LoopUnswitch.cpp revision 2fac1d5d61a83c45dcf44119c41dce15ef10e9dc
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:
12//
13//  for (...)                  if (lic)
14//    A                          for (...)
15//    if (lic)                     A; B; C
16//      B                      else
17//    C                          for (...)
18//                                 A; C
19//
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.
23//
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.
26//
27//===----------------------------------------------------------------------===//
28
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/InlineCost.h"
36#include "llvm/Analysis/InstructionSimplify.h"
37#include "llvm/Analysis/LoopInfo.h"
38#include "llvm/Analysis/LoopPass.h"
39#include "llvm/Analysis/Dominators.h"
40#include "llvm/Analysis/ScalarEvolution.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>
52using namespace llvm;
53
54STATISTIC(NumBranches, "Number of branches unswitched");
55STATISTIC(NumSwitches, "Number of switches unswitched");
56STATISTIC(NumSelects , "Number of selects unswitched");
57STATISTIC(NumTrivial , "Number of unswitches that are trivial");
58STATISTIC(NumSimplify, "Number of simplifications of unswitched code");
59
60// The specific value of 50 here was chosen based only on intuition and a
61// few specific examples.
62static cl::opt<unsigned>
63Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"),
64          cl::init(50), cl::Hidden);
65
66namespace {
67  class LoopUnswitch : public LoopPass {
68    LoopInfo *LI;  // Loop information
69    LPPassManager *LPM;
70
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;
75
76    bool OptimizeForSize;
77    bool redoLoop;
78
79    Loop *currentLoop;
80    DominatorTree *DT;
81    BasicBlock *loopHeader;
82    BasicBlock *loopPreheader;
83
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;
90
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) {
97        initializeLoopUnswitchPass(*PassRegistry::getPassRegistry());
98      }
99
100    bool runOnLoop(Loop *L, LPPassManager &LPM);
101    bool processCurrentLoop();
102
103    /// This transformation requires natural loop information & requires that
104    /// loop preheaders be inserted into the CFG.
105    ///
106    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
107      AU.addRequiredID(LoopSimplifyID);
108      AU.addPreservedID(LoopSimplifyID);
109      AU.addRequired<LoopInfo>();
110      AU.addPreserved<LoopInfo>();
111      AU.addRequiredID(LCSSAID);
112      AU.addPreservedID(LCSSAID);
113      AU.addPreserved<DominatorTree>();
114      AU.addPreserved<ScalarEvolution>();
115    }
116
117  private:
118
119    virtual void releaseMemory() {
120      UnswitchedVals.clear();
121    }
122
123    /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist,
124    /// remove it.
125    void RemoveLoopFromWorklist(Loop *L) {
126      std::vector<Loop*>::iterator I = std::find(LoopProcessWorklist.begin(),
127                                                 LoopProcessWorklist.end(), L);
128      if (I != LoopProcessWorklist.end())
129        LoopProcessWorklist.erase(I);
130    }
131
132    void initLoopData() {
133      loopHeader = currentLoop->getHeader();
134      loopPreheader = currentLoop->getLoopPreheader();
135    }
136
137    /// Split all of the edges from inside the loop to their exit blocks.
138    /// Update the appropriate Phi nodes as we do so.
139    void SplitExitEdges(Loop *L, const SmallVector<BasicBlock *, 8> &ExitBlocks);
140
141    bool UnswitchIfProfitable(Value *LoopCond, Constant *Val);
142    void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val,
143                                  BasicBlock *ExitBlock);
144    void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L);
145
146    void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
147                                              Constant *Val, bool isEqual);
148
149    void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
150                                        BasicBlock *TrueDest,
151                                        BasicBlock *FalseDest,
152                                        Instruction *InsertPt);
153
154    void SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L);
155    void RemoveBlockIfDead(BasicBlock *BB,
156                           std::vector<Instruction*> &Worklist, Loop *l);
157    void RemoveLoopFromHierarchy(Loop *L);
158    bool IsTrivialUnswitchCondition(Value *Cond, Constant **Val = 0,
159                                    BasicBlock **LoopExit = 0);
160
161  };
162}
163char LoopUnswitch::ID = 0;
164INITIALIZE_PASS_BEGIN(LoopUnswitch, "loop-unswitch", "Unswitch loops",
165                      false, false)
166INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
167INITIALIZE_PASS_DEPENDENCY(LoopInfo)
168INITIALIZE_PASS_DEPENDENCY(LCSSA)
169INITIALIZE_PASS_END(LoopUnswitch, "loop-unswitch", "Unswitch loops",
170                      false, false)
171
172Pass *llvm::createLoopUnswitchPass(bool Os) {
173  return new LoopUnswitch(Os);
174}
175
176/// FindLIVLoopCondition - Cond is a condition that occurs in L.  If it is
177/// invariant in the loop, or has an invariant piece, return the invariant.
178/// Otherwise, return null.
179static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) {
180  // We can never unswitch on vector conditions.
181  if (Cond->getType()->isVectorTy())
182    return 0;
183
184  // Constants should be folded, not unswitched on!
185  if (isa<Constant>(Cond)) return 0;
186
187  // TODO: Handle: br (VARIANT|INVARIANT).
