src/Doc/Codegen/Evaluation.thy
author haftmann
Fri May 09 08:13:37 2014 +0200 (2014-05-09)
changeset 56927 4044a7d1720f
parent 55418 9f25e0cca254
child 58100 f54a8a4134d3
permissions -rw-r--r--
hardcoded nbe and sml into value command
haftmann@38510
     1
theory Evaluation
haftmann@38510
     2
imports Setup
haftmann@38510
     3
begin
haftmann@38510
     4
haftmann@40751
     5
section {* Evaluation \label{sec:evaluation} *}
haftmann@38510
     6
haftmann@39599
     7
text {*
haftmann@39599
     8
  Recalling \secref{sec:principle}, code generation turns a system of
haftmann@39599
     9
  equations into a program with the \emph{same} equational semantics.
haftmann@39599
    10
  As a consequence, this program can be used as a \emph{rewrite
haftmann@39599
    11
  engine} for terms: rewriting a term @{term "t"} using a program to a
haftmann@39599
    12
  term @{term "t'"} yields the theorems @{prop "t \<equiv> t'"}.  This
haftmann@39599
    13
  application of code generation in the following is referred to as
haftmann@39599
    14
  \emph{evaluation}.
haftmann@39599
    15
*}
haftmann@38510
    16
haftmann@38510
    17
haftmann@38510
    18
subsection {* Evaluation techniques *}
haftmann@38510
    19
haftmann@39599
    20
text {*
haftmann@40350
    21
  The existing infrastructure provides a rich palette of evaluation
haftmann@39599
    22
  techniques, each comprising different aspects:
haftmann@39599
    23
haftmann@39599
    24
  \begin{description}
haftmann@39599
    25
haftmann@39599
    26
    \item[Expressiveness.]  Depending on how good symbolic computation
haftmann@39599
    27
      is supported, the class of terms which can be evaluated may be
haftmann@39599
    28
      bigger or smaller.
haftmann@38510
    29
haftmann@39599
    30
    \item[Efficiency.]  The more machine-near the technique, the
haftmann@39599
    31
      faster it is.
haftmann@38510
    32
haftmann@39599
    33
    \item[Trustability.]  Techniques which a huge (and also probably
haftmann@39599
    34
      more configurable infrastructure) are more fragile and less
haftmann@39599
    35
      trustable.
haftmann@39599
    36
haftmann@39599
    37
  \end{description}
haftmann@39599
    38
*}
haftmann@38510
    39
haftmann@38510
    40
haftmann@39599
    41
subsubsection {* The simplifier (@{text simp}) *}
haftmann@38510
    42
haftmann@39599
    43
text {*
haftmann@39599
    44
  The simplest way for evaluation is just using the simplifier with
haftmann@39599
    45
  the original code equations of the underlying program.  This gives
haftmann@39599
    46
  fully symbolic evaluation and highest trustablity, with the usual
haftmann@39599
    47
  performance of the simplifier.  Note that for operations on abstract
haftmann@39599
    48
  datatypes (cf.~\secref{sec:invariant}), the original theorems as
haftmann@39599
    49
  given by the users are used, not the modified ones.
haftmann@39599
    50
*}
haftmann@38510
    51
haftmann@38510
    52
haftmann@39599
    53
subsubsection {* Normalization by evaluation (@{text nbe}) *}
haftmann@38510
    54
haftmann@39599
    55
text {*
haftmann@39599
    56
  Normalization by evaluation \cite{Aehlig-Haftmann-Nipkow:2008:nbe}
haftmann@39599
    57
  provides a comparably fast partially symbolic evaluation which
haftmann@39599
    58
  permits also normalization of functions and uninterpreted symbols;
haftmann@39599
    59
  the stack of code to be trusted is considerable.
haftmann@39599
    60
*}
haftmann@38510
    61
haftmann@38510
    62
haftmann@39599
    63
subsubsection {* Evaluation in ML (@{text code}) *}
haftmann@39599
    64
haftmann@39599
    65
text {*
haftmann@39599
    66
  Highest performance can be achieved by evaluation in ML, at the cost
haftmann@39599
    67
  of being restricted to ground results and a layered stack of code to
haftmann@39599
    68
  be trusted, including code generator configurations by the user.
