diff --git a/docs/core-spec/core-spec.mng b/docs/core-spec/core-spec.mng
index e726602..4e134e0 100644
--- a/docs/core-spec/core-spec.mng
+++ b/docs/core-spec/core-spec.mng
@@ -1,464 +1,464 @@
\documentclass{article}
\usepackage{supertabular}
\usepackage{amsmath}
\usepackage{amssymb}
\usepackage{stmaryrd}
\usepackage{xcolor}
\usepackage{fullpage}
\usepackage{multirow}
\usepackage{url}
\newcommand{\ghcfile}[1]{\textsl{#1}}
\newcommand{\arraylabel}[1]{\multicolumn{2}{l}{\!\!\!\!\!\!\!\!\!\text{\underline{#1}:}}}
\input{CoreOtt}
% increase spacing between rules for ease of reading:
\renewcommand{\ottusedrule}[1]{\[#1\]\\[1ex]}
\setlength{\parindent}{0in}
\setlength{\parskip}{1ex}
\newcommand{\gram}[1]{\ottgrammartabular{#1\ottafterlastrule}}
\begin{document}
\begin{center}
\LARGE
System FC, as implemented in GHC\footnote{This
document was originally prepared by Richard Eisenberg (\texttt{eir@cis.upenn.edu}),
but it should be maintained by anyone who edits the functions or data structures
mentioned in this file. Please feel free to contact Richard for more information.}\\
\Large 21 November, 2013
\end{center}
\section{Introduction}
This document presents the typing system of System FC, very closely to how it is
implemented in GHC. Care is taken to include only those checks that are actually
written in the GHC code. It should be maintained along with any changes to this
type system.
Who will use this? Any implementer of GHC who wants to understand more about the
type system can look here to see the relationships among constructors and the
different types used in the implementation of the type system. Note that the
type system here is quite different from that of Haskell---these are the details
of the internal language, only.
At the end of this document is a \emph{hypothetical} operational semantics for GHC.
It is hypothetical because GHC does not strictly implement a concrete operational
semantics anywhere in its code. While all the typing rules can be traced back to
lines of real code, the operational semantics do not, in general, have as clear
a provenance.
There are a number of details elided from this presentation. The goal of the
formalism is to aid in reasoning about type safety, and checks that do not
work toward this goal were omitted. For example, various scoping checks (other
than basic context inclusion) appear in the GHC code but not here.
\section{Grammar}
\subsection{Metavariables}
We will use the following metavariables:
\ottmetavars{}\\
\subsection{Literals}
Literals do not play a major role, so we leave them abstract:
\gram{\ottlit}
We also leave abstract the function \coderef{basicTypes/Literal.lhs}{literalType}
and the judgment \coderef{coreSyn/CoreLint.lhs}{lintTyLit} (written $[[G |-tylit lit : k]]$).
\subsection{Variables}
GHC uses the same datatype to represent term-level variables and type-level
variables:
\gram{
\ottz
}
\gram{
\ottn
}
\subsection{Expressions}
The datatype that represents expressions:
\gram{\otte}
There are a few key invariants about expressions:
\begin{itemize}
\item The right-hand sides of all top-level and recursive $[[let]]$s
must be of lifted type.
\item The right-hand side of a non-recursive $[[let]]$ and the argument
of an application may be of unlifted type, but only if the expression
is ok-for-speculation. See \verb|#let_app_invariant#| in \ghcfile{coreSyn/CoreSyn.lhs}.
\item We allow a non-recursive $[[let]]$ for bind a type variable.
\item The $[[_]]$ case for a $[[case]]$ must come first.
\item The list of case alternatives must be exhaustive.
\item Types and coercions can only appear on the right-hand-side of an application.
\end{itemize}
Bindings for $[[let]]$ statements:
\gram{\ottbinding}
Case alternatives:
\gram{\ottalt}
Constructors as used in patterns:
\gram{\ottKp}
Notes that can be inserted into the AST. We leave these abstract:
\gram{\otttick}
A program is just a list of bindings:
\gram{\ottprogram}
\subsection{Types}
\gram{\ottt}
There are some invariants on types:
\begin{itemize}
\item The type $[[t1]]$ in the form $[[t1 t2]]$ must not be a type constructor
$[[T]]$. It should be another application or a type variable.
