This class is a thin wrapper around libicu regular expressions, so if we don't
have libicu we simply don't compile it at all. This will give people a linker
failure, rather than a nonfunctional class if they try to use GNUstep without
ICU with code that requires it.
The Apple documentation says that this class has a primitive method that takes
a block as an argument and that this method is called by others, so subclasses
can replace that block method without touching the convenience methods. We
mimic this behaviour when compiling with block, but when compiling without them
it's a problem. The current code contains some ugly hacks that will work in
normal usage but break with subclassing when not compiling with blocks.
This commit also includes a partial implementation of NSTextCheckingResult,
implementing the subset of its functionality required for NSRegularExpression
to work.
It also includes numerous fixes to GSICUString. This is heavily used by
NSRegularExpression, to avoid copying strings when mapping between UText for
libicu and NSString for GNUstep.
Note: I don't have a copy of iOS anywhere to test this against, so it's
entirely possible that there are significant discrepancies between this
implementation of NSRegularExpression and the iOS version. This version should
function exactly as the iOS one is described as functioning, but I think we've
all seen that Apple documentation refers more to hopes than facts. Any testing
that someone who does have an ip{hone,od,ad} can do is very welcome.
git-svn-id: svn+ssh://svn.gna.org/svn/gnustep/libs/base/trunk@31635 72102866-910b-0410-8b05-ffd578937521
Note: I'm fairly sure that all of the #import lines in Foundation.h are wrong; we should be using "foo.h" instead of <Foundation/foo.h>, or we will pick up the installed version when including this file while building GNUstep. I don't want to make this change in case it is intentional to work around some broken ObjC compiler somewhere though...
git-svn-id: svn+ssh://svn.gna.org/svn/gnustep/libs/base/trunk@31595 72102866-910b-0410-8b05-ffd578937521
Most of these changes involve simply removing direct manipulation of any runtime structures and replacing them with runtime function calls. For example class->name becomes class_getName(class) and so on.
libobjc2, like the Apple runtime, the NeXT runtime, and every version of the Objective-C spec, calls the class pointer in id isa. A few files now have #define class_pointer isa at the top. This line replaces class_pointer in the old GNU libobjc headers with isa so either class_pointer or isa can be used for accessing the class of an object. Note: object_getClass() should be used in most cases because, in some future version of the runtime, this will skip things like lock classes and other hidden classes (e.g. KVO classes).
All of the old forwarding stuff has been removed. Most of this stuff followed convoluted code paths that ended with an exception. A few simply broke in exciting ways. Hopefully no one has used them for the last ten years or so, but we can bring them back with some #ifndef stuff if they're really needed by anyone.
There is currently a bug in configure, which prevents dladdr() from being detected, so you need to manually tweak config.h to build - I have not fixed the fall-back code in objc-load.m to work with libobjc2, I just added a new version that uses the loader's functionality directly.
Although -base now builds, it builds with a lot of warnings. <string.h> is missing from a lot of files, so memcpy() and strlen() generate implicit function declaration warnings.
Additionally, libobjc2 does still provide the sel_{get,register}_*() functions, but they're wrappers around the newer API ones. These are deprecated and are not exposed in the headers. Although they work, we should be replacing them with the libobjc2 versions as soon as possible.
This incorporates a patch by Eric.
git-svn-id: svn+ssh://svn.gna.org/svn/gnustep/libs/base/trunk@31265 72102866-910b-0410-8b05-ffd578937521
never build the ICU stuff because the corresponding macro wasn't defined).
git-svn-id: svn+ssh://svn.gna.org/svn/gnustep/libs/base/trunk@30995 72102866-910b-0410-8b05-ffd578937521
so people can safely write subclasses using different allocation schemes.
git-svn-id: svn+ssh://svn.gna.org/svn/gnustep/libs/base/trunk@29847 72102866-910b-0410-8b05-ffd578937521
The -pointerValue method now returns the value cast to a pointer, not some random value, as the documentation says it should. This is a change from OpenStep, which said:
> It's an error to send this message to an NSValue that doesn't store a pointer.
The OS X docs now say:
> The receiver's value as a pointer to void. If the receiver was not created to hold a pointer-sized data item, the result is undefined.
This means that any NSNumber created with a word-sized integer should return the same value.
Fixed a number of corner-cases in the compare: implementation caused by incorrect type promotion. The OS X docs say:
> The compare: method follows the standard C rules for type conversion.
The OS X implementation does not do this. We now match Apple's conversion rules bug-for-bug: Every value is stored in the smallest signed type that will hold it, unless there is no unsigned type that can hold it, in which case it is stored in an `unsigned long long`, comparisons between integer and floating point values cast both to a double, comparisons between integer types perform a real comparison (so an unsigned long long is always greater than any negative number, at any precision). The Apple implementation is actually quite sane, it is just completely unrelated to the documentation in any way.
