1 //===- BreakCriticalEdges.cpp - Critical Edge Elimination Pass ------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // BreakCriticalEdges pass - Break all of the critical edges in the CFG by
11 // inserting a dummy basic block. This pass may be "required" by passes that
12 // cannot deal with critical edges. For this usage, the structure type is
13 // forward declared. This pass obviously invalidates the CFG, but can update
14 // forward dominator (set, immediate dominators, tree, and frontier)
17 //===----------------------------------------------------------------------===//
19 #define DEBUG_TYPE "break-crit-edges"
20 #include "llvm/Transforms/Scalar.h"
21 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
22 #include "llvm/Analysis/Dominators.h"
23 #include "llvm/Analysis/LoopInfo.h"
24 #include "llvm/Analysis/ProfileInfo.h"
25 #include "llvm/Function.h"
26 #include "llvm/Instructions.h"
27 #include "llvm/Type.h"
28 #include "llvm/Support/CFG.h"
29 #include "llvm/Support/ErrorHandling.h"
30 #include "llvm/ADT/SmallVector.h"
31 #include "llvm/ADT/Statistic.h"
34 STATISTIC(NumBroken, "Number of blocks inserted");
37 struct BreakCriticalEdges : public FunctionPass {
38 static char ID; // Pass identification, replacement for typeid
39 BreakCriticalEdges() : FunctionPass(ID) {}
41 virtual bool runOnFunction(Function &F);
43 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
44 AU.addPreserved<DominatorTree>();
45 AU.addPreserved<DominanceFrontier>();
46 AU.addPreserved<LoopInfo>();
47 AU.addPreserved<ProfileInfo>();
49 // No loop canonicalization guarantees are broken by this pass.
50 AU.addPreservedID(LoopSimplifyID);
55 char BreakCriticalEdges::ID = 0;
56 INITIALIZE_PASS(BreakCriticalEdges, "break-crit-edges",
57 "Break critical edges in CFG", false, false);
59 // Publically exposed interface to pass...
60 char &llvm::BreakCriticalEdgesID = BreakCriticalEdges::ID;
61 FunctionPass *llvm::createBreakCriticalEdgesPass() {
62 return new BreakCriticalEdges();
65 // runOnFunction - Loop over all of the edges in the CFG, breaking critical
66 // edges as they are found.
68 bool BreakCriticalEdges::runOnFunction(Function &F) {
70 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
71 TerminatorInst *TI = I->getTerminator();
72 if (TI->getNumSuccessors() > 1 && !isa<IndirectBrInst>(TI))
73 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
74 if (SplitCriticalEdge(TI, i, this)) {
83 //===----------------------------------------------------------------------===//
84 // Implementation of the external critical edge manipulation functions
85 //===----------------------------------------------------------------------===//
87 // isCriticalEdge - Return true if the specified edge is a critical edge.
88 // Critical edges are edges from a block with multiple successors to a block
89 // with multiple predecessors.
91 bool llvm::isCriticalEdge(const TerminatorInst *TI, unsigned SuccNum,
92 bool AllowIdenticalEdges) {
93 assert(SuccNum < TI->getNumSuccessors() && "Illegal edge specification!");
94 if (TI->getNumSuccessors() == 1) return false;
96 const BasicBlock *Dest = TI->getSuccessor(SuccNum);
97 const_pred_iterator I = pred_begin(Dest), E = pred_end(Dest);
99 // If there is more than one predecessor, this is a critical edge...
100 assert(I != E && "No preds, but we have an edge to the block?");
101 const BasicBlock *FirstPred = *I;
102 ++I; // Skip one edge due to the incoming arc from TI.
103 if (!AllowIdenticalEdges)
106 // If AllowIdenticalEdges is true, then we allow this edge to be considered
107 // non-critical iff all preds come from TI's block.
109 const BasicBlock *P = *I;
112 // Note: leave this as is until no one ever compiles with either gcc 4.0.1
113 // or Xcode 2. This seems to work around the pred_iterator assert in PR 2207
120 /// CreatePHIsForSplitLoopExit - When a loop exit edge is split, LCSSA form
121 /// may require new PHIs in the new exit block. This function inserts the
122 /// new PHIs, as needed. Preds is a list of preds inside the loop, SplitBB
123 /// is the new loop exit block, and DestBB is the old loop exit, now the
124 /// successor of SplitBB.
125 static void CreatePHIsForSplitLoopExit(SmallVectorImpl<BasicBlock *> &Preds,
127 BasicBlock *DestBB) {
128 // SplitBB shouldn't have anything non-trivial in it yet.
