Re: [PATCH] c++: Incremental fix for g++.dg/gomp/for-21.C [PR84469]

[Jason Merrill <jason@redhat.com>]

On 11/30/22 10:51, Jakub Jelinek wrote:
> On Tue, Nov 29, 2022 at 11:05:33PM +0100, Jakub Jelinek wrote:
>> On Tue, Nov 29, 2022 at 04:38:50PM -0500, Jason Merrill wrote:
>>>> --- gcc/testsuite/g++.dg/gomp/for-21.C.jj	2020-01-12 11:54:37.178401867 +0100
>>>> +++ gcc/testsuite/g++.dg/gomp/for-21.C	2022-11-29 13:06:59.038410557 +0100
>>>> @@ -54,9 +54,9 @@ void
>>>>    f6 (S (&a)[10])
>>>>    {
>>>>      #pragma omp for collapse (2)
>>>> -  for (auto [i, j, k] : a)			// { dg-error "use of 'i' before deduction of 'auto'" "" { target *-*-* } .-1 }
>>>> +  for (auto [i, j, k] : a)			// { dg-error "use of 'i' before deduction of 'auto'" }
>>>>        for (int l = i; l < j; l += k)		// { dg-error "use of 'j' before deduction of 'auto'" }
>>>> -      ;						// { dg-error "use of 'k' before deduction of 'auto'" "" { target *-*-* } .-3 }
>>>> +      ;						// { dg-error "use of 'k' before deduction of 'auto'" "" { target *-*-* } .-1 }
>>>
>>> Hmm, this error is surprising: since the initializer is non-dependent, we
>>> should have deduced immediately.  I'd expect the same error as in the
>>> non-structured-binding cases, "* expression refers to iteration variable".
>>
>> The reason was just to be consistent what is (unfortunately) emitted
>> in the other cases (!processing_template_decl or type dependent).
>> I guess I could try how much work would it be to deduce it sooner, but
>> generally it is pretty corner case, people rarely do this in OpenMP code.
> 
> I had a look at that today, but it would be pretty hard.  The thing is
> we must emit all the associated code for all the range for loops in
> OpenMP loops at a different spot.  So, the only possibility I see would
> be if we during parsing of a range for loop inside of the OpenMP loop nest
> we don't do the cp_finish_omp_range_for stuff to avoid e.g. cp_finish_decl,
> but instead
>    build_x_indirect_ref (input_location, begin, RO_UNARY_STAR,
> 			NULL_TREE, tf_none)
> and if that gives a non-dependent type, temporarily overwrite TREE_TYPE
> of the decl and if it is structured binding, temporarily
> ++processing_template_decl and cp_finish_decomp, then after parsing all the
> associated loop headers we revert that (and ditto for instantiation of
> OpenMP loops).

It looks like we're already deducing the type for the underlying S 
variable in cp_convert_omp_range_for, we just aren't updating the types 
of the individual bindings.

Jason