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/READMES/110.27-README
ViewVC logotype

View of /sml/trunk/READMES/110.27-README

Parent Directory Parent Directory | Revision Log Revision Log


Revision 619 - (download) (annotate)
Wed Apr 19 14:55:09 2000 UTC (19 years, 5 months ago) by dbm
File size: 26731 byte(s)
summary edited; distribution files listed
    
			S  M  L   /   N  J

                  1  1  0  .  2  7      N  E  W  S
			
  		           April 10, 2000

			      WARNING

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

        http://cm.bell-labs.com/cm/cs/what/smlnj/index.html

------------------------------------------------------------------------ 
Summary:

 * This version has some minor tweeks to FLINT (after the major merge
   in 110.26).  Work continues on tuning FLINT and the various optimizations
   it implements.

 * CM has been revised extensively, and the modmap environment mechanism
   supporting stubbified pickles has been reworked completely.  The pathconfig
   file has been simplified.  Installation scripts have been further
   modified.  See src/system/README and the latest version of the
   CM manual at 

    <//http://www.kurims.kyoto-u.ac.jp/~blume/SMLNJ-DEV/manual/index.html>
    <//http://www.kurims.kyoto-u.ac.jp/~blume/SMLNJ-DEV/manual.ps>

   for further information about these changes.

 * MLRISC, and particularly the x86 back end have been modified extensively.

 * There are a few updates to the SML/NJ Library

 * Reported bug fixes:
     1556. (jhr) signal race condition
     Some CM bugs (not recorded)

 * Distribution file names have been simplified.  They no longer start
   with the version number (e.g. "110.27-config.tar.gz" is now
   simply "config.tar.gz").  The boot directory tarballs are now
   "boot.alpha32-unix.tar.gz", etc. (i.e. no version number and the
   "sml." prefix is dropped).  The new install script will restore
   the usual name (e.g. "sml.boot.alpha32-unix" when the tarball is
   unpacked.  [We dropped the initial "sml." for the boot tarballs to
   get the file names under 28 characters because of a limitation of
   the Bell Labs ftp server.]
   The version README file is still named 110.27-README, however.

     110.27-README
     HISTORY
     MLRISC.tar.gz
     boot.alpha32-unix.tar.gz
     boot.hppa-unix.tar.gz
     boot.ppc-unix.tar.gz
     boot.sparc-unix.tar.gz
     boot.x86-unix.tar.gz
     cm.tar.gz
     compiler.tar.gz
     config.tar.gz
     ml-burg.tar.gz
     ml-lex.tar.gz
     ml-yacc.tar.gz
     runtime.tar.gz
     smlnj-lib.tar.gz
     system.tar.gz

======================================================================
Details of changes


======================================================================
FLINT:
======================================================================

Name: Stefan
Date: 2000/04/07 10:00:00 EDT
Tag: monnier-20000406-branch-handling
Description:

Improved handling of branches (mostly those generated from
polymorphic equality), removed switchoff and changed the
default optimization settings (more cpsopt and less flintopt).


======================================================================
MLRISC:
======================================================================

1. Register Allocator

   a. The interface and implementation of the register allocator have been 
      changed slightly to accommodate the possibility of skipping 
      the register allocation phases completely and go directly to 
      memory allocation.  This is needed for C-- use.

   b. I've improved the spill propagation algorithm, using an approximation
      of maximal weighted independent sets.     This affects only the x86
      platform. 

2. MLTREE

   a. Renamed the constructor CALL in MLTREE by popular demand.

3. X86

   a. More assembly output problems involving the indexed addressing mode
      on the x86 have been found and corrected. Thanks to Fermin Reig for the
      fix.

   b.  x86Rewrite bug with MUL3 (found by Lal)

   c.  Added the instructions FSTS, FSTL

   d.  The old code generated for SETcc was completely wrong.  
       The Intel optimization guide is VERY misleading.

   e.  Various fixes related floating point, and extensions.

   f.  Things like

       jmp %eax
       jmp (%eax)

       are now output as

       jmp *%eax
       jmp *(%eax)

    g.  Yet another fix for x86 assembly for idivl, imull, mull and friends.

    h.  I've changed andl to testl in the floating point test sequence
        whenever appropriate.  The Intel optimization guide states that
        testl is perferable to andl.

