LICM.cpp revision 2741c971044d2165be572749b94398043caccfeb
1//===-- LICM.cpp - Loop Invariant Code Motion Pass ------------------------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file was developed by the LLVM research group and is distributed under
6// the University of Illinois Open Source License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This pass performs loop invariant code motion, attempting to remove as much
11// code from the body of a loop as possible.  It does this by either hoisting
12// code into the preheader block, or by sinking code to the exit blocks if it is
13// safe.  This pass also promotes must-aliased memory locations in the loop to
14// live in registers, thus hoisting and sinking "invariant" loads and stores.
15//
16// This pass uses alias analysis for two purposes:
17//
18//  1. Moving loop invariant loads and calls out of loops.  If we can determine
19//     that a load or call inside of a loop never aliases anything stored to,
20//     we can hoist it or sink it like any other instruction.
21//  2. Scalar Promotion of Memory - If there is a store instruction inside of
22//     the loop, we try to move the store to happen AFTER the loop instead of
23//     inside of the loop.  This can only happen if a few conditions are true:
24//       A. The pointer stored through is loop invariant
25//       B. There are no stores or loads in the loop which _may_ alias the
26//          pointer.  There are no calls in the loop which mod/ref the pointer.
27//     If these conditions are true, we can promote the loads and stores in the
28//     loop of the pointer to use a temporary alloca'd variable.  We then use
29//     the mem2reg functionality to construct the appropriate SSA form for the
30//     variable.
31//
32//===----------------------------------------------------------------------===//
33
34#include "llvm/Transforms/Scalar.h"
35#include "llvm/DerivedTypes.h"
36#include "llvm/Instructions.h"
37#include "llvm/Target/TargetData.h"
38#include "llvm/Analysis/LoopInfo.h"
39#include "llvm/Analysis/AliasAnalysis.h"
40#include "llvm/Analysis/AliasSetTracker.h"
41#include "llvm/Analysis/Dominators.h"
42#include "llvm/Support/CFG.h"
43#include "llvm/Transforms/Utils/PromoteMemToReg.h"
44#include "llvm/Transforms/Utils/Local.h"
45#include "Support/CommandLine.h"
46#include "Support/Debug.h"
47#include "Support/Statistic.h"
48#include <algorithm>
49using namespace llvm;
50
51namespace {
52  cl::opt<bool>
53  DisablePromotion("disable-licm-promotion", cl::Hidden,
54                   cl::desc("Disable memory promotion in LICM pass"));
55
56  Statistic<> NumSunk("licm", "Number of instructions sunk out of loop");
57  Statistic<> NumHoisted("licm", "Number of instructions hoisted out of loop");
58  Statistic<> NumMovedLoads("licm", "Number of load insts hoisted or sunk");
59  Statistic<> NumMovedCalls("licm", "Number of call insts hoisted or sunk");
60  Statistic<> NumPromoted("licm",
61                          "Number of memory locations promoted to registers");
62
63  struct LICM : public FunctionPass {
64    virtual bool runOnFunction(Function &F);
65
66    /// This transformation requires natural loop information & requires that
67    /// loop preheaders be inserted into the CFG...
68    ///
69    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
70      AU.setPreservesCFG();
71      AU.addRequiredID(LoopSimplifyID);
72      AU.addRequired<LoopInfo>();
73      AU.addRequired<DominatorTree>();
74      AU.addRequired<DominanceFrontier>();  // For scalar promotion (mem2reg)
75      AU.addRequired<AliasAnalysis>();
76    }
77
78  private:
79    // Various analyses that we use...
80    AliasAnalysis *AA;       // Current AliasAnalysis information
81    LoopInfo      *LI;       // Current LoopInfo
82    DominatorTree *DT;       // Dominator Tree for the current Loop...
83    DominanceFrontier *DF;   // Current Dominance Frontier
84
85    // State that is updated as we process loops
86    bool Changed;            // Set to true when we change anything.
