 const Twine &NameStr = "", ///< Name of the instruction
1394 Instruction *FlagsSource = nullptr
1395 ) : CmpInst(makeCmpResultType(LHS->getType()), Instruction::FCmp, Pred, LHS,
1396 RHS, NameStr, nullptr, FlagsSource) {
1397 AssertOK();
1398 }
1399
1400 /// @returns true if the predicate of this instruction is EQ or NE.
1401 /// Determine if this is an equality predicate.
1402 static bool isEquality(Predicate Pred) {
1403 return Pred == FCMP_OEQ || Pred == FCMP_ONE || Pred == FCMP_UEQ ||
1404 Pred == FCMP_UNE;
1405 }
1406
1407 /// @returns true if the predicate of this instruction is EQ or NE.
1408 /// Determine if this is an equality predicate.
1409 bool isEquality() const { return isEquality(getPredicate()); }
1410
1411 /// @returns true if the predicate of this instruction is commutative.
1412 /// Determine if this is a commutative predicate.
1413 bool isCommutative() const {
1414 return isEquality() ||
1415 getPredicate() == FCMP_FALSE ||
1416 getPredicate() == FCMP_TRUE ||
1417 getPredicate() == FCMP_ORD ||
1418 getPredicate() == FCMP_UNO;
1419 }
1420
1421 /// @returns true if the predicate is relational (not EQ or NE).
1422 /// Determine if this a relational predicate.
1423 bool isRelational() const { return !isEquality(); }
1424
1425 /// Exchange the two operands to this instruction in such a way that it does
1426 /// not modify the semantics of the instruction. The predicate value may be
1427 /// changed to retain the same result if the predicate is order dependent
1428 /// (e.g. ult).
1429 /// Swap operands and adjust predicate.
1430 void swapOperands() {
1431 setPredicate(getSwappedPredicate());
1432 Op<0>().swap(Op<1>());
1433 }
1434
1435 /// Methods for support type inquiry through isa, cast, and dyn_cast:
1436 static bool classof(const Instruction *I) {
1437 return I->getOpcode() == Instruction::FCmp;
1438 }
1439 static bool classof(const Value *V) {
1440 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1441 }
1442};
1443
1444//===----------------------------------------------------------------------===//
1445/// This class represents a function call, abstracting a target
1446/// machine's calling convention. This class uses low bit of the SubClassData
1447/// field to indicate whether or not this is a tail call. The rest of the bits
1448/// hold the calling convention of the call.
1449///
1450class CallInst : public CallBase {
1451 CallInst(const CallInst &CI);
1452
1453 /// Construct a CallInst given a range of arguments.
1454 /// Construct a CallInst from a range of arguments
1455 inline CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1456 ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr,
1457 Instruction *InsertBefore);
1458
1459 inline CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1460 const Twine &NameStr, Instruction *InsertBefore)
1461 : CallInst(Ty, Func, Args, None, NameStr, InsertBefore) {}
1462
1463 /// Construct a CallInst given a range of arguments.
1464 /// Construct a CallInst from a range of arguments
1465 inline CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1466 ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr,
1467 BasicBlock *InsertAtEnd);
1468
1469 explicit CallInst(FunctionType *Ty, Value *F, const Twine &NameStr,
1470 Instruction *InsertBefore);
1471
1472 CallInst(FunctionType *ty, Value *F, const Twine &NameStr,
1473 BasicBlock *InsertAtEnd);
1474
1475 void init(FunctionType *FTy, Value *Func, ArrayRef<Value *> Args,
1476 ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr);
1477 void init(FunctionType *FTy, Value *Func, const Twine &NameStr);
1478
1479 /// Compute the number of operands to allocate.
1480 static int ComputeNumOperands(int NumArgs, int NumBundleInputs = 0) {
1481 // We need one operand for the called function, plus the input operand
1482 // counts provided.
1483 return 1 + NumArgs + NumBundleInputs;
1484 }
1485
1486protected:
1487 // Note: Instruction needs to be a friend here to call cloneImpl.
1488 friend class Instruction;
1489
1490 CallInst *cloneImpl() const;
1491
1492public:
1493 static CallInst *Create(FunctionType *Ty, Value *F, const Twine &NameStr = "",
1494 Instruction *InsertBefore = nullptr) {
1495 return new (ComputeNumOperands(0)) CallInst(Ty, F, NameStr, InsertBefore);
1496 }
1497
1498 static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1499 const Twine &NameStr,
1500 Instruction *InsertBefore = nullptr) {
1501 return new (ComputeNumOperands(Args.size()))
1502 CallInst(Ty, Func, Args, None, NameStr, InsertBefore);
1503 }
1504
1505 static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1506 ArrayRef<OperandBundleDef> Bundles = None,
1507 const Twine &NameStr = "",
1508 Instruction *InsertBefore = nullptr) {
1509 const int NumOperands =
1510 ComputeNumOperands(Args.size(), CountBundleInputs(Bundles));
1511 const unsigned DescriptorBytes = Bundles.size() * sizeof(BundleOpInfo);
1512
1513 return new (NumOperands, DescriptorBytes)
1514 CallInst(Ty, Func, Args, Bundles, NameStr, InsertBefore);
1515 }
1516
1517 static CallInst *Create(FunctionType *Ty, Value *F, const Twine &NameStr,
1518 BasicBlock *InsertAtEnd) {
1519 return new (ComputeNumOperands(0)) CallInst(Ty, F, NameStr, InsertAtEnd);
1520 }
1521
1522 static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1523 const Twine &NameStr, BasicBlock *InsertAtEnd) {
1524 return new (ComputeNumOperands(Args.size()))
1525 CallInst(Ty, Func, Args, None, NameStr, InsertAtEnd);
1526 }
1527
1528 static CallInst *Create(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1529 ArrayRef<OperandBundleDef> Bundles,
1530 const Twine &NameStr, BasicBlock *InsertAtEnd) {
1531 const int NumOperands =
1532 ComputeNumOperands(Args.size(), CountBundleInputs(Bundles));
1533 const unsigned DescriptorBytes = Bundles.size() * sizeof(BundleOpInfo);
1534
1535 return new (NumOperands, DescriptorBytes)
1536 CallInst(Ty, Func, Args, Bundles, NameStr, InsertAtEnd);
1537 }
1538
1539 static CallInst *Create(FunctionCallee Func, const Twine &NameStr = "",
1540 Instruction *InsertBefore = nullptr) {
1541 return Create(Func.getFunctionType(), Func.getCallee(), NameStr,
1542 InsertBefore);
1543 }
1544
1545 static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args,
1546 ArrayRef<OperandBundleDef> Bundles = None,
1547 const Twine &NameStr = "",
1548 Instruction *InsertBefore = nullptr) {
1549 return Create(Func.getFunctionType(), Func.getCallee(), Args, Bundles,
1550 NameStr, InsertBefore);
1551 }
1552
1553 static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args,
1554 const Twine &NameStr,
1555 Instruction *InsertBefore = nullptr) {
1556 return Create(Func.getFunctionType(), Func.getCallee(), Args, NameStr,
1557 InsertBefore);
1558 }
1559
1560 static CallInst *Create(FunctionCallee Func, const Twine &NameStr,
1561 BasicBlock *InsertAtEnd) {
1562 return Create(Func.getFunctionType(), Func.getCallee(), NameStr,
1563 InsertAtEnd);
1564 }
1565
1566 static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args,
1567 const Twine &NameStr, BasicBlock *InsertAtEnd) {
1568 return Create(Func.getFunctionType(), Func.getCallee(), Args, NameStr,
1569 InsertAtEnd);
1570 }
1571
1572 static CallInst *Create(FunctionCallee Func, ArrayRef<Value *> Args,
1573 ArrayRef<OperandBundleDef> Bundles,
1574 const Twine &NameStr, BasicBlock *InsertAtEnd) {
1575 return Create(Func.getFunctionType(), Func.getCallee(), Args, Bundles,
1576 NameStr, InsertAtEnd);
1577 }
1578
1579 /// Create a clone of \p CI with a different set of operand bundles and
1580 /// insert it before \p InsertPt.
1581 ///
1582 /// The returned call instruction is identical \p CI in every way except that
1583 /// the operand bundles for the new instruction are set to the operand bundles
1584 /// in \p Bundles.
1585 static CallInst *Create(CallInst *CI, ArrayRef<OperandBundleDef> Bundles,
1586 Instruction *InsertPt = nullptr);
1587
1588 /// Create a clone of \p CI with a different set of operand bundles and
1589 /// insert it before \p InsertPt.
1590 ///
1591 /// The returned call instruction is identical \p CI in every way except that
1592 /// the operand bundle for the new instruction is set to the operand bundle
1593 /// in \p Bundle.
1594 static CallInst *CreateWithReplacedBundle(CallInst *CI,
1595 OperandBundleDef Bundle,
1596 Instruction *InsertPt = nullptr);
1597
1598 /// Generate the IR for a call to malloc:
1599 /// 1. Compute the malloc call's argument as the specified type's size,
1600 /// possibly multiplied by the array size if the array size is not
1601 /// constant 1.
1602 /// 2. Call malloc with that argument.
1603 /// 3. Bitcast the result of the malloc call to the specified type.
1604 static Instruction *CreateMalloc(Instruction *InsertBefore, Type *IntPtrTy,
1605 Type *AllocTy, Value *AllocSize,
1606 Value *ArraySize = nullptr,
1607 Function *MallocF = nullptr,
1608 const Twine &Name = "");
1609 static Instruction *CreateMalloc(BasicBlock *InsertAtEnd, Type *IntPtrTy,
1610 Type *AllocTy, Value *AllocSize,
1611 Value *ArraySize = nullptr,
1612 Function *MallocF = nullptr,
1613 const Twine &Name = "");
1614 static Instruction *CreateMalloc(Instruction *InsertBefore, Type *IntPtrTy,
1615 Type *AllocTy, Value *AllocSize,
1616 Value *ArraySize = nullptr,
1617 ArrayRef<OperandBundleDef> Bundles = None,
1618 Function *MallocF = nullptr,
1619 const Twine &Name = "");
1620 static Instruction *CreateMalloc(BasicBlock *InsertAtEnd, Type *IntPtrTy,
1621 Type *AllocTy, Value *AllocSize,
1622 Value *ArraySize = nullptr,
1623 ArrayRef<OperandBundleDef> Bundles = None,
1624 Function *MallocF = nullptr,
1625 const Twine &Name = "");
1626 /// Generate the IR for a call to the builtin free function.
1627 static Instruction *CreateFree(Value *Source, Instruction *InsertBefore);
1628 static Instruction *CreateFree(Value *Source, BasicBlock *InsertAtEnd);
1629 static Instruction *CreateFree(Value *Source,
1630 ArrayRef<OperandBundleDef> Bundles,
1631 Instruction *InsertBefore);
1632 static Instruction *CreateFree(Value *Source,
1633 ArrayRef<OperandBundleDef> Bundles,
1634 BasicBlock *InsertAtEnd);
1635
1636 // Note that 'musttail' implies 'tail'.
