Re: [PATCH] arm: Silence gcc warnings about arch ABI drift

[Kent Overstreet]

On Mon, Feb 19, 2024 at 09:52:08AM +0000, Russell King (Oracle) wrote:
> On Mon, Feb 19, 2024 at 04:40:00AM -0500, Kent Overstreet wrote:
> > On Mon, Feb 19, 2024 at 09:26:51AM +0000, Russell King (Oracle) wrote:
> > > On Mon, Feb 19, 2024 at 07:21:11AM +0100, Arnd Bergmann wrote:
> > > > I think these should be addressed in bcachefs instead.
> > > > While it's not the fault of bcachefs that the calling
> > > > convention changed between gcc versions, have a look at
> > > > the actual structure layout:
> > > > 
> > > > struct bch_val {
> > > >         __u64           __nothing[0];
> > > > };
> > > > struct bpos {
> > > >         /*
> > > >          * Word order matches machine byte order - btree code treats a bpos as a
> > > >          * single large integer, for search/comparison purposes
> > > >          *
> > > >          * Note that wherever a bpos is embedded in another on disk data
> > > >          * structure, it has to be byte swabbed when reading in metadata that
> > > >          * wasn't written in native endian order:
> > > >          */
> > > > #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
> > > >         __u32           snapshot;
> > > >         __u64           offset;
> > > >         __u64           inode;
> > > > #elif __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
> > > >         __u64           inode;
> > > >         __u64           offset;         /* Points to end of extent - sectors */
> > > >         __u32           snapshot;
> > > > #else
> > > > #error edit for your odd byteorder.
> > > > #endif
> > > > } __packed
> > > > struct bch_backpointer {
> > > >         struct bch_val          v;
> > > >         __u8                    btree_id;
> > > >         __u8                    level;
> > > >         __u8                    data_type;
> > > >         __u64                   bucket_offset:40;
> > > >         __u32                   bucket_len;
> > > >         struct bpos             pos;
> > > > } __packed __aligned(8);
> > > > 
> > > > This is not something that should ever be passed by value
> > > > into a function.
> > > 
> > > +1 - bcachefs definitely needs fixing. Passing all that as an argument
> > > not only means that it has to be read into registers, but also when
> > > accessing members, it has to be extracted from those registers as well.
> > > 
> > > Passing that by argument is utterly insane.
> > 
> > If the compiler people can't figure out a vaguely efficient way to pass
> > a small struct by value, that's their problem - from the way you
> > describe it, I have to wonder at what insanity is going on there.
> 
> It sounds like you have a personal cruisade here.
> 
> Passing structures on through function arguments is never efficient.
> The entire thing _has_ to be loaded from memory at function call and
> either placed onto the stack (creating an effective memcpy() on every
> function call) or into registers. Fundamentally. It's not something
> compiler people can mess around with.
> 
> Sorry but it's bcachefs that's the problem here, and well done to the
> compiler people for pointing out stupid code.

Eh? Passing via stack copy is normal and expected; you were talking
about something else.

I'm not always trying to write code that will generate the fastest
assembly possible; there aro other considerations. As long a the
compiler is doing something /reasonable/, the code is fine.