(* $Id$ *)

theory integration imports base begin

chapter {* System integration *}

section {* Isar toplevel \label{sec:isar-toplevel} *}

text {* The Isar toplevel may be considered the centeral hub of the

  Isabelle/Isar system, where all key components and sub-systems are

  integrated into a single read-eval-print loop of Isar commands.  We

  shall even incorporate the existing {\ML} toplevel of the compiler

  and run-time system (cf.\ \secref{sec:ML-toplevel}).

  Isabelle/Isar departs from the original ``LCF system architecture''

  where {\ML} was really The Meta Language for defining theories and

  conducting proofs.  Instead, {\ML} now only serves as the

  implementation language for the system (and user extensions), while

  the specific Isar toplevel supports the concepts of theory and proof

  development natively.  This includes the graph structure of theories

  and the block structure of proofs, support for unlimited undo,

  facilities for tracing, debugging, timing, profiling etc.

  \medskip The toplevel maintains an implicit state, which is

  transformed by a sequence of transitions -- either interactively or

  in batch-mode.  In interactive mode, Isar state transitions are

  encapsulated as safe transactions, such that both failure and undo

  are handled conveniently without destroying the underlying draft

  theory (cf.~\secref{sec:context-theory}).  In batch mode,

  transitions operate in a linear (destructive) fashion, such that

  error conditions abort the present attempt to construct a theory or

  proof altogether.

  The toplevel state is a disjoint sum of empty @{text toplevel}, or

  @{text theory}, or @{text proof}.  On entering the main Isar loop we

  start with an empty toplevel.  A theory is commenced by giving a

  @{text \<THEORY>} header; within a theory we may issue theory

  commands such as @{text \<DEFINITION>}, or state a @{text

  \<THEOREM>} to be proven.  Now we are within a proof state, with a

  rich collection of Isar proof commands for structured proof

  composition, or unstructured proof scripts.  When the proof is

  concluded we get back to the theory, which is then updated by

  storing the resulting fact.  Further theory declarations or theorem

  statements with proofs may follow, until we eventually conclude the

  theory development by issuing @{text \<END>}.  The resulting theory

  is then stored within the theory database and we are back to the

  empty toplevel.

  In addition to these proper state transformations, there are also

  some diagnostic commands for peeking at the toplevel state without

  modifying it (e.g.\ \isakeyword{thm}, \isakeyword{term},

  \isakeyword{print-cases}).

*}

text %mlref {*

  \begin{mldecls}

  @{index_ML_type Toplevel.state} \\

  @{index_ML Toplevel.UNDEF: "exn"} \\

  @{index_ML Toplevel.is_toplevel: "Toplevel.state -> bool"} \\

  @{index_ML Toplevel.theory_of: "Toplevel.state -> theory"} \\

  @{index_ML Toplevel.proof_of: "Toplevel.state -> Proof.state"} \\

  @{index_ML Toplevel.debug: "bool ref"} \\

  @{index_ML Toplevel.timing: "bool ref"} \\

  @{index_ML Toplevel.profiling: "int ref"} \\

  \end{mldecls}

  \begin{description}

  \item @{ML_type Toplevel.state} represents Isar toplevel states,

  which are normally manipulated through the concept of toplevel

  transitions only (\secref{sec:toplevel-transition}).  Also note that

  a raw toplevel state is subject to the same linearity restrictions

  as a theory context (cf.~\secref{sec:context-theory}).

  \item @{ML Toplevel.UNDEF} is raised for undefined toplevel

  operations.  Many operations work only partially for certain cases,

  since @{ML_type Toplevel.state} is a sum type.

  \item @{ML Toplevel.is_toplevel}~@{text "state"} checks for an empty

  toplevel state.

  \item @{ML Toplevel.theory_of}~@{text "state"} selects the theory of

  a theory or proof (!), otherwise raises @{ML Toplevel.UNDEF}.

  \item @{ML Toplevel.proof_of}~@{text "state"} selects the Isar proof

  state if available, otherwise raises @{ML Toplevel.UNDEF}.

  \item @{ML "set Toplevel.debug"} makes the toplevel print further

  details about internal error conditions, exceptions being raised

  etc.

  \item @{ML "set Toplevel.timing"} makes the toplevel print timing

  information for each Isar command being executed.

  \item @{ML Toplevel.profiling}~@{verbatim ":="}~@{text "n"} controls

  low-level profiling of the underlying {\ML} runtime system.  For

  Poly/ML, @{text "n = 1"} means time and @{text "n = 2"} space

  profiling.