We now use the same range of reusable objects. Note that there is an error in Cocoa Design Patterns in the description of how Apple's implementation works. Do not use this as a reference.
We now return `nil` when an NSNumber is sent an -init message. This is consistent with Apple's implementation but breaks some things in the GNUstep test suite (which RFM said he will fix).
There is a small change in NSValue.h so that the locale parameter is now an `id` not an `NSString*`. This is because, under recent OS X, it may also be an `NSLocale` instance. I am not sure how much GNUstep supports `NSLocale`, but this change shouldn't affect anything.
The new (private) GSNumberTypes.h file lets you define macros that are instantiated with each of the names of primitive C types. These might be useful for simplifying other classes that have -intValue, -floatValue, and so on methods, such as the `NSCell` family.
The old NSConcreteNumberTemplate and NSConcreteNumber stuff has been removed. The code is now a bit more than 10% of the size of the old NSNumber code, and is hopefully maintainable now, so the next change won't require a complete rewrite.
git-svn-id: svn+ssh://svn.gna.org/svn/gnustep/libs/base/trunk@29618 72102866-910b-0410-8b05-ffd578937521
Also check that NSINTEGER_DEFINED and CGFLOAT_DEFINED are undefined
before defining NSInterger and CGFloat.
git-svn-id: svn+ssh://svn.gna.org/svn/gnustep/libs/base/trunk@29351 72102866-910b-0410-8b05-ffd578937521
As with GSArray, GSSet uses its isa pointer for detecting mutations. This may change as a result of adding KVO notifications, so it might not be the best solution, but I can't currently think of a way we could catch isa changing to [GSMutableSet class] and not changing to a hidden class...
git-svn-id: svn+ssh://svn.gna.org/svn/gnustep/libs/base/trunk@29179 72102866-910b-0410-8b05-ffd578937521
Some of these implementations are not as efficient as they could be (especially the ones that take an NSIndexSet as the first argument). They also don't yet support concurrent enumeration. Apple implements these using Grand Central. We could possibly have a background thread that we send these things to (or use GCD if libdispatch is available). It's not worth spawning a new thread for them, except in exceptional circumstances (and, unfortunately, we can't easily tell how expensive a single iteration of a block is. Possibly we could time one block invocation, and if it's longer than some threshold make it concurrent, but it's probably not worth the effort).
git-svn-id: svn+ssh://svn.gna.org/svn/gnustep/libs/base/trunk@29176 72102866-910b-0410-8b05-ffd578937521
backtrace_symbols() code. Implemented the -callStackSymbols method from
10.5 using this. For this to be enabled, the configure script will
require a small modification, which Gregory will add later.
git-svn-id: svn+ssh://svn.gna.org/svn/gnustep/libs/base/trunk@28662 72102866-910b-0410-8b05-ffd578937521
NSObject up to feature-parity with the OS X 10.5 implementation when using the
new runtime and up to feature-parity with the 10.6 implementation if you are
using the new runtime and compiling with clang.
Also removes the objc_mutex_wibble stuff from NSObject in favour of just using
NSLocks (which, with the new implementation, are now faster than using
objc_mutex_stuff).
git-svn-id: svn+ssh://svn.gna.org/svn/gnustep/libs/base/trunk@28657 72102866-910b-0410-8b05-ffd578937521
* Headers/Foundation/NSLock.h
Completely rewritten implementations of NSLock.h classes. These are now
faster, more complete, OS X-compatible, and most importantly actually
work. The old ones, for example, called functions that were not
implemented on Windows.
* Source/NSThread.m
Call pthread functions directly in NSThread instead of via the libobjc
abstraction layer. Also fixed a few issues, such as GC not being
initialized properly for NSThread subclasses that override -main (Javaism
supported by OS X) and tidies up the code in several places, removing
premature optimizations, especially those that introduce a test for an
unlikely case at the start of a method and thus slow everything down.
As a result of this change, GNUstep now depends on an implementation of
POSIX threads. This is included as standard on all modern UNIX systems,
and as an option on less-modern UNIX systems and non-UNIX systems,
including Windows. If you are building GNUstep on Windows, please install
the pthreads-win32 package, available from:
http://sourceware.org/pthreads-win32/
PLEASE TEST THIS! There may be some code that depended on the old
behaviour. I have been running the new NSLock implementation on FreeBSD
for a few weeks without issue; please report to me any problems that you
have on your platform.
git-svn-id: svn+ssh://svn.gna.org/svn/gnustep/libs/base/trunk@28598 72102866-910b-0410-8b05-ffd578937521
