https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88626
--- Comment #4 from Jan Hubicka <hubicka at gcc dot gnu.org> --- Also note that builtin_constant_p is handled conservatively. If you have if (builtin_constant_p (x)) constant code; else nonconstant code; inliner will know that nonconstant code will be optimized out if x is constant, but it will not know that constant code will be optimized out if x is variable (because it never knows that info, only later optimizations will decide) and it does not know that it is always the case that one of the branches will be optimized out. The last problem I think is most serious about our BCP handling. It may be possible to compute size/time vectors of the two variants and then produce final one that will take constant variant when x is constant and longer of the two variants when x is not, but it will need some work. Honza
