1//===- LoopSimplify.cpp - Loop Canonicalization Pass ----------------------===//
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// This pass performs several transformations to transform natural loops into a
11// simpler form, which makes subsequent analyses and transformations simpler and
12// more effective.
13//
14// Loop pre-header insertion guarantees that there is a single, non-critical
15// entry edge from outside of the loop to the loop header. This simplifies a
16// number of analyses and transformations, such as LICM.
17//
18// Loop exit-block insertion guarantees that all exit blocks from the loop
19// (blocks which are outside of the loop that have predecessors inside of the
20// loop) only have predecessors from inside of the loop (and are thus dominated
21// by the loop header). This simplifies transformations such as store-sinking
22// that are built into LICM.
23//
24// This pass also guarantees that loops will have exactly one backedge.
25//
26// Indirectbr instructions introduce several complications. If the loop
27// contains or is entered by an indirectbr instruction, it may not be possible
28// to transform the loop and make these guarantees. Client code should check
29// that these conditions are true before relying on them.
30//
31// Note that the simplifycfg pass will clean up blocks which are split out but
32// end up being unnecessary, so usage of this pass should not pessimize
33// generated code.
34//
35// This pass obviously modifies the CFG, but updates loop information and
36// dominator information.
37//
38//===----------------------------------------------------------------------===//
39
40#define DEBUG_TYPE "loopsimplify"
41#include "llvm/Transforms/Scalar.h"
42#include "llvm/Constants.h"
43#include "llvm/Instructions.h"
44#include "llvm/Function.h"
45#include "llvm/LLVMContext.h"
46#include "llvm/Type.h"
47#include "llvm/Analysis/AliasAnalysis.h"
48#include "llvm/Analysis/Dominators.h"
49#include "llvm/Analysis/LoopPass.h"
50#include "llvm/Analysis/ScalarEvolution.h"
51#include "llvm/Transforms/Utils/BasicBlockUtils.h"
52#include "llvm/Transforms/Utils/Local.h"
53#include "llvm/Support/CFG.h"
54#include "llvm/ADT/SetOperations.h"
55#include "llvm/ADT/SetVector.h"
56#include "llvm/ADT/Statistic.h"
57#include "llvm/ADT/DepthFirstIterator.h"
58using namespace llvm;
59
60STATISTIC(NumInserted, "Number of pre-header or exit blocks inserted");
61STATISTIC(NumNested , "Number of nested loops split out");
62
63namespace {
64 struct LoopSimplify : public LoopPass {
65 static char ID; // Pass identification, replacement for typeid
66 LoopSimplify() : LoopPass(&ID) {}
67
68 // AA - If we have an alias analysis object to update, this is it, otherwise
69 // this is null.
70 AliasAnalysis *AA;
71 LoopInfo *LI;
72 DominatorTree *DT;
73 Loop *L;
74 virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
75
76 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
77 // We need loop information to identify the loops...
78 AU.addRequiredTransitive<LoopInfo>();
79 AU.addRequiredTransitive<DominatorTree>();
80
81 AU.addPreserved<LoopInfo>();
82 AU.addPreserved<DominatorTree>();
83 AU.addPreserved<DominanceFrontier>();
84 AU.addPreserved<AliasAnalysis>();
85 AU.addPreserved<ScalarEvolution>();
86 AU.addPreservedID(BreakCriticalEdgesID); // No critical edges added.
87 }
88
89 /// verifyAnalysis() - Verify LoopSimplifyForm's guarantees.
90 void verifyAnalysis() const;
91
92 private:
93 bool ProcessLoop(Loop *L, LPPassManager &LPM);
94 BasicBlock *RewriteLoopExitBlock(Loop *L, BasicBlock *Exit);
95 BasicBlock *InsertPreheaderForLoop(Loop *L);
96 Loop *SeparateNestedLoop(Loop *L, LPPassManager &LPM);
97 BasicBlock *InsertUniqueBackedgeBlock(Loop *L, BasicBlock *Preheader);
98 void PlaceSplitBlockCarefully(BasicBlock *NewBB,
99 SmallVectorImpl<BasicBlock*> &SplitPreds,
100 Loop *L);
101 };
102}
103
104char LoopSimplify::ID = 0;
105static RegisterPass<LoopSimplify>
106X("loopsimplify", "Canonicalize natural loops", true);
107
108// Publically exposed interface to pass...
109const PassInfo *const llvm::LoopSimplifyID = &X;
110Pass *llvm::createLoopSimplifyPass() { return new LoopSimplify(); }
111
112/// runOnFunction - Run down all loops in the CFG (recursively, but we could do
113/// it in any convenient order) inserting preheaders...
114///
115bool LoopSimplify::runOnLoop(Loop *l, LPPassManager &LPM) {
116 L = l;
117 bool Changed = false;
118 LI = &getAnalysis<LoopInfo>();
119 AA = getAnalysisIfAvailable<AliasAnalysis>();
120 DT = &getAnalysis<DominatorTree>();
121
122 Changed |= ProcessLoop(L, LPM);
123
124 return Changed;
125}
126
127/// ProcessLoop - Walk the loop structure in depth first order, ensuring that
128/// all loops have preheaders.
129///
130bool LoopSimplify::ProcessLoop(Loop *L, LPPassManager &LPM) {
131 bool Changed = false;
132ReprocessLoop:
133
134 // Check to see that no blocks (other than the header) in this loop that has
135 // predecessors that are not in the loop. This is not valid for natural
136 // loops, but can occur if the blocks are unreachable. Since they are
137 // unreachable we can just shamelessly delete those CFG edges!
138 for (Loop::block_iterator BB = L->block_begin(), E = L->block_end();
139 BB != E; ++BB) {
140 if (*BB == L->getHeader()) continue;
141
142 SmallPtrSet<BasicBlock *, 4> BadPreds;
143 for (pred_iterator PI = pred_begin(*BB), PE = pred_end(*BB); PI != PE; ++PI)
144 if (!L->contains(*PI))
145 BadPreds.insert(*PI);
146
147 // Delete each unique out-of-loop (and thus dead) predecessor.
148 for (SmallPtrSet<BasicBlock *, 4>::iterator I = BadPreds.begin(),
149 E = BadPreds.end(); I != E; ++I) {
150 // Inform each successor of each dead pred.
151 for (succ_iterator SI = succ_begin(*I), SE = succ_end(*I); SI != SE; ++SI)
152 (*SI)->removePredecessor(*I);
153 // Zap the dead pred's terminator and replace it with unreachable.
154 TerminatorInst *TI = (*I)->getTerminator();
155 TI->replaceAllUsesWith(UndefValue::get(TI->getType()));
156 (*I)->getTerminator()->eraseFromParent();
157 new UnreachableInst((*I)->getContext(), *I);
158 Changed = true;
159 }
160 }
161
162 // Does the loop already have a preheader? If so, don't insert one.
163 BasicBlock *Preheader = L->getLoopPreheader();
164 if (!Preheader) {
165 Preheader = InsertPreheaderForLoop(L);
166 if (Preheader) {
167 NumInserted++;
168 Changed = true;
169 }
170 }
171
172 // Next, check to make sure that all exit nodes of the loop only have
173 // predecessors that are inside of the loop. This check guarantees that the
174 // loop preheader/header will dominate the exit blocks. If the exit block has
175 // predecessors from outside of the loop, split the edge now.
