[RFC] c++: fix broken conversion in coroutines

  You can't use CONVERT_EXPR to convert between two class types, and it was
breaking copy elision.

Unfortunately, this patch breaks symmetric-transfer-00-basic.C, where
susp_type is Loopy<int>::handle_type.  How is this supposed to work?

gcc/cp/ChangeLog:

	* coroutines.cc (expand_one_await_expression): Change conversion
	to assert.
---
 gcc/cp/coroutines.cc | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)


base-commit: 43faf3e5445b571731e52faa1be085ecd0a09323
  

Hi Jason,

> On 30 Sep 2022, at 23:06, Jason Merrill <jason@redhat.com> wrote:
> 
> You can't use CONVERT_EXPR to convert between two class types, and it was
> breaking copy elision.
> 
> Unfortunately, this patch breaks symmetric-transfer-00-basic.C, where
> susp_type is Loopy<int>::handle_type.  How is this supposed to work?

We are trying to save a type-erased handle (which the symmetric transfer makes
and indirect call through, nothing else).

so, I suppose the equivalent could be:

conthand = coroutine_handle::from_address (suspend.address())

or, is there some cast version that would be valid here?
Iain

> 
> gcc/cp/ChangeLog:
> 
> 	* coroutines.cc (expand_one_await_expression): Change conversion
> 	to assert.
> ---
> gcc/cp/coroutines.cc | 4 +++-
> 1 file changed, 3 insertions(+), 1 deletion(-)
> 
> diff --git a/gcc/cp/coroutines.cc b/gcc/cp/coroutines.cc
> index eca01abcb7a..568f2edf67d 100644
> --- a/gcc/cp/coroutines.cc
> +++ b/gcc/cp/coroutines.cc
> @@ -1728,7 +1728,9 @@ expand_one_await_expression (tree *stmt, tree *await_expr, void *d)
>     }
>   else
>     {
> -      r = build1_loc (loc, CONVERT_EXPR, void_coro_handle_type, suspend);
> +      gcc_checking_assert (same_type_ignoring_top_level_qualifiers_p
> +			   (void_coro_handle_type, susp_type));
> +      r = suspend;
>       r = build2_loc (loc, INIT_EXPR, void_coro_handle_type, data->conthand, r);
>       r = build1 (CONVERT_EXPR, void_type_node, r);
>       append_to_statement_list (r, &body_list);
> 
> base-commit: 43faf3e5445b571731e52faa1be085ecd0a09323
> -- 
> 2.31.1
>
  

On 9/30/22 18:50, Iain Sandoe wrote:
> Hi Jason,
> 
>> On 30 Sep 2022, at 23:06, Jason Merrill <jason@redhat.com> wrote:
>>
>> You can't use CONVERT_EXPR to convert between two class types, and it was
>> breaking copy elision.
>>
>> Unfortunately, this patch breaks symmetric-transfer-00-basic.C, where
>> susp_type is Loopy<int>::handle_type.  How is this supposed to work?
> 
> We are trying to save a type-erased handle (which the symmetric transfer makes
> and indirect call through, nothing else).

The problem is you're treating one class directly as another class here, 
without the indirection involved in usual type-erasure idioms.

It does seem that the gimplifier handles this fine, but it doesn't 
correspond to anything in the language and much of the front end assumes 
that CONVERT_EXPR is only used for scalars.  VIEW_CONVERT_EXPR would 
better express that we're not doing anything to the value, just cheating 
the type system.  That's still dodgy from a language perspective, but 
probably safe enough in this case.

Note that I was wrong to mention copy elision above; it's irrelevant to 
codegen here since the handle type returns in a register.

> so, I suppose the equivalent could be:
> 
> conthand = coroutine_handle::from_address (suspend.address())

That sounds more correct, yes.

Jason
  

On 10/3/22 23:53, Jason Merrill wrote:
> On 9/30/22 18:50, Iain Sandoe wrote:
>> Hi Jason,
>>
>>> On 30 Sep 2022, at 23:06, Jason Merrill <jason@redhat.com> wrote:
>>>
>>> You can't use CONVERT_EXPR to convert between two class types, and it 
>>> was
>>> breaking copy elision.
>>>
>>> Unfortunately, this patch breaks symmetric-transfer-00-basic.C, where
>>> susp_type is Loopy<int>::handle_type.  How is this supposed to work?
>>
>> We are trying to save a type-erased handle (which the symmetric 
>> transfer makes
>> and indirect call through, nothing else).
> 
> The problem is you're treating one class directly as another class here, 
> without the indirection involved in usual type-erasure idioms.
> 
> It does seem that the gimplifier handles this fine, but it doesn't 
> correspond to anything in the language and much of the front end assumes 
> that CONVERT_EXPR is only used for scalars.  VIEW_CONVERT_EXPR would 
> better express that we're not doing anything to the value, just cheating 
> the type system.  That's still dodgy from a language perspective, but 
> probably safe enough in this case.

So I'm applying this:
  

> On 6 Oct 2022, at 22:44, Jason Merrill <jason@redhat.com> wrote:
> 
> On 10/3/22 23:53, Jason Merrill wrote:
>> On 9/30/22 18:50, Iain Sandoe wrote:
>>> Hi Jason,
>>> 
>>>> On 30 Sep 2022, at 23:06, Jason Merrill <jason@redhat.com> wrote:
>>>> 
>>>> You can't use CONVERT_EXPR to convert between two class types, and it was
>>>> breaking copy elision.
>>>> 
>>>> Unfortunately, this patch breaks symmetric-transfer-00-basic.C, where
>>>> susp_type is Loopy<int>::handle_type.  How is this supposed to work?
>>> 
>>> We are trying to save a type-erased handle (which the symmetric transfer makes
>>> and indirect call through, nothing else).
>> The problem is you're treating one class directly as another class here, without the indirection involved in usual type-erasure idioms.
>> It does seem that the gimplifier handles this fine, but it doesn't correspond to anything in the language and much of the front end assumes that CONVERT_EXPR is only used for scalars.  VIEW_CONVERT_EXPR would better express that we're not doing anything to the value, just cheating the type system.  That's still dodgy from a language perspective, but probably safe enough in this case.
> 
> So I'm applying this:<0001-c-fix-broken-conversion-in-coroutines.patch>

thanks, I have not had any cycles to look at this.

however, when I next do - was planning on looking at the:
cont = handle.from_address(await_suspend().address())
approach, since both .address() and .from_address() are constexpr, cp_fold_function should turn that into essentially a NOP.
Iain