Jeremiah Morris 2009-12-09 16:23:50 PST
Re-enable mouseLocation workaround on 10.4, 10.5
OS X systems before 10.6 have a bug with [NSEvent mouseLocation] if the screen
resolution changes. SDL_QuartzVideo.m contains a workaround for this bug, but
it was placed inside an #ifdef in revision 4762. The comment reads, "I'm
gambling they fixed this by 10.4."
After seeing this bug on several Tiger and Leopard systems (both PPC and
Intel), I can confirm that it's not fixed until 10.6. The workaround doesn't
compile for x86_64/10.6, so I can understand why it was segregated, but it
needs to remain in place for the i386 and ppc versions. The workaround causes
no problems under 10.6, even though it's not necessary there.
Patch is attached with one method of keeping the workaround active on the
systems that need it.
--HG--
branch : SDL-1.2
John Popplewell 2009-12-08 23:05:50 PST
Originally reported by AKFoerster on the mailing list.
Error decoding UTF8 Russian text to UTF-16LE on Windows, but specifically on
platforms without iconv support (the default on Windows).
Valid UTF8 characters are flagged as being overlong and then substituted by the
UNKNOWN_UNICODE character.
After studying the testiconv.c example program, reading the RFCs and putting
some printf statements in SDL_iconv.c the problem is in a test for 'Maximum
overlong sequences', specifically 4.2.1, which is carried out by the following
code:
} else if ( p[0] >= 0xC0 ) {
if ( (p[0] & 0xE0) != 0xC0 ) {
/* Skip illegal sequences
return SDL_ICONV_EILSEQ;
*/
ch = UNKNOWN_UNICODE;
} else {
if ( (p[0] & 0xCE) == 0xC0 ) { <<<<<<<< here
overlong = SDL_TRUE;
}
ch = (Uint32)(p[0] & 0x1F);
left = 1;
}
} else {
Here is the 2-byte encoding of a character in range 00000080 - 000007FF
110xxxxx 10xxxxxx
The line in question is supposed to be checking for an overlong sequence which
would be less than
11000001 10111111
which should be represented as a single byte.
BUT, the mask value (0xCE) is wrong, it isn't checking the top-most bit:
11000001 value
11001110 mask (incorrect)
^
and should be (0xDE):
11000001 value
11011110 mask (correct)
making the above code:
} else if ( p[0] >= 0xC0 ) {
if ( (p[0] & 0xE0) != 0xC0 ) {
/* Skip illegal sequences
return SDL_ICONV_EILSEQ;
*/
ch = UNKNOWN_UNICODE;
} else {
if ( (p[0] & 0xDE) == 0xC0 ) { <<<<<<<< here
overlong = SDL_TRUE;
}
ch = (Uint32)(p[0] & 0x1F);
left = 1;
}
} else {
I can supply a test program and/or a patch if required,
best regards,
John Popplewell
--HG--
branch : SDL-1.2
The OS/2 joystick code included an IBM header and was in violation of the license.
Ryan may put this back someday with a legal header. :)
--HG--
branch : SDL-1.2
Ronald Lamprecht to SDL
Hi,
Sam Lantinga wrote:
The problem with that fix is that it breaks IME events again. Maybe
we can handle keyboard events differently to prevent this issue?
Spending an hour reading MSDN, analysing SDL and another hour testing the reality on XP I am really wondering how patch r4990 could have ever worked in any situation. It's main effect is to break the unicode translation and causing spurious activation events!
Why does TranslateMessage(&msg) nothing useful? Simply because it does not affect "msg" at all! All keyboard events are dispatched without the slightest change (see MSDN). TranslateMessage() just appends additional WM_CHAR, WM_DEADCHAR, WM_SYSCHAR, WM_SYSDEADCHAR event messages to the queue. But I could not find any SDL event handling routine that catches these events and transforms them to proper SDL keyevents while eliminating the corresponding WM_KEYDOWN, etc. events. Thus any IME input like the '@' generated by "Alt + 6(Numpad) + 4(Numpad)" is simply lost.
But the situation is even worse! Up to r4990 the TranslateKey()/ToUnicode() calls did evaluate dead keys and did deliver proper key events for subsequent key strokes like '´' + 'e' resulting in 'é'. ToUnicode() needs proper key state informations to be able to handle these substitutions. But unfortunatly TranslateMessage() needs the same state information and eats it up while generating the WM_CHAR messages :-( Thus the current 1.2.14 breakes the partial IME support of previous releases, too.
