xref: /llvm-project/llvm/lib/Transforms/Scalar/MergedLoadStoreMotion.cpp (revision d34e60ca8532511acb8c93ef26297e349fbec86a)
1 //===- MergedLoadStoreMotion.cpp - merge and hoist/sink load/stores -------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 //! \file
11 //! This pass performs merges of loads and stores on both sides of a
12 //  diamond (hammock). It hoists the loads and sinks the stores.
13 //
14 // The algorithm iteratively hoists two loads to the same address out of a
15 // diamond (hammock) and merges them into a single load in the header. Similar
16 // it sinks and merges two stores to the tail block (footer). The algorithm
17 // iterates over the instructions of one side of the diamond and attempts to
18 // find a matching load/store on the other side. It hoists / sinks when it
19 // thinks it safe to do so.  This optimization helps with eg. hiding load
20 // latencies, triggering if-conversion, and reducing static code size.
21 //
22 // NOTE: This code no longer performs load hoisting, it is subsumed by GVNHoist.
23 //
24 //===----------------------------------------------------------------------===//
25 //
26 //
27 // Example:
28 // Diamond shaped code before merge:
29 //
30 //            header:
31 //                     br %cond, label %if.then, label %if.else
32 //                        +                    +
33 //                       +                      +
34 //                      +                        +
35 //            if.then:                         if.else:
36 //               %lt = load %addr_l               %le = load %addr_l
37 //               <use %lt>                        <use %le>
38 //               <...>                            <...>
39 //               store %st, %addr_s               store %se, %addr_s
40 //               br label %if.end                 br label %if.end
41 //                     +                         +
42 //                      +                       +
43 //                       +                     +
44 //            if.end ("footer"):
45 //                     <...>
46 //
47 // Diamond shaped code after merge:
48 //
49 //            header:
50 //                     %l = load %addr_l
51 //                     br %cond, label %if.then, label %if.else
52 //                        +                    +
53 //                       +                      +
54 //                      +                        +
55 //            if.then:                         if.else:
56 //               <use %l>                         <use %l>
57 //               <...>                            <...>
58 //               br label %if.end                 br label %if.end
59 //                      +                        +
60 //                       +                      +
61 //                        +                    +
62 //            if.end ("footer"):
63 //                     %s.sink = phi [%st, if.then], [%se, if.else]
64 //                     <...>
65 //                     store %s.sink, %addr_s
66 //                     <...>
67 //
68 //
69 //===----------------------- TODO -----------------------------------------===//
70 //
71 // 1) Generalize to regions other than diamonds
72 // 2) Be more aggressive merging memory operations
73 // Note that both changes require register pressure control
74 //
75 //===----------------------------------------------------------------------===//
76 
77 #include "llvm/Transforms/Scalar/MergedLoadStoreMotion.h"
78 #include "llvm/ADT/Statistic.h"
79 #include "llvm/Analysis/AliasAnalysis.h"
80 #include "llvm/Analysis/CFG.h"
81 #include "llvm/Analysis/GlobalsModRef.h"
82 #include "llvm/Analysis/Loads.h"
83 #include "llvm/Analysis/ValueTracking.h"
84 #include "llvm/IR/Metadata.h"
85 #include "llvm/Support/Debug.h"
86 #include "llvm/Support/raw_ostream.h"
87 #include "llvm/Transforms/Scalar.h"
88 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
89 
90 using namespace llvm;
91 
92 #define DEBUG_TYPE "mldst-motion"
93 
94 namespace {
95 //===----------------------------------------------------------------------===//
96 //                         MergedLoadStoreMotion Pass
97 //===----------------------------------------------------------------------===//
98 class MergedLoadStoreMotion {
99   AliasAnalysis *AA = nullptr;
100 
101   // The mergeLoad/Store algorithms could have Size0 * Size1 complexity,
102   // where Size0 and Size1 are the #instructions on the two sides of
103   // the diamond. The constant chosen here is arbitrary. Compiler Time
104   // Control is enforced by the check Size0 * Size1 < MagicCompileTimeControl.
