A template for `jsdoc` that generat the HTML5 public classref.
The script can be run via `npm run docs` to print to stdout.
You can dry run via `npm run docs -- --d dry-run` or write to file via
`npm run docs -- -d /path/to/file.rst`
Also update Makefile in `doc/` and add dry run test to CI.
(cherry picked from commit 472482013e)
This commit also removes the utils.js engine file, moving some of it's
content to config.js and some to engine.js .
(cherry picked from commit 018ee5a4dc)
We used to have it like `$GODOT_VERSION` which caused inconsistencies
between different scons versions when substituting it.
It's now `@GODOT_VERSION@`, which is safe on both scons3 and scons4.
(cherry picked from commit 4404eb57e4)
When using use_static_cpp we want to statically link with atomic as well
to make sure we don't incur any new runtime dependencies.
Scons doesn't quite support this so we do this little trick.
According to the LLVM documentation when using GNU's libstdc++ clang
will not automatically link with -latomic. This is necessary since we
merged c++11 atomics support.
This fixes linking using Clang on Linux
Three canvas resize policies:
- `None`: Godot window settings are ignored.
- `Project`: Godot handles the canvas like a native app (resizing it
when setting the window size).
- `Adaptive`: Canvas size will always adapt to browser window size.
Use `None` if you want to control the canvas size with custom JavaScript
code.
This helps resolve issues where the project ndk version differs from the one pointed by the `ANDROID_NDK_ROOT` environment variable (if it exists).
(cherry picked from commit edeca16fb6)
- Based on C++11's `atomic`
- Reworked `SafeRefCount` (based on the rewrite by @hpvb)
- Replaced free atomic functions by the new `SafeNumeric<T>`
- Replaced wrong cases of `volatile` by the new `SafeFlag`
- Platform-specific implementations no longer needed
Co-authored-by: Hein-Pieter van Braam-Stewart <hp@tmm.cx>
In addition, add support for scaling and applying filter to the splash screen on Android.
One limitation of the api being used is that the splash screen aspect ratio is not maintained when it's scaled up.
- Based on C++11's `thread` and `thread_local`
- No more need to allocate-deallocate or check for null
- No pointer anymore, just a member variable
- Platform-specific implementations no longer needed (except for the few cases of non-portable functions)
- Simpler for `NO_THREADS`
- Thread ids are now the same across platforms (main is 1; others follow)
- Based on C++11's `mutex` and `condition_variable`
- No more need to allocate-deallocate or check for null
- No pointer anymore, just a member variable
- Platform-specific implementations no longer needed
- Simpler for `NO_THREADS`
- Based on C++11's `mutex`
- No more need to allocate-deallocate or check for null
- No pointer anymore, just a member variable
- Platform-specific implementations no longer needed
- Simpler for `NO_THREADS`
- `BinaryMutex` added for special cases as the non-recursive version
- `MutexLock` now takes a reference. At this point the cases of null `Mutex`es are rare. If you ever need that, just don't use `MutexLock`.
- `ScopedMutexLock` is dropped and replaced by `MutexLock`, because they were pretty much the same.
- Based on C++14's `shared_time_mutex`
- No more need to allocate-deallocate or check for null
- No pointer anymore, just a member variable
- Platform-specific implementations no longer needed
- Simpler for `NO_THREADS`
This makes it possibly to run Linux binaries compiled with udev support on
Linux systems which do not provide udev (typically systemd-less distros).
If udev is missing, we fall back to parsing `/dev/input` like when compiled
without udev support (`udev=no`).
Also adding some verbose debug statements to know which method we're using
when debugging Linux joypad issues.
The libudev so wrappers were generated on Mageia 8 with libudev 246.9 using
https://github.com/hpvb/dynload-wrapper:
```
./generate-wrapper.py --include /usr/include/libudev.h --sys-include '<libudev.h>' \
--soname libudev.so.1 --init-name libudev --omit-prefix gnu_ \
--output-header libudev-so_wrap.h --output-implementation libudev-so_wrap.c
```
(cherry picked from commits a10c259c1d
and e26a1f807b)
Edit: Updated to version 0.2 of dynload-wrapper to fix symbols clobbering as
done in #46143.
This is what GitHub Actions now provide and they removed the previous 21.3.6528147.
A bit annoying to have our hand forced this way but it's still 21.x so should be good
to upgrade.
