Re: Should we split the project into two branches?

[H. Nikolaus Schaller]

Am 15.02.2022 um 15:10 schrieb Derek Fawcus <dfawcus+lists-gnustep-discuss@employees.org>:

I don't know about the level of changes needed in GCC, or about the
possibility of using an objC preprocessor, but it was possible to
build and deploy ARC using (including __weak references) code to 10.6.8.

As I recall one had to build against the 10.7 SDK, and then add an
extra package for the __weak support.

So in theory one could do something similar regarding the libobjc1 code.

DF

https://stackoverflow.com/questions/7696201/how-to-deploy-to-snow-leopard-with-arc-enabled
https://www.mikeash.com/pyblog/introducing-plweakcompatibility.html

Interesting. And I still wonder why ARC seems to need any library support.

In my understanding (which is certainly wrong) ARC is automatic reference counting.

Done by a static analysis of the source code. Done during compile time and not

during run time where the libs/frameworks would play a role. This is the reason why

I proposed to think about a preprocessor for standard-gcc to handle these rules and

pass everything unchanged.

AFAIR the most important memory management rules of ObjC 1.0 are:

-alloc and -copy* and -retain return a retained object which must be released before returning

from a method. Objects returned by a method must be (explicitly or implicitly)

retained+autoreleased.

Objects assigned to persistent store (instance vars) must be retained and the previous value

must be released before. This is (not necessarily) usually wrapped by a setter method.

These are not extraordinarily complex so I also tend to think that training is not well enough

if programmers struggle with them. But they obviously do.

Hm. Just some thought: is anyone is looking for a master's or PhD-thesis: would it be

possible to statically analyse retain/release/autorelease code to locate memory leaks

which come from not following these rules?

BTW note: Linux kernel also has trouble with non-automatic reference counting.

retain/release are called get and put there. E.g.:

https://patchwork.kernel.org/project/linux-omap/patch/20211014085719.23739-1-wanjiabing@vivo.com/#24665947

But they don't have the nice autorelease-pool and the rule that most methods return

autoreleased objects and caller must retain them (they often return retained object

and caller must put them)...

--hns