[Vm-dev] ARMv8 JIT much slower to JIT than x86_64

[Eliot Miranda]

Hi All,

    I find this fascinating, and I'll profile as soon as I can. I have two
MacBook Pros, a 2018 2.9GHz Intel Core i9, and a 2021 Apple M1 Max. On
large loads the latter is about -20% faster than the former. I know this
because Tim found a bug in the become primitive with jitted methods that
manifested on the Linux ARMv8 (DUAL_MAPPED_CODE_ZONE regime). SImulating
the re3co9mpilation of all methods in the system that Tim's example did
took about 5 hours on the x86_64 and only 4 on the M1 Max.  But if I
compare the JIT benchmarks on them the speeds swing wildly the other way:

These are the average times to JIT all of COmpiledCode's methods (excluding
subclasses)
Apple M1:
min        max        average
2.417    79.257    9.674 usecs to JIT
2            20          5.938 number of literals
1           106         17.464 number of bytecodes

Apple 2.9GHz Core i9:
min        max        average
0.705     32.497   4.256 usecs to JIT
2            20          5.938 number of literals
1            106        17.464 number of bytecodes

So there must be some really inefficient thing that the Apple ARMv8 JIT is
doing to make it so slow. Interesting :-)

_,,,^..^,,,_
best, Eliot
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.squeakfoundation.org/pipermail/vm-dev/attachments/20230116/a9dae9a2/attachment.html>