1 //===- PRE.cpp - Partial Redundancy Elimination ---------------------------===//
3 // This file implements the well known Partial Redundancy Elimination
4 // optimization, using an SSA formulation based on e-paths. See this paper for
7 // E-path_PRE: partial redundancy elimination made easy
8 // By: Dhananjay M. Dhamdhere In: ACM SIGPLAN Notices. Vol 37, #8, 2002
9 // http://doi.acm.org/10.1145/596992.597004
11 // This file actually implements a sparse version of the algorithm, using SSA
12 // and CFG properties instead of bit-vectors.
14 //===----------------------------------------------------------------------===//
16 #include "llvm/Pass.h"
17 #include "llvm/Function.h"
18 #include "llvm/Type.h"
19 #include "llvm/iPHINode.h"
20 #include "llvm/iMemory.h"
21 #include "llvm/Support/CFG.h"
22 #include "llvm/Analysis/Dominators.h"
23 #include "llvm/Analysis/PostDominators.h"
24 #include "llvm/Analysis/ValueNumbering.h"
25 #include "llvm/Transforms/Scalar.h"
26 #include "Support/Debug.h"
27 #include "Support/DepthFirstIterator.h"
28 #include "Support/PostOrderIterator.h"
29 #include "Support/Statistic.h"
30 #include "Support/hash_set"
33 Statistic<> NumExprsEliminated("pre", "Number of expressions constantified");
34 Statistic<> NumRedundant ("pre", "Number of redundant exprs eliminated");
35 Statistic<> NumInserted ("pre", "Number of expressions inserted");
37 struct PRE : public FunctionPass {
38 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
39 AU.addRequiredID(BreakCriticalEdgesID); // No critical edges for now!
40 AU.addRequired<PostDominatorTree>();
41 AU.addRequired<PostDominanceFrontier>();
42 AU.addRequired<DominatorSet>();
43 AU.addRequired<DominatorTree>();
44 AU.addRequired<DominanceFrontier>();
45 AU.addRequired<ValueNumbering>();
47 virtual bool runOnFunction(Function &F);
50 // Block information - Map basic blocks in a function back and forth to
52 std::vector<BasicBlock*> BlockMapping;
53 hash_map<BasicBlock*, unsigned> BlockNumbering;
55 // ProcessedExpressions - Keep track of which expressions have already been
57 hash_set<Instruction*> ProcessedExpressions;
59 // Provide access to the various analyses used...
61 DominatorTree *DT; PostDominatorTree *PDT;
62 DominanceFrontier *DF; PostDominanceFrontier *PDF;
65 // AvailableBlocks - Contain a mapping of blocks with available expression
66 // values to the expression value itself. This can be used as an efficient
67 // way to find out if the expression is available in the block, and if so,
68 // which version to use. This map is only used while processing a single
71 typedef hash_map<BasicBlock*, Instruction*> AvailableBlocksTy;
72 AvailableBlocksTy AvailableBlocks;
74 bool ProcessBlock(BasicBlock *BB);
76 // Anticipatibility calculation...
77 void MarkPostDominatingBlocksAnticipatible(PostDominatorTree::Node *N,
78 std::vector<char> &AntBlocks,
79 Instruction *Occurance);
80 void CalculateAnticipatiblityForOccurance(unsigned BlockNo,
81 std::vector<char> &AntBlocks,
82 Instruction *Occurance);
83 void CalculateAnticipatibleBlocks(const std::map<unsigned, Instruction*> &D,
84 std::vector<char> &AnticipatibleBlocks);
86 // PRE for an expression
87 void MarkOccuranceAvailableInAllDominatedBlocks(Instruction *Occurance,
88 BasicBlock *StartBlock);
89 void ReplaceDominatedAvailableOccurancesWith(Instruction *NewOcc,
90 DominatorTree::Node *N);
91 bool ProcessExpression(Instruction *I);
94 RegisterOpt<PRE> Z("pre", "Partial Redundancy Elimination");
98 bool PRE::runOnFunction(Function &F) {
99 VN = &getAnalysis<ValueNumbering>();
100 DS = &getAnalysis<DominatorSet>();
101 DT = &getAnalysis<DominatorTree>();
102 DF = &getAnalysis<DominanceFrontier>();
103 PDT = &getAnalysis<PostDominatorTree>();
104 PDF = &getAnalysis<PostDominanceFrontier>();
106 DEBUG(std::cerr << "\n*** Running PRE on func '" << F.getName() << "'...\n");
108 // Number the basic blocks based on a reverse post-order traversal of the CFG
109 // so that all predecessors of a block (ignoring back edges) are visited
110 // before a block is visited.
