LLVM  10.0.0svn
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 
79 #include "llvm/ADT/Statistic.h"
81 #include "llvm/Analysis/CFG.h"
83 #include "llvm/Analysis/Loads.h"
85 #include "llvm/IR/Metadata.h"
86 #include "llvm/Support/Debug.h"
88 #include "llvm/Transforms/Scalar.h"
90 
91 using namespace llvm;
92 
93 #define DEBUG_TYPE "mldst-motion"
94 
95 namespace {
96 //===----------------------------------------------------------------------===//
97 // MergedLoadStoreMotion Pass
98 //===----------------------------------------------------------------------===//
100  AliasAnalysis *AA = nullptr;
101 
102  // The mergeLoad/Store algorithms could have Size0 * Size1 complexity,
103  // where Size0 and Size1 are the #instructions on the two sides of
104  // the diamond. The constant chosen here is arbitrary. Compiler Time
105  // Control is enforced by the check Size0 * Size1 < MagicCompileTimeControl.
106  const int MagicCompileTimeControl = 250;
107 
108  const bool SplitFooterBB;
109 public:
110  MergedLoadStoreMotion(bool SplitFooterBB) : SplitFooterBB(SplitFooterBB) {}
111  bool run(Function &F, AliasAnalysis &AA);
112 
113 private:
114  BasicBlock *getDiamondTail(BasicBlock *BB);
115  bool isDiamondHead(BasicBlock *BB);
116  // Routines for sinking stores
117  StoreInst *canSinkFromBlock(BasicBlock *BB, StoreInst *SI);
118  PHINode *getPHIOperand(BasicBlock *BB, StoreInst *S0, StoreInst *S1);
119  bool isStoreSinkBarrierInRange(const Instruction &Start,
120  const Instruction &End, MemoryLocation Loc);
121  bool canSinkStoresAndGEPs(StoreInst *S0, StoreInst *S1) const;
122  void sinkStoresAndGEPs(BasicBlock *BB, StoreInst *SinkCand,
123  StoreInst *ElseInst);
124  bool mergeStores(BasicBlock *BB);
125 };
126 } // end anonymous namespace
127 
128 ///
129 /// Return tail block of a diamond.
130 ///
131 BasicBlock *MergedLoadStoreMotion::getDiamondTail(BasicBlock *BB) {
132  assert(isDiamondHead(BB) && "Basic block is not head of a diamond");
133  return BB->getTerminator()->getSuccessor(0)->getSingleSuccessor();
134 }
135 
136 ///
137 /// True when BB is the head of a diamond (hammock)
138 ///
139 bool MergedLoadStoreMotion::isDiamondHead(BasicBlock *BB) {
140  if (!BB)
141  return false;
142  auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
143  if (!BI || !BI->isConditional())
144  return false;
145 
146  BasicBlock *Succ0 = BI->getSuccessor(0);
147  BasicBlock *Succ1 = BI->getSuccessor(1);
148 
149  if (!Succ0->getSinglePredecessor())
150  return false;
151  if (!Succ1->getSinglePredecessor())
152  return false;
153 
154  BasicBlock *Succ0Succ = Succ0->getSingleSuccessor();
155  BasicBlock *Succ1Succ = Succ1->getSingleSuccessor();
156  // Ignore triangles.
157  if (!Succ0Succ || !Succ1Succ || Succ0Succ != Succ1Succ)
158  return false;
159  return true;
160 }
161 
162 
163 ///
164 /// True when instruction is a sink barrier for a store
165 /// located in Loc
166 ///
167 /// Whenever an instruction could possibly read or modify the
168 /// value being stored or protect against the store from
169 /// happening it is considered a sink barrier.
170 ///
171 bool MergedLoadStoreMotion::isStoreSinkBarrierInRange(const Instruction &Start,
172  const Instruction &End,
173  MemoryLocation Loc) {
174  for (const Instruction &Inst :
175  make_range(Start.getIterator(), End.getIterator()))
176  if (Inst.mayThrow())
177  return true;
178  return AA->canInstructionRangeModRef(Start, End, Loc, ModRefInfo::ModRef);
179 }
180 
181 ///
182 /// Check if \p BB contains a store to the same address as \p SI
183 ///
184 /// \return The store in \p when it is safe to sink. Otherwise return Null.
