LLVM 20.0.0git
MergedLoadStoreMotion.cpp
Go to the documentation of this file.
1//===- MergedLoadStoreMotion.cpp - merge and hoist/sink load/stores -------===//
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//! \file
10//! This pass performs merges of loads and stores on both sides of a
11// diamond (hammock). It hoists the loads and sinks the stores.
12//
13// The algorithm iteratively hoists two loads to the same address out of a
14// diamond (hammock) and merges them into a single load in the header. Similar
15// it sinks and merges two stores to the tail block (footer). The algorithm
16// iterates over the instructions of one side of the diamond and attempts to
17// find a matching load/store on the other side. New tail/footer block may be
18// insterted if the tail/footer block has more predecessors (not only the two
19// predecessors that are forming the diamond). It hoists / sinks when it thinks
20// it safe to do so. This optimization helps with eg. hiding load latencies,
21// triggering if-conversion, and reducing static code size.
22//
23// NOTE: This code no longer performs load hoisting, it is subsumed by GVNHoist.
24//
25//===----------------------------------------------------------------------===//
26//
27//
28// Example:
29// Diamond shaped code before merge:
30//
31// header:
32// br %cond, label %if.then, label %if.else
33// + +
34// + +
35// + +
36// if.then: if.else:
37// %lt = load %addr_l %le = load %addr_l
38// <use %lt> <use %le>
39// <...> <...>
40// store %st, %addr_s store %se, %addr_s
41// br label %if.end br label %if.end
42// + +
43// + +
44// + +
45// if.end ("footer"):
46// <...>
47//
48// Diamond shaped code after merge:
49//
50// header:
51// %l = load %addr_l
52// br %cond, label %if.then, label %if.else
53// + +
54// + +
55// + +
56// if.then: if.else:
57// <use %l> <use %l>
58// <...> <...>
59// br label %if.end br label %if.end
60// + +
61// + +
62// + +
63// if.end ("footer"):
64// %s.sink = phi [%st, if.then], [%se, if.else]
65// <...>
66// store %s.sink, %addr_s
67// <...>
68//
69//
70//===----------------------- TODO -----------------------------------------===//
71//
72// 1) Generalize to regions other than diamonds
73// 2) Be more aggressive merging memory operations
74// Note that both changes require register pressure control
75//
76//===----------------------------------------------------------------------===//
77
81#include "llvm/IR/IRBuilder.h"
83#include "llvm/Support/Debug.h"
87
88using namespace llvm;
89
90#define DEBUG_TYPE "mldst-motion"
91
92namespace {
93//===----------------------------------------------------------------------===//
94// MergedLoadStoreMotion Pass
95//===----------------------------------------------------------------------===//
96class MergedLoadStoreMotion {
97 AliasAnalysis *AA = nullptr;
98
99 // The mergeLoad/Store algorithms could have Size0 * Size1 complexity,
100 // where Size0 and Size1 are the #instructions on the two sides of
101 // the diamond. The constant chosen here is arbitrary. Compiler Time
102 // Control is enforced by the check Size0 * Size1 < MagicCompileTimeControl.
103 const int MagicCompileTimeControl = 250;
104
105 const bool SplitFooterBB;
106public:
107 MergedLoadStoreMotion(bool SplitFooterBB) : SplitFooterBB(SplitFooterBB) {}
108 bool run(Function &F, AliasAnalysis &AA);
109
110private:
111 BasicBlock *getDiamondTail(BasicBlock *BB);
112 bool isDiamondHead(BasicBlock *BB);
113 // Routines for sinking stores
114 StoreInst *canSinkFromBlock(BasicBlock *BB, StoreInst *SI);
115 PHINode *getPHIOperand(BasicBlock *BB, StoreInst *S0, StoreInst *S1);
116 bool isStoreSinkBarrierInRange(const Instruction &Start,
117 const Instruction &End, MemoryLocation Loc);
118 bool canSinkStoresAndGEPs(StoreInst *S0, StoreInst *S1) const;
119 void sinkStoresAndGEPs(BasicBlock *BB, StoreInst *SinkCand,
120 StoreInst *ElseInst);
121 bool mergeStores(BasicBlock *BB);
122};
123} // end anonymous namespace
124
125///
126/// Return tail block of a diamond.
