[PATCH] D15048: [tsan] Add dispatch_group API interceptors and synchronization

[Kuba Brecka via llvm-commits]

kubabrecka added inline comments.

================
Comment at: lib/tsan/rtl/tsan_libdispatch_mac.cc:193
@@ +192,3 @@
+  WRAP(dispatch_async)(queue, ^(void) {
+    block();
+    WRAP(dispatch_group_leave)(group);
----------------
dvyukov wrote:
> Why don't we need Block_copy here, but need it below? Is the block automatically referenced from captured by another block?
Technically, we don't need Block_copy for correctness, because the function that will store a reference to a block for later (dispatch_async in this case) will do it anyway.  The reason for a Block_copy here is to get rid of a very common false positive, because Block_copy will call malloc and the malloc'd region is then accessed from another thread.  Since this happens *inside* dispatch_async, it's already after we've Release()'d.  So I'm just forcing the allocation to happen early (before `Release()`).

I'm not calling Block_copy here, because `WRAP(dispatch_async)` will call it anyway.  Below, we're calling `REAL(...)`, so we need to Block_copy.


http://reviews.llvm.org/D15048