V10/cmd/sml/doc/NEWS
Standard ML of New Jersey version 0.44, 4 December 1989
New release:
Starting with this release, the compiler source code is covered by an AT&T
copyright notice, which is included in each source code file. The file
doc/COPYRIGHT contains the official copyright notice, license, and
disclaimer. The file doc/release-form contains a release form that our
lawyer would like you to sign if you contribute software or modifications
that are included in the distribution. Whether you sign it as an
individual or get your institution to sign it is your decision, but it is
undoubtedly quicker and simpler to sign as an individual. The release form
is not necessary for simple bug fixes.
Mailing list:
There is now a Standard ML mailing list for messages and queries of general
interest to the ML community. You can join by sending a request to
sml-request@cs.cmu.edu, and you can send a message to the list by mailing
to sml@cs.cmu.edu. This mailing list will be the main vehicle for
distributing announcements about Standard ML of New Jersey.
Modules:
All known major bugs in the module system have been fixed, so please report
any new bugs or old bugs we missed. The full module syntax from the
Definition, Version 3 is now implemented with minor limitations or
extensions, including let structure expressions and local and open
signature specs (see comments in doc/localspec). Include specs work
properly once again.
Local and open specs:
David MacQueen has some comments about "local" and "open" specifications
in signatures; see doc/localspec.
Weak type variable syntax:
Standard ML of New Jersey has a more precise system for weak typing
(i.e. typechecking of programs that use semi-polymorphic reference
variables). Other compilers denote weak type variables with an underscore,
i.e. '_abc, while ours uses an initial integer to denote level of weakness,
i.e. '2abc. See doc/weak-types for a fuller explanation. However,
for compatibility:
The notation for weak type variables using an initial underscore is now
supported for input and output. A weak type variable of the form '_a
is equivalent to the numeric form '1a. On output, weak type variables
are printed using an initial integer. However, if the ref variable
System.Control.weakUnderscore is set to true, then weak type variables
of all levels are printed with an underscore.
Equality types:
Equality types are still broken. Fixing these is the next order of business.
Import:
Nested imports do not work reliably in this version.
Optimizing compiler:
Hints on making your programs run faster are in doc/optimize.
Miscellaneous:
The SPARC code generator has been completely rewritten by John Reppy.
Bruce Duba continues to make improvements to the CPS code generator.
Andrew Appel is re-writing the parser to use David Tarditi's newest
version of the ML parser generator; this re-write is not complete but
the source files are included (along with the working recursive descent
parser) in src/parse/.
Linkdata:
The "linkdata" program, used for making a sharable version of the ML
compiler by putting all the .mo files into the text segment, did not work
on all operating systems. There is a more portable, albeit much less
efficient, version of linkdata called "slinkdata". Look in "src/makeml" to
see how it is used for the NeXT operating system, and you can use it that
way for your operating system too. (Admittedly, this is very minimal
documentation!)
Bug fixes:
See latest doc/bugs file for remaining bugs. This has been brought more or
less up to date. We appreciate your bug reports and hope you will be
patient as we work our way through them. There is a new version of the
bug report format file (doc/bug.form), including a new "Severity:" field
that you should use to indicate how serious the bug is from your point of
view. We get a lot of bug reports (which is good!) but we may not assign
the right priority to them without your help.
Version 1 plan:
Our tentative schedule calls for the release of version 1 of the compiler in
mid-January 1990. There will probably be one more beta release between now
and then. This is a more or less arbitrary deadline, and we will try to fix
as many bugs as possible before version 1 is released. Development of the
compiler will continue as usual following this milestone, and major, widely
publicized releases will occur on a more regular and less frequent schedule,
e.g. perhaps three times a year. Interim developmental versions can be made
available to those who have a special need for them.
------------------------------------------------------------------------------
Standard ML of New Jersey version 0.42, 27 October 1989
Release notes by Andrew Appel, covering changes since version 0.39.
Bug fixes:
Large floating point literals do the right thing now, at least on the Vax.
Duplicate constructor names and field names are again detected.
Printing of records like {1="abc"} works correctly.
Reading 1025 bytes from standard input works.
Large list expressions of the form [(0,0),(0,0),(0,0), ..... (0,0)] no longer
cause the compiler to blow up.
Previous versions have run on Sun-3s with Mach, this version might work
on Vaxes with Mach. The next version should run on anything with Mach.
Compile-time processing of functors has been extensively rewritten,
and should be more robust. There are still some remaining bugs, however,
which should now be easy to iron out.
The Io exception now carries a much more useful string value, containing
all sorts of information.
Other changes since version 0.39 are mostly of interest to implementors.
VALvars and STRvars now have qualids instead of ids, lexer interface
parametrized more cleanly. Page-faulting is no longer used to detect
end-of-memory on most machines.
