Previously we were just passing a slice of the visible data, instead
of using getFullData.
Bug: 50059
Change-Id: I3bbea49132ef4a720a147ba9b170c39a0c00f711
Previously we would reserialize the JSON blob in data-mw even if we
didn't change it, which potentially reordered keys and caused a DOM
diff.
Bug: 50066
Change-Id: If0a5bcc67d3a172de0e8839cfda11efacfbf36ff
Fixes (follows-up I1b48ef5240, I6daff5c596):
* Invalid html passed to jQuery constructor.
* Use prop() instead of attr() for boolean values.
* Use append() instead of html() when appending nodes instead
of parsing html.
* Rename shadowed variable name clash 'mw' to 'mwData'.
* Fix odd construction where we parse '{}' to create an empty
object.
* Have ve.ce.MWReferenceListNode#update perform changes off-document
in a detached tree.
* Fix deep property access that can fail. mwData is set to
either JSON parse of data-mw attr or empty object.
Accessing mwData.attrs.group needs to be guarded by whether
mw.attrs is indeed set.
* Have `mw` and `about` attribtue in references list roundtrip
(especially mw which can data we aren't editing/re-creating).
* Add missing 'refGroup' property to MWReferenceListNode's
data element (similar to what MWReferenceNode already has).
Change-Id: I67e4f378ccd04e97361d8e58ae57db5353075756
Better for performance (lower level, #get calls #toArray when
called without arguments) and more semantically correct.
Change-Id: I8e615674e51fd148367fd397bd169fa203a240ad
At the moment we create an alien meta item if an annotation
wraps contents of zero length, however we also need to handle
nested empty annotations, i.e. where the contents is other
empty annotations (i.e. alien meta items).
As alien meta items simply preserve DOM elements we don't
mind what type of alien meta items they are, so this also fixes
the case of an annotation wrapping another alien meta
e.g. a comment or <meta> tag.
Bug: 48605
Change-Id: I98d88b341efcd6384ee1fda5d6d7e1e1acb107aa
This function builds a transaction that takes a document slice and
inserts it back into the document it came from, applying any changes
that were made.
This makes editing document slices simple:
slicedDoc = doc.getDocumentSlice( captionNode );
// Edit slicedDoc using a surface
tx = ve.dm.Transaction.newFromDocumentInsertion( doc, captionNode, slicedDoc );
surface.change( tx );
Specifically, newFromDocumentInsertion replaces the node's contents
with the document's contents (meaning any changes made to the node in
the meantime are lost). It also merges the stores internal lists
of the two documents and remaps indexes accordingly. This means editing
of references inside of references is supported.
This functionality is not specific to slices, and can also be used to
safely insert data from a paste buffer, with internal list data being
transplanted correctly.
ve.dm.MetaLinearData:
* Make merge( [ undefined, undefined, ... ] ) return undefined rather
than [].
ve.dm.Document:
* In getDocumentSlice, store a pointer to the original dm.Document in
the new one, and also store the length of the internal list. This
allows us to figure out which internal list items the two documents
have in common when we insert the modified slice back into the main
document.
* In getMetadataReplace, optionally take the inserted metadata as a
parameter, to allow for operations that insert both data and metadata.
Per Ed's review, rewrite this function to return null rather than {}
if no metadata needs to be replaced.
ve.dm.InternalList:
* Add method to merge two internal lists
ve.dm.Transaction:
* Remove newFromNodeReplacement and replace it with newFromDocumentInsertion.
* In pushReplace, optionally take the inserted metadata as a parameter.
Change-Id: I786ee7bad796aa54bc242993b4de3ad18ad0773e
Added return boolean to ve.dm.Surface#breakpoint to indicate if
and transactions were pushed to the big stack. Use this value in
MWMetaDialog to see if we need to undo on close.
Added tests for ve.dm.Surface#breakpoint.
Bug: 49630
Change-Id: Ieb2e9e361afe057af93c4d374acc85df58bfb4c3
Slightly more efficient way of seeing if the template has changed,
and results in a cleaner set of data element attributes.
Change-Id: I1507520005bfb8a88bfa6038dac5c3b15506425d
These kinds of empty nodes shouldn't occur since the converter fills
them with empty paragraphs, but selectNodes() should still behave
correctly for them.