105   const int MagicCompileTimeControl = 250;
106 
107 public:
108   bool run(Function &F, AliasAnalysis &AA);
109 
110 private:
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 sinkStore(BasicBlock *BB, StoreInst *SinkCand, StoreInst *ElseInst);
119   bool mergeStores(BasicBlock *BB);
120 };
121 } // end anonymous namespace
122 
123 ///
124 /// Return tail block of a diamond.
125 ///
126 BasicBlock *MergedLoadStoreMotion::getDiamondTail(BasicBlock *BB) {
127   assert(isDiamondHead(BB) && "Basic block is not head of a diamond");
128   return BB->getTerminator()->getSuccessor(0)->getSingleSuccessor();
129 }
130 
131 ///
132 /// True when BB is the head of a diamond (hammock)
133 ///
134 bool MergedLoadStoreMotion::isDiamondHead(BasicBlock *BB) {
135   if (!BB)
136     return false;
137   auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
138   if (!BI || !BI->isConditional())
139     return false;
140 
141   BasicBlock *Succ0 = BI->getSuccessor(0);
142   BasicBlock *Succ1 = BI->getSuccessor(1);
143 
144   if (!Succ0->getSinglePredecessor())
145     return false;
146   if (!Succ1->getSinglePredecessor())
147     return false;
148 
149   BasicBlock *Succ0Succ = Succ0->getSingleSuccessor();
150   BasicBlock *Succ1Succ = Succ1->getSingleSuccessor();
151   // Ignore triangles.
152   if (!Succ0Succ || !Succ1Succ || Succ0Succ != Succ1Succ)
153     return false;
154   return true;
155 }
156 
157 
158 ///
159 /// True when instruction is a sink barrier for a store
160 /// located in Loc
161 ///
162 /// Whenever an instruction could possibly read or modify the
163 /// value being stored or protect against the store from
164 /// happening it is considered a sink barrier.
165 ///
166 bool MergedLoadStoreMotion::isStoreSinkBarrierInRange(const Instruction &Start,
167                                                       const Instruction &End,
168                                                       MemoryLocation Loc) {
169   for (const Instruction &Inst :
170        make_range(Start.getIterator(), End.getIterator()))
171     if (Inst.mayThrow())
172       return true;
173   return AA->canInstructionRangeModRef(Start, End, Loc, ModRefInfo::ModRef);
174 }
175 
176 ///
177 /// Check if \p BB contains a store to the same address as \p SI
178 ///
179 /// \return The store in \p  when it is safe to sink. Otherwise return Null.
180 ///
181 StoreInst *MergedLoadStoreMotion::canSinkFromBlock(BasicBlock *BB1,
182                                                    StoreInst *Store0) {
183   LLVM_DEBUG(dbgs() << "can Sink? : "; Store0->dump(); dbgs() << "\n");
184   BasicBlock *BB0 = Store0->getParent();
185   for (Instruction &Inst : reverse(*BB1)) {
186     auto *Store1 = dyn_cast<StoreInst>(&Inst);
187     if (!Store1)
188       continue;
189 
190     MemoryLocation Loc0 = MemoryLocation::get(Store0);
191     MemoryLocation Loc1 = MemoryLocation::get(Store1);
192     if (AA->isMustAlias(Loc0, Loc1) && Store0->isSameOperationAs(Store1) &&
193         !isStoreSinkBarrierInRange(*Store1->getNextNode(), BB1->back(), Loc1) &&
194         !isStoreSinkBarrierInRange(*Store0->getNextNode(), BB0->back(), Loc0)) {
195       return Store1;
196     }
197   }
198   return nullptr;
199 }
200 
201 ///
202 /// Create a PHI node in BB for the operands of S0 and S1
203 ///
204 PHINode *MergedLoadStoreMotion::getPHIOperand(BasicBlock *BB, StoreInst *S0,
205                                               StoreInst *S1) {
206   // Create a phi if the values mismatch.
