Re: buffer.c/buffer.h: How to add new buffer-local variables?

["Daniel Colascione" <dancol@dancol.org>]

>> From: Andreas Schwab <schwab@linux-m68k.org>
>> Date: Sun, 31 Mar 2019 13:41:51 +0200
>> Cc: Keith David Bershatsky <esq@lawlist.com>, emacs-devel@gnu.org
>>
>> On Mär 31 2019, Eli Zaretskii <eliz@gnu.org> wrote:
>>
>> > So you are saying that the 1024 value is arbitrary and should be
>> enlarged when more local vars are added?
>>
>> No.
>
> Daniel, could you please help me understand why the value 1024 was
> used here:
>
>   static dump_off
>   field_relpos (const void *in_start, const void *in_field)
>   {
>     ptrdiff_t in_start_val = (ptrdiff_t) in_start;
>     ptrdiff_t in_field_val = (ptrdiff_t) in_field;
>     eassert (in_start_val <= in_field_val);
>     ptrdiff_t relpos = in_field_val - in_start_val;
>     eassert (relpos < 1024); /* Sanity check.  */  <<<<<<<<<<<<
>     return (dump_off) relpos;
>   }
>
> The comment says "sanity check", but I would like to understand what
> kind of sanity is being checked here, and what should be done when
> some structure we dump becomes larger than this value.  E.g., is there
> some other limit that requires that offsets of dumped fields never
> exceed 1024 here?  I'd like to document in comments what to do when
> the assertion is violated.

Indeed. That comment could have been a lot better. The general idea here
is this:

When we enter field_relpos, we're in the middle of some code that's
dumping some data structure field-by-field. The object that we're dumping
begins at in_start; and in_field is an interior pointer into that object.
We can't actually check that the two pointers refer to the same object: C
doesn't give us that level of introspection. But if the two pointers point
to addresses that differ by a lot, then the two pointers probably don't
refer to the same object, and in this case, we can fail an assertion. 1024
is probably too conservative here. We probably want to greatly increase
this number (say, to 32k) and also to give it a named constantly, maybe
something like PDUMPER_MAXIMUM_STRUCT_SIZE.

Note that this limit doesn't apply to big variable-length structures like
vectors: we dump these element-by-element instead of treating the whole
thing as one big "object" with a large and variable number of fields.