On Mon, Dec 9, 2024 at 5:27 PM LIU Hao <lh_mo...@126.com> wrote: > > In mingw-w64 it's implemented in 'mingw-w64-crt/math/powi.def.h'. Mostly > it's because rounding > errors accumulate. > > Maybe `powl` shouldn't be implemented that way, however I don't know how > to do that correctly. >
I doubt that I'd be of much use in trying to rewrite it. And as someone who has been relying on others to build gcc for me, my capacity to test is probably also limited. The issue seems to be with powl() specifically. At least, if a similar issue exists with pow() and/or powq(), then I haven't struck it. The debian devs apparently fixed this powl() problem in glibc-2.17, and then made more changes in 2.18 to improve on the performance degradation that 2.17 had suffered as a result of the fix. Correction: Actually, there used to be some off-by-one-ULP errors (not limited to MSWindows) with powq (eg powq(2.0Q, 0.5Q), but they all seem to have been fixed in recent gcc releases. Anyway .... if there's some way of getting this fixed, that would be excellent. Otherwise, I'll just accept that it requires a workaround whenever correctness is expected. Thanks for replying so promptly. _______________________________________________ Mingw-w64-public mailing list Mingw-w64-public@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/mingw-w64-public