1637 enum TailCallKind : unsigned {
1638 TCK_None = 0,
1639 TCK_Tail = 1,
1640 TCK_MustTail = 2,
1641 TCK_NoTail = 3,
1642 TCK_LAST = TCK_NoTail
1643 };
1644
1645 using TailCallKindField = Bitfield::Element<TailCallKind, 0, 2, TCK_LAST>;
1646 static_assert(
1647 Bitfield::areContiguous<TailCallKindField, CallBase::CallingConvField>(),
1648 "Bitfields must be contiguous");
1649
1650 TailCallKind getTailCallKind() const {
1651 return getSubclassData<TailCallKindField>();
1652 }
1653
1654 bool isTailCall() const {
1655 TailCallKind Kind = getTailCallKind();
1656 return Kind == TCK_Tail || Kind == TCK_MustTail;
1657 }
1658
1659 bool isMustTailCall() const { return getTailCallKind() == TCK_MustTail; }
1660
1661 bool isNoTailCall() const { return getTailCallKind() == TCK_NoTail; }
1662
1663 void setTailCallKind(TailCallKind TCK) {
1664 setSubclassData<TailCallKindField>(TCK);
1665 }
1666
1667 void setTailCall(bool IsTc = true) {
1668 setTailCallKind(IsTc ? TCK_Tail : TCK_None);
1669 }
1670
1671 /// Return true if the call can return twice
1672 bool canReturnTwice() const { return hasFnAttr(Attribute::ReturnsTwice); }
1673 void setCanReturnTwice() {
1674 addAttribute(AttributeList::FunctionIndex, Attribute::ReturnsTwice);
1675 }
1676
1677 // Methods for support type inquiry through isa, cast, and dyn_cast:
1678 static bool classof(const Instruction *I) {
1679 return I->getOpcode() == Instruction::Call;
1680 }
1681 static bool classof(const Value *V) {
1682 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1683 }
1684
1685 /// Updates profile metadata by scaling it by \p S / \p T.
1686 void updateProfWeight(uint64_t S, uint64_t T);
1687
1688private:
1689 // Shadow Instruction::setInstructionSubclassData with a private forwarding
1690 // method so that subclasses cannot accidentally use it.
1691 template <typename Bitfield>
1692 void setSubclassData(typename Bitfield::Type Value) {
1693 Instruction::setSubclassData<Bitfield>(Value);
1694 }
1695};
1696
1697CallInst::CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1698 ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr,
1699 BasicBlock *InsertAtEnd)
1700 : CallBase(Ty->getReturnType(), Instruction::Call,
1701 OperandTraits<CallBase>::op_end(this) -
1702 (Args.size() + CountBundleInputs(Bundles) + 1),
1703 unsigned(Args.size() + CountBundleInputs(Bundles) + 1),
1704 InsertAtEnd) {
1705 init(Ty, Func, Args, Bundles, NameStr);
1706}
1707
1708CallInst::CallInst(FunctionType *Ty, Value *Func, ArrayRef<Value *> Args,
1709 ArrayRef<OperandBundleDef> Bundles, const Twine &NameStr,
1710 Instruction *InsertBefore)
1711 : CallBase(Ty->getReturnType(), Instruction::Call,
1712 OperandTraits<CallBase>::op_end(this) -
1713 (Args.size() + CountBundleInputs(Bundles) + 1),
1714 unsigned(Args.size() + CountBundleInputs(Bundles) + 1),
1715 InsertBefore) {
1716 init(Ty, Func, Args, Bundles, NameStr);
1717}
1718
1719//===----------------------------------------------------------------------===//
1720// SelectInst Class
1721//===----------------------------------------------------------------------===//
1722
1723/// This class represents the LLVM 'select' instruction.
1724///
1725class SelectInst : public Instruction {
1726 SelectInst(Value *C, Value *S1, Value *S2, const Twine &NameStr,
1727 Instruction *InsertBefore)
1728 : Instruction(S1->getType(), Instruction::Select,
1729 &Op<0>(), 3, InsertBefore) {
1730 init(C, S1, S2);
1731 setName(NameStr);
1732 }
1733
1734 SelectInst(Value *C, Value *S1, Value *S2, const Twine &NameStr,
1735 BasicBlock *InsertAtEnd)
1736 : Instruction(S1->getType(), Instruction::Select,
1737 &Op<0>(), 3, InsertAtEnd) {
1738 init(C, S1, S2);
1739 setName(NameStr);
1740 }
1741
1742 void init(Value *C, Value *S1, Value *S2) {
1743 assert(!areInvalidOperands(C, S1, S2) && "Invalid operands for select")((!areInvalidOperands(C, S1, S2) && "Invalid operands for select"
) ? static_cast<void> (0) : __assert_fail ("!areInvalidOperands(C, S1, S2) && \"Invalid operands for select\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 1743, __PRETTY_FUNCTION__))
;
1744 Op<0>() = C;
1745 Op<1>() = S1;
1746 Op<2>() = S2;
1747 }
1748
1749protected:
1750 // Note: Instruction needs to be a friend here to call cloneImpl.
1751 friend class Instruction;
1752
1753 SelectInst *cloneImpl() const;
1754
1755public:
1756 static SelectInst *Create(Value *C, Value *S1, Value *S2,
1757 const Twine &NameStr = "",
1758 Instruction *InsertBefore = nullptr,
1759 Instruction *MDFrom = nullptr) {
1760 SelectInst *Sel = new(3) SelectInst(C, S1, S2, NameStr, InsertBefore);
1761 if (MDFrom)
1762 Sel->copyMetadata(*MDFrom);
1763 return Sel;
1764 }
1765
1766 static SelectInst *Create(Value *C, Value *S1, Value *S2,
1767 const Twine &NameStr,
1768 BasicBlock *InsertAtEnd) {
1769 return new(3) SelectInst(C, S1, S2, NameStr, InsertAtEnd);
1770 }
1771
1772 const Value *getCondition() const { return Op<0>(); }
1773 const Value *getTrueValue() const { return Op<1>(); }
1774 const Value *getFalseValue() const { return Op<2>(); }
1775 Value *getCondition() { return Op<0>(); }
1776 Value *getTrueValue() { return Op<1>(); }
1777 Value *getFalseValue() { return Op<2>(); }
1778
1779 void setCondition(Value *V) { Op<0>() = V; }
1780 void setTrueValue(Value *V) { Op<1>() = V; }
1781 void setFalseValue(Value *V) { Op<2>() = V; }
1782
1783 /// Swap the true and false values of the select instruction.
1784 /// This doesn't swap prof metadata.
1785 void swapValues() { Op<1>().swap(Op<2>()); }
1786
1787 /// Return a string if the specified operands are invalid
1788 /// for a select operation, otherwise return null.
1789 static const char *areInvalidOperands(Value *Cond, Value *True, Value *False);
1790
1791 /// Transparently provide more efficient getOperand methods.
1792 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
1793
1794 OtherOps getOpcode() const {
1795 return static_cast<OtherOps>(Instruction::getOpcode());
1796 }
1797
1798 // Methods for support type inquiry through isa, cast, and dyn_cast:
1799 static bool classof(const Instruction *I) {
1800 return I->getOpcode() == Instruction::Select;
1801 }
1802 static bool classof(const Value *V) {
1803 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1804 }
1805};
1806
1807template <>
1808struct OperandTraits<SelectInst> : public FixedNumOperandTraits<SelectInst, 3> {
1809};
1810
1811DEFINE_TRANSPARENT_OPERAND_ACCESSORS(SelectInst, Value)SelectInst::op_iterator SelectInst::op_begin() { return OperandTraits
<SelectInst>::op_begin(this); } SelectInst::const_op_iterator
SelectInst::op_begin() const { return OperandTraits<SelectInst
>::op_begin(const_cast<SelectInst*>(this)); } SelectInst
::op_iterator SelectInst::op_end() { return OperandTraits<
SelectInst>::op_end(this); } SelectInst::const_op_iterator
SelectInst::op_end() const { return OperandTraits<SelectInst
>::op_end(const_cast<SelectInst*>(this)); } Value *SelectInst
::getOperand(unsigned i_nocapture) const { ((i_nocapture <
OperandTraits<SelectInst>::operands(this) && "getOperand() out of range!"
) ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<SelectInst>::operands(this) && \"getOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 1811, __PRETTY_FUNCTION__)); return cast_or_null<Value>
( OperandTraits<SelectInst>::op_begin(const_cast<SelectInst
*>(this))[i_nocapture].get()); } void SelectInst::setOperand
(unsigned i_nocapture, Value *Val_nocapture) { ((i_nocapture <
OperandTraits<SelectInst>::operands(this) && "setOperand() out of range!"
) ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<SelectInst>::operands(this) && \"setOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 1811, __PRETTY_FUNCTION__)); OperandTraits<SelectInst>
::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned SelectInst
::getNumOperands() const { return OperandTraits<SelectInst
>::operands(this); } template <int Idx_nocapture> Use
&SelectInst::Op() { return this->OpFrom<Idx_nocapture
>(this); } template <int Idx_nocapture> const Use &
SelectInst::Op() const { return this->OpFrom<Idx_nocapture
>(this); }
1812
1813//===----------------------------------------------------------------------===//
1814// VAArgInst Class
1815//===----------------------------------------------------------------------===//
1816
1817/// This class represents the va_arg llvm instruction, which returns
1818/// an argument of the specified type given a va_list and increments that list
1819///
1820class VAArgInst : public UnaryInstruction {
1821protected:
1822 // Note: Instruction needs to be a friend here to call cloneImpl.
1823 friend class Instruction;
1824
1825 VAArgInst *cloneImpl() const;
1826
1827public:
1828 VAArgInst(Value *List, Type *Ty, const Twine &NameStr = "",
1829 Instruction *InsertBefore = nullptr)
1830 : UnaryInstruction(Ty, VAArg, List, InsertBefore) {
1831 setName(NameStr);
1832 }
1833
1834 VAArgInst(Value *List, Type *Ty, const Twine &NameStr,
1835 BasicBlock *InsertAtEnd)
1836 : UnaryInstruction(Ty, VAArg, List, InsertAtEnd) {
1837 setName(NameStr);
1838 }
1839
1840 Value *getPointerOperand() { return getOperand(0); }
1841 const Value *getPointerOperand() const { return getOperand(0); }
1842 static unsigned getPointerOperandIndex() { return 0U; }
1843
1844 // Methods for support type inquiry through isa, cast, and dyn_cast:
1845 static bool classof(const Instruction *I) {
1846 return I->getOpcode() == VAArg;
1847 }
1848 static bool classof(const Value *V) {
1849 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1850 }
1851};
1852
1853//===----------------------------------------------------------------------===//
1854// ExtractElementInst Class
1855//===----------------------------------------------------------------------===//
1856
1857/// This instruction extracts a single (scalar)
1858/// element from a VectorType value
1859///
1860class ExtractElementInst : public Instruction {
1861 ExtractElementInst(Value *Vec, Value *Idx, const Twine &NameStr = "",
1862 Instruction *InsertBefore = nullptr);
1863 ExtractElementInst(Value *Vec, Value *Idx, const Twine &NameStr,
1864 BasicBlock *InsertAtEnd);
1865
1866protected:
1867 // Note: Instruction needs to be a friend here to call cloneImpl.