---------------------------------------------------------------------------
Standard ML of New Jersey version 0.39, 8 September 1989
This notes overlap with the 0.36 release notes, since 0.36 was not
widely distributed [neither were 0.37 or 0.38].
Release notes by Andrew Appel.
Changes since version 0.33:
1. Support for Sun-4 (SPARC) thanks to John Reppy at Cornell;
Encore (NS32032), NeXT (also Mach on other 68020 machines), and
Sony NEWS (68020-based workstation), thanks to James O'Toole at MIT.
2. An interpreter for faster compilation (4x) but slower execution (18x).
Set System.Control.interp := true to enable this mode. Use the -i
or -ionly to makeml to set this option as the default.
3. Delay and Force primitives (meant for internal use only) for research
in lazy evaluation.
4. Bug fixes: reading 2049-byte files works now. Separate compilation
shouldn't dump core. Miscellaneous others.
5. A code generator generator implemented by Johan Lilius,
Jari Porra, & Jussi Rintanen of Helsinki University of Technology
is include in the Standard ML Software Library: see lib/twig.
6. Various internal efficiency improvements; the CMACHINE signature was
simplified a bit; the format of generational-gc update lists was changed;
the CPS optimizer was beefed up.
7. Separate compilation has been improved. The compiled version of a
separately compiled file has the extension ".bin", and internally
identifies the version of SML-NJ and also the machine type it
runs on. The command "head -1 foo.bin" will print this information.
8. Many functor bugs have been fixed.
9. An improved version of ML-Yacc by David Tarditi at Princeton is provided,
more efficient and with a better interface, and with sophisticated syntax
error recovery.
Forthcoming in future releases:
- Mach support for Vax and other machines.
- Some work has been done on a MacIntosh II port, and we are hoping for
volunteers to continue this work.
- Our MIPS code generator almost works now (thanks to Norman Ramsey at
Princeton).
-----------------------------------------------------------------------------
What's new in SML of NJ Version 0.33, 1 April 1989
1. CPS code generator
This version uses a new CPS (Continuation Passing Style) code
generator. The CPS code generator differs from the older code
generator in the following ways:
(i) The code produced is about 50% faster.
(ii) The compiler generates code about half as quickly (this should
gradually improve with future releases).
(iii) the new code generator does not use a stack. This makes it easier
to support experimental features like callcc and multiple processes.
It also makes efficient garbage collection even more critical.
(iv) The cps code generator tends to build larger intermediate
representations, so tuning of the runtime system becomes more
important. See the discussion of memory management in doc/INSTALL.
2. The runtime system
The runtime system has been extensively rewritten and cleaned up. The
amount of assembly code has been reduced (only 220 lines for the Vax,
385 for the Sun), and the system call interface has been simplified. The
mechanism for initiating garbage collection has been made more robust.
In the m68 version (for Sun 3), exhaustion of free space is detected
by an explicit comparison rather than by trapping a segmentation fault.
This finally eliminates persistent flakiness caused by the way the 68020
recovers from segmentation faults.
LaTeX source for a new paper describing the runtime system can be
found in doc/papers/runtime. This is still an early draft, but it should
be helpful in understanding internals of the run time system.
3. The pervasive environment
The initial environment has been reorganized to make it possible to
build a system whose pervasives (i.e., built-in bindings) correspond
closely to those described in "The Definition of Standard ML, Version
2" by R. Harper, R. Milner, and M. Tofte (ECS-LFCS-88-62, August 1988,
hereafter referred to as the "Definition"). Note that in our humble
opinion the "standard" environment is somewhat brain-damaged and is
missing many important features (like arrays, bytearrays, etc). See
the file doc/environ for discussion of specific deficiencies of the
environment described in the Definition and our proposed fixes, some
of which are incorporated in the "New Jersey" environment.
The default initial environment set up by the bootstrapping process is
the richer, and more efficient, New Jersey environment. However,
loading the source file src/lib/standard.sml defines a structure named
Standard that provides a close approximation to the "standard"
environment. Opening the structure Standard at the top level replaces
any nonstandard bindings in the New Jersey environment. The New
Jersey environment is still there in the background, so any New Jersey
bindings not rebound in Standard will still be visible. In the future
we may provide a mechanism (e.g. a function switchEnv: string -> unit)
for completely replacing the base environment with a different version
specified by a structure. For example, `switchEnv "Standard"' would
replace the existing base environment with the contents of Standard.
The initial environment also contains:
(1) The signatures ARRAY, BASICIO, BITS, BOOL, BYTEARRAY, GENERAL, IO,
INTEGER, LIST, REF, STRING, and REAL as defined in src/boot/perv.sig,
and the signature SYSTEM as defined in src/boot/system.sig
(2) The structures Array : ARRAY, BasicIO : BASICIO, Bits : BITS, etc.