188
189  // Hoist simple values out.
190  if (L->makeLoopInvariant(Cond, Changed))
191    return Cond;
192
193  if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond))
194    if (BO->getOpcode() == Instruction::And ||
195        BO->getOpcode() == Instruction::Or) {
196      // If either the left or right side is invariant, we can unswitch on this,
197      // which will cause the branch to go away in one loop and the condition to
198      // simplify in the other one.
199      if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed))
200        return LHS;
201      if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed))
202        return RHS;
203    }
204
205  return 0;
206}
207
208bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) {
209  LI = &getAnalysis<LoopInfo>();
210  LPM = &LPM_Ref;
211  DT = getAnalysisIfAvailable<DominatorTree>();
212  currentLoop = L;
213  Function *F = currentLoop->getHeader()->getParent();
214  bool Changed = false;
215  do {
216    assert(currentLoop->isLCSSAForm(*DT));
217    redoLoop = false;
218    Changed |= processCurrentLoop();
219  } while(redoLoop);
220
221  if (Changed) {
222    // FIXME: Reconstruct dom info, because it is not preserved properly.
223    if (DT)
224      DT->runOnFunction(*F);
225  }
226  return Changed;
227}
228
229/// processCurrentLoop - Do actual work and unswitch loop if possible
230/// and profitable.
231bool LoopUnswitch::processCurrentLoop() {
232  bool Changed = false;
233  LLVMContext &Context = currentLoop->getHeader()->getContext();
234
235  // Loop over all of the basic blocks in the loop.  If we find an interior
236  // block that is branching on a loop-invariant condition, we can unswitch this
237  // loop.
238  for (Loop::block_iterator I = currentLoop->block_begin(),
239         E = currentLoop->block_end(); I != E; ++I) {
240    TerminatorInst *TI = (*I)->getTerminator();
241    if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
242      // If this isn't branching on an invariant condition, we can't unswitch
243      // it.
244      if (BI->isConditional()) {
245        // See if this, or some part of it, is loop invariant.  If so, we can
246        // unswitch on it if we desire.
247        Value *LoopCond = FindLIVLoopCondition(BI->getCondition(),
248                                               currentLoop, Changed);
249        if (LoopCond && UnswitchIfProfitable(LoopCond,
250                                             ConstantInt::getTrue(Context))) {
251          ++NumBranches;
252          return true;
253        }
254      }
255    } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
256      Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
257                                             currentLoop, Changed);
258      if (LoopCond && SI->getNumCases() > 1) {
259        // Find a value to unswitch on:
260        // FIXME: this should chose the most expensive case!
261        // FIXME: scan for a case with a non-critical edge?
262        Constant *UnswitchVal = SI->getCaseValue(1);
263        // Do not process same value again and again.
264        if (!UnswitchedVals.insert(UnswitchVal))
265          continue;
266
267        if (UnswitchIfProfitable(LoopCond, UnswitchVal)) {
268          ++NumSwitches;
269          return true;
270        }
271      }
272    }
273
274    // Scan the instructions to check for unswitchable values.
275    for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end();
276         BBI != E; ++BBI)
277      if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) {
278        Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
279                                               currentLoop, Changed);
280        if (LoopCond && UnswitchIfProfitable(LoopCond,
281                                             ConstantInt::getTrue(Context))) {
282          ++NumSelects;
283          return true;
284        }
285      }
286  }
287  return Changed;
288}
289
290/// isTrivialLoopExitBlock - Check to see if all paths from BB exit the
291/// loop with no side effects (including infinite loops).
292///
293/// If true, we return true and set ExitBB to the block we
294/// exit through.
295///
296static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB,
297                                         BasicBlock *&ExitBB,
298                                         std::set<BasicBlock*> &Visited) {
299  if (!Visited.insert(BB).second) {
300    // Already visited. Without more analysis, this could indicate an infinte loop.
301    return false;
302  } else if (!L->contains(BB)) {
303    // Otherwise, this is a loop exit, this is fine so long as this is the
304    // first exit.
305    if (ExitBB != 0) return false;
306    ExitBB = BB;
307    return true;
308  }
309
310  // Otherwise, this is an unvisited intra-loop node.  Check all successors.
311  for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) {
312    // Check to see if the successor is a trivial loop exit.
313    if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited))
314      return false;
315  }
316
317  // Okay, everything after this looks good, check to make sure that this block
318  // doesn't include any side effects.
319  for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
320    if (I->mayHaveSideEffects())
321      return false;
322
323  return true;
324}
325
326/// isTrivialLoopExitBlock - Return true if the specified block unconditionally
327/// leads to an exit from the specified loop, and has no side-effects in the
328/// process.  If so, return the block that is exited to, otherwise return null.
329static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) {
330  std::set<BasicBlock*> Visited;
331  Visited.insert(L->getHeader());  // Branches to header make infinite loops.
332  BasicBlock *ExitBB = 0;
333  if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited))
334    return ExitBB;
335  return 0;
336}
337
338/// IsTrivialUnswitchCondition - Check to see if this unswitch condition is
339/// trivial: that is, that the condition controls whether or not the loop does
340/// anything at all.  If this is a trivial condition, unswitching produces no
341/// code duplications (equivalently, it produces a simpler loop and a new empty
342/// loop, which gets deleted).