1868 friend class Instruction;
1869
1870 ExtractElementInst *cloneImpl() const;
1871
1872public:
1873 static ExtractElementInst *Create(Value *Vec, Value *Idx,
1874 const Twine &NameStr = "",
1875 Instruction *InsertBefore = nullptr) {
1876 return new(2) ExtractElementInst(Vec, Idx, NameStr, InsertBefore);
1877 }
1878
1879 static ExtractElementInst *Create(Value *Vec, Value *Idx,
1880 const Twine &NameStr,
1881 BasicBlock *InsertAtEnd) {
1882 return new(2) ExtractElementInst(Vec, Idx, NameStr, InsertAtEnd);
1883 }
1884
1885 /// Return true if an extractelement instruction can be
1886 /// formed with the specified operands.
1887 static bool isValidOperands(const Value *Vec, const Value *Idx);
1888
1889 Value *getVectorOperand() { return Op<0>(); }
1890 Value *getIndexOperand() { return Op<1>(); }
1891 const Value *getVectorOperand() const { return Op<0>(); }
1892 const Value *getIndexOperand() const { return Op<1>(); }
1893
1894 VectorType *getVectorOperandType() const {
1895 return cast<VectorType>(getVectorOperand()->getType());
1896 }
1897
1898 /// Transparently provide more efficient getOperand methods.
1899 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
1900
1901 // Methods for support type inquiry through isa, cast, and dyn_cast:
1902 static bool classof(const Instruction *I) {
1903 return I->getOpcode() == Instruction::ExtractElement;
1904 }
1905 static bool classof(const Value *V) {
1906 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1907 }
1908};
1909
1910template <>
1911struct OperandTraits<ExtractElementInst> :
1912 public FixedNumOperandTraits<ExtractElementInst, 2> {
1913};
1914
1915DEFINE_TRANSPARENT_OPERAND_ACCESSORS(ExtractElementInst, Value)ExtractElementInst::op_iterator ExtractElementInst::op_begin(
) { return OperandTraits<ExtractElementInst>::op_begin(
this); } ExtractElementInst::const_op_iterator ExtractElementInst
::op_begin() const { return OperandTraits<ExtractElementInst
>::op_begin(const_cast<ExtractElementInst*>(this)); }
ExtractElementInst::op_iterator ExtractElementInst::op_end()
{ return OperandTraits<ExtractElementInst>::op_end(this
); } ExtractElementInst::const_op_iterator ExtractElementInst
::op_end() const { return OperandTraits<ExtractElementInst
>::op_end(const_cast<ExtractElementInst*>(this)); } Value
*ExtractElementInst::getOperand(unsigned i_nocapture) const {
((i_nocapture < OperandTraits<ExtractElementInst>::
operands(this) && "getOperand() out of range!") ? static_cast
<void> (0) : __assert_fail ("i_nocapture < OperandTraits<ExtractElementInst>::operands(this) && \"getOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 1915, __PRETTY_FUNCTION__)); return cast_or_null<Value>
( OperandTraits<ExtractElementInst>::op_begin(const_cast
<ExtractElementInst*>(this))[i_nocapture].get()); } void
ExtractElementInst::setOperand(unsigned i_nocapture, Value *
Val_nocapture) { ((i_nocapture < OperandTraits<ExtractElementInst
>::operands(this) && "setOperand() out of range!")
? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<ExtractElementInst>::operands(this) && \"setOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 1915, __PRETTY_FUNCTION__)); OperandTraits<ExtractElementInst
>::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned
ExtractElementInst::getNumOperands() const { return OperandTraits
<ExtractElementInst>::operands(this); } template <int
Idx_nocapture> Use &ExtractElementInst::Op() { return
this->OpFrom<Idx_nocapture>(this); } template <int
Idx_nocapture> const Use &ExtractElementInst::Op() const
{ return this->OpFrom<Idx_nocapture>(this); }
1916
1917//===----------------------------------------------------------------------===//
1918// InsertElementInst Class
1919//===----------------------------------------------------------------------===//
1920
1921/// This instruction inserts a single (scalar)
1922/// element into a VectorType value
1923///
1924class InsertElementInst : public Instruction {
1925 InsertElementInst(Value *Vec, Value *NewElt, Value *Idx,
1926 const Twine &NameStr = "",
1927 Instruction *InsertBefore = nullptr);
1928 InsertElementInst(Value *Vec, Value *NewElt, Value *Idx, const Twine &NameStr,
1929 BasicBlock *InsertAtEnd);
1930
1931protected:
1932 // Note: Instruction needs to be a friend here to call cloneImpl.
1933 friend class Instruction;
1934
1935 InsertElementInst *cloneImpl() const;
1936
1937public:
1938 static InsertElementInst *Create(Value *Vec, Value *NewElt, Value *Idx,
1939 const Twine &NameStr = "",
1940 Instruction *InsertBefore = nullptr) {
1941 return new(3) InsertElementInst(Vec, NewElt, Idx, NameStr, InsertBefore);
1942 }
1943
1944 static InsertElementInst *Create(Value *Vec, Value *NewElt, Value *Idx,
1945 const Twine &NameStr,
1946 BasicBlock *InsertAtEnd) {
1947 return new(3) InsertElementInst(Vec, NewElt, Idx, NameStr, InsertAtEnd);
1948 }
1949
1950 /// Return true if an insertelement instruction can be
1951 /// formed with the specified operands.
1952 static bool isValidOperands(const Value *Vec, const Value *NewElt,
1953 const Value *Idx);
1954
1955 /// Overload to return most specific vector type.
1956 ///
1957 VectorType *getType() const {
1958 return cast<VectorType>(Instruction::getType());
1959 }
1960
1961 /// Transparently provide more efficient getOperand methods.
1962 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
1963
1964 // Methods for support type inquiry through isa, cast, and dyn_cast:
1965 static bool classof(const Instruction *I) {
1966 return I->getOpcode() == Instruction::InsertElement;
1967 }
1968 static bool classof(const Value *V) {
1969 return isa<Instruction>(V) && classof(cast<Instruction>(V));
1970 }
1971};
1972
1973template <>
1974struct OperandTraits<InsertElementInst> :
1975 public FixedNumOperandTraits<InsertElementInst, 3> {
1976};
1977
1978DEFINE_TRANSPARENT_OPERAND_ACCESSORS(InsertElementInst, Value)InsertElementInst::op_iterator InsertElementInst::op_begin() {
return OperandTraits<InsertElementInst>::op_begin(this
); } InsertElementInst::const_op_iterator InsertElementInst::
op_begin() const { return OperandTraits<InsertElementInst>
::op_begin(const_cast<InsertElementInst*>(this)); } InsertElementInst
::op_iterator InsertElementInst::op_end() { return OperandTraits
<InsertElementInst>::op_end(this); } InsertElementInst::
const_op_iterator InsertElementInst::op_end() const { return OperandTraits
<InsertElementInst>::op_end(const_cast<InsertElementInst
*>(this)); } Value *InsertElementInst::getOperand(unsigned
i_nocapture) const { ((i_nocapture < OperandTraits<InsertElementInst
>::operands(this) && "getOperand() out of range!")
? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<InsertElementInst>::operands(this) && \"getOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 1978, __PRETTY_FUNCTION__)); return cast_or_null<Value>
( OperandTraits<InsertElementInst>::op_begin(const_cast
<InsertElementInst*>(this))[i_nocapture].get()); } void
InsertElementInst::setOperand(unsigned i_nocapture, Value *Val_nocapture
) { ((i_nocapture < OperandTraits<InsertElementInst>
::operands(this) && "setOperand() out of range!") ? static_cast
<void> (0) : __assert_fail ("i_nocapture < OperandTraits<InsertElementInst>::operands(this) && \"setOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 1978, __PRETTY_FUNCTION__)); OperandTraits<InsertElementInst
>::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned
InsertElementInst::getNumOperands() const { return OperandTraits
<InsertElementInst>::operands(this); } template <int
Idx_nocapture> Use &InsertElementInst::Op() { return this
->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture
> const Use &InsertElementInst::Op() const { return this
->OpFrom<Idx_nocapture>(this); }
1979
1980//===----------------------------------------------------------------------===//
1981// ShuffleVectorInst Class
1982//===----------------------------------------------------------------------===//
1983
1984constexpr int UndefMaskElem = -1;
1985
1986/// This instruction constructs a fixed permutation of two
1987/// input vectors.
1988///
1989/// For each element of the result vector, the shuffle mask selects an element
1990/// from one of the input vectors to copy to the result. Non-negative elements
1991/// in the mask represent an index into the concatenated pair of input vectors.
1992/// UndefMaskElem (-1) specifies that the result element is undefined.
1993///
1994/// For scalable vectors, all the elements of the mask must be 0 or -1. This
1995/// requirement may be relaxed in the future.
1996class ShuffleVectorInst : public Instruction {
1997 SmallVector<int, 4> ShuffleMask;
1998 Constant *ShuffleMaskForBitcode;
1999
2000protected:
2001 // Note: Instruction needs to be a friend here to call cloneImpl.
2002 friend class Instruction;
2003
2004 ShuffleVectorInst *cloneImpl() const;
2005
2006public:
2007 ShuffleVectorInst(Value *V1, Value *V2, Value *Mask,
2008 const Twine &NameStr = "",
2009 Instruction *InsertBefor = nullptr);
2010 ShuffleVectorInst(Value *V1, Value *V2, Value *Mask,
2011 const Twine &NameStr, BasicBlock *InsertAtEnd);
2012 ShuffleVectorInst(Value *V1, Value *V2, ArrayRef<int> Mask,
2013 const Twine &NameStr = "",
2014 Instruction *InsertBefor = nullptr);
2015 ShuffleVectorInst(Value *V1, Value *V2, ArrayRef<int> Mask,
2016 const Twine &NameStr, BasicBlock *InsertAtEnd);
2017
2018 void *operator new(size_t s) { return User::operator new(s, 2); }
2019
2020 /// Swap the operands and adjust the mask to preserve the semantics
2021 /// of the instruction.
2022 void commute();
2023
2024 /// Return true if a shufflevector instruction can be
2025 /// formed with the specified operands.
2026 static bool isValidOperands(const Value *V1, const Value *V2,
2027 const Value *Mask);
2028 static bool isValidOperands(const Value *V1, const Value *V2,
2029 ArrayRef<int> Mask);
2030
2031 /// Overload to return most specific vector type.
2032 ///
2033 VectorType *getType() const {
2034 return cast<VectorType>(Instruction::getType());
2035 }
2036
2037 /// Transparently provide more efficient getOperand methods.
2038 DECLARE_TRANSPARENT_OPERAND_ACCESSORS(Value)public: inline Value *getOperand(unsigned) const; inline void
setOperand(unsigned, Value*); inline op_iterator op_begin();
inline const_op_iterator op_begin() const; inline op_iterator
op_end(); inline const_op_iterator op_end() const; protected
: template <int> inline Use &Op(); template <int
> inline const Use &Op() const; public: inline unsigned
getNumOperands() const
;
2039
2040 /// Return the shuffle mask value of this instruction for the given element
2041 /// index. Return UndefMaskElem if the element is undef.
2042 int getMaskValue(unsigned Elt) const { return ShuffleMask[Elt]; }
2043
2044 /// Convert the input shuffle mask operand to a vector of integers. Undefined
2045 /// elements of the mask are returned as UndefMaskElem.
2046 static void getShuffleMask(const Constant *Mask,
2047 SmallVectorImpl<int> &Result);
2048
2049 /// Return the mask for this instruction as a vector of integers. Undefined
2050 /// elements of the mask are returned as UndefMaskElem.