127///
128BasicBlock *MergedLoadStoreMotion::getDiamondTail(BasicBlock *BB) {
129 assert(isDiamondHead(BB) && "Basic block is not head of a diamond");
131}
132
133///
134/// True when BB is the head of a diamond (hammock)
135///
136bool MergedLoadStoreMotion::isDiamondHead(BasicBlock *BB) {
137 if (!BB)
138 return false;
139 auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
140 if (!BI || !BI->isConditional())
141 return false;
142
143 BasicBlock *Succ0 = BI->getSuccessor(0);
144 BasicBlock *Succ1 = BI->getSuccessor(1);
145
146 if (!Succ0->getSinglePredecessor())
147 return false;
148 if (!Succ1->getSinglePredecessor())
149 return false;
150
151 BasicBlock *Succ0Succ = Succ0->getSingleSuccessor();
152 BasicBlock *Succ1Succ = Succ1->getSingleSuccessor();
153 // Ignore triangles.
154 if (!Succ0Succ || !Succ1Succ || Succ0Succ != Succ1Succ)
155 return false;
156 return true;
157}
158
159
160///
161/// True when instruction is a sink barrier for a store
162/// located in Loc
163///
164/// Whenever an instruction could possibly read or modify the
165/// value being stored or protect against the store from
166/// happening it is considered a sink barrier.
167///
168bool MergedLoadStoreMotion::isStoreSinkBarrierInRange(const Instruction &Start,
169 const Instruction &End,
170 MemoryLocation Loc) {
171 for (const Instruction &Inst :
172 make_range(Start.getIterator(), End.getIterator()))
173 if (Inst.mayThrow())
174 return true;
175 return AA->canInstructionRangeModRef(Start, End, Loc, ModRefInfo::ModRef);
176}
177
178///
179/// Check if \p BB contains a store to the same address as \p SI
180///
181/// \return The store in \p when it is safe to sink. Otherwise return Null.
182///
183StoreInst *MergedLoadStoreMotion::canSinkFromBlock(BasicBlock *BB1,
184 StoreInst *Store0) {
185 LLVM_DEBUG(dbgs() << "can Sink? : "; Store0->dump(); dbgs() << "\n");
186 BasicBlock *BB0 = Store0->getParent();
187 for (Instruction &Inst : reverse(*BB1)) {
188 auto *Store1 = dyn_cast<StoreInst>(&Inst);
189 if (!Store1)
190 continue;
191
192 MemoryLocation Loc0 = MemoryLocation::get(Store0);
193 MemoryLocation Loc1 = MemoryLocation::get(Store1);
194
195 if (AA->isMustAlias(Loc0, Loc1) &&
196 !isStoreSinkBarrierInRange(*Store1->getNextNode(), BB1->back(), Loc1) &&
197 !isStoreSinkBarrierInRange(*Store0->getNextNode(), BB0->back(), Loc0) &&
198 Store0->hasSameSpecialState(Store1) &&
200 Store0->getValueOperand()->getType(),
201 Store1->getValueOperand()->getType(),
202 Store0->getDataLayout()))
203 return Store1;
204 }
205 return nullptr;
206}
207
208///
209/// Create a PHI node in BB for the operands of S0 and S1
210///
211PHINode *MergedLoadStoreMotion::getPHIOperand(BasicBlock *BB, StoreInst *S0,
212 StoreInst *S1) {
213 // Create a phi if the values mismatch.
214 Value *Opd1 = S0->getValueOperand();
215 Value *Opd2 = S1->getValueOperand();
216 if (Opd1 == Opd2)
217 return nullptr;
218
219 auto *NewPN = PHINode::Create(Opd1->getType(), 2, Opd2->getName() + ".sink");
220 NewPN->insertBefore(BB->begin());
221 NewPN->applyMergedLocation(S0->getDebugLoc(), S1->getDebugLoc());
222 NewPN->addIncoming(Opd1, S0->getParent());
223 NewPN->addIncoming(Opd2, S1->getParent());
224 return NewPN;
225}
226
227///
228/// Check if 2 stores can be sunk, optionally together with corresponding GEPs.
229///
230bool MergedLoadStoreMotion::canSinkStoresAndGEPs(StoreInst *S0,
231 StoreInst *S1) const {
232 if (S0->getPointerOperand() == S1->getPointerOperand())
233 return true;
234 auto *GEP0 = dyn_cast<GetElementPtrInst>(S0->getPointerOperand());
235 auto *GEP1 = dyn_cast<GetElementPtrInst>(S1->getPointerOperand());
236 return GEP0 && GEP1 && GEP0->isIdenticalTo(GEP1) && GEP0->hasOneUse() &&
237 (GEP0->getParent() == S0->getParent()) && GEP1->hasOneUse() &&
238 (GEP1->getParent() == S1->getParent());
239}
240
241///
242/// Merge two stores to same address and sink into \p BB
243///
244/// Optionally also sinks GEP instruction computing the store address
245///
246void MergedLoadStoreMotion::sinkStoresAndGEPs(BasicBlock *BB, StoreInst *S0,
247 StoreInst *S1) {
248 Value *Ptr0 = S0->getPointerOperand();
249 Value *Ptr1 = S1->getPointerOperand();
250 // Only one definition?
