Fix for PR2484: add an SSE1 pattern for a shuffle we normally prefer to

handle with an SSE2 instruction.

llvm-svn: 73760

llvm/lib/Target/X86/X86InstrSSE.td

@@ -3027,6 +3027,12 @@ def : Pat<(v4i32 (movlp VR128:$src1, VR128:$src2)),
           (MOVLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
 }
 
+// vector_shuffle v1, v2 <4, 5, 2, 3> using SHUFPSrri (we prefer movsd, but
+// fall back to this for SSE1)
+def : Pat<(v4f32 (movlp:$src3 VR128:$src1, (v4f32 VR128:$src2))),
+          (SHUFPSrri VR128:$src2, VR128:$src1, 
+                     (SHUFFLE_get_shuf_imm VR128:$src3))>, Requires<[HasSSE1]>;
+
 // Set lowest element and zero upper elements.
 let AddedComplexity = 15 in
 def : Pat<(v2f64 (movl immAllZerosV_bc, VR128:$src)),

llvm/test/CodeGen/X86/2009-06-18-movlp-shuffle-register.ll

@@ -0,0 +1,9 @@
+; RUN: llvm-as < %s | llc -march=x86 -mattr=+sse,-sse2
+; PR2484
+
+define <4 x float> @f4523(<4 x float> %a,<4 x float> %b) nounwind {
+entry:
+%shuffle = shufflevector <4 x float> %a, <4 x float> %b, <4 x i32> <i32 4,i32
+5,i32 2,i32 3>
+ret <4 x float> %shuffle
+}