343///
344/// If this is a trivial condition, return true, otherwise return false.  When
345/// returning true, this sets Cond and Val to the condition that controls the
346/// trivial condition: when Cond dynamically equals Val, the loop is known to
347/// exit.  Finally, this sets LoopExit to the BB that the loop exits to when
348/// Cond == Val.
349///
350bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val,
351                                       BasicBlock **LoopExit) {
352  BasicBlock *Header = currentLoop->getHeader();
353  TerminatorInst *HeaderTerm = Header->getTerminator();
354  LLVMContext &Context = Header->getContext();
355
356  BasicBlock *LoopExitBB = 0;
357  if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) {
358    // If the header block doesn't end with a conditional branch on Cond, we
359    // can't handle it.
360    if (!BI->isConditional() || BI->getCondition() != Cond)
361      return false;
362
363    // Check to see if a successor of the branch is guaranteed to
364    // exit through a unique exit block without having any
365    // side-effects.  If so, determine the value of Cond that causes it to do
366    // this.
367    if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
368                                             BI->getSuccessor(0)))) {
369      if (Val) *Val = ConstantInt::getTrue(Context);
370    } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
371                                                    BI->getSuccessor(1)))) {
372      if (Val) *Val = ConstantInt::getFalse(Context);
373    }
374  } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) {
375    // If this isn't a switch on Cond, we can't handle it.
376    if (SI->getCondition() != Cond) return false;
377
378    // Check to see if a successor of the switch is guaranteed to go to the
379    // latch block or exit through a one exit block without having any
380    // side-effects.  If so, determine the value of Cond that causes it to do
381    // this.  Note that we can't trivially unswitch on the default case.
382    for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
383      if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
384                                               SI->getSuccessor(i)))) {
385        // Okay, we found a trivial case, remember the value that is trivial.
386        if (Val) *Val = SI->getCaseValue(i);
387        break;
388      }
389  }
390
391  // If we didn't find a single unique LoopExit block, or if the loop exit block
392  // contains phi nodes, this isn't trivial.
393  if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin()))
394    return false;   // Can't handle this.
395
396  if (LoopExit) *LoopExit = LoopExitBB;
397
398  // We already know that nothing uses any scalar values defined inside of this
399  // loop.  As such, we just have to check to see if this loop will execute any
400  // side-effecting instructions (e.g. stores, calls, volatile loads) in the
401  // part of the loop that the code *would* execute.  We already checked the
402  // tail, check the header now.
403  for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I)
404    if (I->mayHaveSideEffects())
405      return false;
406  return true;
407}
408
409/// UnswitchIfProfitable - We have found that we can unswitch currentLoop when
410/// LoopCond == Val to simplify the loop.  If we decide that this is profitable,
411/// unswitch the loop, reprocess the pieces, then return true.
412bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val) {
413
414  initLoopData();
415
416  // If LoopSimplify was unable to form a preheader, don't do any unswitching.
417  if (!loopPreheader)
418    return false;
419
420  Function *F = loopHeader->getParent();
421
422  Constant *CondVal = 0;
423  BasicBlock *ExitBlock = 0;
424  if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) {
425    // If the condition is trivial, always unswitch. There is no code growth
426    // for this case.
427    UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock);
428    return true;
429  }
430
431  // Check to see if it would be profitable to unswitch current loop.
432
433  // Do not do non-trivial unswitch while optimizing for size.
434  if (OptimizeForSize || F->hasFnAttr(Attribute::OptimizeForSize))
435    return false;
436
437  // FIXME: This is overly conservative because it does not take into
438  // consideration code simplification opportunities and code that can
439  // be shared by the resultant unswitched loops.
440  CodeMetrics Metrics;
441  for (Loop::block_iterator I = currentLoop->block_begin(),
442         E = currentLoop->block_end();
443       I != E; ++I)
444    Metrics.analyzeBasicBlock(*I);
445
446  // Limit the number of instructions to avoid causing significant code
447  // expansion, and the number of basic blocks, to avoid loops with
448  // large numbers of branches which cause loop unswitching to go crazy.
449  // This is a very ad-hoc heuristic.
450  if (Metrics.NumInsts > Threshold ||
451      Metrics.NumBlocks * 5 > Threshold ||
452      Metrics.containsIndirectBr || Metrics.isRecursive) {
453    DEBUG(dbgs() << "NOT unswitching loop %"
454          << currentLoop->getHeader()->getName() << ", cost too high: "
455          << currentLoop->getBlocks().size() << "\n");
456    return false;
457  }
458
459  UnswitchNontrivialCondition(LoopCond, Val, currentLoop);
460  return true;
461}
462
463/// CloneLoop - Recursively clone the specified loop and all of its children,
464/// mapping the blocks with the specified map.
465static Loop *CloneLoop(Loop *L, Loop *PL, ValueToValueMapTy &VM,
466                       LoopInfo *LI, LPPassManager *LPM) {
467  Loop *New = new Loop();
468  LPM->insertLoop(New, PL);
469
470  // Add all of the blocks in L to the new loop.
