[Bug tree-optimization/126486] codesize regression by 17.5% for spec2017 507.catcuBSSN_r caused by r15-6657 with O2

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=3D126486

--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Hongtao Liu from comment #5)
> The canonicalization in r15-6657 fold signed integer operation into
> unsigned, overflow for signed integer is UD, and can be simplied in
> tree-affine, but the later one can't, add -fwrapv can also reproduce the
> issue without r15-6657.

I'll note that tree-affine generally ignores UB issues, but it does have
some "tricks" where it can expand widening/shortening conversions.  I also
found the r15-6657 change problematic when working on PR125730 and PR126008,
it's some entangled mess ...=