LoopUnswitch.cpp revision 7b550ccfc5a3346c17e0390a59e2d6d19bc52705
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/ConstantFolding.h"
36#include "llvm/Analysis/InlineCost.h"
37#include "llvm/Analysis/LoopInfo.h"
38#include "llvm/Analysis/LoopPass.h"
39#include "llvm/Analysis/Dominators.h"
40#include "llvm/Transforms/Utils/Cloning.h"
41#include "llvm/Transforms/Utils/Local.h"
42#include "llvm/Transforms/Utils/BasicBlockUtils.h"
43#include "llvm/ADT/Statistic.h"
44#include "llvm/ADT/SmallPtrSet.h"
45#include "llvm/ADT/STLExtras.h"
46#include "llvm/Support/CommandLine.h"
47#include "llvm/Support/Debug.h"
48#include "llvm/Support/raw_ostream.h"
49#include <algorithm>
50#include <set>
51using namespace llvm;
52
53STATISTIC(NumBranches, "Number of branches unswitched");
54STATISTIC(NumSwitches, "Number of switches unswitched");
55STATISTIC(NumSelects , "Number of selects unswitched");
56STATISTIC(NumTrivial , "Number of unswitches that are trivial");
57STATISTIC(NumSimplify, "Number of simplifications of unswitched code");
58
59// The specific value of 50 here was chosen based only on intuition and a
60// few specific examples.
61static cl::opt<unsigned>
62Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"),
63          cl::init(50), cl::Hidden);
64
65namespace {
66  class LoopUnswitch : public LoopPass {
67    LoopInfo *LI;  // Loop information
68    LPPassManager *LPM;
69
70    // LoopProcessWorklist - Used to check if second loop needs processing
71    // after RewriteLoopBodyWithConditionConstant rewrites first loop.
72    std::vector<Loop*> LoopProcessWorklist;
73    SmallPtrSet<Value *,8> UnswitchedVals;
74
75    bool OptimizeForSize;
76    bool redoLoop;
77
78    Loop *currentLoop;
79    DominanceFrontier *DF;
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), DF(NULL), DT(NULL), loopHeader(NULL),
96      loopPreheader(NULL) {}
97
98    bool runOnLoop(Loop *L, LPPassManager &LPM);
99    bool processCurrentLoop();
100
101    /// This transformation requires natural loop information & requires that
102    /// loop preheaders be inserted into the CFG...
103    ///
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>();
112      AU.addPreserved<DominanceFrontier>();
113    }
114
115  private:
116
117    virtual void releaseMemory() {
118      UnswitchedVals.clear();
119    }
120
121    /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist,
122    /// remove it.
123    void RemoveLoopFromWorklist(Loop *L) {
124      std::vector<Loop*>::iterator I = std::find(LoopProcessWorklist.begin(),
125                                                 LoopProcessWorklist.end(), L);
126      if (I != LoopProcessWorklist.end())
127        LoopProcessWorklist.erase(I);
128    }
129
130    void initLoopData() {
131      loopHeader = currentLoop->getHeader();
132      loopPreheader = currentLoop->getLoopPreheader();
133    }
134
135    /// Split all of the edges from inside the loop to their exit blocks.
136    /// Update the appropriate Phi nodes as we do so.
137    void SplitExitEdges(Loop *L, const SmallVector<BasicBlock *, 8> &ExitBlocks);
138
139    bool UnswitchIfProfitable(Value *LoopCond, Constant *Val);
140    void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val,
141                                  BasicBlock *ExitBlock);
142    void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L);
143
144    void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
145                                              Constant *Val, bool isEqual);
146
147    void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
148                                        BasicBlock *TrueDest,
149                                        BasicBlock *FalseDest,
150                                        Instruction *InsertPt);
151
152    void SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L);
153    void RemoveBlockIfDead(BasicBlock *BB,
154                           std::vector<Instruction*> &Worklist, Loop *l);
155    void RemoveLoopFromHierarchy(Loop *L);
156    bool IsTrivialUnswitchCondition(Value *Cond, Constant **Val = 0,
157                                    BasicBlock **LoopExit = 0);
158
159  };
160}
161char LoopUnswitch::ID = 0;
162static RegisterPass<LoopUnswitch> X("loop-unswitch", "Unswitch loops");
163
164Pass *llvm::createLoopUnswitchPass(bool Os) {
165  return new LoopUnswitch(Os);
166}
167
168/// FindLIVLoopCondition - Cond is a condition that occurs in L.  If it is
169/// invariant in the loop, or has an invariant piece, return the invariant.
170/// Otherwise, return null.
171static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) {
172  // Constants should be folded, not unswitched on!
173  if (isa<Constant>(Cond)) return 0;
174
175  // TODO: Handle: br (VARIANT|INVARIANT).
176
177  // Hoist simple values out.
