1 //===-- LoopUnswitch.cpp - Hoist loop-invariant conditionals in loop ------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass transforms loops that contain branches on loop-invariant conditions
11 // to have multiple loops. For example, it turns the left into the right code:
20 // This can increase the size of the code exponentially (doubling it every time
21 // a loop is unswitched) so we only unswitch if the resultant code will be
22 // smaller than a threshold.
24 // This pass expects LICM to be run before it to hoist invariant conditions out
25 // of the loop, to make the unswitching opportunity obvious.
27 //===----------------------------------------------------------------------===//
29 #define DEBUG_TYPE "loop-unswitch"
30 #include "llvm/Transforms/Scalar.h"
31 #include "llvm/Constants.h"
32 #include "llvm/DerivedTypes.h"
33 #include "llvm/Function.h"
34 #include "llvm/Instructions.h"
35 #include "llvm/Analysis/ConstantFolding.h"
36 #include "llvm/Analysis/LoopInfo.h"
37 #include "llvm/Analysis/LoopPass.h"
38 #include "llvm/Analysis/Dominators.h"
39 #include "llvm/Transforms/Utils/Cloning.h"
40 #include "llvm/Transforms/Utils/Local.h"
41 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
42 #include "llvm/ADT/Statistic.h"
43 #include "llvm/ADT/SmallPtrSet.h"
44 #include "llvm/ADT/PostOrderIterator.h"
45 #include "llvm/Support/CommandLine.h"
46 #include "llvm/Support/Compiler.h"
47 #include "llvm/Support/Debug.h"
52 STATISTIC(NumBranches, "Number of branches unswitched");
53 STATISTIC(NumSwitches, "Number of switches unswitched");
54 STATISTIC(NumSelects , "Number of selects unswitched");
55 STATISTIC(NumTrivial , "Number of unswitches that are trivial");
56 STATISTIC(NumSimplify, "Number of simplifications of unswitched code");
60 Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"),
61 cl::init(10), cl::Hidden);
63 class VISIBILITY_HIDDEN LoopUnswitch : public LoopPass {
64 LoopInfo *LI; // Loop information
67 // LoopProcessWorklist - Used to check if second loop needs processing
68 // after RewriteLoopBodyWithConditionConstant rewrites first loop.
69 std::vector<Loop*> LoopProcessWorklist;
70 SmallPtrSet<Value *,8> UnswitchedVals;
74 static char ID; // Pass ID, replacement for typeid
75 LoopUnswitch(bool Os = false) :
76 LoopPass((intptr_t)&ID), OptimizeForSize(Os) {}
78 bool runOnLoop(Loop *L, LPPassManager &LPM);
80 /// This transformation requires natural loop information & requires that
81 /// loop preheaders be inserted into the CFG...
83 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
84 AU.addRequiredID(LoopSimplifyID);
85 AU.addPreservedID(LoopSimplifyID);
86 AU.addRequired<LoopInfo>();
87 AU.addPreserved<LoopInfo>();
88 AU.addRequiredID(LCSSAID);
92 /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist,
94 void RemoveLoopFromWorklist(Loop *L) {
95 std::vector<Loop*>::iterator I = std::find(LoopProcessWorklist.begin(),
96 LoopProcessWorklist.end(), L);
97 if (I != LoopProcessWorklist.end())
98 LoopProcessWorklist.erase(I);
101 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val,Loop *L);
102 unsigned getLoopUnswitchCost(Loop *L, Value *LIC);
103 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val,
104 BasicBlock *ExitBlock);
105 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L);
107 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
108 Constant *Val, bool isEqual);
110 void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
111 BasicBlock *TrueDest,
112 BasicBlock *FalseDest,
113 Instruction *InsertPt);
115 void SimplifyCode(std::vector<Instruction*> &Worklist);
116 void RemoveBlockIfDead(BasicBlock *BB,
117 std::vector<Instruction*> &Worklist);
118 void RemoveLoopFromHierarchy(Loop *L);
120 char LoopUnswitch::ID = 0;
121 RegisterPass<LoopUnswitch> X("loop-unswitch", "Unswitch loops");
124 LoopPass *llvm::createLoopUnswitchPass(bool Os) {
125 return new LoopUnswitch(Os);
128 /// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is
129 /// invariant in the loop, or has an invariant piece, return the invariant.
130 /// Otherwise, return null.
131 static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) {
132 // Constants should be folded, not unswitched on!
133 if (isa<Constant>(Cond)) return false;
135 // TODO: Handle: br (VARIANT|INVARIANT).
136 // TODO: Hoist simple expressions out of loops.
137 if (L->isLoopInvariant(Cond)) return Cond;
139 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond))
140 if (BO->getOpcode() == Instruction::And ||
141 BO->getOpcode() == Instruction::Or) {
142 // If either the left or right side is invariant, we can unswitch on this,
143 // which will cause the branch to go away in one loop and the condition to
144 // simplify in the other one.
