fix more AST updating bugs, correcting miscompilation in PR8041
[llvm.git] / lib / Transforms / Scalar / LoopUnswitch.cpp
blob9afe428ba5691ae4a3d5a7e0d62d91d563ae8df4
1 //===-- LoopUnswitch.cpp - Hoist loop-invariant conditionals in loop ------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This pass transforms loops that contain branches on loop-invariant conditions
11 // to have multiple loops. For example, it turns the left into the right code:
13 // for (...) if (lic)
14 // A for (...)
15 // if (lic) A; B; C
16 // B else
17 // C for (...)
18 // A; C
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/InlineCost.h"
37 #include "llvm/Analysis/InstructionSimplify.h"
38 #include "llvm/Analysis/LoopInfo.h"
39 #include "llvm/Analysis/LoopPass.h"
40 #include "llvm/Analysis/Dominators.h"
41 #include "llvm/Transforms/Utils/Cloning.h"
42 #include "llvm/Transforms/Utils/Local.h"
43 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
44 #include "llvm/ADT/Statistic.h"
45 #include "llvm/ADT/SmallPtrSet.h"
46 #include "llvm/ADT/STLExtras.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/raw_ostream.h"
50 #include <algorithm>
51 #include <set>
52 using namespace llvm;
54 STATISTIC(NumBranches, "Number of branches unswitched");
55 STATISTIC(NumSwitches, "Number of switches unswitched");
56 STATISTIC(NumSelects , "Number of selects unswitched");
57 STATISTIC(NumTrivial , "Number of unswitches that are trivial");
58 STATISTIC(NumSimplify, "Number of simplifications of unswitched code");
60 // The specific value of 50 here was chosen based only on intuition and a
61 // few specific examples.
62 static cl::opt<unsigned>
63 Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"),
64 cl::init(50), cl::Hidden);
66 namespace {
67 class LoopUnswitch : public LoopPass {
68 LoopInfo *LI; // Loop information
69 LPPassManager *LPM;
71 // LoopProcessWorklist - Used to check if second loop needs processing
72 // after RewriteLoopBodyWithConditionConstant rewrites first loop.
73 std::vector<Loop*> LoopProcessWorklist;
74 SmallPtrSet<Value *,8> UnswitchedVals;
76 bool OptimizeForSize;
77 bool redoLoop;
79 Loop *currentLoop;
80 DominatorTree *DT;
81 BasicBlock *loopHeader;
82 BasicBlock *loopPreheader;
84 // LoopBlocks contains all of the basic blocks of the loop, including the
85 // preheader of the loop, the body of the loop, and the exit blocks of the
86 // loop, in that order.
87 std::vector<BasicBlock*> LoopBlocks;
88 // NewBlocks contained cloned copy of basic blocks from LoopBlocks.
89 std::vector<BasicBlock*> NewBlocks;
91 public:
92 static char ID; // Pass ID, replacement for typeid
93 explicit LoopUnswitch(bool Os = false) :
94 LoopPass(ID), OptimizeForSize(Os), redoLoop(false),
95 currentLoop(NULL), DT(NULL), loopHeader(NULL),
96 loopPreheader(NULL) {}
98 bool runOnLoop(Loop *L, LPPassManager &LPM);
99 bool processCurrentLoop();
101 /// This transformation requires natural loop information & requires that
102 /// loop preheaders be inserted into the CFG.
104 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
105 AU.addRequiredID(LoopSimplifyID);
106 AU.addPreservedID(LoopSimplifyID);
107 AU.addRequired<LoopInfo>();
108 AU.addPreserved<LoopInfo>();
109 AU.addRequiredID(LCSSAID);
110 AU.addPreservedID(LCSSAID);
111 AU.addPreserved<DominatorTree>();
114 private:
116 virtual void releaseMemory() {
117 UnswitchedVals.clear();
120 /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist,
121 /// remove it.
122 void RemoveLoopFromWorklist(Loop *L) {
123 std::vector<Loop*>::iterator I = std::find(LoopProcessWorklist.begin(),
124 LoopProcessWorklist.end(), L);
125 if (I != LoopProcessWorklist.end())
126 LoopProcessWorklist.erase(I);
129 void initLoopData() {
130 loopHeader = currentLoop->getHeader();
131 loopPreheader = currentLoop->getLoopPreheader();
134 /// Split all of the edges from inside the loop to their exit blocks.
135 /// Update the appropriate Phi nodes as we do so.
