SLPVectorizer: start constructing chains at stores that are not power of two.

The type <3 x i8> is a common in graphics and we want to be able to vectorize it.

This changes accelerates bullet by 12% and 471_omnetpp by 5%.

llvm-svn: 184317

2 files changed, 63 insertions, 3 deletions

diff --git a/llvm/lib/Transforms/Vectorize/VecUtils.cpp b/llvm/lib/Transforms/Vectorize/VecUtils.cpp
index 2f6b7df21b6..1e97ed400be 100644
--- a/llvm/lib/Transforms/Vectorize/VecUtils.cpp
+++ b/llvm/lib/Transforms/Vectorize/VecUtils.cpp
@@ -104,6 +104,8 @@ bool BoUpSLP::isConsecutiveAccess(Value *A, Value *B) {
 }
 
 bool BoUpSLP::vectorizeStoreChain(ArrayRef<Value *> Chain, int CostThreshold) {
+  unsigned ChainLen = Chain.size();
+    DEBUG(dbgs()<<"SLP: Analyzing a store chain of length " <<ChainLen<< "\n");
   Type *StoreTy = cast<StoreInst>(Chain[0])->getValueOperand()->getType();
   unsigned Sz = DL->getTypeSizeInBits(StoreTy);
   unsigned VF = MinVecRegSize / Sz;
@@ -112,8 +114,8 @@ bool BoUpSLP::vectorizeStoreChain(ArrayRef<Value *> Chain, int CostThreshold) {
 
   bool Changed = false;
   // Look for profitable vectorizable trees at all offsets, starting at zero.
-  for (unsigned i = 0, e = Chain.size(); i < e; ++i) {
-    if (i + VF > e) return Changed;
+  for (unsigned i = 0, e = ChainLen; i < e; ++i) {
+    if (i + VF > e) break;
     DEBUG(dbgs()<<"SLP: Analyzing " << VF << " stores at offset "<< i << "\n");
     ArrayRef<Value *> Operands = Chain.slice(i, VF);
 
@@ -128,7 +130,19 @@ bool BoUpSLP::vectorizeStoreChain(ArrayRef<Value *> Chain, int CostThreshold) {
     }
   }
 
-  return Changed;
+  if (Changed)
+    return true;
+
+  int Cost = getTreeCost(Chain);
+  if (Cost < CostThreshold) {
+    DEBUG(dbgs() << "SLP: Found store chain cost = "<< Cost <<" for size = " <<
+          ChainLen << "\n");
+    Builder.SetInsertPoint(getInsertionPoint(getLastIndex(Chain, ChainLen)));
+    vectorizeTree(Chain, ChainLen);
+    return true;
+  }
+
+  return false;
 }
 
 bool BoUpSLP::vectorizeStores(ArrayRef<StoreInst *> Stores, int costThreshold) {
diff --git a/llvm/test/Transforms/SLPVectorizer/X86/odd_store.ll b/llvm/test/Transforms/SLPVectorizer/X86/odd_store.ll
new file mode 100644
index 00000000000..cb2b6862990
--- /dev/null
+++ b/llvm/test/Transforms/SLPVectorizer/X86/odd_store.ll
@@ -0,0 +1,46 @@
+; RUN: opt < %s -basicaa -slp-vectorizer -dce -S -mtriple=x86_64-apple-macosx10.8.0 -mcpu=corei7-avx | FileCheck %s
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
+target triple = "x86_64-apple-macosx10.8.0"
+
+;int foo(char * restrict A, float * restrict B, float T) {
+;  A[0] = (T * B[10] + 4.0);
+;  A[1] = (T * B[11] + 5.0);
+;  A[2] = (T * B[12] + 6.0);
+;}
+
+;CHECK: @foo
+;CHECK: load <3 x float>
+;CHECK: fmul <3 x float>
+;CHECK: fpext <3 x float>
+;CHECK: fadd <3 x double>
+;CHECK: fptosi <3 x double>
+;CHECK: store <3 x i8>
+;CHECK: ret
+define i32 @foo(i8* noalias nocapture %A, float* noalias nocapture %B, float %T) {
+  %1 = getelementptr inbounds float* %B, i64 10
+  %2 = load float* %1, align 4
+  %3 = fmul float %2, %T
+  %4 = fpext float %3 to double
+  %5 = fadd double %4, 4.000000e+00
+  %6 = fptosi double %5 to i8
+  store i8 %6, i8* %A, align 1
+  %7 = getelementptr inbounds float* %B, i64 11
+  %8 = load float* %7, align 4
+  %9 = fmul float %8, %T
+  %10 = fpext float %9 to double
+  %11 = fadd double %10, 5.000000e+00
+  %12 = fptosi double %11 to i8
+  %13 = getelementptr inbounds i8* %A, i64 1
+  store i8 %12, i8* %13, align 1
+  %14 = getelementptr inbounds float* %B, i64 12
+  %15 = load float* %14, align 4
+  %16 = fmul float %15, %T
+  %17 = fpext float %16 to double
+  %18 = fadd double %17, 6.000000e+00
+  %19 = fptosi double %18 to i8
+  %20 = getelementptr inbounds i8* %A, i64 2
+  store i8 %19, i8* %20, align 1
+  ret i32 undef
+}
+