Re: [C++ Patch] PR 84333 ("[6/7/8 Regression] ICE with ternary operator in template function")

[Jason Merrill <jason at redhat dot com>]

On Tue, Feb 13, 2018 at 10:43 AM, Paolo Carlini
<paolo.carlini@oracle.com> wrote:
> On 13/02/2018 15:35, Jason Merrill wrote:
> Thanks. Unfortunately, I don't think I sent a complete solution for these
> issues. In fact we would still ICE on:
>
> template<typename T> int foo()
> {
>   return sizeof(T) > 1 ? : 1;
> }
>
> because type_contains_placeholder_p doesn't know how to handle a
> TEMPLATE_TYPE_PARM. In fact, cp_save_expr teaches me something:
>
> tree
> cp_save_expr (tree expr)
> {
>   /* There is no reason to create a SAVE_EXPR within a template; if
>      needed, we can create the SAVE_EXPR when instantiating the
>      template.  Furthermore, the middle-end cannot handle C++-specific
>      tree codes.  */
>   if (processing_template_decl)
>     return expr;
>   return save_expr (expr);
> }
>
> would it be correct to just use it? Or we have to do something more complex?
> Note the issue only affects the GNU-extension with omitted middle operand,
> thus I believe we have *some* leeway...

Hmm, yes, that should work.  In a template we're just figuring out the
result type, at instantiation time we'll see the missing middle op
again and can do the right thing.  OK.

Jason