136 void SplitExitEdges(Loop *L, const SmallVector<BasicBlock *, 8> &ExitBlocks);
138 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val);
139 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val,
140 BasicBlock *ExitBlock);
141 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L);
143 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
144 Constant *Val, bool isEqual);
146 void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
147 BasicBlock *TrueDest,
148 BasicBlock *FalseDest,
149 Instruction *InsertPt);
151 void SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L);
152 void RemoveBlockIfDead(BasicBlock *BB,
153 std::vector<Instruction*> &Worklist, Loop *l);
154 void RemoveLoopFromHierarchy(Loop *L);
155 bool IsTrivialUnswitchCondition(Value *Cond, Constant **Val = 0,
156 BasicBlock **LoopExit = 0);
160 char LoopUnswitch::ID = 0;
161 INITIALIZE_PASS(LoopUnswitch, "loop-unswitch", "Unswitch loops", false, false);
163 Pass *llvm::createLoopUnswitchPass(bool Os) {
164 return new LoopUnswitch(Os);
167 /// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is
168 /// invariant in the loop, or has an invariant piece, return the invariant.
169 /// Otherwise, return null.
170 static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) {
171 // We can never unswitch on vector conditions.
172 if (Cond->getType()->isVectorTy())
173 return 0;
175 // Constants should be folded, not unswitched on!
176 if (isa<Constant>(Cond)) return 0;
178 // TODO: Handle: br (VARIANT|INVARIANT).
180 // Hoist simple values out.
181 if (L->makeLoopInvariant(Cond, Changed))
182 return Cond;
184 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond))
185 if (BO->getOpcode() == Instruction::And ||
186 BO->getOpcode() == Instruction::Or) {
187 // If either the left or right side is invariant, we can unswitch on this,
188 // which will cause the branch to go away in one loop and the condition to
189 // simplify in the other one.
190 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed))
191 return LHS;
192 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed))
193 return RHS;
196 return 0;
199 bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) {
200 LI = &getAnalysis<LoopInfo>();
201 LPM = &LPM_Ref;
202 DT = getAnalysisIfAvailable<DominatorTree>();
203 currentLoop = L;
204 Function *F = currentLoop->getHeader()->getParent();
205 bool Changed = false;
206 do {
207 assert(currentLoop->isLCSSAForm(*DT));
208 redoLoop = false;
209 Changed |= processCurrentLoop();
210 } while(redoLoop);
212 if (Changed) {
213 // FIXME: Reconstruct dom info, because it is not preserved properly.
214 if (DT)
215 DT->runOnFunction(*F);
217 return Changed;
220 /// processCurrentLoop - Do actual work and unswitch loop if possible
221 /// and profitable.
222 bool LoopUnswitch::processCurrentLoop() {
223 bool Changed = false;
224 LLVMContext &Context = currentLoop->getHeader()->getContext();
226 // Loop over all of the basic blocks in the loop. If we find an interior
227 // block that is branching on a loop-invariant condition, we can unswitch this
228 // loop.
229 for (Loop::block_iterator I = currentLoop->block_begin(),
230 E = currentLoop->block_end(); I != E; ++I) {
231 TerminatorInst *TI = (*I)->getTerminator();
232 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
233 // If this isn't branching on an invariant condition, we can't unswitch
234 // it.
235 if (BI->isConditional()) {
236 // See if this, or some part of it, is loop invariant. If so, we can
237 // unswitch on it if we desire.
238 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(),
239 currentLoop, Changed);
240 if (LoopCond && UnswitchIfProfitable(LoopCond,
241 ConstantInt::getTrue(Context))) {
242 ++NumBranches;
243 return true;
246 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
247 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
248 currentLoop, Changed);
249 if (LoopCond && SI->getNumCases() > 1) {
250 // Find a value to unswitch on:
251 // FIXME: this should chose the most expensive case!
252 Constant *UnswitchVal = SI->getCaseValue(1);
253 // Do not process same value again and again.
254 if (!UnswitchedVals.insert(UnswitchVal))
255 continue;
257 if (UnswitchIfProfitable(LoopCond, UnswitchVal)) {
258 ++NumSwitches;
259 return true;
264 // Scan the instructions to check for unswitchable values.
265 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end();
266 BBI != E; ++BBI)
267 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) {
268 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
269 currentLoop, Changed);
270 if (LoopCond && UnswitchIfProfitable(LoopCond,
271 ConstantInt::getTrue(Context))) {
272 ++NumSelects;
273 return true;
277 return Changed;
280 /// isTrivialLoopExitBlock - Check to see if all paths from BB exit the
281 /// loop with no side effects (including infinite loops).
283 /// If true, we return true and set ExitBB to the block we
284 /// exit through.
286 static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB,
287 BasicBlock *&ExitBB,
288 std::set<BasicBlock*> &Visited) {
289 if (!Visited.insert(BB).second) {
290 // Already visited. Without more analysis, this could indicate an infinte loop.
291 return false;
292 } else if (!L->contains(BB)) {
293 // Otherwise, this is a loop exit, this is fine so long as this is the
294 // first exit.
