# Wednesday, 24 November 2004
« New Snapshot | Main | New Runtime -> Library Interface & New S... »
Direct ByteBuffer, Inlining & New Snapshot

Michael Koch finished Classpath's implementation of DirectByteBufferImpl. Direct ByteBuffers allow Java code to access unmanaged memory. You can allocate them from Java using ByteBuffer.allocateDirect or from JNI using NewDirectByteBuffer. The default Classpath implementation uses native methods to implement accessing the buffer, but for IKVM I really didn't want to do it that way, because I like to use managed code whenever possible (for both maintainability and portability) and also because native method invocation is fairly slow. I therefore decided to reimplement gnu.classpath.RawData (the Classpath abstract native memory pointer) and java.nio.VMDirectByteBuffer.

At first RawData was a value type containing a single IntPtr field in an attempt to avoid an indirection when loading the pointer. This required that I replaced the implementation of DirectByteBufferImpl.get/put with some handwritten CIL instructions, because the compiled Java code resulted in unnecessarily boxing the RawData struct and that obviously killed performance. Since I was replacing the get and put methods anyway, I also inlined the index check and position update. After some experimentation I discovered that making RawData a value type didn't really help performance (the .NET JIT isn't very good at taking advantage of value type optimization opportunities) so I changed it into a reference type, this removed the need for special casing get and put, but when I removed the handcoded CIL performance went down significantly (approx. 50% slower). For some reason the JIT wasn't inlining the index check method, so I left the handcoded CIL in map.xml.

Here's a random microbenchmark:

import java.nio.*;
class DirectByteBuffer {
  public static void main(String[] args) {
    final int SIZE = 10 * 1024 * 1024;
    for(int j = 0; j < 2; j++) {
      ByteBuffer b = ByteBuffer.allocateDirect(SIZE);
      long start = System.currentTimeMillis();
      for(int i = 0; i < SIZE; i++)
        b.put(i, (byte)i);
      long sum = 0;
      for(int i = 0; i < SIZE; i++)
        sum += b.get(i) & 0xFF;
      long end = System.currentTimeMillis();
      System.out.println(end - start);
      System.out.println("sum = " + sum);


JRE 1.4.2 HotSpot Client VM Average
671 530 541 521 461 545
450 461 450 460 460 456
JRE 1.5 HotSpot Client VM  
401 461 411 481 480 447
410 360 441 410 411 406
JRE 1.5 HotSpot Server VM  
220 120 110 120 120 138
100 110 110 160 100 116
IKVM on .NET 1.1  
201 231 231 240 251 231
210 210 210 221 220 214
IKVM on Mono 1.0.2 (--optimize=all)  
430 391 400 401 390 402
401 390 401 390 401 397

Pretty good results. Most other operations perform similarly, except for compact, copy a large buffer takes about twice as long as on the JRE. The .NET class library really should have a memcpy or memmove method that is properly optimized. Note that there is CIL instruction cpblk to copy memory and I tried that as well, it is about 30% faster than my straightforward C# loop, but it doesn't handle overlapping regions and I didn't want to use an unverfiable instruction anyway (as that would cause IKVM.GNU.Classpath to fail verification and I use verification as simple smoke test during build to catch the most obvious regressions in the compiler).

I've also updated the Japi status results: IKVM vs JDK and JDK vs IKVM.

Other changes:

  • Resync'ed with GNU Classpath cvs.
  • Removed java/net/SocketInputStream.java and java/net/SocketOutputStream.java.
  • Moved contents of java/net/PlainSocketImpl.java to gnu/java/net/PlainSocketImpl.java.
  • Moved contents of java/net/PlainDatagramSocketImpl.java to gnu/java/net/PlainDatagramSocketImpl.java.
  • Added deprecated attribute to the three deprecated String methods.
  • Added deprecated attribute support to map.xml parser.
  • Added deprecated attribute support to ClassFileWriter.
  • Added deprecated attribute support to ikvmstub.
  • Added public API to query deprecatedness of fields, methods, constructors and classes.
  • Changed ikvmstub to emit serialVersionUID fields even if it is private, to make Japicompat status more accurate.
  • Added public API to query the constant value of a field.
  • Added ConstantValueAttribute to persist the constant value of constant instance fields (which are treated the same as static constant fields, by javac).
  • Added support to ikvmstub for constant instance fields.
  • Changed japi build file to generate both forward and backward compatibility reports.
  • Fixed runtime to support loading a pre-compiled Java class by referencing an ikvmstub generated stub class.
  • Fixed module names when generating -Xsave files.
  • Moved JNI method pointer field to proxy class when generating -Xsave files.
  • Replaced gnu.classpath.RawData with an IKVM specific version.
  • Added optimized special case implementation of several java.nio.DirectByteBufferImpl methods to map.xml.
  • Removed CharConversionException handlers from StringHelper (I fixed Classpath's character/string encoders/decoders to no longer throw exceptions).
  • ClassLoaderWrapper has a new method PublishLibraryImplementationHelperType to publish types that live in IKVM.Runtime.dll (this is used for the new gnu.classpath.RawData that lives in IKVM.Runtime.dll because it contains unverifiable code).
  • Improved line number table encoding (more space efficient and simpler code).
  • Implemented JNI methods NewDirectByteBuffer, GetDirectBufferAddress and GetDirectBufferCapacity.
  • Added new method AttributeHelper.SetEditorBrowsableNever to decorate methods with the EditorBrowsableAttribute (to prevent them from showing up in Intellisense).
  • Implemented support to override method implementations in map.xml (not implemented for constructors yet).
  • Renamed all of the IKVM introduced fields and methods to __<xxx>.
  • Added support for several new CIL instructions to remapper.cs.
  • Shadow types (e.g. the .NET java.lang.Object type) now hide all inherited methods and decorate them with EditorBrowsableAttribute(Never) so that they no longer show up in Intellisense.
  • I broke JNI.RegisterNatives in this snapshot (but it's fixed in cvs).

New snapshots: just the binaries and source plus binaries.

Wednesday, 24 November 2004 13:56:56 (W. Europe Standard Time, UTC+01:00)  #    Comments [0] Tracked by:
"Online poker demographics" (Online poker demographics) [Trackback]