176 SmallVector<BasicBlock*, 8> ExitBlocks;
177 L->getExitBlocks(ExitBlocks);
178
179 SetVector<BasicBlock*> ExitBlockSet(ExitBlocks.begin(), ExitBlocks.end());
180 for (SetVector<BasicBlock*>::iterator I = ExitBlockSet.begin(),
181 E = ExitBlockSet.end(); I != E; ++I) {
182 BasicBlock *ExitBlock = *I;
183 for (pred_iterator PI = pred_begin(ExitBlock), PE = pred_end(ExitBlock);
184 PI != PE; ++PI)
185 // Must be exactly this loop: no subloops, parent loops, or non-loop preds
186 // allowed.
187 if (!L->contains(*PI)) {
188 if (RewriteLoopExitBlock(L, ExitBlock)) {
189 NumInserted++;
190 Changed = true;
191 }
192 break;
193 }
194 }
195
196 // If the header has more than two predecessors at this point (from the
197 // preheader and from multiple backedges), we must adjust the loop.
198 BasicBlock *LoopLatch = L->getLoopLatch();
199 if (!LoopLatch) {
200 // If this is really a nested loop, rip it out into a child loop. Don't do
201 // this for loops with a giant number of backedges, just factor them into a
202 // common backedge instead.
203 if (L->getNumBackEdges() < 8) {
204 if (SeparateNestedLoop(L, LPM)) {
205 ++NumNested;
206 // This is a big restructuring change, reprocess the whole loop.
207 Changed = true;
208 // GCC doesn't tail recursion eliminate this.
209 goto ReprocessLoop;
210 }
211 }
212
213 // If we either couldn't, or didn't want to, identify nesting of the loops,
214 // insert a new block that all backedges target, then make it jump to the
215 // loop header.
216 LoopLatch = InsertUniqueBackedgeBlock(L, Preheader);
217 if (LoopLatch) {
218 NumInserted++;
219 Changed = true;
220 }
221 }
222
223 // Scan over the PHI nodes in the loop header. Since they now have only two
224 // incoming values (the loop is canonicalized), we may have simplified the PHI
225 // down to 'X = phi [X, Y]', which should be replaced with 'Y'.
226 PHINode *PN;
227 for (BasicBlock::iterator I = L->getHeader()->begin();
228 (PN = dyn_cast<PHINode>(I++)); )
229 if (Value *V = PN->hasConstantValue(DT)) {
230 if (AA) AA->deleteValue(PN);
231 PN->replaceAllUsesWith(V);
232 PN->eraseFromParent();
233 }
234
235 // If this loop has muliple exits and the exits all go to the same
236 // block, attempt to merge the exits. This helps several passes, such
237 // as LoopRotation, which do not support loops with multiple exits.
238 // SimplifyCFG also does this (and this code uses the same utility
239 // function), however this code is loop-aware, where SimplifyCFG is
240 // not. That gives it the advantage of being able to hoist
241 // loop-invariant instructions out of the way to open up more
242 // opportunities, and the disadvantage of having the responsibility
243 // to preserve dominator information.
244 bool UniqueExit = true;
245 if (!ExitBlocks.empty())
246 for (unsigned i = 1, e = ExitBlocks.size(); i != e; ++i)
247 if (ExitBlocks[i] != ExitBlocks[0]) {
248 UniqueExit = false;
249 break;
250 }
251 if (UniqueExit) {
252 SmallVector<BasicBlock*, 8> ExitingBlocks;
253 L->getExitingBlocks(ExitingBlocks);
254 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
255 BasicBlock *ExitingBlock = ExitingBlocks[i];
256 if (!ExitingBlock->getSinglePredecessor()) continue;
257 BranchInst *BI = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
258 if (!BI || !BI->isConditional()) continue;
259 CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition());
260 if (!CI || CI->getParent() != ExitingBlock) continue;
261
262 // Attempt to hoist out all instructions except for the
263 // comparison and the branch.
264 bool AllInvariant = true;
265 for (BasicBlock::iterator I = ExitingBlock->begin(); &*I != BI; ) {
266 Instruction *Inst = I++;
267 if (Inst == CI)
268 continue;
269 if (!L->makeLoopInvariant(Inst, Changed,
270 Preheader ? Preheader->getTerminator() : 0)) {
271 AllInvariant = false;
272 break;
273 }
274 }
275 if (!AllInvariant) continue;
276
277 // The block has now been cleared of all instructions except for
278 // a comparison and a conditional branch. SimplifyCFG may be able
279 // to fold it now.
280 if (!FoldBranchToCommonDest(BI)) continue;
281
282 // Success. The block is now dead, so remove it from the loop,
283 // update the dominator tree and dominance frontier, and delete it.
284 assert(pred_begin(ExitingBlock) == pred_end(ExitingBlock));
285 Changed = true;
286 LI->removeBlock(ExitingBlock);
287
288 DominanceFrontier *DF = getAnalysisIfAvailable<DominanceFrontier>();
289 DomTreeNode *Node = DT->getNode(ExitingBlock);
290 const std::vector<DomTreeNodeBase<BasicBlock> *> &Children =
291 Node->getChildren();
292 while (!Children.empty()) {
293 DomTreeNode *Child = Children.front();
294 DT->changeImmediateDominator(Child, Node->getIDom());
295 if (DF) DF->changeImmediateDominator(Child->getBlock(),
296 Node->getIDom()->getBlock(),
297 DT);
298 }
299 DT->eraseNode(ExitingBlock);
300 if (DF) DF->removeBlock(ExitingBlock);
301
302 BI->getSuccessor(0)->removePredecessor(ExitingBlock);
303 BI->getSuccessor(1)->removePredecessor(ExitingBlock);
304 ExitingBlock->eraseFromParent();
305 }
306 }
307
| 1//===- LoopSimplify.cpp - Loop Canonicalization Pass ----------------------===//
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// This pass performs several transformations to transform natural loops into a
11// simpler form, which makes subsequent analyses and transformations simpler and
12// more effective.
13//
14// Loop pre-header insertion guarantees that there is a single, non-critical
15// entry edge from outside of the loop to the loop header. This simplifies a
16// number of analyses and transformations, such as LICM.
17//
18// Loop exit-block insertion guarantees that all exit blocks from the loop
19// (blocks which are outside of the loop that have predecessors inside of the
20// loop) only have predecessors from inside of the loop (and are thus dominated
21// by the loop header). This simplifies transformations such as store-sinking
22// that are built into LICM.
23//
24// This pass also guarantees that loops will have exactly one backedge.
25//
26// Indirectbr instructions introduce several complications. If the loop
27// contains or is entered by an indirectbr instruction, it may not be possible
28// to transform the loop and make these guarantees. Client code should check
29// that these conditions are true before relying on them.
30//
31// Note that the simplifycfg pass will clean up blocks which are split out but
32// end up being unnecessary, so usage of this pass should not pessimize
33// generated code.
34//
35// This pass obviously modifies the CFG, but updates loop information and
36// dominator information.