Change-Id: Ia37f3db1c2a84b842e2311cf70642fa66af04d91
* Move them from ve.example.js to ve.dm.example.js
** Also move lookupNode() and createDomElement() there
** Delete ve.example.js because there's nothing left in it
* Make main example document implicit, but allow override
* Specify nodes as arrays and do the lookup later
* Specify range and mode separately rather than calling selectNodes()
during construction
* Construct statically rather than in a function
* Use expect( cases.length );
Change-Id: I620e949c5e612b32eaa57c5d9b60cc91f9ddbf02
Otherwise fun exceptions occur when you have a <references /> tag
with no <ref>s, for instance.
Also disregard the internalList in the data->DOM conversion, to prevent
a nasty interaction where the whitespace information on the last element
is considered invalid because it doesn't match the internalList's.
Plus test updates from hell because this touches ve.dm.example.data
Change-Id: I62881d9fc27fa081123856d1b35a6021af469271
Objective:
* Allow editing reference groups and names in the reference dialog
Bonus:
* Modify attribute transaction builder to support multiple attribute
changes in a single transaction
Changes:
ve.ui.MWReferenceDialog.js
* Load ref name and group from model
* Save ref name and group, if changed, to model
ve.ui.ListAction.js, ve.ui.Transaction.test.js, ve.ce.ResizableNode.js
* Update use of newFromAttributeChange to newFromAttributeChanges
ve.dm.SurfaceFragment.test.js
* Add test for new changeAttributes method
ve.dm.InternalList.js
* Missing new line at end of file
ve.dm.Transaction.js
* Change newFromAttributeChange to accept an list of attribute changes and
produce a single transaction that applies one or more attribute changes
at once
ve.dm.SurfaceFragment.js
* Fix bug in getCoveredNodes where the wrong mode name was being used
* Add changeAttributes method, which applies attributes to all covered
nodes and allows filtering of which types of nodes the attributes are
applied to
ve.dm.MWReferenceNode.js
* Actually write key and group back to DOM
* Separate onRoot functionality into addToInternalList so it can be called
separately (similarly onUnroot/removeFromInternalList)
ve.ce.MWReferenceListNode.js
* Clone internal item CE node before appending to avoid rendering bug.
*.php
* Add links to messages and sort them
Change-Id: Ic4121e4fcfc09265d5863af6f078cdeb77926c8e
Per the bug report, it's useless because Firefox does not preserve order
in the .attributes array (but sorts it alphabetically instead), and so we
actually get the same exact behavior regardless of whether we use ordering
cleverness or just a straight-up plain object: order is preserved in Chrome,
and alphabetized in Firefox.
Bug: 48980
Change-Id: I1463d06db4900083dd4c565292bbabef09194b43
Objectives:
* Rename just about every use of "template" to "transclusion"
* Make a proper data structure for transclusions
* Abstract away template data
* Use more template data in the user interface
* Allow adding parameters
* Allow removing templates, parameters and content
Changes:
ve.ui.Dialog.css
* Add rule to place add param controls on a single line
ve.ui.MWTemplateDialogs.js
* Move template spec loading into transclusion class
* Add remove button for parts and parameters
* Add parameter adding form
* Use template data for labels and descriptions
ve.dm.*
* Add new transclusion data structures
*.php
* Add links to new files
*.*
* Rename all things "template" to "transclusion"
Bug: 39598
Bug: 49403
Change-Id: I3bcf924a3e179cb65f19e833277a39dfd3dad8bd
We don't need to decompose and rebuild alien meta items
as they are uneditable, so just store the dom node and return that
in toDomElements.
Update test cases and several tests now need to use
assert.deepEqualWithDomElements.
Change-Id: I4d2eed267dac7855cb929a17d7aef566eaf0e287
Objective:
* Allow opening reference dialog with arbitrary selection
* Auto-insert reference when selection is not a reference node
Changes:
ve.init.mw.ViewPageTarget.js
* Added reference button to toolbar
ve.init.Target.js
* Add getToolbarSubset so we can exclude the reference button from
the toolbar in the reference dialog (nested references are not
allowed).
ve.ui.MWReferenceDialog.js
* Stop storing referenceNode (not needed)
* Only store internalItem on open if there's a focused node that's a
reference
* Use wrapper paragraph when creating a new reference
* Create new reference on dialog close if required
ve.dm.InternalList.js
* Major rewrite to support key less references.
