When promoting an alloca to registers discard any lifetime intrinsics.
[oota-llvm.git] / lib / Transforms / Utils / LoopSimplify.cpp
index 57bab60e6b1c96dbec597805a01906f8eb6bf06e..f02ffd20bca99b956f11690d98774ef908ef4823 100644 (file)
 //
 //===----------------------------------------------------------------------===//
 
-#define DEBUG_TYPE "loopsimplify"
+#define DEBUG_TYPE "loop-simplify"
 #include "llvm/Transforms/Scalar.h"
 #include "llvm/Constants.h"
 #include "llvm/Instructions.h"
+#include "llvm/IntrinsicInst.h"
 #include "llvm/Function.h"
 #include "llvm/LLVMContext.h"
 #include "llvm/Type.h"
 #include "llvm/Analysis/AliasAnalysis.h"
 #include "llvm/Analysis/Dominators.h"
+#include "llvm/Analysis/InstructionSimplify.h"
 #include "llvm/Analysis/LoopPass.h"
 #include "llvm/Analysis/ScalarEvolution.h"
 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
 #include "llvm/Transforms/Utils/Local.h"
 #include "llvm/Support/CFG.h"
+#include "llvm/Support/Debug.h"
 #include "llvm/ADT/SetOperations.h"
 #include "llvm/ADT/SetVector.h"
 #include "llvm/ADT/Statistic.h"
@@ -63,24 +66,27 @@ STATISTIC(NumNested  , "Number of nested loops split out");
 namespace {
   struct LoopSimplify : public LoopPass {
     static char ID; // Pass identification, replacement for typeid
-    LoopSimplify() : LoopPass(&ID) {}
+    LoopSimplify() : LoopPass(ID) {
+      initializeLoopSimplifyPass(*PassRegistry::getPassRegistry());
+    }
 
     // AA - If we have an alias analysis object to update, this is it, otherwise
     // this is null.
     AliasAnalysis *AA;
     LoopInfo *LI;
     DominatorTree *DT;
+    ScalarEvolution *SE;
     Loop *L;
     virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
 
     virtual void getAnalysisUsage(AnalysisUsage &AU) const {
       // We need loop information to identify the loops...
-      AU.addRequiredTransitive<LoopInfo>();
-      AU.addRequiredTransitive<DominatorTree>();
+      AU.addRequired<DominatorTree>();
+      AU.addPreserved<DominatorTree>();
 
+      AU.addRequired<LoopInfo>();
       AU.addPreserved<LoopInfo>();
-      AU.addPreserved<DominatorTree>();
-      AU.addPreserved<DominanceFrontier>();
+
       AU.addPreserved<AliasAnalysis>();
       AU.addPreserved<ScalarEvolution>();
       AU.addPreservedID(BreakCriticalEdgesID);  // No critical edges added.
@@ -102,11 +108,15 @@ namespace {
 }
 
 char LoopSimplify::ID = 0;
-static RegisterPass<LoopSimplify>
-X("loopsimplify", "Canonicalize natural loops", true);
-
-// Publically exposed interface to pass...
-const PassInfo *const llvm::LoopSimplifyID = &X;
+INITIALIZE_PASS_BEGIN(LoopSimplify, "loop-simplify",
+                "Canonicalize natural loops", true, false)
+INITIALIZE_PASS_DEPENDENCY(DominatorTree)
+INITIALIZE_PASS_DEPENDENCY(LoopInfo)
+INITIALIZE_PASS_END(LoopSimplify, "loop-simplify",
+                "Canonicalize natural loops", true, false)
+
+// Publicly exposed interface to pass...
+char &llvm::LoopSimplifyID = LoopSimplify::ID;
 Pass *llvm::createLoopSimplifyPass() { return new LoopSimplify(); }
 
