[llvm] r202067 - SLPVectorizer: Try vectorizing 'splat' stores

Arnold Schwaighofer aschwaighofer at apple.com
Mon Feb 24 11:52:29 PST 2014


Author: arnolds
Date: Mon Feb 24 13:52:29 2014
New Revision: 202067

URL: http://llvm.org/viewvc/llvm-project?rev=202067&view=rev
Log:
SLPVectorizer: Try vectorizing 'splat' stores

Vectorize sequential stores of a broadcasted value.
5% on eon.

radar://16124699

Modified:
    llvm/trunk/lib/Transforms/Vectorize/SLPVectorizer.cpp
    llvm/trunk/test/Transforms/SLPVectorizer/X86/tiny-tree.ll

Modified: llvm/trunk/lib/Transforms/Vectorize/SLPVectorizer.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Vectorize/SLPVectorizer.cpp?rev=202067&r1=202066&r2=202067&view=diff
==============================================================================
--- llvm/trunk/lib/Transforms/Vectorize/SLPVectorizer.cpp (original)
+++ llvm/trunk/lib/Transforms/Vectorize/SLPVectorizer.cpp Mon Feb 24 13:52:29 2014
@@ -1101,11 +1101,15 @@ bool BoUpSLP::isFullyVectorizableTinyTre
   if (VectorizableTree.size() != 2)
     return false;
 
+  // Handle splat stores.
+  if (!VectorizableTree[0].NeedToGather && isSplat(VectorizableTree[1].Scalars))
+    return true;
+
   // Gathering cost would be too much for tiny trees.
-  if (VectorizableTree[0].NeedToGather || VectorizableTree[1].NeedToGather) 
-    return false; 
+  if (VectorizableTree[0].NeedToGather || VectorizableTree[1].NeedToGather)
+    return false;
 
-  return true; 
+  return true;
 }
 
 int BoUpSLP::getTreeCost() {

Modified: llvm/trunk/test/Transforms/SLPVectorizer/X86/tiny-tree.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/SLPVectorizer/X86/tiny-tree.ll?rev=202067&r1=202066&r2=202067&view=diff
==============================================================================
--- llvm/trunk/test/Transforms/SLPVectorizer/X86/tiny-tree.ll (original)
+++ llvm/trunk/test/Transforms/SLPVectorizer/X86/tiny-tree.ll Mon Feb 24 13:52:29 2014
@@ -138,3 +138,18 @@ for.body:
 for.end:                                          ; preds = %for.body, %entry
   ret void
 }
+
+
+; CHECK-LABEL: store_splat
+; CHECK: store <4 x float>
+define void @store_splat(float*, float) {
+  %3 = getelementptr inbounds float* %0, i64 0
+  store float %1, float* %3, align 4
+  %4 = getelementptr inbounds float* %0, i64 1
+  store float %1, float* %4, align 4
+  %5 = getelementptr inbounds float* %0, i64 2
+  store float %1, float* %5, align 4
+  %6 = getelementptr inbounds float* %0, i64 3
+  store float %1, float* %6, align 4
+  ret void
+}





More information about the llvm-commits mailing list