agda/notes/typechecking/algorithm.tex

\documentclass[a4paper,11pt]{article}

\usepackage[OT1]{fontenc}
\usepackage[latin1]{inputenc}
\usepackage{amsmath,amssymb,amsfonts}

\usepackage{proof}

\newcommand\keyword[1]{\mathbf{#1}}
\newcommand\Coloneqq{\mathrel{::=}}
\newcommand\OR{~~|~~}
\newcommand\Hid[1]{\{#1\}}
\newcommand\lam[1]{\lambda#1.\,}
\newcommand\hlam[1]{\lam{\Hid{#1}}}
\newcommand\tlam[2]{\lam{(#1:#2)}}
\newcommand\thlam[2]{\lam{\Hid{#1:#2}}}
\newcommand\ePi[3]{(#1:#2)\to#3}
\newcommand\ehPi[3]{\{#1:#2\}\to#3}
\newcommand\vPi[2]{\Pi#1:#2.\,}
\newcommand\vhPi[2]{\Pi\{#1:#2\}.\,}
\newcommand\vPiTel[1]{\Pi#1.\,}
\newcommand\vhPiTel[1]{\vPiTel{\{#1\}}}
\newcommand\Let[2]{\keyword{let}~#1~\keyword{in}~#2}
\newcommand\Set[1]{\mathsf{Set}_{#1}}
\newcommand\Prop{\mathsf{Prop}}
\newcommand\el{\mathsf{El}}
\newcommand\El[1]{\el_{#1}\,}
\newcommand\lub{\sqcup}

\newcommand\APP[2]{\mathsf{app}(#1,#2)}
\newcommand\HAPP[2]{\mathsf{happ}(#1,#2)}
\newcommand\Subst[3]{#1[#2/#3]}

\newcommand\GetSort[1]{\mathsf{sortof}(#1)}

% Judgement forms
\renewcommand\Check[5]{#1\,;\,#2\vdash#3\uparrow#4:#5}
\newcommand\Infer[5]{#1\,;\,#2\vdash#3\downarrow#4:#5}
\newcommand\IsType[4]{#1\,;\,#2\vdash#3\uparrow#4~\mathbf{type}}
\newcommand\Equal[5]{#1\,;\,#2\vdash#3=#4:#5}
\newcommand\TEqual[4]{#1\,;\,#2\vdash#3=#4}
\newcommand\Expand[6]{#1\,;\,#2\vdash#3:#4\prec#5:#6}
\newcommand\CheckDecl[4]{#1\,;\,#2\vdash#3\to#4}

\newcommand\AddGlobalMeta[4]{#1\,;\,#2\vdash{#3}:#4}
\newcommand\AddLocalMeta[4]{#1\,;\,#2\vdash{#3}:#4}

\title{Agda II Type Checking Algorithm}
\author{Ulf Norell}

\begin{document}
\maketitle

\section{Introduction}

    Write something nice here.

\section{Syntax}

\subsection{Expressions}

    Expressions have been scope checked, but not type checked. Hence the mix
    between terms, types and sorts.

    \[\begin{array}{lcl}
	e & \Coloneqq & x \OR c \OR l \OR ? \OR \_ \\
	  & \OR & \lam xe \OR \hlam xe \OR \tlam xee \OR \thlam xee \\
	  & \OR & e\,e \OR e\,\Hid e \OR \Let{\vec\delta}e \\
	  & \OR & \ePi xee \OR \ehPi xee \OR \Set n \OR \Prop \\
	l & \Coloneqq & \mathit{integer} \OR \mathit{float} \OR \mathit{string} \OR \mathit{character} \\
    \end{array}\]

    Constants ($c$) are names of constructors, defined functions, postulates and datatypes.

\subsection{Declarations}

    \[\begin{array}{lcl}
	\delta & \Coloneqq & \ldots
    \end{array}\]

\subsection{Terms}

    Terms are type checked versions of expressions (that aren't types). The
    type checking algorithm produces one of these when type checking. The
    implementation uses deBruijn variables, but to simplify the presentation
    we don't do that here.

    \[\begin{array}{lcl}
	s,t & \Coloneqq & x \OR c \OR l \OR ?_i \\
	    & \OR & \lam xt \OR \hlam xt \OR s\,t \OR s\,\Hid t
    \end{array}\]

    Worth noting is that meta variables are now named and that there are no
    typed lambdas left.

    Terms are supposed to always be on normal form. We do some work in the
    rules to ensure that this is the case.

\subsection{Types and Sorts}

    After type checking we distinguish between terms, types and sorts.

    \[\begin{array}{lcl}
	A,B & \Coloneqq & \El\alpha t \OR \vPi xAB \OR \vhPi xAB \OR \alpha \\
	\alpha,\beta & \Coloneqq & \Set n \OR \Prop \\
    \end{array}\]

    In some presentation of the system design we had type and sort meta
    variables. I will try to do without them. What this means is that we can't,
    for instance, infer the type of a domain-free lambda by introducing a meta
    variable for the domain type.

