Fixes#6320.
Part 1 of moving from lerna to a simpler, maintained monorepo solution,
using TypeScript Projects and NPM Workspaces.
There is more work to be done here. At this point, KeymanWeb builds and
runs without errors, but the built file is substantially different,
mostly in include order.
Using TypeScript Projects, we move away from the need to run build
scripts in various locations for almost all the Typescript modules.
TODO: Embedded versions and tests have not been verified.
TODO: developer/server is not yet verified.
TODO: developer/js (needs a rename!) is not yet verified.
TODO: Currently, the predictive-text folder needs refactoring to move
the construction of the worker wrapper out of the Predictive Text build
and into the final assembly of keymanweb.js (as it should be valid to
run it as a separate .js anyway).
TODO: Most of the `<reference>` paths need to be re-verified. Ideally
there should be no references outside the current module for any given
.ts.
TODO: The embedded vs browser vs node (headless) builds should be tidied
up for consistency so that it's obvious what depends on what. This is
currently messiest in the predictive-text folder, where the output names
diverge from the filenames and the various files are mixed in the same
folder (as evidenced by the exclusions listed in each tsconfig.json).
TODO: `npm install` should be removed from most build scripts and
instead `npm ci` (#6196) should be run only once from the top-level
folder for any given build. I've had eliminated side-effects from the
`install` action for npm, which makes it easier to reason about state.
TODO: verify_npm_setup and related functions can probably be eliminated.
TODO: most of the build scripts should be largely eliminated for web.
TODO: several ts projects use inconsistent output folders.
TODO: it may be possible to generate a .d.ts for models/types so that
we can use a consistent reference for those as well.
TODO: build.sh, tsconfig.json should always be in the module's top-level
folder, not in a subfolder such as src (e.g. see input-processor/src,
keyboard-processor/src, web/source).
TODO: resources/web-environment should be in common/web.
TODO: other js node_modules imports should be wrapped like es6-shim.
TODO: fix up the publish code for npm modules
TODO: eliminate version numbers from package.json if possible?
Whew, that's most of the stuff I noticed!
Summary of these changes:
Two targeted improvements; more optimization is possible but IMO this is
a significant enough improvement to go through test:
1. In oskView.ts, avoid calling `VisualKeyboard.refreshLayout` twice
from `OSKView.refreshLayout`. This roughly halves the time spent in
`VisualKeyboard.refreshLayout` for a keyboard switch.
2. In modelManager.ts, we avoid a very expensive `unloadModel` /
`loadModel` sequence from `registerModel` by checking to see if the
model spec we are passed is already registered. This avoids an
expensive callback to Keyman for Android which causes the banner to
'bounce'.
Longer notes:
So `VisualKeyboard.refreshLayout` is called 20+ times during a keyboard
switch. It is a very expensive call. It is the bulk of the time spent
when you press the globe key in Keyman for Android. This seems ripe for
optimization; see the following functions in keyboard.html:
* `setKeymanLanguage` -> `setActiveKeyboard` -> 6 calls
* `enableSuggestions` -> `registerModel` -> 8 calls
* `stateChange` -> 6+ calls
`OSKView.refreshLayout` always seems to call
`VisualKeyboard.refreshLayout` twice: first with `setSize`, and then
again itself.
Keyman for Android and KeymanWeb seem to fight over who is responsible
for selecting models (KMW first calls `loadModel`, then Keyman for
Android calls `unloadModel`, `loadModel`, through `enableSuggestions`).
The fix for rotation (`correctOSKTextSize()`), as called from
`stateChange()`, is expensive as well, because it rebuilds the entire
keyboard from scratch.
Along with that, we have the size of the webview changing as the model
is unloaded and reloaded and lots of back-and-forth between the
KeymanWeb and KMEA which seems unnecessary.
Shouldn't KeymanWeb be responsible for loading and unloading models,
once they are all registered? We should only need to register the models
once at page load time (when config changes, do we just reload the
page?).
