Re: [Patch, Fortran] PR 85599: warn about short-circuiting of logical expressions for non-pure functions

[Janus Weil <janus@gcc.gnu.org>]

> 2018-07-16 21:50 GMT+02:00 Thomas Koenig <tkoenig@netcologne.de>:
>> What I would suggest is to enable -Wfunction-eliminiation with
>> -Wextra and also use that for your new warning.
>
> Thanks for the comments. Makes sense. Updated patch attached.


Huh, after actually trying -Wfunction-elimination, I'm not so sure any
more if it really makes sense to throw the new warnings in there,
mainly for two reasons:

a) -Wfunction-elimination is slightly different, in that it warns
about removing *any* kind of function, not just impure ones. This
makes it pretty verbose, and I'm not sure why one would need to know
if a pure function call is removed.
b) It gives warnings on places that I don't understand. Simple example:

subroutine sub(r)
   implicit none
   real, intent(in) :: r
   if ((abs(r) > 1e6) .or. (abs(r) < 1e-6)) then
      print *, "rrr"
   end if
end

Compiling this with "gfortran-8 -O1 -Wfunction-elimination" gives me:

    if ((abs(r) > 1e6) .or. (abs(r) < 1e-6)) then
        1
Warning: Removing call to function ‘abs’ at (1) [-Wfunction-elimination]


Why can I remove the call to ABS there? If I read the dump correctly,
then the two calls to ABS are optimized into a single one, which is
used for both comparisons via a temporary. Is that what the warning is
trying to tell me? And if yes, why should I care (if the function is
pure)? The middle-end certainly does tons of optimizations that
rearrange various expressions, without warning me about it ...

In other words: Does it make sense to tone down
-Wfunction-elimination, by only warning about impure functions? (We
certainly have the diagnostic capabilites for this.) If not, I'd
rather have a separate flag for the new warnings.
-Wfunction-elimination is just too noisy for my taste in its current
form.

Cheers,
Janus