The purpose of this patch is twofold:
- Initialize the unicode-category data *lazily* during text-extraction, since this is completely unused during general parsing/rendering.
- Stop exposing this data in the API, since it's unused on the main-thread and it seems like it was *accidentally* included.
Obviously these changes are API-observable, but hopefully no user is depending on this. Furthermore, it's trivial for a user to re-create this unicode-category data manually with a regular expression (from the exposed `unicode` property).
Currently, during text-extraction, we're repeatedly normalizing and (when necessary) reversing the unicode-values every time. This seems a little unnecessary, since the result won't change, hence this patch moves that into the `Glyph`-instance and makes it *lazily* initialized.
Taking the `tracemonkey.pdf` document as an example: When extracting the text-content there's a total of 69236 characters but only 595 unique `Glyph`-instances, which mean a 99.1 percent cache hit-rate. Generally speaking, the longer a PDF document is the more beneficial this should be.
*Please note:* The old code is fast enough that it unfortunately seems difficult to measure a (clear) performance improvement with this patch, so I completely understand if it's deemed an unnecessary change.
Given the sheer number of heuristics added to this method over the years, moving the *valid* unicode found case to the top should improve readability of the code.
This patch first of all makes `isOffscreenCanvasSupported` configurable, defaulting to `true` in browsers and `false` in Node.js environments, with a new `getDocument` parameter. While you normally want to use this, in order to improve performance, it should still be possible for users to control it (similar to e.g. `isEvalSupported`).
The specific problem, as reported in issue 14952, is that the SVG back-end doesn't support the new ImageMask data-format that's introduced in PR 14754. In particular:
- When the SVG back-end is used in Node.js environments, this patch will "just work" without the user needing to make any code changes.
- If the SVG back-end is used in browsers, this patch will require that `isOffscreenCanvasSupported: false` is added to the `getDocument`-call.
OperatorList.addOp can trigger a flush if it's required, hence the values passed to it must
be correctly initialized in order to avoid some wrong values in the renderer.
Because of that a clip path was considered as empty, nothing was clipped, hence the wrong
rendering in bug 1791583.
Most of the `String.prototype.search` call-sites found throughout the code-base is actually not necessary, since we usually only want a *boolean*, and those can be replaced with `RegExp.prototype.test` instead.
In the referenced PDF document the fonts have /CIDToGIDMap-entries that cannot be loaded. Hence, only when `ignoreErrors` is set, we'll now ignore these corrupt /CIDToGIDMap-entries and fallback to simply assume that no such data is available.
Given that this is *clearly* a case of a corrupt PDF document, there's no guarantee that this will "fix" things in the general case since a /CIDToGIDMap may be *required* in order for some composite fonts to render correctly. However, attempting to render *something* is surely better than skipping a font altogether.
Note how we currently throw a "raw" Error, which is problematical since all of the `PartialEvaluator.loadFont` call-sites expect a Promise to be returned. Furthermore, this also means that we don't benefit from the fallback code-path that now exists below.
*Please note:* Unfortunately I don't have a test-case that fails without this patch, since it's something I happened to notice when reading the code while working on another patch.
This extends PR 13461, by also building a fallback bounding box for Type3 fonts that contain a much too small /FontBBox-entry.
*Please note:* While this patch improves things overall, copy-and-pasting still doesn't work perfectly for this document. In particular the lowercase letter "c" cannot be selected/copied, however this can be reproduced in both Adobe Reader and PDFium (in Google Chrome) too, which is caused by a lack of proper /ToUnicode-data in the PDF document.
In the `src/display/canvas.js` code the `d1` operator will be used to set the clipping region, and it obviously cannot be empty since that prevents the Type3-glyph from rendering.
Also, the patch removes an outdated comment; refer to PR 12718.
This limits the heuristics for handling of incomplete path operators, see PR 9838, to only apply to *sequences* of such operators. In practice a couple of invalid path operators are (hopefully) unlikely to completely break rendering, whereas a sequence of them will easily lead to fairly chaotic rendering artifacts.
The current `lastModified`-getter, which only contains a time-stamp, is a fairly crude way of detecting if the stored data has actually been changed. In particular, when the `getRawValue`-method is used, the `lastModified`-getter doesn't cope with data being modified from the "outside".
