Re: Excessive use of `eassert`

[Paul Eggert]

On 2024-01-22 05:20, Stefan Monnier wrote:
> Compared to the patch I proposed this has the downside that it
> duplicates the logic between `lisp_h_builtin_lisp_symbol` and
> `make_lisp_symbol`
Yes, when --enable-checking is used the tradeoff is: would we rather 
omit all runtime checking in make_lisp_symbol (the patch you proposed), 
or omit it only for builtin symbols (the patch I installed)?

For builtin symbols like Qnil and Qt the runtime checking is not all 
that useful - if these symbols' data items are improperly aligned Emacs 
will crash early anyway. For non-builtin symbols the runtime checking is 
arguably useful, to catch (presumably rare) alignment bugs in the memory 
allocator.

If you'd rather have the simpler solution that doesn't catch alignment 
bugs in non-builtin symbols, I could prepare a patch along those lines. 
Any such alignment bugs are pretty unlikely, after all.


>  (worse: the logic is actually not the same, so it's
> > not obvious that it ends up returning the same thing).

I used that logic to convince gcc to optimize builtin symbols like Qt 
and Qnil into integer constants at the machine level even when -O0 is 
used. But if nobody needs that sort of performance with -O0, we could go 
back to the old way of doing things. (It's obvious to *me* that the two 
formulations are equivalent but hey! only four people in the world 
understand lisp.h and nobody knows which four people they are....)