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/Function.h"
33 #include "llvm/Instructions.h"
34 #include "llvm/Analysis/LoopInfo.h"
35 #include "llvm/Transforms/Utils/Cloning.h"
36 #include "llvm/Transforms/Utils/Local.h"
37 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
38 #include "llvm/ADT/Statistic.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/CommandLine.h"
47 Statistic<> NumBranches("loop-unswitch", "Number of branches unswitched");
48 Statistic<> NumSwitches("loop-unswitch", "Number of switches unswitched");
49 Statistic<> NumSelects ("loop-unswitch", "Number of selects unswitched");
50 Statistic<> NumTrivial ("loop-unswitch",
51 "Number of unswitches that are trivial");
52 Statistic<> NumSimplify("loop-unswitch",
53 "Number of simplifications of unswitched code");
55 Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"),
56 cl::init(10), cl::Hidden);
58 class LoopUnswitch : public FunctionPass {
59 LoopInfo *LI; // Loop information
61 virtual bool runOnFunction(Function &F);
62 bool visitLoop(Loop *L);
64 /// This transformation requires natural loop information & requires that
65 /// loop preheaders be inserted into the CFG...
67 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
68 AU.addRequiredID(LoopSimplifyID);
69 AU.addPreservedID(LoopSimplifyID);
70 AU.addRequired<LoopInfo>();
71 AU.addPreserved<LoopInfo>();
75 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val,Loop *L);
76 unsigned getLoopUnswitchCost(Loop *L, Value *LIC);
77 void VersionLoop(Value *LIC, Constant *OnVal,
78 Loop *L, Loop *&Out1, Loop *&Out2);
79 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val,
80 bool EntersWhenTrue, BasicBlock *ExitBlock);
81 BasicBlock *SplitEdge(BasicBlock *From, BasicBlock *To);
82 BasicBlock *SplitBlock(BasicBlock *Old, Instruction *SplitPt);
83 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,Constant *Val,
85 bool TryToRemoveEdge(TerminatorInst *TI, unsigned SuccNo,
86 std::vector<Instruction*> &Worklist);
88 RegisterOpt<LoopUnswitch> X("loop-unswitch", "Unswitch loops");
91 FunctionPass *llvm::createLoopUnswitchPass() { return new LoopUnswitch(); }
93 bool LoopUnswitch::runOnFunction(Function &F) {
95 LI = &getAnalysis<LoopInfo>();
97 // Transform all the top-level loops. Copy the loop list so that the child
98 // can update the loop tree if it needs to delete the loop.
99 std::vector<Loop*> SubLoops(LI->begin(), LI->end());
100 for (unsigned i = 0, e = SubLoops.size(); i != e; ++i)
101 Changed |= visitLoop(SubLoops[i]);
107 /// LoopValuesUsedOutsideLoop - Return true if there are any values defined in
108 /// the loop that are used by instructions outside of it.
109 static bool LoopValuesUsedOutsideLoop(Loop *L) {
110 // We will be doing lots of "loop contains block" queries. Loop::contains is
111 // linear time, use a set to speed this up.
112 std::set<BasicBlock*> LoopBlocks;
114 for (Loop::block_iterator BB = L->block_begin(), E = L->block_end();
116 LoopBlocks.insert(*BB);
118 for (Loop::block_iterator BB = L->block_begin(), E = L->block_end();
120 for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end(); I != E; ++I)
121 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
123 BasicBlock *UserBB = cast<Instruction>(*UI)->getParent();
124 if (!LoopBlocks.count(UserBB))
131 /// isTrivialLoopExitBlock - Check to see if all paths from BB either:
132 /// 1. Exit the loop with no side effects.
133 /// 2. Branch to the latch block with no side-effects.
135 /// If these conditions are true, we return true and set ExitBB to the block we
138 static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB,
140 std::set<BasicBlock*> &Visited) {
141 if (!Visited.insert(BB).second) {
142 // Already visited and Ok, end of recursion.
