Disable stack coloring with register for now. It's not able to set kill markers.
[llvm/avr.git] / lib / Transforms / Utils / SimplifyCFG.cpp
blob0f9493dac264dbe9afc3d7bcaea8a286693d5e16
1 //===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===//
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 // Peephole optimize the CFG.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "simplifycfg"
15 #include "llvm/Transforms/Utils/Local.h"
16 #include "llvm/Constants.h"
17 #include "llvm/Instructions.h"
18 #include "llvm/IntrinsicInst.h"
19 #include "llvm/LLVMContext.h"
20 #include "llvm/Type.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/GlobalVariable.h"
23 #include "llvm/Support/CFG.h"
24 #include "llvm/Support/Debug.h"
25 #include "llvm/Support/raw_ostream.h"
26 #include "llvm/Analysis/ConstantFolding.h"
27 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
28 #include "llvm/ADT/SmallVector.h"
29 #include "llvm/ADT/SmallPtrSet.h"
30 #include "llvm/ADT/Statistic.h"
31 #include <algorithm>
32 #include <functional>
33 #include <set>
34 #include <map>
35 using namespace llvm;
37 STATISTIC(NumSpeculations, "Number of speculative executed instructions");
39 /// SafeToMergeTerminators - Return true if it is safe to merge these two
40 /// terminator instructions together.
41 ///
42 static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) {
43 if (SI1 == SI2) return false; // Can't merge with self!
45 // It is not safe to merge these two switch instructions if they have a common
46 // successor, and if that successor has a PHI node, and if *that* PHI node has
47 // conflicting incoming values from the two switch blocks.
48 BasicBlock *SI1BB = SI1->getParent();
49 BasicBlock *SI2BB = SI2->getParent();
50 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB));
52 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I)
53 if (SI1Succs.count(*I))
54 for (BasicBlock::iterator BBI = (*I)->begin();
55 isa<PHINode>(BBI); ++BBI) {
56 PHINode *PN = cast<PHINode>(BBI);
57 if (PN->getIncomingValueForBlock(SI1BB) !=
58 PN->getIncomingValueForBlock(SI2BB))
59 return false;
62 return true;
65 /// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will
66 /// now be entries in it from the 'NewPred' block. The values that will be
67 /// flowing into the PHI nodes will be the same as those coming in from
68 /// ExistPred, an existing predecessor of Succ.
69 static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred,
70 BasicBlock *ExistPred) {
71 assert(std::find(succ_begin(ExistPred), succ_end(ExistPred), Succ) !=
72 succ_end(ExistPred) && "ExistPred is not a predecessor of Succ!");
73 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do
75 PHINode *PN;
76 for (BasicBlock::iterator I = Succ->begin();
77 (PN = dyn_cast<PHINode>(I)); ++I)
78 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred);
81 /// CanPropagatePredecessorsForPHIs - Return true if we can fold BB, an
82 /// almost-empty BB ending in an unconditional branch to Succ, into succ.
83 ///
84 /// Assumption: Succ is the single successor for BB.
85 ///
86 static bool CanPropagatePredecessorsForPHIs(BasicBlock *BB, BasicBlock *Succ) {
87 assert(*succ_begin(BB) == Succ && "Succ is not successor of BB!");
89 DEBUG(errs() << "Looking to fold " << BB->getName() << " into "
90 << Succ->getName() << "\n");
91 // Shortcut, if there is only a single predecessor it must be BB and merging
92 // is always safe
93 if (Succ->getSinglePredecessor()) return true;
95 typedef SmallPtrSet<Instruction*, 16> InstrSet;
96 InstrSet BBPHIs;
98 // Make a list of all phi nodes in BB
99 BasicBlock::iterator BBI = BB->begin();
100 while (isa<PHINode>(*BBI)) BBPHIs.insert(BBI++);
102 // Make a list of the predecessors of BB
103 typedef SmallPtrSet<BasicBlock*, 16> BlockSet;
104 BlockSet BBPreds(pred_begin(BB), pred_end(BB));
106 // Use that list to make another list of common predecessors of BB and Succ
107 BlockSet CommonPreds;
108 for (pred_iterator PI = pred_begin(Succ), PE = pred_end(Succ);
109 PI != PE; ++PI)
110 if (BBPreds.count(*PI))
111 CommonPreds.insert(*PI);
113 // Shortcut, if there are no common predecessors, merging is always safe
114 if (CommonPreds.empty())
115 return true;
117 // Look at all the phi nodes in Succ, to see if they present a conflict when
118 // merging these blocks
119 for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) {
120 PHINode *PN = cast<PHINode>(I);
122 // If the incoming value from BB is again a PHINode in
123 // BB which has the same incoming value for *PI as PN does, we can
124 // merge the phi nodes and then the blocks can still be merged
125 PHINode *BBPN = dyn_cast<PHINode>(PN->getIncomingValueForBlock(BB));
126 if (BBPN && BBPN->getParent() == BB) {
127 for (BlockSet::iterator PI = CommonPreds.begin(), PE = CommonPreds.end();
128 PI != PE; PI++) {
129 if (BBPN->getIncomingValueForBlock(*PI)
130 != PN->getIncomingValueForBlock(*PI)) {
131 DEBUG(errs() << "Can't fold, phi node " << PN->getName() << " in "
132 << Succ->getName() << " is conflicting with "
133 << BBPN->getName() << " with regard to common predecessor "
134 << (*PI)->getName() << "\n");
135 return false;
138 // Remove this phinode from the list of phis in BB, since it has been
139 // handled.
140 BBPHIs.erase(BBPN);
141 } else {
142 Value* Val = PN->getIncomingValueForBlock(BB);
143 for (BlockSet::iterator PI = CommonPreds.begin(), PE = CommonPreds.end();
144 PI != PE; PI++) {
145 // See if the incoming value for the common predecessor is equal to the
146 // one for BB, in which case this phi node will not prevent the merging
147 // of the block.
148 if (Val != PN->getIncomingValueForBlock(*PI)) {
149 DEBUG(errs() << "Can't fold, phi node " << PN->getName() << " in "
150 << Succ->getName() << " is conflicting with regard to common "
151 << "predecessor " << (*PI)->getName() << "\n");
152 return false;
158 // If there are any other phi nodes in BB that don't have a phi node in Succ
159 // to merge with, they must be moved to Succ completely. However, for any
160 // predecessors of Succ, branches will be added to the phi node that just
161 // point to itself. So, for any common predecessors, this must not cause
162 // conflicts.
163 for (InstrSet::iterator I = BBPHIs.begin(), E = BBPHIs.end();
164 I != E; I++) {
165 PHINode *PN = cast<PHINode>(*I);
166 for (BlockSet::iterator PI = CommonPreds.begin(), PE = CommonPreds.end();
167 PI != PE; PI++)
168 if (PN->getIncomingValueForBlock(*PI) != PN) {
169 DEBUG(errs() << "Can't fold, phi node " << PN->getName() << " in "
170 << BB->getName() << " is conflicting with regard to common "
171 << "predecessor " << (*PI)->getName() << "\n");
172 return false;
176 return true;
179 /// TryToSimplifyUncondBranchFromEmptyBlock - BB contains an unconditional
180 /// branch to Succ, and contains no instructions other than PHI nodes and the
181 /// branch. If possible, eliminate BB.
182 static bool TryToSimplifyUncondBranchFromEmptyBlock(BasicBlock *BB,
183 BasicBlock *Succ) {
184 // Check to see if merging these blocks would cause conflicts for any of the
185 // phi nodes in BB or Succ. If not, we can safely merge.
186 if (!CanPropagatePredecessorsForPHIs(BB, Succ)) return false;
188 DOUT << "Killing Trivial BB: \n" << *BB;
190 if (isa<PHINode>(Succ->begin())) {
191 // If there is more than one pred of succ, and there are PHI nodes in
192 // the successor, then we need to add incoming edges for the PHI nodes
194 const SmallVector<BasicBlock*, 16> BBPreds(pred_begin(BB), pred_end(BB));
196 // Loop over all of the PHI nodes in the successor of BB.
197 for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) {
198 PHINode *PN = cast<PHINode>(I);
199 Value *OldVal = PN->removeIncomingValue(BB, false);
200 assert(OldVal && "No entry in PHI for Pred BB!");
202 // If this incoming value is one of the PHI nodes in BB, the new entries
203 // in the PHI node are the entries from the old PHI.
204 if (isa<PHINode>(OldVal) && cast<PHINode>(OldVal)->getParent() == BB) {
205 PHINode *OldValPN = cast<PHINode>(OldVal);
206 for (unsigned i = 0, e = OldValPN->getNumIncomingValues(); i != e; ++i)
207 // Note that, since we are merging phi nodes and BB and Succ might
208 // have common predecessors, we could end up with a phi node with
209 // identical incoming branches. This will be cleaned up later (and
210 // will trigger asserts if we try to clean it up now, without also
211 // simplifying the corresponding conditional branch).
212 PN->addIncoming(OldValPN->getIncomingValue(i),
213 OldValPN->getIncomingBlock(i));
214 } else {
215 // Add an incoming value for each of the new incoming values.
216 for (unsigned i = 0, e = BBPreds.size(); i != e; ++i)
217 PN->addIncoming(OldVal, BBPreds[i]);
222 if (isa<PHINode>(&BB->front())) {
223 SmallVector<BasicBlock*, 16>
224 OldSuccPreds(pred_begin(Succ), pred_end(Succ));
226 // Move all PHI nodes in BB to Succ if they are alive, otherwise
227 // delete them.
228 while (PHINode *PN = dyn_cast<PHINode>(&BB->front())) {
229 if (PN->use_empty()) {
230 // Just remove the dead phi. This happens if Succ's PHIs were the only
231 // users of the PHI nodes.
232 PN->eraseFromParent();
233 continue;
236 // The instruction is alive, so this means that BB must dominate all
237 // predecessors of Succ (Since all uses of the PN are after its
238 // definition, so in Succ or a block dominated by Succ. If a predecessor
239 // of Succ would not be dominated by BB, PN would violate the def before
240 // use SSA demand). Therefore, we can simply move the phi node to the
241 // next block.
242 Succ->getInstList().splice(Succ->begin(),
243 BB->getInstList(), BB->begin());
245 // We need to add new entries for the PHI node to account for
246 // predecessors of Succ that the PHI node does not take into
247 // account. At this point, since we know that BB dominated succ and all
248 // of its predecessors, this means that we should any newly added
249 // incoming edges should use the PHI node itself as the value for these
250 // edges, because they are loop back edges.
251 for (unsigned i = 0, e = OldSuccPreds.size(); i != e; ++i)
252 if (OldSuccPreds[i] != BB)
253 PN->addIncoming(PN, OldSuccPreds[i]);
257 // Everything that jumped to BB now goes to Succ.
258 BB->replaceAllUsesWith(Succ);
259 if (!Succ->hasName()) Succ->takeName(BB);
260 BB->eraseFromParent(); // Delete the old basic block.
261 return true;
264 /// GetIfCondition - Given a basic block (BB) with two predecessors (and
265 /// presumably PHI nodes in it), check to see if the merge at this block is due
266 /// to an "if condition". If so, return the boolean condition that determines
267 /// which entry into BB will be taken. Also, return by references the block
268 /// that will be entered from if the condition is true, and the block that will
269 /// be entered if the condition is false.