  \end{description}

*}

subsection {* Toplevel transitions \label{sec:toplevel-transition} *}

text {*

  An Isar toplevel transition consists of a partial function on the

  toplevel state, with additional information for diagnostics and

  error reporting: there are fields for command name, source position,

  optional source text, as well as flags for interactive-only commands

  (which issue a warning in batch-mode), printing of result state,

  etc.

  The operational part is represented as the sequential union of a

  list of partial functions, which are tried in turn until the first

  one succeeds.  This acts like an outer case-expression for various

  alternative state transitions.  For example, \isakeyword{qed} acts

  differently for a local proofs vs.\ the global ending of the main

  proof.

  Toplevel transitions are composed via transition transformers.

  Internally, Isar commands are put together from an empty transition

  extended by name and source position (and optional source text).  It

  is then left to the individual command parser to turn the given

  concrete syntax into a suitable transition transformer that adjoin

  actual operations on a theory or proof state etc.

*}

text %mlref {*

  \begin{mldecls}

  @{index_ML Toplevel.print: "Toplevel.transition -> Toplevel.transition"} \\

  @{index_ML Toplevel.no_timing: "Toplevel.transition -> Toplevel.transition"} \\

  @{index_ML Toplevel.keep: "(Toplevel.state -> unit) ->

  Toplevel.transition -> Toplevel.transition"} \\

  @{index_ML Toplevel.theory: "(theory -> theory) ->

  Toplevel.transition -> Toplevel.transition"} \\

  @{index_ML Toplevel.theory_to_proof: "(theory -> Proof.state) ->

  Toplevel.transition -> Toplevel.transition"} \\

  @{index_ML Toplevel.proof: "(Proof.state -> Proof.state) ->

  Toplevel.transition -> Toplevel.transition"} \\

  @{index_ML Toplevel.proofs: "(Proof.state -> Proof.state Seq.seq) ->

  Toplevel.transition -> Toplevel.transition"} \\

  @{index_ML Toplevel.end_proof: "(bool -> Proof.state -> Proof.context) ->

  Toplevel.transition -> Toplevel.transition"} \\

  \end{mldecls}

  \begin{description}

  \item @{ML Toplevel.print}~@{text "tr"} sets the print flag, which

  causes the toplevel loop to echo the result state (in interactive

  mode).

  \item @{ML Toplevel.no_timing}~@{text "tr"} indicates that the

  transition should never show timing information, e.g.\ because it is

  a diagnostic command.

  \item @{ML Toplevel.keep}~@{text "tr"} adjoins a diagnostic

  function.

  \item @{ML Toplevel.theory}~@{text "tr"} adjoins a theory

  transformer.

  \item @{ML Toplevel.theory_to_proof}~@{text "tr"} adjoins a global

  goal function, which turns a theory into a proof state.  The theory

  may be changed before entering the proof; the generic Isar goal

  setup includes an argument that specifies how to apply the proven

  result to the theory, when the proof is finished.

  \item @{ML Toplevel.proof}~@{text "tr"} adjoins a deterministic

  proof command, with a singleton result.

  \item @{ML Toplevel.proofs}~@{text "tr"} adjoins a general proof

  command, with zero or more result states (represented as a lazy

  list).

  \item @{ML Toplevel.end_proof}~@{text "tr"} adjoins a concluding

  proof command, that returns the resulting theory, after storing the

  resulting facts in the context etc.

  \end{description}

*}

subsection {* Toplevel control *}

text {*

  There are a few special control commands that modify the behavior

  the toplevel itself, and only make sense in interactive mode.  Under

  normal circumstances, the user encounters these only implicitly as

  part of the protocol between the Isabelle/Isar system and a

  user-interface such as ProofGeneral.

  \begin{description}

  \item \isacommand{undo} follows the three-level hierarchy of empty

  toplevel vs.\ theory vs.\ proof: undo within a proof reverts to the

  previous proof context, undo after a proof reverts to the theory

  before the initial goal statement, undo of a theory command reverts

  to the previous theory value, undo of a theory header discontinues

  the current theory development and removes it from the theory

  database (\secref{sec:theory-database}).

  \item \isacommand{kill} aborts the current level of development:

  kill in a proof context reverts to the theory before the initial

  goal statement, kill in a theory context aborts the current theory

  development, removing it from the database.

  \item \isacommand{exit} drops out of the Isar toplevel into the

  underlying {\ML} toplevel (\secref{sec:ML-toplevel}).  The Isar

  toplevel state is preserved and may be continued later.

  \item \isacommand{quit} terminates the Isabelle/Isar process without

  saving.