 /// runOnLoop - Run down all loops in the CFG (recursively, but we could do
@@ -118,6 +128,7 @@ bool LoopSimplify::runOnLoop(Loop *l, LPPassManager &LPM) {
   LI = &getAnalysis<LoopInfo>();
   AA = getAnalysisIfAvailable<AliasAnalysis>();
   DT = &getAnalysis<DominatorTree>();
+  SE = getAnalysisIfAvailable<ScalarEvolution>();
 
   Changed |= ProcessLoop(L, LPM);
 
@@ -131,7 +142,7 @@ bool LoopSimplify::ProcessLoop(Loop *L, LPPassManager &LPM) {
   bool Changed = false;
 ReprocessLoop:
 
-  // Check to see that no blocks (other than the header) in this loop that has
+  // Check to see that no blocks (other than the header) in this loop have
   // predecessors that are not in the loop.  This is not valid for natural
   // loops, but can occur if the blocks are unreachable.  Since they are
   // unreachable we can just shamelessly delete those CFG edges!
@@ -139,14 +150,21 @@ ReprocessLoop:
        BB != E; ++BB) {
     if (*BB == L->getHeader()) continue;
 
-    SmallPtrSet<BasicBlock *, 4> BadPreds;
-    for (pred_iterator PI = pred_begin(*BB), PE = pred_end(*BB); PI != PE; ++PI)
-      if (!L->contains(*PI))
-        BadPreds.insert(*PI);
+    SmallPtrSet<BasicBlock*, 4> BadPreds;
+    for (pred_iterator PI = pred_begin(*BB),
+         PE = pred_end(*BB); PI != PE; ++PI) {
+      BasicBlock *P = *PI;
+      if (!L->contains(P))
+        BadPreds.insert(P);
+    }
 
     // Delete each unique out-of-loop (and thus dead) predecessor.
-    for (SmallPtrSet<BasicBlock *, 4>::iterator I = BadPreds.begin(),
+    for (SmallPtrSet<BasicBlock*, 4>::iterator I = BadPreds.begin(),
          E = BadPreds.end(); I != E; ++I) {
+
+      DEBUG(dbgs() << "LoopSimplify: Deleting edge from dead predecessor "
+                   << (*I)->getName() << "\n");
+
       // Inform each successor of each dead pred.
       for (succ_iterator SI = succ_begin(*I), SE = succ_end(*I); SI != SE; ++SI)
         (*SI)->removePredecessor(*I);
@@ -159,12 +177,32 @@ ReprocessLoop:
     }
   }
 
