qpdf/TODO

Documentation
=============

* See #530 -- add an appendix explaining PDF encryption in general
  plus how it's handled by qpdf.


Document-level work
===================

* QPDFPageCopier -- object for moving pages around within files or
  between files and performing various transformations

  * Handle all the stuff of pages and split-pages
  * Do n-up, booklet, collation
  * Look through cli and see what else...flatten-*?
  * See comments in QPDFPageDocumentHelper.hh for addPage -- search
    for "a future version".
  * Make it efficient for bulk operations
  * Make certain doc-level features selectable
  * qpdf.cc should do all its page operations, including
    overlay/underlay, splitting, and merging, using this
  * There should also be example code

* After doc-level checks are in, call --check on the output files in
  the "Copy Annotations" tests.

* Document-level checks. For example, for forms, make sure all form
  fields point to an annotation on exactly one page as well as that
  all widget annotations are associated with a form field. Hook this
  into QPDFPageCopier as well as the doc helpers. Make sure it is
  called from --check.

* See also issues tagged with "pages"

* Add flags to CLI to select which document-level options to
  preserve or not preserve. We will probably need a pair of mutually
  exclusive, repeatable options with a way to specify all, none, only
  {x,y}, or all but {x,y}.

* If a page contains a reference a file attachment annotation, when
  that page is copied, if the file attachment appears in the top-level
  EmbeddedFiles tree, that entry should be preserved in the
  destination file. Otherwise, we probably will require the use of
  --copy-attachments-from to preserve these. What will the strategy be
  for deduplicating in the automatic case?

Text Appearance Streams
=======================

This is a list of known issues with text appearance streams and things
we might do about it.

* For variable text, the spec says to pull any resources from /DR that
  are referenced in /DA but if the resource dictionary already has
  that resource, just use the one that's there. The current code looks
  only for /Tf and adds it if needed. We might want to instead merge
  /DR with resources and then remove anything that's unreferenced. We
  have all the code required for that in ResourceFinder except
  TfFinder also gets the font size, which ResourceFinder doesn't do.

* There are things we are missing because we don't look at font
  metrics. The code from TextBuilder (work) has almost everything in
  it that is required. Once we have knowledge of character widths, we
  can support quadding and multiline text fields (/Ff 4096), and we
  can potentially squeeze text to fit into a field. For multiline,
  first squeeze vertically down to the font height, then squeeze
  horizontally with Tz. For single line, squeeze horizontally with Tz.
  If we use Tz, issue a warning.

* When mapping characters to widths, we will need to care about
  character encoding. For built-in fonts, we can create a map from
  Unicode code point to width and then go from the font's encoding to
  unicode to the width. Get rid of "ugly switch statements" in
  QUtil.cc and replace with static map initializers. See
  misc/character-encoding/ (not on github) and font metric information
  for the 14 standard fonts in my local pdf-spec directory.

Fuzz Errors
===========

* https://bugs.chromium.org/p/oss-fuzz/issues/detail?id=<N>

* Ignoring these:
  * Out of memory in dct: 35001, 32516

GitHub Actions
==============

* Actions are triggered on push to main and master. When we eventually
  rename master to main, make sure the reference to master is removed
  from .github/workflows/*.yml.

* At the time of migrating from Azure Pipelines to GitHub Actions
  (2020-10), there was no standard test result publisher (to replace
  the PublishTestResults@2 task). There are some third-party actions,
  but I'd rather not depend on them. Keep an eye open for this coming
  to GitHub Actions.

External Libraries
==================

Current state (10.0.2):

* qpdf/external-libs repository builds external-libs on a schedule.
  It detects and downloads the latest versions of zlib, jpeg, and
  openssl and creates source and binary distribution zip files in an
  artifact called "distribution".

* Releases in qpdf/external-libs are made manually. They contain
  qpdf-external-libs-{bin,src}.zip.

* The qpdf build finds the latest non-prerelease release and downloads
  the qpdf-external-libs-*.zip files from the releases in the setup
  stage.