4. Alpha 

    a. Some extra patterns related to loads with signed/zero extension
       provided by Fermin.
    b.  Added the instructions LDBU, LDWU, STB, STW as per Fermin's suggestion.
    c.  Added a new mode byteWordLoadStores to the functor parameter to Alpha()
    d.  Added reassociation code for address computation.

5. PA-RISC 

   a.  B label should not be a delay slot candidate!  Why did this work?
   b.  ADDT(32, REG(32, r), LI n) now generates one instruction instead of two,
       as it should be.
   c.  The assembly syntax for fstds and fstdd was wrong.
   d.  Added the composite instruction COMICLR/LDO, which is the immediate
       operand variant of COMCLR/LDO.
   e.  Long jumps in span dependence resolution used to depend on the existence
       of the base pointer in the SML/NJ runtime.  

        A jump to a long label L was expanded into the following sequence:
     
      LDIL %hi(L-8192), %r29
      LDO  %lo(L-8192)(%r29), %r29
      ADD  %r29, baseptr, %r29
      BV,n %r0(%r29)

         I've changed it so that the following sequence of instructions 
       are generated, which doesn't mention the base pointer at all:

         BL,n  L', %r29           /* branch and link, L' + 4 -> %r29 */
    L':  ADDIL L-(L'+4), %r29     /* Compute address of L */ 
         BV,n  %r0(%r29)          /* Jump */ 

6. Generic MLRISC

   a.  shuffle.sml rewritten to be slightly more efficient
   b.  DIV bug in mltree-simplify fixed (found by Fermin)

7. Assembly Output

    a.  When generating assemby, resolve the value of client defined constants, 
        instead of generating symbolic values.  This is controlled by the 
        new flag "asm-resolve-constants", which is default to true.

    b.  Added a new flag 

          "asm-indent-copies" (default to false)

       When this flag is on, parallel copies will be indented an extra level.


8. Machine Descriptions/Generation

    a. The precedence parser was slightly broken when parsing infixr symbols.
    b. The type generalizing code had the bound variables reversed, resulting
       in a problem during arity raising. 
    c. Various fixes in machine descriptions.

======================================================================
CPS->MLRISC Code Generation
======================================================================

   This release contains *MAJOR* changes to the way code is generated from CPS
   in the module mlriscGen, and in various backend modules.

1. Forward propagation fix.

   There was a bug in forward propagation introduced at about the same time
   as the MLRISC x86 backend, which prohibits coalescing to be 
   performed effectively in loops. 

   Effect: speed up of loops in RISC architectures.
           By itself, this actually slowed down certain benchmarks on the x86.

2. Forward propagating addresses from consing.

   I've changed the way consing code is generated.  Basically I separated 
   out the initialization part:

        store tag,   offset(allocptr)
        store elem1, offset+4(allocptr)
        store elem2, offset+8(allocptr)
        ...
        store elemn, offset+4n(allocptr)

   and the address computation part:

        celladdr <- offset+4+alloctpr

   and move the address computation part

   Effect:  register pressure is generally lower as a result.  This 
            makes compilation of certain expressions much faster, such as
            long lists with non-trivial elements.

             [(0,0), (0,0), .... (0,0)]

3. Base pointer elimination.

    As part of the linkage mechanism, we generate the sequence:

     L:  ...  <- start of the code fragment
 
     L1:
         base pointer <- linkreg - L1 + L

     The base pointer was then used for computing relocatable addresses
   in the code fragment.  Frequently (such as in lots of continuations) 
   this is not needed.  We now eliminate this sequence whenever possible.

     For compile time efficiency, I'm using a very stupid local heuristic.  
   But in general, this should be done as a control flow analysis.

   Effect:  Smaller code size.  Speed up of most programs.


4. Frequency annotations

     I've added an annotation that states that all call gc blocks have zero
   execution frequencies.  This improves register allocation on the x86.
    
BENCHMARKS
==========

   I've only perform the comparison on 110.25.

   The platforms are:
  
    HPPA  A four processor HP machine (E9000) with 5G of memory.
    X86   A 300Hhz Pentium II with 128M of memory, and 
    SPARC An Ultra sparc 2 with 512M of memory.