  \end{description}

*}

section {* ML toplevel \label{sec:ML-toplevel} *}

text {*

  The {\ML} toplevel provides a read-compile-eval-print loop for {\ML}

  values, types, structures, and functors.  {\ML} declarations operate

  on the global system state, which consists of the compiler

  environment plus the values of {\ML} reference variables.  There is

  no clean way to undo {\ML} declarations, except for reverting to a

  previously saved state of the whole Isabelle process.  {\ML} input

  is either read interactively from a TTY, or from a string (usually

  within a theory text), or from a source file (usually loaded from a

  theory).

  Whenever the {\ML} toplevel is active, the current Isabelle theory

  context is passed as an internal reference variable.  Thus {\ML}

  code may access the theory context during compilation, it may even

  change the value of a theory being under construction --- while

  observing the usual linearity restrictions

  (cf.~\secref{sec:context-theory}).

*}

text %mlref {*

  \begin{mldecls}

  @{index_ML context: "theory -> unit"} \\

  @{index_ML the_context: "unit -> theory"} \\

  @{index_ML "Context.>> ": "(theory -> theory) -> unit"} \\

  \end{mldecls}

  \begin{description}

  \item @{ML context}~@{text thy} sets the {\ML} theory context to

  @{text thy}.  This is usually performed automatically by the system,

  when dropping out of the interactive Isar toplevel into {\ML}, or

  when Isar invokes {\ML} to process code from a string or a file.

  \item @{ML "the_context ()"} refers to the theory context of the

  {\ML} toplevel --- at compile time!  {\ML} code needs to take care

  to refer to @{ML "the_context ()"} correctly.  Recall that

  evaluation of a function body is delayed until actual runtime.

  Moreover, persistent {\ML} toplevel bindings to an unfinished theory

  should be avoided: code should either project out the desired

  information immediately, or produce an explicit @{ML_type

  theory_ref} (cf.\ \secref{sec:context-theory}).

  \item @{ML "Context.>>"}~@{text f} applies theory transformation

  @{text f} to the current theory of the {\ML} toplevel.  In order to

  work as expected, the theory should be still under construction, and

  the Isar language element that invoked the {\ML} compiler in the

  first place should be ready to accept the changed theory value

  (e.g.\ \isakeyword{ML-setup}, but not plain \isakeyword{ML}).

  Otherwise the theory becomes stale!

  \end{description}

  It is very important to note that the above functions are really

  restricted to the compile time, even though the {\ML} compiler is

  invoked at runtime!  The majority of {\ML} code uses explicit

  functional arguments of a theory or proof context instead.  Thus it

  may be invoked for an arbitrary context later on, without having to

  worry about any operational details.

  \bigskip

  \begin{mldecls}

  @{index_ML Isar.main: "unit -> unit"} \\

  @{index_ML Isar.loop: "unit -> unit"} \\

  @{index_ML Isar.state: "unit -> Toplevel.state"} \\

  @{index_ML Isar.context: "unit -> Proof.context"} \\

  @{index_ML Isar.exn: "unit -> (exn * string) option"} \\

  \end{mldecls}

  \begin{description}

  \item @{ML "Isar.main ()"} invokes the Isar toplevel from {\ML},

  initializing an empty toplevel state.

  \item @{ML "Isar.loop ()"} continues the Isar toplevel with the

  current state, after having dropped out of the Isar toplevel loop.

  \item @{ML "Isar.state ()"} and @{ML "Isar.exn ()"} get current

  toplevel state and error condition, respectively.  This only works

  after having dropped out of the Isar toplevel loop.

  \item @{ML "Isar.context ()"} produces the proof context from @{ML

  "Isar.state ()"}, analogous to @{ML Context.proof_of}

  (\secref{sec:generic-context}).

  \end{description}

*}

section {* Theory database \label{sec:theory-database} *}

text {*

  The theory database maintains a collection of theories, together

  with some administrative information about their original sources,

  which are held in an external store (i.e.\ some directory within the

  regular file system).

  The theory database is organized as a directed acyclic graph;

  entries are referenced by theory name.  Although some additional

  interfaces allow to include a directory specification as well, this

  is only a hint to the underlying theory loader.  The internal theory

  name space is flat!

  Theory @{text A} is associated with the main theory file @{text

  A}\verb,.thy,, which needs to be accessible through the theory

  loader path.  Any number of additional {\ML} source files may be

  associated with each theory, by declaring these dependencies in the

  theory header as @{text \<USES>}, and loading them consecutively

  within the theory context.  The system keeps track of incoming {\ML}

  sources and associates them with the current theory.  The file

  @{text A}\verb,.ML, is loaded after a theory has been concluded, in

  order to support legacy proof {\ML} proof scripts.

