Inlining calls without concrete types can slow things down a lot

[simonster]

In situations where a function is inlined without concrete type information, we can end up with multiple dynamic (jl_apply_generic) function calls when we only needed one before inlining, which can slow things down a lot. Consider:

g(x) = x + x + x + x + x + x + x + x

function h()
    x = Any[1][1]
    for i = 1:1000000
        g(x)
    end
end

julia> @time h()
 350.321 milliseconds (2000 k allocations: 62500 KB, 2.18% gc time)

After adding @noinline to g(x):

julia> @time h()
  17.324 milliseconds (5 allocations: 224 bytes)

[yuyichao]

Should be type inference not codegen?

[ArchRobison]

Looks like I ran into this too.

It seems like jl_apply_generic should be split into two parts: one that does the lookup and one that does the call. Then the lookup portion could be subjected to CSE or PRE.

[KristofferC]

FWIW, g doesn't get automatically inlined anymore.

[vtjnash]

Confirmed—nothing to do here anymore