https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107591
--- Comment #17 from Jakub Jelinek <jakub at gcc dot gnu.org> --- (In reply to Andrew Macleod from comment #16) > (In reply to Jakub Jelinek from comment #15) > > I've screwed up the real_value_negate calls, they need to assign the result. > > > > Anyway, yet another option would be for cdce to ask the ranger if the sqrt > > argument can be smaller than -0.0 (and only if it can, emit the comparison). > > Then we don't need to deal with removing the comparison again. > > Of course, such removals might be still useful for user code. > > even checking the global range from cdce at this point should show that its > >= -0.0... Without larger changes in tree-call-cdce.cc, perhaps we could change gen_one_condition to check the global or even non-global frange (but in the latter case the question is at what stmt) and if the range indicates that tcode of arg against (float_type) lbub (aka lbub_real_cst) is always true or always false, simply emit an always true or always false condition (though the domain ranges aren't exact but rounded up/down to integers and I think even there is sometimes an off by one conservatively, so probably only always true conditions) rather than actual arg comparison. Even better would be not to emit any and/or propagate a stmt to it, but am not sure about either (the code treats nconds == 0 differently and there are separators in between sometimes; and for stmt, it uses some code for better placement of a check, but not sure if say a range could be true on the sqrt etc. call and then not on the placement of the condition which it picks). In any case, not much familiar with the range_of_expr etc. APIs.