185 ///
186 StoreInst *MergedLoadStoreMotion::canSinkFromBlock(BasicBlock *BB1,
187  StoreInst *Store0) {
188  LLVM_DEBUG(dbgs() << "can Sink? : "; Store0->dump(); dbgs() << "\n");
189  BasicBlock *BB0 = Store0->getParent();
190  for (Instruction &Inst : reverse(*BB1)) {
191  auto *Store1 = dyn_cast<StoreInst>(&Inst);
192  if (!Store1)
193  continue;
194 
195  MemoryLocation Loc0 = MemoryLocation::get(Store0);
196  MemoryLocation Loc1 = MemoryLocation::get(Store1);
197  if (AA->isMustAlias(Loc0, Loc1) && Store0->isSameOperationAs(Store1) &&
198  !isStoreSinkBarrierInRange(*Store1->getNextNode(), BB1->back(), Loc1) &&
199  !isStoreSinkBarrierInRange(*Store0->getNextNode(), BB0->back(), Loc0)) {
200  return Store1;
201  }
202  }
203  return nullptr;
204 }
205 
206 ///
207 /// Create a PHI node in BB for the operands of S0 and S1
208 ///
209 PHINode *MergedLoadStoreMotion::getPHIOperand(BasicBlock *BB, StoreInst *S0,
210  StoreInst *S1) {
211  // Create a phi if the values mismatch.
212  Value *Opd1 = S0->getValueOperand();
213  Value *Opd2 = S1->getValueOperand();
214  if (Opd1 == Opd2)
215  return nullptr;
216 
217  auto *NewPN = PHINode::Create(Opd1->getType(), 2, Opd2->getName() + ".sink",
218  &BB->front());
219  NewPN->applyMergedLocation(S0->getDebugLoc(), S1->getDebugLoc());
220  NewPN->addIncoming(Opd1, S0->getParent());
221  NewPN->addIncoming(Opd2, S1->getParent());
222  return NewPN;
223 }
224 
225 ///
226 /// Check if 2 stores can be sunk together with corresponding GEPs
227 ///
228 bool MergedLoadStoreMotion::canSinkStoresAndGEPs(StoreInst *S0,
229  StoreInst *S1) const {
230  auto *A0 = dyn_cast<Instruction>(S0->getPointerOperand());
231  auto *A1 = dyn_cast<Instruction>(S1->getPointerOperand());
232  return A0 && A1 && A0->isIdenticalTo(A1) && A0->hasOneUse() &&
233  (A0->getParent() == S0->getParent()) && A1->hasOneUse() &&
234  (A1->getParent() == S1->getParent()) && isa<GetElementPtrInst>(A0);
235 }
236 
237 ///
238 /// Merge two stores to same address and sink into \p BB
239 ///
240 /// Also sinks GEP instruction computing the store address
241 ///
242 void MergedLoadStoreMotion::sinkStoresAndGEPs(BasicBlock *BB, StoreInst *S0,
243  StoreInst *S1) {
244  // Only one definition?
245  auto *A0 = dyn_cast<Instruction>(S0->getPointerOperand());
246  auto *A1 = dyn_cast<Instruction>(S1->getPointerOperand());
247  LLVM_DEBUG(dbgs() << "Sink Instruction into BB \n"; BB->dump();
248  dbgs() << "Instruction Left\n"; S0->dump(); dbgs() << "\n";
249  dbgs() << "Instruction Right\n"; S1->dump(); dbgs() << "\n");
250  // Hoist the instruction.
251  BasicBlock::iterator InsertPt = BB->getFirstInsertionPt();
252  // Intersect optional metadata.
253  S0->andIRFlags(S1);
255 
256  // Create the new store to be inserted at the join point.
257  StoreInst *SNew = cast<StoreInst>(S0->clone());
258  Instruction *ANew = A0->clone();
259  SNew->insertBefore(&*InsertPt);
260  ANew->insertBefore(SNew);
261 
262  assert(S0->getParent() == A0->getParent());
263  assert(S1->getParent() == A1->getParent());
264 
265  // New PHI operand? Use it.
266  if (PHINode *NewPN = getPHIOperand(BB, S0, S1))
267  SNew->setOperand(0, NewPN);
268  S0->eraseFromParent();
269  S1->eraseFromParent();
270  A0->replaceAllUsesWith(ANew);
271  A0->eraseFromParent();
272  A1->replaceAllUsesWith(ANew);
273  A1->eraseFromParent();
274 }
275 
276 ///
277 /// True when two stores are equivalent and can sink into the footer
278 ///
279 /// Starting from a diamond head block, iterate over the instructions in one
280 /// successor block and try to match a store in the second successor.