(cherry picked from commit c730da8b20)
By generating stubs using https://github.com/hpvb/dynload-wrapper we
can dynamically load libpulse and libasound on systems where it is available.
Both are still a build-time requirement but no longer a run-time dependency.
For maintenance purposes the wrappers should not need to be re-generated
unless we want to bump pulse or asound to an incompatible version. It is
unlikely we will want to do this any time soon.
cherry-pick from 09f82fa6ea
`OS.get_screen_scale` will now return the `window.devicePixelRatio`
value, `OS.get_screen_dpi` uses CSS media queries to find approximate
DPI value for the current display.
`OS.get_screen_size` also return the actual screen size (not the CSS
pixel size).
Issues addressed:
a) Axis mappings were including virtual mouse axes on NVIDIA Shield TV.
The virtual mouse axes have the same axis numbers as the normal analog stick numbers. This was completely breaking joypad support on NVIDIA Shield TV.
b) Joypads were being tracked in a List with the index in the list being treated as the Godot device id.
If a device were to be removed, any device later in the list would be shifted, potentially causing future events with the shifted joypads to have incorrect IDs according to the Godot engine.
c) Unnecessary events were being sent to the Godot engine.
A check was added (per Joystick) that will prevent sending events for all axes when only a single axis value changed.
A similar check was added for "HATs".
See #45712
This is meant for users making custom builds to match the options used on
optimized, official builds.
This enables, on the platforms which support them:
- `use_static_cpp=yes` (portable binaries for Linux and Windows)
- `use_lto=yes` (link time optimizations - note: requires a lot of RAM!)
- `debug_symbols=no` (no debug symbols, smaller binaries)
Also abort when using MSVC with `production=yes`, as:
- It cannot optimize the GDScript VM like GCC or Clang do, leading to
significant performance drops.
- Its LTO support is unreliable, at least used to trigger crashes last
we tried it extensively.
All options can still be overridden if specified, and the `dev=yes` option
was changed to also support overrides.
(cherry picked from commit db26871210)
We used to only persist specific sub-folder of /home/web_user/ when
running the Web Editor. This resulted in bad UX about default project
creation path etc.
This PR makes the whole folder persistent, move the zip preloading to a
different folder (to avoid persisting it), and automatically prompt the
user to import it if present.
The canvas_id is `#`-prefixed to work with emscripten as a CSS selector.
When comparing to an event target ID (e.g. when checking if the canvas
is fullscreen, or is locking the mouse) we need to skip the first char
(the hash).
This has been enabled for years in official binaries, and users making custom builds
may end up not enabling it unknowingly, so it's best if we default to the same as
what official builds do.
The original reason for having it opt-in was likely the addition of a dependency on
libudev, but that should be fairly ubiquitous by now.
(cherry picked from commit e8b69fccbe)
This enables `-static-libgcc -static-libstdc++` which help make custom Linux
builds more portable (official builds have been using this option for years).
For some obscure reason Ubuntu 18.04 i386 crashes when using the option for
i386 builds, so let's play it safe and enable for x86_64 only for now.
(cherry picked from commit 1ebd66daff)
There are no guarantees that joypads are in event0-event32
range. Some devices, such as laptops with detachable keyboards
and wacom can reserve events all the way up to 32.
Some udev rules with e.g. custom controller firmwares may
load the device as /dev/input/eventX, where X is greater than
32.
This patch uses POSIX dirent to enumerate the event devices, so
entries outside 0-32 range are not skipped.
(cherry picked from commit 01c030f9b7)
Moved AppDelegate push notifications methods implementation to 'GODOT_ENABLE_PUSH_NOTIFICATIONS'
which can be used in plugins to implement APNS plugins.
(cherry picked from commit 366ce084f4)
No longer use emscripten functions for gamepads, implement them as
library functions in library_godot_display.js instead.
This allows us to do a better job at "guessing" vendorId, productId, OS,
etc. thus allowing us to better find the remapping for the controller.
Moved previously builtin modules 'GameCenter', 'AppStore', 'iCloud' to separate modules to be represented as plugin.
Modified 'ARKit' and 'Camera' to not be builtin into engine and work as plugin.
Changed platform code so it's not affected by the move.
Modified Xcode project file to remove parameters that doesn't make any effect.
Added basic '.gdip' plugin config file.