* Add new method for creating a transaction to insert a new iternal
item. Also returns the index of the new item to be passed to the
reference node.
Change-Id: I839ae165c299248484ce93d4ab087318a95fbb94
Previously we populated the reference body into all <ref> tags
with the same key. Now we store an internal attribute marking
which element originally had the data.
If that tag is deleted the body is moved to the first <ref> tag
with that name.
Change-Id: If9f12bfb699e6ce85bb8f7d2ea9e6df528610a3d
The easy part is getting the correct numbers from the InternalList
and generating the ordered list HTML. The tricky part is connecting
up the events to make sure the renumberings/list generations are
triggered when required.
InternalList can emit an update event on document transaction, which
triggers the renumbering/relisting if any references have been
added or deleted during that transaction.
ve.ce.MWReferenceListNode also listens to changes on the
InternalListNode (i.e. changes to the contents of the references)
and always triggers a rebuild.
Change-Id: I1b48ef5240e433c3b314259aa68cde13841ea98b
Due to the "es5: true" jshint option we enabled, these
warnings were surpressed. I've disabled the option since
we no longer require it. It was enabled in 07c86fc to fix
a bug with jshint. This bug has now been fixed.
Change-Id: I55b7d031eb5581af5f733f050cf2ea98dacb2af6
Also keep items in the order they appear in the document
and grouped by group and key.
Additions and removals are triggered by the new root/unroot events.
Change-Id: Ia3e90ccfdab88f352b89992b90554e5f03ff9952
The only reason it doesn't save correctly right now is because Parsoid
doesn't serialize reference edits correctly.
Change-Id: Ia0f272c07cc28ee829372eb848f23aec99eb92f0
A document slice is a document built from a data slice of an existing
document. It's completely separate from the original document and has
its own store and internalList. The new document's data also contains
the entirety of the original document's internal list. It's possible
to create a document slice of data located inside the internal list,
in which case the resulting document will contain that data twice (one
mutable copy at the top level, and one immutable copy in the internal
list).
ve.dm.Document.js:
* Optionally take an internalList in the constructor. This allows us to
create a document with a clone of an existing internalList rather than
an empty one.
* Add edgeMetadata flag to getFullData()
ve.dm.IndexValueStore.js, ve.dm.InternalList.js:
* Make these classes cloneable
Change-Id: I93e06f764ace16aee9df941b07f8c2bff1a28e2b
Getting & setting the cursor is done with byte offsets
instead of data model offset (characters) so we need to
be able to convert between the two as well as splitting
characters.
TODO: The regex only works on surrogate pairs, not
yet combining accents.
fixupInsertion will combine a combining mark with the
character to its left it it can.
Bug: 48630
Change-Id: I8d936fb15d82f73cd45fac142c540a7950850d55
Right now the internal list is very much HTML-based, so it's not
populated at all when creating a document from data. So hack this in by
specifying what should be in the internal list as a property of the
example document's data array.
Change-Id: I51c7b7b4dcd9fd3333777c1287b7ba544887aa32
* VisualEditor.hooks.php
** export URL to magnify clip as a configuration variable so thumbnail image can render correctly
* ve.ce.Surface.css
** add CSS styling to make image captions look the same way in edit mode as they look in the view mode
* ve.ce.ParagraphNode.js
** add CSS class ve-ce-generated-wrapper if it is a generated wrapper
* ve.ce.MWImageCaptionNode.js
** make image caption rendering match the view mode
Change-Id: I0cd1b25e8f8355e0500aabc90e7c4cdf591545f3
Previous check wouldn't make sense, cause the last offset in the data
could be that one that makes data balanced (and j is increased always
after iteration).
Change-Id: Ie9498d0ac9e3417d09b8b3043bf3281e7dfbf9db
Rather than using namespaced linmod attributes, store the preserved
HTML attributes in the .htmlAttributes property of the linear model
element, in a nested structure to allow for easier treatment of child
nodes. Also added attribute order preservation by storing attributes
as an object plus an array of keys.
ve.ce.Node.js:
* Remove html/* attribute synchronization. Doesn't make sense any more
because these things aren't in the attributes object any more. I don't
think it ever made sense because these attributes were never supposed
to be changed anyway.
ve.ce.View.js:
* Replace renderAttributes() with a simple wrapper around
renderHtmlAttributeList()
ve.dm.Converter.js:
* Add buildHtmlAttributeList() and renderHtmlAttributes() for building
and rendering HTML attribute lists
ve.dm.Model.js:
* Add getter for .htmlAttributes
ve.dm.Node.js:
* Drop .htmlAttributes on clone, and remove logic dropping html/*
ve.ui.MWCategoryWidget.js:
* Remove html/0/about hack, was already unnecessary and now doesn't
work any more
tests/:
* UPDATE ALL THE TESTS
Change-Id: I620573afd70d36ade6b80413075b6e1f4a435abe
Escapes }} and | by wrapping them in <nowiki> tags, and steps around
<nowiki> tags so as to not double-wrap things.