471  for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
472       I != E; ++I)
473    if (LI->getLoopFor(*I) == L)
474      New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), LI->getBase());
475
476  // Add all of the subloops to the new loop.
477  for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
478    CloneLoop(*I, New, VM, LI, LPM);
479
480  return New;
481}
482
483/// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values
484/// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest.  Insert the
485/// code immediately before InsertPt.
486void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
487                                                  BasicBlock *TrueDest,
488                                                  BasicBlock *FalseDest,
489                                                  Instruction *InsertPt) {
490  // Insert a conditional branch on LIC to the two preheaders.  The original
491  // code is the true version and the new code is the false version.
492  Value *BranchVal = LIC;
493  if (!isa<ConstantInt>(Val) ||
494      Val->getType() != Type::getInt1Ty(LIC->getContext()))
495    BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val);
496  else if (Val != ConstantInt::getTrue(Val->getContext()))
497    // We want to enter the new loop when the condition is true.
498    std::swap(TrueDest, FalseDest);
499
500  // Insert the new branch.
501  BranchInst *BI = BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt);
502
503  // If either edge is critical, split it. This helps preserve LoopSimplify
504  // form for enclosing loops.
505  SplitCriticalEdge(BI, 0, this);
506  SplitCriticalEdge(BI, 1, this);
507}
508
509/// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable
510/// condition in it (a cond branch from its header block to its latch block,
511/// where the path through the loop that doesn't execute its body has no
512/// side-effects), unswitch it.  This doesn't involve any code duplication, just
513/// moving the conditional branch outside of the loop and updating loop info.
514void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond,
515                                            Constant *Val,
516                                            BasicBlock *ExitBlock) {
517  DEBUG(dbgs() << "loop-unswitch: Trivial-Unswitch loop %"
518        << loopHeader->getName() << " [" << L->getBlocks().size()
519        << " blocks] in Function " << L->getHeader()->getParent()->getName()
520        << " on cond: " << *Val << " == " << *Cond << "\n");
521
522  // First step, split the preheader, so that we know that there is a safe place
523  // to insert the conditional branch.  We will change loopPreheader to have a
524  // conditional branch on Cond.
525  BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, this);
526
527  // Now that we have a place to insert the conditional branch, create a place
528  // to branch to: this is the exit block out of the loop that we should
529  // short-circuit to.
530
531  // Split this block now, so that the loop maintains its exit block, and so
532  // that the jump from the preheader can execute the contents of the exit block
533  // without actually branching to it (the exit block should be dominated by the
534  // loop header, not the preheader).
535  assert(!L->contains(ExitBlock) && "Exit block is in the loop?");
536  BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), this);
537
538  // Okay, now we have a position to branch from and a position to branch to,
539  // insert the new conditional branch.
540  EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH,
541                                 loopPreheader->getTerminator());
542  LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L);
543  loopPreheader->getTerminator()->eraseFromParent();
544
545  // We need to reprocess this loop, it could be unswitched again.
546  redoLoop = true;
547
548  // Now that we know that the loop is never entered when this condition is a
549  // particular value, rewrite the loop with this info.  We know that this will
550  // at least eliminate the old branch.
551  RewriteLoopBodyWithConditionConstant(L, Cond, Val, false);
552  ++NumTrivial;
553}
554
555/// SplitExitEdges - Split all of the edges from inside the loop to their exit
556/// blocks.  Update the appropriate Phi nodes as we do so.
557void LoopUnswitch::SplitExitEdges(Loop *L,
558                                const SmallVector<BasicBlock *, 8> &ExitBlocks){
559
560  for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
561    BasicBlock *ExitBlock = ExitBlocks[i];
562    SmallVector<BasicBlock *, 4> Preds(pred_begin(ExitBlock),
563                                       pred_end(ExitBlock));
564
565    // Although SplitBlockPredecessors doesn't preserve loop-simplify in
566    // general, if we call it on all predecessors of all exits then it does.
567    if (!ExitBlock->isLandingPad()) {
568      SplitBlockPredecessors(ExitBlock, Preds, ".us-lcssa", this);
569    } else {
570      SmallVector<BasicBlock*, 2> NewBBs;
571      SplitLandingPadPredecessors(ExitBlock, Preds, ".us-lcssa", ".us-lcssa",
572                                  this, NewBBs);
573    }
574  }
575}
576
577/// UnswitchNontrivialCondition - We determined that the loop is profitable
578/// to unswitch when LIC equal Val.  Split it into loop versions and test the
579/// condition outside of either loop.  Return the loops created as Out1/Out2.
580void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val,
581                                               Loop *L) {
582  Function *F = loopHeader->getParent();
583  DEBUG(dbgs() << "loop-unswitch: Unswitching loop %"
584        << loopHeader->getName() << " [" << L->getBlocks().size()
585        << " blocks] in Function " << F->getName()
586        << " when '" << *Val << "' == " << *LIC << "\n");
587
588  if (ScalarEvolution *SE = getAnalysisIfAvailable<ScalarEvolution>())
589    SE->forgetLoop(L);
590
591  LoopBlocks.clear();
592  NewBlocks.clear();
593
594  // First step, split the preheader and exit blocks, and add these blocks to
595  // the LoopBlocks list.