    The reason for not having type meta variables is that I'm not sure how they
    interact with coercions.  Depending on the order you solve constraints you
    might end up with different instantiations (since different coercions were
    applied).  It might be that it doesn't matter, but until we're sure I
    prefer to err on the side of caution.

    If $x\notin\mathit{FV}(B)$ I will sometimes write $A\to B$ for $\vPi xAB$.

\section{Judgement forms}

    In the judgement forms below $\Sigma$ is the signature and contains the
    type and definition (if any) of the constants currently in scope. $\Gamma$
    is the context and contains the types of the bound variables.

    \[\begin{array}{ll}
	\Check\Sigma\Gamma etA	    & \mbox{Type checking, computes $t$.} \\
	\Infer\Sigma\Gamma etA	    & \mbox{Type inference, computes $t$ and $A$.} \\
	\IsType\Sigma\Gamma eA	    & \mbox{Checking that $e$ is a type, computes $A$.} \\
	\Equal\Sigma\Gamma stA	    & \mbox{Typed conversion.} \\
%	\TEqual\Sigma\Gamma AB	    & \mbox{Type conversion.} \\
	\Expand\Sigma\Gamma sAtB    & \mbox{Coercing conversion, computes $t$.} \\
	\CheckDecl\Sigma\Gamma\delta{\Sigma'} & \mbox{Checking declarations, computes $\Sigma'$.}
    \end{array}\]

    The only non-standard judgement is the coercing conversion which replaces
    the convertibility judgement for types. The purpose of this judgement form
    is to insert things that have been hidden. For instance, suppose that
    $f:\vhPi A{\Set0}\vPi xAA$ in $\Sigma$ and we want to check
    $\Check\Sigma\Gamma ft{\vPi xBB}$. This should succeed with $t =
    f\,\Hid{B}$. The reason it does succeed is because we can coerce $f$ to
    have the desired type:
    \[
	\Expand\Sigma\Gamma f{\vhPi A{\Set0}\vPi xAA}{f\,\Hid{B}}{\vPi xBB}
    \]

\section{Judgements}

\subsection{Checking}

    Type checking is used only as a last resort. If we can infer the type,
    that's what we do.

    \[\begin{array}{c}
	\infer{ \Check\Sigma\Gamma etB }
	{ \Infer\Sigma\Gamma esA
	& \Expand\Sigma\Gamma sAtB
	}
    \end{array}\]

    The coercing conversion inserts any hidden lambdas or applications that are
    missing from $s$.

    We can't infer the type of domain free lambdas.

    \[\begin{array}{c}
	\infer{ \Check\Sigma\Gamma{\lam xe}{\lam xt}{\vPi xAB} }
	      { \Check\Sigma{\Gamma,x:A}etB }
	\\{}\\
	\infer{ \Check\Sigma\Gamma{\hlam xe}{\hlam xt}{\vhPi xAB} }
	      { \Check\Sigma{\Gamma,x:A}etB }
    \end{array}\]

    If we're checking a non-hidden lambda against a hidden function type we
    should insert the appropriate number of hidden lambdas. There is some
    abuse of notation to make the rule more readable: If $\Delta =
    (x_1:A_1)\ldots(x_n:A_n)$, then $\hlam\Delta t$ means $\hlam{x_1}\ldots\hlam{x_n}t$ and
    $\vhPiTel\Delta B$ means $\vhPi{x_1}{A_1}\ldots\vhPi{x_n}{A_n}B$.

    \[\begin{array}{c}
	\infer{ \Check\Sigma\Gamma{\lam xe}{\hlam\Delta\lam xt}{\vhPiTel\Delta\vPi xAB} }
	{ \Check\Sigma{\Gamma,\Delta,x:A} etB
	}
    \end{array}\]

    The type of meta variables can't be inferred either.

    \[\begin{array}{ccc}
	\infer[\AddGlobalMeta\Sigma\Gamma{?_i}A]{\Check\Sigma\Gamma{{?}}{{?_i}}A}{}
	    &&
	\infer[\AddLocalMeta\Sigma\Gamma{?_i}A]{\Check\Sigma\Gamma\_{{?_i}}A}{}
    \end{array}\]

    Let bindings can be inferred only if the body can be inferred, so we need a
    checking rule in case it can't.

    \[\begin{array}{c}
	\infer{ \Check\Sigma\Gamma{\Let\delta e}tA }
	{ \CheckDecl\Sigma\Gamma\delta{\Sigma'}
	& \Check{\Sigma'}\Gamma etA
	}
    \end{array}\]

    An alternative approach would be to infer the type of everything, inserting
    meta variables when we don't know.  This would require type and sort meta
    variables, though.

