This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: match.pd: Optimize (x & y) ^ (x | y)
- From: Marek Polacek <polacek at redhat dot com>
- To: Marc Glisse <marc dot glisse at inria dot fr>
- Cc: Richard Biener <rguenther at suse dot de>, Jakub Jelinek <jakub at redhat dot com>, GCC Patches <gcc-patches at gcc dot gnu dot org>
- Date: Fri, 12 Jun 2015 11:02:04 +0200
- Subject: Re: match.pd: Optimize (x & y) ^ (x | y)
- Authentication-results: sourceware.org; auth=none
- References: <20150611110432 dot GY2756 at redhat dot com> <20150611110905 dot GW10247 at tucnak dot redhat dot com> <20150611120246 dot GZ2756 at redhat dot com> <alpine dot DEB dot 2 dot 20 dot 1506112127430 dot 1589 at laptop-mg dot saclay dot inria dot fr> <F754CE2F-C241-4380-996E-93B9EA4A03F0 at suse dot de> <alpine dot DEB dot 2 dot 20 dot 1506120834490 dot 1593 at laptop-mg dot saclay dot inria dot fr>
On Fri, Jun 12, 2015 at 08:44:45AM +0200, Marc Glisse wrote:
> On Fri, 12 Jun 2015, Richard Biener wrote:
>
> >>Not judging at all whether it is desirable or not, but you might have
> >>hit the issue that when you want several convert?, you need to use the
> >>spelling convert1?, convert2?, and it stops there, while here you would
> >>probably want at least 4 (maybe 6?) for this case. You might be able to
> >>work around it with a user-defined predicate, but I keep getting errors
> >>like
> >>generic-match.c:6655:16: error: redeclaration of âtree_node* o20_pops [2]â
I remember trying convert1?, convert2?, and so on. But I gave up soon.
> >>If you want to reproduce the error (this is probably not good as is, it
> >>is only provided as a reproducer)
> >>
> >>(match (nopand @0 @1)
> >> (bit_and (convert1? @0) (convert2? @1))
> >> (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
> >> && tree_nop_conversion_p (type, TREE_TYPE (@1)))))
> >>(match (nopior @0 @1)
> >> (bit_ior (convert1? @0) (convert2? @1))
> >> (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
> >> && tree_nop_conversion_p (type, TREE_TYPE (@1)))))
> >>(simplify
> >> (bit_xor:c (convert1? (nopand@2 @0 @1))
> >> (convert2? (nopior@3 @0 @1)))
> >> (if (tree_nop_conversion_p (type, TREE_TYPE (@2))
> >> && tree_nop_conversion_p (type, TREE_TYPE (@3)))
> >> (bit_xor (convert @0) (convert @1))))
> >>
> >>
> >>fold-const.c traditionally avoided the combinatorial explosion by using
> >>strip_nops.
> >
> >Yeah. We can probably special case conditional conversions in code
> >generation instead of lowering it. And then go the full way and special
> >case nop conversions so you can avoid writing the predicate as well.
>
> Without special casing, I currently have:
>
> (match (nopcvt @0)
> (convert? @0)
> (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))))
> (simplify
> (bit_xor:c (convert1? (bit_and@2 (nopcvt @0) (nopcvt @1)))
> (convert2? (bit_ior:c (nopcvt @0) (nopcvt @1))))
> (if (tree_nop_conversion_p (type, TREE_TYPE (@2)))
> (bit_xor (convert @0) (convert @1))))
>
> which simplifies Jakub's testcase without exploding the size of *-match.c,
> but it is still not very satisfying.
Yeah, imagine if we'd have to change every pattern like that :-(.
Marek