Re: [PATCH] middle-end IFN_ASSUME support [PR106654]

[Richard Biener <rguenther@suse.de>]

[-- Attachment #1: Type: text/plain, Size: 3113 bytes --]

On Wed, 12 Oct 2022, Andrew MacLeod wrote:

> 
> On 10/12/22 10:39, Jakub Jelinek wrote:
> > On Wed, Oct 12, 2022 at 10:31:00AM -0400, Andrew MacLeod wrote:
> >> I presume you are looking to get this working for this release, making the
> >> priority high? :-)
> > Yes.  So that we can claim we actually support C++23 Portable Assumptions
> > and OpenMP assume directive's hold clauses for something non-trivial so
> > people won't be afraid to actually use it.
> > Of course, first the posted patch needs to be reviewed and only once it gets
> > in, the ranger/GORI part can follow.  As the latter is only an optimization,
> > it can be done incrementally.
> 
> I will start poking at something to find ranges for parameters from the return
> backwards.

If the return were

  if (return_val)
    return return_val;

you could use path-ranger with the parameter SSA default defs as
"interesting".  So you "only" need to somehow interpret the return
statement as such and do path rangers compute_ranges () 

> 
> >> Intersection I believe...?  I think the value from the assume's should add
> >> restrictions to the range..
> > Sure, sorry.
> >
> >> I figured as much, I was just wondering if there might be some way to
> >> "simplify" certain things by processing it and turning each parameter query
> >> into a smaller function returning the range we determined from the main
> >> one...   but perhaps that is more complicated.
> > We don't really know what the condition is, it can be pretty arbitrary
> > expression (well, e.g. for C++ conditional expression, so say
> > [[assume (var = foo ())]];
> > is not valid but
> > [[assume ((var = foo ()))]];
> > is.  And with GNU statement expressions it can do a lot of stuff and until
> > we e.g. inline into it and optimize it a little, we don't really know what
> > it will be like.
> >
> >  
> 
> No, I just meant that once we finally process the complicated function, and
> decide the final range we are storing is for x_1 is say [20,30], we could
> replace the assume call site with something like
> 
>   int assume03_x (x) { if (x>= 20 || x <= 30) return x; gcc_unreachable(); }
> 
> then at call sites:
> 
>    x_5 = assume03_x(x_3);
> 
> For that matter, once all the assume functions have been processed, we could
> textually replace the assume call with an expression which represents the
> determined range...  Kind of our own mini inlining?  Maybe thats even better
> than adding any kind of support in fold_using_range..   just let things
> naturally fall into place?
> 
> .ASSUME_blah ( , , x_4);
> 
> where if x is determined to be [20, 30][50,60] could be textually "expanded"
> in the IL with
> 
>   if (x<20 || x>60 || (x>30 && x < 50)) gcc_unreachcable();
> 
> for each of the parameters?   If we processed this like early inlining, we
> could maybe expose the entire thing to optimization that way?
> 
> Andrew
> 
> 
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE Software Solutions Germany GmbH, Frankenstrasse 146, 90461 Nuernberg,
Germany; GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman;
HRB 36809 (AG Nuernberg)