These contain the New Jersey versions of bindings.
(3) The structure NewJersey, which contains the complete New Jersey initial
environment except for the signatures in (1) above and the structure
NewJersey itself.
It follows that even if Standard has been defined and opened, all New Jersey
bindings are available via the structure NewJersey, or alternatively via
the structures Array, IO, etc.
Here are some differences between the New Jersey and Standard environments:
a. The functions +, /, *, -, div, mod are an order of magnitude
slower in Standard than in the New Jersey environment. This is
because they are required to raise distinct exceptions (according
to the Definition), and because the treatment of negative arguments
of div and mod doesn't match what the hardware provides; that
complicates their implementation. The NewJersey functions all raise
Overflow on results that won't fit in 31 bits (and Div on zero divisor
for div and mod).
b. The functions div and mod return different results.
c. input and output are curried in NewJersey (and in the IO structure),
but not in Standard.
d. The New Jersey environment has many more functions in it, as described
in the appendix of our reference manual or the signatures in the files
src/boot/perv.sig and src/boot/system.sig.
The New Jersey environment has changed slightly from previous releases,
mostly to make the it more compatible with the Definition [IMPORTANT --
THESE CHANGES MAY BREAK SOME PROGRAMS]:
a. The exception Io_failure has been renamed Io for consistency
with the Definition.
b. The exception Float has been renamed Real. Float is still defined
as a synonym for Real.
c. The identifier "length" is no longer overloaded: use "size" or
"String.size" for string lengths, "length" or "List.length" for
list lengths, and "Array.length" for array lengths.
d. The functions "cd" and "system" are now components of the
System structure.
e. Various components of the System structure have been moved
into the System.Unsafe substructure, and are not recommended for
casual use.
f. The print functions (Integer.print, Real.print, String.print,
and the overloaded print), now return unit rather than their
argument. This is because we found that in our code virtually
every call of print was followed by "; ()".
We plan to provide a definition of the NewJersey structure in terms of
the "standard" environment (analogous to the definition of Standard in
lib/src/standard.sml). This structure would enable any Standard ML
system to run programs based on the New Jersey environment. Such
programs should not reference the System structure, of course.
4. Command line arguments and Shell environment
System.argv : unit -> string list
System.environ : unit -> string list
are new functions that return the list of command line arguments and
the Unix shell environment, respectively.
5. First-class continuations: callcc, throw, and cont (experimental)
These are new experimental primitives to allow one to capture and use
"first-class continuations". See the file doc/callcc for
documentation of these new facilities.
6. Separate compilation (experimental)
We have integrated code written by Nick Rothwell and Mads Tofte of
Edinburgh that provides a preliminary, experimental form of secure (type
checked) separate compilation. There is a special declaration form:
import "filename"
where "filename" is a string literal designating either a source file
filename.sml or a binary file filename.bin. If only the source file
exists in the current directory, then it is compiled and the binary
version is created. If the binary version exists, it is loaded
instead of the source file.
As a consequence of adding this new facility, the identifier "import"
has become a reserved word. This may cause incompatibilities with
existing code.
7. mlyacc
The new directory src/lib/mlyacc contains a yacc-like parser generator
written in ML, with documentation. The parser generator is used much
like the Unix yacc utility, and the grammar specification language is
essentially that of yacc. The parser generator was written by David
Tarditi of Princeton, based on an earlier effort by Nick Rothwell.
Other, more sophisticated, interfaces to the parser generator will
probably be provided in future releases. As before, the directory
src/lib/lexgen contains a lexical analyzer generator.
8. Fixes
A number of things that have been broken are now fixed. This includes
the garbage collector for the Sun 3, floating point operations on the
Sun 3, exportFn, sqrt, and execute. Unfortunately, many old bugs
remain and the error messages are just as bad as ever. We expect the
next release to fix many of these problems. Bug reports are welcome as
ever, as are bug fixes.
9. Things to come
John Reppy has ported SML of NJ to the Sun 4 (SPARC) architecture and
his code generator will be included in the next release. A port to
the MIPS architecture (DECstation 3100 under Ultrix) is in the works
at Princeton, and a NeXT port is likely. A port to the Mac II under
the native Mac OS may also appear soon.
We are working on a new licensing scheme similar to that used by the
Free Software Foundation. We hope to have this approved within three
months, and we will keep you informed of new developments.
------------------------------------------------------------------------------
New features of SML of NJ Version 0.24 (November 22, 1988)
1. Profiler
A call-count and statistical profiler has been implemented. See
operating instructions in the README file, and the paper provided in
src/doc/papers/profiling.
