The presence of those abis cause them to be included in the set of `p_features` passed to the `gdextension_export_plugin#_export_file(...)` method, which caused them to be lumped in the `features_wo_arch` set.
When trying to find the gdextension library path, we use a predicate with the following logic:
```
[features_wo_arch, arch_tag](String p_feature) { return features_wo_arch.has(p_feature) || (p_feature == arch_tag); }
```
For a `gdextension` config file like the one below, this causes the first android entry (`android.armeabi-v7a = ...`) to always be returned regardless of archs since it always satisfies the predicate.
```
[configuration]
entry_symbol = "example_library_init"
compatibility_minimum = 4.1
[libraries]
linux.x86_64 = "res://libgdexample.so"
android.armeabi-v7a = "res://libgdexample.android.template_release.armeabi-v7a.so"
android.arm32 = "res://libgdexample.android.template_release.armeabi-v7a.so"
android.x86 = "res://x86/libgdexample.android.template_release.x86.so"
android.x86_32 = "res://x86/libgdexample.android.template_release.x86.so"
android.x86_64 = "res://libgdexample.android.template_release.x86_64.so"
android.arm64-v8a = "res://libgdexample.android.template_release.arm64-v8a.so"
android.arm64 = "res://libgdexample.android.template_release.arm64-v8a.so"
```
(cherry picked from commit dce2686e52)
We made a mistake when cherry-picking #80612 with 269b115d9c,
where the global flag was defaulted to false to preserve the 4.1-stable behavior for desktop
platforms, but we forgot that the refactoring removed the force disabling of exceptions for
Android, iOS, and Web.
This reintroduces this behavior so it should be back to the same as in 4.1/4.1.1, and the
export templates should get back to their original size.
Only difference, the old code used to keep exceptions for the Web editor, but I see no reason
for it, so I disable them like with the templates.
Notably fixes issues with `is_action_just_*` queries in `_physics_process`
for TouchScreenButton.
Fixes#66318.
Fixes#82396.
(cherry picked from commit 5137497c18)
An NVIDIA profile is applied to the current executable to disable
threaded OpenGL optimizations on Windows (see #71472). But because the
application is only added to the profile upon the profile creation,
newer executables won't be added to the profile (e.g. if the profile is
created on first launch of Godot_v4.1-stable_win64.exe, when users
update the editor and launch Godot_v4.2-stable_win64.exe, the profile
will never be applied to this new executable).
This patch fixes that scenario by splitting creating the profile (if it
doesn't exist) and adding the application (if it doesn't have a profile
applied) into two separate steps.
Applications that have been manually added to a different profile aren't
overriden to avoid confusing users who know what they're doing.
(cherry picked from commit 6263774aec)
This fixes multiple issues/inconsistencies around `get_compiler_version()`:
* With no shell allocated, launching the compiler could fail even
with proper paths being set.
* The return value was described as "an array of version numbers as ints",
but the function actually returned a `Dictionary` (or `None`).
* Not all calls were properly handling a `None` return value in case of errors.
On Windows this broke compiling for me since #81869 with default settings.
* Some calls defined inconsistent defaults/fallbacks (`0` or `-1`).
(cherry picked from commit 426e18fd37)
When a hardware keyboard is connected, all key events come through so we can route them directly to the engine.
This is not the case for soft keyboards, for which the current logic was designed as it requires extra processing.
(cherry picked from commit 87102e358d)
Code now always double quotes the filename to use as command line
argument when calling explorer.exe. In particular, commas in a filename
would be interpreted by explorer.exe as separators for commands.
Similarly a trim_suffix for "file://" is assumed to be a mistake, this
could potentially be a PREfix that we want to strip, but never a suffix.
Since it didn't seem needed in the end, we removed it.
(cherry picked from commit 5608b1d411)
This means no CPU occlusion culling (and not compiling Embree), unless
you compile custom export templates with `module_raycast_enabled=yes`.
This reduces the memory footprint significantly, and binary size.
Fixes#70621.
Co-authored-by: Hugo Locurcio <hugo.locurcio@hugo.pro>
(cherry picked from commit 8c40edf8c0)
It wrongly returned 20 on array buffers, which used to be the enumerator
value of Godot 3.x's type PoolByteArray, and now is the value of type Color,
while it should return 29 which is the enumerator value for PackedByteArray.
(cherry picked from commit c662491bd4)
During GDC and general testing on Steam Deck units, we found that single
gamepads would often register inputs twice under certain circumstances.
This was caused by SteamInput creating a new virtual device, which Godot
registers as a second gamepad. This resulted in two gamepad devices
reporting the same button presses, often leading to buggy input response
on games with no multi-device logic and other-wise could cause intended
Steam rebindings to not work as intended (for example, swapping o and x
on a playstation pad if that feature isn't supported by the game.)
SDL gets around this by taking in a list of devices that are to be
ignored. When valve sees a controller that wants to be rebound via
SteamInput, they push a new VID/PID entry onto the environment
variable `SDL_GAMECONTROLLER_IGNORE_DEVICES` for the original gamepad
so that all game inputs can be read from the virtual gamepad instead.
This leverages the same logic as we are already using SDL gamepad
related HID mappings.
(cherry picked from commit 8de98dbf21)
Upon investigating the extremely slow MSVC build times in #80513, I noticed
that while Godot policy is to never use exceptions, we weren't enforcing it
with compiler flags, and thus still included exception handling code and
stack unwinding.
This is wasteful on multiple aspects:
- Binary size: Around 20% binary size reduction with exceptions disabled
for both MSVC and GCC binaries.
- Compile time:
* More than 50% build time reduction with MSVC.
* 10% to 25% build time reduction with GCC + LTO.
- Performance: Possibly, needs to be benchmarked.
Since users may want to re-enable exceptions in their own thirdparty code
or the libraries they compile with Godot, this behavior can be toggled with
the `disable_exceptions` SCons option, which defaults to true.
(cherry picked from commit 3907e53ff6)
Body length cannot be reliably retrieved from the web.
Reading the "content-length" value will return a meaningless value when
the response is compressed, as reading will return uncompressed chunks
in any case, resulting in a mismatch between the detected body size and
the actual size returned by repeatedly calling read_response_body_chunk.
Additionally, while "content-length" is considered a safe CORS header,
"content-encoding" is not, so using the "content-encoding" to decide if
"content-length" is meaningful is not an option either.
We simply must accept the fact that browsers are awful when it comes to
networking APIs.
(cherry picked from commit f4713d235a)
We can't rely on the error code from `gio` or `kioclient5`, in my
rudimentary testing they return `1` for both missing files and other
situations like not having a Trash can on the mounted volume.
Fixes#79108.
(cherry picked from commit a6e75f3971)
jsjtxietian
Fredia Huya-Kouadio
jitspoe
scgm0
thfrwn
Ben Rog-Wilhelm
bruvzg
Mikael Klasson
Rémi Verschelde
Alexander Hartmann
Max Hilbrunner
floatingpointer
Aitor Guevara
Adam Scott
SysError99
Mario Liebisch
Brecht Kuppens
Hugo Locurcio
PorkrollPosadist
Tyler
Fabio Alessandrelli
Emmanouil Papadeas
Eoin O'Neill
ocean (they/them)
kleonc