144 } else if (!L->contains(BB)) {
145 // Otherwise, this is a loop exit, this is fine so long as this is the
147 if (ExitBB != 0) return false;
152 // Otherwise, this is an unvisited intra-loop node. Check all successors.
153 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) {
154 // Check to see if the successor is a trivial loop exit.
155 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited))
159 // Okay, everything after this looks good, check to make sure that this block
160 // doesn't include any side effects.
161 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
162 if (I->mayWriteToMemory())
168 static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) {
169 std::set<BasicBlock*> Visited;
170 Visited.insert(L->getHeader()); // Branches to header are ok.
171 Visited.insert(BB); // Don't revisit BB after we do.
172 BasicBlock *ExitBB = 0;
173 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited))
178 /// IsTrivialUnswitchCondition - Check to see if this unswitch condition is
179 /// trivial: that is, that the condition controls whether or not the loop does
180 /// anything at all. If this is a trivial condition, unswitching produces no
181 /// code duplications (equivalently, it produces a simpler loop and a new empty
182 /// loop, which gets deleted).
184 /// If this is a trivial condition, return ConstantBool::True if the loop body
185 /// runs when the condition is true, False if the loop body executes when the
186 /// condition is false. Otherwise, return null to indicate a complex condition.
187 static bool IsTrivialUnswitchCondition(Loop *L, Value *Cond,
189 bool *EntersWhenTrue = 0,
190 BasicBlock **LoopExit = 0) {
191 BasicBlock *Header = L->getHeader();
192 TerminatorInst *HeaderTerm = Header->getTerminator();
194 BasicBlock *LoopExitBB = 0;
195 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) {
196 // If the header block doesn't end with a conditional branch on Cond, we
198 if (!BI->isConditional() || BI->getCondition() != Cond)
201 // Check to see if a successor of the branch is guaranteed to go to the
202 // latch block or exit through a one exit block without having any
203 // side-effects. If so, determine the value of Cond that causes it to do
205 if ((LoopExitBB = isTrivialLoopExitBlock(L, BI->getSuccessor(0)))) {
206 if (Val) *Val = ConstantBool::False;
207 } else if ((LoopExitBB = isTrivialLoopExitBlock(L, BI->getSuccessor(1)))) {
208 if (Val) *Val = ConstantBool::True;
210 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) {
211 // If this isn't a switch on Cond, we can't handle it.
212 if (SI->getCondition() != Cond) return false;
214 // Check to see if a successor of the switch is guaranteed to go to the
215 // latch block or exit through a one exit block without having any
216 // side-effects. If so, determine the value of Cond that causes it to do
217 // this. Note that we can't trivially unswitch on the default case.
218 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
219 if ((LoopExitBB = isTrivialLoopExitBlock(L, SI->getSuccessor(i)))) {
220 // Okay, we found a trivial case, remember the value that is trivial.
221 if (Val) *Val = SI->getCaseValue(i);
222 if (EntersWhenTrue) *EntersWhenTrue = false;
228 return false; // Can't handle this.
230 if (LoopExit) *LoopExit = LoopExitBB;
232 // We already know that nothing uses any scalar values defined inside of this
233 // loop. As such, we just have to check to see if this loop will execute any
234 // side-effecting instructions (e.g. stores, calls, volatile loads) in the
235 // part of the loop that the code *would* execute. We already checked the
236 // tail, check the header now.
237 for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I)
238 if (I->mayWriteToMemory())
243 /// getLoopUnswitchCost - Return the cost (code size growth) that will happen if
244 /// we choose to unswitch the specified loop on the specified value.
246 unsigned LoopUnswitch::getLoopUnswitchCost(Loop *L, Value *LIC) {
247 // If the condition is trivial, always unswitch. There is no code growth for
249 if (IsTrivialUnswitchCondition(L, LIC))
253 // FIXME: this is brain dead. It should take into consideration code
255 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
258 // Do not include empty blocks in the cost calculation. This happen due to
259 // loop canonicalization and will be removed.