87    BasicBlock *Preheader;   // The preheader block of the current loop...
88    Loop *CurLoop;           // The current loop we are working on...
89    AliasSetTracker *CurAST; // AliasSet information for the current loop...
90
91    /// visitLoop - Hoist expressions out of the specified loop...
92    ///
93    void visitLoop(Loop *L, AliasSetTracker &AST);
94
95    /// SinkRegion - Walk the specified region of the CFG (defined by all blocks
96    /// dominated by the specified block, and that are in the current loop) in
97    /// reverse depth first order w.r.t the DominatorTree.  This allows us to
98    /// visit uses before definitions, allowing us to sink a loop body in one
99    /// pass without iteration.
100    ///
101    void SinkRegion(DominatorTree::Node *N);
102
103    /// HoistRegion - Walk the specified region of the CFG (defined by all
104    /// blocks dominated by the specified block, and that are in the current
105    /// loop) in depth first order w.r.t the DominatorTree.  This allows us to
106    /// visit definitions before uses, allowing us to hoist a loop body in one
107    /// pass without iteration.
108    ///
109    void HoistRegion(DominatorTree::Node *N);
110
111    /// inSubLoop - Little predicate that returns true if the specified basic
112    /// block is in a subloop of the current one, not the current one itself.
113    ///
114    bool inSubLoop(BasicBlock *BB) {
115      assert(CurLoop->contains(BB) && "Only valid if BB is IN the loop");
116      for (Loop::iterator I = CurLoop->begin(), E = CurLoop->end(); I != E; ++I)
117        if ((*I)->contains(BB))
118          return true;  // A subloop actually contains this block!
119      return false;
120    }
121
122    /// isExitBlockDominatedByBlockInLoop - This method checks to see if the
123    /// specified exit block of the loop is dominated by the specified block
124    /// that is in the body of the loop.  We use these constraints to
125    /// dramatically limit the amount of the dominator tree that needs to be
126    /// searched.
127    bool isExitBlockDominatedByBlockInLoop(BasicBlock *ExitBlock,
128                                           BasicBlock *BlockInLoop) const {
129      // If the block in the loop is the loop header, it must be dominated!
130      BasicBlock *LoopHeader = CurLoop->getHeader();
131      if (BlockInLoop == LoopHeader)
132        return true;
133
134      DominatorTree::Node *BlockInLoopNode = DT->getNode(BlockInLoop);
135      DominatorTree::Node *IDom            = DT->getNode(ExitBlock);
136
137      // Because the exit block is not in the loop, we know we have to get _at
138      // least_ its immediate dominator.
139      do {
140        // Get next Immediate Dominator.
141        IDom = IDom->getIDom();
142
143        // If we have got to the header of the loop, then the instructions block
144        // did not dominate the exit node, so we can't hoist it.
145        if (IDom->getBlock() == LoopHeader)
146          return false;
147
148      } while (IDom != BlockInLoopNode);
149
150      return true;
151    }
152
153    /// sink - When an instruction is found to only be used outside of the loop,
154    /// this function moves it to the exit blocks and patches up SSA form as
155    /// needed.
156    ///
157    void sink(Instruction &I);
158
159    /// hoist - When an instruction is found to only use loop invariant operands
160    /// that is safe to hoist, this instruction is called to do the dirty work.
161    ///
162    void hoist(Instruction &I);
163
164    /// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it
165    /// is not a trapping instruction or if it is a trapping instruction and is
166    /// guaranteed to execute.
167    ///
168    bool isSafeToExecuteUnconditionally(Instruction &I);
169
170    /// pointerInvalidatedByLoop - Return true if the body of this loop may
171    /// store into the memory location pointed to by V.
172    ///
173    bool pointerInvalidatedByLoop(Value *V) {
174      // Check to see if any of the basic blocks in CurLoop invalidate *V.