112 BlockMapping.reserve(F.size());
114 ReversePostOrderTraversal<Function*> RPOT(&F);
115 DEBUG(std::cerr << "Block order: ");
116 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(),
117 E = RPOT.end(); I != E; ++I) {
118 // Keep track of mapping...
120 BlockNumbering.insert(std::make_pair(BB, BlockMapping.size()));
121 BlockMapping.push_back(BB);
122 DEBUG(std::cerr << BB->getName() << " ");
124 DEBUG(std::cerr << "\n");
127 // Traverse the current function depth-first in dominator-tree order. This
128 // ensures that we see all definitions before their uses (except for PHI
129 // nodes), allowing us to hoist dependent expressions correctly.
130 bool Changed = false;
131 for (unsigned i = 0, e = BlockMapping.size(); i != e; ++i)
132 Changed |= ProcessBlock(BlockMapping[i]);
135 BlockMapping.clear();
136 BlockNumbering.clear();
137 ProcessedExpressions.clear();
142 // ProcessBlock - Process any expressions first seen in this block...
144 bool PRE::ProcessBlock(BasicBlock *BB) {
145 bool Changed = false;
147 // PRE expressions first defined in this block...
148 Instruction *PrevInst = 0;
149 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); )
150 if (ProcessExpression(I)) {
151 // The current instruction may have been deleted, make sure to back up to
165 void PRE::MarkPostDominatingBlocksAnticipatible(PostDominatorTree::Node *N,
166 std::vector<char> &AntBlocks,
167 Instruction *Occurance) {
168 unsigned BlockNo = BlockNumbering[N->getNode()];
170 if (AntBlocks[BlockNo]) return; // Already known to be anticipatible??
172 // Check to see if any of the operands are defined in this block, if so, the
173 // entry of this block does not anticipate the expression. This computes
175 for (unsigned i = 0, e = Occurance->getNumOperands(); i != e; ++i)
176 if (Instruction *I = dyn_cast<Instruction>(Occurance->getOperand(i)))
177 if (I->getParent() == N->getNode()) // Operand is defined in this block!
180 if (isa<LoadInst>(Occurance))
181 return; // FIXME: compute transparency for load instructions using AA
183 // Insert block into AntBlocks list...
184 AntBlocks[BlockNo] = true;
186 for (PostDominatorTree::Node::iterator I = N->begin(), E = N->end(); I != E;
188 MarkPostDominatingBlocksAnticipatible(*I, AntBlocks, Occurance);
191 void PRE::CalculateAnticipatiblityForOccurance(unsigned BlockNo,
192 std::vector<char> &AntBlocks,
193 Instruction *Occurance) {
194 if (AntBlocks[BlockNo]) return; // Block already anticipatible!
196 BasicBlock *BB = BlockMapping[BlockNo];
198 // For each occurance, mark all post-dominated blocks as anticipatible...
199 MarkPostDominatingBlocksAnticipatible(PDT->getNode(BB), AntBlocks,
202 // Next, mark any blocks in the post-dominance frontier as anticipatible iff
203 // all successors are anticipatible.
205 PostDominanceFrontier::iterator PDFI = PDF->find(BB);
206 if (PDFI != DF->end())
207 for (std::set<BasicBlock*>::iterator DI = PDFI->second.begin();
208 DI != PDFI->second.end(); ++DI) {
209 BasicBlock *PDFBlock = *DI;
210 bool AllSuccessorsAnticipatible = true;
211 for (succ_iterator SI = succ_begin(PDFBlock), SE = succ_end(PDFBlock);
213 if (!AntBlocks[BlockNumbering[*SI]]) {
214 AllSuccessorsAnticipatible = false;
218 if (AllSuccessorsAnticipatible)
219 CalculateAnticipatiblityForOccurance(BlockNumbering[PDFBlock],
220 AntBlocks, Occurance);
225 void PRE::CalculateAnticipatibleBlocks(const std::map<unsigned,
227 std::vector<char> &AntBlocks) {
228 // Initialize to zeros...
