On Tue, Jul 03, 2018 at 08:17:25AM -0700, Richard Henderson wrote: > Beginning with John Arbuckle's fdiv test case, clean up some > of the fp helpers. As with fdiv, fre and fresqrt are missing > divide-by-zero exceptions. > > I've also noticed that load/store were using arithmetic conversions > to/from float32. These should be using the non-arithmetic algorithms > listed in the manual. > > Aside from cleaning up all of the rest of the helpers, I believe that > the implementation of the single-precision operations are incorrect. > They are currently implemented with the double-precision operation > followed by a round-to-single. This causes incorrect results via > double rounding. I believe better results could be had by using > these non-arithmetic converters to produce float32 operands, use the > proper float32 softfloat operations, and then convert back. > > Anyway, all of this has been broken long enough that it'll need to > wait til next devel cycle before anything further gets done.
Applied to ppc-for-3.1 (newly created).
--
David Gibson | I'll have my music baroque, and my code
david AT gibson.dropbear.id.au | minimalist, thank you. NOT _the_ _other_
| _way_ _around_!
http://www.ozlabs.org/~dgibson
signature.asc
Description: PGP signature