175      return CurAST->getAliasSetForPointer(V, 0).isMod();
176    }
177
178    bool canSinkOrHoistInst(Instruction &I);
179    bool isLoopInvariantInst(Instruction &I);
180    bool isNotUsedInLoop(Instruction &I);
181
182    /// PromoteValuesInLoop - Look at the stores in the loop and promote as many
183    /// to scalars as we can.
184    ///
185    void PromoteValuesInLoop();
186
187    /// findPromotableValuesInLoop - Check the current loop for stores to
188    /// definite pointers, which are not loaded and stored through may aliases.
189    /// If these are found, create an alloca for the value, add it to the
190    /// PromotedValues list, and keep track of the mapping from value to
191    /// alloca...
192    ///
193    void findPromotableValuesInLoop(
194                   std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues,
195                                    std::map<Value*, AllocaInst*> &Val2AlMap);
196  };
197
198  RegisterOpt<LICM> X("licm", "Loop Invariant Code Motion");
199}
200
201FunctionPass *llvm::createLICMPass() { return new LICM(); }
202
203/// runOnFunction - For LICM, this simply traverses the loop structure of the
204/// function, hoisting expressions out of loops if possible.
205///
206bool LICM::runOnFunction(Function &) {
207  Changed = false;
208
209  // Get our Loop and Alias Analysis information...
210  LI = &getAnalysis<LoopInfo>();
211  AA = &getAnalysis<AliasAnalysis>();
212  DF = &getAnalysis<DominanceFrontier>();
213  DT = &getAnalysis<DominatorTree>();
214
215  // Hoist expressions out of all of the top-level loops.
216  for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I) {
217    AliasSetTracker AST(*AA);
218    visitLoop(*I, AST);
219  }
220  return Changed;
221}
222
223
224/// visitLoop - Hoist expressions out of the specified loop...
225///
226void LICM::visitLoop(Loop *L, AliasSetTracker &AST) {
227  // Recurse through all subloops before we process this loop...
228  for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) {
229    AliasSetTracker SubAST(*AA);
230    visitLoop(*I, SubAST);
231
232    // Incorporate information about the subloops into this loop...
233    AST.add(SubAST);
234  }
235  CurLoop = L;
236  CurAST = &AST;
237
238  // Get the preheader block to move instructions into...
239  Preheader = L->getLoopPreheader();
240  assert(Preheader&&"Preheader insertion pass guarantees we have a preheader!");
241
242  // Loop over the body of this loop, looking for calls, invokes, and stores.
243  // Because subloops have already been incorporated into AST, we skip blocks in
244  // subloops.
245  //
246  for (std::vector<BasicBlock*>::const_iterator I = L->getBlocks().begin(),
247         E = L->getBlocks().end(); I != E; ++I)
248    if (LI->getLoopFor(*I) == L)        // Ignore blocks in subloops...
249      AST.add(**I);                     // Incorporate the specified basic block
250
251  // We want to visit all of the instructions in this loop... that are not parts
252  // of our subloops (they have already had their invariants hoisted out of
253  // their loop, into this loop, so there is no need to process the BODIES of
254  // the subloops).
255  //
256  // Traverse the body of the loop in depth first order on the dominator tree so
257  // that we are guaranteed to see definitions before we see uses.  This allows
258  // us to sink instructions in one pass, without iteration.  AFter sinking
259  // instructions, we perform another pass to hoist them out of the loop.
260  //
261  SinkRegion(DT->getNode(L->getHeader()));
262  HoistRegion(DT->getNode(L->getHeader()));
263
264  // Now that all loop invariants have been removed from the loop, promote any
265  // memory references to scalars that we can...
266  if (!DisablePromotion)
267    PromoteValuesInLoop();
268
269  // Clear out loops state information for the next iteration
270  CurLoop = 0;
271  Preheader = 0;
272}
273
274/// SinkRegion - Walk the specified region of the CFG (defined by all blocks
275/// dominated by the specified block, and that are in the current loop) in
276/// reverse depth first order w.r.t the DominatorTree.  This allows us to visit
277/// uses before definitions, allowing us to sink a loop body in one pass without
278/// iteration.
