theory Advanced

imports Base First_Steps

begin

(*<*)

setup{*

open_file_with_prelude 

  "Advanced_Code.thy"

  ["theory Advanced", "imports Base First_Steps", "begin"]

*}

(*>*)

chapter {* Advanced Isabelle\label{chp:advanced} *}

text {*

   \begin{flushright}

  {\em All things are difficult before they are easy.} \\[1ex]

  proverb

  \end{flushright}

  \medskip

  While terms, types and theorems are the most basic data structures in

  Isabelle, there are a number of layers built on top of them. Most of these

  layers are concerned with storing and manipulating data. Handling them

  properly is an essential skill for programming on the ML-level of Isabelle. 

  The most basic layer are theories. They contain global data and

  can be seen as the ``long-term memory'' of Isabelle. There is usually only

  one theory active at each moment. Proof contexts and local theories, on the

  other hand, store local data for a task at hand. They act like the

  ``short-term memory'' and there can be many of them that are active in

  parallel.

*}

section {* Theories and Setups\label{sec:theories} *}

text {*

  Theories, as said above, are the most basic layer of abstraction in

  Isabelle. They record information about definitions, syntax declarations, axioms,

  theorems and much more.  For example, if a definition is made, it

  must be stored in a theory in order to be usable later on. Similar

  with proofs: once a proof is finished, the proved theorem needs to

  be stored in the theorem database of the theory in order to be

  usable. All relevant data of a theory can be queried with the

  Isabelle command \isacommand{print\_theory}.

  \begin{isabelle}

  \isacommand{print\_theory}\\

  @{text "> names: Pure Code_Generator HOL \<dots>"}\\

  @{text "> classes: Inf < type \<dots>"}\\

  @{text "> default sort: type"}\\

  @{text "> syntactic types: #prop \<dots>"}\\

  @{text "> logical types: 'a \<times> 'b \<dots>"}\\

  @{text "> type arities: * :: (random, random) random \<dots>"}\\

  @{text "> logical constants: == :: 'a \<Rightarrow> 'a \<Rightarrow> prop \<dots>"}\\

  @{text "> abbreviations: \<dots>"}\\

  @{text "> axioms: \<dots>"}\\

  @{text "> oracles: \<dots>"}\\

  @{text "> definitions: \<dots>"}\\

  @{text "> theorems: \<dots>"}

  \end{isabelle}

  Functions acting on theories often end with the suffix @{text "_global"},

  for example the function @{ML read_term_global in Syntax} in the structure

  @{ML_struct Syntax}. The reason is to set them syntactically apart from

  functions acting on contexts or local theories, which will be discussed in

  the next sections. There is a tendency amongst Isabelle developers to prefer

  ``non-global'' operations, because they have some advantages, as we will also

  discuss later. However, some basic understanding of theories is still necessary

  for effective Isabelle programming.

  An important Isabelle command with theories is \isacommand{setup}. In the

  previous chapters we used it already to make a theorem attribute known

  to Isabelle and to register a theorem under a name. What happens behind the

  scenes is that \isacommand{setup} expects a function of type @{ML_type

  "theory -> theory"}: the input theory is the current theory and the output

  the theory where the attribute has been registered or the theorem has been

  stored.  This is a fundamental principle in Isabelle. A similar situation

  arises with declaring a constant, which can be done on the ML-level with 

  function @{ML_ind declare_const in Sign} from the structure @{ML_struct

  Sign}. To see how \isacommand{setup} works, consider the following code:

*}  

ML{*let

  val thy = @{theory}

  val bar_const = ((@{binding "BAR"}, @{typ "nat"}), NoSyn)

in 

  Sign.declare_const @{context} bar_const thy  

end*}

text {*

  If you simply run this code\footnote{Recall that ML-code needs to be enclosed in

  \isacommand{ML}~@{text "\<verbopen> \<dots> \<verbclose>"}.} with the

  intention of declaring a constant @{text "BAR"} having type @{typ nat}, then 

  indeed you obtain a theory as result. But if you query the

  constant on the Isabelle level using the command \isacommand{term}

  \begin{isabelle}

  \isacommand{term}~@{text BAR}\\

  @{text "> \"BAR\" :: \"'a\""}

  \end{isabelle}

  you can see that you do \emph{not} obtain the expected constant of type @{typ nat}, but a free 

  variable (printed in blue) of polymorphic type. The problem is that the 

  ML-expression above did not ``register'' the declaration with the current theory. 