251 LLVM_DEBUG(dbgs() << "Sink Instruction into BB \n"; BB->dump();
252 dbgs() << "Instruction Left\n"; S0->dump(); dbgs() << "\n";
253 dbgs() << "Instruction Right\n"; S1->dump(); dbgs() << "\n");
254 // Hoist the instruction.
256 // Intersect optional metadata.
257 S0->andIRFlags(S1);
259 S0->applyMergedLocation(S0->getDebugLoc(), S1->getDebugLoc());
260 S0->mergeDIAssignID(S1);
261
262 // Insert bitcast for conflicting typed stores (or just use original value if
263 // same type).
264 IRBuilder<> Builder(S0);
265 auto Cast = Builder.CreateBitOrPointerCast(S0->getValueOperand(),
266 S1->getValueOperand()->getType());
267 S0->setOperand(0, Cast);
268
269 // Create the new store to be inserted at the join point.
270 StoreInst *SNew = cast<StoreInst>(S0->clone());
271 SNew->insertBefore(InsertPt);
272 // New PHI operand? Use it.
273 if (PHINode *NewPN = getPHIOperand(BB, S0, S1))
274 SNew->setOperand(0, NewPN);
275 S0->eraseFromParent();
276 S1->eraseFromParent();
277
278 if (Ptr0 != Ptr1) {
279 auto *GEP0 = cast<GetElementPtrInst>(Ptr0);
280 auto *GEP1 = cast<GetElementPtrInst>(Ptr1);
281 Instruction *GEPNew = GEP0->clone();
282 GEPNew->insertBefore(SNew);
283 GEPNew->applyMergedLocation(GEP0->getDebugLoc(), GEP1->getDebugLoc());
284 SNew->setOperand(1, GEPNew);
285 GEP0->replaceAllUsesWith(GEPNew);
286 GEP0->eraseFromParent();
287 GEP1->replaceAllUsesWith(GEPNew);
288 GEP1->eraseFromParent();
289 }
290}
291
292///
293/// True when two stores are equivalent and can sink into the footer
294///
295/// Starting from a diamond head block, iterate over the instructions in one
296/// successor block and try to match a store in the second successor.
297///
298bool MergedLoadStoreMotion::mergeStores(BasicBlock *HeadBB) {
299
300 bool MergedStores = false;
301 BasicBlock *TailBB = getDiamondTail(HeadBB);
302 BasicBlock *SinkBB = TailBB;
303 assert(SinkBB && "Footer of a diamond cannot be empty");
304
305 succ_iterator SI = succ_begin(HeadBB);
306 assert(SI != succ_end(HeadBB) && "Diamond head cannot have zero successors");
307 BasicBlock *Pred0 = *SI;
308 ++SI;
309 assert(SI != succ_end(HeadBB) && "Diamond head cannot have single successor");
310 BasicBlock *Pred1 = *SI;
311 // tail block of a diamond/hammock?
312 if (Pred0 == Pred1)
313 return false; // No.
314 // bail out early if we can not merge into the footer BB
315 if (!SplitFooterBB && TailBB->hasNPredecessorsOrMore(3))
316 return false;
317 // #Instructions in Pred1 for Compile Time Control
318 auto InstsNoDbg = Pred1->instructionsWithoutDebug();
319 int Size1 = std::distance(InstsNoDbg.begin(), InstsNoDbg.end());
320 int NStores = 0;
321
322 for (BasicBlock::reverse_iterator RBI = Pred0->rbegin(), RBE = Pred0->rend();
323 RBI != RBE;) {
324
325 Instruction *I = &*RBI;
326 ++RBI;
327
328 // Don't sink non-simple (atomic, volatile) stores.
329 auto *S0 = dyn_cast<StoreInst>(I);
330 if (!S0 || !S0->isSimple())
331 continue;
332
333 ++NStores;
334 if (NStores * Size1 >= MagicCompileTimeControl)
335 break;
336 if (StoreInst *S1 = canSinkFromBlock(Pred1, S0)) {
337 if (!canSinkStoresAndGEPs(S0, S1))
338 // Don't attempt to sink below stores that had to stick around
339 // But after removal of a store and some of its feeding
340 // instruction search again from the beginning since the iterator
341 // is likely stale at this point.
