RE: [PATCH] Use rtx_refs_may_alias_p instead of alias_sets_conflict_p in

["Bingfeng Mei" <bmei at broadcom dot com>]

Hi,
I created a new may_alias_p function as Richard proposed. The function
disables offset-based memory disambiguation and TBAA, it is also slightly
more conservative than true/output/anti_dependence. For example, 
write_dependence checks: 
  if (GET_MODE (x) == BLKmode && GET_CODE (XEXP (x, 0)) == SCRATCH)
    return 1;
  if (GET_MODE (mem) == BLKmode && GET_CODE (XEXP (mem, 0)) == SCRATCH)
    return 1; 

true_depdence checks both above and following (isn't above 
statement redundant then?).
  if (mem_mode == BLKmode || GET_MODE (x) == BLKmode)
    return 1;
may_alias_p takes more conservative check of the latter one only. 

Since insn_alias_sets_conflict_p is misnamed and only used once by ddg.c,
it & friends (walk_mems_1, walk_mems_2) are moved into ddg.c and
renamed as insns_may_alias_p. 

The patch is bootstrapped on x86_64-unknown-linux-gnu and tested.
OK for trunk? 

Thanks,
Bingfeng


2010-08-10  Bingfeng Mei  <bmei@broadcom.com>

	* ddg.c (walk_mems_2): Moved from alias.c, use may_alias_p instead of
	alias_sets_conflict_p.
	(walk_mems_1): Moved from alias.c.
	(insns_may_alias_p): New function, originally insn_alias_sets_conflict_p
	in alias.c. 
        (add_inter_loop_mem_dep): Use insns_may_alias_p now.
        * cse.c (cse_insn): New argument in calling nonoverlapping_memrefs_p.
        * alias.c (walk_mems_2): Moved to ddg.c.
	(walk_mems_1): Ditto.
	(insn_alias_sets_conflict_p): Renamed to insns_may_alias_p and moved
	to ddg.c.
	(nonoverlapping_memrefs_p): Add flag to guard offset-based memory
	disambiguation.
	*(may_alias_p): New function to check whether two memory expression
	may alias or not. Currently used in buidling inter-iteration memory
	dependence.
	*alias.h (nonoverlapping_memrefs_p): New flag as third argument.
	(insn_alias_sets_conflict_p): Removed
	*rtl.h (may_alias_p): New function prototype.

> -----Original Message-----
> From: Richard Guenther [mailto:richard.guenther@gmail.com]
> Sent: 06 August 2010 11:10
> To: Bingfeng Mei
> Cc: Michael Matz; Diego Novillo; gcc-patches@gcc.gnu.org
> Subject: Re: [PATCH] Use rtx_refs_may_alias_p instead of
> alias_sets_conflict_p in
> 
> On Thu, Aug 5, 2010 at 7:00 PM, Bingfeng Mei <bmei@broadcom.com> wrote:
> > Hi,
> > If we are going to use true_dependence function etc in ddg.c,
> > we should be able to disable loop variant stuff (currently
> > I can only think of offset-based memory disambiguation and TBAA).
> > I added a flag loop_invariant to relevant functions in attached
> > patch. The patch is tested and bootstrapped. Is this OK?
> >
> > Next, I need to adapt ddg.c to use these XXX_dependence functions,
> > move and rename insn_alias_sets_conflict_p & friends to ddg.c.
> > The original add_inter_loop_mem_dep is over-simplified.
> > For example, it doesn't consider parallel patterns with more than
> > one memory expressions.
> 
> +  if (!loop_invariant &&
> 
> &&s go to the next line.
> 
>    if (DIFFERENT_ALIAS_SETS_P (x, mem))
>      return 0;
> 
> -  if (nonoverlapping_memrefs_p (mem, x))
> +  if (nonoverlapping_memrefs_p (mem, x, loop_invariant))
>      return 0;
> 
> The DIFFERENT_ALIAS_SETS_P would also need guarding.
> 
> I think that instead of changing all present functions it would be
> better to create a new entry to the alias oracle for inter-iteration
> queries, as for example you can't distinguish true and anti dependence
> (which is also why TBAA isn't valid here).  In fact the implementation
> would be the same for all of RW, WR and WW dependence types.
> 
> The query would exactly match what Zdenek thinks of a
> may-alias query, so I'd suggest to simply name it may_alias_p.
> 
> Richard.
> 
> > Thanks
> > Bingfeng
> >
> >> -----Original Message-----
> >> From: Michael Matz [mailto:matz@suse.de]
> >> Sent: 04 August 2010 17:15
> >> To: Bingfeng Mei
> >> Cc: Richard Guenther; Diego Novillo; gcc-patches@gcc.gnu.org
> >> Subject: RE: [PATCH] Use rtx_refs_may_alias_p instead of
> >> alias_sets_conflict_p in
> >>
> >> Hi,
> >>
> >> On Wed, 4 Aug 2010, Bingfeng Mei wrote:
> >>
> >> > Does rtx_refs_may_alias_p include offset based disambiguation
> >> (with/without
> >> > tbaa flag).
> >>
> >> Actually it doesn't (we switch off recognizing SSA names as being
> equal
> >> when called from RTL land). ?I forgot that this was recently changed.
> >>
> >> Unfortunately I was wrong with my assertion that TBAA can be used
> cross
> >> iteration. ?The catch is the new memory model, with it's asymmetric
> >> dependencies: you may move a load before a non-conflicting (type
> >> wise) store, but you may not move a store before a non-conflicting
> load.
> >>
> >> In cross-iteration situations you are posed with both cases, hence
> you
> >> can't really use type-based disambiguation (perhaps in some few
> special
> >> cases one might, but not generally).
> >>
> >> Basically for cross iteration analysis you can only use points-to
> >> information (or generally everything that is loop invariant for the
> >> loop
> >> in question, but that isn't implemented).
> >>
> >> > If yes, this patch is not correct, I need to work out a better one.
> >>
> >> It would be better to use the four different predicates we have in
> RTL
> >> land as Richard suggested, they do the right thing for each
> dependency
> >> type (switching off TBAA when appropriate).
> >>
> >>
> >> Ciao,
> >> Michael.
> >
> >

Attachment: patch_maybe_alias_p;size=7694;creation-date="Mon,
Description: patch_maybe_alias_p