Order dependent function compilation with inlining

[raminammour]

Hello,

I know that there are some issues open about order dependent inference behavior, I am not sure if this is related, but it may help with debugging. Feel free to close if this is duplicate or expected. I was producing a simple example to show how inlining can be bad sometimes and found this behavior peculiar:

function g(x,s)
    f(x,s[1])
end

@inline function f(x,s)
    s1=s
    for i in x
        s1 += i
    end
    s1
end

#As expected, inlining causes a bad type instability
for i in 3:7
    x=rand(10^i);@time g(x,Any[0.])
end
  0.000033 seconds (2.00 k allocations: 31.344 KiB)
  0.000288 seconds (20.00 k allocations: 312.594 KiB)
  0.002709 seconds (200.00 k allocations: 3.052 MiB)
  0.046871 seconds (2.00 M allocations: 30.518 MiB, 30.53% gc time)
  0.381641 seconds (20.00 M allocations: 305.176 MiB, 37.92% gc time)

@noinline function f(x,s)
    s1=s
    for i in x
        s1 += i
    end
    s1
end

# Again expected, not inlining ~ function barrier --> helps
for i in 3:7
    x=rand(10^i);@time g(x,Any[0.])
end
  0.000006 seconds (2 allocations: 112 bytes)
  0.000014 seconds (2 allocations: 112 bytes)
  0.000132 seconds (2 allocations: 112 bytes)
  0.001321 seconds (2 allocations: 112 bytes)
  0.012353 seconds (2 allocations: 112 bytes)

# Now **redefine f** with inline
@inline function f(x,s)
    s1=s
    for i in x
        s1 += i
    end
    s1
end

# Surprising, same behavior as with no-inline
for i in 3:7
    x=rand(10^i);@time g(x,Any[0.])
end
  0.007094 seconds (4.13 k allocations: 218.143 KiB, 99.71% compilation time)
  0.000014 seconds (2 allocations: 112 bytes)
  0.000132 seconds (2 allocations: 112 bytes)
  0.001481 seconds (2 allocations: 112 bytes)
  0.018629 seconds (2 allocations: 112 bytes)

It looks as if g is not recompiled (though the first line shows it being recompiled), if we redefine g it works as expected. Is it expected with inlining that the function is not recompiled?

Cheers, and great JuliaCon!

Julia Version 1.6.1
Commit 6aaedecc44 (2021-04-23 05:59 UTC)
Platform Info:
  OS: macOS (x86_64-apple-darwin18.7.0)
  CPU: Intel(R) Core(TM) i9-9880H CPU @ 2.30GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-11.0.1 (ORCJIT, skylake)

[kimikage]

Is this a kind of "so-called" #35800?