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]

Re: Reorganize -ffast-math code.


In article <200103081946.UAA01461@ulmo> you write:
>Don't we have the same problem with IA64? I read the beast spec a
>while ago and seem to remember that some extended precision is still
>there (82 bits or something, can't find where I put the manual...).

There is no such problem with IA-64.  IA-64 has all of the necessary
single/double/double-extended FP instructions that x86 is lacking, and hence
does not suffer from this FP rounding/truncation/excess-precision problem.

IA-64 does have 82-bit FP registers, but that is a hardware implementation
detail invisible to people using high level languages like C and Fortran.
One extra bit is for encoding NatVals (Not a Thing Values).  These are
non-numeric values obtained by failed speculative loads, and indicate that
the value must be re-read from memory.  One extra bit is for correct rounding
of divide/sqrt sequences that use reciprocal approximation plus Newton-Raphson
iteration.

Jim


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