bug#31118: 27.0.50; Can't load/compile websocket in 32bit master

[Stefan Monnier]

> Do you mean 'read' should call 'display-warning' for now? I suppose we could
> make read-integer-overflow-as-float be a three-state variable: either (1)
> signal an error, or (2) call display-warning and yield a float, or (3)
> silently yield a float, and have (2) be the default for now. However,
> I worry that (2) might lead to further problems, e.g., display-warning is
> Lisp code that might in turn call 'read' and loop recursively.

We already solved this problem for old-style backquotes: don't emit the
warning directly, instead set a variable and let the caller check the
value of the variable to emit a warning if needed.

> The code we're talking about had nonportable code like (eq desktop
> #xffffffff) that doesn't work on platforms with 30-bit fixnums anyway.

I don't disagree with your change.  But the problems encountered with
code in GNU ELPA makes me think that we want to do it in two steps to
reduce the backward compatibility pain.

> Admittedly --with-wide-hit is a ~30% CPU performance hit on my circa-2010
> AMD desktop.

4 out of 5 of my machines predate your AMD desktop, and I use a 32bit
build on them.

> If defaulting to --with-wide-int is too drastic, I hope that the
> already-existing read-integer-overflow-as-float flag is enough
> backstop for people who want to run nonportable code on platforms with
> 30-bit fixnums.

If we want to signal an error by default, then I think it's important
for this error to give clear steps to set read-integer-overflow-as-float
(and in a message that's not too prone to disappear before the user
gets to see and act on it).


        Stefan