When rendering big PDF pages at high zoom levels, we currently fall back
to CSS zoom to avoid rendering canvases with too many pixels. This
causes zoomed in PDF to look blurry, and the text to be potentially
unreadable.
This commit adds support for rendering _part_ of a page (called
`PDFPageDetailView` in the code), so that we can render portion of a
page in a smaller canvas without hiting the maximun canvas size limit.
Specifically, we render an area of that page that is slightly larger
than the area that is visible on the screen (100% larger in each
direction, unless we have to limit it due to the maximum canvas size).
As the user scrolls around the page, we re-render a new area centered
around what is currently visible.
If either of the factory-urls are missing or invalid, the fallback value would currently become `useWorkerFetch === null`.
While that is obviously a falsy value, which means that the code still works as intended, we should ensure that this is consistent.
Currently we have three separate and virtually identical message handlers for this data, which can easily be combined into a single message handler instead.
Given that we now have a few different factory-url parameters, we introduce a helper function for parsing them.
*Please note:* These parameters have always been documented as requiring a trailing slash[1], which we can now easily enforce during the `getDocument`-call.
---
[1] I recall that we've occasionally seen issues because users miss that detail, and the new Error should hopefully be more easily actionable than one thrown during rendering/parsing.
After PR 2317, which landed in 2012, we'd immediately clean-up after rendering for pages with large image resources. This had the effect that re-rendering, e.g. after zooming, would force us to re-parse the entire page which could easily lead to bad performance.
In PR 16108, which landed in 2023, we tried to lessen the impact of that by slightly delaying clean-up however that's obviously not a perfect solution (and it increased the complexity of the relevant code).
Furthermore, the condition for this "immediate" clean-up seems a bit arbitrary to me since a page could easily contain a large number of smaller images whose total size vastly exceeds the threshold.
Hence this patch, which suggests that we remove the conditional and delayed clean-up after rendering. Compared to the situation back in 2012, a number of things have improved since:
- We have *multiple* caches for repeated image-resources on the worker-thread[1], which helps reduce overall memory usage and improves performance.
- We downsize huge images on the worker-thread, which means that the images we're using on the main-thread cannot be arbitrarily large.
- The amount of available RAM on devices should be a lot higher, since more than a decade has passed.
A future improvement here, for more resource constrained environments, could be to instead clean-up when actually needed using e.g. `WeakRef`s (see issue 18148).
---
[1] More specifically:
- `LocalImageCache`, which caches image-data by /Name and /Ref on the `PartialEvaluator.prototype.getOperatorList` level.
- `RegionalImageCache`, which caches image-data by /Ref on the `PartialEvaluator`-instance (i.e. at the page) level.
- `GlobalImageCache`, which caches image-data by /Ref globally at the document level.
- Most of the these are only used in the `src/display/api.js` file, and this leads to slightly shorter code.
- A number of unit-tests need a `BaseCanvasFactory`-instance, however that one is available through the `PDFDocumentProxy`-instance nowadays.
- For other unit-tests the remaining necessary default Factory-definitions can be moved into the `test/unit/test_utils.js` file.
Given that we nowadays provide default Node.js versions of these Factory-parameters it no longer seems necessary to mention that environment specifically.
In order to fix bug 1935076, we'll have to add a pure js fallback in case wasm is disabled
or simd isn't supported. Unfortunately, this fallback will take some space.
So, the main goal of this patch is to reduce the overall size (by ~93k).
As a side effect, it should make easier to use an other wasm file (which must export
_jp2_decode, _malloc and _free).
Rather than having to manually implement the exception-handling for the "DocException" message, we can instead re-use (and slightly extend) the existing `wrapReason` function since that one already does what we need.
Furthermore, we can also simplify handling of the "PasswordRequest" message a little bit and again re-use the `wrapReason` function.
Finally, the patch makes the following smaller changes:
- Avoid needlessly re-creating exceptions in the `wrapReason` function.
- Use a slightly shorter parameter name in the `wrapReason` function.
- Remove the unused entries in the `CallbackKind`/`StreamKind` enumerations.
The `Path2D` glyph-objects are cached on the `FontFaceObject`-instance, so we can save a little bit of memory by removing the raw path-strings once they're no longer needed.
This slightly shortens the code, in various `destroy`-methods, which cannot hurt.
Also, use pre-processor checks to simplify `PDFDocumentLoadingTask.destroy` in the Firefox PDF Viewer since the `PDFWorker.fromPort`-method isn't used there.
The `Array` type takes one parameter that describes the possible types
of the inner elements. This parameter may exist of pipes to indicate
multiple possible types. However, the current type definition provides
multiple parameters; this is incorrect syntax, and TypeScript 5.7+ now
fails on this due to stricter syntax validation.
This commit fixes the issue by changing the type definition to the
proper syntax, which together with the accompanying comment about the
contents of the fingerprints array should document it sufficiently.
Given that there are multiple issues with `ImageDecoder` in Chromium browsers, affecting both BMP and JPEG images, for now we (by default) disable that functionality there to avoid problems.
This also means that we can remove the previously added, and separate, `isChrome` API-option.
This allows end-users to forcibly disable `ImageDecoder` usage, even if the browser appears to support it (similar to the pre-existing option for `OffscreenCanvas`).
- Over time the number and size of these factories have increased, especially the `DOMFilterFactory` class, and this split should thus aid readability/maintainability of the code.
- By introducing a couple of new import maps we can avoid bundling the `DOMCMapReaderFactory`/`DOMStandardFontDataFactory` classes in the Firefox PDF Viewer, since they are dead code there given that worker-thread fetching is always being used.