\item The form $[[T ]]$ (\texttt{TyConApp})
does \emph{not} need to be saturated.
\item A saturated application of $[[(->) t1 t2]]$ should be represented as
$[[t1 -> t2]]$. This is a different point in the grammar, not just pretty-printing.
The constructor for a saturated $[[(->)]]$ is \texttt{FunTy}.
\item A type-level literal is represented in GHC with a different datatype than
a term-level literal, but we are ignoring this distinction here.
\end{itemize}
\subsection{Coercions}
\gram{\ottg}
Invariants on coercions:
\begin{itemize}
\item $[[_R]]$ is used; never $[[_R _Nom]]$.
\item If $[[_R]]$ is applied to some coercions, at least one of which is not
reflexive, use $[[T_R ]]$, never $[[_R g1 g2]] \ldots$.
\item The $[[T]]$ in $[[T_R ]]$ is never a type synonym, though it could
be a type function.
\end{itemize}
Roles label what equality relation a coercion is a witness of. Nominal equality
means that two types are identical (have the same name); representational equality
means that two types have the same representation (introduced by newtypes); and
phantom equality includes all types. See \url{http://ghc.haskell.org/trac/ghc/wiki/Roles}
for more background.
\gram{\ottR}
Is it a left projection or a right projection?
\gram{\ottLorR}
Axioms:
\gram{
\ottC\ottinterrule
\ottaxBranch
}
The definition for $[[axBranch]]$ above does not include the list of
incompatible branches (field \texttt{cab\_incomps} of \texttt{CoAxBranch}),
as that would unduly clutter this presentation. Instead, as the list
of incompatible branches can be computed at any time, it is checked for
in the judgment $[[no_conflict]]$. See Section~\ref{sec:no_conflict}.
Axiom rules, produced by the type-nats solver:
\gram{\ottmu}
-\label{sec:axiom_rule}
+\label{sec:axiom_rules}
An axiom rule $[[mu]] = [[M(I, role_list, R')]]$ is an axiom name $[[M]]$, with a
type arity $[[I]]$, a list of roles $[[role_list]]$ for its coercion parameters,
and an output role $[[R']]$. The definition within GHC also includes a field named
$[[coaxrProves]]$ which computes the output coercion from a list of types and
a list of coercions. This is elided in this presentation, as we simply identify
axiom rules by their names $[[M]]$. See also \coderef{typecheck/TcTypeNats.lhs}{mkBinAxiom}
and \coderef{typecheck/TcTypeNats.lhs}{mkAxiom1}.
\subsection{Type constructors}
Type constructors in GHC contain \emph{lots} of information. We leave most of it out
for this formalism:
\gram{\ottT}
We include some representative primitive type constructors. There are many more in \ghcfile{prelude/TysPrim.lhs}.
\gram{\ottH}
\section{Contexts}
The functions in \ghcfile{coreSyn/CoreLint.lhs} use the \texttt{LintM} monad.
This monad contains a context with a set of bound variables $[[G]]$. The
formalism treats $[[G]]$ as an ordered list, but GHC uses a set as its
representation.
\gram{
\ottG
}
We assume the Barendregt variable convention that all new variables are
fresh in the context. In the implementation, of course, some work is done
to guarantee this freshness. In particular, adding a new type variable to
the context sometimes requires creating a new, fresh variable name and then
applying a substitution. We elide these details in this formalism, but
see \coderef{types/Type.lhs}{substTyVarBndr} for details.
\section{Typing judgments}
The following functions are used from GHC. Their names are descriptive, and they
are not formalized here: \coderef{types/TyCon.lhs}{tyConKind},
\coderef{types/TyCon.lhs}{tyConArity}, \coderef{basicTypes/DataCon.lhs}{dataConTyCon}, \coderef{types/TyCon.lhs}{isNewTyCon}, \coderef{basicTypes/DataCon.lhs}{dataConRepType}.
\subsection{Program consistency}
Check the entire bindings list in a context including the whole list. We extract
the actual variables (with their types/kinds) from the bindings, check for duplicates,
and then check each binding.