281 ///
282 bool MergedLoadStoreMotion::mergeStores(BasicBlock *HeadBB) {
283 
284  bool MergedStores = false;
285  BasicBlock *TailBB = getDiamondTail(HeadBB);
286  BasicBlock *SinkBB = TailBB;
287  assert(SinkBB && "Footer of a diamond cannot be empty");
288 
289  succ_iterator SI = succ_begin(HeadBB);
290  assert(SI != succ_end(HeadBB) && "Diamond head cannot have zero successors");
291  BasicBlock *Pred0 = *SI;
292  ++SI;
293  assert(SI != succ_end(HeadBB) && "Diamond head cannot have single successor");
294  BasicBlock *Pred1 = *SI;
295  // tail block of a diamond/hammock?
296  if (Pred0 == Pred1)
297  return false; // No.
298  // bail out early if we can not merge into the footer BB
299  if (!SplitFooterBB && TailBB->hasNPredecessorsOrMore(3))
300  return false;
301  // #Instructions in Pred1 for Compile Time Control
302  auto InstsNoDbg = Pred1->instructionsWithoutDebug();
303  int Size1 = std::distance(InstsNoDbg.begin(), InstsNoDbg.end());
304  int NStores = 0;
305 
306  for (BasicBlock::reverse_iterator RBI = Pred0->rbegin(), RBE = Pred0->rend();
307  RBI != RBE;) {
308 
309  Instruction *I = &*RBI;
310  ++RBI;
311 
312  // Don't sink non-simple (atomic, volatile) stores.
313  auto *S0 = dyn_cast<StoreInst>(I);
314  if (!S0 || !S0->isSimple())
315  continue;
316 
317  ++NStores;
318  if (NStores * Size1 >= MagicCompileTimeControl)
319  break;
320  if (StoreInst *S1 = canSinkFromBlock(Pred1, S0)) {
321  if (!canSinkStoresAndGEPs(S0, S1))
322  // Don't attempt to sink below stores that had to stick around
323  // But after removal of a store and some of its feeding
324  // instruction search again from the beginning since the iterator
325  // is likely stale at this point.
326  break;
327 
328  if (SinkBB == TailBB && TailBB->hasNPredecessorsOrMore(3)) {
329  // We have more than 2 predecessors. Insert a new block
330  // postdominating 2 predecessors we're going to sink from.
331  SinkBB = SplitBlockPredecessors(TailBB, {Pred0, Pred1}, ".sink.split");
332  if (!SinkBB)
333  break;
334  }
335 
336  MergedStores = true;
337  sinkStoresAndGEPs(SinkBB, S0, S1);
338  RBI = Pred0->rbegin();
339  RBE = Pred0->rend();
340  LLVM_DEBUG(dbgs() << "Search again\n"; Instruction *I = &*RBI; I->dump());
341  }
342  }
343  return MergedStores;
344 }
345 
346 bool MergedLoadStoreMotion::run(Function &F, AliasAnalysis &AA) {
347  this->AA = &AA;
348 
349  bool Changed = false;
350  LLVM_DEBUG(dbgs() << "Instruction Merger\n");
351 
352  // Merge unconditional branches, allowing PRE to catch more
353  // optimization opportunities.
354  // This loop doesn't care about newly inserted/split blocks
355  // since they never will be diamond heads.
356  for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE;) {
357  BasicBlock *BB = &*FI++;
358 
359  // Hoist equivalent loads and sink stores
360  // outside diamonds when possible
361  if (isDiamondHead(BB)) {
362  Changed |= mergeStores(BB);
363  }
364  }
365  return Changed;
366 }
367 
368 namespace {
369 class MergedLoadStoreMotionLegacyPass : public FunctionPass {
370  const bool SplitFooterBB;
371 public:
372  static char ID; // Pass identification, replacement for typeid
373  MergedLoadStoreMotionLegacyPass(bool SplitFooterBB = false)
374  : FunctionPass(ID), SplitFooterBB(SplitFooterBB) {
377  }
378 
379  ///
380  /// Run the transformation for each function
381  ///
382  bool runOnFunction(Function &F) override {
383  if (skipFunction(F))
384  return false;
385  MergedLoadStoreMotion Impl(SplitFooterBB);
386  return Impl.run(F, getAnalysis<AAResultsWrapperPass>().getAAResults());
387  }
388 
389 private:
390  void getAnalysisUsage(AnalysisUsage &AU) const override {
391  if (!SplitFooterBB)
392  AU.setPreservesCFG();
395  }
396 };
397 
399 } // anonymous namespace
400 
401 ///
402 /// createMergedLoadStoreMotionPass - The public interface to this file.
