[ARM][ISel] Improve the lowering of vector loads.

When vectors are built from a single value, the ARM lowering issues a
scalar_to_vector node.
This node is then always morphed into a move from the general purpose unit to
the vector unit.
When the value comes from a load, this can be simplified into a vector load to
the right lane.

This patch changes the lowering of insert_vector_elt to expose a vector
friendly pattern in this situation.

This is a step toward fixing <rdar://problem/14170854>.

llvm-svn: 186999

2 files changed, 17 insertions, 1 deletions

diff --git a/llvm/lib/Target/ARM/ARMISelLowering.cpp b/llvm/lib/Target/ARM/ARMISelLowering.cpp
index 569beadc60e..ea13c88be82 100644
--- a/llvm/lib/Target/ARM/ARMISelLowering.cpp
+++ b/llvm/lib/Target/ARM/ARMISelLowering.cpp
@@ -4674,7 +4674,9 @@ SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
   if (ValueCounts.size() == 0)
     return DAG.getUNDEF(VT);
 
-  if (isOnlyLowElement)
+  // Loads are better lowered with insert_vector_elt/ARMISD::BUILD_VECTOR.
+  // Keep going if we are hitting this case.
+  if (isOnlyLowElement && !ISD::isNormalLoad(Value.getNode()))
     return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
 
   unsigned EltSize = VT.getVectorElementType().getSizeInBits();
diff --git a/llvm/test/CodeGen/ARM/vector-DAGCombine.ll b/llvm/test/CodeGen/ARM/vector-DAGCombine.ll
index 6d586f24264..3e138199e6f 100644
--- a/llvm/test/CodeGen/ARM/vector-DAGCombine.ll
+++ b/llvm/test/CodeGen/ARM/vector-DAGCombine.ll
@@ -184,3 +184,17 @@ entry:
 
 ; Function Attrs: nounwind readnone
 declare <8 x i16> @llvm.arm.neon.vmullu.v8i16(<8 x i8>, <8 x i8>)
+
+; Check that (insert_vector_elt (load)) => (vector_load).
+; Thus, check that scalar_to_vector do not interfer with that.
+define <8 x i16> @t4(i8* nocapture %sp0) {
+; CHECK: t4
+; CHECK: vld1.32 {{{d[0-9]+}}[0]}, [r0]
+entry:
+  %pix_sp0.0.cast = bitcast i8* %sp0 to i32*
+  %pix_sp0.0.copyload = load i32* %pix_sp0.0.cast, align 1
+  %vec = insertelement <2 x i32> undef, i32 %pix_sp0.0.copyload, i32 0
+  %0 = bitcast <2 x i32> %vec to <8 x i8>
+  %vmull.i = tail call <8 x i16> @llvm.arm.neon.vmullu.v8i16(<8 x i8> %0, <8 x i8> %0)
+  ret <8 x i16> %vmull.i
+}