596  BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, this);
597  LoopBlocks.push_back(NewPreheader);
598
599  // We want the loop to come after the preheader, but before the exit blocks.
600  LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
601
602  SmallVector<BasicBlock*, 8> ExitBlocks;
603  L->getUniqueExitBlocks(ExitBlocks);
604
605  // Split all of the edges from inside the loop to their exit blocks.  Update
606  // the appropriate Phi nodes as we do so.
607  SplitExitEdges(L, ExitBlocks);
608
609  // The exit blocks may have been changed due to edge splitting, recompute.
610  ExitBlocks.clear();
611  L->getUniqueExitBlocks(ExitBlocks);
612
613  // Add exit blocks to the loop blocks.
614  LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end());
615
616  // Next step, clone all of the basic blocks that make up the loop (including
617  // the loop preheader and exit blocks), keeping track of the mapping between
618  // the instructions and blocks.
619  NewBlocks.reserve(LoopBlocks.size());
620  ValueToValueMapTy VMap;
621  for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
622    BasicBlock *NewBB = CloneBasicBlock(LoopBlocks[i], VMap, ".us", F);
623    NewBlocks.push_back(NewBB);
624    VMap[LoopBlocks[i]] = NewBB;  // Keep the BB mapping.
625    LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], NewBB, L);
626  }
627
628  // Splice the newly inserted blocks into the function right before the
629  // original preheader.
630  F->getBasicBlockList().splice(NewPreheader, F->getBasicBlockList(),
631                                NewBlocks[0], F->end());
632
633  // Now we create the new Loop object for the versioned loop.
634  Loop *NewLoop = CloneLoop(L, L->getParentLoop(), VMap, LI, LPM);
635  Loop *ParentLoop = L->getParentLoop();
636  if (ParentLoop) {
637    // Make sure to add the cloned preheader and exit blocks to the parent loop
638    // as well.
639    ParentLoop->addBasicBlockToLoop(NewBlocks[0], LI->getBase());
640  }
641
642  for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
643    BasicBlock *NewExit = cast<BasicBlock>(VMap[ExitBlocks[i]]);
644    // The new exit block should be in the same loop as the old one.
645    if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i]))
646      ExitBBLoop->addBasicBlockToLoop(NewExit, LI->getBase());
647
648    assert(NewExit->getTerminator()->getNumSuccessors() == 1 &&
649           "Exit block should have been split to have one successor!");
650    BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0);
651
652    // If the successor of the exit block had PHI nodes, add an entry for
653    // NewExit.
654    PHINode *PN;
655    for (BasicBlock::iterator I = ExitSucc->begin(); isa<PHINode>(I); ++I) {
656      PN = cast<PHINode>(I);
657      Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]);
658      ValueToValueMapTy::iterator It = VMap.find(V);
659      if (It != VMap.end()) V = It->second;
660      PN->addIncoming(V, NewExit);
661    }
662
663    if (LandingPadInst *LPad = NewExit->getLandingPadInst()) {
664      PN = PHINode::Create(LPad->getType(), 0, "",
665                           ExitSucc->getFirstInsertionPt());
666
667      for (pred_iterator I = pred_begin(ExitSucc), E = pred_end(ExitSucc);
668           I != E; ++I) {
669        BasicBlock *BB = *I;
670        LandingPadInst *LPI = BB->getLandingPadInst();
671        LPI->replaceAllUsesWith(PN);
672        PN->addIncoming(LPI, BB);
673      }
674    }
675  }
676
677  // Rewrite the code to refer to itself.
678  for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i)
679    for (BasicBlock::iterator I = NewBlocks[i]->begin(),
680           E = NewBlocks[i]->end(); I != E; ++I)
681      RemapInstruction(I, VMap,RF_NoModuleLevelChanges|RF_IgnoreMissingEntries);
682
683  // Rewrite the original preheader to select between versions of the loop.
684  BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator());
685  assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] &&
686         "Preheader splitting did not work correctly!");
687
688  // Emit the new branch that selects between the two versions of this loop.
689  EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR);
690  LPM->deleteSimpleAnalysisValue(OldBR, L);
691  OldBR->eraseFromParent();
692
693  LoopProcessWorklist.push_back(NewLoop);
694  redoLoop = true;
695
696  // Keep a WeakVH holding onto LIC.  If the first call to RewriteLoopBody
697  // deletes the instruction (for example by simplifying a PHI that feeds into
698  // the condition that we're unswitching on), we don't rewrite the second
699  // iteration.
700  WeakVH LICHandle(LIC);
701
702  // Now we rewrite the original code to know that the condition is true and the
703  // new code to know that the condition is false.
704  RewriteLoopBodyWithConditionConstant(L, LIC, Val, false);
705
706  // It's possible that simplifying one loop could cause the other to be
707  // changed to another value or a constant.  If its a constant, don't simplify
708  // it.