* To upgrade to a new version of external-libs, create a new release
  of qpdf/external-libs (see README-maintainer in external-libs) from
  the distribution artifact of the most recent successful build after
  ensuring that it works.

Desired state:

* The qpdf/external-libs repository should create release candidates.
  Ideally, every scheduled run would make its zip files available. A
  personal access token with actions:read scope for the
  qpdf/external-libs repository is required to download the artifact
  from an action run, and qpdf/qpdf's secrets.GITHUB_TOKEN doesn't
  have this access. We could create a service account for this
  purpose. As an alternative, we could have a draft release in
  qpdf/external-libs that the qpdf/external-libs build could update
  with each candidate. It may also be possible to solve this by
  developing a simple GitHub app.

* Scheduled runs of the qpdf build in the qpdf/qpdf repository (not a
  fork or pull request) could download external-libs from the release
  candidate area instead of the latest stable release. Pushes to the
  build branch should still use the latest release so it always
  matches the main branch.

* Periodically, we would create a release of external-libs from the
  release candidate zip files. This could be done safely because we
  know the latest qpdf works with it. This could be done at least
  before every release of qpdf, but potentially it could be done at
  other times, such as when a new dependency version is available or
  after some period of time.

Other notes:

* The external-libs branch in qpdf/qpdf was never documented. We might
  be able to get away with deleting it.

* See README-maintainer in qpdf/external-libs for information on
  creating a release. This could be at least partially scripted in a
  way that works for the qpdf/qpdf repository as well since they are
  very similar.

ABI Changes
===========

This is a list of changes to make next time there is an ABI change.
Comments appear in the code prefixed by "ABI"

* Search for ABI to find items not listed here.
* Merge two versions of QPDFObjectHandle::makeDirect per comment
* After removing legacy QPDFNameTreeObjectHelper and
  QPDFNumberTreeObjectHelper constructors, NNTreeImpl can switch to
  having a QPDF reference and assume that the reference is always
  valid.
* Use `= delete` and `= default` for constructors and destructors
  where possible
* Consider having setters return Class& where possible to allow for
  use of fluent interfaces
* Consider having addArrayItem, replaceKey, etc. return the new value
  so you can say
  auto oh = dict.replaceKey("/Key", QPDFObjectHandle::newSomething());
  But this has to be clean with respect to fluent interfaces, so we
  might need something slightly different.
* Added QPDFObjectHandle::ParserCallbacks::handleWarning but had to
  revert because it was not binary compatible. Consider re-adding. The
  commit that added this comment includes the reverting of the change.
  The previous commit removes the code that was calling and using
  handleWarning.
* Make it easier to deal with objects that should be indirect. Search
  for makeIndirectObject in the code to find patterns. For example, it
  would be nice to have a one-liner for the case of one or all
  dictionary values or array items being replaced with an indirect
  objects if direct. Maybe we want a version of copyForeignObject that
  takes the foreign qpdf and converts the source object to indirect
  before copying, though maybe we don't because it could cause
  multiple copies to be made...usually it's better to handle that
  explicitly.
* Deal with weak cryptographic algorithms:
  * Github issue #576
  * Add something to QPDFWriter that you must call in order to allow
    creation of files with insecure crypto. Maybe
    QPDFWriter::allowWeakCrypto. Call this when --allow-weak-crypto is
    passed and probably also when copying encryption by default from
    an input file. There should be some API change so that, when
    people recompile with qpdf 11, their code won't suddenly stop
    working. Getting this right will take careful consideration of the
    developer and user experience. We don't want to create a situation
    where exactly the same code fails to work in 11 but worked on 10.
    See #576 for latest notes.
  * Change deterministic id to use something other than MD5 but allow
    the old way for compatibility -- maybe rename the method to force
    the developer to make a choice
  * Find other uses of MD5 and find the ones that are discretionary,
    if any
  * Have QPDFWriter raise an exception if it's about to write using
    weak crypto and hasn't been given permission
  * Search for --allow-weak-crypto in the manual and in qpdf.cc's help
    information
  * Update the ref.weak-crypto section of the manual

Page splitting/merging
======================

 * Update page splitting and merging to handle document-level
   constructs with page impact such as interactive forms and article
   threading. Check keys in the document catalog for others, such as
   outlines, page labels, thumbnails, and zones. For threads,
   Subramanyam provided a test file; see ../misc/article-threads.pdf.
   Email Q-Count: 431864 from 2009-11-03.