Happy new year to the wonderful Godot community!
2020 has been a tough year for most of us personally, but a good year for
Godot development nonetheless with a huge amount of work done towards Godot
4.0 and great improvements backported to the long-lived 3.2 branch.
We've had close to 400 contributors to engine code this year, authoring near
7,000 commit! (And that's only for the `master` branch and for the engine code,
there's a lot more when counting docs, demos and other first-party repos.)
Here's to a great year 2021 for all Godot users 🎆
(cherry picked from commit b5334d14f7)
The previously used tool, `jarsigner` has been deprecated in favor of `apksigner` which is bundled with the Android SDK.
The logic is refactored accordingly and a few editor settings have been deprecated in the process as they're no longer necessary.
Note: As a side effect, specifying the Android SDK path is now required. The docs will be updated to reflect that change.
Otherwise we can get situations where platform-specific opts with the same name
can override each other depending on the order at which platforms are parsed,
as was the case with `use_static_cpp` in Linux/Windows.
Fixes#44304.
This also has the added benefit that the `scons --help` output will now only
include the options which are relevant for the selected (or detected) platform.
(cherry picked from commit 0f84d8dc49)
Fixes#42232.
And fixes memory leak with use of DirAccess, and harmonize the use of
the sanitized pkg name.
(cherry picked from commits 707a62783b
and 02d228554b)
Working, with emscripten > 2.0.9
Yes, the unreleased version. 2.0.9 works, but throws and error due to a
bug in emscripten with the thirdparty ENet library.
The issue is fixed upstream so newer releases will work.
This was caused by the devicePixelRatio being applied twice, once by the
HTML code, once by the OS code.
More specifically, OS.get_window_size() would return the canvas element
size, while OS.set_window_size() would set the element size to the
specified value times the devicePixelRatio.
Calling OS.set_window_size(OS.get_window_size()) would reapply the
devicePixelRatio every time.
This commit changes the behaviour so that OS.set_window_size() do not
apply the devicePixelRatio to the canvas element size, by it divides the
CSS size instead.
Applies to javascript files inside the platform library folder, the
exposed Engine code, and any javascript files in modules.
Files ending with ".externs.js" will be ignored, you can create a
".eslintignore" file to specify extra files to be ignored.
Initial work to make liniting easier.
This includes:
- Rename http_request.js to library_godot_http_request.js.
- Rename externs.js to engine.externs.js.
- New library_godot_runtime.js (GodotRuntime) wraps around emscripten
functions.
- Refactor of XMLHttpRequest handler in engine/preloader.js.
- Few fixes to bugs spotted by early stage linting.
For some reason the `-target` option on the `LINKFLAGS` was causing a weird
issue where osxcross' clang wrapper would attempt using the system `/bin/ld`
instead of the osxcross version (which is Apple's `ld64`).
The error message would be:
```
/bin/ld: unrecognized option '-dynamic'
```
Also removed from `CCFLAGS` for consistency, it seems to work fine with only
`-mmacosx-version-min`.
(cherry picked from commit dbbbb53927)
This makes these platform behave as MacOS in that regard and also fixes the editor window appearing in some cases even when --no-window has been passed.
Some controllers (notably those made by 8bitdo) do not always emit an event to zero out a D-pad axis before flipping direction. For example, when rolling around aggressively the D-pad of an 8bitdo SN30 Pro/Pro+, the following may be observed:
```
ABS_HAT0X : -1
ABS_HAT0Y : -1
ABS_HAT0Y : 0
ABS_HAT0Y : 1
ABS_HAT0X : 1
```
Notable here is that no event for `ABS_HAT0X: 0` is emitted between the events for `ABS_HAT0X: -1` and `ABS_HAT0X: 1`. Consequently, the game engine believes that both the negative _and_ positive x-axis directions of the D-pad are activated simultaneously (i.e `is_joy_button_pressed()` returns `true` for both `JOY_BUTTON_DPAD_LEFT` and `JOY_BUTTON_DPAD_RIGHT`), which should be impossible.
This issue is _not_ reproducible on all controllers. The Xbox One controller in particular will not exhibit this problem (it always emits zeroing out events for an axis before flipping direction).
The fix is to always zero out the opposite direction on the D-pad axis in question when processing an event with a nonzero value. This unfortunately wastes a small number of CPU cycles on controllers that behave nicely.
