Revert the switch of loop-idiom to use the new dependence analysis.

The new analysis is not yet ready for prime time. It has a *critical*
flawed assumption, and some troubling shortages of testing. Until it's
been hammered into better shape, let's stick with the working code. This
should be easy to revert itself when the analysis is ready.

Fixes PR14241, a miscompile of any memcpy-able loop which uses a pointer
as the induction mechanism. If you have been seeing miscompiles in this
revision range, you really want to test with this backed out. The
results of this miscompile are a bit subtle as they can lead to
downstream passes concluding things are impossible which are in fact
possible.

Thanks to David Blaikie for the majority of the reduction of this
miscompile. I'll be checking in the test case in a non-revert commit.

Revesions reverted here:

r167045: LoopIdiom: Fix a serious missed optimization: we only turned
         top-level loops into memmove.
r166877: LoopIdiom: Add checks to avoid turning memmove into an infinite
         loop.
r166875: LoopIdiom: Recognize memmove loops.
r166874: LoopIdiom: Replace custom dependence analysis with
         DependenceAnalysis.
llvm-svn: 167286

1 files changed, 88 insertions, 69 deletions

diff --git a/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp b/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp
index 7b42559ddeb..a44e798f121 100644
--- a/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp
+++ b/llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp
@@ -16,7 +16,7 @@
 // TODO List:
 //
 // Future loop memory idioms to recognize:
-//   memcmp, strlen, etc.
+//   memcmp, memmove, strlen, etc.
 // Future floating point idioms to recognize in -ffast-math mode:
 //   fpowi
 // Future integer operation idioms to recognize:
@@ -48,7 +48,6 @@
 #include "llvm/Module.h"
 #include "llvm/ADT/Statistic.h"
 #include "llvm/Analysis/AliasAnalysis.h"
-#include "llvm/Analysis/DependenceAnalysis.h"
 #include "llvm/Analysis/LoopPass.h"
 #include "llvm/Analysis/ScalarEvolutionExpander.h"
 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
@@ -60,9 +59,8 @@
 #include "llvm/Transforms/Utils/Local.h"
 using namespace llvm;
 
-STATISTIC(NumMemSet, "Number of memsets formed from loop stores");
-STATISTIC(NumMemCpy, "Number of memcpys formed from loop load+stores");
-STATISTIC(NumMemMove, "Number of memmoves formed from loop load+stores");
+STATISTIC(NumMemSet, "Number of memset's formed from loop stores");
+STATISTIC(NumMemCpy, "Number of memcpy's formed from loop load+stores");
 
 namespace {
   class LoopIdiomRecognize : public LoopPass {
@@ -108,8 +106,6 @@ namespace {
       AU.addPreserved<AliasAnalysis>();
       AU.addRequired<ScalarEvolution>();
       AU.addPreserved<ScalarEvolution>();
-      AU.addRequired<DependenceAnalysis>();
-      AU.addPreserved<DependenceAnalysis>();
       AU.addPreserved<DominatorTree>();
       AU.addRequired<DominatorTree>();
       AU.addRequired<TargetLibraryInfo>();
@@ -126,7 +122,6 @@ INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
 INITIALIZE_PASS_DEPENDENCY(LCSSA)
 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
-INITIALIZE_PASS_DEPENDENCY(DependenceAnalysis)
 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
 INITIALIZE_PASS_END(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms",
                     false, false)
@@ -168,6 +163,15 @@ static void deleteDeadInstruction(Instruction *I, ScalarEvolution &SE,
   } while (!NowDeadInsts.empty());
 }
 
+/// deleteIfDeadInstruction - If the specified value is a dead instruction,
+/// delete it and any recursively used instructions.
+static void deleteIfDeadInstruction(Value *V, ScalarEvolution &SE,
+                                    const TargetLibraryInfo *TLI) {
+  if (Instruction *I = dyn_cast<Instruction>(V))
+    if (isInstructionTriviallyDead(I, TLI))
+      deleteDeadInstruction(I, SE, TLI);
+}
+
 bool LoopIdiomRecognize::runOnLoop(Loop *L, LPPassManager &LPM) {
   CurLoop = L;
 
@@ -178,7 +182,7 @@ bool LoopIdiomRecognize::runOnLoop(Loop *L, LPPassManager &LPM) {
 
   // Disable loop idiom recognition if the function's name is a common idiom.
   StringRef Name = L->getHeader()->getParent()->getName();
-  if (Name == "memset" || Name == "memcpy" || Name == "memmove")
+  if (Name == "memset" || Name == "memcpy")
     return false;
 
   // The trip count of the loop must be analyzable.
@@ -364,6 +368,40 @@ processLoopMemSet(MemSetInst *MSI, const SCEV *BECount) {
                                  MSI, Ev, BECount);
 }
 
