3 ms·

Right, "just make your stack frame a multiple of 16 bytes" was wrong; I should have said "just make your stack frame 16 * k - 4 bytes long". This is basically f

by ssp 17y ago

Right, "just make your stack frame a multiple of 16 bytes" was wrong; I should have said "just make your stack frame 16 * k - 4 bytes long". This is basically free because the size of the stack frame is known at compile time. Realigning the stack at runtime though, is expensive and complicated.

jedbrown 17y ago
Aligning the stack at runtime is one instruction in the preamble so it's not a big deal, especially if the function is big enough to have non-register SSE locals. Also, when using VLAs, it's not optional. I use SSE all the time and have no problems with the alignment requirements, but it's not really a performance issue and I'm doubtful that it was worth the time for people to learn and deal with this idiosyncrasy.
- jedbrown 17y ago
  Actually, come to think of it, when the function doesn't call anything else, locals will be indexed off the base pointer and the stack pointer will not be explicitly moved. So it costs two extra instructions to get aligned locals in a function that needs non-register locals but doesn't call any other functions. Of course these are 1/.33 (latency/inverse throughput) instructions.
- ssp 17y ago
  You need to restore the stack to what it was before on return. Normally you can easily do that because you know the stack pointer delta at compile time. But when realigning at runtime, you don't, so you need to compute the difference and then store that somewhere, and on return you have to read that back and add it to the stack pointer. Or is there some trick that I'm missing?
  - jedbrown 17y ago
    It's in the frame pointer. Edit: You're right, optimized code might not save the frame pointer and this wouldn't provide any benefit if your frame had unknown size. But usually the preamble saves the old ESP in EBP and it is restored at the end. Maybe there is another alternative.
    - ssp 17y ago
      Hmm, I was actually about to edit my comment to say that you are right because usually you'll store the old stack pointer as the new base pointer. I guess the answer is that realigning at runtime costs a register, but most code has to pay that cost anyway.

by ssp 17y ago

jedbrown 17y ago

Aligning the stack at runtime is one instruction in the preamble so it's not a big deal, especially if the function is big enough to have non-register SSE locals. Also, when using VLAs, it's not optional. I use SSE all the time and have no problems with the alignment requirements, but it's not really a performance issue and I'm doubtful that it was worth the time for people to learn and deal with this idiosyncrasy.

jedbrown 17y ago
Actually, come to think of it, when the function doesn't call anything else, locals will be indexed off the base pointer and the stack pointer will not be explicitly moved. So it costs two extra instructions to get aligned locals in a function that needs non-register locals but doesn't call any other functions. Of course these are 1/.33 (latency/inverse throughput) instructions.
ssp 17y ago
You need to restore the stack to what it was before on return. Normally you can easily do that because you know the stack pointer delta at compile time. But when realigning at runtime, you don't, so you need to compute the difference and then store that somewhere, and on return you have to read that back and add it to the stack pointer. Or is there some trick that I'm missing?
- jedbrown 17y ago
  It's in the frame pointer. Edit: You're right, optimized code might not save the frame pointer and this wouldn't provide any benefit if your frame had unknown size. But usually the preamble saves the old ESP in EBP and it is restored at the end. Maybe there is another alternative.
  - ssp 17y ago
    Hmm, I was actually about to edit my comment to say that you are right because usually you'll store the old stack pointer as the new base pointer. I guess the answer is that realigning at runtime costs a register, but most code has to pay that cost anyway.