Features Download
From: Andy Wingo <wingo <at> pobox.com>
Subject: compiler.texi: Compiling to the virtual machine
Newsgroups: gmane.lisp.guile.devel
Date: Sunday 11th January 2009 17:37:32 UTC (over 8 years ago)

@c -*-texinfo-*-
@c This is part of the GNU Guile Reference Manual.
@c Copyright (C)  2008
@c   Free Software Foundation, Inc.
@c See the file guile.texi for copying conditions.

@node Compiling to the Virtual Machine
@section Compiling to the Virtual Machine

Compilers have a mystique about them that is attractive and
off-putting at the same time. They are attractive because they are
magical -- they transform inert text into live results, like throwing
the switch on Frankenstein's monster. However, this magic is perceived
by many to be impenetrable.

This section aims to pay attention to the small man behind the

@xref{Read/Load/Eval/Compile}, if you're lost and you just wanted to
know how to compile your .scm file.

* Compiler Tower::                   
* The Scheme Compiler::                   
* GHIL::                 
* GLIL::                
* Object Code::                   
* Extending the Compiler::
@end menu

@node Compiler Tower
@subsection Compiler Tower

Guile's compiler is quite simple, actually -- its @emph{compilers}, to
put it more accurately. Guile defines a tower of languages, starting
at Scheme and progressively simplifying down to languages that
resemble the VM instruction set (@pxref{Instruction Set}).

Each language knows how to compile to the next, so each step is simple
and understandable. Furthermore, this set of languages is not
hardcoded into Guile, so it is possible for the user to add new
high-level languages, new passes, or even different compilation

Languages are registered in the module, @code{(system base language)}:

(use-modules (system base language))
@end example

They are registered with the @code{define-language} form.

@deffn {Scheme Syntax} define-language @
name title version reader printer @
[parser=#f] [read-file=#f] [compilers='()] [evaluator=#f]
Define a language.

This syntax defines a @code{#} object, bound to @var{name}
in the current environment. In addition, the language will be added to
the global language set. For example, this is the language definition
for Scheme:

(define-language scheme
  #:title	"Guile Scheme"
  #:version	"0.5"
  #:reader	read
  #:read-file	read-file
  #:compilers   `((,ghil . ,compile-ghil))
  #:evaluator	(lambda (x module) (primitive-eval x))
  #:printer	write)
@end example

In this example, from @code{(language scheme spec)}, @code{read-file}
reads expressions from a port and wraps them in a @code{begin} block.
@end deffn

The interesting thing about having languages defined this way is that
they present a uniform interface to the read-eval-print loop. This
allows the user to change the current language of the REPL:

$ guile
Guile Scheme interpreter 0.5 on Guile 1.9.0
Copyright (C) 2001-2008 Free Software Foundation, Inc.

Enter `,help' for help.
scheme@@(guile-user)> ,language ghil
Guile High Intermediate Language (GHIL) interpreter 0.3 on Guile 1.9.0
Copyright (C) 2001-2008 Free Software Foundation, Inc.

Enter `,help' for help.
@end example

Languages can be looked up by name, as they were above.

@deffn {Scheme Procedure} lookup-language name
Looks up a language named @var{name}, autoloading it if necessary.

Languages are autoloaded by looking for a variable named @var{name} in
a module named @code{(language @var{name} spec)}.

The language object will be returned, or @code{#f} if there does not
exist a language with that name.
@end deffn

Defining languages this way allows us to programmatically determine
the necessary steps for compiling code from one language to another.

@deffn {Scheme Procedure} lookup-compilation-order from to
Recursively traverses the set of languages to which @var{from} can
compile, depth-first, and return the first path that can transform
@var{from} to @var{to}. Returns @code{#f} if no path is found.

This function memoizes its results in a cache that is invalidated by
subsequent calls to @code{define-language}, so it should be quite
@end deffn

There is a notion of a ``current language'', which is maintained in
the @code{*current-language*} fluid. This language is normally Scheme,
and may be rebound by the user. The runtime compilation interfaces
(@pxref{Read/Load/Eval/Compile}) also allow you to choose other source
and target languages.

The normal tower of languages when compiling Scheme goes like this:

@item Scheme, which we know and love
@item Guile High Intermediate Language (GHIL)
@item Guile Low Intermediate Language (GLIL)
@item Object code
@end itemize

Object code may be serialized to disk directly, though it has a cookie
and version prepended to the front. But when compiling Scheme at
runtime, you want a Scheme value, e.g. a compiled procedure. For this
reason, so as not to break the abstraction, Guile defines a fake
language, @code{value}. Compiling to @code{value} loads the object
code into a procedure, and wakes the sleeping giant.