279///
280void LICM::SinkRegion(DominatorTree::Node *N) {
281  assert(N != 0 && "Null dominator tree node?");
282  BasicBlock *BB = N->getBlock();
283
284  // If this subregion is not in the top level loop at all, exit.
285  if (!CurLoop->contains(BB)) return;
286
287  // We are processing blocks in reverse dfo, so process children first...
288  const std::vector<DominatorTree::Node*> &Children = N->getChildren();
289  for (unsigned i = 0, e = Children.size(); i != e; ++i)
290    SinkRegion(Children[i]);
291
292  // Only need to process the contents of this block if it is not part of a
293  // subloop (which would already have been processed).
294  if (inSubLoop(BB)) return;
295
296  for (BasicBlock::iterator II = BB->end(); II != BB->begin(); ) {
297    Instruction &I = *--II;
298
299    // Check to see if we can sink this instruction to the exit blocks
300    // of the loop.  We can do this if the all users of the instruction are
301    // outside of the loop.  In this case, it doesn't even matter if the
302    // operands of the instruction are loop invariant.
303    //
304    if (canSinkOrHoistInst(I) && isNotUsedInLoop(I)) {
305      ++II;
306      sink(I);
307    }
308  }
309}
310
311
312/// HoistRegion - Walk the specified region of the CFG (defined by all blocks
313/// dominated by the specified block, and that are in the current loop) in depth
314/// first order w.r.t the DominatorTree.  This allows us to visit definitions
315/// before uses, allowing us to hoist a loop body in one pass without iteration.
316///
317void LICM::HoistRegion(DominatorTree::Node *N) {
318  assert(N != 0 && "Null dominator tree node?");
319  BasicBlock *BB = N->getBlock();
320
321  // If this subregion is not in the top level loop at all, exit.
322  if (!CurLoop->contains(BB)) return;
323
324  // Only need to process the contents of this block if it is not part of a
325  // subloop (which would already have been processed).
326  if (!inSubLoop(BB))
327    for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) {
328      Instruction &I = *II++;
329
330      // Try hoisting the instruction out to the preheader.  We can only do this
331      // if all of the operands of the instruction are loop invariant and if it
332      // is safe to hoist the instruction.
333      //
334      if (isLoopInvariantInst(I) && canSinkOrHoistInst(I) &&
335          isSafeToExecuteUnconditionally(I))
336          hoist(I);
337      }
338
339  const std::vector<DominatorTree::Node*> &Children = N->getChildren();
340  for (unsigned i = 0, e = Children.size(); i != e; ++i)
341    HoistRegion(Children[i]);
342}
343
344/// canSinkOrHoistInst - Return true if the hoister and sinker can handle this
345/// instruction.
346///
347bool LICM::canSinkOrHoistInst(Instruction &I) {
348  // Loads have extra constraints we have to verify before we can hoist them.
349  if (LoadInst *LI = dyn_cast<LoadInst>(&I)) {
350    if (LI->isVolatile())
351      return false;        // Don't hoist volatile loads!
352
353    // Don't hoist loads which have may-aliased stores in loop.
354    return !pointerInvalidatedByLoop(LI->getOperand(0));
355  } else if (CallInst *CI = dyn_cast<CallInst>(&I)) {
356    // Handle obvious cases efficiently.
357    if (Function *Callee = CI->getCalledFunction()) {
358      if (AA->doesNotAccessMemory(Callee))
359        return true;
360      else if (AA->onlyReadsMemory(Callee)) {
361        // If this call only reads from memory and there are no writes to memory
362        // in the loop, we can hoist or sink the call as appropriate.
363        bool FoundMod = false;
364        for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
365             I != E; ++I) {
366          AliasSet &AS = *I;
367          if (!AS.isForwardingAliasSet() && AS.isMod()) {
368            FoundMod = true;
369            break;
370          }
371        }
372        if (!FoundMod) return true;
373      }
374    }
375
376    // FIXME: This should use mod/ref information to see if we can hoist or sink
377    // the call.
