1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=-sse4.1 | FileCheck %s --check-prefix=NO_SSE_41
3; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.1 | FileCheck %s --check-prefix=SSE_41
4
5; PR20472 ( http://llvm.org/bugs/show_bug.cgi?id=20472 )
6; When sexting a trunc'd vector value, we can't eliminate the zext.
7; If we don't have SSE4.1, use punpck.
8; If we have SSE4.1, use pmovzx because it combines the load op.
9; There may be a better way to do this using pshufb + pmovsx,
10; but that is beyond our current codegen capabilities.
11
12define <4 x i32> @trunc_sext(<4 x i16>* %in) {
13; NO_SSE_41-LABEL: trunc_sext:
14; NO_SSE_41:       # %bb.0:
15; NO_SSE_41-NEXT:    movq {{.*#+}} xmm0 = mem[0],zero
16; NO_SSE_41-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
17; NO_SSE_41-NEXT:    pslld $24, %xmm0
18; NO_SSE_41-NEXT:    psrad $24, %xmm0
19; NO_SSE_41-NEXT:    retq
20;
21; SSE_41-LABEL: trunc_sext:
22; SSE_41:       # %bb.0:
23; SSE_41-NEXT:    pmovzxwd {{.*#+}} xmm0 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero
24; SSE_41-NEXT:    pslld $24, %xmm0
25; SSE_41-NEXT:    psrad $24, %xmm0
26; SSE_41-NEXT:    retq
27  %load = load <4 x i16>, <4 x i16>* %in
28  %trunc = trunc <4 x i16> %load to <4 x i8>
29  %sext = sext <4 x i8> %trunc to <4 x i32>
30  ret <4 x i32> %sext
31}
32
33