TODO

* Decree that versions older than 2 years are not supported;
  yet, keep supporting an upgrade from whatever vendors provide
  — so push them to provide recent stuff.
** Already, swank-asdf doesn't support anything older than 2.014.6
   (as provided by the first 2011 iteration of Quicklisp), and
   I (fare) am pushing for swank-asdf to not support anything older
   than 3.1.2 (since no maintained implementation has anything older,
   with the LispWorks 7.0 release being last to upgrade, in May 2015;
   or CLISP hg in May 2016, though it hasn't been released since 2010).
** Maybe remove old versions from upgrade tests; or maybe not:
   if we support old versions in any way,
   it should be for upgrade only, and "punt" is acceptable.
** Upgrade tests might be split in two test sets: short and long.
** Make tests stateless and/or make any modifications private,
   so that they can be run in parallel.

* Have a better story for linking C static or dynamic libraries.
** Some groundwork has been done in cffi-toolchain, but an interface must be designed.
** Then you'll have to migrate all the code in Quicklisp to use your new interface.
** See commonqt's qt.asd and qtools's qt-libs.asd for instances of ugliness.

* Testing support
** Replace test-op with run-test-op and test-report-op ?

* Implement a style-checker, that will issue STYLE-WARNINGs
  when you don't follow all the recommended guidelines.

* Moving toward cross-compilation (ASDF 3.4? 4?)
   This would allow to get most of the benefits of XCVB
   (reproducibility, scalability, applicability to mobile platforms, etc.),
   except without the backward incompatibility of XCVB.
   NB: the MOCL author would benefit a lot, and might chip in.
** Need to further split asdf:perform in two parts:
   A part that runs on the host and specifies (as SEXP or string to read)
   forms that must be evaluated on the target machine, not the host:
   (asdf:perform-forms target operation component &key input-files output-files)
   The default implementation can fallback to running the
   (asdf:perform operation component) on the target.
   To ensure isolation, it would run in an isolated process in which
   just the specially crafted operation and component objects are defined,
   though the defsystem-depends-on dependencies are also loaded;
   the action would be run in a temporary directory, that would contain
   only a read-only copy of the declared input-files, and some writable
   output directory for the output files, that would afterwards be
   copied to their actual destination.

   ;; null means the target is the current host.
   ;; Or should the base class be plan-traversal or such?
   (defmethod perform-form ((target null) operation component)
     ;; :host means execute on the host. () is empty options.
     `(:host () (asdf:perform ',op ',comp)))

   ;; external-target means a cross compiler that we spawn.
   (defmethod perform-form ((target external-target) operation component)
     `(:target ;; who is to execute that? the target Lisp.
       ;; options for the performing system:
       (:depends-on ,(defsystem-dependencies
                      (component-system component)))
       ;; The actual forms to evaluate
       (asdf:perform ,(fake-operation-form operation)
                     ,(fake-component-form component))))

   By specifying :host-common-lisp you can have forms run on the host.
   By specifying :run-program, you can spawn a subprocess without
   having to first spawn a Lisp that will call run-program. etc.

   The other way around the default perform method could be:

   (defmethod perform ((o t) (c t))
     (let ((forms (perform-forms nil o c)))
       (if (equal forms
                  `((asdf:perform ',o ',c)))
           (error "Missing perform method on ..." ...)
           (map () 'eval-thunk forms))))

   Note that in the general case, there are multiple forms,
   that must be provided on the command line somehow
   (see the lisp-invocation system and also how cl-launch does it),
   with packages necessary to read latter forms being
   created by former forms, so you cannot read everything in one go.
   ASDF or at least UIOP should probably be provided in the target
   image by default, anyway, without needing to be explicitly specified.
   A more declarative interface to needing the defsystem-deps
   is probably useful.
** A real cross-compiling backend that works the way of XCVB
   would need to compute a complete graph for the plan,
   which could be achieved by inheriting from the code in POIU.
   Thinking about it, the cross-compiling target object
   could actually be the same as the plan object.
   The basic case of in-image execution would need to be in ASDF,
   while the actual cross-compilation case would be in an extension.
** Audit all the current specializers of perform, and offer them a new way;
   they can keep the perform method to run on older versions of ASDF.
** Modify SBCL and other implementations so they provide UIOP
   independently from ASDF?
** The default perform-plan method could iterate over pre-computed forms,
   in a way that does not involve the ASDF object model anymore,
   making ASDF friendlier to self-upgrades.
   Then, ASDF wouldn't need to be concatenated during upgrade
   — but should still concatenate for bootstrap purposes,
   when require'ing it as an implementation-provided module,
   compiling from ASDF 2, or otherwise distributing as a single file.

