Re: [PATCH 0/4] math: Fix inaccurate sin/cos/tan for large arguments (BZ 34376)
Adhemerval Zanella Netto <[email protected]>
| Newsgroups | gmane.comp.lib.glibc.alpha |
|---|---|
| Organization | Linaro |
| Message-ID | <[email protected]> |
On 17/07/26 08:18, Stefan Kanthak wrote: > "Adhemerval Zanella" <[email protected]> wrote: > >> The fix restores the fdlibm e_rem_pio2.c [...] > > Don't you have (or want a better) alternative? > How big is this code, and how fast^Wslow does it run? Not right now where we at near the hard freeze for 2.44 release. > > JFTR: fdlibm was written for 32-bit machines, it doesn't use 64-bit integer > arithmetic supported since 20 years on mainstream processors, or via > (hopefully properly optimising) compilers on 32-bit processors since > the last millenium. > > If it has more than about 60 instructions on x86-64 or about 140 instructions > on x86/x87: DUMP IT! > > Take a look at <https://skanthak.hier-im-netz.de/ieee-754.html#payne-hanek> > for comparison. This is essentially a new implementation and this is material for next release.