https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96811
kargl at gcc dot gnu.org changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |kargl at gcc dot gnu.org
--- Comment #2 from kargl at gcc dot gnu.org ---
(In reply to Tobias Burnus from comment #0)
> Fabio Azevedo has some observations at
> https://gcc.gnu.org/pipermail/fortran/2020-August/054931.html
>
> In particular:
>
> The only difference in behaviour I see is that when the exponent is
> negative, libgcc2 calculates 1/(x**(-n)) while libgfortran calculates
> (1/x)**(-n). The first version being more accurate.
> [...]
> On the top of that x**n with x real for negative exponents appears to yield
> results more compatible with 1/(x**(-n)) than (1/x)**(-n).
>
> libgcc/libgcc2.c has:
>
> NAME (TYPE x, int m)
> {
> unsigned int n = m < 0 ? -m : m;
> TYPE y = n % 2 ? x : 1;
> while (n >>= 1)
> {
> x = x * x;
> if (n % 2)
> y = y * x;
> }
> return m < 0 ? 1/y : y; // <<< division done at the end.
> }
>
>
> While libgfortran/m4/pow.m4 has (for noninteger x):
>
For real x, gfortran does not generate a call to libgfortran function.
% cat a.f90
function foo(x,n)
real foo, x
integer n
foo = x**n
end
% cat a.f90.004.original
foo (real(kind=4) & restrict x, integer(kind=4) & restrict n)
{
real(kind=4) __result_foo;
{
real(kind=4) D.3806;
D.3806 = *x;
__result_foo = __builtin_powif (D.3806, *n);
}
return __result_foo;
}
For complex or integer x, gfortran does generate the call.
With the obvious change to a.f90,
% cat a.f90.004t.original
foo (complex(kind=4) & restrict x, integer(kind=4) & restrict n)
{
complex(kind=4) __result_foo;
{
complex(kind=4) D.3806;
D.3806 = *x;
__result_foo = _gfortran_pow_c4_i4 (D.3806, *n);
}
return __result_foo;
}
Richard's observation about rounding error is probably sufficient
reason to prefer libgcc2.c.