295 if (ExitBB != 0) return false;
296 ExitBB = BB;
297 return true;
300 // Otherwise, this is an unvisited intra-loop node. Check all successors.
301 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) {
302 // Check to see if the successor is a trivial loop exit.
303 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited))
304 return false;
307 // Okay, everything after this looks good, check to make sure that this block
308 // doesn't include any side effects.
309 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
310 if (I->mayHaveSideEffects())
311 return false;
313 return true;
316 /// isTrivialLoopExitBlock - Return true if the specified block unconditionally
317 /// leads to an exit from the specified loop, and has no side-effects in the
318 /// process. If so, return the block that is exited to, otherwise return null.
319 static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) {
320 std::set<BasicBlock*> Visited;
321 Visited.insert(L->getHeader()); // Branches to header make infinite loops.
322 BasicBlock *ExitBB = 0;
323 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited))
324 return ExitBB;
325 return 0;
328 /// IsTrivialUnswitchCondition - Check to see if this unswitch condition is
329 /// trivial: that is, that the condition controls whether or not the loop does
330 /// anything at all. If this is a trivial condition, unswitching produces no
331 /// code duplications (equivalently, it produces a simpler loop and a new empty
332 /// loop, which gets deleted).
334 /// If this is a trivial condition, return true, otherwise return false. When
335 /// returning true, this sets Cond and Val to the condition that controls the
336 /// trivial condition: when Cond dynamically equals Val, the loop is known to
337 /// exit. Finally, this sets LoopExit to the BB that the loop exits to when
338 /// Cond == Val.
340 bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val,
341 BasicBlock **LoopExit) {
342 BasicBlock *Header = currentLoop->getHeader();
343 TerminatorInst *HeaderTerm = Header->getTerminator();
344 LLVMContext &Context = Header->getContext();
346 BasicBlock *LoopExitBB = 0;
347 if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) {
348 // If the header block doesn't end with a conditional branch on Cond, we
349 // can't handle it.
350 if (!BI->isConditional() || BI->getCondition() != Cond)
351 return false;
353 // Check to see if a successor of the branch is guaranteed to
354 // exit through a unique exit block without having any
355 // side-effects. If so, determine the value of Cond that causes it to do
356 // this.
357 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
358 BI->getSuccessor(0)))) {
359 if (Val) *Val = ConstantInt::getTrue(Context);
360 } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
361 BI->getSuccessor(1)))) {
362 if (Val) *Val = ConstantInt::getFalse(Context);
364 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) {
365 // If this isn't a switch on Cond, we can't handle it.
366 if (SI->getCondition() != Cond) return false;
368 // Check to see if a successor of the switch is guaranteed to go to the
369 // latch block or exit through a one exit block without having any
370 // side-effects. If so, determine the value of Cond that causes it to do
371 // this. Note that we can't trivially unswitch on the default case.
372 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
373 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
374 SI->getSuccessor(i)))) {
375 // Okay, we found a trivial case, remember the value that is trivial.
376 if (Val) *Val = SI->getCaseValue(i);
377 break;
381 // If we didn't find a single unique LoopExit block, or if the loop exit block
382 // contains phi nodes, this isn't trivial.
383 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin()))
384 return false; // Can't handle this.
386 if (LoopExit) *LoopExit = LoopExitBB;
388 // We already know that nothing uses any scalar values defined inside of this
389 // loop. As such, we just have to check to see if this loop will execute any
390 // side-effecting instructions (e.g. stores, calls, volatile loads) in the
391 // part of the loop that the code *would* execute. We already checked the
392 // tail, check the header now.
393 for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I)
394 if (I->mayHaveSideEffects())
395 return false;
396 return true;
399 /// UnswitchIfProfitable - We have found that we can unswitch currentLoop when
400 /// LoopCond == Val to simplify the loop. If we decide that this is profitable,
401 /// unswitch the loop, reprocess the pieces, then return true.
402 bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val) {
404 initLoopData();
406 // If LoopSimplify was unable to form a preheader, don't do any unswitching.
407 if (!loopPreheader)
408 return false;
410 Function *F = loopHeader->getParent();
412 Constant *CondVal = 0;
413 BasicBlock *ExitBlock = 0;
414 if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) {
415 // If the condition is trivial, always unswitch. There is no code growth
416 // for this case.
417 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock);
418 return true;
421 // Check to see if it would be profitable to unswitch current loop.
423 // Do not do non-trivial unswitch while optimizing for size.
424 if (OptimizeForSize || F->hasFnAttr(Attribute::OptimizeForSize))
425 return false;
427 // FIXME: This is overly conservative because it does not take into
428 // consideration code simplification opportunities and code that can
429 // be shared by the resultant unswitched loops.