Perhaps this strangeness can be explained by example:
@code{compile-file} defaults to compiling to object code, because it
produces object code that has to live in the barren world outside the
Guile runtime; but @code{compile} defaults to compiling to
@code{value}, as its product re-enters the Guile world.

Indeed, the process of compilation can circulate through these
different worlds indefinitely, as shown by the following quine:

((lambda (x) ((compile x) x)) '(lambda (x) ((compile x) x)))
@end example

@node The Scheme Compiler
@subsection The Scheme Compiler

The job of the Scheme compiler is to expand all macros and to resolve
all symbols to lexical variables. Its target language, GHIL, is fairly
close to Scheme itself, so this process is not very complicated.

The Scheme compiler is driven by a table of @dfn{translators},
declared with the @code{define-scheme-translator} form, defined in the
module, @code{(language scheme compile-ghil)}.

@deffn {Scheme Syntax} define-scheme-translator head clause1 clause2...
The best documentation of this form is probably an example. Here is
the translator for @code{if}:

(define-scheme-translator if
  ;; (if TEST THEN [ELSE])
  ((,test ,then)
   (make-ghil-if e l (retrans test) (retrans then) (retrans '(begin))))
  ((,test ,then ,else)
   (make-ghil-if e l (retrans test) (retrans then) (retrans else))))
@end example

The match syntax is from the @code{pmatch} macro, defined in
@code{(system base pmatch)}. The result of a clause should be a valid
GHIL value. If no clause matches, a syntax error is signalled.

In the body of the clauses, the following bindings are introduced:
@item @code{e}, the current environment
@item @code{l}, the current source location (or @code{#f})
@item @code{retrans}, a procedure that may be called to compile
@end itemize

Note that translators are looked up by @emph{value}, not by name. That
is to say, the translator is keyed under the @emph{value} of
@code{if}, which normally prints as @code{#}.
@end deffn

Users can extend the compiler by defining new translators.
Additionally, some forms can be inlined directly to
instructions -- @xref{Inlined Scheme Instructions}, for a list. The
actual inliners are defined in @code{(language scheme inline)}:

@deffn {Scheme Syntax} define-inline head arity1 result1 arity2 result2...
Defines an inliner for @code{head}. As in
@code{define-scheme-translator}, inliners are keyed by value and not
by name.

Expressions are matched on their arities. For example:

(define-inline eq?
  (x y) (eq? x y))
@end example

This inlines calls to the Scheme procedure, @code{eq?}, to the
instruction @code{eq?}.

A more complicated example would be:

(define-inline +
  () 0
  (x) x
  (x y) (add x y)
  (x y . rest) (add x (+ y . rest)))
@end example
@end deffn

Compilers take two arguments, an expression and an environment, and
return two values as well: an expression in the target language, and
an environment suitable for the target language. The format of the
environment is language-dependent.

For Scheme, an environment may be one of three things:
@item @code{#f}, in which case compilation is performed in the context
of the current module;
@item a module, which specifies the context of the compilation; or
@item a @dfn{compile environment}, which specifies lexical variables
as well.
@end itemize

The format of a compile environment for scheme is @code{(@var{module}
@var{lexicals} . @var{externals})}, though users are strongly
discouraged from constructing these environments themselves. Instead,
if you need this functionality -- as in GOOPS' dynamic method compiler
-- capture an environment with @code{compile-time-environment}, then
pass that environment to @code{compile}.

@deffn {Scheme Procedure} compile-time-environment
A special function known to the compiler that, when compiled, will
return a representation of the lexical environment in place at compile
time. Useful for supporting some forms of dynamic compilation. Returns
@code{#f} if called from the interpreter.
@end deffn

@node GHIL
@subsection GHIL

Guile High Intermediate Language (GHIL) is a structured intermediate
language that is close in expressive power to Scheme. It is an
expanded, pre-analyzed Scheme.

