Bug Summary

File:llvm/lib/Transforms/Scalar/LICM.cpp
Warning:line 1113, column 22
Called C++ object pointer is null

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name 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 -mthread-model posix -mframe-pointer=none -fmath-errno -fno-rounding-math -masm-verbose -mconstructor-aliases -munwind-tables -fuse-init-array -target-cpu x86-64 -dwarf-column-info -debugger-tuning=gdb -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-10/lib/clang/10.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-10~+201911111502510600c19528f1809/build-llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-10~+201911111502510600c19528f1809/build-llvm/include -I /build/llvm-toolchain-snapshot-10~+201911111502510600c19528f1809/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-10/lib/clang/10.0.0/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-comment -std=c++14 -fdeprecated-macro -fdebug-compilation-dir /build/llvm-toolchain-snapshot-10~+201911111502510600c19528f1809/build-llvm/lib/Transforms/Scalar -fdebug-prefix-map=/build/llvm-toolchain-snapshot-10~+201911111502510600c19528f1809=. -ferror-limit 19 -fmessage-length 0 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -fobjc-runtime=gcc -fdiagnostics-show-option -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -o /tmp/scan-build-2019-12-07-102640-14763-1 -x c++ /build/llvm-toolchain-snapshot-10~+201911111502510600c19528f1809/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/CaptureTracking.h"
39#include "llvm/Analysis/ConstantFolding.h"
40#include "llvm/Analysis/GlobalsModRef.h"
41#include "llvm/Analysis/GuardUtils.h"
42#include "llvm/Analysis/Loads.h"
43#include "llvm/Analysis/LoopInfo.h"
44#include "llvm/Analysis/LoopIterator.h"
45#include "llvm/Analysis/LoopPass.h"
46#include "llvm/Analysis/MemoryBuiltins.h"
47#include "llvm/Analysis/MemorySSA.h"
48#include "llvm/Analysis/MemorySSAUpdater.h"
49#include "llvm/Analysis/OptimizationRemarkEmitter.h"
50#include "llvm/Analysis/ScalarEvolution.h"
51#include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
52#include "llvm/Analysis/TargetLibraryInfo.h"
53#include "llvm/Analysis/ValueTracking.h"
54#include "llvm/IR/CFG.h"
55#include "llvm/IR/Constants.h"
56#include "llvm/IR/DataLayout.h"
57#include "llvm/IR/DebugInfoMetadata.h"
58#include "llvm/IR/DerivedTypes.h"
59#include "llvm/IR/Dominators.h"
60#include "llvm/IR/Instructions.h"
61#include "llvm/IR/IntrinsicInst.h"
62#include "llvm/IR/LLVMContext.h"
63#include "llvm/IR/Metadata.h"
64#include "llvm/IR/PatternMatch.h"
65#include "llvm/IR/PredIteratorCache.h"
66#include "llvm/Support/CommandLine.h"
67#include "llvm/Support/Debug.h"
68#include "llvm/Support/raw_ostream.h"
69#include "llvm/Transforms/Scalar.h"
70#include "llvm/Transforms/Scalar/LoopPassManager.h"
71#include "llvm/Transforms/Utils/BasicBlockUtils.h"
72#include "llvm/Transforms/Utils/Local.h"
73#include "llvm/Transforms/Utils/LoopUtils.h"
74#include "llvm/Transforms/Utils/SSAUpdater.h"
75#include <algorithm>
76#include <utility>
77using namespace llvm;
78
79#define DEBUG_TYPE"licm" "licm"
80
81STATISTIC(NumCreatedBlocks, "Number of blocks created")static llvm::Statistic NumCreatedBlocks = {"licm", "NumCreatedBlocks"
, "Number of blocks created"}
;
82STATISTIC(NumClonedBranches, "Number of branches cloned")static llvm::Statistic NumClonedBranches = {"licm", "NumClonedBranches"
, "Number of branches cloned"}
;
83STATISTIC(NumSunk, "Number of instructions sunk out of loop")static llvm::Statistic NumSunk = {"licm", "NumSunk", "Number of instructions sunk out of loop"
}
;
84STATISTIC(NumHoisted, "Number of instructions hoisted out of loop")static llvm::Statistic NumHoisted = {"licm", "NumHoisted", "Number of instructions hoisted out of loop"
}
;
85STATISTIC(NumMovedLoads, "Number of load insts hoisted or sunk")static llvm::Statistic NumMovedLoads = {"licm", "NumMovedLoads"
, "Number of load insts hoisted or sunk"}
;
86STATISTIC(NumMovedCalls, "Number of call insts hoisted or sunk")static llvm::Statistic NumMovedCalls = {"licm", "NumMovedCalls"
, "Number of call insts hoisted or sunk"}
;
87STATISTIC(NumPromoted, "Number of memory locations promoted to registers")static llvm::Statistic NumPromoted = {"licm", "NumPromoted", "Number of memory locations promoted to registers"
}
;
88
89/// Memory promotion is enabled by default.
90static cl::opt<bool>
91 DisablePromotion("disable-licm-promotion", cl::Hidden, cl::init(false),
92 cl::desc("Disable memory promotion in LICM pass"));
93
94static cl::opt<bool> ControlFlowHoisting(
95 "licm-control-flow-hoisting", cl::Hidden, cl::init(false),
96 cl::desc("Enable control flow (and PHI) hoisting in LICM"));
97
98static cl::opt<uint32_t> MaxNumUsesTraversed(
99 "licm-max-num-uses-traversed", cl::Hidden, cl::init(8),
100 cl::desc("Max num uses visited for identifying load "
101 "invariance in loop using invariant start (default = 8)"));
102
103// Default value of zero implies we use the regular alias set tracker mechanism
104// instead of the cross product using AA to identify aliasing of the memory
105// location we are interested in.
106static cl::opt<int>
107LICMN2Theshold("licm-n2-threshold", cl::Hidden, cl::init(0),
108 cl::desc("How many instruction to cross product using AA"));
109
110// Experimental option to allow imprecision in LICM in pathological cases, in
111// exchange for faster compile. This is to be removed if MemorySSA starts to
112// address the same issue. This flag applies only when LICM uses MemorySSA
113// instead on AliasSetTracker. LICM calls MemorySSAWalker's
114// getClobberingMemoryAccess, up to the value of the Cap, getting perfect
115// accuracy. Afterwards, LICM will call into MemorySSA's getDefiningAccess,
116// which may not be precise, since optimizeUses is capped. The result is
117// correct, but we may not get as "far up" as possible to get which access is
118// clobbering the one queried.
119cl::opt<unsigned> llvm::SetLicmMssaOptCap(
120 "licm-mssa-optimization-cap", cl::init(100), cl::Hidden,
121 cl::desc("Enable imprecision in LICM in pathological cases, in exchange "
122 "for faster compile. Caps the MemorySSA clobbering calls."));
123
124// Experimentally, memory promotion carries less importance than sinking and
125// hoisting. Limit when we do promotion when using MemorySSA, in order to save
126// compile time.
127cl::opt<unsigned> llvm::SetLicmMssaNoAccForPromotionCap(
128 "licm-mssa-max-acc-promotion", cl::init(250), cl::Hidden,
129 cl::desc("[LICM & MemorySSA] When MSSA in LICM is disabled, this has no "
130 "effect. When MSSA in LICM is enabled, then this is the maximum "
131 "number of accesses allowed to be present in a loop in order to "
132 "enable memory promotion."));
133
134static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI);
135static bool isNotUsedOrFreeInLoop(const Instruction &I, const Loop *CurLoop,
136 const LoopSafetyInfo *SafetyInfo,
137 TargetTransformInfo *TTI, bool &FreeInLoop);
138static void hoist(Instruction &I, const DominatorTree *DT, const Loop *CurLoop,
139 BasicBlock *Dest, ICFLoopSafetyInfo *SafetyInfo,
140 MemorySSAUpdater *MSSAU, ScalarEvolution *SE,
141 OptimizationRemarkEmitter *ORE);
142static bool sink(Instruction &I, LoopInfo *LI, DominatorTree *DT,
143 const Loop *CurLoop, ICFLoopSafetyInfo *SafetyInfo,
144 MemorySSAUpdater *MSSAU, OptimizationRemarkEmitter *ORE);
145static bool isSafeToExecuteUnconditionally(Instruction &Inst,
146 const DominatorTree *DT,
147 const Loop *CurLoop,
148 const LoopSafetyInfo *SafetyInfo,
149 OptimizationRemarkEmitter *ORE,
150 const Instruction *CtxI = nullptr);
151static bool pointerInvalidatedByLoop(MemoryLocation MemLoc,
152 AliasSetTracker *CurAST, Loop *CurLoop,
153 AliasAnalysis *AA);
154static bool pointerInvalidatedByLoopWithMSSA(MemorySSA *MSSA, MemoryUse *MU,
155 Loop *CurLoop,
156 SinkAndHoistLICMFlags &Flags);
157static Instruction *CloneInstructionInExitBlock(
158 Instruction &I, BasicBlock &ExitBlock, PHINode &PN, const LoopInfo *LI,
159 const LoopSafetyInfo *SafetyInfo, MemorySSAUpdater *MSSAU);
160
161static void eraseInstruction(Instruction &I, ICFLoopSafetyInfo &SafetyInfo,
162 AliasSetTracker *AST, MemorySSAUpdater *MSSAU);
163
164static void moveInstructionBefore(Instruction &I, Instruction &Dest,
165 ICFLoopSafetyInfo &SafetyInfo,
166 MemorySSAUpdater *MSSAU, ScalarEvolution *SE);
167
168namespace {
169struct LoopInvariantCodeMotion {
170 using ASTrackerMapTy = DenseMap<Loop *, std::unique_ptr<AliasSetTracker>>;
171 bool runOnLoop(Loop *L, AliasAnalysis *AA, LoopInfo *LI, DominatorTree *DT,
172 TargetLibraryInfo *TLI, TargetTransformInfo *TTI,
173 ScalarEvolution *SE, MemorySSA *MSSA,
174 OptimizationRemarkEmitter *ORE, bool DeleteAST);
175
176 ASTrackerMapTy &getLoopToAliasSetMap() { return LoopToAliasSetMap; }
177 LoopInvariantCodeMotion(unsigned LicmMssaOptCap,
178 unsigned LicmMssaNoAccForPromotionCap)
179 : LicmMssaOptCap(LicmMssaOptCap),
180 LicmMssaNoAccForPromotionCap(LicmMssaNoAccForPromotionCap) {}
181
182private:
183 ASTrackerMapTy LoopToAliasSetMap;
184 unsigned LicmMssaOptCap;
185 unsigned LicmMssaNoAccForPromotionCap;
186
187 std::unique_ptr<AliasSetTracker>
188 collectAliasInfoForLoop(Loop *L, LoopInfo *LI, AliasAnalysis *AA);
189 std::unique_ptr<AliasSetTracker>
190 collectAliasInfoForLoopWithMSSA(Loop *L, AliasAnalysis *AA,
191 MemorySSAUpdater *MSSAU);
192};
193
194struct LegacyLICMPass : public LoopPass {
195 static char ID; // Pass identification, replacement for typeid
196 LegacyLICMPass(
197 unsigned LicmMssaOptCap = SetLicmMssaOptCap,
198 unsigned LicmMssaNoAccForPromotionCap = SetLicmMssaNoAccForPromotionCap)
199 : LoopPass(ID), LICM(LicmMssaOptCap, LicmMssaNoAccForPromotionCap) {
200 initializeLegacyLICMPassPass(*PassRegistry::getPassRegistry());
201 }
202
203 bool runOnLoop(Loop *L, LPPassManager &LPM) override {
204 if (skipLoop(L)) {
205 // If we have run LICM on a previous loop but now we are skipping
206 // (because we've hit the opt-bisect limit), we need to clear the
207 // loop alias information.
208 LICM.getLoopToAliasSetMap().clear();
209 return false;
210 }
211
212 auto *SE = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>();
213 MemorySSA *MSSA = EnableMSSALoopDependency
214 ? (&getAnalysis<MemorySSAWrapperPass>().getMSSA())
215 : nullptr;
216 // For the old PM, we can't use OptimizationRemarkEmitter as an analysis
217 // pass. Function analyses need to be preserved across loop transformations
218 // but ORE cannot be preserved (see comment before the pass definition).
219 OptimizationRemarkEmitter ORE(L->getHeader()->getParent());
220 return LICM.runOnLoop(L,
221 &getAnalysis<AAResultsWrapperPass>().getAAResults(),
222 &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(),
223 &getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
224 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(
225 *L->getHeader()->getParent()),
226 &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
227 *L->getHeader()->getParent()),
228 SE ? &SE->getSE() : nullptr, MSSA, &ORE, false);
229 }
230
231 /// This transformation requires natural loop information & requires that
232 /// loop preheaders be inserted into the CFG...
233 ///
234 void getAnalysisUsage(AnalysisUsage &AU) const override {
235 AU.addPreserved<DominatorTreeWrapperPass>();
236 AU.addPreserved<LoopInfoWrapperPass>();
237 AU.addRequired<TargetLibraryInfoWrapperPass>();
238 if (EnableMSSALoopDependency) {
239 AU.addRequired<MemorySSAWrapperPass>();
240 AU.addPreserved<MemorySSAWrapperPass>();
241 }
242 AU.addRequired<TargetTransformInfoWrapperPass>();
243 getLoopAnalysisUsage(AU);
244 }
245
246 using llvm::Pass::doFinalization;
247
248 bool doFinalization() override {
249 auto &AliasSetMap = LICM.getLoopToAliasSetMap();
250 // All loops in the AliasSetMap should be cleaned up already. The only case
251 // where we fail to do so is if an outer loop gets deleted before LICM
252 // visits it.
