[Bug tree-optimization/126464] [16/17 Regression] wrong code with FP ranges since r16-1108

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126464

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
           Assignee|unassigned at gcc dot gnu.org      |jakub at gcc dot gnu.org

--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Created attachment 65161
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=65161&action=edit
gcc17-pr126464.patch

Untested fix.
This perhaps extends it too much in some cases, e.g. that float [-inf, -inf]
on reverse from double extends to [-inf, -0x0.ffffffp+128] when for IEEE double
to float and rounding to nearest only [-inf, -0x0.ffffff8p+128] would be
strictly required.  But am not sure how to handle that portably...
Anyway, the problem is that any doiuble smaller or equal to that
-0x0.ffffff8p+128 when cast to float results in -inf.