pkgsrc-Changes archive

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Old Index]

CVS commit: pkgsrc/lang/jamvm



Module Name:    pkgsrc
Committed By:   dholland
Date:           Sun Sep  7 06:41:44 UTC 2008

Modified Files:
        pkgsrc/lang/jamvm: Makefile PLIST distinfo
        pkgsrc/lang/jamvm/patches: patch-aa

Log Message:
Update to 1.5.1 and make it build on amd64. Add destdir support.

JamVM 1.5.1 released on 10th March 2008
=======================================

A summary of changes since 1.5.0:

- Implemented a new internal VM symbol framework.  This ensures that
  all class, method, field names, etc. are unique thus removing the
  need to use strcmp for comparison.

- Implemented a new internal VM exception mechanism for frequent
  exceptions thrown within the VM.

- Hand-coded support for AMD64 ABI.  Previously AMD64 required libffi
  but this is not individually packaged on some Linux distributions.
  On a random set of signatures the hand-coded assembler is 4.6 times
  faster (and jBYTEmark FP Index is 30% faster).

- Ported to Darwin/ARM (i.e. the iPhone).  Built using unofficial
  toolchain and tested on jailbroken iPhone (1.1.3 firmware)

- Class initialisation/first active use fixes

  - initialise if first active use is object allocation via reflection
    (newInstance)

  - initialise if first active use is object allocation via JNI
    (AllocObject/NewObject)

  - The wait in step 2 of the initialisation sequence must be
    non-interruptible (not specified in the JVM or Java spec).

  - If allocation of the ExceptionInInitializerError fails,
    OutOfMemoryError should be thrown not the original exception

  - Don't resolve if initialiser throws ExceptionInInitializerError
    (subsequent resolution will not see the class is erroneous)

- JNI_OnLoad fixes

  - initialise JNI local refs before calling

  - if JNI_FindClass is called from JNI_OnLoad, ensure correct class
    loader context is used.

- JNI_OnUnload fixes

  - initialise JNI local refs before calling

  - If a library defines JNI_OnUnload, and the class loader which
    loaded it is GCed JNI_OnUnload is called and the library is
    unloaded.  Previously this was done within the GC.  If OnUnload
    allocated objects it could corrupt the heap.  Dummy unloader
    objects are now created, and JNI_OnUnload is called from the
    finalize method by the finalizer thread.

- Fix variable argument passing of floats When calling Java methods from
  native code

- Rework VM initialisation/start-up sequence

- Fix for JNI_ThrowNew and exceptions loaded by non-bootstrap classloader

- Fix for JNI_GetMethodID and object/class initialiser methods
  (do not search class hierarchy)

- During resolution throw IllegalAccessError and not IllegalAccessException

- handle loadClass returning null without throwing exception

- Fix for running Jar files where the main class is package-protected

- Interpeter inlining changes

  - Do not inline class initialisers (<clinit>).  No performance
    improvement as they are only ran once.  Reduces memory usage

  - Fix for class unloading when inlining disabled (-Xnoinlining)

- Bump java.specification.version to 1.5

- Improve error messages if VM aborts during initialisation

- During configure, do not disable zip/jar file support if zlib library
  or header can't be found (now aborts).  Users never saw the warning
  and they ended up with a build which "didn't work".

- Various compiler warnings.

JamVM 1.5.0 released on 28th October 2007
=========================================

A summary of changes since 1.4.5:

- Substantially modified the interpreter to implement inline-threading
  (also known as super-instructions or code-copying).  This copies
  code sequences together to produce native code that executes without
  the normal interpreter dispatch overhead (similar to a simple JIT).

  - Supported on i386, AMD64, PowerPC and ARM, with or without stack-
    caching.  It is enabled by default on i386, AMD64 and PowerPC.

  - Performance improvement upto 300% on Pentium 4 (NetBurst), and upto
    200% on Athlon 64.  Less on Core 2 Duo (upto 70% faster as it has
    indirect branch prediction), and PowerPC which was already very
    optimised (upto 30% faster).

