The scroll bar always spans the content area of the screen: the
position and height depend only on the theme and not on who is
drawing it.
Note that the coordinates passed were wrong in most cases, so this
commit fixes the scroll bar positioning for several dialogs.
If the application in question daemonizes, it will continue running
no matter whether we start it with system() or execlp(). So I don't
see a reason for this feature to exist and removing it means less
code paths to worry about.
The labels were longer than the space before the RGBA controls and
the fact that these are colors is already clear from the context
(such as having an RGBA control after it ;).
I tried to update the translated versions of these labels as well.
However, since I don't speak most of these languages, it is possible
the result is grammatically incorrect. If this is the case, please
mail me a correction.
It was only used to share implementation between IconButton and Link.
However, there was only one non-trivial method, handleTS(), and that
method used a different code path for each use case: doubleClick was
always false for IconButton and always true for Link. So the total
amount of code was actually reduced by eliminating this code sharing.
The main motivation for this split is that I can now freely refactor
Link without having to worry about IconButton.
Each part of the code deals with either Links or IconButtons, but
not both: the base class is only used to share implementation and
not interface. Make this explicit by doing private inheritance.
Nowhere in the code do we actually mix IconButtons and Links (the other
Button subclass), so I'm thinking of breaking up this class hierarchy
or at least making the inheritance private.
Also switched to C++11 style loops.
Previously, the links would scroll when the cursor was about to move
out of screen. By scrolling earlier, the user gets a view of the next
row before it becomes the current row. This allows a longer reaction
time to switch from vertical to horizontal navigation when looking for
a particular link in the grid.
This reduces the number of required preprocessor directives, leading to
more readable code and more code being examined by the compiler (useful
to spot problems). Since the method is inlined, the compiler should be
able to eliminate the same amount of code that the preprocessor would,
only at a later stage of the compilation.
This fixes a bug with the captured background being wrong when using
double buffering. Also it ensures that for example the clock in the
status bar is updated when the context menu is open.
All of the entries in the context menu affect sections and links, so
the context menu should be considered part of the main menu, not of
the global / background context.
Writing hours and minutes separately while the string representation
is being constructed could lead to an incorrect result on an hour
boundary. Avoid this by using an atomic timestamp.
With GCC 4.8.1 on MIPS this atomic timestamp is lock free.
SDL does not guard against the final callback still executing when the
call to remove the timer returns. So we have to make sure the object
we access on the callback is not prematurely destructed. This commit
uses shared_ptr and weak_ptr to ensure that.
Note that we sort-of have a singleton again, only now it is private
and safely destructed.
Removed the suspend check: the best thing we can do after oversleeping
is the same as when we're woken right on time: fetch the time and
reschedule for the next minute boundary.
Don't create a new timer on every callback; instead return the next
interval to SDL.
Don't make Clock a singleton. While there should be no reason to
instantiate this class more than once, there is no problem with doing
that either. Removing the singleton makes it easier to control access
to the instance. It also avoids the rather nasty construct that was
used to delete it.
Make sure the timer callback function is a proper C function, since
SDL is a C library. This requires some trickery to be able to call
a private method from the callback, but I found a way using an
intermediate nested class. The compiler should be able to inline this
to eliminate any overhead.
Also some minor cleanups.
This method was never called.
And I cannot really think of a scenario in which it is useful to wait
for any button to be released: a particular button or all buttons I can
imagine, but not any button.
Instead of having a list and wrapping between beginning and end, always
put the current section in the middle and show the previous and next
sections using wrap-around.
Made sure it fills the space between the top and bottom bar. Use a one
pixel margin; I tried without margins but it didn't look nice.
Also cleaned up the paint code a bit.
Replaced Font constructor with factory method, so that if the TTF
cannot be loaded, the Font object is not constructed. The normal C++
way of handling this is with exceptions, but we're compiling with
-fno-exceptions.
Originally the font implementation was based on SFont, but it was
recently replaced by an SDL_ttf based implementation, so the name
no longer made sense.