  The basic internal actions of the theory database are @{text

  "update"}, @{text "outdate"}, and @{text "remove"}:

  \begin{itemize}

  \item @{text "update A"} introduces a link of @{text "A"} with a

  @{text "theory"} value of the same name; it asserts that the theory

  sources are now consistent with that value;

  \item @{text "outdate A"} invalidates the link of a theory database

  entry to its sources, but retains the present theory value;

  \item @{text "remove A"} deletes entry @{text "A"} from the theory

  database.

  \end{itemize}

  These actions are propagated to sub- or super-graphs of a theory

  entry as expected, in order to preserve global consistency of the

  state of all loaded theories with the sources of the external store.

  This implies certain causalities between actions: @{text "update"}

  or @{text "outdate"} of an entry will @{text "outdate"} all

  descendants; @{text "remove"} will @{text "remove"} all descendants.

  \medskip There are separate user-level interfaces to operate on the

  theory database directly or indirectly.  The primitive actions then

  just happen automatically while working with the system.  In

  particular, processing a theory header @{text "\<THEORY> A

  \<IMPORTS> B\<^sub>1 \<dots> B\<^sub>n \<BEGIN>"} ensures that the

  sub-graph of the collective imports @{text "B\<^sub>1 \<dots> B\<^sub>n"}

  is up-to-date, too.  Earlier theories are reloaded as required, with

  @{text update} actions proceeding in topological order according to

  theory dependencies.  There may be also a wave of implied @{text

  outdate} actions for derived theory nodes until a stable situation

  is achieved eventually.

*}

text %mlref {*

  \begin{mldecls}

  @{index_ML theory: "string -> theory"} \\

  @{index_ML use_thy: "string -> unit"} \\

  @{index_ML update_thy: "string -> unit"} \\

  @{index_ML touch_thy: "string -> unit"} \\

  @{index_ML remove_thy: "string -> unit"} \\[1ex]

  @{index_ML ThyInfo.begin_theory}@{verbatim ": ... -> bool -> theory"} \\

  @{index_ML ThyInfo.end_theory: "theory -> theory"} \\

  @{index_ML ThyInfo.register_theory: "theory -> unit"} \\[1ex]

  @{verbatim "datatype action = Update | Outdate | Remove"} \\

  @{index_ML ThyInfo.add_hook: "(ThyInfo.action -> string -> unit) -> unit"} \\

  \end{mldecls}

  \begin{description}

  \item @{ML theory}~@{text A} retrieves the theory value presently

  associated with name @{text A}.  Note that the result might be

  outdated.

  \item @{ML use_thy}~@{text A} loads theory @{text A} if it is absent

  or out-of-date.  It ensures that all parent theories are available

  as well, but does not reload them if older versions are already

  present.

  \item @{ML update_thy} is similar to @{ML use_thy}, but ensures that

  theory @{text A} and all ancestors are fully up-to-date.

  \item @{ML touch_thy}~@{text A} performs and @{text outdate} action

  on theory @{text A} and all descendants.

  \item @{ML remove_thy}~@{text A} deletes theory @{text A} and all

  descendants from the theory database.

  \item @{ML ThyInfo.begin_theory} is the basic operation behind a

  @{text \<THEORY>} header declaration.  The boolean argument

  indicates the strictness of treating ancestors: for @{ML true} (as

  in interactive mode) like @{ML update_thy}, and for @{ML false} (as

  in batch mode) like @{ML use_thy}.  This is {\ML} functions is

  normally not invoked directly.

  \item @{ML ThyInfo.end_theory} concludes the loading of a theory

  proper.  An attached theory {\ML} file may be still loaded later on.

  This is function is normally not invoked directly.

  \item @{ML ThyInfo.register_theory}~@{text "text thy"} registers an

  existing theory value with the theory loader database.  There is no

  management of associated sources.

  \item @{ML "ThyInfo.add_hook"}~@{text f} registers function @{text

  f} as a hook for theory database actions.  The function will be

  invoked with the action and theory name being involved; thus derived

  actions may be performed in associated system components, e.g.\

  maintaining the state of an editor for the theory sources.

  The kind and order of actions occurring in practice depends both on

  user interactions and the internal process of resolving theory

  imports.  Hooks should not rely on a particular policy here!  Any

  exceptions raised by the hook are ignored.

  \end{description}

*}

end