How to get C51 to generate a DJNZ loop?

I want a fast and simple delay routine:

#pragma OT(8,SPEED)    // Switch to speed optimization for this routine
void my_delay(BYTE count)
{
        register BYTE i;
        for (i=count; i; i--);
}

Is the best so far and becomes:

B02:0xB1F1  EF       MOV      A,R7
B02:0xB1F2  6003     JZ       B02:B1F7
B02:0xB1F4  1F       DEC      R7
B02:0xB1F5  80FA     SJMP     I2C_delay(B02:B1F1)
B02:0xB1F7  22       RET

but I was hoping there was a way to get C51 to do just a DJNZ instead of 4 instructions for each loop.

Is there a magic coding style or optimization level that will generate the DJNZ?