[Bug tree-optimization/24021] VRP does not work with floating points

["rguenther at suse dot de" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=24021

--- Comment #17 from rguenther at suse dot de <rguenther at suse dot de> ---
On Thu, 10 Feb 2022, jakub at gcc dot gnu.org wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=24021
> 
> --- Comment #16 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
> But just tracking those fpclassify/signbit properties wouldn't be enough,
> because in many cases e.g. whether something can be infinite or not will depend
> on more precise value ranges.
> If we track just a bitmask, can the value be:
> zero
> subnormal
> normal
> infinite
> qNaN
> sNaN
> have positive signbit
> have negative signbit
> then even on simple multiplication or addition we'll need to assume from normal
> * normal or normal + normal that it can be infinite.  When we know that
> one operand is [-15.123 - epsilon, 23.152 + epsilon] and the other is
> [256.0 - epsilon, 512.0 + epsilon], we actually can find out the result will
> not be infinite etc.
> But sure, tracking in a bitmask the above properties in addition to some
> approximate range is useful.

Yes.  What I was trying to say is that while the value-range
propagation process should likely track actual FP value ranges it
remains to be seen whether we need to store those into the SSA name
info or whether tracking a set of flags is good enough in practice
(given that match.pd should generally not invoke ranger in
resolving mode but look at what's in SSA annotations).