[cfe-dev] [RFC] Delayed target-specific diagnostic when compiling for the devices.

[John McCall via cfe-dev]

On 17 Jan 2019, at 10:20, Finkel, Hal J. wrote:

> On 1/17/19 4:01 AM, Alexey Bataev wrote:
>> We cannot build the call graph during parsing. The function may be 
>> parsed already, but it may be called later.
>
>
> Calling this "during parsing" maybe a bit over simplistic - it also
> needs to deal with template instantiation, etc. However, the
> parse-before-call is not a problem because all functions need to be
> declared before being called, so we always have a way to record the
> information (that seems to be the logic of the current approach).

It probably makes more sense to walk functions retroactively looking for
declaration references, but that's still quite feasible.

>> I've come to an idea how to reuse the existing solution for Cuda for 
>> our purposes without the redesign and significant changes. Need some 
>> time to try to implement a prototype.
>> Just one question though. Should I reuse the existing classes/objects 
>> for Cuda or better to add the specific ones for OpenMP, but almost 
>> with the same behavior? Or just rename the existing Cuda classes to 
>> be more common, like CudaDeferredDiags to something like 
>> DeviceDeferredDiags?
>
>
> I definitely vote for renaming and generalizing the current
> functionality unless there's some strong reason not to do so.

I agree.  If the functionality is generalized to apply to not just
CUDA, it should be renamed to something non-CUDA-specific.

John.