**I have verified this issue is also reproducible in the stable 3.2 branch**
(cherry picked from commit dd021099ff)
`debug_symbols=yes` will now behave like `debug_symbols=full` did
before. The difference in compressed file sizes is not that large,
which means there isn't much point in having two different values.
This helps make the buildsystem easier to understand.
(cherry picked from commit ff1f0d2cb5)
Rewrote AudioDriverJavaScript to support multiple processor nodes.
The old (and deprecated) ScriptProcessorNode when threads are not
available, and the new AudioWorklet API when threads are enabled.
The new implementation uses two ring buffers and a shared state to
communicated with the AudioWorklet thread.
The audio.worklet.js JavaScript file is always added to the export
template, but only really used (and download) in the thread build.
The API is implemented in javascript, and generates C functions that can
be called from godot.
This allows much cleaner code replacing all `EM_ASM` calls in our C++
code with plain C function calls.
This also gets rid of few hacks and comes with few optimizations (e.g.
custom cursor shapes should be much faster now).
The size of the audio buffer was incorrectly doubled when creating the
script processor.
latencyHint is expressed in seconds, not milliseconds.
Additionally, on some browsers it actually affect the performance and
stability of the audio driver.
For this reason it has been completely disabled (interactive) and a not
has been left for future reference.
minizip documentation describes tm_mon as expecting the number of months
since January - [0, 11], but the month returned by OS.get_date() is in
the range of [1, 12].
(cherry picked from commit 5fe902244a)
Batching is mostly separated into a common template which can be used with multiple backends (GLES2 and GLES3 here). Only necessary specifics are in the backend files.
Batching is extended to cover more primitives.
A new editor plugin, specific to HTML5, that provide some extra features
needed to make the editor usable on that platform.
For now, it adds a "Download project sources" option in the "Tool" menu,
so the user can download the work done as a zip file (from the browser
storage).
This should be made available in emscripten in a decent way.
Possibly after unmount, to free the database lock and allow performing
operations on it from javascript after the Emscripten Runtime has
exited.
This fixes a "random" deadlock when quitting the editor.
I still haven't figure out the root cause, but having a bigger seems to
greatly mitigate the issue.
The new pool size (pre-allocated threads) is now 8.
Backport for X11 Display Server fixes on 3.2 branch.
1. Implement SAVE_TARGETS mechanism
Allows sending the clipboard content to the clipboard manager on exit to
keep the content when using a clipboard manager that doesn't
automatically makes a backup when copying.
MULTIPLE selection mechanism also had to be implemented, because in this
case, the clipboard manager might request multiple selection targets at
once.
Known use case: Ubuntu with XFCE4
2. Implement INCR mechanism
Allows pasting from x11 clipboard to receive data incrementally, which
is required when handling data size > 256KB.
No longer parse emscripten/emsdk config to detect emcc/node paths.
Use WhereIs to find "emcc" and "node", look for "node_modules" in "emcc"
path.
(cherry picked from commit 7998745237)
This should fix some of the audio stuttering issues when the HTML5
export is compiled with threads support.
The API should be ported to AudioWorklet to (hopefully) be perfect.
That though, cannot be backported to 3.2 due to extra restriction of
AudioWorklet (which only runs in SecureContext, and needs a polyfill for
Safari).
This allow the page to be considered a SecureContext if the address is
localhost (127.0.0.1/::1) and let Firefox (and future Chrome versions)
enable extra features needed for the HTML5 threaded export.
Fixes: #28683, #28621, #28596 and maybe others
For iOS we enable pvrtc feature by default for both GLES2/GLES3
Etc1 for iOS doesn't have any sense, so it disabled.
Fixed checks in export editor.
Fixed pvrtc ability detection in GLES2 driver.
Fixed pvrtc encoding procedure.
The engine now expects to emscripten FS to be setup and sync-ed before
main is called. This is exposed via `Module["initFS"]` which also allows
to setup multiple persistence paths (internal use only for now).
Additionally, FS syncing is done **once** for every loop if at least one
file in a persistent path was open for writing and closed, and if the FS
is not syncing already.
This should potentially fix issues reported by users where "autosave"
would not work on the web (never calling `syncfs` because of too many
writes).
All Linux distros, and FreeBSD and OpenBSD seem to have libXrandr.so.2,
but for some reason recent NetBSD versions seem to have libXrandr.so.3 now.