37//
38//===----------------------------------------------------------------------===//
39
40#define DEBUG_TYPE "loopsimplify"
41#include "llvm/Transforms/Scalar.h"
42#include "llvm/Constants.h"
43#include "llvm/Instructions.h"
44#include "llvm/Function.h"
45#include "llvm/LLVMContext.h"
46#include "llvm/Type.h"
47#include "llvm/Analysis/AliasAnalysis.h"
48#include "llvm/Analysis/Dominators.h"
49#include "llvm/Analysis/LoopPass.h"
50#include "llvm/Analysis/ScalarEvolution.h"
51#include "llvm/Transforms/Utils/BasicBlockUtils.h"
52#include "llvm/Transforms/Utils/Local.h"
53#include "llvm/Support/CFG.h"
54#include "llvm/ADT/SetOperations.h"
55#include "llvm/ADT/SetVector.h"
56#include "llvm/ADT/Statistic.h"
57#include "llvm/ADT/DepthFirstIterator.h"
58using namespace llvm;
59
60STATISTIC(NumInserted, "Number of pre-header or exit blocks inserted");
61STATISTIC(NumNested , "Number of nested loops split out");
62
63namespace {
64 struct LoopSimplify : public LoopPass {
65 static char ID; // Pass identification, replacement for typeid
66 LoopSimplify() : LoopPass(&ID) {}
67
68 // AA - If we have an alias analysis object to update, this is it, otherwise
69 // this is null.
70 AliasAnalysis *AA;
71 LoopInfo *LI;
72 DominatorTree *DT;
73 Loop *L;
74 virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
75
76 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
77 // We need loop information to identify the loops...
78 AU.addRequiredTransitive<LoopInfo>();
79 AU.addRequiredTransitive<DominatorTree>();
80
81 AU.addPreserved<LoopInfo>();
82 AU.addPreserved<DominatorTree>();
83 AU.addPreserved<DominanceFrontier>();
84 AU.addPreserved<AliasAnalysis>();
85 AU.addPreserved<ScalarEvolution>();
86 AU.addPreservedID(BreakCriticalEdgesID); // No critical edges added.
87 }
88
89 /// verifyAnalysis() - Verify LoopSimplifyForm's guarantees.
90 void verifyAnalysis() const;
91
92 private:
93 bool ProcessLoop(Loop *L, LPPassManager &LPM);
94 BasicBlock *RewriteLoopExitBlock(Loop *L, BasicBlock *Exit);
95 BasicBlock *InsertPreheaderForLoop(Loop *L);
96 Loop *SeparateNestedLoop(Loop *L, LPPassManager &LPM);
97 BasicBlock *InsertUniqueBackedgeBlock(Loop *L, BasicBlock *Preheader);
98 void PlaceSplitBlockCarefully(BasicBlock *NewBB,
99 SmallVectorImpl<BasicBlock*> &SplitPreds,
100 Loop *L);
101 };
102}
103
104char LoopSimplify::ID = 0;
105static RegisterPass<LoopSimplify>
106X("loopsimplify", "Canonicalize natural loops", true);
107
108// Publically exposed interface to pass...
109const PassInfo *const llvm::LoopSimplifyID = &X;
110Pass *llvm::createLoopSimplifyPass() { return new LoopSimplify(); }
111
112/// runOnFunction - Run down all loops in the CFG (recursively, but we could do
113/// it in any convenient order) inserting preheaders...
114///
115bool LoopSimplify::runOnLoop(Loop *l, LPPassManager &LPM) {
116 L = l;
117 bool Changed = false;
118 LI = &getAnalysis<LoopInfo>();
119 AA = getAnalysisIfAvailable<AliasAnalysis>();
120 DT = &getAnalysis<DominatorTree>();
121
122 Changed |= ProcessLoop(L, LPM);
123
124 return Changed;
125}
126
127/// ProcessLoop - Walk the loop structure in depth first order, ensuring that
128/// all loops have preheaders.
129///
130bool LoopSimplify::ProcessLoop(Loop *L, LPPassManager &LPM) {
131 bool Changed = false;
132ReprocessLoop:
133
134 // Check to see that no blocks (other than the header) in this loop that has
135 // predecessors that are not in the loop. This is not valid for natural
136 // loops, but can occur if the blocks are unreachable. Since they are
137 // unreachable we can just shamelessly delete those CFG edges!
138 for (Loop::block_iterator BB = L->block_begin(), E = L->block_end();
139 BB != E; ++BB) {
140 if (*BB == L->getHeader()) continue;
141
142 SmallPtrSet<BasicBlock *, 4> BadPreds;
143 for (pred_iterator PI = pred_begin(*BB), PE = pred_end(*BB); PI != PE; ++PI)
144 if (!L->contains(*PI))
145 BadPreds.insert(*PI);
146
147 // Delete each unique out-of-loop (and thus dead) predecessor.
148 for (SmallPtrSet<BasicBlock *, 4>::iterator I = BadPreds.begin(),
149 E = BadPreds.end(); I != E; ++I) {
150 // Inform each successor of each dead pred.
151 for (succ_iterator SI = succ_begin(*I), SE = succ_end(*I); SI != SE; ++SI)
152 (*SI)->removePredecessor(*I);
153 // Zap the dead pred's terminator and replace it with unreachable.
154 TerminatorInst *TI = (*I)->getTerminator();
155 TI->replaceAllUsesWith(UndefValue::get(TI->getType()));
156 (*I)->getTerminator()->eraseFromParent();
157 new UnreachableInst((*I)->getContext(), *I);
158 Changed = true;
159 }
160 }
161
162 // Does the loop already have a preheader? If so, don't insert one.
163 BasicBlock *Preheader = L->getLoopPreheader();
164 if (!Preheader) {
165 Preheader = InsertPreheaderForLoop(L);
166 if (Preheader) {
167 NumInserted++;
168 Changed = true;
169 }
170 }
171
172 // Next, check to make sure that all exit nodes of the loop only have
173 // predecessors that are inside of the loop. This check guarantees that the
174 // loop preheader/header will dominate the exit blocks. If the exit block has
175 // predecessors from outside of the loop, split the edge now.
176 SmallVector<BasicBlock*, 8> ExitBlocks;
177 L->getExitBlocks(ExitBlocks);
178
179 SetVector<BasicBlock*> ExitBlockSet(ExitBlocks.begin(), ExitBlocks.end());
180 for (SetVector<BasicBlock*>::iterator I = ExitBlockSet.begin(),
181 E = ExitBlockSet.end(); I != E; ++I) {
182 BasicBlock *ExitBlock = *I;
183 for (pred_iterator PI = pred_begin(ExitBlock), PE = pred_end(ExitBlock);
184 PI != PE; ++PI)
185 // Must be exactly this loop: no subloops, parent loops, or non-loop preds
186 // allowed.
187 if (!L->contains(*PI)) {
188 if (RewriteLoopExitBlock(L, ExitBlock)) {
189 NumInserted++;
190 Changed = true;
191 }
192 break;
193 }
194 }
195
196 // If the header has more than two predecessors at this point (from the
197 // preheader and from multiple backedges), we must adjust the loop.
198 BasicBlock *LoopLatch = L->getLoopLatch();
199 if (!LoopLatch) {
200 // If this is really a nested loop, rip it out into a child loop. Don't do
201 // this for loops with a giant number of backedges, just factor them into a
202 // common backedge instead.
203 if (L->getNumBackEdges() < 8) {
204 if (SeparateNestedLoop(L, LPM)) {
205 ++NumNested;
206 // This is a big restructuring change, reprocess the whole loop.