207   Value *Opd1 = S0->getValueOperand();
208   Value *Opd2 = S1->getValueOperand();
209   if (Opd1 == Opd2)
210     return nullptr;
211 
212   auto *NewPN = PHINode::Create(Opd1->getType(), 2, Opd2->getName() + ".sink",
213                                 &BB->front());
214   NewPN->addIncoming(Opd1, S0->getParent());
215   NewPN->addIncoming(Opd2, S1->getParent());
216   return NewPN;
217 }
218 
219 ///
220 /// Merge two stores to same address and sink into \p BB
221 ///
222 /// Also sinks GEP instruction computing the store address
223 ///
224 bool MergedLoadStoreMotion::sinkStore(BasicBlock *BB, StoreInst *S0,
225                                       StoreInst *S1) {
226   // Only one definition?
227   auto *A0 = dyn_cast<Instruction>(S0->getPointerOperand());
228   auto *A1 = dyn_cast<Instruction>(S1->getPointerOperand());
229   if (A0 && A1 && A0->isIdenticalTo(A1) && A0->hasOneUse() &&
230       (A0->getParent() == S0->getParent()) && A1->hasOneUse() &&
231       (A1->getParent() == S1->getParent()) && isa<GetElementPtrInst>(A0)) {
232     LLVM_DEBUG(dbgs() << "Sink Instruction into BB \n"; BB->dump();
233                dbgs() << "Instruction Left\n"; S0->dump(); dbgs() << "\n";
234                dbgs() << "Instruction Right\n"; S1->dump(); dbgs() << "\n");
235     // Hoist the instruction.
236     BasicBlock::iterator InsertPt = BB->getFirstInsertionPt();
237     // Intersect optional metadata.
238     S0->andIRFlags(S1);
239     S0->dropUnknownNonDebugMetadata();
240 
241     // Create the new store to be inserted at the join point.
242     StoreInst *SNew = cast<StoreInst>(S0->clone());
243     Instruction *ANew = A0->clone();
244     SNew->insertBefore(&*InsertPt);
245     ANew->insertBefore(SNew);
246 
247     assert(S0->getParent() == A0->getParent());
248     assert(S1->getParent() == A1->getParent());
249 
250     // New PHI operand? Use it.
251     if (PHINode *NewPN = getPHIOperand(BB, S0, S1))
252       SNew->setOperand(0, NewPN);
253     S0->eraseFromParent();
254     S1->eraseFromParent();
255     A0->replaceAllUsesWith(ANew);
256     A0->eraseFromParent();
257     A1->replaceAllUsesWith(ANew);
258     A1->eraseFromParent();
259     return true;
260   }
261   return false;
262 }
263 
264 ///
265 /// True when two stores are equivalent and can sink into the footer
266 ///
267 /// Starting from a diamond tail block, iterate over the instructions in one
268 /// predecessor block and try to match a store in the second predecessor.
269 ///
270 bool MergedLoadStoreMotion::mergeStores(BasicBlock *T) {
271 
272   bool MergedStores = false;
273   assert(T && "Footer of a diamond cannot be empty");
274 
275   pred_iterator PI = pred_begin(T), E = pred_end(T);
276   assert(PI != E);
277   BasicBlock *Pred0 = *PI;
278   ++PI;
279   BasicBlock *Pred1 = *PI;
280   ++PI;
281   // tail block  of a diamond/hammock?
282   if (Pred0 == Pred1)
283     return false; // No.
284   if (PI != E)
285     return false; // No. More than 2 predecessors.
286 
287   // #Instructions in Succ1 for Compile Time Control
288   int Size1 = distance(Pred1->instructionsWithoutDebug());
289   int NStores = 0;
290 
291   for (BasicBlock::reverse_iterator RBI = Pred0->rbegin(), RBE = Pred0->rend();
292        RBI != RBE;) {
293 
294     Instruction *I = &*RBI;
295     ++RBI;
296 
297     // Don't sink non-simple (atomic, volatile) stores.
