Re: [PATCH 5/5] cbtree.h: define cb_init() in terms of CBTREE_INIT

[Jeff King <peff@peff.net>]

On Mon, Sep 27, 2021 at 01:02:35PM +0200, Ævar Arnfjörð Bjarmason wrote:

> >>  static inline void cb_init(struct cb_tree *t)
> >>  {
> >> -	t->root = NULL;
> >> +	struct cb_tree blank = CBTREE_INIT;
> >
> > This could be
> >
> > 	static const struct cb_tree blank = CBTREE_INIT;
> 
> *nod*...
> [...]
> ...but to both this & the above my reply in the side-thread at
> https://lore.kernel.org/git/87h7e61duk.fsf@evledraar.gmail.com/
> applies. I.e. this is just following a pattern I got from Jeff King &
> used in bd4232fac33 (Merge branch 'ab/struct-init', 2021-07-16).

I'm not sure how a compiler would react to the "static const" thing. I
tested the compiler output for the "auto" struct case you've written
here, and at least gcc and clang are smart enough to just initialize the
pointed-to struct directly, with no extra copy.

For a "static const" I'm not sure if they'd end up with the same code,
or if they'd allocate a struct in the data segment and just memcpy()
into place. A non-const static would perhaps push it in the direction of the
data/memcpy thing, though the compiler should be well aware that the
struct is never changed nor aliased, and thus we're always writing the
INIT values.

I suspect the performance is not that different either way (the big
thing to avoid is initializing an auto struct on the fly and then
copying from it, but this is a pretty easy optimization for compilers to
get right).

> >> +	memcpy(t, &blank, sizeof(*t));
> >
> > Is
> > 	*t = blank;
> >
> > not a thing in C?

It would be fine to use struct assignment here, and should be equivalent
in most compilers. They know about memcpy() and will inline it as
appropriate.

I think some C programmers tend to prefer memcpy() just because that's
how they think. It also wasn't legal in old K&R compilers, but as far as
I know was in C89.

You have to take care with assignment of flex-structs, of course, but
you also have to do so with memcpy(), too. :)

> FWIW with "const" in general I don't use it as much as I'd personally
> prefer, see e.g. [1] for one recent discussion, but maybe there wouldn't
> be any push-back in this case...

This isn't a parameter, so I don't think that discussion applies. _If_
you are going to make it a static, I think a const makes sense here (but
probably does nothing beyond signaling your intention, because the
compiler can see that it is never modified), but I wouldn't bother with
either.

-Peff