229 AntBlocks.resize(BlockMapping.size());
231 // Loop over all of the expressions...
232 for (std::map<unsigned, Instruction*>::const_iterator I = Defs.begin(),
233 E = Defs.end(); I != E; ++I)
234 CalculateAnticipatiblityForOccurance(I->first, AntBlocks, I->second);
237 /// MarkOccuranceAvailableInAllDominatedBlocks - Add entries to AvailableBlocks
238 /// for all nodes dominated by the occurance to indicate that it is now the
239 /// available occurance to use in any of these blocks.
241 void PRE::MarkOccuranceAvailableInAllDominatedBlocks(Instruction *Occurance,
243 // FIXME: There are much more efficient ways to get the blocks dominated
244 // by a block. Use them.
246 DominatorTree::Node *N = DT->getNode(Occurance->getParent());
247 for (df_iterator<DominatorTree::Node*> DI = df_begin(N), E = df_end(N);
249 AvailableBlocks[(*DI)->getNode()] = Occurance;
252 /// ReplaceDominatedAvailableOccurancesWith - This loops over the region
253 /// dominated by N, replacing any available expressions with NewOcc.
254 void PRE::ReplaceDominatedAvailableOccurancesWith(Instruction *NewOcc,
255 DominatorTree::Node *N) {
256 BasicBlock *BB = N->getNode();
257 Instruction *&ExistingAvailableVal = AvailableBlocks[BB];
259 // If there isn't a definition already active in this node, make this the new
260 // active definition...
261 if (ExistingAvailableVal == 0) {
262 ExistingAvailableVal = NewOcc;
264 for (DominatorTree::Node::iterator I = N->begin(), E = N->end(); I != E;++I)
265 ReplaceDominatedAvailableOccurancesWith(NewOcc, *I);
267 // If there is already an active definition in this block, replace it with
268 // NewOcc, and force it into all dominated blocks.
269 DEBUG(std::cerr << " Replacing dominated occ %"
270 << ExistingAvailableVal->getName() << " with %" << NewOcc->getName()
272 assert(ExistingAvailableVal != NewOcc && "NewOcc already inserted??");
273 ExistingAvailableVal->replaceAllUsesWith(NewOcc);
276 assert(ExistingAvailableVal->getParent() == BB &&
277 "OldOcc not defined in current block?");
278 BB->getInstList().erase(ExistingAvailableVal);
280 // Mark NewOCC as the Available expression in all blocks dominated by BB
281 for (df_iterator<DominatorTree::Node*> DI = df_begin(N), E = df_end(N);
283 AvailableBlocks[(*DI)->getNode()] = NewOcc;
288 /// ProcessExpression - Given an expression (instruction) process the
289 /// instruction to remove any partial redundancies induced by equivalent
290 /// computations. Note that we only need to PRE each expression once, so we
291 /// keep track of whether an expression has been PRE'd already, and don't PRE an
292 /// expression again. Expressions may be seen multiple times because process
293 /// the entire equivalence class at once, which may leave expressions later in
294 /// the control path.
296 bool PRE::ProcessExpression(Instruction *Expr) {
297 if (Expr->mayWriteToMemory() || Expr->getType() == Type::VoidTy ||
299 return false; // Cannot move expression
300 if (ProcessedExpressions.count(Expr)) return false; // Already processed.
302 // Ok, this is the first time we have seen the expression. Build a set of
303 // equivalent expressions using SSA def/use information. We consider
304 // expressions to be equivalent if they are the same opcode and have
305 // equivalent operands. As a special case for SSA, values produced by PHI
306 // nodes are considered to be equivalent to all of their operands.
