[Bug fortran/40643] maxloc/minloc: Wrong result for NaN at position 1

["jakub at gcc dot gnu dot org" <gcc-bugzilla@gcc.gnu.org>]

------- Comment #2 from jakub at gcc dot gnu dot org  2009-07-10 16:31 -------
Created an attachment (id=18174)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=18174&action=view)
gcc45-pr40643.patch

Inline minmaxloc patch.
Is this the behavior we want?
If NaNs aren't supported and array size is known, no mask, then inline
maxloc/minloc should be quite a bit faster (condition should be just val <
limit
instead of val < limit || (pos == 0 && val == limit)).  On the other side, if
array size isn't known at compile time (or non-scalar mask is used) and NaNs
must be supported, the loop is more expensive.

I haven't touched minmaxval inline expansion yet, there I'm afraid for floating
point without -ffast-math the loop will have to be quite a bit slower than what
we do ATM, as we'd need to return -inf for maxval on array full of -infs (or
full of nans with at least one -inf), or nan for array full of nans, but for
empty array or completely masked away array -huge.

The patch also disables the inline expansion for -O0, I think at -O0 we should
aim at generating smaller and more easily debuggable code.  Probably we
shouldn't be expanding this inline for optimize_size either.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40643