145 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed))
147 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed))
154 bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) {
155 assert(L->isLCSSAForm());
156 LI = &getAnalysis<LoopInfo>();
158 bool Changed = false;
160 // Loop over all of the basic blocks in the loop. If we find an interior
161 // block that is branching on a loop-invariant condition, we can unswitch this
163 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
165 TerminatorInst *TI = (*I)->getTerminator();
166 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
167 // If this isn't branching on an invariant condition, we can't unswitch
169 if (BI->isConditional()) {
170 // See if this, or some part of it, is loop invariant. If so, we can
171 // unswitch on it if we desire.
172 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(), L, Changed);
173 if (LoopCond && UnswitchIfProfitable(LoopCond, ConstantInt::getTrue(),
179 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
180 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), L, Changed);
181 if (LoopCond && SI->getNumCases() > 1) {
182 // Find a value to unswitch on:
183 // FIXME: this should chose the most expensive case!
184 Constant *UnswitchVal = SI->getCaseValue(1);
185 // Do not process same value again and again.
186 if (!UnswitchedVals.insert(UnswitchVal))
189 if (UnswitchIfProfitable(LoopCond, UnswitchVal, L)) {
196 // Scan the instructions to check for unswitchable values.
197 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end();
199 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) {
200 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), L, Changed);
201 if (LoopCond && UnswitchIfProfitable(LoopCond, ConstantInt::getTrue(),
209 assert(L->isLCSSAForm());
214 /// isTrivialLoopExitBlock - Check to see if all paths from BB either:
215 /// 1. Exit the loop with no side effects.
216 /// 2. Branch to the latch block with no side-effects.
218 /// If these conditions are true, we return true and set ExitBB to the block we
221 static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB,
223 std::set<BasicBlock*> &Visited) {
224 if (!Visited.insert(BB).second) {
225 // Already visited and Ok, end of recursion.
227 } else if (!L->contains(BB)) {
228 // Otherwise, this is a loop exit, this is fine so long as this is the
230 if (ExitBB != 0) return false;
235 // Otherwise, this is an unvisited intra-loop node. Check all successors.
236 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) {
237 // Check to see if the successor is a trivial loop exit.
238 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited))
242 // Okay, everything after this looks good, check to make sure that this block
243 // doesn't include any side effects.
244 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
245 if (I->mayWriteToMemory())
251 /// isTrivialLoopExitBlock - Return true if the specified block unconditionally
252 /// leads to an exit from the specified loop, and has no side-effects in the
253 /// process. If so, return the block that is exited to, otherwise return null.
254 static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) {
255 std::set<BasicBlock*> Visited;
256 Visited.insert(L->getHeader()); // Branches to header are ok.
257 BasicBlock *ExitBB = 0;
258 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited))
263 /// IsTrivialUnswitchCondition - Check to see if this unswitch condition is
264 /// trivial: that is, that the condition controls whether or not the loop does
265 /// anything at all. If this is a trivial condition, unswitching produces no
266 /// code duplications (equivalently, it produces a simpler loop and a new empty
267 /// loop, which gets deleted).
269 /// If this is a trivial condition, return true, otherwise return false. When
270 /// returning true, this sets Cond and Val to the condition that controls the
271 /// trivial condition: when Cond dynamically equals Val, the loop is known to
272 /// exit. Finally, this sets LoopExit to the BB that the loop exits to when
275 static bool IsTrivialUnswitchCondition(Loop *L, Value *Cond, Constant **Val = 0,
276 BasicBlock **LoopExit = 0) {
277 BasicBlock *Header = L->getHeader();
278 TerminatorInst *HeaderTerm = Header->getTerminator();
280 BasicBlock *LoopExitBB = 0;
281 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) {
282 // If the header block doesn't end with a conditional branch on Cond, we
284 if (!BI->isConditional() || BI->getCondition() != Cond)
287 // Check to see if a successor of the branch is guaranteed to go to the
288 // latch block or exit through a one exit block without having any
289 // side-effects. If so, determine the value of Cond that causes it to do
291 if ((LoopExitBB = isTrivialLoopExitBlock(L, BI->getSuccessor(0)))) {
292 if (Val) *Val = ConstantInt::getTrue();
293 } else if ((LoopExitBB = isTrivialLoopExitBlock(L, BI->getSuccessor(1)))) {
294 if (Val) *Val = ConstantInt::getFalse();
296 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) {
297 // If this isn't a switch on Cond, we can't handle it.
298 if (SI->getCondition() != Cond) return false;
300 // Check to see if a successor of the switch is guaranteed to go to the
301 // latch block or exit through a one exit block without having any
302 // side-effects. If so, determine the value of Cond that causes it to do
303 // this. Note that we can't trivially unswitch on the default case.
304 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
305 if ((LoopExitBB = isTrivialLoopExitBlock(L, SI->getSuccessor(i)))) {
306 // Okay, we found a trivial case, remember the value that is trivial.