178  if (L->makeLoopInvariant(Cond, Changed))
179    return Cond;
180
181  if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond))
182    if (BO->getOpcode() == Instruction::And ||
183        BO->getOpcode() == Instruction::Or) {
184      // If either the left or right side is invariant, we can unswitch on this,
185      // which will cause the branch to go away in one loop and the condition to
186      // simplify in the other one.
187      if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed))
188        return LHS;
189      if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed))
190        return RHS;
191    }
192
193  return 0;
194}
195
196bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) {
197  LI = &getAnalysis<LoopInfo>();
198  LPM = &LPM_Ref;
199  DF = getAnalysisIfAvailable<DominanceFrontier>();
200  DT = getAnalysisIfAvailable<DominatorTree>();
201  currentLoop = L;
202  Function *F = currentLoop->getHeader()->getParent();
203  bool Changed = false;
204  do {
205    assert(currentLoop->isLCSSAForm());
206    redoLoop = false;
207    Changed |= processCurrentLoop();
208  } while(redoLoop);
209
210  if (Changed) {
211    // FIXME: Reconstruct dom info, because it is not preserved properly.
212    if (DT)
213      DT->runOnFunction(*F);
214    if (DF)
215      DF->runOnFunction(*F);
216  }
217  return Changed;
218}
219
220/// processCurrentLoop - Do actual work and unswitch loop if possible
221/// and profitable.
222bool LoopUnswitch::processCurrentLoop() {
223  bool Changed = false;
224  LLVMContext &Context = currentLoop->getHeader()->getContext();
225
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();
231       I != E; ++I) {
232    TerminatorInst *TI = (*I)->getTerminator();
233    if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
234      // If this isn't branching on an invariant condition, we can't unswitch
235      // it.
236      if (BI->isConditional()) {
237        // See if this, or some part of it, is loop invariant.  If so, we can
238        // unswitch on it if we desire.
239        Value *LoopCond = FindLIVLoopCondition(BI->getCondition(),
240                                               currentLoop, Changed);
241        if (LoopCond && UnswitchIfProfitable(LoopCond,
242                                             ConstantInt::getTrue(Context))) {
243          ++NumBranches;
244          return true;
245        }
246      }
247    } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
248      Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
249                                             currentLoop, Changed);
250      if (LoopCond && SI->getNumCases() > 1) {
251        // Find a value to unswitch on:
252        // FIXME: this should chose the most expensive case!
253        Constant *UnswitchVal = SI->getCaseValue(1);
254        // Do not process same value again and again.
255        if (!UnswitchedVals.insert(UnswitchVal))
256          continue;
257
258        if (UnswitchIfProfitable(LoopCond, UnswitchVal)) {
259          ++NumSwitches;
260          return true;
261        }
262      }
263    }
264
265    // Scan the instructions to check for unswitchable values.
266    for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end();
267         BBI != E; ++BBI)
268      if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) {
269        Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
270                                               currentLoop, Changed);
271        if (LoopCond && UnswitchIfProfitable(LoopCond,
272                                             ConstantInt::getTrue(Context))) {
273          ++NumSelects;
274          return true;
275        }
276      }
277  }
278  return Changed;
279}
280
281/// isTrivialLoopExitBlock - Check to see if all paths from BB either:
282///   1. Exit the loop with no side effects.
283///   2. Branch to the latch block with no side-effects.
284///
285/// If these conditions are true, we return true and set ExitBB to the block we
286/// exit through.
287///
288static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB,
289                                         BasicBlock *&ExitBB,
290                                         std::set<BasicBlock*> &Visited) {
291  if (!Visited.insert(BB).second) {
292    // Already visited and Ok, end of recursion.
293    return true;
294  } else if (!L->contains(BB)) {
295    // Otherwise, this is a loop exit, this is fine so long as this is the
296    // first exit.
297    if (ExitBB != 0) return false;
298    ExitBB = BB;
299    return true;
300  }
301
302  // Otherwise, this is an unvisited intra-loop node.  Check all successors.
303  for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) {
304    // Check to see if the successor is a trivial loop exit.
305    if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited))
306      return false;
307  }
308
309  // Okay, everything after this looks good, check to make sure that this block
310  // doesn't include any side effects.
311  for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
312    if (I->mayHaveSideEffects())
313      return false;
314
315  return true;
316}
317
318/// isTrivialLoopExitBlock - Return true if the specified block unconditionally
319/// leads to an exit from the specified loop, and has no side-effects in the
320/// process.  If so, return the block that is exited to, otherwise return null.
321static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) {
322  std::set<BasicBlock*> Visited;
323  Visited.insert(L->getHeader());  // Branches to header are ok.
324  BasicBlock *ExitBB = 0;
325  if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited))
326    return ExitBB;
327  return 0;
328}
329
330/// IsTrivialUnswitchCondition - Check to see if this unswitch condition is
331/// trivial: that is, that the condition controls whether or not the loop does
332/// anything at all.  If this is a trivial condition, unswitching produces no
333/// code duplications (equivalently, it produces a simpler loop and a new empty
334/// loop, which gets deleted).