2051 void getShuffleMask(SmallVectorImpl<int> &Result) const {
2052 Result.assign(ShuffleMask.begin(), ShuffleMask.end());
2053 }
2054
2055 /// Return the mask for this instruction, for use in bitcode.
2056 ///
2057 /// TODO: This is temporary until we decide a new bitcode encoding for
2058 /// shufflevector.
2059 Constant *getShuffleMaskForBitcode() const { return ShuffleMaskForBitcode; }
2060
2061 static Constant *convertShuffleMaskForBitcode(ArrayRef<int> Mask,
2062 Type *ResultTy);
2063
2064 void setShuffleMask(ArrayRef<int> Mask);
2065
2066 ArrayRef<int> getShuffleMask() const { return ShuffleMask; }
2067
2068 /// Return true if this shuffle returns a vector with a different number of
2069 /// elements than its source vectors.
2070 /// Examples: shufflevector <4 x n> A, <4 x n> B, <1,2,3>
2071 /// shufflevector <4 x n> A, <4 x n> B, <1,2,3,4,5>
2072 bool changesLength() const {
2073 unsigned NumSourceElts = cast<VectorType>(Op<0>()->getType())
2074 ->getElementCount()
2075 .getKnownMinValue();
2076 unsigned NumMaskElts = ShuffleMask.size();
2077 return NumSourceElts != NumMaskElts;
2078 }
2079
2080 /// Return true if this shuffle returns a vector with a greater number of
2081 /// elements than its source vectors.
2082 /// Example: shufflevector <2 x n> A, <2 x n> B, <1,2,3>
2083 bool increasesLength() const {
2084 unsigned NumSourceElts =
2085 cast<FixedVectorType>(Op<0>()->getType())->getNumElements();
2086 unsigned NumMaskElts = ShuffleMask.size();
2087 return NumSourceElts < NumMaskElts;
2088 }
2089
2090 /// Return true if this shuffle mask chooses elements from exactly one source
2091 /// vector.
2092 /// Example: <7,5,undef,7>
2093 /// This assumes that vector operands are the same length as the mask.
2094 static bool isSingleSourceMask(ArrayRef<int> Mask);
2095 static bool isSingleSourceMask(const Constant *Mask) {
2096 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((Mask->getType()->isVectorTy() && "Shuffle needs vector constant."
) ? static_cast<void> (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2096, __PRETTY_FUNCTION__))
;
2097 SmallVector<int, 16> MaskAsInts;
2098 getShuffleMask(Mask, MaskAsInts);
2099 return isSingleSourceMask(MaskAsInts);
2100 }
2101
2102 /// Return true if this shuffle chooses elements from exactly one source
2103 /// vector without changing the length of that vector.
2104 /// Example: shufflevector <4 x n> A, <4 x n> B, <3,0,undef,3>
2105 /// TODO: Optionally allow length-changing shuffles.
2106 bool isSingleSource() const {
2107 return !changesLength() && isSingleSourceMask(ShuffleMask);
2108 }
2109
2110 /// Return true if this shuffle mask chooses elements from exactly one source
2111 /// vector without lane crossings. A shuffle using this mask is not
2112 /// necessarily a no-op because it may change the number of elements from its
2113 /// input vectors or it may provide demanded bits knowledge via undef lanes.
2114 /// Example: <undef,undef,2,3>
2115 static bool isIdentityMask(ArrayRef<int> Mask);
2116 static bool isIdentityMask(const Constant *Mask) {
2117 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((Mask->getType()->isVectorTy() && "Shuffle needs vector constant."
) ? static_cast<void> (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2117, __PRETTY_FUNCTION__))
;
2118 SmallVector<int, 16> MaskAsInts;
2119 getShuffleMask(Mask, MaskAsInts);
2120 return isIdentityMask(MaskAsInts);
2121 }
2122
2123 /// Return true if this shuffle chooses elements from exactly one source
2124 /// vector without lane crossings and does not change the number of elements
2125 /// from its input vectors.
2126 /// Example: shufflevector <4 x n> A, <4 x n> B, <4,undef,6,undef>
2127 bool isIdentity() const {
2128 return !changesLength() && isIdentityMask(ShuffleMask);
2129 }
2130
2131 /// Return true if this shuffle lengthens exactly one source vector with
2132 /// undefs in the high elements.
2133 bool isIdentityWithPadding() const;
2134
2135 /// Return true if this shuffle extracts the first N elements of exactly one
2136 /// source vector.
2137 bool isIdentityWithExtract() const;
2138
2139 /// Return true if this shuffle concatenates its 2 source vectors. This
2140 /// returns false if either input is undefined. In that case, the shuffle is
2141 /// is better classified as an identity with padding operation.
2142 bool isConcat() const;
2143
2144 /// Return true if this shuffle mask chooses elements from its source vectors
2145 /// without lane crossings. A shuffle using this mask would be
2146 /// equivalent to a vector select with a constant condition operand.
2147 /// Example: <4,1,6,undef>
2148 /// This returns false if the mask does not choose from both input vectors.
2149 /// In that case, the shuffle is better classified as an identity shuffle.
2150 /// This assumes that vector operands are the same length as the mask
2151 /// (a length-changing shuffle can never be equivalent to a vector select).
2152 static bool isSelectMask(ArrayRef<int> Mask);
2153 static bool isSelectMask(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-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2154, __PRETTY_FUNCTION__))
;
2155 SmallVector<int, 16> MaskAsInts;
2156 getShuffleMask(Mask, MaskAsInts);
2157 return isSelectMask(MaskAsInts);
2158 }
2159
2160 /// Return true if this shuffle chooses elements from its source vectors
2161 /// without lane crossings and all operands have the same number of elements.
2162 /// In other words, this shuffle is equivalent to a vector select with a
2163 /// constant condition operand.
2164 /// Example: shufflevector <4 x n> A, <4 x n> B, <undef,1,6,3>
2165 /// This returns false if the mask does not choose from both input vectors.
2166 /// In that case, the shuffle is better classified as an identity shuffle.
2167 /// TODO: Optionally allow length-changing shuffles.
2168 bool isSelect() const {
2169 return !changesLength() && isSelectMask(ShuffleMask);
2170 }
2171
2172 /// Return true if this shuffle mask swaps the order of elements from exactly
2173 /// one source vector.
2174 /// Example: <7,6,undef,4>
2175 /// This assumes that vector operands are the same length as the mask.
2176 static bool isReverseMask(ArrayRef<int> Mask);
2177 static bool isReverseMask(const Constant *Mask) {
2178 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((Mask->getType()->isVectorTy() && "Shuffle needs vector constant."
) ? static_cast<void> (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2178, __PRETTY_FUNCTION__))
;
2179 SmallVector<int, 16> MaskAsInts;
2180 getShuffleMask(Mask, MaskAsInts);
2181 return isReverseMask(MaskAsInts);
2182 }
2183
2184 /// Return true if this shuffle swaps the order of elements from exactly
2185 /// one source vector.
2186 /// Example: shufflevector <4 x n> A, <4 x n> B, <3,undef,1,undef>
2187 /// TODO: Optionally allow length-changing shuffles.
2188 bool isReverse() const {
2189 return !changesLength() && isReverseMask(ShuffleMask);
2190 }
2191
2192 /// Return true if this shuffle mask chooses all elements with the same value
2193 /// as the first element of exactly one source vector.
2194 /// Example: <4,undef,undef,4>
2195 /// This assumes that vector operands are the same length as the mask.
2196 static bool isZeroEltSplatMask(ArrayRef<int> Mask);
2197 static bool isZeroEltSplatMask(const Constant *Mask) {
2198 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((Mask->getType()->isVectorTy() && "Shuffle needs vector constant."
) ? static_cast<void> (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2198, __PRETTY_FUNCTION__))
;
2199 SmallVector<int, 16> MaskAsInts;
2200 getShuffleMask(Mask, MaskAsInts);
2201 return isZeroEltSplatMask(MaskAsInts);
2202 }
2203
2204 /// Return true if all elements of this shuffle are the same value as the
2205 /// first element of exactly one source vector without changing the length
2206 /// of that vector.
2207 /// Example: shufflevector <4 x n> A, <4 x n> B, <undef,0,undef,0>
2208 /// TODO: Optionally allow length-changing shuffles.
2209 /// TODO: Optionally allow splats from other elements.
2210 bool isZeroEltSplat() const {
2211 return !changesLength() && isZeroEltSplatMask(ShuffleMask);
2212 }
2213
2214 /// Return true if this shuffle mask is a transpose mask.
2215 /// Transpose vector masks transpose a 2xn matrix. They read corresponding
2216 /// even- or odd-numbered vector elements from two n-dimensional source
2217 /// vectors and write each result into consecutive elements of an
2218 /// n-dimensional destination vector. Two shuffles are necessary to complete
2219 /// the transpose, one for the even elements and another for the odd elements.
2220 /// This description closely follows how the TRN1 and TRN2 AArch64
2221 /// instructions operate.
2222 ///
2223 /// For example, a simple 2x2 matrix can be transposed with:
2224 ///
2225 /// ; Original matrix
2226 /// m0 = < a, b >
2227 /// m1 = < c, d >
2228 ///
2229 /// ; Transposed matrix
2230 /// t0 = < a, c > = shufflevector m0, m1, < 0, 2 >
2231 /// t1 = < b, d > = shufflevector m0, m1, < 1, 3 >
2232 ///
2233 /// For matrices having greater than n columns, the resulting nx2 transposed
2234 /// matrix is stored in two result vectors such that one vector contains
2235 /// interleaved elements from all the even-numbered rows and the other vector
2236 /// contains interleaved elements from all the odd-numbered rows. For example,
2237 /// a 2x4 matrix can be transposed with:
2238 ///
2239 /// ; Original matrix
2240 /// m0 = < a, b, c, d >
2241 /// m1 = < e, f, g, h >
2242 ///
2243 /// ; Transposed matrix
2244 /// t0 = < a, e, c, g > = shufflevector m0, m1 < 0, 4, 2, 6 >
2245 /// t1 = < b, f, d, h > = shufflevector m0, m1 < 1, 5, 3, 7 >
2246 static bool isTransposeMask(ArrayRef<int> Mask);
2247 static bool isTransposeMask(const Constant *Mask) {
2248 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((Mask->getType()->isVectorTy() && "Shuffle needs vector constant."
) ? static_cast<void> (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2248, __PRETTY_FUNCTION__))
;
2249 SmallVector<int, 16> MaskAsInts;
2250 getShuffleMask(Mask, MaskAsInts);
2251 return isTransposeMask(MaskAsInts);
2252 }
2253
2254 /// Return true if this shuffle transposes the elements of its inputs without
2255 /// changing the length of the vectors. This operation may also be known as a
2256 /// merge or interleave. See the description for isTransposeMask() for the
2257 /// exact specification.
2258 /// Example: shufflevector <4 x n> A, <4 x n> B, <0,4,2,6>
2259 bool isTranspose() const {
2260 return !changesLength() && isTransposeMask(ShuffleMask);
2261 }
2262
2263 /// Return true if this shuffle mask is an extract subvector mask.
2264 /// A valid extract subvector mask returns a smaller vector from a single
2265 /// source operand. The base extraction index is returned as well.