709  if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop &&
710      LICHandle && !isa<Constant>(LICHandle))
711    RewriteLoopBodyWithConditionConstant(NewLoop, LICHandle, Val, true);
712}
713
714/// RemoveFromWorklist - Remove all instances of I from the worklist vector
715/// specified.
716static void RemoveFromWorklist(Instruction *I,
717                               std::vector<Instruction*> &Worklist) {
718  std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(),
719                                                     Worklist.end(), I);
720  while (WI != Worklist.end()) {
721    unsigned Offset = WI-Worklist.begin();
722    Worklist.erase(WI);
723    WI = std::find(Worklist.begin()+Offset, Worklist.end(), I);
724  }
725}
726
727/// ReplaceUsesOfWith - When we find that I really equals V, remove I from the
728/// program, replacing all uses with V and update the worklist.
729static void ReplaceUsesOfWith(Instruction *I, Value *V,
730                              std::vector<Instruction*> &Worklist,
731                              Loop *L, LPPassManager *LPM) {
732  DEBUG(dbgs() << "Replace with '" << *V << "': " << *I);
733
734  // Add uses to the worklist, which may be dead now.
735  for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
736    if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
737      Worklist.push_back(Use);
738
739  // Add users to the worklist which may be simplified now.
740  for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
741       UI != E; ++UI)
742    Worklist.push_back(cast<Instruction>(*UI));
743  LPM->deleteSimpleAnalysisValue(I, L);
744  RemoveFromWorklist(I, Worklist);
745  I->replaceAllUsesWith(V);
746  I->eraseFromParent();
747  ++NumSimplify;
748}
749
750/// RemoveBlockIfDead - If the specified block is dead, remove it, update loop
751/// information, and remove any dead successors it has.
752///
753void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB,
754                                     std::vector<Instruction*> &Worklist,
755                                     Loop *L) {
756  if (pred_begin(BB) != pred_end(BB)) {
757    // This block isn't dead, since an edge to BB was just removed, see if there
758    // are any easy simplifications we can do now.
759    if (BasicBlock *Pred = BB->getSinglePredecessor()) {
760      // If it has one pred, fold phi nodes in BB.
761      while (isa<PHINode>(BB->begin()))
762        ReplaceUsesOfWith(BB->begin(),
763                          cast<PHINode>(BB->begin())->getIncomingValue(0),
764                          Worklist, L, LPM);
765
766      // If this is the header of a loop and the only pred is the latch, we now
767      // have an unreachable loop.
768      if (Loop *L = LI->getLoopFor(BB))
769        if (loopHeader == BB && L->contains(Pred)) {
770          // Remove the branch from the latch to the header block, this makes
771          // the header dead, which will make the latch dead (because the header
772          // dominates the latch).
773          LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L);
774          Pred->getTerminator()->eraseFromParent();
775          new UnreachableInst(BB->getContext(), Pred);
776
777          // The loop is now broken, remove it from LI.
778          RemoveLoopFromHierarchy(L);
779
780          // Reprocess the header, which now IS dead.
781          RemoveBlockIfDead(BB, Worklist, L);
782          return;
783        }
784
785      // If pred ends in a uncond branch, add uncond branch to worklist so that
786      // the two blocks will get merged.
787      if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator()))
788        if (BI->isUnconditional())
789          Worklist.push_back(BI);
790    }
791    return;
792  }
793
794  DEBUG(dbgs() << "Nuking dead block: " << *BB);
795
796  // Remove the instructions in the basic block from the worklist.
797  for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
798    RemoveFromWorklist(I, Worklist);
799
800    // Anything that uses the instructions in this basic block should have their
801    // uses replaced with undefs.
802    // If I is not void type then replaceAllUsesWith undef.
803    // This allows ValueHandlers and custom metadata to adjust itself.
804    if (!I->getType()->isVoidTy())
805      I->replaceAllUsesWith(UndefValue::get(I->getType()));
806  }
807
808  // If this is the edge to the header block for a loop, remove the loop and
809  // promote all subloops.
810  if (Loop *BBLoop = LI->getLoopFor(BB)) {
811    if (BBLoop->getLoopLatch() == BB) {
812      RemoveLoopFromHierarchy(BBLoop);
813      if (currentLoop == BBLoop) {
814        currentLoop = 0;
815        redoLoop = false;
816      }
817    }
818  }
819
820  // Remove the block from the loop info, which removes it from any loops it
821  // was in.
822  LI->removeBlock(BB);
823
824
825  // Remove phi node entries in successors for this block.
826  TerminatorInst *TI = BB->getTerminator();
827  SmallVector<BasicBlock*, 4> Succs;
828  for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
829    Succs.push_back(TI->getSuccessor(i));
830    TI->getSuccessor(i)->removePredecessor(BB);
831  }
832
833  // Unique the successors, remove anything with multiple uses.
834  array_pod_sort(Succs.begin(), Succs.end());
835  Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end());
836
837  // Remove the basic block, including all of the instructions contained in it.
838  LPM->deleteSimpleAnalysisValue(BB, L);
839  BB->eraseFromParent();
840  // Remove successor blocks here that are not dead, so that we know we only
841  // have dead blocks in this list.  Nondead blocks have a way of becoming dead,
842  // then getting removed before we revisit them, which is badness.