The code still has a lot of overlap with the other methods of Link and
LinkApp, but at least it is in the same place now.
Since this was the last outside use, setIconPath() could be declared
as 'protected'.
This could be used to create platform-independent packages. It can
be useful for instance in the case where the executable is a script,
or when the package points to an executable that is present on all
systems on which OPKs are supposed to run.
The empty "" filter was previously matching all the files; since
a few commits it now matches the files without extensions. The
new "*" filter allows again to match all files.
tolower() will trigger an assertion failure in the case where the
string contains UTF-8 codes. This is not a problem as only the
file extension needs to be processed, and that one should contain
only ASCII.
These are not actual OPK files, but metadata (I assume) storage that
Mac OS X adds to vfat file systems. The current version of libopk has
extremely poor error handling, so ignoring these files avoids a crash.
But even when libopk is fixed, not trying to open these files as OPKs
will be useful since it saves time.
When a app is launched with a file selected with the built-in file
selector, we memorize the directory on which the file was found,
so that a future launch of that application will directly open
that directory.
The filename passed to it should contain the filename of the squashfs
image, the character '#', and the filename of the .png file contained
inside the package.
E.g.:
/tmp/package.opk#icon.png
Console apps are always launched on tty1. However it can occur that
the console displayed on the framebuffer is another one, so we have
to explicitely activate tty1.
Use unsigned ints for width and height, like SDL does.
We're using the SDL typedefs there after all.
Include argument names in header when type name does tell everything.
Use whitespace in argument lists.
Fixes bug introduced in 57ad81e3df.
This bug would prevent applications from launching, so it's pretty serious.
I did test before committing, but apparently I didn't test the right binary.
LinkApp now calls its launch method via the delegate mechanism instead of
via an overridden virtual method. This should make it easier to move the
launch method out of LinkApp and into the GMenu2X main class.
This functionality was half-removed already: some code missing, other
essential parts commented out.
I don't see the point of this either: RGB graphics are created for a certain
gamma value; if that differs from the screen gamma it is up to the application
to do something about it: gamma correct the graphics and/or change the screen
gamma. It doesn't make sense to make this user configurable.
Working directory was only configurable in the ini files, not in the GUI.
Many GP2X/Dingoo applications break if the working directory is changed
because they use relative paths for locating their data files.
I don't see a reason to keep this feature, since I don't know of a realistic
scenario where it would be useful. In any case, the functionality can be
re-created by using a wrapper script to change the working directory based on
a parameter value.
Initialize in constructor and clean up in destructor instead of having separate init() and deinit() methods.
Don't close file descriptor if open failed (ts_fd == -1).
Renamed initialized() to available(), since it tests whether the touch screen was found, not whether an initialization was attempted.
Improved code layout and minor other cleanups.
There is no need for the overhead of std::vector if we know the number
of elements at compile time.
This shaves off about 2 kB of the stripped binary size.
Do not test whether config file exists; just open it and handle failure.
Config file was not closed on errors, fixed now.
Make sure unconfigured entries of the button map are properly initialized.
Ignore lines with invalid keywords instead of aborting parsing.
Made log messages more specific.
The desired backlight level depends on the amount of light coming in from
the environment, not on the application that is run. Setting the brightness
when launching an application discards what is probably the right value
for this location (indoor/outdoor) and time of day. So we decided that this
is not a desirable feature and therefore removed it.
Instead of having fscanf() read into a string buffer (with the risk of
overflowing it if sysfs returns an unexpectedly long string) and converting
the value to integer with atoi() afterwards, just use fscanf() to read
an integer.
It is already constrained on editing and when actually applying the frequency.
The string value constrain is problematic because it happens in a class that
has no knowledge of the actual CPU limits, so it can only contrain to
hardcoded limits.
With this change you'll actually be able to run apps at 432 MHz.
The maximum frequency for applications was set to 430 MHz before,
while some Dingoos can handle higher frequencies. Now the limit is set to
the theoretical maximum frequency of 500 MHz.
