A DESCRIPTION OF THE REQUEST :
Hotspot generates the following machine code when using DirectFloatBuffer.put(pos+7, value)) (native byte order):

0x00b4397b: mov edx, ebx
  0x00b4397d: add edx, 0x07 ;*iadd
  0x00b43980: test edx, edx
  0x00b43982: jl 0x00B44107 ;*iflt
                                        ; - java.nio.Buffer::checkIndex@1 (line 513)
                                        ; - java.nio.DirectFloatBufferU::put@6 (line 254)

The "test edx, edx" is not neccessary because all flags are already set by the "add edx, 0x07" instruction.

A rule should be added to the macro assembler (atleast for the server compiler) that removes the test instruction when the preceeding instruction is a add or sub on the same register.

PS: The above disassembly was obtained with -XX+PrintAssembly and my own disassembler.dll based on libdisasm.


JUSTIFICATION :
Performance improvements in code that builds DirectXXBuffers (like OpenGL applications).


---------- BEGIN SOURCE ----------
import java.nio.*;

static class Test {
void test(FloatBuffer buf, float x, float y) {
    int pos = buf.position();
    buf.position(pos + 8);
    buf.put(pos, x).put(pos+1, y).put(pos+2, x+1).put(pos+3, y);
    y++;
    buf.put(pos+4, x).put(pos+5, y).put(pos+6, x+1).put(pos+7, y);
}

public static void main(String[] args) {
   FloatBuffer buf = ByteBuffer.allocateDirect(4*8*1000000).order(ByteOrder.nativeByteOrder()).asFloatBuffer();
   for(int i=0 ; i<1000000 ; i++) {
       test(buf, i*17, i*5);
   }
}

}
---------- END SOURCE ----------