129 assert(SplitBB->getFirstNonPHI() == SplitBB->getTerminator() &&
130 "SplitBB has non-PHI nodes!");
132 // For each PHI in the destination block...
133 for (BasicBlock::iterator I = DestBB->begin();
134 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
135 unsigned Idx = PN->getBasicBlockIndex(SplitBB);
136 Value *V = PN->getIncomingValue(Idx);
137 // If the input is a PHI which already satisfies LCSSA, don't create
139 if (const PHINode *VP = dyn_cast<PHINode>(V))
140 if (VP->getParent() == SplitBB)
142 // Otherwise a new PHI is needed. Create one and populate it.
143 PHINode *NewPN = PHINode::Create(PN->getType(), "split",
144 SplitBB->getTerminator());
145 for (unsigned i = 0, e = Preds.size(); i != e; ++i)
146 NewPN->addIncoming(V, Preds[i]);
147 // Update the original PHI.
148 PN->setIncomingValue(Idx, NewPN);
152 /// SplitCriticalEdge - If this edge is a critical edge, insert a new node to
153 /// split the critical edge. This will update DominatorTree and
154 /// DominatorFrontier information if it is available, thus calling this pass
155 /// will not invalidate either of them. This returns the new block if the edge
156 /// was split, null otherwise.
158 /// If MergeIdenticalEdges is true (not the default), *all* edges from TI to the
159 /// specified successor will be merged into the same critical edge block.
160 /// This is most commonly interesting with switch instructions, which may
161 /// have many edges to any one destination. This ensures that all edges to that
162 /// dest go to one block instead of each going to a different block, but isn't
163 /// the standard definition of a "critical edge".
165 /// It is invalid to call this function on a critical edge that starts at an
166 /// IndirectBrInst. Splitting these edges will almost always create an invalid
167 /// program because the address of the new block won't be the one that is jumped
170 BasicBlock *llvm::SplitCriticalEdge(TerminatorInst *TI, unsigned SuccNum,
171 Pass *P, bool MergeIdenticalEdges) {
172 if (!isCriticalEdge(TI, SuccNum, MergeIdenticalEdges)) return 0;
174 assert(!isa<IndirectBrInst>(TI) &&
175 "Cannot split critical edge from IndirectBrInst");
177 BasicBlock *TIBB = TI->getParent();
178 BasicBlock *DestBB = TI->getSuccessor(SuccNum);
180 // Create a new basic block, linking it into the CFG.
181 BasicBlock *NewBB = BasicBlock::Create(TI->getContext(),
182 TIBB->getName() + "." + DestBB->getName() + "_crit_edge");
183 // Create our unconditional branch.
184 BranchInst::Create(DestBB, NewBB);
186 // Branch to the new block, breaking the edge.
187 TI->setSuccessor(SuccNum, NewBB);
189 // Insert the block into the function... right after the block TI lives in.
190 Function &F = *TIBB->getParent();
191 Function::iterator FBBI = TIBB;
192 F.getBasicBlockList().insert(++FBBI, NewBB);
194 // If there are any PHI nodes in DestBB, we need to update them so that they
195 // merge incoming values from NewBB instead of from TIBB.
196 if (PHINode *APHI = dyn_cast<PHINode>(DestBB->begin())) {
197 // This conceptually does:
198 // foreach (PHINode *PN in DestBB)
199 // PN->setIncomingBlock(PN->getIncomingBlock(TIBB), NewBB);
200 // but is optimized for two cases.
202 if (APHI->getNumIncomingValues() <= 8) { // Small # preds case.
204 for (BasicBlock::iterator I = DestBB->begin(); isa<PHINode>(I); ++I) {
205 // We no longer enter through TIBB, now we come in through NewBB.
206 // Revector exactly one entry in the PHI node that used to come from
207 // TIBB to come from NewBB.
208 PHINode *PN = cast<PHINode>(I);
210 // Reuse the previous value of BBIdx if it lines up. In cases where we
211 // have multiple phi nodes with *lots* of predecessors, this is a speed
212 // win because we don't have to scan the PHI looking for TIBB. This
213 // happens because the BB list of PHI nodes are usually in the same
215 if (PN->getIncomingBlock(BBIdx) != TIBB)
216 BBIdx = PN->getBasicBlockIndex(TIBB);
217 PN->setIncomingBlock(BBIdx, NewBB);
220 // However, the foreach loop is slow for blocks with lots of predecessors
221 // because PHINode::getIncomingBlock is O(n) in # preds. Instead, walk
222 // the user list of TIBB to find the PHI nodes.