For the menu itself the maximum frequency is set to a safer 408 MHz,
to avoid over-eager users from putting their system in an unusable state
that is difficult to recover from.
Also the frequency settings are now done in 24 MHz increments, to match
the limitations of the current JZ4740 kernel. Note that only new entries
are rounded to 24 MHz multiples, existing entries stay as they are.
There was a lot of gp2x-specific code which was built for
all platform. For instance, the code was compiled for all
the platforms with the constant "TARGET_GP2X" defined.
This obviously had to be fixed.
This reverts commit 439d25c292.
Refusing to load files of the Default skin for a different one is a bad idea; it requires
that skin to provide every single PNG file needed by gmenu2x.
It ensures that GMenu2X can restart after a SDL program segfaulted or crashed.
It also allow GMenu2X to exit without clearing the screen, so that the "loading" screen appears
until the launched binary starts using the framebuffer.
writeSkinConfig() had a very naughty bug when RGBA values were written
to config file unpadded. This lead to a problem that, for example,
0x00000080 was written as 0x00080 in config file (Somehow gmenu2x
reads and re-saves config on selection). This, in turn, led
to problem when reading skin config file in strtorgba function,
because it parses the color string in token of two symbols
(substr(0,2) and etc).
This eliminates the build-time dependency on SDL_image and the run-time
dependencies on all libraries used by SDL_image except SDL and libpng.
In addition we now let libpng convert to ARGB format while decoding the
image, rather than letting SDL convert the surface afterwards.
A bug in SDL_ConvertSurface() leaves the per-surface alpha undefined when
converting from RGBA to RGBA. This can cause problems if the surface is
later converted to a format without an alpha channel, such as the display
format.
Instead of having the copy constructor convert to display format, the new
copy constructor preserves the pixel format and a separate method was
introduced to convert surfaces to display format.
The code was made more robust as well: it should no longer crash if the
wallpaper cannot be loaded.
I cheated a bit by declaring ASFont as friend, but all other outside
access now happens via methods.
I removed the "saveScreenshot" method since the code calling it is
commented out and I never heard anyone complain about missing this
feature.
If loading fails, the factory method returns NULL, while previously the
constructor would create a Surface object with a NULL "raw" field.
However, since most of the methods dereference "raw" without checking,
such a Surface would likely crash the application when used.
Use real double buffering instead.
I checked the SDL code and if the hardware cannot provide double buffering
it will use a shadow surface to ensure that a frame is not displayed until
it has been fully painted.
Also disable mouse cursor before opening the output surface. The reason it
was disabled after the surface was opened is that SDL on GP2X has a bug.
However, this means the cursor is visible for a short time during startup
which looks ugly.
The unnecessary alpha channel might harm performance but is not a big deal.
However, the fact that the alpha flag (bool) is somehow automatically
converted by C++ into a skin name (string) causes the entire screen to
stay black.
The black screen was fixed by removing the alpha flag.
For optimum blitting performance, the wallpaper surface is then converted
to the pixel format of the frame buffer.
Bug was introduced in 52c89d6005.
Before this commit loadPNG() could return any surface format and the caller
was responsible for converting it to the desired format. However, in practice
all callers want a surface with an alpha channel and SDL only supports that
with 32bpp surfaces, so RGBA or a permutation thereof. So I changed the
contract for loadPNG() so it is required to return an RGBA surface, and
removed the conversion code in the callers.
The next step is to replace IMG_Load() by a function that calls libpng
directly and loads a fixed 32bpp pixel format. That way, we can drop the
SDL_image dependency and avoid unnecessary pixel format conversions.
Renamed methods that draw a single line from write() to writeLine().
There is now only one write() method left: the public method.
Pass surface to draw on as wrapped Surface instead of SDL_Surface.
At the end of the call chain we still use SDL directly though.
Removed all methods that are never called.
Made methods that are only called by Surface itself private.
One ugly thing remaining is outside access to the "raw" field.