253 assert(all_of(AliasSetMap,((all_of(AliasSetMap, [](LoopInvariantCodeMotion::ASTrackerMapTy
::value_type &KV) { return !KV.first->getParentLoop();
}) && "Didn't free loop alias sets") ? static_cast<
void> (0) : __assert_fail ("all_of(AliasSetMap, [](LoopInvariantCodeMotion::ASTrackerMapTy::value_type &KV) { return !KV.first->getParentLoop(); }) && \"Didn't free loop alias sets\""
, "/build/llvm-toolchain-snapshot-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 257, __PRETTY_FUNCTION__))
254 [](LoopInvariantCodeMotion::ASTrackerMapTy::value_type &KV) {((all_of(AliasSetMap, [](LoopInvariantCodeMotion::ASTrackerMapTy
::value_type &KV) { return !KV.first->getParentLoop();
}) && "Didn't free loop alias sets") ? static_cast<
void> (0) : __assert_fail ("all_of(AliasSetMap, [](LoopInvariantCodeMotion::ASTrackerMapTy::value_type &KV) { return !KV.first->getParentLoop(); }) && \"Didn't free loop alias sets\""
, "/build/llvm-toolchain-snapshot-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 257, __PRETTY_FUNCTION__))
255 return !KV.first->getParentLoop();((all_of(AliasSetMap, [](LoopInvariantCodeMotion::ASTrackerMapTy
::value_type &KV) { return !KV.first->getParentLoop();
}) && "Didn't free loop alias sets") ? static_cast<
void> (0) : __assert_fail ("all_of(AliasSetMap, [](LoopInvariantCodeMotion::ASTrackerMapTy::value_type &KV) { return !KV.first->getParentLoop(); }) && \"Didn't free loop alias sets\""
, "/build/llvm-toolchain-snapshot-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 257, __PRETTY_FUNCTION__))
256 }) &&((all_of(AliasSetMap, [](LoopInvariantCodeMotion::ASTrackerMapTy
::value_type &KV) { return !KV.first->getParentLoop();
}) && "Didn't free loop alias sets") ? static_cast<
void> (0) : __assert_fail ("all_of(AliasSetMap, [](LoopInvariantCodeMotion::ASTrackerMapTy::value_type &KV) { return !KV.first->getParentLoop(); }) && \"Didn't free loop alias sets\""
, "/build/llvm-toolchain-snapshot-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 257, __PRETTY_FUNCTION__))
257 "Didn't free loop alias sets")((all_of(AliasSetMap, [](LoopInvariantCodeMotion::ASTrackerMapTy
::value_type &KV) { return !KV.first->getParentLoop();
}) && "Didn't free loop alias sets") ? static_cast<
void> (0) : __assert_fail ("all_of(AliasSetMap, [](LoopInvariantCodeMotion::ASTrackerMapTy::value_type &KV) { return !KV.first->getParentLoop(); }) && \"Didn't free loop alias sets\""
, "/build/llvm-toolchain-snapshot-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 257, __PRETTY_FUNCTION__))
;
258 AliasSetMap.clear();
259 return false;
260 }
261
262private:
263 LoopInvariantCodeMotion LICM;
264
265 /// cloneBasicBlockAnalysis - Simple Analysis hook. Clone alias set info.
266 void cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To,
267 Loop *L) override;
268
269 /// deleteAnalysisValue - Simple Analysis hook. Delete value V from alias
270 /// set.
271 void deleteAnalysisValue(Value *V, Loop *L) override;
272
273 /// Simple Analysis hook. Delete loop L from alias set map.
274 void deleteAnalysisLoop(Loop *L) override;
275};
276} // namespace
277
278PreservedAnalyses LICMPass::run(Loop &L, LoopAnalysisManager &AM,
279 LoopStandardAnalysisResults &AR, LPMUpdater &) {
280 const auto &FAM =
281 AM.getResult<FunctionAnalysisManagerLoopProxy>(L, AR).getManager();
282 Function *F = L.getHeader()->getParent();
283
284 auto *ORE = FAM.getCachedResult<OptimizationRemarkEmitterAnalysis>(*F);
285 // FIXME: This should probably be optional rather than required.
286 if (!ORE)
287 report_fatal_error("LICM: OptimizationRemarkEmitterAnalysis not "
288 "cached at a higher level");
289
290 LoopInvariantCodeMotion LICM(LicmMssaOptCap, LicmMssaNoAccForPromotionCap);
291 if (!LICM.runOnLoop(&L, &AR.AA, &AR.LI, &AR.DT, &AR.TLI, &AR.TTI, &AR.SE,
292 AR.MSSA, ORE, true))
293 return PreservedAnalyses::all();
294
295 auto PA = getLoopPassPreservedAnalyses();
296
297 PA.preserve<DominatorTreeAnalysis>();
298 PA.preserve<LoopAnalysis>();
299 if (AR.MSSA)
300 PA.preserve<MemorySSAAnalysis>();
301
302 return PA;
303}
304
305char LegacyLICMPass::ID = 0;
306INITIALIZE_PASS_BEGIN(LegacyLICMPass, "licm", "Loop Invariant Code Motion",static void *initializeLegacyLICMPassPassOnce(PassRegistry &
Registry) {
307 false, false)static void *initializeLegacyLICMPassPassOnce(PassRegistry &
Registry) {
308INITIALIZE_PASS_DEPENDENCY(LoopPass)initializeLoopPassPass(Registry);
309INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry);
310INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry);
311INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)initializeMemorySSAWrapperPassPass(Registry);
312INITIALIZE_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)); }
313 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)); }
314
315Pass *llvm::createLICMPass() { return new LegacyLICMPass(); }
316Pass *llvm::createLICMPass(unsigned LicmMssaOptCap,
317 unsigned LicmMssaNoAccForPromotionCap) {
318 return new LegacyLICMPass(LicmMssaOptCap, LicmMssaNoAccForPromotionCap);
319}
320
321/// Hoist expressions out of the specified loop. Note, alias info for inner
322/// loop is not preserved so it is not a good idea to run LICM multiple
323/// times on one loop.
324/// We should delete AST for inner loops in the new pass manager to avoid
325/// memory leak.
326///
327bool LoopInvariantCodeMotion::runOnLoop(
328 Loop *L, AliasAnalysis *AA, LoopInfo *LI, DominatorTree *DT,
329 TargetLibraryInfo *TLI, TargetTransformInfo *TTI, ScalarEvolution *SE,
330 MemorySSA *MSSA, OptimizationRemarkEmitter *ORE, bool DeleteAST) {
331 bool Changed = false;
332
333 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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 333, __PRETTY_FUNCTION__))
;
334
335 // If this loop has metadata indicating that LICM is not to be performed then
336 // just exit.
337 if (hasDisableLICMTransformsHint(L)) {
338 return false;
339 }
340
341 std::unique_ptr<AliasSetTracker> CurAST;
342 std::unique_ptr<MemorySSAUpdater> MSSAU;
343 bool NoOfMemAccTooLarge = false;
344 unsigned LicmMssaOptCounter = 0;
345
346 if (!MSSA) {
347 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)
;
348 CurAST = collectAliasInfoForLoop(L, LI, AA);
349 } else {
350 LLVM_DEBUG(dbgs() << "LICM: Using MemorySSA.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "LICM: Using MemorySSA.\n"; } } while
(false)
;
351 MSSAU = std::make_unique<MemorySSAUpdater>(MSSA);
352
353 unsigned AccessCapCount = 0;
354 for (auto *BB : L->getBlocks()) {
355 if (auto *Accesses = MSSA->getBlockAccesses(BB)) {
356 for (const auto &MA : *Accesses) {
357 (void)MA;
358 AccessCapCount++;
359 if (AccessCapCount > LicmMssaNoAccForPromotionCap) {
360 NoOfMemAccTooLarge = true;
361 break;
362 }
363 }
364 }
365 if (NoOfMemAccTooLarge)
366 break;
367 }
368 }
369
370 // Get the preheader block to move instructions into...
371 BasicBlock *Preheader = L->getLoopPreheader();
372
373 // Compute loop safety information.
374 ICFLoopSafetyInfo SafetyInfo(DT);
375 SafetyInfo.computeLoopSafetyInfo(L);
376
377 // We want to visit all of the instructions in this loop... that are not parts
378 // of our subloops (they have already had their invariants hoisted out of
379 // their loop, into this loop, so there is no need to process the BODIES of
380 // the subloops).
381 //
382 // Traverse the body of the loop in depth first order on the dominator tree so
383 // that we are guaranteed to see definitions before we see uses. This allows
384 // us to sink instructions in one pass, without iteration. After sinking
385 // instructions, we perform another pass to hoist them out of the loop.
386 SinkAndHoistLICMFlags Flags = {NoOfMemAccTooLarge, LicmMssaOptCounter,
387 LicmMssaOptCap, LicmMssaNoAccForPromotionCap,
388 /*IsSink=*/true};
389 if (L->hasDedicatedExits())
390 Changed |= sinkRegion(DT->getNode(L->getHeader()), AA, LI, DT, TLI, TTI, L,
391 CurAST.get(), MSSAU.get(), &SafetyInfo, Flags, ORE);
392 Flags.IsSink = false;
393 if (Preheader)
394 Changed |=
395 hoistRegion(DT->getNode(L->getHeader()), AA, LI, DT, TLI, L,
396 CurAST.get(), MSSAU.get(), SE, &SafetyInfo, Flags, ORE);
397
398 // Now that all loop invariants have been removed from the loop, promote any
399 // memory references to scalars that we can.
400 // Don't sink stores from loops without dedicated block exits. Exits
401 // containing indirect branches are not transformed by loop simplify,
402 // make sure we catch that. An additional load may be generated in the
403 // preheader for SSA updater, so also avoid sinking when no preheader
404 // is available.
405 if (!DisablePromotion && Preheader && L->hasDedicatedExits() &&
406 !NoOfMemAccTooLarge) {
407 // Figure out the loop exits and their insertion points
408 SmallVector<BasicBlock *, 8> ExitBlocks;
409 L->getUniqueExitBlocks(ExitBlocks);
410
411 // We can't insert into a catchswitch.
412 bool HasCatchSwitch = llvm::any_of(ExitBlocks, [](BasicBlock *Exit) {
413 return isa<CatchSwitchInst>(Exit->getTerminator());
414 });
415
416 if (!HasCatchSwitch) {
417 SmallVector<Instruction *, 8> InsertPts;
418 SmallVector<MemoryAccess *, 8> MSSAInsertPts;
419 InsertPts.reserve(ExitBlocks.size());
420 if (MSSAU)
421 MSSAInsertPts.reserve(ExitBlocks.size());
422 for (BasicBlock *ExitBlock : ExitBlocks) {
423 InsertPts.push_back(&*ExitBlock->getFirstInsertionPt());
424 if (MSSAU)
425 MSSAInsertPts.push_back(nullptr);
426 }
427
428 PredIteratorCache PIC;
429
430 bool Promoted = false;
431
432 // Build an AST using MSSA.
433 if (!CurAST.get())
434 CurAST = collectAliasInfoForLoopWithMSSA(L, AA, MSSAU.get());
435
436 // Loop over all of the alias sets in the tracker object.
437 for (AliasSet &AS : *CurAST) {
438 // We can promote this alias set if it has a store, if it is a "Must"
439 // alias set, if the pointer is loop invariant, and if we are not
440 // eliminating any volatile loads or stores.
441 if (AS.isForwardingAliasSet() || !AS.isMod() || !AS.isMustAlias() ||
442 !L->isLoopInvariant(AS.begin()->getValue()))
443 continue;
444
445 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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 447, __PRETTY_FUNCTION__))
446 !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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 447, __PRETTY_FUNCTION__))
447 "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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 447, __PRETTY_FUNCTION__))
;
448
449 SmallSetVector<Value *, 8> PointerMustAliases;
450 for (const auto &ASI : AS)
451 PointerMustAliases.insert(ASI.getValue());
452
453 Promoted |= promoteLoopAccessesToScalars(
454 PointerMustAliases, ExitBlocks, InsertPts, MSSAInsertPts, PIC, LI,
455 DT, TLI, L, CurAST.get(), MSSAU.get(), &SafetyInfo, ORE);
456 }
457
458 // Once we have promoted values across the loop body we have to
459 // recursively reform LCSSA as any nested loop may now have values defined
460 // within the loop used in the outer loop.
461 // FIXME: This is really heavy handed. It would be a bit better to use an
462 // SSAUpdater strategy during promotion that was LCSSA aware and reformed
463 // it as it went.
464 if (Promoted)
465 formLCSSARecursively(*L, *DT, LI, SE);
466
467 Changed |= Promoted;
468 }
469 }
470
471 // Check that neither this loop nor its parent have had LCSSA broken. LICM is
472 // specifically moving instructions across the loop boundary and so it is
473 // especially in need of sanity checking here.
474 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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 474, __PRETTY_FUNCTION__))
;
475 assert((!L->getParentLoop() || L->getParentLoop()->isLCSSAForm(*DT)) &&(((!L->getParentLoop() || L->getParentLoop()->isLCSSAForm
(*DT)) && "Parent loop not left in LCSSA form after LICM!"
) ? static_cast<void> (0) : __assert_fail ("(!L->getParentLoop() || L->getParentLoop()->isLCSSAForm(*DT)) && \"Parent loop not left in LCSSA form after LICM!\""
, "/build/llvm-toolchain-snapshot-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 476, __PRETTY_FUNCTION__))
476 "Parent loop not left in LCSSA form after LICM!")(((!L->getParentLoop() || L->getParentLoop()->isLCSSAForm
(*DT)) && "Parent loop not left in LCSSA form after LICM!"
) ? static_cast<void> (0) : __assert_fail ("(!L->getParentLoop() || L->getParentLoop()->isLCSSAForm(*DT)) && \"Parent loop not left in LCSSA form after LICM!\""
, "/build/llvm-toolchain-snapshot-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 476, __PRETTY_FUNCTION__))
;
477
478 // If this loop is nested inside of another one, save the alias information
479 // for when we process the outer loop.
480 if (!MSSAU.get() && CurAST.get() && L->getParentLoop() && !DeleteAST)
481 LoopToAliasSetMap[L] = std::move(CurAST);
482
483 if (MSSAU.get() && VerifyMemorySSA)
484 MSSAU->getMemorySSA()->verifyMemorySSA();
485
486 if (Changed && SE)
487 SE->forgetLoopDispositions(L);
488 return Changed;
489}
490
491/// Walk the specified region of the CFG (defined by all blocks dominated by
492/// the specified block, and that are in the current loop) in reverse depth
493/// first order w.r.t the DominatorTree. This allows us to visit uses before
494/// definitions, allowing us to sink a loop body in one pass without iteration.
495///
496bool llvm::sinkRegion(DomTreeNode *N, AliasAnalysis *AA, LoopInfo *LI,
497 DominatorTree *DT, TargetLibraryInfo *TLI,
498 TargetTransformInfo *TTI, Loop *CurLoop,
499 AliasSetTracker *CurAST, MemorySSAUpdater *MSSAU,
500 ICFLoopSafetyInfo *SafetyInfo,
501 SinkAndHoistLICMFlags &Flags,
502 OptimizationRemarkEmitter *ORE) {
503
504 // Verify inputs.
505 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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 507, __PRETTY_FUNCTION__))
506 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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 507, __PRETTY_FUNCTION__))
507 "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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 507, __PRETTY_FUNCTION__))
;
508 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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 509, __PRETTY_FUNCTION__))
509 "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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 509, __PRETTY_FUNCTION__))
;
510
511 // We want to visit children before parents. We will enque all the parents
512 // before their children in the worklist and process the worklist in reverse
513 // order.
514 SmallVector<DomTreeNode *, 16> Worklist = collectChildrenInLoop(N, CurLoop);
515
516 bool Changed = false;
517 for (DomTreeNode *DTN : reverse(Worklist)) {
518 BasicBlock *BB = DTN->getBlock();
519 // Only need to process the contents of this block if it is not part of a
520 // subloop (which would already have been processed).
521 if (inSubLoop(BB, CurLoop, LI))
522 continue;
523
524 for (BasicBlock::iterator II = BB->end(); II != BB->begin();) {
525 Instruction &I = *--II;
526
527 // If the instruction is dead, we would try to sink it because it isn't
528 // used in the loop, instead, just delete it.
529 if (isInstructionTriviallyDead(&I, TLI)) {
530 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)
;
531 salvageDebugInfo(I);
532 ++II;
533 eraseInstruction(I, *SafetyInfo, CurAST, MSSAU);
534 Changed = true;
535 continue;
536 }
537
538 // Check to see if we can sink this instruction to the exit blocks
539 // of the loop. We can do this if the all users of the instruction are
540 // outside of the loop. In this case, it doesn't even matter if the
541 // operands of the instruction are loop invariant.
