Eliminate some invalidations from loading FixedPointNumbers [tweaked #35733] #35803
Conversation
|
Why is Statistics testing ? PR to statistics: https://github.com/JuliaLang/Statistics.jl/pull/36 |
| @@ -3,6 +3,8 @@ | |||
| # promote Bool to any other numeric type | |||
| promote_rule(::Type{Bool}, ::Type{T}) where {T<:Number} = T | |||
|
|
|||
| convert(::Type{Bool}, x::Bool) = x # prevent invalidation | |||
There was a problem hiding this comment.
This one might not be necessary either, after the Union{} stuff is fixed. It's just a really common case that would otherwise fall back to convert(::Type{T}, x::T).
| @@ -49,6 +49,7 @@ Char | |||
| (::Type{AbstractChar})(x::Number) = Char(x) | |||
| (::Type{T})(x::AbstractChar) where {T<:Union{Number,AbstractChar}} = T(codepoint(x)) | |||
| (::Type{T})(x::T) where {T<:AbstractChar} = x | |||
| AbstractChar(x::AbstractChar) = x | |||
| @@ -179,6 +180,7 @@ convert(::Type{AbstractChar}, x::Number) = Char(x) # default to Char | |||
| convert(::Type{T}, x::Number) where {T<:AbstractChar} = T(x) | |||
| convert(::Type{T}, x::AbstractChar) where {T<:Number} = T(x) | |||
| convert(::Type{T}, c::AbstractChar) where {T<:AbstractChar} = T(c) | |||
| convert(::Type{AbstractChar}, c::AbstractChar) = c | |||
There was a problem hiding this comment.
...possibly same for everything up to the reduce_empty stuff.
| @@ -306,25 +306,29 @@ with reduction `op` over an empty array with element type of `T`. | |||
|
|
|||
| If not defined, this will throw an `ArgumentError`. | |||
| """ | |||
| reduce_empty(op, T) = _empty_reduce_error() | |||
| reduce_empty(::typeof(+), T) = zero(T) | |||
| reduce_empty(op, ::Type{T}) where T = _empty_reduce_error() | |||
There was a problem hiding this comment.
AFAICT this is not a signature-based invalidation, this is one that came from jl_method_table_disable which seems to have Base.delete_method as its only caller. I don't have a very good handle on what is going on here, sorry.
One interesting thing is that the inferred return type is Union{} (even before loading FixedPointNumbers):
julia> code_typed(Base.reduce_empty, (Base.BottomRF{typeof(max)}, Type{VersionNumber}); optimize=false)
1-element Array{Any,1}:
CodeInfo(
1 ─ %1 = Base.getproperty(op, :rf)::Core.Compiler.Const(max, false)
│ Base.reduce_empty(%1, T)::Union{}
└── Core.Compiler.Const(:(return %2), false)::Union{}
) => Union{}
|
I thought this PR would be more trivial than it was. Now when it starts generating discussion I think it is best to just close it to keep the focus in #35733. |
#35733 but with the controversial
::Union{}methods removed. Still provides a significant speedup in the test case posted in that PR.No impact on sysimage size.