Re: [PATCH] c++: consider built-in operator candidates first

[Jason Merrill <jason@redhat.com>]

On 9/20/21 15:32, Patrick Palka wrote:
> On Mon, 20 Sep 2021, Jason Merrill wrote:
> 
>> On 9/20/21 12:46, Patrick Palka wrote:
>>> During operator overload resolution, we currently consider non-member
>>> candidates before built-in candidates.  This didn't make a difference
>>> before r12-3346, but after this change add_candidates will avoid
>>> computing excess argument conversions if we've already seen a strictly
>>> viable candidate, so it's better to consider built-in candidates first.
>>
>> Doesn't r12-3346 stop considering conversions after it sees a bad one, and
>> later return to the bad candidate if there is no strictly viable candidate?
>> How does this patch change that?
> 
> Yes, but add_candidates also looks for a strictly viable candidate among
> the already-considered candidates in the 'candidates' list via the line:
> 
>    bool seen_strictly_viable = any_strictly_viable (*candidates);
> 
> So by considering the built-in candidates first, the subsequent call to
> add_candidates that considers the non-member functions in will be aware
> of any (built-in) strictly viable candidate.

Ah, I get it, the problem is that the first add_candidates can't see any 
strictly-viable candidates.

>> Depending on the order of the candidates seems fragile.
> 
> Yeah.. :/  I guess in general it'd be better to build up the entire
> overload set first and then call add_candidates exactly once (which
> would also make the perfect candidate optimization more consistent/effective).
> But I'm not sure if we can easily build up such an overload set in this
> case since built-in candidates are represented and handled differently
> than non-built-in candidates..

Or as another way of getting the same effect, add another possible value 
of shortcut_bad_convs to mean leave bad candidates incomplete in the 
candidates list, and then once we're done adding candidates and still 
don't have a viable one, we can go back and finish processing the bad 
candidates?

Either way, this could also help when there are both member and 
non-member candidates for the operator.

> FWIW, although the test case added by this patch is contrived, this
> opportunity was found in the real world by instrumenting the 'bad_fns'
> mechanism added by r12-3346 to look for situations where we still end up
> using it (and thus end up redundantly considering some candidates twice),
> and this built-in operator situation was the most common in the
> codebases that I tested (although still quite rare in the codebases that
> I tested).

>>
>>> Bootstrapped and regtested on x86_64-pc-linux-gnu, does this look OK for
>>> trunk?
>>>
>>> gcc/cp/ChangeLog:
>>>
>>> 	* call.c (add_operator_candidates): Consider built-in operator
>>> 	candidates before considering non-member candidates.
>>>
>>> gcc/testsuite/ChangeLog:
>>>
>>> 	* g++.dg/template/conv17.C: Extend test.
>>> ---
>>>    gcc/cp/call.c                          | 13 +++++++------
>>>    gcc/testsuite/g++.dg/template/conv17.C |  7 +++++++
>>>    2 files changed, 14 insertions(+), 6 deletions(-)
>>>
>>> diff --git a/gcc/cp/call.c b/gcc/cp/call.c
>>> index c5601d96ab8..c0da083758f 100644
>>> --- a/gcc/cp/call.c
>>> +++ b/gcc/cp/call.c
>>> @@ -6321,7 +6321,6 @@ add_operator_candidates (z_candidate **candidates,
>>>    			 vec<tree, va_gc> *arglist,
>>>    			 int flags, tsubst_flags_t complain)
>>>    {
>>> -  z_candidate *start_candidates = *candidates;
>>>      bool ismodop = code2 != ERROR_MARK;
>>>      tree fnname = ovl_op_identifier (ismodop, ismodop ? code2 : code);
>>>    @@ -6333,6 +6332,12 @@ add_operator_candidates (z_candidate **candidates,
>>>      if (rewritten && code != EQ_EXPR && code != SPACESHIP_EXPR)
>>>        flags &= ~LOOKUP_REWRITTEN;
>>>    +  /* Add built-in candidates to the candidate set.  The standard says to
>>> +     rewrite built-in candidates, too, but there's no point.  */
>>> +  if (!rewritten)
>>> +    add_builtin_candidates (candidates, code, code2, fnname, arglist,
>>> +			    flags, complain);
>>> +
>>>      bool memonly = false;
>>>      switch (code)
>>>        {
>>> @@ -6352,6 +6357,7 @@ add_operator_candidates (z_candidate **candidates,
>>>        /* Add namespace-scope operators to the list of functions to
>>>         consider.  */
>>> +  z_candidate *start_candidates = *candidates;
>>>      if (!memonly)
>>>        {
>>>          tree fns = lookup_name (fnname, LOOK_where::BLOCK_NAMESPACE);
>>> @@ -6423,11 +6429,6 @@ add_operator_candidates (z_candidate **candidates,
>>>        if (!rewritten)
>>>        {
>>> -      /* The standard says to rewrite built-in candidates, too,
>>> -	 but there's no point.  */
>>> -      add_builtin_candidates (candidates, code, code2, fnname, arglist,
>>> -			      flags, complain);
>>> -
>>>          /* Maybe add C++20 rewritten comparison candidates.  */
>>>          tree_code rewrite_code = ERROR_MARK;
>>>          if (cxx_dialect >= cxx20
>>> diff --git a/gcc/testsuite/g++.dg/template/conv17.C
>>> b/gcc/testsuite/g++.dg/template/conv17.C
>>> index f0f10f2ef4f..87ecefb8de3 100644
>>> --- a/gcc/testsuite/g++.dg/template/conv17.C
>>> +++ b/gcc/testsuite/g++.dg/template/conv17.C
>>> @@ -61,3 +61,10 @@ concept E = requires { T().h(nullptr); };
>>>      static_assert(!E<C>);
>>>    #endif
>>> +
>>> +// Verify that the strictly viable built-in operator+ candidate precludes
>>> +// us from computing all argument conversions for the below non-strictly
>>> +// viable non-member candidate.
>>> +enum N { n };
>>> +int operator+(N&, B);
>>> +int f = n + 42;
>>>
>>
>>
>