Specialize MI if necessary

[vchuravy]

@vtjnash is there a way to speedup this query? For identity(::Int) this currently takes ~500ns

The goal here is to compute the mi we would have invoked.

[vchuravy]

So this is now ~70ns

[vtjnash]

!isdispatchtuple(mi.specTypes) is true, you are guaranteed to get back the exact same mi object

[vchuravy]

Okay using !== instead of != gets me the performance back.

julia> @benchmark GPUCompiler.methodinstance(typeof(identity), Tuple{Int,})
BenchmarkTools.Trial: 10000 samples with 968 evaluations.
 Range (min … max):  78.574 ns … 118.419 ns  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     84.008 ns               ┊ GC (median):    0.00%
 Time  (mean ± σ):   83.506 ns ±   2.955 ns  ┊ GC (mean ± σ):  0.00% ± 0.00%

  ▅                        █▄ ▂▂      ▂                         
  █▁▂▂▁▁▂▄▂▃▂▁▁▂▃▆▂▅▄▂▂▃▃▂▃██▂██▅▂▂▂▂▂█▂▄▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▂▃ ▃
  78.6 ns         Histogram: frequency by time         90.7 ns <

 Memory estimate: 0 bytes, allocs estimate: 0.

julia> @benchmark GPUCompiler.methodinstance(typeof(+), Tuple{Int,Int})
BenchmarkTools.Trial: 10000 samples with 999 evaluations.
 Range (min … max):  12.633 ns … 31.822 ns  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     12.893 ns              ┊ GC (median):    0.00%
 Time  (mean ± σ):   12.903 ns ±  0.353 ns  ┊ GC (mean ± σ):  0.00% ± 0.00%

                                 █             ▅ ▁             
  ▂▂▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▂▃█▄▁▂▁▁▁▁▃█▂▇▁▄█▂█▁▃▃▁▃▂▂▁▁▂▂ ▃
  12.6 ns         Histogram: frequency by time          13 ns <

 Memory estimate: 0 bytes, allocs estimate: 0.

With != it was:

julia> @benchmark GPUCompiler.methodinstance(typeof(identity), Tuple{Int,})
BenchmarkTools.Trial: 10000 samples with 759 evaluations.
 Range (min … max):  154.545 ns … 283.768 ns  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     171.700 ns               ┊ GC (median):    0.00%
 Time  (mean ± σ):   174.559 ns ±  14.674 ns  ┊ GC (mean ± σ):  0.00% ± 0.00%

         ▁▅▄▅█                                                   
  ▁▁▁▁▂▁▄█████▇▇▃▂▁▁▂▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁ ▂
  155 ns           Histogram: frequency by time          256 ns <

 Memory estimate: 0 bytes, allocs estimate: 0.

[vtjnash]

With isdispatchtuple, that should not be reachable

[vtjnash]

Could I interest you in fixing this in JuliaInterpreter, now that you know all of the secrets?

[vchuravy]

I could fix this in Base for Base.method_instance, but other than that I currently trying to not do code development and focus on my thesis