2266 static bool isExtractSubvectorMask(ArrayRef<int> Mask, int NumSrcElts,
2267 int &Index);
2268 static bool isExtractSubvectorMask(const Constant *Mask, int NumSrcElts,
2269 int &Index) {
2270 assert(Mask->getType()->isVectorTy() && "Shuffle needs vector constant.")((Mask->getType()->isVectorTy() && "Shuffle needs vector constant."
) ? static_cast<void> (0) : __assert_fail ("Mask->getType()->isVectorTy() && \"Shuffle needs vector constant.\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2270, __PRETTY_FUNCTION__))
;
2271 SmallVector<int, 16> MaskAsInts;
2272 getShuffleMask(Mask, MaskAsInts);
2273 return isExtractSubvectorMask(MaskAsInts, NumSrcElts, Index);
2274 }
2275
2276 /// Return true if this shuffle mask is an extract subvector mask.
2277 bool isExtractSubvectorMask(int &Index) const {
2278 int NumSrcElts =
2279 cast<FixedVectorType>(Op<0>()->getType())->getNumElements();
2280 return isExtractSubvectorMask(ShuffleMask, NumSrcElts, Index);
2281 }
2282
2283 /// Change values in a shuffle permute mask assuming the two vector operands
2284 /// of length InVecNumElts have swapped position.
2285 static void commuteShuffleMask(MutableArrayRef<int> Mask,
2286 unsigned InVecNumElts) {
2287 for (int &Idx : Mask) {
2288 if (Idx == -1)
2289 continue;
2290 Idx = Idx < (int)InVecNumElts ? Idx + InVecNumElts : Idx - InVecNumElts;
2291 assert(Idx >= 0 && Idx < (int)InVecNumElts * 2 &&((Idx >= 0 && Idx < (int)InVecNumElts * 2 &&
"shufflevector mask index out of range") ? static_cast<void
> (0) : __assert_fail ("Idx >= 0 && Idx < (int)InVecNumElts * 2 && \"shufflevector mask index out of range\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2292, __PRETTY_FUNCTION__))
2292 "shufflevector mask index out of range")((Idx >= 0 && Idx < (int)InVecNumElts * 2 &&
"shufflevector mask index out of range") ? static_cast<void
> (0) : __assert_fail ("Idx >= 0 && Idx < (int)InVecNumElts * 2 && \"shufflevector mask index out of range\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2292, __PRETTY_FUNCTION__))
;
2293 }
2294 }
2295
2296 // Methods for support type inquiry through isa, cast, and dyn_cast:
2297 static bool classof(const Instruction *I) {
2298 return I->getOpcode() == Instruction::ShuffleVector;
2299 }
2300 static bool classof(const Value *V) {
2301 return isa<Instruction>(V) && classof(cast<Instruction>(V));
2302 }
2303};
2304
2305template <>
2306struct OperandTraits<ShuffleVectorInst>
2307 : public FixedNumOperandTraits<ShuffleVectorInst, 2> {};
2308
2309DEFINE_TRANSPARENT_OPERAND_ACCESSORS(ShuffleVectorInst, Value)ShuffleVectorInst::op_iterator ShuffleVectorInst::op_begin() {
return OperandTraits<ShuffleVectorInst>::op_begin(this
); } ShuffleVectorInst::const_op_iterator ShuffleVectorInst::
op_begin() const { return OperandTraits<ShuffleVectorInst>
::op_begin(const_cast<ShuffleVectorInst*>(this)); } ShuffleVectorInst
::op_iterator ShuffleVectorInst::op_end() { return OperandTraits
<ShuffleVectorInst>::op_end(this); } ShuffleVectorInst::
const_op_iterator ShuffleVectorInst::op_end() const { return OperandTraits
<ShuffleVectorInst>::op_end(const_cast<ShuffleVectorInst
*>(this)); } Value *ShuffleVectorInst::getOperand(unsigned
i_nocapture) const { ((i_nocapture < OperandTraits<ShuffleVectorInst
>::operands(this) && "getOperand() out of range!")
? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<ShuffleVectorInst>::operands(this) && \"getOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2309, __PRETTY_FUNCTION__)); return cast_or_null<Value>
( OperandTraits<ShuffleVectorInst>::op_begin(const_cast
<ShuffleVectorInst*>(this))[i_nocapture].get()); } void
ShuffleVectorInst::setOperand(unsigned i_nocapture, Value *Val_nocapture
) { ((i_nocapture < OperandTraits<ShuffleVectorInst>
::operands(this) && "setOperand() out of range!") ? static_cast
<void> (0) : __assert_fail ("i_nocapture < OperandTraits<ShuffleVectorInst>::operands(this) && \"setOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2309, __PRETTY_FUNCTION__)); OperandTraits<ShuffleVectorInst
>::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned
ShuffleVectorInst::getNumOperands() const { return OperandTraits
<ShuffleVectorInst>::operands(this); } template <int
Idx_nocapture> Use &ShuffleVectorInst::Op() { return this
->OpFrom<Idx_nocapture>(this); } template <int Idx_nocapture
> const Use &ShuffleVectorInst::Op() const { return this
->OpFrom<Idx_nocapture>(this); }
2310
2311//===----------------------------------------------------------------------===//
2312// ExtractValueInst Class
2313//===----------------------------------------------------------------------===//
2314
2315/// This instruction extracts a struct member or array
2316/// element value from an aggregate value.
2317///
2318class ExtractValueInst : public UnaryInstruction {
2319 SmallVector<unsigned, 4> Indices;
2320
2321 ExtractValueInst(const ExtractValueInst &EVI);
2322
2323 /// Constructors - Create a extractvalue instruction with a base aggregate
2324 /// value and a list of indices. The first ctor can optionally insert before
2325 /// an existing instruction, the second appends the new instruction to the
2326 /// specified BasicBlock.
2327 inline ExtractValueInst(Value *Agg,
2328 ArrayRef<unsigned> Idxs,
2329 const Twine &NameStr,
2330 Instruction *InsertBefore);
2331 inline ExtractValueInst(Value *Agg,
2332 ArrayRef<unsigned> Idxs,
2333 const Twine &NameStr, BasicBlock *InsertAtEnd);
2334
2335 void init(ArrayRef<unsigned> Idxs, const Twine &NameStr);
2336
2337protected:
2338 // Note: Instruction needs to be a friend here to call cloneImpl.
2339 friend class Instruction;
2340
2341 ExtractValueInst *cloneImpl() const;
2342
2343public:
2344 static ExtractValueInst *Create(Value *Agg,
2345 ArrayRef<unsigned> Idxs,
2346 const Twine &NameStr = "",
2347 Instruction *InsertBefore = nullptr) {
2348 return new
2349 ExtractValueInst(Agg, Idxs, NameStr, InsertBefore);
2350 }
2351
2352 static ExtractValueInst *Create(Value *Agg,
2353 ArrayRef<unsigned> Idxs,
2354 const Twine &NameStr,
2355 BasicBlock *InsertAtEnd) {
2356 return new ExtractValueInst(Agg, Idxs, NameStr, InsertAtEnd);
2357 }
2358
2359 /// Returns the type of the element that would be extracted
2360 /// with an extractvalue instruction with the specified parameters.
2361 ///
2362 /// Null is returned if the indices are invalid for the specified type.
2363 static Type *getIndexedType(Type *Agg, ArrayRef<unsigned> Idxs);
2364
2365 using idx_iterator = const unsigned*;
2366
2367 inline idx_iterator idx_begin() const { return Indices.begin(); }
2368 inline idx_iterator idx_end() const { return Indices.end(); }
2369 inline iterator_range<idx_iterator> indices() const {
2370 return make_range(idx_begin(), idx_end());
2371 }
2372
2373 Value *getAggregateOperand() {
2374 return getOperand(0);
2375 }
2376 const Value *getAggregateOperand() const {
2377 return getOperand(0);
2378 }
2379 static unsigned getAggregateOperandIndex() {
2380 return 0U; // get index for modifying correct operand
2381 }
2382
2383 ArrayRef<unsigned> getIndices() const {
2384 return Indices;
2385 }
2386
2387 unsigned getNumIndices() const {
2388 return (unsigned)Indices.size();
2389 }
2390
2391 bool hasIndices() const {
2392 return true;
2393 }
2394
2395 // Methods for support type inquiry through isa, cast, and dyn_cast:
2396 static bool classof(const Instruction *I) {
2397 return I->getOpcode() == Instruction::ExtractValue;
2398 }
2399 static bool classof(const Value *V) {
2400 return isa<Instruction>(V) && classof(cast<Instruction>(V));
2401 }
2402};
2403
2404ExtractValueInst::ExtractValueInst(Value *Agg,
2405 ArrayRef<unsigned> Idxs,
2406 const Twine &NameStr,
2407 Instruction *InsertBefore)
2408 : UnaryInstruction(checkGEPType(getIndexedType(Agg->getType(), Idxs)),
2409 ExtractValue, Agg, InsertBefore) {
2410 init(Idxs, NameStr);
2411}
2412
2413ExtractValueInst::ExtractValueInst(Value *Agg,
2414 ArrayRef<unsigned> Idxs,
2415 const Twine &NameStr,
2416 BasicBlock *InsertAtEnd)
2417 : UnaryInstruction(checkGEPType(getIndexedType(Agg->getType(), Idxs)),
2418 ExtractValue, Agg, InsertAtEnd) {
2419 init(Idxs, NameStr);
2420}
2421
2422//===----------------------------------------------------------------------===//
2423// InsertValueInst Class
2424//===----------------------------------------------------------------------===//
2425
2426/// This instruction inserts a struct field of array element
2427/// value into an aggregate value.
2428///
2429class InsertValueInst : public Instruction {
2430 SmallVector<unsigned, 4> Indices;
2431
2432 InsertValueInst(const InsertValueInst &IVI);
2433
2434 /// Constructors - Create a insertvalue instruction with a base aggregate
2435 /// value, a value to insert, and a list of indices. The first ctor can
2436 /// optionally insert before an existing instruction, the second appends
2437 /// the new instruction to the specified BasicBlock.
2438 inline InsertValueInst(Value *Agg, Value *Val,
2439 ArrayRef<unsigned> Idxs,
2440 const Twine &NameStr,
2441 Instruction *InsertBefore);
2442 inline InsertValueInst(Value *Agg, Value *Val,
2443 ArrayRef<unsigned> Idxs,
2444 const Twine &NameStr, BasicBlock *InsertAtEnd);
2445
2446 /// Constructors - These two constructors are convenience methods because one
2447 /// and two index insertvalue instructions are so common.
2448 InsertValueInst(Value *Agg, Value *Val, unsigned Idx,
2449 const Twine &NameStr = "",
2450 Instruction *InsertBefore = nullptr);
2451 InsertValueInst(Value *Agg, Value *Val, unsigned Idx, const Twine &NameStr,
2452 BasicBlock *InsertAtEnd);
2453
2454 void init(Value *Agg, Value *Val, ArrayRef<unsigned> Idxs,
2455 const Twine &NameStr);
2456
2457protected:
2458 // Note: Instruction needs to be a friend here to call cloneImpl.
