The stream returned by a call to ResourceManager::getVolumeFile
either MUST (when returning an I/O stream from a Common::FSNode)
or must NOT (when returning a Common::File *) be deleted by the
caller, depending upon some internal implementation details of
ResourceSource that should never have been exposed to callers.
FSNode streams that should have been deleted were not being
deleted all the time, which leaked and eventually caused ScummVM
to run out of FDs.
This commit improves this situation by shielding callers from
these internal details by centralizing the destruction logic in
one place, so FSNode read streams stop being leaked and callers
no longer need to know stuff about the internals of the
ResourceSource they are trying to read in order to avoid leaking
or breaking the volume file cache.
Fixes Trac#9782.
Save game metadata validity checks in SCI32 should all exist within
kCheckSaveGame32 since this allows most games to recover
successfully from an attempt to load an invalid save game. If
gamestate_restore fails, the game will usually crash because the
engine is left in an inconsistent state (game scripts have cleaned
up objects in preparation for a game load that is no longer
happening).
This "fix" is more of a hack, in the interest of making the game
completable. The root cause is a combination of two problems in
the game scripts:
1. Blink::init expects to receive either 0 or 2 arguments, but
it assumes that if it received *any* arguments, it must have
received 2 arguments. This assumption is wrong, though,
because--
2. soTorinWhoAreYou::changeState(0) calls
poPecandEyes::setCycle(Blink) without including a second
argument (the blink speed).
This ends up with the second parameter being some garbage, and
that garbage gets sent to kRandom which then complains about
receiving garbage.
The correct fix for this would be to fix soTorinWhoAreYou (in
script 51400) to pass a second argument to setCycle, but there are
not enough obvious spare bytes for a quick and easy patch, so this
workaround will have to do for now.
Fixes Trac#9779.
The runtime code for this had previously relied on hot patching
volume file offsets at the moment that a resource was loaded,
instead of correcting file offsets when reading audio maps. The
code added for sanity checking audio volumes started to report
warnings because the offsets being received were for the original
uncompressed audio volume, which (naturally) is larger than the
compressed audio volume.
This commit also deduplicates code between addResource and
updateResource, and tweaks a validation-related error message for
improved clarity.
Fixes Trac#9764.
While earlier commits had fixed handling of audio resources in
audio bundles to match SSCI, audio *patches* in SCI16 were still
being treated like standard resource patches. They are now
special-cased in the resource manager, just like SCI32.
Incidentally, while fixing the problem with audio patches, I also
noticed that the patch resource fixes for SQ5/German were very
similar to the special-case operations for resources in SCI32,
though using an odd heuristic. After investigating, it appears
that Sierra SCI1.1 works mostly like SCI32, and not like what
was there from SCI View. So, the old special-case code is deleted
and the special-case code for SCI32 has been expanded to cover
SCI1.1. (SSCI prior to 1.1 do not appear to do this
special-casing.)
Fixes Trac#9773.
1. The chunk number was hard-coded to zero and inaccessible.
2. Running ResourceManager::getVolumeFile for a chunk resource
would always return nullptr instead of a stream of the chunk,
which made it impossible to generically validate that resources
being added were within bounds of the container file (or, in
this case, container chunk).
Simple assertions in the resource manager are not sufficient to
track down resource corruption issues, and some error conditions
that were being checked already were either ignored or only raised
as warnings that casual users would be unlikely to see.
Ideally, error handling in ResourceManager would be improved to
the point where errors would correctly propagate out of it (so the
warning dialogue could be displayed from outside), but right now
error codes are dropped all over the place, and it would take more
effort to fix that without much benefit for the current situation.
If/until someone has the energy to fix that, the warning dialogue
is simply shown from ResourceManager::scanNewSources.
Refs Trac#9764.
Updating the cursor's texture is not necessary if it is not going to be drawn.