\subsection{Inference}

    Inferring the type of a variable or a constant amounts to looking it up in
    the context or signature. This will never fail, since the expressions have
    been scope checked prior to type checking.

    \[\begin{array}{ccc}
	\infer{\Infer\Sigma\Gamma xx{\Gamma(x)}}{} &&
	\infer{\Infer\Sigma\Gamma cc{\Sigma(x)}}{}
    \end{array}\]

    Literals have predefined types.
    \[\begin{array}{c}
	\infer{\Infer\Sigma\Gamma ll{\mathsf{typeof}(l)}}{}
    \end{array}\]

    There are three rules for application. The first two are for the easy cases
    where all implicit arguments have been made explicit.
    \[\begin{array}{c}
	\infer{ \Infer\Sigma\Gamma{e_1\,e_2}{\APP st}{\Subst Btx} }
	{ \Infer\Sigma\Gamma{e_1}s{\vPi xAB}
	& \Check\Sigma\Gamma{e_2}tA
	}
	\\{}\\
	\infer{ \Infer\Sigma\Gamma{e_1\,\Hid{e_2}}{\HAPP st}{\Subst Btx} }
	{ \Infer\Sigma\Gamma{e_1}s{\vhPi xAB}
	& \Check\Sigma\Gamma{e_2}tA
	}
    \end{array}\]
    The functions $\APP --$ and $\HAPP --$ perform $\beta$-reductions and
    definition unfolding if necessary, to make sure that terms are on normal
    form. The third rule handles the case when you apply a function expecting
    hidden arguments to a non-hidden argument, in which case we have to fill in
    the hidden arguments with meta variables.
    \[\begin{array}{c}
	\infer[\AddLocalMeta\Sigma\Gamma{\vec ?}\Delta]
	{ \Infer\Sigma\Gamma{e_1\,e_2}{\APP{\HAPP{s}{\vec{?}}}t}{B[\vec{?}/\Delta,t/x]} }
	{ \Infer\Sigma\Gamma{e_1}s{\vhPiTel\Delta\vPi xAB}
	& \Check\Sigma\Gamma{e_2}t{\Subst A{\vec ?}\Delta}
	}
    \end{array}\]
    A consequence of these rules is that when you give a hidden argument
    explicitly it is always interpreted as the left-most hidden argument, so
    $f\,\Hid{x}\,y$ is the same as $f\,\Hid{x}\,\Hid{\_}\,y$ for an $f$ of the
    appropriate type.

    The inference rule for let is the same as the checking rule.
    \[\begin{array}{c}
	\infer{ \Infer\Sigma\Gamma{\Let\delta e}tA }
	{ \CheckDecl\Sigma\Gamma\delta{\Sigma'}
	& \Infer{\Sigma'}\Gamma etA
	}
    \end{array}\]

\subsection{Computing sorts}

    Types contain enough information to retrieve the sort.

    \[\begin{array}{lcl}
	\GetSort{\El\alpha t} & = & \alpha \\
	\GetSort{\vPi xAB} & = & \GetSort A\lub\GetSort B \\
	\GetSort{\vhPi xAB} & = & \GetSort A\lub\GetSort B \\
	\GetSort{\Set n} & = & \Set{n+1} \\
	\GetSort{\Prop} & = & \Set1 \\
	{}\\
	\Set n\lub\Set m & = & \Set{\mathsf{max}(n,m)} \\
	\Prop\lub\Prop & = & \Prop \\
	\Prop\lub\Set n & = & \Set1\lub\Set n \\
	\Set n\lub\Prop & = & \Set n\lub\Set 1 \\
    \end{array}\]

    In PTS terms we have the rule $(\alpha,\beta,\alpha\lub\beta)$.
    We might want to consider having $(\Set0,\Prop,\Prop)$ as well.

\subsection{Is type}

    The {\em is type} judgement checks that an expression is a valid type and
    returns that type.  It could also compute its sort but since we can easily get the
    sort of a type, it isn't necessary.

    \[\begin{array}{c}
	\infer
	{ \IsType\Sigma\Gamma e{\El\alpha t} }
	{ \Infer\Sigma\Gamma et\alpha }
    \end{array}\]
    
    \[\begin{array}{c}
	\infer
	{ \IsType\Sigma\Gamma{\ePi x{e_1}{e_2}}{\vPi xAB} }
	{ \IsType\Sigma\Gamma{e_1}A
	& \IsType\Sigma{\Gamma,x:A}{e_2}B
	} \\{}\\
	\infer
	{ \IsType\Sigma\Gamma{\ehPi x{e_1}{e_2}}{\vhPi xAB} }
	{ \IsType\Sigma\Gamma{e_1}A
	& \IsType\Sigma{\Gamma,x:A}{e_2}B
	}
    \end{array}\]

    \[\begin{array}{c}
	\infer
	{ \IsType\Sigma\Gamma\alpha{\GetSort\alpha} }
	{}
    \end{array}\]

\subsection{Coercing conversion}

    The coercing conversion $\Expand\Sigma\Gamma sAtB$ computes a $t$ of type
    $B$ given an $s$ of type $A$, by adding hidden applications or lambdas to
    $s$ until the types $A$ and $B$ match.

