[release-0.4] backport ARM ABI

[tkelman]

Backporting #14194 again, but this time in a way that should build correctly on release-0.4.

[tkelman]

A colleague hit a segfault where the root cause was calling float("0"). ccall ABI fixes it.

runbenchmarks("all", vs = "JuliaLang/julia:v0.4.3")

[jrevels]

The format allowed for vs is "user/repo:branch" or "user/repo@sha". It'd be nice if it accepted tags, though...

Also, I think you were looking for the keyword ALL, which is a placeholder that matches all benchmarks:

runbenchmarks(ALL, vs = "JuliaLang/julia@a2f713dea5ac6320d8dcf2835ac4a37ea751af05")

[nanosoldier]

Your benchmark job has completed - possible performance regressions were detected. A full report can be found here. cc @jrevels

[tkelman]

are those regressions serious enough to worry about? what's the best way to replicate locally? this makes the arm build work much much better on release-0.4

[jrevels]

The timing regressions are all from benchmarks that have proven very noisy in the past. The ~83x allocation count regression on the in-place insertion sort benchmark is weird, though, and worth checking out. I could tell you how to do it using BaseBenchmarks.jl, but those instructions will be obsolete soon, so it'll probably be easier just to check it manually:

list = rand(MersenneTwister(1), 50000)
@time sort!(list, alg = InsertionSort)

I'm currently building this branch locally to see if I can reproduce the regression using the newer benchmark tools, I'll report back if I find anything relevant.

Edit: I can't reproduce the regression locally using the new tools, must've been a fluke.

[tkelman]

Thanks @jrevels!