Re: Excessive use of `eassert`

[Alan Mackenzie <acm@muc.de>]

Hello, Paul.

On Sat, Jan 20, 2024 at 17:41:57 -0800, Paul Eggert wrote:
> On 2024-01-19 11:42, Alan Mackenzie wrote:

> > On Fri, Jan 19, 2024 at 17:02:06 +0200, Eli Zaretskii wrote:

> >> make_fixnum is a trivial bit-shuffling, whereas make_lisp_symbol is
> >> much trickier.  Perhaps especially so now that we have
> >> symbols-with-positions as well as bare symbols.

> > Not really.  Symbols with positions don't belong in the obarray.  If they
> > somehow get there, then that's a bug to be fixed.

> The problem here isn't calling make_lisp_symbol to make symbols with 
> positions. It's that Qnil expands to builtin_lisp_symbol (0) which calls 
> make_lisp_symbol (&lispsym[0]) which calls XSYMBOL, and XSYMBOL is 
> significantly slower now that we have symbols-with-positions, even when 
> it is applied to a bare symbol - and this is particularly true with 
> --enable-checking and the eassert Stefan mentioned.

> I looked into this a bit and installed the attached patches which should 
> speed things slightly even with a default build. The main goal was to 
> speed up debugging builds, though.

I don't think the first patch is correct.  With it applied, the code
no longer signals the error of a symbol with position being processed
when symbols_with_pos_enabled is false.  That's in the debug build, of
course.

The second patch seems OK.

> When I built with --enable-checking this seemed to help significantly on 
> Ubuntu 23.10 x86-64 with GCC 13.2 -O2 (at least looking at the machine 
> code; I didn't benchmark). This depends on compiler and platform so it'd 
> be helpful if Stefan could try it out on his machine and see whether it 
> helps his cases' performance.

The "significantly" here means the 0.4% speed up in the default build
that you mention in the first patch, does it?

-- 
Alan Mackenzie (Nuremberg, Germany).