On 2012-06-28 09:20, Jakub Jelinek wrote:
> Perhaps the problem is then that the permutation is much more expensive
> for even/odd. With even/odd the f2 routine is:
...
> vpshufb %xmm2, %xmm5, %xmm5
> vpshufb %xmm1, %xmm4, %xmm4
> vpor %xmm4, %xmm5, %xmm4
...
> and with lo/hi it is:
> vshufps $221, %xmm2, %xmm3, %xmm2
Hmm. That second has a reformatting delay.
Last week when I pulled the mulv4si3 routine out to i386.c,
I experimented with a few different options, including that
interleave+shufps sequence seen here for lo/hi. See the
comment there discussing options and timing.
This also shows a deficiency in our vec_perm logic:
0L 0H 2L 2H 1L 1H 3L 3H
0H 2H 0H 2H 1H 3H 1H 3H 2*pshufd
0H 1H 2H 3H punpckldq
without the permutation constants in memory.
r~