2459 friend class Instruction;
2460
2461 InsertValueInst *cloneImpl() const;
2462
2463public:
2464 // allocate space for exactly two operands
2465 void *operator new(size_t s) {
2466 return User::operator new(s, 2);
2467 }
2468
2469 static InsertValueInst *Create(Value *Agg, Value *Val,
2470 ArrayRef<unsigned> Idxs,
2471 const Twine &NameStr = "",
2472 Instruction *InsertBefore = nullptr) {
2473 return new InsertValueInst(Agg, Val, Idxs, NameStr, InsertBefore);
2474 }
2475
2476 static InsertValueInst *Create(Value *Agg, Value *Val,
2477 ArrayRef<unsigned> Idxs,
2478 const Twine &NameStr,
2479 BasicBlock *InsertAtEnd) {
2480 return new InsertValueInst(Agg, Val, Idxs, NameStr, InsertAtEnd);
2481 }
2482
2483 /// Transparently provide more efficient getOperand methods.
2484 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
;
2485
2486 using idx_iterator = const unsigned*;
2487
2488 inline idx_iterator idx_begin() const { return Indices.begin(); }
2489 inline idx_iterator idx_end() const { return Indices.end(); }
2490 inline iterator_range<idx_iterator> indices() const {
2491 return make_range(idx_begin(), idx_end());
2492 }
2493
2494 Value *getAggregateOperand() {
2495 return getOperand(0);
2496 }
2497 const Value *getAggregateOperand() const {
2498 return getOperand(0);
2499 }
2500 static unsigned getAggregateOperandIndex() {
2501 return 0U; // get index for modifying correct operand
2502 }
2503
2504 Value *getInsertedValueOperand() {
2505 return getOperand(1);
2506 }
2507 const Value *getInsertedValueOperand() const {
2508 return getOperand(1);
2509 }
2510 static unsigned getInsertedValueOperandIndex() {
2511 return 1U; // get index for modifying correct operand
2512 }
2513
2514 ArrayRef<unsigned> getIndices() const {
2515 return Indices;
2516 }
2517
2518 unsigned getNumIndices() const {
2519 return (unsigned)Indices.size();
2520 }
2521
2522 bool hasIndices() const {
2523 return true;
2524 }
2525
2526 // Methods for support type inquiry through isa, cast, and dyn_cast:
2527 static bool classof(const Instruction *I) {
2528 return I->getOpcode() == Instruction::InsertValue;
2529 }
2530 static bool classof(const Value *V) {
2531 return isa<Instruction>(V) && classof(cast<Instruction>(V));
2532 }
2533};
2534
2535template <>
2536struct OperandTraits<InsertValueInst> :
2537 public FixedNumOperandTraits<InsertValueInst, 2> {
2538};
2539
2540InsertValueInst::InsertValueInst(Value *Agg,
2541 Value *Val,
2542 ArrayRef<unsigned> Idxs,
2543 const Twine &NameStr,
2544 Instruction *InsertBefore)
2545 : Instruction(Agg->getType(), InsertValue,
2546 OperandTraits<InsertValueInst>::op_begin(this),
2547 2, InsertBefore) {
2548 init(Agg, Val, Idxs, NameStr);
2549}
2550
2551InsertValueInst::InsertValueInst(Value *Agg,
2552 Value *Val,
2553 ArrayRef<unsigned> Idxs,
2554 const Twine &NameStr,
2555 BasicBlock *InsertAtEnd)
2556 : Instruction(Agg->getType(), InsertValue,
2557 OperandTraits<InsertValueInst>::op_begin(this),
2558 2, InsertAtEnd) {
2559 init(Agg, Val, Idxs, NameStr);
2560}
2561
2562DEFINE_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-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2562, __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-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2562, __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); }
2563
2564//===----------------------------------------------------------------------===//
2565// PHINode Class
2566//===----------------------------------------------------------------------===//
2567
2568// PHINode - The PHINode class is used to represent the magical mystical PHI
2569// node, that can not exist in nature, but can be synthesized in a computer
2570// scientist's overactive imagination.
2571//
2572class PHINode : public Instruction {
2573 /// The number of operands actually allocated. NumOperands is
2574 /// the number actually in use.
2575 unsigned ReservedSpace;
2576
2577 PHINode(const PHINode &PN);
2578
2579 explicit PHINode(Type *Ty, unsigned NumReservedValues,
2580 const Twine &NameStr = "",
2581 Instruction *InsertBefore = nullptr)
2582 : Instruction(Ty, Instruction::PHI, nullptr, 0, InsertBefore),
2583 ReservedSpace(NumReservedValues) {
2584 setName(NameStr);
2585 allocHungoffUses(ReservedSpace);
2586 }
2587
2588 PHINode(Type *Ty, unsigned NumReservedValues, const Twine &NameStr,
2589 BasicBlock *InsertAtEnd)
2590 : Instruction(Ty, Instruction::PHI, nullptr, 0, InsertAtEnd),
2591 ReservedSpace(NumReservedValues) {
2592 setName(NameStr);
2593 allocHungoffUses(ReservedSpace);
2594 }
2595
2596protected:
2597 // Note: Instruction needs to be a friend here to call cloneImpl.
2598 friend class Instruction;
2599
2600 PHINode *cloneImpl() const;
2601
2602 // allocHungoffUses - this is more complicated than the generic
2603 // User::allocHungoffUses, because we have to allocate Uses for the incoming
2604 // values and pointers to the incoming blocks, all in one allocation.
2605 void allocHungoffUses(unsigned N) {
2606 User::allocHungoffUses(N, /* IsPhi */ true);
2607 }
2608
2609public:
2610 /// Constructors - NumReservedValues is a hint for the number of incoming
2611 /// edges that this phi node will have (use 0 if you really have no idea).
2612 static PHINode *Create(Type *Ty, unsigned NumReservedValues,
2613 const Twine &NameStr = "",
2614 Instruction *InsertBefore = nullptr) {
2615 return new PHINode(Ty, NumReservedValues, NameStr, InsertBefore);
2616 }
2617
2618 static PHINode *Create(Type *Ty, unsigned NumReservedValues,
2619 const Twine &NameStr, BasicBlock *InsertAtEnd) {
2620 return new PHINode(Ty, NumReservedValues, NameStr, InsertAtEnd);
2621 }
2622
2623 /// Provide fast operand accessors
2624 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
;
2625
2626 // Block iterator interface. This provides access to the list of incoming
2627 // basic blocks, which parallels the list of incoming values.
2628
2629 using block_iterator = BasicBlock **;
2630 using const_block_iterator = BasicBlock * const *;
2631
2632 block_iterator block_begin() {
2633 return reinterpret_cast<block_iterator>(op_begin() + ReservedSpace);
2634 }
2635
2636 const_block_iterator block_begin() const {
2637 return reinterpret_cast<const_block_iterator>(op_begin() + ReservedSpace);
2638 }
2639
2640 block_iterator block_end() {
2641 return block_begin() + getNumOperands();
2642 }
2643
2644 const_block_iterator block_end() const {
2645 return block_begin() + getNumOperands();
2646 }
2647
2648 iterator_range<block_iterator> blocks() {
2649 return make_range(block_begin(), block_end());
2650 }
2651
2652 iterator_range<const_block_iterator> blocks() const {
2653 return make_range(block_begin(), block_end());
2654 }
2655
2656 op_range incoming_values() { return operands(); }
2657
2658 const_op_range incoming_values() const { return operands(); }
2659
2660 /// Return the number of incoming edges
2661 ///
2662 unsigned getNumIncomingValues() const { return getNumOperands(); }
2663
2664 /// Return incoming value number x
2665 ///
2666 Value *getIncomingValue(unsigned i) const {
2667 return getOperand(i);
2668 }
2669 void setIncomingValue(unsigned i, Value *V) {
2670 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-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2670, __PRETTY_FUNCTION__))
;
2671 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-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2672, __PRETTY_FUNCTION__))
2672 "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-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2672, __PRETTY_FUNCTION__))
;
2673 setOperand(i, V);
2674 }
2675
2676 static unsigned getOperandNumForIncomingValue(unsigned i) {
2677 return i;
2678 }
2679
2680 static unsigned getIncomingValueNumForOperand(unsigned i) {
2681 return i;
2682 }
2683
2684 /// Return incoming basic block number @p i.
2685 ///
2686 BasicBlock *getIncomingBlock(unsigned i) const {
2687 return block_begin()[i];
2688 }
2689
2690 /// Return incoming basic block corresponding
2691 /// to an operand of the PHI.
2692 ///
2693 BasicBlock *getIncomingBlock(const Use &U) const {
2694 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-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2694, __PRETTY_FUNCTION__))
;
2695 return getIncomingBlock(unsigned(&U - op_begin()));
2696 }
2697
2698 /// Return incoming basic block corresponding
2699 /// to value use iterator.
2700 ///
2701 BasicBlock *getIncomingBlock(Value::const_user_iterator I) const {
2702 return getIncomingBlock(I.getUse());
2703 }
2704
2705 void setIncomingBlock(unsigned i, BasicBlock *BB) {
2706 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-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2706, __PRETTY_FUNCTION__))
;
2707 block_begin()[i] = BB;
2708 }
2709
2710 /// Replace every incoming basic block \p Old to basic block \p New.
2711 void replaceIncomingBlockWith(const BasicBlock *Old, BasicBlock *New) {
2712 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-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2712, __PRETTY_FUNCTION__))
;
2713 for (unsigned Op = 0, NumOps = getNumOperands(); Op != NumOps; ++Op)
2714 if (getIncomingBlock(Op) == Old)
2715 setIncomingBlock(Op, New);
2716 }
2717
2718 /// Add an incoming value to the end of the PHI list
2719 ///
2720 void addIncoming(Value *V, BasicBlock *BB) {
2721 if (getNumOperands() == ReservedSpace)
2722 growOperands(); // Get more space!
2723 // Initialize some new operands.
2724 setNumHungOffUseOperands(getNumOperands() + 1);
2725 setIncomingValue(getNumOperands() - 1, V);
2726 setIncomingBlock(getNumOperands() - 1, BB);
2727 }
2728
2729 /// Remove an incoming value. This is useful if a
2730 /// predecessor basic block is deleted. The value removed is returned.
2731 ///
2732 /// If the last incoming value for a PHI node is removed (and DeletePHIIfEmpty
2733 /// is true), the PHI node is destroyed and any uses of it are replaced with
2734 /// dummy values. The only time there should be zero incoming values to a PHI
2735 /// node is when the block is dead, so this strategy is sound.
2736 ///
2737 Value *removeIncomingValue(unsigned Idx, bool DeletePHIIfEmpty = true);
2738
2739 Value *removeIncomingValue(const BasicBlock *BB, bool DeletePHIIfEmpty=true) {
2740 int Idx = getBasicBlockIndex(BB);
2741 assert(Idx >= 0 && "Invalid basic block argument to remove!")((Idx >= 0 && "Invalid basic block argument to remove!"
) ? static_cast<void> (0) : __assert_fail ("Idx >= 0 && \"Invalid basic block argument to remove!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2741, __PRETTY_FUNCTION__))
;
2742 return removeIncomingValue(Idx, DeletePHIIfEmpty);
2743 }
2744
2745 /// Return the first index of the specified basic
2746 /// block in the value list for this PHI. Returns -1 if no instance.
