Andrew Gierth <andrew@tao11.riddles.org.uk> writes:
> "Tom" == Tom Lane <tgl@sss.pgh.pa.us> writes:
> Tom> I'm inclined to think that we need to make ecpglib.h's
> Tom> bool-related definitions exactly match c.h,
> I'm wondering whether we should actually go the opposite way and say
> that c.h's "bool" definition should be backend only, and that in
> frontend code we should define a PG_bool type or something of that ilk
> for when we want "PG's 1-byte bool" and otherwise let the platform
> define "bool" however it wants.
> And we certainly shouldn't be defining "bool" in something that's going
> to be included in the user's code the way that ecpglib.h is.
The trouble here is the hazard of creating an ABI break, if we modify
ecpglib.h in a way that causes its "bool" references to be interpreted
differently than they were before. I don't think we want that (although
I suspect we have inadvertently caused ABI breaks already on platforms
where this matters).
In practice, since v11 on every modern platform, the exported ecpglib
functions have supposed that "bool" is _Bool, because they were compiled
in files that included c.h before ecpglib.h. I assert furthermore that
clients might well have included <stdbool.h> before ecpglib.h and thereby
been fully compatible with that. If we start having ecpglib.h include
<stdbool.h> itself, we'll just be eliminating a minor header inclusion
order hazard. It's also rather hard to argue that including <stdbool.h>
automatically is really likely to break anything that was including
ecpglib.h already, since that file was already usurping those symbols.
Except on platforms where sizeof(_Bool) isn't 1, but things are already
pretty darn broken there.
I think it's possible to construct a counterexample that will fail
for *anything* we can do here. I'm not inclined to uglify things like
mad to reduce the problem space from 0.1% to 0.01% of use-cases, or
whatever the numbers would be in practice.
regards, tom lane