* Implementation bugs
** CMUCL has bugs:
*** In ASDF 2.32.13, sibling-dependencies is renamed to sideway-dependencies
    in defclass component, a superclass of system, and upgrade from
    2.32 or earlier thus fails with
    Error in function KERNEL:CLASS-TYPEP:  Class is currently invalid:
    #<KERNEL::STANDARD-CLASS ASDF/SYSTEM:SYSTEM {58398145}>
    in a direct call from ASDF/FIND-SYSTEM:LOCATE-SYSTEM,
    which suggests CMUCL's PCL fails to finalize classes properly.
*** Non-default optimize settings may or may not subtly break PCL.
    For the longest time (from 2.26.136 to 3.1.0.42), ASDF had:
    (declaim (optimize (speed 1) (safety 3) (debug 3)))
    This form may or may not work around that breakage in some cases:
    (setf c::top-level-lambda-max 0)
*** At some point there was massive breakage due to uninitialized slots
    as revealed in 3.1.0.37 with
    make t l=cmucl t=test-encodings.script
    which was a different PCL bug from the above (without same workaround)
    but possibly also triggered by the non-standard declaim.
*** `#5(1 ,@\`(2 3)))` returns `#(1 2 3)`,
     rather than `#(1 2 3 2 3 2 3 2 3)` or even better `#(1 2 3 3 3)`.
*** It purports to support Unicode, but it has only 16-bit characters.
   Note that (subtypep 'character 'base-char) says T -- everything is a base char(!)
*** In general, running upgrade tests triggers massive CLOS breakage on CMUCL.

** SCL has bugs:
*** it doesn't like run-program from a modified directory?
*** it somehow pushes `:non-base-chars-exist-p` even though `+non-base-chars-exist-p+` is `NIL`???

** XCL has bad bugs:
*** make-pathname doesn't handle :type nil properly and
    has massive lossage in logical-pathname support.
*** If using block () and return in search-for-system-definition
    instead of return-from, XCL mistakenly has the return jump
    to the end of the try function or something.
*** It also takes a long time for errors to propagate into a debugging REPL —
     Maybe O(n^3) behavior or worse in walking the C++ stack?
*** The backtrace frames may contains unprintable objects.
*** make-pathname fails to copy the version from the defaults.
*** (make-pathname :host "ASDFTEST" :directory '(:absolute "system2" "module4")
      :name nil :type nil) fails to handle the logical-pathname-ness of ASDFTEST.
*** `#5(1 ,@`(2 3)))` returns #(1 2 3),
     rather than #(1 2 3 2 3 2 3 2 3) or even better #(1 2 3 3 3).
*** XCL recognizes :unspecific for a pathname type,
   but prints the namestring with a dot separator!
   (make-pathname :name "foo" :type :unspecific) ==> #P"foo."
    If bug is ever fixed, upgrade *unspecific-pathname-type* in uiop/pathname.
*** XCL doesn't properly process files the name of which contains a *
*** XCL fails to recognize that specialized method parameters are implicitly ignorable.