 * bookmarks (outlines) 12.3.3
   * support bookmarks when merging
   * prune bookmarks that don't point to a surviving page when merging
     or splitting
   * make sure conflicting named destinations work possibly test by
     including the same file by two paths in a merge
   * see also comments in issue 343

   Note: original implementation of bookmark preservation for split
   pages caused a very high performance hit. The problem was
   introduced in 313ba081265f69ac9a0324f9fe87087c72918191 and reverted
   in the commit that adds this paragraph. The revert includes marking
   a few tests cases as $td->EXPECT_FAILURE. When properly coded, the
   test cases will need to be adjusted to only include the parts of
   the outlines that are actually copied. The tests in question are
   "split page with outlines". When implementing properly, ensure that
   the performance is not adversely affected by timing split-pages on
   a large file with complex outlines such as the PDF specification.

   When pruning outlines, keep all outlines in the hierarchy that are
   above an outline for a page we care about. If one of the ancestor
   outlines points to a non-existent page, clear its dest. If an
   outline does not have any children that point to pages in the
   document, just omit it.

   Possible strategy:
   * resolve all named destinations to explicit destinations
   * concatenate top-level outlines
   * prune outlines whose dests don't point to a valid page
   * recompute all /Count fields

   Test files
   * page-labels-and-outlines.pdf: old file with both page labels and
     outlines. All destinations are explicit destinations. Each page
     has Potato and a number. All titles are feline names.
   * outlines-with-actions.pdf: mixture of explicit destinations,
     named destinations, goto actions with explicit destinations, and
     goto actions with named destinations; uses /Dests key in names
     dictionary. Each page has Salad and a number. All titles are
     silly words. One destination is an indirect object.
   * outlines-with-old-root-dests.pdf: like outlines-with-actions
     except it uses the PDF-1.1 /Dests dictionary for named
     destinations, and each page has Soup and a number. Also pages are
     numbered with upper-case Roman numerals starting with 0. All
     titles are silly words preceded by a bullet.

   If outline handling is significantly improved, see
   ../misc/bad-outlines/bad-outlines.pdf and email:
   https://mail.google.com/mail/u/0/#search/rfc822msgid%3A02aa01d3d013%249f766990%24de633cb0%24%40mono.hr)

 * Form fields: should be similar to outlines.

Performance
===========

As described in https://github.com/qpdf/qpdf/issues/401, there was
great performance degradation between qpdf 7.1.1 and 9.1.1. Doing a
bisect between dac65a21fb4fa5f871e31c314280b75adde89a6c and
release-qpdf-7.1.1, I found several commits that damaged performance.
I fixed some of them to improve performance by about 70% (as measured
by saying that old times were 170% of new times). The remaining
commits that broke performance either can't be correct because they
would re-introduce an old bug or aren't worth correcting because of
the high value they offer relative to a relatively low penalty. For
historical reference, here are the commits. The numbers are the time
in seconds on the machine I happened to be using of splitting the
first 100 pages of PDF32000_2008.pdf 20 times and taking an average
duration.

