I've guessed whether break or [[fallthrough]] is appropriate. In cases
where it looked particularly suspicious, I added a FIXME comment. All
of these changes should be reviewed by someone familiar with the code.
Moved MS DIB parser out of ICO file reader and made it available for
artwork and layout images.
Added more efficient I/O and better error checking for JPEG file loading
(MAME will no longer exit immediately on a bad JPEG file).
Made caller responsible for opening files for loading images, to avoid
decompressing images used in ZIP/7z artwork multiple times.
Added support for JPEG and Windows DIB to picture_image_device.
Added support for SVG image files in external artwork.
Added support for using I/O port value for animation state and masking
animation state values.
Made bounds elements more flexible in layouts.
Reworked headers to reduce dependencies.
Updated layout file format documentation.
The global_alloc/global_free functions have outlived their usefulness.
They don't allow consistently overriding the default memory allocation
behaviour because they aren't used consistently, and we don't have
standard library allocator wrappers for them that we'd need to use them
consistently with all the standard library containers we're using. If
you need to change the default allocator behaviour, you can override the
new/delete operators, and there are ways to get more fine-grained
control that way. We're already doing that to pre-fill memory in debug
builds.
Code was already starting to depend on global_alloc/global_free wrapping
new/delete. For example some parts of the code (including the UI and
Windows debugger) was putting the result of global_alloc in a
std::unique_ptr wrappers without custom deleters, and the SPU sound
device was assuming it could use global_free to release memory allocated
with operator new. There was also code misunderstanding the behaviour
of global_alloc, for example the GROM port cartridge code was checking
for nullptr when a failure will actually throw std::bad_alloc.
As well as substituting new/delete, I've made several things use smart
pointers to reduce the chance of leaks, and fixed a couple of leaks,
too.
* psixptty: Fix for FreeBSD
We need to check for __FreeBSD__ not __FreeBSD_kernel__
* FreeBSD: Use gl includes from pkg-config
FreeBSD doesn't install the includes from 3rdparty software in a default
location so use pkg-config to get the correct path.
* FreeBSD: genieos is named simple bsd
Made const-qualified pixel accessors (pix, pixt, raw_pixptr) return
const-qualified references/pointers to pixesl, and added non-const
versions. This makes bitmap more like standard library containers where
const protects the content as well as the dimensions.
Made the templated pixt accessor protected - having it public makes it
too easy to inadvertently get a pointer to the wrong location.
Removed the pix(8|16|32|64) accessors from the specific bitmaps. You
could only use the "correct" one anyway, and having the "incorrect" ones
available prevented explicit instantiations of the class template
because the static assertions would fail. You can still see the pixel
type in the bitmap class names, and you can't assign the result of
&pix(y, x) to the wrong kind of pointer without a cast.
Added fill member functions to the specific bitmap template, and added
a explicit instantiations. This allows the bitmap size check to be
skipped on most bitmap fills, although the clipping check is still
there. Also fixed a couple of places that were trying to fill an
indexed 16-bit bitmap with rgb_t::black() exposed by this (replaced with
zero to get the same net effect). The explicit template instantiations
in the .cpp file mean the compiler can inline the function if necessary,
but don't need to generate a local out-of-line body if it chooses not
to.
Extended the size of the fill value parameter in the base bitmap class
to 64 bits so it works correctly for 64-bit bitmaps.
Fixed places where IE15 and VGM visualiser weren't accounting for row
bytes potentially being larger than width.
Fixed an off-by-one in an HP-DIO card where it was treating the Topcat
cursor right edge as exclusive.
Updated everything to work with the API changes, reduced the scope of
many variables, added more const, and replaced a few fill/copy loops
with stuff from <algorithm>.
The previous behaviour was unintuitive - parsing an XML file and writing
it out immediately would produce invalid XML if the file contained any
characters that needed escaping. It makes far more sense to escape on
writing rather than expecting the user to escape input.
-Add preliminary support for visibility toggles to artwork system.
This allows the user to show/hide related elements in a view, with
nesting. The view can specify whether elements are shown or hidden by
default. Settings are saved per host window/screen per view. There is
no way to set the initial visibility state on the command line. Legacy
"Space Invaders cabinet model" layers are mapped onto visibility
toggles.
This is not stable yet. In particular, the XML element/attribute names
have not been finalised. The new features have not been added to
complay.py to prevent them from being used before they're finalised.
-----------------------
07536: [Graphics] Prescale option does not work properly on D3D renderer (Ryan Holtz)
-renderer/d3d: Removed old StretchRect code. All drivers these days punt it to a shader backend anyway, and it's causing issues with -prescale. [Ryan Holtz]
- Memory references in expressions no longer default to the console's visible CPU if no device name was specified, except when entered through the console itself. Expressions in view windows now use the context of the currently selected device instead.
- The pcatmem debug command and similar qt mouseover function now produce an error message if the initial address translation fails.
Related internal changes (nw)
- The debugger_cpu class no longer interprets memory accesses. The existing routines have been moved into symbol_table (which used to invoke them as callbacks), and reimplemented in most other places. Thecode duplication is a bit messy, but could be potentially improved in the future with new utility classes.
- The cheat engine no longer needs to hook into the debugger_cpu class or instantiate a dummy instance of it.
- The inclusion of debug/express.h within emu.h has been undone. Some debugging structures now need unique_ptr to wrap the resulting incomplete classes; hopefully the performance impact of this is negligible. Another direct consequence is that the breakpoint, watchpoint and registerpoint classes are no longer inside device_debug and have their own source file.
- The breakpoint list is now a std::multimap, using the addresses as keys to hopefully expedite lookup.
- The visible CPU pointer has been removed from the debugger_cpu class, being now considered a property of the console instead.
- Many minor bits of code have been simplified.