GHIL is ``structured'' in the sense that its representation is based
on records, not S-expressions. This gives a rigidity to the language
that ensures that compiling to a lower-level language only requires a
limited set of transformations. Practically speaking, consider the
GHIL type, @code{}, which has fields named @code{env},
@code{loc}, and @code{exp}. Instances of this type are records created
via @code{make-ghil-quote}, and whose fields are accessed as
@code{ghil-quote-env}, @code{ghil-quote-loc}, and
@code{ghil-quote-exp}. There is also a predicate, @code{ghil-quote?}.
@xref{Records}, for more information on records.

Expressions of GHIL name their environments explicitly, and all
variables are referenced by identity in addition to by name.
@code{(language ghil)} defines a number of routines to deal explicitly
with variables and environments:

@deftp {Scheme Variable}  [table='()]
A toplevel environment. The @var{table} holds all toplevel variables
that have been resolved in this environment.
@end deftp
@deftp {Scheme Variable}  parent [table='()] [variables='()]
A lexical environment. @var{parent} will be the enclosing lexical
environment, or a toplevel environment. @var{table} holds an alist
mapping symbols to variables bound in this environment, while
@var{variables} holds a cumulative list of all variables ever defined
in this environment.

Lexical environments correspond to procedures. Bindings introduced
e.g. by Scheme's @code{let} add to the bindings in a lexical
environment. An example of a case in which a variable might be in
@var{variables} but not in @var{table} would be a variable that is in
the same procedure, but is out of scope.
@end deftp
@deftp {Scheme Variable}  env name kind [index=#f]
A variable. @var{kind} is one of @code{argument}, @code{local},
@code{external}, @code{toplevel}, @code{public}, or @code{private};
see the procedures below for more information. @var{index} is used in
@end deftp

@deffn {Scheme Procedure} ghil-var-is-bound? env sym
Recursively look up a variable named @var{sym} in @var{env}, and
return it or @code{#f} if none is found.
@end deffn
@deffn {Scheme Procedure} ghil-var-for-ref! env sym
Recursively look up a variable named @var{sym} in @var{env}, and
return it. If the symbol was not bound, return a new toplevel
@end deffn
@deffn {Scheme Procedure} ghil-var-for-set! env sym
Like @code{ghil-var-for-ref!}, except that the returned variable will
be marked as @code{external}. @xref{Variables and the VM}.
@end deffn
@deffn {Scheme Procedure} ghil-var-define! toplevel-env sym
Return an existing or new toplevel variable named @var{sym}.
@var{toplevel-env} must be a toplevel environment.
@end deffn
@deffn {Scheme Procedure} ghil-var-at-module! env modname sym interface?
Return a variable that will be resolved at runtime with respect to a
specific module named @var{modname}. If @var{interface?} is true, the
variable will be of type @code{public}, otherwise @code{private}.
@end deffn
@deffn {Scheme Procedure} call-with-ghil-environment env syms func
Bind @var{syms} to fresh variables within a new lexical environment
whose parent is @var{env}, and call @var{func} as @code{(@var{func}
@var{new-env} @var{new-vars})}.
@end deffn
@deffn {Scheme Procedure} call-with-ghil-bindings env syms func
Like @code{call-with-ghil-environment}, except the existing
environment @var{env} is re-used. For that reason, @var{func} is
invoked as @code{(@var{func} @var{new-vars})}
@end deffn

In the aforementioned @code{} type, the @var{env} slot
holds a pointer to the environment in which the expression occurs. The
@var{loc} slot holds source location information, so that errors
corresponding to this expression can be mapped back to the initial
expression in the higher-level language, e.g. Scheme. @xref{Compiled
Procedures}, for more information on source location objects.

GHIL also has a declarative serialization format, which makes writing
and reading it a tractable problem for the human mind. Since all GHIL
language constructs contain @code{env} and @code{loc} pointers, they
are left out of the serialization. (Serializing @code{env} structures
would be difficult, as they are often circular.) What is left is the
type of expression, and the remaining slots defined in the expression

For example, an S-expression representation of the @code{}
expression would be:

(quote 3)
@end example

It's deceptively like Scheme. The general rule is, for a type defined
as @code{ env loc @var{slot1} @var{slot2}...}, the
S-expression representation will be @code{(@var{foo} @var{slot1}
@var{slot2}...)}. Users may program with this format directly at the

scheme@@(guile-user)> ,language ghil
Guile High Intermediate Language (GHIL) interpreter 0.3 on Guile 1.9.0
Copyright (C) 2001-2008 Free Software Foundation, Inc.

