Ben,

There is nothing wrong with the djnz loops.

Your problem has to be, as others have suggested, the routine's interface with the rest of the program.

That means stack problems (unlikely since its only a single level call and return) or interrupt problems.

When the values of that DJNZ routine affect other code, it can only be due to two situations. First the delay duration is wrong due to the loop timing calculation or some dependent device specification. The sinister and illogical connection is when interrupts are occurring and changing the count-down values by using the same register, by changing the register bank selection, or by changing chip memory locations where the registers exist.

You might post the purpose of the delay routine and the crystal on the board so others can ponder whether the delay is sufficient.

-aka Jay on vacation still