430 CodeMetrics Metrics;
431 for (Loop::block_iterator I = currentLoop->block_begin(),
432 E = currentLoop->block_end();
433 I != E; ++I)
434 Metrics.analyzeBasicBlock(*I);
436 // Limit the number of instructions to avoid causing significant code
437 // expansion, and the number of basic blocks, to avoid loops with
438 // large numbers of branches which cause loop unswitching to go crazy.
439 // This is a very ad-hoc heuristic.
440 if (Metrics.NumInsts > Threshold ||
441 Metrics.NumBlocks * 5 > Threshold ||
442 Metrics.containsIndirectBr || Metrics.isRecursive) {
443 DEBUG(dbgs() << "NOT unswitching loop %"
444 << currentLoop->getHeader()->getName() << ", cost too high: "
445 << currentLoop->getBlocks().size() << "\n");
446 return false;
449 UnswitchNontrivialCondition(LoopCond, Val, currentLoop);
450 return true;
453 // RemapInstruction - Convert the instruction operands from referencing the
454 // current values into those specified by VMap.
456 static inline void RemapInstruction(Instruction *I,
457 ValueMap<const Value *, Value*> &VMap) {
458 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
459 Value *Op = I->getOperand(op);
460 ValueMap<const Value *, Value*>::iterator It = VMap.find(Op);
461 if (It != VMap.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, ValueMap<const Value*, Value*> &VM,
469 LoopInfo *LI, LPPassManager *LPM) {
470 Loop *New = new Loop();
471 LPM->insertLoop(New, PL);
473 // Add all of the blocks in L to the new loop.
474 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
475 I != E; ++I)
476 if (LI->getLoopFor(*I) == L)
477 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), LI->getBase());
479 // Add all of the subloops to the new loop.
480 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
481 CloneLoop(*I, New, VM, LI, LPM);
483 return New;
486 /// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values
487 /// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the
488 /// code immediately before InsertPt.
489 void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
490 BasicBlock *TrueDest,
491 BasicBlock *FalseDest,
492 Instruction *InsertPt) {
493 // Insert a conditional branch on LIC to the two preheaders. The original
494 // code is the true version and the new code is the false version.
495 Value *BranchVal = LIC;
496 if (!isa<ConstantInt>(Val) ||
497 Val->getType() != Type::getInt1Ty(LIC->getContext()))
498 BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val, "tmp");
499 else if (Val != ConstantInt::getTrue(Val->getContext()))
500 // We want to enter the new loop when the condition is true.
501 std::swap(TrueDest, FalseDest);
503 // Insert the new branch.
504 BranchInst *BI = BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt);
506 // If either edge is critical, split it. This helps preserve LoopSimplify
507 // form for enclosing loops.
508 SplitCriticalEdge(BI, 0, this);
509 SplitCriticalEdge(BI, 1, this);
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,
519 BasicBlock *ExitBlock) {
520 DEBUG(dbgs() << "loop-unswitch: Trivial-Unswitch loop %"
521 << loopHeader->getName() << " [" << L->getBlocks().size()
522 << " blocks] in Function " << L->getHeader()->getParent()->getName()
523 << " on cond: " << *Val << " == " << *Cond << "\n");
525 // First step, split the preheader, so that we know that there is a safe place
526 // to insert the conditional branch. We will change loopPreheader to have a
527 // conditional branch on Cond.
528 BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, this);
530 // Now that we have a place to insert the conditional branch, create a place
531 // to branch to: this is the exit block out of the loop that we should
532 // short-circuit to.
534 // Split this block now, so that the loop maintains its exit block, and so
535 // that the jump from the preheader can execute the contents of the exit block
536 // without actually branching to it (the exit block should be dominated by the
537 // loop header, not the preheader).
538 assert(!L->contains(ExitBlock) && "Exit block is in the loop?");
539 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), this);
541 // Okay, now we have a position to branch from and a position to branch to,
542 // insert the new conditional branch.
543 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH,
544 loopPreheader->getTerminator());
545 LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L);
546 loopPreheader->getTerminator()->eraseFromParent();
548 // We need to reprocess this loop, it could be unswitched again.
549 redoLoop = true;
551 // Now that we know that the loop is never entered when this condition is a
552 // particular value, rewrite the loop with this info. We know that this will
553 // at least eliminate the old branch.
554 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false);
555 ++NumTrivial;
558 /// SplitExitEdges - Split all of the edges from inside the loop to their exit
559 /// blocks. Update the appropriate Phi nodes as we do so.
