GH-126892: Reset warmup counters when JIT compiling code #126893
Conversation
brandtbucher
added
performance
There was a problem hiding this comment.
The warmup numbers are going to be wrong once #126816 is merged.
Do you want to get that merged first, then update this. Or the other way around?
|
I'll do that then this. |
|
Fixed the tests (I also needed to change the logic a bit since the bytecode re-specializes before we JIT a new trace). |
| __run_using_command=[_strace_binary] + strace_flags) | ||
| __run_using_command=[_strace_binary] + strace_flags, | ||
| # Don't want to trace our JIT's own mmap and mprotect calls: | ||
| PYTHON_JIT="0", |
There was a problem hiding this comment.
Guessing this showed up in the test_fileio test test_syscalls_read?
I have been keeping an eye on / working through a couple other issues where mmap(NULL) and mprotect show up, trying to figure out if just excluding those two from that test would be good/effective (memory allocation happening in reading files is fine / not what the test is focused on).
There was a problem hiding this comment.
Yeah, I was wondering if this was susceptible to catching badly-timed arena allocations and such too. Might make sense to have an option to ignore the mmap/munmap/mprotect family of functions instead of this ad-hoc fix.
I'm just not familiar with how this is being used, and didn't want to change the meaning/behavior of the tests too dramatically.
19% fewer traces created, no change in uops executed. No performance impact.