(cherry picked from commit 413ff7938d)
Add __NetBSD__ to `platform_config.h` so that it can find `alloca`
and use the proper `pthread_setname_np` format.
Rename RANDOM_MAX to avoid conflict with NetBSD stdlib.
Fixes#42145.
(cherry picked from commit 5f4d64f4f3)
By default 'add_ios_framework' would not embed a framework to save previous behavior.
New 'add_ios_embedded_framework' would embed framework on export.
Previously `joypad_count` was used as the index into the d_joypads array when initializing a new gamepad.
This caused the accidental override of an already connected device when a gamepad with a lower id was disconnected and connected again.
fixes#17566
(cherry picked from commit 802a0316c5)
It's too hard to get compatibility between GNU and BSD sed,
so let's just use perl oneliners.
And improve it to also remove trailing tabs, not just spaces.
(cherry picked from commit c71e189efd)
This gesture recognizer will prevent GLView from processing unwanted gestures.
Emulates UIScrollView behavior
Fires delayed touches on significant movement
1. Disable virtual keyboard focus adjustment on Android
The default adjustment setting was causing the view to pan down in order
to adjust the focus on the text content.
We don't need any focus adjustment since we're using a fixed size window
for our application.
Documentation:
https://developer.android.com/reference/android/view/WindowManager.LayoutParams#SOFT_INPUT_ADJUST_NOTHING
2. Fix virtual keyboard height regression
Disabling virtual keyboard focus adjustement caused get_keyboard_height
to always return 0 because it was calculated when the view is resized.
In order to fix it, a PopupWindow is now created on top of the main view
and is set for focus adjustments so the keyboard size can be calculated
based on this popup without affecting the main view.
Its last use was removed in Godot 3.0, so it no longer makes sense to define.
Also removed `D3D_DEBUG_INFO` for Windows as it's likely a left over from a
long time ago pre-opensourcing when Godot had some form of Direct3D 9 support?
(cherry picked from commit dcf902df85)
Depending on the device implementation, editor actions could be
received with different action ids or not at all for multi-line.
Added a parameter to virtual keyboards to properly handle single-line
and multi-line cases in all situations.
Single-line:
Input type set to text without multiline to make sure actions are sent.
IME options are set to DONE action to force action id consistency.
Multi-line:
Input type set to text and multiline to make sure enter triggers new lines.
Actions are disabled by the multiline flag, so '\n' characters are
handled in text changed callbacks.
This reverts commit 920639511d.
The changes are good, this revert is only done for release management reasons
as we want this feature to get more testing before making it in a stable build,
but a 3.2.3 release is imminent to handle some regressions in 3.2.2.
This will be re-committed in a 3.2-based feature branch, and we'll merge it
again once we're confident about it (probably for 3.2.4).
The Online Tutorials section of InputMap in the editor's built-in
documentation viewer contains this link:
docs.godotengine.org/en/latest/tutorials/inputs/inputevent.html#inputmap
The macOS implementation for opening a link percent-encodes it before
sending it to the browser, resulting in a 404. This is to fix#13422
where filenames with special characters could not be opened in Finder.
However, this breaks URLS so I added a check to see if the resource
scheme is file:// and if so, only then is it escaped. This allows other
schemes like `http`, `ftp`, and `mailto` to be used.
(cherry picked from commit b8e6ff9a7f)
Fixes a small memory leak reported by lsan:
```
Direct leak of 73 byte(s) in 1 object(s) allocated from:
#0 0x7f29825f3e70 in malloc (/lib64/liblsan.so.0+0xee70)
#1 0x7f29824a5729 in XGetWindowProperty (/lib64/libX11.so.6+0x29729)
```
The Haiku port now resides at:
https://github.com/godotengine/godot-haiku-platform
While we're happy to support as many Free and Open Source platforms as we can,
we need to put the main focus on the platforms that we can reasonably maintain,
and for Haiku we have been lacking maintainers ever since the port was first
merged in 2015.
The Haiku code has not been compiling and much less working at least since the
release of Godot 3.0, and while some attempts have been made at fixing things,
it's still not functional today in the `3.2` branch (and much less in `master`,
understandably).
Having it in an external repository should hopefully enable Haiku contributors
to fix issues in their own time, and possibly tag versions compatible with
past Godot releases once they are ready.