335///
336/// If this is a trivial condition, return true, otherwise return false.  When
337/// returning true, this sets Cond and Val to the condition that controls the
338/// trivial condition: when Cond dynamically equals Val, the loop is known to
339/// exit.  Finally, this sets LoopExit to the BB that the loop exits to when
340/// Cond == Val.
341///
342bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val,
343                                       BasicBlock **LoopExit) {
344  BasicBlock *Header = currentLoop->getHeader();
345  TerminatorInst *HeaderTerm = Header->getTerminator();
346  LLVMContext &Context = Header->getContext();
347
348  BasicBlock *LoopExitBB = 0;
349  if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) {
350    // If the header block doesn't end with a conditional branch on Cond, we
351    // can't handle it.
352    if (!BI->isConditional() || BI->getCondition() != Cond)
353      return false;
354
355    // Check to see if a successor of the branch is guaranteed to go to the
356    // latch block or exit through a one exit block without having any
357    // side-effects.  If so, determine the value of Cond that causes it to do
358    // this.
359    if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
360                                             BI->getSuccessor(0)))) {
361      if (Val) *Val = ConstantInt::getTrue(Context);
362    } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
363                                                    BI->getSuccessor(1)))) {
364      if (Val) *Val = ConstantInt::getFalse(Context);
365    }
366  } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) {
367    // If this isn't a switch on Cond, we can't handle it.
368    if (SI->getCondition() != Cond) return false;
369
370    // Check to see if a successor of the switch is guaranteed to go to the
371    // latch block or exit through a one exit block without having any
372    // side-effects.  If so, determine the value of Cond that causes it to do
373    // this.  Note that we can't trivially unswitch on the default case.
374    for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
375      if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
376                                               SI->getSuccessor(i)))) {
377        // Okay, we found a trivial case, remember the value that is trivial.
378        if (Val) *Val = SI->getCaseValue(i);
379        break;
380      }
381  }
382
383  // If we didn't find a single unique LoopExit block, or if the loop exit block
384  // contains phi nodes, this isn't trivial.
385  if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin()))
386    return false;   // Can't handle this.
387
388  if (LoopExit) *LoopExit = LoopExitBB;
389
390  // We already know that nothing uses any scalar values defined inside of this
391  // loop.  As such, we just have to check to see if this loop will execute any
392  // side-effecting instructions (e.g. stores, calls, volatile loads) in the
393  // part of the loop that the code *would* execute.  We already checked the
394  // tail, check the header now.
395  for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I)
396    if (I->mayHaveSideEffects())
397      return false;
398  return true;
399}
400
401/// UnswitchIfProfitable - We have found that we can unswitch currentLoop when
402/// LoopCond == Val to simplify the loop.  If we decide that this is profitable,
403/// unswitch the loop, reprocess the pieces, then return true.
404bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val){
405
406  initLoopData();
407  Function *F = loopHeader->getParent();
408
409  // If LoopSimplify was unable to form a preheader, don't do any unswitching.
410  if (!loopPreheader)
411    return false;
412
413  // If the condition is trivial, always unswitch.  There is no code growth for
414  // this case.
415  if (!IsTrivialUnswitchCondition(LoopCond)) {
416    // Check to see if it would be profitable to unswitch current loop.
417
418    // Do not do non-trivial unswitch while optimizing for size.
419    if (OptimizeForSize || F->hasFnAttr(Attribute::OptimizeForSize))
420      return false;
421
422    // FIXME: This is overly conservative because it does not take into
423    // consideration code simplification opportunities and code that can
424    // be shared by the resultant unswitched loops.
425    CodeMetrics Metrics;
426    for (Loop::block_iterator I = currentLoop->block_begin(),
427           E = currentLoop->block_end();
428         I != E; ++I)
429      Metrics.analyzeBasicBlock(*I);
430
431    // Limit the number of instructions to avoid causing significant code
432    // expansion, and the number of basic blocks, to avoid loops with
433    // large numbers of branches which cause loop unswitching to go crazy.
434    // This is a very ad-hoc heuristic.
435    if (Metrics.NumInsts > Threshold ||
436        Metrics.NumBlocks * 5 > Threshold ||
437        Metrics.NeverInline) {
438      DEBUG(errs() << "NOT unswitching loop %"
439            << currentLoop->getHeader()->getName() << ", cost too high: "
440            << currentLoop->getBlocks().size() << "\n");
441      return false;
442    }
443  }
444
445  Constant *CondVal;
446  BasicBlock *ExitBlock;
447  if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) {
448    UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock);
449  } else {
450    UnswitchNontrivialCondition(LoopCond, Val, currentLoop);
451  }
452
453  return true;
454}
455
456// RemapInstruction - Convert the instruction operands from referencing the
457// current values into those specified by ValueMap.