+
+/// mayLoopAccessLocation - Return true if the specified loop might access the
+/// specified pointer location, which is a loop-strided access.  The 'Access'
+/// argument specifies what the verboten forms of access are (read or write).
+static bool mayLoopAccessLocation(Value *Ptr,AliasAnalysis::ModRefResult Access,
+                                  Loop *L, const SCEV *BECount,
+                                  unsigned StoreSize, AliasAnalysis &AA,
+                                  Instruction *IgnoredStore) {
+  // Get the location that may be stored across the loop.  Since the access is
+  // strided positively through memory, we say that the modified location starts
+  // at the pointer and has infinite size.
+  uint64_t AccessSize = AliasAnalysis::UnknownSize;
+
+  // If the loop iterates a fixed number of times, we can refine the access size
+  // to be exactly the size of the memset, which is (BECount+1)*StoreSize
+  if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
+    AccessSize = (BECst->getValue()->getZExtValue()+1)*StoreSize;
+
+  // TODO: For this to be really effective, we have to dive into the pointer
+  // operand in the store.  Store to &A[i] of 100 will always return may alias
+  // with store of &A[100], we need to StoreLoc to be "A" with size of 100,
+  // which will then no-alias a store to &A[100].
+  AliasAnalysis::Location StoreLoc(Ptr, AccessSize);
+
+  for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E;
+       ++BI)
+    for (BasicBlock::iterator I = (*BI)->begin(), E = (*BI)->end(); I != E; ++I)
+      if (&*I != IgnoredStore &&
+          (AA.getModRefInfo(I, StoreLoc) & Access))
+        return true;
+
+  return false;
+}
+
 /// getMemSetPatternValue - If a strided store of the specified value is safe to
 /// turn into a memset_pattern16, return a ConstantArray of 16 bytes that should
 /// be passed in.  Otherwise, return null.
@@ -436,18 +474,6 @@ processLoopStridedStore(Value *DestPtr, unsigned StoreSize,
     return false;
   }
 
-  // Make sure the store has no dependencies (i.e. other loads and stores) in
-  // the loop.
-  DependenceAnalysis &DA = getAnalysis<DependenceAnalysis>();
-  for (Loop::block_iterator BI = CurLoop->block_begin(),
-       BE = CurLoop->block_end(); BI != BE; ++BI)
-    for (BasicBlock::iterator I = (*BI)->begin(), E = (*BI)->end(); I != E; ++I)
-      if (&*I != TheStore && I->mayReadOrWriteMemory()) {
-        OwningPtr<Dependence> D(DA.depends(TheStore, I, true));
-        if (D)
-          return false;
-      }
-
   // The trip count of the loop and the base pointer of the addrec SCEV is
   // guaranteed to be loop invariant, which means that it should dominate the
   // header.  This allows us to insert code for it in the preheader.
@@ -466,6 +492,15 @@ processLoopStridedStore(Value *DestPtr, unsigned StoreSize,
                            Preheader->getTerminator());
 
 
+  if (mayLoopAccessLocation(BasePtr, AliasAnalysis::ModRef,
+                            CurLoop, BECount,
+                            StoreSize, getAnalysis<AliasAnalysis>(), TheStore)){
+    Expander.clear();
+    // If we generated new code for the base pointer, clean up.
+    deleteIfDeadInstruction(BasePtr, *SE, TLI);
+    return false;
+  }
+
   // Okay, everything looks good, insert the memset.
 
   // The # stored bytes is (BECount+1)*Size.  Expand the trip count out to
@@ -523,46 +558,11 @@ processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
                            const SCEVAddRecExpr *LoadEv,
                            const SCEV *BECount) {
   // If we're not allowed to form memcpy, we fail.
-  if (!TLI->has(LibFunc::memcpy) || !TLI->has(LibFunc::memmove))
+  if (!TLI->has(LibFunc::memcpy))
     return false;
 
   LoadInst *LI = cast<LoadInst>(SI->getValueOperand());
 