+  // If there are exiting blocks with branches on undef, resolve the undef in
+  // the direction which will exit the loop. This will help simplify loop
+  // trip count computations.
+  SmallVector<BasicBlock*, 8> ExitingBlocks;
+  L->getExitingBlocks(ExitingBlocks);
+  for (SmallVectorImpl<BasicBlock *>::iterator I = ExitingBlocks.begin(),
+       E = ExitingBlocks.end(); I != E; ++I)
+    if (BranchInst *BI = dyn_cast<BranchInst>((*I)->getTerminator()))
+      if (BI->isConditional()) {
+        if (UndefValue *Cond = dyn_cast<UndefValue>(BI->getCondition())) {
+
+          DEBUG(dbgs() << "LoopSimplify: Resolving \"br i1 undef\" to exit in "
+                       << (*I)->getName() << "\n");
+
+          BI->setCondition(ConstantInt::get(Cond->getType(),
+                                            !L->contains(BI->getSuccessor(0))));
+          Changed = true;
+        }
+      }
+
   // Does the loop already have a preheader?  If so, don't insert one.
   BasicBlock *Preheader = L->getLoopPreheader();
   if (!Preheader) {
     Preheader = InsertPreheaderForLoop(L);
     if (Preheader) {
-      NumInserted++;
+      ++NumInserted;
       Changed = true;
     }
   }
@@ -187,7 +225,7 @@ ReprocessLoop:
       // allowed.
       if (!L->contains(*PI)) {
         if (RewriteLoopExitBlock(L, ExitBlock)) {
-          NumInserted++;
+          ++NumInserted;
           Changed = true;
         }
         break;
@@ -216,7 +254,7 @@ ReprocessLoop:
     // loop header.
     LoopLatch = InsertUniqueBackedgeBlock(L, Preheader);
     if (LoopLatch) {
-      NumInserted++;
+      ++NumInserted;
       Changed = true;
     }
   }
@@ -227,8 +265,9 @@ ReprocessLoop:
   PHINode *PN;
   for (BasicBlock::iterator I = L->getHeader()->begin();
        (PN = dyn_cast<PHINode>(I++)); )
-    if (Value *V = PN->hasConstantValue(DT)) {
+    if (Value *V = SimplifyInstruction(PN, 0, DT)) {
       if (AA) AA->deleteValue(PN);
+      if (SE) SE->forgetValue(PN);
       PN->replaceAllUsesWith(V);
       PN->eraseFromParent();
     }
@@ -250,8 +289,6 @@ ReprocessLoop:
         break;
       }
   if (UniqueExit) {
-    SmallVector<BasicBlock*, 8> ExitingBlocks;
-    L->getExitingBlocks(ExitingBlocks);
     for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
       BasicBlock *ExitingBlock = ExitingBlocks[i];
       if (!ExitingBlock->getSinglePredecessor()) continue;
@@ -265,6 +302,9 @@ ReprocessLoop:
       bool AllInvariant = true;
       for (BasicBlock::iterator I = ExitingBlock->begin(); &*I != BI; ) {
         Instruction *Inst = I++;
+        // Skip debug info intrinsics.
+        if (isa<DbgInfoIntrinsic>(Inst))
+          continue;
         if (Inst == CI)
           continue;
         if (!L->makeLoopInvariant(Inst, Changed,
@@ -281,24 +321,22 @@ ReprocessLoop:
       if (!FoldBranchToCommonDest(BI)) continue;
 
       // Success. The block is now dead, so remove it from the loop,
-      // update the dominator tree and dominance frontier, and delete it.
+      // update the dominator tree and delete it.
+      DEBUG(dbgs() << "LoopSimplify: Eliminating exiting block "
+                   << ExitingBlock->getName() << "\n");
+
       assert(pred_begin(ExitingBlock) == pred_end(ExitingBlock));
       Changed = true;
       LI->removeBlock(ExitingBlock);
 
-      DominanceFrontier *DF = getAnalysisIfAvailable<DominanceFrontier>();
       DomTreeNode *Node = DT->getNode(ExitingBlock);
       const std::vector<DomTreeNodeBase<BasicBlock> *> &Children =
         Node->getChildren();
       while (!Children.empty()) {
         DomTreeNode *Child = Children.front();
         DT->changeImmediateDominator(Child, Node->getIDom());
-        if (DF) DF->changeImmediateDominator(Child->getBlock(),
-                                             Node->getIDom()->getBlock(),
-                                             DT);
       }
       DT->eraseNode(ExitingBlock);
-      if (DF) DF->removeBlock(ExitingBlock);
 
       BI->getSuccessor(0)->removePredecessor(ExitingBlock);
       BI->getSuccessor(1)->removePredecessor(ExitingBlock);
@@ -319,22 +357,27 @@ BasicBlock *LoopSimplify::InsertPreheaderForLoop(Loop *L) {
   // Compute the set of predecessors of the loop that are not in the loop.
   SmallVector<BasicBlock*, 8> OutsideBlocks;
   for (pred_iterator PI = pred_begin(Header), PE = pred_end(Header);
-       PI != PE; ++PI)
-    if (!L->contains(*PI)) {         // Coming in from outside the loop?
+       PI != PE; ++PI) {
+    BasicBlock *P = *PI;
+    if (!L->contains(P)) {         // Coming in from outside the loop?
       // If the loop is branched to from an indirect branch, we won't
       // be able to fully transform the loop, because it prohibits
       // edge splitting.
-      if (isa<IndirectBrInst>((*PI)->getTerminator())) return 0;
+      if (isa<IndirectBrInst>(P->getTerminator())) return 0;
 