Commits that broke performance:

* d0e99f195a987c483bbb6c5449cf39bee34e08a1 -- object description and
  context: 0.39 -> 0.45
* a01359189b32c60c2d55b039f7aefd6c3ce0ebde (minus 313ba08) -- fix
  dangling references: 0.55 -> 0.6
* e5f504b6c5dc34337cc0b316b4a7b1fca7e614b1 -- sparse array: 0.6 -> 0.62

Other intermediate steps that were previously fixed:

* 313ba081265f69ac9a0324f9fe87087c72918191 -- copy outlines into
  split: 0.55 -> 4.0
* a01359189b32c60c2d55b039f7aefd6c3ce0ebde -- fix dangling references:
  4.0 -> 9.0

This commit fixed the awful problem introduced in 313ba081:

* a5a016cdd26a8e5c99e5f019bc30d1bdf6c050a2 -- revert outline
  preservation: 9.0 -> 0.6

Note that the fix dangling references commit had a much worse impact
prior to removing the outline preservation, so I also measured its
impact in isolation.

A few important lessons:

* Indirection through PointerHolder<Members> is expensive, and should
  not be used for things that are created and destroyed frequently
  such as QPDFObjectHandle and QPDFObject.
* Traversal of objects is expensive and should be avoided where
  possible.

Future ideas:

* Look at places in the code where object traversal is being done and,
  where possible, try to avoid it entirely or at least avoid ever
  traversing the same objects multiple times.
* Avoid attaching too much metadata to objects and object handles
  since those have to get copied around a lot.

Also, it turns out that PointerHolder is more performant than
std::shared_ptr.

Analytics
=========

Consider features that make it easier to detect certain patterns in
PDF files. The information below could be computed using an external
program that reads the existing json, but if it's useful enough, we
could add it directly to the json output.

 * Add to "pages" in the json:
   * "inheritsresources": bool; whether there are any inherited
     attributes from ancestor page tree nodes
   * "sharedresources": a list of indirect objects that are
     "/Resources" dictionaries or "XObject" resource dictionary subkeys
     of either the page itself or of any form XObject referenced by the
     page.

 * Add to "objectinfo" in json: "directpagerefcount": the number of
   pages that directly reference this object (i.e., you can find an
   indirect reference to the object in the page dictionary without
   traversing over any indirect objects)

General
=======

NOTE: Some items in this list refer to files in my personal home
directory or that are otherwise not publicly accessible. This includes
things sent to me by email that are specifically not public. Even so,
I find it useful to make reference to them in this list.

 * Get rid of remaining assert() calls from non-test code.

 * Consider updating the fuzzer with code that exercises
   copyAnnotations, file attachments, and name and number trees. Check
   fuzzer coverage.

 * Add code for creation of a file attachment annotation. It should
   also be possible to create a widget annotation and a form field.
   Update the pdf-attach-file.cc example with new APIs when ready.

 * If I do more with json, take a look at this C++ header-only JSON
   library: https://github.com/nlohmann/json/releases

 * Flattening of form XObjects seems like something that would be
   useful in the library. We are seeing more cases of completely valid
   PDF files with form XObjects that cause problems in other software.
   Flattening of form XObjects could be a useful way to work around
   those issues or to prepare files for additional processing, making
   it possible for users of the qpdf library to not be concerned about
   form XObjects. This could be done recursively; i.e., we could have a
   method to embed a form XObject into whatever contains it, whether
   that is a form XObject or a page. This would require more
   significant interpretation of the content stream. We would need a
   test file in which the placement of the form XObject has to be in
   the right place, e.g., the form XObject partially obscures earlier
   code and is partially obscured by later code. Keys in the resource
   dictionary may need to be changed -- create test cases with lots of
   duplicated/overlapping keys.

 * Figure out how to render Gajić correctly in the PDF version of the
   qpdf manual.

 * Investigate whether there is a way to automate the memory checker
   tests for Windows.

 * Part of closed_file_input_source.cc is disabled on Windows because
   of odd failures. It might be worth investigating so we can fully
   exercise this in the test suite. That said, ClosedFileInputSource
   is exercised elsewhere in qpdf's test suite, so this is not that
   pressing.