308 std::vector<Value*> Values;
309 VN->getEqualNumberNodes(Expr, Values);
312 // FIXME: This should handle PHI nodes correctly. To do this, we need to
313 // consider expressions of the following form equivalent to this set of
316 // If an operand is a PHI node, add any occurances of the expression with the
317 // PHI operand replaced with the PHI node operands. This is only valid if the
318 // PHI operand occurances exist in blocks post-dominated by the incoming edge
322 // We have to be careful to handle expression definitions which dominated by
323 // other expressions. These can be directly eliminated in favor of their
324 // dominating value. Keep track of which blocks contain definitions (the key)
325 // and if a block contains a definition, which instruction it is.
327 std::map<unsigned, Instruction*> Definitions;
328 Definitions.insert(std::make_pair(BlockNumbering[Expr->getParent()], Expr));
330 bool Changed = false;
332 // Look at all of the equal values. If any of the values is not an
333 // instruction, replace all other expressions immediately with it (it must be
334 // an argument or a constant or something). Otherwise, convert the list of
335 // values into a list of expression (instruction) definitions ordering
336 // according to their dominator tree ordering.
338 Value *NonInstValue = 0;
339 for (unsigned i = 0, e = Values.size(); i != e; ++i)
340 if (Instruction *I = dyn_cast<Instruction>(Values[i])) {
341 Instruction *&BlockInst = Definitions[BlockNumbering[I->getParent()]];
342 if (BlockInst && BlockInst != I) { // Eliminate direct redundancy
343 if (DS->dominates(I, BlockInst)) { // I dom BlockInst
344 BlockInst->replaceAllUsesWith(I);
345 BlockInst->getParent()->getInstList().erase(BlockInst);
346 } else { // BlockInst dom I
347 I->replaceAllUsesWith(BlockInst);
348 I->getParent()->getInstList().erase(I);
355 NonInstValue = Values[i];
358 std::vector<Value*>().swap(Values); // Done with the values list
361 // This is the good, though unlikely, case where we find out that this
362 // expression is equal to a constant or argument directly. We can replace
363 // this and all of the other equivalent instructions with the value
366 for (std::map<unsigned, Instruction*>::iterator I = Definitions.begin(),
367 E = Definitions.end(); I != E; ++I) {
368 Instruction *Inst = I->second;
369 // Replace the value with the specified non-instruction value.
370 Inst->replaceAllUsesWith(NonInstValue); // Fixup any uses
371 Inst->getParent()->getInstList().erase(Inst); // Erase the instruction
373 NumExprsEliminated += Definitions.size();
374 return true; // Program modified!
377 // There are no expressions equal to this one. Exit early.
378 assert(!Definitions.empty() && "no equal expressions??");
380 if (Definitions.size() == 1) {
381 ProcessedExpressions.insert(Definitions.begin()->second);
385 DEBUG(std::cerr << "\n====--- Expression: " << Expr);
386 const Type *ExprType = Expr->getType();
388 // AnticipatibleBlocks - Blocks where the current expression is anticipatible.
389 // This is logically std::vector<bool> but using 'char' for performance.
390 std::vector<char> AnticipatibleBlocks;
392 // Calculate all of the blocks which the current expression is anticipatible.
393 CalculateAnticipatibleBlocks(Definitions, AnticipatibleBlocks);
395 // Print out anticipatible blocks...
396 DEBUG(std::cerr << "AntBlocks: ";
397 for (unsigned i = 0, e = AnticipatibleBlocks.size(); i != e; ++i)
398 if (AnticipatibleBlocks[i])
399 std::cerr << BlockMapping[i]->getName() <<" ";
404 // AvailabilityFrontier - Calculates the availability frontier for the current
405 // expression. The availability frontier contains the blocks on the dominance
406 // frontier of the current available expressions, iff they anticipate a
407 // definition of the expression.
408 hash_set<unsigned> AvailabilityFrontier;
410 Instruction *NonPHIOccurance = 0;
412 while (!Definitions.empty() || !AvailabilityFrontier.empty()) {
413 if (!Definitions.empty() &&
414 (AvailabilityFrontier.empty() ||
415 Definitions.begin()->first < *AvailabilityFrontier.begin())) {
416 Instruction *Occurance = Definitions.begin()->second;
417 BasicBlock *BB = Occurance->getParent();
418 Definitions.erase(Definitions.begin());
420 DEBUG(std::cerr << "PROCESSING Occurance: " << Occurance);
422 // Check to see if there is already an incoming value for this block...