\ottdefnlintCoreBindings{}
Here is the definition of $[[vars_of]]$, taken from \coderef{coreSyn/CoreSyn.lhs}{bindersOf}:
\[
\begin{array}{ll}
[[vars_of n = e]] &= [[n]] \\
[[vars_of rec ]] &= [[]]
\end{array}
\]
\subsection{Binding consistency}
\ottdefnlintXXbind{}
\ottdefnlintSingleBinding{}
In the GHC source, this function contains a number of other checks, such
as for strictness and exportability. See the source code for further information.
\subsection{Expression typing}
\ottdefnlintCoreExpr{}
\begin{itemize}
\item Some explication of \ottdrulename{Tm\_LetRec} is helpful: The idea behind the
second premise ($[[]]$) is that we wish
to check each substituted type $[[s'i]]$ in a context containing all the types
that come before it in the list of bindings. The $[[G'i]]$ are contexts
containing the names and kinds of all type variables (and term variables,
for that matter) up to the $i$th binding. This logic is extracted from
\coderef{coreSyn/CoreLint.lhs}{lintAndScopeIds}.
\item There is one more case for $[[G |-tm e : t]]$, for type expressions.
This is included in the GHC code but is elided
here because the case is never used in practice. Type expressions
can only appear in arguments to functions, and these are handled
in \ottdrulename{Tm\_AppType}.
\item The GHC source code checks all arguments in an application expression
all at once using \coderef{coreSyn/CoreSyn.lhs}{collectArgs}
and \coderef{coreSyn/CoreLint.lhs}{lintCoreArgs}. The operation
has been unfolded for presentation here.
\item If a $[[tick]]$ contains breakpoints, the GHC source performs additional
(scoping) checks.
\item The rule for $[[case]]$ statements also checks to make sure that
the alternatives in the $[[case]]$ are well-formed with respect to the
invariants listed above. These invariants do not affect the type or
evaluation of the expression, so the check is omitted here.
\item The GHC source code for \ottdrulename{Tm\_Var} contains checks for
a dead id and for one-tuples. These checks are omitted here.
\end{itemize}
\subsection{Kinding}
\ottdefnlintType{}
\subsection{Kind validity}
\ottdefnlintKind{}
\subsection{Coercion typing}
In the coercion typing judgment, the $\#$ marks are left off the equality
operators to reduce clutter. This is not actually inconsistent, because
the GHC function that implements this check, \texttt{lintCoercion}, actually
returns four separate values (the kind, the two types, and the role), not
a type with head $[[(~#)]]$ or $[[(~R#)]]$. Note that the difference between
these two forms of equality is interpreted in the rules \ottdrulename{Co\_CoVarCoNom}
and \ottdrulename{Co\_CoVarCoRepr}.
\ottdefnlintCoercion{}
In \ottdrulename{Co\_AxiomInstCo}, the use of $[[inits]]$ creates substitutions from
the first $i$ mappings in $[[ si] // i /> ]]$. This has the effect of
folding the substitution over the kinds for kind-checking.
See Section~\ref{sec:tyconroles} for more information about $[[tyConRolesX]]$, and
see Section~\ref{sec:axiom_rules} for more information about $[[coaxrProves]]$.
\subsection{Name consistency}
There are two very similar checks for names, one declared as a local function:
\ottdefnlintSingleBindingXXlintBinder{}
\ottdefnlintBinder{}
\subsection{Substitution consistency}
\ottdefncheckTyKind{}
\subsection{Case alternative consistency}
\ottdefnlintCoreAlt{}
\subsection{Telescope substitution}
\ottdefnapplyTys{}
\subsection{Case alternative binding consistency}
\ottdefnlintAltBinders{}
\subsection{Arrow kinding}
\ottdefnlintArrow{}
\subsection{Type application kinding}
\ottdefnlintXXapp{}
\subsection{Sub-kinding}
\ottdefnisSubKind{}
\subsection{Roles}
\label{sec:tyconroles}
During type-checking, role inference is carried out, assigning roles to the
arguments of every type constructor. The function $[[tyConRoles]]$ extracts these
roles. Also used in other judgments is $[[tyConRolesX]]$, which is the same as
$[[tyConRoles]]$, but with an arbitrary number of $[[Nom]]$ at the end, to account
for potential oversaturation.
The checks encoded in the following
judgments are run from \coderef{typecheck/TcTyClsDecls.lhs}{checkValidTyCon}
when \texttt{-dcore-lint} is set.
\ottdefncheckValidRoles{}
\ottdefncheckXXdcXXroles{}
In the following judgment, the role $[[R]]$ is an \emph{input}, not an output.
