Comment by chubot | Hacker News Reader

Comment by chubot | Hacker News Reader

Parent Full thread

Ah thanks, I think this is the terminology I mentioned in this comment!

https://news.ycombinator.com/item?id=16777516

that is, "system" VMs vs. "process" VMs. I don't really like those terms, but the distinction is a good one. I think the author must have been at the VM summer school I linked there.

-----

BTW this 2015 paper is saying that the labeled gotos technique isn't a big deal on Haswell:

Branch prediction and the performance of interpreters -- Don't trust folklore

https://scholar.google.com/scholar?cluster=20257610669850946...

In other words the branch predictors improved enough on Intel such that it's not a big optimization. But I don't like how they didn't mention other CPU architectures. Intel is not all we care about!

Also, it would be nice to have an update on this 2015 paper to account for 2017-2018 Spectre/Meltdown mitigations...

The first part of the paper compares token threading vs naive switch in cpython 3.2 which has a compile flag to disable token threading[1]. I were referring to direct threading but the difference is probably minor or irrelevant.

on x64, token threading is

    ; goto tokens[program[ip]]
    mov r0, [program_instructions + ip]
    jmp [tokens + r0]

while direct threading is

    ; goto program[ip]
    jmp [program_instructions + ip]

the paper says

Nehalem shows a few outstanding speedups (in the 30 %– 40 % range), as well as Sandy Bridge to a lesser extent, but the average speedups (geomean of individual speedups) for Nehalem, Sandy Bridge, and Haswell are respectively 10.1 %, 4.2 %, and 2.8 % with a few outstanding values for each microarchitecture. The benefits of threaded code decreases with each new generation of microarchitecture.

but looking at the graphic, it seems a few benchmarks still had a +5% speed boost on Haswell.

It would also have been interesting if the paper provided the generated code by gcc for the switch statement. If the switch statement's cases are linear (0, 1, 2, 3, ...) and if there are more than 4 cases, then gcc token threads the code.[2]

[1] https://hg.python.org/cpython/file/v3.3.2/Python/ceval.c#l82... [2] https://godbolt.org/g/yRjFto

Reply on news.ycombinator.com