272 static Value *GetIfCondition(BasicBlock *BB,
273 BasicBlock *&IfTrue, BasicBlock *&IfFalse) {
274 assert(std::distance(pred_begin(BB), pred_end(BB)) == 2 &&
275 "Function can only handle blocks with 2 predecessors!");
276 BasicBlock *Pred1 = *pred_begin(BB);
277 BasicBlock *Pred2 = *++pred_begin(BB);
279 // We can only handle branches. Other control flow will be lowered to
280 // branches if possible anyway.
281 if (!isa<BranchInst>(Pred1->getTerminator()) ||
282 !isa<BranchInst>(Pred2->getTerminator()))
283 return 0;
284 BranchInst *Pred1Br = cast<BranchInst>(Pred1->getTerminator());
285 BranchInst *Pred2Br = cast<BranchInst>(Pred2->getTerminator());
287 // Eliminate code duplication by ensuring that Pred1Br is conditional if
288 // either are.
289 if (Pred2Br->isConditional()) {
290 // If both branches are conditional, we don't have an "if statement". In
291 // reality, we could transform this case, but since the condition will be
292 // required anyway, we stand no chance of eliminating it, so the xform is
293 // probably not profitable.
294 if (Pred1Br->isConditional())
295 return 0;
297 std::swap(Pred1, Pred2);
298 std::swap(Pred1Br, Pred2Br);
301 if (Pred1Br->isConditional()) {
302 // If we found a conditional branch predecessor, make sure that it branches
303 // to BB and Pred2Br. If it doesn't, this isn't an "if statement".
304 if (Pred1Br->getSuccessor(0) == BB &&
305 Pred1Br->getSuccessor(1) == Pred2) {
306 IfTrue = Pred1;
307 IfFalse = Pred2;
308 } else if (Pred1Br->getSuccessor(0) == Pred2 &&
309 Pred1Br->getSuccessor(1) == BB) {
310 IfTrue = Pred2;
311 IfFalse = Pred1;
312 } else {
313 // We know that one arm of the conditional goes to BB, so the other must
314 // go somewhere unrelated, and this must not be an "if statement".
315 return 0;
318 // The only thing we have to watch out for here is to make sure that Pred2
319 // doesn't have incoming edges from other blocks. If it does, the condition
320 // doesn't dominate BB.
321 if (++pred_begin(Pred2) != pred_end(Pred2))
322 return 0;
324 return Pred1Br->getCondition();
327 // Ok, if we got here, both predecessors end with an unconditional branch to
328 // BB. Don't panic! If both blocks only have a single (identical)
329 // predecessor, and THAT is a conditional branch, then we're all ok!
330 if (pred_begin(Pred1) == pred_end(Pred1) ||
331 ++pred_begin(Pred1) != pred_end(Pred1) ||
332 pred_begin(Pred2) == pred_end(Pred2) ||
333 ++pred_begin(Pred2) != pred_end(Pred2) ||
334 *pred_begin(Pred1) != *pred_begin(Pred2))
335 return 0;
337 // Otherwise, if this is a conditional branch, then we can use it!
338 BasicBlock *CommonPred = *pred_begin(Pred1);
339 if (BranchInst *BI = dyn_cast<BranchInst>(CommonPred->getTerminator())) {
340 assert(BI->isConditional() && "Two successors but not conditional?");
341 if (BI->getSuccessor(0) == Pred1) {
342 IfTrue = Pred1;
343 IfFalse = Pred2;
344 } else {
345 IfTrue = Pred2;
346 IfFalse = Pred1;
348 return BI->getCondition();
350 return 0;
353 /// DominatesMergePoint - If we have a merge point of an "if condition" as
354 /// accepted above, return true if the specified value dominates the block. We
355 /// don't handle the true generality of domination here, just a special case
356 /// which works well enough for us.
358 /// If AggressiveInsts is non-null, and if V does not dominate BB, we check to
359 /// see if V (which must be an instruction) is cheap to compute and is
360 /// non-trapping. If both are true, the instruction is inserted into the set
361 /// and true is returned.
362 static bool DominatesMergePoint(Value *V, BasicBlock *BB,
363 std::set<Instruction*> *AggressiveInsts) {
364 Instruction *I = dyn_cast<Instruction>(V);
365 if (!I) {
366 // Non-instructions all dominate instructions, but not all constantexprs
367 // can be executed unconditionally.
368 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V))
369 if (C->canTrap())
370 return false;
371 return true;
373 BasicBlock *PBB = I->getParent();
375 // We don't want to allow weird loops that might have the "if condition" in
376 // the bottom of this block.
377 if (PBB == BB) return false;
379 // If this instruction is defined in a block that contains an unconditional
380 // branch to BB, then it must be in the 'conditional' part of the "if
381 // statement".
382 if (BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator()))
383 if (BI->isUnconditional() && BI->getSuccessor(0) == BB) {
384 if (!AggressiveInsts) return false;
385 // Okay, it looks like the instruction IS in the "condition". Check to
386 // see if its a cheap instruction to unconditionally compute, and if it
387 // only uses stuff defined outside of the condition. If so, hoist it out.
388 if (!I->isSafeToSpeculativelyExecute())
389 return false;
391 switch (I->getOpcode()) {
392 default: return false; // Cannot hoist this out safely.
393 case Instruction::Load: {
394 // We have to check to make sure there are no instructions before the
395 // load in its basic block, as we are going to hoist the loop out to
396 // its predecessor.
397 BasicBlock::iterator IP = PBB->begin();
398 while (isa<DbgInfoIntrinsic>(IP))
399 IP++;
400 if (IP != BasicBlock::iterator(I))
401 return false;
402 break;
404 case Instruction::Add:
405 case Instruction::Sub:
406 case Instruction::And:
407 case Instruction::Or:
408 case Instruction::Xor:
409 case Instruction::Shl:
410 case Instruction::LShr:
411 case Instruction::AShr:
412 case Instruction::ICmp:
413 break; // These are all cheap and non-trapping instructions.
416 // Okay, we can only really hoist these out if their operands are not
417 // defined in the conditional region.
418 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
419 if (!DominatesMergePoint(*i, BB, 0))
420 return false;
421 // Okay, it's safe to do this! Remember this instruction.
422 AggressiveInsts->insert(I);
425 return true;
428 /// GatherConstantSetEQs - Given a potentially 'or'd together collection of
429 /// icmp_eq instructions that compare a value against a constant, return the
430 /// value being compared, and stick the constant into the Values vector.
431 static Value *GatherConstantSetEQs(Value *V, std::vector<ConstantInt*> &Values){
432 if (Instruction *Inst = dyn_cast<Instruction>(V)) {
433 if (Inst->getOpcode() == Instruction::ICmp &&
434 cast<ICmpInst>(Inst)->getPredicate() == ICmpInst::ICMP_EQ) {
435 if (ConstantInt *C = dyn_cast<ConstantInt>(Inst->getOperand(1))) {
436 Values.push_back(C);
437 return Inst->getOperand(0);
438 } else if (ConstantInt *C = dyn_cast<ConstantInt>(Inst->getOperand(0))) {
439 Values.push_back(C);
440 return Inst->getOperand(1);
442 } else if (Inst->getOpcode() == Instruction::Or) {
443 if (Value *LHS = GatherConstantSetEQs(Inst->getOperand(0), Values))
444 if (Value *RHS = GatherConstantSetEQs(Inst->getOperand(1), Values))
445 if (LHS == RHS)
446 return LHS;
449 return 0;
452 /// GatherConstantSetNEs - Given a potentially 'and'd together collection of
453 /// setne instructions that compare a value against a constant, return the value
454 /// being compared, and stick the constant into the Values vector.
455 static Value *GatherConstantSetNEs(Value *V, std::vector<ConstantInt*> &Values){
456 if (Instruction *Inst = dyn_cast<Instruction>(V)) {
457 if (Inst->getOpcode() == Instruction::ICmp &&
458 cast<ICmpInst>(Inst)->getPredicate() == ICmpInst::ICMP_NE) {
459 if (ConstantInt *C = dyn_cast<ConstantInt>(Inst->getOperand(1))) {
460 Values.push_back(C);
461 return Inst->getOperand(0);
462 } else if (ConstantInt *C = dyn_cast<ConstantInt>(Inst->getOperand(0))) {
463 Values.push_back(C);
464 return Inst->getOperand(1);
466 } else if (Inst->getOpcode() == Instruction::And) {
467 if (Value *LHS = GatherConstantSetNEs(Inst->getOperand(0), Values))
468 if (Value *RHS = GatherConstantSetNEs(Inst->getOperand(1), Values))
469 if (LHS == RHS)
470 return LHS;
473 return 0;
476 /// GatherValueComparisons - If the specified Cond is an 'and' or 'or' of a
477 /// bunch of comparisons of one value against constants, return the value and
478 /// the constants being compared.
479 static bool GatherValueComparisons(Instruction *Cond, Value *&CompVal,
480 std::vector<ConstantInt*> &Values) {
481 if (Cond->getOpcode() == Instruction::Or) {
482 CompVal = GatherConstantSetEQs(Cond, Values);
484 // Return true to indicate that the condition is true if the CompVal is
485 // equal to one of the constants.
486 return true;
487 } else if (Cond->getOpcode() == Instruction::And) {
488 CompVal = GatherConstantSetNEs(Cond, Values);
490 // Return false to indicate that the condition is false if the CompVal is
491 // equal to one of the constants.
492 return false;
494 return false;
497 static void EraseTerminatorInstAndDCECond(TerminatorInst *TI) {
498 Instruction* Cond = 0;
499 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
500 Cond = dyn_cast<Instruction>(SI->getCondition());
501 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
502 if (BI->isConditional())
503 Cond = dyn_cast<Instruction>(BI->getCondition());
506 TI->eraseFromParent();
507 if (Cond) RecursivelyDeleteTriviallyDeadInstructions(Cond);
510 /// isValueEqualityComparison - Return true if the specified terminator checks
511 /// to see if a value is equal to constant integer value.
512 static Value *isValueEqualityComparison(TerminatorInst *TI) {
513 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
514 // Do not permit merging of large switch instructions into their
515 // predecessors unless there is only one predecessor.
516 if (SI->getNumSuccessors() * std::distance(pred_begin(SI->getParent()),
517 pred_end(SI->getParent())) > 128)
518 return 0;
520 return SI->getCondition();
522 if (BranchInst *BI = dyn_cast<BranchInst>(TI))
523 if (BI->isConditional() && BI->getCondition()->hasOneUse())
524 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition()))
525 if ((ICI->getPredicate() == ICmpInst::ICMP_EQ ||
526 ICI->getPredicate() == ICmpInst::ICMP_NE) &&
527 isa<ConstantInt>(ICI->getOperand(1)))
528 return ICI->getOperand(0);
529 return 0;
532 /// GetValueEqualityComparisonCases - Given a value comparison instruction,
533 /// decode all of the 'cases' that it represents and return the 'default' block.
