It's not clear to me why specializing nsCOMPtr<nsISupports> is
particularly useful, and this removes a lot of code, bringing us closer
to unifying nsCOMPtr and RefPtr.
The changes in other places are needed because we used to allow
nsCOMPtr<nsISupports> to implicitly coerce to nsISupports even for
rvalue references.
0:47.88 /home/emilio/src/moz/gecko-4/netwerk/protocol/http/nsHttpActivityDistributor.cpp:121:11: error: conversion function from 'nsCOMPtr<nsISupports>' to 'nsISupports *' invokes a deleted function
0:47.88 nsCOMPtr<nsISupports>(do_QueryObject(channel)), aActivityType,
0:47.88 ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
0:47.88 /home/emilio/src/moz/gecko-4/obj-debug/dist/include/nsCOMPtr.h:764:3: note: 'operator nsISupports *' has been explicitly marked deleted here
0:47.88 operator T*() const&& = delete;
0:47.88 ^
0:47.88 /home/emilio/src/moz/gecko-4/netwerk/protocol/http/nsHttpActivityDistributor.cpp:29:57: note: passing argument to parameter 'aHttpChannel' here
0:47.88 nsHttpActivityDistributor::ObserveActivity(nsISupports* aHttpChannel,
0:47.88 ^
Which is not allowed for a good reason, generally.
Differential Revision: https://phabricator.services.mozilla.com/D179124
This is the same as the other patch, except that it is applied to the
case where the QI returns an nsresult.
In addition, I marked the WithError helper class as being stack-only.
Depends on D7553
Differential Revision: https://phabricator.services.mozilla.com/D7554
This patch adds a static assert to enforce that do_QueryInterface is
not used to go from a class to a base class, because that can be done
more efficiently via a static_cast. This is done by putting the type
of the source into the nsQueryInterface type. Once that is done, it is
easy to check the source and destination type. This has to be done
carefully so that in non-DEBUG builds, where NSCAP_FEATURE_USE_BASE is
defined, we don't generate any additional code.
The first step is to rename nsQueryInterface to
nsQueryInterfaceISupports. In DEBUG builds, I then add a new subtype
nsQueryInterface<T>, where T is the type of the object we are going to
QI. This class is a thin wrapper around nsQueryInterfaceISupports that
only forwards operations to the base class.
The main bit of trickery here is PointedToType<T>, which is needed to
get the type parameter for nsQueryInterface. This dereferences the
type, gets the address of it, then does RemovePointer. This is needed
because a wide variety of pointer types are passed in to
do_QueryInterface, including RefPtr<>, nsCOMPtr<>, raw pointers, and
OwningNonNull<>. PointedToType<RefPtr<T>> is equal to T,
PointedToType<T*> is equal to T, and so on.
In NSCAP_FEATURE_USE_BASE builds (opt), we only use
nsQueryInterfaceISupports, but in debug builds we use
nsQueryInterface<> where possible. The latter is also used for the
nsCOMPtr<nsISupports> overload, because we can always QI to
nsISupports, because that is sometimes used for canonicalization.
Another gross bit is that Assert_NoQueryNeeded can no longer use
nsCOMPtr<>, because it works by QIing T to T, which is banned by the
static analysis. Instead I had to reimplement it by hand.
Depends on D7527
Differential Revision: https://phabricator.services.mozilla.com/D7553
nsQueryReferent is defined as an nsCOMPtr_helper, which implies that
calling its operator() method requires a virtual call. While
nsQueryReferent is marked `final`, compiler inlining decisions make it
impossible to de-virtualize the call to operator(). However, we have
many other classes returned by do_* functions that nsCOMPtr handles
directly, requiring no extra virtual calls, and we can give
nsQueryReferent the same treatment.
this change to me and since he is on vacation until monday and I need this
for an M11 bug, I am checking it in. I am hoping it is the right
thing to do.
a=brendan.