[Bug target/65832] Inefficient vector construction

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65832

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
typedef int v4si __attribute__((vector_size(16)));

v4si bar (int *i, int *j, int *k, int *l)
{
  return (v4si) { *i, *j, *k, *l };
}

looks reasonable (no spills at least, stray move for the return value).

        movd    (%rsi), %xmm0
        movd    (%rdi), %xmm3
        movd    (%rcx), %xmm1
        movd    (%rdx), %xmm2
        punpckldq       %xmm0, %xmm3
        punpckldq       %xmm1, %xmm2
        movdqa  %xmm3, %xmm0
        punpcklqdq      %xmm2, %xmm0

With -mavx2 we get

        vmovd   (%rdx), %xmm2
        vmovd   (%rdi), %xmm3
        vpinsrd $1, (%rcx), %xmm2, %xmm1
        vpinsrd $1, (%rsi), %xmm3, %xmm0
        vpunpcklqdq     %xmm1, %xmm0, %xmm0