To fix these issues[1], and to prevent any future bugs in this code, this patch introduces a new `AnnotationStorage.hash`-getter which computes a hash of the currently stored data. To simplify things this re-uses the existing `MurmurHash3_64`-implementation, which required moving that file into the `src/shared/`-folder, since its performance should be good enough here.
---
[1] Given how the `AnnotationStorage.lastModified`-getter was used, this would have been limited to *printing* of forms.
As it turns out, most of the code-paths in the `PDFImage`-class won't actually pass the TypedArray (containing the image-data) to the `ColorSpace`-code. Hence we *generally* don't need to force the image-data to be a `Uint8ClampedArray`, and can just as well directly use a `Uint8Array` instead.
In the following cases we're returning the data without any `ColorSpace`-parsing, and the exact TypedArray used shouldn't matter:
- b72a448327/src/core/image.js (L714)
- b72a448327/src/core/image.js (L751)
In the following cases the image-data is only used *internally*, and again the exact TypedArray used shouldn't matter:
- b72a448327/src/core/image.js (L762) with the actual image-data being defined (as `Uint8ClampedArray`) further below
- b72a448327/src/core/image.js (L837)
*Please note:* This is tagged `api-minor` because it's API-observable, given that *some* image/mask-data will now be returned as `Uint8Array` rather than using `Uint8ClampedArray` unconditionally. However, that seems like a small price to pay to (slightly) reduce memory usage during image-conversion.
While working on PR 14825, I couldn't help noticing that the code to increment the `count` for cached ImageMasks was repeated multiple times. Hence it makes sense, as far as I'm concerned, to move this into a helper function instead.
Currently we only insert optionalContent-data into the operatorList the first time that an image is parsed, which will (in hindsight) obviously cause problems for cached images.
Hence we also need to insert the optionalContent-data in the various worker-thread image caches, such that it can be accessed in the fast-paths that are used to skip re-parsing of images.
In order to reduce the amount of repeated code, this patch also adds a new `OperatorList`-method that takes care of inserting the necessary data in the operatorList.
In the referenced PDF document the fonts have /Encoding-entries that are Streams (containing completely bogus data), which are thus obviously not valid here.
Hence, only when `ignoreErrors` is set, we'll now ignore these corrupt /Encoding-entries and fallback to the existing code to try and infer a usable encoding.
Given that this is *clearly* a case of corrupt PDF documents, there's no guarantee that this will "fix" all such cases, however it's the best that we do here and shouldn't really be worse than ignoring an entire font.
- most of the time the current transform is a scaling one (modulo translation),
hence it's possible to avoid to apply the transform on each bbox and then apply
it a posteriori;
- compute the bbox when it's possible in the worker.
- it's the second part of the fix for https://bugzilla.mozilla.org/show_bug.cgi?id=857031;
- some image masks can be used several times but at different positions;
- an image need to be pre-process before to be rendered:
* rescale it;
* use the fill color/pattern.
- the two operations above are time consuming so we can cache the generated canvas;
- the cache key is based on the current transform matrix (without the translation part)
and the current fill color when it isn't a pattern.
- the rendering of the pdf in the above bug is really faster than without this patch.
- it aims to partially fix performance issue reported: https://bugzilla.mozilla.org/show_bug.cgi?id=857031;
- the idea is too avoid to use byte arrays but use ImageBitmap which are a way faster to draw:
* an ImageBitmap is Transferable which means that it can be built in the worker instead of in the main thread:
- this is achieved in using an OffscreenCanvas when it's available, there is a bug to enable them
for pdf.js: https://bugzilla.mozilla.org/show_bug.cgi?id=1763330;
- or in using createImageBitmap: in Firefox a task is sent to the main thread to build the bitmap so
it's slightly slower than using an OffscreenCanvas.
* it's transfered from the worker to the main thread by "reference";
* the byte buffers used to create the image data have a very short lifetime and ergo the memory used is globally
less than before.
- Use the localImageCache for the mask;
- Fix the pdf issue4436r.pdf: it was expected to have a binary stream for the image;
- Move the singlePixel trick from operator_list to image: this way we can use this trick even if it isn't in a set
as defined in operator_list.