2. Runtime reorganization
The runtime system has been reorganized to reduce the need for assembly
language and provide a standard interface for Unix system calls and
C code. It should become possible to add special purpose C code to
the runtime system, for instance to provide an X windows interface.
Many of the assembly language functions have been replaced by more
portible C implementations and uses of the system call interface.
See src/doc/VERSION.24 for some more details.
3. Memory Management tuning
The variable
System.Control.Runtime.ratio : int ref
can be used to control the garbage collection behavior to some extent,
in conjunction with the Unix limit variable (in the csh). See the
discussion in README.
4. Simple gnu emacs sml mode
A simple gnu emacs sml mode was contributed by Mitch Wand. It can be
found in src/lib/sml.el. If you have a better mode or make improvements
on this version, please send them in to macqueen@research.att.com for
inclusion in future releases.
--------------------------------------------------------------------------
New features of SML of NJ Version 0.22 (October 10, 1988)
This file describes the most notable changes since version 0.20. (Version 0.21
was a local version that was not distributed.)
1. Bugs
A few bugs have been fixed; exportFn works again; the system runs under the
latest version (4.0) of Sun Unix on a Sun-3. See src/doc/bugs for the
remaining known bugs. The next release with many more bug fixes will
probably follow fairly shortly.
2. Sharable text segments
We now have an option to run sml with the compiler read-only and sharable
(in the Unix text segment). This should help on systems where several
people are running sml at once, and also improves garbage-collector performance.
See src/doc/INSTALL for more info.
3. CPS code generator
We are including the source code for the new cps code generator with
this release for those who would like an early look at it. This code
generator is still under development and works only on Vaxes at this
point.
4. Installation
The maker script for building the interactive image has been reorganized
and new scripts maker.batch and maker.run have been added to simplify
making batch systems and compiling the runtime system. See src/doc/INSTALL
and src/doc/BATCHINSTALL for details.
5. Lexer generator
The directory src/lib is a new directory for contributed software.
The programs included here do not come under the terms of the AT&T
license. The first contributed software is a lexical analyzer
generator similar to the Unix lex utility. It is in src/lib/lexgen.
lexgen has been used to generate a new lexical analyzer for the
compiler. The specification is in src/lex/ml.lex.
----------------------------------------------------------------------------
New features of SML of NJ Version 0.20
This file describes the most notable changes since version 0.18. (Version 0.19
was a local version that was not distributed.)
1. Bugs
Quite a few bugs have been fixed, mostly relating to type checking and modules.
See src/doc/bugs for the remaining known bugs.
2. Reference types
The typing of references and arrays is less restrictive. For
instance, the following is now permitted, whereas it used to cause an
error message:
let val r = ref [] in r := [1]; !r end;
Thus locally declared reference variables like r in this example do not
need to be explicitly constrained to have a monotype.
3. Equality types
Equality types are now implemented. Equality types are those
"hereditarily concrete" types whose values admit structural equality.
The implementation of equality types involves the new signature
specification keyword "eqtype" indicating that the specified type
constructor admits equality, and special equality type variables
starting with double single quotes (e.g. ''a) that are restricted to
range over equality types.
An equality type is one containing only type constructors that "support
equality". The following type constructors support equality:
int, string, real
any datatype whose constuctor argument types are equality types
(disregarding parameter type variables)
record type constructors
type constructors from functor parameters that are specified as eqtypes
In addition, any reference or array type is an equality type.
Abstract type constructors, "->", and exn do not support equality.
4. Include specifications in signatures
Include specifications allow one to include the specifications of a previously
defined signature in a new signature. The syntax is
include SIG
where SIG is the name of a signature. For example, given the declaration
signature AA =
sig
type t
val x : t
end
the declaration
signature BB =
sig
include AA
val y : t -> int
end
is equivalent to
signature BB =
sig
type t
val x : t
val y : t -> int
end
5. Free identifiers in signatures and structures
Structure definitions can now contain free identifiers of any sort
(values, constructors, exceptions, types), not just free signatures,
structures and functors. Signature definitions can contain free
structure and type identifiers, as well as free signature identifiers.
6. Printing of signatures and structures
Signatures defined at top level are printed. Structures defined at
top level have their signatures printed (currently in full -- in the
future only the name of the signature will be printed if it has one).
Printing of signatures can be turned off by setting the flag
System.Control.Print.signatures to false.
7. Memory management improvements
The runtime system manages memory better. The system automatically
resizes the heap on startup and during collections, allowing the
system to perform more gracefully on systems with limited memory.
The behavior and type of the flag controlling gc messages has changed:
System.Control.Runtime.gcmessages: int ref
when 0, no messages are printed
when 1, only major collections are reported
when 2, major collections and heap resizings are reported
when 3, minor and major collections and heap resizings are reported