307 if (Val) *Val = SI->getCaseValue(i);
312 // If we didn't find a single unique LoopExit block, or if the loop exit block
313 // contains phi nodes, this isn't trivial.
314 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin()))
315 return false; // Can't handle this.
317 if (LoopExit) *LoopExit = LoopExitBB;
319 // We already know that nothing uses any scalar values defined inside of this
320 // loop. As such, we just have to check to see if this loop will execute any
321 // side-effecting instructions (e.g. stores, calls, volatile loads) in the
322 // part of the loop that the code *would* execute. We already checked the
323 // tail, check the header now.
324 for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I)
325 if (I->mayWriteToMemory())
330 /// getLoopUnswitchCost - Return the cost (code size growth) that will happen if
331 /// we choose to unswitch the specified loop on the specified value.
333 unsigned LoopUnswitch::getLoopUnswitchCost(Loop *L, Value *LIC) {
334 // If the condition is trivial, always unswitch. There is no code growth for
336 if (IsTrivialUnswitchCondition(L, LIC))
339 // FIXME: This is really overly conservative. However, more liberal
340 // estimations have thus far resulted in excessive unswitching, which is bad
341 // both in compile time and in code size. This should be replaced once
342 // someone figures out how a good estimation.
343 return L->getBlocks().size();
346 // FIXME: this is brain dead. It should take into consideration code
348 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
351 // Do not include empty blocks in the cost calculation. This happen due to
352 // loop canonicalization and will be removed.
353 if (BB->begin() == BasicBlock::iterator(BB->getTerminator()))
356 // Count basic blocks.
363 /// UnswitchIfProfitable - We have found that we can unswitch L when
364 /// LoopCond == Val to simplify the loop. If we decide that this is profitable,
365 /// unswitch the loop, reprocess the pieces, then return true.
366 bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val,Loop *L){
367 // Check to see if it would be profitable to unswitch this loop.
368 unsigned Cost = getLoopUnswitchCost(L, LoopCond);
370 // Do not do non-trivial unswitch while optimizing for size.
371 if (Cost && OptimizeForSize)
374 if (Cost > Threshold) {
375 // FIXME: this should estimate growth by the amount of code shared by the
376 // resultant unswitched loops.
378 DOUT << "NOT unswitching loop %"
379 << L->getHeader()->getName() << ", cost too high: "
380 << L->getBlocks().size() << "\n";
384 // If this is a trivial condition to unswitch (which results in no code
385 // duplication), do it now.
387 BasicBlock *ExitBlock;
388 if (IsTrivialUnswitchCondition(L, LoopCond, &CondVal, &ExitBlock)) {
389 UnswitchTrivialCondition(L, LoopCond, CondVal, ExitBlock);
391 UnswitchNontrivialCondition(LoopCond, Val, L);
397 // RemapInstruction - Convert the instruction operands from referencing the
398 // current values into those specified by ValueMap.
400 static inline void RemapInstruction(Instruction *I,
401 DenseMap<const Value *, Value*> &ValueMap) {
402 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
403 Value *Op = I->getOperand(op);
404 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(Op);
405 if (It != ValueMap.end()) Op = It->second;
406 I->setOperand(op, Op);
410 // CloneDomInfo - NewBB is cloned from Orig basic block. Now clone Dominator Info.
412 // If Orig block's immediate dominator is mapped in VM then use corresponding
413 // immediate dominator from the map. Otherwise Orig block's dominator is also
414 // NewBB's dominator.
416 // OrigPreheader is loop pre-header before this pass started
417 // updating CFG. NewPrehader is loops new pre-header. However, after CFG
418 // manipulation, loop L may not exist. So rely on input parameter NewPreheader.
419 void CloneDomInfo(BasicBlock *NewBB, BasicBlock *Orig,
420 BasicBlock *NewPreheader, BasicBlock *OrigPreheader,
421 BasicBlock *OrigHeader,
422 DominatorTree *DT, DominanceFrontier *DF,
423 DenseMap<const Value*, Value*> &VM) {
425 // If NewBB alreay has found its place in domiantor tree then no need to do
427 if (DT->getNode(NewBB))
430 // If Orig does not have any immediate domiantor then its clone, NewBB, does
431 // not need any immediate dominator.
432 DomTreeNode *OrigNode = DT->getNode(Orig);
435 DomTreeNode *OrigIDomNode = OrigNode->getIDom();
439 BasicBlock *OrigIDom = NULL;
441 // If Orig is original loop header then its immediate dominator is
443 if (Orig == OrigHeader)
444 OrigIDom = NewPreheader;
446 // If Orig is new pre-header then its immediate dominator is
447 // original pre-header.
448 else if (Orig == NewPreheader)
449 OrigIDom = OrigPreheader;
451 // Other as DT to find Orig's immediate dominator.
453 OrigIDom = OrigIDomNode->getBlock();
455 // Initially use Orig's immediate dominator as NewBB's immediate dominator.