260 if (BB->begin() == BasicBlock::iterator(BB->getTerminator()))
263 // Count basic blocks.
270 /// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is
271 /// invariant in the loop, or has an invariant piece, return the invariant.
272 /// Otherwise, return null.
273 static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) {
274 // Constants should be folded, not unswitched on!
275 if (isa<Constant>(Cond)) return false;
277 // TODO: Handle: br (VARIANT|INVARIANT).
278 // TODO: Hoist simple expressions out of loops.
279 if (L->isLoopInvariant(Cond)) return Cond;
281 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond))
282 if (BO->getOpcode() == Instruction::And ||
283 BO->getOpcode() == Instruction::Or) {
284 // If either the left or right side is invariant, we can unswitch on this,
285 // which will cause the branch to go away in one loop and the condition to
286 // simplify in the other one.
287 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed))
289 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed))
296 bool LoopUnswitch::visitLoop(Loop *L) {
297 bool Changed = false;
299 // Recurse through all subloops before we process this loop. Copy the loop
300 // list so that the child can update the loop tree if it needs to delete the
302 std::vector<Loop*> SubLoops(L->begin(), L->end());
303 for (unsigned i = 0, e = SubLoops.size(); i != e; ++i)
304 Changed |= visitLoop(SubLoops[i]);
306 // Loop over all of the basic blocks in the loop. If we find an interior
307 // block that is branching on a loop-invariant condition, we can unswitch this
309 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
311 TerminatorInst *TI = (*I)->getTerminator();
312 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
313 // If this isn't branching on an invariant condition, we can't unswitch
315 if (BI->isConditional()) {
316 // See if this, or some part of it, is loop invariant. If so, we can
317 // unswitch on it if we desire.
318 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(), L, Changed);
319 if (LoopCond && UnswitchIfProfitable(LoopCond, ConstantBool::True, L)) {
324 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
325 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), L, Changed);
326 if (LoopCond && SI->getNumCases() > 1) {
327 // Find a value to unswitch on:
328 // FIXME: this should chose the most expensive case!
329 Constant *UnswitchVal = SI->getCaseValue(1);
330 if (UnswitchIfProfitable(LoopCond, UnswitchVal, L)) {
337 // Scan the instructions to check for unswitchable values.
338 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end();
340 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) {
341 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), L, Changed);
342 if (LoopCond && UnswitchIfProfitable(LoopCond, ConstantBool::True, L)) {
352 /// UnswitchIfProfitable - We have found that we can unswitch L when
353 /// LoopCond == Val to simplify the loop. If we decide that this is profitable,
354 /// unswitch the loop, reprocess the pieces, then return true.
355 bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val,Loop *L){
356 // Check to see if it would be profitable to unswitch this loop.
357 if (getLoopUnswitchCost(L, LoopCond) > Threshold) {
358 // FIXME: this should estimate growth by the amount of code shared by the
359 // resultant unswitched loops.
361 DEBUG(std::cerr << "NOT unswitching loop %"
362 << L->getHeader()->getName() << ", cost too high: "
363 << L->getBlocks().size() << "\n");
367 // If this loop has live-out values, we can't unswitch it. We need something
368 // like loop-closed SSA form in order to know how to insert PHI nodes for
370 if (LoopValuesUsedOutsideLoop(L)) {
371 DEBUG(std::cerr << "NOT unswitching loop %" << L->getHeader()->getName()
372 << ", a loop value is used outside loop!\n");
376 //std::cerr << "BEFORE:\n"; LI->dump();
377 Loop *NewLoop1 = 0, *NewLoop2 = 0;
379 // If this is a trivial condition to unswitch (which results in no code
380 // duplication), do it now.