       // Keep track of it.
-      OutsideBlocks.push_back(*PI);
+      OutsideBlocks.push_back(P);
     }
+  }
 
   // Split out the loop pre-header.
   BasicBlock *NewBB =
     SplitBlockPredecessors(Header, &OutsideBlocks[0], OutsideBlocks.size(),
                            ".preheader", this);
 
+  DEBUG(dbgs() << "LoopSimplify: Creating pre-header " << NewBB->getName()
+               << "\n");
+
   // Make sure that NewBB is put someplace intelligent, which doesn't mess up
   // code layout too horribly.
   PlaceSplitBlockCarefully(NewBB, OutsideBlocks, L);
@@ -347,19 +390,23 @@ BasicBlock *LoopSimplify::InsertPreheaderForLoop(Loop *L) {
 /// outside of the loop.
 BasicBlock *LoopSimplify::RewriteLoopExitBlock(Loop *L, BasicBlock *Exit) {
   SmallVector<BasicBlock*, 8> LoopBlocks;
-  for (pred_iterator I = pred_begin(Exit), E = pred_end(Exit); I != E; ++I)
-    if (L->contains(*I)) {
+  for (pred_iterator I = pred_begin(Exit), E = pred_end(Exit); I != E; ++I) {
+    BasicBlock *P = *I;
+    if (L->contains(P)) {
       // Don't do this if the loop is exited via an indirect branch.
-      if (isa<IndirectBrInst>((*I)->getTerminator())) return 0;
+      if (isa<IndirectBrInst>(P->getTerminator())) return 0;
 
-      LoopBlocks.push_back(*I);
+      LoopBlocks.push_back(P);
     }
+  }
 
   assert(!LoopBlocks.empty() && "No edges coming in from outside the loop?");
   BasicBlock *NewBB = SplitBlockPredecessors(Exit, &LoopBlocks[0], 
                                              LoopBlocks.size(), ".loopexit",
                                              this);
 
+  DEBUG(dbgs() << "LoopSimplify: Creating dedicated exit block "
+               << NewBB->getName() << "\n");
   return NewBB;
 }
 
@@ -385,11 +432,11 @@ static void AddBlockAndPredsToSet(BasicBlock *InputBB, BasicBlock *StopBlock,
 /// FindPHIToPartitionLoops - The first part of loop-nestification is to find a
 /// PHI node that tells us how to partition the loops.
 static PHINode *FindPHIToPartitionLoops(Loop *L, DominatorTree *DT,
-                                        AliasAnalysis *AA) {
+                                        AliasAnalysis *AA, LoopInfo *LI) {
   for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ) {
     PHINode *PN = cast<PHINode>(I);
     ++I;
-    if (Value *V = PN->hasConstantValue(DT)) {
+    if (Value *V = SimplifyInstruction(PN, 0, DT)) {
       // This is a degenerate PHI already, don't modify it!
       PN->replaceAllUsesWith(V);
       if (AA) AA->deleteValue(PN);
@@ -463,7 +510,7 @@ void LoopSimplify::PlaceSplitBlockCarefully(BasicBlock *NewBB,
 /// created.
 ///
 Loop *LoopSimplify::SeparateNestedLoop(Loop *L, LPPassManager &LPM) {
-  PHINode *PN = FindPHIToPartitionLoops(L, DT, AA);
+  PHINode *PN = FindPHIToPartitionLoops(L, DT, AA, LI);
   if (PN == 0) return 0;  // No known way to partition.
 