456 BasicBlock *NewIDom = OrigIDom;
457 DenseMap<const Value*, Value*>::iterator I = VM.find(OrigIDom);
459 // if (!DT->getNode(OrigIDom))
460 // CloneDomInfo(NewIDom, OrigIDom, NewPreheader, OrigPreheader,
461 // OrigHeader, DT, DF, VM);
463 NewIDom = cast<BasicBlock>(I->second);
465 // If NewIDom does not have corresponding dominatore tree node then
467 if (!DT->getNode(NewIDom))
468 CloneDomInfo(NewIDom, OrigIDom, NewPreheader, OrigPreheader,
469 OrigHeader, DT, DF, VM);
471 // if (NewBB == NewIDom) {
472 // DT->addNewBlock(NewBB, OrigIDom);
473 // DT->changeImmediateDominator(NewBB, NewIDom);
475 DT->addNewBlock(NewBB, NewIDom);
477 DominanceFrontier::DomSetType NewDFSet;
479 DominanceFrontier::iterator DFI = DF->find(Orig);
480 if ( DFI != DF->end()) {
481 DominanceFrontier::DomSetType S = DFI->second;
482 for (DominanceFrontier::DomSetType::iterator I = S.begin(), E = S.end();
485 DenseMap<const Value*, Value*>::iterator I = VM.find(BB);
487 NewDFSet.insert(cast<BasicBlock>(I->second));
492 DF->addBasicBlock(NewBB, NewDFSet);
496 /// CloneLoop - Recursively clone the specified loop and all of its children,
497 /// mapping the blocks with the specified map.
498 static Loop *CloneLoop(Loop *L, Loop *PL, DenseMap<const Value*, Value*> &VM,
499 LoopInfo *LI, LPPassManager *LPM) {
500 Loop *New = new Loop();
502 LPM->insertLoop(New, PL);
504 // Add all of the blocks in L to the new loop.
505 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
507 if (LI->getLoopFor(*I) == L)
508 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), *LI);
510 // Add all of the subloops to the new loop.
511 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
512 CloneLoop(*I, New, VM, LI, LPM);
517 /// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values
518 /// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the
519 /// code immediately before InsertPt.
520 void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
521 BasicBlock *TrueDest,
522 BasicBlock *FalseDest,
523 Instruction *InsertPt) {
524 // Insert a conditional branch on LIC to the two preheaders. The original
525 // code is the true version and the new code is the false version.
526 Value *BranchVal = LIC;
527 if (!isa<ConstantInt>(Val) || Val->getType() != Type::Int1Ty)
528 BranchVal = new ICmpInst(ICmpInst::ICMP_EQ, LIC, Val, "tmp", InsertPt);
529 else if (Val != ConstantInt::getTrue())
530 // We want to enter the new loop when the condition is true.
531 std::swap(TrueDest, FalseDest);
533 // Insert the new branch.
534 BranchInst *BRI = new BranchInst(TrueDest, FalseDest, BranchVal, InsertPt);
536 // Update dominator info.
537 // BranchVal is a new preheader so it dominates true and false destination
539 if (DominatorTree *DT = getAnalysisToUpdate<DominatorTree>()) {
540 DT->changeImmediateDominator(TrueDest, BRI->getParent());
541 DT->changeImmediateDominator(FalseDest, BRI->getParent());
543 // No need to update DominanceFrontier. BRI->getParent() dominated TrueDest
544 // and FalseDest anyway. Now it immediately dominates them.
548 /// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable
549 /// condition in it (a cond branch from its header block to its latch block,
550 /// where the path through the loop that doesn't execute its body has no
551 /// side-effects), unswitch it. This doesn't involve any code duplication, just
552 /// moving the conditional branch outside of the loop and updating loop info.
553 void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond,
555 BasicBlock *ExitBlock) {
556 DOUT << "loop-unswitch: Trivial-Unswitch loop %"
557 << L->getHeader()->getName() << " [" << L->getBlocks().size()
558 << " blocks] in Function " << L->getHeader()->getParent()->getName()
559 << " on cond: " << *Val << " == " << *Cond << "\n";
561 // First step, split the preheader, so that we know that there is a safe place
562 // to insert the conditional branch. We will change 'OrigPH' to have a
563 // conditional branch on Cond.
564 BasicBlock *OrigPH = L->getLoopPreheader();
565 BasicBlock *NewPH = SplitEdge(OrigPH, L->getHeader(), this);
567 // Now that we have a place to insert the conditional branch, create a place
568 // to branch to: this is the exit block out of the loop that we should
571 // Split this block now, so that the loop maintains its exit block, and so
572 // that the jump from the preheader can execute the contents of the exit block
573 // without actually branching to it (the exit block should be dominated by the
574 // loop header, not the preheader).
575 assert(!L->contains(ExitBlock) && "Exit block is in the loop?");
576 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), this);
578 // Okay, now we have a position to branch from and a position to branch to,
579 // insert the new conditional branch.