(cherry picked from commit efcc508ee5)
The missing file type in file attributes was causing the file to lose
executable permissions when unzipped with some softwares.
(cherry picked from commit 4501771fd8)
- Refactored the Engine code, splitted across files.
- Use MODULARIZE option to build emscripten code into it's own closure.
- Optional closure compiler run for JS and generated code.
- Enable lto support (saves ~2MiB in release).
- Can now build with tools=yes (not much to see yet).
- Dropped some deprecated code for older toolchains.
- Add onExit, and onExecute JS function.
- Add files drag and drop support.
- Add support for low precessor usage mode (via offscreen render, swap).
Make the first iteration as soon as the canvas is setup, avoiding a
black screen before the next animation frame is requested.
Also create OS and do setup before syncing FS to avoid crash when IDBFS
access is denied.
The previous code for OS_Windows::get_ticks_usec() multiplied the tick count by 1000000 before dividing by ticks_per_second. The ticks is counted in a 64 bit integer and is susceptible to overflow when a machine has been running for a long period of time (days) with a high frequency timer.
This PR separates the overall calculation into one for seconds and one for the remainder, removing the possibility of overflow due to the multiplier.
(cherry picked from commit db9fa88160)
Configured for a max line length of 120 characters.
psf/black is very opinionated and purposely doesn't leave much room for
configuration. The output is mostly OK so that should be fine for us,
but some things worth noting:
- Manually wrapped strings will be reflowed, so by using a line length
of 120 for the sake of preserving readability for our long command
calls, it also means that some manually wrapped strings are back on
the same line and should be manually merged again.
- Code generators using string concatenation extensively look awful,
since black puts each operand on a single line. We need to refactor
these generators to use more pythonic string formatting, for which
many options are available (`%`, `format` or f-strings).
- CI checks and a pre-commit hook will be added to ensure that future
buildsystem changes are well-formatted.
(cherry picked from commit cd4e46ee65)
The value of this, does not include the layout.
The code has extra logic to map the unicode value to our keylist,
supporting ASCII and Latin-1.
(cherry picked from commit 0a35b97b62
with help of https://github.com/Faless/godot/tree/js/keyCode)
Each driver used to define the (same) project settings value, but the
setting names are not driver specific. Ovverriding is still possible via
platform tags.
(cherry picked from commit 90c7102b51)
Affects per-pixel transparency
The current method renders to the screen by copying the GLES output to a
DIB for transparency using the CPU instead of rendering directly to the
window via the GPU. This is slower and also forces the window to be borderless
as WS_EX_LAYERED affects the non-client region as well.
This change uses DWMEnableBlurBehindWindow which allows using the standard
glClearColor() background alpha and is also performed through the GPU,
eliminating CPU bottlenecks
Similar to https://github.com/godotengine/godot/pull/36557
At least in chrome, the following error is printed for each mouse wheel
rotation:
[Intervention] Unable to preventDefault inside passive event listener due to target being treated as passive. See https://www.chromestatus.com/features/6662647093133312
This PR moves the handler to the canvas and thereby fixes the error.
Tested on: Chrome and Firefox (MacOS), Firefox, Chrome(Android), Safari (IPad + MacOS)
(cherry picked from commit b1e8ac7b08)
The previous logic used the 'tools' directory within the Android sdk to validate it. That directory was recently deprecated and removed from the Android sdk folder (https://developer.android.com/studio/releases/sdk-tools)
(cherry picked from commit 328354f878)
The issue was caused by PR #36906 which changes prevented the generated shared libraries from being stripped.
Since the change is only needed for development (debugging) purposes, it's commented out by default.
(cherry picked from commit 2f38cfd9ab)
This is the only location in the codebase where it's being used, so no need to make the main lib have a dependency on it.
(cherry picked from commit c591cb8fda)
Right now, games only work on devices when exported with FullAOT+Interpreter.
There are some issues left that need to addressed for FullAOT alone. Right now,
it's giving issues with the Godot.NativeCalls static constructor.
With the NDK installed locally, gradle plugin 3.6.0 seems to enforce
a specific older NDK version, and will fail building if you don't have
it installed with:
```
No version of NDK matched the requested version 20.0.5594570.
Versions available locally: 21.0.6113669
```
Upstream issue: https://github.com/gradle/gradle/issues/12440
(cherry picked from commit ba2ec53a26)