- it aims to fix issue #14627;
- the basic idea of the recent text refactoring was to only consider the rendered visible whitespaces.
But sometimes, the heuristics aren't correct and although some whitespaces are in the text stream
they weren't in the text chunks because they were too small. Hence we added some exceptions, for example,
we always add a whitespace when it is between two non-whitespace chars but only when in the same Tj.
So basically, this patch removes the constraint to have the chars in the same Tj
(in using a circular buffer to save the two last chars) but don't add a space when the visible space is really
too small (hence `NOT_A_SPACE_FACTOR`).
Note that the Prettier update made it possible to move a couple of comments after `default:`-cases back to their original/intended positions, please see https://prettier.io/blog/2022/03/16/2.6.0.html
This patch removes the existing `forEach` methods, in favor of making the classes properly iterable instead. Given that the classes are using a `Set` respectively a `Map` internally, implementing this is very easy/efficient and allows us to simplify some existing code.
The situation described in issue 14626 seems like a fairly special case, and it thus seem reasonable that we simply follow the same pattern as elsewhere in the `PartialEvaluator` when the `stopAtErrors` API-option is being used.
The call-sites are replaced by direct `typeof`-checks instead, which removes unnecessary function calls. Note that in the `src/`-folder we already had more `typeof`-cases than `isString`-calls.
The call-sites are replaced by direct `typeof`-checks instead, which removes unnecessary function calls. Note that in the `src/`-folder we already had more `typeof`-cases than `isNum`-calls.
These changes were *mostly* done using regular expression search-and-replace, with two exceptions:
- In `Font._charToGlyph` we no longer unconditionally update the `width`, since that seems completely unnecessary.
- In `PDFDocument.documentInfo`, when parsing custom entries, we now do the `typeof`-check once.
Unless you actually need to check that something is both a `Name` and also of the *correct* type, using `instanceof Name` directly should be a tiny bit more efficient since it avoids one function call and an unnecessary `undefined` check.
This patch uses ESLint to enforce this, since we obviously still want to keep the `isName` helper function for where it makes sense.
Unless you actually need to check that something is both a `Dict` and also of the *correct* type, using `instanceof Dict` directly should be a tiny bit more efficient since it avoids one function call and an unnecessary `undefined` check.
This patch uses ESLint to enforce this, since we obviously still want to keep the `isDict` helper function for where it makes sense.
This helper function is not really needed, since it's just a wrapper around a simple `instanceof` check, and it only adds unnecessary indirection in the code.
At this point all the various Stream-classes extends an abstract base-class, hence this helper function is no longer necessary and only adds unnecessary indirection in the code.
This appears to be consistent with the behaviour in both Adobe Reader and PDFium (in Google Chrome); this is essentially the same approach as used for a single decimal point in PR 9827.
With these changes, we'll now *always* replace all whitespaces with standard spaces (0x20). This behaviour is already, since many years, the default in both the viewer and the browser-tests.
- it aims to fix#14497;
- previously, only rotations with an angle 0, 90, 180 or 270 were taken into account;
- so generalize to any angle but keep the fast path for 0, 90, ... because they're likely more common than anything else.
After the changes in PR 14428 we can *directly*, and more efficiently, handle whitespace conversion in `PartialEvaluator.getTextContent` when the `normalizeWhitespace` option is being used.
This way we no longer need a separate helper function for this, and can avoid having to (again) iterate through the text and checking each character. Finally, this also removes the need for using a regular expression on e.g. all non-ASCII text.
In corrupt PDF documents Type3 fonts may introduce circular dependencies, thus resulting in the affected font(s) never loading and parsing/rendering never completing.
Note that I've not seen any real-world examples of this kind of font corruption, but the attached PDF document was rather found in https://github.com/pdf-association/safedocs/tree/main/Miscellaneous%20Targeted%20Test%20PDFs
*Please note:* That repository contains a number of reduced test-cases that are specifically intended to test interoperability (between PDF viewer) and parsing/rendering for various kinds of strange/corrupt PDF documents.
Some of the test-cases found there may thus not make sense to try and "fix" upfront, in my opinion, unless the problems are also found in real-world PDF documents.
*Please note:* These changes will primarily benefit longer documents, somewhat at the expense of e.g. one-page documents.