560 void LoopUnswitch::SplitExitEdges(Loop *L,
561 const SmallVector<BasicBlock *, 8> &ExitBlocks){
563 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
564 BasicBlock *ExitBlock = ExitBlocks[i];
565 SmallVector<BasicBlock *, 4> Preds(pred_begin(ExitBlock),
566 pred_end(ExitBlock));
567 SplitBlockPredecessors(ExitBlock, Preds.data(), Preds.size(),
568 ".us-lcssa", this);
572 /// UnswitchNontrivialCondition - We determined that the loop is profitable
573 /// to unswitch when LIC equal Val. Split it into loop versions and test the
574 /// condition outside of either loop. Return the loops created as Out1/Out2.
575 void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val,
576 Loop *L) {
577 Function *F = loopHeader->getParent();
578 DEBUG(dbgs() << "loop-unswitch: Unswitching loop %"
579 << loopHeader->getName() << " [" << L->getBlocks().size()
580 << " blocks] in Function " << F->getName()
581 << " when '" << *Val << "' == " << *LIC << "\n");
583 LoopBlocks.clear();
584 NewBlocks.clear();
586 // First step, split the preheader and exit blocks, and add these blocks to
587 // the LoopBlocks list.
588 BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, this);
589 LoopBlocks.push_back(NewPreheader);
591 // We want the loop to come after the preheader, but before the exit blocks.
592 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
594 SmallVector<BasicBlock*, 8> ExitBlocks;
595 L->getUniqueExitBlocks(ExitBlocks);
597 // Split all of the edges from inside the loop to their exit blocks. Update
598 // the appropriate Phi nodes as we do so.
599 SplitExitEdges(L, ExitBlocks);
601 // The exit blocks may have been changed due to edge splitting, recompute.
602 ExitBlocks.clear();
603 L->getUniqueExitBlocks(ExitBlocks);
605 // Add exit blocks to the loop blocks.
606 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end());
608 // Next step, clone all of the basic blocks that make up the loop (including
609 // the loop preheader and exit blocks), keeping track of the mapping between
610 // the instructions and blocks.
611 NewBlocks.reserve(LoopBlocks.size());
612 ValueMap<const Value*, Value*> VMap;
613 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
614 BasicBlock *NewBB = CloneBasicBlock(LoopBlocks[i], VMap, ".us", F);
615 NewBlocks.push_back(NewBB);
616 VMap[LoopBlocks[i]] = NewBB; // Keep the BB mapping.
617 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], NewBB, L);
620 // Splice the newly inserted blocks into the function right before the
621 // original preheader.
622 F->getBasicBlockList().splice(NewPreheader, F->getBasicBlockList(),
623 NewBlocks[0], F->end());
625 // Now we create the new Loop object for the versioned loop.
626 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), VMap, LI, LPM);
627 Loop *ParentLoop = L->getParentLoop();
628 if (ParentLoop) {
629 // Make sure to add the cloned preheader and exit blocks to the parent loop
630 // as well.
631 ParentLoop->addBasicBlockToLoop(NewBlocks[0], LI->getBase());
634 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
635 BasicBlock *NewExit = cast<BasicBlock>(VMap[ExitBlocks[i]]);
636 // The new exit block should be in the same loop as the old one.
637 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i]))
638 ExitBBLoop->addBasicBlockToLoop(NewExit, LI->getBase());
640 assert(NewExit->getTerminator()->getNumSuccessors() == 1 &&
641 "Exit block should have been split to have one successor!");
642 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0);
644 // If the successor of the exit block had PHI nodes, add an entry for
645 // NewExit.
646 PHINode *PN;
647 for (BasicBlock::iterator I = ExitSucc->begin(); isa<PHINode>(I); ++I) {
648 PN = cast<PHINode>(I);
649 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]);
650 ValueMap<const Value *, Value*>::iterator It = VMap.find(V);
651 if (It != VMap.end()) V = It->second;
652 PN->addIncoming(V, NewExit);
656 // Rewrite the code to refer to itself.
657 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i)
658 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
659 E = NewBlocks[i]->end(); I != E; ++I)
660 RemapInstruction(I, VMap);
662 // Rewrite the original preheader to select between versions of the loop.
663 BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator());
664 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] &&
665 "Preheader splitting did not work correctly!");
667 // Emit the new branch that selects between the two versions of this loop.
668 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR);
669 LPM->deleteSimpleAnalysisValue(OldBR, L);
670 OldBR->eraseFromParent();
672 LoopProcessWorklist.push_back(NewLoop);
673 redoLoop = true;
675 // Keep a WeakVH holding onto LIC. If the first call to RewriteLoopBody
676 // deletes the instruction (for example by simplifying a PHI that feeds into
677 // the condition that we're unswitching on), we don't rewrite the second
678 // iteration.
679 WeakVH LICHandle(LIC);
681 // Now we rewrite the original code to know that the condition is true and the
682 // new code to know that the condition is false.