  This is what the command \isacommand{setup} is for. The constant is properly 

  declared with

*}

setup %gray {* fn thy => 

let

  val bar_const = ((@{binding "BAR"}, @{typ "nat"}), NoSyn)

  val (_, thy') = Sign.declare_const @{context} bar_const thy

in 

  thy'

end *}

text {* 

  where the declaration is actually applied to the current theory and

  \begin{isabelle}

  \isacommand{term}~@{text [quotes] "BAR"}\\

  @{text "> \"BAR\" :: \"nat\""}

  \end{isabelle}

  now returns a (black) constant with the type @{typ nat}, as expected.

  In a sense, \isacommand{setup} can be seen as a transaction that

  takes the current theory @{text thy}, applies an operation, and

  produces a new current theory @{text thy'}. This means that we have

  to be careful to apply operations always to the most current theory,

  not to a \emph{stale} one. Consider again the function inside the

  \isacommand{setup}-command:

  \begin{isabelle}

  \begin{graybox}

  \isacommand{setup}~@{text "\<verbopen>"} @{ML

"fn thy => 

let

  val bar_const = ((@{binding \"BAR\"}, @{typ \"nat\"}), NoSyn)

  val (_, thy') = Sign.declare_const @{context} bar_const thy

in

  thy

end"}~@{text "\<verbclose>"}\isanewline

  @{text "> ERROR: \"Stale theory encountered\""}

  \end{graybox}

  \end{isabelle}

  This time we erroneously return the original theory @{text thy}, instead of

  the modified one @{text thy'}. Such buggy code will always result into 

  a runtime error message about stale theories.

  However, sometimes it does make sense to work with two theories at the same

  time, especially in the context of parsing and typing. In the code below we

  use in Line 3 the function @{ML_ind copy in Theory} from the structure

  @{ML_struct Theory} for obtaining a new theory that contains the same

  data, but is unrelated to the existing theory.

*}

setup %graylinenos {* fn thy => 

let

  val tmp_thy = Theory.copy thy

  val foo_const = ((@{binding "FOO"}, @{typ "nat \<Rightarrow> nat"}), NoSyn)

  val (_, tmp_thy') = Sign.declare_const @{context} foo_const tmp_thy

  val trm1 = Syntax.read_term_global tmp_thy' "FOO baz"

  val trm2 = Syntax.read_term_global thy "FOO baz"

  val _ = writeln (@{make_string} trm1 ^ "\n" ^ @{make_string} trm2)

in

  thy

end *}

text {*

  That means we can make changes to the theory @{text tmp_thy} without

  affecting the current theory @{text thy}. In this case we declare in @{text

  "tmp_thy"} the constant @{text FOO} (Lines 4 and 5). The point of this code

  is that we next, in Lines 6 and 7, parse a string to become a term (both

  times the string is @{text [quotes] "FOO baz"}). But since we parse the string

  once in the context of the theory @{text tmp_thy'} in which @{text FOO} is

  declared to be a constant of type @{typ "nat \<Rightarrow>nat"} and once in the context 

  of @{text thy} where it is not, we obtain two different terms, namely 

  \begin{isabelle}

  \begin{graybox}

  @{text "> Const (\"Advanced.FOO\", \"nat \<Rightarrow> nat\") $ Free (\"baz\", \"nat\")"}\isanewline

  @{text "> Free (\"FOO\", \"'a \<Rightarrow> 'b\") $ Free (\"baz\", \"'a\")"}

  \end{graybox}

  \end{isabelle}

  There are two reasons for parsing a term in a temporary theory. One is to

  obtain fully qualified names for constants and the other is appropriate type 

  inference. This is relevant in situations where definitions are made later, 

  but parsing and type inference has to take already proceed as if the definitions 

  were already made.