403 ///
405  return new MergedLoadStoreMotionLegacyPass(SplitFooterBB);
406 }
407 
408 INITIALIZE_PASS_BEGIN(MergedLoadStoreMotionLegacyPass, "mldst-motion",
409  "MergedLoadStoreMotion", false, false)
411 INITIALIZE_PASS_END(MergedLoadStoreMotionLegacyPass, "mldst-motion",
413 
416  MergedLoadStoreMotion Impl(Options.SplitFooterBB);
417  auto &AA = AM.getResult<AAManager>(F);
418  if (!Impl.run(F, AA))
419  return PreservedAnalyses::all();
420 
422  if (!Options.SplitFooterBB)
423  PA.preserveSet<CFGAnalyses>();
424  PA.preserve<GlobalsAA>();
425  return PA;
426 }
Legacy wrapper pass to provide the GlobalsAAResult object.
The access may reference and may modify the value stored in memory.
Value * getValueOperand()
Definition: Instructions.h:409
SymbolTableList< Instruction >::iterator eraseFromParent()
This method unlinks &#39;this&#39; from the containing basic block and deletes it.
Definition: Instruction.cpp:67
AnalysisUsage & addPreserved()
Add the specified Pass class to the set of analyses preserved by this pass.
static PassRegistry * getPassRegistry()
getPassRegistry - Access the global registry object, which is automatically initialized at applicatio...
bool isSameOperationAs(const Instruction *I, unsigned flags=0) const
This function determines if the specified instruction executes the same operation as the current one...
NodeTy * getNextNode()
Get the next node, or nullptr for the list tail.
Definition: ilist_node.h:288
This class represents lattice values for constants.
Definition: AllocatorList.h:23
This is the interface for a simple mod/ref and alias analysis over globals.
void dropUnknownNonDebugMetadata(ArrayRef< unsigned > KnownIDs)
Drop all unknown metadata except for debug locations.
Definition: Metadata.cpp:1195
BasicBlock * getSuccessor(unsigned Idx) const
Return the specified successor. This instruction must be a terminator.
iterator end()
Definition: Function.h:682
INITIALIZE_PASS_BEGIN(MergedLoadStoreMotionLegacyPass, "mldst-motion", "MergedLoadStoreMotion", false, false) INITIALIZE_PASS_END(MergedLoadStoreMotionLegacyPass
This file contains the declarations for metadata subclasses.
F(f)
reverse_iterator rend()
Definition: BasicBlock.h:280
reverse_iterator rbegin()
Definition: BasicBlock.h:278
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.cpp:144
bool isIdenticalTo(const Instruction *I) const
Return true if the specified instruction is exactly identical to the current one. ...
void dump() const
Support for debugging, callable in GDB: V->dump()
Definition: AsmWriter.cpp:4428
AnalysisUsage & addRequired()
#define INITIALIZE_PASS_DEPENDENCY(depName)
Definition: PassSupport.h:50
FunctionPass * createMergedLoadStoreMotionPass(bool SplitFooterBB=false)
createMergedLoadStoreMotionPass - The public interface to this file.
mldst motion
mldst MergedLoadStoreMotion
Interval::succ_iterator succ_begin(Interval *I)
succ_begin/succ_end - define methods so that Intervals may be used just like BasicBlocks can with the...
Definition: Interval.h:102
auto reverse(ContainerTy &&C, typename std::enable_if< has_rbegin< ContainerTy >::value >::type *=nullptr) -> decltype(make_range(C.rbegin(), C.rend()))
Definition: STLExtras.h:261
Instruction * clone() const
Create a copy of &#39;this&#39; instruction that is identical in all ways except the following: ...
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:245
void andIRFlags(const Value *V)
Logical &#39;and&#39; of any supported wrapping, exact, and fast-math flags of V and this instruction...
const BasicBlock * getSingleSuccessor() const
Return the successor of this block if it has a single successor.
