Editor module specific rules¶
The editor module has not been maintained aggressively about a decade. Therefore, this module needs to be treated as a young module or in a transition period to align the behavior to the other browsers and take modern C++ style.
Undoubtedly, this editor module is under rewritten for modern and optimized for current specs. Additionally, this module does really complicated things which may cause security issues. Therefore, there are specific rules:
Treat other browsers behavior as standard if the behavior is reasonable¶
The editing behavior is not standardized since as you see too many lines in the editor classes, the number of cases which need to handle edge cases is crazy and that makes it impossible to standardize. Additionally, our editor behavior is not so stable. Some behaviors were aligned to Internet Explorer, some other behaviors were not for making “better” UX for users of email composer and HTML composer which were in SeaMonkey, and the other browser engines (Blink and WebKit) have same roots but the behavior is different from both IE and Gecko.
Therefore, there were no reference behavior.
In these days, compatibility between browsers becomes more important, and fortunately, the behavior of Blink (Chrome/Chromium) which has the biggest market share is more reasonable than ours in a lot of cases. Therefore, if we get web-compat issue reports, we should align the behavior to Blink in theory.
However, if Blink’s behavior is also odd, this is the worst case. In this case, we should try to align the behavior to WebKit if and only if WebKit’s behavior is different from Blink and reasonable, or doing something “better” for hiding the issue from web-apps and file an issue to the Editing Working Group with creating a “tentative” web-platform test.
Don’t make methods of editor classes public if they are used only by helper classes¶
Although this is a silly rule. Of course, APIs of editor classes need to be public for the other
modules. However, the other methods which are used only by helper classes in the editor module –the
methods may be crashed if called by the other modules because editor classes store and guarantee the
colleagues (e.g., Selection
) when it starts to handle an edit action (edit command or
operation)– does not want to do it for the performance reason. Therefore, such methods are now
declared as protected methods and the caller classes are registered as friends.
For solving this issue, we could split the editor classes one is exported and the other is not exposed, and make the former to proxies and own the latter. However, this approach might cause performance regressions and requires a lot of line changes (at least each method definition and warning messages at the caller sides). Tracked in bug 1555916.
Steps to handle one editor command or operation¶
One edit command or operation is called “edit action” in the editor module. Handling it starts
when an XPCOM method or a public method which is named as *AsAction
. Those methods create
AutoEditActionDataSetter
in the stack first, then, call one of CanHandle()
,
CanHandleAndMaybeDispatchBeforeInputEvent()
or CanHandleAndFlushPendingNotifications()
.
If CanHandleAndMaybeDispatchBeforeInputEvent()
causes dispatching beforeinput
event and if
the event is consumed by the web app, it returns NS_ERROR_EDITOR_ACTION_CANCELED
. In this case,
the method can do anything due to the beforeinput
event definition.
At this time, AutoEditActionDataSetter
stores Selection
etc which are required for handling
the edit action in it and set EditorBase::mEditActionData
to its address. Then all methods of
editor can access the objects via the pointer (typically wrapped in inline methods) and the lifetime
of the objects are guaranteed.
Then, the methods call one or more edit-sub action handlers. E.g., when user types a character
with a non-collapsed selection range, editor needs to delete the selected content first and insert
the character there. For implementing this behavior, “insert text” edit action handler needs to call
“delete selection” sub-action handler and “insert text” sub-action handler. The sub-edit action
handlers are named as *AsSubAction
.
The callers of edit sub-action handlers or the handlers themselves create AutoPlaceholderBatch
in the stack. This creates a placeholder transaction to make all transactions undoable with one
“undo” command.
Then, each edit sub-action handler creates AutoEditSubActionNotifier
in the stack and if it’s
the topmost edit sub-action handling, OnStartToHandleTopLevelEditSubAction()
is called at the
creation and OnEndHandlingTopLevelEditSubAction()
is called at the destruction. The latter will
clean up the modified range, e.g., remove unnecessary empty nodes.
Finally, the edit sub-actions does something while AutoEditSubActionNotifier
is alive. Helper
methods of edit sub-action handlers are typically named as *WithTransaction
because they are
done with transaction classes for making everything undoable.
Don’t update Selection immediately¶
Changing the ranges of Selection
is expensive (due ot validating new range, notifying new
selected or unselected frames, notifying selection listeners, etc), and retrieving current
Selection
ranges at staring to handle something makes the code statefull which is harder to
debug when you investigate a bug. Therefore, each method should return new caret position or
update ranges given as in/out parameter of AutoRangeArray
. Result<CaretPoint, nsresult>
is a good result type for the former, and the latter is useful style if the method needs to keep
Selection
similar to given ranges, e.g., when paragraphs around selection are changed to
different type of blocks. Finally, edit sub-action handler methods should update Selection
before destroying AutoEditSubActionNotifier
whose post-processing requires Selection
.
