On 10/20/23 00:34, waffl3x wrote:
Based on what you've said, I assume that OFFSET_REF handles static
member functions that are overloaded. But as I've said this seems to
contradict the comments I'm reading, so I'm not sure that I'm
understanding you correctly.
That's right. For instance,
struct A {
static void g();
static void g(int);
};
void (*p)(int) = &A::g; // cp_build_addr_expr_1 gets an OFFSET_REF
I think we need the OFFSET_REF for an explicit-object member function
because it expresses that the code satisfies the requirement "If the
operand names an explicit object member function, the operand shall be a
qualified-id."
I do agree here, but it does reinforce that OFFSET_REF is no longer
just for members represented by pointer to member type. So that might
be something to take into consideration.
An OFFSET_REF that isn't type_unknown_p, agreed.
It might simplify things to remove the optimization in build_offset_ref
so we get an OFFSET_REF even for a single static member function, and
add support for that to cp_build_addr_expr_1.
I don't think this should be necessary, the "right thing" should just
be done for explicit-object member functions. With all the stuff going
on here that I missed I'm starting to wonder how function overloads
ever worked at all in my patch. On the other hand though, this
optimization probably could be documented better, but I very well might
have missed it even if it were.
Hell, maybe it needs a greater redesign altogether, it seems strange to
me to bundle overload information in with a construct for a specific
expression. (Assuming that's whats happening of course, I still don't
fully understand it.) It's not like this has rules unique to it for how
overload resolution is decided, right? Initializing a param/variable of
pointer to function type with an overloaded function resolves that with
similar rules, I think? Maybe it is a little different now that I write
it out loud.
I wasn't going to finish my musings about that, but it made me realize
that it might not actually be correct for address of explicit-object
member functions to be wrapped by OFFSET_REF. I mean surely it's fine
because based on what you've said static member functions are also
wrapped by OFFSET_REF, so it's likely fully implemented, especially
considering things worked before. But now that there are 2 different
varieties of class members that the address of them can be taken, it
might make sense to split things up a bit? Then again, why were static
member functions ever handled the same way? Taking the address of other
static members isn't handled in the same way here is it?
Functions are different because of overloading; in general we can't
decide what an expression that names a function actually means until we
have enough context to decide which function, exactly. So we represent
the id-expression largely as lookup+syntax until overload resolution
turns it into a specific function. The type_unknown_p check earlier in
cp_build_addr_expr_1 is for that case.
Yeah this all makes sense, but that's why I was confused by the
following documentation from cp-tree.def.
```
/* An OFFSET_REF is used in two situations:
1. An expression of the form `A::m' where `A' is a class and `m' is
a non-static member. In this case, operand 0 will be a TYPE
(corresponding to `A') and operand 1 will be a FIELD_DECL,
BASELINK, or TEMPLATE_ID_EXPR (corresponding to `m').
The expression is a pointer-to-member if its address is taken,
but simply denotes a member of the object if its address is not
taken.
```
Yeah, I'll adjust that statement to be more conditional. Is this clearer?
1. An expression of the form `A::m' where `A' is a class and `m' is
a non-static member or an overload set. In this case, operand 0
will be a TYPE (corresponding to `A') and operand 1 will be a
FIELD_DECL, BASELINK, or TEMPLATE_ID_EXPR (corresponding to
`m').
The expression is a pointer-to-member if its address is taken
(and
if, after any overload resolution, 'm' does not designate a
static or explicit object member function). It simply denotes a
member of the object if its address is not taken.
An OFFSET_REF that isn't type_unknown_p, agreed.
But I suppose that's what this might have been referring to.
So is that the case then? OFFSET_REF might be for a regular address of
member expression unless it's type_unknown_p?
If it's type_unknown_p we don't know which member it is, so we don't
know whether taking its address gives a PMF or a regular pointer.
If it's not type_unknown_p, we know which member it is, and we currently
skip building it at all for static members, so taking its address always
gives a pointer to member. explicit object member functions will make
that assumption no longer valid.
An id-expression that names a single non-template function
(!really_overloaded_fn) is handled somewhat differently, as we don't
need to defer everything. But that means various special-case code.
Currently build_offset_ref special-cases &A::f for a single static
member function, but we can't use the same special case for single
explicit object member functions because we need to distinguish between
&A::f and &f somehow to check the requirement I quoted above. So it
seems to me we'll need to add support for single explicit object member
functions in the OFFSET_REF handling in cp_build_addr_expr_1. And I
thought if we're doing that, perhaps we want to move the single static
handling over there as well, but that's not necessary.
I'm done for today but it does seem like that special case is what is
causing my crash. I found that it only jumps into the section that it
crashes in when there are no overloads. I'm kinda close to fixing it
probably but I've spent too long on it for today, it's melting together.
Currently build_offset_ref special-cases &A::f for a single static
member function, but we can't use the same special case for single
explicit object member functions because we need to distinguish between
&A::f and &f somehow to check the requirement I quoted above.
I don't understand what this means exactly, under what circumstances
would &f find the member function. Oh, I guess while in the body of
it's class, I hadn't considered that. Is that what you're referring to?
Right:
struct A {
void g(this A&);
A() {
&A::g; // ok
&g; // same error as for an implicit object member function
}
};
Well either way, I'm going to pick back up here tomorrow and see if I
can finish figuring out exactly whats causing the problems. I'm pretty
certain it's the special case that causes it, I'm just not sure what
part exactly is causing the difference, and whether there aren't more
hiding issues. But for now I need to get some rest. Thanks for the
detailed responses, they helped a lot, this part of the code has been
particularly difficult to figure out.
You're welcome, thanks for your interest.
Jason
