- Fixed a bug where pushing items over queue's size left it in a corrupted state
- For non-trivial types, have clear() and pop() run destructors
- Added emplace(args...)
- Added empty()
FixedSizeQueue has semantics of a circular buffer,
so pushing items continuously is expected to keep overwriting oldest elements gracefully.
Tests have been updated to verify correctness of a previously bugged behaviour
and to verify correctness of destructing non-trivial types
API has been made stricter, layers are now managed with shared pointers,
so using them temporarily increased their reference counters.
Additionally, any s_layers map has been guarded by a read/write lock,
as concurrent write/reads to it were possible.
At its only usage point, its return value is stored into a u32, and the
default implementation returns 0xFFFFFFFF (-1), which would be an
unsigned integer. Given all of the bits are used to determine a color,
it makes slightly more sense to treat this as an unsigned value as
opposed to a signed one.
We're allowed (by the standard) to forward declare types within
std::vector, so we can replace direct includes with forward declarations
and then include the types where they're directly needed.
While we're at it, we can remove an unused inclusion of <cstring>, given
nothing in the header uses anything from it. This also revealed an
indirect inclusion, which this also resolves.
Previously u32 was being used for part of the interface and unsigned int
was being used for other parts. This makes the interface fully consistent by
using only one type.
We opt for u32 here given they communicate the same thing (for platforms
we care about where int is 32-bit), while also being less to read.
While we're at it, we can also default the constructor and destructor of
inheriting classes in their respective cpp file to prevent the
construction and destruction of non-trivial types being inlined into
other regions of code.
Previously these functions were declared without the static specifier,
giving them external linkage, which isn't really ideal.
Instead, we can place these functions up by the relevant file-scope
variables and place them inside an anonymous namespace with said variables,
giving them internal linkage.
Avoids the use of the null pointer to represent an empty string.
Instead, we can simply pass an empty string_view instance. Using
std::string_view enforces this invariant at the API level.
Due to the lack of cast here, this will actually print out the ascii
value, rather than the character itself, due to promoting to integral
values. Instead, we can eliminate the use of character operands and just
print the value itself directly, given it's equivalent behavior with
less code.
Allows these arrays to be placed within the read-only segment (and
enforces the immutability in the code itself). While we're at it, we can
make use of std::array here.
Now that the std::map less-than comparitor is capable of being used with
heterogenous lookup, we're able to convert many of the querying
functions that took std::string references over to std::string_view.
Now these functions may be used without potentially allocating a
std::string instance unnecessarily.
Previously, when performing find() operations or indexing operations on
the section map, it would need to operate on a std::string key.
This means cases like:
map.find(some_string_view)
aren't usable, which kind of sucks, especially given for most cases, we
use regular string literals to perform operations in calling code.
However, since C++14, it's possible to use heterogenous lookup to avoid
needing to construct exact key types. In otherwords, we can perform the
above or use string literals without constructing a std::string instance
around them implicitly.
We simply need to specify a member type within our comparison struct
named is_transparent, to allow std::map to perform automatic type
deduction.
We also slightly alter the algorithm to an equivalent compatible with
std::string_view (which need not be null-terminated), as strcasecmp
requires null-terminated strings.
While we're at it, we can also provide a helper function to the struct
for comparing string equality rather than only less than. This allows
removing other usages of strcasecmp in other functions, allowing for the
transition of them to std::string_view.
fmt diverges from printf in that '.' as a precision specifier may only
be used for floating-point values (makes sense, given it's indicating
precision after the decimal point).
Begins the transition to using fmt for string formatting where
applicable. Given fmt supports formatting std::string instances out of
the box, we can remove now-unnecessary calls to .c_str() and .data().
Note that this change does not touch the actual logging subsystem aside
from converting the final StringFromFormat call in the process over to
fmt::format. Given our logging system is heavily used throughout the
entire codebase, and converting that over will be quite a large change
by itself, this will be tackled near the end of the conversion process.
Avoids dragging in IniFile, EXI device and SI device headers in this header which is
quite widely used throughout the codebase.
This also uncovered a few cases where indirect inclusions were being
relied upon, which this also fixes.
Allows for both string types to be non-allocating. We can't remove the
const char* overload in this case due to the fact that pointers can
implicitly convert to bool, so if we removed the overload all const
char arrays passed in would begin executing the bool overload instead of
the string_view overload, which is definitely not what we want to occur.
Since C++17, non-member std::size() is present in the standard library
which also operates on regular C arrays. Given that, we can just replace
usages of ArraySize with that where applicable.
In many cases, we can just change the actual C array ArraySize() was
called on into a std::array and just use its .size() member function
instead.
In some other cases, we can collapse the loops they were used in, into a
ranged-for loop, eliminating the need for en explicit bounds query.
Prior to this commit, the emitter would emit a 7-byte instruction when
loading a 32-bit immediate to a 64-bit register.
0: 48 c7 c0 ff ff ff 7f mov rax,0x7fffffff
With this change, it will check if it can instead emit a load to a
32-bit register, which takes only 5 or 6 bytes.
0: b8 ff ff ff 7f mov eax,0x7fffffff
std::call_once is guaranteed to execute the given callable object
exactly once. This guarantee holds even if the function is called
concurrently from several threads.
Given that, we can replace the mutex and boolean flag with
std::call_once and a std::once_flag to perform the same behavior.
Previously, every entry pair within the map would be copied. The reason
for this is subtle.
A std::map's internal entry type is defined as:
std::pair<const Key, Value>
but the loop was declaring it as:
std::pair<Key, Value>
These two types aren't synonymous with one another and so the compiler
is required to always perform a copy.
Using structured bindings avoids this (as would plain auto or correcting
the explicit type), while also allowing the use of more appropriate
names compared to first and second.
std::function is allowed to heap allocate in order to hold any necessary
bound data in order to execute properly (e.g. lambdas with captures), so
this avoids unnecessary reallocating.
While current usages of ParseLine aren't problematic, this is still a
public function that can be used for other purposes. Essentially makes
the function handle potential external inputs a little nicer.
We can just utilize map's insert_or_assign() function and check the
return value to determine whether or not we need to insert the key into
the keys_order vector.
All supported platforms now have easy access to a compiler with C++17
support.
C++17 potentially allows for some nice cleanups and removes the need
for standard library backports (optional/variant).
See discussion at https://dolp.in/pr6264#discussion_r158134178
Different address spaces can be chosen in the memory view panel.
* Effective (or virtual): Probably the view people mostly want. Address
translation goes through MMU.
* Auxiliary: ARAM address space. Does not display anything in Wii mode.
* Physical: Physical address space. Only supports mem1 and mem2 (wii
mode) so far.
MemoryWatcher only works on Linux and affects emulation determinism due
to scheduling additional events, which causes NetPlay to desync.
Considering that this interface is a rather specialized use case, the
communication with it is kinda crappy *and* it's affecting emulation, I
think it's best to just axe it and come up with a better implementation
of the functionality.
* The high half of regOp is immediately overwritten so the value in it is irrelevant.
* MOVSD produces an unnecessary dependency on the high half of regOp.
* MOVAPS is implemented as a register rename on modern microarchitectures.
There is no reason to use MOVAPD over MOVAPS, for two reasons:
* There has never been a microarchitecture with separate single and double domains.
* MOVAPD is one byte longer than MOVAPS
This sends arbitrary packets in chunks to be reassembled at the other
end, allowing large data transfers to be speed-limited and interleaved
with other packets being sent. It also enables tracking the progress of
large data transfers.
Stenzek