[AArch64] Add inlines for signbit

  Add AArch64 inlines for __signbit, __signbitf and __signbitl which default to the GCC builtin
functions.

Is there a reason this couldn't be done by default in math.h similar to isgreater (unlike
__builtin_isinf et al, GCC implements signbit efficiently and correctly).

OK for commit?

2015-04-15  Wilco Dijkstra  <wdijkstr@arm.com>

	* sysdeps/aarch64/fpu/bits/mathinline.h: New file.
	(__signbit): New function, use __builtin_signbit.
	(__signbitf): Likewise.  (__signbitl): Likewise.

---
 sysdeps/aarch64/fpu/bits/mathinline.h | 49 +++++++++++++++++++++++++++++++++++
 1 file changed, 49 insertions(+)
 create mode 100644 sysdeps/aarch64/fpu/bits/mathinline.h
  

On Wed, 15 Apr 2015, Wilco Dijkstra wrote:

> Is there a reason this couldn't be done by default in math.h similar to isgreater (unlike
> __builtin_isinf et al, GCC implements signbit efficiently and correctly).

It is, however, not type-generic in GCC 
<https://gcc.gnu.org/bugzilla/show_bug.cgi?id=36757>, so you'd still need 
to have a macro expansion checking sizeof.
  

> Joseph Myers wrote:
> On Wed, 15 Apr 2015, Wilco Dijkstra wrote:
> 
> > Is there a reason this couldn't be done by default in math.h similar to isgreater (unlike
> > __builtin_isinf et al, GCC implements signbit efficiently and correctly).
> 
> It is, however, not type-generic in GCC
> <https://gcc.gnu.org/bugzilla/show_bug.cgi?id=36757>, so you'd still need
> to have a macro expansion checking sizeof.

Yes indeed. However the issue is how to deal with the targets that currently
define inlines. Would it be reasonable just to leave them and add a signbit
expansion for GCC >= 3.0 that directly uses the builtins?
(if we define __signbit(f/l) to be __builtin_signbit(f/l), things may go 
wrong as the various mathinlines don't do an undef first).

Eg. in math.h:

#if defined __USE_ISOC99 && __GNUC_PREREQ(2,97)
# undef signbit
# ifdef __NO_LONG_DOUBLE_MATH
#  define signbit(x) \
     (sizeof (x) == sizeof (float) ? __builtin_signbitf (x) : __builtin_signbit (x))
# else
...

Wilco
  

On Mon, 27 Apr 2015, Wilco Dijkstra wrote:

> > Joseph Myers wrote:
> > On Wed, 15 Apr 2015, Wilco Dijkstra wrote:
> > 
> > > Is there a reason this couldn't be done by default in math.h similar to isgreater (unlike
> > > __builtin_isinf et al, GCC implements signbit efficiently and correctly).
> > 
> > It is, however, not type-generic in GCC
> > <https://gcc.gnu.org/bugzilla/show_bug.cgi?id=36757>, so you'd still need
> > to have a macro expansion checking sizeof.
> 
> Yes indeed. However the issue is how to deal with the targets that currently
> define inlines. Would it be reasonable just to leave them and add a signbit
> expansion for GCC >= 3.0 that directly uses the builtins?
> (if we define __signbit(f/l) to be __builtin_signbit(f/l), things may go 
> wrong as the various mathinlines don't do an undef first).

I'd think defining __signbit to __builtin_signbit, etc., would be 
appropriate (with conditionals in the mathinline.h files to disable their 
local definitions for GCC versions where __builtin_signbit works and is at 
least as good for that architecture).  It would also cover cases within 
glibc that directly use the __signbit etc. names (although arguably those 
should move to using the type-generic macros; likewise direct uses of e.g. 
__finite, __isnan).