382 bool EntersWhenTrue = true;
383 BasicBlock *ExitBlock;
384 if (IsTrivialUnswitchCondition(L, LoopCond, &CondVal,
385 &EntersWhenTrue, &ExitBlock)) {
386 UnswitchTrivialCondition(L, LoopCond, CondVal, EntersWhenTrue, ExitBlock);
389 VersionLoop(LoopCond, Val, L, NewLoop1, NewLoop2);
392 //std::cerr << "AFTER:\n"; LI->dump();
394 // Try to unswitch each of our new loops now!
395 if (NewLoop1) visitLoop(NewLoop1);
396 if (NewLoop2) visitLoop(NewLoop2);
400 /// SplitBlock - Split the specified block at the specified instruction - every
401 /// thing before SplitPt stays in Old and everything starting with SplitPt moves
402 /// to a new block. The two blocks are joined by an unconditional branch and
403 /// the loop info is updated.
405 BasicBlock *LoopUnswitch::SplitBlock(BasicBlock *Old, Instruction *SplitPt) {
406 BasicBlock::iterator SplitIt = SplitPt;
407 while (isa<PHINode>(SplitIt))
409 BasicBlock *New = Old->splitBasicBlock(SplitIt, Old->getName()+".split");
411 // The new block lives in whichever loop the old one did.
412 if (Loop *L = LI->getLoopFor(Old))
413 L->addBasicBlockToLoop(New, *LI);
419 BasicBlock *LoopUnswitch::SplitEdge(BasicBlock *BB, BasicBlock *Succ) {
420 TerminatorInst *LatchTerm = BB->getTerminator();
421 unsigned SuccNum = 0;
422 for (unsigned i = 0, e = LatchTerm->getNumSuccessors(); ; ++i) {
423 assert(i != e && "Didn't find edge?");
424 if (LatchTerm->getSuccessor(i) == Succ) {
430 // If this is a critical edge, let SplitCriticalEdge do it.
431 if (SplitCriticalEdge(BB->getTerminator(), SuccNum, this))
432 return LatchTerm->getSuccessor(SuccNum);
434 // If the edge isn't critical, then BB has a single successor or Succ has a
435 // single pred. Split the block.
436 BasicBlock::iterator SplitPoint;
437 if (BasicBlock *SP = Succ->getSinglePredecessor()) {
438 // If the successor only has a single pred, split the top of the successor
440 assert(SP == BB && "CFG broken");
441 return SplitBlock(Succ, Succ->begin());
443 // Otherwise, if BB has a single successor, split it at the bottom of the
445 assert(BB->getTerminator()->getNumSuccessors() == 1 &&
446 "Should have a single succ!");
447 return SplitBlock(BB, BB->getTerminator());
453 // RemapInstruction - Convert the instruction operands from referencing the
454 // current values into those specified by ValueMap.
456 static inline void RemapInstruction(Instruction *I,
457 std::map<const Value *, Value*> &ValueMap) {
458 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
459 Value *Op = I->getOperand(op);
460 std::map<const Value *, Value*>::iterator It = ValueMap.find(Op);
461 if (It != ValueMap.end()) Op = It->second;
462 I->setOperand(op, Op);
466 /// CloneLoop - Recursively clone the specified loop and all of its children,
467 /// mapping the blocks with the specified map.
468 static Loop *CloneLoop(Loop *L, Loop *PL, std::map<const Value*, Value*> &VM,
470 Loop *New = new Loop();
473 PL->addChildLoop(New);
475 LI->addTopLevelLoop(New);
477 // Add all of the blocks in L to the new loop.
478 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
480 if (LI->getLoopFor(*I) == L)
481 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), *LI);
483 // Add all of the subloops to the new loop.
484 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
485 CloneLoop(*I, New, VM, LI);
490 /// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values
491 /// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the
492 /// code immediately before InsertPt.
493 static void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
494 BasicBlock *TrueDest,
495 BasicBlock *FalseDest,
496 Instruction *InsertPt) {
497 // Insert a conditional branch on LIC to the two preheaders. The original
498 // code is the true version and the new code is the false version.