The existing `PDFDocumentProxy.getStats` function, which in the default viewer is called for each rendered page, requires a round-trip to the worker-thread in order to obtain the current document stats. In the default viewer, we currently make one such API-call for *every rendered* page.
This patch proposes replacing that method with a *synchronous* `PDFDocumentProxy.stats` getter instead, combined with re-factoring the worker-thread code by adding a `DocStats`-class to track Stream/Font-types and *only send* them to the main-thread *the first time* that a type is encountered.
Note that in practice most PDF documents only use a fairly limited number of Stream/Font-types, which means that in longer documents most of the `PDFDocumentProxy.getStats`-calls will return the same data.[1]
This re-factoring will obviously benefit longer document the most[2], and could actually be seen as a regression for one-page documents, since in practice there'll usually be a couple of "DocStats" messages sent during the parsing of the first page. However, if the user zooms/rotates the document (which causes re-rendering), note that even a one-page document would start to benefit from these changes.
Another benefit of having the data available/cached in the API is that unless the document stats change during parsing, repeated `PDFDocumentProxy.stats`-calls will return *the same identical* object.
This is something that we can easily take advantage of in the default viewer, by now *only* reporting "documentStats" telemetry[3] when the data actually have changed rather than once per rendered page (again beneficial in longer documents).
---
[1] Furthermore, the maximium number of `StreamType`/`FontType` are `10` respectively `12`, which means that regardless of the complexity and page count in a PDF document there'll never be more than twenty-two "DocStats" messages sent; see 41ac3f0c07/src/shared/util.js (L206-L232)
[2] One example is the `pdf.pdf` document in the test-suite, where rendering all of its 1310 pages only result in a total of seven "DocStats" messages being sent from the worker-thread.
[3] Reporting telemetry, in Firefox, includes using `JSON.stringify` on the data and then sending an event to the `PdfStreamConverter.jsm`-code.
In that code the event is handled and `JSON.parse` is used to retrieve the data, and in the "documentStats"-case we'll then iterate through the data to avoid double-reporting telemetry; see https://searchfox.org/mozilla-central/rev/8f4c180b87e52f3345ef8a3432d6e54bd1eb18dc/toolkit/components/pdfjs/content/PdfStreamConverter.jsm#515-549
- it aims to fix https://bugzilla.mozilla.org/show_bug.cgi?id=931481;
- real space chars are pushed in the chunk but when there is an extra spacing, the next char position must be compared with the previous one;
- for example, an extra spacing can cancel a space so visually there are no space.
In `beginGroup` we create a new canvas that is the size of the
bounding box and we translate it to the offset. This means we don't need to
also apply the bounding box during `paintFormXObjectBegin`.
This improves #6961 quite a bit, but it still is missing the indention
in the ruler.
- PR #13257 fixed a lot of issues but not all and this patch aims to fix almost all remaining issues.
- the idea in this new patch is to compare position of new glyph with the last position where a glyph has been drawn;
- no space are "drawn": it just moves the cursor but they aren't added in the chunk;
- so this way a space followed by a cursor move can be treated as only one space: it helps to merge all spaces into one.
- to make difference between real spaces and tracking ones, we used a factor of the space width (from the font)
- it was a pretty good idea in general but it fails with some fonts where space was too big:
- in Poppler, they're using a factor of the font size: this is an excellent idea (<= 0.1 * fontSize implies tracking space).
This is unfortunately *yet another* bug in the `preEvaluateFont`-implementation, and I've lost count of the number of times I've had to tweak this code over the years :-(
I really cannot help thinking that PR 4423 was way too simplistic, since it missed a bunch of cases that leads to broken font rendering in many PDF documents.
Fixes https://bugzilla.mozilla.org/show_bug.cgi?id=1734802
In this particular case the `CMap`-data that we create contains only numbers, but no strings, which causes `PartialEvaluator.readToUnicode` to create a ToUnicode-map with only empty strings.
*Please note:* This is yet another case where I don't know if it's necessarily the best and most correct solution, but it does fix the referenced issue.
While some of the output looks worse to my eye, this behavior more
closely matches what I see when I open the PDFs in Adobe acrobat.
