I've recently read about CPU instruction reordering for efficiency. But I'm not able to understand how CPU reorders its instructions. I mean compile time reordering is thinkable since the compiler can foresee the upcoming code. But for a CPU which reads instruction one after the other, how does it see upcoming instructions to reorder them
问题:
回答1:
Instructions are fetched in program order into an instruction queue; from the queue they are decoded and moved into reservation stations. These reservation stations effectively do the reordering: instructions are dispatched for execution to execution units as their arguments become available and the time all the arguments become available generally does not correspond to the order in the instruction queue/memory.
For an example, using the Tomasulo Algorithm, see these two videos:
Issue (and register renaming): https://youtu.be/I2qMY0XvYHA?list=PLAwxTw4SYaPkNw98-MFodLzKgi6bYGjZs
Dispatch/reordering: https://youtu.be/bEB7sZTP8zc?list=PLAwxTw4SYaPkNw98-MFodLzKgi6bYGjZs
回答2:
The instructions are decoded in order, but they then go into a collection of "in progress" instructions. Instructions can make forward progress if their dependencies are met.
For example, say the instructions are:
- Load register A from memory.
- Load register B from memory.
- Increment register A.
- Increment register B.
It may be that the last two instructions are in progress at the same time and if the memory read for register B completes first (maybe it was already in the L1 cache), then the increment of register B will take place before the increment of register A. (Though, of course, after that instruction is decoded.)