580 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH,
581 OrigPH->getTerminator());
582 OrigPH->getTerminator()->eraseFromParent();
584 // We need to reprocess this loop, it could be unswitched again.
587 // Now that we know that the loop is never entered when this condition is a
588 // particular value, rewrite the loop with this info. We know that this will
589 // at least eliminate the old branch.
590 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false);
594 /// VersionLoop - We determined that the loop is profitable to unswitch when LIC
595 /// equal Val. Split it into loop versions and test the condition outside of
596 /// either loop. Return the loops created as Out1/Out2.
597 void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val,
599 Function *F = L->getHeader()->getParent();
600 DOUT << "loop-unswitch: Unswitching loop %"
601 << L->getHeader()->getName() << " [" << L->getBlocks().size()
602 << " blocks] in Function " << F->getName()
603 << " when '" << *Val << "' == " << *LIC << "\n";
605 // LoopBlocks contains all of the basic blocks of the loop, including the
606 // preheader of the loop, the body of the loop, and the exit blocks of the
607 // loop, in that order.
608 std::vector<BasicBlock*> LoopBlocks;
610 // First step, split the preheader and exit blocks, and add these blocks to
611 // the LoopBlocks list.
612 BasicBlock *OrigHeader = L->getHeader();
613 BasicBlock *OrigPreheader = L->getLoopPreheader();
614 BasicBlock *NewPreheader = SplitEdge(OrigPreheader, L->getHeader(), this);
615 LoopBlocks.push_back(NewPreheader);
617 // We want the loop to come after the preheader, but before the exit blocks.
618 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
620 std::vector<BasicBlock*> ExitBlocks;
621 L->getUniqueExitBlocks(ExitBlocks);
623 // Split all of the edges from inside the loop to their exit blocks. Update
624 // the appropriate Phi nodes as we do so.
625 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
626 BasicBlock *ExitBlock = ExitBlocks[i];
627 std::vector<BasicBlock*> Preds(pred_begin(ExitBlock), pred_end(ExitBlock));
629 for (unsigned j = 0, e = Preds.size(); j != e; ++j) {
630 BasicBlock* MiddleBlock = SplitEdge(Preds[j], ExitBlock, this);
631 BasicBlock* StartBlock = Preds[j];
632 BasicBlock* EndBlock;
633 if (MiddleBlock->getSinglePredecessor() == ExitBlock) {
634 EndBlock = MiddleBlock;
635 MiddleBlock = EndBlock->getSinglePredecessor();;
637 EndBlock = ExitBlock;
640 std::set<PHINode*> InsertedPHIs;
641 PHINode* OldLCSSA = 0;
642 for (BasicBlock::iterator I = EndBlock->begin();
643 (OldLCSSA = dyn_cast<PHINode>(I)); ++I) {
644 Value* OldValue = OldLCSSA->getIncomingValueForBlock(MiddleBlock);
645 PHINode* NewLCSSA = new PHINode(OldLCSSA->getType(),
646 OldLCSSA->getName() + ".us-lcssa",
647 MiddleBlock->getTerminator());
648 NewLCSSA->addIncoming(OldValue, StartBlock);
649 OldLCSSA->setIncomingValue(OldLCSSA->getBasicBlockIndex(MiddleBlock),
651 InsertedPHIs.insert(NewLCSSA);
654 BasicBlock::iterator InsertPt = EndBlock->begin();
655 while (dyn_cast<PHINode>(InsertPt)) ++InsertPt;
656 for (BasicBlock::iterator I = MiddleBlock->begin();
657 (OldLCSSA = dyn_cast<PHINode>(I)) && InsertedPHIs.count(OldLCSSA) == 0;
659 PHINode *NewLCSSA = new PHINode(OldLCSSA->getType(),
660 OldLCSSA->getName() + ".us-lcssa",
662 OldLCSSA->replaceAllUsesWith(NewLCSSA);
663 NewLCSSA->addIncoming(OldLCSSA, MiddleBlock);
668 // The exit blocks may have been changed due to edge splitting, recompute.
670 L->getUniqueExitBlocks(ExitBlocks);
672 // Add exit blocks to the loop blocks.
673 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end());
675 // Next step, clone all of the basic blocks that make up the loop (including
676 // the loop preheader and exit blocks), keeping track of the mapping between
677 // the instructions and blocks.
678 std::vector<BasicBlock*> NewBlocks;
679 NewBlocks.reserve(LoopBlocks.size());
680 DenseMap<const Value*, Value*> ValueMap;
681 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
682 BasicBlock *New = CloneBasicBlock(LoopBlocks[i], ValueMap, ".us", F);
683 NewBlocks.push_back(New);
684 ValueMap[LoopBlocks[i]] = New; // Keep the BB mapping.
687 // Update dominator info
688 DominanceFrontier *DF = getAnalysisToUpdate<DominanceFrontier>();
689 if (DominatorTree *DT = getAnalysisToUpdate<DominatorTree>())
690 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
691 BasicBlock *LBB = LoopBlocks[i];
692 BasicBlock *NBB = NewBlocks[i];
693 CloneDomInfo(NBB, LBB, NewPreheader, OrigPreheader,
694 OrigHeader, DT, DF, ValueMap);
697 // Splice the newly inserted blocks into the function right before the
698 // original preheader.