207 Changed = true;
208 // GCC doesn't tail recursion eliminate this.
209 goto ReprocessLoop;
210 }
211 }
212
213 // If we either couldn't, or didn't want to, identify nesting of the loops,
214 // insert a new block that all backedges target, then make it jump to the
215 // loop header.
216 LoopLatch = InsertUniqueBackedgeBlock(L, Preheader);
217 if (LoopLatch) {
218 NumInserted++;
219 Changed = true;
220 }
221 }
222
223 // Scan over the PHI nodes in the loop header. Since they now have only two
224 // incoming values (the loop is canonicalized), we may have simplified the PHI
225 // down to 'X = phi [X, Y]', which should be replaced with 'Y'.
226 PHINode *PN;
227 for (BasicBlock::iterator I = L->getHeader()->begin();
228 (PN = dyn_cast<PHINode>(I++)); )
229 if (Value *V = PN->hasConstantValue(DT)) {
230 if (AA) AA->deleteValue(PN);
231 PN->replaceAllUsesWith(V);
232 PN->eraseFromParent();
233 }
234
235 // If this loop has muliple exits and the exits all go to the same
236 // block, attempt to merge the exits. This helps several passes, such
237 // as LoopRotation, which do not support loops with multiple exits.
238 // SimplifyCFG also does this (and this code uses the same utility
239 // function), however this code is loop-aware, where SimplifyCFG is
240 // not. That gives it the advantage of being able to hoist
241 // loop-invariant instructions out of the way to open up more
242 // opportunities, and the disadvantage of having the responsibility
243 // to preserve dominator information.
244 bool UniqueExit = true;
245 if (!ExitBlocks.empty())
246 for (unsigned i = 1, e = ExitBlocks.size(); i != e; ++i)
247 if (ExitBlocks[i] != ExitBlocks[0]) {
248 UniqueExit = false;
249 break;
250 }
251 if (UniqueExit) {
252 SmallVector<BasicBlock*, 8> ExitingBlocks;
253 L->getExitingBlocks(ExitingBlocks);
254 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
255 BasicBlock *ExitingBlock = ExitingBlocks[i];
256 if (!ExitingBlock->getSinglePredecessor()) continue;
257 BranchInst *BI = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
258 if (!BI || !BI->isConditional()) continue;
259 CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition());
260 if (!CI || CI->getParent() != ExitingBlock) continue;
261
262 // Attempt to hoist out all instructions except for the
263 // comparison and the branch.
264 bool AllInvariant = true;
265 for (BasicBlock::iterator I = ExitingBlock->begin(); &*I != BI; ) {
266 Instruction *Inst = I++;
267 if (Inst == CI)
268 continue;
269 if (!L->makeLoopInvariant(Inst, Changed,
270 Preheader ? Preheader->getTerminator() : 0)) {
271 AllInvariant = false;
272 break;
273 }
274 }
275 if (!AllInvariant) continue;
276
277 // The block has now been cleared of all instructions except for
278 // a comparison and a conditional branch. SimplifyCFG may be able
279 // to fold it now.
280 if (!FoldBranchToCommonDest(BI)) continue;
281
282 // Success. The block is now dead, so remove it from the loop,
283 // update the dominator tree and dominance frontier, and delete it.
284 assert(pred_begin(ExitingBlock) == pred_end(ExitingBlock));
285 Changed = true;
286 LI->removeBlock(ExitingBlock);
287
288 DominanceFrontier *DF = getAnalysisIfAvailable<DominanceFrontier>();
289 DomTreeNode *Node = DT->getNode(ExitingBlock);
290 const std::vector<DomTreeNodeBase<BasicBlock> *> &Children =
291 Node->getChildren();
292 while (!Children.empty()) {
293 DomTreeNode *Child = Children.front();
294 DT->changeImmediateDominator(Child, Node->getIDom());
295 if (DF) DF->changeImmediateDominator(Child->getBlock(),
296 Node->getIDom()->getBlock(),
297 DT);
298 }
299 DT->eraseNode(ExitingBlock);
300 if (DF) DF->removeBlock(ExitingBlock);
301
302 BI->getSuccessor(0)->removePredecessor(ExitingBlock);
303 BI->getSuccessor(1)->removePredecessor(ExitingBlock);
304 ExitingBlock->eraseFromParent();
305 }
306 }
307
|
308 return Changed;
309}
310
311/// InsertPreheaderForLoop - Once we discover that a loop doesn't have a
312/// preheader, this method is called to insert one. This method has two phases:
313/// preheader insertion and analysis updating.
314///
315BasicBlock *LoopSimplify::InsertPreheaderForLoop(Loop *L) {
316 BasicBlock *Header = L->getHeader();
317
318 // Compute the set of predecessors of the loop that are not in the loop.
319 SmallVector<BasicBlock*, 8> OutsideBlocks;
320 for (pred_iterator PI = pred_begin(Header), PE = pred_end(Header);
321 PI != PE; ++PI)
322 if (!L->contains(*PI)) { // Coming in from outside the loop?
323 // If the loop is branched to from an indirect branch, we won't
324 // be able to fully transform the loop, because it prohibits
325 // edge splitting.
326 if (isa<IndirectBrInst>((*PI)->getTerminator())) return 0;
327
328 // Keep track of it.
329 OutsideBlocks.push_back(*PI);
330 }
331
332 // Split out the loop pre-header.
333 BasicBlock *NewBB =
334 SplitBlockPredecessors(Header, &OutsideBlocks[0], OutsideBlocks.size(),
335 ".preheader", this);
336
337 // Make sure that NewBB is put someplace intelligent, which doesn't mess up
338 // code layout too horribly.
339 PlaceSplitBlockCarefully(NewBB, OutsideBlocks, L);
340
341 return NewBB;
342}
343
344/// RewriteLoopExitBlock - Ensure that the loop preheader dominates all exit
345/// blocks. This method is used to split exit blocks that have predecessors
346/// outside of the loop.
347BasicBlock *LoopSimplify::RewriteLoopExitBlock(Loop *L, BasicBlock *Exit) {
348 SmallVector<BasicBlock*, 8> LoopBlocks;
349 for (pred_iterator I = pred_begin(Exit), E = pred_end(Exit); I != E; ++I)
350 if (L->contains(*I)) {
351 // Don't do this if the loop is exited via an indirect branch.
352 if (isa<IndirectBrInst>((*I)->getTerminator())) return 0;
353
354 LoopBlocks.push_back(*I);
355 }
356
357 assert(!LoopBlocks.empty() && "No edges coming in from outside the loop?");
358 BasicBlock *NewBB = SplitBlockPredecessors(Exit, &LoopBlocks[0],
359 LoopBlocks.size(), ".loopexit",
360 this);
361
362 return NewBB;
363}
364
365/// AddBlockAndPredsToSet - Add the specified block, and all of its
366/// predecessors, to the specified set, if it's not already in there. Stop
367/// predecessor traversal when we reach StopBlock.
