Previously, the supplied pointer was converted to an array index, then
passed to valptridx::idx for validation. If the index_type is smaller
than std::size_t, this would truncate the value before validation.
Certain out-of-range indexes would be in-range after truncation, and
incorrectly not be reported.
Reorder the check to validate the index against the array size before
truncation.
Players are not robots, but the show-path cheat tried to pretend they
are, and triggered a BUG warning[1] when looking up robot information
from the player object. Fix this by passing in the robot_info when the
caller is providing a robot, and a named `nullptr` value (as
`create_path_unused_robot_info`) when the caller is providing a
non-robot object.
[1] As below, but repeated many times
```
similar/main/ai.cpp:1905: BUG: object 0x555555858550 has type 4, expected 2
similar/main/ai.cpp:1974: BUG: object 0x555555858550 has type 4, expected 2
```
Zero the entire array, then overwrite the leading portion with the
received data. This permits the compiler to do a fixed number of large
stores, instead of a variable number of small stores.
The type aliases are sufficient. Individual bytebuffer_t
implementations can define a `static constexpr` member `endian` from the
type alias and rely on `std::integral_constant<T, V>::operator()`
instead of defining a `static` method just to return an instance of the
`std::integral_constant`.
Use `object &` instead of `vmobjptr_t`. This should generate equivalent
code, but produce smaller debug information and may require less
inlining by the compiler.
Every caller passes `1`, so remove the parameter and always use `1`.
For backward compatibility with the previous network protocol, continue
to send a `1` in the network message.
This allows taking the input by-value instead of by-reference, while
still protecting against unwanted type conversion.
Add some basic static_assert tests that the swapped values are correct.
Commit f1606f7747 ("Simplify test for
__builtin_bswap16") changed the SConstruct test to either define both
DXX_HAVE_BUILTIN_BSWAP and DXX_HAVE_BUILTIN_BSWAP16 or to define neither
of them. Follow up that commit by removing the definition of
DXX_HAVE_BUILTIN_BSWAP16 and redirecting uses of it to
DXX_HAVE_BUILTIN_BSWAP.
The conditional definition of a D2X flag causes a technical ODR
violation. It is legal, though not useful, to define this flag in D1X
and in common code. Remove the preprocessor guard so that D2X and
common code use the same definition for the enum.
Split check_index_range to check_index_range+check_index_range_size.
Redirect check_explicit_index_range_ref to check_index_range_size, so
that the index_type is not truncated and then extended.
Use a single constructor that accepts anything convertible to both of
the required types, rather than special constructors for:
- Accepting a qualified_segment
- Accepting a variant of susegment with compatible const qualifiers
- Accepting a type T that converts to qualified_segment
This reduces the number of constructors to consider, which improves
error messages when an invalid input is used.
v29_trigger and v30_trigger define a field `time`. v29_trigger never
initializes it. v30_trigger initializes it from the uninitialized
v29_trigger in legacy mode, and from a file field otherwise. No program
logic ever reads this member, so remove it.
When a file name is available, use SDL_RWFromFile + Mix_LoadMUSType_RW,
rather than calling Mix_LoadMUS(filename). This allows the calling code
to be more consistent.
SDL1 defines Mix_LoadMUS_RW as:
```
Mix_LoadMUS_RW(SDL_RWops *rw)
Mix_LoadMUSType_RW(rw, MUS_NONE, SDL_FALSE);
```
SDL2 defines Mix_LoadMUS_RW as:
```
Mix_LoadMUS_RW(SDL_RWops *src, int freesrc)
Mix_LoadMUSType_RW(src, MUS_NONE, freesrc);
```
The version with freesrc is preferable, and Rebirth used it to set
freesrc=SDL_TRUE in SDL2 mode. For SDL1, Rebirth used special logic
in the reset() call to emulate setting freesrc=SDL_TRUE. SDL1 also
exposes Mix_LoadMUSType_RW, which allows the caller to set freesrc in
both SDL1 and SDL2. Switch to that, so that the same code is used in
SDL1 and SDL2.