542 //
543 bool FreeInLoop = false;
544 if (isNotUsedOrFreeInLoop(I, CurLoop, SafetyInfo, TTI, FreeInLoop) &&
545 canSinkOrHoistInst(I, AA, DT, CurLoop, CurAST, MSSAU, true, &Flags,
546 ORE) &&
547 !I.mayHaveSideEffects()) {
548 if (sink(I, LI, DT, CurLoop, SafetyInfo, MSSAU, ORE)) {
549 if (!FreeInLoop) {
550 ++II;
551 eraseInstruction(I, *SafetyInfo, CurAST, MSSAU);
552 }
553 Changed = true;
554 }
555 }
556 }
557 }
558 if (MSSAU && VerifyMemorySSA)
559 MSSAU->getMemorySSA()->verifyMemorySSA();
560 return Changed;
561}
562
563namespace {
564// This is a helper class for hoistRegion to make it able to hoist control flow
565// in order to be able to hoist phis. The way this works is that we initially
566// start hoisting to the loop preheader, and when we see a loop invariant branch
567// we make note of this. When we then come to hoist an instruction that's
568// conditional on such a branch we duplicate the branch and the relevant control
569// flow, then hoist the instruction into the block corresponding to its original
570// block in the duplicated control flow.
571class ControlFlowHoister {
572private:
573 // Information about the loop we are hoisting from
574 LoopInfo *LI;
575 DominatorTree *DT;
576 Loop *CurLoop;
577 MemorySSAUpdater *MSSAU;
578
579 // A map of blocks in the loop to the block their instructions will be hoisted
580 // to.
581 DenseMap<BasicBlock *, BasicBlock *> HoistDestinationMap;
582
583 // The branches that we can hoist, mapped to the block that marks a
584 // convergence point of their control flow.
585 DenseMap<BranchInst *, BasicBlock *> HoistableBranches;
586
587public:
588 ControlFlowHoister(LoopInfo *LI, DominatorTree *DT, Loop *CurLoop,
589 MemorySSAUpdater *MSSAU)
590 : LI(LI), DT(DT), CurLoop(CurLoop), MSSAU(MSSAU) {}
591
592 void registerPossiblyHoistableBranch(BranchInst *BI) {
593 // We can only hoist conditional branches with loop invariant operands.
594 if (!ControlFlowHoisting || !BI->isConditional() ||
595 !CurLoop->hasLoopInvariantOperands(BI))
596 return;
597
598 // The branch destinations need to be in the loop, and we don't gain
599 // anything by duplicating conditional branches with duplicate successors,
600 // as it's essentially the same as an unconditional branch.
601 BasicBlock *TrueDest = BI->getSuccessor(0);
602 BasicBlock *FalseDest = BI->getSuccessor(1);
603 if (!CurLoop->contains(TrueDest) || !CurLoop->contains(FalseDest) ||
604 TrueDest == FalseDest)
605 return;
606
607 // We can hoist BI if one branch destination is the successor of the other,
608 // or both have common successor which we check by seeing if the
609 // intersection of their successors is non-empty.
610 // TODO: This could be expanded to allowing branches where both ends
611 // eventually converge to a single block.
612 SmallPtrSet<BasicBlock *, 4> TrueDestSucc, FalseDestSucc;
613 TrueDestSucc.insert(succ_begin(TrueDest), succ_end(TrueDest));
614 FalseDestSucc.insert(succ_begin(FalseDest), succ_end(FalseDest));
615 BasicBlock *CommonSucc = nullptr;
616 if (TrueDestSucc.count(FalseDest)) {
617 CommonSucc = FalseDest;
618 } else if (FalseDestSucc.count(TrueDest)) {
619 CommonSucc = TrueDest;
620 } else {
621 set_intersect(TrueDestSucc, FalseDestSucc);
622 // If there's one common successor use that.
623 if (TrueDestSucc.size() == 1)
624 CommonSucc = *TrueDestSucc.begin();
625 // If there's more than one pick whichever appears first in the block list
626 // (we can't use the value returned by TrueDestSucc.begin() as it's
627 // unpredicatable which element gets returned).
628 else if (!TrueDestSucc.empty()) {
629 Function *F = TrueDest->getParent();
630 auto IsSucc = [&](BasicBlock &BB) { return TrueDestSucc.count(&BB); };
631 auto It = std::find_if(F->begin(), F->end(), IsSucc);
632 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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 632, __PRETTY_FUNCTION__))
;
633 CommonSucc = &*It;
634 }
635 }
636 // The common successor has to be dominated by the branch, as otherwise
637 // there will be some other path to the successor that will not be
638 // controlled by this branch so any phi we hoist would be controlled by the
639 // wrong condition. This also takes care of avoiding hoisting of loop back
640 // edges.
641 // TODO: In some cases this could be relaxed if the successor is dominated
642 // by another block that's been hoisted and we can guarantee that the
643 // control flow has been replicated exactly.
644 if (CommonSucc && DT->dominates(BI, CommonSucc))
645 HoistableBranches[BI] = CommonSucc;
646 }
647
648 bool canHoistPHI(PHINode *PN) {
649 // The phi must have loop invariant operands.
650 if (!ControlFlowHoisting || !CurLoop->hasLoopInvariantOperands(PN))
651 return false;
652 // We can hoist phis if the block they are in is the target of hoistable
653 // branches which cover all of the predecessors of the block.
654 SmallPtrSet<BasicBlock *, 8> PredecessorBlocks;
655 BasicBlock *BB = PN->getParent();
656 for (BasicBlock *PredBB : predecessors(BB))
657 PredecessorBlocks.insert(PredBB);
658 // If we have less predecessor blocks than predecessors then the phi will
659 // have more than one incoming value for the same block which we can't
660 // handle.
661 // TODO: This could be handled be erasing some of the duplicate incoming
662 // values.
663 if (PredecessorBlocks.size() != pred_size(BB))
664 return false;
665 for (auto &Pair : HoistableBranches) {
666 if (Pair.second == BB) {
667 // Which blocks are predecessors via this branch depends on if the
668 // branch is triangle-like or diamond-like.
669 if (Pair.first->getSuccessor(0) == BB) {
670 PredecessorBlocks.erase(Pair.first->getParent());
671 PredecessorBlocks.erase(Pair.first->getSuccessor(1));
672 } else if (Pair.first->getSuccessor(1) == BB) {
673 PredecessorBlocks.erase(Pair.first->getParent());
674 PredecessorBlocks.erase(Pair.first->getSuccessor(0));
675 } else {
676 PredecessorBlocks.erase(Pair.first->getSuccessor(0));
677 PredecessorBlocks.erase(Pair.first->getSuccessor(1));
678 }
679 }
680 }
681 // PredecessorBlocks will now be empty if for every predecessor of BB we
682 // found a hoistable branch source.
683 return PredecessorBlocks.empty();
684 }
685
686 BasicBlock *getOrCreateHoistedBlock(BasicBlock *BB) {
687 if (!ControlFlowHoisting)
688 return CurLoop->getLoopPreheader();
689 // If BB has already been hoisted, return that
690 if (HoistDestinationMap.count(BB))
691 return HoistDestinationMap[BB];
692
693 // Check if this block is conditional based on a pending branch
694 auto HasBBAsSuccessor =
695 [&](DenseMap<BranchInst *, BasicBlock *>::value_type &Pair) {
696 return BB != Pair.second && (Pair.first->getSuccessor(0) == BB ||
697 Pair.first->getSuccessor(1) == BB);
698 };
699 auto It = std::find_if(HoistableBranches.begin(), HoistableBranches.end(),
700 HasBBAsSuccessor);
701
702 // If not involved in a pending branch, hoist to preheader
703 BasicBlock *InitialPreheader = CurLoop->getLoopPreheader();
704 if (It == HoistableBranches.end()) {
705 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)
706 << " 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)
707 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "LICM using " << InitialPreheader
->getName() << " as hoist destination for " <<
BB->getName() << "\n"; } } while (false)
;
708 HoistDestinationMap[BB] = InitialPreheader;
709 return InitialPreheader;
710 }
711 BranchInst *BI = It->first;
712 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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 714, __PRETTY_FUNCTION__))
713 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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 714, __PRETTY_FUNCTION__))
714 "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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 714, __PRETTY_FUNCTION__))
;
715
716 LLVMContext &C = BB->getContext();
717 BasicBlock *TrueDest = BI->getSuccessor(0);
718 BasicBlock *FalseDest = BI->getSuccessor(1);
719 BasicBlock *CommonSucc = HoistableBranches[BI];
720 BasicBlock *HoistTarget = getOrCreateHoistedBlock(BI->getParent());
721
722 // Create hoisted versions of blocks that currently don't have them
723 auto CreateHoistedBlock = [&](BasicBlock *Orig) {
724 if (HoistDestinationMap.count(Orig))
725 return HoistDestinationMap[Orig];
726 BasicBlock *New =
727 BasicBlock::Create(C, Orig->getName() + ".licm", Orig->getParent());
728 HoistDestinationMap[Orig] = New;
729 DT->addNewBlock(New, HoistTarget);
730 if (CurLoop->getParentLoop())
731 CurLoop->getParentLoop()->addBasicBlockToLoop(New, *LI);
732 ++NumCreatedBlocks;
733 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)
734 << " 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)
735 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "LICM created " << New->
getName() << " as hoist destination for " << Orig
->getName() << "\n"; } } while (false)
;
736 return New;
737 };
738 BasicBlock *HoistTrueDest = CreateHoistedBlock(TrueDest);
739 BasicBlock *HoistFalseDest = CreateHoistedBlock(FalseDest);
740 BasicBlock *HoistCommonSucc = CreateHoistedBlock(CommonSucc);
741
742 // Link up these blocks with branches.
743 if (!HoistCommonSucc->getTerminator()) {
744 // The new common successor we've generated will branch to whatever that
745 // hoist target branched to.
746 BasicBlock *TargetSucc = HoistTarget->getSingleSuccessor();
747 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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 747, __PRETTY_FUNCTION__))
;
748 HoistCommonSucc->moveBefore(TargetSucc);
749 BranchInst::Create(TargetSucc, HoistCommonSucc);
750 }
751 if (!HoistTrueDest->getTerminator()) {
752 HoistTrueDest->moveBefore(HoistCommonSucc);
753 BranchInst::Create(HoistCommonSucc, HoistTrueDest);
754 }
755 if (!HoistFalseDest->getTerminator()) {
756 HoistFalseDest->moveBefore(HoistCommonSucc);
757 BranchInst::Create(HoistCommonSucc, HoistFalseDest);
758 }
759
760 // If BI is being cloned to what was originally the preheader then
761 // HoistCommonSucc will now be the new preheader.
762 if (HoistTarget == InitialPreheader) {
763 // Phis in the loop header now need to use the new preheader.
764 InitialPreheader->replaceSuccessorsPhiUsesWith(HoistCommonSucc);
765 if (MSSAU)
766 MSSAU->wireOldPredecessorsToNewImmediatePredecessor(
767 HoistTarget->getSingleSuccessor(), HoistCommonSucc, {HoistTarget});
768 // The new preheader dominates the loop header.
769 DomTreeNode *PreheaderNode = DT->getNode(HoistCommonSucc);
770 DomTreeNode *HeaderNode = DT->getNode(CurLoop->getHeader());
771 DT->changeImmediateDominator(HeaderNode, PreheaderNode);
772 // The preheader hoist destination is now the new preheader, with the
773 // exception of the hoist destination of this branch.
774 for (auto &Pair : HoistDestinationMap)
775 if (Pair.second == InitialPreheader && Pair.first != BI->getParent())
776 Pair.second = HoistCommonSucc;
777 }
778
779 // Now finally clone BI.
780 ReplaceInstWithInst(
781 HoistTarget->getTerminator(),
782 BranchInst::Create(HoistTrueDest, HoistFalseDest, BI->getCondition()));
783 ++NumClonedBranches;
784
785 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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 786, __PRETTY_FUNCTION__))
786 "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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 786, __PRETTY_FUNCTION__))
;
787 return HoistDestinationMap[BB];
788 }
789};
790} // namespace
791
792
793/// Return true if we know how to rewrite all uses of the given alloca after
794/// hoisting it out of the loop. The main concerns are a) potential captures
795/// and b) invariant.start markers which don't capture, but are no longer
796/// valid w/o a corresponding invariant.end.
797static bool canRewriteUsesOfAlloca(AllocaInst &AI) {
798 // TODO: This looks a lot like capture tracking, but we need to remove any
799 // invariant starts if we extend the lifetime of the alloca by hoisting it.
800 // We should probably refactor capture tracking into a form which allows us
801 // to reuse the relevant bits and remove the duplicated logic here.
802
803 SmallVector<Use *, 16> Worklist;
804 for (Use &U : AI.uses())
805 Worklist.push_back(&U);
806
807 unsigned NumUsesExplored = 0;
808 while (!Worklist.empty()) {
809 Use *U = Worklist.pop_back_val();
810 Instruction *I = cast<Instruction>(U->getUser());
811 NumUsesExplored++;
812 if (NumUsesExplored > DefaultMaxUsesToExplore)
813 return false;
814 // Non capturing, terminating uses
815 if (isa<LoadInst>(I) ||
816 (isa<StoreInst>(I) && U->getOperandNo() == 1))
817 continue;
818 // Non capturing, non-terminating
819 if (!isa<BitCastInst>(I) && !isa<GetElementPtrInst>(I))
820 return false;
821 for (Use &U : I->uses())
822 Worklist.push_back(&U);
823 }
824 return true;
825}
826
827/// Walk the specified region of the CFG (defined by all blocks dominated by
828/// the specified block, and that are in the current loop) in depth first
829/// order w.r.t the DominatorTree. This allows us to visit definitions before
830/// uses, allowing us to hoist a loop body in one pass without iteration.
831///
832bool llvm::hoistRegion(DomTreeNode *N, AliasAnalysis *AA, LoopInfo *LI,
833 DominatorTree *DT, TargetLibraryInfo *TLI, Loop *CurLoop,
834 AliasSetTracker *CurAST, MemorySSAUpdater *MSSAU,
835 ScalarEvolution *SE, ICFLoopSafetyInfo *SafetyInfo,
836 SinkAndHoistLICMFlags &Flags,
837 OptimizationRemarkEmitter *ORE) {
838 // Verify inputs.
839 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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 841, __PRETTY_FUNCTION__))
840 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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 841, __PRETTY_FUNCTION__))
841 "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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 841, __PRETTY_FUNCTION__))
;
842 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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 843, __PRETTY_FUNCTION__))
843 "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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 843, __PRETTY_FUNCTION__))
;
844
845 ControlFlowHoister CFH(LI, DT, CurLoop, MSSAU);
846
847 // Keep track of instructions that have been hoisted, as they may need to be
848 // re-hoisted if they end up not dominating all of their uses.
849 SmallVector<Instruction *, 16> HoistedInstructions;
850
851 // For PHI hoisting to work we need to hoist blocks before their successors.
852 // We can do this by iterating through the blocks in the loop in reverse
853 // post-order.
854 LoopBlocksRPO Worklist(CurLoop);
855 Worklist.perform(LI);
856 bool Changed = false;
857 for (BasicBlock *BB : Worklist) {
858 // Only need to process the contents of this block if it is not part of a
859 // subloop (which would already have been processed).
