From d11c5d08a5f4f030d6e357378d0d46d93efd9a59 Mon Sep 17 00:00:00 2001
From: Benjamin Kramer <benny.kra@googlemail.com>
Date: Sat, 27 Oct 2012 14:25:51 +0000
Subject: [PATCH] LoopIdiom: Recognize memmove loops.

This turns loops like
  for (unsigned i = 0; i != n; ++i)
    p[i] = p[i+1];
into memmove, which has a highly optimized implementation in most libcs.

This was really easy with the new DependenceAnalysis :)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@166875 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Transforms/Scalar/LoopIdiomRecognize.cpp | 34 ++++++++++++++------
 test/Transforms/LoopIdiom/basic.ll           | 22 +++++++++++++
 2 files changed, 46 insertions(+), 10 deletions(-)

diff --git a/lib/Transforms/Scalar/LoopIdiomRecognize.cpp b/lib/Transforms/Scalar/LoopIdiomRecognize.cpp
index 3050b1e0c13..495d403e549 100644
--- a/lib/Transforms/Scalar/LoopIdiomRecognize.cpp
+++ b/lib/Transforms/Scalar/LoopIdiomRecognize.cpp
@@ -16,7 +16,7 @@
 // TODO List:
 //
 // Future loop memory idioms to recognize:
-//   memcmp, memmove, strlen, etc.
+//   memcmp, strlen, etc.
 // Future floating point idioms to recognize in -ffast-math mode:
 //   fpowi
 // Future integer operation idioms to recognize:
@@ -60,8 +60,9 @@
 #include "llvm/Transforms/Utils/Local.h"
 using namespace llvm;
 
-STATISTIC(NumMemSet, "Number of memset's formed from loop stores");
-STATISTIC(NumMemCpy, "Number of memcpy's formed from loop load+stores");
+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");
 
 namespace {
   class LoopIdiomRecognize : public LoopPass {
@@ -532,6 +533,7 @@ processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
   // 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)
@@ -552,8 +554,14 @@ processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
   // 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)
-    return false;
+  if (Dep) {
+    // If there is a dependence but the direction is positive we can still
+    // safely turn this into memmove.
+    if (Dep->getLevels() != 1 ||
+        Dep->getDirection(1) != 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
@@ -590,12 +598,19 @@ processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
   Value *NumBytes =
     Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator());
 
-  CallInst *NewCall =
-    Builder.CreateMemCpy(StoreBasePtr, LoadBasePtr, NumBytes,
-                         std::min(SI->getAlignment(), LI->getAlignment()));
+  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;
+  }
   NewCall->setDebugLoc(SI->getDebugLoc());
 
-  DEBUG(dbgs() << "  Formed memcpy: " << *NewCall << "\n"
+  DEBUG(dbgs() << "  Formed " << (isMemcpySafe ? "memcpy: " : "memmove: ")
+               << *NewCall << "\n"
                << "    from load ptr=" << *LoadEv << " at: " << *LI << "\n"
                << "    from store ptr=" << *StoreEv << " at: " << *SI << "\n");
 
@@ -603,6 +618,5 @@ 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;
 }
diff --git a/test/Transforms/LoopIdiom/basic.ll b/test/Transforms/LoopIdiom/basic.ll
index 46ab7e5542b..5afc405fa6b 100644
--- a/test/Transforms/LoopIdiom/basic.ll
+++ b/test/Transforms/LoopIdiom/basic.ll
@@ -383,4 +383,26 @@ for.end:                                          ; preds = %for.inc
 
 }
 
+@p = common global [1024 x i8] zeroinitializer, align 16
 
+define void @test15(i32 %n) nounwind {
+entry:
+  %cmp6 = icmp eq i32 %n, 0
+  br i1 %cmp6, label %for.end, label %for.body
+
+for.body:                                         ; preds = %entry, %for.body
+  %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
+  %indvars.iv.next = add i64 %indvars.iv, 1
+  %arrayidx = getelementptr inbounds [1024 x i8]* @p, i64 0, i64 %indvars.iv.next
+  %0 = load i8* %arrayidx, align 1
+  %arrayidx2 = getelementptr inbounds [1024 x i8]* @p, i64 0, i64 %indvars.iv
+  store i8 %0, i8* %arrayidx2, align 1
+  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
+  %exitcond = icmp eq i32 %lftr.wideiv, %n
+  br i1 %exitcond, label %for.end, label %for.body
+
+for.end:                                          ; preds = %for.body, %entry
+  ret void
+; CHECK: @test15
+; CHECK: call void @llvm.memmove.p0i8.p0i8.i64(i8* getelementptr inbounds ([1024 x i8]* @p, i32 0, i32 0), i8* getelementptr inbounds ([1024 x i8]* @p, i64 0, i64 1),
+}