Re: [PATCH] c++: refine CWG 2369 satisfaction vs non-dep convs [PR99599]

[Jason Merrill <jason@redhat.com>]

On 8/21/23 21:51, Patrick Palka wrote:
> Bootstrapped and regtested on x86_64-pc-linux-gnu, does this look like
> a reasonable approach?  I didn't observe any compile time/memory impact
> of this change.
> 
> -- >8 --
> 
> As described in detail in the PR, CWG 2369 has the surprising
> consequence of introducing constraint recursion in seemingly valid and
> innocent code.
> 
> This patch attempts to fix this surpising behavior for the majority
> of problematic use cases.  Rather than checking satisfaction before
> _all_ non-dependent conversions, as specified by the CWG issue,
> this patch makes us first check "safe" non-dependent conversions,
> then satisfaction, then followed by "unsafe" non-dependent conversions.
> In this case, a conversion is "safe" if computing it is guaranteed
> to not induce template instantiation.  This patch heuristically
> determines "safety" by checking for a constructor template or conversion
> function template in the (class) parm or arg types respectively.
> If neither type has such a member, then computing the conversion
> should not induce instantiation (modulo satisfaction checking of
> non-template constructor and conversion functions I suppose).
> 
> +	  /* We're checking only non-instantiating conversions.
> +	     A conversion may instantiate only if it's to/from a
> +	     class type that has a constructor template/conversion
> +	     function template.  */
> +	  tree parm_nonref = non_reference (parm);
> +	  tree type_nonref = non_reference (type);
> +
> +	  if (CLASS_TYPE_P (parm_nonref))
> +	    {
> +	      if (!COMPLETE_TYPE_P (parm_nonref)
> +		  && CLASSTYPE_TEMPLATE_INSTANTIATION (parm_nonref))
> +		return unify_success (explain_p);
> +
> +	      tree ctors = get_class_binding (parm_nonref,
> +					      complete_ctor_identifier);
> +	      for (tree ctor : lkp_range (ctors))
> +		if (TREE_CODE (ctor) == TEMPLATE_DECL)
> +		  return unify_success (explain_p);

Today we discussed maybe checking CLASSTYPE_NON_AGGREGATE?

Also, instantiation can also happen when checking for conversion to a 
pointer or reference to base class.

Jason