Re: [PATCH] FIx endless match.pd recursion on cst1 + cst2 + cst3 (PR tree-optimization/84334)

[Jakub Jelinek <jakub@redhat.com>]

On Wed, Feb 14, 2018 at 12:09:57PM +0100, Richard Biener wrote:
> On Tue, 13 Feb 2018, Marc Glisse wrote:
> 
> > On Tue, 13 Feb 2018, Richard Biener wrote:
> > 
> > > On February 13, 2018 6:51:29 PM GMT+01:00, Jakub Jelinek <jakub@redhat.com>
> > > wrote:
> > > > Hi!
> > > > 
> > > > On the following testcase, we recurse infinitely, because
> > > > we have float re-association enabled, but also rounding-math, so
> > > > we try to optimize (cst1 + cst2) + cst3 as (cst2 + cst3) + cst1
> > > > but (cst2 + cst3) doesn't simplify and we try again and optimize
> > > > it as (cst3 + cst1) + cst2 and then (cst1 + cst2) + cst3 and so on
> > > > forever.  If @0 is not a CONSTANT_CLASS_P, there is not a problem,
> > > > if it is, the code just checks if we can actually simplify the
> > > > operation between cst2 and cst3 into a constant.
> > > 
> > > Is there a reason to try simplifying at all for constant @0?
> > 
> > Yes. cst2+cst3 might simplify (the operation happens to be exact and not
> > require rounding), which leaves us with only one addition instead of 2.
> > 
> > On the other hand, mixing -frounding-math with reassociation seems strange to
> > me, and likely not worth optimizing for.
> 
> ./cc1 -quiet t.c -O -frounding-math -fassociative-math
> cc1: warning: -fassociative-math disabled; other options take precedence

You need
./cc1 -quiet t.c -O -fassociative-math -fno-trapping-math -fno-signed-zeros -frounding-math

> So _maybe_ we should disable these patterns for !flag_associative_math
> when dealing with FP?

We do, this is in block with:
 /* We can't reassociate floating-point unless -fassociative-math
    or fixed-point plus or minus because of saturation to +-Inf.  */
 (if ((!FLOAT_TYPE_P (type) || flag_associative_math)
      && !FIXED_POINT_TYPE_P (type))

But that doesn't mean you can't request associative math and rounding math
at the same time.

	Jakub