2747 ///
2748 int getBasicBlockIndex(const BasicBlock *BB) const {
2749 for (unsigned i = 0, e = getNumOperands(); i != e; ++i)
2750 if (block_begin()[i] == BB)
2751 return i;
2752 return -1;
2753 }
2754
2755 Value *getIncomingValueForBlock(const BasicBlock *BB) const {
2756 int Idx = getBasicBlockIndex(BB);
2757 assert(Idx >= 0 && "Invalid basic block argument!")((Idx >= 0 && "Invalid basic block argument!") ? static_cast
<void> (0) : __assert_fail ("Idx >= 0 && \"Invalid basic block argument!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2757, __PRETTY_FUNCTION__))
;
2758 return getIncomingValue(Idx);
2759 }
2760
2761 /// Set every incoming value(s) for block \p BB to \p V.
2762 void setIncomingValueForBlock(const BasicBlock *BB, Value *V) {
2763 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-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2763, __PRETTY_FUNCTION__))
;
2764 bool Found = false;
2765 for (unsigned Op = 0, NumOps = getNumOperands(); Op != NumOps; ++Op)
2766 if (getIncomingBlock(Op) == BB) {
2767 Found = true;
2768 setIncomingValue(Op, V);
2769 }
2770 (void)Found;
2771 assert(Found && "Invalid basic block argument to set!")((Found && "Invalid basic block argument to set!") ? static_cast
<void> (0) : __assert_fail ("Found && \"Invalid basic block argument to set!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2771, __PRETTY_FUNCTION__))
;
2772 }
2773
2774 /// If the specified PHI node always merges together the
2775 /// same value, return the value, otherwise return null.
2776 Value *hasConstantValue() const;
2777
2778 /// Whether the specified PHI node always merges
2779 /// together the same value, assuming undefs are equal to a unique
2780 /// non-undef value.
2781 bool hasConstantOrUndefValue() const;
2782
2783 /// If the PHI node is complete which means all of its parent's predecessors
2784 /// have incoming value in this PHI, return true, otherwise return false.
2785 bool isComplete() const {
2786 return llvm::all_of(predecessors(getParent()),
2787 [this](const BasicBlock *Pred) {
2788 return getBasicBlockIndex(Pred) >= 0;
2789 });
2790 }
2791
2792 /// Methods for support type inquiry through isa, cast, and dyn_cast:
2793 static bool classof(const Instruction *I) {
2794 return I->getOpcode() == Instruction::PHI;
2795 }
2796 static bool classof(const Value *V) {
2797 return isa<Instruction>(V) && classof(cast<Instruction>(V));
2798 }
2799
2800private:
2801 void growOperands();
2802};
2803
2804template <>
2805struct OperandTraits<PHINode> : public HungoffOperandTraits<2> {
2806};
2807
2808DEFINE_TRANSPARENT_OPERAND_ACCESSORS(PHINode, Value)PHINode::op_iterator PHINode::op_begin() { return OperandTraits
<PHINode>::op_begin(this); } PHINode::const_op_iterator
PHINode::op_begin() const { return OperandTraits<PHINode>
::op_begin(const_cast<PHINode*>(this)); } PHINode::op_iterator
PHINode::op_end() { return OperandTraits<PHINode>::op_end
(this); } PHINode::const_op_iterator PHINode::op_end() const {
return OperandTraits<PHINode>::op_end(const_cast<PHINode
*>(this)); } Value *PHINode::getOperand(unsigned i_nocapture
) const { ((i_nocapture < OperandTraits<PHINode>::operands
(this) && "getOperand() out of range!") ? static_cast
<void> (0) : __assert_fail ("i_nocapture < OperandTraits<PHINode>::operands(this) && \"getOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2808, __PRETTY_FUNCTION__)); return cast_or_null<Value>
( OperandTraits<PHINode>::op_begin(const_cast<PHINode
*>(this))[i_nocapture].get()); } void PHINode::setOperand(
unsigned i_nocapture, Value *Val_nocapture) { ((i_nocapture <
OperandTraits<PHINode>::operands(this) && "setOperand() out of range!"
) ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<PHINode>::operands(this) && \"setOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2808, __PRETTY_FUNCTION__)); OperandTraits<PHINode>::
op_begin(this)[i_nocapture] = Val_nocapture; } unsigned PHINode
::getNumOperands() const { return OperandTraits<PHINode>
::operands(this); } template <int Idx_nocapture> Use &
PHINode::Op() { return this->OpFrom<Idx_nocapture>(this
); } template <int Idx_nocapture> const Use &PHINode
::Op() const { return this->OpFrom<Idx_nocapture>(this
); }
2809
2810//===----------------------------------------------------------------------===//
2811// LandingPadInst Class
2812//===----------------------------------------------------------------------===//
2813
2814//===---------------------------------------------------------------------------
2815/// The landingpad instruction holds all of the information
2816/// necessary to generate correct exception handling. The landingpad instruction
2817/// cannot be moved from the top of a landing pad block, which itself is
2818/// accessible only from the 'unwind' edge of an invoke. This uses the
2819/// SubclassData field in Value to store whether or not the landingpad is a
2820/// cleanup.
2821///
2822class LandingPadInst : public Instruction {
2823 using CleanupField = BoolBitfieldElementT<0>;
2824
2825 /// The number of operands actually allocated. NumOperands is
2826 /// the number actually in use.
2827 unsigned ReservedSpace;
2828
2829 LandingPadInst(const LandingPadInst &LP);
2830
2831public:
2832 enum ClauseType { Catch, Filter };
2833
2834private:
2835 explicit LandingPadInst(Type *RetTy, unsigned NumReservedValues,
2836 const Twine &NameStr, Instruction *InsertBefore);
2837 explicit LandingPadInst(Type *RetTy, unsigned NumReservedValues,
2838 const Twine &NameStr, BasicBlock *InsertAtEnd);
2839
2840 // Allocate space for exactly zero operands.
2841 void *operator new(size_t s) {
2842 return User::operator new(s);
2843 }
2844
2845 void growOperands(unsigned Size);
2846 void init(unsigned NumReservedValues, const Twine &NameStr);
2847
2848protected:
2849 // Note: Instruction needs to be a friend here to call cloneImpl.
2850 friend class Instruction;
2851
2852 LandingPadInst *cloneImpl() const;
2853
2854public:
2855 /// Constructors - NumReservedClauses is a hint for the number of incoming
2856 /// clauses that this landingpad will have (use 0 if you really have no idea).
2857 static LandingPadInst *Create(Type *RetTy, unsigned NumReservedClauses,
2858 const Twine &NameStr = "",
2859 Instruction *InsertBefore = nullptr);
2860 static LandingPadInst *Create(Type *RetTy, unsigned NumReservedClauses,
2861 const Twine &NameStr, BasicBlock *InsertAtEnd);
2862
2863 /// Provide fast operand accessors
2864 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
;
2865
2866 /// Return 'true' if this landingpad instruction is a
2867 /// cleanup. I.e., it should be run when unwinding even if its landing pad
2868 /// doesn't catch the exception.
2869 bool isCleanup() const { return getSubclassData<CleanupField>(); }
2870
2871 /// Indicate that this landingpad instruction is a cleanup.
2872 void setCleanup(bool V) { setSubclassData<CleanupField>(V); }
2873
2874 /// Add a catch or filter clause to the landing pad.
2875 void addClause(Constant *ClauseVal);
2876
2877 /// Get the value of the clause at index Idx. Use isCatch/isFilter to
2878 /// determine what type of clause this is.
2879 Constant *getClause(unsigned Idx) const {
2880 return cast<Constant>(getOperandList()[Idx]);
2881 }
2882
2883 /// Return 'true' if the clause and index Idx is a catch clause.
2884 bool isCatch(unsigned Idx) const {
2885 return !isa<ArrayType>(getOperandList()[Idx]->getType());
2886 }
2887
2888 /// Return 'true' if the clause and index Idx is a filter clause.
2889 bool isFilter(unsigned Idx) const {
2890 return isa<ArrayType>(getOperandList()[Idx]->getType());
2891 }
2892
2893 /// Get the number of clauses for this landing pad.
2894 unsigned getNumClauses() const { return getNumOperands(); }
2895
2896 /// Grow the size of the operand list to accommodate the new
2897 /// number of clauses.
2898 void reserveClauses(unsigned Size) { growOperands(Size); }
2899
2900 // Methods for support type inquiry through isa, cast, and dyn_cast:
2901 static bool classof(const Instruction *I) {
2902 return I->getOpcode() == Instruction::LandingPad;
2903 }
2904 static bool classof(const Value *V) {
2905 return isa<Instruction>(V) && classof(cast<Instruction>(V));
2906 }
2907};
2908
2909template <>
2910struct OperandTraits<LandingPadInst> : public HungoffOperandTraits<1> {
2911};
2912
2913DEFINE_TRANSPARENT_OPERAND_ACCESSORS(LandingPadInst, Value)LandingPadInst::op_iterator LandingPadInst::op_begin() { return
OperandTraits<LandingPadInst>::op_begin(this); } LandingPadInst
::const_op_iterator LandingPadInst::op_begin() const { return
OperandTraits<LandingPadInst>::op_begin(const_cast<
LandingPadInst*>(this)); } LandingPadInst::op_iterator LandingPadInst
::op_end() { return OperandTraits<LandingPadInst>::op_end
(this); } LandingPadInst::const_op_iterator LandingPadInst::op_end
() const { return OperandTraits<LandingPadInst>::op_end
(const_cast<LandingPadInst*>(this)); } Value *LandingPadInst
::getOperand(unsigned i_nocapture) const { ((i_nocapture <
OperandTraits<LandingPadInst>::operands(this) &&
"getOperand() out of range!") ? static_cast<void> (0) :
__assert_fail ("i_nocapture < OperandTraits<LandingPadInst>::operands(this) && \"getOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2913, __PRETTY_FUNCTION__)); return cast_or_null<Value>
( OperandTraits<LandingPadInst>::op_begin(const_cast<
LandingPadInst*>(this))[i_nocapture].get()); } void LandingPadInst
::setOperand(unsigned i_nocapture, Value *Val_nocapture) { ((
i_nocapture < OperandTraits<LandingPadInst>::operands
(this) && "setOperand() out of range!") ? static_cast
<void> (0) : __assert_fail ("i_nocapture < OperandTraits<LandingPadInst>::operands(this) && \"setOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2913, __PRETTY_FUNCTION__)); OperandTraits<LandingPadInst
>::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned
LandingPadInst::getNumOperands() const { return OperandTraits
<LandingPadInst>::operands(this); } template <int Idx_nocapture
> Use &LandingPadInst::Op() { return this->OpFrom<
Idx_nocapture>(this); } template <int Idx_nocapture>
const Use &LandingPadInst::Op() const { return this->
OpFrom<Idx_nocapture>(this); }
2914
2915//===----------------------------------------------------------------------===//
2916// ReturnInst Class
2917//===----------------------------------------------------------------------===//
2918
2919//===---------------------------------------------------------------------------
2920/// Return a value (possibly void), from a function. Execution
2921/// does not continue in this function any longer.
