On Windows, IUknown::AddRef and IUnknown::Release are defined to
return ULONG, which is defined as unsigned long. The existing
code relies on the PR_BYTES_PER_LONG macro which is #defined in
prtypes.h to always be 4 on Windows. If we remove the prtypes.h
inclusion from nscore.h, in the places where prtypes.h is not
included by other things, we will fall into the other path in the
#ifdef condition which defines nsrefcnt to be unsigned int, which
causes the signature of AddRef and Release to change, which
results in linker errors.
There are a couple of reasons why this change is correct:
1. On both Win32 and Win64, the size of the long type is 4 bytes,
always. And prtypes.h doesn't detect the size of longs, it just
assumes it, so this check is really a tautology.
2. If the size of the long type changes on Windows, sometime,
the return value of AddRef and Release on Windows should change
size as well in order to maintain binary compatibility with
IUnknown, and before this change, the code would fail to hold
that promise.
--HG--
extra : rebase_source : 8333582b494aaca17c91cf35287a084083219050
This makes our builds a bit more reproducible. Without it when processing
#include "IPC/IPCMessageUtils.h
we end up in
#include "ShadowLayers.h"
which in the final XUL translates into the symbol (run nm -a to see it):
- 00 0000 SOL ../../../../gfx/layers/IPC/ShadowLayers.h
now when some file in gfx/layers/ipc does
#include "ShadowLayres.h"
the file found is the local one with the lowercase directory name and we end up
with
- 00 0000 SOL ../../../../gfx/layers/ipc/ShadowLayers.h
--HG--
extra : rebase_source : 96c68f6a3448eedaa280039960df911a9c707cac
This patch actually converts over most of the existing native CC classes
to use the participant. This is done by converting them to use
the newly generalized nsCycleCollectingAutoRefCnt instead of their usual
ref count.
This is mostly a matter of defining new macros defining AddRef() and Release()
for native CC classes with the CC ref count.
nsTimeout is left as a legacy native CC class because it is a weird class,
but hopefully eventually it too can be converted over. See bug 774874.
We add a new field to purple buffer entries, to store the participant for
non-nsISupports classes. For nsISupports, we store NULL instead. The participant
has to be passed into Suspect2.
In the cycle collector itself, we generalize canonicalization to handle the
possibility that we can just grab the participant directly from a field,
rather than needing to QI something.
Most of the patch is just routing around this extra pointer.