368static void AddBlockAndPredsToSet(BasicBlock *InputBB, BasicBlock *StopBlock,
369 std::set<BasicBlock*> &Blocks) {
370 std::vector<BasicBlock *> WorkList;
371 WorkList.push_back(InputBB);
372 do {
373 BasicBlock *BB = WorkList.back(); WorkList.pop_back();
374 if (Blocks.insert(BB).second && BB != StopBlock)
375 // If BB is not already processed and it is not a stop block then
376 // insert its predecessor in the work list
377 for (pred_iterator I = pred_begin(BB), E = pred_end(BB); I != E; ++I) {
378 BasicBlock *WBB = *I;
379 WorkList.push_back(WBB);
380 }
381 } while(!WorkList.empty());
382}
383
384/// FindPHIToPartitionLoops - The first part of loop-nestification is to find a
385/// PHI node that tells us how to partition the loops.
386static PHINode *FindPHIToPartitionLoops(Loop *L, DominatorTree *DT,
387 AliasAnalysis *AA) {
388 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ) {
389 PHINode *PN = cast<PHINode>(I);
390 ++I;
391 if (Value *V = PN->hasConstantValue(DT)) {
392 // This is a degenerate PHI already, don't modify it!
393 PN->replaceAllUsesWith(V);
394 if (AA) AA->deleteValue(PN);
395 PN->eraseFromParent();
396 continue;
397 }
398
399 // Scan this PHI node looking for a use of the PHI node by itself.
400 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
401 if (PN->getIncomingValue(i) == PN &&
402 L->contains(PN->getIncomingBlock(i)))
403 // We found something tasty to remove.
404 return PN;
405 }
406 return 0;
407}
408
409// PlaceSplitBlockCarefully - If the block isn't already, move the new block to
410// right after some 'outside block' block. This prevents the preheader from
411// being placed inside the loop body, e.g. when the loop hasn't been rotated.
412void LoopSimplify::PlaceSplitBlockCarefully(BasicBlock *NewBB,
413 SmallVectorImpl<BasicBlock*> &SplitPreds,
414 Loop *L) {
415 // Check to see if NewBB is already well placed.
416 Function::iterator BBI = NewBB; --BBI;
417 for (unsigned i = 0, e = SplitPreds.size(); i != e; ++i) {
418 if (&*BBI == SplitPreds[i])
419 return;
420 }
421
422 // If it isn't already after an outside block, move it after one. This is
423 // always good as it makes the uncond branch from the outside block into a
424 // fall-through.
425
426 // Figure out *which* outside block to put this after. Prefer an outside
427 // block that neighbors a BB actually in the loop.
428 BasicBlock *FoundBB = 0;
429 for (unsigned i = 0, e = SplitPreds.size(); i != e; ++i) {
430 Function::iterator BBI = SplitPreds[i];
431 if (++BBI != NewBB->getParent()->end() &&
432 L->contains(BBI)) {
433 FoundBB = SplitPreds[i];
434 break;
435 }
436 }
437
438 // If our heuristic for a *good* bb to place this after doesn't find
439 // anything, just pick something. It's likely better than leaving it within
440 // the loop.
441 if (!FoundBB)
442 FoundBB = SplitPreds[0];
443 NewBB->moveAfter(FoundBB);
444}
445
446
447/// SeparateNestedLoop - If this loop has multiple backedges, try to pull one of
448/// them out into a nested loop. This is important for code that looks like
449/// this:
450///
451/// Loop:
452/// ...
453/// br cond, Loop, Next
454/// ...
455/// br cond2, Loop, Out
456///
457/// To identify this common case, we look at the PHI nodes in the header of the
458/// loop. PHI nodes with unchanging values on one backedge correspond to values
459/// that change in the "outer" loop, but not in the "inner" loop.
460///
461/// If we are able to separate out a loop, return the new outer loop that was
462/// created.
463///
464Loop *LoopSimplify::SeparateNestedLoop(Loop *L, LPPassManager &LPM) {
465 PHINode *PN = FindPHIToPartitionLoops(L, DT, AA);
466 if (PN == 0) return 0; // No known way to partition.
467
468 // Pull out all predecessors that have varying values in the loop. This
469 // handles the case when a PHI node has multiple instances of itself as
470 // arguments.
471 SmallVector<BasicBlock*, 8> OuterLoopPreds;
472 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
473 if (PN->getIncomingValue(i) != PN ||
474 !L->contains(PN->getIncomingBlock(i)))
475 OuterLoopPreds.push_back(PN->getIncomingBlock(i));
476
477 BasicBlock *Header = L->getHeader();
478 BasicBlock *NewBB = SplitBlockPredecessors(Header, &OuterLoopPreds[0],
479 OuterLoopPreds.size(),
480 ".outer", this);
481
482 // Make sure that NewBB is put someplace intelligent, which doesn't mess up
483 // code layout too horribly.
484 PlaceSplitBlockCarefully(NewBB, OuterLoopPreds, L);
485
486 // Create the new outer loop.
487 Loop *NewOuter = new Loop();
488
489 // Change the parent loop to use the outer loop as its child now.
490 if (Loop *Parent = L->getParentLoop())
491 Parent->replaceChildLoopWith(L, NewOuter);
492 else
493 LI->changeTopLevelLoop(L, NewOuter);
494
495 // L is now a subloop of our outer loop.
496 NewOuter->addChildLoop(L);
497
498 // Add the new loop to the pass manager queue.
499 LPM.insertLoopIntoQueue(NewOuter);
500
501 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
502 I != E; ++I)
503 NewOuter->addBlockEntry(*I);
504
505 // Now reset the header in L, which had been moved by
506 // SplitBlockPredecessors for the outer loop.
507 L->moveToHeader(Header);
508
509 // Determine which blocks should stay in L and which should be moved out to
510 // the Outer loop now.
511 std::set<BasicBlock*> BlocksInL;
512 for (pred_iterator PI = pred_begin(Header), E = pred_end(Header); PI!=E; ++PI)
513 if (DT->dominates(Header, *PI))
514 AddBlockAndPredsToSet(*PI, Header, BlocksInL);
515
516
517 // Scan all of the loop children of L, moving them to OuterLoop if they are
518 // not part of the inner loop.
519 const std::vector<Loop*> &SubLoops = L->getSubLoops();
520 for (size_t I = 0; I != SubLoops.size(); )
521 if (BlocksInL.count(SubLoops[I]->getHeader()))
522 ++I; // Loop remains in L
523 else
524 NewOuter->addChildLoop(L->removeChildLoop(SubLoops.begin() + I));
525
526 // Now that we know which blocks are in L and which need to be moved to
527 // OuterLoop, move any blocks that need it.
528 for (unsigned i = 0; i != L->getBlocks().size(); ++i) {
529 BasicBlock *BB = L->getBlocks()[i];
530 if (!BlocksInL.count(BB)) {
531 // Move this block to the parent, updating the exit blocks sets
532 L->removeBlockFromLoop(BB);
533 if ((*LI)[BB] == L)
534 LI->changeLoopFor(BB, NewOuter);
535 --i;
536 }
537 }
538
539 return NewOuter;
540}
541
542
543
544/// InsertUniqueBackedgeBlock - This method is called when the specified loop
545/// has more than one backedge in it. If this occurs, revector all of these
546/// backedges to target a new basic block and have that block branch to the loop
547/// header. This ensures that loops have exactly one backedge.
548///
549BasicBlock *
550LoopSimplify::InsertUniqueBackedgeBlock(Loop *L, BasicBlock *Preheader) {
551 assert(L->getNumBackEdges() > 1 && "Must have > 1 backedge!");
552
553 // Get information about the loop
554 BasicBlock *Header = L->getHeader();
555 Function *F = Header->getParent();
556
557 // Unique backedge insertion currently depends on having a preheader.