Definition: BasicBlock.cpp:275
An instruction for storing to memory.
Definition: Instructions.h:320
iterator begin()
Definition: Function.h:680
Interval::succ_iterator succ_end(Interval *I)
Definition: Interval.h:105
static bool runOnFunction(Function &F, bool PostInlining)
static MemoryLocation get(const LoadInst *LI)
Return a location with information about the memory reference by the given instruction.
BasicBlock * SplitBlockPredecessors(BasicBlock *BB, ArrayRef< BasicBlock *> Preds, const char *Suffix, DominatorTree *DT=nullptr, 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...
A set of analyses that are preserved following a run of a transformation pass.
Definition: PassManager.h:153
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:223
const BasicBlock * getSinglePredecessor() const
Return the predecessor of this block if it has a single predecessor block.
Definition: BasicBlock.cpp:240
void insertBefore(Instruction *InsertPos)
Insert an unlinked instruction into a basic block immediately before the specified instruction...
Definition: Instruction.cpp:73
LLVM Basic Block Representation.
Definition: BasicBlock.h:57
Conditional or Unconditional Branch instruction.
iterator_range< filter_iterator< BasicBlock::const_iterator, std::function< bool(const Instruction &)> > > instructionsWithoutDebug() const
Return a const iterator range over the instructions in the block, skipping any debug instructions...
Definition: BasicBlock.cpp:94
const Instruction & front() const
Definition: BasicBlock.h:285
A manager for alias analyses.
Represent the analysis usage information of a pass.
const Instruction & back() const
Definition: BasicBlock.h:287
Analysis pass providing a never-invalidated alias analysis result.
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:284
self_iterator getIterator()
Definition: ilist_node.h:81
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: PassManager.h:159
INITIALIZE_PASS_END(RegBankSelect, DEBUG_TYPE, "Assign register bank of generic virtual registers", false, false) RegBankSelect
Representation for a specific memory location.
iterator_range< T > make_range(T x, T y)
Convenience function for iterating over sub-ranges.
bool hasNPredecessorsOrMore(unsigned N) const
Return true if this block has N predecessors or more.
Definition: BasicBlock.cpp:271
Iterator for intrusive lists based on ilist_node.
void applyMergedLocation(const DILocation *LocA, const DILocation *LocB)
Merge 2 debug locations and apply it to the Instruction.
Definition: DebugInfo.cpp:689
static PHINode * Create(Type *Ty, unsigned NumReservedValues, const Twine &NameStr="", Instruction *InsertBefore=nullptr)
Constructors - NumReservedValues is a hint for the number of incoming edges that this phi node will h...
void setPreservesCFG()
This function should be called by the pass, iff they do not:
Definition: Pass.cpp:301
void setOperand(unsigned i, Value *Val)
Definition: User.h:174
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:132
Represents analyses that only rely on functions&#39; control flow.
Definition: PassManager.h:114
This pass performs merges of loads and stores on both sides of a.
const DebugLoc & getDebugLoc() const
Return the debug location for this node as a DebugLoc.
Definition: Instruction.h:331
void preserveSet()
Mark an analysis set as preserved.
Definition: PassManager.h:189
StringRef getName() const
Return a constant reference to the value&#39;s name.
Definition: Value.cpp:214
const Function * getParent() const
Return the enclosing method, or null if none.
Definition: BasicBlock.h:106
#define I(x, y, z)
Definition: MD5.cpp:58
LLVM_NODISCARD std::enable_if<!is_simple_type< Y >::value, typename cast_retty< X, const Y >::ret_type >::type dyn_cast(const Y &Val)
Definition: Casting.h:332
void preserve()
Mark an analysis as preserved.
Definition: PassManager.h:174
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
LLVM Value Representation.
Definition: Value.h:73
A container for analyses that lazily runs them and caches their results.
A wrapper pass to provide the legacy pass manager access to a suitably prepared AAResults object...
bool isSimple() const
Definition: Instructions.h:401
#define LLVM_DEBUG(X)
Definition: Debug.h:122
void initializeMergedLoadStoreMotionLegacyPassPass(PassRegistry &)
Value * getPointerOperand()
Definition: Instructions.h:412
const BasicBlock * getParent() const
Definition: Instruction.h:66