2004-11-09 Paul Eggert <eggert@cs.ucla.edu.
[BZ #535]
* time/difftime.c: Fix a double-rounding bug on hosts with
64-bit time_t and long double being IEEE double. Also, port
to more valid C99 hosts, even those that have padding bits.
Don't include <values.h> since it is marked as an obsolescent
interface. Include <limits.h>, <float.h>, and <stdint.h> instead.
(TYPE_BITS, TYPE_FLOATING, TYPE_SIGNED): New macros.
(subtract): New static function, that works correctly without
double-rounding, even on hosts with 64-bit time_t. Also cater
to hosts with padding bits.
(__difftime): Use it. Use DBL_MANT_DIG and LDBL_MANT_DIG to
determine whether floating types are wide enough: the old
test (which used sizeof) could in theory report the wrong results
on hosts with padding bits in floating-point values.
29311370cd