Performance regression in tuples getting copied to functions that only read from them. #15277
Description
There is a performance regression in what I believe is tuple indexing (now I believe it is function call overhead) now I believe it is tuples sometimes getting copied into functions that previously got passed by reference, from the giant codegen rewrite PR.
Test script (note the @noinline):
@noinline function get_idx_tuple(n::NTuple, i::Int)
@inbounds v = n[i]
return v
end
function bench{N, T}(tuple::NTuple{N, T})
s = 0.0
t_tup = @elapsed for n in 1:10^5
for i in 1:N
s += get_idx_tuple(tuple, i)
end
end
return t_tup
end
tuple = (rand(75)...)
bench(tuple)This typically looks like this for different sizes of the tuple:
Full bisect log: https://gist.github.com/KristofferC/1ba13fe374897e4f7f98
Relevant:
# possible first bad commit: [5c5d48369a1a8f9a5959699ec44cab393255d0c5] [wip codegen rewrite] fix another type information mismatch error in assignment
# possible first bad commit: [548075efe73083f6a58b97814ea558bdbc0db41d] [wip codegen rewrite] fix mixed-type intrinsic usage (such as `bswap(Float64)`)
# possible first bad commit: [322878d93879f81cdf3d6a783376ad149b10d65b] [wip codegen rewrite] fix select_value validation and missing value copy in variable assignment
# possible first bad commit: [e4dcff39221571b8fc59d25873de092edaf46768] [wip codegen rewrite] fix isVolatile marking and closure memloc allocation marking
# possible first bad commit: [8b98aa0fe4056915b69176c0eca23391a77f2f58] [wip codegen rewrite] fix marking of varinfo.value when type_is_ghost and recursive constant struct creation
# possible first bad commit: [ffc200620862e04d6e4d5b9363ea03130276deca] [wip codegen rewrite] cleanup closure env handling to occur early
# possible first bad commit: [ae1a623634aed1912b8fb12a624529f3108fe418] [wip codegen rewrite] use correct macro test for jl_is_immutable
# possible first bad commit: [fec6aa9048d2e54f3c178b379db96074f1d867be] [wip] codegen rewrite + rewrite intrinsics to accomodate
cc @vtjnash