Fixing ADL, Round 2
Yesterday I blogged about a Clang patch I made that I claimed was an implementation of Herb Sutter’s proposal P0934R0 “A Modest Proposal: Fixing ADL” (February 2018). (You can find the patch here, or try it live on Godbolt as part of my experimental P1144 branch.)
When the post hit Reddit, commenter sphere991 pointed out that I had utterly missed about two-thirds of Herb’s proposal —
Herb’s proposal has two parts: narrowing the set of associated namespaces and narrowing the set of functions looked up.
Here’s the “narrowing” wording from P0934:
When considering an associated namespace, the lookup is the same as the lookup performed when the associated namespace is used as a qualifier, except that:
… Any function that does not have a parameter type of (possibly cv-qualified)
T
or accessible base ofT
, pointer toT
or accessible base ofT
, or reference toT
or accessible base ofT
, in the parameter position ofT
, is ignored.
As far as I can tell, Herb’s wording includes instantiated functions, so an ADL call
to distance(list.begin(), list.end())
would still happily find std::distance<std::list::iterator>
.
(There was some disagreement about this in the Reddit comments.)
My -Wadl
patch does not implement this narrowed form of lookup.
But there is also a third part to P0934, missed by both myself and sphere991, which expands the lookup set (compared to how it is today)!
P0934 replaces the phrase “lookup” with the phrase “ordinary unqualified lookup” in two places, which seems to have the effect of circumventing C++’s current “name-hiding” mechanics in ADL calls. This test case illustrates the proposed change:
namespace A {
class A {};
inline void h(A&) {
puts("A::h");
}
}
namespace B {
inline void h(const A::A&) {
puts("B::h");
}
void calls_Ah(A::A parm) {
using B::h; // supplementing
h(parm);
}
void calls_Bh(A::A parm) {
void h(const A::A&); // hiding
h(parm);
}
}
Right now, the declaration of B::h
inside calls_Bh
more or less “resets” the overload set available
to ADL calls from within calls_Bh
, “hiding” the better-matching A::h
from view.
Herb’s paper claims (see the example involving NS::g
on page 14) to effectively
eliminate this “hiding” mechanism for ADL calls, which means that both of the functions
above would end up calling A::h
because it is the best match. If you deliberately
wanted the worse-matching B::h
, you’d have to call it by its qualified name.
I don’t claim to understand why this change would be good. I almost suspect that it would be super bad because it would interfere with lookups of class member functions. But I suspect I’m misunderstanding it. I’m certainly fuzzy enough on it that I don’t think I could implement it!