 * If possible, consider adding CCITT3, CCITT4, or any other easy
   filters. For some reference code that we probably can't use but may
   be handy anyway, see
   http://partners.adobe.com/public/developer/ps/sdk/index_archive.html

 * If possible, support the following types of broken files:

    - Files that have no whitespace token after "endobj" such that
      endobj collides with the start of the next object

    - See ../misc/broken-files

    - See ../misc/bad-files-issue-476. This directory contains a
      snapshot of the google doc and linked PDF files from issue #476.
      Please see the issue for details.

 * Additional form features
   * set value from CLI? Specify title, and provide way to
     disambiguate, probably by giving objgen of field

 * replace mode: --replace-object, --replace-stream-raw,
   --replace-stream-filtered
   * update first paragraph of QPDF JSON in the manual to mention this
   * object numbers are not preserved by write, so object ID lookup
     has to be done separately for each invocation
   * you don't have to specify length for streams
   * you only have to specify filtering for streams if providing raw data

 * Pl_TIFFPredictor is pretty slow.

 * Support for handling file names with Unicode characters in Windows
   is incomplete. qpdf seems to support them okay from a functionality
   standpoint, and the right thing happens if you pass in UTF-8
   encoded filenames to QPDF library routines in Windows (they are
   converted internally to wchar_t*), but file names are encoded in
   UTF-8 on output, which doesn't produce nice error messages or
   output on Windows in some cases.

 * If we ever wanted to do anything more with character encoding, see
   ../misc/character-encoding/, which includes machine-readable dump
   of table D.2 in the ISO-32000 PDF spec. This shows the mapping
   between Unicode, StandardEncoding, WinAnsiEncoding,
   MacRomanEncoding, and PDFDocEncoding.

 * Some test cases on bad files fail because qpdf is unable to find
   the root dictionary when it fails to read the trailer. Recovery
   could find the root dictionary and even the info dictionary in
   other ways. In particular, issue-202.pdf can be opened by evince,
   and there's no real reason that qpdf couldn't be made to be able to
   recover that file as well.

 * Audit every place where qpdf allocates memory to see whether there
   are cases where malicious inputs could cause qpdf to attempt to
   grab very large amounts of memory. Certainly there are cases like
   this, such as if a very highly compressed, very large image stream
   is requested in a buffer. Hopefully normal input to output
   filtering doesn't ever try to do this. QPDFWriter should be checked
   carefully too. See also bugs/private/from-email-663916/

 * Interactive form modification:
   https://github.com/qpdf/qpdf/issues/213 contains a good discussion
   of some ideas for adding methods to modify annotations and form
   fields if we want to make it easier to support modifications to
   interactive forms. Some of the ideas have been implemented, and
   some of the probably never will be implemented, but it's worth a
   read if there is an intention to work on this. In the issue, search
   for "Regarding write functionality", and read that comment and the
   responses to it.

 * Look at ~/Q/pdf-collection/forms-from-appian/

 * Consider adding "uninstall" target to makefile. It should only
   uninstall what it installed, which means that you must run
   uninstall from the version you ran install with. It would only be
   supported for the toolchains that support the install target
   (libtool).

 * Provide support in QPDFWriter for writing incremental updates.
   Provide support in qpdf for preserving incremental updates.  The
   goal should be that QDF mode should be fully functional for files
   with incremental updates including fix_qdf.

   Note that there's nothing that says an indirect object in one
   update can't refer to an object that doesn't appear until a later
   update.  This means that QPDF has to treat indirect null objects
   differently from how it does now.  QPDF drops indirect null objects
   that appear as members of arrays or dictionaries.  For arrays, it's
   handled in QPDFWriter where we make indirect nulls direct.  This is
   in a single if block, and nothing else in the code cares about it.
   We could just remove that if block and not break anything except a
   few test cases that exercise the current behavior.  For
   dictionaries, it's more complicated.  In this case,
   QPDF_Dictionary::getKeys() ignores all keys with null values, and
   hasKey() returns false for keys that have null values.  We would
   probably want to make QPDF_Dictionary able to handle the special
   case of keys that are indirect nulls and basically never have it
   drop any keys that are indirect objects.