558 if (!Preheader)
559 return 0;
560
561 // Figure out which basic blocks contain back-edges to the loop header.
562 std::vector<BasicBlock*> BackedgeBlocks;
563 for (pred_iterator I = pred_begin(Header), E = pred_end(Header); I != E; ++I)
564 if (*I != Preheader) BackedgeBlocks.push_back(*I);
565
566 // Create and insert the new backedge block...
567 BasicBlock *BEBlock = BasicBlock::Create(Header->getContext(),
568 Header->getName()+".backedge", F);
569 BranchInst *BETerminator = BranchInst::Create(Header, BEBlock);
570
571 // Move the new backedge block to right after the last backedge block.
572 Function::iterator InsertPos = BackedgeBlocks.back(); ++InsertPos;
573 F->getBasicBlockList().splice(InsertPos, F->getBasicBlockList(), BEBlock);
574
575 // Now that the block has been inserted into the function, create PHI nodes in
576 // the backedge block which correspond to any PHI nodes in the header block.
577 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
578 PHINode *PN = cast<PHINode>(I);
579 PHINode *NewPN = PHINode::Create(PN->getType(), PN->getName()+".be",
580 BETerminator);
581 NewPN->reserveOperandSpace(BackedgeBlocks.size());
582 if (AA) AA->copyValue(PN, NewPN);
583
584 // Loop over the PHI node, moving all entries except the one for the
585 // preheader over to the new PHI node.
586 unsigned PreheaderIdx = ~0U;
587 bool HasUniqueIncomingValue = true;
588 Value *UniqueValue = 0;
589 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
590 BasicBlock *IBB = PN->getIncomingBlock(i);
591 Value *IV = PN->getIncomingValue(i);
592 if (IBB == Preheader) {
593 PreheaderIdx = i;
594 } else {
595 NewPN->addIncoming(IV, IBB);
596 if (HasUniqueIncomingValue) {
597 if (UniqueValue == 0)
598 UniqueValue = IV;
599 else if (UniqueValue != IV)
600 HasUniqueIncomingValue = false;
601 }
602 }
603 }
604
605 // Delete all of the incoming values from the old PN except the preheader's
606 assert(PreheaderIdx != ~0U && "PHI has no preheader entry??");
607 if (PreheaderIdx != 0) {
608 PN->setIncomingValue(0, PN->getIncomingValue(PreheaderIdx));
609 PN->setIncomingBlock(0, PN->getIncomingBlock(PreheaderIdx));
610 }
611 // Nuke all entries except the zero'th.
612 for (unsigned i = 0, e = PN->getNumIncomingValues()-1; i != e; ++i)
613 PN->removeIncomingValue(e-i, false);
614
615 // Finally, add the newly constructed PHI node as the entry for the BEBlock.
616 PN->addIncoming(NewPN, BEBlock);
617
618 // As an optimization, if all incoming values in the new PhiNode (which is a
619 // subset of the incoming values of the old PHI node) have the same value,
620 // eliminate the PHI Node.
621 if (HasUniqueIncomingValue) {
622 NewPN->replaceAllUsesWith(UniqueValue);
623 if (AA) AA->deleteValue(NewPN);
624 BEBlock->getInstList().erase(NewPN);
625 }
626 }
627
628 // Now that all of the PHI nodes have been inserted and adjusted, modify the
629 // backedge blocks to just to the BEBlock instead of the header.
630 for (unsigned i = 0, e = BackedgeBlocks.size(); i != e; ++i) {
631 TerminatorInst *TI = BackedgeBlocks[i]->getTerminator();
632 for (unsigned Op = 0, e = TI->getNumSuccessors(); Op != e; ++Op)
633 if (TI->getSuccessor(Op) == Header)
634 TI->setSuccessor(Op, BEBlock);
635 }
636
637 //===--- Update all analyses which we must preserve now -----------------===//
638
639 // Update Loop Information - we know that this block is now in the current
640 // loop and all parent loops.
641 L->addBasicBlockToLoop(BEBlock, LI->getBase());
642
643 // Update dominator information
644 DT->splitBlock(BEBlock);
645 if (DominanceFrontier *DF = getAnalysisIfAvailable<DominanceFrontier>())
646 DF->splitBlock(BEBlock);
647
648 return BEBlock;
649}
650
651void LoopSimplify::verifyAnalysis() const {
652 // It used to be possible to just assert L->isLoopSimplifyForm(), however
653 // with the introduction of indirectbr, there are now cases where it's
654 // not possible to transform a loop as necessary. We can at least check
655 // that there is an indirectbr near any time there's trouble.
656
657 // Indirectbr can interfere with preheader and unique backedge insertion.
658 if (!L->getLoopPreheader() || !L->getLoopLatch()) {
659 bool HasIndBrPred = false;
660 for (pred_iterator PI = pred_begin(L->getHeader()),
661 PE = pred_end(L->getHeader()); PI != PE; ++PI)
662 if (isa<IndirectBrInst>((*PI)->getTerminator())) {
663 HasIndBrPred = true;
664 break;
665 }
666 assert(HasIndBrPred &&
667 "LoopSimplify has no excuse for missing loop header info!");
668 }
669
670 // Indirectbr can interfere with exit block canonicalization.
671 if (!L->hasDedicatedExits()) {
672 bool HasIndBrExiting = false;
673 SmallVector<BasicBlock*, 8> ExitingBlocks;
674 L->getExitingBlocks(ExitingBlocks);
675 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i)
676 if (isa<IndirectBrInst>((ExitingBlocks[i])->getTerminator())) {
677 HasIndBrExiting = true;
678 break;
679 }
680 assert(HasIndBrExiting &&
681 "LoopSimplify has no excuse for missing exit block info!");
682 }
683}
| 314 return Changed;
315}
316
317/// InsertPreheaderForLoop - Once we discover that a loop doesn't have a
318/// preheader, this method is called to insert one. This method has two phases:
319/// preheader insertion and analysis updating.
320///
321BasicBlock *LoopSimplify::InsertPreheaderForLoop(Loop *L) {
322 BasicBlock *Header = L->getHeader();
323
324 // Compute the set of predecessors of the loop that are not in the loop.
325 SmallVector<BasicBlock*, 8> OutsideBlocks;
326 for (pred_iterator PI = pred_begin(Header), PE = pred_end(Header);
327 PI != PE; ++PI)
328 if (!L->contains(*PI)) { // Coming in from outside the loop?
329 // If the loop is branched to from an indirect branch, we won't
330 // be able to fully transform the loop, because it prohibits
331 // edge splitting.
332 if (isa<IndirectBrInst>((*PI)->getTerminator())) return 0;
333
334 // Keep track of it.
335 OutsideBlocks.push_back(*PI);
336 }
337
338 // Split out the loop pre-header.
339 BasicBlock *NewBB =
340 SplitBlockPredecessors(Header, &OutsideBlocks[0], OutsideBlocks.size(),
341 ".preheader", this);
342
343 // Make sure that NewBB is put someplace intelligent, which doesn't mess up
344 // code layout too horribly.
345 PlaceSplitBlockCarefully(NewBB, OutsideBlocks, L);
346
347 return NewBB;
348}
349
350/// RewriteLoopExitBlock - Ensure that the loop preheader dominates all exit
351/// blocks. This method is used to split exit blocks that have predecessors
352/// outside of the loop.