378
379    return false;
380  }
381
382  return isa<BinaryOperator>(I) || isa<ShiftInst>(I) || isa<CastInst>(I) ||
383         isa<SelectInst>(I) ||
384         isa<GetElementPtrInst>(I) || isa<VANextInst>(I) || isa<VAArgInst>(I);
385}
386
387/// isNotUsedInLoop - Return true if the only users of this instruction are
388/// outside of the loop.  If this is true, we can sink the instruction to the
389/// exit blocks of the loop.
390///
391bool LICM::isNotUsedInLoop(Instruction &I) {
392  for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E; ++UI) {
393    Instruction *User = cast<Instruction>(*UI);
394    if (PHINode *PN = dyn_cast<PHINode>(User)) {
395      // PHI node uses occur in predecessor blocks!
396      for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
397        if (PN->getIncomingValue(i) == &I)
398          if (CurLoop->contains(PN->getIncomingBlock(i)))
399            return false;
400    } else if (CurLoop->contains(User->getParent())) {
401      return false;
402    }
403  }
404  return true;
405}
406
407
408/// isLoopInvariantInst - Return true if all operands of this instruction are
409/// loop invariant.  We also filter out non-hoistable instructions here just for
410/// efficiency.
411///
412bool LICM::isLoopInvariantInst(Instruction &I) {
413  // The instruction is loop invariant if all of its operands are loop-invariant
414  for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
415    if (!CurLoop->isLoopInvariant(I.getOperand(i)))
416      return false;
417
418  // If we got this far, the instruction is loop invariant!
419  return true;
420}
421
422/// sink - When an instruction is found to only be used outside of the loop,
423/// this function moves it to the exit blocks and patches up SSA form as needed.
424/// This method is guaranteed to remove the original instruction from its
425/// position, and may either delete it or move it to outside of the loop.
426///
427void LICM::sink(Instruction &I) {
428  DEBUG(std::cerr << "LICM sinking instruction: " << I);
429
430  std::vector<BasicBlock*> ExitBlocks;
431  CurLoop->getExitBlocks(ExitBlocks);
432
433  if (isa<LoadInst>(I)) ++NumMovedLoads;
434  else if (isa<CallInst>(I)) ++NumMovedCalls;
435  ++NumSunk;
436  Changed = true;
437
438  // The case where there is only a single exit node of this loop is common
439  // enough that we handle it as a special (more efficient) case.  It is more
440  // efficient to handle because there are no PHI nodes that need to be placed.
441  if (ExitBlocks.size() == 1) {
442    if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[0], I.getParent())) {
443      // Instruction is not used, just delete it.
444      CurAST->deleteValue(&I);
445      I.getParent()->getInstList().erase(&I);
446    } else {
447      // Move the instruction to the start of the exit block, after any PHI
448      // nodes in it.
449      I.getParent()->getInstList().remove(&I);
450
451      BasicBlock::iterator InsertPt = ExitBlocks[0]->begin();
452      while (isa<PHINode>(InsertPt)) ++InsertPt;
453      ExitBlocks[0]->getInstList().insert(InsertPt, &I);
454    }
455  } else if (ExitBlocks.size() == 0) {
456    // The instruction is actually dead if there ARE NO exit blocks.
457    CurAST->deleteValue(&I);
458    I.getParent()->getInstList().erase(&I);
459  } else {
460    // Otherwise, if we have multiple exits, use the PromoteMem2Reg function to
461    // do all of the hard work of inserting PHI nodes as necessary.  We convert
462    // the value into a stack object to get it to do this.
463
464    // Firstly, we create a stack object to hold the value...
465    AllocaInst *AI = new AllocaInst(I.getType(), 0, I.getName(),
466                                   I.getParent()->getParent()->front().begin());
467
468    // Secondly, insert load instructions for each use of the instruction
469    // outside of the loop.