Fixes glDrawArrays sometimes failing due to using a framebuffer with an
incomplete color attachment. In SCI32 games, the framebuffer is incomplete
because the engine does not define pixel data for the cursor.
In SSCI, mouse events are received through a hardware interrupt
and the cursor is drawn directly to the graphics card by the
interrupt handler. This allows game scripts to omit calls to
kFrameOut without stopping the mouse cursor from being redrawn
in response to mouse movement.
ScummVM, in contrast, needs to poll for events and submit screen
updates explicitly from the main thread. Submitting screen updates
may block on vsync, which means that this call should really only
be made once per frame, just after the game has finished updating
its back buffer. The closest signal in SCI32 for having completed
drawing a frame is the kFrameOut call, so this is where the
update is submitted (by calling OSystem::updateScreen).
The problem with the approach in ScummVM is that, even though the
mouse position is being updated (by calls to kGetEvent) and drawn
to the backend's back buffer (by GfxCursor32::drawToHardware),
OSystem::updateScreen is never called during game loops that omit
calls to kFrameOut.
This commit introduces a workaround that looks at the number of
times kGetEvent is called between calls to kFrameOut. If the
number of kGetEvent calls is higher than usual (where "usual"
seems to be 0 or 1), we assume that the game is running one of
these tight event loops, and kGetEvent starts calling
OSystem::updateScreen until the next kFrameOut call. We also
then start throttling the calls to kGetEvent to keep CPU usage
down. This fixes at least two such known loops:
1. When interacting with the menu bar at the top of the screen
in LSL6hires;
2. When restoring a game in Phant2 and sitting on the "click mouse"
screen.
A similar workaround may also be needed for kGetTime, though loops
around kGetTime should preferably be replaced using a script patch
to call kWait instead.
This may be masking another bug, but at least what happens in
Phant1 is that the mouse does not get redrawn after being moved
programmatically to the fast-forward button when deviceMoved
returns early.
Videos in GK2 use this flag (e.g. the chapter 6 intro).
Now that GfxPalette32::updateHardware no longer calls
OSystem::updateScreen (only GfxFrameout::frameOut does this now),
every time a palette swap occurs during playback, there is a frame
of blackness that should not exist. This is because the order of
operation is:
1. Send black palette
2. Call frameOut (which updates the screen)
3. Send new, correct palette
4. No frameOut (so the screen is not updated with the correct
palette)
OSystem::updateScreen cannot be called multiple times for the same
frame due to vsync, but also, there does not appear to be any
reason to send a black palette, since it seems to be intended to
avoid temporarily rendering video frames with the wrong palette on
a hardware device that cannot guarantee simultaneous application
of a new palette and new pixel data. ScummVM does not have such
a problem, so this feature appears to be unnecessary for us.
For the moment, this 'feature' remains hidden behind an ifdef,
instead of being removed entirely, to avoid potential confusion
when comparing VMD code from SSCI.
More than one call to OSystem::updateScreen per frame on systems
with vsync ruins performance because the call is blocked until
the next vsync interval.
This also fixes bad rendering performance with the OpenGL backend.
The approach to video benchmarking used by SCI engine does not
translate very well to modern video devices -- it will either be
so slow that the games think the system is not capable of showing
normal visual effects, or so fast that the benchmarks overflow
their counters. So, game scripts that perform video benchmarking
are now patched to unconditionally return the highest speed value.
A pleasant but subtle side-effect of this change is that the extra
time sitting at a blank screen before the start of a game (while
benchmarks ran) is now gone.
Fixes Trac#9741.
There were several issues.
The first one was introduced recently and caused the preferred target
to be used as a game ID, which resulted in an error when this is not
a valid game ID. Thus this fixes bug #9754.
The other issues were here since the auto-detect command was added and
caused other command line options, suh as the path, to be lost. This
usually resulted in a failure to start the game as the data files could
not be found (unless the ID happened to be the same name as a target
previously added). This also caused a reappearance of the old bug