Change-Id: Ia50906524c1fcf55c9f390a114856bc7f20b2d3a
In stopWrapping we assign any left over whitespace to the paragraph
in position 3, however we weren't clearing this whitespace buffer
if an inline content node followed it.
Change-Id: I8b3ee3915044abd6bafda386430bf7f992ca4aa8
The way it operated was evil. It did a depth-first search from the root,
finding the node using reference equality. For documents with deep
structures, this could take a long time. Inez did some profiling and
found it was called tens of millions of times on a complex document.
Kill getOffsetFromNode() and move its functionality to getOffset().
The logic has been completely rewritten: getOffset() now traverses
up from the node rather than down from the root, and pretty much does
the reverse of what getNodeFromOffset() does. This should be much more
efficient even without offset caching in the node objects (which we may
still implement later).
Change-Id: I125f9fa423c40db6472e2c4a7c94214218ba3bc7
For nodes that handle their own children (as well as leaf nodes and
meta items), store the first child's attributes in html/0-0/*, the
second child's attributes in html/0-1/*, the second element's third
child's fourth child's attributes in html/1-2-3/* , etc.
This obsoletes the ad-hoc code that basically did the same thing in
MWInlineImageNode.
Change-Id: If5abd2d5d9c361b359617ff4b0f3d6ba4c9b0142
Store the HTML as a domElements array like everywhere else, rather than
as a string. Also disable HTML attribute preservation because there's
no point doing that when we're already preserving all of the HTML.
Also fixed a misnamed attribute (<li li="foo"> --> <li id="foo">) in
the test case.
Change-Id: I36bf8bade8118e07a75eb6f3a2427a00ef4915d7
It now allows you to specify which attributes to preserve in various
ways rather than just setting true or false.
Removed unused factory methods that exposed the old value.
Change-Id: I914164adcf1f0e48fa3fa85277e68c72dbad393e
For extension-specific types such as mw: , we require that all
types be matched. But we want MWTemplateNode to match
anything with an mw:Object/Template type, even if it also has
other types (like mw:WikiLink/Category in our test case).
Hack this into MWTemplate by matching on /^mw:/ then using
a matchFunction to assert that mw:Object/Template is in the
typeof attribute.
Update the test case. Because it's now a template, there's a bunch
of store stuff involved. Remove the other test case for about
group forcing because it's now a duplicate of this one.
Change-Id: Iacbe952a66d610c19b46bd76b84c50488857ac29
When converting an element that starts an about group with at
least one other element in it, we now only consider models that
support about grouping. This prevents the first node from being
converted to something else and leaving the others hanging.
In practical terms, this means that elements like
<link rel="mw:WikiLink/Category" typeof="mw:Object/Template">
get alienated and pull in the rest of their about group, rather than
being converted to a category or alienMeta or whatever and
leaving the other elements to be converted normally.
Added a test case that asserts this. Really the result should be an
MWtemplate rather than an alien, but that's a separate issue.
Also removed superfluous mustMatchAll checks; we've already
filtered the array by the time we get there.
Change-Id: I522ba4c56d5bc52c7e9aab1e2535385540c1315d
This was broken for both normal elements, where the meta item is
inside the element, and wrapper paragraphs, where the meta item
gets moved outside the wrapper.
Bug: 47712
Change-Id: I42daaf142e548e5b221ff0a52df0ad24ec6a4fd0
Specifically by looking for "data-ve-changed",
"ChangeMarker*" and internal.changed.
Various tests, test counters and unused variables also
affected.
Bug: 45061
Change-Id: Ibd1ee68e0d650979d40574eff9cebded1a28499f
Remove all manual changes to SF ranges as these are not
undoable. Instead change translate range to default to
outer expand and build functionality around that behaviour
never changing.
