bug#41321: 27.0.91; Emacs aborts due to invalid pseudovector objects

[Pip Cet]

On Sat, May 30, 2020 at 4:31 PM Paul Eggert <eggert@cs.ucla.edu> wrote:
> On 5/29/20 2:01 PM, Pip Cet wrote:
> > (1) says:
> > It’s an invalid optimization, since pointers can address the
> > middle of Lisp_Object data.
> >
> > That may be true (we still haven't observed it),
>
> I observed it earlier, in code that iterated through a Lisp vector;

Sorry, I must have missed that.

> at the
> machine level the only pointer was into the middle of that vector. Addresses 
> of
> Lisp_Vector elements are not GCALIGNED on x86 and other platforms.

True.

> > 1. malloc() % GCALIGNMENT != 0. Never happens, as far as I can tell.
>
> Although that's true of all current Emacs porting targets as far as I know, 
> I'd
> rather not hardwire this into the code, as neither POSIX nor the C standard
> require it. This is why the comment refers to platforms where malloc() % 8 != > 0
> as "oddball hosts".

But we can't figure out what alignment malloc guarantees, on practical
hosts. To say we assume a malloc alignment of 8 is much better than to
say we assume one of alignof (max_align_t), which is false on many
systems.

> > 2. A Lisp object requires greater alignment than malloc() gives it.
> > IIRC, there was at least one RISC architecture whose specification
>
> We don't need anything that obscure. Just use __int128 on x86 with glibc 2.24.
> On that platform __int128's alignment is 16, malloc's is 8.

Sorry, but I think a type that is actually used by Emacs is less
obscure than __float128 (which I think you mean; __int128 doesn't
exist on x86), nevermind the question of whether the alignment of that
should have been 16, since it works just fine misaligned (except when
AC is set, but that's no longer x86-as-we-know-and-hate-it).

> > I'm not saying it's the best solution, but I would prefer simply
> > defining LISP_ALIGNMENT to be 8 to either patch.
>
> That's not correct for !USE_LSB_TAG, where LISP_ALIGNMENT is merely the native
> alignment (there's no need to align objects to 8 because the tags are at the
> high end).

How is it incorrect? Suboptimal, maybe, though there's a performance
improvement keeping things you access together in the same cache line.

There's no need to align anything (non-SIMD) to anything on x86
without AC set, it's just good for performance; and that performance
improvement applies whether or not Lisp_Objects are natively 64-bit or
2x32-bit.