534 static BasicBlock *
535 GetValueEqualityComparisonCases(TerminatorInst *TI,
536 std::vector<std::pair<ConstantInt*,
537 BasicBlock*> > &Cases) {
538 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
539 Cases.reserve(SI->getNumCases());
540 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i)
541 Cases.push_back(std::make_pair(SI->getCaseValue(i), SI->getSuccessor(i)));
542 return SI->getDefaultDest();
545 BranchInst *BI = cast<BranchInst>(TI);
546 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition());
547 Cases.push_back(std::make_pair(cast<ConstantInt>(ICI->getOperand(1)),
548 BI->getSuccessor(ICI->getPredicate() ==
549 ICmpInst::ICMP_NE)));
550 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ);
554 /// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries
555 /// in the list that match the specified block.
556 static void EliminateBlockCases(BasicBlock *BB,
557 std::vector<std::pair<ConstantInt*, BasicBlock*> > &Cases) {
558 for (unsigned i = 0, e = Cases.size(); i != e; ++i)
559 if (Cases[i].second == BB) {
560 Cases.erase(Cases.begin()+i);
561 --i; --e;
565 /// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as
566 /// well.
567 static bool
568 ValuesOverlap(std::vector<std::pair<ConstantInt*, BasicBlock*> > &C1,
569 std::vector<std::pair<ConstantInt*, BasicBlock*> > &C2) {
570 std::vector<std::pair<ConstantInt*, BasicBlock*> > *V1 = &C1, *V2 = &C2;
572 // Make V1 be smaller than V2.
573 if (V1->size() > V2->size())
574 std::swap(V1, V2);
576 if (V1->size() == 0) return false;
577 if (V1->size() == 1) {
578 // Just scan V2.
579 ConstantInt *TheVal = (*V1)[0].first;
580 for (unsigned i = 0, e = V2->size(); i != e; ++i)
581 if (TheVal == (*V2)[i].first)
582 return true;
585 // Otherwise, just sort both lists and compare element by element.
586 std::sort(V1->begin(), V1->end());
587 std::sort(V2->begin(), V2->end());
588 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size();
589 while (i1 != e1 && i2 != e2) {
590 if ((*V1)[i1].first == (*V2)[i2].first)
591 return true;
592 if ((*V1)[i1].first < (*V2)[i2].first)
593 ++i1;
594 else
595 ++i2;
597 return false;
600 /// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a
601 /// terminator instruction and its block is known to only have a single
602 /// predecessor block, check to see if that predecessor is also a value
603 /// comparison with the same value, and if that comparison determines the
604 /// outcome of this comparison. If so, simplify TI. This does a very limited
605 /// form of jump threading.
606 static bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI,
607 BasicBlock *Pred) {
608 Value *PredVal = isValueEqualityComparison(Pred->getTerminator());
609 if (!PredVal) return false; // Not a value comparison in predecessor.
611 Value *ThisVal = isValueEqualityComparison(TI);
612 assert(ThisVal && "This isn't a value comparison!!");
613 if (ThisVal != PredVal) return false; // Different predicates.
615 // Find out information about when control will move from Pred to TI's block.
616 std::vector<std::pair<ConstantInt*, BasicBlock*> > PredCases;
617 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(),
618 PredCases);
619 EliminateBlockCases(PredDef, PredCases); // Remove default from cases.
621 // Find information about how control leaves this block.
622 std::vector<std::pair<ConstantInt*, BasicBlock*> > ThisCases;
623 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases);
624 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases.
626 // If TI's block is the default block from Pred's comparison, potentially
627 // simplify TI based on this knowledge.
628 if (PredDef == TI->getParent()) {
629 // If we are here, we know that the value is none of those cases listed in
630 // PredCases. If there are any cases in ThisCases that are in PredCases, we
631 // can simplify TI.
632 if (ValuesOverlap(PredCases, ThisCases)) {
633 if (isa<BranchInst>(TI)) {
634 // Okay, one of the successors of this condbr is dead. Convert it to a
635 // uncond br.
636 assert(ThisCases.size() == 1 && "Branch can only have one case!");
637 // Insert the new branch.
638 Instruction *NI = BranchInst::Create(ThisDef, TI);
640 // Remove PHI node entries for the dead edge.
641 ThisCases[0].second->removePredecessor(TI->getParent());
643 DOUT << "Threading pred instr: " << *Pred->getTerminator()
644 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n";
646 EraseTerminatorInstAndDCECond(TI);
647 return true;
649 } else {
650 SwitchInst *SI = cast<SwitchInst>(TI);
651 // Okay, TI has cases that are statically dead, prune them away.
652 SmallPtrSet<Constant*, 16> DeadCases;
653 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
654 DeadCases.insert(PredCases[i].first);
656 DOUT << "Threading pred instr: " << *Pred->getTerminator()
657 << "Through successor TI: " << *TI;
659 for (unsigned i = SI->getNumCases()-1; i != 0; --i)
660 if (DeadCases.count(SI->getCaseValue(i))) {
661 SI->getSuccessor(i)->removePredecessor(TI->getParent());
662 SI->removeCase(i);
665 DOUT << "Leaving: " << *TI << "\n";
666 return true;
670 } else {
671 // Otherwise, TI's block must correspond to some matched value. Find out
672 // which value (or set of values) this is.
673 ConstantInt *TIV = 0;
674 BasicBlock *TIBB = TI->getParent();
675 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
676 if (PredCases[i].second == TIBB) {
677 if (TIV == 0)
678 TIV = PredCases[i].first;
679 else
680 return false; // Cannot handle multiple values coming to this block.
682 assert(TIV && "No edge from pred to succ?");
684 // Okay, we found the one constant that our value can be if we get into TI's
685 // BB. Find out which successor will unconditionally be branched to.
686 BasicBlock *TheRealDest = 0;
687 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i)
688 if (ThisCases[i].first == TIV) {
689 TheRealDest = ThisCases[i].second;
690 break;
693 // If not handled by any explicit cases, it is handled by the default case.
694 if (TheRealDest == 0) TheRealDest = ThisDef;
696 // Remove PHI node entries for dead edges.
697 BasicBlock *CheckEdge = TheRealDest;
698 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI)
699 if (*SI != CheckEdge)
700 (*SI)->removePredecessor(TIBB);
701 else
702 CheckEdge = 0;
704 // Insert the new branch.
705 Instruction *NI = BranchInst::Create(TheRealDest, TI);
707 DOUT << "Threading pred instr: " << *Pred->getTerminator()
708 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n";
710 EraseTerminatorInstAndDCECond(TI);
711 return true;
713 return false;
716 namespace {
717 /// ConstantIntOrdering - This class implements a stable ordering of constant
718 /// integers that does not depend on their address. This is important for
719 /// applications that sort ConstantInt's to ensure uniqueness.
720 struct ConstantIntOrdering {
721 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const {
722 return LHS->getValue().ult(RHS->getValue());
727 /// FoldValueComparisonIntoPredecessors - The specified terminator is a value
728 /// equality comparison instruction (either a switch or a branch on "X == c").
729 /// See if any of the predecessors of the terminator block are value comparisons
730 /// on the same value. If so, and if safe to do so, fold them together.
731 static bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI) {
732 BasicBlock *BB = TI->getParent();
733 Value *CV = isValueEqualityComparison(TI); // CondVal
734 assert(CV && "Not a comparison?");
735 bool Changed = false;
737 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
738 while (!Preds.empty()) {
739 BasicBlock *Pred = Preds.pop_back_val();
741 // See if the predecessor is a comparison with the same value.
742 TerminatorInst *PTI = Pred->getTerminator();
743 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal
745 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) {
746 // Figure out which 'cases' to copy from SI to PSI.
747 std::vector<std::pair<ConstantInt*, BasicBlock*> > BBCases;
748 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases);
750 std::vector<std::pair<ConstantInt*, BasicBlock*> > PredCases;
751 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases);
753 // Based on whether the default edge from PTI goes to BB or not, fill in
754 // PredCases and PredDefault with the new switch cases we would like to
755 // build.
756 SmallVector<BasicBlock*, 8> NewSuccessors;
758 if (PredDefault == BB) {
759 // If this is the default destination from PTI, only the edges in TI
760 // that don't occur in PTI, or that branch to BB will be activated.
761 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
762 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
763 if (PredCases[i].second != BB)
764 PTIHandled.insert(PredCases[i].first);
765 else {
766 // The default destination is BB, we don't need explicit targets.
767 std::swap(PredCases[i], PredCases.back());
768 PredCases.pop_back();
769 --i; --e;
772 // Reconstruct the new switch statement we will be building.
773 if (PredDefault != BBDefault) {
774 PredDefault->removePredecessor(Pred);
775 PredDefault = BBDefault;
776 NewSuccessors.push_back(BBDefault);
778 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
779 if (!PTIHandled.count(BBCases[i].first) &&
780 BBCases[i].second != BBDefault) {
781 PredCases.push_back(BBCases[i]);
782 NewSuccessors.push_back(BBCases[i].second);
785 } else {
786 // If this is not the default destination from PSI, only the edges
787 // in SI that occur in PSI with a destination of BB will be
788 // activated.
789 std::set<ConstantInt*, ConstantIntOrdering> PTIHandled;
790 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
791 if (PredCases[i].second == BB) {
792 PTIHandled.insert(PredCases[i].first);
793 std::swap(PredCases[i], PredCases.back());
794 PredCases.pop_back();
795 --i; --e;
798 // Okay, now we know which constants were sent to BB from the
799 // predecessor. Figure out where they will all go now.
800 for (unsigned i = 0, e = BBCases.size(); i != e; ++i)
801 if (PTIHandled.count(BBCases[i].first)) {
802 // If this is one we are capable of getting...
803 PredCases.push_back(BBCases[i]);
804 NewSuccessors.push_back(BBCases[i].second);
805 PTIHandled.erase(BBCases[i].first);// This constant is taken care of
808 // If there are any constants vectored to BB that TI doesn't handle,
809 // they must go to the default destination of TI.
810 for (std::set<ConstantInt*, ConstantIntOrdering>::iterator I =
811 PTIHandled.begin(),
812 E = PTIHandled.end(); I != E; ++I) {
813 PredCases.push_back(std::make_pair(*I, BBDefault));
814 NewSuccessors.push_back(BBDefault);
818 // Okay, at this point, we know which new successor Pred will get. Make
819 // sure we update the number of entries in the PHI nodes for these
820 // successors.
821 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i)
822 AddPredecessorToBlock(NewSuccessors[i], Pred, BB);
824 // Now that the successors are updated, create the new Switch instruction.
825 SwitchInst *NewSI = SwitchInst::Create(CV, PredDefault,
826 PredCases.size(), PTI);
827 for (unsigned i = 0, e = PredCases.size(); i != e; ++i)
828 NewSI->addCase(PredCases[i].first, PredCases[i].second);
830 EraseTerminatorInstAndDCECond(PTI);
832 // Okay, last check. If BB is still a successor of PSI, then we must
833 // have an infinite loop case. If so, add an infinitely looping block
834 // to handle the case to preserve the behavior of the code.
