Exceptions come in two categories: traps and faults. Faults represent a condition that makes the processor unable to continue execution. Traps are the conditions that cause an exception as part of the execution. Both report the address of the next instruction to be executed. The difference is that to cause a trap, an instruction has been executed, so the reported address is the address of the instruction after the one causing the trap.
The debug exception you are talking about is also a trap. That is not a bug, it is a feature.
When you single step to a program you expect that each instruction is executed exactly once, and that you can execute exactly one instruction at a time. What the processor thus does is that at the end of each instruction, it checks wether it is single stepping and cause int #1 if necessary.
You can not compute the address of the previous instruction directly by looking at the machine code, that's not guaranteed to work. Consider this to be the previously executed instruction:
If you only know where this instruction ends you can not accurately determine what instruction it is - is it one byte (90, NOP), is it two bytes (CD 90 - int 0x91) or four (mov ax, 0x90CD) or even more?
However since you are single stepping, you can know the exact location - something must have just set the trap flag, or you have had a previous debugging exception.
I still fail to see why you would want the location of the previously executed instruction - you can only do things more often than they are supposed to take place - if e.g. I keep repeating MOV EAX, [EAX] I probably get a pagefault somewhere in the near future...