*}

section {* Contexts *}

text {*

  Contexts are arguably more important than theories, even though they only 

  contain ``short-term memory data''. The reason is that a vast number of

  functions in Isabelle depend in one way or another on contexts. Even such

  mundane operations like printing out a term make essential use of contexts.

  For this consider the following contrived proof-snippet whose purpose is to 

  fix two variables:

*}

lemma "True"

proof -

  txt_raw {*\mbox{}\\[-7mm]*} 

  ML_prf {* Variable.dest_fixes @{context} *} 

  txt_raw {*\mbox{}\\[-7mm]\mbox{}*}

 fix x y  

  txt_raw {*\mbox{}\\[-7mm]*}

  ML_prf {* Variable.dest_fixes @{context} *} 

  txt_raw {*\mbox{}\\[-7mm] \ldots*}(*<*)oops(*>*)

text {*

  The interesting point is that we injected ML-code before and after

  the variables are fixed. For this remember that ML-code inside a proof

  needs to be enclosed inside \isacommand{ML\_prf}~@{text "\<verbopen> \<dots> \<verbclose>"},

  not \isacommand{ML}~@{text "\<verbopen> \<dots> \<verbclose>"}. The function 

  @{ML_ind dest_fixes in Variable} from the structure @{ML_struct Variable} takes 

  a context and returns all its currently fixed variable (names). That 

  means a context has a dataslot containing information about fixed variables.

  The ML-antiquotation @{text "@{context}"} points to the context that is

  active at that point of the theory. Consequently, in the first call to 

  @{ML dest_fixes in Variable} this dataslot is  empty; in the second it is 

  filled with @{text x} and @{text y}. What is interesting is that contexts

  can be stacked. For this consider the following proof fragment:

*}

lemma "True"

proof -

  fix x y

  { fix z w

  txt_raw {*\mbox{}\\[-7mm]*}

  ML_prf {* Variable.dest_fixes @{context} *} 

  txt_raw {*\mbox{}\\[-7mm]\mbox{}*}

 }

  txt_raw {*\mbox{}\\[-7mm]*}

  ML_prf {* Variable.dest_fixes @{context} *} 

  txt_raw {*\mbox{}\\[-7mm] \ldots*}(*<*)oops(*>*)

text {*

  Here the first time we call @{ML dest_fixes in Variable} we have four fixes variables;

  the second time we get only the fixes variables @{text x} and @{text y} as answer, since

  @{text z} and @{text w} are not anymore in the scope of the context. 

  This means the curly-braces act on the Isabelle level as opening and closing statements 

  for a context. The above proof fragment corresponds roughly to the following ML-code

*}

ML{*val ctxt0 = @{context};

val ([x, y], ctxt1) = Variable.add_fixes ["x", "y"] ctxt0;

val ([z, w], ctxt2) = Variable.add_fixes ["z", "w"] ctxt1*}

text {*

  where the function @{ML_ind add_fixes in Variable} fixes a list of variables

  specified by strings. Let us come back to the point about printing terms. Consider

  printing out the term \mbox{@{text "(x, y, z, w)"}} using our function @{ML_ind pretty_term}.

  This function takes a term and a context as argument. Notice how the printing

  of the term changes according to which context is used.

  \begin{isabelle}

  \begin{graybox}

  @{ML "let

  val trm = @{term \"(x, y, z, w)\"}

in

  pwriteln (Pretty.chunks 

    [ pretty_term ctxt0 trm,

      pretty_term ctxt1 trm,

      pretty_term ctxt2 trm ])

end"}\\

  \setlength{\fboxsep}{0mm}

  @{text ">"}~@{text "("}\colorbox{gray!5}{\raisebox{0mm}[3mm][1mm]{@{text x}}}@{text ","}~%

  \colorbox{gray!5}{\raisebox{0mm}[3mm][1mm]{@{text y}}}@{text ","}~%

  \colorbox{gray!5}{\raisebox{0mm}[3mm][1mm]{@{text z}}}@{text ","}~%

  \colorbox{gray!5}{\raisebox{0mm}[3mm][1mm]{@{text w}}}@{text ")"}\\

  @{text ">"}~@{text "("}\colorbox{gray!20}{\raisebox{0mm}[3mm][1mm]{@{text x}}}@{text ","}~%

