[PATCH] D23242: [CUDA] Raise an error if a wrong-side call is codegen'ed.

[Justin Lebar via cfe-commits]

jlebar marked 3 inline comments as done.
jlebar added a comment.

> Should be add few tests for calling device functions from host-side global initializers? Perhaps for device->host, too, as there may be unexpected interplay with constructor emptiness checks.


Hm, these seem completely broken, before and after this patch.  I'll handle it separately if that's OK.


================
Comment at: clang/include/clang/Sema/Sema.h:9162
@@ -9161,1 +9161,3 @@
 
+  /// Check whether we're allowed to call Callee from the current context.
+  ///
----------------
tra wrote:
> \p Callee
> 
This file isn't at all consistent about using this.  For example, we're missing three `\p`s on IsAllowedCUDACall, two on IdentifyCUDAPreference just above, and four on maybeAddCUDAHostDeviceAttrs.

If we were to do it here, we'd need to add `\p` to three places; not just `Callee`, but also `CFP_Never` and `CFP_WrongSide`.

I'm highly skeptical that this ends up being beneficial overall to readers (i.e. that the clutter in source code outweighs the slightly prettier formatting on doxygen).


https://reviews.llvm.org/D23242