As translate range is always outer I don't think we need to
check for start and end crossing over?
Added more undo tests to assert these selections are maintained
properly, and added the test case to 'update' for when and undo
point is overwritten.
Insert content now results in a selection over the inserted
content. Most usages were expecting this anyway and were
followed up with an adjustRange(-length,0) which is no longer
necessary.
Noticed that the link inspector case was never being triggered
as word boundary was always expanding to at least one char (mainly
for Hanzi selection). This doesn't make much sense as single
spaces get auto selected so removed this functionality.
Split collapseRange out into collapseRangeToStart and
collapseRangeToEnd as this may be required to get the old
behaviour (range moves to end after insert).
Change-Id: I3dc0b4d00d37bad1ca3076a69b41c5f0b3fa0570
As jQuery hash problems in some cases converting HTML, it is
easier just to store the original DOM elements.
The bulk of this commit is fixing the tests as although we have an
assertion for comparing DOM elements, we don't have one for comparing
objects or arrays which may contain DOM elements.
ve.js
* copyObject & copyArray fixed to run cloneNode(true) on any item
of type Node.
* Added callback function so an object/array can be copied with
modifications.
ve.qunit.js
* Added deepEqualWithDomElements: Using the new copyObect/Array
callback, we can copy the incoming object and convert any nodes
to node summaries, then just run the normal deep equal comparison.
ve.dm.AlienNodes.js
* Instead of storing HMTL we store cloned DOM elements which we can
send straight back to the converter without any processing.
ve.dm.example.js
* Updated tests to expect DOM elements instead of HTML.
ve.dm.Converter.test.js
* Updated tests to use deepEqualWithDomElements
Bug: 47737
Change-Id: I3df8f49b170c31da9610129d53cf8cb65dd5d5f8
In this case, selectNodes() returns the paragraph, which is not a content
node, and so newFromContentBranchConversion() decides to do nothing at all.
This change makes newFromContentBranchConversion() more intelligent about
finding the content branch to operate on, fixing cases such as these
where selectNodes() returns the content branch itself rather than one of
its children.
Bug: 41203
Change-Id: I710fbf184ef5ef84d9c2f5bca2b115e0660f5b8f
fixUpInsertion now returns an object with both data and offset
which allows offset changes.
Within fixUpInsertion we lazy-generate first/lastChildStack which
is a list of parent nodes for which the current node is the first/last
child. Whenever we try to close off a node we check these stacks and
if they are populated we instead use a recursive call to start
fixUpInsertion again but with the offset shifted by 1.
Bug: 46799
Change-Id: Ic51dd03725c11f1f7e279929534ee3afea14d662
Because we have a node for <table>, we also need one for <caption>,
otherwise we'll try to alienate it and fail.
Added the test case as a separate example document so Ed can use it
for his tests.
Removed test case asserting <caption> is alienated.
Change-Id: I3a917db58e6c0eb97899b214b07d01fc8d86b56d
In most places we call .contains we already know the index, so we
can avoid store lookups by using .containsIndex.
Change-Id: I45a9a421473f9bec479ab8ccceceb162b7004c3a
The fixUpStack is actually redundant code and closingStack
and openingStack handle all our cases. It was causing the
insertion to try to correct balance itself in the middle
of inserting two paragraphs, causing the creation on an
empty paragraph between them.
Added a test case for the fix and other cases to make
sure removing fixUpStack hasn't caused problems.
Bug: 46800
Change-Id: I35e54165709ac56e8116359a7c3b487eecf08ff7
To help the selective serialiser we can return the original
HTML for generated content if it is unmodified.
As the output of toDomElements now depends on changes
to the dataElement we now have a 'modify' function in
the some test cases.
We also now have 'storeItems' to assert that the index-value
store is correctly populated and for loading values back
into the store for toDomElements tests.
Also make 'mw' an attribute and remove 'about' property.
Bug: 47394
Change-Id: I2bbb5d2d6a90c4eb87fa129671112c92a9b931e7
Also add detection for whether the browser is actually broken (most are,
but some, like Opera, aren't), treat <textarea> and <listing> in addition
to <pre>, and fix a bug where the function would crash if the <pre> was
empty (because .firstChild was undefined/null).