  \colorbox{gray!20}{\raisebox{0mm}[3mm][1mm]{@{text y}}}@{text ","}~%

  \colorbox{gray!5}{\raisebox{0mm}[3mm][1mm]{@{text z}}}@{text ","}~%

  \colorbox{gray!5}{\raisebox{0mm}[3mm][1mm]{@{text w}}}@{text ")"}\\

  @{text ">"}~@{text "("}\colorbox{gray!20}{\raisebox{0mm}[3mm][1mm]{@{text x}}}@{text ","}~%

  \colorbox{gray!20}{\raisebox{0mm}[3mm][1mm]{@{text y}}}@{text ","}~%

  \colorbox{gray!20}{\raisebox{0mm}[3mm][1mm]{@{text z}}}@{text ","}~%

  \colorbox{gray!20}{\raisebox{0mm}[3mm][1mm]{@{text w}}}@{text ")"}

  \end{graybox}

  \end{isabelle}

  The term we are printing is in all three cases the same---a tuple containing

  four free variables. As you can see, however, in case of @{ML "ctxt0"} all

  variables are highlighted indicating a problem, while in case of @{ML

  "ctxt1"} @{text x} and @{text y} are printed as normal (blue) free

  variables, but not @{text z} and @{text w}. In the last case all variables

  are printed as expected. The point of this example is that the context

  contains the information which variables are fixed, and designates all other

  free variables as being alien or faulty.  Therefore the highlighting. 

  While this seems like a minor detail, the concept of making the context aware 

  of fixed variables is actually quite useful. For example it prevents us from 

  fixing a variable twice.

  @{ML_response_fake [gray, display]

  "@{context}

|> Variable.add_fixes [\"x\", \"x\"]" 

  "ERROR: Duplicate fixed variable(s): \"x\""}

  More importantly it also allows us to easily create fresh free variables avoiding any 

  clashes with fixed variables. In Line~3 below we fix the variable @{text x} in the context

  @{text ctxt1}. Next we want to create two fresh variables of type @{typ nat}

  as variants of the string @{text [quotes] "x"} (Lines 6 and 7).

  @{ML_response_fake [display, gray, linenos]

  "let

  val ctxt0 = @{context}

  val (_, ctxt1) = Variable.add_fixes [\"x\"] ctxt0

  val frees = replicate 2 (\"x\", @{typ nat})

in

  (Variable.variant_frees ctxt0 [] frees,

   Variable.variant_frees ctxt1 [] frees)

end"

  "([(\"x\", \"nat\"), (\"xa\", \"nat\")], 

 [(\"xa\", \"nat\"), (\"xb\", \"nat\")])"}

  As you can see, if we create the fresh variables with the context @{text ctxt0},

  then we obtain @{text "x"} and @{text "xa"}; but in @{text ctxt1} we obtain @{text "xa"}

  and @{text "xb"} avoiding @{text x}, which was fixed in this context.

  Often one has the problem that given some terms, create fresh variables

  avoiding any variable occurring in those terms. For this you can use the

  function @{ML_ind declare_term in Variable} from the structure @{ML_struct Variable}.

  @{ML_response_fake [gray, display]

  "let

  val ctxt1 = Variable.declare_term @{term \"(x, xa)\"} @{context}

  val frees = replicate 2 (\"x\", @{typ nat})

in

  Variable.variant_frees ctxt1 [] frees

end"

  "[(\"xb\", \"nat\"), (\"xc\", \"nat\")]"}

  The result is @{text xb} and @{text xc} for the names of the fresh

  variables, since @{text x} and @{text xa} occur in the term we declared. 