298     auto *S0 = dyn_cast<StoreInst>(I);
299     if (!S0 || !S0->isSimple())
300       continue;
301 
302     ++NStores;
303     if (NStores * Size1 >= MagicCompileTimeControl)
304       break;
305     if (StoreInst *S1 = canSinkFromBlock(Pred1, S0)) {
306       bool Res = sinkStore(T, S0, S1);
307       MergedStores |= Res;
308       // Don't attempt to sink below stores that had to stick around
309       // But after removal of a store and some of its feeding
310       // instruction search again from the beginning since the iterator
311       // is likely stale at this point.
312       if (!Res)
313         break;
314       RBI = Pred0->rbegin();
315       RBE = Pred0->rend();
316       LLVM_DEBUG(dbgs() << "Search again\n"; Instruction *I = &*RBI; I->dump());
317     }
318   }
319   return MergedStores;
320 }
321 
322 bool MergedLoadStoreMotion::run(Function &F, AliasAnalysis &AA) {
323   this->AA = &AA;
324 
325   bool Changed = false;
326   LLVM_DEBUG(dbgs() << "Instruction Merger\n");
327 
328   // Merge unconditional branches, allowing PRE to catch more
329   // optimization opportunities.
330   for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE;) {
331     BasicBlock *BB = &*FI++;
332 
333     // Hoist equivalent loads and sink stores
334     // outside diamonds when possible
335     if (isDiamondHead(BB)) {
336       Changed |= mergeStores(getDiamondTail(BB));
337     }
338   }
339   return Changed;
340 }
341 
342 namespace {
343 class MergedLoadStoreMotionLegacyPass : public FunctionPass {
344 public:
345   static char ID; // Pass identification, replacement for typeid
346   MergedLoadStoreMotionLegacyPass() : FunctionPass(ID) {
347     initializeMergedLoadStoreMotionLegacyPassPass(
348         *PassRegistry::getPassRegistry());
349   }
350 
351   ///
352   /// Run the transformation for each function
353   ///
354   bool runOnFunction(Function &F) override {
355     if (skipFunction(F))
356       return false;
357     MergedLoadStoreMotion Impl;
358     return Impl.run(F, getAnalysis<AAResultsWrapperPass>().getAAResults());
359   }
360 
361 private:
362   void getAnalysisUsage(AnalysisUsage &AU) const override {
363     AU.setPreservesCFG();
364     AU.addRequired<AAResultsWrapperPass>();
365     AU.addPreserved<GlobalsAAWrapperPass>();
366   }
367 };
368 
369 char MergedLoadStoreMotionLegacyPass::ID = 0;
370 } // anonymous namespace
371 
372 ///
373 /// createMergedLoadStoreMotionPass - The public interface to this file.
374 ///
375 FunctionPass *llvm::createMergedLoadStoreMotionPass() {
376   return new MergedLoadStoreMotionLegacyPass();
377 }
378 
379 INITIALIZE_PASS_BEGIN(MergedLoadStoreMotionLegacyPass, "mldst-motion",
380                       "MergedLoadStoreMotion", false, false)
381 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
382 INITIALIZE_PASS_END(MergedLoadStoreMotionLegacyPass, "mldst-motion",
383                     "MergedLoadStoreMotion", false, false)
384 
385 PreservedAnalyses
386 MergedLoadStoreMotionPass::run(Function &F, FunctionAnalysisManager &AM) {
387   MergedLoadStoreMotion Impl;
388   auto &AA = AM.getResult<AAManager>(F);
389   if (!Impl.run(F, AA))
390     return PreservedAnalyses::all();
391 
392   PreservedAnalyses PA;
393   PA.preserveSet<CFGAnalyses>();
394   PA.preserve<GlobalsAA>();
395   return PA;
396 }
397