Fixes: #4706, #9713, #8245, #1344
Following the STR in the issue, this patch reduces the number of `PartialEvaluator.getTextContent`-related `postMessage`-calls by approximately 78 percent.[1]
Note that by enforcing a relatively low value when batching text chunks, we should thus improve worst-case scenarios while not negatively affect all `textLayer` building.
While working on these changes I noticed, thanks to our unit-tests, that the implementation of the `appendEOL` function unfortunately means that the number and content of the textItems could actually be affected by the particular chunking used.
That seems *extremely* unfortunate, since in practice this means that the particular chunking used is thus observable through the API. Obviously that should be a completely internal implementation detail, which is why this patch also modifies `appendEOL` to mitigate that.[2]
Given that this patch adds a *minimum* batch size in `enqueueChunk`, there's obviously nothing preventing it from becoming a lot larger then the limit (depending e.g. on the PDF structure and the CPU load/speed).
While sending more text chunks at once isn't an issue in itself, it could become problematic at the main-thread during `textLayer` building. Note how both the `PartialEvaluator` and `CanvasGraphics` implementations utilize `Date.now()`-checks, to prevent long-running parsing/rendering from "hanging" the respective thread. In the `textLayer` building we don't utilize such a construction[3], and streaming of textContent is thus essentially acting as a *simple* stand-in for that functionality.
Hence why we want to avoid choosing a too large minimum batch size, since that could thus indirectly affect main-thread performance negatively.
---
[1] While it'd be possible to go even lower, that'd likely require more invasive re-factoring/changes to the `PartialEvaluator.getTextContent`-code to ensure that the batches don't become too large.
[2] This should also, as far as I can tell, explain some of the regressions observed in the "enhance" text-selection tests back in PR 13257.
Looking closer at the `appendEOL` function it should potentially be changed even more, however that should probably not be done here.
[3] I'd really like to avoid implementing something like that for the `textLayer` building as well, given that it'd require adding a fair bit of complexity.
Something that I *just* realized is that while PR 13854 fixed an issue as reported, it could still cause bugs in other similarily broken documents since we'll not insert a matching endMarkedContent-operator in the operatorList.
Currently, in the `PartialEvaluator`, we only support Optional Content in Form-/XObjects. Hence this patch adds support for Image-/XObjects as well, which looks like a simple oversight in PR 12095 since the canvas-implementation already contains the necessary code to support this.
This way there cannot be any *incorrect* cache hits, since Refs are guaranteed to be unique.
Please note that the reason for caching by Ref rather than doing something along the lines of the `localShadingPatternCache` (which uses a `Map` directly), is that TilingPatterns are streams and those cannot be cached on the `XRef`-instance (this way we avoid unnecessary parsing).
The PDF in bug 1721949 uses many unique pattern objects
that references the same shading many times. This caused
a new canvas pattern to be created and cached many times
driving up memory use.
To fix, I've changed the cache in the worker to key off the
shading object and instead send the shading and matrix
separately. While that worked well to fix the above bug,
there could be PDFs that use many shading that could
cause memory issues, so I've also added a LRU cache
on the main thread for canvas patterns. This should prevent
memory use from getting too high.
- All the scale factors in for the substitution font were wrong because of different glyph positions between Liberation and the other ones:
- regenerate all the factors
- Text may have polish chars for example and in this case the glyph widths were wrong:
- treat substitution font as a composite one
- add a map glyphIndex to unicode for Liberation in order to generate width array for cid font
- In order to better compute text fields size, use line height with no gaps (and consequently guessed height for text are slightly better in general).
- Fix default background color in fields.
Bug 1721218 has a shading pattern that was used thousands of times.
To improve performance of this PDF:
- add a cache for patterns in the evaluator and only send the IR form once
to the main thread (this also makes caching in canvas easier)
- cache the created canvas radial/axial patterns
- for shading fill radial/axial use the pattern directly instead of creating temporary
canvas
When working on PR 13612, I mostly prioritized a simple solution that didn't require touching a lot of code. However, while working on PR 13735 I started to realize that the static `Name.empty` construction really wasn't a good idea.
In particular, having a special `Name`-instance where the `name`-property isn't actually a String is confusing (to put it mildly) and can easily lead to issues elsewhere. The only reason for not simply allowing the `name`-property to be an *empty* string, in PR 13612, was to avoid having to touch a lot of existing code. However, it turns out that this is only limited to a few methods in the `PartialEvaluator` and a few of the `BaseLocalCache`-implementations, all of which can be easily re-factored to handle *empty* `Name`-instances.
