I made the implicit destructor always be created in __generic address space.

Looks good!

I made the check into an assert as suggested.

Added a comment explaining what the check is meant for.

I've reverted to using Constructors and prioritizing based on which constructor is the most qualified.

I've added a check in the constructor for CastOperator that removes qualifiers for non-class, non-array types.
I had to add a check for ObjectiveC, since they seem to use qualifiers for a lot of language features,
but maybe a better solution would be to remove specific qualifiers instead?

Cleaned up the test by renaming the struct and making the test compile.

I've cleaned up the check. The qualifiers of SrcType are removed prior to this function, so only DestType needs to have the address space removed.

Made the check more general, it's no longer OpenCL specific and no longer restricted to just constructors, since the check only requires them to be methods.

Relaxed the checks in the codegen test because of failures on platforms with slightly different IR being generated.

Fixed the code style and added some comments to the tests.

Added a codegen test and clarified what some parts are testing.

I tried to add a special case for 3 and 4 element vectors, but that caused issues in codgen, so I've left it as a FIXME for now.

Removed the address space related changes. This revision now only allows for reinterpret_cast between vectors of the same size.

Fixed formatting.

Fixed merge error

Added a link to the C++ for OpenCL specification and cleaned up the text according to the requests.

Improved the description and example per the feedback.

I noticed that I could move the new check above the diagnostic for program scope variables, and the check during parsing can then be removed, while maintaining good diagnostics.

The only ones that change (in the test cases at least) are as follows:
Here Old is with the current change, and New is with the call to diagnoseOpenCLTypes at parsing removed.

Renamed test as suggested.

Ran git-clang-format.

Inlined the thread_local check and moved static clk_event_t into the appropriate test.
I did not move the thread_local check since NewVD had the wrong value for TSCS, so D was required.

Restricted the checks to C++ for OpenCL and added another test for OpenCL vector types.

Added more exhaustive tests, as well as fixed the diagnostic to allow reference types.

I have amended the relevant test to validate this behaviour in C++ for OpenCL.

The code now directly adds the __generic address space to the pointee type.

I've removed the comments calling for a fix because I no longer feel that this approach needs that. Given the code that already exists, and without changing too much of it, adding address space deduction in both cases seems like the right choice.

I have added a check in deduceOpenCLAddressSpace() to check if the type is dependent, and not deduce the address space if it is. This is a big change in the behaviour of address space deduction, meaning that template variables only receive address spaces when they are being specialized. This is good, but caused a lot of changes in the current test file, which has been included in this patch.

Disregard the last comment, rolling back that diff for now.

Added a guard in deduceOpenCLAddressSpace to stop it from deducing address space for dependent types, as requested.

Essentially, in each pass the type for the variable starts off being extract from the SourceTypeInfo, but since the address space isn't specified, this type is then missing an address space.
Adding the address space deduction then reintroduces the address space to the type.