ObjFW  History of src/runtime/class.m of 4d82cae50d9b5688

Apparently the optimizer crashes on some recursive functions. file: [4cb79aa414] check-in: [5c0117205d] user: js, branch: trunk, size: 22277 [annotate] [blame] [check-ins using] [diff]

This already makes tagged pointers work in the non-assembly lookup. file: [02bad041ab] check-in: [8707dd73c4] user: js, branch: tagged-pointers, size: 21822 [annotate] [blame] [check-ins using] [diff]

Static selectors are initialized by a global constructor that might run
too late: If a class has a +[load] that triggers an +[initialize], this
could result in trying to call +[initialize] with the selector in
class.m still being uninitialized. When this happens, it results in a
weird looking crash in an entirely unrelated method. This is because
when uninitialized, the selector name has not been replaced with a
selector UID yet, meaning the pointer to the selector name is treated as
the selector UID. As the dispatch only considers the low 16 bits of the
selector UID for performance reasons, it will just take the low 16 bits
of the pointer to the selector name as the UID without any complaints.
This can then result in calling a random method on that class which then
crashes. file: [132bb904b5] check-in: [3b3729a316] user: js, branch: trunk, size: 21746 [annotate] [blame] [check-ins using] [diff]

This is more compatible with the various conventions used on various
systems. file: [15508d0871] check-in: [79fe29dbf4] user: js, branch: trunk, size: 21132 [annotate] [blame] [check-ins using] [diff]

This now calls +[initialize] several times on the same class if it is
unimplemented in one or more of the classes in the chain, to match the
behavior of the Apple runtime. While the behavior before was safer, it
can lead to testing against the ObjFW runtime and everything working,
but then failing when trying to use it with the Apple runtime. file: [889cba075c] check-in: [ede088a30d] user: js, branch: trunk, size: 21133 [annotate] [blame] [check-ins using] [diff]

ObjFW_RT used to be a separate project that followed the BSD style, as
it was written in pure C, while ObjFW's style is based on the BSD style
with changes to make it a better fit for Objective-C. This commit
changes ObjFW_RT to use the same style as ObjFW. file: [66376a6ccb] check-in: [ef6d69931e] user: js, branch: trunk, size: 20544 [annotate] [blame] [check-ins using] [diff]

A plugin which calls it might be loaded while already being
multi-threaded. file: [241b0f4d77] check-in: [1ac944d7c1] user: js, branch: trunk, size: 19581 [annotate] [blame] [check-ins using] [diff]

This currently uses __saveds on all exported functions for simplicity.
As an optimization, __saveds can be removed from some functions later
after careful verification that no code path starting from that function
accesses any data. file: [973ed6f5e5] check-in: [482698c5be] user: js, branch: trunk, size: 20233 [annotate] [blame] [check-ins using] [diff]

This results in stdio being linked in when building an Amiga library. file: [b4196ed858] check-in: [2ae3bc33e1] user: js, branch: trunk, size: 20084 [annotate] [blame] [check-ins using] [diff]

As we define the ABI, we can just replace BOOL with bool everywhere,
including in ObjFW itself. For the Apple platforms where BOOL and bool
are different, this is not a problem as BOOL and bool are passed and
returned the same way in the ABI.

This still defines BOOL to bool for compatibility, except on AmigaOS and
Wii, which both have its own BOOL type. file: [8659e5eda9] check-in: [fe2cbe0021] user: js, branch: trunk, size: 19487 [annotate] [blame] [check-ins using] [diff]

Only restore superclass to the name of the superclass if the class has
actually been set up. file: [16946c32d6] check-in: [8651ed04a0] user: js, branch: trunk, size: 19487 [annotate] [blame] [check-ins using] [diff]

This also renames ObjFW-RT to ObjFW_RT, as hyphens should not be in
framework names. file: [2ddf79d34d] check-in: [0f42b0eafd] user: js, branch: trunk, size: 19447 [annotate] [blame] [check-ins using] [diff]

This is required to create a .library on MorphOS. file: [de3e6ed813] check-in: [79011c5f56] user: js, branch: trunk, size: 19492 [annotate] [blame] [check-ins using] [diff]

Casts are now written like types in variable declarations. file: [bb8140f5e4] check-in: [4af49a13c3] user: js, branch: trunk, size: 19431 [annotate] [blame] [check-ins using] [diff]

Forgot to add 2017, even though I already did quite some changes in
2017. file: [a5e62b6d57] check-in: [44f45c2e35] user: js, branch: trunk, size: 19418 [annotate] [blame] [check-ins using] [diff]

Also moves out some stuff to functions to be reusable by
class_addMethod(). file: [c8cb26401f] check-in: [d4cae4c963] user: js, branch: trunk, size: 19328 [annotate] [blame] [check-ins using] [diff]

(U)INT_FAST*_MAX is broken on Android, which makes (u)int_fast*_t
dangerous to use while it does not provide any real speed benefit. file: [ecd4869918] check-in: [d3158d091f] user: js, branch: trunk, size: 18120 [annotate] [blame] [check-ins using] [diff]

While at it, also update the mail address. file: [2f6309b70c] check-in: [2a27cf3000] user: js, branch: trunk, size: 18140 [annotate] [blame] [check-ins using] [diff]

