+; SSE2-LABEL: e:
+; SSE2: movdqa {{[0-9]*}}(%rsp), %xmm0
+; SSE2-NEXT: pshufd {{.*}} # [[X1:xmm[0-9]+]] = xmm0[1,1,3,3]
+; SSE2-NEXT: movdqa {{[0-9]*}}(%rsp), %[[X2:xmm[0-9]+]]
+; SSE2-NEXT: pmuludq %[[X2]], %xmm0
+; SSE2-NEXT: pshufd {{.*}} # [[X2]] = [[X2]][1,1,3,3]
+; SSE2-NEXT: pmuludq %[[X1]], %[[X2]]
+; SSE2-NEXT: shufps {{.*}} # xmm0 = xmm0[0,2],[[X2]][0,2]
+; SSE2-NEXT: shufps {{.*}} # xmm0 = xmm0[0,2,1,3]
+; SSE2-NEXT: addq ${{[0-9]+}}, %rsp
+; SSE2-NEXT: retq
+;
+; SSE41-LABEL: e:
+; SSE41: pmulld {{[0-9]+}}(%rsp), %xmm
+; SSE41-NEXT: addq ${{[0-9]+}}, %rsp
+; SSE41-NEXT: retq
+entry:
+ ; Use a call to force spills.
+ call void @foo()
+ %A = mul <4 x i32> %i, %j
+ ret <4 x i32> %A