[Bug target/52607] v4df __builtin_shuffle with {0,2,1,3} or {1,3,0,2}

["marc.glisse at normalesup dot org" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52607

--- Comment #21 from Marc Glisse <marc.glisse at normalesup dot org> 2012-03-27 18:21:39 UTC ---
(In reply to comment #20)
> I don't like much the calls to ix86_expand_vec_perm_const_1, if you are looking
> for exactly two insn permutations,

Actually, it isn't just 2 insn. The call in expand_vec_perm_vperm2f128_merge
can take 3, and the calls in expand_vec_perm_perm_blend(...,true) up to 4 (this
is how I get a maximum of 9 insn, 1+2*4). But some more splits of
ix86_expand_vec_perm_const_1 to avoid recursive calls should be doable, if you
don't like the recursion.

> then really the two insn permutation
> functions should be groupped together into expand_vec_perm_2 and you should
> call that instead, or if it is 1 or 2, then expand_vec_perm_1 ||
> expand_vec_perm_2.

Yes, this grouping by size makes sense, whether it ends up being used or not.
Although there are expanders in the "3" category that occasionally get lucky
and generate only 2 :-)

> expand_vec_perm_vperm2f128_merge has probably swapped the meaning of dfirst and
> dsecond permutations when it first performs the dsecond permutation.

If you are just talking of the naming of the variables, yes, I completely agree
they should be swapped (or given more explicit names, like swap_lanes and
dintra).