C optimization: conditional store to avoid dirtying a cache line

Question

In the libuv source, I found this code:

  /* The if statement lets the compiler compile it to a conditional store.
   * Avoids dirtying a cache line.
   */
  if          


        

Answer 1

"to cache" means to hide something. The function of a Cache in computing is to hide the distance to main memory, by preempting main memory accesses as much as possible.

That only works if you used the data before, you have not pushed it out of cache yet, and nobody else took it away before you. Any other actor (other CPU, IO-Bus, ...) must be able to get the current value and change it, even if you have it cached. This task is accomplished using cache coherency protocols. Higher coherency means higher cost.

What your code tries to do, is make the compiler emit a conditional move, so the CPU checks for 0, and only writes if not 0. There's a whole family of conditional move instructions in the Intel/AMD IS and many others.

So now, step for step:

• Test for 0: If your CPU does not have a copy of the tested data, it must ask for one. That's far worse than before. Let's hope you don't hit main memory.
• Prepare to write a value:
  • You own the data: Wonderful, you are already done.
  • You don't own the data: The Cache calls its brethren and higher layers to inform them that it now owns this piece. Nobody else can retain a copy.
• Write a value: The cache saves the change and marks the cacheline (Lowest granularity of cache) dirty, need to write back.

So, is it worth it? It depends.

Aside: Why provide conditional store instruction, if you can synthesise them using a conditional jump and a store? The advantages are using less instructions and no risk of flushing the instruction pipeline (partly executed following instructions). UPDATE: Looks like they cannot move from register/immediate to memory on x86/x86_64.