bug in logarithm()

[Martin Koeppe]

Hi Jeff, hi Patrick,

On Wed, 17 Oct 2007, Jeff Johnston wrote:
> Patrick Mansfield wrote:
>> On Wed, Oct 17, 2007 at 12:28:47PM +0200, Martin Koeppe wrote:
>> 
>>> I'm using newlib on a i386 architecture, and now I wonder if it's possible 
>>> to configure libm in such a way that the fast asm versions of e.g. log() 
>>> are used. In my current newlib-libm.a there are the entry points for 
>>> _f_log() etc., but apparently log() doesn't make use of _f_log().
>>> 
>>> (How) Is it possible to configure newlib's libm to make use of _f_log 
>>> within log()?
>>> 
>>> Thank you very much
>>> 
>> 
>> They aren't compatible per comments in the code. You could call them
>> directly. AFAIUI there should be prototypes and proper names (without
>> underscore prefix, like #define f_log _f_log) for them in newlib, but I
>> don't see any.
>> 
>> -- Patrick Mansfield
>> 
> Actually, they are meant to be called when the user specifies -ffast-math on 
> the compilation.
>
> The definitions used to be part of libc/include/machine/fastmath.h with an 
> i386 check.  This got inadvertently left out when we initially made the 
> transition of newlib from internal to external repositories.
>
> Anyway, I have added libc/machine/i386/machine/fastmath.h based on the older 
> sources plus a few changes.  If you compile your code with the -ffast-math 
> option, <math.h> will bring in <machine/fastmath.h> and you will get access 
> to the few fast math routines available for x86, including log().

I use newlib's libm, and only the libm, not the whole libc+libm, on 
Interix, as replacement for the buggy standard libm there. So I don't 
want to use any different header files if at all possible.

I can see from the code that e.g. log() and _f_log() don't provide the 
same functionality, e.g. errno is set by the former, but not by the 
latter. My question/wishlist item was more that I would like to 
compile newlib with some option like -ffast-math, or better a 
configure option, just like --enable-newlib-hw-fp, so that log() 
checks the argument and sets errno, but then calls _f_log() to 
calculate the logarithm instead of using the Newton algorithm.

User programs then still could be further optimized by skipping errno 
processing and call _f_log() directly with -ffast-math if necessary, 
but even without that it would be reasonably fast.


BTW: Are there any snapshots available for easy building?

Martin