If the file is accessible via PhysFS, the next branch, based on
PHYSFS_openRead, will find it and use it. If it is not accessible via
PhysFS, then attempting to resolve its path would fail. Thus, this
attempt is redundant regardless of whether the file is reachable by
PhysFS. Remove it.
The contents of the output buffer are undefined if PHYSFSX_getRealPath
fails, so mark the function as [[nodiscard]] and modify all callers to
check that the function succeeded.
Rework the error paths to return path-specific status codes so that the
caller can report exactly which step caused an HMP file to be rejected.
On error, print this reason numerically and, if the reason was a PhysFS
error, also print the PhysFS error code numerically and symbolically.
If the type is not None, an earlier statement will have already
returned. Thus, the type can be assumed to be None if these test
statements are reached.
Previously, if the filename had no dots, then mix_play_file would refuse
to play it at all. This was excessive, as only the check for hmp needs
a filename with a dot in it. Change the logic so that a dot-less file
still is assumed not to be an hmp, but instead of returning, will fall
through and try the other load types.
Return the type for compatibility, but set the music type immediately.
Change callers to return without using a switch to set the music type
based on the returned value. This allows the callers to exit early on a
successful load.
Remove the `switch` in the caller, since now every path that can set
`current_music_type` will either (a) set it to None because the load
failed or (b) set it to non-None and return before reaching the
`switch`.
Return the type for compatibility, but set the music type immediately.
Change callers to return without using a switch to set the music type
based on the returned value. This allows the callers to exit early on a
successful load.
- Centralize the default value for the hook function.
- Move some static functions to the anonymous namespace.
- Define functions for setting the mixer parameters to ADLMIDI mode and
SDL_mixer mode, so that these modes can be set from more places.
AlumiuN reports that mingw32-w64-gcc-8.1.0 incorrectly reports
`ascending` as unused-but-set. This is clearly not true. Reorder the
code to avoid saving `ascending`, and instead use the result of
`detail::get_xrange_ascending` directly. This also improves the error
message in the DXX_ALWAYS_ERROR_FUNCTION path.
Reported-by: AlumiuN <https://github.com/dxx-rebirth/dxx-rebirth/issues/626>
If two or more events are delivered in the same loop, the previous
implementation would count joystick motion multiple times. Fix this by
moving the joystick interpretation to occur once, after all the events
have been processed.
- Use std::integral_constant instead of a static function that returns
the value
- Remove unused protocol_family
- Replace the enum with a typedef for the one type that the enum was
used to define
If exactly one object will always be needed, use an overload that
returns the object id. Otherwise, use an overload that only returns
whether at least one object was created. This simplifies callers that
always request exactly one object.
If at least two robots would be dropped, and a drop failed, then
object_create_robot_egg would report failure to the caller. Callers
that check the return code treat any failure as total failure, but that
is not guaranteed to be true. If the game successfully dropped one
robot and failed when dropping a second, then the caller would receive a
status of failure.
Fix this by returning a status of whether at least one object was
created.
- Change D1X to use D2X rules regarding Vulcan cannon pickup. The D1X
rules were confusing at best, and seem outright wrong in some ways.
- When a Vulcan cannon was picked up, it was treated as containing not
less than VULCAN_WEAPON_AMMO_AMOUNT rounds, regardless of what it
actually contained. D2X respected the actual contained count, even
when running in D1X emulation mode.
- In D1X single player, if the Vulcan cannon was not picked up, then
it could be treated as Vulcan ammo. If at least 1 unit of
ammunition was added to the player, the entire powerup would be
consumed, regardless of how much ammunition remained.
- In D2X single player, if the Vulcan cannon was not picked up, then
ammunition would be taken from it, but the powerup would only be
consumed if all its ammunition was taken.
- In D1X multiplayer, a player who already had a Vulcan cannon could
not get anything from touching a new cannon, and the cannon would
not be changed.
- In D2X multiplayer, a player who already had a Vulcan cannon could
take ammunition from the cannon, but the cannon could not be drained
below VULCAN_AMMO_AMOUNT. If the cannon had VULCAN_AMMO_AMOUNT or
less, then the player could not take ammunition. If the cannon had
more, the player could drain it to that level.