haftmann@38510
    69
haftmann@39599
    70
  Evaluation is carried out in a target language \emph{Eval} which
haftmann@39599
    71
  inherits from \emph{SML} but for convenience uses parts of the
haftmann@39599
    72
  Isabelle runtime environment.  The soundness of computation carried
haftmann@39609
    73
  out there depends crucially on the correctness of the code
haftmann@39643
    74
  generator setup; this is one of the reasons why you should not use
haftmann@39609
    75
  adaptation (see \secref{sec:adaptation}) frivolously.
haftmann@39599
    76
*}
haftmann@38510
    77
haftmann@38510
    78
haftmann@39599
    79
subsection {* Aspects of evaluation *}
haftmann@38510
    80
haftmann@38510
    81
text {*
haftmann@39599
    82
  Each of the techniques can be combined with different aspects.  The
haftmann@39599
    83
  most important distinction is between dynamic and static evaluation.
haftmann@39599
    84
  Dynamic evaluation takes the code generator configuration \qt{as it
haftmann@39599
    85
  is} at the point where evaluation is issued.  Best example is the
haftmann@39599
    86
  @{command_def value} command which allows ad-hoc evaluation of
haftmann@39599
    87
  terms:
haftmann@38510
    88
*}
haftmann@38510
    89
haftmann@38510
    90
value %quote "42 / (12 :: rat)"
haftmann@38510
    91
haftmann@38510
    92
text {*
haftmann@56927
    93
  \noindent @{command value} tries first to evaluate using ML, falling
haftmann@56927
    94
  back to normalization by evaluation if this fails.
haftmann@38510
    95
bulwahn@43656
    96
  To employ dynamic evaluation in the document generation, there is also
haftmann@56927
    97
  a @{text value} antiquotation with the same evaluation techniques 
haftmann@56927
    98
  as @{command value}.
bulwahn@43656
    99
haftmann@39599
   100
  Static evaluation freezes the code generator configuration at a
haftmann@39599
   101
  certain point and uses this context whenever evaluation is issued
haftmann@39599
   102
  later on.  This is particularly appropriate for proof procedures
haftmann@39599
   103
  which use evaluation, since then the behaviour of evaluation is not
haftmann@39599
   104
  changed or even compromised later on by actions of the user.
haftmann@39599
   105
haftmann@39599
   106
  As a technical complication, terms after evaluation in ML must be
haftmann@39599
   107
  turned into Isabelle's internal term representation again.  Since
haftmann@39599
   108
  this is also configurable, it is never fully trusted.  For this
haftmann@39599
   109
  reason, evaluation in ML comes with further aspects:
haftmann@39599
   110
haftmann@39599
   111
  \begin{description}
haftmann@39599
   112
haftmann@39599
   113
    \item[Plain evaluation.]  A term is normalized using the provided
haftmann@39599
   114
      term reconstruction from ML to Isabelle; for applications which
haftmann@39599
   115
      do not need to be fully trusted.
haftmann@39599
   116
haftmann@39599
   117
    \item[Property conversion.]  Evaluates propositions; since these
haftmann@39599
   118
      are monomorphic, the term reconstruction is fixed once and for all
haftmann@39599
   119
      and therefore trustable.
haftmann@39599
   120
haftmann@39599
   121
    \item[Conversion.]  Evaluates an arbitrary term @{term "t"} first
haftmann@39599
   122
      by plain evaluation and certifies the result @{term "t'"} by
haftmann@39599
   123
      checking the equation @{term "t \<equiv> t'"} using property
haftmann@39599
   124
      conversion.
haftmann@39599
   125
haftmann@39599
   126
  \end{description}
haftmann@39599
   127
haftmann@39599
   128
  \noindent The picture is further complicated by the roles of
haftmann@39599
   129
  exceptions.  Here three cases have to be distinguished:
haftmann@39599
   130
haftmann@39599
   131
  \begin{itemize}
haftmann@39599
   132
haftmann@39599
   133
    \item Evaluation of @{term t} terminates with a result @{term
haftmann@39599
   134
      "t'"}.
haftmann@39599
   135
haftmann@39599
   136
    \item Evaluation of @{term t} terminates which en exception
haftmann@40350
   137
      indicating a pattern match failure or a non-implemented
haftmann@39599
   138
      function.  As sketched in \secref{sec:partiality}, this can be
haftmann@39599
   139
      interpreted as partiality.
