This is the mail archive of the gcc@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: real.c on unicosmk



On Thursday, May 23, 2002, at 02:00 , Richard Henderson wrote:
> We no longer trust any assembler.  It is easier to emit all
> floating point values with their integer bit-wise representation.

Great! I've been having so many problems, especially on MIPS systems,
where there used to be floating point constants output with too many
digits by GCC. When properly converted to the nearest machine number,
it would have been correct, but the assembler only looks at the first
digits and truncates the extras causing a double rounding, which results
in the largest denormal positive number instead of the smallest normal 
one.
Of course this results in exceptions at run-time, and wrong results if
using the mode where denormals flush to zero....

I fully agree we always should output values in integer bit-wise
representation.

Thanks for fixing this!

   -Geert


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]