   // Pull out all predecessors that have varying values in the loop.  This
@@ -480,6 +527,14 @@ Loop *LoopSimplify::SeparateNestedLoop(Loop *L, LPPassManager &LPM) {
       OuterLoopPreds.push_back(PN->getIncomingBlock(i));
     }
 
+  DEBUG(dbgs() << "LoopSimplify: Splitting out a new outer loop\n");
+
+  // If ScalarEvolution is around and knows anything about values in
+  // this loop, tell it to forget them, because we're about to
+  // substantially change it.
+  if (SE)
+    SE->forgetLoop(L);
+
   BasicBlock *Header = L->getHeader();
   BasicBlock *NewBB = SplitBlockPredecessors(Header, &OuterLoopPreds[0],
                                              OuterLoopPreds.size(),
@@ -515,10 +570,11 @@ Loop *LoopSimplify::SeparateNestedLoop(Loop *L, LPPassManager &LPM) {
   // Determine which blocks should stay in L and which should be moved out to
   // the Outer loop now.
   std::set<BasicBlock*> BlocksInL;
-  for (pred_iterator PI = pred_begin(Header), E = pred_end(Header); PI!=E; ++PI)
-    if (DT->dominates(Header, *PI))
-      AddBlockAndPredsToSet(*PI, Header, BlocksInL);
-
+  for (pred_iterator PI=pred_begin(Header), E = pred_end(Header); PI!=E; ++PI) {
+    BasicBlock *P = *PI;
+    if (DT->dominates(Header, P))
+      AddBlockAndPredsToSet(P, Header, BlocksInL);
+  }
 
   // Scan all of the loop children of L, moving them to OuterLoop if they are
   // not part of the inner loop.
@@ -566,14 +622,24 @@ LoopSimplify::InsertUniqueBackedgeBlock(Loop *L, BasicBlock *Preheader) {
 
   // Figure out which basic blocks contain back-edges to the loop header.
   std::vector<BasicBlock*> BackedgeBlocks;
-  for (pred_iterator I = pred_begin(Header), E = pred_end(Header); I != E; ++I)
-    if (*I != Preheader) BackedgeBlocks.push_back(*I);
+  for (pred_iterator I = pred_begin(Header), E = pred_end(Header); I != E; ++I){
+    BasicBlock *P = *I;
+
+    // Indirectbr edges cannot be split, so we must fail if we find one.
+    if (isa<IndirectBrInst>(P->getTerminator()))
+      return 0;
+
+    if (P != Preheader) BackedgeBlocks.push_back(P);
+  }
 
   // Create and insert the new backedge block...
   BasicBlock *BEBlock = BasicBlock::Create(Header->getContext(),
                                            Header->getName()+".backedge", F);
   BranchInst *BETerminator = BranchInst::Create(Header, BEBlock);
 
+  DEBUG(dbgs() << "LoopSimplify: Inserting unique backedge block "
+               << BEBlock->getName() << "\n");
+
   // Move the new backedge block to right after the last backedge block.
   Function::iterator InsertPos = BackedgeBlocks.back(); ++InsertPos;
   F->getBasicBlockList().splice(InsertPos, F->getBasicBlockList(), BEBlock);
@@ -582,9 +648,8 @@ LoopSimplify::InsertUniqueBackedgeBlock(Loop *L, BasicBlock *Preheader) {
   // the backedge block which correspond to any PHI nodes in the header block.
   for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
     PHINode *PN = cast<PHINode>(I);
-    PHINode *NewPN = PHINode::Create(PN->getType(), PN->getName()+".be",
-                                     BETerminator);
-    NewPN->reserveOperandSpace(BackedgeBlocks.size());
+    PHINode *NewPN = PHINode::Create(PN->getType(), BackedgeBlocks.size(),
+                                     PN->getName()+".be", BETerminator);
     if (AA) AA->copyValue(PN, NewPN);
 
     // Loop over the PHI node, moving all entries except the one for the
@@ -648,8 +713,6 @@ LoopSimplify::InsertUniqueBackedgeBlock(Loop *L, BasicBlock *Preheader) {
 
   // Update dominator information
   DT->splitBlock(BEBlock);
-  if (DominanceFrontier *DF = getAnalysisIfAvailable<DominanceFrontier>())
-    DF->splitBlock(BEBlock);
 
   return BEBlock;
 }