haftmann@39599
   140
     
haftmann@39643
   141
    \item Evaluation raises any other kind of exception.
haftmann@39599
   142
     
haftmann@39599
   143
  \end{itemize}
haftmann@39599
   144
haftmann@39599
   145
  \noindent For conversions, the first case yields the equation @{term
haftmann@39599
   146
  "t = t'"}, the second defaults to reflexivity @{term "t = t"}.
haftmann@39643
   147
  Exceptions of the third kind are propagated to the user.
haftmann@39599
   148
haftmann@39599
   149
  By default return values of plain evaluation are optional, yielding
haftmann@40350
   150
  @{text "SOME t'"} in the first case, @{text "NONE"} in the
haftmann@40350
   151
  second, and propagating the exception in the third case.  A strict
haftmann@39599
   152
  variant of plain evaluation either yields @{text "t'"} or propagates
haftmann@51713
   153
  any exception, a liberal variant captures any exception in a result
haftmann@39599
   154
  of type @{text "Exn.result"}.
haftmann@39599
   155
  
haftmann@39599
   156
  For property conversion (which coincides with conversion except for
haftmann@39599
   157
  evaluation in ML), methods are provided which solve a given goal by
haftmann@39599
   158
  evaluation.
haftmann@38510
   159
*}
haftmann@38510
   160
haftmann@38510
   161
haftmann@39599
   162
subsection {* Schematic overview *}
haftmann@39599
   163
haftmann@38510
   164
text {*
haftmann@39693
   165
  \newcommand{\ttsize}{\fontsize{5.8pt}{8pt}\selectfont}
haftmann@39693
   166
  \fontsize{9pt}{12pt}\selectfont
haftmann@39599
   167
  \begin{tabular}{ll||c|c|c}
haftmann@39599
   168
    & & @{text simp} & @{text nbe} & @{text code} \tabularnewline \hline \hline
haftmann@56927
   169
    \multirow{4}{1ex}{\rotatebox{90}{dynamic}}
haftmann@39693
   170
    & plain evaluation & & & \ttsize@{ML "Code_Evaluation.dynamic_value"} \tabularnewline \cline{2-5}
haftmann@39599
   171
    & evaluation method & @{method code_simp} & @{method normalization} & @{method eval} \tabularnewline
haftmann@39693
   172
    & property conversion & & & \ttsize@{ML "Code_Runtime.dynamic_holds_conv"} \tabularnewline \cline{2-5}
haftmann@41184
   173
    & conversion & \ttsize@{ML "Code_Simp.dynamic_conv"} & \ttsize@{ML "Nbe.dynamic_conv"}
haftmann@41184
   174
      & \ttsize@{ML "Code_Evaluation.dynamic_conv"} \tabularnewline \hline \hline
haftmann@39693
   175
    \multirow{3}{1ex}{\rotatebox{90}{static}}
haftmann@41184
   176
    & plain evaluation & & & \ttsize@{ML "Code_Evaluation.static_value"} \tabularnewline \cline{2-5}
haftmann@39599
   177
    & property conversion & &
haftmann@39693
   178
      & \ttsize@{ML "Code_Runtime.static_holds_conv"} \tabularnewline \cline{2-5}
haftmann@41184
   179
    & conversion & \ttsize@{ML "Code_Simp.static_conv"}
haftmann@41184
   180
      & \ttsize@{ML "Nbe.static_conv"}
haftmann@41184
   181
      & \ttsize@{ML "Code_Evaluation.static_conv"}
haftmann@39599
   182
  \end{tabular}
haftmann@39599
   183
*}
haftmann@39599
   184
haftmann@39599
   185
haftmann@52287
   186
subsection {* Preprocessing HOL terms into evaluable shape *}
haftmann@52287
   187
haftmann@52287
   188
text {*
haftmann@52287
   189
  When integration decision procedures developed inside HOL into HOL itself,
haftmann@52287
   190
  it is necessary to somehow get from the Isabelle/ML representation to
haftmann@52287
   191
  the representation used by the decision procedure itself (``reification'').
haftmann@52287
   192
  One option is to hardcode it using code antiquotations (see \secref{sec:code_antiq}).