683 RewriteLoopBodyWithConditionConstant(L, LIC, Val, false);
685 // It's possible that simplifying one loop could cause the other to be
686 // changed to another value or a constant. If its a constant, don't simplify
687 // it.
688 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop &&
689 LICHandle && !isa<Constant>(LICHandle))
690 RewriteLoopBodyWithConditionConstant(NewLoop, LICHandle, Val, true);
693 /// RemoveFromWorklist - Remove all instances of I from the worklist vector
694 /// specified.
695 static void RemoveFromWorklist(Instruction *I,
696 std::vector<Instruction*> &Worklist) {
697 std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(),
698 Worklist.end(), I);
699 while (WI != Worklist.end()) {
700 unsigned Offset = WI-Worklist.begin();
701 Worklist.erase(WI);
702 WI = std::find(Worklist.begin()+Offset, Worklist.end(), I);
706 /// ReplaceUsesOfWith - When we find that I really equals V, remove I from the
707 /// program, replacing all uses with V and update the worklist.
708 static void ReplaceUsesOfWith(Instruction *I, Value *V,
709 std::vector<Instruction*> &Worklist,
710 Loop *L, LPPassManager *LPM) {
711 DEBUG(dbgs() << "Replace with '" << *V << "': " << *I);
713 // Add uses to the worklist, which may be dead now.
714 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
715 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
716 Worklist.push_back(Use);
718 // Add users to the worklist which may be simplified now.
719 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
720 UI != E; ++UI)
721 Worklist.push_back(cast<Instruction>(*UI));
722 LPM->deleteSimpleAnalysisValue(I, L);
723 RemoveFromWorklist(I, Worklist);
724 I->replaceAllUsesWith(V);
725 I->eraseFromParent();
726 ++NumSimplify;
729 /// RemoveBlockIfDead - If the specified block is dead, remove it, update loop
730 /// information, and remove any dead successors it has.
732 void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB,
733 std::vector<Instruction*> &Worklist,
734 Loop *L) {
735 if (pred_begin(BB) != pred_end(BB)) {
736 // This block isn't dead, since an edge to BB was just removed, see if there
737 // are any easy simplifications we can do now.
738 if (BasicBlock *Pred = BB->getSinglePredecessor()) {
739 // If it has one pred, fold phi nodes in BB.
740 while (isa<PHINode>(BB->begin()))
741 ReplaceUsesOfWith(BB->begin(),
742 cast<PHINode>(BB->begin())->getIncomingValue(0),
743 Worklist, L, LPM);
745 // If this is the header of a loop and the only pred is the latch, we now
746 // have an unreachable loop.
747 if (Loop *L = LI->getLoopFor(BB))
748 if (loopHeader == BB && L->contains(Pred)) {
749 // Remove the branch from the latch to the header block, this makes
750 // the header dead, which will make the latch dead (because the header
751 // dominates the latch).
752 LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L);
753 Pred->getTerminator()->eraseFromParent();
754 new UnreachableInst(BB->getContext(), Pred);
756 // The loop is now broken, remove it from LI.
757 RemoveLoopFromHierarchy(L);
759 // Reprocess the header, which now IS dead.
760 RemoveBlockIfDead(BB, Worklist, L);
761 return;
764 // If pred ends in a uncond branch, add uncond branch to worklist so that
765 // the two blocks will get merged.
766 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator()))
767 if (BI->isUnconditional())
768 Worklist.push_back(BI);
770 return;
773 DEBUG(dbgs() << "Nuking dead block: " << *BB);
775 // Remove the instructions in the basic block from the worklist.
776 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
777 RemoveFromWorklist(I, Worklist);
779 // Anything that uses the instructions in this basic block should have their
780 // uses replaced with undefs.
781 // If I is not void type then replaceAllUsesWith undef.
782 // This allows ValueHandlers and custom metadata to adjust itself.
783 if (!I->getType()->isVoidTy())
784 I->replaceAllUsesWith(UndefValue::get(I->getType()));
787 // If this is the edge to the header block for a loop, remove the loop and
788 // promote all subloops.
789 if (Loop *BBLoop = LI->getLoopFor(BB)) {
790 if (BBLoop->getLoopLatch() == BB)
791 RemoveLoopFromHierarchy(BBLoop);
794 // Remove the block from the loop info, which removes it from any loops it
795 // was in.
796 LI->removeBlock(BB);
799 // Remove phi node entries in successors for this block.
800 TerminatorInst *TI = BB->getTerminator();
801 SmallVector<BasicBlock*, 4> Succs;
802 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
803 Succs.push_back(TI->getSuccessor(i));
804 TI->getSuccessor(i)->removePredecessor(BB);
807 // Unique the successors, remove anything with multiple uses.
