Optimize codegen for VSETLNi{8,16,32} operating on Q registers. Degenerate to a VSETLN on D registers, instead of an (INSERT_SUBREG (VSETLN (EXTRACT_SUBREG ))) sequence to help the register coalescer.

llvm-svn: 163298

2 files changed, 36 insertions, 1 deletions

diff --git a/llvm/test/CodeGen/ARM/integer_insertelement.ll b/llvm/test/CodeGen/ARM/integer_insertelement.ll
new file mode 100644
index 00000000000..4f2d7e3f73e
--- /dev/null
+++ b/llvm/test/CodeGen/ARM/integer_insertelement.ll
@@ -0,0 +1,35 @@
+; RUN: llc %s -o - -march=arm -mattr=+neon | FileCheck %s
+
+; This test checks that when inserting one (integer) element into a vector,
+; the vector is not spuriously copied. "vorr dX, dY, dY" is the way of moving
+; one DPR to another that we check for.
+
+; CHECK: @f
+; CHECK-NOT: vorr d
+; CHECK: vmov s
+; CHECK-NOT: vorr d
+; CHECK: mov pc, lr
+define <4 x i32> @f(<4 x i32> %in) {
+  %1 = insertelement <4 x i32> %in, i32 255, i32 3
+  ret <4 x i32> %1
+}
+
+; CHECK: @g
+; CHECK-NOT: vorr d
+; CHECK: vmov.16 d
+; CHECK-NOT: vorr d
+; CHECK: mov pc, lr
+define <8 x i16> @g(<8 x i16> %in) {
+  %1 = insertelement <8 x i16> %in, i16 255, i32 7
+  ret <8 x i16> %1
+}
+
+; CHECK: @h
+; CHECK-NOT: vorr d
+; CHECK: vmov.8 d
+; CHECK-NOT: vorr d
+; CHECK: mov pc, lr
+define <16 x i8> @h(<16 x i8> %in) {
+  %1 = insertelement <16 x i8> %in, i8 255, i32 15
+  ret <16 x i8> %1
+}
diff --git a/llvm/test/CodeGen/ARM/vget_lane.ll b/llvm/test/CodeGen/ARM/vget_lane.ll
index 1fc885d6137..2ed65c9aeed 100644
--- a/llvm/test/CodeGen/ARM/vget_lane.ll
+++ b/llvm/test/CodeGen/ARM/vget_lane.ll
@@ -200,7 +200,7 @@ define <8 x i16> @vsetQ_lane16(<8 x i16>* %A, i16 %B) nounwind {
 
 define <4 x i32> @vsetQ_lane32(<4 x i32>* %A, i32 %B) nounwind {
 ;CHECK: vsetQ_lane32:
-;CHECK: vmov.32
+;CHECK: vmov s
 	%tmp1 = load <4 x i32>* %A
 	%tmp2 = insertelement <4 x i32> %tmp1, i32 %B, i32 1
 	ret <4 x i32> %tmp2