860 if (inSubLoop(BB, CurLoop, LI))
861 continue;
862
863 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E;) {
864 Instruction &I = *II++;
865 // Try constant folding this instruction. If all the operands are
866 // constants, it is technically hoistable, but it would be better to
867 // just fold it.
868 if (Constant *C = ConstantFoldInstruction(
869 &I, I.getModule()->getDataLayout(), TLI)) {
870 LLVM_DEBUG(dbgs() << "LICM folding inst: " << I << " --> " << *Cdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "LICM folding inst: " << I <<
" --> " << *C << '\n'; } } while (false)
871 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "LICM folding inst: " << I <<
" --> " << *C << '\n'; } } while (false)
;
872 if (CurAST)
873 CurAST->copyValue(&I, C);
874 // FIXME MSSA: Such replacements may make accesses unoptimized (D51960).
875 I.replaceAllUsesWith(C);
876 if (isInstructionTriviallyDead(&I, TLI))
877 eraseInstruction(I, *SafetyInfo, CurAST, MSSAU);
878 Changed = true;
879 continue;
880 }
881
882 // Try hoisting the instruction out to the preheader. We can only do
883 // this if all of the operands of the instruction are loop invariant and
884 // if it is safe to hoist the instruction.
885 // TODO: It may be safe to hoist if we are hoisting to a conditional block
886 // and we have accurately duplicated the control flow from the loop header
887 // to that block.
888 if (CurLoop->hasLoopInvariantOperands(&I) &&
889 canSinkOrHoistInst(I, AA, DT, CurLoop, CurAST, MSSAU, true, &Flags,
890 ORE) &&
891 isSafeToExecuteUnconditionally(
892 I, DT, CurLoop, SafetyInfo, ORE,
893 CurLoop->getLoopPreheader()->getTerminator())) {
894 hoist(I, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), SafetyInfo,
895 MSSAU, SE, ORE);
896 HoistedInstructions.push_back(&I);
897 Changed = true;
898 continue;
899 }
900
901 // Attempt to remove floating point division out of the loop by
902 // converting it to a reciprocal multiplication.
903 if (I.getOpcode() == Instruction::FDiv &&
904 CurLoop->isLoopInvariant(I.getOperand(1)) &&
905 I.hasAllowReciprocal()) {
906 auto Divisor = I.getOperand(1);
907 auto One = llvm::ConstantFP::get(Divisor->getType(), 1.0);
908 auto ReciprocalDivisor = BinaryOperator::CreateFDiv(One, Divisor);
909 ReciprocalDivisor->setFastMathFlags(I.getFastMathFlags());
910 SafetyInfo->insertInstructionTo(ReciprocalDivisor, I.getParent());
911 ReciprocalDivisor->insertBefore(&I);
912
913 auto Product =
914 BinaryOperator::CreateFMul(I.getOperand(0), ReciprocalDivisor);
915 Product->setFastMathFlags(I.getFastMathFlags());
916 SafetyInfo->insertInstructionTo(Product, I.getParent());
917 Product->insertAfter(&I);
918 I.replaceAllUsesWith(Product);
919 eraseInstruction(I, *SafetyInfo, CurAST, MSSAU);
920
921 hoist(*ReciprocalDivisor, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB),
922 SafetyInfo, MSSAU, SE, ORE);
923 HoistedInstructions.push_back(ReciprocalDivisor);
924 Changed = true;
925 continue;
926 }
927
928 auto IsInvariantStart = [&](Instruction &I) {
929 using namespace PatternMatch;
930 return I.use_empty() &&
931 match(&I, m_Intrinsic<Intrinsic::invariant_start>());
932 };
933 auto MustExecuteWithoutWritesBefore = [&](Instruction &I) {
934 return SafetyInfo->isGuaranteedToExecute(I, DT, CurLoop) &&
935 SafetyInfo->doesNotWriteMemoryBefore(I, CurLoop);
936 };
937 if ((IsInvariantStart(I) || isGuard(&I)) &&
938 CurLoop->hasLoopInvariantOperands(&I) &&
939 MustExecuteWithoutWritesBefore(I)) {
940 hoist(I, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), SafetyInfo,
941 MSSAU, SE, ORE);
942 HoistedInstructions.push_back(&I);
943 Changed = true;
944 continue;
945 }
946
947 if (isa<AllocaInst>(&I) &&
948 SafetyInfo->isGuaranteedToExecute(I, DT, CurLoop) &&
949 canRewriteUsesOfAlloca(cast<AllocaInst>(I))) {
950 hoist(I, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), SafetyInfo,
951 MSSAU, SE, ORE);
952 HoistedInstructions.push_back(&I);
953 Changed = true;
954 continue;
955 }
956
957 if (PHINode *PN = dyn_cast<PHINode>(&I)) {
958 if (CFH.canHoistPHI(PN)) {
959 // Redirect incoming blocks first to ensure that we create hoisted
960 // versions of those blocks before we hoist the phi.
961 for (unsigned int i = 0; i < PN->getNumIncomingValues(); ++i)
962 PN->setIncomingBlock(
963 i, CFH.getOrCreateHoistedBlock(PN->getIncomingBlock(i)));
964 hoist(*PN, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), SafetyInfo,
965 MSSAU, SE, ORE);
966 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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 966, __PRETTY_FUNCTION__))
;
967 Changed = true;
968 continue;
969 }
970 }
971
972 // Remember possibly hoistable branches so we can actually hoist them
973 // later if needed.
974 if (BranchInst *BI = dyn_cast<BranchInst>(&I))
975 CFH.registerPossiblyHoistableBranch(BI);
976 }
977 }
978
979 // If we hoisted instructions to a conditional block they may not dominate
980 // their uses that weren't hoisted (such as phis where some operands are not
981 // loop invariant). If so make them unconditional by moving them to their
982 // immediate dominator. We iterate through the instructions in reverse order
983 // which ensures that when we rehoist an instruction we rehoist its operands,
984 // and also keep track of where in the block we are rehoisting to to make sure
985 // that we rehoist instructions before the instructions that use them.
986 Instruction *HoistPoint = nullptr;
987 if (ControlFlowHoisting) {
988 for (Instruction *I : reverse(HoistedInstructions)) {
989 if (!llvm::all_of(I->uses(),
990 [&](Use &U) { return DT->dominates(I, U); })) {
991 BasicBlock *Dominator =
992 DT->getNode(I->getParent())->getIDom()->getBlock();
993 if (!HoistPoint || !DT->dominates(HoistPoint->getParent(), Dominator)) {
994 if (HoistPoint)
995 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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 996, __PRETTY_FUNCTION__))
996 "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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 996, __PRETTY_FUNCTION__))
;
997 HoistPoint = Dominator->getTerminator();
998 }
999 LLVM_DEBUG(dbgs() << "LICM rehoisting to "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "LICM rehoisting to " << HoistPoint
->getParent()->getName() << ": " << *I <<
"\n"; } } while (false)
1000 << HoistPoint->getParent()->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "LICM rehoisting to " << HoistPoint
->getParent()->getName() << ": " << *I <<
"\n"; } } while (false)
1001 << ": " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "LICM rehoisting to " << HoistPoint
->getParent()->getName() << ": " << *I <<
"\n"; } } while (false)
;
1002 moveInstructionBefore(*I, *HoistPoint, *SafetyInfo, MSSAU, SE);
1003 HoistPoint = I;
1004 Changed = true;
1005 }
1006 }
1007 }
1008 if (MSSAU && VerifyMemorySSA)
1009 MSSAU->getMemorySSA()->verifyMemorySSA();
1010
1011 // Now that we've finished hoisting make sure that LI and DT are still
1012 // valid.
1013#ifdef EXPENSIVE_CHECKS
1014 if (Changed) {
1015 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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1016, __PRETTY_FUNCTION__))
1016 "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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1016, __PRETTY_FUNCTION__))
;
1017 LI->verify(*DT);
1018 }
1019#endif
1020
1021 return Changed;
1022}
1023
1024// Return true if LI is invariant within scope of the loop. LI is invariant if
1025// CurLoop is dominated by an invariant.start representing the same memory
1026// location and size as the memory location LI loads from, and also the
1027// invariant.start has no uses.
1028static bool isLoadInvariantInLoop(LoadInst *LI, DominatorTree *DT,
1029 Loop *CurLoop) {
1030 Value *Addr = LI->getOperand(0);
1031 const DataLayout &DL = LI->getModule()->getDataLayout();
1032 const uint32_t LocSizeInBits = DL.getTypeSizeInBits(LI->getType());
1033
1034 // if the type is i8 addrspace(x)*, we know this is the type of
1035 // llvm.invariant.start operand
1036 auto *PtrInt8Ty = PointerType::get(Type::getInt8Ty(LI->getContext()),
1037 LI->getPointerAddressSpace());
1038 unsigned BitcastsVisited = 0;
1039 // Look through bitcasts until we reach the i8* type (this is invariant.start
1040 // operand type).
1041 while (Addr->getType() != PtrInt8Ty) {
1042 auto *BC = dyn_cast<BitCastInst>(Addr);
1043 // Avoid traversing high number of bitcast uses.
1044 if (++BitcastsVisited > MaxNumUsesTraversed || !BC)
1045 return false;
1046 Addr = BC->getOperand(0);
1047 }
1048
1049 unsigned UsesVisited = 0;
1050 // Traverse all uses of the load operand value, to see if invariant.start is
1051 // one of the uses, and whether it dominates the load instruction.
1052 for (auto *U : Addr->users()) {
1053 // Avoid traversing for Load operand with high number of users.
1054 if (++UsesVisited > MaxNumUsesTraversed)
1055 return false;
1056 IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
1057 // If there are escaping uses of invariant.start instruction, the load maybe
1058 // non-invariant.
1059 if (!II || II->getIntrinsicID() != Intrinsic::invariant_start ||
1060 !II->use_empty())
1061 continue;
1062 unsigned InvariantSizeInBits =
1063 cast<ConstantInt>(II->getArgOperand(0))->getSExtValue() * 8;
1064 // Confirm the invariant.start location size contains the load operand size
1065 // in bits. Also, the invariant.start should dominate the load, and we
1066 // should not hoist the load out of a loop that contains this dominating
1067 // invariant.start.
1068 if (LocSizeInBits <= InvariantSizeInBits &&
1069 DT->properlyDominates(II->getParent(), CurLoop->getHeader()))
1070 return true;
1071 }
1072
1073 return false;
1074}
1075
1076namespace {
1077/// Return true if-and-only-if we know how to (mechanically) both hoist and
1078/// sink a given instruction out of a loop. Does not address legality
1079/// concerns such as aliasing or speculation safety.
1080bool isHoistableAndSinkableInst(Instruction &I) {
1081 // Only these instructions are hoistable/sinkable.
1082 return (isa<LoadInst>(I) || isa<StoreInst>(I) || isa<CallInst>(I) ||
2
Assuming 'I' is not a 'LoadInst'
3
Assuming 'I' is not a 'StoreInst'
4
Assuming 'I' is not a 'CallInst'
15
Returning the value 1, which participates in a condition later
1083 isa<FenceInst>(I) || isa<CastInst>(I) ||
5
Assuming 'I' is not a 'FenceInst'
6
Assuming 'I' is not a 'CastInst'
1084 isa<UnaryOperator>(I) || isa<BinaryOperator>(I) ||
7
Assuming 'I' is not a 'UnaryOperator'
8
Assuming 'I' is not a 'BinaryOperator'
1085 isa<SelectInst>(I) || isa<GetElementPtrInst>(I) || isa<CmpInst>(I) ||
9
Assuming 'I' is not a 'SelectInst'
10
Assuming 'I' is not a 'GetElementPtrInst'
11
Assuming 'I' is not a 'CmpInst'
1086 isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
12
Assuming 'I' is not a 'InsertElementInst'
13
Assuming 'I' is not a 'ExtractElementInst'
1087 isa<ShuffleVectorInst>(I) || isa<ExtractValueInst>(I) ||
14
Assuming 'I' is a 'ShuffleVectorInst'
1088 isa<InsertValueInst>(I));
1089}
1090/// Return true if all of the alias sets within this AST are known not to
1091/// contain a Mod, or if MSSA knows thare are no MemoryDefs in the loop.
1092bool isReadOnly(AliasSetTracker *CurAST, const MemorySSAUpdater *MSSAU,
1093 const Loop *L) {
1094 if (CurAST) {
1095 for (AliasSet &AS : *CurAST) {
1096 if (!AS.isForwardingAliasSet() && AS.isMod()) {
1097 return false;
1098 }
1099 }
1100 return true;
1101 } else { /*MSSAU*/
1102 for (auto *BB : L->getBlocks())
1103 if (MSSAU->getMemorySSA()->getBlockDefs(BB))
1104 return false;
1105 return true;
1106 }
1107}
1108
1109/// Return true if I is the only Instruction with a MemoryAccess in L.
1110bool isOnlyMemoryAccess(const Instruction *I, const Loop *L,
1111 const MemorySSAUpdater *MSSAU) {
1112 for (auto *BB : L->getBlocks())
31
Assuming '__begin1' is not equal to '__end1'
1113 if (auto *Accs = MSSAU->getMemorySSA()->getBlockAccesses(BB)) {
32
Called C++ object pointer is null
1114 int NotAPhi = 0;
1115 for (const auto &Acc : *Accs) {
1116 if (isa<MemoryPhi>(&Acc))
1117 continue;
1118 const auto *MUD = cast<MemoryUseOrDef>(&Acc);
1119 if (MUD->getMemoryInst() != I || NotAPhi++ == 1)
1120 return false;
1121 }
1122 }
1123 return true;
1124}
1125}
1126
1127bool llvm::canSinkOrHoistInst(Instruction &I, AAResults *AA, DominatorTree *DT,
1128 Loop *CurLoop, AliasSetTracker *CurAST,
1129 MemorySSAUpdater *MSSAU,
1130 bool TargetExecutesOncePerLoop,
1131 SinkAndHoistLICMFlags *Flags,
1132 OptimizationRemarkEmitter *ORE) {
1133 // If we don't understand the instruction, bail early.
1134 if (!isHoistableAndSinkableInst(I))
1
Calling 'isHoistableAndSinkableInst'
16
Returning from 'isHoistableAndSinkableInst'
17
Taking false branch
1135 return false;
1136
1137 MemorySSA *MSSA = MSSAU ? MSSAU->getMemorySSA() : nullptr;
18
Assuming 'MSSAU' is null
19
'?' condition is false
1138 if (MSSA
19.1
'MSSA' is null
)
20
Taking false branch
1139 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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1139, __PRETTY_FUNCTION__))
;
1140
1141 // Loads have extra constraints we have to verify before we can hoist them.
1142 if (LoadInst *LI
21.1
'LI' is null
= dyn_cast<LoadInst>(&I)) {
21
Assuming the object is not a 'LoadInst'
22
Taking false branch
1143 if (!LI->isUnordered())
1144 return false; // Don't sink/hoist volatile or ordered atomic loads!
1145
1146 // Loads from constant memory are always safe to move, even if they end up
1147 // in the same alias set as something that ends up being modified.
