Use isb for normal cpu pause on aarch64

[gbaraldi]

Following rust-lang/rust@c064b65, it seems that isb is probably best for this. wfe is still good for some specific spins we do because we explicitly call jl_cpu_wake on them, but in most cases we just call jl_cpu_pause which technically can take unbounded time in aarch64, though it seems to be about a microsecond in the M1 which is far too long for some things.

In order to keep the nice pause/wakecombo I added jl_cpu_suspend (Open to bikeshedding) to be the actual long wait that needs to be woken up, while pause now behaves more like x86 pause taking a couple nanoseconds to execute.

[gbaraldi]

1.9

julia> @benchmark ccall(:jl_cpu_pause, Cvoid, ())
BenchmarkTools.Trial: 3769 samples with 1000 evaluations.
 Range (min … max):  369.250 ns …  4.283 μs  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):       1.333 μs              ┊ GC (median):    0.00%
 Time  (mean ± σ):     1.326 μs ± 79.711 ns  ┊ GC (mean ± σ):  0.00% ± 0.00%

                                                           ▂█▄
  ▃▁▁▁▃▃▁▃▁▁▁▃▁▃▁▄▄▄▁▁▃▄▃▃▁▄▁▁▁▄▃▃▁▁▃▅▄▄▄▃▁▃▅▄▄▅▄▄▄▅▆▆▆▅▇▇████ █
  369 ns        Histogram: log(frequency) by time      1.36 μs <

 Memory estimate: 0 bytes, allocs estimate: 0.

PR

julia> @benchmark ccall(:jl_cpu_pause, Cvoid, ())
BenchmarkTools.Trial: 10000 samples with 998 evaluations.
 Range (min … max):  16.449 ns … 56.696 ns  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     16.575 ns              ┊ GC (median):    0.00%
 Time  (mean ± σ):   16.742 ns ±  1.451 ns  ┊ GC (mean ± σ):  0.00% ± 0.00%

   ▄▇ █▆ ▅▄              ▁ ▂                                  ▂
  ▆██▁██▁██▁█▅▁▃▃▁▆▄▁▃▃▁▇█▁██▁▇▇▄▆▆▇▁▆▆▁▆▆▁▆▆▁▆█▁▇▇▁▇▆▁▅▅▁▅▄▃ █
  16.4 ns      Histogram: log(frequency) by time      18.1 ns <

 Memory estimate: 0 bytes, allocs estimate: 0.

For reference on an 12th gen intel

@benchmark ccall(:jl_cpu_pause,Cvoid, ())
BenchmarkTools.Trial: 10000 samples with 992 evaluations.
 Range (min … max):  36.617 ns … 46.337 ns  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     36.830 ns              ┊ GC (median):    0.00%
 Time  (mean ± σ):   37.402 ns ±  0.923 ns  ┊ GC (mean ± σ):  0.00% ± 0.00%

  █  ▅                     ▇  ▄                               ▁
  █▅▃█▅▂▃▄▂▃▃▄▅▃▃▄▅▆▆▆▆▆▆▅▅█▆▅█▆▄▄▆▆▅▄▆▅▅▅▆▅▆▅▅▅▅▃▃▅▅▄█▆▇▅▅▄▅ █
  36.6 ns      Histogram: log(frequency) by time      40.3 ns <

 Memory estimate: 0 bytes, allocs estimate: 0.

[gbaraldi]

.NET does some more fancier things for this dotnet/coreclr#13670. But that can probably a separate PR.

[d-netto]

If we're changing jl_cpu_pause, then the exponential backoff algorithm from #48600 might need some adjustments as well.

[gbaraldi]

Maybe, this behaves a lot more like the x86 one so it's probably a bit easier to share code.

[vchuravy]

This makes sense from my side, and the fact that other languages like Rust have done that is encouraging.

I am not necessarily familiar enough with AArch64 , so I am hoping @yuyichao can share some advice.

[vchuravy]

Also cc: @kpamnany

[kpamnany]

I'm no expert on AMD architecture, but this seems like a good thing.