Standard ML of New Jersey
Version 110.25, December 6, 1999

/cm/cs/what/smlnj/index.html

Warning

This version is intended for compiler hackers. We are in the midst of substantial structural changes, and this is a snapshot.

Summary:

This version has a variety of changes to CM, MLRISC, the x86 back end, Library, and some crucial bug fixes. In no specific order:

Polyequal

The compiler will now print a warning each time it generates a call to polyequal. Such calls, which are known to be expensive in practice, can most of the time be eliminated using a type constraint (but see bug1534). The warning messages can be turned off using: 
	Compiler.Control.CG.polyEqWarn := false;

Bug 1507

Bug 1507 which easily ranks as the most difficult and mysterious bug yet, is fixed. See the bugs files for the saga of message related to the bug.

comp-lib

The comp-lib directory is gone. Pretty printing is done using the SML/NJ Library version, and most variants of integer sets were replaced with red-black trees. Whatever was not directly supported by the SML/NJ Library was moved to compiler/MiscUtil/library.

MLRISC

MLRISC GC type

MLRISC Code Generator (mlriscGen)

MLRISC Register Allocator

MLRISC clusters/flowgraphs

There is now a new mechanism for incrementally adding/replacing code.

ML-Yacc

The generated initialization code now spills less. The old generated code caused problems on certain architectures (HP).

Intel IA32

This version uses a new register allocator for the x86. Unlike the previous allocator which targets pseudo memory registers and coalesces memory using a two phase scheme, this one directly controls memory coalescing in one single (but more complex) phase. Compilation of the compiler has sped up by 20-24% As a side benefit, the benchmarks are also running faster on the average. Here are the numbers: 
                     Compilation Time        Run time
                  110.24   New   Speedup    110.24  New    Speedup
      barnesHut   4.395   4.257   3.25%     2.985   2.760    8.15%
          boyer  17.417   6.587  164.42%    0.252   0.252    0.00%
   count-graphs   1.710   1.527  12.01%    21.675  21.242    2.04%
            fft   1.090   1.033   5.48%     0.875   0.808    8.25%
    knuthBendix   5.118   3.755  36.31%     0.820   0.788    4.02%
         lexgen   9.780   7.698  27.04%     0.815   0.693   17.55%
           life   1.065   0.940  13.30%     0.110   0.095   15.79%
          logic   2.862   2.532  13.03%     4.777   4.278   11.65%
     mandelbrot   0.120   0.107  12.50%     0.615   0.515   19.42%
         mlyacc  31.312  26.233  19.36%     0.463   0.415   11.65%
        nucleic   5.138   5.222  -1.60%     0.128   0.128    0.00%
  ratio-regions  13.518   5.102  164.98%   96.717  95.265    1.52%
            ray   1.872   1.673  11.85%     3.102   3.133   -1.01%
         simple   8.480   7.127  18.99%     2.723   2.362   15.31%
            tsp   1.387   1.167  18.86%     6.352   6.428   -1.19%
           vliw  29.178  22.705  28.51%     1.595   1.513    5.40%
        Average                  34.27%                      7.41%

Each number represents an average of 6 runs. These numbers are collected on a dual processor (400Mhz Pentium II 512K) Linux box with 1G of memory, and @SMLalloc=256k as the allocation parameter.

The generated code of the new allocator seems to be just as compact as the old one. The heap image sizes generated by the two allocators are identical given the same set of compiler source code as input.

CM

The "sml" command accepts command-line parameters. These parameters must be names of files that contain either ML source code or CM library descriptions. The distinction is made based upon the filename suffix (.sml and .sig = ML source, .cm = CM description). Files are processed left-to-right. ML source code files are "use"d, CM files are either "CM.autoload"ed or given to "CM.make". "CM.autoload" is used until the first occurence of the "-m" flag which switches to "CM.make". The "-a" flag restores the default behavior ("CM.autoload").

CM properly handles the exception that the elaborator throws in cases of semantic errors. As a result, you don't see an unhandled exception at top-level anymore, and CM.keep_going also works.

Filters are treated "correctly" during cutoff recompilation. If you have an ML source that exports two things (say, structures A and B) but one of them gets filtered out by an export filter, then the rest of the system does not get "infected" by changes to the part that it cannot see. So if B is filtered away, then for the rest of the system, your source really looks like it only exports A. Modifications to B do no longer lead to recompilation of other modules if they cannot actually "see" B because of some export filter.

This behavior was always the intended one, but it has never actually worked so far (not under the old CM and not under the new one). The fix actually required some tweaking of the pickler.

Structure CM re-shaped and distinction between "minimal-cm.cm" and "full-cm.cm": The new "full" structure CM is cleaned up quite a bit and uses sub-structures to organize things. This will probably be fleshed out more in the future. Full documentation of this can be found at: 

       http://www.kurims.kyoto-u.ac.jp/~blume/SMLNJ-DEV/manual.ps or
       http://www.kurims.kyoto-u.ac.jp/~blume/SMLNJ-DEV/manual/index.html
The library host-cm.cm has been eliminated. Its place is taken by two _new_ libraries: full-cm.cm and minimal-cm.cm. The latter is initially registered at toplevel.

CM's master-slave protocol made more robust against unexpected failures: This mainly robustifies CM's master-slave protocol against unexpected deaths of slave processes. This is important because rsh tends to die when you press ^C -- at least on some systems. CM will now gracefully continue to utilize remaining compile servers in the event of some other servers going down. If the last servers disappears, operation will automatically switch back to non-parallel mode.

CM's internal cache logic cleaned up: It will also avoid spurious error messages about files belonging to multiple groups.

CM's "Tools" structure (library "cm-tools.cm") cleaned up and made more flexible.

Pre-loading of libraries during bootstrap can now be customized. For "makeml", the relevant file is in src/system/preloads.standard; for "install.sh" it is "config/preloads".

Library-related changes

Reorganization of the libraries that make up the compiler viscomp-lib.cm was split into a core part (viscomp-core.cm) and various machine-dependent siblings (viscomp-.cm). The dependence of host-cmb.cm on target-compilers.cm is eliminated so that you don't suck in stuff you don't need when you load it. This cleans up some of the .cm files quite nicely (all that "#if" goo is gone except in one place where structures CM and CMB are built. Also, the "LIGHT" variable is treated quite nicely in just one place. On the downside (if you consider this a downside), there are quite a few additional .cm files (and libraries). In particular, there is now one library for each architecture (called -compiler.cm) and one library for each CMB (-.cm). As a result, you can re-target either summarily by loading target-compilers.cm (which is just the sum of the above) or by loading precisely the one CMB or the one Compiler that you need.

All libraries except (viscomp-* and MLRISC-*) updated to eliminate all incomplete match warnings during compilation.

Installation scripts

The scripts "makeml" and "installml" are modified to automatically patch relevant "pathconfig" files.

Default "bin" dir is now called sml.bin.-; default "boot" dir is now called sml.boot.-. Scripts are updated accordingly. (This seems to be more consistent with other naming conventions.)

The "-bare" option to "makeml" works again. Pre-loading of libraries into a "bare" system can be controlled using the src/system/preloads.bare file.

Lal George
Last modified: Mon Dec 6 13:59:56 EST 1999