699 F->getBasicBlockList().splice(LoopBlocks[0], F->getBasicBlockList(),
700 NewBlocks[0], F->end());
702 // Now we create the new Loop object for the versioned loop.
703 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), ValueMap, LI, LPM);
704 Loop *ParentLoop = L->getParentLoop();
706 // Make sure to add the cloned preheader and exit blocks to the parent loop
708 ParentLoop->addBasicBlockToLoop(NewBlocks[0], *LI);
711 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
712 BasicBlock *NewExit = cast<BasicBlock>(ValueMap[ExitBlocks[i]]);
713 // The new exit block should be in the same loop as the old one.
714 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i]))
715 ExitBBLoop->addBasicBlockToLoop(NewExit, *LI);
717 assert(NewExit->getTerminator()->getNumSuccessors() == 1 &&
718 "Exit block should have been split to have one successor!");
719 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0);
721 // If the successor of the exit block had PHI nodes, add an entry for
724 for (BasicBlock::iterator I = ExitSucc->begin();
725 (PN = dyn_cast<PHINode>(I)); ++I) {
726 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]);
727 DenseMap<const Value *, Value*>::iterator It = ValueMap.find(V);
728 if (It != ValueMap.end()) V = It->second;
729 PN->addIncoming(V, NewExit);
733 // Rewrite the code to refer to itself.
734 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i)
735 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
736 E = NewBlocks[i]->end(); I != E; ++I)
737 RemapInstruction(I, ValueMap);
739 // Rewrite the original preheader to select between versions of the loop.
740 BranchInst *OldBR = cast<BranchInst>(OrigPreheader->getTerminator());
741 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] &&
742 "Preheader splitting did not work correctly!");
744 // Emit the new branch that selects between the two versions of this loop.
745 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR);
746 OldBR->eraseFromParent();
748 LoopProcessWorklist.push_back(NewLoop);
751 // Now we rewrite the original code to know that the condition is true and the
752 // new code to know that the condition is false.
753 RewriteLoopBodyWithConditionConstant(L , LIC, Val, false);
755 // It's possible that simplifying one loop could cause the other to be
756 // deleted. If so, don't simplify it.
757 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop)
758 RewriteLoopBodyWithConditionConstant(NewLoop, LIC, Val, true);
761 /// RemoveFromWorklist - Remove all instances of I from the worklist vector
763 static void RemoveFromWorklist(Instruction *I,
764 std::vector<Instruction*> &Worklist) {
765 std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(),
767 while (WI != Worklist.end()) {
768 unsigned Offset = WI-Worklist.begin();
770 WI = std::find(Worklist.begin()+Offset, Worklist.end(), I);
774 /// ReplaceUsesOfWith - When we find that I really equals V, remove I from the
775 /// program, replacing all uses with V and update the worklist.
776 static void ReplaceUsesOfWith(Instruction *I, Value *V,
777 std::vector<Instruction*> &Worklist) {
778 DOUT << "Replace with '" << *V << "': " << *I;
780 // Add uses to the worklist, which may be dead now.
781 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
782 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
783 Worklist.push_back(Use);
785 // Add users to the worklist which may be simplified now.
786 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
788 Worklist.push_back(cast<Instruction>(*UI));
789 I->replaceAllUsesWith(V);
790 I->eraseFromParent();
791 RemoveFromWorklist(I, Worklist);
795 /// RemoveBlockIfDead - If the specified block is dead, remove it, update loop
796 /// information, and remove any dead successors it has.
798 void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB,
799 std::vector<Instruction*> &Worklist) {
800 if (pred_begin(BB) != pred_end(BB)) {
801 // This block isn't dead, since an edge to BB was just removed, see if there
802 // are any easy simplifications we can do now.
803 if (BasicBlock *Pred = BB->getSinglePredecessor()) {
804 // If it has one pred, fold phi nodes in BB.
805 while (isa<PHINode>(BB->begin()))
806 ReplaceUsesOfWith(BB->begin(),
807 cast<PHINode>(BB->begin())->getIncomingValue(0),
810 // If this is the header of a loop and the only pred is the latch, we now
811 // have an unreachable loop.
812 if (Loop *L = LI->getLoopFor(BB))
813 if (L->getHeader() == BB && L->contains(Pred)) {
814 // Remove the branch from the latch to the header block, this makes
815 // the header dead, which will make the latch dead (because the header
816 // dominates the latch).
817 Pred->getTerminator()->eraseFromParent();
818 new UnreachableInst(Pred);
820 // The loop is now broken, remove it from LI.