342 break;
343
344 if (SinkBB == TailBB && TailBB->hasNPredecessorsOrMore(3)) {
345 // We have more than 2 predecessors. Insert a new block
346 // postdominating 2 predecessors we're going to sink from.
347 SinkBB = SplitBlockPredecessors(TailBB, {Pred0, Pred1}, ".sink.split");
348 if (!SinkBB)
349 break;
350 }
351
352 MergedStores = true;
353 sinkStoresAndGEPs(SinkBB, S0, S1);
354 RBI = Pred0->rbegin();
355 RBE = Pred0->rend();
356 LLVM_DEBUG(dbgs() << "Search again\n"; Instruction *I = &*RBI; I->dump());
357 }
358 }
359 return MergedStores;
360}
361
362bool MergedLoadStoreMotion::run(Function &F, AliasAnalysis &AA) {
363 this->AA = &AA;
364
365 bool Changed = false;
366 LLVM_DEBUG(dbgs() << "Instruction Merger\n");
367
368 // Merge unconditional branches, allowing PRE to catch more
369 // optimization opportunities.
370 // This loop doesn't care about newly inserted/split blocks
371 // since they never will be diamond heads.
373 // Hoist equivalent loads and sink stores
374 // outside diamonds when possible
375 if (isDiamondHead(&BB))
376 Changed |= mergeStores(&BB);
377 return Changed;
378}
379
382 MergedLoadStoreMotion Impl(Options.SplitFooterBB);
383 auto &AA = AM.getResult<AAManager>(F);
384 if (!Impl.run(F, AA))
385 return PreservedAnalyses::all();
386
388 if (!Options.SplitFooterBB)
390 return PA;
391}
392
394 raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) {
396 OS, MapClassName2PassName);
397 OS << '<';
398 OS << (Options.SplitFooterBB ? "" : "no-") << "split-footer-bb";
399 OS << '>';
400}
static const LLT S1
#define LLVM_DEBUG(X)
Definition: Debug.h:101
bool End
Definition: ELF_riscv.cpp:480
This is the interface for a simple mod/ref and alias analysis over globals.
#define F(x, y, z)
Definition: MD5.cpp:55
#define I(x, y, z)
Definition: MD5.cpp:58
This pass performs merges of loads and stores on both sides of a.
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
raw_pwrite_stream & OS
A manager for alias analyses.
A container for analyses that lazily runs them and caches their results.
Definition: PassManager.h:253
PassT::Result & getResult(IRUnitT &IR, ExtraArgTs... ExtraArgs)
Get the result of an analysis pass for a given IR unit.
Definition: PassManager.h:405
LLVM Basic Block Representation.
Definition: BasicBlock.h:61
iterator begin()
Instruction iterator methods.
Definition: BasicBlock.h:448
const_iterator getFirstInsertionPt() const
Returns an iterator to the first instruction in this block that is suitable for inserting a non-PHI i...
Definition: BasicBlock.cpp:416
reverse_iterator rbegin()
Definition: BasicBlock.h:464
iterator_range< filter_iterator< BasicBlock::const_iterator, std::function< bool(const Instruction &)> > > instructionsWithoutDebug(bool SkipPseudoOp=true) const
Return a const iterator range over the instructions in the block, skipping any debug instructions.
Definition: BasicBlock.cpp:250
const BasicBlock * getSinglePredecessor() const
Return the predecessor of this block if it has a single predecessor block.
Definition: BasicBlock.cpp:459
InstListType::reverse_iterator reverse_iterator
Definition: BasicBlock.h:179
const BasicBlock * getSingleSuccessor() const
Return the successor of this block if it has a single successor.
Definition: BasicBlock.cpp:489
reverse_iterator rend()
Definition: BasicBlock.h:466
InstListType::iterator iterator
Instruction iterators...
Definition: BasicBlock.h:177
const Instruction * getTerminator() const LLVM_READONLY
Returns the terminator instruction if the block is well formed or null if the block is not well forme...
Definition: BasicBlock.h:239
bool hasNPredecessorsOrMore(unsigned N) const
Return true if this block has N predecessors or more.
Definition: BasicBlock.cpp:485
const Instruction & back() const
Definition: BasicBlock.h:473
Represents analyses that only rely on functions' control flow.
Definition: Analysis.h:72
static bool isBitOrNoopPointerCastable(Type *SrcTy, Type *DestTy, const DataLayout &DL)
Check whether a bitcast, inttoptr, or ptrtoint cast between these types is valid and a no-op.
This provides a uniform API for creating instructions and inserting them into a basic block: either a...