499 Value *BranchVal = LIC;
500 if (!isa<ConstantBool>(Val)) {
501 BranchVal = BinaryOperator::createSetEQ(LIC, Val, "tmp", InsertPt);
502 } else if (Val != ConstantBool::True) {
503 // We want to enter the new loop when the condition is true.
504 std::swap(TrueDest, FalseDest);
507 // Insert the new branch.
508 new BranchInst(TrueDest, FalseDest, BranchVal, InsertPt);
512 /// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable
513 /// condition in it (a cond branch from its header block to its latch block,
514 /// where the path through the loop that doesn't execute its body has no
515 /// side-effects), unswitch it. This doesn't involve any code duplication, just
516 /// moving the conditional branch outside of the loop and updating loop info.
517 void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond,
518 Constant *Val, bool EntersWhenTrue,
519 BasicBlock *ExitBlock) {
520 DEBUG(std::cerr << "loop-unswitch: Trivial-Unswitch loop %"
521 << L->getHeader()->getName() << " [" << L->getBlocks().size()
522 << " blocks] in Function " << L->getHeader()->getParent()->getName()
523 << " on cond: " << *Val << (EntersWhenTrue ? " == " : " != ") <<
526 // First step, split the preheader, so that we know that there is a safe place
527 // to insert the conditional branch. We will change 'OrigPH' to have a
528 // conditional branch on Cond.
529 BasicBlock *OrigPH = L->getLoopPreheader();
530 BasicBlock *NewPH = SplitEdge(OrigPH, L->getHeader());
532 // Now that we have a place to insert the conditional branch, create a place
533 // to branch to: this is the exit block out of the loop that we should
536 // Split this block now, so that the loop maintains its exit block, and so
537 // that the jump from the preheader can execute the contents of the exit block
538 // without actually branching to it (the exit block should be dominated by the
539 // loop header, not the preheader).
540 assert(!L->contains(ExitBlock) && "Exit block is in the loop?");
541 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin());
543 // Okay, now we have a position to branch from and a position to branch to,
544 // insert the new conditional branch.
546 BasicBlock *TrueDest = NewPH, *FalseDest = NewExit;
547 if (!EntersWhenTrue) std::swap(TrueDest, FalseDest);
548 EmitPreheaderBranchOnCondition(Cond, Val, TrueDest, FalseDest,
549 OrigPH->getTerminator());
551 OrigPH->getTerminator()->eraseFromParent();
553 // Now that we know that the loop is never entered when this condition is a
554 // particular value, rewrite the loop with this info. We know that this will
555 // at least eliminate the old branch.
556 RewriteLoopBodyWithConditionConstant(L, Cond, Val, EntersWhenTrue);
561 /// VersionLoop - We determined that the loop is profitable to unswitch when LIC
562 /// equal Val. Split it into loop versions and test the condition outside of
563 /// either loop. Return the loops created as Out1/Out2.
564 void LoopUnswitch::VersionLoop(Value *LIC, Constant *Val, Loop *L,
565 Loop *&Out1, Loop *&Out2) {
566 Function *F = L->getHeader()->getParent();
568 DEBUG(std::cerr << "loop-unswitch: Unswitching loop %"
569 << L->getHeader()->getName() << " [" << L->getBlocks().size()
570 << " blocks] in Function " << F->getName()
571 << " when '" << *Val << "' == " << *LIC << "\n");
573 // LoopBlocks contains all of the basic blocks of the loop, including the
574 // preheader of the loop, the body of the loop, and the exit blocks of the
575 // loop, in that order.
576 std::vector<BasicBlock*> LoopBlocks;
578 // First step, split the preheader and exit blocks, and add these blocks to
579 // the LoopBlocks list.
580 BasicBlock *OrigPreheader = L->getLoopPreheader();
581 LoopBlocks.push_back(SplitEdge(OrigPreheader, L->getHeader()));
583 // We want the loop to come after the preheader, but before the exit blocks.