Change-Id: I541b57e9fd5c9c42d19d0a59f6e29fb43d35c9b6
addSet:
* Instead of indexing items in the store, just union the indexStore arrays
removeSet/removeNotInSet:
* difference or intersect the indexStore arrays
filter:
* push indices into the result set instead of values
simpleArrayUnion/Intersect/Difference have been created as utilities
in ve. They are prefixed 'simple' because they use object keys to
do fast in-array comparisons. This means they are limited to string
values or values which will compare as strings (e.g. numbers).
Change-Id: I079cbdfece4f6d80ec0afd61959913f13217fcb3
By removing the transaction listeners from surface fragments we
no longer have to make sure they are always manually destroyed.
In order to retain the functionality of having fragments update
with transactions elsewhere we keep a pointer to a place in the
new complete history stack in the surface. The complete history
stack records all transactions, even undone ones.
Whenever getRange is called we replay all transactions in the
complete history (in the correct order) since the fragment was
last updated.
Also in this commit:
* Updated Format/IndentationAction to test undo(). This increases
coverage of surface fragment behaviour.
* .range is always accessed by .getRange now, although as an
optimisation we can use the noCopy mode when we a sure the
returned range will not be modified.
* Added undo test to .update (previously .onTransact)
Bug: 47343
Change-Id: I9e9818da1baa8319a3002f6d74fd1aad6732a8f5
This was broken, especially in wrappers.
Changed the wrapping algorithm so that meta items are placed outside
wrappers if possible. On the left-hand side, this is already the case:
we don't open wrappers for meta items. On the right-hand side, this is
accomplished by buffering the meta items and only inserting them when
we encounter either real text (not whitespace) or the end of the wrapper.
If we're interrupted by real text, we insert the meta items with the
unmodified whitespace. If we're interrupted by the end of the wrapper,
we insert the meta items outside of the wrapper with whitespace stripped.
Internally, this is done by stripping the whitespace into the whitespace[0]
of the meta item to its right. Then when we output the meta items, we
either decide to 'restore' the whitespace, or to 'fixup' by also setting
whitespace[3] on the element before the whitespace.
Change-Id: Ibeea2a9906c4aae9fe6d284613edd6ec853ca5e7
So. It turns out that the design of SurfaceFragment is a little -
shall we say - wonky.
One of the best things about ve.dm.SurfaceFragment is its magical
ability to retain the intention of its range, even as transactions
are being processed. This ability is granted by each fragment
listening to the surface's change event, and responding by using
translateRange for each transaction that gets processed. Surface
fragments also have these clever methods that allow you to get a
fragment based on another, which makes adjusting the range easy to do
inline without having to manually store multiple fragments or
modifying the original.
This sounded good, and we seemed to all be convinced it was well
designed. But if you add a console.log( 'hello' ); to the first line
of ve.dm.SurfaceFragment.prototype.onTransact, and then start using
the bold tool on various selections of text, you will find that there
may indeed be a flaw. What you will probably realize is that the
number of times that particular line of code is being called is
disturbingly large, and increases each time you do just about anything
in the editor. What's going on? How did we get here? Read on…
It turns out that fragments are immortal. We create them, they listen
to the surface's transact event, we are done with them, but the
surface keeps on emitting events to the now long forgotten about
fragments. They continue to build up over time, never go out of scope,
and bloat the hell out of our program.
The same ended up being true of toolbars - and each time the context
menu fired up a new one the old one was left in limbo, still
responding to events, still taking up memory, but not being visible to
the user.
All of this immortality was causing strange and difficult to track
down problems. This patch fixes this by introducing a destroy method.
This method unbinds events, allowing the object to finally fall out of
scope and die - and more importantly stop receiving notifications of
changes.
This is a hack, but Ed will no doubt get this situation sorted out
properly by making fragments lazy-evaluate their selections by only
storing an identifier of the most recent transaction they were based
on, see bug 47343.
Change-Id: I18bb986001a44732a7871b9d79dc3015eedfb168
Before, it took an array of objects and translated those to indexes
using the store. Literally every caller outside of the test suite got
an array of indexes from the linear model, translated those to objects,
then passed them into the AnnotationSet constructor which translated
them right back to indexes.
The previous behavior was kind of ridiculous on its face, but the
reason we found it is because Inez was investigating the performance
degradation when bolding a line and found that half of it was due
to the hundreds of ve.getHash() calls caused by this behavior.
Change-Id: I38df8ae9f6392849dacf477ea2f804283c964417
