Re: RFA: Add ! LDBL_EQ_DBL implementations of cosl, sinl and tanl

[Corinna Vinschen <vinschen at redhat dot com>]

On Mar  2 13:18, Craig Howland wrote:
> On 03/02/2015 12:52 PM, Nick Clifton wrote:
> >Hi Jeff, Hi Corinna,
> >
> >   Attached is a patch that implements cosl(), sinl() and tanl() when
> >   long double is not the same as double.  In the course of implementing
> >   these functions I found that I also needed to implement the fabls()
> >   and floorl() functions as well as several internal library support
> >   routines.
> >
> >   Tested in the usual way with an x86_64-linux-gnu toolchain with
> >   direct linking to the math functions.
> >
> >   OK to apply ?
> >
> >Cheers
> >   Nick
> >
> >newlib/ChangeLog
> >2015-03-02  Nick Clifton  <...>
> >
> >	* libm/common/cosl.c (cosl): Add implementation for when long
> >	double is not the same as double.
> >	* libm/common/fabls.c (fabls): Likewise.
> >	* libm/common/floorl.c (floorl): Likewise.
> >	* libm/common/sinl.c (sinl): Likewise.
> >	* libm/common/tanl.c (tanl): Likewise.
> >	* libm/common/fdlibmh.h: Add prototypes for __ieee754_rem_pio2l,
> >	__kernel_cosl, __kernel_cosl,  __kernel_sinl and __kernel_tanl.
> >	* libm/math/e_rem_pio2l.c: New file.  Implements
> >	__ieee754_rem_pio2l.
> >	* libm/math/k_cosl.c: New file.  Implements __kernel_cosl.
> >	* libm/math/k_sinl.c: New file.  Implements __kernel_sinl.
> >	* libm/math/k_tanl.c: New file.  Implements __kernel_tanl.
> >	* libm/math/Makefile.am (src): Add k_cosl.c, k_sinl.c, k_tanl.c
> >	and e_rem_pio2l.c.
> >	* libm/math/Makefile.in: Regenerate.

First of all, thanks for this patch, Nick.

> math.h needs to be changed, too, to take the new routines out of the long
> double==double checks.

That's right, a math.h patch is missing.

> These routines are not fully C99/POSIX-compliant in terms of floating point
> exceptions and error returns if IEC 60559 is being used.  (See, for example,
> http://pubs.opengroup.org/onlinepubs/9699919799/functions/cosl.html, under
> the "MX" markings, for the error-return considerations.)

Neither are all our non-long double functions.  Have a look at
libm/math/s_cos.c, for instance.

> Do we want to start out with this consideration missing?  I would expect
> that the vast majority of people probably do not care about the floating
> point exception aspect.  Setting errno, however, is quite a different
> question, for which I have no idea how to gauge interest.

Setting errno should be easy to add:

- cosl, sinl and tanl just need an extra check for Inf to set errno to
  EDOM.

- tanl needs an extra check if the result of a non-Inf arg is Inf to
  return ERANGE.

- fabsl doesn't define errors>

- floorl doesn't produce errors (see Linux man page),

Nick, any problem to add this?

However, that also means our long double functions will behave
differently than our non-long double functions.  Is anybody willing to
making a sweep through our non-long double functions to add errno
setting where missing?


Thanks,
Corinna

-- 
Corinna Vinschen
Cygwin Maintainer
Red Hat

Attachment: pgpYtGoCdqajc.pgp
Description: PGP signature