Enter `,help' for help.
ghil@@(guile-user)> (call (ref +) (quote 32) (quote 10))
@result{} 42
@end example

For convenience, some slots are serialized as rest arguments; those
are noted below. The other caveat is that variables are serialized as
their names only, and not their identities.

@deftp {Scheme Variable}  env loc
The unspecified value.
@end deftp
@deftp {Scheme Variable}  env loc exp
A quoted expression.

Note that unlike in Scheme, there are no self-quoting expressions; all
constants must come from @code{quote} expressions.
@end deftp
@deftp {Scheme Variable}  env loc exp
A quasiquoted expression. The expression is treated as a constant,
except for embedded @code{unquote} and @code{unquote-splicing} forms.
@end deftp
@deftp {Scheme Variable}  env loc exp
Like Scheme's @code{unquote}; only valid within a quasiquote.
@end deftp
@deftp {Scheme Variable}  env loc exp
Like Scheme's @code{unquote-splicing}; only valid within a quasiquote.
@end deftp
@deftp {Scheme Variable}  env loc var
A variable reference. Note that for purposes of serialization,
@var{var} is serialized as its name, as a symbol.
@end deftp
@deftp {Scheme Variable}  env loc var val
A variable mutation. @var{var} is serialized as a symbol.
@end deftp
@deftp {Scheme Variable}  env loc var val
A toplevel variable definition. See @code{ghil-var-define!}.
@end deftp
@deftp {Scheme Variable}  env loc test then else
A conditional. Note that @var{else} is not optional.
@end deftp
@deftp {Scheme Variable}  env loc . exps
Like Scheme's @code{and}.
@end deftp
@deftp {Scheme Variable}  env loc . exps
Like Scheme's @code{or}.
@end deftp
@deftp {Scheme Variable}  env loc . body
Like Scheme's @code{begin}.
@end deftp
@deftp {Scheme Variable}  env loc vars exprs . body
Like a deconstructed @code{let}: each element of @var{vars} will be
bound to the corresponding GHIL expression in @var{exprs}.

Note that for purposes of the serialization format, @var{exprs} are
evaluated before the new bindings are added to the environment. For
@code{letrec} semantics, there also exists a @code{bindrec} parse
flavor. This is useful for writing GHIL at the REPL, but the
serializer does not currently have the cleverness needed to determine
whether a @code{} has @code{let} or @code{letrec}
semantics, and thus only serializes @code{} as @code{bind}.
@end deftp
@deftp {Scheme Variable}  env loc vars rest producer . body
Like Scheme's @code{receive} -- binds the values returned by
applying @code{producer}, which should be a thunk, to the
@code{lambda}-like bindings described by @var{vars} and @var{rest}.
@end deftp
@deftp {Scheme Variable}  env loc vars rest meta . body
A closure. @var{vars} is the argument list, serialized as a list of
symbols. @var{rest} is a boolean, which is @code{#t} iff the last
argument is a rest argument. @var{meta} is an association list of
properties. The actual @var{body} should be a list of GHIL
@end deftp
@deftp {Scheme Variable}  env loc proc . args
A procedure call.
@end deftp
@deftp {Scheme Variable}  env loc producer consumer
Like Scheme's @code{call-with-values}.
@end deftp
@deftp {Scheme Variable}  env loc op . args
An inlined VM instruction. @var{op} should be the instruction name as
a symbol, and @var{args} should be its arguments, as GHIL expressions.
@end deftp
@deftp {Scheme Variable}  env loc . values
Like Scheme's @code{values}.
@end deftp
@deftp {Scheme Variable}  env loc . values
@var{values} are as in the Scheme expression, @code{(apply values .
@end deftp
@deftp {Scheme Variable}  env loc
Produces, at runtime, a reification of the environment at compile
time. Used in the implementation of Scheme's
@end deftp

GHIL implements a compiler to GLIL that recursively traverses GHIL
expressions, writing out GLIL expressions into a linear list. The
compiler also keeps some state as to whether the current expression is
in tail context, and whether its value will be used in future
computations. This state allows the compiler not to emit code for
constant expressions that will not be used (e.g. docstrings), and to
perform tail calls when in tail position.

Just as the Scheme to GHIL compiler introduced new hidden state---the
environment---the GHIL to GLIL compiler introduces more state, the
stack. While not represented explicitly, the stack is present in the
compilation of each GHIL expression: compiling a GHIL expression
should leave the runtime value stack in the same state. For example,
if the intermediate value stack has two elements before evaluating an
@code{if} expression, it should have two elements after that

Interested readers are encouraged to read the implementation in
@code{(language ghil compile-glil)} for more details.

