Discussion:
[thread] Ambiguity in variadic wait_for_all
(too old to reply)
Gavin Lambert
2017-02-03 01:38:41 UTC
Permalink
A colleague recently ran into an interesting issue when upgrading some
working code from Boost 1.55 to 1.60/1.63 (and from C++03 to C++11).

The exact details are unimportant, but it was trying to get the address
of boost::wait_for_all(Iterator, Iterator), for use in a Boost.Bind call.

In 1.55, this worked fine, although of course it had to supply template
parameters that proved the parameters were not futures (thus causing
enable_if<is_future_type<F1>> to fail, disabling the two-future-args
overload of the function.

In 1.60/1.63, the variadic method is encountered instead, which lacks
any enable_if condition and thus the ambiguity between the two remains
and the compiler errors out. (When calling it directly, the variadic
one manages to get disabled as long as the iterator type lacks a wait()
method, but this doesn't apply when just getting the method address, as
it's not part of the signature.)

Should the variadic form also have enable_if? Or is this considered
user error for trying to use Bind in a variadic-compatible compiler
instead of using a lambda?
Vicente J. Botet Escriba
2017-02-07 17:51:38 UTC
Permalink
Post by Gavin Lambert
A colleague recently ran into an interesting issue when upgrading some
working code from Boost 1.55 to 1.60/1.63 (and from C++03 to C++11).
The exact details are unimportant, but it was trying to get the
address of boost::wait_for_all(Iterator, Iterator), for use in a
Boost.Bind call.
In 1.55, this worked fine, although of course it had to supply
template parameters that proved the parameters were not futures (thus
causing enable_if<is_future_type<F1>> to fail, disabling the
two-future-args overload of the function.
In 1.60/1.63, the variadic method is encountered instead, which lacks
any enable_if condition and thus the ambiguity between the two remains
and the compiler errors out. (When calling it directly, the variadic
one manages to get disabled as long as the iterator type lacks a
wait() method, but this doesn't apply when just getting the method
address, as it's not part of the signature.)
Should the variadic form also have enable_if? Or is this considered
user error for trying to use Bind in a variadic-compatible compiler
instead of using a lambda?
Hi,

I will suggest you to replace the call to bind by lambdas whenever you can.

You have made two changes IIUC, boost version and compiler version, so I
don't know if this is a regression in boost 1.53 or a behavior that
different between c++02 and C++11. I suspect that it is the change to
C++11.

Curiously I added enable if only for the function with two arguments,
but not for the others.
template<typename F1,typename F2>
typename boost::enable_if<is_future_type<F1>,void>::type
wait_for_all(F1& f1,F2& f2)

When I added the C++11 version I surely forget to add the enable_if.

Please, could you try to add the enable_if to the variadic template and
tell me if this would fix your issue?

template<typename F1, typename... Fs>
*typename boost::enable_if<is_future_type<F1>,void>::type*
wait_for_all(F1& f1, Fs&... fs)


Best,
Vicente
Gavin Lambert
2017-02-07 23:09:25 UTC
Permalink
Post by Vicente J. Botet Escriba
You have made two changes IIUC, boost version and compiler version, so I
don't know if this is a regression in boost 1.53 or a behavior that
different between c++02 and C++11. I suspect that it is the change to
C++11.
Sorry, I gave a misleading impression there. Both Boost versions were
tested with the same compiler (VS2015), so the issue was introduced by
the Boost upgrade only.

Explicitly disabling the variadic version by defining
BOOST_NO_CXX11_VARIADIC_TEMPLATES also avoids the issue, so I'm pretty
sure it's just the introduction of that which did it.
Post by Vicente J. Botet Escriba
Curiously I added enable if only for the function with two arguments,
but not for the others.
That's correct -- only the two-argument form is ambiguous with the
iterator overload.

Although having said that, the iterator form takes a single template
parameter and the future form takes two template parameters, so the
compiler ought to be able to tell them apart (since the template
parameters were explicitly provided), unless it's trying to infer the
second parameter.
Post by Vicente J. Botet Escriba
Please, could you try to add the enable_if to the variadic template and
tell me if this would fix your issue?
template<typename F1, typename... Fs>
*typename boost::enable_if<is_future_type<F1>,void>::type*
wait_for_all(F1& f1, Fs&... fs)
Yes, making that change resolves the ambiguity and the original code
compiles again.

Loading...