** GCL is almost working again; but implementation bugs remain.
   See Francois-Rene Rideau's messages on gcl-devel starting November 2013.
*** Missing support for many UIOP functions:
   combine-fasls, get-optimization-settings...
*** ASDF may now be included for (require "asdf") and (require "ASDF"),
    but how to compile it properly?
*** GCL fails to properly compile ASDF with the default compilation settings,
   but works with those in script-support.lisp.
   Symptoms when trying to load asdf.o:
   Signalled by "AN ANONYMOUS FUNCTION".
   INTERNAL-SIMPLE-ERROR: The package named ASDF/INTERFACE, does not exist
*** Another GCL compiler bug:
   when I changed the definition of getcwd from
   `(let ((*default-pathname-defaults* #p"")) (truename #p""))` to
   `(let ((*default-pathname-defaults* *nil-pathname*)) (truename *nil-pathname*))`
   to guard against e.g. a logical-pathname context while loading asdf
   and parsing `#p""`, calls to `getcwd` result in a segfault.
*** An another bug: gcl refuses dynamic-extent declaration on functions.
   ```uiop/stream.lisp:         #-gcl (declare (dynamic-extent ,@(when
   before `(#',beforef)) ,@(when after `(#',afterf))))```
*** `(typep p 'logical-pathname)` should be T if p has a logical-pathname host.
*** `apropos` is case-sensitive and returns a same symbol many times
   (presumably once per package where it is present,
   instead of just once for its home package)
*** `compile-file` fails to return proper secondary values in case of non-style `WARNING`.
*** `(pathname-directory #p"foo/")` is incorrectly `("foo")` instead of `(:RELATIVE "foo")`
*** Do whatever it takes to pass the asdf tests, add the above?
*** Trying to `uiop:slurp-stream-forms` from a stream with `#+(or) :foo`
   (or `read-file-forms` from an file with same) results in an error,
   rather than `nil`. This is probably a bug in `#+` processing.
   Unhappily, debian creates such a file in
   `/etc/common-lisp/asdf-output-translations.conf.d/01-common-lisp-controller.conf`
*** Tests that try to catch an error fail (but catching a warning succeeds),
   which suggests brokenness in handler-bind and/or error.
*** `COMPILE-FILE*` fails due to handler-bind error in
   `with-muffled-compiler-conditions` or so.
*** `#5(1 ,@`(2 3)))` returns `#(1 2 3)`,
     rather than `#(1 2 3 2 3 2 3 2 3)` or even better `#(1 2 3 3 3)`.
*** `(DIRECTORY #p"*.*")` fails to match files with pathname-type `NIL`.
*** GCL doesn't properly process files the name of which contains a `*`

** ABCL has a few bugs.
*** ABCL fails the timestamp propagation test.
*** `#5(1 ,@`(2 3)))` returns #(1 2 3),
   rather than #(1 2 3 2 3 2 3 2 3) or even better #(1 2 3 3 3).
*** ABCL doesn't properly process files the name of which contains a *

** ECL has issues with its bundles.
   On MacOS X, test-bundle.script fails.
*** ECL doesn't properly process files the name of which contains a *

** Allegro fails to parse argv the correct Windows way wrt C++ escaping conventions.
   Is this (still) the case? What about LispWorks? Other implementations?

* Design & Implement some out-of-line configuration mechanism for various options?
  i.e. let configuration files override some variables around some actions.
** Could be used to override new backward-incompatible defaults
   (e.g. for operation propagation, deferred warnings, etc.)
** Use the above to bind variable around performing of actions.

* Improve robustness
** Fix bugs on http://bugs.launchpad.net/asdf/+bugs
*** fix directory-files to not return directories on CCL, etc. Add tests.
*** have a better defsystem form verifier - see lp#1007335
*** have a function verify-strict-asd that can verify a asd is pure lp#541562
   Then if it passes, use load-strict-asd.
   If not, optionally issue a warning. Start migrating people toward that.
** Include tests for prebuilt-system in test-bundle and in test-program.
** Test that a package-system secondary system is properly superseded
   when the primary system is redefined.

* Learn to use cl-grid-test
** So we can easily make sure ASDF changes don't break stuff,
   and that breakage gets fixed quickly.

* Diagnostics
** have a mode to explain WHY a component needs to be recompiled.
** A ``dry run'' of an operation can be made with the following form:
     (let ((asdf:*verbose-out* *standard-output*))
       (loop :for (op . comp) :in
         (asdf/plan:plan-actions
          (asdf/plan:make-plan nil '<operation-name> '<system-name> :force t))
         :do (asdf:explain op comp)))
   What would be a nice interface for this functionality?

* have with-input-file use the encodings mechanism?
** Import asdf-encodings into UIOP? Sounds a lot of code for no users.

* have a single test .asd that tests as many features as possible, use it for upgrade test.

* operation cleanup?