843  //
844  for (unsigned i = 0; i != Succs.size(); ++i)
845    if (pred_begin(Succs[i]) != pred_end(Succs[i])) {
846      // One exception is loop headers.  If this block was the preheader for a
847      // loop, then we DO want to visit the loop so the loop gets deleted.
848      // We know that if the successor is a loop header, that this loop had to
849      // be the preheader: the case where this was the latch block was handled
850      // above and headers can only have two predecessors.
851      if (!LI->isLoopHeader(Succs[i])) {
852        Succs.erase(Succs.begin()+i);
853        --i;
854      }
855    }
856
857  for (unsigned i = 0, e = Succs.size(); i != e; ++i)
858    RemoveBlockIfDead(Succs[i], Worklist, L);
859}
860
861/// RemoveLoopFromHierarchy - We have discovered that the specified loop has
862/// become unwrapped, either because the backedge was deleted, or because the
863/// edge into the header was removed.  If the edge into the header from the
864/// latch block was removed, the loop is unwrapped but subloops are still alive,
865/// so they just reparent loops.  If the loops are actually dead, they will be
866/// removed later.
867void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) {
868  LPM->deleteLoopFromQueue(L);
869  RemoveLoopFromWorklist(L);
870}
871
872// RewriteLoopBodyWithConditionConstant - We know either that the value LIC has
873// the value specified by Val in the specified loop, or we know it does NOT have
874// that value.  Rewrite any uses of LIC or of properties correlated to it.
875void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
876                                                        Constant *Val,
877                                                        bool IsEqual) {
878  assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?");
879
880  // FIXME: Support correlated properties, like:
881  //  for (...)
882  //    if (li1 < li2)
883  //      ...
884  //    if (li1 > li2)
885  //      ...
886
887  // FOLD boolean conditions (X|LIC), (X&LIC).  Fold conditional branches,
888  // selects, switches.
889  std::vector<Instruction*> Worklist;
890  LLVMContext &Context = Val->getContext();
891
892
893  // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC
894  // in the loop with the appropriate one directly.
895  if (IsEqual || (isa<ConstantInt>(Val) &&
896      Val->getType()->isIntegerTy(1))) {
897    Value *Replacement;
898    if (IsEqual)
899      Replacement = Val;
900    else
901      Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()),
902                                     !cast<ConstantInt>(Val)->getZExtValue());
903
904    for (Value::use_iterator UI = LIC->use_begin(), E = LIC->use_end();
905         UI != E; ++UI) {
906      Instruction *U = dyn_cast<Instruction>(*UI);
907      if (!U || !L->contains(U))
908        continue;
909      Worklist.push_back(U);
910    }
911
912    for (std::vector<Instruction*>::iterator UI = Worklist.begin();
913         UI != Worklist.end(); ++UI)
914      (*UI)->replaceUsesOfWith(LIC, Replacement);
915
916    SimplifyCode(Worklist, L);
917    return;
918  }
919
920  // Otherwise, we don't know the precise value of LIC, but we do know that it
921  // is certainly NOT "Val".  As such, simplify any uses in the loop that we
922  // can.  This case occurs when we unswitch switch statements.
923  for (Value::use_iterator UI = LIC->use_begin(), E = LIC->use_end();
924       UI != E; ++UI) {
925    Instruction *U = dyn_cast<Instruction>(*UI);
926    if (!U || !L->contains(U))
927      continue;
928
929    Worklist.push_back(U);
930
931    // TODO: We could do other simplifications, for example, turning
932    // 'icmp eq LIC, Val' -> false.
933
934    // If we know that LIC is not Val, use this info to simplify code.
935    SwitchInst *SI = dyn_cast<SwitchInst>(U);
936    if (SI == 0 || !isa<ConstantInt>(Val)) continue;
937
938    unsigned DeadCase = SI->findCaseValue(cast<ConstantInt>(Val));
939    if (DeadCase == 0) continue;  // Default case is live for multiple values.
940
941    // Found a dead case value.  Don't remove PHI nodes in the
942    // successor if they become single-entry, those PHI nodes may
943    // be in the Users list.
944
945    BasicBlock *Switch = SI->getParent();
946    BasicBlock *SISucc = SI->getSuccessor(DeadCase);
947    BasicBlock *Latch = L->getLoopLatch();
948    if (!SI->findCaseDest(SISucc)) continue;  // Edge is critical.
949    // If the DeadCase successor dominates the loop latch, then the
950    // transformation isn't safe since it will delete the sole predecessor edge
951    // to the latch.
952    if (Latch && DT->dominates(SISucc, Latch))
953      continue;
954
955    // FIXME: This is a hack.  We need to keep the successor around
956    // and hooked up so as to preserve the loop structure, because
957    // trying to update it is complicated.  So instead we preserve the
958    // loop structure and put the block on a dead code path.
959    SplitEdge(Switch, SISucc, this);
960    // Compute the successors instead of relying on the return value
961    // of SplitEdge, since it may have split the switch successor
962    // after PHI nodes.
963    BasicBlock *NewSISucc = SI->getSuccessor(DeadCase);
964    BasicBlock *OldSISucc = *succ_begin(NewSISucc);
965    // Create an "unreachable" destination.