223 SmallPtrSet<PHINode*, 16> UpdatedPHIs;
225 for (Value::use_iterator UI = TIBB->use_begin(), E = TIBB->use_end();
227 Value::use_iterator Use = UI++;
228 if (PHINode *PN = dyn_cast<PHINode>(*Use)) {
229 // Remove one entry from each PHI.
230 if (PN->getParent() == DestBB && UpdatedPHIs.insert(PN))
231 PN->setOperand(Use.getOperandNo(), NewBB);
237 // If there are any other edges from TIBB to DestBB, update those to go
238 // through the split block, making those edges non-critical as well (and
239 // reducing the number of phi entries in the DestBB if relevant).
240 if (MergeIdenticalEdges) {
241 for (unsigned i = SuccNum+1, e = TI->getNumSuccessors(); i != e; ++i) {
242 if (TI->getSuccessor(i) != DestBB) continue;
244 // Remove an entry for TIBB from DestBB phi nodes.
245 DestBB->removePredecessor(TIBB);
247 // We found another edge to DestBB, go to NewBB instead.
248 TI->setSuccessor(i, NewBB);
254 // If we don't have a pass object, we can't update anything...
255 if (P == 0) return NewBB;
257 DominatorTree *DT = P->getAnalysisIfAvailable<DominatorTree>();
258 DominanceFrontier *DF = P->getAnalysisIfAvailable<DominanceFrontier>();
259 LoopInfo *LI = P->getAnalysisIfAvailable<LoopInfo>();
260 ProfileInfo *PI = P->getAnalysisIfAvailable<ProfileInfo>();
262 // If we have nothing to update, just return.
263 if (DT == 0 && DF == 0 && LI == 0 && PI == 0)
266 // Now update analysis information. Since the only predecessor of NewBB is
267 // the TIBB, TIBB clearly dominates NewBB. TIBB usually doesn't dominate
268 // anything, as there are other successors of DestBB. However, if all other
269 // predecessors of DestBB are already dominated by DestBB (e.g. DestBB is a
270 // loop header) then NewBB dominates DestBB.
271 SmallVector<BasicBlock*, 8> OtherPreds;
273 // If there is a PHI in the block, loop over predecessors with it, which is
274 // faster than iterating pred_begin/end.
275 if (PHINode *PN = dyn_cast<PHINode>(DestBB->begin())) {
276 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
277 if (PN->getIncomingBlock(i) != NewBB)
278 OtherPreds.push_back(PN->getIncomingBlock(i));
280 for (pred_iterator I = pred_begin(DestBB), E = pred_end(DestBB);
284 OtherPreds.push_back(P);
288 bool NewBBDominatesDestBB = true;
290 // Should we update DominatorTree information?
292 DomTreeNode *TINode = DT->getNode(TIBB);
294 // The new block is not the immediate dominator for any other nodes, but
295 // TINode is the immediate dominator for the new node.
297 if (TINode) { // Don't break unreachable code!
298 DomTreeNode *NewBBNode = DT->addNewBlock(NewBB, TIBB);
299 DomTreeNode *DestBBNode = 0;
301 // If NewBBDominatesDestBB hasn't been computed yet, do so with DT.
302 if (!OtherPreds.empty()) {
303 DestBBNode = DT->getNode(DestBB);
304 while (!OtherPreds.empty() && NewBBDominatesDestBB) {
305 if (DomTreeNode *OPNode = DT->getNode(OtherPreds.back()))
306 NewBBDominatesDestBB = DT->dominates(DestBBNode, OPNode);
307 OtherPreds.pop_back();
312 // If NewBBDominatesDestBB, then NewBB dominates DestBB, otherwise it
313 // doesn't dominate anything.
314 if (NewBBDominatesDestBB) {
315 if (!DestBBNode) DestBBNode = DT->getNode(DestBB);
316 DT->changeImmediateDominator(DestBBNode, NewBBNode);
321 // Should we update DominanceFrontier information?
323 // If NewBBDominatesDestBB hasn't been computed yet, do so with DF.
324 if (!OtherPreds.empty()) {
325 // FIXME: IMPLEMENT THIS!
326 llvm_unreachable("Requiring domfrontiers but not idom/domtree/domset."
327 " not implemented yet!");
330 // Since the new block is dominated by its only predecessor TIBB,
331 // it cannot be in any block's dominance frontier. If NewBB dominates
332 // DestBB, its dominance frontier is the same as DestBB's, otherwise it is
334 DominanceFrontier::DomSetType NewDFSet;
335 if (NewBBDominatesDestBB) {
336 DominanceFrontier::iterator I = DF->find(DestBB);
337 if (I != DF->end()) {
338 DF->addBasicBlock(NewBB, I->second);
340 if (I->second.count(DestBB)) {
341 // However NewBB's frontier does not include DestBB.