835 BasicBlock *InfLoopBlock = 0;
836 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i)
837 if (NewSI->getSuccessor(i) == BB) {
838 if (InfLoopBlock == 0) {
839 // Insert it at the end of the function, because it's either code,
840 // or it won't matter if it's hot. :)
841 InfLoopBlock = BasicBlock::Create("infloop", BB->getParent());
842 BranchInst::Create(InfLoopBlock, InfLoopBlock);
844 NewSI->setSuccessor(i, InfLoopBlock);
847 Changed = true;
850 return Changed;
853 // isSafeToHoistInvoke - If we would need to insert a select that uses the
854 // value of this invoke (comments in HoistThenElseCodeToIf explain why we
855 // would need to do this), we can't hoist the invoke, as there is nowhere
856 // to put the select in this case.
857 static bool isSafeToHoistInvoke(BasicBlock *BB1, BasicBlock *BB2,
858 Instruction *I1, Instruction *I2) {
859 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
860 PHINode *PN;
861 for (BasicBlock::iterator BBI = SI->begin();
862 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
863 Value *BB1V = PN->getIncomingValueForBlock(BB1);
864 Value *BB2V = PN->getIncomingValueForBlock(BB2);
865 if (BB1V != BB2V && (BB1V==I1 || BB2V==I2)) {
866 return false;
870 return true;
873 /// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and
874 /// BB2, hoist any common code in the two blocks up into the branch block. The
875 /// caller of this function guarantees that BI's block dominates BB1 and BB2.
876 static bool HoistThenElseCodeToIf(BranchInst *BI) {
877 // This does very trivial matching, with limited scanning, to find identical
878 // instructions in the two blocks. In particular, we don't want to get into
879 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As
880 // such, we currently just scan for obviously identical instructions in an
881 // identical order.
882 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination.
883 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination
885 BasicBlock::iterator BB1_Itr = BB1->begin();
886 BasicBlock::iterator BB2_Itr = BB2->begin();
888 Instruction *I1 = BB1_Itr++, *I2 = BB2_Itr++;
889 while (isa<DbgInfoIntrinsic>(I1))
890 I1 = BB1_Itr++;
891 while (isa<DbgInfoIntrinsic>(I2))
892 I2 = BB2_Itr++;
893 if (I1->getOpcode() != I2->getOpcode() || isa<PHINode>(I1) ||
894 !I1->isIdenticalTo(I2) ||
895 (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2)))
896 return false;
898 // If we get here, we can hoist at least one instruction.
899 BasicBlock *BIParent = BI->getParent();
901 do {
902 // If we are hoisting the terminator instruction, don't move one (making a
903 // broken BB), instead clone it, and remove BI.
904 if (isa<TerminatorInst>(I1))
905 goto HoistTerminator;
907 // For a normal instruction, we just move one to right before the branch,
908 // then replace all uses of the other with the first. Finally, we remove
909 // the now redundant second instruction.
910 BIParent->getInstList().splice(BI, BB1->getInstList(), I1);
911 if (!I2->use_empty())
912 I2->replaceAllUsesWith(I1);
913 BB2->getInstList().erase(I2);
915 I1 = BB1_Itr++;
916 while (isa<DbgInfoIntrinsic>(I1))
917 I1 = BB1_Itr++;
918 I2 = BB2_Itr++;
919 while (isa<DbgInfoIntrinsic>(I2))
920 I2 = BB2_Itr++;
921 } while (I1->getOpcode() == I2->getOpcode() && I1->isIdenticalTo(I2));
923 return true;
925 HoistTerminator:
926 // It may not be possible to hoist an invoke.
927 if (isa<InvokeInst>(I1) && !isSafeToHoistInvoke(BB1, BB2, I1, I2))
928 return true;
930 // Okay, it is safe to hoist the terminator.
931 Instruction *NT = I1->clone(BB1->getContext());
932 BIParent->getInstList().insert(BI, NT);
933 if (NT->getType() != Type::VoidTy) {
934 I1->replaceAllUsesWith(NT);
935 I2->replaceAllUsesWith(NT);
936 NT->takeName(I1);
939 // Hoisting one of the terminators from our successor is a great thing.
940 // Unfortunately, the successors of the if/else blocks may have PHI nodes in
941 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI
942 // nodes, so we insert select instruction to compute the final result.
943 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects;
944 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) {
945 PHINode *PN;
946 for (BasicBlock::iterator BBI = SI->begin();
947 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
948 Value *BB1V = PN->getIncomingValueForBlock(BB1);
949 Value *BB2V = PN->getIncomingValueForBlock(BB2);
950 if (BB1V != BB2V) {
951 // These values do not agree. Insert a select instruction before NT
952 // that determines the right value.
953 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)];
954 if (SI == 0)
955 SI = SelectInst::Create(BI->getCondition(), BB1V, BB2V,
956 BB1V->getName()+"."+BB2V->getName(), NT);
957 // Make the PHI node use the select for all incoming values for BB1/BB2
958 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
959 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2)
960 PN->setIncomingValue(i, SI);
965 // Update any PHI nodes in our new successors.
966 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI)
967 AddPredecessorToBlock(*SI, BIParent, BB1);
969 EraseTerminatorInstAndDCECond(BI);
970 return true;
973 /// SpeculativelyExecuteBB - Given a conditional branch that goes to BB1
974 /// and an BB2 and the only successor of BB1 is BB2, hoist simple code
975 /// (for now, restricted to a single instruction that's side effect free) from
976 /// the BB1 into the branch block to speculatively execute it.
977 static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *BB1) {
978 // Only speculatively execution a single instruction (not counting the
979 // terminator) for now.
980 Instruction *HInst = NULL;
981 Instruction *Term = BB1->getTerminator();
982 for (BasicBlock::iterator BBI = BB1->begin(), BBE = BB1->end();
983 BBI != BBE; ++BBI) {
984 Instruction *I = BBI;
985 // Skip debug info.
986 if (isa<DbgInfoIntrinsic>(I)) continue;
987 if (I == Term) break;
989 if (!HInst)
990 HInst = I;
991 else
992 return false;
994 if (!HInst)
995 return false;
997 // Be conservative for now. FP select instruction can often be expensive.
998 Value *BrCond = BI->getCondition();
999 if (isa<Instruction>(BrCond) &&
1000 cast<Instruction>(BrCond)->getOpcode() == Instruction::FCmp)
1001 return false;
1003 // If BB1 is actually on the false edge of the conditional branch, remember
1004 // to swap the select operands later.
1005 bool Invert = false;
1006 if (BB1 != BI->getSuccessor(0)) {
1007 assert(BB1 == BI->getSuccessor(1) && "No edge from 'if' block?");
1008 Invert = true;
1011 // Turn
1012 // BB:
1013 // %t1 = icmp
1014 // br i1 %t1, label %BB1, label %BB2
1015 // BB1:
1016 // %t3 = add %t2, c
1017 // br label BB2
1018 // BB2:
1019 // =>
1020 // BB:
1021 // %t1 = icmp
1022 // %t4 = add %t2, c
1023 // %t3 = select i1 %t1, %t2, %t3
1024 switch (HInst->getOpcode()) {
1025 default: return false; // Not safe / profitable to hoist.
1026 case Instruction::Add:
1027 case Instruction::Sub:
1028 // Not worth doing for vector ops.
1029 if (isa<VectorType>(HInst->getType()))
1030 return false;
1031 break;
1032 case Instruction::And:
1033 case Instruction::Or:
1034 case Instruction::Xor:
1035 case Instruction::Shl:
1036 case Instruction::LShr:
1037 case Instruction::AShr:
1038 // Don't mess with vector operations.
1039 if (isa<VectorType>(HInst->getType()))
1040 return false;
1041 break; // These are all cheap and non-trapping instructions.
1044 // If the instruction is obviously dead, don't try to predicate it.
1045 if (HInst->use_empty()) {
1046 HInst->eraseFromParent();
1047 return true;
1050 // Can we speculatively execute the instruction? And what is the value
1051 // if the condition is false? Consider the phi uses, if the incoming value
1052 // from the "if" block are all the same V, then V is the value of the
1053 // select if the condition is false.
1054 BasicBlock *BIParent = BI->getParent();
1055 SmallVector<PHINode*, 4> PHIUses;
1056 Value *FalseV = NULL;
1058 BasicBlock *BB2 = BB1->getTerminator()->getSuccessor(0);
1059 for (Value::use_iterator UI = HInst->use_begin(), E = HInst->use_end();
1060 UI != E; ++UI) {
1061 // Ignore any user that is not a PHI node in BB2. These can only occur in
1062 // unreachable blocks, because they would not be dominated by the instr.
1063 PHINode *PN = dyn_cast<PHINode>(UI);
1064 if (!PN || PN->getParent() != BB2)
1065 return false;
1066 PHIUses.push_back(PN);
1068 Value *PHIV = PN->getIncomingValueForBlock(BIParent);
1069 if (!FalseV)
1070 FalseV = PHIV;
1071 else if (FalseV != PHIV)
1072 return false; // Inconsistent value when condition is false.
1075 assert(FalseV && "Must have at least one user, and it must be a PHI");
1077 // Do not hoist the instruction if any of its operands are defined but not
1078 // used in this BB. The transformation will prevent the operand from
1079 // being sunk into the use block.
1080 for (User::op_iterator i = HInst->op_begin(), e = HInst->op_end();
1081 i != e; ++i) {
1082 Instruction *OpI = dyn_cast<Instruction>(*i);
1083 if (OpI && OpI->getParent() == BIParent &&
1084 !OpI->isUsedInBasicBlock(BIParent))
1085 return false;
1088 // If we get here, we can hoist the instruction. Try to place it
1089 // before the icmp instruction preceding the conditional branch.
1090 BasicBlock::iterator InsertPos = BI;
1091 if (InsertPos != BIParent->begin())
1092 --InsertPos;
1093 // Skip debug info between condition and branch.
1094 while (InsertPos != BIParent->begin() && isa<DbgInfoIntrinsic>(InsertPos))
1095 --InsertPos;
1096 if (InsertPos == BrCond && !isa<PHINode>(BrCond)) {
1097 SmallPtrSet<Instruction *, 4> BB1Insns;
1098 for(BasicBlock::iterator BB1I = BB1->begin(), BB1E = BB1->end();
1099 BB1I != BB1E; ++BB1I)
1100 BB1Insns.insert(BB1I);
1101 for(Value::use_iterator UI = BrCond->use_begin(), UE = BrCond->use_end();
1102 UI != UE; ++UI) {
1103 Instruction *Use = cast<Instruction>(*UI);
1104 if (BB1Insns.count(Use)) {
1105 // If BrCond uses the instruction that place it just before
1106 // branch instruction.
1107 InsertPos = BI;
1108 break;
1111 } else
1112 InsertPos = BI;
1113 BIParent->getInstList().splice(InsertPos, BB1->getInstList(), HInst);
1115 // Create a select whose true value is the speculatively executed value and
1116 // false value is the previously determined FalseV.