1148 if (AA->pointsToConstantMemory(LI->getOperand(0)))
1149 return true;
1150 if (LI->hasMetadata(LLVMContext::MD_invariant_load))
1151 return true;
1152
1153 if (LI->isAtomic() && !TargetExecutesOncePerLoop)
1154 return false; // Don't risk duplicating unordered loads
1155
1156 // This checks for an invariant.start dominating the load.
1157 if (isLoadInvariantInLoop(LI, DT, CurLoop))
1158 return true;
1159
1160 bool Invalidated;
1161 if (CurAST)
1162 Invalidated = pointerInvalidatedByLoop(MemoryLocation::get(LI), CurAST,
1163 CurLoop, AA);
1164 else
1165 Invalidated = pointerInvalidatedByLoopWithMSSA(
1166 MSSA, cast<MemoryUse>(MSSA->getMemoryAccess(LI)), CurLoop, *Flags);
1167 // Check loop-invariant address because this may also be a sinkable load
1168 // whose address is not necessarily loop-invariant.
1169 if (ORE && Invalidated && CurLoop->isLoopInvariant(LI->getPointerOperand()))
1170 ORE->emit([&]() {
1171 return OptimizationRemarkMissed(
1172 DEBUG_TYPE"licm", "LoadWithLoopInvariantAddressInvalidated", LI)
1173 << "failed to move load with loop-invariant address "
1174 "because the loop may invalidate its value";
1175 });
1176
1177 return !Invalidated;
1178 } else if (CallInst *CI
23.1
'CI' is null
= dyn_cast<CallInst>(&I)) {
23
Assuming the object is not a 'CallInst'
24
Taking false branch
1179 // Don't sink or hoist dbg info; it's legal, but not useful.
1180 if (isa<DbgInfoIntrinsic>(I))
1181 return false;
1182
1183 // Don't sink calls which can throw.
1184 if (CI->mayThrow())
1185 return false;
1186
1187 using namespace PatternMatch;
1188 if (match(CI, m_Intrinsic<Intrinsic::assume>()))
1189 // Assumes don't actually alias anything or throw
1190 return true;
1191
1192 if (match(CI, m_Intrinsic<Intrinsic::experimental_widenable_condition>()))
1193 // Widenable conditions don't actually alias anything or throw
1194 return true;
1195
1196 // Handle simple cases by querying alias analysis.
1197 FunctionModRefBehavior Behavior = AA->getModRefBehavior(CI);
1198 if (Behavior == FMRB_DoesNotAccessMemory)
1199 return true;
1200 if (AliasAnalysis::onlyReadsMemory(Behavior)) {
1201 // A readonly argmemonly function only reads from memory pointed to by
1202 // it's arguments with arbitrary offsets. If we can prove there are no
1203 // writes to this memory in the loop, we can hoist or sink.
1204 if (AliasAnalysis::onlyAccessesArgPointees(Behavior)) {
1205 // TODO: expand to writeable arguments
1206 for (Value *Op : CI->arg_operands())
1207 if (Op->getType()->isPointerTy()) {
1208 bool Invalidated;
1209 if (CurAST)
1210 Invalidated = pointerInvalidatedByLoop(
1211 MemoryLocation(Op, LocationSize::unknown(), AAMDNodes()),
1212 CurAST, CurLoop, AA);
1213 else
1214 Invalidated = pointerInvalidatedByLoopWithMSSA(
1215 MSSA, cast<MemoryUse>(MSSA->getMemoryAccess(CI)), CurLoop,
1216 *Flags);
1217 if (Invalidated)
1218 return false;
1219 }
1220 return true;
1221 }
1222
1223 // If this call only reads from memory and there are no writes to memory
1224 // in the loop, we can hoist or sink the call as appropriate.
1225 if (isReadOnly(CurAST, MSSAU, CurLoop))
1226 return true;
1227 }
1228
1229 // FIXME: This should use mod/ref information to see if we can hoist or
1230 // sink the call.
1231
1232 return false;
1233 } else if (auto *FI
25.1
'FI' is non-null
= dyn_cast<FenceInst>(&I)) {
25
Assuming the object is a 'FenceInst'
26
Taking true branch
1234 // Fences alias (most) everything to provide ordering. For the moment,
1235 // just give up if there are any other memory operations in the loop.
1236 if (CurAST) {
27
Assuming 'CurAST' is null
28
Taking false branch
1237 auto Begin = CurAST->begin();
1238 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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1238, __PRETTY_FUNCTION__))
;
1239 if (std::next(Begin) != CurAST->end())
1240 // constant memory for instance, TODO: handle better
1241 return false;
1242 auto *UniqueI = Begin->getUniqueInstruction();
1243 if (!UniqueI)
1244 // other memory op, give up
1245 return false;
1246 (void)FI; // suppress unused variable warning
1247 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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1247, __PRETTY_FUNCTION__))
;
1248 return true;
1249 } else // MSSAU
1250 return isOnlyMemoryAccess(FI, CurLoop, MSSAU);
29
Passing null pointer value via 3rd parameter 'MSSAU'
30
Calling 'isOnlyMemoryAccess'
1251 } else if (auto *SI = dyn_cast<StoreInst>(&I)) {
1252 if (!SI->isUnordered())
1253 return false; // Don't sink/hoist volatile or ordered atomic store!
1254
1255 // We can only hoist a store that we can prove writes a value which is not
1256 // read or overwritten within the loop. For those cases, we fallback to
1257 // load store promotion instead. TODO: We can extend this to cases where
1258 // there is exactly one write to the location and that write dominates an
1259 // arbitrary number of reads in the loop.
1260 if (CurAST) {
1261 auto &AS = CurAST->getAliasSetFor(MemoryLocation::get(SI));
1262
1263 if (AS.isRef() || !AS.isMustAlias())
1264 // Quick exit test, handled by the full path below as well.
1265 return false;
1266 auto *UniqueI = AS.getUniqueInstruction();
1267 if (!UniqueI)
1268 // other memory op, give up
1269 return false;
1270 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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1270, __PRETTY_FUNCTION__))
;
1271 return true;
1272 } else { // MSSAU
1273 if (isOnlyMemoryAccess(SI, CurLoop, MSSAU))
1274 return true;
1275 // If there are more accesses than the Promotion cap, give up, we're not
1276 // walking a list that long.
1277 if (Flags->NoOfMemAccTooLarge)
1278 return false;
1279 // Check store only if there's still "quota" to check clobber.
1280 if (Flags->LicmMssaOptCounter >= Flags->LicmMssaOptCap)
1281 return false;
1282 // If there are interfering Uses (i.e. their defining access is in the
1283 // loop), or ordered loads (stored as Defs!), don't move this store.
1284 // Could do better here, but this is conservatively correct.
1285 // TODO: Cache set of Uses on the first walk in runOnLoop, update when
1286 // moving accesses. Can also extend to dominating uses.
1287 auto *SIMD = MSSA->getMemoryAccess(SI);
1288 for (auto *BB : CurLoop->getBlocks())
1289 if (auto *Accesses = MSSA->getBlockAccesses(BB)) {
1290 for (const auto &MA : *Accesses)
1291 if (const auto *MU = dyn_cast<MemoryUse>(&MA)) {
1292 auto *MD = MU->getDefiningAccess();
1293 if (!MSSA->isLiveOnEntryDef(MD) &&
1294 CurLoop->contains(MD->getBlock()))
1295 return false;
1296 // Disable hoisting past potentially interfering loads. Optimized
1297 // Uses may point to an access outside the loop, as getClobbering
1298 // checks the previous iteration when walking the backedge.
1299 // FIXME: More precise: no Uses that alias SI.
1300 if (!Flags->IsSink && !MSSA->dominates(SIMD, MU))
1301 return false;
1302 } else if (const auto *MD = dyn_cast<MemoryDef>(&MA)) {
1303 if (auto *LI = dyn_cast<LoadInst>(MD->getMemoryInst())) {
1304 (void)LI; // Silence warning.
1305 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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1305, __PRETTY_FUNCTION__))
;
1306 return false;
1307 }
1308 // Any call, while it may not be clobbering SI, it may be a use.
1309 if (auto *CI = dyn_cast<CallInst>(MD->getMemoryInst())) {
1310 // Check if the call may read from the memory locattion written
1311 // to by SI. Check CI's attributes and arguments; the number of
1312 // such checks performed is limited above by NoOfMemAccTooLarge.
1313 ModRefInfo MRI = AA->getModRefInfo(CI, MemoryLocation::get(SI));
1314 if (isModOrRefSet(MRI))
1315 return false;
1316 }
1317 }
1318 }
1319
1320 auto *Source = MSSA->getSkipSelfWalker()->getClobberingMemoryAccess(SI);
1321 Flags->LicmMssaOptCounter++;
1322 // If there are no clobbering Defs in the loop, store is safe to hoist.
1323 return MSSA->isLiveOnEntryDef(Source) ||
1324 !CurLoop->contains(Source->getBlock());
1325 }
1326 }
1327
1328 assert(!I.mayReadOrWriteMemory() && "unhandled aliasing")((!I.mayReadOrWriteMemory() && "unhandled aliasing") ?
static_cast<void> (0) : __assert_fail ("!I.mayReadOrWriteMemory() && \"unhandled aliasing\""
, "/build/llvm-toolchain-snapshot-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1328, __PRETTY_FUNCTION__))
;
1329
1330 // We've established mechanical ability and aliasing, it's up to the caller
1331 // to check fault safety
1332 return true;
1333}
1334
1335/// Returns true if a PHINode is a trivially replaceable with an
1336/// Instruction.
1337/// This is true when all incoming values are that instruction.
1338/// This pattern occurs most often with LCSSA PHI nodes.
1339///
1340static bool isTriviallyReplaceablePHI(const PHINode &PN, const Instruction &I) {
1341 for (const Value *IncValue : PN.incoming_values())
1342 if (IncValue != &I)
1343 return false;
1344
1345 return true;
1346}
1347
1348/// Return true if the instruction is free in the loop.
1349static bool isFreeInLoop(const Instruction &I, const Loop *CurLoop,
1350 const TargetTransformInfo *TTI) {
1351
1352 if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(&I)) {
1353 if (TTI->getUserCost(GEP) != TargetTransformInfo::TCC_Free)
1354 return false;
1355 // For a GEP, we cannot simply use getUserCost because currently it
1356 // optimistically assume that a GEP will fold into addressing mode
1357 // regardless of its users.
1358 const BasicBlock *BB = GEP->getParent();
1359 for (const User *U : GEP->users()) {
1360 const Instruction *UI = cast<Instruction>(U);
1361 if (CurLoop->contains(UI) &&
1362 (BB != UI->getParent() ||
1363 (!isa<StoreInst>(UI) && !isa<LoadInst>(UI))))
1364 return false;
1365 }
1366 return true;
1367 } else
1368 return TTI->getUserCost(&I) == TargetTransformInfo::TCC_Free;
1369}
1370
1371/// Return true if the only users of this instruction are outside of
1372/// the loop. If this is true, we can sink the instruction to the exit
1373/// blocks of the loop.
1374///
1375/// We also return true if the instruction could be folded away in lowering.
1376/// (e.g., a GEP can be folded into a load as an addressing mode in the loop).
1377static bool isNotUsedOrFreeInLoop(const Instruction &I, const Loop *CurLoop,
1378 const LoopSafetyInfo *SafetyInfo,
1379 TargetTransformInfo *TTI, bool &FreeInLoop) {
1380 const auto &BlockColors = SafetyInfo->getBlockColors();
1381 bool IsFree = isFreeInLoop(I, CurLoop, TTI);
1382 for (const User *U : I.users()) {
1383 const Instruction *UI = cast<Instruction>(U);
1384 if (const PHINode *PN = dyn_cast<PHINode>(UI)) {
1385 const BasicBlock *BB = PN->getParent();
1386 // We cannot sink uses in catchswitches.
1387 if (isa<CatchSwitchInst>(BB->getTerminator()))
1388 return false;
1389
1390 // We need to sink a callsite to a unique funclet. Avoid sinking if the
1391 // phi use is too muddled.
1392 if (isa<CallInst>(I))
1393 if (!BlockColors.empty() &&
1394 BlockColors.find(const_cast<BasicBlock *>(BB))->second.size() != 1)
1395 return false;
1396 }
1397
1398 if (CurLoop->contains(UI)) {
1399 if (IsFree) {
1400 FreeInLoop = true;
1401 continue;
1402 }
1403 return false;
1404 }
1405 }
1406 return true;
1407}
1408
1409static Instruction *CloneInstructionInExitBlock(
1410 Instruction &I, BasicBlock &ExitBlock, PHINode &PN, const LoopInfo *LI,
1411 const LoopSafetyInfo *SafetyInfo, MemorySSAUpdater *MSSAU) {
1412 Instruction *New;
1413 if (auto *CI = dyn_cast<CallInst>(&I)) {
1414 const auto &BlockColors = SafetyInfo->getBlockColors();
1415
1416 // Sinking call-sites need to be handled differently from other
1417 // instructions. The cloned call-site needs a funclet bundle operand
1418 // appropriate for its location in the CFG.
1419 SmallVector<OperandBundleDef, 1> OpBundles;
1420 for (unsigned BundleIdx = 0, BundleEnd = CI->getNumOperandBundles();
1421 BundleIdx != BundleEnd; ++BundleIdx) {
1422 OperandBundleUse Bundle = CI->getOperandBundleAt(BundleIdx);
1423 if (Bundle.getTagID() == LLVMContext::OB_funclet)
1424 continue;
1425
1426 OpBundles.emplace_back(Bundle);
1427 }
1428
1429 if (!BlockColors.empty()) {
1430 const ColorVector &CV = BlockColors.find(&ExitBlock)->second;
1431 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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1431, __PRETTY_FUNCTION__))
;
1432 BasicBlock *BBColor = CV.front();
1433 Instruction *EHPad = BBColor->getFirstNonPHI();
1434 if (EHPad->isEHPad())
1435 OpBundles.emplace_back("funclet", EHPad);
1436 }
1437
1438 New = CallInst::Create(CI, OpBundles);
1439 } else {
1440 New = I.clone();
1441 }
1442
1443 ExitBlock.getInstList().insert(ExitBlock.getFirstInsertionPt(), New);
1444 if (!I.getName().empty())
1445 New->setName(I.getName() + ".le");
1446
1447 if (MSSAU && MSSAU->getMemorySSA()->getMemoryAccess(&I)) {
1448 // Create a new MemoryAccess and let MemorySSA set its defining access.
1449 MemoryAccess *NewMemAcc = MSSAU->createMemoryAccessInBB(
1450 New, nullptr, New->getParent(), MemorySSA::Beginning);
1451 if (NewMemAcc) {
1452 if (auto *MemDef = dyn_cast<MemoryDef>(NewMemAcc))
1453 MSSAU->insertDef(MemDef, /*RenameUses=*/true);
1454 else {
1455 auto *MemUse = cast<MemoryUse>(NewMemAcc);
1456 MSSAU->insertUse(MemUse, /*RenameUses=*/true);
1457 }
1458 }
1459 }
1460
1461 // Build LCSSA PHI nodes for any in-loop operands. Note that this is
1462 // particularly cheap because we can rip off the PHI node that we're
1463 // replacing for the number and blocks of the predecessors.
1464 // OPT: If this shows up in a profile, we can instead finish sinking all
1465 // invariant instructions, and then walk their operands to re-establish
1466 // LCSSA. That will eliminate creating PHI nodes just to nuke them when
1467 // sinking bottom-up.
