1; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=-sse4.1 | FileCheck %s --check-prefix=NO_SSE_41
2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.1 | FileCheck %s --check-prefix=SSE_41
3
4; PR20472 ( http://llvm.org/bugs/show_bug.cgi?id=20472 )
5; When sexting a trunc'd vector value, we can't eliminate the zext.
6; If we don't have SSE4.1, use punpck.
7; If we have SSE4.1, use pmovzx because it combines the load op.
8; There may be a better way to do this using pshufb + pmovsx,
9; but that is beyond our current codegen capabilities.
10
11define <4 x i32> @trunc_sext(<4 x i16>* %in) {
12; NO_SSE_41-LABEL: trunc_sext:
13; NO_SSE_41:       # BB#0:
14; NO_SSE_41-NEXT:    movq {{.*#+}} xmm0 = mem[0],zero
15; NO_SSE_41-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
16; NO_SSE_41-NEXT:    pslld $24, %xmm0
17; NO_SSE_41-NEXT:    psrad $24, %xmm0
18; NO_SSE_41-NEXT:    retq
19;
20; SSE_41-LABEL: trunc_sext:
21; SSE_41:       # BB#0:
22; SSE_41-NEXT:    pmovzxwd {{.*#+}} xmm0 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero
23; SSE_41-NEXT:    pslld $24, %xmm0
24; SSE_41-NEXT:    psrad $24, %xmm0
25; SSE_41-NEXT:    retq
26  %load = load <4 x i16>, <4 x i16>* %in
27  %trunc = trunc <4 x i16> %load to <4 x i8>
28  %sext = sext <4 x i8> %trunc to <4 x i32>
29  ret <4 x i32> %sext
30}
31
32