Block args set by fields are static and never change. Inputs that do
change are always set onto args. With these two assumptions we can
reuse the same objects for each execution of the same block instead of
constantly creating them and letting them be garbage collected.
Use a private StackFrame class to help internally manage use of Stack
Frame values and memory. Create params, reported, and executionContext
on demand.
Explain how BlocksExecuteCache helps speed up execute by storing values from
Blocks and one-time derived values on an object that is released when Blocks is
edited in the editor. The one time created cache object for a block then
simplifies the complexity of later javascript operations to use these stored
values.
The new `maybeFormatMessage` function detects whether its argument
looks like a message descriptor object and, if so, will call
`formatMessage` on it. This is now used for all user-visible text fields
in extensions.
Also, messages may use "select" to check the target type with a message
like this: '{targetType, select, stage {text for stage} sprite {text for
sprite} other {text for other}'. Note that the "other" clause is
required by `formatMessage`.
A LOOP block is like a conditional, but the LOOP block will be
re-evaluated after any child branch runs.
Also:
- Support using '---' as a block separator
- Refactor common code from `_registerExtensionPrimitives` and
`_refreshExtensionPrimitives` into new `_fillExtensionCategory`
- Improve error reporting during block conversion
This involves adding a new opcode, event_whenstageclicked, and adding a
method to the blocks container to update the opcode pair depending on
whether the target is the stage. This method is then called in two
places: first after the sb2 import parses the blocks (not done in the
sb2 parsing itself because no other blocks are target specific) and then
again when blocks are shared between targets.
Also added tests for the block container method, and a fixture project
that tests the opcode splitting on sb2 import.
The previous logic seemed to be expecting `peekStack` to return 0 when
the stack was empty, but its return type is `?string` and it returns
`undefined` when the stack is empty. The `!== 0` check would never pass,
then, leading to problems with (for example) the "say/think and wait"
blocks.
Towards fixing #865. This adds an IO class for detecting the mouse wheel
being scrolled. Basic tests are included; they mock the runtime to see
what blocks are activated by scrolling.
Add a Blocks cache available only to execute. This cache lets execute
get a blocks inputs, fields, opcode, and mutation in one request and an
object execute can further modify to store derivative values it use
this and every later execute iteration to perform its duties quicker.
Before calling execute, if a thread's target is null, retire that
thread.
This saves repeatedly checking if the thread's target is null in
recursive calls where even if scratch-gui or blocks, or some other
related library set the target to null, that will not happen during
block execution. It will happen at some time outside of the sequencer
letting the sequencer check once instead of execute checking at every
recursive level.
Store the thread's blocks at blockContainer letting execute quickly
determine the block source. Monitor threads are a monitor thread. They
do not become a target thread suddenly.
When execute calls itself to step into the next stack level, pass
RECURSIVE to the recursiveCall level. This argument provides
opportunity to recursed calls in execute to reduce the time needed for
some checks.
- Reduce time for checking around setting
thread.requestScriptGlowInFrame
- Call thread.pushReportedValue at end of execute when in a recursive
call
A boolean check is faster than a type lookup, command blocks will only
be executed in non-recursive calls. This saves a minor amount of time
for any result reporting blocks like blocks that get variables or add
two inputs together.
If the block is not returning a promise and execute is a recursed call,
that implies that the block is neither a hat or that the thread is not
at the top. That reduces handleReport to pushReportedValue. Using the
recursiveCall argument the final block calling handleReport can
shortcut the extra work in handleReport and reduce it to immediately
calling pushReportedValue.
- Add stackFrame.justReported
Report the block result to a known key `justReported` instead of a
dynamic key on `reported`. Assuming blocks with a promised value are
relatively infrequent the most common recursive input flow immediately
reads the value "just" reported. In the assumed uncommon case of a
promised thread status, empty the already argValues assigned values
onto the currentStackFrame's reported member. In the next execute call
on this stackFrame, values assigned to reported are read back off onto
argValues, and execute will returned to the assumed common case. This
is a safe assumption since a thread in the promise state will not exit
that state until the next frame when javascript has a chance to call
the resolve handle, setting the thread's state back to another
executable state.
Using direct assignment to `justReported` saves building an object
dynamically. Instead of always building `reported` and `argValues` only
`argValues` is built until a promised state is reached. This also gives
a known time when `reported` is used, allowing cleanup of a
stackFrame's reported to only happen when it was used to persist
already reported values.
* localize the block and menu strings in the pen extension
* adds .tx/config to be able to push translations to transifex
* includes format-message to localize strings and extracting them.
* add setLocale function to VM to allow GUI to pass in locale data.
* refresh block definitions when the locale changes.
### Still to be decided
For now just extracting messages from the pen extension into their own file. We’ll need to decide if each category gets its own file, or group all the strings into one resource.