342 DominanceFrontier::iterator NF = DF->find(NewBB);
343 DF->removeFromFrontier(NF, DestBB);
347 DF->addBasicBlock(NewBB, DominanceFrontier::DomSetType());
349 DominanceFrontier::DomSetType NewDFSet;
350 NewDFSet.insert(DestBB);
351 DF->addBasicBlock(NewBB, NewDFSet);
355 // Update LoopInfo if it is around.
357 if (Loop *TIL = LI->getLoopFor(TIBB)) {
358 // If one or the other blocks were not in a loop, the new block is not
359 // either, and thus LI doesn't need to be updated.
360 if (Loop *DestLoop = LI->getLoopFor(DestBB)) {
361 if (TIL == DestLoop) {
362 // Both in the same loop, the NewBB joins loop.
363 DestLoop->addBasicBlockToLoop(NewBB, LI->getBase());
364 } else if (TIL->contains(DestLoop)) {
365 // Edge from an outer loop to an inner loop. Add to the outer loop.
366 TIL->addBasicBlockToLoop(NewBB, LI->getBase());
367 } else if (DestLoop->contains(TIL)) {
368 // Edge from an inner loop to an outer loop. Add to the outer loop.
369 DestLoop->addBasicBlockToLoop(NewBB, LI->getBase());
371 // Edge from two loops with no containment relation. Because these
372 // are natural loops, we know that the destination block must be the
373 // header of its loop (adding a branch into a loop elsewhere would
374 // create an irreducible loop).
375 assert(DestLoop->getHeader() == DestBB &&
376 "Should not create irreducible loops!");
377 if (Loop *P = DestLoop->getParentLoop())
378 P->addBasicBlockToLoop(NewBB, LI->getBase());
381 // If TIBB is in a loop and DestBB is outside of that loop, split the
382 // other exit blocks of the loop that also have predecessors outside
383 // the loop, to maintain a LoopSimplify guarantee.
384 if (!TIL->contains(DestBB) &&
385 P->mustPreserveAnalysisID(LoopSimplifyID)) {
386 assert(!TIL->contains(NewBB) &&
387 "Split point for loop exit is contained in loop!");
389 // Update LCSSA form in the newly created exit block.
390 if (P->mustPreserveAnalysisID(LCSSAID)) {
391 SmallVector<BasicBlock *, 1> OrigPred;
392 OrigPred.push_back(TIBB);
393 CreatePHIsForSplitLoopExit(OrigPred, NewBB, DestBB);
396 // For each unique exit block...
397 SmallVector<BasicBlock *, 4> ExitBlocks;
398 TIL->getExitBlocks(ExitBlocks);
399 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
400 // Collect all the preds that are inside the loop, and note
401 // whether there are any preds outside the loop.
402 SmallVector<BasicBlock *, 4> Preds;
403 bool HasPredOutsideOfLoop = false;
404 BasicBlock *Exit = ExitBlocks[i];
405 for (pred_iterator I = pred_begin(Exit), E = pred_end(Exit);
408 if (TIL->contains(P))
411 HasPredOutsideOfLoop = true;
413 // If there are any preds not in the loop, we'll need to split
414 // the edges. The Preds.empty() check is needed because a block
415 // may appear multiple times in the list. We can't use
416 // getUniqueExitBlocks above because that depends on LoopSimplify
417 // form, which we're in the process of restoring!
418 if (!Preds.empty() && HasPredOutsideOfLoop) {
419 BasicBlock *NewExitBB =
420 SplitBlockPredecessors(Exit, Preds.data(), Preds.size(),
422 if (P->mustPreserveAnalysisID(LCSSAID))
423 CreatePHIsForSplitLoopExit(Preds, NewExitBB, Exit);
427 // LCSSA form was updated above for the case where LoopSimplify is
428 // available, which means that all predecessors of loop exit blocks
429 // are within the loop. Without LoopSimplify form, it would be
430 // necessary to insert a new phi.
431 assert((!P->mustPreserveAnalysisID(LCSSAID) ||
432 P->mustPreserveAnalysisID(LoopSimplifyID)) &&
433 "SplitCriticalEdge doesn't know how to update LCCSA form "
434 "without LoopSimplify!");
438 // Update ProfileInfo if it is around.
440 PI->splitEdge(TIBB, DestBB, NewBB, MergeIdenticalEdges);