[PATCH] D62962: Clang implementation of sizeless types

[JF Bastien via Phabricator via cfe-commits]

jfb added inline comments.


================
Comment at: test/Sema/sizeless-1.c:66
+  _Static_assert(__atomic_always_lock_free(1, &local_int8) == __atomic_always_lock_free(1, incomplete_ptr), "");
+  _Static_assert(__atomic_always_lock_free(2, &local_int8) == __atomic_always_lock_free(2, incomplete_ptr), "");
+
----------------
rsandifo-arm wrote:
> jfb wrote:
> > I expect sizeless types are never lock free.
> Yeah, that's right.  But AIUI `__atomic_always_lock_free(N, P)` asks whether an N-byte access at P is lock-free (where P can be null to query standard alignment).  So the question isn't whether sizeless types are lock-free, but whether an N-byte access is lock-free given the alignment guarantees of P.  For this line the answer would be yes if `&local_int8` was aligned to a 2-byte boundary.
> 
> The query isn't really that interesting for sizeless types.  The reason for having it is that `IntExprEvaluator::VisitBuiltinCallExpr` says that 1-byte accesses are lock-free if the target has lock-free accesses, whatever P happens to be.  But for larger N it punts when P is a pointer to incomplete type, because in that case it knows nothing about the alignment of P.  The test is enforcing this behaviour for sizeless types too.
Ah yes, that makes sense. Thanks!


Repository:
  rC Clang

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D62962/new/

https://reviews.llvm.org/D62962