Home My Page Projects Code Snippets Project Openings SML/NJ
Summary Activity Forums Tracker Lists Tasks Docs Surveys News SCM Files

SCM Repository

[smlnj] View of /sml/trunk/src/cm/Doc/bt-03-diffs.tex
ViewVC logotype

View of /sml/trunk/src/cm/Doc/bt-03-diffs.tex

Parent Directory Parent Directory | Revision Log Revision Log


Revision 840 - (download) (as text) (annotate)
Fri Jun 15 19:05:19 2001 UTC (18 years, 11 months ago) by blume
File size: 8927 byte(s)
c-calls (and NLFFI) implementation for Sparc
% -*- latex -*-

\section{Differences between CMB and CM}

In this section we discuss why compiling the compiler is different
from compiling other ML programs.  Each of the following sub-sections
focuses on one particular aspect.

\subsection{Code sharing}

CM keeps compiled code within the same directory tree that contains
the corresponding ML source code.  Thus, there is a fixed function
that maps the names of source files to the names of corresponding
binfiles and the names of CM description files (for libraries) to
their corresponding stable files.

As a result, these files will be shared between programs that use the
same libraries.  Moreover, CM will let different programs that are
loaded into an interactive session at the same time share their
in-memory copies of common compiled modules.  (There is also an issue of
state-sharing, but this does not concern CMB because the bootstrap
compiler only compiles code without linking it.)

Sharing of code is useful for ordinary usage, but when compiling the
compiler itself, it is not desirable.  During bootstrap compilation,
it is often the case that several different versions of compiled code
have to coexist.  Some, or even all of these versions can differ
significantly from those of the currently running system.

Therefore, CMB keeps binfiles and stable files in separate directory
trees.  The names of the directories where these trees are rooted at
are constructed from three parts; the binfile directory's name is {\tt
$u$.bin.{\it arch}-{\it os}} and the stablefile directory's name is
{\tt $u$.boot.{\it arch}-{\it os}}.  As mentioned before, {\it arch}
is a string describing the current CPU architecture and {\it os} is a
string describing the current operating system kind. Component $u$ is
a string that can be selected freely when the bootstrap compiler is
invoked.  When using {\tt CMB.make} it defaults to {\tt sml},
otherwise it is the argument given to {\tt CMB.make'}.  (The $u$
component is kept variable to make it possible to keep and use several
compiled versions of the system at the same time.)

The auxiliary script {\tt makeml} (which is responsible for
bootstrapping a new system) also accepts a parameters to select $u$.
If the parameter is missing, it defaults to {\tt sml} (in accordance
with {\tt CMB.make}'s behavior).

\subsection{Init library}

The {\em init library} ({\tt \$smlnj/internal/init.cmi}) is a library
that is used implicitly by all programs.  This library is ``special''
in several ways and cannot be described using an ordinary CM
description file.  It is the bootstrap compiler's responsibility to
properly prepare a stable version.

Ordinary programs (those managed by CM) do not have to worry about the
special aspects of how to construct this library; they just have to be
able to use its stable version.

There are several reasons why the library cannot be described as an
ordinary CM library:

\begin{itemize}
\item The library exports the {\em pervasive environment} which
normally is imported implicitly by every compilation unit.  Within the
init library, no pervasive environment is available yet.
\item One binding in the above-mentioned pervasive environment is a
binding for {\tt structure \_Core}.  The symbol {\tt \_Core} is not a
legal SML identifier, and the bootstrap compiler has to take special
action to create a binding for it anyway.
\item One of the compilation units in this library is merely a
placeholder which at link time has to be replaced by the SML/NJ
runtime system (which is written in C).
\end{itemize}

\subsubsection{Linkage to runtime system}

The ML source file {\tt dummy.sml} (located in directory {\tt
src/system/smlnj/init}) contains a carefully constructed module whose
signature matches that of the runtime system's binary API.  This file
is being compiled as part of constructing the init library, but it has
been marked specially as {\em runtime system placeholder}.  During
compilation, this file pretends to {\em be} the runtime system; other
modules that use the runtime system ``think'' they are using {\tt
dummy.sml}.

At link time (i.e., bootstrap time---when {\tt makeml} is run), the
boot loader will ignore {\tt dummy.sml} and use the actual runtime
system in its place.  This trick makes it possible that (from the
point of view of all other modules) using services from the runtime
system appears to be no different than using services from an ordinary
ML compilation unit.

\subsection{BOOTLIST file}

Linking of SML/NJ programs involves executing the code of each of the
concerned compilation units.  The code of each compilation unit is
technically a closed function; all its imports have been turned into
arguments and all exports have been turned into return values.