584 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
586 std::vector<BasicBlock*> ExitBlocks;
587 L->getExitBlocks(ExitBlocks);
588 std::sort(ExitBlocks.begin(), ExitBlocks.end());
589 ExitBlocks.erase(std::unique(ExitBlocks.begin(), ExitBlocks.end()),
592 // Split all of the edges from inside the loop to their exit blocks. This
593 // unswitching trivial: no phi nodes to update.
594 unsigned NumBlocks = L->getBlocks().size();
596 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
597 BasicBlock *ExitBlock = ExitBlocks[i];
598 std::vector<BasicBlock*> Preds(pred_begin(ExitBlock), pred_end(ExitBlock));
600 for (unsigned j = 0, e = Preds.size(); j != e; ++j) {
601 assert(L->contains(Preds[j]) &&
602 "All preds of loop exit blocks must be the same loop!");
603 SplitEdge(Preds[j], ExitBlock);
607 // The exit blocks may have been changed due to edge splitting, recompute.
609 L->getExitBlocks(ExitBlocks);
610 std::sort(ExitBlocks.begin(), ExitBlocks.end());
611 ExitBlocks.erase(std::unique(ExitBlocks.begin(), ExitBlocks.end()),
614 // Add exit blocks to the loop blocks.
615 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end());
617 // Next step, clone all of the basic blocks that make up the loop (including
618 // the loop preheader and exit blocks), keeping track of the mapping between
619 // the instructions and blocks.
620 std::vector<BasicBlock*> NewBlocks;
621 NewBlocks.reserve(LoopBlocks.size());
622 std::map<const Value*, Value*> ValueMap;
623 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
624 BasicBlock *New = CloneBasicBlock(LoopBlocks[i], ValueMap, ".us", F);
625 NewBlocks.push_back(New);
626 ValueMap[LoopBlocks[i]] = New; // Keep the BB mapping.
629 // Splice the newly inserted blocks into the function right before the
630 // original preheader.
631 F->getBasicBlockList().splice(LoopBlocks[0], F->getBasicBlockList(),
632 NewBlocks[0], F->end());
634 // Now we create the new Loop object for the versioned loop.
635 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), ValueMap, LI);
636 Loop *ParentLoop = L->getParentLoop();
638 // Make sure to add the cloned preheader and exit blocks to the parent loop
640 ParentLoop->addBasicBlockToLoop(NewBlocks[0], *LI);
643 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
644 BasicBlock *NewExit = cast<BasicBlock>(ValueMap[ExitBlocks[i]]);
645 // The new exit block should be in the same loop as the old one.
646 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i]))
647 ExitBBLoop->addBasicBlockToLoop(NewExit, *LI);
649 assert(NewExit->getTerminator()->getNumSuccessors() == 1 &&
650 "Exit block should have been split to have one successor!");
651 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0);
653 // If the successor of the exit block had PHI nodes, add an entry for
656 for (BasicBlock::iterator I = ExitSucc->begin();
657 (PN = dyn_cast<PHINode>(I)); ++I) {
658 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]);
659 std::map<const Value *, Value*>::iterator It = ValueMap.find(V);
660 if (It != ValueMap.end()) V = It->second;
661 PN->addIncoming(V, NewExit);
665 // Rewrite the code to refer to itself.
666 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i)
667 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
668 E = NewBlocks[i]->end(); I != E; ++I)
669 RemapInstruction(I, ValueMap);
671 // Rewrite the original preheader to select between versions of the loop.
672 BranchInst *OldBR = cast<BranchInst>(OrigPreheader->getTerminator());
673 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] &&
674 "Preheader splitting did not work correctly!");
676 // Emit the new branch that selects between the two versions of this loop.
677 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR);
678 OldBR->eraseFromParent();
680 // Now we rewrite the original code to know that the condition is true and the
681 // new code to know that the condition is false.
