Date: Thu, 6 Mar 2003 22:34:15 -0800 From: David Schultz <das@FreeBSD.ORG> To: standards@FreeBSD.ORG Cc: obrien@FreeBSD.ORG Subject: gdtoa import this weekend, future directions Message-ID: <20030307063415.GA60214@HAL9000.homeunix.com> In-Reply-To: <20030221085508.GA55786@HAL9000.homeunix.com> References: <20030221085508.GA55786@HAL9000.homeunix.com>
next in thread | previous in thread | raw e-mail | index | archive | help
This took me longer than expected due to obscure bugs in the
vendor source, but I've finally pulled everything together. To
recap, gdtoa is a library for string->float and float->string
conversion. Our libc presently uses its predecessor from 3BSD,
which lacks support for hexadecimal floating point constants,
floats, and long doubles.
I'd like to do the import this weekend. I've tested extensively
on i386, and I've verified buildworlds and basic functionality on
ia64 and sparc64. I do not plan to immediately integrate long
double support into printf because there are multiple outstanding
bugs in our printf implementation's floating point support, and I
haven't decided whether to repair or replace yet. As it is, my
test cases are crashing in printf on ia64 without any of my
changes. The initial import will only add strtof() and strtold()
and the associated documentation.
As per Garret's suggestion, I would like to soon add some
interfaces that are not standardized, but perhaps not all of the
features gdtoa provides. (In particular, strtor{f,d,ld} can be
emulated by setting the rounding direction before calling the
standard functions, so maybe we don't need to bother with those
for now.) I would like to add:
int strtofi(const char *nptr, char **endptr, float *lo, float *hi);
int strtodi(const char *nptr, char **endptr, double *lo, double *hi);
int strtoldi(const char *nptr, char **endptr, long double *lo,
long double *hi);
These routines are like strtod(), but they efficiently compute two
floating point numbers that tightly bound the actual value of the
parsed string. The vendor implementation returns an integer
constant that classifies the number as zero, normal, NaN, etc.
Since we already have fpclassify() and a separate set of manifest
constants for that, I suggest defining the return value of our
version to be zero if the converted value can be represented
exactly (in which case 'hi' is also zero) and nonzero otherwise.
(I have chosen different function names to reflect that.)
Alternatively, we could just add more manifest constants and use
the vendor's version. Comments and suggestions are appreciated.
The code is in http://www.FreeBSD.ORG/~das/gdtoa/ . The files are
described below:
gdtoa.tgz latest vendor source; will appear on the
netlib mirrors within a few days
gdtoa.diff patches against vendor source
libc_gdtoa.diff patches against the rest of the tree:
do a 'mkdir lib/libc/{gdtoa,powerpc/stdlib}' first
To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-standards" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20030307063415.GA60214>