Don’t add new things into OnEndHandlingTopLevelEditSubAction()¶
When the topmost edit sub-action is handled, OnEndHandlingTopLevelEditSubAction
is called and
it cleans up something in (or around) the modified range. However, this “post-processing” approach
makes it harder to change the behavior for fixing web-compat issues. For example, it deletes empty
nodes in the range, but if only some empty nodes are inserted intentionally, it doesn’t have the
details and may unexpectedly delete necessary empty nodes.
Instead, new things should be done immediately at or after modifying the DOM tree, and if it
requires to disable the post-processing, add new bool
flag to
EditorBase::TopLevelEditSubActionData
and when it’s set, make
OnEndHandlingTopLevelEditSubAction
stop doing something.
Don’t use NS_WARN_IF for checking NS_FAILED, isErr() and Failed()¶
The warning messages like NS_FAILED(rv)
does not help except the line number, and in the cases
of that we get web-compat reports, somewhere in the editor modules may get unexpected result. For
saving the investigation time of web-compat issues, each failure should warn which method call
failed, for example:
nsresult rv = DoSomething();
if (NS_FAILED(rv)) {
NS_WARNING("HTMLEditor::DoSomething() failed");
return rv;
}
These warnings will let you know the stack of failure in debug build. In other words, when you investigate a web-compat issue in editor, you should do the steps to reproduce in debug build first. Then, you’d see failure point stack in the terminal.
Return NS_ERROR_EDITOR_DESTROYED when editor gets destroyed¶
The most critical error while an editor class method is running is what the editor instance is
destroyed by the web app. This can be checked with a call of EditorBase::Destroyed()
and
if it returns true
, methods should return NS_ERROR_EDITOR_DESTROYED
with stopping handling
anything. Then, all callers which handle the error result properly will stop handling too.
Finally, public methods should return EditorBase::ToGenericNSResult(rv)
instead of exposing
an internal error of the editor module.
Note that destroying the editor is intentional thing for the web app. Thus we should not throw exception for this failure reason. Therefore, the public methods shouldn’t return error.
When you make a method return NS_ERROR_EDITOR_DESTROYED
properly, you should mark the method
as [[nodiscard]]
. In other words, if you see [[nodiscard]]
in method definition and it
returns nsresult
or Result<*, nsresult>
, the method callers do not need to check
Destroyed()
by themselves.
Use reference instead of pointer as far as possible¶
When you create or redesign a method, it should take references instead of pointers if they take. This rule forces that the caller to do null-check and this avoids a maybe unexpected case like:
inline bool IsBRElement(const nsINode* aNode) {
return aNode && aNode->IsHTMLElement(nsGkAtoms::br);
}
void DoSomethingExceptIfBRElement(const nsINode* aNode) {
if (IsBRElement(aNode)) {
return;
}
// Do something for non-BR element node.
}
In this case, DoSomethingExceptIfBRElement
expects that aNode
is never nullptr
but it
could be at least in build time. Using reference fixes this mistake at build time.
Use EditorUtils
or HTMLEditUtils
for stateless methods¶
When you create a new static method to the editor classes or a new inline method in cpp file which
defines the editor classes, please check if it’s a common method which may be used from other
places in the editor module. If it’s possible to be used only in HTMLEditor
or its helper
classes, the method should be in HTMLEditUtils
. If it’s possible be used in EditorBase
or
TextEditor
or their helper classes, it should be in EditorUtils
.
Don’t use bool argument¶
If you create a new method which take one or more bool
arguments, use enum class
instead
since true
or false
in the caller side is not easy to read. For example, you must not
be able to understand what this example mean:
if (IsEmpty(aNode, true)) {
For avoiding this issue, you should create new enum class
for each. E.g.,
if (IsEmpty(aNode, TreatSingleBR::AsVisible)) {
Basically, both enum class
name and its value names explains what it means fluently. However, if
it’s impossible, use No
and Yes
for the value like:
if (DoSomething(aNode, OnlyIfEmpty::Yes)) {
Don’t use out parameters¶
In most cases, editor methods meet error of low level APIs, thus editor methods usually return error
code. On the other hand, a lot of code need to return computed things, e.g., new caret position,
whether it’s handled, ignored or canceled, a target node looked for, etc. We used nsresult
for
the return value type and out parameters for the other results, but it makes callers scattering a
lot of auto variables and reusing them makes the code harder to understand.
Now we can use mozilla::Result<Foo, nsresult>
instead.