xref: /llvm-project/llvm/lib/Transforms/Scalar/MergedLoadStoreMotion.cpp (revision 34b139594aa20fe712bc2ad68544632b3e4d8512)
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 
78 #include "llvm/Transforms/Scalar/MergedLoadStoreMotion.h"
79 #include "llvm/Analysis/AliasAnalysis.h"
80 #include "llvm/Analysis/GlobalsModRef.h"
81 #include "llvm/IR/IRBuilder.h"
82 #include "llvm/IR/Instructions.h"
83 #include "llvm/Support/Debug.h"
84 #include "llvm/Support/raw_ostream.h"
85 #include "llvm/Transforms/Scalar.h"
86 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
87 #include "llvm/Transforms/Utils/Local.h"
88 
89 using namespace llvm;
90 
91 #define DEBUG_TYPE "mldst-motion"
92 
93 namespace {
94 //===----------------------------------------------------------------------===//
95 //                         MergedLoadStoreMotion Pass
96 //===----------------------------------------------------------------------===//
97 class MergedLoadStoreMotion {
98   AliasAnalysis *AA = nullptr;
99 
100   // The mergeLoad/Store algorithms could have Size0 * Size1 complexity,
101   // where Size0 and Size1 are the #instructions on the two sides of
102   // the diamond. The constant chosen here is arbitrary. Compiler Time
103   // Control is enforced by the check Size0 * Size1 < MagicCompileTimeControl.
104   const int MagicCompileTimeControl = 250;
105 
106   const bool SplitFooterBB;
107 public:
108   MergedLoadStoreMotion(bool SplitFooterBB) : SplitFooterBB(SplitFooterBB) {}
109   bool run(Function &F, AliasAnalysis &AA);
110 
111 private:
112   BasicBlock *getDiamondTail(BasicBlock *BB);
113   bool isDiamondHead(BasicBlock *BB);
114   // Routines for sinking stores
115   StoreInst *canSinkFromBlock(BasicBlock *BB, StoreInst *SI);
116   PHINode *getPHIOperand(BasicBlock *BB, StoreInst *S0, StoreInst *S1);
117   bool isStoreSinkBarrierInRange(const Instruction &Start,
118                                  const Instruction &End, MemoryLocation Loc);
119   bool canSinkStoresAndGEPs(StoreInst *S0, StoreInst *S1) const;
120   void sinkStoresAndGEPs(BasicBlock *BB, StoreInst *SinkCand,
121                          StoreInst *ElseInst);
122   bool mergeStores(BasicBlock *BB);
123 };
124 } // end anonymous namespace
125 
126 ///
127 /// Return tail block of a diamond.
128 ///
129 BasicBlock *MergedLoadStoreMotion::getDiamondTail(BasicBlock *BB) {
130   assert(isDiamondHead(BB) && "Basic block is not head of a diamond");
131   return BB->getTerminator()->getSuccessor(0)->getSingleSuccessor();
132 }
133 
134 ///
135 /// True when BB is the head of a diamond (hammock)
136 ///
137 bool MergedLoadStoreMotion::isDiamondHead(BasicBlock *BB) {
138   if (!BB)
139     return false;
140   auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
141   if (!BI || !BI->isConditional())
142     return false;
143 
144   BasicBlock *Succ0 = BI->getSuccessor(0);
145   BasicBlock *Succ1 = BI->getSuccessor(1);
146 
147   if (!Succ0->getSinglePredecessor())
148     return false;
149   if (!Succ1->getSinglePredecessor())
150     return false;
151 
152   BasicBlock *Succ0Succ = Succ0->getSingleSuccessor();
153   BasicBlock *Succ1Succ = Succ1->getSingleSuccessor();
154   // Ignore triangles.
155   if (!Succ0Succ || !Succ1Succ || Succ0Succ != Succ1Succ)
156     return false;
157   return true;
158 }
159 
160 
161 ///
162 /// True when instruction is a sink barrier for a store
163 /// located in Loc
164 ///
165 /// Whenever an instruction could possibly read or modify the
166 /// value being stored or protect against the store from
167 /// happening it is considered a sink barrier.
168 ///
169 bool MergedLoadStoreMotion::isStoreSinkBarrierInRange(const Instruction &Start,
170                                                       const Instruction &End,
171                                                       MemoryLocation Loc) {
172   for (const Instruction &Inst :
173        make_range(Start.getIterator(), End.getIterator()))
174     if (Inst.mayThrow())
175       return true;
176   return AA->canInstructionRangeModRef(Start, End, Loc, ModRefInfo::ModRef);
177 }
178 
179 ///
180 /// Check if \p BB contains a store to the same address as \p SI
181 ///
182 /// \return The store in \p  when it is safe to sink. Otherwise return Null.