** Refactor operation classes in action.lisp to not use slots for metadata
*** See commit message for 8e68ac229c19b7b611caf942a7f0998c88047a79
*** The problem: slots will NOT get properly upgraded with new initform.
*** Bad solution: A U-I-F-R-C method might do but would be much more complex to maintain,
not just for ASDF but potentially for anyone who defines extensions to ASDF.
*** The solution: use defmethod self-operation, not a self-operation slot.
This solution is trivially backward-compatible with older variants of ASDF.
*** Migration: find everyone on Quicklisp that defines an operation, and have them use defmethod.
Document all that in asdf.texinfo. Yet, unless noone is affected in Quicklisp,
don't modify ASDF classes themselves until a year or two afterwards, for compatibility purposes;
and/or when ASDF classes themselves need to be modified for the sake of some fix.
** To allow semantically distinct operations of the same class:
You'd need to have a protocol to canonicalize them
in the *OPERATIONS* memoization table, not by class name,
but by CONS of the class name and some CANONICAL-OPERATION-INITARGS.
The latter would be a generic function called on the initargs,
where any parasite initargs such as FORCE and FORCE-NOT have been removed,
since they below to the PLAN, not the OPERATION:
the OPERATE protocol would be refined to explicit split
arguments to be passed to MAKE-PLAN or to MAKE-OPERATION.
The default method for CANONICAL-OPERATION-INITARGS
would SORT (a plist->alist of) the initargs,
and that would replace the current ORIGINAL-INITARGS slot.
For this scheme to work even in presence of undisciplined users
using MAKE-INSTANCE on an operation class,
the OPERATION class would have an extra slot EFFECTIVE-OPERATION,
uninitialized by default (nil or unbound), whose accessor initializes it
if it's uninitialized, by looking up a canonical instance in *OPERATIONS*,
and if unfound registering the current operation as canonical.
Then, each component's COMPONENT-OPERATION-TIME hash-table
would be indexed by canonicalized operation object
rather than by operation class,
and POIU would have to be changed accordingly.
Of course, this entire cleanup is incompatible
with how SWANK and GBBopen currently abuse slots of operation,
so these would have to be fixed first.
And that's why I didn't do it.
It looks like SWANK can be fixed soon, though, so we'll see.

* Hunt systems that use obsolete APIs

* Get rid of component-properties
** directly use component-properties:
*** yaclml.asd => misguided :features for version comparison
*** amazon-ecs-20110418-git/amazon-ecs.asd => :website property
*** hemlock.asd => bad implementation of latin1 encodings
** .asd use :properties
*** arnesi, arnesi+ => :features, misguided for version comparison and pseudo #+features
*** lkcas, thopter => :long-name
*** cl-irc cliki-bot rss cl-syslog com.informatimago.clext
   com.informatimago.clisp com.informatimago.clmisc
   com.informatimago.common-lisp.arithmetic
   com.informatimago.common-lisp.bank
   com.informatimago.common-lisp.cesarum
   com.informatimago.common-lisp com.informatimago.common-lisp.csv
   com.informatimago.common-lisp.cxx
   com.informatimago.common-lisp.data-encoding
   com.informatimago.common-lisp.diagram
   com.informatimago.common-lisp.ed
   com.informatimago.common-lisp.graphviz
   com.informatimago.common-lisp.heap
   com.informatimago.common-lisp.html-base
   com.informatimago.common-lisp.html-generator
   com.informatimago.common-lisp.html-parser
   com.informatimago.common-lisp.http
   com.informatimago.common-lisp.interactive
   com.informatimago.common-lisp.invoice
   com.informatimago.common-lisp.lisp
   com.informatimago.common-lisp.lisp.ibcl
   com.informatimago.common-lisp.lisp.stepper
   com.informatimago.common-lisp.lisp-reader
   com.informatimago.common-lisp.lisp-sexp
   com.informatimago.common-lisp.lisp-text
   com.informatimago.common-lisp.parser
   com.informatimago.common-lisp.picture
   com.informatimago.common-lisp.regexp
   com.informatimago.common-lisp.rfc2822
   com.informatimago.common-lisp.rfc3548
   com.informatimago.common-lisp.telnet
   com.informatimago.common-lisp.unix
   linc
   com.informatimago.lispdoc
   com.informatimago.lua
   com.informatimago.cocoa-playground
   com.informatimago.objcl
   com.informatimago.rdp
   com.informatimago.rdp.basic
   com.informatimago.rdp.basic.example
   com.informatimago.rdp.example
   com.informatimago.susv3
   com.informatimago.common-lisp.tools.make-depends
   com.informatimago.xcode
   spartns
   xlunit
   => (uninterned!)
   #:author-email #:date
   (#:albert #:output-dirs)
   (#:albert #:formats)
   (#:albert #:docbook #:template)
   (#:albert #:docbook #:bgcolor)
   (#:albert #:docbook #:textcolor)
   (#:albert #:docbook #:dtd)
*** portableaserve
   =>
   ("system" "author" "email")
   ("albert" "presentation" "output-dir")