   If we make a change to have qpdf preserve indirect references to
   null objects, we have to note this in ChangeLog and in the release
   notes since this will change output files.  We did this before when
   we stopped flattening scalar references, so this is probably not a
   big deal.  We also have to make sure that the testing for this
   handles non-trivial cases of the targets of indirect nulls being
   replaced by real objects in an update.  I'm not sure how this plays
   with linearization, if at all.  For cases where incremental updates
   are not being preserved as incremental updates and where the data
   is being folded in (as is always the case with qpdf now), none of
   this should make any difference in the actual semantics of the
   files.

 * When decrypting files with /R=6, hash_V5 is called more than once
   with the same inputs.  Caching the results or refactoring to reduce
   the number of identical calls could improve performance for
   workloads that involve processing large numbers of small files.

 * Consider adding a method to balance the pages tree.  It would call
   pushInheritedAttributesToPage, construct a pages tree from scratch,
   and replace the /Pages key of the root dictionary with the new
   tree.

 * Secure random number generation could be made more efficient by
   using a local static to ensure a single random device or crypt
   provider as long as this can be done in a thread-safe fashion.  In
   the initial implementation, this is being skipped to avoid having
   to add any dependencies on threading libraries.

 * Study what's required to support savable forms that can be saved by
   Adobe Reader. Does this require actually signing the document with
   an Adobe private key? Search for "Digital signatures" in the PDF
   spec, and look at ~/Q/pdf-collection/form-with-full-save.pdf, which
   came from Adobe's example site. See also
   ../misc/digital-sign-from-trueroad/. If digital signatures are
   implemented, update the docs on crypto providers, which mention
   that this may happen in the future.

 * Provide APIs for embedded files.  See *attachments*.pdf in test
   suite.  The private method findAttachmentStreams finds at least
   cases for modern versions of Adobe Reader (>= 1.7, maybe earlier).
   PDF Reference 1.7 section 3.10, "File Specifications", discusses
   this.

   A sourceforge user asks if qpdf can handle extracting and embedded
   resources and references these tools, which may be useful as a
   reference.

   http://multivalent.sourceforge.net/Tools/pdf/Extract.html
   http://multivalent.sourceforge.net/Tools/pdf/Embed.html

 * Qpdf does not honor /EFF when adding new file attachments. When it
   encrypts, it never generates streams with explicit crypt filters.
   Prior to 10.2, there was an incorrect attempt to treat /EFF as a
   default value for decrypting file attachment streams, but it is not
   supposed to mean that. Instead, it is intended for conforming
   writers to obey this when adding new attachments. Qpdf is not a
   conforming writer in that respect.

 * The second xref stream for linearized files has to be padded only
   because we need file_size as computed in pass 1 to be accurate.  If
   we were not allowing writing to a pipe, we could seek back to the
   beginning and fill in the value of /L in the linearization
   dictionary as an optimization to alleviate the need for this
   padding.  Doing so would require us to pad the /L value
   individually and also to save the file descriptor and determine
   whether it's seekable.  This is probably not worth bothering with.

 * The whole xref handling code in the QPDF object allows the same
   object with more than one generation to coexist, but a lot of logic
   assumes this isn't the case.  Anything that creates mappings only
   with the object number and not the generation is this way,
   including most of the interaction between QPDFWriter and QPDF.  If
   we wanted to allow the same object with more than one generation to
   coexist, which I'm not sure is allowed, we could fix this by
   changing xref_table.  Alternatively, we could detect and disallow
   that case.  In fact, it appears that Adobe reader and other PDF
   viewing software silently ignores objects of this type, so this is
   probably not a big deal.

 * Based on an idea suggested by user "Atom Smasher", consider
   providing some mechanism to recover earlier versions of a file
   embedded prior to appended sections.

