[Cfp-interest 1952] Re: inexact exception

[Jim Thomas]

Paul,

F.10 has two binding tables. The requirement for full support of the “inexact” floating-point exception is intended to apply only to the functions in the first binding table, not to the IEC 60559-recommended mathematical functions like hypot in the second table that are not required to be correctly rounded. Clarification is needed. Good catch.

Jim Thomas

> On Apr 7, 2021, at 8:12 AM, Paul Zimmermann <Paul.Zimmermann at inria.fr> wrote:
> 
>      Hello,
> 
> question: should the inexact exception be correctly raised (or not) for
> mathematical functions?
> 
> In http://www.open-std.org/jtc1/sc22/wg14/www/docs/n2596.pdf, section F.10,
> note 10 says:
> 
> "Whether or when library functions not bound to operations in IEC 60559 raise the "inexact" floating-point exception is unspecified, unless stated otherwise."
> 
> Since hypot is described in F.10.4.4, I understand it is bound to operations
> in IEC 60559, thus it should correctly raise inexact.
> 
> $ cat hypot.c
> #include <stdio.h>
> #include <fenv.h>
> #include <math.h>
> 
> int main()
> {
> double x, y, z;
> x = 1235216565974040.0;
> y = x + 1.0;
> printf ("x=%.6f y=%.6f\n", x, y);
> printf ("inexact: %d\n", fetestexcept (FE_INEXACT));
> z = hypot (x, y);
> printf ("z=%.6f\n", z);
> printf ("inexact: %d\n", fetestexcept (FE_INEXACT));
> }
> 
> $ gcc -fno-builtin hypot.c -lm
> $ ./a.out
> x=1235216565974040.000000 y=1235216565974041.000000
> inexact: 0
> z=1746860020068409.000000
> inexact: 32
> 
> x is a Pythagorean number, such that x^2+(x+1)^2 is an exact square,
> thus we should expect that the inexact flag remains zero.
> 
> Is my reasoning correct?
> 
> Best regards,
> Paul
> _______________________________________________
> Cfp-interest mailing list
> Cfp-interest at oakapple.net
> http://mailman.oakapple.net/mailman/listinfo/cfp-interest