Optimize and fix handling of QtMessageHandlers

A function may almost always have static storage duration, but that
does not necessarily mean that we can store and load pointers to them
without memory ordering. Play it safe and use store-release and
load-acquire for them (which combines to ordered for the fetchAndSet
call in qInstall*Handler(), as we don't know what the caller will do
with the returned function pointer).

Also change the initial value of the atomic pointer to nullptr.
Nullptr already signified the default handler in qInstall*Handler(),
so the API doesn't change. But by using nullptr to mean default, we
place these variables in the BSS segment instead of TEXT, save dynamic
init, or at least a relocation, and we dodge the smelly comparison of
function pointers, using comparison against nullptr instead.

Also, as a drive-by, put the call to ungrabMessageHandler() in a
scope-guard. Both the message handler, as well as the Qt code calling
it (toLocal8Bit()!), may throw, and that would stop all further
logging. In Qt 5.9, we can't use qScopeGuard(), yet, so use a local
struct calling ungrabMessageHandler() in its dtor.

The code still has one problem: When a logging action is underway, and
another thread exchanges the message handler, we might still execute
code in the old handler. This is probably not a problem in practice,
since no-one will use a dynamically-compiled function for logging
(right? :), but should probably be documented or fixed. This patch
does not address this issue, though.

Change-Id: I21aa907288b9c8c6646787b4001002d145b114a5
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
(cherry picked from commit cd401b74a13cd9d9a47d977f195c7985cf725d55)
(cherry picked from commit ea16c860bd75a35134ebb1d4f3be5db58f4a4e21)

0 files changed, 0 insertions, 0 deletions