   I used the following parameters for the SML benchmarks:

             @SMLalloc
     HPPA    256k
     SPARC   512k
     X86     256k

COMPILATION TIME
----------------
   Here are the numbers comparing the compilation times of the compilers.
   I've only compared 110.25 compiling the new sources versus
   a fixpoint version of the new compiler compiling the same.

                 110.25                                  New
           Total  Time in RA  Spill+Reload   Total  Time In RA Spill+Reload
     HPPA   627s    116s        2684+3584     599s    95s       1003+1879
     SPARC  892s    173s        2891+3870     708s    116s      1004+1880
     X86    999s    315s       94006+130691   987s    296s    108877+141957

               110.25         New
            Code Size      Code Size
     HPPA   8596736         8561421
     SPARC  8974299         8785143
     X86    9029180         8716783

   So in summary, things are at least as good as before.   Dramatic 
   reduction in compilation is obtained on the Sparc; I can't explain it, 
   but it is reproducible.  Perhaps someone should try to reproduce this 
   on their own machines.

SML BENCHMARKS
--------------

    On the average, all benchmarks perform at least as well as before. 

      HPPA         Compilation Time     Spill+Reload      Run Time
                 110.25  New            110.25    New   110.25  New    

      barnesHut  3.158  3.015  4.75%    1+1       0+0   2.980  2.922   2.00%
          boyer  6.152  5.708  7.77%    0+0       0+0   0.218  0.213   2.34%
   count-graphs  1.168  1.120  4.32%    0+0       0+0  22.705 23.073  -1.60%
            fft  0.877  0.792 10.74%    1+3       1+3   0.602  0.587   2.56%
    knuthBendix  3.180  2.857 11.32%    0+0       0+0   0.675  0.662   2.02%
         lexgen  6.190  5.290 17.01%    0+0       0+0   0.913  0.788  15.86%
           life  0.803  0.703 14.22%   25+25      0+0   0.153  0.140   9.52%
          logic  2.048  2.007  2.08%    6+6       1+1   4.133  4.008   3.12%
     mandelbrot  0.077  0.080 -4.17%    0+0       0+0   0.765  0.712   7.49%
         mlyacc 22.932 20.937  9.53%  154+181    32+57  0.468  0.430   8.91%
        nucleic  5.183  5.060  2.44%    2+2       0+0   0.125  0.120   4.17%
  ratio-regions  3.357  3.142  6.84%    0+0       0+0  116.225 113.173 2.70%
            ray  1.283  1.290 -0.52%    0+0       0+0   2.887  2.855   1.11%
         simple  6.307  6.032  4.56%   28+30      5+7   3.705  3.658   1.28%
            tsp  0.888  0.862  3.09%    0+0       0+0   7.040  6.893   2.13%
           vliw 24.378 23.455  3.94%  106+127    25+45  2.758  2.707   1.91%
  --------------------------------------------------------------------------
   Average                     6.12%                                   4.09%

      SPARC        Compilation Time     Spill+Reload      Run Time
                 110.25  New            110.25    New   110.25  New    

      barnesHut  3.778  3.592  5.20%    2+2       0+0   3.648  3.453    5.65%
          boyer  6.632  6.110  8.54%    0+0       0+0   0.258  0.242    6.90%
   count-graphs  1.435  1.325  8.30%    0+0       0+0  33.672 34.737   -3.07%
            fft  0.980  0.940  4.26%    3+9       2+6   0.838  0.827    1.41%
    knuthBendix  3.590  3.138 14.39%    0+0       0+0   0.962  0.967   -0.52%
         lexgen  6.593  6.072  8.59%    1+1       0+0   1.077  1.078   -0.15%
           life  0.972  0.868 11.90%   26+26      0+0   0.143  0.140    2.38%
          logic  2.525  2.387  5.80%    7+7       1+1   5.625  5.158    9.05%
     mandelbrot  0.090  0.093 -3.57%    0+0       0+0   0.855  0.728   17.39%
         mlyacc 26.732 23.827 12.19%  162+189    32+57  0.550  0.560   -1.79%
        nucleic  6.233  6.197  0.59%    3+3       0+0   0.163  0.173   -5.77%
  ratio-regions  3.780  3.507  7.79%    0+0       0+0 133.993 131.035   2.26%
            ray  1.595  1.550  2.90%    1+1       0+0   3.440  3.418    0.63%
         simple  6.972  6.487  7.48%   29+32      5+7   3.523  3.525   -0.05%
            tsp  1.115  1.063  4.86%    0+0       0+0   7.393  7.265    1.77%
           vliw 27.765 24.818 11.87%  110+135    25+45  2.265  2.135    6.09%
  ----------------------------------------------------------------------------
   Average                     6.94%                                    2.64%