  Note that @{ML_ind declare_term in Variable} does not fix the

  variables; it just makes them ``known'' to the context. You can see

  that if you print out a declared term.

  \begin{isabelle}

  \begin{graybox}

  @{ML "let

  val trm = @{term \"P x y z\"}

  val ctxt1 = Variable.declare_term trm @{context}

in

  pwriteln (pretty_term ctxt1 trm)

end"}\\

  \setlength{\fboxsep}{0mm}

  @{text ">"}~\colorbox{gray!5}{\raisebox{0mm}[3mm][1mm]{@{text P}}}~%

  \colorbox{gray!5}{\raisebox{0mm}[3mm][1mm]{@{text x}}}~%

  \colorbox{gray!5}{\raisebox{0mm}[3mm][1mm]{@{text y}}}~%

  \colorbox{gray!5}{\raisebox{0mm}[3mm][1mm]{@{text z}}}

  \end{graybox}

  \end{isabelle}

  All variables are highligted, indicating that they are not

  fixed. However, declaring a term is helpful when parsing terms using

  the function @{ML_ind read_term in Syntax} from the structure

  @{ML_struct Syntax}. Consider the following code:

  @{ML_response_fake [gray, display]

  "let

  val ctxt0 = @{context}

  val ctxt1 = Variable.declare_term @{term \"x::nat\"} ctxt0

in

  (Syntax.read_term ctxt0 \"x\", 

   Syntax.read_term ctxt1 \"x\") 

end"

  "(Free (\"x\", \"'a\"), Free (\"x\", \"nat\"))"}

  Parsing the string in the context @{text "ctxt0"} results in a free variable 

  with a default polymorphic type, but in case of @{text "ctxt1"} we obtain a

  free variable of type @{typ nat} as previously declared. Which

  type the parsed term receives depends upon the last declaration that

  is made, as the next example illustrates.

  @{ML_response_fake [gray, display]

  "let

  val ctxt1 = Variable.declare_term @{term \"x::nat\"} @{context}

  val ctxt2 = Variable.declare_term @{term \"x::int\"} ctxt1

in

  (Syntax.read_term ctxt1 \"x\", 

   Syntax.read_term ctxt2 \"x\") 

end"

  "(Free (\"x\", \"nat\"), Free (\"x\", \"int\"))"}

  \begin{isabelle}

  \begin{graybox}

  \begin{linenos}

  @{ML "let

  val ctxt0 = @{context}

  val (_, ctxt1) = Variable.add_fixes [\"x\", \"y\", \"z\"] ctxt0 

  val foo_trm = @{term \"P x y z\"}

in

  singleton (Variable.export_terms ctxt1 ctxt0) foo_trm

  |> pretty_term ctxt0

  |> pwriteln

end"}

  \end{linenos}

  \setlength{\fboxsep}{0mm}

  @{text ">"}~\colorbox{gray!5}{\raisebox{0mm}[3mm][1mm]{@{text P}}}~%

  @{text "?x ?y ?z"}

  \end{graybox}

  \end{isabelle}

  In Line 3 we fix the variables @{term x}, @{term y} and @{term z} in

  context @{text ctxt1}.  The function @{ML_ind export_terms in

  terms between contexts. Transferring means to turn all (free)

  variables that are fixed in one context, but not in the other, into

  schematic variables. In our example, we are transferring the term

  @{text "P x y z"} from context @{text "ctxt1"} to @{text "ctxt0"},

  which means @{term x}, @{term y} and @{term z} become schematic

  variables (as can be seen by the leading question marks). Note that

  the variable @{text P} stays a free variable, since it not fixed in

  @{text ctxt1}; it is even highlighed, because @{text "ctxt0"} does

  not know about it. Note also that in Line 6 we had to use the

  function @{ML_ind singleton}, because the function @{ML_ind

  export_terms in Variable} normally works over lists of terms.

  The case of transferring theorems is even more useful. The reason is

  that the generalisation of fixed variables to schematic variables is

  @{ML_response_fake [display, gray]

  "let

  val thy = @{theory}

  val ctxt0 = @{context}

  val foo_thm = Skip_Proof.make_thm thy @{prop \"P x y z x y z\"}

in

  singleton (Proof_Context.export ctxt1 ctxt0) foo_thm

end"

*}

text {*

*}

(*

ML{*Proof_Context.debug := true*}

ML{*Proof_Context.verbose := true*}

*)

(*

lemma "True"

proof -

  { -- "\<And>x. _"

    fix x

    have "B x" sorry

    thm this

  }

  thm this

  { -- "A \<Longrightarrow> _"

    assume A

    have B sorry

    thm this

  }

  thm this

  { -- "\<And>x. x = _ \<Longrightarrow> _"