458//
459static inline void RemapInstruction(Instruction *I,
460                                    DenseMap<const Value *, Value*> &ValueMap) {
461  for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
462    Value *Op = I->getOperand(op);
463    DenseMap<const Value *, Value*>::iterator It = ValueMap.find(Op);
464    if (It != ValueMap.end()) Op = It->second;
465    I->setOperand(op, Op);
466  }
467}
468
469/// CloneLoop - Recursively clone the specified loop and all of its children,
470/// mapping the blocks with the specified map.
471static Loop *CloneLoop(Loop *L, Loop *PL, DenseMap<const Value*, Value*> &VM,
472                       LoopInfo *LI, LPPassManager *LPM) {
473  Loop *New = new Loop();
474
475  LPM->insertLoop(New, PL);
476
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();
479       I != E; ++I)
480    if (LI->getLoopFor(*I) == L)
481      New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), LI->getBase());
482
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, LPM);
486
487  return New;
488}
489
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.
493void LoopUnswitch::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<ConstantInt>(Val) ||
501      Val->getType() != Type::getInt1Ty(LIC->getContext()))
502    BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val, "tmp");
503  else if (Val != ConstantInt::getTrue(Val->getContext()))
504    // We want to enter the new loop when the condition is true.
505    std::swap(TrueDest, FalseDest);
506
507  // Insert the new branch.
508  BranchInst *BI = BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt);
509
510  // If either edge is critical, split it. This helps preserve LoopSimplify
511  // form for enclosing loops.
512  SplitCriticalEdge(BI, 0, this);
513  SplitCriticalEdge(BI, 1, this);
514}
515
516/// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable
517/// condition in it (a cond branch from its header block to its latch block,
518/// where the path through the loop that doesn't execute its body has no
519/// side-effects), unswitch it.  This doesn't involve any code duplication, just
520/// moving the conditional branch outside of the loop and updating loop info.
521void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond,
522                                            Constant *Val,
523                                            BasicBlock *ExitBlock) {
524  DEBUG(errs() << "loop-unswitch: Trivial-Unswitch loop %"
525        << loopHeader->getName() << " [" << L->getBlocks().size()
526        << " blocks] in Function " << L->getHeader()->getParent()->getName()
527        << " on cond: " << *Val << " == " << *Cond << "\n");
528
529  // First step, split the preheader, so that we know that there is a safe place
530  // to insert the conditional branch.  We will change loopPreheader to have a
531  // conditional branch on Cond.
532  BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, this);
533
534  // Now that we have a place to insert the conditional branch, create a place
535  // to branch to: this is the exit block out of the loop that we should
536  // short-circuit to.
537
538  // Split this block now, so that the loop maintains its exit block, and so
539  // that the jump from the preheader can execute the contents of the exit block
540  // without actually branching to it (the exit block should be dominated by the
541  // loop header, not the preheader).
542  assert(!L->contains(ExitBlock) && "Exit block is in the loop?");
543  BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), this);
544
545  // Okay, now we have a position to branch from and a position to branch to,
546  // insert the new conditional branch.
547  EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH,
548                                 loopPreheader->getTerminator());
549  LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L);
550  loopPreheader->getTerminator()->eraseFromParent();
551
552  // We need to reprocess this loop, it could be unswitched again.
553  redoLoop = true;
554
555  // Now that we know that the loop is never entered when this condition is a
556  // particular value, rewrite the loop with this info.  We know that this will
557  // at least eliminate the old branch.
558  RewriteLoopBodyWithConditionConstant(L, Cond, Val, false);
559  ++NumTrivial;
560}
561
562/// SplitExitEdges - Split all of the edges from inside the loop to their exit
563/// blocks.  Update the appropriate Phi nodes as we do so.
564void LoopUnswitch::SplitExitEdges(Loop *L,
565                                const SmallVector<BasicBlock *, 8> &ExitBlocks)
566{
567
568  for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
569    BasicBlock *ExitBlock = ExitBlocks[i];
570    SmallVector<BasicBlock *, 4> Preds(pred_begin(ExitBlock),
571                                       pred_end(ExitBlock));
572    SplitBlockPredecessors(ExitBlock, Preds.data(), Preds.size(),
573                           ".us-lcssa", this);
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(errs() << "loop-unswitch: Unswitching loop %"
584        << loopHeader->getName() << " [" << L->getBlocks().size()
585        << " blocks] in Function " << F->getName()
586        << " when '" << *Val << "' == " << *LIC << "\n");
587
588  LoopBlocks.clear();
589  NewBlocks.clear();
590
591  // First step, split the preheader and exit blocks, and add these blocks to
592  // the LoopBlocks list.
593  BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, this);
594  LoopBlocks.push_back(NewPreheader);
595
596  // We want the loop to come after the preheader, but before the exit blocks.
