Re: Floating-point constant folding in Emacs byte compiler

[Eli Zaretskii]

> From: Robert Pluim <address@hidden>
> Cc: address@hidden
> Date: Mon, 26 Mar 2018 17:57:01 +0200
> 
> >> >> I imagine that on a 32-bit version of emacs, (byte-compile (lambda (x)
> >> >> (* #x10000 #x10000))) similarly produces a constant-zero function,
> >> >
> >> > The result depends on whether Emacs was build --with-wide-int.  If it
> >> > was, the result is #[(x) "\300\207" [4294967296] 1], if it wasn't, I
> >> > get #[(x) "\300\207" [0] 1].
> >> 
> >> Hmm, would it be worthwhile to have Emacs signal overflow in such a
> >> situation (perhaps controlled by a configuration variable) so we could
> >> fix such issues?
> >
> > In which of these two cases do you see overflow?
> >
> 
> Does (* #x10000 #x10000) not overflow on 32-bit?

Only if Emacs was not built --with-wide-int.

> > I also don't think I understand the utility of signaling an overflow
> > error from the byte compiler that happened because it did constant
> > folding.  What else except bug reports could this yield?
> 
> I was proposing it as a debug option, not a generally enabled one. Of
> course if there's no overflow, it would be kind of useless :-)

If we want the byte code to be portable, the results of compiling
cannot depend on the machine on which the code was byte-compiled.