All-in-all, I think that this patch is even an *overall* improvement since we're now validating (what should always be) `Name`-data better in the `PartialEvaluator`.
This is what I ought to have done from the start, sorry about the code churn here!
- font line height is taken into account by acrobat when it isn't with masterpdfeditor: I extracted a font from a pdf, modified some ascent/descent properties thanks to ttx and the reinjected the font in the pdf: only Acrobat is taken it into account. So in this patch, line heights for some substituted fonts are added.
- it seems that Acrobat is using a line height of 1.2 when the line height in the font is not enough (it's the only way I found to fix correctly bug 1718741).
- don't use flex in wrapper container (which was causing an horizontal overflow in the above bug).
- consequently, the above fixes introduced a lot of small regressions, so in order to see real improvements on reftests, I fixed the regressions in this patch:
- replace margin by padding in some case where padding is a part of a container dimensions;
- remove some flex display: some containers are wrongly sized when rendered;
- set letter-spacing to 0.01px: it helps to be sure that text is not broken because of not enough width in Firefox.
Given that we're not imposing any font-type restrictions[1] in the non-/FontDescriptor case, it's not really clear to me why we'd actually need to do that in the general case.
Please note that there's some *expected* movement, all of which should be improvements, in the `fips197.pdf` file with this patch.
---
[1] With the exception of Type3-fonts, of course.
*Sorry about the churn here, since the change that I made in PR 13516 was not very smart.*
With the current code, it's now *impossible* for a user to actually control the `useSystemFonts` option manually. To prevent outright breakage we obviously still need to default to setting `useSystemFonts = false` when `disableFontFace === true`, however that should be possible for an API consumer to override.
Rather than having to create and check a *separate* `ToUnicodeMap` to handle these cases, we can simply use the `fallbackToUnicode`-data (when it exists) to directly supplement *missing* /ToUnicode entires in the regular `ToUnicodeMap` instead.
We must force-fetch standard font data, when `disableFontFace = true` is set in the API, since otherwise rendering in e.g. the viewer is still broken (same as before PR 12726 landed).
*Please note:* We still need to also load standard font data for patterns and/or some text-rendering modes, however that will require larger changes so I figured that it cannot hurt to submit *this* patch right now.
*This implementation is basically a copy of the pre-existing `builtInCMapCache` implementation.*
For some, badly generated, PDF documents it's possible that we'll end up having to fetch the *same* standard font data over and over (which is obviously inefficient).
While not common, it's certainly possible that a PDF document uses *custom* font names where the actual font then references one of the standard fonts; see e.g. issue 11399 for one such example.
Note that I did suggest adding worker-thread caching of standard font data in PR 12726, however it wasn't deemed necessary at the time. Now that we have a real-world example that benefit from caching, I think that we should simply implement this now.
- Some js files contain scale factors for each glyph in order to rescale Liberation to have a final font with the correct width.
- A lot of XFA have some containers where their dimensions are based on their text content, so using default font from browser can lead to an almost unreadable pdf.
This patch uses the new option added in PR 12726 to *also* allow fetching binary CMap data directly in the worker-thread in browsers.
Given that these changes remove the need to transfer data between threads for the default (browser) use-case, we can also revert the changes in PR 11118 since that simplifies the overall implementation.
This is necessary now, since with the previous patch the /FontBBox potentially depends on the contents of the /CharProcs-streams.
Note that if `getOperatorList` is called *before* `getTextContent`, this patch doesn't matter since the font is already fully loaded/parsed. However, for e.g. the `text` test-cases this is necessary to ensure correct reference images.
For Type3 fonts where the /CharProcs-streams of the individual glyph starts with a `d1` operator, we can use that to build a fallback bounding box for the font and thus improve text-selection in some cases.
The building of glyph paths, in the `FontRendererFactory`, can fail in various ways for corrupt font data. However, we're currently not attempting to handle any such errors in the evaluator, which means that a single broken glyph *can* prevent an entire page from rendering.
