Performance regression returning anonymous function from struct

[imLew]

Sorry if the title isn't too clear, I'm not exactly sure what is going on. I experienced a huge performance decrease in a project I'm working on after upgrading to julia 1.6. I posted about it here, and following @KristofferC's advice tried reverting fd8f97e, which solved the problem.

A MWE that shows the problem is

using BenchmarkTools

struct Model
    ϕ  # feature maps
end

y(ϕ) = x -> ϕ(x)
y(model::Model) = x -> model.ϕ(x)
y(model::Model, x) =  model.ϕ(x)

ϕ = x -> x
M = Model(ϕ);
x = rand(100);

y(model::Model) = x -> model.ϕ(x)
@benchmark y($ϕ).($x) .* 1 
@benchmark y($M).($x) .* 1  # <---- This takes much longer with fd8f97e than without
@benchmark y($M, $x) .* 1  

This is an minimal as I could make it. The regression occurs when using the value of y($M).($x) (hence the .* 1). For julia installed from the Arch User Repository the results of the benchmark are:

julia> @benchmark y($M).($x) .* 1
BenchmarkTools.Trial:
  memory estimate:  277.09 KiB
  allocs estimate:  4286
  --------------
  minimum time:     3.082 ms (0.00% GC)
  median time:      3.914 ms (0.00% GC)
  mean time:        4.008 ms (0.42% GC)
  maximum time:     7.155 ms (41.82% GC)
  --------------
  samples:          1245
  evals/sample:     1

julia> versioninfo()
Julia Version 1.6.0
Commit f9720dc2eb* (2021-03-24 12:55 UTC)
Platform Info:
  OS: Linux (x86_64-pc-linux-gnu)
  CPU: Intel(R) Core(TM) i7-1065G7 CPU @ 1.30GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-11.1.0 (ORCJIT, icelake-client)

For julia master:

julia> @benchmark y($M).($x) .* 1
BenchmarkTools.Trial:
  memory estimate:  285.91 KiB
  allocs estimate:  4444
  --------------
  minimum time:     3.011 ms (0.00% GC)
  median time:      3.393 ms (0.00% GC)
  mean time:        3.549 ms (0.40% GC)
  maximum time:     6.644 ms (38.91% GC)
  --------------
  samples:          1407
  evals/sample:     1

julia> versioninfo()
Julia Version 1.7.0-DEV.1009
Commit 578d51857b* (2021-04-25 19:27 UTC)
Platform Info:
  OS: Linux (x86_64-pc-linux-gnu)
  CPU: Intel(R) Core(TM) i7-1065G7 CPU @ 1.30GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-11.0.1 (ORCJIT, icelake-client)

And finally for tag v1.6.0 with commit fd8f97e reverted:

julia> @benchmark y($M).($x) .* 1
BenchmarkTools.Trial:
  memory estimate:  4.38 KiB
  allocs estimate:  209
  --------------
  minimum time:     5.251 μs (0.00% GC)
  median time:      6.084 μs (0.00% GC)
  mean time:        6.344 μs (1.10% GC)
  maximum time:     150.130 μs (94.01% GC)
  --------------
  samples:          10000
  evals/sample:     6

julia> versioninfo()
Julia Version 1.6.1
Commit 3121d06bd6 (2021-04-25 18:53 UTC)
Platform Info:
  OS: Linux (x86_64-pc-linux-gnu)
  CPU: Intel(R) Core(TM) i7-1065G7 CPU @ 1.30GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-11.0.1 (ORCJIT, icelake-client)

I also tried reverting fd8f97e on master but there was a conflict in the affected code that I could not resolve.

[vtjnash]

seems just as bad or worse now:

julia> @benchmark y($M).($x) .* 1  # <---- This takes much longer with fd8f97e than without
BenchmarkTools.Trial: 4185 samples with 1 evaluation.
 Range (min … max):  1.146 ms …  33.064 ms  ┊ GC (min … max): 0.00% … 96.04%
 Time  (median):     1.179 ms               ┊ GC (median):    0.00%
 Time  (mean ± σ):   1.192 ms ± 525.705 μs  ┊ GC (mean ± σ):  0.64% ±  1.48%

            ▁ ▅▄▄▄▃▃▂▂▅▃▃▄▃▃     ▁ ▁▅██▆▇▄▅▁                   
  ▁▁▂▃▄▅▆█▆█████████████████▇█▇▇██████████████▇▆▅▄▃▂▃▂▃▃▃▂▂▂▂ ▅
  1.15 ms         Histogram: frequency by time        1.22 ms <

 Memory estimate: 60.28 KiB, allocs estimate: 1424.