For ordinary programs, this process is under control of CM; CM will
take care of properly passing the exports of one compilation unit to
the imports of the next.

When booting a stand-alone program, though, there is no CM available
yet.  Thus, executing module-level code and passing exports to imports
has to be done by the (bare) runtime system.  The runtime system
understands enough about the layout of binfiles and library files so
that it can do that---provided there is a special {\em bootlist}
file that contains instructions about which modules to load in what
order.

The bootlist mechanism is not restricted to building SML/NJ.  Ordinary
ML code can also be turned into stand-alone programs, and as far as the
runtime system is concerned, the mechanisms are the same.  The
bootlist file used by such ordinary stand-alone ML programs will be
constructed by CM; only in the case of bootstrapping SML/NJ itself it
will be constructed by CMB.

The name of the bootlist file is {\tt BOOTLIST}, and it is located at
the root of the directory tree that contains stable files (i.e., its
name is {\tt $u$.boot.{\it arch}-{\it os}/BOOTLIST}).

\subsection{PIDMAP file}

The last file to be loaded by the bootstrap process contains
module-level code which will trigger the self-initialization of the
interactive system---including CM.  One job of CM is to manage sharing
of link-time state (i.e., dynamic state created by module-level code
at link time).  Link-time state of a module used by the interactive
system should be shared with any program using the same module.  The
file {\tt $u$.boot.{\it arch}-{\it os}/PIDMAP} contains information
that enables CM to relate existing link-time state to particular
library modules and also to identify any link-time state that will
never be shared and which can therefore be dropped.  It is CMB's
responsibility to construct the {\tt PIDMAP} file.

\subsection{Cross-compiling}

Several different versions of the bootstrap compiler can
coexist---each being responsible for targeting another CPU-OS
combination.  Structure {\tt CMB} is the default bootstrap compiler
that targets the current system; it is exported from {\tt
\$smlnj/cmb.cm}.  The following table lists the names of other
structures---those corresponding to various cross-compilers.  All
these structures share the same signature.

The table also shows the names of libraries that the structures are
exported from as well as those {\it arch} and {\it os} strings that
are used to name binfile- and stablefile-directory.

\begin{small}
\begin{center}
\begin{tabular}{p{2.2in}||p{1.5in}|l|l|l|l}
library & structure & architecture & OS & {\it arch} & {\it os} \\
\hline\hline
{\tt \$smlnj/cmb.cm} \newline
{\tt \$smlnj/cmb/current.cm} & {\tt CMB} & current & current & & \\
\hline\hline
{\tt \$smlnj/cmb/alpha32-unix.cm} & {\tt Alpha32UnixCMB} &
  Alpha & Unix & {\tt alpha32} & {\tt unix} \\
\hline
{\tt \$smlnj/cmb/hppa-unix.cm} & {\tt HPPAUnixCMB} &
  HP-PA & Unix & {\tt hppa} & {\tt unix} \\
\hline
{\tt \$smlnj/cmb/ppc-macos.cm} & {\tt PPCMacOSCMB} &
  Power-PC & Mac-OS & {\tt ppc} & {\tt macos} \\
\hline
{\tt \$smlnj/cmb/ppc-unix.cm} & {\tt PPCUnixCMB} &
  Power-PC & Unix & {\tt ppc} & {\tt unix} \\
\hline
{\tt \$smlnj/cmb/sparc-unix.cm} & {\tt SparcUnixCMB} &
  Sparc & Unix & {\tt sparc} & {\tt unix} \\
\hline
{\tt \$smlnj/cmb/x86-unix.cm} & {\tt X86UnixCMB} &
  Intel x86 & Unix & {\tt x86} & {\tt unix} \\
\hline
{\tt \$smlnj/cmb/x86-win32.cm} & {\tt X86Win32CMB} &
  Intel x86 & Win32 & {\tt x86} & {\tt win32} \\
\hline\hline
{\tt \$smlnj/cmb/all.cm} & all of the above (except {\tt CMB}) & & & & \\
\end{tabular}
\end{center}
\end{small}

As an example, consider targeting a Sparc/Unix system.  The first step
is to load the library that exports the corresponding cross-compiler:

\begin{verbatim}
  CM.autoload "$smlnj/cmb/sparc-unix.cm";
\end{verbatim}
% $

Once this is done, run the equivalent of {\tt CMB.make}:

\begin{verbatim}
  SparcUnixCMB.make ();
\end{verbatim}

This will recompile the compiler, producing object code for a Sparc.
Binfiles will be stored under {\tt sml.bin.sparc-unix} and stable
libraries under {\tt sml.boot.sparc-unix}.

root@smlnj-gforge.cs.uchicago.edu
ViewVC Help
Powered by ViewVC 1.0.0