1468 for (User::op_iterator OI = New->op_begin(), OE = New->op_end(); OI != OE;
1469 ++OI)
1470 if (Instruction *OInst = dyn_cast<Instruction>(*OI))
1471 if (Loop *OLoop = LI->getLoopFor(OInst->getParent()))
1472 if (!OLoop->contains(&PN)) {
1473 PHINode *OpPN =
1474 PHINode::Create(OInst->getType(), PN.getNumIncomingValues(),
1475 OInst->getName() + ".lcssa", &ExitBlock.front());
1476 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
1477 OpPN->addIncoming(OInst, PN.getIncomingBlock(i));
1478 *OI = OpPN;
1479 }
1480 return New;
1481}
1482
1483static void eraseInstruction(Instruction &I, ICFLoopSafetyInfo &SafetyInfo,
1484 AliasSetTracker *AST, MemorySSAUpdater *MSSAU) {
1485 if (AST)
1486 AST->deleteValue(&I);
1487 if (MSSAU)
1488 MSSAU->removeMemoryAccess(&I);
1489 SafetyInfo.removeInstruction(&I);
1490 I.eraseFromParent();
1491}
1492
1493static void moveInstructionBefore(Instruction &I, Instruction &Dest,
1494 ICFLoopSafetyInfo &SafetyInfo,
1495 MemorySSAUpdater *MSSAU,
1496 ScalarEvolution *SE) {
1497 SafetyInfo.removeInstruction(&I);
1498 SafetyInfo.insertInstructionTo(&I, Dest.getParent());
1499 I.moveBefore(&Dest);
1500 if (MSSAU)
1501 if (MemoryUseOrDef *OldMemAcc = cast_or_null<MemoryUseOrDef>(
1502 MSSAU->getMemorySSA()->getMemoryAccess(&I)))
1503 MSSAU->moveToPlace(OldMemAcc, Dest.getParent(), MemorySSA::End);
1504 if (SE)
1505 SE->forgetValue(&I);
1506}
1507
1508static Instruction *sinkThroughTriviallyReplaceablePHI(
1509 PHINode *TPN, Instruction *I, LoopInfo *LI,
1510 SmallDenseMap<BasicBlock *, Instruction *, 32> &SunkCopies,
1511 const LoopSafetyInfo *SafetyInfo, const Loop *CurLoop,
1512 MemorySSAUpdater *MSSAU) {
1513 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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1514, __PRETTY_FUNCTION__))
1514 "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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1514, __PRETTY_FUNCTION__))
;
1515 BasicBlock *ExitBlock = TPN->getParent();
1516 Instruction *New;
1517 auto It = SunkCopies.find(ExitBlock);
1518 if (It != SunkCopies.end())
1519 New = It->second;
1520 else
1521 New = SunkCopies[ExitBlock] = CloneInstructionInExitBlock(
1522 *I, *ExitBlock, *TPN, LI, SafetyInfo, MSSAU);
1523 return New;
1524}
1525
1526static bool canSplitPredecessors(PHINode *PN, LoopSafetyInfo *SafetyInfo) {
1527 BasicBlock *BB = PN->getParent();
1528 if (!BB->canSplitPredecessors())
1529 return false;
1530 // It's not impossible to split EHPad blocks, but if BlockColors already exist
1531 // it require updating BlockColors for all offspring blocks accordingly. By
1532 // skipping such corner case, we can make updating BlockColors after splitting
1533 // predecessor fairly simple.
1534 if (!SafetyInfo->getBlockColors().empty() && BB->getFirstNonPHI()->isEHPad())
1535 return false;
1536 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
1537 BasicBlock *BBPred = *PI;
1538 if (isa<IndirectBrInst>(BBPred->getTerminator()))
1539 return false;
1540 }
1541 return true;
1542}
1543
1544static void splitPredecessorsOfLoopExit(PHINode *PN, DominatorTree *DT,
1545 LoopInfo *LI, const Loop *CurLoop,
1546 LoopSafetyInfo *SafetyInfo,
1547 MemorySSAUpdater *MSSAU) {
1548#ifndef NDEBUG
1549 SmallVector<BasicBlock *, 32> ExitBlocks;
1550 CurLoop->getUniqueExitBlocks(ExitBlocks);
1551 SmallPtrSet<BasicBlock *, 32> ExitBlockSet(ExitBlocks.begin(),
1552 ExitBlocks.end());
1553#endif
1554 BasicBlock *ExitBB = PN->getParent();
1555 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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1555, __PRETTY_FUNCTION__))
;
1556
1557 // Split predecessors of the loop exit to make instructions in the loop are
1558 // exposed to exit blocks through trivially replaceable PHIs while keeping the
1559 // loop in the canonical form where each predecessor of each exit block should
1560 // be contained within the loop. For example, this will convert the loop below
1561 // from
1562 //
1563 // LB1:
1564 // %v1 =
1565 // br %LE, %LB2
1566 // LB2:
1567 // %v2 =
1568 // br %LE, %LB1
1569 // LE:
1570 // %p = phi [%v1, %LB1], [%v2, %LB2] <-- non-trivially replaceable
1571 //
1572 // to
1573 //
1574 // LB1:
1575 // %v1 =
1576 // br %LE.split, %LB2
1577 // LB2:
1578 // %v2 =
1579 // br %LE.split2, %LB1
1580 // LE.split:
1581 // %p1 = phi [%v1, %LB1] <-- trivially replaceable
1582 // br %LE
1583 // LE.split2:
1584 // %p2 = phi [%v2, %LB2] <-- trivially replaceable
1585 // br %LE
1586 // LE:
1587 // %p = phi [%p1, %LE.split], [%p2, %LE.split2]
1588 //
1589 const auto &BlockColors = SafetyInfo->getBlockColors();
1590 SmallSetVector<BasicBlock *, 8> PredBBs(pred_begin(ExitBB), pred_end(ExitBB));
1591 while (!PredBBs.empty()) {
1592 BasicBlock *PredBB = *PredBBs.begin();
1593 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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1594, __PRETTY_FUNCTION__))
1594 "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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1594, __PRETTY_FUNCTION__))
;
1595 if (PN->getBasicBlockIndex(PredBB) >= 0) {
1596 BasicBlock *NewPred = SplitBlockPredecessors(
1597 ExitBB, PredBB, ".split.loop.exit", DT, LI, MSSAU, true);
1598 // Since we do not allow splitting EH-block with BlockColors in
1599 // canSplitPredecessors(), we can simply assign predecessor's color to
1600 // the new block.
1601 if (!BlockColors.empty())
1602 // Grab a reference to the ColorVector to be inserted before getting the
1603 // reference to the vector we are copying because inserting the new
1604 // element in BlockColors might cause the map to be reallocated.
1605 SafetyInfo->copyColors(NewPred, PredBB);
1606 }
1607 PredBBs.remove(PredBB);
1608 }
1609}
1610
1611/// When an instruction is found to only be used outside of the loop, this
1612/// function moves it to the exit blocks and patches up SSA form as needed.
1613/// This method is guaranteed to remove the original instruction from its
1614/// position, and may either delete it or move it to outside of the loop.
1615///
1616static bool sink(Instruction &I, LoopInfo *LI, DominatorTree *DT,
1617 const Loop *CurLoop, ICFLoopSafetyInfo *SafetyInfo,
1618 MemorySSAUpdater *MSSAU, OptimizationRemarkEmitter *ORE) {
1619 LLVM_DEBUG(dbgs() << "LICM sinking instruction: " << I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "LICM sinking instruction: " <<
I << "\n"; } } while (false)
;
1620 ORE->emit([&]() {
1621 return OptimizationRemark(DEBUG_TYPE"licm", "InstSunk", &I)
1622 << "sinking " << ore::NV("Inst", &I);
1623 });
1624 bool Changed = false;
1625 if (isa<LoadInst>(I))
1626 ++NumMovedLoads;
1627 else if (isa<CallInst>(I))
1628 ++NumMovedCalls;
1629 ++NumSunk;
1630
1631 // Iterate over users to be ready for actual sinking. Replace users via
1632 // unreachable blocks with undef and make all user PHIs trivially replaceable.
1633 SmallPtrSet<Instruction *, 8> VisitedUsers;
1634 for (Value::user_iterator UI = I.user_begin(), UE = I.user_end(); UI != UE;) {
1635 auto *User = cast<Instruction>(*UI);
1636 Use &U = UI.getUse();
1637 ++UI;
1638
1639 if (VisitedUsers.count(User) || CurLoop->contains(User))
1640 continue;
1641
1642 if (!DT->isReachableFromEntry(User->getParent())) {
1643 U = UndefValue::get(I.getType());
1644 Changed = true;
1645 continue;
1646 }
1647
1648 // The user must be a PHI node.
1649 PHINode *PN = cast<PHINode>(User);
1650
1651 // Surprisingly, instructions can be used outside of loops without any
1652 // exits. This can only happen in PHI nodes if the incoming block is
1653 // unreachable.
1654 BasicBlock *BB = PN->getIncomingBlock(U);
1655 if (!DT->isReachableFromEntry(BB)) {
1656 U = UndefValue::get(I.getType());
1657 Changed = true;
1658 continue;
1659 }
1660
1661 VisitedUsers.insert(PN);
1662 if (isTriviallyReplaceablePHI(*PN, I))
1663 continue;
1664
1665 if (!canSplitPredecessors(PN, SafetyInfo))
1666 return Changed;
1667
1668 // Split predecessors of the PHI so that we can make users trivially
1669 // replaceable.
1670 splitPredecessorsOfLoopExit(PN, DT, LI, CurLoop, SafetyInfo, MSSAU);
1671
1672 // Should rebuild the iterators, as they may be invalidated by
1673 // splitPredecessorsOfLoopExit().
1674 UI = I.user_begin();
1675 UE = I.user_end();
1676 }
1677
1678 if (VisitedUsers.empty())
1679 return Changed;
1680
1681#ifndef NDEBUG
1682 SmallVector<BasicBlock *, 32> ExitBlocks;
1683 CurLoop->getUniqueExitBlocks(ExitBlocks);
1684 SmallPtrSet<BasicBlock *, 32> ExitBlockSet(ExitBlocks.begin(),
1685 ExitBlocks.end());
1686#endif
1687
1688 // Clones of this instruction. Don't create more than one per exit block!
1689 SmallDenseMap<BasicBlock *, Instruction *, 32> SunkCopies;
1690
1691 // If this instruction is only used outside of the loop, then all users are
1692 // PHI nodes in exit blocks due to LCSSA form. Just RAUW them with clones of
1693 // the instruction.
1694 SmallSetVector<User*, 8> Users(I.user_begin(), I.user_end());
1695 for (auto *UI : Users) {
1696 auto *User = cast<Instruction>(UI);
1697
1698 if (CurLoop->contains(User))
1699 continue;
1700
1701 PHINode *PN = cast<PHINode>(User);
1702 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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1703, __PRETTY_FUNCTION__))
1703 "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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1703, __PRETTY_FUNCTION__))
;
1704 // The PHI must be trivially replaceable.
1705 Instruction *New = sinkThroughTriviallyReplaceablePHI(
1706 PN, &I, LI, SunkCopies, SafetyInfo, CurLoop, MSSAU);
1707 PN->replaceAllUsesWith(New);
1708 eraseInstruction(*PN, *SafetyInfo, nullptr, nullptr);
1709 Changed = true;
1710 }
1711 return Changed;
1712}
1713
1714/// When an instruction is found to only use loop invariant operands that
1715/// is safe to hoist, this instruction is called to do the dirty work.
1716///
1717static void hoist(Instruction &I, const DominatorTree *DT, const Loop *CurLoop,
1718 BasicBlock *Dest, ICFLoopSafetyInfo *SafetyInfo,
1719 MemorySSAUpdater *MSSAU, ScalarEvolution *SE,
1720 OptimizationRemarkEmitter *ORE) {
1721 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)
1722 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "LICM hoisting to " << Dest
->getName() << ": " << I << "\n"; } } while
(false)
;
1723 ORE->emit([&]() {
1724 return OptimizationRemark(DEBUG_TYPE"licm", "Hoisted", &I) << "hoisting "
1725 << ore::NV("Inst", &I);
1726 });
1727
1728 // Metadata can be dependent on conditions we are hoisting above.
1729 // Conservatively strip all metadata on the instruction unless we were
1730 // guaranteed to execute I if we entered the loop, in which case the metadata
1731 // is valid in the loop preheader.
1732 if (I.hasMetadataOtherThanDebugLoc() &&
1733 // The check on hasMetadataOtherThanDebugLoc is to prevent us from burning
1734 // time in isGuaranteedToExecute if we don't actually have anything to
1735 // drop. It is a compile time optimization, not required for correctness.
1736 !SafetyInfo->isGuaranteedToExecute(I, DT, CurLoop))
1737 I.dropUnknownNonDebugMetadata();
1738
1739 if (isa<PHINode>(I))
1740 // Move the new node to the end of the phi list in the destination block.
1741 moveInstructionBefore(I, *Dest->getFirstNonPHI(), *SafetyInfo, MSSAU, SE);
1742 else
1743 // Move the new node to the destination block, before its terminator.
1744 moveInstructionBefore(I, *Dest->getTerminator(), *SafetyInfo, MSSAU, SE);
1745
1746 // Apply line 0 debug locations when we are moving instructions to different
1747 // basic blocks because we want to avoid jumpy line tables.
1748 if (const DebugLoc &DL = I.getDebugLoc())
1749 I.setDebugLoc(DebugLoc::get(0, 0, DL.getScope(), DL.getInlinedAt()));
1750
1751 if (isa<LoadInst>(I))
1752 ++NumMovedLoads;
1753 else if (isa<CallInst>(I))
1754 ++NumMovedCalls;
1755 ++NumHoisted;
1756}
1757
1758/// Only sink or hoist an instruction if it is not a trapping instruction,
1759/// or if the instruction is known not to trap when moved to the preheader.
1760/// or if it is a trapping instruction and is guaranteed to execute.
1761static bool isSafeToExecuteUnconditionally(Instruction &Inst,
1762 const DominatorTree *DT,
1763 const Loop *CurLoop,
1764 const LoopSafetyInfo *SafetyInfo,
1765 OptimizationRemarkEmitter *ORE,
1766 const Instruction *CtxI) {
1767 if (isSafeToSpeculativelyExecute(&Inst, CtxI, DT))
1768 return true;
1769
1770 bool GuaranteedToExecute =
1771 SafetyInfo->isGuaranteedToExecute(Inst, DT, CurLoop);
1772
1773 if (!GuaranteedToExecute) {
1774 auto *LI = dyn_cast<LoadInst>(&Inst);
1775 if (LI && CurLoop->isLoopInvariant(LI->getPointerOperand()))
1776 ORE->emit([&]() {
1777 return OptimizationRemarkMissed(
1778 DEBUG_TYPE"licm", "LoadWithLoopInvariantAddressCondExecuted", LI)
1779 << "failed to hoist load with loop-invariant address "
1780 "because load is conditionally executed";
1781 });
1782 }
1783
1784 return GuaranteedToExecute;
1785}
1786
1787namespace {
1788class LoopPromoter : public LoadAndStorePromoter {
1789 Value *SomePtr; // Designated pointer to store to.
1790 const SmallSetVector<Value *, 8> &PointerMustAliases;
1791 SmallVectorImpl<BasicBlock *> &LoopExitBlocks;
1792 SmallVectorImpl<Instruction *> &LoopInsertPts;
1793 SmallVectorImpl<MemoryAccess *> &MSSAInsertPts;
1794 PredIteratorCache &PredCache;
1795 AliasSetTracker &AST;
1796 MemorySSAUpdater *MSSAU;
1797 LoopInfo &LI;
1798 DebugLoc DL;
1799 int Alignment;
1800 bool UnorderedAtomic;
1801 AAMDNodes AATags;
1802 ICFLoopSafetyInfo &SafetyInfo;
1803
1804 Value *maybeInsertLCSSAPHI(Value *V, BasicBlock *BB) const {
1805 if (Instruction *I = dyn_cast<Instruction>(V))
1806 if (Loop *L = LI.getLoopFor(I->getParent()))
1807 if (!L->contains(BB)) {
1808 // We need to create an LCSSA PHI node for the incoming value and
1809 // store that.