597  LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
598
599  SmallVector<BasicBlock*, 8> ExitBlocks;
600  L->getUniqueExitBlocks(ExitBlocks);
601
602  // Split all of the edges from inside the loop to their exit blocks.  Update
603  // the appropriate Phi nodes as we do so.
604  SplitExitEdges(L, ExitBlocks);
605
606  // The exit blocks may have been changed due to edge splitting, recompute.
607  ExitBlocks.clear();
608  L->getUniqueExitBlocks(ExitBlocks);
609
610  // Add exit blocks to the loop blocks.
611  LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end());
612
613  // Next step, clone all of the basic blocks that make up the loop (including
614  // the loop preheader and exit blocks), keeping track of the mapping between
615  // the instructions and blocks.
616  NewBlocks.reserve(LoopBlocks.size());
617  DenseMap<const Value*, Value*> ValueMap;
618  for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
619    BasicBlock *New = CloneBasicBlock(LoopBlocks[i], ValueMap, ".us", F);
620    NewBlocks.push_back(New);
621    ValueMap[LoopBlocks[i]] = New;  // Keep the BB mapping.
622    LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], New, L);
623  }
624
625  // Splice the newly inserted blocks into the function right before the
626  // original preheader.
627  F->getBasicBlockList().splice(LoopBlocks[0], F->getBasicBlockList(),
628                                NewBlocks[0], F->end());
629
630  // Now we create the new Loop object for the versioned loop.
631  Loop *NewLoop = CloneLoop(L, L->getParentLoop(), ValueMap, LI, LPM);
632  Loop *ParentLoop = L->getParentLoop();
633  if (ParentLoop) {
634    // Make sure to add the cloned preheader and exit blocks to the parent loop
635    // as well.
636    ParentLoop->addBasicBlockToLoop(NewBlocks[0], LI->getBase());
637  }
638
639  for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
640    BasicBlock *NewExit = cast<BasicBlock>(ValueMap[ExitBlocks[i]]);
641    // The new exit block should be in the same loop as the old one.
642    if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i]))
643      ExitBBLoop->addBasicBlockToLoop(NewExit, LI->getBase());
644
645    assert(NewExit->getTerminator()->getNumSuccessors() == 1 &&
646           "Exit block should have been split to have one successor!");
647    BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0);
648
649    // If the successor of the exit block had PHI nodes, add an entry for
650    // NewExit.
651    PHINode *PN;
652    for (BasicBlock::iterator I = ExitSucc->begin();
653         (PN = dyn_cast<PHINode>(I)); ++I) {
654      Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]);
655      DenseMap<const Value *, Value*>::iterator It = ValueMap.find(V);
656      if (It != ValueMap.end()) V = It->second;
657      PN->addIncoming(V, NewExit);
658    }
659  }
660
661  // Rewrite the code to refer to itself.
662  for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i)
663    for (BasicBlock::iterator I = NewBlocks[i]->begin(),
664           E = NewBlocks[i]->end(); I != E; ++I)
665      RemapInstruction(I, ValueMap);
666
667  // Rewrite the original preheader to select between versions of the loop.
668  BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator());
669  assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] &&
670         "Preheader splitting did not work correctly!");
671
672  // Emit the new branch that selects between the two versions of this loop.
673  EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR);
674  LPM->deleteSimpleAnalysisValue(OldBR, L);
675  OldBR->eraseFromParent();
676
677  LoopProcessWorklist.push_back(NewLoop);
678  redoLoop = true;
679
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
684  // It's possible that simplifying one loop could cause the other to be
685  // deleted.  If so, don't simplify it.
686  if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop)
687    RewriteLoopBodyWithConditionConstant(NewLoop, LIC, Val, true);
688
689}
690
691/// RemoveFromWorklist - Remove all instances of I from the worklist vector
692/// specified.
693static void RemoveFromWorklist(Instruction *I,
694                               std::vector<Instruction*> &Worklist) {
695  std::vector<Instruction*>::iterator WI = std::find(Worklist.begin(),
696                                                     Worklist.end(), I);
697  while (WI != Worklist.end()) {
698    unsigned Offset = WI-Worklist.begin();
699    Worklist.erase(WI);
700    WI = std::find(Worklist.begin()+Offset, Worklist.end(), I);
701  }
702}
703
704/// ReplaceUsesOfWith - When we find that I really equals V, remove I from the
705/// program, replacing all uses with V and update the worklist.
706static void ReplaceUsesOfWith(Instruction *I, Value *V,
707                              std::vector<Instruction*> &Worklist,
708                              Loop *L, LPPassManager *LPM) {
709  DEBUG(errs() << "Replace with '" << *V << "': " << *I);
710
711  // Add uses to the worklist, which may be dead now.
712  for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
713    if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
714      Worklist.push_back(Use);
715
716  // Add users to the worklist which may be simplified now.
