wide function pointer type

[Martin Uecker]

Am Sonntag, den 10.10.2021, 11:57 -0700 schrieb Daniel Colascione:
> 
> On October 10, 2021 11:47:18 Martin Uecker <ma.uecker@gmail.com> wrote:
> 
> > Am Sonntag, den 10.10.2021, 11:17 -0700 schrieb Daniel Colascione:
> > > On October 10, 2021 11:05:07 Martin Uecker <ma.uecker@gmail.com> wrote:
> > > 
> > > > Am Sonntag, den 10.10.2021, 10:49 -0700 schrieb Daniel Colascione:
> > > > > On October 10, 2021 10:44:32 Martin Uecker via Libffi-discuss
> > > > > <libffi-discuss@sourceware.org> wrote:
> > > > > 
> > > > > > Am Sonntag, den 10.10.2021, 10:01 -0700 schrieb Kaz Kylheku (libffi):
> > > > > > > On 2021-10-10 04:32, Martin Uecker via Libffi-discuss wrote:
> > > > > > > > Hi all,
> > > > > > > > 
> > > > > > > > I will propose a wide function pointer type (actually
> > > > > > > > a wide function type) to WG14 for C23 as a common
> > > > > > > > type for callbacks, closures, which now require an
> > > > > > > > additional void pointer argument in C APIs. This
> > > > > > > > is intended to be compatible with ABIs with now
> > > > > > > > use a static chain register.
> > > > > > > 
> > > > > > > Opposed. There is nothing wrong with separate arguments
> > > > > > > for function pointer and context.
> > > > > > 
> > > > > > Noted.  Your argument sbelow all boil down to the point
> > > > > > that there are cases where it might not be the ideal
> > > > > > choice. But nobody forces anyone to use it.
> > > > > "I like it" is not by itself a rationale for including something in a core
> > > > > language standard. This is not something that should be in C. Why,
> > > > > precisely, can't you just define a struct with two fields and make that
> > > > > your closure type?
> > > > 
> > > > Sure. There are at least five reasons:
> > > > 
> > > > - one now does has to do it manually which is
> > > > inconvenient because it has to be done for each type,
> > > > although this a common pattern which is trivial
> > > > to automate.
> > > > 
> > > > - APIs are often unsafe because they have to use
> > > > void pointer to be generic. When the void pointer
> > > > can be packaged into the wide pointer this problem
> > > > goes away.
> > > > 
> > > > - APIs are often different for no good reason.
> > > > Having a standard approach would make this more
> > > > canonical, hence simpler and less error prone,
> > > > and also more compatible - requiring less adapter
> > > > code.
> > > > 
> > > > - There are APIs where a pointer to a
> > > > function is expected but there is no data pointer.
> > > > Then other languages have a problem interfacing
> > > > with such APIs.   (or why does libffi generate
> > > > trampolines?)
> > > > 
> > > > - Finally, C might get lambda expressions
> > > > which are far more useful with such a pointer type.
> > > > (for nested function you could avoid executable
> > > > stack).
> > > > 
> > > > 
> > > > All five are good reasons to put a standard
> > > > approach in the core language in my opinion.
> > > > 
> > > > 
> > > > But I am actually not interested in this discussion
> > > > and do not have time for it.
> > > > 
> > > > I am more interested in constructive technical
> > > > feedback.
> > > 
> > > Your proposal doesn't actually solve any of these problems though, and as
> > > we agree, users can already trivially make their own closure types.
> > 
> > I wonder why you think it does not address
> > these problems?
> 
> Because there's no difference between your proposal and a simple user 
> struct that does the same thing.

I see at least the following differences:

- Simple user code does not know how to load the static chain
pointer into its ABI-specified register in a portable way.

- A simple user struct works exactly for one function type,
so such a struct has to be created for all APIs separately.

- One can not easily convert a regular pointer to a
simple user struct (this would require at least inserting
a branch at it each call site) 

- A simple user struct can be defined in different ways,
so there is no standard.

- If a library defines a simple struct and another library
another simple struct, then these two structs are not
compatible even if defined identically.

- Other languages or interface generating tools will not
know about the simple user struct, but will certainly
know about a standardized type.


>  The language could get lambdas one day 
> (though I doubt it), but if it does, each lambda will likely have a unique 
> and anonymous type (like in C++) and won't fit in your wide function 
> pointer struct anyway.

I think the lambda proposal has a good chance, although
I am not sure it will make it already into C23.
There would be a conversion rule (as mentioned in
the proposal).

C++ has std::function for a similar purpose.

> > > What *would* be interesting is a new standard library facility for making
> > > libffi-style closures and a standard struct type for packaging the things.
> > > *That's* something that could be justified for inclusion in a standard
> > > library, since it's not something that programs can make on their own
> > > without the kind of platform specific glue that libffi provides.
> > 
> > I also would love to have this, but see below...
> > 
> > > Basically,
> > > the interface would look something like this:
> > > 
> > > typedef opaque_mumble stdclosure_t;
> > > 
> > > // Returns a callable function pointer or NULL on error with errno set
> > > void* stdclosure_init(stdclosure_t* closure, void* fnptr, void* data);
> > 
> > A void* might not be big enough for a callable function
> > pointer an all architectures supported by C so this
> > needs to return something like
> > 
> > void (*)();
> > 
> > 
> > (for POSIX this is not a problem)
> 
> Yeah, although at this point, I'm in favor of standardizing things that are 
> de facto universal anyway. POSIX for example mandates that NULL is all zero 
> bits and that function pointers fit in void*, and I think it's time for C 
> to just acknowledge the same.

I tend to agree, but I do not see this happening.

> > 
> > > void stsclosure_destroy(stdclosure_t* closure);
> > > 
> > > The function pointer returned by stdclosure_init would, when called, call
> > > FNPTR with an extra initial parameter (or trailing parameter or something)
> > > of type void* and value DATA and all the remaining parameters with which it
> > > was called.
> > > 
> > > This facility would actually probably have to be more complicated in real
> > > life due to varying platform ABIs, but you get the idea. This is something
> > > that could belong in the standard because it'd provide a generic
> > > abstraction for something every platform can do but that every platform has
> > > to do in a slightly different way.
> > 
> > One problem could be that this requires generating
> > trampolines at run-time (and it is not clear we
> > can do this everywhere where C is supported) or
> > having a fixed set of preallocated trampolines.
> 
> Or remapping a fixed set of trampolines, like libffi does on some 
> platforms.

This should be quite expensive though, if you use lots
of closures.

>  The new stdclosure feature would have to be optional, yes, but 
> in practice it would be supported everywhere people write programs except 
> maybe a few microcontrollers and DSPs.

Yes, you should try to propose this.

> > The other limitation (which the new type would
> > avoid) is the need for dynamic resource allocation,
> > which is sometimes not desirable.
> > 
> > But I would love to see such a proposal
> > brought forward.

Martin