   ("albert" "presentation" "formats")
   ("albert" "docbook" "dtd")
   ("albert" "docbook" "template")
*** com.clearly-useful.generic-collection-interface => :com.clearly-useful
*** metatilities
   => :ait-timeout :system-applicable-p
*** ucw ucw-core
   =>
   version

* ASDF4: search for this tag, rename things (incompatibly, thus) and cleanup code.
** Migrate from component-depends-on to action-depends-on
*** I contend a future version of ASDF will replace
`(component-depends-on operation component)`
with `(action-depends-on plan operation component)`.
This allows for different normalization strategies for dependencies
(including strategies that avoid resolving things to NIL),
a reified context for featurep checks, etc.
*** Easy but long backward-compatible transition:
**** get all users to use the new gf and keep extending both new and old gf,
   meanwhile the new gf has an around method that actually calls the old gf
   except in testing mode and/or if we can check that they're using the new convention
**** when everyone has migrated, remove the old mode and the short-circuit.
*** However, we cannot deprecate component-depends-on yet — not until we have
   some transition in place to a better interface.

* Documentation
** See message from 2014-01-27 05:26:44 GMT for stuff to document.
   http://thread.gmane.org/gmane.lisp.asdf.devel/3675/focus=3695
** style guide for .asd files

* UIOP
** deferred-warnings support probably needs to be redone better.
*** implement deferred warnings support on lispworks
*** turn undefined function style-warnings into full warnings(?)
*** work with cl-test-grid to eliminate deferred warnings from quicklisp libraries
*** https://www.mail-archive.com/asdf-devel@common-lisp.net/msg03608.html
*** czak@google.com rewrote part of deferred warnings support.
   Ask him to release the code he implemented at Google under QITAB.

** Ensure all run-program issues on Windows are solved.
*** SBCL cannot do raw CMD.EXE command-lines
*** We don't seem to properly escape CRLF in escape-windows-token,
   or maybe it is, but CMD.EXE does the wrong thing in them,
   so the semantics depend on whether it's used,
   or maybe it's CCL eating them while parsing, or maybe not, etc.
   In any case, this requires clarification.

** define-package cannot deal with symbols moving "forward".
*** document that THOU SHALT NOT USE :RECYCLE with packages previously
   defined by define-package when building from scratch.
*** Maybe add a check? But how to tell whether you're building from scratch?
*** Add some :moved-to directive to define-package, such that
   (:moved-to forward-package symbols...)
   will declare symbols as moved forward: when the symbol already exists,
   it is moved to the forward package if that package already exists,
   after the package created if it doesn't exist yet (!)
*** There again, a check that a forward-package is not backward
   would be very nice.

* Faster source-registry:
** In addition and/or as a substitute to the .cl-source-registry.cache,
   that is meant to be semi-automatically managed, there could be
   a cl-source-registry.conf meant for manual management:
   when recursing into a source-registry :tree, if such file is present
   (or if not, if a hidden .cl-source-registry.conf is present instead?),
   its contents is read as a (:source-registry ...) specification, and
   replaces the actual or notional (:tree ...) for the current directory;
   it may then include :file entries as well as :directory and :tree entries,
   whereby the programmer can explicitly give a definitive list of
   systems exported by his software, while excluding any test system
   that he doesn't want to export. This means that developers have
   both a way of speeding up the build of their software and of
   avoiding pollution by test systems that should remain private,
   and that they can otherwise explicitly enable when they need them.
** The .cl-source-registry.cache should avoid recursing into
   directories that themselves already have a cache or a .conf file,
   but instead store the name of these directories,
   so these files will be read recursively by asdf, allowing for
   semi-modular updates.

* Properly deprecate all that needs to go
** Get lisa, readable, bourbaki to NOT include ASDF 1 anymore in their sources.
** Identify all the things in ASDF that we want to deprecate, and make
   sure we deprecate them all in 3.4.
** Implement support for deprecating old packages,
   by having all functions and macros under the old package name be
   proxies to those in the new package name?
   (Exception: special variables can't be proxies, have to be imported.)
** Delete asdf-driver.asd, nobody has used it in years
** Eradicate Quicklisp systems that depend on asdf-package-system, make them depend on asdf3.1 instead.
** Deprecate and remove asdf-binary-locations (ABL) compatibility altogether.
  No one has been using ASDF1 for years.