1117 SelectInst *SI;
1118 if (Invert)
1119 SI = SelectInst::Create(BrCond, FalseV, HInst,
1120 FalseV->getName() + "." + HInst->getName(), BI);
1121 else
1122 SI = SelectInst::Create(BrCond, HInst, FalseV,
1123 HInst->getName() + "." + FalseV->getName(), BI);
1125 // Make the PHI node use the select for all incoming values for "then" and
1126 // "if" blocks.
1127 for (unsigned i = 0, e = PHIUses.size(); i != e; ++i) {
1128 PHINode *PN = PHIUses[i];
1129 for (unsigned j = 0, ee = PN->getNumIncomingValues(); j != ee; ++j)
1130 if (PN->getIncomingBlock(j) == BB1 ||
1131 PN->getIncomingBlock(j) == BIParent)
1132 PN->setIncomingValue(j, SI);
1135 ++NumSpeculations;
1136 return true;
1139 /// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch
1140 /// across this block.
1141 static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) {
1142 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
1143 unsigned Size = 0;
1145 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1146 if (isa<DbgInfoIntrinsic>(BBI))
1147 continue;
1148 if (Size > 10) return false; // Don't clone large BB's.
1149 ++Size;
1151 // We can only support instructions that do not define values that are
1152 // live outside of the current basic block.
1153 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end();
1154 UI != E; ++UI) {
1155 Instruction *U = cast<Instruction>(*UI);
1156 if (U->getParent() != BB || isa<PHINode>(U)) return false;
1159 // Looks ok, continue checking.
1162 return true;
1165 /// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value
1166 /// that is defined in the same block as the branch and if any PHI entries are
1167 /// constants, thread edges corresponding to that entry to be branches to their
1168 /// ultimate destination.
1169 static bool FoldCondBranchOnPHI(BranchInst *BI) {
1170 BasicBlock *BB = BI->getParent();
1171 LLVMContext &Context = BB->getContext();
1172 PHINode *PN = dyn_cast<PHINode>(BI->getCondition());
1173 // NOTE: we currently cannot transform this case if the PHI node is used
1174 // outside of the block.
1175 if (!PN || PN->getParent() != BB || !PN->hasOneUse())
1176 return false;
1178 // Degenerate case of a single entry PHI.
1179 if (PN->getNumIncomingValues() == 1) {
1180 FoldSingleEntryPHINodes(PN->getParent());
1181 return true;
1184 // Now we know that this block has multiple preds and two succs.
1185 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false;
1187 // Okay, this is a simple enough basic block. See if any phi values are
1188 // constants.
1189 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1190 ConstantInt *CB;
1191 if ((CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i))) &&
1192 CB->getType() == Type::Int1Ty) {
1193 // Okay, we now know that all edges from PredBB should be revectored to
1194 // branch to RealDest.
1195 BasicBlock *PredBB = PN->getIncomingBlock(i);
1196 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue());
1198 if (RealDest == BB) continue; // Skip self loops.
1200 // The dest block might have PHI nodes, other predecessors and other
1201 // difficult cases. Instead of being smart about this, just insert a new
1202 // block that jumps to the destination block, effectively splitting
1203 // the edge we are about to create.
1204 BasicBlock *EdgeBB = BasicBlock::Create(RealDest->getName()+".critedge",
1205 RealDest->getParent(), RealDest);
1206 BranchInst::Create(RealDest, EdgeBB);
1207 PHINode *PN;
1208 for (BasicBlock::iterator BBI = RealDest->begin();
1209 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
1210 Value *V = PN->getIncomingValueForBlock(BB);
1211 PN->addIncoming(V, EdgeBB);
1214 // BB may have instructions that are being threaded over. Clone these
1215 // instructions into EdgeBB. We know that there will be no uses of the
1216 // cloned instructions outside of EdgeBB.
1217 BasicBlock::iterator InsertPt = EdgeBB->begin();
1218 std::map<Value*, Value*> TranslateMap; // Track translated values.
1219 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) {
1220 if (PHINode *PN = dyn_cast<PHINode>(BBI)) {
1221 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB);
1222 } else {
1223 // Clone the instruction.
1224 Instruction *N = BBI->clone(Context);
1225 if (BBI->hasName()) N->setName(BBI->getName()+".c");
1227 // Update operands due to translation.
1228 for (User::op_iterator i = N->op_begin(), e = N->op_end();
1229 i != e; ++i) {
1230 std::map<Value*, Value*>::iterator PI =
1231 TranslateMap.find(*i);
1232 if (PI != TranslateMap.end())
1233 *i = PI->second;
1236 // Check for trivial simplification.
1237 if (Constant *C = ConstantFoldInstruction(N, Context)) {
1238 TranslateMap[BBI] = C;
1239 delete N; // Constant folded away, don't need actual inst
1240 } else {
1241 // Insert the new instruction into its new home.
1242 EdgeBB->getInstList().insert(InsertPt, N);
1243 if (!BBI->use_empty())
1244 TranslateMap[BBI] = N;
1249 // Loop over all of the edges from PredBB to BB, changing them to branch
1250 // to EdgeBB instead.
1251 TerminatorInst *PredBBTI = PredBB->getTerminator();
1252 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i)
1253 if (PredBBTI->getSuccessor(i) == BB) {
1254 BB->removePredecessor(PredBB);
1255 PredBBTI->setSuccessor(i, EdgeBB);
1258 // Recurse, simplifying any other constants.
1259 return FoldCondBranchOnPHI(BI) | true;
1263 return false;
1266 /// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry
1267 /// PHI node, see if we can eliminate it.
1268 static bool FoldTwoEntryPHINode(PHINode *PN) {
1269 // Ok, this is a two entry PHI node. Check to see if this is a simple "if
1270 // statement", which has a very simple dominance structure. Basically, we
1271 // are trying to find the condition that is being branched on, which
1272 // subsequently causes this merge to happen. We really want control
1273 // dependence information for this check, but simplifycfg can't keep it up
1274 // to date, and this catches most of the cases we care about anyway.
1276 BasicBlock *BB = PN->getParent();
1277 BasicBlock *IfTrue, *IfFalse;
1278 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse);
1279 if (!IfCond) return false;
1281 // Okay, we found that we can merge this two-entry phi node into a select.
1282 // Doing so would require us to fold *all* two entry phi nodes in this block.
1283 // At some point this becomes non-profitable (particularly if the target
1284 // doesn't support cmov's). Only do this transformation if there are two or
1285 // fewer PHI nodes in this block.
1286 unsigned NumPhis = 0;
1287 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I)
1288 if (NumPhis > 2)
1289 return false;
1291 DEBUG(errs() << "FOUND IF CONDITION! " << *IfCond << " T: "
1292 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n");
1294 // Loop over the PHI's seeing if we can promote them all to select
1295 // instructions. While we are at it, keep track of the instructions
1296 // that need to be moved to the dominating block.
1297 std::set<Instruction*> AggressiveInsts;
1299 BasicBlock::iterator AfterPHIIt = BB->begin();
1300 while (isa<PHINode>(AfterPHIIt)) {
1301 PHINode *PN = cast<PHINode>(AfterPHIIt++);
1302 if (PN->getIncomingValue(0) == PN->getIncomingValue(1)) {
1303 if (PN->getIncomingValue(0) != PN)
1304 PN->replaceAllUsesWith(PN->getIncomingValue(0));
1305 else
1306 PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
1307 } else if (!DominatesMergePoint(PN->getIncomingValue(0), BB,
1308 &AggressiveInsts) ||
1309 !DominatesMergePoint(PN->getIncomingValue(1), BB,
1310 &AggressiveInsts)) {
1311 return false;
1315 // If we all PHI nodes are promotable, check to make sure that all
1316 // instructions in the predecessor blocks can be promoted as well. If
1317 // not, we won't be able to get rid of the control flow, so it's not
1318 // worth promoting to select instructions.
1319 BasicBlock *DomBlock = 0, *IfBlock1 = 0, *IfBlock2 = 0;
1320 PN = cast<PHINode>(BB->begin());
1321 BasicBlock *Pred = PN->getIncomingBlock(0);
1322 if (cast<BranchInst>(Pred->getTerminator())->isUnconditional()) {
1323 IfBlock1 = Pred;
1324 DomBlock = *pred_begin(Pred);
1325 for (BasicBlock::iterator I = Pred->begin();
1326 !isa<TerminatorInst>(I); ++I)
1327 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1328 // This is not an aggressive instruction that we can promote.
1329 // Because of this, we won't be able to get rid of the control
1330 // flow, so the xform is not worth it.
1331 return false;
1335 Pred = PN->getIncomingBlock(1);
1336 if (cast<BranchInst>(Pred->getTerminator())->isUnconditional()) {
1337 IfBlock2 = Pred;
1338 DomBlock = *pred_begin(Pred);
1339 for (BasicBlock::iterator I = Pred->begin();
1340 !isa<TerminatorInst>(I); ++I)
1341 if (!AggressiveInsts.count(I) && !isa<DbgInfoIntrinsic>(I)) {
1342 // This is not an aggressive instruction that we can promote.
1343 // Because of this, we won't be able to get rid of the control
1344 // flow, so the xform is not worth it.
1345 return false;
1349 // If we can still promote the PHI nodes after this gauntlet of tests,
1350 // do all of the PHI's now.
1352 // Move all 'aggressive' instructions, which are defined in the
1353 // conditional parts of the if's up to the dominating block.
1354 if (IfBlock1) {
1355 DomBlock->getInstList().splice(DomBlock->getTerminator(),
1356 IfBlock1->getInstList(),
1357 IfBlock1->begin(),
1358 IfBlock1->getTerminator());
1360 if (IfBlock2) {
1361 DomBlock->getInstList().splice(DomBlock->getTerminator(),
1362 IfBlock2->getInstList(),
1363 IfBlock2->begin(),
1364 IfBlock2->getTerminator());
1367 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) {
1368 // Change the PHI node into a select instruction.
1369 Value *TrueVal =
1370 PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse);
1371 Value *FalseVal =
1372 PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue);
1374 Value *NV = SelectInst::Create(IfCond, TrueVal, FalseVal, "", AfterPHIIt);
1375 PN->replaceAllUsesWith(NV);
1376 NV->takeName(PN);
1378 BB->getInstList().erase(PN);
1380 return true;
1383 /// isTerminatorFirstRelevantInsn - Return true if Term is very first
1384 /// instruction ignoring Phi nodes and dbg intrinsics.
1385 static bool isTerminatorFirstRelevantInsn(BasicBlock *BB, Instruction *Term) {
1386 BasicBlock::iterator BBI = Term;
1387 while (BBI != BB->begin()) {
1388 --BBI;
1389 if (!isa<DbgInfoIntrinsic>(BBI))
1390 break;
1393 if (isa<PHINode>(BBI) || &*BBI == Term || isa<DbgInfoIntrinsic>(BBI))
1394 return true;
1395 return false;
1398 /// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes
1399 /// to two returning blocks, try to merge them together into one return,
1400 /// introducing a select if the return values disagree.