1810 PHINode *PN = PHINode::Create(I->getType(), PredCache.size(BB),
1811 I->getName() + ".lcssa", &BB->front());
1812 for (BasicBlock *Pred : PredCache.get(BB))
1813 PN->addIncoming(I, Pred);
1814 return PN;
1815 }
1816 return V;
1817 }
1818
1819public:
1820 LoopPromoter(Value *SP, ArrayRef<const Instruction *> Insts, SSAUpdater &S,
1821 const SmallSetVector<Value *, 8> &PMA,
1822 SmallVectorImpl<BasicBlock *> &LEB,
1823 SmallVectorImpl<Instruction *> &LIP,
1824 SmallVectorImpl<MemoryAccess *> &MSSAIP, PredIteratorCache &PIC,
1825 AliasSetTracker &ast, MemorySSAUpdater *MSSAU, LoopInfo &li,
1826 DebugLoc dl, int alignment, bool UnorderedAtomic,
1827 const AAMDNodes &AATags, ICFLoopSafetyInfo &SafetyInfo)
1828 : LoadAndStorePromoter(Insts, S), SomePtr(SP), PointerMustAliases(PMA),
1829 LoopExitBlocks(LEB), LoopInsertPts(LIP), MSSAInsertPts(MSSAIP),
1830 PredCache(PIC), AST(ast), MSSAU(MSSAU), LI(li), DL(std::move(dl)),
1831 Alignment(alignment), UnorderedAtomic(UnorderedAtomic), AATags(AATags),
1832 SafetyInfo(SafetyInfo) {}
1833
1834 bool isInstInList(Instruction *I,
1835 const SmallVectorImpl<Instruction *> &) const override {
1836 Value *Ptr;
1837 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1838 Ptr = LI->getOperand(0);
1839 else
1840 Ptr = cast<StoreInst>(I)->getPointerOperand();
1841 return PointerMustAliases.count(Ptr);
1842 }
1843
1844 void doExtraRewritesBeforeFinalDeletion() override {
1845 // Insert stores after in the loop exit blocks. Each exit block gets a
1846 // store of the live-out values that feed them. Since we've already told
1847 // the SSA updater about the defs in the loop and the preheader
1848 // definition, it is all set and we can start using it.
1849 for (unsigned i = 0, e = LoopExitBlocks.size(); i != e; ++i) {
1850 BasicBlock *ExitBlock = LoopExitBlocks[i];
1851 Value *LiveInValue = SSA.GetValueInMiddleOfBlock(ExitBlock);
1852 LiveInValue = maybeInsertLCSSAPHI(LiveInValue, ExitBlock);
1853 Value *Ptr = maybeInsertLCSSAPHI(SomePtr, ExitBlock);
1854 Instruction *InsertPos = LoopInsertPts[i];
1855 StoreInst *NewSI = new StoreInst(LiveInValue, Ptr, InsertPos);
1856 if (UnorderedAtomic)
1857 NewSI->setOrdering(AtomicOrdering::Unordered);
1858 NewSI->setAlignment(MaybeAlign(Alignment));
1859 NewSI->setDebugLoc(DL);
1860 if (AATags)
1861 NewSI->setAAMetadata(AATags);
1862
1863 if (MSSAU) {
1864 MemoryAccess *MSSAInsertPoint = MSSAInsertPts[i];
1865 MemoryAccess *NewMemAcc;
1866 if (!MSSAInsertPoint) {
1867 NewMemAcc = MSSAU->createMemoryAccessInBB(
1868 NewSI, nullptr, NewSI->getParent(), MemorySSA::Beginning);
1869 } else {
1870 NewMemAcc =
1871 MSSAU->createMemoryAccessAfter(NewSI, nullptr, MSSAInsertPoint);
1872 }
1873 MSSAInsertPts[i] = NewMemAcc;
1874 MSSAU->insertDef(cast<MemoryDef>(NewMemAcc), true);
1875 // FIXME: true for safety, false may still be correct.
1876 }
1877 }
1878 }
1879
1880 void replaceLoadWithValue(LoadInst *LI, Value *V) const override {
1881 // Update alias analysis.
1882 AST.copyValue(LI, V);
1883 }
1884 void instructionDeleted(Instruction *I) const override {
1885 SafetyInfo.removeInstruction(I);
1886 AST.deleteValue(I);
1887 if (MSSAU)
1888 MSSAU->removeMemoryAccess(I);
1889 }
1890};
1891
1892
1893/// Return true iff we can prove that a caller of this function can not inspect
1894/// the contents of the provided object in a well defined program.
1895bool isKnownNonEscaping(Value *Object, const TargetLibraryInfo *TLI) {
1896 if (isa<AllocaInst>(Object))
1897 // Since the alloca goes out of scope, we know the caller can't retain a
1898 // reference to it and be well defined. Thus, we don't need to check for
1899 // capture.
1900 return true;
1901
1902 // For all other objects we need to know that the caller can't possibly
1903 // have gotten a reference to the object. There are two components of
1904 // that:
1905 // 1) Object can't be escaped by this function. This is what
1906 // PointerMayBeCaptured checks.
1907 // 2) Object can't have been captured at definition site. For this, we
1908 // need to know the return value is noalias. At the moment, we use a
1909 // weaker condition and handle only AllocLikeFunctions (which are
1910 // known to be noalias). TODO
1911 return isAllocLikeFn(Object, TLI) &&
1912 !PointerMayBeCaptured(Object, true, true);
1913}
1914
1915} // namespace
1916
1917/// Try to promote memory values to scalars by sinking stores out of the
1918/// loop and moving loads to before the loop. We do this by looping over
1919/// the stores in the loop, looking for stores to Must pointers which are
1920/// loop invariant.
1921///
1922bool llvm::promoteLoopAccessesToScalars(
1923 const SmallSetVector<Value *, 8> &PointerMustAliases,
1924 SmallVectorImpl<BasicBlock *> &ExitBlocks,
1925 SmallVectorImpl<Instruction *> &InsertPts,
1926 SmallVectorImpl<MemoryAccess *> &MSSAInsertPts, PredIteratorCache &PIC,
1927 LoopInfo *LI, DominatorTree *DT, const TargetLibraryInfo *TLI,
1928 Loop *CurLoop, AliasSetTracker *CurAST, MemorySSAUpdater *MSSAU,
1929 ICFLoopSafetyInfo *SafetyInfo, OptimizationRemarkEmitter *ORE) {
1930 // Verify inputs.
1931 assert(LI != nullptr && DT != nullptr && CurLoop != nullptr &&((LI != nullptr && DT != nullptr && CurLoop !=
nullptr && CurAST != nullptr && SafetyInfo !=
nullptr && "Unexpected Input to promoteLoopAccessesToScalars"
) ? static_cast<void> (0) : __assert_fail ("LI != nullptr && DT != nullptr && CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr && \"Unexpected Input to promoteLoopAccessesToScalars\""
, "/build/llvm-toolchain-snapshot-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1933, __PRETTY_FUNCTION__))
1932 CurAST != nullptr && SafetyInfo != nullptr &&((LI != nullptr && DT != nullptr && CurLoop !=
nullptr && CurAST != nullptr && SafetyInfo !=
nullptr && "Unexpected Input to promoteLoopAccessesToScalars"
) ? static_cast<void> (0) : __assert_fail ("LI != nullptr && DT != nullptr && CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr && \"Unexpected Input to promoteLoopAccessesToScalars\""
, "/build/llvm-toolchain-snapshot-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1933, __PRETTY_FUNCTION__))
1933 "Unexpected Input to promoteLoopAccessesToScalars")((LI != nullptr && DT != nullptr && CurLoop !=
nullptr && CurAST != nullptr && SafetyInfo !=
nullptr && "Unexpected Input to promoteLoopAccessesToScalars"
) ? static_cast<void> (0) : __assert_fail ("LI != nullptr && DT != nullptr && CurLoop != nullptr && CurAST != nullptr && SafetyInfo != nullptr && \"Unexpected Input to promoteLoopAccessesToScalars\""
, "/build/llvm-toolchain-snapshot-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 1933, __PRETTY_FUNCTION__))
;
1934
1935 Value *SomePtr = *PointerMustAliases.begin();
1936 BasicBlock *Preheader = CurLoop->getLoopPreheader();
1937
1938 // It is not safe to promote a load/store from the loop if the load/store is
1939 // conditional. For example, turning:
1940 //
1941 // for () { if (c) *P += 1; }
1942 //
1943 // into:
1944 //
1945 // tmp = *P; for () { if (c) tmp +=1; } *P = tmp;
1946 //
1947 // is not safe, because *P may only be valid to access if 'c' is true.
1948 //
1949 // The safety property divides into two parts:
1950 // p1) The memory may not be dereferenceable on entry to the loop. In this
1951 // case, we can't insert the required load in the preheader.
1952 // p2) The memory model does not allow us to insert a store along any dynamic
1953 // path which did not originally have one.
1954 //
1955 // If at least one store is guaranteed to execute, both properties are
1956 // satisfied, and promotion is legal.
1957 //
1958 // This, however, is not a necessary condition. Even if no store/load is
1959 // guaranteed to execute, we can still establish these properties.
1960 // We can establish (p1) by proving that hoisting the load into the preheader
1961 // is safe (i.e. proving dereferenceability on all paths through the loop). We
1962 // can use any access within the alias set to prove dereferenceability,
1963 // since they're all must alias.
1964 //
1965 // There are two ways establish (p2):
1966 // a) Prove the location is thread-local. In this case the memory model
1967 // requirement does not apply, and stores are safe to insert.
1968 // b) Prove a store dominates every exit block. In this case, if an exit
1969 // blocks is reached, the original dynamic path would have taken us through
1970 // the store, so inserting a store into the exit block is safe. Note that this
1971 // is different from the store being guaranteed to execute. For instance,
1972 // if an exception is thrown on the first iteration of the loop, the original
1973 // store is never executed, but the exit blocks are not executed either.
1974
1975 bool DereferenceableInPH = false;
1976 bool SafeToInsertStore = false;
1977
1978 SmallVector<Instruction *, 64> LoopUses;
1979
1980 // We start with an alignment of one and try to find instructions that allow
1981 // us to prove better alignment.
1982 unsigned Alignment = 1;
1983 // Keep track of which types of access we see
1984 bool SawUnorderedAtomic = false;
1985 bool SawNotAtomic = false;
1986 AAMDNodes AATags;
1987
1988 const DataLayout &MDL = Preheader->getModule()->getDataLayout();
1989
1990 bool IsKnownThreadLocalObject = false;
1991 if (SafetyInfo->anyBlockMayThrow()) {
1992 // If a loop can throw, we have to insert a store along each unwind edge.
1993 // That said, we can't actually make the unwind edge explicit. Therefore,
1994 // we have to prove that the store is dead along the unwind edge. We do
1995 // this by proving that the caller can't have a reference to the object
1996 // after return and thus can't possibly load from the object.
1997 Value *Object = GetUnderlyingObject(SomePtr, MDL);
1998 if (!isKnownNonEscaping(Object, TLI))
1999 return false;
2000 // Subtlety: Alloca's aren't visible to callers, but *are* potentially
2001 // visible to other threads if captured and used during their lifetimes.
2002 IsKnownThreadLocalObject = !isa<AllocaInst>(Object);
2003 }
2004
2005 // Check that all of the pointers in the alias set have the same type. We
2006 // cannot (yet) promote a memory location that is loaded and stored in
2007 // different sizes. While we are at it, collect alignment and AA info.
2008 for (Value *ASIV : PointerMustAliases) {
2009 // Check that all of the pointers in the alias set have the same type. We
2010 // cannot (yet) promote a memory location that is loaded and stored in
2011 // different sizes.
2012 if (SomePtr->getType() != ASIV->getType())
2013 return false;
2014
2015 for (User *U : ASIV->users()) {
2016 // Ignore instructions that are outside the loop.
2017 Instruction *UI = dyn_cast<Instruction>(U);
2018 if (!UI || !CurLoop->contains(UI))
2019 continue;
2020
2021 // If there is an non-load/store instruction in the loop, we can't promote
2022 // it.
2023 if (LoadInst *Load = dyn_cast<LoadInst>(UI)) {
2024 if (!Load->isUnordered())
2025 return false;
2026
2027 SawUnorderedAtomic |= Load->isAtomic();
2028 SawNotAtomic |= !Load->isAtomic();
2029
2030 unsigned InstAlignment = Load->getAlignment();
2031 if (!InstAlignment)
2032 InstAlignment =
2033 MDL.getABITypeAlignment(Load->getType());
2034
2035 // Note that proving a load safe to speculate requires proving
2036 // sufficient alignment at the target location. Proving it guaranteed
2037 // to execute does as well. Thus we can increase our guaranteed
2038 // alignment as well.
2039 if (!DereferenceableInPH || (InstAlignment > Alignment))
2040 if (isSafeToExecuteUnconditionally(*Load, DT, CurLoop, SafetyInfo,
2041 ORE, Preheader->getTerminator())) {
2042 DereferenceableInPH = true;
2043 Alignment = std::max(Alignment, InstAlignment);
2044 }
2045 } else if (const StoreInst *Store = dyn_cast<StoreInst>(UI)) {
2046 // Stores *of* the pointer are not interesting, only stores *to* the
2047 // pointer.
2048 if (UI->getOperand(1) != ASIV)
2049 continue;
2050 if (!Store->isUnordered())
2051 return false;
2052
2053 SawUnorderedAtomic |= Store->isAtomic();
2054 SawNotAtomic |= !Store->isAtomic();
2055
2056 // If the store is guaranteed to execute, both properties are satisfied.
2057 // We may want to check if a store is guaranteed to execute even if we
2058 // already know that promotion is safe, since it may have higher
2059 // alignment than any other guaranteed stores, in which case we can
2060 // raise the alignment on the promoted store.
2061 unsigned InstAlignment = Store->getAlignment();
2062 if (!InstAlignment)
2063 InstAlignment =
2064 MDL.getABITypeAlignment(Store->getValueOperand()->getType());
2065
2066 if (!DereferenceableInPH || !SafeToInsertStore ||
2067 (InstAlignment > Alignment)) {
2068 if (SafetyInfo->isGuaranteedToExecute(*UI, DT, CurLoop)) {
2069 DereferenceableInPH = true;
2070 SafeToInsertStore = true;
2071 Alignment = std::max(Alignment, InstAlignment);
2072 }
2073 }
2074
2075 // If a store dominates all exit blocks, it is safe to sink.