808 array_pod_sort(Succs.begin(), Succs.end());
809 Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end());
811 // Remove the basic block, including all of the instructions contained in it.
812 LPM->deleteSimpleAnalysisValue(BB, L);
813 BB->eraseFromParent();
814 // Remove successor blocks here that are not dead, so that we know we only
815 // have dead blocks in this list. Nondead blocks have a way of becoming dead,
816 // then getting removed before we revisit them, which is badness.
818 for (unsigned i = 0; i != Succs.size(); ++i)
819 if (pred_begin(Succs[i]) != pred_end(Succs[i])) {
820 // One exception is loop headers. If this block was the preheader for a
821 // loop, then we DO want to visit the loop so the loop gets deleted.
822 // We know that if the successor is a loop header, that this loop had to
823 // be the preheader: the case where this was the latch block was handled
824 // above and headers can only have two predecessors.
825 if (!LI->isLoopHeader(Succs[i])) {
826 Succs.erase(Succs.begin()+i);
827 --i;
831 for (unsigned i = 0, e = Succs.size(); i != e; ++i)
832 RemoveBlockIfDead(Succs[i], Worklist, L);
835 /// RemoveLoopFromHierarchy - We have discovered that the specified loop has
836 /// become unwrapped, either because the backedge was deleted, or because the
837 /// edge into the header was removed. If the edge into the header from the
838 /// latch block was removed, the loop is unwrapped but subloops are still alive,
839 /// so they just reparent loops. If the loops are actually dead, they will be
840 /// removed later.
841 void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) {
842 LPM->deleteLoopFromQueue(L);
843 RemoveLoopFromWorklist(L);
846 // RewriteLoopBodyWithConditionConstant - We know either that the value LIC has
847 // the value specified by Val in the specified loop, or we know it does NOT have
848 // that value. Rewrite any uses of LIC or of properties correlated to it.
849 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
850 Constant *Val,
851 bool IsEqual) {
852 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?");
854 // FIXME: Support correlated properties, like:
855 // for (...)
856 // if (li1 < li2)
857 // ...
858 // if (li1 > li2)
859 // ...
861 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches,
862 // selects, switches.
863 std::vector<User*> Users(LIC->use_begin(), LIC->use_end());
864 std::vector<Instruction*> Worklist;
865 LLVMContext &Context = Val->getContext();
868 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC
869 // in the loop with the appropriate one directly.
870 if (IsEqual || (isa<ConstantInt>(Val) &&
871 Val->getType()->isIntegerTy(1))) {
872 Value *Replacement;
873 if (IsEqual)
874 Replacement = Val;
875 else
876 Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()),
877 !cast<ConstantInt>(Val)->getZExtValue());
879 for (unsigned i = 0, e = Users.size(); i != e; ++i)
880 if (Instruction *U = cast<Instruction>(Users[i])) {
881 if (!L->contains(U))
882 continue;
883 U->replaceUsesOfWith(LIC, Replacement);
884 Worklist.push_back(U);
886 SimplifyCode(Worklist, L);
887 return;
890 // Otherwise, we don't know the precise value of LIC, but we do know that it
891 // is certainly NOT "Val". As such, simplify any uses in the loop that we
892 // can. This case occurs when we unswitch switch statements.
893 for (unsigned i = 0, e = Users.size(); i != e; ++i) {
894 Instruction *U = cast<Instruction>(Users[i]);
895 if (!L->contains(U))
896 continue;
898 Worklist.push_back(U);
900 // TODO: We could do other simplifications, for example, turning
901 // 'icmp eq LIC, Val' -> false.
903 // If we know that LIC is not Val, use this info to simplify code.
904 SwitchInst *SI = dyn_cast<SwitchInst>(U);
905 if (SI == 0 || !isa<ConstantInt>(Val)) continue;
907 unsigned DeadCase = SI->findCaseValue(cast<ConstantInt>(Val));
908 if (DeadCase == 0) continue; // Default case is live for multiple values.
910 // Found a dead case value. Don't remove PHI nodes in the
911 // successor if they become single-entry, those PHI nodes may
912 // be in the Users list.
914 // FIXME: This is a hack. We need to keep the successor around
915 // and hooked up so as to preserve the loop structure, because
916 // trying to update it is complicated. So instead we preserve the
917 // loop structure and put the block on a dead code path.
918 BasicBlock *Switch = SI->getParent();
919 SplitEdge(Switch, SI->getSuccessor(DeadCase), this);
920 // Compute the successors instead of relying on the return value
921 // of SplitEdge, since it may have split the switch successor
922 // after PHI nodes.
923 BasicBlock *NewSISucc = SI->getSuccessor(DeadCase);
924 BasicBlock *OldSISucc = *succ_begin(NewSISucc);
925 // Create an "unreachable" destination.