1401 static bool SimplifyCondBranchToTwoReturns(BranchInst *BI) {
1402 assert(BI->isConditional() && "Must be a conditional branch");
1403 BasicBlock *TrueSucc = BI->getSuccessor(0);
1404 BasicBlock *FalseSucc = BI->getSuccessor(1);
1405 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator());
1406 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator());
1408 // Check to ensure both blocks are empty (just a return) or optionally empty
1409 // with PHI nodes. If there are other instructions, merging would cause extra
1410 // computation on one path or the other.
1411 if (!isTerminatorFirstRelevantInsn(TrueSucc, TrueRet))
1412 return false;
1413 if (!isTerminatorFirstRelevantInsn(FalseSucc, FalseRet))
1414 return false;
1416 // Okay, we found a branch that is going to two return nodes. If
1417 // there is no return value for this function, just change the
1418 // branch into a return.
1419 if (FalseRet->getNumOperands() == 0) {
1420 TrueSucc->removePredecessor(BI->getParent());
1421 FalseSucc->removePredecessor(BI->getParent());
1422 ReturnInst::Create(0, BI);
1423 EraseTerminatorInstAndDCECond(BI);
1424 return true;
1427 // Otherwise, figure out what the true and false return values are
1428 // so we can insert a new select instruction.
1429 Value *TrueValue = TrueRet->getReturnValue();
1430 Value *FalseValue = FalseRet->getReturnValue();
1432 // Unwrap any PHI nodes in the return blocks.
1433 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue))
1434 if (TVPN->getParent() == TrueSucc)
1435 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent());
1436 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue))
1437 if (FVPN->getParent() == FalseSucc)
1438 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent());
1440 // In order for this transformation to be safe, we must be able to
1441 // unconditionally execute both operands to the return. This is
1442 // normally the case, but we could have a potentially-trapping
1443 // constant expression that prevents this transformation from being
1444 // safe.
1445 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue))
1446 if (TCV->canTrap())
1447 return false;
1448 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue))
1449 if (FCV->canTrap())
1450 return false;
1452 // Okay, we collected all the mapped values and checked them for sanity, and
1453 // defined to really do this transformation. First, update the CFG.
1454 TrueSucc->removePredecessor(BI->getParent());
1455 FalseSucc->removePredecessor(BI->getParent());
1457 // Insert select instructions where needed.
1458 Value *BrCond = BI->getCondition();
1459 if (TrueValue) {
1460 // Insert a select if the results differ.
1461 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) {
1462 } else if (isa<UndefValue>(TrueValue)) {
1463 TrueValue = FalseValue;
1464 } else {
1465 TrueValue = SelectInst::Create(BrCond, TrueValue,
1466 FalseValue, "retval", BI);
1470 Value *RI = !TrueValue ?
1471 ReturnInst::Create(BI) :
1472 ReturnInst::Create(TrueValue, BI);
1474 DOUT << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:"
1475 << "\n " << *BI << "NewRet = " << *RI
1476 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc;
1478 EraseTerminatorInstAndDCECond(BI);
1480 return true;
1483 /// FoldBranchToCommonDest - If this basic block is ONLY a setcc and a branch,
1484 /// and if a predecessor branches to us and one of our successors, fold the
1485 /// setcc into the predecessor and use logical operations to pick the right
1486 /// destination.
1487 bool llvm::FoldBranchToCommonDest(BranchInst *BI) {
1488 BasicBlock *BB = BI->getParent();
1489 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition());
1490 if (Cond == 0) return false;
1493 // Only allow this if the condition is a simple instruction that can be
1494 // executed unconditionally. It must be in the same block as the branch, and
1495 // must be at the front of the block.
1496 BasicBlock::iterator FrontIt = BB->front();
1497 // Ignore dbg intrinsics.
1498 while(isa<DbgInfoIntrinsic>(FrontIt))
1499 ++FrontIt;
1500 if ((!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) ||
1501 Cond->getParent() != BB || &*FrontIt != Cond || !Cond->hasOneUse()) {
1502 return false;
1505 // Make sure the instruction after the condition is the cond branch.
1506 BasicBlock::iterator CondIt = Cond; ++CondIt;
1507 // Ingore dbg intrinsics.
1508 while(isa<DbgInfoIntrinsic>(CondIt))
1509 ++CondIt;
1510 if (&*CondIt != BI) {
1511 assert (!isa<DbgInfoIntrinsic>(CondIt) && "Hey do not forget debug info!");
1512 return false;
1515 // Cond is known to be a compare or binary operator. Check to make sure that
1516 // neither operand is a potentially-trapping constant expression.
1517 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(0)))
1518 if (CE->canTrap())
1519 return false;
1520 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Cond->getOperand(1)))
1521 if (CE->canTrap())
1522 return false;
1525 // Finally, don't infinitely unroll conditional loops.
1526 BasicBlock *TrueDest = BI->getSuccessor(0);
1527 BasicBlock *FalseDest = BI->getSuccessor(1);
1528 if (TrueDest == BB || FalseDest == BB)
1529 return false;
1531 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
1532 BasicBlock *PredBlock = *PI;
1533 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator());
1535 // Check that we have two conditional branches. If there is a PHI node in
1536 // the common successor, verify that the same value flows in from both
1537 // blocks.
1538 if (PBI == 0 || PBI->isUnconditional() ||
1539 !SafeToMergeTerminators(BI, PBI))
1540 continue;
1542 Instruction::BinaryOps Opc;
1543 bool InvertPredCond = false;
1545 if (PBI->getSuccessor(0) == TrueDest)
1546 Opc = Instruction::Or;
1547 else if (PBI->getSuccessor(1) == FalseDest)
1548 Opc = Instruction::And;
1549 else if (PBI->getSuccessor(0) == FalseDest)
1550 Opc = Instruction::And, InvertPredCond = true;
1551 else if (PBI->getSuccessor(1) == TrueDest)
1552 Opc = Instruction::Or, InvertPredCond = true;
1553 else
1554 continue;
1556 DOUT << "FOLDING BRANCH TO COMMON DEST:\n" << *PBI << *BB;
1558 // If we need to invert the condition in the pred block to match, do so now.
1559 if (InvertPredCond) {
1560 Value *NewCond =
1561 BinaryOperator::CreateNot(BI->getParent()->getContext(),
1562 PBI->getCondition(),
1563 PBI->getCondition()->getName()+".not", PBI);
1564 PBI->setCondition(NewCond);
1565 BasicBlock *OldTrue = PBI->getSuccessor(0);
1566 BasicBlock *OldFalse = PBI->getSuccessor(1);
1567 PBI->setSuccessor(0, OldFalse);
1568 PBI->setSuccessor(1, OldTrue);
1571 // Clone Cond into the predecessor basic block, and or/and the
1572 // two conditions together.
1573 Instruction *New = Cond->clone(BB->getContext());
1574 PredBlock->getInstList().insert(PBI, New);
1575 New->takeName(Cond);
1576 Cond->setName(New->getName()+".old");
1578 Value *NewCond = BinaryOperator::Create(Opc, PBI->getCondition(),
1579 New, "or.cond", PBI);
1580 PBI->setCondition(NewCond);
1581 if (PBI->getSuccessor(0) == BB) {
1582 AddPredecessorToBlock(TrueDest, PredBlock, BB);
1583 PBI->setSuccessor(0, TrueDest);
1585 if (PBI->getSuccessor(1) == BB) {
1586 AddPredecessorToBlock(FalseDest, PredBlock, BB);
1587 PBI->setSuccessor(1, FalseDest);
1589 return true;
1591 return false;
1594 /// SimplifyCondBranchToCondBranch - If we have a conditional branch as a
1595 /// predecessor of another block, this function tries to simplify it. We know
1596 /// that PBI and BI are both conditional branches, and BI is in one of the
1597 /// successor blocks of PBI - PBI branches to BI.
1598 static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) {
1599 assert(PBI->isConditional() && BI->isConditional());
1600 BasicBlock *BB = BI->getParent();
1601 LLVMContext &Context = BB->getContext();
1603 // If this block ends with a branch instruction, and if there is a
1604 // predecessor that ends on a branch of the same condition, make
1605 // this conditional branch redundant.
1606 if (PBI->getCondition() == BI->getCondition() &&
1607 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
1608 // Okay, the outcome of this conditional branch is statically
1609 // knowable. If this block had a single pred, handle specially.
1610 if (BB->getSinglePredecessor()) {
1611 // Turn this into a branch on constant.
1612 bool CondIsTrue = PBI->getSuccessor(0) == BB;
1613 BI->setCondition(ConstantInt::get(Type::Int1Ty, CondIsTrue));
1614 return true; // Nuke the branch on constant.
1617 // Otherwise, if there are multiple predecessors, insert a PHI that merges
1618 // in the constant and simplify the block result. Subsequent passes of
1619 // simplifycfg will thread the block.
1620 if (BlockIsSimpleEnoughToThreadThrough(BB)) {
1621 PHINode *NewPN = PHINode::Create(Type::Int1Ty,
1622 BI->getCondition()->getName() + ".pr",
1623 BB->begin());
1624 // Okay, we're going to insert the PHI node. Since PBI is not the only
1625 // predecessor, compute the PHI'd conditional value for all of the preds.
1626 // Any predecessor where the condition is not computable we keep symbolic.
1627 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
1628 if ((PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) &&
1629 PBI != BI && PBI->isConditional() &&
1630 PBI->getCondition() == BI->getCondition() &&
1631 PBI->getSuccessor(0) != PBI->getSuccessor(1)) {
1632 bool CondIsTrue = PBI->getSuccessor(0) == BB;
1633 NewPN->addIncoming(ConstantInt::get(Type::Int1Ty,
1634 CondIsTrue), *PI);
1635 } else {
1636 NewPN->addIncoming(BI->getCondition(), *PI);
1639 BI->setCondition(NewPN);
1640 return true;
1644 // If this is a conditional branch in an empty block, and if any
1645 // predecessors is a conditional branch to one of our destinations,
1646 // fold the conditions into logical ops and one cond br.
1647 BasicBlock::iterator BBI = BB->begin();
1648 // Ignore dbg intrinsics.
1649 while (isa<DbgInfoIntrinsic>(BBI))
1650 ++BBI;
1651 if (&*BBI != BI)
1652 return false;
1655 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(BI->getCondition()))
1656 if (CE->canTrap())
1657 return false;
1659 int PBIOp, BIOp;
1660 if (PBI->getSuccessor(0) == BI->getSuccessor(0))
1661 PBIOp = BIOp = 0;
1662 else if (PBI->getSuccessor(0) == BI->getSuccessor(1))
1663 PBIOp = 0, BIOp = 1;
1664 else if (PBI->getSuccessor(1) == BI->getSuccessor(0))
1665 PBIOp = 1, BIOp = 0;
1666 else if (PBI->getSuccessor(1) == BI->getSuccessor(1))
1667 PBIOp = BIOp = 1;
1668 else
1669 return false;
1671 // Check to make sure that the other destination of this branch
1672 // isn't BB itself. If so, this is an infinite loop that will
1673 // keep getting unwound.
