[Bug fortran/52473] CSHIFT slow - inline it?

[tkoenig at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=52473

--- Comment #8 from Thomas Koenig <tkoenig at gcc dot gnu.org> ---
(In reply to Jerry DeLisle from comment #7)
> (In reply to Thomas Koenig from comment #6)
> > Created attachment 41508 [details]
> > What an unrolled cshift could look like
> > 
> > This is what an unrolled version of cshift could look like,
> > for a simple one-dimensional case.
> > 
> > If the shifts are constants, all the if statements are
> > optimized away at compile-time, so it is quite efficient.
> 
> How often or how important is this feature? Would it improve any known
> benchmark or real world application in a significant manner?

There's protein.f90 - see PR 52934.

Calculating a numerical derivative for a periodic function as

diff = (cshift(a,1) - cshift(a,-1))/(2*h)

seems natural, or the second derivative as

d2 = (cshift(a,1) - 2*a + cshift(a,-1))/(h**2)

Of course, to be really useful, this would also have to be
extended to eoshift...