Improve codegen for BUILD_VECTORs on ARM.

If we have a BUILD_VECTOR that is mostly a constant splat, it is often better to splat that constant then insertelement the non-constant lanes instead of insertelementing every lane from an undef base.



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@163304 91177308-0d34-0410-b5e6-96231b3b80d8

1 files changed, 34 insertions, 0 deletions

diff --git a/test/CodeGen/ARM/vdup.ll b/test/CodeGen/ARM/vdup.ll
index 05332e4d8c..a8c224b438 100644
--- a/test/CodeGen/ARM/vdup.ll
+++ b/test/CodeGen/ARM/vdup.ll
@@ -261,3 +261,37 @@ define void @redundantVdup(<8 x i8>* %ptr) nounwind {
   store <8 x i8> %2, <8 x i8>* %ptr, align 8
   ret void
 }
+
+define <4 x i32> @tdupi(i32 %x, i32 %y) {
+;CHECK: tdupi
+;CHECK: vdup.32
+  %1 = insertelement <4 x i32> undef, i32 %x, i32 0
+  %2 = insertelement <4 x i32> %1, i32 %x, i32 1
+  %3 = insertelement <4 x i32> %2, i32 %x, i32 2
+  %4 = insertelement <4 x i32> %3, i32 %y, i32 3
+  ret <4 x i32> %4
+}
+
+define <4 x float> @tdupf(float %x, float %y) {
+;CHECK: tdupf
+;CHECK: vdup.32
+  %1 = insertelement <4 x float> undef, float %x, i32 0
+  %2 = insertelement <4 x float> %1, float %x, i32 1
+  %3 = insertelement <4 x float> %2, float %x, i32 2
+  %4 = insertelement <4 x float> %3, float %y, i32 3
+  ret <4 x float> %4
+}
+
+; This test checks that when splatting an element from a vector into another,
+; the value isn't moved out to GPRs first.
+define <4 x i32> @tduplane(<4 x i32> %invec) {
+;CHECK: tduplane
+;CHECK-NOT: vmov {{.*}}, d16[1]
+;CHECK: vdup.32 {{.*}}, d16[1]
+  %in = extractelement <4 x i32> %invec, i32 1
+  %1 = insertelement <4 x i32> undef, i32 %in, i32 0
+  %2 = insertelement <4 x i32> %1, i32 %in, i32 1
+  %3 = insertelement <4 x i32> %2, i32 %in, i32 2
+  %4 = insertelement <4 x i32> %3, i32 255, i32 3
+  ret <4 x i32> %4
+}