Martin,

the reason for my post was to give the HotSpot guys some hint for additional enhancement.

BTW, I have used the recent UTF_8 decoder for my reference.
... But I have experienced some errors:
http://bugs.sun.com/bugdatabase/view_bug.do?bug_id=6795537
http://bugs.sun.com/bugdatabase/view_bug.do?bug_id=6798514

Most of them I have repaired yet, and fortunately I could enhance the speed by another ~20 % (on my machine, should be evaluated).

-Ulf


Am 03.02.2009 00:02, Martin Buchholz schrieb:
Ulf, the new UTF_8 decoder in very recent jdks is much faster,
and the primary trick that was used to achieve that
was reducing the size of the bytecode in the hot methods
and moving cold code to separate methods.

Martin

On Mon, Feb 2, 2009 at 14:35, Ulf Zibis <[email protected]> wrote:
Hi all,

I have experienced a strange speed dependency of HotSpot compiled code after
little change in code, which is not part of the "hot" loop.
Is there any explanation for this behavior?

Code before little change:

class UTF_8_new extends Unicode {

      // ....

      private static CoderResult malformed(ByteBuffer src, int sp,
                                           CharBuffer dst, int dp, int nb) {
          src.position(sp -= nb - src.arrayOffset());
          CoderResult cr = malformedN(src, nb);
          updatePositions(src, sp, dst, dp);
          return cr;
      }

      // ....

          while (sp < sl) {
              // ....
                  int b2 = sa[sp++];
                  int b3 = sa[sp++];
                  int b4 = sa[sp++];
                  if (isMalformed4(b1, b2))
                      return malformed(src, sp-2, dst, dp, 2);
                  if (isNotContinuation(b3))
                      return malformed(src, sp-1, dst, dp, 3);
                  if (isNotContinuation(b4))
                      return malformed(src, sp, dst, dp, 4);
                  if (dp >= dl - 1)
                      return overflow(src, sp-4, dst, dp);
                  int uc = (b1 << 18) ^ (b2 << 12) ^ (b3 << 06) ^ b4 ^
0x00381f80;
              // ....
          }

The above code has following output (note gain against UTF_8_70$Decoder):
time for sun.nio.cs.UTF_8_60$Decoder: 2701 ms
time for sun.nio.cs.UTF_8_70$Decoder: 1984 ms
time for sun.nio.cs.UTF_8_new$Decoder: 1720 ms  // gain: 264 ms
time for sun.nio.cs.UTF_8_last$Decoder: 2110 ms


Following small code change made the "hot" part of the loop much slower:

class UTF_8_new extends Unicode {

      // ....

      private static CoderResult malformed(ByteBuffer src, int sp,
                                           CharBuffer dst, int dp, int nb) {
          src.position(sp - src.arrayOffset());    // removed subtraction of
nb
          CoderResult cr = malformedN(src, nb);
          updatePositions(src, sp, dst, dp);
          return cr;
      }

      // ....

          while (sp < sl) {
              // ....
                  int b2 = sa[sp++];
                  int b3 = sa[sp++];
                  int b4 = sa[sp++];
                  if (isMalformed4(b1, b2))
                      return malformed(src, sp-4, dst, dp, 2);
                  if (isNotContinuation(b3))
                      return malformed(src, sp-4, dst, dp, 3);
                  if (isNotContinuation(b4))
                      return malformed(src, sp-4, dst, dp, 4);
                  if (dp >= dl - 1)
                      return overflow(src, sp-4, dst, dp);
                  int uc = (b1 << 18) ^ (b2 << 12) ^ (b3 << 06) ^ b4 ^
0x00381f80;
              // ....
          }

time for sun.nio.cs.UTF_8_60$Decoder: 2731 ms
time for sun.nio.cs.UTF_8_70$Decoder: 1981 ms
time for sun.nio.cs.UTF_8_new$Decoder: 1872 ms  // gain: 109 ms
time for sun.nio.cs.UTF_8_last$Decoder: 2094 ms

For complete sources compare following revisions (... and run
UTF_8Benchmark.java ):
https://java-nio-charset-enhanced.dev.java.net/source/browse/java-nio-charset-enhanced/trunk/test/sun/nio/cs/?diff_format=s&rev=613
https://java-nio-charset-enhanced.dev.java.net/source/browse/java-nio-charset-enhanced/trunk/test/sun/nio/cs/?diff_format=s&rev=614


-Ulf







Reply via email to