This is the mail archive of the
gcc-bugs@gcc.gnu.org
mailing list for the GCC project.
[Bug target/52607] v4df __builtin_shuffle with {0,2,1,3} or {1,3,0,2}
- From: "marc.glisse at normalesup dot org" <gcc-bugzilla at gcc dot gnu dot org>
- To: gcc-bugs at gcc dot gnu dot org
- Date: Sun, 18 Mar 2012 18:58:44 +0000
- Subject: [Bug target/52607] v4df __builtin_shuffle with {0,2,1,3} or {1,3,0,2}
- Auto-submitted: auto-generated
- References: <bug-52607-4@http.gcc.gnu.org/bugzilla/>
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52607
--- Comment #6 from Marc Glisse <marc.glisse at normalesup dot org> 2012-03-18 18:58:44 UTC ---
Created attachment 26912
--> http://gcc.gnu.org/bugzilla/attachment.cgi?id=26912
generic shuffle of a single v8sf
An additional function (I should find better names...) to handle generic
shuffles of a single v8sf in 4 instructions. Only tested on {6,2,3,3,5,2,3,7}.
By the way, expand_vec_perm_vperm2f128_vblend2 does vpermilpd+vperm2f128 in
this order, but it would be better to do it in the reverse order (adapting the
mask), because it is common to need several __builtin_shuffle(x,*) and the
vperm2f128 can then be shared.
I also noticed while experimenting that -mavx2 generates vpermd instead of
vpermps (the vpermq->vpermpd change didn't affect that).