423 AvailableBlocksTy::iterator LBI = AvailableBlocks.find(BB);
424 if (LBI != AvailableBlocks.end()) {
425 // Yes, there is a dominating definition for this block. Replace this
426 // occurance with the incoming value.
427 if (LBI->second != Occurance) {
428 DEBUG(std::cerr << " replacing with: " << LBI->second);
429 Occurance->replaceAllUsesWith(LBI->second);
430 BB->getInstList().erase(Occurance); // Delete instruction
435 ProcessedExpressions.insert(Occurance);
436 if (!isa<PHINode>(Occurance))
437 NonPHIOccurance = Occurance; // Keep an occurance of this expr
439 // Okay, there is no incoming value for this block, so this expression
440 // is a new definition that is good for this block and all blocks
441 // dominated by it. Add this information to the AvailableBlocks map.
443 MarkOccuranceAvailableInAllDominatedBlocks(Occurance, BB);
445 // Update the dominance frontier for the definitions so far... if a node
446 // in the dominator frontier now has all of its predecessors available,
447 // and the block is in an anticipatible region, we can insert a PHI node
449 DominanceFrontier::iterator DFI = DF->find(BB);
450 if (DFI != DF->end()) {
451 for (std::set<BasicBlock*>::iterator DI = DFI->second.begin();
452 DI != DFI->second.end(); ++DI) {
453 BasicBlock *DFBlock = *DI;
454 unsigned DFBlockID = BlockNumbering[DFBlock];
455 if (AnticipatibleBlocks[DFBlockID]) {
456 // Check to see if any of the predecessors of this block on the
457 // frontier are not available...
458 bool AnyNotAvailable = false;
459 for (pred_iterator PI = pred_begin(DFBlock),
460 PE = pred_end(DFBlock); PI != PE; ++PI)
461 if (!AvailableBlocks.count(*PI)) {
462 AnyNotAvailable = true;
466 // If any predecessor blocks are not available, add the node to
467 // the current expression dominance frontier.
468 if (AnyNotAvailable) {
469 AvailabilityFrontier.insert(DFBlockID);
471 // This block is no longer in the availability frontier, it IS
473 AvailabilityFrontier.erase(DFBlockID);
475 // If all of the predecessor blocks are available (and the block
476 // anticipates a definition along the path to the exit), we need
477 // to insert a new PHI node in this block. This block serves as
478 // a new definition for the expression, extending the available
481 PHINode *PN = new PHINode(ExprType, Expr->getName()+".pre",
483 ProcessedExpressions.insert(PN);
485 DEBUG(std::cerr << " INSERTING PHI on frontier: " << PN);
487 // Add the incoming blocks for the PHI node
488 for (pred_iterator PI = pred_begin(DFBlock),
489 PE = pred_end(DFBlock); PI != PE; ++PI)
491 PN->addIncoming(AvailableBlocks[*PI], *PI);
492 else // edge from the current block
493 PN->addIncoming(PN, DFBlock);
495 Instruction *&BlockOcc = Definitions[DFBlockID];
497 DEBUG(std::cerr <<" PHI superceeds occurance: "<<BlockOcc);
498 BlockOcc->replaceAllUsesWith(PN);
499 BlockOcc->getParent()->getInstList().erase(BlockOcc);
510 // Otherwise we must be looking at a node in the availability frontier!
511 unsigned AFBlockID = *AvailabilityFrontier.begin();
512 AvailabilityFrontier.erase(AvailabilityFrontier.begin());
513 BasicBlock *AFBlock = BlockMapping[AFBlockID];
515 // We eliminate the partial redundancy on this frontier by inserting a PHI
516 // node into this block, merging any incoming available versions into the
517 // PHI and inserting a new computation into predecessors without an
518 // incoming value. Note that we would have to insert the expression on
519 // the edge if the predecessor didn't anticipate the expression and we
520 // didn't break critical edges.