183 ///
184 StoreInst *MergedLoadStoreMotion::canSinkFromBlock(BasicBlock *BB1,
185                                                    StoreInst *Store0) {
186   LLVM_DEBUG(dbgs() << "can Sink? : "; Store0->dump(); dbgs() << "\n");
187   BasicBlock *BB0 = Store0->getParent();
188   for (Instruction &Inst : reverse(*BB1)) {
189     auto *Store1 = dyn_cast<StoreInst>(&Inst);
190     if (!Store1)
191       continue;
192 
193     MemoryLocation Loc0 = MemoryLocation::get(Store0);
194     MemoryLocation Loc1 = MemoryLocation::get(Store1);
195 
196     if (AA->isMustAlias(Loc0, Loc1) &&
197         !isStoreSinkBarrierInRange(*Store1->getNextNode(), BB1->back(), Loc1) &&
198         !isStoreSinkBarrierInRange(*Store0->getNextNode(), BB0->back(), Loc0) &&
199         Store0->hasSameSpecialState(Store1) &&
200         CastInst::isBitOrNoopPointerCastable(
201             Store0->getValueOperand()->getType(),
202             Store1->getValueOperand()->getType(),
203             Store0->getDataLayout()))
204       return Store1;
205   }
206   return nullptr;
207 }
208 
209 ///
210 /// Create a PHI node in BB for the operands of S0 and S1
211 ///
212 PHINode *MergedLoadStoreMotion::getPHIOperand(BasicBlock *BB, StoreInst *S0,
213                                               StoreInst *S1) {
214   // Create a phi if the values mismatch.
215   Value *Opd1 = S0->getValueOperand();
216   Value *Opd2 = S1->getValueOperand();
217   if (Opd1 == Opd2)
218     return nullptr;
219 
220   auto *NewPN = PHINode::Create(Opd1->getType(), 2, Opd2->getName() + ".sink");
221   NewPN->insertBefore(BB->begin());
222   NewPN->applyMergedLocation(S0->getDebugLoc(), S1->getDebugLoc());
223   NewPN->addIncoming(Opd1, S0->getParent());
224   NewPN->addIncoming(Opd2, S1->getParent());
225   return NewPN;
226 }
227 
228 ///
229 /// Check if 2 stores can be sunk, optionally together with corresponding GEPs.
230 ///
231 bool MergedLoadStoreMotion::canSinkStoresAndGEPs(StoreInst *S0,
232                                                  StoreInst *S1) const {
233   if (S0->getPointerOperand() == S1->getPointerOperand())
234     return true;
235   auto *GEP0 = dyn_cast<GetElementPtrInst>(S0->getPointerOperand());
236   auto *GEP1 = dyn_cast<GetElementPtrInst>(S1->getPointerOperand());
237   return GEP0 && GEP1 && GEP0->isIdenticalTo(GEP1) && GEP0->hasOneUse() &&
238          (GEP0->getParent() == S0->getParent()) && GEP1->hasOneUse() &&
239          (GEP1->getParent() == S1->getParent());
240 }
241 
242 ///
243 /// Merge two stores to same address and sink into \p BB
244 ///
245 /// Optionally also sinks GEP instruction computing the store address
246 ///
247 void MergedLoadStoreMotion::sinkStoresAndGEPs(BasicBlock *BB, StoreInst *S0,
248                                               StoreInst *S1) {
249   Value *Ptr0 = S0->getPointerOperand();
250   Value *Ptr1 = S1->getPointerOperand();
251   // Only one definition?
252   LLVM_DEBUG(dbgs() << "Sink Instruction into BB \n"; BB->dump();
253              dbgs() << "Instruction Left\n"; S0->dump(); dbgs() << "\n";
254              dbgs() << "Instruction Right\n"; S1->dump(); dbgs() << "\n");
255   // Hoist the instruction.
256   BasicBlock::iterator InsertPt = BB->getFirstInsertionPt();
257   // Intersect optional metadata.
258   S0->andIRFlags(S1);
259 
260   combineMetadataForCSE(S0, S1, true);
261   S0->applyMergedLocation(S0->getDebugLoc(), S1->getDebugLoc());
262   S0->mergeDIAssignID(S1);
263 
264   // Insert bitcast for conflicting typed stores (or just use original value if
265   // same type).
266   IRBuilder<> Builder(S0);
267   auto Cast = Builder.CreateBitOrPointerCast(S0->getValueOperand(),
268                                              S1->getValueOperand()->getType());
269   S0->setOperand(0, Cast);
270 
271   // Create the new store to be inserted at the join point.
272   StoreInst *SNew = cast<StoreInst>(S0->clone());
273   SNew->insertBefore(InsertPt);
274   // New PHI operand? Use it.
275   if (PHINode *NewPN = getPHIOperand(BB, S0, S1))
276     SNew->setOperand(0, NewPN);
277   S0->eraseFromParent();
278   S1->eraseFromParent();
279 
280   if (Ptr0 != Ptr1) {
281     auto *GEP0 = cast<GetElementPtrInst>(Ptr0);
282     auto *GEP1 = cast<GetElementPtrInst>(Ptr1);
283     Instruction *GEPNew = GEP0->clone();
284     GEPNew->insertBefore(SNew->getIterator());
285     GEPNew->applyMergedLocation(GEP0->getDebugLoc(), GEP1->getDebugLoc());
286     SNew->setOperand(1, GEPNew);
287     GEP0->replaceAllUsesWith(GEPNew);
288     GEP0->eraseFromParent();
289     GEP1->replaceAllUsesWith(GEPNew);
290     GEP1->eraseFromParent();
291   }
292 }
293 
294 ///
295 /// True when two stores are equivalent and can sink into the footer
296 ///
297 /// Starting from a diamond head block, iterate over the instructions in one
298 /// successor block and try to match a store in the second successor.
