bug#65051: internal_equal manipulates symbols with position without checking symbols-with-pos-enabled.

[Alan Mackenzie]

Hello, Mattias.

On Mon, Aug 07, 2023 at 10:58:41 +0200, Mattias Engdegård wrote:
> 6 aug. 2023 kl. 17.02 skrev Alan Mackenzie <acm@muc.de>:

> >>> together with an unsigned integer called the @dfn{position}.  These
> >>> -objects are intended for use by the byte compiler, which records in
> >>> -them the position of each symbol occurrence and uses those positions
> >>> -in warning and error messages.
> >>> +objects are stored internally much like vectors

> >> Not sure why we want to say how they are stored here. They can be
> >> stored in bubble memory for all the user cares.

> > The point is, they are _not_ stored in the obarray.  Eli specifically
> > asked me to clarify this point, yesterday.

> Oh that part is perfectly fine (thank you), we just don't need to say
> that the sympos objects are stored "like vectors" -- that just confuses
> the reader.

Why not?  It's true, and I doubt it will cause confusion.  I think we
need to say something positive in that place (since we're following it
with a negative).  Perhaps you could suggest an alternative.

> >>> +When @code{symbols-with-pos-enabled} is @code{nil}, any symbols with
> >>> +position continue to exist, but do not behave as symbols, or have the
> >>> +other useful properties outlined in the previous paragraph.  @code{eq}
> >>> +returns @code{t} when given identical arguments, and @code{equal}
> >>> +returns @code{t} when given arguments with @code{equal} components.

> >> Since the components are bare symbols and fixnums, equality and
> >> identity for them are equivalent, right?

> > No.  If there are two distinct SWPs with the same bare symbol and the
> > same position, they should be equal, but not eq.  But the real point is
> > to contrast how equal and eq work when symbols-with-pos-enabled is nil
> > with when it is non-nil.

> I meant that the components of equal sympos objects aren't merely equal
> but identical. (This is a very minor quibble; you can keep the text if
> you like.)

The current proposed text has a more subtle intention.  It says that eq
and equal behave just like they always have done for everything when
symbols-with-pos-enabled is nil.

> >> OK. This reduces the number of branches in the hot path for ordinary
> >> (non-sympos) code by one while adding one to sym-pos code, and that
> >> should be a fair trade-off. The new branch should be well-predicted but
> >> is still consuming resources.

> > I did some simple timings on the old and new code, and the new code is
> > not slower.

> This is not easy to measure and details matter, but as I said -- there
> is no reason to believe that your changes should be a regression in the
> important measure, rather the opposite.

Agreed.

> >>> + if (SYMBOL_WITH_POS_P(o1)) /* symbols_with_pos_enabled is false. */
> >>> +   return (internal_equal (XSYMBOL_WITH_POS (o1)->sym,
> >>> +                           XSYMBOL_WITH_POS (o2)->sym,
> >>> +                           equal_kind, depth + 1, ht)
> >>> +           && internal_equal (XSYMBOL_WITH_POS (o1)->pos,
> >>> +                              XSYMBOL_WITH_POS (o2)->pos,
> >>> +                              equal_kind, depth + 1, ht));

> >> Why recurse here if the components are a bare symbol and a fixnum,
> >> respectively?

> > Maybe in case they might somehow be something else?

> No, we must be able to assume that internal invariants hold when we
> offer no way for them to be violated. Let's just change the calls to
> BASE_EQ and be done with it.

OK, I think you're right, here, I'll change that.

> >> However we should make an effort to prevent the compiler from
> >> optimising (eq X X) -> t etc, which it is completely entitled to doing,
> >> ....

> > Why?  (eq X X) is t in all circumstances, whether X is a symbol, a cons
> > structure, or anything else.  What am I missing, here?

> If the compiler transforms (eq foo1 foo1) into t then the test won't
> actually exercise the implementation of `eq`.

Ah!  You're talking about the tests.  OK.  In my tests, I timed (equal a
b) where a and b were variables which were either equal or not.

> >> .... and also test both the interpreted and compiled version of `eq`
> >> and `equal`.

> > They're the same code in both cases.  I'm missing something here, too, I
> > think.

> Byte-code doesn't call Feq, it uses its own implementation. They should
> work identically but as we are checking edge cases here we'd better be
> sure about that.

> >> The test bytecomp--eq-symbols-with-pos-enabled already does most of
> >> this for a different reason. Perhaps it can be extended to cover
> >> `equal` as well?

> > I don't have such a test in my repository anywhere.  Are you sure you
> > wrote it right?

> It was added in 44d7fd3805.

OK.  That commit is recent, then?  If so, I'll see it soon.

-- 
Alan Mackenzie (Nuremberg, Germany).