https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91940
--- Comment #12 from Jorn Wolfgang Rennecke ---
(In reply to Jakub Jelinek from comment #11)
But the condition I quoted rejects the recognition of a bswap16 with
non-promoted
arguments.
vectorizable_bswap doesn't do anything for processors that
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91940
--- Comment #11 from Jakub Jelinek ---
Only 16-bit byteswap has the r>> 8 canonical form, larger byteswaps don't.
Larger byteswaps certainly aren't rotates, but just permutes. So, if the
vectorizer doesn't try that already, it should try to vect
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91940
--- Comment #10 from Jorn Wolfgang Rennecke ---
Even if you add support for V2HI bswap, it won't help vectorization without
support for V4QI vectors and permutations, because vectorizable_bswap won't
recognize the bswap capability of the target a
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91940
Jorn Wolfgang Rennecke changed:
What|Removed |Added
CC||amylaar at gcc dot gnu.org
---
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91940
Jakub Jelinek changed:
What|Removed |Added
Status|ASSIGNED|RESOLVED
Resolution|---
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91940
--- Comment #7 from Marc Glisse ---
(In reply to Jakub Jelinek from comment #1)
> The loop with the rotate is vectorized, while the one with __builtin_bswap16
> is not.
It is a bit surprising that we do not canonicalize one to the other somewher
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91940
--- Comment #6 from Jakub Jelinek ---
Author: jakub
Date: Wed Oct 2 10:18:50 2019
New Revision: 276442
URL: https://gcc.gnu.org/viewcvs?rev=276442&root=gcc&view=rev
Log:
PR tree-optimization/91940
* tree-vect-patterns.c: Include
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91940
--- Comment #5 from Jakub Jelinek ---
Created attachment 46985
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=46985&action=edit
gcc10-pr91940.patch
Full untested patch.
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91940
--- Comment #4 from Richard Biener ---
Looks good from a quick look.
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91940
--- Comment #3 from Jakub Jelinek ---
Untested WIP patch that does both.
If it finds vectorize_bswap will work (the corresponding permutation is
supported), it will just undo the promotion, if target supports vector rotates,
will use vector rotat
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91940
--- Comment #2 from Richard Biener ---
Another option is to elide the promotion?
int foo (unsigned short x)
{
return __builtin_bswap16 (x);
}
return (int) __builtin_bswap16 ((int) x);
but BUILT_IN_BSWAP16 is BT_FN_UINT16_UINT16, not sure w
11 matches
Mail list logo
