On Wed, 2019-03-13 at 14:38 +0000, Joseph Myers wrote: > > ------------------------------------------------------------------- > --- > On Wed, 13 Mar 2019, Jakub Jelinek wrote: > > > If the finite only doesn't buy anything, then another option is to drop the > > math-finite.h stuff or portions thereof. > > Well, finite-only entry points avoid wrappers for the scalar functions - > it's just that suitable optimized implementations could avoid the wrappers > in all cases without needing a separate finite-only variant. It's not > clear that adding wrappers in this case for scalar functions to avoid them > for vector functions is a good idea. And regardless of the merits of a > particular set of entry points, I think requiring the same set of variants > for both vector and scalar functions is flawed; the headers should be able > to declare a scalar variant to be used under certain conditions without > requiring a corresponding vector variant, or of course the other way > round.
If some targets don't need _finite variants it might be useful to tell the compiler not to generate them. For example, on aarch64 we know that __exp_finite and __expf_finite are just aliases of exp and expf, so if we could turn off the math-finite.h functionality for those functions on this target, GCC would also not generate calls to the vector versions of exp_finite and expf_finite. This doesn't directly address the issue of the scalar and vector variants being independent of each other but it does make the problem moot in this specific case. Steve Ellcey sell...@marvell.com