2922///
2923class ReturnInst : public Instruction {
2924 ReturnInst(const ReturnInst &RI);
2925
2926private:
2927 // ReturnInst constructors:
2928 // ReturnInst() - 'ret void' instruction
2929 // ReturnInst( null) - 'ret void' instruction
2930 // ReturnInst(Value* X) - 'ret X' instruction
2931 // ReturnInst( null, Inst *I) - 'ret void' instruction, insert before I
2932 // ReturnInst(Value* X, Inst *I) - 'ret X' instruction, insert before I
2933 // ReturnInst( null, BB *B) - 'ret void' instruction, insert @ end of B
2934 // ReturnInst(Value* X, BB *B) - 'ret X' instruction, insert @ end of B
2935 //
2936 // NOTE: If the Value* passed is of type void then the constructor behaves as
2937 // if it was passed NULL.
2938 explicit ReturnInst(LLVMContext &C, Value *retVal = nullptr,
2939 Instruction *InsertBefore = nullptr);
2940 ReturnInst(LLVMContext &C, Value *retVal, BasicBlock *InsertAtEnd);
2941 explicit ReturnInst(LLVMContext &C, BasicBlock *InsertAtEnd);
2942
2943protected:
2944 // Note: Instruction needs to be a friend here to call cloneImpl.
2945 friend class Instruction;
2946
2947 ReturnInst *cloneImpl() const;
2948
2949public:
2950 static ReturnInst* Create(LLVMContext &C, Value *retVal = nullptr,
2951 Instruction *InsertBefore = nullptr) {
2952 return new(!!retVal) ReturnInst(C, retVal, InsertBefore);
2953 }
2954
2955 static ReturnInst* Create(LLVMContext &C, Value *retVal,
2956 BasicBlock *InsertAtEnd) {
2957 return new(!!retVal) ReturnInst(C, retVal, InsertAtEnd);
2958 }
2959
2960 static ReturnInst* Create(LLVMContext &C, BasicBlock *InsertAtEnd) {
2961 return new(0) ReturnInst(C, InsertAtEnd);
2962 }
2963
2964 /// Provide fast operand accessors
2965 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
;
2966
2967 /// Convenience accessor. Returns null if there is no return value.
2968 Value *getReturnValue() const {
2969 return getNumOperands() != 0 ? getOperand(0) : nullptr;
2970 }
2971
2972 unsigned getNumSuccessors() const { return 0; }
2973
2974 // Methods for support type inquiry through isa, cast, and dyn_cast:
2975 static bool classof(const Instruction *I) {
2976 return (I->getOpcode() == Instruction::Ret);
2977 }
2978 static bool classof(const Value *V) {
2979 return isa<Instruction>(V) && classof(cast<Instruction>(V));
2980 }
2981
2982private:
2983 BasicBlock *getSuccessor(unsigned idx) const {
2984 llvm_unreachable("ReturnInst has no successors!")::llvm::llvm_unreachable_internal("ReturnInst has no successors!"
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2984)
;
2985 }
2986
2987 void setSuccessor(unsigned idx, BasicBlock *B) {
2988 llvm_unreachable("ReturnInst has no successors!")::llvm::llvm_unreachable_internal("ReturnInst has no successors!"
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2988)
;
2989 }
2990};
2991
2992template <>
2993struct OperandTraits<ReturnInst> : public VariadicOperandTraits<ReturnInst> {
2994};
2995
2996DEFINE_TRANSPARENT_OPERAND_ACCESSORS(ReturnInst, Value)ReturnInst::op_iterator ReturnInst::op_begin() { return OperandTraits
<ReturnInst>::op_begin(this); } ReturnInst::const_op_iterator
ReturnInst::op_begin() const { return OperandTraits<ReturnInst
>::op_begin(const_cast<ReturnInst*>(this)); } ReturnInst
::op_iterator ReturnInst::op_end() { return OperandTraits<
ReturnInst>::op_end(this); } ReturnInst::const_op_iterator
ReturnInst::op_end() const { return OperandTraits<ReturnInst
>::op_end(const_cast<ReturnInst*>(this)); } Value *ReturnInst
::getOperand(unsigned i_nocapture) const { ((i_nocapture <
OperandTraits<ReturnInst>::operands(this) && "getOperand() out of range!"
) ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<ReturnInst>::operands(this) && \"getOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2996, __PRETTY_FUNCTION__)); return cast_or_null<Value>
( OperandTraits<ReturnInst>::op_begin(const_cast<ReturnInst
*>(this))[i_nocapture].get()); } void ReturnInst::setOperand
(unsigned i_nocapture, Value *Val_nocapture) { ((i_nocapture <
OperandTraits<ReturnInst>::operands(this) && "setOperand() out of range!"
) ? static_cast<void> (0) : __assert_fail ("i_nocapture < OperandTraits<ReturnInst>::operands(this) && \"setOperand() out of range!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 2996, __PRETTY_FUNCTION__)); OperandTraits<ReturnInst>
::op_begin(this)[i_nocapture] = Val_nocapture; } unsigned ReturnInst
::getNumOperands() const { return OperandTraits<ReturnInst
>::operands(this); } template <int Idx_nocapture> Use
&ReturnInst::Op() { return this->OpFrom<Idx_nocapture
>(this); } template <int Idx_nocapture> const Use &
ReturnInst::Op() const { return this->OpFrom<Idx_nocapture
>(this); }
2997
2998//===----------------------------------------------------------------------===//
2999// BranchInst Class
3000//===----------------------------------------------------------------------===//
3001
3002//===---------------------------------------------------------------------------
3003/// Conditional or Unconditional Branch instruction.
3004///
3005class BranchInst : public Instruction {
3006 /// Ops list - Branches are strange. The operands are ordered:
3007 /// [Cond, FalseDest,] TrueDest. This makes some accessors faster because
3008 /// they don't have to check for cond/uncond branchness. These are mostly
3009 /// accessed relative from op_end().
3010 BranchInst(const BranchInst &BI);
3011 // BranchInst constructors (where {B, T, F} are blocks, and C is a condition):
3012 // BranchInst(BB *B) - 'br B'
3013 // BranchInst(BB* T, BB *F, Value *C) - 'br C, T, F'
3014 // BranchInst(BB* B, Inst *I) - 'br B' insert before I
3015 // BranchInst(BB* T, BB *F, Value *C, Inst *I) - 'br C, T, F', insert before I
3016 // BranchInst(BB* B, BB *I) - 'br B' insert at end
3017 // BranchInst(BB* T, BB *F, Value *C, BB *I) - 'br C, T, F', insert at end
3018 explicit BranchInst(BasicBlock *IfTrue, Instruction *InsertBefore = nullptr);
3019 BranchInst(BasicBlock *IfTrue, BasicBlock *IfFalse, Value *Cond,
3020 Instruction *InsertBefore = nullptr);
3021 BranchInst(BasicBlock *IfTrue, BasicBlock *InsertAtEnd);
3022 BranchInst(BasicBlock *IfTrue, BasicBlock *IfFalse, Value *Cond,
3023 BasicBlock *InsertAtEnd);
3024
3025 void AssertOK();
3026
3027protected:
3028 // Note: Instruction needs to be a friend here to call cloneImpl.
3029 friend class Instruction;
3030
3031 BranchInst *cloneImpl() const;
3032
3033public:
3034 /// Iterator type that casts an operand to a basic block.
3035 ///
3036 /// This only makes sense because the successors are stored as adjacent
3037 /// operands for branch instructions.
3038 struct succ_op_iterator
3039 : iterator_adaptor_base<succ_op_iterator, value_op_iterator,
3040 std::random_access_iterator_tag, BasicBlock *,
3041 ptrdiff_t, BasicBlock *, BasicBlock *> {
3042 explicit succ_op_iterator(value_op_iterator I) : iterator_adaptor_base(I) {}
3043
3044 BasicBlock *operator*() const { return cast<BasicBlock>(*I); }
3045 BasicBlock *operator->() const { return operator*(); }
3046 };
3047
3048 /// The const version of `succ_op_iterator`.
3049 struct const_succ_op_iterator
3050 : iterator_adaptor_base<const_succ_op_iterator, const_value_op_iterator,
3051 std::random_access_iterator_tag,
3052 const BasicBlock *, ptrdiff_t, const BasicBlock *,
3053 const BasicBlock *> {
3054 explicit const_succ_op_iterator(const_value_op_iterator I)
3055 : iterator_adaptor_base(I) {}
3056
3057 const BasicBlock *operator*() const { return cast<BasicBlock>(*I); }
3058 const BasicBlock *operator->() const { return operator*(); }
3059 };
3060
3061 static BranchInst *Create(BasicBlock *IfTrue,
3062 Instruction *InsertBefore = nullptr) {
3063 return new(1) BranchInst(IfTrue, InsertBefore);
3064 }
3065
3066 static BranchInst *Create(BasicBlock *IfTrue, BasicBlock *IfFalse,
3067 Value *Cond, Instruction *InsertBefore = nullptr) {
3068 return new(3) BranchInst(IfTrue, IfFalse, Cond, InsertBefore);
3069 }
3070
3071 static BranchInst *Create(BasicBlock *IfTrue, BasicBlock *InsertAtEnd) {
3072 return new(1) BranchInst(IfTrue, InsertAtEnd);
3073 }
3074
3075 static BranchInst *Create(BasicBlock *IfTrue, BasicBlock *IfFalse,
3076 Value *Cond, BasicBlock *InsertAtEnd) {
3077 return new(3) BranchInst(IfTrue, IfFalse, Cond, InsertAtEnd);
3078 }
3079
3080 /// Transparently provide more efficient getOperand methods.
3081 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
;
3082
3083 bool isUnconditional() const { return getNumOperands() == 1; }
3084 bool isConditional() const { return getNumOperands() == 3; }
3085
3086 Value *getCondition() const {
3087 assert(isConditional() && "Cannot get condition of an uncond branch!")((isConditional() && "Cannot get condition of an uncond branch!"
) ? static_cast<void> (0) : __assert_fail ("isConditional() && \"Cannot get condition of an uncond branch!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 3087, __PRETTY_FUNCTION__))
;
3088 return Op<-3>();
3089 }
3090
3091 void setCondition(Value *V) {
3092 assert(isConditional() && "Cannot set condition of unconditional branch!")((isConditional() && "Cannot set condition of unconditional branch!"
) ? static_cast<void> (0) : __assert_fail ("isConditional() && \"Cannot set condition of unconditional branch!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 3092, __PRETTY_FUNCTION__))
;
3093 Op<-3>() = V;
3094 }
3095
3096 unsigned getNumSuccessors() const { return 1+isConditional(); }
3097
3098 BasicBlock *getSuccessor(unsigned i) const {
3099 assert(i < getNumSuccessors() && "Successor # out of range for Branch!")((i < getNumSuccessors() && "Successor # out of range for Branch!"
) ? static_cast<void> (0) : __assert_fail ("i < getNumSuccessors() && \"Successor # out of range for Branch!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 3099, __PRETTY_FUNCTION__))
;
3100 return cast_or_null<BasicBlock>((&Op<-1>() - i)->get());
3101 }
3102
3103 void setSuccessor(unsigned idx, BasicBlock *NewSucc) {
3104 assert(idx < getNumSuccessors() && "Successor # out of range for Branch!")((idx < getNumSuccessors() && "Successor # out of range for Branch!"
) ? static_cast<void> (0) : __assert_fail ("idx < getNumSuccessors() && \"Successor # out of range for Branch!\""
, "/build/llvm-toolchain-snapshot-12.0.0~++20201102111116+1ed2ca68191/llvm/include/llvm/IR/Instructions.h"
, 3104, __PRETTY_FUNCTION__))
;
3105