Switch macOS task mechanism back to ASM #39679
Conversation
5327824 rearranged the ifdefs here and accidentally switched macOS to unwind based switching, which is much slower.
Keno
added
performance
|
This was intentional. In benchmarks, it was somewhat slower than the best, but nowhere near how abysmal linux was. |
|
I don't think slowing down the task switch path with something that needs to read unwind tables is an acceptable solution to this. Can you explain what the problem is with using the ASM implementation? |
|
It reads them pretty fast (around the cost of the kernel context switch we also already do). Perhaps it shouldn't do either one, but people haven't generally complained about it much. |
|
From offline discussion with @vtjnash, this change was intentional, because GDB like to crash when it encounters setjmp/longjmp, so this needs to be enabled for effective debugging of tasks. However, we already turn it off on Linux for performance reasons and I noticed additional performance concerns (over and above the 5x slow down reported in the additional PR) with the upcoming switch to the LLVM unwind library. In the fullness of time, we should switch to something even faster on all platforms, which I wrote down over in #39680. |
Keno
changed the title
5327824 rearranged the ifdefs here and switched macOS to unwind based switching, which is somewhat slower. That was deemed acceptable at the time, because it work around some GDB misbehavior in setjmp. However, we're about to update to the LLVM libunwind fork, which appears to be even slower here, so switch back to ASM to avoid that performance regression.
5327824 rearranged the ifdefs here and switched macOS to unwind based switching, which is somewhat slower. That was deemed acceptable at the time, because it work around some GDB misbehavior in setjmp. However, we're about to update to the LLVM libunwind fork, which appears to be even slower here, so switch back to ASM to avoid that performance regression.
5327824 rearranged the ifdefs here and
accidentally switched macOS to unwind based switching, which is much slower.