* It seems <limits.h> is required for SIZE_MAX.
* Use "exec ./tests" inside adb shell. It seems that either new versions
of ADB or Android don't return otherwise. file: [775c018011] check-in: [05a247fbe4] user: js, branch: trunk, size: 18130 [annotate] [blame] [check-ins using] [diff]

* dtable.m is now the old sparse array, using IMP as type for values and
thus not violating the C standard anymore (functions may not be stored
in void*)
* New sparsearray.m which can work with any size, based on the sparse
array from the fast path of class.m
* Fast path of class.m now uses the new sparsearray.m file: [79ebd88d66] check-in: [1ebb9eb7b3] user: js, branch: trunk, size: 18110 [annotate] [blame] [check-ins using] [diff]

It is too big already for inlining to make any sense. file: [ae710a7cad] check-in: [e23441b121] user: js, branch: trunk, size: 19237 [annotate] [blame] [check-ins using] [diff]

Also cleans up runtime/hashtable.m a little. file: [e257cf48b1] check-in: [428520105b] user: js, branch: trunk, size: 19276 [annotate] [blame] [check-ins using] [diff]

No functional change intended. file: [cf051d40ec] check-in: [9da9b2b64c] user: js, branch: trunk, size: 19093 [annotate] [blame] [check-ins using] [diff]

This matches the behaviour of the Apple runtime. file: [dff5c1b4fe] check-in: [52c0e9f94e] user: js, branch: trunk, size: 18038 [annotate] [blame] [check-ins using] [diff]

This was never used and is not part of the Apple runtime either. file: [3f90a22e50] check-in: [7c0fa1fa85] user: js, branch: trunk, size: 17709 [annotate] [blame] [check-ins using] [diff]

It behaves like in the Apple runtime now. file: [e18e9ca229] check-in: [3cd9d0ad69] user: js, branch: trunk, size: 17886 [annotate] [blame] [check-ins using] [diff]

This should improve the performance for the GCC ABI, as
objc_classname_to_class() is used for all sorts of class lookups, e.g.
objc_lookup_class().

As this performance improvement needs RAM, it is only used after 128
calls into objc_classname_to_class(), so that if the ObjFW ABI is used
and the user does not call into objc_getClass() or similar in a loop, no
memory is wasted.

Runtime internal usage of objc_classname_to_class() does not use the
fast path and does not count as a call into objc_classname_to_class().
The reason for this is that if the runtime calls
objc_classname_to_class(), it already has the lock and thus the
performance gain would be small, but it would waste memory. file: [e9b6e354b6] check-in: [84a724dd4b] user: js, branch: trunk, size: 15844 [annotate] [blame] [check-ins using] [diff]

This makes it possible to find aliases in the classes hashtable and not
call objc_free_class() on them. file: [5994e9e90d] check-in: [69d9f76520] user: js, branch: trunk, size: 12594 [annotate] [blame] [check-ins using] [diff]

The only places where BOOL is left are those where they are required by
the ABI. file: [d8be8e78d1] check-in: [c5ef582958] user: js, branch: trunk, size: 12447 [annotate] [blame] [check-ins using] [diff]

Inside object_{get,set}Class, a cast is used to get rid of the warning
now and all other direct isa references were removed. file: [640206d718] check-in: [2347371351] user: js, branch: trunk, size: 12439 [annotate] [blame] [check-ins using] [diff]

This makes it easier to use runtime functions in portable code. file: [72c72347b1] check-in: [120caad331] user: js, branch: trunk, size: 11993 [annotate] [blame] [check-ins using] [diff]

ERROR seems to be defined on MinGW32 and it's likely that it's also used
on other systems, thus it's better to prefix it with OBJC_. file: [0f8bf1f4e6] check-in: [345eb93240] user: js, branch: trunk, size: 11633 [annotate] [blame] [check-ins using] [diff]

For consistency. file: [db289f8cb1] check-in: [963db3089f] user: js, branch: runtime, size: 11574 [annotate] [blame] [check-ins using] [diff]

This is no longer done in objc_msg_lookup() and thus allows direct class
references. file: [16b932ec92] check-in: [dcf845546a] user: js, branch: runtime, size: 11576 [annotate] [blame] [check-ins using] [diff]

It's required by Clang for @compatibility_alias, even with the fragile
ABI. This might be a bug in Clang, as _np means non-portable and Clang
should not just emit references to such symbols. file: [886ace82f1] check-in: [8525f9600e] user: js, branch: runtime, size: 11088 [annotate] [blame] [check-ins using] [diff]

We now have OBJC_CLASS_INFO_{SETUP,LOADED,INITIALIZED}. file: [106999ef41] check-in: [5a719c9b1f] user: js, branch: runtime, size: 9487 [annotate] [blame] [check-ins using] [diff]

It's better to handle the class like it does not even exist in case it
is not completely loaded. Throwing an error there introcudes trouble
with static instances which try to get the class before initializing the
static instances, so objc_lookup_class() should never throw an error. file: [227edc0aad] check-in: [0e35ae63a5] user: js, branch: runtime, size: 9246 [annotate] [blame] [check-ins using] [diff]