821 RemoveLoopFromHierarchy(L);
823 // Reprocess the header, which now IS dead.
824 RemoveBlockIfDead(BB, Worklist);
828 // If pred ends in a uncond branch, add uncond branch to worklist so that
829 // the two blocks will get merged.
830 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator()))
831 if (BI->isUnconditional())
832 Worklist.push_back(BI);
837 DOUT << "Nuking dead block: " << *BB;
839 // Remove the instructions in the basic block from the worklist.
840 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
841 RemoveFromWorklist(I, Worklist);
843 // Anything that uses the instructions in this basic block should have their
844 // uses replaced with undefs.
846 I->replaceAllUsesWith(UndefValue::get(I->getType()));
849 // If this is the edge to the header block for a loop, remove the loop and
850 // promote all subloops.
851 if (Loop *BBLoop = LI->getLoopFor(BB)) {
852 if (BBLoop->getLoopLatch() == BB)
853 RemoveLoopFromHierarchy(BBLoop);
856 // Remove the block from the loop info, which removes it from any loops it
861 // Remove phi node entries in successors for this block.
862 TerminatorInst *TI = BB->getTerminator();
863 std::vector<BasicBlock*> Succs;
864 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
865 Succs.push_back(TI->getSuccessor(i));
866 TI->getSuccessor(i)->removePredecessor(BB);
869 // Unique the successors, remove anything with multiple uses.
870 std::sort(Succs.begin(), Succs.end());
871 Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end());
873 // Remove the basic block, including all of the instructions contained in it.
874 BB->eraseFromParent();
876 // Remove successor blocks here that are not dead, so that we know we only
877 // have dead blocks in this list. Nondead blocks have a way of becoming dead,
878 // then getting removed before we revisit them, which is badness.
880 for (unsigned i = 0; i != Succs.size(); ++i)
881 if (pred_begin(Succs[i]) != pred_end(Succs[i])) {
882 // One exception is loop headers. If this block was the preheader for a
883 // loop, then we DO want to visit the loop so the loop gets deleted.
884 // We know that if the successor is a loop header, that this loop had to
885 // be the preheader: the case where this was the latch block was handled
886 // above and headers can only have two predecessors.
887 if (!LI->isLoopHeader(Succs[i])) {
888 Succs.erase(Succs.begin()+i);
893 for (unsigned i = 0, e = Succs.size(); i != e; ++i)
894 RemoveBlockIfDead(Succs[i], Worklist);
897 /// RemoveLoopFromHierarchy - We have discovered that the specified loop has
898 /// become unwrapped, either because the backedge was deleted, or because the
899 /// edge into the header was removed. If the edge into the header from the
900 /// latch block was removed, the loop is unwrapped but subloops are still alive,
901 /// so they just reparent loops. If the loops are actually dead, they will be
903 void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) {
904 LPM->deleteLoopFromQueue(L);
905 RemoveLoopFromWorklist(L);
910 // RewriteLoopBodyWithConditionConstant - We know either that the value LIC has
911 // the value specified by Val in the specified loop, or we know it does NOT have
912 // that value. Rewrite any uses of LIC or of properties correlated to it.
913 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
916 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?");
918 // FIXME: Support correlated properties, like:
925 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches,
926 // selects, switches.
927 std::vector<User*> Users(LIC->use_begin(), LIC->use_end());
928 std::vector<Instruction*> Worklist;
930 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC
931 // in the loop with the appropriate one directly.
932 if (IsEqual || (isa<ConstantInt>(Val) && Val->getType() == Type::Int1Ty)) {
937 Replacement = ConstantInt::get(Type::Int1Ty,
938 !cast<ConstantInt>(Val)->getZExtValue());
940 for (unsigned i = 0, e = Users.size(); i != e; ++i)
941 if (Instruction *U = cast<Instruction>(Users[i])) {
942 if (!L->contains(U->getParent()))
944 U->replaceUsesOfWith(LIC, Replacement);
945 Worklist.push_back(U);
948 // Otherwise, we don't know the precise value of LIC, but we do know that it
949 // is certainly NOT "Val". As such, simplify any uses in the loop that we
950 // can. This case occurs when we unswitch switch statements.
951 for (unsigned i = 0, e = Users.size(); i != e; ++i)
952 if (Instruction *U = cast<Instruction>(Users[i])) {
953 if (!L->contains(U->getParent()))
956 Worklist.push_back(U);
958 // If we know that LIC is not Val, use this info to simplify code.
959 if (SwitchInst *SI = dyn_cast<SwitchInst>(U)) {
960 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) {
961 if (SI->getCaseValue(i) == Val) {
962 // Found a dead case value. Don't remove PHI nodes in the
963 // successor if they become single-entry, those PHI nodes may
964 // be in the Users list.
966 // FIXME: This is a hack. We need to keep the successor around
967 // and hooked up so as to preserve the loop structure, because
968 // trying to update it is complicated. So instead we preserve the
969 // loop structure and put the block on an dead code path.