522 PHINode *PN = new PHINode(ExprType, Expr->getName()+".PRE",
524 DEBUG(std::cerr << "INSERTING PHI for PR: " << PN);
526 // If there is a pending occurance in this block, make sure to replace it
527 // with the PHI node...
528 std::map<unsigned, Instruction*>::iterator EDFI =
529 Definitions.find(AFBlockID);
530 if (EDFI != Definitions.end()) {
531 // There is already an occurance in this block. Replace it with PN and
533 Instruction *OldOcc = EDFI->second;
534 DEBUG(std::cerr << " Replaces occurance: " << OldOcc);
535 OldOcc->replaceAllUsesWith(PN);
536 AFBlock->getInstList().erase(OldOcc);
537 Definitions.erase(EDFI);
541 for (pred_iterator PI = pred_begin(AFBlock), PE = pred_end(AFBlock);
543 BasicBlock *Pred = *PI;
544 AvailableBlocksTy::iterator LBI = AvailableBlocks.find(Pred);
545 if (LBI != AvailableBlocks.end()) { // If there is a available value
546 PN->addIncoming(LBI->second, Pred); // for this pred, use it.
547 } else { // No available value yet...
548 unsigned PredID = BlockNumbering[Pred];
550 // Is the predecessor the same block that we inserted the PHI into?
552 if (Pred == AFBlock) {
553 // Yes, reuse the incoming value here...
554 PN->addIncoming(PN, Pred);
556 // No, we must insert a new computation into this block and add it
557 // to the definitions list...
558 assert(NonPHIOccurance && "No non-phi occurances seen so far???");
559 Instruction *New = NonPHIOccurance->clone();
560 New->setName(NonPHIOccurance->getName() + ".PRE-inserted");
561 ProcessedExpressions.insert(New);
563 DEBUG(std::cerr << " INSERTING OCCURANCE: " << New);
565 // Insert it into the bottom of the predecessor, right before the
566 // terminator instruction...
567 Pred->getInstList().insert(Pred->getTerminator(), New);
569 // Make this block be the available definition for any blocks it
570 // dominates. The ONLY case that this can affect more than just the
571 // block itself is when we are moving a computation to a loop
572 // header. In all other cases, because we don't have critical
573 // edges, the node is guaranteed to only dominate itself.
575 ReplaceDominatedAvailableOccurancesWith(New, DT->getNode(Pred));
577 // Add it as an incoming value on this edge to the PHI node
578 PN->addIncoming(New, Pred);
579 NonPHIOccurance = New;
585 // Find out if there is already an available value in this block. If so,
586 // we need to replace the available value with the PHI node. This can
587 // only happen when we just inserted a PHI node on a backedge.
589 AvailableBlocksTy::iterator LBBlockAvailableValIt =
590 AvailableBlocks.find(AFBlock);
591 if (LBBlockAvailableValIt != AvailableBlocks.end()) {
592 if (LBBlockAvailableValIt->second->getParent() == AFBlock) {
593 Instruction *OldVal = LBBlockAvailableValIt->second;
594 OldVal->replaceAllUsesWith(PN); // Use the new PHI node now
596 DEBUG(std::cerr << " PHI replaces available value: %"
597 << OldVal->getName() << "\n");
599 // Loop over all of the blocks dominated by this PHI node, and change
600 // the AvailableBlocks entries to be the PHI node instead of the old
602 MarkOccuranceAvailableInAllDominatedBlocks(PN, AFBlock);
604 AFBlock->getInstList().erase(OldVal); // Delete old instruction!
606 // The resultant PHI node is a new definition of the value!
607 Definitions.insert(std::make_pair(AFBlockID, PN));
609 // If the value is not defined in this block, that means that an
610 // inserted occurance in a predecessor is now the live value for the
611 // region (occurs when hoisting loop invariants, f.e.). In this case,
612 // the PHI node should actually just be removed.
613 assert(PN->use_empty() && "No uses should exist for dead PHI node!");
614 PN->getParent()->getInstList().erase(PN);
617 // The resultant PHI node is a new definition of the value!
618 Definitions.insert(std::make_pair(AFBlockID, PN));
623 AvailableBlocks.clear();