When erasing and inserting many items this caused the hashmap capacity
to grow more than it should which resulted in performances issues (and
possibly memory issues as well). The issue was reported on IRC today
with the wintermute engine.
The cos/sin table class now has an at() function for indexing
safely into its internal array.
This allows the checking and computing of the correct
indexes to be done internally.
The indexing in allows cos/sine of 0 to 2pi to be
obtained.
The values returned by getTable are the same as before.
Comments that describe the values that the table
contains has been modified to line up with what
the code is doing.
This commit introduces the following changes:
1. Graphics::loadThumbnail()
Now returns a boolean and takes a new argument skipThumbnail which
defaults to false. In case of true, loadThumbnail() reads past the
thumbnail data in the input stream instead of actually loading the
thumbnail. This simplifies savegame handling where, up until now,
many engines always read the whole savegame metadata (including
the thumbnail) and then threw away the thumbnail when not needed
(which is in almost all cases, the most common exception being
MetaEngine::querySaveMetaInfos() which is responsible for loading
savegame metadata for displaying it in the GUI launcher.
2. readSavegameHeader()
Engines which already implement such a method (name varies) now take
a new argument skipThumbnail (default: true) which is passed
through to loadThumbnail(). This means that the default case for
readSavegameHeader() is now _not_ loading the thumbnail from a
savegame and just reading past it. In those cases, e.g.
querySaveMetaInfos(), where we actually are interested in loading
the thumbnail readSavegameHeader() needs to explicitely be called
with skipThumbnail == false.
Engines whose readSavegameHeader() (name varies) already takes an
argument loadThumbnail have been adapted to have a similar
prototype and semantics.
I.e. readSaveHeader(in, loadThumbnail, header) now is
readSaveHeader(in, header, skipThumbnail).
3. Error handling
Engines which previously did not check the return value of
readSavegameHeader() (name varies) now do so ensuring that possibly
broken savegames (be it a broken thumbnail or something else) don't
make it into the GUI launcher list in the first place.
Compilers optimise these back into shifts on architectures where
shifts of negative integers work the same as mul/div, so this
solves the UB without actually causing any performance issue.
Custom deleters of ScopedPtr are not currently fully conforming to
C++11's support for custom deleters in std::unique_ptr for the
sake of simplicity of implementation. Unlike in the standard
library, plain functions and lvalue references are not supported,
nor may custom deleters be passed to the constructor at runtime.
This can be improved in the future, if necessary, by doing what
standard library implementations usually do and creating a Pair
class that uses the Empty Base Optimization idiom to avoid extra
storage overhead of the deleter instance when it is not needed, as
in typical standard library implementations, plus some additional
type traits to support the necessary metaprogramming for the
different type overloads.
This improves the output of static assertions in all compilers,
and prevents problems in MSVC 2015 where the mechanism for
triggering a compilation error in C++98 mode may cause errors when
that compiler builds in release mode.
Fixes Trac#10154.
Normally with SDL, a mouse motion event will be sent after the
system mouse cursor has been moved by a call to
SDL_WarpMouseInWindow, but if the system cursor cannot be moved
(e.g. because the window does not have mouse focus), games still
need to receive these mouse events so they can successfully update
the mouse position internally. Otherwise, games continue to think
the mouse is still in the original position and will continue to
try to perform whatever action is associated with that mouse
position.
Refs Trac#9689.
There is no particular reason why backends that don't need to
calculate screen dimensions in advance should still need to
implement initSizeHint at this point.
Although the previous count-constructor would never make a copy of
a member at runtime, Array<T>::reserve *may* copy-construct, so
the compiler would forbid creation of arrays of NonCopyable objects
even when the array was created only once and then never resized
(and thus never actually tried to perform a copy-construction).