926 BasicBlock *Abort = BasicBlock::Create(Context, "us-unreachable",
927 Switch->getParent(),
928 OldSISucc);
929 new UnreachableInst(Context, Abort);
930 // Force the new case destination to branch to the "unreachable"
931 // block while maintaining a (dead) CFG edge to the old block.
932 NewSISucc->getTerminator()->eraseFromParent();
933 BranchInst::Create(Abort, OldSISucc,
934 ConstantInt::getTrue(Context), NewSISucc);
935 // Release the PHI operands for this edge.
936 for (BasicBlock::iterator II = NewSISucc->begin();
937 PHINode *PN = dyn_cast<PHINode>(II); ++II)
938 PN->setIncomingValue(PN->getBasicBlockIndex(Switch),
939 UndefValue::get(PN->getType()));
940 // Tell the domtree about the new block. We don't fully update the
941 // domtree here -- instead we force it to do a full recomputation
942 // after the pass is complete -- but we do need to inform it of
943 // new blocks.
944 if (DT)
945 DT->addNewBlock(Abort, NewSISucc);
948 SimplifyCode(Worklist, L);
951 /// SimplifyCode - Okay, now that we have simplified some instructions in the
952 /// loop, walk over it and constant prop, dce, and fold control flow where
953 /// possible. Note that this is effectively a very simple loop-structure-aware
954 /// optimizer. During processing of this loop, L could very well be deleted, so
955 /// it must not be used.
957 /// FIXME: When the loop optimizer is more mature, separate this out to a new
958 /// pass.
960 void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) {
961 while (!Worklist.empty()) {
962 Instruction *I = Worklist.back();
963 Worklist.pop_back();
965 // Simple constant folding.
966 if (Constant *C = ConstantFoldInstruction(I)) {
967 ReplaceUsesOfWith(I, C, Worklist, L, LPM);
968 continue;
971 // Simple DCE.
972 if (isInstructionTriviallyDead(I)) {
973 DEBUG(dbgs() << "Remove dead instruction '" << *I);
975 // Add uses to the worklist, which may be dead now.
976 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
977 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
978 Worklist.push_back(Use);
979 LPM->deleteSimpleAnalysisValue(I, L);
980 RemoveFromWorklist(I, Worklist);
981 I->eraseFromParent();
982 ++NumSimplify;
983 continue;
986 // See if instruction simplification can hack this up. This is common for
987 // things like "select false, X, Y" after unswitching made the condition be
988 // 'false'.
989 if (Value *V = SimplifyInstruction(I)) {
990 ReplaceUsesOfWith(I, V, Worklist, L, LPM);
991 continue;
994 // Special case hacks that appear commonly in unswitched code.
995 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
996 if (BI->isUnconditional()) {
997 // If BI's parent is the only pred of the successor, fold the two blocks
998 // together.
999 BasicBlock *Pred = BI->getParent();
1000 BasicBlock *Succ = BI->getSuccessor(0);
1001 BasicBlock *SinglePred = Succ->getSinglePredecessor();
1002 if (!SinglePred) continue; // Nothing to do.
1003 assert(SinglePred == Pred && "CFG broken");
1005 DEBUG(dbgs() << "Merging blocks: " << Pred->getName() << " <- "
1006 << Succ->getName() << "\n");
1008 // Resolve any single entry PHI nodes in Succ.
1009 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin()))
1010 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM);
1012 // Move all of the successor contents from Succ to Pred.
1013 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(),
1014 Succ->end());
1015 LPM->deleteSimpleAnalysisValue(BI, L);
1016 BI->eraseFromParent();
1017 RemoveFromWorklist(BI, Worklist);
1019 // If Succ has any successors with PHI nodes, update them to have
1020 // entries coming from Pred instead of Succ.
1021 Succ->replaceAllUsesWith(Pred);
1023 // Remove Succ from the loop tree.
1024 LI->removeBlock(Succ);
1025 LPM->deleteSimpleAnalysisValue(Succ, L);
1026 Succ->eraseFromParent();
1027 ++NumSimplify;
1028 continue;
1031 if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){
1032 // Conditional branch. Turn it into an unconditional branch, then
1033 // remove dead blocks.
1034 continue; // FIXME: Enable.
1036 DEBUG(dbgs() << "Folded branch: " << *BI);
1037 BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue());
1038 BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue());
1039 DeadSucc->removePredecessor(BI->getParent(), true);
1040 Worklist.push_back(BranchInst::Create(LiveSucc, BI));
1041 LPM->deleteSimpleAnalysisValue(BI, L);
1042 BI->eraseFromParent();
1043 RemoveFromWorklist(BI, Worklist);
1044 ++NumSimplify;
1046 RemoveBlockIfDead(DeadSucc, Worklist, L);
1048 continue;