- This patch has been successfully tested, by running `$ ./mach test toolkit/components/pdfjs/`, in a local Firefox artifact-build.
*Note:* This patch reduces the size of the `gulp mozcentral` output by `1.3` kilo-bytes, which isn't a lot but still cannot hurt.
This way we can directly throw Errors, rather than having to "manually" return rejected Promises, which is ever so slightly shorter.
Also, since `useWorkerFetch` is always true in MOZCENTRAL builds these message-handlers should not be invoked there.
We can convert the handler to an `async` function, which removes the need to create a temporary Promise here.
Given the age of this code it shouldn't hurt to simplify it a little bit.
- Change the "message" event handler function to a private method.
- Remove the "message" event listener with an `AbortSignal`.
- Extend the `LoopbackPort`-class with `AbortSignal` support.
Given that the `WorkerTransport`-constructor will access all possible factory-instances, let's ensure that the `transportFactory`-object always has a consistent shape regardless of other options.
Also, since `useWorkerFetch` is always true in MOZCENTRAL builds we never need to create `CMapReaderFactory`/`StandardFontDataFactory`-instances there.
This unifies the various factory-options, since it's consistent with `CMapReaderFactory`/`StandardFontDataFactory`, and ensures that any needed parameters will always be consistently provided when creating `CanvasFactory`/`FilterFactory`-instances.
As shown in the modified example this may simplify some custom implementations, since we now provide the ability to access the `CanvasFactory`-instance used with a particular `getDocument`-invocation.
This code is quite old and has been moved/re-factored a few times over the years, however we can simplify this even further since we don't actually need a function to determine what NetworkStream-implementation to use.
This is possible thanks to features, i.e. private fields and in particular static initialization blocks, that didn't exist back when we started using classes in the code-base.
In *hindsight* this seems like a better idea, since it avoids the need to manually pass `isEditing` around as a boolean value.
Note that `RenderingIntentFlag` is *internal* functionality, not exposed in the official API, which means that it can be extended and modified as necessary.
Right now, editable annotations are using their own canvas when they're drawn, but
it induces several issues:
- if the annotation has to be composed with the page then the canvas must be correctly
composed with its parent. That means we should move the canvas under canvasWrapper
and we should extract composing info from the drawing instructions...
Currently it's the case with highlight annotations.
- we use some extra memory for those canvas even if the user will never edit them, which
the case for example when opening a pdf in Fenix.
So with this patch, all the editable annotations are drawn on the canvas. When the
user switches to editing mode, then the pages with some editable annotations are redrawn but
without them: they'll be replaced by their counterpart in the annotation editor layer.
Errors related to this `requestAnimationFrame` show up intermittently when running the integration-tests on the bots, however I've been unable to reproduce it locally.
Hence I cannot guarantee that it's enough to fix the timing issues, however this should be generally safe since the `requestAnimationFrame` invokes the `_next`-method and the first thing that one does is check that rendering hasn't been cancelled.
This avoids having to repeat the same code multiple times, since besides resolving the promise we also need to send the "configure" message to the worker-thread.
The feature-testing on the worker-thread has been simplified in previous pull requests, which means that we can simplify this main-thread handler as well.
This helps reduce overall indentation in the method, thus leading to slightly less code.
Also, remove an old comment referring to Chrome 15 since that's no longer relevant now.
- Move the definition of the `loadingParams` Object, to simplify the code.
- Add a unit-test, since none existed and the viewer depends on this functionality.
This is very old code, and predates e.g. the introduction of JavaScript classes, which creates unnecessarily unwieldy code in the viewer.
By introducing a new `TextLayer` class in the API, similar to how e.g. the `AnnotationLayer` looks, we're able to keep most parameters on the class-instance itself. This removes the need to manually track them in the viewer, and simplifies the call-sites.
This also removes the `numTextDivs` parameter from the "textlayerrendered" event, since that's only added to support default-viewer functionality that no longer exists.
Finally we try, as far as possible, to polyfill the old `renderTextLayer` and `updateTextLayer` functions since they are exposed in the library API.
For *simple* invocations of `renderTextLayer` the behaviour should thus be the same, with only a warning printed in the console.
For images that failed to decode once we want to avoid a pointless round-trip to the main-thread, which could otherwise happen for globally cached images.
- These changes will allow a simpler way of implementing PR 17770.
- The /Lang attribute is fetched lazily, with the first `getTextContent` invocation. Given the existing worker-thread caching, this will thus only need to be done *once* per PDF document (and most PDFs don't included this data).
- This makes the /Lang attribute *directly available* in the `textLayer`, which has the following advantages:
- We don't need to block, and thus delay, overall viewer initialization on fetching it (nor pass it around throughout the viewer).
- Third-party users of the `textLayer` will automatically benefit from this, once we start actually using the /Lang attribute in PR 17770.
*Please note:* This also, importantly, means that the `text` reference-tests will then cover this code (which wouldn't otherwise have been the case).
- The `stopAtErrors` API option, which is the inverse of the "internal" `ignoreErrors` option, is explicitly documented as applying to *parsing* (i.e. the worker-thread) while the `FontFaceObject` class is used during rendering (i.e. the main-thread); see b6765403a1/src/display/api.js (L164-L167)
- A glyph that fails in the `FontRendererFactory`, on the worker-thread, will already cause (overall) parsing to stop when `ignoreErrors === false` hence checking the option on the main-thread as well seems redundant; see b6765403a1/src/core/evaluator.js (L4527-L4533)
- Removing this option simplifies the code, and slightly reduces the number of options that we need to handle in the main-thread code.