1674 if (PBI->getSuccessor(PBIOp) == BB)
1675 return false;
1677 // Do not perform this transformation if it would require
1678 // insertion of a large number of select instructions. For targets
1679 // without predication/cmovs, this is a big pessimization.
1680 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp);
1682 unsigned NumPhis = 0;
1683 for (BasicBlock::iterator II = CommonDest->begin();
1684 isa<PHINode>(II); ++II, ++NumPhis)
1685 if (NumPhis > 2) // Disable this xform.
1686 return false;
1688 // Finally, if everything is ok, fold the branches to logical ops.
1689 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1);
1691 DOUT << "FOLDING BRs:" << *PBI->getParent()
1692 << "AND: " << *BI->getParent();
1695 // If OtherDest *is* BB, then BB is a basic block with a single conditional
1696 // branch in it, where one edge (OtherDest) goes back to itself but the other
1697 // exits. We don't *know* that the program avoids the infinite loop
1698 // (even though that seems likely). If we do this xform naively, we'll end up
1699 // recursively unpeeling the loop. Since we know that (after the xform is
1700 // done) that the block *is* infinite if reached, we just make it an obviously
1701 // infinite loop with no cond branch.
1702 if (OtherDest == BB) {
1703 // Insert it at the end of the function, because it's either code,
1704 // or it won't matter if it's hot. :)
1705 BasicBlock *InfLoopBlock = BasicBlock::Create("infloop", BB->getParent());
1706 BranchInst::Create(InfLoopBlock, InfLoopBlock);
1707 OtherDest = InfLoopBlock;
1710 DOUT << *PBI->getParent()->getParent();
1712 // BI may have other predecessors. Because of this, we leave
1713 // it alone, but modify PBI.
1715 // Make sure we get to CommonDest on True&True directions.
1716 Value *PBICond = PBI->getCondition();
1717 if (PBIOp)
1718 PBICond = BinaryOperator::CreateNot(Context, PBICond,
1719 PBICond->getName()+".not",
1720 PBI);
1721 Value *BICond = BI->getCondition();
1722 if (BIOp)
1723 BICond = BinaryOperator::CreateNot(Context, BICond,
1724 BICond->getName()+".not",
1725 PBI);
1726 // Merge the conditions.
1727 Value *Cond = BinaryOperator::CreateOr(PBICond, BICond, "brmerge", PBI);
1729 // Modify PBI to branch on the new condition to the new dests.
1730 PBI->setCondition(Cond);
1731 PBI->setSuccessor(0, CommonDest);
1732 PBI->setSuccessor(1, OtherDest);
1734 // OtherDest may have phi nodes. If so, add an entry from PBI's
1735 // block that are identical to the entries for BI's block.
1736 PHINode *PN;
1737 for (BasicBlock::iterator II = OtherDest->begin();
1738 (PN = dyn_cast<PHINode>(II)); ++II) {
1739 Value *V = PN->getIncomingValueForBlock(BB);
1740 PN->addIncoming(V, PBI->getParent());
1743 // We know that the CommonDest already had an edge from PBI to
1744 // it. If it has PHIs though, the PHIs may have different
1745 // entries for BB and PBI's BB. If so, insert a select to make
1746 // them agree.
1747 for (BasicBlock::iterator II = CommonDest->begin();
1748 (PN = dyn_cast<PHINode>(II)); ++II) {
1749 Value *BIV = PN->getIncomingValueForBlock(BB);
1750 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent());
1751 Value *PBIV = PN->getIncomingValue(PBBIdx);
1752 if (BIV != PBIV) {
1753 // Insert a select in PBI to pick the right value.
1754 Value *NV = SelectInst::Create(PBICond, PBIV, BIV,
1755 PBIV->getName()+".mux", PBI);
1756 PN->setIncomingValue(PBBIdx, NV);
1760 DOUT << "INTO: " << *PBI->getParent();
1762 DOUT << *PBI->getParent()->getParent();
1764 // This basic block is probably dead. We know it has at least
1765 // one fewer predecessor.
1766 return true;
1770 /// SimplifyCFG - This function is used to do simplification of a CFG. For
1771 /// example, it adjusts branches to branches to eliminate the extra hop, it
1772 /// eliminates unreachable basic blocks, and does other "peephole" optimization
1773 /// of the CFG. It returns true if a modification was made.
1775 /// WARNING: The entry node of a function may not be simplified.
1777 bool llvm::SimplifyCFG(BasicBlock *BB) {
1778 bool Changed = false;
1779 Function *M = BB->getParent();
1781 assert(BB && BB->getParent() && "Block not embedded in function!");
1782 assert(BB->getTerminator() && "Degenerate basic block encountered!");
1783 assert(&BB->getParent()->getEntryBlock() != BB &&
1784 "Can't Simplify entry block!");
1786 // Remove basic blocks that have no predecessors... or that just have themself
1787 // as a predecessor. These are unreachable.
1788 if (pred_begin(BB) == pred_end(BB) || BB->getSinglePredecessor() == BB) {
1789 DOUT << "Removing BB: \n" << *BB;
1790 DeleteDeadBlock(BB);
1791 return true;
1794 // Check to see if we can constant propagate this terminator instruction
1795 // away...
1796 Changed |= ConstantFoldTerminator(BB);
1798 // If there is a trivial two-entry PHI node in this basic block, and we can
1799 // eliminate it, do so now.
1800 if (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
1801 if (PN->getNumIncomingValues() == 2)
1802 Changed |= FoldTwoEntryPHINode(PN);
1804 // If this is a returning block with only PHI nodes in it, fold the return
1805 // instruction into any unconditional branch predecessors.
1807 // If any predecessor is a conditional branch that just selects among
1808 // different return values, fold the replace the branch/return with a select
1809 // and return.
1810 if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) {
1811 if (isTerminatorFirstRelevantInsn(BB, BB->getTerminator())) {
1812 // Find predecessors that end with branches.
1813 SmallVector<BasicBlock*, 8> UncondBranchPreds;
1814 SmallVector<BranchInst*, 8> CondBranchPreds;
1815 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
1816 TerminatorInst *PTI = (*PI)->getTerminator();
1817 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) {
1818 if (BI->isUnconditional())
1819 UncondBranchPreds.push_back(*PI);
1820 else
1821 CondBranchPreds.push_back(BI);
1825 // If we found some, do the transformation!
1826 if (!UncondBranchPreds.empty()) {
1827 while (!UncondBranchPreds.empty()) {
1828 BasicBlock *Pred = UncondBranchPreds.pop_back_val();
1829 DOUT << "FOLDING: " << *BB
1830 << "INTO UNCOND BRANCH PRED: " << *Pred;
1831 Instruction *UncondBranch = Pred->getTerminator();
1832 // Clone the return and add it to the end of the predecessor.
1833 Instruction *NewRet = RI->clone(BB->getContext());
1834 Pred->getInstList().push_back(NewRet);
1836 BasicBlock::iterator BBI = RI;
1837 if (BBI != BB->begin()) {
1838 // Move region end info into the predecessor.
1839 if (DbgRegionEndInst *DREI = dyn_cast<DbgRegionEndInst>(--BBI))
1840 DREI->moveBefore(NewRet);
1843 // If the return instruction returns a value, and if the value was a
1844 // PHI node in "BB", propagate the right value into the return.
1845 for (User::op_iterator i = NewRet->op_begin(), e = NewRet->op_end();
1846 i != e; ++i)
1847 if (PHINode *PN = dyn_cast<PHINode>(*i))
1848 if (PN->getParent() == BB)
1849 *i = PN->getIncomingValueForBlock(Pred);
1851 // Update any PHI nodes in the returning block to realize that we no
1852 // longer branch to them.
1853 BB->removePredecessor(Pred);
1854 Pred->getInstList().erase(UncondBranch);
1857 // If we eliminated all predecessors of the block, delete the block now.
1858 if (pred_begin(BB) == pred_end(BB))
1859 // We know there are no successors, so just nuke the block.
1860 M->getBasicBlockList().erase(BB);
1862 return true;
1865 // Check out all of the conditional branches going to this return
1866 // instruction. If any of them just select between returns, change the
1867 // branch itself into a select/return pair.
1868 while (!CondBranchPreds.empty()) {
1869 BranchInst *BI = CondBranchPreds.pop_back_val();
1871 // Check to see if the non-BB successor is also a return block.
1872 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) &&
1873 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) &&
1874 SimplifyCondBranchToTwoReturns(BI))
1875 return true;
1878 } else if (isa<UnwindInst>(BB->begin())) {
1879 // Check to see if the first instruction in this block is just an unwind.
1880 // If so, replace any invoke instructions which use this as an exception
1881 // destination with call instructions, and any unconditional branch
1882 // predecessor with an unwind.
1884 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
1885 while (!Preds.empty()) {
1886 BasicBlock *Pred = Preds.back();
1887 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator())) {
1888 if (BI->isUnconditional()) {
1889 Pred->getInstList().pop_back(); // nuke uncond branch
1890 new UnwindInst(Pred); // Use unwind.
1891 Changed = true;
1893 } else if (InvokeInst *II = dyn_cast<InvokeInst>(Pred->getTerminator()))
1894 if (II->getUnwindDest() == BB) {
1895 // Insert a new branch instruction before the invoke, because this
1896 // is now a fall through...
1897 BranchInst *BI = BranchInst::Create(II->getNormalDest(), II);
1898 Pred->getInstList().remove(II); // Take out of symbol table
1900 // Insert the call now...
1901 SmallVector<Value*,8> Args(II->op_begin()+3, II->op_end());
1902 CallInst *CI = CallInst::Create(II->getCalledValue(),
1903 Args.begin(), Args.end(),
1904 II->getName(), BI);
1905 CI->setCallingConv(II->getCallingConv());
1906 CI->setAttributes(II->getAttributes());
1907 // If the invoke produced a value, the Call now does instead
1908 II->replaceAllUsesWith(CI);
1909 delete II;
1910 Changed = true;
1913 Preds.pop_back();
1916 // If this block is now dead, remove it.
1917 if (pred_begin(BB) == pred_end(BB)) {
1918 // We know there are no successors, so just nuke the block.
1919 M->getBasicBlockList().erase(BB);
1920 return true;
1923 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1924 if (isValueEqualityComparison(SI)) {
1925 // If we only have one predecessor, and if it is a branch on this value,
1926 // see if that predecessor totally determines the outcome of this switch.
1927 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
1928 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred))
1929 return SimplifyCFG(BB) || 1;
1931 // If the block only contains the switch, see if we can fold the block
1932 // away into any preds.
1933 BasicBlock::iterator BBI = BB->begin();
1934 // Ignore dbg intrinsics.
1935 while (isa<DbgInfoIntrinsic>(BBI))
1936 ++BBI;
1937 if (SI == &*BBI)
1938 if (FoldValueComparisonIntoPredecessors(SI))
1939 return SimplifyCFG(BB) || 1;
1941 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
1942 if (BI->isUnconditional()) {
1943 BasicBlock::iterator BBI = BB->getFirstNonPHI();
1945 BasicBlock *Succ = BI->getSuccessor(0);
1946 // Ignore dbg intrinsics.
