Release build slower than debug on ARM (Yocto cross-compile)

embedded
1

Hey everybody,

I have written an application for some networking tasks that runs in the user space of my Yocto-Linux.
I noticed a dip in performance running the --release version.
Afterwards, I’ve tried different opt-levels and explicitly setting target-cpu, but without improvement.

I realise it is probably a "me problem", since i have a hard time finding any sources talking about problems with the optimization. If you got any pointer on how to pinpoint the issue, I would greatly appreciate the help. Thanks in advance!

Context:

  • Target: Rock Pi E (Arm Cortex‑A53 64‑bit)
  • Build system: Yocto, using a sourced SDK
  • Cross-compiling Rust (not building natively on the device)
2

Can you share your code? Have you tried profiling to find out which parts of the code have their performance increased from not using --release?

3

I do find it interesting that this is the second case reported here in a short time (you'd have to find the other post); is it possible there's been a regression?

4

Do you mean this?

5

No, it was more recent than that one. I don't believe it's been closed yet... let me try and find it.

Edit: here it is: Chess engine is faster with debug-assertions = true

6

For the chess engine, it was finally just an -O3 vs -O2 issue. But the OP said that they tested with different opt-levels already. Of course, it is not really nice that -O3 can be 20 percent slower than -O3 for a few apps, but I have currently not the skills, time, and motivation to debug that.