470    while (!I.use_empty()) {
471      Instruction *U = cast<Instruction>(I.use_back());
472
473      // If the user is a PHI Node, we actually have to insert load instructions
474      // in all predecessor blocks, not in the PHI block itself!
475      if (PHINode *UPN = dyn_cast<PHINode>(U)) {
476        // Only insert into each predecessor once, so that we don't have
477        // different incoming values from the same block!
478        std::map<BasicBlock*, Value*> InsertedBlocks;
479        for (unsigned i = 0, e = UPN->getNumIncomingValues(); i != e; ++i)
480          if (UPN->getIncomingValue(i) == &I) {
481            BasicBlock *Pred = UPN->getIncomingBlock(i);
482            Value *&PredVal = InsertedBlocks[Pred];
483            if (!PredVal) {
484              // Insert a new load instruction right before the terminator in
485              // the predecessor block.
486              PredVal = new LoadInst(AI, "", Pred->getTerminator());
487            }
488
489            UPN->setIncomingValue(i, PredVal);
490          }
491
492      } else {
493        LoadInst *L = new LoadInst(AI, "", U);
494        U->replaceUsesOfWith(&I, L);
495      }
496    }
497
498    // Thirdly, insert a copy of the instruction in each exit block of the loop
499    // that is dominated by the instruction, storing the result into the memory
500    // location.  Be careful not to insert the instruction into any particular
501    // basic block more than once.
502    std::set<BasicBlock*> InsertedBlocks;
503    BasicBlock *InstOrigBB = I.getParent();
504
505    for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
506      BasicBlock *ExitBlock = ExitBlocks[i];
507
508      if (isExitBlockDominatedByBlockInLoop(ExitBlock, InstOrigBB)) {
509        // If we haven't already processed this exit block, do so now.
510        if (InsertedBlocks.insert(ExitBlock).second) {
511          // Insert the code after the last PHI node...
512          BasicBlock::iterator InsertPt = ExitBlock->begin();
513          while (isa<PHINode>(InsertPt)) ++InsertPt;
514
515          // If this is the first exit block processed, just move the original
516          // instruction, otherwise clone the original instruction and insert
517          // the copy.
518          Instruction *New;
519          if (InsertedBlocks.size() == 1) {
520            I.getParent()->getInstList().remove(&I);
521            ExitBlock->getInstList().insert(InsertPt, &I);
522            New = &I;
523          } else {
524            New = I.clone();
525            New->setName(I.getName()+".le");
526            ExitBlock->getInstList().insert(InsertPt, New);
527          }
528
529          // Now that we have inserted the instruction, store it into the alloca
530          new StoreInst(New, AI, InsertPt);
531        }
532      }
533    }
534
535    // If the instruction doesn't dominate any exit blocks, it must be dead.
536    if (InsertedBlocks.empty()) {
537      CurAST->deleteValue(&I);
538      I.getParent()->getInstList().erase(&I);
539    }
540
541    // Finally, promote the fine value to SSA form.
542    std::vector<AllocaInst*> Allocas;
543    Allocas.push_back(AI);
544    PromoteMemToReg(Allocas, *DT, *DF, AA->getTargetData());
545  }
546}
547
548/// hoist - When an instruction is found to only use loop invariant operands
549/// that is safe to hoist, this instruction is called to do the dirty work.
550///
551void LICM::hoist(Instruction &I) {
552  DEBUG(std::cerr << "LICM hoisting to" << Preheader->getName()
553                  << ": " << I);
554
555  // Remove the instruction from its current basic block... but don't delete the
556  // instruction.
557  I.getParent()->getInstList().remove(&I);
558
559  // Insert the new node in Preheader, before the terminator.
560  Preheader->getInstList().insert(Preheader->getTerminator(), &I);
561
562  if (isa<LoadInst>(I)) ++NumMovedLoads;
563  else if (isa<CallInst>(I)) ++NumMovedCalls;
564  ++NumHoisted;
565  Changed = true;
566}
567
568/// isSafeToExecuteUnconditionally - Only sink or hoist an instruction if it is
569/// not a trapping instruction or if it is a trapping instruction and is
570/// guaranteed to execute.