  There is no ABL to be compatible with —
  is anyone still using the compatibility mode?
  Maybe issue a warning, then a cerror, before to remove the code?



* Fix plan
** Goal: Implement proper incrementality for defsystem-depends-on.
Fix the bug wherein ASDF fails to properly handle incrementality for `defsystem-depends-on`.
https://bugs.launchpad.net/asdf/+bug/1500578
Some discussion about ASDF at
http://ccl.clozure.com/irc-logs/lisp/lisp-2016-10.txt

** DONE Items
*** Add `DEFINE-OP` for loading system definitions.
*** A `DEFINE-OP` will depend on other `DEFINE-OP`s
*** A `DEFINE-OP` depends on all `LOAD-OP` in the file.
Whether via defsystem-depends-on or manual `LOAD-OP`s.
*** Store timestamps in `DEFINE-OP` entries, not `*DEFINED-SYSTEMS*`.
*** In find-system, add condition that all dependencies are up-to-date.
*** This necessitates another layer of visit states,
that do not plan, just check planning...
*** `LOAD-ASD` calls `PERFORM DEFINE-OP` (not the other way around)
*** Systems probably need depend on `(DEFINE-OP . PRIMARY-SYSTEM)`
*** These dynamically-discovered dependencies are stored in a slot of `SYSTEM`.
*** Have a special kind of definition-depends-on dependency.
Before we reuse a node from a previous session,
we must make sure it is still properly defined.
This matters because we do reuse (at least *some* nodes, e.g. system nodes)
from previous sessions, and the control structure may have changed
(i.e. the system hasn't been redefined, but one of its defsystem-depends-on
dependencies has changed).
*** Handle incomplete definitions properly
Don't just let the incompletely parsed system be registered
in a false positive success as if it were an empty system.
Solution: It is registered, but considered out-of-date,
so next attempt to use it will cause a refresh attempt.
https://bugs.launchpad.net/asdf/+bug/1515372
*** Handle nested calls to `OPERATE`
We must accept that ASDF isn't really plan-then-perform,
as the ASDF1 model claimed to be.
It's an arbitrary nesting of plan-and-perform, like a parenthesization
with plan (steps) as left parenthesis and perform (steps) as right parenthesis,
and with operate being a maximal balanced span of plan and perform steps.
*** What goes in a shared dynamic `*SESSION*`, what in a private lexical `PLAN`?
Memoization CACHE of various functions, Status of VISITED-NODES, FORCING parameters.
The session will contain not just the current `*CACHE*`, but also
all or most of the graph traversal, thus the `:force` options, etc.
The `*PLAN*` may or may not contain some work queueing.
But even, e.g. a table of "nodes still directly blocking us"
is shared at the shared traversal level.
*** It is an error for a nested `OPERATE` to contradict session flags.
This raises the question of `:force` and `:force-not` flags.
An actual operate cannot contradict them and cause things to be performed
with a wrong subset of dependencies loaded.
Yet, `REQUIRED-COMPONENTS` (that does cause anything to be performed —
except via nested `OPERATE` from `defsystem-depends-on`)
must be able to skip some dependencies
(though maybe only _in addition_ to the global ones).
*** What timestamp for `forced` and `forced-not` actions?
If an action is forced-not, it should be returned with an action stamp from this cache,
and not with `nil` as is currently wrongly done
(unless of course the cache entry is missing,
at which point `nil` is the correct answer).
Conversely, a force means the stamp should be `t`
(indicating it needs to be done in the future)
(that part asdf gets right).
*** Handle planning things that are now done but were done in a previous phase of the build,
so dependency on them should trigger the build.
*** Protocol to not re-perform an action twice
In PERFORM-PLAN, check that an action wasn't already DONE-P due to
being needed in a previous phase of recursively calling OPERATE,
before to PERFORM it.
To correctly propagate timestamps in across nested calls,
the graph traversal and timestamp propagation is shared at the session level
between all nested calls to `OPERATE`.
During a session and across phases, every action has a status made of
a stamp and three bits: KEEP-P, DONE-P, NEED-P.
The stamp also survives across sessions, in `COMPONENT-OPERATION-TIMES`.
*** Detect circularities across nested calls to `OPERATE`
This means the circularity detection data is part of the session.
Notably, the visiting list and set.
The top of the list can also serve to record the dependencies from `OPERATE` to `DEPEND-OP`.
*** In UIOP, RECORD-DEPENDENCY must not cross to an upper plan and/or it should
record the dependency in said upper plan (that matters because
parallel plans doing extra checking, and this is a problem
whereby some links to extra actions are left unresolved).