      X86          Compilation Time     Spill+Reload      Run Time
                 110.25  New            110.25    New   110.25  New    

      barnesHut  5.530  5.420  2.03%  593+893   597+915   3.532  3.440   2.66%
          boyer  8.768  7.747 13.19%  493+199   301+289   0.327  0.297  10.11%
   count-graphs  2.040  2.010  1.49%  298+394   315+457  26.578 28.660  -7.26%
            fft  1.327  1.302  1.92%  112+209   115+210   1.055  0.962   9.71%
    knuthBendix  5.218  5.475 -4.69%  451+598   510+650   0.928  0.932  -0.36%
         lexgen  9.970  9.623  3.60% 1014+841  1157+885   0.947  0.928   1.97%
           life  1.183  1.183  0.00%  162+182   145+148   0.127  0.103  22.58%
          logic  3.285  3.512 -6.45%  514+684   591+836   5.682  5.577   1.88%
     mandelbrot  0.147  0.143  2.33%   38+41     33+54    0.703  0.690   1.93%
         mlyacc 35.457 32.763  8.22% 3496+4564 3611+4860  0.552  0.550   0.30%
        nucleic  7.100  6.888  3.07%  239+168   201+158   0.175  0.173   0.96%
  ratio-regions  6.388  6.843 -6.65% 1182+257   981+300  120.142 120.345 -0.17%
            ray  2.332  2.338 -0.29%  346+398   402+494   3.593  3.540   1.51%
         simple  9.912  9.903  0.08% 1475+941  1579+1168  3.057  3.178  -3.83%
            tsp  1.623  1.532  5.98%  266+200   250+211   8.045  7.878   2.12%
           vliw 33.947 35.470 -4.29% 2629+2774 2877+3171  2.072  1.890   9.61%
  ----------------------------------------------------------------------------
   Average                     1.22%                                     3.36%



Aliasing
---------
This update contains a rewritten (and hopefully more correct) module 
for extracting aliasing information from CPS.  

   To turn on this feature:  

        Compiler.Control.CG.memDisambiguate := true

   To pretty print the region information with assembly

       Compiler.Control.MLRISC.getFlag "asm-show-region" := true;

   To control how many levels of aliasing information are printed, use:

       Compiler.Control.MLRISC.getInt "points-to-show-level" := n

   The default of n is 3.

======================================================================
Boot code and glue scripts
======================================================================

Size info in BOOTLIST

  The BOOTLIST file now has an optional first line that specifies an
  upper bound on the number of boot files and an upper bound on the
  length of each individual name.  With this, there are no longer
  hard-wired restrictions on these values in the runtime system.
  (If the specification is missing in BOOTLIST, the runtime system
  falls back to its old behavior, i.e., hard-wired defaults.)

Allocation-size heuristics in .run-sml

  The .run-sml scripts tries to read processor cache size from
  /proc/cpuinfo.  This works on Linux and is important for small-cache
  Celeron systems that suffer badly when allocation size is set too
  high.

Install script

  - Written in a more modular fashion (using shell functions).
  - Made more robust.
  - Automagically fetches archive files over the network if they do not
    exist locally.  Thus, you only need to fetch config.tar.gz yourself.
    Unpack it and go!
    (Requires "wget" or "lynx" to be installed on the system and a
    live connection to the internet.  Moreover, the contents of
    config/srcarchiveurl must be set properly.)
    For CVS users, this may be convenient when fetching new sets of binfiles.
  - Handles archive files with or without version number and compressed
    with one of "gzip", "compress", or "bzip2".  Recognized suffixes are
    ".tar.gz", ".tgz", ".tar", ".tar.Z", and ".tar.bz2".

