https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109008
--- Comment #17 from Richard Biener <rguenth at gcc dot gnu.org> --- (In reply to Jakub Jelinek from comment #15) > I believe worst case the above would do (for IEEE quad) something like 16 > checks because of 16-bit exponent, roughly one range_arithmetic, one (for > mult/div 2?) frange_arithmetic and real_equal + real_isfinite each, usually > much less. Plus if we narrow it down further up to mantissa precision bits > more. But at least for the latter we could have some hard limit, don't go > further than say 30 bits at most or so. > Once/if it makes correctly, we can just grab statistics across some math > heavy project (suggestions welcome). I think FP _compares_ are not so common, but nevertheless I'd suggest to simply use SPEC CPU 2017 FP and/or polyhedron