682 RewriteLoopBodyWithConditionConstant(L, LIC, Val, false);
683 RewriteLoopBodyWithConditionConstant(NewLoop, LIC, Val, true);
688 /// RemoveFromWorklist - Remove all instances of I from the worklist vector
690 static void RemoveFromWorklist(Instruction *I,
691 std::vector<Instruction*> &Worklist) {
692 std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(),
694 while (WI != Worklist.end()) {
695 unsigned Offset = WI-Worklist.begin();
697 WI = std::find(Worklist.begin()+Offset, Worklist.end(), I);
701 /// ReplaceUsesOfWith - When we find that I really equals V, remove I from the
702 /// program, replacing all uses with V and update the worklist.
703 static void ReplaceUsesOfWith(Instruction *I, Value *V,
704 std::vector<Instruction*> &Worklist) {
705 DEBUG(std::cerr << "Replace with '" << *V << "': " << *I);
707 // Add uses to the worklist, which may be dead now.
708 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
709 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
710 Worklist.push_back(Use);
712 // Add users to the worklist which may be simplified now.
713 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
715 Worklist.push_back(cast<Instruction>(*UI));
716 I->replaceAllUsesWith(V);
717 I->eraseFromParent();
718 RemoveFromWorklist(I, Worklist);
722 /// TryToRemoveEdge - Determine whether this is a case where we're smart enough
723 /// to remove the specified edge from the CFG and know how to update loop
724 /// information. If it is, update SSA and the loop information for the future
725 /// change, then return true. If not, return false.
726 bool LoopUnswitch::TryToRemoveEdge(TerminatorInst *TI, unsigned DeadSuccNo,
727 std::vector<Instruction*> &Worklist) {
728 BasicBlock *BB = TI->getParent(), *Succ = TI->getSuccessor(DeadSuccNo);
729 Loop *BBLoop = LI->getLoopFor(BB);
730 Loop *SuccLoop = LI->getLoopFor(Succ);
732 // If this edge is not in a loop, or if this edge is leaving a loop to a
733 // non-loop area, this is trivial.
735 Succ->removePredecessor(BB, true);
742 // RewriteLoopBodyWithConditionConstant - We know either that the value LIC has
743 // the value specified by Val in the specified loop, or we know it does NOT have
744 // that value. Rewrite any uses of LIC or of properties correlated to it.
745 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
748 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?");
750 // FIXME: Support correlated properties, like:
757 // NotVal - If Val is a bool, this contains its inverse.
758 Constant *NotVal = 0;
759 if (ConstantBool *CB = dyn_cast<ConstantBool>(Val))
760 NotVal = ConstantBool::get(!CB->getValue());
762 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches,
763 // selects, switches.
764 std::vector<User*> Users(LIC->use_begin(), LIC->use_end());
766 std::vector<Instruction*> Worklist;
768 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC
769 // in the loop with the appropriate one directly.
770 if (IsEqual || NotVal) {
771 Value *Replacement = NotVal ? NotVal : Val;
773 for (unsigned i = 0, e = Users.size(); i != e; ++i)
774 if (Instruction *U = cast<Instruction>(Users[i])) {
775 if (!L->contains(U->getParent()))
777 U->replaceUsesOfWith(LIC, Replacement);
778 Worklist.push_back(U);
781 // Otherwise, we don't know the precise value of LIC, but we do know that it
782 // is certainly NOT "Val". As such, simplify any uses in the loop that we
783 // can. This case occurs when we unswitch switch statements.
784 for (unsigned i = 0, e = Users.size(); i != e; ++i)
785 if (Instruction *U = cast<Instruction>(Users[i])) {
786 if (!L->contains(U->getParent()))
789 Worklist.push_back(U);
791 // If we know that LIC is not Val, use this info to simplify code.
792 if (SwitchInst *SI = dyn_cast<SwitchInst>(U)) {
793 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) {
794 if (SI->getCaseValue(i) == Val) {
795 // Found a dead case value. Don't remove PHI nodes in the
796 // successor if they become single-entry, those PHI nodes may
797 // be in the Users list.