Definition: IRBuilder.h:2686
Instruction * clone() const
Create a copy of 'this' instruction that is identical in all ways except the following:
void mergeDIAssignID(ArrayRef< const Instruction * > SourceInstructions)
Merge the DIAssignID metadata from this instruction and those attached to instructions in SourceInstr...
Definition: DebugInfo.cpp:953
void insertBefore(Instruction *InsertPos)
Insert an unlinked instruction into a basic block immediately before the specified instruction.
Definition: Instruction.cpp:97
const DebugLoc & getDebugLoc() const
Return the debug location for this node as a DebugLoc.
Definition: Instruction.h:466
void andIRFlags(const Value *V)
Logical 'and' of any supported wrapping, exact, and fast-math flags of V and this instruction.
InstListType::iterator eraseFromParent()
This method unlinks 'this' from the containing basic block and deletes it.
Definition: Instruction.cpp:92
BasicBlock * getSuccessor(unsigned Idx) const LLVM_READONLY
Return the specified successor. This instruction must be a terminator.
bool hasSameSpecialState(const Instruction *I2, bool IgnoreAlignment=false) const LLVM_READONLY
This function determines if the speficied instruction has the same "special" characteristics as the c...
void dropUnknownNonDebugMetadata(ArrayRef< unsigned > KnownIDs=std::nullopt)
Drop all unknown metadata except for debug locations.
Definition: Metadata.cpp:1596
void applyMergedLocation(DILocation *LocA, DILocation *LocB)
Merge 2 debug locations and apply it to the Instruction.
Definition: DebugInfo.cpp:949
const DataLayout & getDataLayout() const
Get the data layout of the module this instruction belongs to.
Definition: Instruction.cpp:74
LLVM_DUMP_METHOD void dump() const
Representation for a specific memory location.
static MemoryLocation get(const LoadInst *LI)
Return a location with information about the memory reference by the given instruction.
void printPipeline(raw_ostream &OS, function_ref< StringRef(StringRef)> MapClassName2PassName)
PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM)
static PHINode * Create(Type *Ty, unsigned NumReservedValues, const Twine &NameStr="", InsertPosition InsertBefore=nullptr)
Constructors - NumReservedValues is a hint for the number of incoming edges that this phi node will h...
A set of analyses that are preserved following a run of a transformation pass.
Definition: Analysis.h:111
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: Analysis.h:117
void preserveSet()
Mark an analysis set as preserved.
Definition: Analysis.h:146
An instruction for storing to memory.
Definition: Instructions.h:290
bool isSimple() const
Definition: Instructions.h:366
Value * getValueOperand()
Definition: Instructions.h:374
Value * getPointerOperand()
Definition: Instructions.h:377
StringRef - Represent a constant reference to a string, i.e.
Definition: StringRef.h:50
void setOperand(unsigned i, Value *Val)
Definition: User.h:174
LLVM Value Representation.
Definition: Value.h:74
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:255
StringRef getName() const
Return a constant reference to the value's name.
Definition: Value.cpp:309
void dump() const
Support for debugging, callable in GDB: V->dump()
Definition: AsmWriter.cpp:5266
An efficient, type-erasing, non-owning reference to a callable.
const ParentTy * getParent() const
Definition: ilist_node.h:32
NodeTy * getNextNode()
Get the next node, or nullptr for the list tail.
Definition: ilist_node.h:353
This class implements an extremely fast bulk output stream that can only output to a stream.
Definition: raw_ostream.h:52
This is an optimization pass for GlobalISel generic memory operations.
Definition: AddressRanges.h:18
iterator_range< T > make_range(T x, T y)
Convenience function for iterating over sub-ranges.
iterator_range< early_inc_iterator_impl< detail::IterOfRange< RangeT > > > make_early_inc_range(RangeT &&Range)
Make a range that does early increment to allow mutation of the underlying range without disrupting i...
Definition: STLExtras.h:656
auto reverse(ContainerTy &&C)
Definition: STLExtras.h:419
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:163
RNSuccIterator< NodeRef, BlockT, RegionT > succ_begin(NodeRef Node)
BasicBlock * SplitBlockPredecessors(BasicBlock *BB, ArrayRef< BasicBlock * > Preds, const char *Suffix, DominatorTree *DT, LoopInfo *LI=nullptr, MemorySSAUpdater *MSSAU=nullptr, bool PreserveLCSSA=false)
This method introduces at least one new basic block into the function and moves some of the predecess...
RNSuccIterator< NodeRef, BlockT, RegionT > succ_end(NodeRef Node)
A CRTP mix-in to automatically provide informational APIs needed for passes.
Definition: PassManager.h:69