2076 // As explained above, if an exit block was executed, a dominating
2077 // store must have been executed at least once, so we are not
2078 // introducing stores on paths that did not have them.
2079 // Note that this only looks at explicit exit blocks. If we ever
2080 // start sinking stores into unwind edges (see above), this will break.
2081 if (!SafeToInsertStore)
2082 SafeToInsertStore = llvm::all_of(ExitBlocks, [&](BasicBlock *Exit) {
2083 return DT->dominates(Store->getParent(), Exit);
2084 });
2085
2086 // If the store is not guaranteed to execute, we may still get
2087 // deref info through it.
2088 if (!DereferenceableInPH) {
2089 DereferenceableInPH = isDereferenceableAndAlignedPointer(
2090 Store->getPointerOperand(), Store->getValueOperand()->getType(),
2091 MaybeAlign(Store->getAlignment()), MDL,
2092 Preheader->getTerminator(), DT);
2093 }
2094 } else
2095 return false; // Not a load or store.
2096
2097 // Merge the AA tags.
2098 if (LoopUses.empty()) {
2099 // On the first load/store, just take its AA tags.
2100 UI->getAAMetadata(AATags);
2101 } else if (AATags) {
2102 UI->getAAMetadata(AATags, /* Merge = */ true);
2103 }
2104
2105 LoopUses.push_back(UI);
2106 }
2107 }
2108
2109 // If we found both an unordered atomic instruction and a non-atomic memory
2110 // access, bail. We can't blindly promote non-atomic to atomic since we
2111 // might not be able to lower the result. We can't downgrade since that
2112 // would violate memory model. Also, align 0 is an error for atomics.
2113 if (SawUnorderedAtomic && SawNotAtomic)
2114 return false;
2115
2116 // If we're inserting an atomic load in the preheader, we must be able to
2117 // lower it. We're only guaranteed to be able to lower naturally aligned
2118 // atomics.
2119 auto *SomePtrElemType = SomePtr->getType()->getPointerElementType();
2120 if (SawUnorderedAtomic &&
2121 Alignment < MDL.getTypeStoreSize(SomePtrElemType))
2122 return false;
2123
2124 // If we couldn't prove we can hoist the load, bail.
2125 if (!DereferenceableInPH)
2126 return false;
2127
2128 // We know we can hoist the load, but don't have a guaranteed store.
2129 // Check whether the location is thread-local. If it is, then we can insert
2130 // stores along paths which originally didn't have them without violating the
2131 // memory model.
2132 if (!SafeToInsertStore) {
2133 if (IsKnownThreadLocalObject)
2134 SafeToInsertStore = true;
2135 else {
2136 Value *Object = GetUnderlyingObject(SomePtr, MDL);
2137 SafeToInsertStore =
2138 (isAllocLikeFn(Object, TLI) || isa<AllocaInst>(Object)) &&
2139 !PointerMayBeCaptured(Object, true, true);
2140 }
2141 }
2142
2143 // If we've still failed to prove we can sink the store, give up.
2144 if (!SafeToInsertStore)
2145 return false;
2146
2147 // Otherwise, this is safe to promote, lets do it!
2148 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)
2149 << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "LICM: Promoting value stored to in loop: "
<< *SomePtr << '\n'; } } while (false)
;
2150 ORE->emit([&]() {
2151 return OptimizationRemark(DEBUG_TYPE"licm", "PromoteLoopAccessesToScalar",
2152 LoopUses[0])
2153 << "Moving accesses to memory location out of the loop";
2154 });
2155 ++NumPromoted;
2156
2157 // Grab a debug location for the inserted loads/stores; given that the
2158 // inserted loads/stores have little relation to the original loads/stores,
2159 // this code just arbitrarily picks a location from one, since any debug
2160 // location is better than none.
2161 DebugLoc DL = LoopUses[0]->getDebugLoc();
2162
2163 // We use the SSAUpdater interface to insert phi nodes as required.
2164 SmallVector<PHINode *, 16> NewPHIs;
2165 SSAUpdater SSA(&NewPHIs);
2166 LoopPromoter Promoter(SomePtr, LoopUses, SSA, PointerMustAliases, ExitBlocks,
2167 InsertPts, MSSAInsertPts, PIC, *CurAST, MSSAU, *LI, DL,
2168 Alignment, SawUnorderedAtomic, AATags, *SafetyInfo);
2169
2170 // Set up the preheader to have a definition of the value. It is the live-out
2171 // value from the preheader that uses in the loop will use.
2172 LoadInst *PreheaderLoad = new LoadInst(
2173 SomePtr->getType()->getPointerElementType(), SomePtr,
2174 SomePtr->getName() + ".promoted", Preheader->getTerminator());
2175 if (SawUnorderedAtomic)
2176 PreheaderLoad->setOrdering(AtomicOrdering::Unordered);
2177 PreheaderLoad->setAlignment(MaybeAlign(Alignment));
2178 PreheaderLoad->setDebugLoc(DL);
2179 if (AATags)
2180 PreheaderLoad->setAAMetadata(AATags);
2181 SSA.AddAvailableValue(Preheader, PreheaderLoad);
2182
2183 if (MSSAU) {
2184 MemoryAccess *PreheaderLoadMemoryAccess = MSSAU->createMemoryAccessInBB(
2185 PreheaderLoad, nullptr, PreheaderLoad->getParent(), MemorySSA::End);
2186 MemoryUse *NewMemUse = cast<MemoryUse>(PreheaderLoadMemoryAccess);
2187 MSSAU->insertUse(NewMemUse, /*RenameUses=*/true);
2188 }
2189
2190 if (MSSAU && VerifyMemorySSA)
2191 MSSAU->getMemorySSA()->verifyMemorySSA();
2192 // Rewrite all the loads in the loop and remember all the definitions from
2193 // stores in the loop.
2194 Promoter.run(LoopUses);
2195
2196 if (MSSAU && VerifyMemorySSA)
2197 MSSAU->getMemorySSA()->verifyMemorySSA();
2198 // If the SSAUpdater didn't use the load in the preheader, just zap it now.
2199 if (PreheaderLoad->use_empty())
2200 eraseInstruction(*PreheaderLoad, *SafetyInfo, CurAST, MSSAU);
2201
2202 return true;
2203}
2204
2205/// Returns an owning pointer to an alias set which incorporates aliasing info
2206/// from L and all subloops of L.
2207/// FIXME: In new pass manager, there is no helper function to handle loop
2208/// analysis such as cloneBasicBlockAnalysis, so the AST needs to be recomputed
2209/// from scratch for every loop. Hook up with the helper functions when
2210/// available in the new pass manager to avoid redundant computation.
2211std::unique_ptr<AliasSetTracker>
2212LoopInvariantCodeMotion::collectAliasInfoForLoop(Loop *L, LoopInfo *LI,
2213 AliasAnalysis *AA) {
2214 std::unique_ptr<AliasSetTracker> CurAST;
2215 SmallVector<Loop *, 4> RecomputeLoops;
2216 for (Loop *InnerL : L->getSubLoops()) {
2217 auto MapI = LoopToAliasSetMap.find(InnerL);
2218 // If the AST for this inner loop is missing it may have been merged into
2219 // some other loop's AST and then that loop unrolled, and so we need to
2220 // recompute it.
2221 if (MapI == LoopToAliasSetMap.end()) {
2222 RecomputeLoops.push_back(InnerL);
2223 continue;
2224 }
2225 std::unique_ptr<AliasSetTracker> InnerAST = std::move(MapI->second);
2226
2227 if (CurAST) {
2228 // What if InnerLoop was modified by other passes ?
2229 // Once we've incorporated the inner loop's AST into ours, we don't need
2230 // the subloop's anymore.
2231 CurAST->add(*InnerAST);
2232 } else {
2233 CurAST = std::move(InnerAST);
2234 }
2235 LoopToAliasSetMap.erase(MapI);
2236 }
2237 if (!CurAST)
2238 CurAST = std::make_unique<AliasSetTracker>(*AA);
2239
2240 // Add everything from the sub loops that are no longer directly available.
2241 for (Loop *InnerL : RecomputeLoops)
2242 for (BasicBlock *BB : InnerL->blocks())
2243 CurAST->add(*BB);
2244
2245 // And merge in this loop (without anything from inner loops).
2246 for (BasicBlock *BB : L->blocks())
2247 if (LI->getLoopFor(BB) == L)
2248 CurAST->add(*BB);
2249
2250 return CurAST;
2251}
2252
2253std::unique_ptr<AliasSetTracker>
2254LoopInvariantCodeMotion::collectAliasInfoForLoopWithMSSA(
2255 Loop *L, AliasAnalysis *AA, MemorySSAUpdater *MSSAU) {
2256 auto *MSSA = MSSAU->getMemorySSA();
2257 auto CurAST = std::make_unique<AliasSetTracker>(*AA, MSSA, L);
2258 CurAST->addAllInstructionsInLoopUsingMSSA();
2259 return CurAST;
2260}
2261
2262/// Simple analysis hook. Clone alias set info.
2263///
2264void LegacyLICMPass::cloneBasicBlockAnalysis(BasicBlock *From, BasicBlock *To,
2265 Loop *L) {
2266 auto ASTIt = LICM.getLoopToAliasSetMap().find(L);
2267 if (ASTIt == LICM.getLoopToAliasSetMap().end())
2268 return;
2269
2270 ASTIt->second->copyValue(From, To);
2271}
2272
2273/// Simple Analysis hook. Delete value V from alias set
2274///
2275void LegacyLICMPass::deleteAnalysisValue(Value *V, Loop *L) {
2276 auto ASTIt = LICM.getLoopToAliasSetMap().find(L);
2277 if (ASTIt == LICM.getLoopToAliasSetMap().end())
2278 return;
2279
2280 ASTIt->second->deleteValue(V);
2281}
2282
2283/// Simple Analysis hook. Delete value L from alias set map.
2284///
2285void LegacyLICMPass::deleteAnalysisLoop(Loop *L) {
2286 if (!LICM.getLoopToAliasSetMap().count(L))
2287 return;
2288
2289 LICM.getLoopToAliasSetMap().erase(L);
2290}
2291
2292static bool pointerInvalidatedByLoop(MemoryLocation MemLoc,
2293 AliasSetTracker *CurAST, Loop *CurLoop,
2294 AliasAnalysis *AA) {
2295 // First check to see if any of the basic blocks in CurLoop invalidate *V.
2296 bool isInvalidatedAccordingToAST = CurAST->getAliasSetFor(MemLoc).isMod();
2297
2298 if (!isInvalidatedAccordingToAST || !LICMN2Theshold)
2299 return isInvalidatedAccordingToAST;
2300
2301 // Check with a diagnostic analysis if we can refine the information above.
2302 // This is to identify the limitations of using the AST.
2303 // The alias set mechanism used by LICM has a major weakness in that it
2304 // combines all things which may alias into a single set *before* asking
2305 // modref questions. As a result, a single readonly call within a loop will
2306 // collapse all loads and stores into a single alias set and report
2307 // invalidation if the loop contains any store. For example, readonly calls
2308 // with deopt states have this form and create a general alias set with all
2309 // loads and stores. In order to get any LICM in loops containing possible
2310 // deopt states we need a more precise invalidation of checking the mod ref
2311 // info of each instruction within the loop and LI. This has a complexity of
2312 // O(N^2), so currently, it is used only as a diagnostic tool since the
2313 // default value of LICMN2Threshold is zero.
2314
2315 // Don't look at nested loops.
2316 if (CurLoop->begin() != CurLoop->end())
2317 return true;
2318
2319 int N = 0;
2320 for (BasicBlock *BB : CurLoop->getBlocks())
2321 for (Instruction &I : *BB) {
2322 if (N >= LICMN2Theshold) {
2323 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)
2324 << *(MemLoc.Ptr) << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "Alasing N2 threshold exhausted for "
<< *(MemLoc.Ptr) << "\n"; } } while (false)
;
2325 return true;
2326 }
2327 N++;
2328 auto Res = AA->getModRefInfo(&I, MemLoc);
2329 if (isModSet(Res)) {
2330 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
)
2331 << *(MemLoc.Ptr) << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("licm")) { dbgs() << "Aliasing failed on " << I <<
" for " << *(MemLoc.Ptr) << "\n"; } } while (false
)
;
2332 return true;
2333 }
2334 }
2335 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)
;
2336 return false;
2337}
2338
2339static bool pointerInvalidatedByLoopWithMSSA(MemorySSA *MSSA, MemoryUse *MU,
2340 Loop *CurLoop,
2341 SinkAndHoistLICMFlags &Flags) {
2342 // For hoisting, use the walker to determine safety
2343 if (!Flags.IsSink) {
2344 MemoryAccess *Source;
2345 // See declaration of SetLicmMssaOptCap for usage details.
2346 if (Flags.LicmMssaOptCounter >= Flags.LicmMssaOptCap)
2347 Source = MU->getDefiningAccess();
2348 else {
2349 Source = MSSA->getSkipSelfWalker()->getClobberingMemoryAccess(MU);
2350 Flags.LicmMssaOptCounter++;
2351 }
2352 return !MSSA->isLiveOnEntryDef(Source) &&
2353 CurLoop->contains(Source->getBlock());
2354 }
2355
2356 // For sinking, we'd need to check all Defs below this use. The getClobbering
2357 // call will look on the backedge of the loop, but will check aliasing with
2358 // the instructions on the previous iteration.
2359 // For example:
2360 // for (i ... )
2361 // load a[i] ( Use (LoE)
2362 // store a[i] ( 1 = Def (2), with 2 = Phi for the loop.
2363 // i++;
2364 // The load sees no clobbering inside the loop, as the backedge alias check
2365 // does phi translation, and will check aliasing against store a[i-1].
2366 // However sinking the load outside the loop, below the store is incorrect.
2367
2368 // For now, only sink if there are no Defs in the loop, and the existing ones
2369 // precede the use and are in the same block.
2370 // FIXME: Increase precision: Safe to sink if Use post dominates the Def;
2371 // needs PostDominatorTreeAnalysis.
2372 // FIXME: More precise: no Defs that alias this Use.
2373 if (Flags.NoOfMemAccTooLarge)
2374 return true;
2375 for (auto *BB : CurLoop->getBlocks())
2376 if (auto *Accesses = MSSA->getBlockDefs(BB))
2377 for (const auto &MA : *Accesses)
2378 if (const auto *MD = dyn_cast<MemoryDef>(&MA))
2379 if (MU->getBlock() != MD->getBlock() ||
2380 !MSSA->locallyDominates(MD, MU))
2381 return true;
2382 return false;
2383}
2384
2385/// Little predicate that returns true if the specified basic block is in
2386/// a subloop of the current one, not the current one itself.
2387///
2388static bool inSubLoop(BasicBlock *BB, Loop *CurLoop, LoopInfo *LI) {
2389 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-10~+201911111502510600c19528f1809/llvm/lib/Transforms/Scalar/LICM.cpp"
, 2389, __PRETTY_FUNCTION__))
;
2390 return LI->getLoopFor(BB) != CurLoop;
2391}