To address this we simply have to pass along, and check, the existing `ignoreErrors` option in `PartialEvaluator.buildFontPaths` similar to the rest of the `PartialEvaluator` code.
As can be seen in PR 13371, some of the `no-var` changes in the `PartialEvaluator.{getOperatorList, getTextContent}` methods caused errors in `gulp server`-mode.
However, there's a handful of instances of `var` in other methods which should be completely *safe* to convert since there's no strange scope-issues present in that code.
This reverts commit 0ef9b5aafc, since it cases a lot of warnings (see below) *locally* with e.g. the document from issue 9627.
Strangely enough, this only occurs with `gulp server`-mode and the actual builds are apparently fine. It seems that this *may* be some unfortunate interaction with the old Babel-plugin that's used together with SystemJS.
```
Warning: getTextContent - ignoring ExtGState: "FormatError: ExtGState should be a dictionary.".
```
Rather than taking the risk that this could actually cover a more serious bug, and since I cannot immediately figure out what's wrong, it thus seem safest to revert this for now and we can (carefully) revisit this once SystemJS has been removed (see PR 12563).
Compared to other data-structures, such as e.g. `Dict`s, we're purposely *not* caching Streams on the `XRef`-instance.[1]
The, somewhat unfortunate, effect of Streams not being cached is that repeatedly getting the *same* Stream-data requires re-parsing/re-initializing of a bunch of data; see `XRef.fetch` and related methods.
For the font-parsing in particular we're currently fetching the `toUnicode`-data, which is very often a Stream, in `PartialEvaluator.preEvaluateFont` and then *again* in `PartialEvaluator.extractDataStructures` soon afterwards.
By instead letting `PartialEvaluator.preEvaluateFont` export the "raw" `toUnicode`-data, we can avoid *some* unnecessary re-parsing/re-initializing when handling fonts.
*Please note:* In this particular case, given that `PartialEvaluator.preEvaluateFont` only accesses the "raw" `toUnicode` data, exporting a Stream should be safe.
---
[1] The reasons for this include:
- Streams, especially `DecodeStream`-instances, can become *very* large once read. Hence caching them really isn't a good idea simply because of the (potential) memory impact of doing so.
- Attempting to read from the *same* Stream-instance more than once won't work, unless it's `reset` in between, since using any method such as e.g. `getBytes` always starts at the current data position.
- Given that parsing, even in the worker-thread, is now fairly asynchronous it's generally impossible to assert that any one Stream-instance isn't being accessed "concurrently" by e.g. different `getOperatorList` calls. Hence `reset`-ing a cached Stream-instance isn't going to work in the general case.
Rather than re-fetching/re-parsing these properties immediately in `PartialEvaluator.translateFont`, we can simply export them instead. (Obviously the effect will be really tiny, but there is less parsing overall this way.)
*Please note:* While I don't have a document that this patches fixes, the current code is however not entirely correct as far as I can tell.
Looking at how the `Widths` array is parsed in `PartialEvaluator.extractWidths`, it's clear that the implementation in `PartialEvaluator.preEvaluateFont` is a bit too simplistic. In particular, by only wrapping the data into a TypedArray, there's no attempt to handle *indirect* objects which could potentially lead to colliding `hash`es being computed.
To hopefully help prevent any future bugs, make sure that composite/non-composite fonts cannot accidentally get matching `hash`es. Given the differences between those font types, that's very unlikely to be useful or even correct in general.
Without this some *composite* fonts may incorrectly end up with matching `hash`es, thus breaking rendering since we'll not actually try to load/parse some of the fonts.
*Please note:* Given that the document, in the referenced issue, doesn't embed *any* of its fonts there's no guarantee that it renders correctly in all configurations even with this patch.
- `FontFlags`, is used in both `src/core/fonts.js` and `src/core/evaluator.js`.
- `getFontType`, same as the above.
- `MacStandardGlyphOrdering`, is a fairly large data-structure and `src/core/fonts.js` is already a *very* large file.
- `recoverGlyphName`, a dependency of `type1FontGlyphMapping`; please see below.
- `SEAC_ANALYSIS_ENABLED`, is used by both `Type1Font`, `CFFFont`, and unit-tests; please see below.
- `type1FontGlyphMapping`, is used by both `Type1Font` and `CFFFont` which a later patch will move to their own files.
