Re: [PATCH v2] c++: tweaks for explicit conversion fns diagnostic

[Marek Polacek <polacek@redhat.com>]

On Tue, Aug 29, 2023 at 04:42:33PM -0400, Jason Merrill wrote:
> On 8/28/23 19:24, Marek Polacek wrote:
> > On Fri, Aug 25, 2023 at 08:34:37PM -0400, Jason Merrill wrote:
> > > On 8/25/23 19:37, Marek Polacek wrote:
> > > > Bootstrapped/regtested on x86_64-pc-linux-gnu, ok for trunk?
> > > > 
> > > > -- >8 --
> > > > 
> > > > 1) When saying that a conversion is erroneous because it would use
> > > > an explicit constructor, it might be nice to show where exactly
> > > > the explicit constructor is located.  For example, with this patch:
> > > > 
> > > > [...]
> > > > explicit.C:4:12: note: 'S::S(int)' declared here
> > > >       4 |   explicit S(int) { }
> > > >         |            ^
> > > > 
> > > > 2) When a conversion doesn't work out merely because the conversion
> > > > function necessary to do the conversion couldn't be used because
> > > > it was marked explicit, it would be useful to the user to say so,
> > > > rather than just saying "cannot convert".  For example, with this patch:
> > > > 
> > > > explicit.C:13:12: error: cannot convert 'S' to 'bool' in initialization
> > > >      13 |   bool b = S{1};
> > > >         |            ^~~~
> > > >         |            |
> > > >         |            S
> > > > explicit.C:5:12: note: explicit conversion function was not considered
> > > >       5 |   explicit operator bool() const { return true; }
> > > >         |            ^~~~~~~~
> > > > 
> > > > gcc/cp/ChangeLog:
> > > > 
> > > > 	* call.cc (convert_like_internal): Show where the conversion function
> > > > 	was declared.
> > > > 	(maybe_show_nonconverting_candidate): New.
> > > > 	* cp-tree.h (maybe_show_nonconverting_candidate): Declare.
> > > > 	* typeck.cc (convert_for_assignment): Call it.
> > > > 
> > > > gcc/testsuite/ChangeLog:
> > > > 
> > > > 	* g++.dg/diagnostic/explicit.C: New test.
> > > > ---
> > > >    gcc/cp/call.cc                             | 41 +++++++++++++++++++---
> > > >    gcc/cp/cp-tree.h                           |  1 +
> > > >    gcc/cp/typeck.cc                           |  5 +++
> > > >    gcc/testsuite/g++.dg/diagnostic/explicit.C | 16 +++++++++
> > > >    4 files changed, 59 insertions(+), 4 deletions(-)
> > > >    create mode 100644 gcc/testsuite/g++.dg/diagnostic/explicit.C
> > > > 
> > > > diff --git a/gcc/cp/call.cc b/gcc/cp/call.cc
> > > > index 23e458d3252..09ebcf6a115 100644
> > > > --- a/gcc/cp/call.cc
> > > > +++ b/gcc/cp/call.cc
> > > > @@ -8459,12 +8459,21 @@ convert_like_internal (conversion *convs, tree expr, tree fn, int argnum,
> > > >    		if (pedwarn (loc, 0, "converting to %qT from initializer list "
> > > >    			     "would use explicit constructor %qD",
> > > >    			     totype, convfn))
> > > > -		  inform (loc, "in C++11 and above a default constructor "
> > > > -			  "can be explicit");
> > > > +		  {
> > > > +		    inform (loc, "in C++11 and above a default constructor "
> > > > +			    "can be explicit");
> > > > +		    inform (DECL_SOURCE_LOCATION (convfn), "%qD declared here",
> > > > +			    convfn);
> > > 
> > > I'd swap these two informs.
> > 
> > Done.
> > > > +++ b/gcc/testsuite/g++.dg/diagnostic/explicit.C
> > > > @@ -0,0 +1,16 @@
> > > > +// { dg-do compile { target c++11 } }
> > > > +
> > > > +struct S {
> > > > +  explicit S(int) { }
> > > > +  explicit operator bool() const { return true; } // { dg-message "explicit conversion function was not considered" }
> > > > +  explicit operator int() const { return 42; } // { dg-message "explicit conversion function was not considered" }
> > > > +};
> > > > +
> > > > +void
> > > > +g ()
> > > > +{
> > > > +  S s = {1}; // { dg-error "would use explicit constructor" }
> > > > +  bool b = S{1}; // { dg-error "cannot convert .S. to .bool. in initialization" }
> > > > +  int i;
> > > > +  i = S{2}; // { dg-error "cannot convert .S. to .int. in assignment" }
> > > > +}
> > > 
> > > Let's also test other copy-initialization contexts: parameter passing,
> > > return, throw, aggregate member initialization.
> > 
> > Done except for throw.  To handle arg passing I moved the call to
> > maybe_show_nonconverting_candidate one line down.  I guess a testcase
> > for throw would be
> > 
> > struct T {
> >    T() { } // #1
> >    explicit T(const T&) { } // #2
> > };
> > 
> > void
> > g ()
> > {
> >    T t{};
> >    throw t;
> > }
> > 
> > but #2 isn't a viable candidate so this would take more effort to handle.
> 
> True, copy-initialization is different when the types are the same.
> 
> > We just say about #1 that "candidate expects 0 arguments, 1 provided".
> > 
> > clang++ says
> > 
> > e.C:3:12: note: explicit constructor is not a candidate
> >      3 |   explicit T(const T&) { }
> >        |            ^
> 
> That would be better; in add_candidates when we see an explicit constructor
> we could add it to bad_fns instead of ignoring it.  But that doesn't need to
> be part of this patch.
 
I created https://gcc.gnu.org/PR111230 for that.

> > +void
> > +maybe_show_nonconverting_candidate (tree to, tree from, tree arg, int flags)
> > +{
> > +  if (!(flags & LOOKUP_ONLYCONVERTING))
> > +    return;
> > +
> > +  conversion_obstack_sentinel cos;
> > +  conversion *c = implicit_conversion (to, from, arg, /*c_cast_p=*/false,
> > +				       flags & ~LOOKUP_ONLYCONVERTING, tf_none);
> > +  if (c && !c->bad_p && c->user_conv_p)
> > +    /* Ay, the conversion would have worked in copy-init context.  */
> 
> s/copy/direct/
> 
> OK with that change.

Duh, of course.  I'm pushing the patch with that fixed.