https://gcc.gnu.org/bugzilla/show_bug.cgi?id=55814
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Component|middle-end |fortran
--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
So the testcase with the loop boils down to optimizing
_Bool foo (int *p, int n, _Bool q)
{
_Bool any = false;
for (i = 0; i < n; ++i)
any |= p[i];
return any || q;
}
where we want swapping of the operands (and its computations) of || (or &&).
Nothing in GCC does this kind of transform and I don't see where it would
fit best. I think that
logical function bar(a,b,c)
logical, intent(in) :: a, b
logical, intent(in), dimension(3) :: c
bar = a .and. b .and. any(c)
end
is translated in a bad way by the fortran FE, not honoring left-to-right
evaluation. So "optimized source" is pessimized here. Even when writing
logical function bar(a,b,c)
logical, intent(in) :: a, b
logical, intent(in), dimension(32) :: c
bar = any (c) .and. (a .and. b)
end
I get the same -- any () is evaluated first and only then the scalar variant.
The original C testcase shows we eventually miss a scalar-to-CFG pass on GIMPLE
before RTL expansion so we can run sinking on that.
That said, the any() thing is a frontend issue IMHO.