- Replace all that with a simplified version of the D2X rules:
- If the player does not have the cannon, the cannon is picked up
and removed from the mine. The player takes as much of its
ammunition as possible. If the cannon was well stocked, and the
player was nearly full, some ammunition will be lost. This is
unfortunate, but the game has always had this rule, and changing
it would require dropping one or more Vulcan Ammo packs to
represent the untaken ammunition.
- If the player already had that cannon, then the player takes as
much ammunition as the cannon has, while not exceeding the
ammunition cap. Other players, if any, are updated about how much
ammunition remains in the cannon. The cannon remains in the mine.
- Backport to D1X the network message for updating the contained
ammunition in a vulcan cannon. zico added the basic feature in
7684ce92, but only applied it to D2X. With the change to let D1X
multiplayer take ammunition from the cannon, D1X now needs the same
feature.
- Remove the special case to delete an empty cannon. Instead, let the
cannon remain in the mine without ammunition. This allows a player in
single player mode to leave behind a backup cannon, which could be
useful if the player is killed and wishes to rearm before returning to
the death site. Similarly, under the new rule that players in
multiplayer can drain the cannon down to 0 ammunition, this removal
allows the cannon to remain behind for someone else to take, rather
than allowing it to be deleted by a player who already had an instance
of it.
clang chokes on use of a `constexpr bool` that is initialized separately
from its definition. gcc allows this. Move the value computation into
a constexpr helper function, so that the variable can be defined and
initialized in the same statement.
Instead of creating the powerup from a player, then overwriting the
location and velocity of the powerup, and fixing up its segment, create
the powerup directly where it should be, with the intended velocity.
Iterating over it returns each side number in turn. This allows
converting many loops of the form:
```
for (int i = 0; i < MAX_SIDES_PER_SEGMENT; ++i)
```
to the compact form:
```
for (const auto i : MAX_SIDES_PER_SEGMENT)
```
The compact form brings the usual benefit of range-based for: delegating
iteration to the compiler prevents the loop body from skipping a step,
and makes clear in the code that this is the case.
Some flags merit a type other than int8_t. Begin moving flags out to
distinct variables with their own type.
Add static_assert checks that the ABI relevant structures do not change.
clang warns that an implicit copy constructor is deprecated in the
presence of an explicitly defaulted copy-assignment operator.
Add an explicitly defaulted default constructor too, since the
explicitly defaulted copy constructor suppresses generation of an
implicit default constructor, and the default constructor is needed.
Commit 4d6affd74d ("Fix SDL-only briefing crash") added a hack to
prevent a crash when the briefing window is automatically resized.
However, a logic error in that hack causes some subcanvases that were
not wrong to be truncated. Change the logic to fix those subcanvases,
such as the one used for rendering a background in the High Scores
window. The new logic still avoids the crash in in SDL-only mode.
Fixes: 4d6affd74d ("Fix SDL-only briefing crash")
On some systems, such as the Apple M1 (ARM-based Macs),
dylibbundler is in $PATH, but not in a directory found by the limited
path that SCons uses for Command() invocations. This causes the SConf
test to succeed (because it uses the full $PATH), but the Command() to
fail (because it uses a reduced path). Fix this by using the user's
$PATH, if defined, for this Command() invocation.
Along the way, rework tool_bundle.py to use SCons' Node() objects in
place of direct path manipulation. This lets SCons produce better error
messages in some cases.
Reported-by: Kreeblah <https://github.com/dxx-rebirth/dxx-rebirth/pull/610>
If std::common_type<range::index_type>... finds a common index_type
among all the zipped ranges, pass that common type through as
zip<...>::index_type. Otherwise, set zip<...>::index_type to void.
This allows enumerate(zip(...))) to report a better index_type.
Trim the bitmap's current width and height to be not more than its
parent. Otherwise, an automatic screen resize, such as triggered when
playing the robot briefings in Descent 2, will give the window a new
smaller bm_data, but continue to use the old dimensions. This then
causes load_briefing_screen -> gr_bitmap_scale_to -> scale_line to write
off the end of the smaller bm_data, causing memory corruption and
eventually a crash.