* Cache Block Inputs & Procedure Definitions
* @mzgoddard requested changes on this pull request
* Move all caching into a single reusable field _cache.
* Invalidate 'all' caches on any change.
* Use 'typeof' instead of hasOwnProperty.
* Take caching out of Block and use lookup instead.
Before: `_removeThread` stops a thread and immediately removes it from
the runtime's thread array. If this happens while iterating over the
thread array, such as in the case of clone deletion, some elements of
the array can be missed.
After: `_removeThread` has been renamed to `_stopThread`. It still stops
the thread immediately but it no longer removes the thread from the
runtime's thread array. Instead, `_stopThread` marks the thread for
later removal. Such threads are removed at the top of the next `_step`.
return false from isActiveThread if thread's status is done even when
its contained in the runtime threads list. As it is done, the thread
will not be run until either its replaced by a new copy of the thread
or the thread is removed from the list and another is added at a later
time.
blockUtility as an object literal inside execute creates 11 objects,
one for the object, and 10 for the function closures. As a separate
object and allocated once, setting its sequencer and thread members,
block functions can share the same util object. Extract blockUtility to
cut down on allocations.
`handleReport` inside `src/engine/execute.js`'s `execute` method needs
to allocate a closure to be able to refer to the higher scoped
variables. `execute` is called frequently that this has a noticable
impact on memory allocation and later collection. Extract handleReport
and add arguments to prevent the allocation.
Testing with implicitly casted `value` and `value.then` is deoptimizing
isPromise, which deoptimizes execute. In this case deoptimizing means
that the JavaScript VM cannot compile the functions into a form that
can be run faster.
Remove indexOf tests for thread existence in doneThreads. Maintain a
list of null and thread objects mirroring the index position of threads
that are done in runtime.threads. Filter out null values after
filtering out done threads from runtime.threads.
This has a small side effect that threads that normally became DONE and
were added to doneThreads before being removed or restarted will not be
in doneThreads in this version. Restarted threads (_restartThread) do
not appear in this version but new copies will be in Runtime
this.threads supporting glow and monitor updates. Removed threads
(_removeThread) do not appear in this version if they are removed after
they were seen as DONE by the prior version. Threads removed before
they are seen as DONE do not appear in doneThreads in the prior or this
version.
Threads that are removed before normally becoming DONE do not appear in
doneThreads in either case. Threads that are restarted before the loop
checks if the thread is done do not appear in doneThreads in either
case.
Due to a typo (I believe) we were overwriting a horizontal hat block's
fields list when collecting hat block inputs. Now we collect inputs into
a temporary object in this case.
Dynamic menus are not yet supported for extension menus, though they are
part of the extension spec. This change provides more thorough feedback
if an extension tries to register an unsupported dynamic menu.
- Remove WeDo 2 extension from the runtime's default block packages list
- The WeDo 2.0 extension now calls its own `connect` method on startup
I also renamed `EXTENSION_NAME` to `EXTENSION_ID` for consistency with
the rest of the extension system.
This prevents generation of invalid XML due to characters like '<' or
'>' in fields' default values. Unfortunately the value comes back in its
escaped form, so there's still more work to be done.
The linter is correct that `devObject[func].call(devObject, args)` is unnecessary because it's equivalent to `devObject[func](args)`.
@tmickel probably either meant to allow passing an array of arguments to be applied, or to call the function with the provided argument. Since we probably want to be able to use multi-argument functions, use `apply` and fix the one place that uses `ioQuery` with an argument.
Problem:
When a clone is deleted, its thread is removed from the thread
list; but the index "i" in the execution for-loop will still +1,
making the next thread skipped.
Changes:
Added a flag "isKilled" on a thread;
Set the flag when a thread is removed from the thread list;
reduce the "i" counter if the thread of the current loop is removed.
the right quote is mission in blockToXML, and it will result in the following code execution errors
vm.on('workspaceUpdate', function (data) {
workspace.clear();
var dom = Blockly.Xml.textToDom(data.xml);
Blockly.Xml.domToWorkspace(dom, workspace);
});
* Add `getTargetIdForDrawableId` to translate between renderer picks and VM targets
* Add `startDrag` to stop sprite motion while dragging
* Add `stopDrag` to return sprite motion after dragging
And it starts to get a little less elegant :/
Wondering if this should not be handled better in another part of the
codebase?
We don't want to be duplicating existing code stepping functionality
locally at the end of the promise script really... What do you think?
The `allStacksAndOwnersDo` function in Scratch 2.0 runtime iterates
targets in reverse and projects sometimes rely on that for correct
initialization. If, for example, each sprite runs a "go back 999 layers"
or "go to front" block as its first action, the order of execution will
determine the ordering of the layers.
This change makes Scratch 3.0 match the Scratch 2.0 execution order.