353BasicBlock *LoopSimplify::RewriteLoopExitBlock(Loop *L, BasicBlock *Exit) {
354 SmallVector<BasicBlock*, 8> LoopBlocks;
355 for (pred_iterator I = pred_begin(Exit), E = pred_end(Exit); I != E; ++I)
356 if (L->contains(*I)) {
357 // Don't do this if the loop is exited via an indirect branch.
358 if (isa<IndirectBrInst>((*I)->getTerminator())) return 0;
359
360 LoopBlocks.push_back(*I);
361 }
362
363 assert(!LoopBlocks.empty() && "No edges coming in from outside the loop?");
364 BasicBlock *NewBB = SplitBlockPredecessors(Exit, &LoopBlocks[0],
365 LoopBlocks.size(), ".loopexit",
366 this);
367
368 return NewBB;
369}
370
371/// AddBlockAndPredsToSet - Add the specified block, and all of its
372/// predecessors, to the specified set, if it's not already in there. Stop
373/// predecessor traversal when we reach StopBlock.
374static void AddBlockAndPredsToSet(BasicBlock *InputBB, BasicBlock *StopBlock,
375 std::set<BasicBlock*> &Blocks) {
376 std::vector<BasicBlock *> WorkList;
377 WorkList.push_back(InputBB);
378 do {
379 BasicBlock *BB = WorkList.back(); WorkList.pop_back();
380 if (Blocks.insert(BB).second && BB != StopBlock)
381 // If BB is not already processed and it is not a stop block then
382 // insert its predecessor in the work list
383 for (pred_iterator I = pred_begin(BB), E = pred_end(BB); I != E; ++I) {
384 BasicBlock *WBB = *I;
385 WorkList.push_back(WBB);
386 }
387 } while(!WorkList.empty());
388}
389
390/// FindPHIToPartitionLoops - The first part of loop-nestification is to find a
391/// PHI node that tells us how to partition the loops.
392static PHINode *FindPHIToPartitionLoops(Loop *L, DominatorTree *DT,
393 AliasAnalysis *AA) {
394 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ) {
395 PHINode *PN = cast<PHINode>(I);
396 ++I;
397 if (Value *V = PN->hasConstantValue(DT)) {
398 // This is a degenerate PHI already, don't modify it!
399 PN->replaceAllUsesWith(V);
400 if (AA) AA->deleteValue(PN);
401 PN->eraseFromParent();
402 continue;
403 }
404
405 // Scan this PHI node looking for a use of the PHI node by itself.
406 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
407 if (PN->getIncomingValue(i) == PN &&
408 L->contains(PN->getIncomingBlock(i)))
409 // We found something tasty to remove.
410 return PN;
411 }
412 return 0;
413}
414
415// PlaceSplitBlockCarefully - If the block isn't already, move the new block to
416// right after some 'outside block' block. This prevents the preheader from
417// being placed inside the loop body, e.g. when the loop hasn't been rotated.
418void LoopSimplify::PlaceSplitBlockCarefully(BasicBlock *NewBB,
419 SmallVectorImpl<BasicBlock*> &SplitPreds,
420 Loop *L) {
421 // Check to see if NewBB is already well placed.
422 Function::iterator BBI = NewBB; --BBI;
423 for (unsigned i = 0, e = SplitPreds.size(); i != e; ++i) {
424 if (&*BBI == SplitPreds[i])
425 return;
426 }
427
428 // If it isn't already after an outside block, move it after one. This is
429 // always good as it makes the uncond branch from the outside block into a
430 // fall-through.
431
432 // Figure out *which* outside block to put this after. Prefer an outside
433 // block that neighbors a BB actually in the loop.
434 BasicBlock *FoundBB = 0;
435 for (unsigned i = 0, e = SplitPreds.size(); i != e; ++i) {
436 Function::iterator BBI = SplitPreds[i];
437 if (++BBI != NewBB->getParent()->end() &&
438 L->contains(BBI)) {
439 FoundBB = SplitPreds[i];
440 break;
441 }
442 }
443
444 // If our heuristic for a *good* bb to place this after doesn't find
445 // anything, just pick something. It's likely better than leaving it within
446 // the loop.
447 if (!FoundBB)
448 FoundBB = SplitPreds[0];
449 NewBB->moveAfter(FoundBB);
450}
451
452
453/// SeparateNestedLoop - If this loop has multiple backedges, try to pull one of
454/// them out into a nested loop. This is important for code that looks like
455/// this:
456///
457/// Loop:
458/// ...
459/// br cond, Loop, Next
460/// ...
461/// br cond2, Loop, Out
462///
463/// To identify this common case, we look at the PHI nodes in the header of the
464/// loop. PHI nodes with unchanging values on one backedge correspond to values
465/// that change in the "outer" loop, but not in the "inner" loop.
466///
467/// If we are able to separate out a loop, return the new outer loop that was
468/// created.
469///
470Loop *LoopSimplify::SeparateNestedLoop(Loop *L, LPPassManager &LPM) {
471 PHINode *PN = FindPHIToPartitionLoops(L, DT, AA);
472 if (PN == 0) return 0; // No known way to partition.
473
474 // Pull out all predecessors that have varying values in the loop. This
475 // handles the case when a PHI node has multiple instances of itself as
476 // arguments.
477 SmallVector<BasicBlock*, 8> OuterLoopPreds;
478 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
479 if (PN->getIncomingValue(i) != PN ||
480 !L->contains(PN->getIncomingBlock(i)))
481 OuterLoopPreds.push_back(PN->getIncomingBlock(i));
482
483 BasicBlock *Header = L->getHeader();
484 BasicBlock *NewBB = SplitBlockPredecessors(Header, &OuterLoopPreds[0],
485 OuterLoopPreds.size(),
486 ".outer", this);
487
488 // Make sure that NewBB is put someplace intelligent, which doesn't mess up
489 // code layout too horribly.
490 PlaceSplitBlockCarefully(NewBB, OuterLoopPreds, L);
491
492 // Create the new outer loop.
493 Loop *NewOuter = new Loop();
494
495 // Change the parent loop to use the outer loop as its child now.
496 if (Loop *Parent = L->getParentLoop())
497 Parent->replaceChildLoopWith(L, NewOuter);
498 else
499 LI->changeTopLevelLoop(L, NewOuter);
500
501 // L is now a subloop of our outer loop.
502 NewOuter->addChildLoop(L);
503
504 // Add the new loop to the pass manager queue.
505 LPM.insertLoopIntoQueue(NewOuter);
506
507 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
508 I != E; ++I)
509 NewOuter->addBlockEntry(*I);
510
511 // Now reset the header in L, which had been moved by
512 // SplitBlockPredecessors for the outer loop.
513 L->moveToHeader(Header);
514
515 // Determine which blocks should stay in L and which should be moved out to
516 // the Outer loop now.
517 std::set<BasicBlock*> BlocksInL;
518 for (pred_iterator PI = pred_begin(Header), E = pred_end(Header); PI!=E; ++PI)
519 if (DT->dominates(Header, *PI))
520 AddBlockAndPredsToSet(*PI, Header, BlocksInL);
521
522
523 // Scan all of the loop children of L, moving them to OuterLoop if they are
524 // not part of the inner loop.