The metavariable $[[O]]$ denotes a \emph{role context}, as shown here:
\gram{\ottO}
\ottdefncheckXXtyXXroles{}
These judgments depend on a sub-role relation:
\ottdefnltRole{}
\subsection{Branched axiom conflict checking}
\label{sec:no_conflict}
The following judgment is used within \ottdrulename{Co\_AxiomInstCo} to make
sure that a type family application cannot unify with any previous branch
in the axiom. The actual code scans through only those branches that are
flagged as incompatible. These branches are stored directly in the
$[[axBranch]]$. However, it is cleaner in this presentation to simply
check for compatibility here.
\ottdefncheckXXnoXXconflict{}
The judgment $[[apart]]$ checks to see whether two lists of types are surely
apart. $[[apart( , )]]$, where $[[
]]$ is a list of types and $[[ ]]$ is a list of type
\emph{patterns} (as in a type family equation), first flattens the $[[ ]]$ using \coderef{types/FamInstEnv.lhs}{flattenTys} and then checks to
see if \coderef{types/Unify.lhs}{tcUnifyTysFG} returns \texttt{SurelyApart}.
Flattening takes all type family applications and replaces them with fresh variables,
taking care to map identical type family applications to the same fresh variable.
The algorithm $[[unify]]$ is implemented in \coderef{types/Unify.lhs}{tcUnifyTys}.
It performs a standard unification, returning a substitution upon success.
\section{Operational semantics}
\subsection{Disclaimer}
GHC does not implement an operational semantics in any concrete form. Most
of the rules below are implied by algorithms in, for example, the simplifier
and optimizer. Yet, there is no one place in GHC that states these rules,
analogously to \texttt{CoreLint.lhs}.
Nevertheless, these rules are included in this document to help the reader
understand System FC.
\subsection{The context $[[S]]$}
We use a context $[[S]]$ to keep track of the values of variables in a (mutually)
recursive group. Its definition is as follows:
\[
[[S]] \quad ::= \quad [[ empty ]] \ |\ [[S]], [[ [n |-> e] ]]
\]
The presence of the context $[[S]]$ is solely to deal with recursion. If your
use of FC does not require modeling recursion, you will not need to track $[[S]]$.
\subsection{Operational semantics rules}
\ottdefnstep{}
\subsection{Notes}
\begin{itemize}
\item The \ottdrulename{S\_LetRec} and \ottdrulename{S\_LetRecReturn} rules
implement recursion. \ottdrulename{S\_LetRec} adds to the context $[[S]]$ bindings
for all of the mutually recursive equations. Then, after perhaps many steps,
when the body of the $[[let]]\ [[rec]]$ contains no variables that are bound
in the $[[let]]\ [[rec]]$, the context is popped.
\item In the $[[case]]$ rules, a constructor $[[K]]$ is written taking three
lists of arguments: two lists of types and a list of terms. The types passed
in are the universally and, respectively, existentially quantified type variables
to the constructor. The terms are the regular term arguments stored in an
algebraic datatype. Coercions (say, in a GADT) are considered term arguments.
\item The rule \ottdrulename{S\_CasePush} is the most complex rule.
\begin{itemize}
\item The logic in this rule is implemented in \coderef{coreSyn/CoreSubst.lhs}{exprIsConApp\_maybe}.
\item The $[[coercionKind]]$ function (\coderef{types/Coercion.lhs}{coercionKind})
extracts the two types (and their kind) from
a coercion. It does not require a typing context, as it does not \emph{check} the
coercion, just extracts its types.
\item The $[[dataConRepType]]$ function (\coderef{basicTypes/DataCon.lhs}{dataConRepType}) extracts the full type of a data constructor. Following the notation for
constructor expressions, the parameters to the constructor are broken into three
groups: universally quantified types, existentially quantified types, and terms.
\item The substitutions in the last premise to the rule are unusual: they replace
\emph{type} variables with \emph{coercions}. This substitution is called lifting
and is implemented in \coderef{types/Coercion.lhs}{liftCoSubst}. The notation is
essentially a pun on the fact that types and coercions have such similar structure.
\item Note that the types $[[ ]]$---the existentially quantified
types---do not change during this step.
\end{itemize}
\end{itemize}
\end{document}