The key state race condition between ToUnicode() and TranslateMessage() requires to avoid any ToUnicode() usage for receiving proper WM_CHAR, etc. messages generated by TranslateMessage(). (Yes - the '@' and 'é' appear as WM_CHAR messages when unicode is switched off).
The spurious SDL activation events are *not* caused by additional WM_ACTIVATE Windows messages! Besides DIB_HandleMessage() SDL_PrivateAppActive() is called by another source which I am not yet aware of - any hints?
Thus I do strongly recommend the deletion of the TranslateMessage(&msg) call as a quick fix.
A proper support of unicode and IME requires a clean SDL keyboard input concept first. Which SDL keyboards events should be transmitted to the app when the user presses '´' + 'e' ? Within the current unicode handling the first key stroke is hidden. Even though ToUnicode() delivers the proper key SDL does ignore it in TranslateKey(). Just the composed key event is transmitted to the app. That is what you expect for text input, but the app can no longer use keys like '^' as a key button because it will never receive a key event for it!
With a given concept it seems to be necessary to regenerate SDL key events out of the WM_CHAR, etc. events and to drop all related direct WM_KEYDOWN, etc. events while the remaining basic WM_KEYDOWN, etc. events would still have to result in SDL key events.
Anyway the source of the spurious WM_ACTIVATE should be located to avoid future trouble.
Greets,
Ronald
--HG--
branch : SDL-1.2
Another one for the "How did this ever work?" file: when
DX5_HandleMessage is called with WM_ACTIVATEAPP, it goes past the end
of the 2-element SDL_DIdev array and if there doesn't happen to be a 0
in the memory next to it, crashes. Patch against SVN attached.
--HG--
branch : SDL-1.2
The compiler complains that long is an incompatible pointer type for
[gl_context setValues: forParameter:] and CGLSetParameter. This only shows up
on x86_64, but it still might be an issue.
--HG--
branch : SDL-1.2
C.W. Betts 2009-10-19 19:16:41 PDT
In 64-bit libraries, the fork reference is a different size than in 32-bit.
Apple has tried to help this transition by making a typedef called FSIORefNum,
but it is only available in Mac OS X 10.5 headers or later. I first noticed
this bug in Mac OS X's CD player code, and the patch is set to fix it
accordingly. This might also be an issue in SDL_QuartzVideo.m
--HG--
branch : SDL-1.2
I guess I don't have to mention that, but ALSA_WaitAudio() is
pure evil, and just broken. getpid() returns pids, not tids. You get
those via gettid(). But I'd recommend dropping that entirely. It's
just broken.
--HG--
branch : SDL-1.2
In ALSA_OpenAudio(): instead of setting period_size+n_periods OR
buffer_size I'd recommend copying the hwparams stuff before you do
this, then first try period_size+n_periods, and then apply it with
snd_pcm_hw_params() and check if that works. If it didn't you should
take the copy of hwparams and try setting buffer_size and apply that
via snd_pcm_hw_params() and check if that worked. And if that failed
too, then take the copy and don't apply neither period nor buffer
settings and see if that works.
--HG--
branch : SDL-1.2
Ludwig Nussel 2009-10-18 06:31:52 PDT
an mprotect call was added to fix bug 528. However that results in a buffer
that allows writing and code execution. Ie the no-execute security features of
modern operating systems are defeated this way. Two mprotect calls are needed.
One to make the buffer executable but not writeable when done and another one
to make the buffer writeable again if the content needs to be changed.
--HG--
branch : SDL-1.2
Ludwig Nussel 2009-10-18 05:34:18 PDT
src/joystick/linux/SDL_sysjoystick.c has some problems:
- test_bit() might break with strict aliasing
- test_bit() assumes array is Uint32 but its actually "unsigned long"
on 64bit systems sizeof(long) != sizeof(Uint32).
- the keybit array is too small
- the arrays are unitialized so the number of
detected buttons is quite random
--HG--
branch : SDL-1.2
For some people setting the period size works better (and is what SDL 1.2.13 did), but for most people it's the same or worse. You can use an environment variable to pick which one you want.
--HG--
branch : SDL-1.2
Fix assertion triggered in pulsaudio by returning "unknown" instead
of NULL pointer when guessing application name.
- Use getprogname on NetBSD as it wont return NULL
--HG--
branch : SDL-1.2
Apply patch to first unlock the CD before trying to eject it. Since
SDL doesn't seem to have lock/unlock primitives, this looks like the
right thing to do.
Fixes audio/tcd's eject function too.
--HG--
branch : SDL-1.2