@node GLIL
@subsection GLIL

Guile Low Intermediate Language (GLIL) is a structured intermediate
language whose expressions closely mirror the functionality of Guile's
VM instruction set.

Its expression types are defined in @code{(language glil)}, and as
with GHIL, some of its fields parse as rest arguments.

@deftp {Scheme Variable}  nargs nrest nlocs nexts meta . body
A unit of code that at runtime will correspond to a compiled
procedure. @var{nargs} @var{nrest} @var{nlocs}, and @var{nexts}
collectively define the program's arity; see @ref{Compiled
Procedures}, for more information. @var{meta} should be an alist of
properties, as in @code{}. @var{body} is a list of GLIL
@end deftp
@deftp {Scheme Variable}  . vars
An advisory expression that notes a liveness extent for a set of
variables. @var{vars} is a list of @code{(@var{name} @var{type}
@var{index})}, where @var{type} should be either @code{argument},
@code{local}, or @code{external}.

@code{} expressions end up being serialized as part of a
program's metadata and do not form part of a program's code path.
@end deftp
@deftp {Scheme Variable}  vars rest
A multiple-value binding of the values on the stack to @var{vars}. Iff
@var{rest} is true, the last element of @var{vars} will be treated as
a rest argument.

In addition to pushing a binding annotation on the stack, like
@code{}, an expression is emitted at compilation time to
make sure that there are enough values available to bind. See the
notes on @code{truncate-values} in @ref{Procedural Instructions}, for
more information.
@end deftp
@deftp {Scheme Variable} 
Closes the liveness extent of the most recently encountered
@code{} or @code{} expression. As GLIL
expressions are compiled, a parallel stack of live bindings is
maintained; this expression pops off the top element from that stack.

Bindings are written into the program's metadata so that debuggers and
other tools can determine the set of live local variables at a given
offset within a VM program.
@end deftp
@deftp {Scheme Variable}  loc
Records source information for the preceding expression. @var{loc}
should be a vector, @code{#(@var{line} @var{column} @var{filename})}.
@end deftp
@deftp {Scheme Variable} 
Pushes the unspecified value on the stack.
@end deftp
@deftp {Scheme Variable}  obj
Pushes a constant value onto the stack. @var{obj} must be a number,
string, symbol, keyword, boolean, character, or a pair or vector or
list thereof, or the empty list.
@end deftp
@deftp {Scheme Variable}  op index
Accesses an argument on the stack. If @var{op} is @code{ref}, the
argument is pushed onto the stack; if it is @code{set}, the argument
is set from the top value on the stack, which is popped off.
@end deftp
@deftp {Scheme Variable}  op index
Like @code{}, but for local variables. @xref{Stack
Layout}, for more information.
@end deftp
@deftp {Scheme Variable}  op depth index
Accesses a heap-allocated variable, addressed by @var{depth}, the nth
enclosing environment, and @var{index}, the variable's position within
the environment. @var{op} is @code{ref} or @code{set}.
@end deftp
@deftp {Scheme Variable}  op name
Accesses a toplevel variable. @var{op} may be @code{ref}, @code{set},
or @code{define}.
@end deftp
@deftp {Scheme Variable}  op mod name public?
Accesses a variable within a specific module. See
@code{ghil-var-at-module!}, for more information.
@end deftp
@deftp {Scheme Variable}  label
Creates a new label. @var{label} can be any Scheme value, and should
be unique.
@end deftp
@deftp {Scheme Variable}  inst label
Branch to a label. @var{label} should be a @code{}.
@code{inst} is a branching instruction: @code{br-if}, @code{br}, etc.
@end deftp
@deftp {Scheme Variable}  inst nargs
This expression is probably misnamed, as it does not correspond to
function calls. @code{} invokes the VM instruction named
@var{inst}, noting that it is called with @var{nargs} stack arguments.
The arguments should be pushed on the stack already. What happens to
the stack afterwards depends on the instruction.
@end deftp
@deftp {Scheme Variable}  nargs ra
Performs a multiple-value call. @var{ra} is a @code{}
corresponding to the multiple-value return address for the call. See
the notes on @code{mv-call} in @ref{Procedural Instructions}, for more
@end deftp

Users may enter in GLIL at the REPL as well, though there is a bit
more bookkeeping to do. Since GLIL needs the set of variables to be
declared explicitly in a @code{}, GLIL expressions must
be wrapped in a thunk that declares the arity of the expression:

scheme@@(guile-user)> ,language glil
Guile Lowlevel Intermediate Language (GLIL) interpreter 0.3 on Guile 1.9.0
Copyright (C) 2001-2008 Free Software Foundation, Inc.

