Fix SROA to avoid unnecessary scalar conversions for 1-element vectors.

When a 1-element vector alloca is promoted, a store instruction can often be
rewritten without converting the value to a scalar and using an insertelement
instruction to stuff it into the new alloca.  This patch just adds a check
to skip that conversion when it is unnecessary.  This turns out to be really
important for some ARM Neon operations where <1 x i64> is used to get around
the fact that i64 is not a legal type.

llvm-svn: 184870

2 files changed, 40 insertions, 15 deletions

diff --git a/llvm/lib/Transforms/Scalar/SROA.cpp b/llvm/lib/Transforms/Scalar/SROA.cpp
index d073e789dcb..7fb1dbd0771 100644
--- a/llvm/lib/Transforms/Scalar/SROA.cpp
+++ b/llvm/lib/Transforms/Scalar/SROA.cpp
@@ -2591,22 +2591,23 @@ private:
 
   bool rewriteVectorizedStoreInst(Value *V,
                                   StoreInst &SI, Value *OldOp) {
-    unsigned BeginIndex = getIndex(BeginOffset);
-    unsigned EndIndex = getIndex(EndOffset);
-    assert(EndIndex > BeginIndex && "Empty vector!");
-    unsigned NumElements = EndIndex - BeginIndex;
-    assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
-    Type *PartitionTy
-      = (NumElements == 1) ? ElementTy
-                           : VectorType::get(ElementTy, NumElements);
-    if (V->getType() != PartitionTy)
-      V = convertValue(TD, IRB, V, PartitionTy);
-
-    // Mix in the existing elements.
-    Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
-                                       "load");
-    V = insertVector(IRB, Old, V, BeginIndex, "vec");
+    if (V->getType() != VecTy) {
+      unsigned BeginIndex = getIndex(BeginOffset);
+      unsigned EndIndex = getIndex(EndOffset);
+      assert(EndIndex > BeginIndex && "Empty vector!");
+      unsigned NumElements = EndIndex - BeginIndex;
+      assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
+      Type *PartitionTy
+        = (NumElements == 1) ? ElementTy
+        : VectorType::get(ElementTy, NumElements);
+      if (V->getType() != PartitionTy)
+        V = convertValue(TD, IRB, V, PartitionTy);
 
+      // Mix in the existing elements.
+      Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
+                                         "load");
+      V = insertVector(IRB, Old, V, BeginIndex, "vec");
+    }
     StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
     Pass.DeadInsts.insert(&SI);
 
diff --git a/llvm/test/Transforms/ScalarRepl/vector_promote.ll b/llvm/test/Transforms/ScalarRepl/vector_promote.ll
index 5c82ae4d196..03ef475c3ed 100644
--- a/llvm/test/Transforms/ScalarRepl/vector_promote.ll
+++ b/llvm/test/Transforms/ScalarRepl/vector_promote.ll
@@ -111,3 +111,27 @@ entry:
 ; CHECK-NOT: alloca
 ; CHECK: and i192
 }
+
+; When promoting an alloca to a 1-element vector type, instructions that
+; produce that same vector type should not be changed to insert one element
+; into a new vector. <rdar://problem/14249078>
+define <1 x i64> @test8(<1 x i64> %a) {
+entry:
+  %a.addr = alloca <1 x i64>, align 8
+  %__a = alloca <1 x i64>, align 8
+  %tmp = alloca <1 x i64>, align 8
+  store <1 x i64> %a, <1 x i64>* %a.addr, align 8
+  %0 = load <1 x i64>* %a.addr, align 8
+  store <1 x i64> %0, <1 x i64>* %__a, align 8
+  %1 = load <1 x i64>* %__a, align 8
+  %2 = bitcast <1 x i64> %1 to <8 x i8>
+  %3 = bitcast <8 x i8> %2 to <1 x i64>
+  %vshl_n = shl <1 x i64> %3, <i64 4>
+  store <1 x i64> %vshl_n, <1 x i64>* %tmp
+  %4 = load <1 x i64>* %tmp
+  ret <1 x i64> %4
+; CHECK: @test8
+; CHECK-NOT: alloca
+; CHECK-NOT: insertelement
+; CHECK: ret <1 x i64>
+}