Hi On 04/30, Jeff Law wrote: > On 4/30/19 8:00 AM, Jakub Jelinek wrote: > > On Tue, Apr 30, 2019 at 07:57:20AM -0600, Jeff Law wrote: > >>> Just curious, do we want to add math identities like above to match.pd ? > >> I'd think so. > >> > >> > >>> In practice, I am not sure how often we'd see "tanh * cosh" instead > >>> of sinh directly in source, > >> We're often surprised when what ultimately shows up in sources :-) And > >> a transformation that allows us to turn two transcendentals and a > >> multiplication into a single transcendental is going to be a big win. > >
I wonder why these kind of optimization are not in the EasyHacks page. It is somewhat easy to add it and can give huge performance improvements. There is a blogpost that I wrote about the previous patches I submitted here. I tried to be didadic, discussing from the mathematical standpoint, to floating point numbers, and GCC implementation. Although there may be several english errors, I think it is still useful: https://flusp.ime.usp.br/gcc/2019/03/26/making-gcc-optimize-some-trigonometric-functions/ > > I guess an important question is if such transformations need to be guarded > > by some -ffast-math suboptions, whether those transformations work properly > > for signed zeros, NaNs, sNaNs, infinities, guarantee the same ulp, have the > > same -ferrno-math behavior etc. > Yes. If you look at the discussion with Gualiano for the first set in > this space that was the hardest part. And for Barbara's we know there's > at least one issue in this space as well -- in particular it looks like > the long double handling is incorrect in the sign bit for very small > inputs. The question is precisely where that's gone wrong :-) > > jeff Giuliano.