717  for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
718       UI != E; ++UI)
719    Worklist.push_back(cast<Instruction>(*UI));
720  LPM->deleteSimpleAnalysisValue(I, L);
721  RemoveFromWorklist(I, Worklist);
722  I->replaceAllUsesWith(V);
723  I->eraseFromParent();
724  ++NumSimplify;
725}
726
727/// RemoveBlockIfDead - If the specified block is dead, remove it, update loop
728/// information, and remove any dead successors it has.
729///
730void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB,
731                                     std::vector<Instruction*> &Worklist,
732                                     Loop *L) {
733  if (pred_begin(BB) != pred_end(BB)) {
734    // This block isn't dead, since an edge to BB was just removed, see if there
735    // are any easy simplifications we can do now.
736    if (BasicBlock *Pred = BB->getSinglePredecessor()) {
737      // If it has one pred, fold phi nodes in BB.
738      while (isa<PHINode>(BB->begin()))
739        ReplaceUsesOfWith(BB->begin(),
740                          cast<PHINode>(BB->begin())->getIncomingValue(0),
741                          Worklist, L, LPM);
742
743      // If this is the header of a loop and the only pred is the latch, we now
744      // have an unreachable loop.
745      if (Loop *L = LI->getLoopFor(BB))
746        if (loopHeader == BB && L->contains(Pred)) {
747          // Remove the branch from the latch to the header block, this makes
748          // the header dead, which will make the latch dead (because the header
749          // dominates the latch).
750          LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L);
751          Pred->getTerminator()->eraseFromParent();
752          new UnreachableInst(BB->getContext(), Pred);
753
754          // The loop is now broken, remove it from LI.
755          RemoveLoopFromHierarchy(L);
756
757          // Reprocess the header, which now IS dead.
758          RemoveBlockIfDead(BB, Worklist, L);
759          return;
760        }
761
762      // If pred ends in a uncond branch, add uncond branch to worklist so that
763      // the two blocks will get merged.
764      if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator()))
765        if (BI->isUnconditional())
766          Worklist.push_back(BI);
767    }
768    return;
769  }
770
771  DEBUG(errs() << "Nuking dead block: " << *BB);
772
773  // Remove the instructions in the basic block from the worklist.
774  for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
775    RemoveFromWorklist(I, Worklist);
776
777    // Anything that uses the instructions in this basic block should have their
778    // uses replaced with undefs.
779    // If I is not void type then replaceAllUsesWith undef.
780    // This allows ValueHandlers and custom metadata to adjust itself.
781    if (!I->getType()->isVoidTy())
782      I->replaceAllUsesWith(UndefValue::get(I->getType()));
783  }
784
785  // If this is the edge to the header block for a loop, remove the loop and
786  // promote all subloops.
787  if (Loop *BBLoop = LI->getLoopFor(BB)) {
788    if (BBLoop->getLoopLatch() == BB)
789      RemoveLoopFromHierarchy(BBLoop);
790  }
791
792  // Remove the block from the loop info, which removes it from any loops it
793  // was in.
794  LI->removeBlock(BB);
795
796
797  // Remove phi node entries in successors for this block.
798  TerminatorInst *TI = BB->getTerminator();
799  SmallVector<BasicBlock*, 4> Succs;
800  for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
801    Succs.push_back(TI->getSuccessor(i));
802    TI->getSuccessor(i)->removePredecessor(BB);
803  }
804
805  // Unique the successors, remove anything with multiple uses.
806  array_pod_sort(Succs.begin(), Succs.end());
807  Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end());
808
809  // Remove the basic block, including all of the instructions contained in it.
810  LPM->deleteSimpleAnalysisValue(BB, L);
811  BB->eraseFromParent();
812  // Remove successor blocks here that are not dead, so that we know we only
813  // have dead blocks in this list.  Nondead blocks have a way of becoming dead,
814  // then getting removed before we revisit them, which is badness.
815  //
816  for (unsigned i = 0; i != Succs.size(); ++i)
817    if (pred_begin(Succs[i]) != pred_end(Succs[i])) {
818      // One exception is loop headers.  If this block was the preheader for a
819      // loop, then we DO want to visit the loop so the loop gets deleted.
820      // We know that if the successor is a loop header, that this loop had to
821      // be the preheader: the case where this was the latch block was handled
822      // above and headers can only have two predecessors.
823      if (!LI->isLoopHeader(Succs[i])) {
824        Succs.erase(Succs.begin()+i);
825        --i;
826      }
827    }
828
829  for (unsigned i = 0, e = Succs.size(); i != e; ++i)
830    RemoveBlockIfDead(Succs[i], Worklist, L);
831}
832
833/// RemoveLoopFromHierarchy - We have discovered that the specified loop has
834/// become unwrapped, either because the backedge was deleted, or because the
835/// edge into the header was removed.  If the edge into the header from the
836/// latch block was removed, the loop is unwrapped but subloops are still alive,
837/// so they just reparent loops.  If the loops are actually dead, they will be
838/// removed later.