Enter `,help' for help.
glil@@(guile-user)> (program 0 0 0 0 () (const 3) (call return 0))
@result{} 3
@end example

Just as in all of Guile's compilers, an environment is passed to the
GLIL-to-object code compiler, and one is returned as well, along with
the object code.

@node Object Code
@subsection Object Code

Object code is the serialization of the raw instruction stream of a
program, ready for interpretation by the VM. Procedures related to
object code are defined in the @code{(system vm objcode)} module.

@deffn {Scheme Procedure} objcode? obj
@deffnx {C Function} scm_objcode_p (obj)
Returns @code{#f} iff @var{obj} is object code, @code{#f} otherwise.
@end deffn

@deffn {Scheme Procedure} bytecode->objcode bytecode nlocs nexts
@deffnx {C Function} scm_bytecode_to_objcode (bytecode, nlocs, nexts)
Makes a bytecode object from @var{bytecode}, which should be a
@code{u8vector}. @var{nlocs} and @var{nexts} denote the number of
stack and heap variables to reserve when this objcode is executed.
@end deffn

@deffn {Scheme Variable} load-objcode file
@deffnx {C Function} scm_load_objcode (file)
Load object code from a file named @var{file}. The file will be mapped
into memory via @code{mmap}, so this is a very fast operation.

On disk, object code has an eight-byte cookie prepended to it, so that
we will not execute arbitrary garbage. In addition, two more bytes are
reserved for @var{nlocs} and @var{nexts}.
@end deffn

@deffn {Scheme Variable} objcode->u8vector objcode
@deffnx {C Function} scm_objcode_to_u8vector (objcode)
Copy object code out to a @code{u8vector} for analysis by Scheme. The
ten-byte header is included.
@end deffn

@deffn {Scheme Variable} objcode->program objcode [external='()]
@deffnx {C Function} scm_objcode_to_program (objcode, external)
Load up object code into a Scheme program. The resulting program will
be a thunk that captures closure variables from @var{external}.
@end deffn

Object code from a file may be disassembled at the REPL via the
meta-command @code{,disassemble-file}, abbreviated as @code{,xx}.
Programs may be disassembled via @code{,disassemble}, abbreviated as

Compiling object code to the fake language, @code{value}, is performed
via loading objcode into a program, then executing that thunk with
respect to the compilation environment. Normally the environment
propagates through the compiler transparently, but users may specify
the compilation environment manually as well:

@deffn {Scheme Procedure} make-objcode-env module externals
Make an object code environment. @var{module} should be a Scheme
module, and @var{externals} should be a list of external variables.
@code{#f} is also a valid object code environment.
@end deffn

@node Extending the Compiler
@subsection Extending the Compiler

At this point, we break with the impersonal tone of the rest of the
manual, and make an intervention. Admit it: if you've read this far
into the compiler internals manual, you are a junkie. Perhaps a course
at your university left you unsated, or perhaps you've always harbored
a sublimated desire to hack the holy of computer science holies: a
compiler. Well you're in good company, and in a good position. Guile's
compiler needs your help.

There are many possible avenues for improving Guile's compiler.
Probably the most important improvement, speed-wise, will be some form
of native compilation, both just-in-time and ahead-of-time. This could
be done in many ways. Probably the easiest strategy would be to extend
the compiled procedure structure to include a pointer to a native code
vector, and compile from bytecode to native code at runtime after a
procedure is called a certain number of times.

The name of the game is a profiling-based harvest of the low-hanging
fruit, running programs of interest under a system-level profiler and
determining which improvements would give the most bang for the buck.
There are many well-known efficiency hacks in the literature: Dybvig's
letrec optimization, individual boxing of heap-allocated values (and
then store the boxes on the stack directory), optimized case-lambda
expressions, stack underflow and overflow handlers, etc. Highly
recommended papers: Dybvig's HOCS, Ghuloum's compiler paper.

The compiler also needs help at the top end, enhancing the Scheme that
it knows to also understand R6RS, and adding new high-level compilers:
Emacs Lisp, Lua, JavaScript...

CD: 4ms