Currently we have two separate image-caches on the worker-thread:
- A local one, which is unique to each `PartialEvaluator.getOperatorList` invocation. This one caches both names *and* references, since image-resources may be accessed in either way.
- A global one, which applies to the entire PDF documents and all its pages. This one only caches references, since nothing else would work.
This patch introduces a third image-cache, which essentially sits "between" the two existing ones. The new `RegionalImageCache`[1] will be usable throughout a `PartialEvaluator` instance, and consequently it *only* caches references, which thus allows us to keep track of repeated image-resources found in e.g. different /Form and /SMask objects.
---
[1] For lack of a better word, since naming things is hard...
This patch extends PR 16115 to work in all browsers, regardless of their `OffscreenCanvas` support, such that transfer functions will be applied to general rendering (and not just image data).
In order to do this we introduce the `BaseFilterFactory` that is then extended in browsers/Node.js environments, similar to all the other factories used in the API, such that we always have the necessary factory available in `src/display/canvas.js`.
These changes help simplify the existing `putBinaryImageData` function, and the new method can easily be stubbed-out in the Firefox PDF Viewer.
*Please note:* This patch removes the old *partial* transfer function support, which only applied to image data, from Node.js environments since the `node-canvas` package currently doesn't support filters. However, this should hopefully be fine given that:
- Transfer functions are not very commonly used in PDF documents.
- Browsers in general, and Firefox in particular, are the *primary* development target for the PDF.js library.
- The FAQ only lists Node.js as *mostly* supported, see https://github.com/mozilla/pdf.js/wiki/Frequently-Asked-Questions#faq-support
This simply extends the approach in PR 10727 to also cover Patterns, which shouldn't be a common occurrence in Type3 fonts (since this is the first issue we've seen).
The idea is to encode large image in BMP format (which is very simple and doesn't
require to compute any checksums) and then use createImageBitmap with a BMP blob
(which doesn't suffer of the Canvas/ImageData limits).
From a performance point of view, it isn't crazy (generating a large blob + decoding
it on the main thread is really not ideal) but at least we've something to display
which is a way better than a blank page (and one can notice that most of the time is
spent in decoding the image from the pdf stream).
PDF 32000-1:2008 7.10.5.1 "Type 4 (PostScript Calculator) Functions"
defers to the PostScript Language Reference for the description of these
functions. The PostScript Language Reference, third edition chapter 8
"Operators" defines the `angle` type as a "number of degrees". Section
8.1 defines "angle `sin` real", "angle `cos` real", and "num den `atan`
angle". The documentation for `atan` further states that it will return
an angle in degrees between 0 and 360.
Handle these operators correctly in `PostScriptEvaluator.execute`.
Convert the inputs to `sin` and `cos` from degrees to radians for use
with `Math.sin` and `Math.cos`. Correctly pop two values from the stack
for `atan`, use `Math.atan2`, and convert from radians to (positive)
degrees.
I noticed several 'Path not found' errors because of a field called #subform[2].
From the XFA specs, the hash is used for a class of elements in the template tree.
When we're looking for a node in the datasets tree, it doesn't make sense to search
for a class. Hence the path element starting with a hash are just skipped.
*Please note:* I cannot reproduce the problem reported in bug 1811668, regarding the context menu, and in any case it's not clear that that part is even a PDF Viewer bug.
Looking at bug 1811668 I couldn't help but noticing that the textLayer isn't correct, and it's unfortunately once again a problem with the `adjustType1ToUnicode` function. That's intended to help improve text-selection for fonts without a /ToUnicode-entry, and in many cases it does help (the original PR fixed lots of issues) however it's also caused some problems.
In order to improve text-selection in bug 1811668, we'll now properly ignore fonts that have a predefined *named* encoding specified since that's really the intention with PR 14050.
The JBIG2 images in this PDF document are corrupt enough that even Adobe Reader warns about it when opening the file.
*Please note:* I don't really know the JBIG2 image format at all, however from a very brief look at the specification it seems that integers should be 32-bit.
The relevant TrueType font is missing both /ToUnicode *and* /Encoding entires, either of which would have prevented the (current) broken textLayer rendering.
My first idea was that we could use the `post` table in the TrueType font, see https://developer.apple.com/fonts/TrueType-Reference-Manual/RM06/Chap6post.html, to get the actual glyphNames and amend the fallback ToUnicode-map that way. Unfortunately that didn't work, since the `post` table only contained ".notdef" and "" (i.e. empty string) entries.
Instead we try to use the `name` table in the TrueType font, see https://developer.apple.com/fonts/TrueType-Reference-Manual/RM06/Chap6name.html, to determine if the platform is Windows and thus fallback to generate a ToUnicode-map from the `WinAnsiEncoding`.
Note how all over the `src/core/annotation.js`-code we're assuming that if an `appearance`-entry exists it's also a Stream. However, we're not actually checking that thoroughly enough which causes issues in some badly generated PDF documents.