 * From a suggestion in bug 3152169, consider having an option to
   re-encode inline images with an ASCII encoding.

 * From github issue 2, provide more in-depth output for examining
   hint stream contents. Consider adding on option to provide a
   human-readable dump of linearization hint tables. This should
   include improving the 'overflow reading bit stream' message as
   reported in issue #2. There are multiple calls to stopOnError in
   the linearization checking code. Ideally, these should not
   terminate checking. It would require re-acquiring an understanding
   of all that code to make the checks more robust. In particular,
   it's hard to look at the code and quickly determine what is a true
   logic error and what could happen because of malformed user input.
   See also ../misc/linearization-errors.

 * If I ever decide to make appearance stream-generation aware of
   fonts or font metrics, see email from Tobias with Message-ID
   <5C3C9C6C.8000102@thax.hardliners.org> dated 2019-01-14.

 * Consider creating a sanitizer to make it easier for people to send
   broken files. Now that we have json mode, this is probably no
   longer worth doing. Here is the previous idea, possibly implemented
   by making it possible to run the lexer (tokenizer) over a whole
   file. Make it possible to replace all strings in a file lexically
   even on badly broken files. Ideally this should work files that are
   lacking xref, have broken links, duplicated dictionary keys, syntax
   errors, etc., and ideally it should work with encrypted files if
   possible. This should go through the streams and strings and
   replace them with fixed or random characters, preferably, but not
   necessarily, in a manner that works with fonts. One possibility
   would be to detect whether a string contains characters with normal
   encoding, and if so, use 0x41. If the string uses character maps,
   use 0x01. The output should otherwise be unrelated to the input.
   This could be built after the filtering and tokenizer rewrite and
   should be done in a manner that takes advantage of the other
   lexical features. This sanitizer should also clear metadata and
   replace images. If I ever do this, the file from issue #494 would
   be a great one to look at.

 * Here are some notes about having stream data providers modify
   stream dictionaries. I had wanted to add this functionality to make
   it more efficient to create stream data providers that may
   dynamically decide what kind of filters to use and that may end up
   modifying the dictionary conditionally depending on the original
   stream data. Ultimately I decided not to implement this feature.
   This paragraph describes why.

   * When writing, the way objects are placed into the queue for
     writing strongly precludes creation of any new indirect objects,
     or even changing which indirect objects are referenced from which
     other objects, because we sometimes write as we are traversing
     and enqueuing objects. For non-linearized files, there is a risk
     that an indirect object that used to be referenced would no
     longer be referenced, and whether it was already written to the
     output file would be based on an accident of where it was
     encountered when traversing the object structure. For linearized
     files, the situation is considerably worse. We decide which
     section of the file to write an object to based on a mapping of
     which objects are used by which other objects. Changing this
     mapping could cause an object to appear in the wrong section, to
     be written even though it is unreferenced, or to be entirely
     omitted since, during linearization, we don't enqueue new objects
     as we traverse for writing.

   * There are several places in QPDFWriter that query a stream's
     dictionary in order to prepare for writing or to make decisions
     about certain aspects of the writing process. If the stream data
     provider has the chance to modify the dictionary, every piece of
     code that gets stream data would have to be aware of this. This
     would potentially include end user code. For example, any code
     that called getDict() on a stream before installing a stream data
     provider and expected that dictionary to be valid would
     potentially be broken. As implemented right now, you must perform
     any modifications on the dictionary in advance and provided
     /Filter and /DecodeParms at the time you installed the stream
     data provider. This means that some computations would have to be
     done more than once, but for linearized files, stream data
     providers are already called more than once. If the work done by
     a stream data provider is especially expensive, it can implement
     its own cache.

   The example examples/pdf-custom-filter.cc demonstrates the use of
   custom stream filters. This includes a custom pipeline, a custom
   stream filter, as well as modification of a stream's dictionary to
   include creation of a new stream that is referenced from
   /DecodeParms.