299 ///
300 bool MergedLoadStoreMotion::mergeStores(BasicBlock *HeadBB) {
301 
302   bool MergedStores = false;
303   BasicBlock *TailBB = getDiamondTail(HeadBB);
304   BasicBlock *SinkBB = TailBB;
305   assert(SinkBB && "Footer of a diamond cannot be empty");
306 
307   succ_iterator SI = succ_begin(HeadBB);
308   assert(SI != succ_end(HeadBB) && "Diamond head cannot have zero successors");
309   BasicBlock *Pred0 = *SI;
310   ++SI;
311   assert(SI != succ_end(HeadBB) && "Diamond head cannot have single successor");
312   BasicBlock *Pred1 = *SI;
313   // tail block  of a diamond/hammock?
314   if (Pred0 == Pred1)
315     return false; // No.
316   // bail out early if we can not merge into the footer BB
317   if (!SplitFooterBB && TailBB->hasNPredecessorsOrMore(3))
318     return false;
319   // #Instructions in Pred1 for Compile Time Control
320   auto InstsNoDbg = Pred1->instructionsWithoutDebug();
321   int Size1 = std::distance(InstsNoDbg.begin(), InstsNoDbg.end());
322   int NStores = 0;
323 
324   for (BasicBlock::reverse_iterator RBI = Pred0->rbegin(), RBE = Pred0->rend();
325        RBI != RBE;) {
326 
327     Instruction *I = &*RBI;
328     ++RBI;
329 
330     // Don't sink non-simple (atomic, volatile) stores.
331     auto *S0 = dyn_cast<StoreInst>(I);
332     if (!S0 || !S0->isSimple())
333       continue;
334 
335     ++NStores;
336     if (NStores * Size1 >= MagicCompileTimeControl)
337       break;
338     if (StoreInst *S1 = canSinkFromBlock(Pred1, S0)) {
339       if (!canSinkStoresAndGEPs(S0, S1))
340         // Don't attempt to sink below stores that had to stick around
341         // But after removal of a store and some of its feeding
342         // instruction search again from the beginning since the iterator
343         // is likely stale at this point.
344         break;
345 
346       if (SinkBB == TailBB && TailBB->hasNPredecessorsOrMore(3)) {
347         // We have more than 2 predecessors. Insert a new block
348         // postdominating 2 predecessors we're going to sink from.
349         SinkBB = SplitBlockPredecessors(TailBB, {Pred0, Pred1}, ".sink.split");
350         if (!SinkBB)
351           break;
352       }
353 
354       MergedStores = true;
355       sinkStoresAndGEPs(SinkBB, S0, S1);
356       RBI = Pred0->rbegin();
357       RBE = Pred0->rend();
358       LLVM_DEBUG(dbgs() << "Search again\n"; Instruction *I = &*RBI; I->dump());
359     }
360   }
361   return MergedStores;
362 }
363 
364 bool MergedLoadStoreMotion::run(Function &F, AliasAnalysis &AA) {
365   this->AA = &AA;
366 
367   bool Changed = false;
368   LLVM_DEBUG(dbgs() << "Instruction Merger\n");
369 
370   // Merge unconditional branches, allowing PRE to catch more
371   // optimization opportunities.
372   // This loop doesn't care about newly inserted/split blocks
373   // since they never will be diamond heads.
374   for (BasicBlock &BB : make_early_inc_range(F))
375     // Hoist equivalent loads and sink stores
376     // outside diamonds when possible
377     if (isDiamondHead(&BB))
378       Changed |= mergeStores(&BB);
379   return Changed;
380 }
381 
382 PreservedAnalyses
383 MergedLoadStoreMotionPass::run(Function &F, FunctionAnalysisManager &AM) {
384   MergedLoadStoreMotion Impl(Options.SplitFooterBB);
385   auto &AA = AM.getResult<AAManager>(F);
386   if (!Impl.run(F, AA))
387     return PreservedAnalyses::all();
388 
389   PreservedAnalyses PA;
390   if (!Options.SplitFooterBB)
391     PA.preserveSet<CFGAnalyses>();
392   return PA;
393 }
394 
395 void MergedLoadStoreMotionPass::printPipeline(
396     raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) {
397   static_cast<PassInfoMixin<MergedLoadStoreMotionPass> *>(this)->printPipeline(
398       OS, MapClassName2PassName);
399   OS << '<';
400   OS << (Options.SplitFooterBB ? "" : "no-") << "split-footer-bb";
401   OS << '>';
402 }
403