525 const std::vector<Loop*> &SubLoops = L->getSubLoops();
526 for (size_t I = 0; I != SubLoops.size(); )
527 if (BlocksInL.count(SubLoops[I]->getHeader()))
528 ++I; // Loop remains in L
529 else
530 NewOuter->addChildLoop(L->removeChildLoop(SubLoops.begin() + I));
531
532 // Now that we know which blocks are in L and which need to be moved to
533 // OuterLoop, move any blocks that need it.
534 for (unsigned i = 0; i != L->getBlocks().size(); ++i) {
535 BasicBlock *BB = L->getBlocks()[i];
536 if (!BlocksInL.count(BB)) {
537 // Move this block to the parent, updating the exit blocks sets
538 L->removeBlockFromLoop(BB);
539 if ((*LI)[BB] == L)
540 LI->changeLoopFor(BB, NewOuter);
541 --i;
542 }
543 }
544
545 return NewOuter;
546}
547
548
549
550/// InsertUniqueBackedgeBlock - This method is called when the specified loop
551/// has more than one backedge in it. If this occurs, revector all of these
552/// backedges to target a new basic block and have that block branch to the loop
553/// header. This ensures that loops have exactly one backedge.
554///
555BasicBlock *
556LoopSimplify::InsertUniqueBackedgeBlock(Loop *L, BasicBlock *Preheader) {
557 assert(L->getNumBackEdges() > 1 && "Must have > 1 backedge!");
558
559 // Get information about the loop
560 BasicBlock *Header = L->getHeader();
561 Function *F = Header->getParent();
562
563 // Unique backedge insertion currently depends on having a preheader.
564 if (!Preheader)
565 return 0;
566
567 // Figure out which basic blocks contain back-edges to the loop header.
568 std::vector<BasicBlock*> BackedgeBlocks;
569 for (pred_iterator I = pred_begin(Header), E = pred_end(Header); I != E; ++I)
570 if (*I != Preheader) BackedgeBlocks.push_back(*I);
571
572 // Create and insert the new backedge block...
573 BasicBlock *BEBlock = BasicBlock::Create(Header->getContext(),
574 Header->getName()+".backedge", F);
575 BranchInst *BETerminator = BranchInst::Create(Header, BEBlock);
576
577 // Move the new backedge block to right after the last backedge block.
578 Function::iterator InsertPos = BackedgeBlocks.back(); ++InsertPos;
579 F->getBasicBlockList().splice(InsertPos, F->getBasicBlockList(), BEBlock);
580
581 // Now that the block has been inserted into the function, create PHI nodes in
582 // the backedge block which correspond to any PHI nodes in the header block.
583 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
584 PHINode *PN = cast<PHINode>(I);
585 PHINode *NewPN = PHINode::Create(PN->getType(), PN->getName()+".be",
586 BETerminator);
587 NewPN->reserveOperandSpace(BackedgeBlocks.size());
588 if (AA) AA->copyValue(PN, NewPN);
589
590 // Loop over the PHI node, moving all entries except the one for the
591 // preheader over to the new PHI node.
592 unsigned PreheaderIdx = ~0U;
593 bool HasUniqueIncomingValue = true;
594 Value *UniqueValue = 0;
595 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
596 BasicBlock *IBB = PN->getIncomingBlock(i);
597 Value *IV = PN->getIncomingValue(i);
598 if (IBB == Preheader) {
599 PreheaderIdx = i;
600 } else {
601 NewPN->addIncoming(IV, IBB);
602 if (HasUniqueIncomingValue) {
603 if (UniqueValue == 0)
604 UniqueValue = IV;
605 else if (UniqueValue != IV)
606 HasUniqueIncomingValue = false;
607 }
608 }
609 }
610
611 // Delete all of the incoming values from the old PN except the preheader's
612 assert(PreheaderIdx != ~0U && "PHI has no preheader entry??");
613 if (PreheaderIdx != 0) {
614 PN->setIncomingValue(0, PN->getIncomingValue(PreheaderIdx));
615 PN->setIncomingBlock(0, PN->getIncomingBlock(PreheaderIdx));
616 }
617 // Nuke all entries except the zero'th.
618 for (unsigned i = 0, e = PN->getNumIncomingValues()-1; i != e; ++i)
619 PN->removeIncomingValue(e-i, false);
620
621 // Finally, add the newly constructed PHI node as the entry for the BEBlock.
622 PN->addIncoming(NewPN, BEBlock);
623
624 // As an optimization, if all incoming values in the new PhiNode (which is a
625 // subset of the incoming values of the old PHI node) have the same value,
626 // eliminate the PHI Node.
627 if (HasUniqueIncomingValue) {
628 NewPN->replaceAllUsesWith(UniqueValue);
629 if (AA) AA->deleteValue(NewPN);
630 BEBlock->getInstList().erase(NewPN);
631 }
632 }
633
634 // Now that all of the PHI nodes have been inserted and adjusted, modify the
635 // backedge blocks to just to the BEBlock instead of the header.
636 for (unsigned i = 0, e = BackedgeBlocks.size(); i != e; ++i) {
637 TerminatorInst *TI = BackedgeBlocks[i]->getTerminator();
638 for (unsigned Op = 0, e = TI->getNumSuccessors(); Op != e; ++Op)
639 if (TI->getSuccessor(Op) == Header)
640 TI->setSuccessor(Op, BEBlock);
641 }
642
643 //===--- Update all analyses which we must preserve now -----------------===//
644
645 // Update Loop Information - we know that this block is now in the current
646 // loop and all parent loops.
647 L->addBasicBlockToLoop(BEBlock, LI->getBase());
648
649 // Update dominator information
650 DT->splitBlock(BEBlock);
651 if (DominanceFrontier *DF = getAnalysisIfAvailable<DominanceFrontier>())
652 DF->splitBlock(BEBlock);
653
654 return BEBlock;
655}
656
657void LoopSimplify::verifyAnalysis() const {
658 // It used to be possible to just assert L->isLoopSimplifyForm(), however
659 // with the introduction of indirectbr, there are now cases where it's
660 // not possible to transform a loop as necessary. We can at least check
661 // that there is an indirectbr near any time there's trouble.
662
663 // Indirectbr can interfere with preheader and unique backedge insertion.
664 if (!L->getLoopPreheader() || !L->getLoopLatch()) {
665 bool HasIndBrPred = false;
666 for (pred_iterator PI = pred_begin(L->getHeader()),
667 PE = pred_end(L->getHeader()); PI != PE; ++PI)
668 if (isa<IndirectBrInst>((*PI)->getTerminator())) {
669 HasIndBrPred = true;
670 break;
671 }
672 assert(HasIndBrPred &&
673 "LoopSimplify has no excuse for missing loop header info!");
674 }
675
676 // Indirectbr can interfere with exit block canonicalization.
677 if (!L->hasDedicatedExits()) {
678 bool HasIndBrExiting = false;
679 SmallVector<BasicBlock*, 8> ExitingBlocks;
680 L->getExitingBlocks(ExitingBlocks);
681 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i)
682 if (isa<IndirectBrInst>((ExitingBlocks[i])->getTerminator())) {
683 HasIndBrExiting = true;
684 break;
685 }
686 assert(HasIndBrExiting &&
687 "LoopSimplify has no excuse for missing exit block info!");
688 }
689}
|