haftmann@52287
   193
  Another option is to use pre-existing infrastructure in HOL:
haftmann@52287
   194
  @{ML "Reification.conv"} and @{ML "Reification.tac"}
haftmann@52287
   195
haftmann@52287
   196
  An simplistic example:
haftmann@52287
   197
*}
haftmann@52287
   198
haftmann@52287
   199
datatype %quote form_ord = T | F | Less nat nat
haftmann@52287
   200
  | And form_ord form_ord | Or form_ord form_ord | Neg form_ord
haftmann@52287
   201
haftmann@52287
   202
primrec %quote interp :: "form_ord \<Rightarrow> 'a::order list \<Rightarrow> bool"
haftmann@52287
   203
where
haftmann@52287
   204
  "interp T vs \<longleftrightarrow> True"
haftmann@52287
   205
| "interp F vs \<longleftrightarrow> False"
haftmann@52287
   206
| "interp (Less i j) vs \<longleftrightarrow> vs ! i < vs ! j"
haftmann@52287
   207
| "interp (And f1 f2) vs \<longleftrightarrow> interp f1 vs \<and> interp f2 vs"
haftmann@52287
   208
| "interp (Or f1 f2) vs \<longleftrightarrow> interp f1 vs \<or> interp f2 vs"
haftmann@52287
   209
| "interp (Neg f) vs \<longleftrightarrow> \<not> interp f vs"
haftmann@52287
   210
haftmann@52287
   211
text {*
haftmann@52287
   212
  The datatype @{type form_ord} represents formulae whose semantics is given by
haftmann@52287
   213
  @{const interp}.  Note that values are represented by variable indices (@{typ nat})
haftmann@52287
   214
  whose concrete values are given in list @{term vs}.
haftmann@52287
   215
*}
haftmann@52287
   216
haftmann@52287
   217
ML (*<*) {* *}
haftmann@52287
   218
schematic_lemma "thm": fixes x y z :: "'a::order" shows "x < y \<and> x < z \<equiv> ?P"
haftmann@52287
   219
ML_prf 
haftmann@52287
   220
(*>*) {* val thm =
haftmann@52287
   221
  Reification.conv @{context} @{thms interp.simps} @{cterm "x < y \<and> x < z"} *} (*<*)
haftmann@52287
   222
by (tactic {* ALLGOALS (rtac thm) *})
haftmann@52287
   223
(*>*) 
haftmann@52287
   224
haftmann@52287
   225
text {*
haftmann@52287
   226
  By virtue of @{fact interp.simps}, @{ML "Reification.conv"} provides a conversion
haftmann@52287
   227
  which, for this concrete example, yields @{thm thm [no_vars]}.  Note that the argument
haftmann@52287
   228
  to @{const interp} does not contain any free variables and can this be evaluated
haftmann@52287
   229
  using evaluation.
haftmann@52287
   230
haftmann@52287
   231
  A less meager example can be found in the AFP, session @{text "Regular-Sets"},
haftmann@52287
   232
  theory @{text Regexp_Method}.
haftmann@52287
   233
*}
haftmann@52287
   234
haftmann@52287
   235
haftmann@39599
   236
subsection {* Intimate connection between logic and system runtime *}
haftmann@39599
   237
haftmann@39609
   238
text {*
haftmann@39609
   239
  The toolbox of static evaluation conversions forms a reasonable base
haftmann@39609
   240
  to interweave generated code and system tools.  However in some
haftmann@39609
   241
  situations more direct interaction is desirable.
haftmann@39609
   242
*}
haftmann@39599
   243
haftmann@39599
   244
haftmann@52287
   245
subsubsection {* Static embedding of generated code into system runtime -- the @{text code} antiquotation \label{sec:code_antiq} *}
haftmann@39599
   246
haftmann@39599
   247
text {*
haftmann@39609
   248
  The @{text code} antiquotation allows to include constants from
haftmann@39609
   249
  generated code directly into ML system code, as in the following toy
haftmann@39609
   250
  example:
haftmann@38510
   251
*}
haftmann@38510
   252
haftmann@38510
   253
datatype %quote form = T | F | And form form | Or form form (*<*)
haftmann@38510
   254
haftmann@39745
   255
(*>*) ML %quotett {*
haftmann@38510
   256
  fun eval_form @{code T} = true
haftmann@38510
   257
    | eval_form @{code F} = false
haftmann@38510
   258
    | eval_form (@{code And} (p, q)) =
haftmann@38510
   259
        eval_form p andalso eval_form q
haftmann@38510
   260
    | eval_form (@{code Or} (p, q)) =
haftmann@38510
   261
        eval_form p orelse eval_form q;
haftmann@38510
   262
*}
haftmann@38510
   263
haftmann@38510
   264
text {*
haftmann@39609
   265
  \noindent @{text code} takes as argument the name of a constant;
haftmann@39609
   266
  after the whole ML is read, the necessary code is generated
haftmann@39609
   267
  transparently and the corresponding constant names are inserted.