971 BasicBlock* Old = SI->getParent();
972 BasicBlock* Split = SplitBlock(Old, SI, this);
974 Instruction* OldTerm = Old->getTerminator();
975 new BranchInst(Split, SI->getSuccessor(i),
976 ConstantInt::getTrue(), OldTerm);
978 Old->getTerminator()->eraseFromParent();
982 for (BasicBlock::iterator II = SI->getSuccessor(i)->begin();
983 (PN = dyn_cast<PHINode>(II)); ++II) {
984 Value *InVal = PN->removeIncomingValue(Split, false);
985 PN->addIncoming(InVal, Old);
994 // TODO: We could do other simplifications, for example, turning
995 // LIC == Val -> false.
999 SimplifyCode(Worklist);
1002 /// SimplifyCode - Okay, now that we have simplified some instructions in the
1003 /// loop, walk over it and constant prop, dce, and fold control flow where
1004 /// possible. Note that this is effectively a very simple loop-structure-aware
1005 /// optimizer. During processing of this loop, L could very well be deleted, so
1006 /// it must not be used.
1008 /// FIXME: When the loop optimizer is more mature, separate this out to a new
1011 void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist) {
1012 while (!Worklist.empty()) {
1013 Instruction *I = Worklist.back();
1014 Worklist.pop_back();
1016 // Simple constant folding.
1017 if (Constant *C = ConstantFoldInstruction(I)) {
1018 ReplaceUsesOfWith(I, C, Worklist);
1023 if (isInstructionTriviallyDead(I)) {
1024 DOUT << "Remove dead instruction '" << *I;
1026 // Add uses to the worklist, which may be dead now.
1027 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
1028 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
1029 Worklist.push_back(Use);
1030 I->eraseFromParent();
1031 RemoveFromWorklist(I, Worklist);
1036 // Special case hacks that appear commonly in unswitched code.
1037 switch (I->getOpcode()) {
1038 case Instruction::Select:
1039 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(0))) {
1040 ReplaceUsesOfWith(I, I->getOperand(!CB->getZExtValue()+1), Worklist);
1044 case Instruction::And:
1045 if (isa<ConstantInt>(I->getOperand(0)) &&
1046 I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS
1047 cast<BinaryOperator>(I)->swapOperands();
1048 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
1049 if (CB->getType() == Type::Int1Ty) {
1050 if (CB->isOne()) // X & 1 -> X
1051 ReplaceUsesOfWith(I, I->getOperand(0), Worklist);
1053 ReplaceUsesOfWith(I, I->getOperand(1), Worklist);
1057 case Instruction::Or:
1058 if (isa<ConstantInt>(I->getOperand(0)) &&
1059 I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS
1060 cast<BinaryOperator>(I)->swapOperands();
1061 if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
1062 if (CB->getType() == Type::Int1Ty) {
1063 if (CB->isOne()) // X | 1 -> 1
1064 ReplaceUsesOfWith(I, I->getOperand(1), Worklist);
1066 ReplaceUsesOfWith(I, I->getOperand(0), Worklist);
1070 case Instruction::Br: {
1071 BranchInst *BI = cast<BranchInst>(I);
1072 if (BI->isUnconditional()) {
1073 // If BI's parent is the only pred of the successor, fold the two blocks
1075 BasicBlock *Pred = BI->getParent();
1076 BasicBlock *Succ = BI->getSuccessor(0);
1077 BasicBlock *SinglePred = Succ->getSinglePredecessor();
1078 if (!SinglePred) continue; // Nothing to do.
1079 assert(SinglePred == Pred && "CFG broken");
1081 DOUT << "Merging blocks: " << Pred->getName() << " <- "
1082 << Succ->getName() << "\n";
1084 // Resolve any single entry PHI nodes in Succ.
1085 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin()))
1086 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist);
1088 // Move all of the successor contents from Succ to Pred.
1089 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(),
1091 BI->eraseFromParent();
1092 RemoveFromWorklist(BI, Worklist);
1094 // If Succ has any successors with PHI nodes, update them to have
1095 // entries coming from Pred instead of Succ.
1096 Succ->replaceAllUsesWith(Pred);
1098 // Remove Succ from the loop tree.
1099 LI->removeBlock(Succ);
1100 Succ->eraseFromParent();
1102 } else if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){
1103 // Conditional branch. Turn it into an unconditional branch, then
1104 // remove dead blocks.
1105 break; // FIXME: Enable.
1107 DOUT << "Folded branch: " << *BI;
1108 BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue());
1109 BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue());
1110 DeadSucc->removePredecessor(BI->getParent(), true);
1111 Worklist.push_back(new BranchInst(LiveSucc, BI));
1112 BI->eraseFromParent();
1113 RemoveFromWorklist(BI, Worklist);
1116 RemoveBlockIfDead(DeadSucc, Worklist);