966    BasicBlock *Abort = BasicBlock::Create(Context, "us-unreachable",
967                                           Switch->getParent(),
968                                           OldSISucc);
969    new UnreachableInst(Context, Abort);
970    // Force the new case destination to branch to the "unreachable"
971    // block while maintaining a (dead) CFG edge to the old block.
972    NewSISucc->getTerminator()->eraseFromParent();
973    BranchInst::Create(Abort, OldSISucc,
974                       ConstantInt::getTrue(Context), NewSISucc);
975    // Release the PHI operands for this edge.
976    for (BasicBlock::iterator II = NewSISucc->begin();
977         PHINode *PN = dyn_cast<PHINode>(II); ++II)
978      PN->setIncomingValue(PN->getBasicBlockIndex(Switch),
979                           UndefValue::get(PN->getType()));
980    // Tell the domtree about the new block. We don't fully update the
981    // domtree here -- instead we force it to do a full recomputation
982    // after the pass is complete -- but we do need to inform it of
983    // new blocks.
984    if (DT)
985      DT->addNewBlock(Abort, NewSISucc);
986  }
987
988  SimplifyCode(Worklist, L);
989}
990
991/// SimplifyCode - Okay, now that we have simplified some instructions in the
992/// loop, walk over it and constant prop, dce, and fold control flow where
993/// possible.  Note that this is effectively a very simple loop-structure-aware
994/// optimizer.  During processing of this loop, L could very well be deleted, so
995/// it must not be used.
996///
997/// FIXME: When the loop optimizer is more mature, separate this out to a new
998/// pass.
999///
1000void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) {
1001  while (!Worklist.empty()) {
1002    Instruction *I = Worklist.back();
1003    Worklist.pop_back();
1004
1005    // Simple DCE.
1006    if (isInstructionTriviallyDead(I)) {
1007      DEBUG(dbgs() << "Remove dead instruction '" << *I);
1008
1009      // Add uses to the worklist, which may be dead now.
1010      for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
1011        if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
1012          Worklist.push_back(Use);
1013      LPM->deleteSimpleAnalysisValue(I, L);
1014      RemoveFromWorklist(I, Worklist);
1015      I->eraseFromParent();
1016      ++NumSimplify;
1017      continue;
1018    }
1019
1020    // See if instruction simplification can hack this up.  This is common for
1021    // things like "select false, X, Y" after unswitching made the condition be
1022    // 'false'.
1023    if (Value *V = SimplifyInstruction(I, 0, 0, DT))
1024      if (LI->replacementPreservesLCSSAForm(I, V)) {
1025        ReplaceUsesOfWith(I, V, Worklist, L, LPM);
1026        continue;
1027      }
1028
1029    // Special case hacks that appear commonly in unswitched code.
1030    if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1031      if (BI->isUnconditional()) {
1032        // If BI's parent is the only pred of the successor, fold the two blocks
1033        // together.
1034        BasicBlock *Pred = BI->getParent();
1035        BasicBlock *Succ = BI->getSuccessor(0);
1036        BasicBlock *SinglePred = Succ->getSinglePredecessor();
1037        if (!SinglePred) continue;  // Nothing to do.
1038        assert(SinglePred == Pred && "CFG broken");
1039
1040        DEBUG(dbgs() << "Merging blocks: " << Pred->getName() << " <- "
1041              << Succ->getName() << "\n");
1042
1043        // Resolve any single entry PHI nodes in Succ.
1044        while (PHINode *PN = dyn_cast<PHINode>(Succ->begin()))
1045          ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM);
1046
1047        // If Succ has any successors with PHI nodes, update them to have
1048        // entries coming from Pred instead of Succ.
1049        Succ->replaceAllUsesWith(Pred);
1050
1051        // Move all of the successor contents from Succ to Pred.
1052        Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(),
1053                                   Succ->end());
1054        LPM->deleteSimpleAnalysisValue(BI, L);
1055        BI->eraseFromParent();
1056        RemoveFromWorklist(BI, Worklist);
1057
1058        // Remove Succ from the loop tree.
1059        LI->removeBlock(Succ);
1060        LPM->deleteSimpleAnalysisValue(Succ, L);
1061        Succ->eraseFromParent();
1062        ++NumSimplify;
1063        continue;
1064      }
1065
1066      if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){
1067        // Conditional branch.  Turn it into an unconditional branch, then
1068        // remove dead blocks.
1069        continue;  // FIXME: Enable.
1070
1071        DEBUG(dbgs() << "Folded branch: " << *BI);
1072        BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue());
1073        BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue());
1074        DeadSucc->removePredecessor(BI->getParent(), true);
1075        Worklist.push_back(BranchInst::Create(LiveSucc, BI));
1076        LPM->deleteSimpleAnalysisValue(BI, L);
1077        BI->eraseFromParent();
1078        RemoveFromWorklist(BI, Worklist);
1079        ++NumSimplify;
1080
1081        RemoveBlockIfDead(DeadSucc, Worklist, L);
1082      }
1083      continue;
1084    }
1085  }
1086}
1087