571///
572bool LICM::isSafeToExecuteUnconditionally(Instruction &Inst) {
573  // If it is not a trapping instruction, it is always safe to hoist.
574  if (!Inst.isTrapping()) return true;
575
576  // Otherwise we have to check to make sure that the instruction dominates all
577  // of the exit blocks.  If it doesn't, then there is a path out of the loop
578  // which does not execute this instruction, so we can't hoist it.
579
580  // If the instruction is in the header block for the loop (which is very
581  // common), it is always guaranteed to dominate the exit blocks.  Since this
582  // is a common case, and can save some work, check it now.
583  if (Inst.getParent() == CurLoop->getHeader())
584    return true;
585
586  // Get the exit blocks for the current loop.
587  std::vector<BasicBlock*> ExitBlocks;
588  CurLoop->getExitBlocks(ExitBlocks);
589
590  // For each exit block, get the DT node and walk up the DT until the
591  // instruction's basic block is found or we exit the loop.
592  for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
593    if (!isExitBlockDominatedByBlockInLoop(ExitBlocks[i], Inst.getParent()))
594      return false;
595
596  return true;
597}
598
599
600/// PromoteValuesInLoop - Try to promote memory values to scalars by sinking
601/// stores out of the loop and moving loads to before the loop.  We do this by
602/// looping over the stores in the loop, looking for stores to Must pointers
603/// which are loop invariant.  We promote these memory locations to use allocas
604/// instead.  These allocas can easily be raised to register values by the
605/// PromoteMem2Reg functionality.
606///
607void LICM::PromoteValuesInLoop() {
608  // PromotedValues - List of values that are promoted out of the loop.  Each
609  // value has an alloca instruction for it, and a canonical version of the
610  // pointer.
611  std::vector<std::pair<AllocaInst*, Value*> > PromotedValues;
612  std::map<Value*, AllocaInst*> ValueToAllocaMap; // Map of ptr to alloca
613
614  findPromotableValuesInLoop(PromotedValues, ValueToAllocaMap);
615  if (ValueToAllocaMap.empty()) return;   // If there are values to promote...
616
617  Changed = true;
618  NumPromoted += PromotedValues.size();
619
620  // Emit a copy from the value into the alloca'd value in the loop preheader
621  TerminatorInst *LoopPredInst = Preheader->getTerminator();
622  for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) {
623    // Load from the memory we are promoting...
624    LoadInst *LI = new LoadInst(PromotedValues[i].second,
625                                PromotedValues[i].second->getName()+".promoted",
626                                LoopPredInst);
627    // Store into the temporary alloca...
628    new StoreInst(LI, PromotedValues[i].first, LoopPredInst);
629  }
630
631  // Scan the basic blocks in the loop, replacing uses of our pointers with
632  // uses of the allocas in question.
633  //
634  const std::vector<BasicBlock*> &LoopBBs = CurLoop->getBlocks();
635  for (std::vector<BasicBlock*>::const_iterator I = LoopBBs.begin(),
636         E = LoopBBs.end(); I != E; ++I) {
637    // Rewrite all loads and stores in the block of the pointer...
638    for (BasicBlock::iterator II = (*I)->begin(), E = (*I)->end();
639         II != E; ++II) {
640      if (LoadInst *L = dyn_cast<LoadInst>(II)) {
641        std::map<Value*, AllocaInst*>::iterator
642          I = ValueToAllocaMap.find(L->getOperand(0));
643        if (I != ValueToAllocaMap.end())
644          L->setOperand(0, I->second);    // Rewrite load instruction...
645      } else if (StoreInst *S = dyn_cast<StoreInst>(II)) {
646        std::map<Value*, AllocaInst*>::iterator
647          I = ValueToAllocaMap.find(S->getOperand(1));
648        if (I != ValueToAllocaMap.end())
649          S->setOperand(1, I->second);    // Rewrite store instruction...
650      }
651    }
652  }
653
654  // Now that the body of the loop uses the allocas instead of the original
655  // memory locations, insert code to copy the alloca value back into the
656  // original memory location on all exits from the loop.  Note that we only
657  // want to insert one copy of the code in each exit block, though the loop may
658  // exit to the same block more than once.