-  // Make sure the load and the store have no dependencies (i.e. other loads and
-  // stores) in the loop. We ignore the direct dependency between SI and LI here
-  // and check it later.
-  DependenceAnalysis &DA = getAnalysis<DependenceAnalysis>();
-  bool isMemcpySafe = true;
-  for (Loop::block_iterator BI = CurLoop->block_begin(),
-       BE = CurLoop->block_end(); BI != BE; ++BI)
-    for (BasicBlock::iterator I = (*BI)->begin(), E = (*BI)->end(); I != E; ++I)
-      if (&*I != SI && &*I != LI && I->mayReadOrWriteMemory()) {
-        // First, check if there is a dependence of the store.
-        OwningPtr<Dependence> DS(DA.depends(SI, I, true));
-        if (DS)
-          return false;
-        // If the scanned instructon may modify memory then we also have to
-        // check for dependencys on the load.
-        if (I->mayWriteToMemory()) {
-          OwningPtr<Dependence> DL(DA.depends(I, LI, true));
-          if (DL)
-            return false;
-        }
-      }
-
-  // Now check the dependency between SI and LI. If there is no dependency we
-  // can safely emit a memcpy.
-  OwningPtr<Dependence> Dep(DA.depends(SI, LI, true));
-  if (Dep) {
-    // If there is a dependence but the direction is positive (or none) we can
-    // still safely turn this into memmove.
-    unsigned Direction = Dep->getDirection(Dep->getLevels());
-    if (Direction != Dependence::DVEntry::NONE &&
-        Direction != Dependence::DVEntry::GT)
-      return false;
-    isMemcpySafe = false;
-  }
-
   // The trip count of the loop and the base pointer of the addrec SCEV is
   // guaranteed to be loop invariant, which means that it should dominate the
   // header.  This allows us to insert code for it in the preheader.
@@ -571,16 +571,41 @@ processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
   SCEVExpander Expander(*SE, "loop-idiom");
 
   // Okay, we have a strided store "p[i]" of a loaded value.  We can turn
-  // this into a memcpy in the loop preheader now if we want.
+  // this into a memcpy in the loop preheader now if we want.  However, this
+  // would be unsafe to do if there is anything else in the loop that may read
+  // or write the memory region we're storing to.  This includes the load that
+  // feeds the stores.  Check for an alias by generating the base address and
+  // checking everything.
   Value *StoreBasePtr =
     Expander.expandCodeFor(StoreEv->getStart(),
                            Builder.getInt8PtrTy(SI->getPointerAddressSpace()),
                            Preheader->getTerminator());
+
+  if (mayLoopAccessLocation(StoreBasePtr, AliasAnalysis::ModRef,
+                            CurLoop, BECount, StoreSize,
+                            getAnalysis<AliasAnalysis>(), SI)) {
+    Expander.clear();
+    // If we generated new code for the base pointer, clean up.
+    deleteIfDeadInstruction(StoreBasePtr, *SE, TLI);
+    return false;
+  }
+
+  // For a memcpy, we have to make sure that the input array is not being
+  // mutated by the loop.
   Value *LoadBasePtr =
     Expander.expandCodeFor(LoadEv->getStart(),
                            Builder.getInt8PtrTy(LI->getPointerAddressSpace()),
                            Preheader->getTerminator());
 
+  if (mayLoopAccessLocation(LoadBasePtr, AliasAnalysis::Mod, CurLoop, BECount,
+                            StoreSize, getAnalysis<AliasAnalysis>(), SI)) {
+    Expander.clear();
+    // If we generated new code for the base pointer, clean up.
+    deleteIfDeadInstruction(LoadBasePtr, *SE, TLI);
+    deleteIfDeadInstruction(StoreBasePtr, *SE, TLI);
+    return false;
+  }
+
   // Okay, everything is safe, we can transform this!
 
 
@@ -598,19 +623,12 @@ processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
   Value *NumBytes =
     Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator());
 
-  CallInst *NewCall;
-  unsigned Align = std::min(SI->getAlignment(), LI->getAlignment());
-  if (isMemcpySafe) {
-    NewCall = Builder.CreateMemCpy(StoreBasePtr, LoadBasePtr, NumBytes, Align);
-    ++NumMemCpy;
-  } else {
-    NewCall = Builder.CreateMemMove(StoreBasePtr, LoadBasePtr, NumBytes, Align);
-    ++NumMemMove;
-  }
+  CallInst *NewCall =
+    Builder.CreateMemCpy(StoreBasePtr, LoadBasePtr, NumBytes,
+                         std::min(SI->getAlignment(), LI->getAlignment()));
   NewCall->setDebugLoc(SI->getDebugLoc());
 
-  DEBUG(dbgs() << "  Formed " << (isMemcpySafe ? "memcpy: " : "memmove: ")
-               << *NewCall << "\n"
+  DEBUG(dbgs() << "  Formed memcpy: " << *NewCall << "\n"
                << "    from load ptr=" << *LoadEv << " at: " << *LI << "\n"
                << "    from store ptr=" << *StoreEv << " at: " << *SI << "\n");
 
@@ -618,5 +636,6 @@ processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
   // Okay, the memset has been formed.  Zap the original store and anything that
   // feeds into it.
   deleteDeadInstruction(SI, *SE, TLI);
+  ++NumMemCpy;
   return true;
 }