Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
improve performance issue of
@nospecialize-d keyword func call (#47059
) This commit tries to fix and improve performance for calling keyword funcs whose arguments types are not fully known but `@nospecialize`-d. The final result would look like (this particular example is taken from our Julia-level compiler implementation): ```julia abstract type CallInfo end struct NoCallInfo <: CallInfo end struct NewInstruction stmt::Any type::Any info::CallInfo line::Int32 flag::UInt8 function NewInstruction(@nospecialize(stmt), @nospecialize(type), @nospecialize(info::CallInfo), line::Int32, flag::UInt8) return new(stmt, type, info, line, flag) end end @nospecialize function NewInstruction(newinst::NewInstruction; stmt=newinst.stmt, type=newinst.type, info::CallInfo=newinst.info, line::Int32=newinst.line, flag::UInt8=newinst.flag) return NewInstruction(stmt, type, info, line, flag) end @Specialize using BenchmarkTools struct VirtualKwargs stmt::Any type::Any info::CallInfo end vkws = VirtualKwargs(nothing, Any, NoCallInfo()) newinst = NewInstruction(nothing, Any, NoCallInfo(), zero(Int32), zero(UInt8)) runner(newinst, vkws) = NewInstruction(newinst; vkws.stmt, vkws.type, vkws.info) @benchmark runner($newinst, $vkws) ``` > on master ``` BenchmarkTools.Trial: 10000 samples with 186 evaluations. Range (min … max): 559.898 ns … 4.173 μs ┊ GC (min … max): 0.00% … 85.29% Time (median): 605.608 ns ┊ GC (median): 0.00% Time (mean ± σ): 638.170 ns ± 125.080 ns ┊ GC (mean ± σ): 0.06% ± 0.85% █▇▂▆▄ ▁█▇▄▂ ▂ ██████▅██████▇▇▇██████▇▇▇▆▆▅▄▅▄▂▄▄▅▇▆▆▆▆▆▅▆▆▄▄▅▅▄▃▄▄▄▅▃▅▅▆▅▆▆ █ 560 ns Histogram: log(frequency) by time 1.23 μs < Memory estimate: 32 bytes, allocs estimate: 2. ``` > on this commit ```julia BenchmarkTools.Trial: 10000 samples with 1000 evaluations. Range (min … max): 3.080 ns … 83.177 ns ┊ GC (min … max): 0.00% … 0.00% Time (median): 3.098 ns ┊ GC (median): 0.00% Time (mean ± σ): 3.118 ns ± 0.885 ns ┊ GC (mean ± σ): 0.00% ± 0.00% ▂▅▇█▆▅▄▂ ▂▄▆▆▇████████▆▃▃▃▃▃▃▃▃▃▃▂▂▂▂▂▂▂▂▂▁▁▂▂▂▁▂▂▂▂▂▂▁▁▂▁▂▂▂▂▂▂▂▂▂ ▃ 3.08 ns Histogram: frequency by time 3.19 ns < Memory estimate: 0 bytes, allocs estimate: 0. ``` So for this particular case it achieves roughly 200x speed up. This is because this commit allows inlining of a call to keyword sorter as well as removal of `NamedTuple` call. Especially this commit is composed of the following improvements: - Add early return case for `structdiff`: This change improves the return type inference for a case when compared `NamedTuple`s are type unstable but there is no difference in their names, e.g. given two `NamedTuple{(:a,:b),T} where T<:Tuple{Any,Any}`s. And in such case the optimizer will remove `structdiff` and succeeding `pairs` calls, letting the keyword sorter to be inlined. - Tweak the core `NamedTuple{names}(args::Tuple)` constructor so that it directly forms `:splatnew` allocation rather than redirects to the general `NamedTuple` constructor, that could be confused for abstract input tuple type. - Improve `nfields_tfunc` accuracy as for abstract `NamedTuple` types. This improvement lets `inline_splatnew` to handle more abstract `NamedTuple`s, especially whose names are fully known but its fields tuple type is abstract. Those improvements are combined to allow our SROA pass to optimize away `NamedTuple` and `tuple` calls generated for keyword argument handling. E.g. the IR for the example `NewInstruction` constructor is now fairly optimized, like: ```julia julia> Base.code_ircode((NewInstruction,Any,Any,CallInfo)) do newinst, stmt, type, info NewInstruction(newinst; stmt, type, info) end |> only 2 1 ─ %1 = Base.getfield(_2, :line)::Int32 │╻╷ Type##kw │ %2 = Base.getfield(_2, :flag)::UInt8 ││┃ getproperty │ %3 = %new(Main.NewInstruction, _3, _4, _5, %1, %2)::NewInstructionstruction └── return %3 │ => NewInstruction ```
- Loading branch information