798 SI->getSuccessor(i)->removePredecessor(SI->getParent(), true);
805 // TODO: We could do other simplifications, for example, turning
806 // LIC == Val -> false.
810 // Okay, now that we have simplified some instructions in the loop, walk over
811 // it and constant prop, dce, and fold control flow where possible. Note that
812 // this is effectively a very simple loop-structure-aware optimizer.
813 while (!Worklist.empty()) {
814 Instruction *I = Worklist.back();
817 // Simple constant folding.
818 if (Constant *C = ConstantFoldInstruction(I)) {
819 ReplaceUsesOfWith(I, C, Worklist);
824 if (isInstructionTriviallyDead(I)) {
825 DEBUG(std::cerr << "Remove dead instruction '" << *I);
827 // Add uses to the worklist, which may be dead now.
828 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
829 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
830 Worklist.push_back(Use);
831 I->eraseFromParent();
832 RemoveFromWorklist(I, Worklist);
837 // Special case hacks that appear commonly in unswitched code.
838 switch (I->getOpcode()) {
839 case Instruction::Select:
840 if (ConstantBool *CB = dyn_cast<ConstantBool>(I->getOperand(0))) {
841 ReplaceUsesOfWith(I, I->getOperand(!CB->getValue()+1), Worklist);
845 case Instruction::And:
846 if (isa<ConstantBool>(I->getOperand(0))) // constant -> RHS
847 cast<BinaryOperator>(I)->swapOperands();
848 if (ConstantBool *CB = dyn_cast<ConstantBool>(I->getOperand(1))) {
849 if (CB->getValue()) // X & 1 -> X
850 ReplaceUsesOfWith(I, I->getOperand(0), Worklist);
852 ReplaceUsesOfWith(I, I->getOperand(1), Worklist);
856 case Instruction::Or:
857 if (isa<ConstantBool>(I->getOperand(0))) // constant -> RHS
858 cast<BinaryOperator>(I)->swapOperands();
859 if (ConstantBool *CB = dyn_cast<ConstantBool>(I->getOperand(1))) {
860 if (CB->getValue()) // X | 1 -> 1
861 ReplaceUsesOfWith(I, I->getOperand(1), Worklist);
863 ReplaceUsesOfWith(I, I->getOperand(0), Worklist);
867 case Instruction::Br: {
868 BranchInst *BI = cast<BranchInst>(I);
869 if (BI->isUnconditional()) {
870 // If BI's parent is the only pred of the successor, fold the two blocks
872 BasicBlock *Pred = BI->getParent();
873 BasicBlock *Succ = BI->getSuccessor(0);
874 BasicBlock *SinglePred = Succ->getSinglePredecessor();
875 if (!SinglePred) continue; // Nothing to do.
876 assert(SinglePred == Pred && "CFG broken");
878 DEBUG(std::cerr << "Merging blocks: " << Pred->getName() << " <- "
879 << Succ->getName() << "\n");
881 // Resolve any single entry PHI nodes in Succ.
882 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin()))
883 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist);
885 // Move all of the successor contents from Succ to Pred.
886 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(),
888 BI->eraseFromParent();
889 RemoveFromWorklist(BI, Worklist);
891 // If Succ has any successors with PHI nodes, update them to have
892 // entries coming from Pred instead of Succ.
893 Succ->replaceAllUsesWith(Pred);
895 // Remove Succ from the loop tree.
896 LI->removeBlock(Succ);
897 Succ->eraseFromParent();
900 } else if (ConstantBool *CB = dyn_cast<ConstantBool>(BI->getCondition())){
901 // Conditional branch.
902 if (TryToRemoveEdge(BI, CB->getValue(), Worklist)) {
903 DEBUG(std::cerr << "Folded branch: " << *BI);
904 new BranchInst(BI->getSuccessor(!CB->getValue()), BI);
905 BI->eraseFromParent();
906 RemoveFromWorklist(BI, Worklist);