1947 while (isa<DbgInfoIntrinsic>(BBI))
1948 ++BBI;
1949 if (BBI->isTerminator() && // Terminator is the only non-phi instruction!
1950 Succ != BB) // Don't hurt infinite loops!
1951 if (TryToSimplifyUncondBranchFromEmptyBlock(BB, Succ))
1952 return true;
1954 } else { // Conditional branch
1955 if (isValueEqualityComparison(BI)) {
1956 // If we only have one predecessor, and if it is a branch on this value,
1957 // see if that predecessor totally determines the outcome of this
1958 // switch.
1959 if (BasicBlock *OnlyPred = BB->getSinglePredecessor())
1960 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred))
1961 return SimplifyCFG(BB) || 1;
1963 // This block must be empty, except for the setcond inst, if it exists.
1964 // Ignore dbg intrinsics.
1965 BasicBlock::iterator I = BB->begin();
1966 // Ignore dbg intrinsics.
1967 while (isa<DbgInfoIntrinsic>(I))
1968 ++I;
1969 if (&*I == BI) {
1970 if (FoldValueComparisonIntoPredecessors(BI))
1971 return SimplifyCFG(BB) | true;
1972 } else if (&*I == cast<Instruction>(BI->getCondition())){
1973 ++I;
1974 // Ignore dbg intrinsics.
1975 while (isa<DbgInfoIntrinsic>(I))
1976 ++I;
1977 if(&*I == BI) {
1978 if (FoldValueComparisonIntoPredecessors(BI))
1979 return SimplifyCFG(BB) | true;
1984 // If this is a branch on a phi node in the current block, thread control
1985 // through this block if any PHI node entries are constants.
1986 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition()))
1987 if (PN->getParent() == BI->getParent())
1988 if (FoldCondBranchOnPHI(BI))
1989 return SimplifyCFG(BB) | true;
1991 // If this basic block is ONLY a setcc and a branch, and if a predecessor
1992 // branches to us and one of our successors, fold the setcc into the
1993 // predecessor and use logical operations to pick the right destination.
1994 if (FoldBranchToCommonDest(BI))
1995 return SimplifyCFG(BB) | 1;
1998 // Scan predecessor blocks for conditional branches.
1999 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
2000 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
2001 if (PBI != BI && PBI->isConditional())
2002 if (SimplifyCondBranchToCondBranch(PBI, BI))
2003 return SimplifyCFG(BB) | true;
2005 } else if (isa<UnreachableInst>(BB->getTerminator())) {
2006 // If there are any instructions immediately before the unreachable that can
2007 // be removed, do so.
2008 Instruction *Unreachable = BB->getTerminator();
2009 while (Unreachable != BB->begin()) {
2010 BasicBlock::iterator BBI = Unreachable;
2011 --BBI;
2012 // Do not delete instructions that can have side effects, like calls
2013 // (which may never return) and volatile loads and stores.
2014 if (isa<CallInst>(BBI) && !isa<DbgInfoIntrinsic>(BBI)) break;
2016 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
2017 if (SI->isVolatile())
2018 break;
2020 if (LoadInst *LI = dyn_cast<LoadInst>(BBI))
2021 if (LI->isVolatile())
2022 break;
2024 // Delete this instruction
2025 BB->getInstList().erase(BBI);
2026 Changed = true;
2029 // If the unreachable instruction is the first in the block, take a gander
2030 // at all of the predecessors of this instruction, and simplify them.
2031 if (&BB->front() == Unreachable) {
2032 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB));
2033 for (unsigned i = 0, e = Preds.size(); i != e; ++i) {
2034 TerminatorInst *TI = Preds[i]->getTerminator();
2036 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
2037 if (BI->isUnconditional()) {
2038 if (BI->getSuccessor(0) == BB) {
2039 new UnreachableInst(TI);
2040 TI->eraseFromParent();
2041 Changed = true;
2043 } else {
2044 if (BI->getSuccessor(0) == BB) {
2045 BranchInst::Create(BI->getSuccessor(1), BI);
2046 EraseTerminatorInstAndDCECond(BI);
2047 } else if (BI->getSuccessor(1) == BB) {
2048 BranchInst::Create(BI->getSuccessor(0), BI);
2049 EraseTerminatorInstAndDCECond(BI);
2050 Changed = true;
2053 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
2054 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i)
2055 if (SI->getSuccessor(i) == BB) {
2056 BB->removePredecessor(SI->getParent());
2057 SI->removeCase(i);
2058 --i; --e;
2059 Changed = true;
2061 // If the default value is unreachable, figure out the most popular
2062 // destination and make it the default.
2063 if (SI->getSuccessor(0) == BB) {
2064 std::map<BasicBlock*, unsigned> Popularity;
2065 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i)
2066 Popularity[SI->getSuccessor(i)]++;
2068 // Find the most popular block.
2069 unsigned MaxPop = 0;
2070 BasicBlock *MaxBlock = 0;
2071 for (std::map<BasicBlock*, unsigned>::iterator
2072 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) {
2073 if (I->second > MaxPop) {
2074 MaxPop = I->second;
2075 MaxBlock = I->first;
2078 if (MaxBlock) {
2079 // Make this the new default, allowing us to delete any explicit
2080 // edges to it.
2081 SI->setSuccessor(0, MaxBlock);
2082 Changed = true;
2084 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from
2085 // it.
2086 if (isa<PHINode>(MaxBlock->begin()))
2087 for (unsigned i = 0; i != MaxPop-1; ++i)
2088 MaxBlock->removePredecessor(SI->getParent());
2090 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i)
2091 if (SI->getSuccessor(i) == MaxBlock) {
2092 SI->removeCase(i);
2093 --i; --e;
2097 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) {
2098 if (II->getUnwindDest() == BB) {
2099 // Convert the invoke to a call instruction. This would be a good
2100 // place to note that the call does not throw though.
2101 BranchInst *BI = BranchInst::Create(II->getNormalDest(), II);
2102 II->removeFromParent(); // Take out of symbol table
2104 // Insert the call now...
2105 SmallVector<Value*, 8> Args(II->op_begin()+3, II->op_end());
2106 CallInst *CI = CallInst::Create(II->getCalledValue(),
2107 Args.begin(), Args.end(),
2108 II->getName(), BI);
2109 CI->setCallingConv(II->getCallingConv());
2110 CI->setAttributes(II->getAttributes());
2111 // If the invoke produced a value, the Call does now instead.
2112 II->replaceAllUsesWith(CI);
2113 delete II;
2114 Changed = true;
2119 // If this block is now dead, remove it.
2120 if (pred_begin(BB) == pred_end(BB)) {
2121 // We know there are no successors, so just nuke the block.
2122 M->getBasicBlockList().erase(BB);
2123 return true;
2128 // Merge basic blocks into their predecessor if there is only one distinct
2129 // pred, and if there is only one distinct successor of the predecessor, and
2130 // if there are no PHI nodes.
2132 if (MergeBlockIntoPredecessor(BB))
2133 return true;
2135 // Otherwise, if this block only has a single predecessor, and if that block
2136 // is a conditional branch, see if we can hoist any code from this block up
2137 // into our predecessor.
2138 pred_iterator PI(pred_begin(BB)), PE(pred_end(BB));
2139 BasicBlock *OnlyPred = *PI++;
2140 for (; PI != PE; ++PI) // Search all predecessors, see if they are all same
2141 if (*PI != OnlyPred) {
2142 OnlyPred = 0; // There are multiple different predecessors...
2143 break;
2146 if (OnlyPred)
2147 if (BranchInst *BI = dyn_cast<BranchInst>(OnlyPred->getTerminator()))
2148 if (BI->isConditional()) {
2149 // Get the other block.
2150 BasicBlock *OtherBB = BI->getSuccessor(BI->getSuccessor(0) == BB);
2151 PI = pred_begin(OtherBB);
2152 ++PI;
2154 if (PI == pred_end(OtherBB)) {
2155 // We have a conditional branch to two blocks that are only reachable
2156 // from the condbr. We know that the condbr dominates the two blocks,
2157 // so see if there is any identical code in the "then" and "else"
2158 // blocks. If so, we can hoist it up to the branching block.
2159 Changed |= HoistThenElseCodeToIf(BI);
2160 } else {
2161 BasicBlock* OnlySucc = NULL;
2162 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB);
2163 SI != SE; ++SI) {
2164 if (!OnlySucc)
2165 OnlySucc = *SI;
2166 else if (*SI != OnlySucc) {
2167 OnlySucc = 0; // There are multiple distinct successors!
2168 break;
2172 if (OnlySucc == OtherBB) {
2173 // If BB's only successor is the other successor of the predecessor,
2174 // i.e. a triangle, see if we can hoist any code from this block up
2175 // to the "if" block.
2176 Changed |= SpeculativelyExecuteBB(BI, BB);
2181 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
2182 if (BranchInst *BI = dyn_cast<BranchInst>((*PI)->getTerminator()))
2183 // Change br (X == 0 | X == 1), T, F into a switch instruction.
2184 if (BI->isConditional() && isa<Instruction>(BI->getCondition())) {
2185 Instruction *Cond = cast<Instruction>(BI->getCondition());
2186 // If this is a bunch of seteq's or'd together, or if it's a bunch of
2187 // 'setne's and'ed together, collect them.
2188 Value *CompVal = 0;
2189 std::vector<ConstantInt*> Values;
2190 bool TrueWhenEqual = GatherValueComparisons(Cond, CompVal, Values);
2191 if (CompVal && CompVal->getType()->isInteger()) {
2192 // There might be duplicate constants in the list, which the switch
2193 // instruction can't handle, remove them now.
2194 std::sort(Values.begin(), Values.end(), ConstantIntOrdering());
2195 Values.erase(std::unique(Values.begin(), Values.end()), Values.end());
2197 // Figure out which block is which destination.
2198 BasicBlock *DefaultBB = BI->getSuccessor(1);
2199 BasicBlock *EdgeBB = BI->getSuccessor(0);
2200 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB);
2202 // Create the new switch instruction now.
2203 SwitchInst *New = SwitchInst::Create(CompVal, DefaultBB,
2204 Values.size(), BI);
2206 // Add all of the 'cases' to the switch instruction.
2207 for (unsigned i = 0, e = Values.size(); i != e; ++i)
2208 New->addCase(Values[i], EdgeBB);
2210 // We added edges from PI to the EdgeBB. As such, if there were any
2211 // PHI nodes in EdgeBB, they need entries to be added corresponding to
2212 // the number of edges added.
2213 for (BasicBlock::iterator BBI = EdgeBB->begin();
2214 isa<PHINode>(BBI); ++BBI) {
2215 PHINode *PN = cast<PHINode>(BBI);
2216 Value *InVal = PN->getIncomingValueForBlock(*PI);
2217 for (unsigned i = 0, e = Values.size()-1; i != e; ++i)
2218 PN->addIncoming(InVal, *PI);
2221 // Erase the old branch instruction.
2222 EraseTerminatorInstAndDCECond(BI);
2223 return true;
2227 return Changed;