Re: IEEE arithmetic handling

eggert@twinsun.com (Paul Eggert) writes:
> But that conflicts with IEEE Std 754-1985, section 5.6, which requires
> that converting a number from binary to decimal and back be the identity
> if the proper precision and rounding is used. The Fortran standard says
> -0.0 must be output as 0.0; this loses information.


This and other recent similar comments about compilers and IEEE reminded
me of this note from the back of a recent technical report here (How To
Read Floating Point Numbers Accurately, William D. Clinger, University of
Oregon CIS-TR-90-01, June 1990)


The IEEE standard explicitly states that, in high level
languages, the destination of an arithmetic operation
may be determined by the compiler, and hence may be
beyond the control of programmers. In other words, the
compiler -- not the programmer who uses the compiler --
is regarded as the client of the standard. Thus the
error bounds guaranteed by the IEEE standard may not be
relied upon by programmers who work in high level
languages.


Modulo arguments about whether FORTRAN is a high-level language :-), it
seems to me that this pretty much answers the question of whether FORTRAN
arithmetic is IEEE compliant. It's not, but it needn't be, and it may be
unrealistic to expect it to be, given the goals of HLL design.


Bart Massey
bart@cs.uoregon.edu
--

Return to the comp.compilers page.
Search the comp.compilers archives again.