659  //
660  std::set<BasicBlock*> ProcessedBlocks;
661
662  std::vector<BasicBlock*> ExitBlocks;
663  CurLoop->getExitBlocks(ExitBlocks);
664  for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
665    if (ProcessedBlocks.insert(ExitBlocks[i]).second) {
666      // Copy all of the allocas into their memory locations...
667      BasicBlock::iterator BI = ExitBlocks[i]->begin();
668      while (isa<PHINode>(*BI))
669        ++BI;             // Skip over all of the phi nodes in the block...
670      Instruction *InsertPos = BI;
671      for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i) {
672        // Load from the alloca...
673        LoadInst *LI = new LoadInst(PromotedValues[i].first, "", InsertPos);
674        // Store into the memory we promoted...
675        new StoreInst(LI, PromotedValues[i].second, InsertPos);
676      }
677    }
678
679  // Now that we have done the deed, use the mem2reg functionality to promote
680  // all of the new allocas we just created into real SSA registers...
681  //
682  std::vector<AllocaInst*> PromotedAllocas;
683  PromotedAllocas.reserve(PromotedValues.size());
684  for (unsigned i = 0, e = PromotedValues.size(); i != e; ++i)
685    PromotedAllocas.push_back(PromotedValues[i].first);
686  PromoteMemToReg(PromotedAllocas, *DT, *DF, AA->getTargetData());
687}
688
689/// findPromotableValuesInLoop - Check the current loop for stores to definite
690/// pointers, which are not loaded and stored through may aliases.  If these are
691/// found, create an alloca for the value, add it to the PromotedValues list,
692/// and keep track of the mapping from value to alloca...
693///
694void LICM::findPromotableValuesInLoop(
695                   std::vector<std::pair<AllocaInst*, Value*> > &PromotedValues,
696                             std::map<Value*, AllocaInst*> &ValueToAllocaMap) {
697  Instruction *FnStart = CurLoop->getHeader()->getParent()->begin()->begin();
698
699  // Loop over all of the alias sets in the tracker object...
700  for (AliasSetTracker::iterator I = CurAST->begin(), E = CurAST->end();
701       I != E; ++I) {
702    AliasSet &AS = *I;
703    // We can promote this alias set if it has a store, if it is a "Must" alias
704    // set, if the pointer is loop invariant, if if we are not eliminating any
705    // volatile loads or stores.
706    if (!AS.isForwardingAliasSet() && AS.isMod() && AS.isMustAlias() &&
707        !AS.isVolatile() && CurLoop->isLoopInvariant(AS.begin()->first)) {
708      assert(AS.begin() != AS.end() &&
709             "Must alias set should have at least one pointer element in it!");
710      Value *V = AS.begin()->first;
711
712      // Check that all of the pointers in the alias set have the same type.  We
713      // cannot (yet) promote a memory location that is loaded and stored in
714      // different sizes.
715      bool PointerOk = true;
716      for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I)
717        if (V->getType() != I->first->getType()) {
718          PointerOk = false;
719          break;
720        }
721
722      if (PointerOk) {
723        const Type *Ty = cast<PointerType>(V->getType())->getElementType();
724        AllocaInst *AI = new AllocaInst(Ty, 0, V->getName()+".tmp", FnStart);
725        PromotedValues.push_back(std::make_pair(AI, V));
726
727        for (AliasSet::iterator I = AS.begin(), E = AS.end(); I != E; ++I)
728          ValueToAllocaMap.insert(std::make_pair(I->first, AI));
729
730        DEBUG(std::cerr << "LICM: Promoting value: " << *V << "\n");
731      }
732    }
733  }
734}
735