Re: static class member as interrupt handler works, but not if class is templated

[David Brown]

On 12/04/2021 14:12, Klaus Rudolph via Gcc-help wrote:
> Am 12.04.21 um 13:53 schrieb Matthijs Kooijman:
>>>> - figure out, where the special handling of the __vector_10 seems to
>>>> happen, and why it is not happening in the class template case. This
>>>> might help diagnose if and where to fix it within the compiler.
>>>
>>> That is compiler internals... yes, if it is a compiler bug, it is the
>>> way to have a solution. But in that case, it seems to be a generic
>>> problem for gcc as attributes are not assigned to any templated class
>>> member functions. No idea if this is related to the target ( avr ) or
>>> generic for all platforms. But I never did any change inside the
>>> compiler. Any help is welcome!
>>
>> My suspiciou would be that this is a generic gcc problem, where the
>> "asm" attribute is not honoured for template functions. It probably also
>> makes some sense, since a template is intended to be instantiated
>> multiple times, and each instantiation gets a name that is generated
>> based on (I believe) the template arguments passed, so I suspect that
>> the "generate a name for this template instantiation" code doesn't look
>> at the asm attribute.
>>
>> Also note that *if* it would, then each instantiation would use the same
>> name and multiple instanations would result in duplicate symbols. If you
>> would bring this up as a gcc bug, I wouldn't be surprised that it would
>> be closed as wontfix for this reason.
> 
> I disagree as a template "always" would be instantiated multiple times.
> And even if it would be, the linker will fire a error message, as it
> sees multiple definitions. So there is no "general" problem in
> "renaming" a templated function. It simply *can* work.

Certainly templates can be instantiated zero, once, or multiple times.
And certainly some templates are intended only to be instantiated a
single time.  But it would surely be difficult to have a feature that
only works when there is a single instantiation.

There is also the issue of linkage of the names here.  A template
instantiation has function/method names that are mangled with the
template type, template parameters, parameter types, etc.  These have a
particular kind of linkage that lets the toolchain (compiler, assembler
and linker) know that they can be defined in more than one unit, and at
link time one of them (by unspecified choice) can be picked for the
final binary.  An assembly-defined specific name for an interrupt
vector, on the other hand, is a C external linkage name that can only be
defined once.  I don't see that these two uses could easily be combined.

It seems natural to me that a feature which can only be relevant to a
single instantiation of a template should be connected to an
instantiation, not a definition.  This is particularly true when the
details of the attribute you want - the "vector" assembly name - are
dependent on a non-type parameter for the template.

> 
> But it looks that not only the "renaming" stuff did not work, all the
> flags are not handled with the templated function. Looks like that the
> asm declaration did not find its target :-)
> 
>>
>> Another workaround that I think hasn't been suggested yet, would be to
>> just define a global `__vector_10` function and from that just call your
>> templated static member.
> 
> That is the classical way "we" all work. And it optimizes well in that
> case as the code from the static templated member is fully inlined. But
> it is still a workaround and it simply breaks standard c++ coding. Yes,
> we can write C with classes, but I simply dislike :-)
> 

Unfortunately, C++ can't do everything - even with gcc extensions.
There are all sorts of things that could be useful to have, but simply
are not practical or possible to implement.

One that I would like is a way to have a class or template instantiation
that is at file/namespace scope, but which is constructed before first
use rather than before main() starts.  But it should not have the
overhead of run-time checks or indirect access (so no static class
member).  I think it is fair to say it is not going to happen - so I
need manual "init" functions with lists of "init" method calls for each
object.

>> Combined with the `always_inline` attribute,
>> you can ensure that the call is inlined and there is no runtime overhead
>> (with LTO, this probably already happens when there's just a single call
>> to the member).
> 
> It already optimizes well in O2 if templated member function and free
> handler definition is in same translation unit.
> 
> Klaus
> 
>