text‹
Code generation by shallow embedding (cf.~\secref{sec:principle})
allows to choose code equations anddatatype constructors freely,
given that some very basic syntactic properties are met; this
flexibility opens up mechanisms for refinement which allow to extend
the scope and quality of generated code dramatically. ›
subsection‹Program refinement›
text‹
Program refinement works by choosing appropriate code equations
explicitly (cf.~\secref{sec:equations}); as example, we use Fibonacci
numbers: ›
text‹ \noindent The runtime of the corresponding code grows exponential due to two recursive calls: ›
text %quote ‹
@{code_stmts fib constant: fib (Haskell)} ›
text‹ \noindent A more efficient implementation would use dynamic
programming, e.g.~sharing of common intermediate results between
recursive calls. This idea is expressed by an auxiliary operation
which computes a Fibonacci number and its successor simultaneously: ›
text‹ \noindent The resulting code shows only linear growth of runtime: ›
text %quote ‹
@{code_stmts fib constant: fib fib_step (Haskell)} ›
subsection‹Datatype refinement›
text‹
Selecting specific code equations \emph{and} datatype constructors
leads todatatype refinement. As an example, we will develop an
alternative representation of the queue example given in \secref{sec:queue_example}. The amortised representation is
convenient for generating code but exposes its \qt{implementation}
details, which may be cumbersome when proving theorems about it.
Therefore, here is a simple, straightforward representation of
queues: ›
primrec %quote enqueue :: "'a \ 'a queue \ 'a queue"where "enqueue x (Queue xs) = Queue (xs @ [x])"
fun %quote dequeue :: "'a queue \ 'a option \ 'a queue"where "dequeue (Queue []) = (None, Queue [])"
| "dequeue (Queue (x # xs)) = (Some x, Queue xs)"
text‹ \noindent This we can use directly for proving; for executing,
we provide an alternative characterisation: ›
definition %quote AQueue :: "'a list \ 'a list \ 'a queue"where "AQueue xs ys = Queue (ys @ rev xs)"
code_datatype %quote AQueue
text‹ \noindent Here we define a \qt{constructor} 🍋‹AQueue› which is defined in terms of ‹Queue›and interprets its arguments
according to what the \emph{content} of an amortised queue is supposed to be.
The prerequisite fordatatype constructors is only syntactical: a
constructor must be of type ‹τ = …==> κ α🚫1 … α🚫n›where‹{α🚫1, …, α🚫n}›is exactly the set of \emph{all} type variables in ‹τ›; then‹κ›is its corresponding datatype. The
HOL datatype package by default registers any new datatypewith its
constructors, but this may be changed using @{command_def code_datatype}; the currently chosen constructors can be inspected using the @{command print_codesetup} command.
Equipped with this, we are able to prove the following equations for our primitive queue operations which \qt{implement} the simple
queues in an amortised fashion: ›
text‹ \noindent It is good style, although no absolute requirement, to
provide code equations for the original artefacts of the implemented
type, if possible; in our case, these are the datatype constructor 🍋‹Queue›and the case combinator 🍋‹case_queue›: ›
text‹
The same techniques can also be applied totypes which are not
specified as datatypes, e.g.~type 🍋‹int›is originally specified
as quotient type by means of @{command_def typedef}, but for code
generation constants allowing construction of binary numeral values
are used as constructors for🍋‹int›.
This approach however fails if the representation of a type demands
invariants; this issue is discussed in the nextsection. ›
text‹ Datatype representation involving invariants require a dedicated setupfor the type and its primitive operations. As a running
example, we implement a type 🍋‹'a dlist\ of lists consisting
of distinct elements.
The specification of 🍋‹'a dlist\ itself can be found in theory 🍋‹HOL-Library.Dlist›.
The first step isto decide on which representation the abstract
type (in our example 🍋‹'a dlist\) should be implemented.
Here we choose 🍋‹'a list\. Then a conversion from the concrete
type to the abstract type must be specified, here: ›
text %quote ‹ 🚫‹Dlist› ›
text‹ \noindentNext follows the specification of a suitable \emph{projection},
i.e.~a conversion from abstract to concrete type: ›
text %quote ‹ 🚫‹list_of_dlist› ›
text‹ \noindent This projection must be specified such that the following \emph{abstract datatype certificate} can be proven: ›
text‹ \noindentNote that so far the invariant on representations
(🚫‹distinct›) has never been mentioned explicitly:
the invariant is only referred to implicitly: all values in
set 🍋‹{xs. list_of_dlist (Dlist xs) = xs}› are invariant, andin our example this is exactly 🍋‹{xs. distinct xs}›.
The primitive operations on 🍋‹'a dlist\ are specified
indirectly using the projection 🍋‹list_of_dlist›. For
the empty ‹dlist›, 🍋‹Dlist.empty›, we finally want
the code equation ›
text %quote ‹ 🍋‹Dlist.empty = Dlist []› ›
text‹ \noindent This we haveto prove indirectly as follows: ›
lemma %quote [code]: "list_of_dlist Dlist.empty = []" by (fact list_of_dlist_empty)
text‹ \noindent This equation logically encodes both the desired code
equation and that the expression 🍋‹Dlist›is applied to obeys
the implicit invariant. Equations for insertion and removal are
similar: ›
lemma %quote [code]: "list_of_dlist (Dlist.insert x dxs) = List.insert x (list_of_dlist dxs)" by (fact list_of_dlist_insert)
lemma %quote [code]: "list_of_dlist (Dlist.remove x dxs) = remove1 x (list_of_dlist dxs)" by (fact list_of_dlist_remove)
text‹ \noindentThen the corresponding code is as follows: ›
text %quote ‹
@{code_stmts Dlist.empty Dlist.insert Dlist.remove list_of_dlist (SML)} ›
text‹ To reduce manual work fordatatype refinement, @{command_def lift_definition} is a valuable tool. See the corresponding sectionin🍋‹"isabelle-isar-ref"›.
See further 🍋‹"Haftmann-Kraus-Kuncar-Nipkow:2013:data_refinement"› for the meta theory of datatype refinement involving invariants.
Typical data structures implemented by representations involving
invariants are available in the library, theory🍋‹HOL-Library.Mapping›
specifies key-value-mappings (type 🍋‹('a, 'b) mapping›);
these can be implemented by red-black-trees (theory🍋‹HOL-Library.RBT›). ›
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung und die Messung sind noch experimentell.