PIDMAP file

  There is a file called PIDMAP in the bootfile directory.
  It is used to minimize the amount of dynamic state that needs to be
  stowed away for the purpose of sharing between interactive system
  and user code.

Building standalone programs

  The command ml-build can be used to build standalone programs.
  ml-build takes three arguments:

     1. the name of the CM library that implements and exports the "main"
        function of your program
     2. the name of the "main" function of your program as exported by 1.
        (The function must have a type that makes it suitable as an argument
         to SMLofNJ.exportFn.)
     3. the name of the heapfile to be generated

Other build scripts

  ml-{lex,yacc} build scripts now make use of the new mechanism for
  building standalone programs.

Fixpoint script

  I added a re-written version of Dave's fixpt script to src/system.
  Changes relative to the original version:
    - sh-ified (not everybody has ksh)
    - automatically figures out which architecture it runs on
    - uses ./makeml a bit more cleverly
    - never invokes ./installml (and, thus, does not clobber your
      good and working installation of sml in case something goes wrong)
    - accepts max iteration count using option "-iter <n>"
    - accepts a "base" name using option "-base <base>"

  It does not build any extraneous heap images but directly rebuilds
  bin- and boot-hierarchies using makeml's "-rebuild" switch. Finally,
  it can incorporate existing bin- and boot- hierarchies.  For example,
  suppose the base is set to "sml" (which is the default).  Then it
  successively builds

          sml.bin.<arch>-unix and sml.boot.<arch>-unix
  then    sml1.bin.<arch>-unix and sml1.boot.<arch>-unix
  then    sml2.bin.<arch>-unix and sml2.boot.<arch>-unix
  ...
  then    sml<n>.bin.<arch>-unix and sml<n>.boot.<arch>-unix

  and so on.  If any of these already exist, it will just use what's
  there.  In particular, many people will have the initial set of bin
  and boot files around, so this saves time for at least one full
  rebuild.  Having sets of the form <base><k>.{bin,boot}.<arch>-unix for
  <k>=1,2,... is normally not a good idea when invoking fixpt.  However,
  they might be the result of an earlier partial run of fixpt (which
  perhaps got accidentially killed).  In this case, fixpt will quickly
  move through what exists before continuing where it left off earlier,
  and, thus, saves a lot of time.

Runtime system code

  - fixed several gcc -Wall warnings that were caused by missing header
    files, missing initializations, etc., in runtime (not all warnings
    eliminated, though)
  - hand to "un-fix" some of them later because they broke the HPPA compile

======================================================================
CM
======================================================================

Several manual updates

  I always try to keep the manual in sync with CM's latest features.

Bootstrap compilation

  No more "CMB.deliver"

    - All work is done by CMB.make (as it used to be in the old CM).
    - CMB.make can be used even with existing bootfiles, i.e., bootfiles do
      not have to be removed beforehand.
    - In "paranoid mode" CM checks a stable libraries CRC checksum to
      verify that it is "valid".  (In "normal mode", such checks do not
      occur.)  Paranoid mode is used for bootstrap compilation.  This is
      what makes it possible to re-use existing bootfiles.

  Initial glue code (init.cmi)

    - treated as a genuine library now
    - there are no more "built-in" modules

CM API

  CM.Anchor.anchor instead of CM.Anchor.{set,cancel}
     - Upon request by Elsa.  Anchors now controlled by get-set-pair
       like most other CM state variables.

  CM tools:
     - It is now possible to have tools that accept additional
       "command line" parameters (specified in the .cm file at each
       instance where the tool's class is used).

     - The parser understands named parameters and recursive options.

     - new "make" and "shell" tools added
       * facilitate fairly seemless hookup to portions of code
         managed using Makefiles or Shell scripts.

     - There are no classes "shared" or "private" anymore.  Instead,
       the sharing annotation is now a parameter to the "sml" class.

     - Tools.registerStdShellCmdTool (from smlnj/cm/tool.cm) takes an
       additional argument called "template" which is an optional
       string that specifiel the layout of the tool command line.  See
       the CM manual for explanation.

