[cfe-commits] [PATCH] OpenCL error for elements in a struct

[Tanya Lattner]

FYI, I'm retracting this patch.

Thanks,
Tanya

On Jun 14, 2012, at 4:08 AM, Anton Lokhmotov <Anton.Lokhmotov at arm.com> wrote:

> Hi Tanya,
> 
>> OpenCL does not allow elements of a struct to be in different address
>> spaces. The attached patch checks for this and throws an error. It
>> includes a test case.
> 
> I can see how your patch attempts to directly support the following
> restriction in the spec:
> 
> "Elements of a struct or union must belong to the same address space.
> Declaring a struct
> or union whose elements are in different address spaces is illegal."
> 
> (Seems to be unchanged since OpenCL v1.0; 6.8.o in OpenCL v1.1).
> 
> However, I think it's a flaw in the spec as elements of a struct or union
> should not have any address space qualifiers.
> 
> In fact, there's an assertion in clang/lib/Sema/SemaExprMember.cpp:1611
> which on:
> 
> __kernel void abc()
> {
>  struct __e {
>    __local int a;
>    __local int b;
>  } x;
> 
> x.a = 1;
> x.b = 2;
> }
> 
> produces:
> 
> clang::ExprResult BuildFieldReferenceExpr(clang::Sema&, clang::Expr*, bool,
> const clang::CXXScopeSpec&, clang::FieldDecl*, clang::DeclAccessPair, const
> clang::DeclarationNameInfo&): Assertion `!MemberQuals.hasAddressSpace()'
> failed.
> Stack dump:
> 0.      11.cl:8:6: current parser token '='
> 1.      11.cl:2:1: parsing function body 'abc'
> 2.      11.cl:2:1: in compound statement ('{}')
> Aborted
> 
> So I'm not sure this check should be implemented, although better error
> handling (than just an assertion) of address space qualifiers on struct and
> union members would be most welcome!
> 
> Best, Anton.
> 
> 
>