839void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) {
840  LPM->deleteLoopFromQueue(L);
841  RemoveLoopFromWorklist(L);
842}
843
844// RewriteLoopBodyWithConditionConstant - We know either that the value LIC has
845// the value specified by Val in the specified loop, or we know it does NOT have
846// that value.  Rewrite any uses of LIC or of properties correlated to it.
847void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
848                                                        Constant *Val,
849                                                        bool IsEqual) {
850  assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?");
851
852  // FIXME: Support correlated properties, like:
853  //  for (...)
854  //    if (li1 < li2)
855  //      ...
856  //    if (li1 > li2)
857  //      ...
858
859  // FOLD boolean conditions (X|LIC), (X&LIC).  Fold conditional branches,
860  // selects, switches.
861  std::vector<User*> Users(LIC->use_begin(), LIC->use_end());
862  std::vector<Instruction*> Worklist;
863  LLVMContext &Context = Val->getContext();
864
865
866  // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC
867  // in the loop with the appropriate one directly.
868  if (IsEqual || (isa<ConstantInt>(Val) &&
869      Val->getType() == Type::getInt1Ty(Val->getContext()))) {
870    Value *Replacement;
871    if (IsEqual)
872      Replacement = Val;
873    else
874      Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()),
875                                     !cast<ConstantInt>(Val)->getZExtValue());
876
877    for (unsigned i = 0, e = Users.size(); i != e; ++i)
878      if (Instruction *U = cast<Instruction>(Users[i])) {
879        if (!L->contains(U->getParent()))
880          continue;
881        U->replaceUsesOfWith(LIC, Replacement);
882        Worklist.push_back(U);
883      }
884  } else {
885    // Otherwise, we don't know the precise value of LIC, but we do know that it
886    // is certainly NOT "Val".  As such, simplify any uses in the loop that we
887    // can.  This case occurs when we unswitch switch statements.
888    for (unsigned i = 0, e = Users.size(); i != e; ++i)
889      if (Instruction *U = cast<Instruction>(Users[i])) {
890        if (!L->contains(U->getParent()))
891          continue;
892
893        Worklist.push_back(U);
894
895        // If we know that LIC is not Val, use this info to simplify code.
896        if (SwitchInst *SI = dyn_cast<SwitchInst>(U)) {
897          for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) {
898            if (SI->getCaseValue(i) == Val) {
899              // Found a dead case value.  Don't remove PHI nodes in the
900              // successor if they become single-entry, those PHI nodes may
901              // be in the Users list.
902
903              // FIXME: This is a hack.  We need to keep the successor around
904              // and hooked up so as to preserve the loop structure, because
905              // trying to update it is complicated.  So instead we preserve the
906              // loop structure and put the block on a dead code path.
907              BasicBlock *Switch = SI->getParent();
908              SplitEdge(Switch, SI->getSuccessor(i), this);
909              // Compute the successors instead of relying on the return value
910              // of SplitEdge, since it may have split the switch successor
911              // after PHI nodes.
912              BasicBlock *NewSISucc = SI->getSuccessor(i);
913              BasicBlock *OldSISucc = *succ_begin(NewSISucc);
914              // Create an "unreachable" destination.
915              BasicBlock *Abort = BasicBlock::Create(Context, "us-unreachable",
916                                                     Switch->getParent(),
917                                                     OldSISucc);
918              new UnreachableInst(Context, Abort);
919              // Force the new case destination to branch to the "unreachable"
920              // block while maintaining a (dead) CFG edge to the old block.
921              NewSISucc->getTerminator()->eraseFromParent();
922              BranchInst::Create(Abort, OldSISucc,
923                                 ConstantInt::getTrue(Context), NewSISucc);
924              // Release the PHI operands for this edge.
925              for (BasicBlock::iterator II = NewSISucc->begin();
926                   PHINode *PN = dyn_cast<PHINode>(II); ++II)
927                PN->setIncomingValue(PN->getBasicBlockIndex(Switch),
928                                     UndefValue::get(PN->getType()));
929              // Tell the domtree about the new block. We don't fully update the
930              // domtree here -- instead we force it to do a full recomputation
931              // after the pass is complete -- but we do need to inform it of
932              // new blocks.
933              if (DT)
934                DT->addNewBlock(Abort, NewSISucc);
935              break;
936            }
937          }
938        }
939
940        // TODO: We could do other simplifications, for example, turning
941        // LIC == Val -> false.
942      }
943  }
944
945  SimplifyCode(Worklist, L);
946}
947
948/// SimplifyCode - Okay, now that we have simplified some instructions in the
949/// loop, walk over it and constant prop, dce, and fold control flow where
950/// possible.  Note that this is effectively a very simple loop-structure-aware
951/// optimizer.  During processing of this loop, L could very well be deleted, so
952/// it must not be used.