     - A special-purpose tool can be "registered" by simply dropping
       the corresponding <...>-tool.cm (and/or <...>-ext.cm) into the
       same directory where the .cm file lives that uses this tool.
       (The behavior/misfeature until now was to look for the tool
       description files in the current working directory.)  As
       before, tool description files could also be anchored -- in
       which case they can live anywhere they like.  Following the
       recent e-mail discussion, this change should make it easier to
       have special-purpose tools that are shipped together with the
       sources of the program that uses them.
       Bug: such a tool does not get un-registered after being done

Library names

  Library names have been completely re-organized.
     Many libraries have been consolidated so that they share the same
     path anchor.  For example, all MLRISC-related libraries are
     anchored at MLRISC, most libraries that are SML/NJ-specific are
     under "smlnj".  Notice that names like host-cmb.cm or
     host-compiler.cm no longer exist.  See system/README for a
     complete description of the new naming scheme.  Quick reference:

        host-cmb.cm        -> smlnj/cmb.cm
        host-compiler.cm   -> smlnj/compiler.cm
        full-cm.cm         -> smlnj/cm.cm
        <arch>-<os>.cm     -> smlnj/cmb/<arch>-<os>.cm
        <arch>-compiler.cm -> smlnj/compiler/<arch>.cm

CM bug fixes

   - exceptions in user code are being passed through (i.e., reach top level)
   - more bugs in paranoia mode fixed
   - bug related to checking group owners fixed
   - better error handling (suppresses many followup-messages)

Internals

   "Global" modmap:
     CM now maintains one "global" modmap that is used for all stable
     libraries.  The use of such a global modmap maximizes sharing and
     minimizes the need for re-traversing parts of environments during
     modmap construction.  (However, this has minor impact since modmap
     construction seems to account for just one percent or less of total
     compile time.)

======================================================================
Compiler Internals
======================================================================

Environment data structures: major changes

  No CMStaticEnv anymore.
    - no CMEnv, no "BareEnvironment" (actually, _only_ BareEnvironment,
      but it is called Environment), no conversions between different
      kinds of static environments

    - There is still a notion of a "modmap", but such modmaps are generated
      on demand at the time when they are needed.  This sounds slow, but I
      sped up the code that generates modmaps enough for this not to lead to
      a slowdown of the compiler (at least I didn't detect any).

    - To facilitate rapid modmap generation, static environments now
      contain an (optional) "modtree" structure.  Modtree annotations are
      constructed by the unpickler during unpickling.  (This means that
      the elaborator does not have to worry about modtrees at all.)
      Modtrees have the advantage that they are compositional in the same
      way as the environment data structure itself is compositional.
      As a result, modtrees never hang on to parts of an environment that
      has already been rendered "stale" by filtering or rebinding.

   -  all files that I touched now compile without warnings (other than
     "polyEqual").

   - compiler now tends to run "leaner" (i.e., ties up less memory in
     redundant modmaps)

Stats phase "genmap" added

  - measures time spent during on-the-fly modmap generation

Changes on behalf of CM

  Compiler.CMSA eliminated
     - No longer supported by CM anyway.

  Fixed bugs in pickler that kept biting Stefan
     - past refs to past refs (was caused by the possibility that
       ad-hoc sharing is more discriminating than hash-cons sharing)
     - integer overflow on LargeInt.minInt

  Handling of "core" environment:

    I eliminated coreEnv from compInfo.  Access to the "Core"
    structure is now done via the ordinary static environment that is
    context to each compilation unit.

    To this end, I arranged that instead of "structure Core" a
    "structure _Core" is bound in the pervasive environment.  Core
    access is done via _Core (which can never be accidentially rebound
    because _Core is not a legal surface-syntax symbol).

    The current solution is much cleaner because the core environment
    is now simply part of the pervasive environment which is part of
    every compilation unit's context anyway.  In particular, this
    eliminates all special-case handling that was necessary until now
    in order to deal with dynamic and symbolic parts of the core
    environment.

    Remaining hackery (to bind the "magic" symbol _Core) is localized
    in the compilation mananger's bootstrap compiler (actually: in the
    "init group" handling).  See the comments in
    src/system/smlnj/init/init.cmi for more details.

    I also tried to track down all mentions of "Core" (as string
    argument to Symbol.strSymbol) in the compiler and replaced them
    with a reference to the new CoreSym.coreSym.  Seems cleaner since
    the actual name appears in one place only.

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