Re: [PATCH, Fix PR 31098] a*C == b*C is not foldded to a == b with --fast-math/floating point

[Pranav Bhandarkar <pranav dot bhandarkar at celunite dot com>]

> Do we really need all "fast math" flags set to enable this opt?  E.g.
> surely we don't care about !flag_signed_zeros since you disable the code
> when "c" is zero.  Perhaps other sub-bits of "fast math" can also be
> ignored for this purpose.
I had put the fast math check for the sake of being consistent with the 
PR description.
>  E.g. try testing various
> different permutations of changing CMP or OP.
Yes, I have tried a number of other permutations with good results.

> Or how about (a OP c) CMP
> (c OP b) when OP is commutative?
Yes, I tried this too and you are right the expressions are 
canonicalized by this point so this works anyways.
However, adding that testcase is a good idea and I shall post that here 
in a short while alongwith some of the other
permutations that I have tried.

Thanks,
Pranav

Pranav Bhandarkar
GNU Tools
Celunite, Inc.
(www.celunite.com)
+91-98220 82881



Kaveh R. GHAZI wrote:
> On Sun, 7 Oct 2007, Pranav Bhandarkar wrote:
>
>   
> > Hi,
> > Could somebody please review this patch which is a proposed fix for PR
> > 31098.
> >
> > 2007-10-07  Pranav Bhandarkar  <pranav.bhandarkar@celunite.com>
> >
> >    * fold-const.c (fold_comparison): fold (a OP c) CMP (b OP c) -> a
> >    CMP b, iff c is not zero
> >    * testsuite/pr31098.c: New
> >     
>
>   
> > +      /* (a OP c) CMP (b OP c) -> a CMP b, iff c is not zero */
> > +      if (fast_math_flags_set_p () && (TREE_CODE (arg0) == TREE_CODE (arg1))
> >     
>
> Do we really need all "fast math" flags set to enable this opt?  E.g.
> surely we don't care about !flag_signed_zeros since you disable the code
> when "c" is zero.  Perhaps other sub-bits of "fast math" can also be
> ignored for this purpose.
>
> Also I would suggest a few more testcases (not necessarily more test
> files, you can put them all in the same file).  E.g. try testing various
> different permutations of changing CMP or OP.  Or how about (a OP c) CMP
> (c OP b) when OP is commutative?  I believe the exprs are canonicalized at
> this point so adding a testcase for one side reversed should just work.
>
> 		Thanks,
> 		--Kaveh
> --
> Kaveh R. Ghazi			ghazi@caip.rutgers.edu
>
>