953///
954/// FIXME: When the loop optimizer is more mature, separate this out to a new
955/// pass.
956///
957void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) {
958  while (!Worklist.empty()) {
959    Instruction *I = Worklist.back();
960    Worklist.pop_back();
961
962    // Simple constant folding.
963    if (Constant *C = ConstantFoldInstruction(I)) {
964      ReplaceUsesOfWith(I, C, Worklist, L, LPM);
965      continue;
966    }
967
968    // Simple DCE.
969    if (isInstructionTriviallyDead(I)) {
970      DEBUG(errs() << "Remove dead instruction '" << *I);
971
972      // Add uses to the worklist, which may be dead now.
973      for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
974        if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
975          Worklist.push_back(Use);
976      LPM->deleteSimpleAnalysisValue(I, L);
977      RemoveFromWorklist(I, Worklist);
978      I->eraseFromParent();
979      ++NumSimplify;
980      continue;
981    }
982
983    // Special case hacks that appear commonly in unswitched code.
984    switch (I->getOpcode()) {
985    case Instruction::Select:
986      if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(0))) {
987        ReplaceUsesOfWith(I, I->getOperand(!CB->getZExtValue()+1), Worklist, L,
988                          LPM);
989        continue;
990      }
991      break;
992    case Instruction::And:
993      if (isa<ConstantInt>(I->getOperand(0)) &&
994          // constant -> RHS
995          I->getOperand(0)->getType() == Type::getInt1Ty(I->getContext()))
996        cast<BinaryOperator>(I)->swapOperands();
997      if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
998        if (CB->getType() == Type::getInt1Ty(I->getContext())) {
999          if (CB->isOne())      // X & 1 -> X
1000            ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM);
1001          else                  // X & 0 -> 0
1002            ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM);
1003          continue;
1004        }
1005      break;
1006    case Instruction::Or:
1007      if (isa<ConstantInt>(I->getOperand(0)) &&
1008          // constant -> RHS
1009          I->getOperand(0)->getType() == Type::getInt1Ty(I->getContext()))
1010        cast<BinaryOperator>(I)->swapOperands();
1011      if (ConstantInt *CB = dyn_cast<ConstantInt>(I->getOperand(1)))
1012        if (CB->getType() == Type::getInt1Ty(I->getContext())) {
1013          if (CB->isOne())   // X | 1 -> 1
1014            ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM);
1015          else                  // X | 0 -> X
1016            ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM);
1017          continue;
1018        }
1019      break;
1020    case Instruction::Br: {
1021      BranchInst *BI = cast<BranchInst>(I);
1022      if (BI->isUnconditional()) {
1023        // If BI's parent is the only pred of the successor, fold the two blocks
1024        // together.
1025        BasicBlock *Pred = BI->getParent();
1026        BasicBlock *Succ = BI->getSuccessor(0);
1027        BasicBlock *SinglePred = Succ->getSinglePredecessor();
1028        if (!SinglePred) continue;  // Nothing to do.
1029        assert(SinglePred == Pred && "CFG broken");
1030
1031        DEBUG(errs() << "Merging blocks: " << Pred->getName() << " <- "
1032              << Succ->getName() << "\n");
1033
1034        // Resolve any single entry PHI nodes in Succ.
1035        while (PHINode *PN = dyn_cast<PHINode>(Succ->begin()))
1036          ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM);
1037
1038        // Move all of the successor contents from Succ to Pred.
1039        Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(),
1040                                   Succ->end());
1041        LPM->deleteSimpleAnalysisValue(BI, L);
1042        BI->eraseFromParent();
1043        RemoveFromWorklist(BI, Worklist);
1044
1045        // If Succ has any successors with PHI nodes, update them to have
1046        // entries coming from Pred instead of Succ.
1047        Succ->replaceAllUsesWith(Pred);
1048
1049        // Remove Succ from the loop tree.
1050        LI->removeBlock(Succ);
1051        LPM->deleteSimpleAnalysisValue(Succ, L);
1052        Succ->eraseFromParent();
1053        ++NumSimplify;
1054      } else if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){
1055        // Conditional branch.  Turn it into an unconditional branch, then
1056        // remove dead blocks.
1057        break;  // FIXME: Enable.
1058
1059        DEBUG(errs() << "Folded branch: " << *BI);
1060        BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue());
1061        BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue());
1062        DeadSucc->removePredecessor(BI->getParent(), true);
1063        Worklist.push_back(BranchInst::Create(LiveSucc, BI));
1064        LPM->deleteSimpleAnalysisValue(BI, L);
1065        BI->eraseFromParent();
1066        RemoveFromWorklist(BI, Worklist);
1067        ++NumSimplify;
1068
1069        RemoveBlockIfDead(DeadSucc, Worklist, L);
1070      }
1071      break;
1072    }
1073    }
1074  }
1075}
1076