haftmann@39609
   268
  This technique also allows to use pattern matching on constructors
haftmann@39643
   269
  stemming from compiled datatypes.  Note that the @{text code}
haftmann@39643
   270
  antiquotation may not refer to constants which carry adaptations;
haftmann@39643
   271
  here you have to refer to the corresponding adapted code directly.
haftmann@38510
   272
haftmann@39609
   273
  For a less simplistic example, theory @{text Approximation} in
haftmann@39609
   274
  the @{text Decision_Procs} session is a good reference.
haftmann@38510
   275
*}
haftmann@38510
   276
haftmann@38510
   277
haftmann@39599
   278
subsubsection {* Static embedding of generated code into system runtime -- @{text code_reflect} *}
haftmann@39599
   279
haftmann@39609
   280
text {*
haftmann@39609
   281
  The @{text code} antiquoation is lightweight, but the generated code
haftmann@39609
   282
  is only accessible while the ML section is processed.  Sometimes this
haftmann@39609
   283
  is not appropriate, especially if the generated code contains datatype
haftmann@39609
   284
  declarations which are shared with other parts of the system.  In these
haftmann@39609
   285
  cases, @{command_def code_reflect} can be used:
haftmann@39609
   286
*}
haftmann@39609
   287
haftmann@39609
   288
code_reflect %quote Sum_Type
haftmann@39609
   289
  datatypes sum = Inl | Inr
blanchet@55418
   290
  functions "Sum_Type.sum.projl" "Sum_Type.sum.projr"
haftmann@39609
   291
haftmann@39609
   292
text {*
haftmann@39609
   293
  \noindent @{command_def code_reflect} takes a structure name and
haftmann@39609
   294
  references to datatypes and functions; for these code is compiled
haftmann@39609
   295
  into the named ML structure and the \emph{Eval} target is modified
haftmann@39609
   296
  in a way that future code generation will reference these
haftmann@39609
   297
  precompiled versions of the given datatypes and functions.  This
haftmann@39609
   298
  also allows to refer to the referenced datatypes and functions from
haftmann@39609
   299
  arbitrary ML code as well.
haftmann@39609
   300
haftmann@39609
   301
  A typical example for @{command code_reflect} can be found in the
haftmann@39609
   302
  @{theory Predicate} theory.
haftmann@39609
   303
*}
haftmann@39609
   304
haftmann@39599
   305
haftmann@39599
   306
subsubsection {* Separate compilation -- @{text code_reflect} *}
haftmann@39599
   307
haftmann@39609
   308
text {*
haftmann@39609
   309
  For technical reasons it is sometimes necessary to separate
haftmann@39609
   310
  generation and compilation of code which is supposed to be used in
haftmann@39609
   311
  the system runtime.  For this @{command code_reflect} with an
haftmann@39609
   312
  optional @{text "file"} argument can be used:
haftmann@39609
   313
*}
haftmann@39609
   314
haftmann@39609
   315
code_reflect %quote Rat
haftmann@39609
   316
  datatypes rat = Frct
haftmann@39609
   317
  functions Fract
haftmann@39609
   318
    "(plus :: rat \<Rightarrow> rat \<Rightarrow> rat)" "(minus :: rat \<Rightarrow> rat \<Rightarrow> rat)"
haftmann@39609
   319
    "(times :: rat \<Rightarrow> rat \<Rightarrow> rat)" "(divide :: rat \<Rightarrow> rat \<Rightarrow> rat)"
haftmann@39609
   320
  file "examples/rat.ML"
haftmann@39609
   321
haftmann@39609
   322
text {*
haftmann@39643
   323
  \noindent This merely generates the referenced code to the given
haftmann@39609
   324
  file which can be included into the system runtime later on.
haftmann@39609
   325
*}
haftmann@39599
   326
haftmann@38510
   327
end
haftmann@46522
   328