  - Tested on gcc 3.4 (i386, AMD64, PowerPC), 4.0 (i386), 4.1
    (i386, AMD64, PowerPC, ARM), 4.2 (i386, AMD64)

  - Several new command line options to control inlining:

      -Xreplication : determines whether duplicate code sequences
       are replicated or shared.  Can be set to 'none' (no replication),
       'always' (all sequences are copies) or a threshold value (e.g.
       10, when sharing of a sequence reaches the threshold the sequence
       is replicated).

      -Xcodemem : the maximum amount of memory for super-instructions.
       Once the maximum is reached no new super-instructions are generated
       but existing super-instructions are reused (class unloading will
       also free unused sequences).  Can be set to a value or 'unlimited'.
       This option can be used to limit code memory on systems with little
       RAM (i.e. embedded).

      -Xshowreloc : debugging option, which shows which opcodes were
       determined to be relocatable, and for opcodes which aren't why
       they aren't relocatable.  When using stack-caching there are
       three versions of each opcode (for 3 stack-cache levels).

      -Xnoinlining : turns off inlining (equivalent to setting codemem
       to zero).

  - Opcode relocatability is determined by default at build time, but
    this doesn't work when cross-compiling (so inlining is disabled).
    Relocatability can be determined at runtime using configure option
    --enable-runtime-reloc-checks, but this increases executable size
    by approx 30%.

- Command line options -version and -showversion now shows build
  information.  This includes the execution engine (e.g. inline-
  threaded interpreter with stack-caching), the gcc version which
  was used to compile JamVM (useful for debugging) and the "built in"
  boot library path and classpath.

- Ported to ARM systems using EABI.  This is a full port, with hand-
  coded assembler to handle the construction of a call-frame for calling
  JNI native methods.

- Refactored GC to remove all possible calls which may deadlock
  with threads suspended in "unsafe" operations.  This includes
  use of malloc/realloc/free and pthreads operations.

- In allocClass() check if gcMalloc() returns NULL (i.e. OOM).

- Copy Sun's behaviour when dealing with an empty property key or no
  equals after key.

- Add java.util.concurrent.atomic.AtomicLong.VMSupportsCS8.

- In method and field access checks, put back in access check for
  class (removed previously because it caused regressions, since
  fixed in Classpath).

- Enable suspend on thread creation (parent thread created thread with
  suspension disabled).  Effects threads which immediately call a long
  running native method.

- Fixed race condition in thread creation and compaction (start function
  obtains class reference when it is not suspendable, if compaction
  occurs, the reference may be threaded, or moved).

- Fix potential deadlock in threadInterrupt under Linuxthreads.

- Protect lock operation in resetPeakThreadsCount().

- In createJavaThread(), re-enable suspension if pthread_create fails.

- Fix race-condition in thread deletion.

- Fix bug in hash table "scavenge".  This could lead to entries not being
  freed, and exhaustion of Monitors on embedded systems.

- Removed use of malloc/free in thread dump.  Removes potential deadlock
  with threads suspended in malloc/realloc/free.

- Added symbolic link from lib/rt.jar to Classpath's glibj.zip (fixes
  programs such as ecj which locates rt.jar via java.home).

- Bumped Java compatible version to 1.5.0 as Classpath is now the generics
  branch (since Classpath 0.95).


To generate a diff of this commit:
cvs rdiff -r1.6 -r1.7 pkgsrc/lang/jamvm/Makefile
cvs rdiff -r1.3 -r1.4 pkgsrc/lang/jamvm/PLIST
cvs rdiff -r1.4 -r1.5 pkgsrc/lang/jamvm/distinfo
cvs rdiff -r1.3 -r1.4 pkgsrc/lang/jamvm/patches/patch-aa

Please note that diffs are not public domain; they are subject to the
copyright notices on the relevant files.



Home | Main Index | Thread Index | Old Index