*Please note:* The reduced test-case is *not* a perfect reproduction of the original PDF document, since this one fails to open in e.g. Adobe Reader, but I do believe that it captures the most important points here.
For corrupt *and* encrypted PDF documents, it's possible that only some trailer dictionaries actually contain an /Encrypt-entry. Previously we'd could easily miss that, since we generally pick the first not obviously corrupt trailer dictionary, and the solution implemented here is to simply pre-parse all trailer dictionaries to see if there's any /Encrypt-entries.
When trying to find incomplete objects, i.e. those missing the "endobj"-string at the end, there's unfortunately a number of possible operators that we need to check for. Otherwise we could miss e.g. the "trailer" at the end of a corrupt PDF document, which is why the referenced document didn't work.
Currently we do all searching on the "raw" bytes of the PDF document, for efficiency, however this doesn't really work when we need to check for *multiple* potential command-strings. To keep the complexity manageable we'll instead use regular expressions here, but we can at least avoid creating lots of substrings thanks to the `RegExp.lastIndex` property; which is well supported across browsers according to https://developer.mozilla.org/en-US/docs/Web/JavaScript/Reference/Global_Objects/RegExp/lastIndex#browser_compatibility
Note that this repeated regular expression usage could perhaps be slightly less efficient than the old code, however this method is only invoked for corrupt PDF documents.
Previously we'd abort all parsing if an Error was encountered, despite the fact that multiple `startXRefQueue`-entries may be available and that continued parsing could thus eventually be able to find usable data.
Note that in the referenced PDF document the `startxref`-operator, at the end of the file, points to a position in the middle of an arbitrary `stream` which is why things break.
This can't be a particularly common feature, since we've supported Optional Content for over two years and this is the very first TilingPattern-case we've seen.
Given that this PDF document is an interesting test-case for performance reasons, w.r.t. inline image caching, it probably can't hurt to add it to the test-suite to make it more readily available.
Considering the contents of that PDF document I'm not sure if we can include it directly in the repository, hence why a *linked* test-case was choosen here.
- For text fields
* when printing, we generate a fake font which contains some widths computed thanks to
an OffscreenCanvas and its method measureText.
In order to avoid to have to layout the glyphs ourselves, we just render all of them
in one call in the showText method in using the system sans-serif/monospace fonts.
* when saving, we continue to create the appearance streams if the fonts contain the char
but when a char is missing, we just set, in the AcroForm dict, the flag /NeedAppearances
to true and remove the appearance stream. This way, we let the different readers handle
the rendering of the strings.
- For FreeText annotations
* when printing, we use the same trick as for text fields.
* there is no need to save an appearance since Acrobat is able to infer one from the
Content entry.
*Please note:* This only fixes the "wrong letter" part of bug 1799927.
It appears that the simple `computeAdler32` function, used when caching inline images, generates hash collisions for some (very short) TypedArrays. In this case that leads to some of the "letters", which are actually inline images, being rendered incorrectly.
Rather than switching to another hashing algorithm, e.g. the `MurmurHash3_64` class, we simply cache using a stringified version of the inline image data as the cacheKey to prevent any future collisions. While this will (naturally) lead to slightly higher peak memory usage, it'll however be limited to the current `Parser`-instance which means that it's not persistent.
One small benefit of these changes is that we can avoid creating lots of `Stream`-instances for already cached inline images.
When a form isn't changed, we used the appearances we had in the file, but when
/NeedAppearances is true, all the appearances have to be regenerated whatever they're.
In the referenced PDF document there are "numbers" which consist only of `-.`, and while that's obviously not valid Adobe Reader seems to handle it just fine.
Letting this method ignore more invalid "numbers" was suggested during the review of PR 14543, so let's simply relax our the validation here.
After PR 14311, and follow-up patches, we no longer require that the /Count entry (in the /Pages dictionary) is either present or even valid in order to parse/render a PDF document.
Hence it seems strange to keep this requirement for *corrupt* PDF documents, when trying to find a usable `trailer` in the `XRef.indexObjects` method.
When we fail to find a usable PDF document `trailer` *and* there were errors during parsing, try and fallback to a *previous* generation as a last resort during fetching of uncompressed references.
*Please note:* This will not affect "normal" PDF documents, with valid /XRef data, and even most *corrupt* documents should be completely unaffected by these changes.
*Please note:* I don't really know what I'm doing here, however the patch appears to fix the referenced issue when comparing the rendering with Adobe Reader (with the caveat that I don't speak the language in question).
*Fixes a regression from PR 15246, sorry about that!*
The return value of all `Annotation.getOperatorList` methods was changed in PR 15246, however I missed updating the error code-path in `Page.getOperatorList` which thus breaks all operatorList-parsing for pages with corrupt Annotations.