Maybe in the future, record the dependency in the session, not in a plan.

** TODO items
*** find-system should probably cause a dependency on define-op, especially inside another define-op
*** Is it worth it trying to optimize away a dependency on a define-op
when an action already depends on another operation that "obviously" depends on that define-op?
*** Document DEFINE-OP and the design in asdf.texinfo and/or in some article.

** Future developments
*** Grovel all .asd files in Quicklisp and collect all violations of declarativeness,
where declarativeness is having only defsystem forms. Specially classify in-package,
maybe also defpackage. Then we can assess the impact of requiring declarativeness.
*** Also grovel Quicklisp for all extensions to ASDF:
see if there are new methods on perform, input-files, output-files, operate, etc.
*** To allow defsystem-depends-on to another secondary system within the
same file, I can imagine some system to dynamically create a separate action
(DEFINE-OP foo/x) for each form defsystem foo/x, and have it depend on
all the preceding actions operated in the .asd file.
Plus a last action for dangling load-system or operate statements?
Or one action for every operate statement?
There would be no action of the entire load-asd, but a separate mechanism
to detect circularity?
What about dangling statements? They would be a bit like xcvb weakly-ordered
run-time dependencies that CAN have circularities? Ouch. Really ugly.
Each statement would need its own action? Ouch.
*** Have load-asd clear any invalidated previous secondary system
before loading, which supposes remembering what these systems are.
Then, you're not allowed to forward-reference a secondary system that
hasn't been defined yet while you're loading the .asd, whereas you
can load the asd if you aren't loading it yet.
*** Maybe split DEFINE-OP into PREPARE-DEFINE-OP and DEFINE-OP, so that
secondary systems can be properly loaded?
*** Add mapping from files to action that creates it and/or uses it?
Then we can detect conflict between multiple actions creating one file.
Forbid direct use of files not listed as inputs? What about transitive inputs?
Build in a container that has only the proper files...
*** What data structure for scheduling planned actions?
A doubly-linked to remove entire chunks of the plan after a sub-operate?
That won't do, because later-discovered defsystem dependency can refer to
systems that were planned but not performed, so there need be no contiguity
in the chunks that are performed by subcalls to operate.
Actually, if we otherwise forbid reference to another system
from non-system components, then we could move things with system granularity,
which is one thing; but we still need a general mechanism at the system level.
So, whether it's an explicit queue or an implicit traversal of the dependency graph,
we must check whether an action has already been done before we try to perform it.
And an implicit traversal has the advantage that you can avoid entire subtrees
when you find that the current node has already been performed during
a previous phase of the plan.
In conclusion: "just" traverse the graph at the last minute, and
perform things serially, and/or if you want some parallelism,
build a queue dynamically at the last minute for the "now" jobs
while keeping most of the plan as graph structures.

*** Merging in parts of POIU?
ASDF internals are nothing to be proud of:
ASDF1 was optimized for "smallest code that kind of works in the usual case",
and ASDF3 was optimized for "smallest code that fixes ASDF1 in the general case
while mostly maintaining backward compatibility".

Because of its optimization constraint, ASDF1 was using and abusing lists a lot,
rather than defining nice data structures.
ASDF3 uses somewhat more structured data, with a few more classes, and
algorithms that are O(n) instead of O(n**3) in simple cases and
exponential in carefully crafted examples;
yet it has shunned any definition of more advanced or general data structures.
By contrast, POIU has a dequeue to represent queued compilation jobs, and
a general graph representation for the action graph
(though the original POIU was also O(n**3) because it lacked PREPARE-OP).

I could import the POIU queue and graph representations into ASDF, and
that would add about 100 lines of code to UIOP for the queue, and
200 lines to ASDF for the graph, maybe a bit more if the code is
generalized, commented and largely put in UIOP.

Then, there's the support for actual forking and using forks,
that would be about 400 lines added to UIOP, and
the using it in ASDF to consume the plan, which would be under 200 lines.
But those parts are probably better left out of ASDF itself.
That said, it all was a lot when ASDF was < 1000 lines of code,
but isn't all that much now that it is > 12000 lines of code.