%     In the following two rules $C$ should not be a hidden function space
%     ($\vhPi xAB$).
    \[\begin{array}{c}
	\infer[\AddLocalMeta\Sigma\Gamma{?_i}A]
	{ \Expand\Sigma\Gamma s{\vhPi xAB}tC }
	{ \Expand\Sigma\Gamma{\HAPP s{?_i}}{\Subst B{?_i}x}tC }
	\\{}\\
	\infer
	{ \Expand\Sigma\Gamma sC{\hlam xt}{\vhPi xAB} }
	{ \Expand\Sigma{\Gamma,x:A}sCtB }
    \end{array}\]
    The first rule applies when $C$ is not a hidden function space, and the
    second rule is applicable for any $C$. This has the effect of
    $\eta$-expanding functions with hidden arguments. This allows, for
    instance, $s:\vhPiTel{A,B:\Set0}A\to B\to A$ to be coerced to $\hlam
    As\,\Hid A\,\Hid A : \vhPi A{\Set0}A\to A\to A$.

    If both types are normal function spaces we $\eta$-expand.
    \[\begin{array}{c}
	\infer
	{ \Expand\Sigma\Gamma s{\vPi xAB}{\lam ys'}{\vPi y{A'}B'} }
	{ \Expand\Sigma{\Gamma,y:A'}y{A'}tA
	& \Expand\Sigma{\Gamma,y:A'}{s\,t}{\Subst Btx}{s'}{B'}
	}
    \end{array}\]
    This allows us to perform coercions in higher order functions. For instance, 
    \[
	\Expand\Sigma\Gamma f{(B\to B)\to C}
	    {\lam xf\,(x\,\Hid{B})}
	    {(\vhPi A{\Set0}A\to A)\to C}
    \]
    The last two cases are when the types are $\el$s or sorts. In neither case
    are there any coercions.
    \[\begin{array}{ccc}
	\infer
	{ \Expand\Sigma\Gamma t\alpha t\alpha }
	{}
	&&
	\infer
	{ \Expand\Sigma\Gamma s{\El\alpha t_1}s{\El\alpha t_2} }
	{ \Equal\Sigma\Gamma{t_1}{t_2}\alpha }
    \end{array}\]

\subsection{Conversion}

    The conversion checking is type directed. This gives us $\eta$-equality for
    functions in a nice way. It also makes it possible to introduce proof
    irrelevance with a rule like this:
    \[\left(\begin{array}{c}
	\infer
	{ \Equal\Sigma\Gamma pqP }
	{ \GetSort{P} = \Prop }
    \end{array}\right)\]
    We don't do that at this point though, but only make use of the types in the
    function case:
    \[\begin{array}{c}
	\infer
	{ \Equal\Sigma\Gamma st{\vPi xAB} }
	{ \Equal\Sigma{\Gamma,x:A}{\APP sx}{\APP tx}B
	}
	\\{}\\
	\infer
	{ \Equal\Sigma\Gamma st{\vhPi xAB} }
	{ \Equal\Sigma{\Gamma,x:A}{\HAPP sx}{\HAPP tx}B
	}
    \end{array}\]

    There are a number of notation abuses in the following two rules. Firstly,
    $\Equal\Sigma\Gamma{\vec s}{\vec t}\Delta$ denotes the extension of the
    conversion judgement to sequences of terms. I am also a bit sloppy with the
    hiding: in $\vPiTel\Delta A$, $\Delta$ can contain both hidden and non-hidden
    things. Consequently when I say $x\,\vec s$ it includes hidden applications.
    \[\begin{array}{c}
	\infer
	{ \Equal\Sigma\Gamma{x\,\vec s}{x\,\vec t}A }
	{ x:\vPiTel\Delta A'\in\Gamma
	& \Equal\Sigma\Gamma{\vec s}{\vec t}\Delta
	}
	\\{}\\
	\infer
	{ \Equal\Sigma\Gamma{c\,\vec s}{c\,\vec t}A }
	{ c:\vPiTel\Delta A'\in\Sigma
	& \Equal\Sigma\Gamma{\vec s}{\vec t}\Delta
	}
    \end{array}\]
    

\subsection{Declarations}

\end{document}