Unexpected recompilation when REPL is loaded

[christiangnrd]

When run in the REPL, things that should have been precompiled seem to be ignored.

1.12:

ChrichriMBP:~/.julia/dev/Metal$ j --proj -e"using REPL, Metal; @time Metal.@sync @metal identity(nothing);"
  0.719506 seconds (3.76 M allocations: 184.910 MiB, 4.25% gc time, 84.25% compilation time: <1% of which was recompilation)
ChrichriMBP:~/.julia/dev/Metal$ j --proj -e"using Metal; @time Metal.@sync @metal identity(nothing);"
  0.167388 seconds (15.88 k allocations: 1.108 MiB, 31.50% compilation time: <1% of which was recompilation)

With --trace-compile-timing. Notice the line that takes 556ms:

Details

ChrichriMBP:~/.julia/dev/Metal$ j --proj --trace-compile=stderr --trace-compile-timing -e"using REPL, Metal; @time Metal.@sync @metal identity(nothing);"
#=    8.6 ms =# precompile(Tuple{typeof(Base.vect), Array{String, 1}, Vararg{Array{String, 1}}})
#=    6.0 ms =# precompile(Tuple{typeof(Base.iterate), Array{Array{String, 1}, 1}})
#=    3.4 ms =# precompile(Tuple{typeof(Base.vect), Base.Dict{String, Any}, Vararg{Any}})
#=    3.1 ms =# precompile(Tuple{typeof(Base.vect), Array{String, 1}, Vararg{Any}})
#=    7.4 ms =# precompile(Tuple{typeof(Base.getindex), Type{Union{Nothing, Array{String, 1}}}, Array{String, 1}, Nothing})
#=    2.5 ms =# precompile(Tuple{typeof(Base.iterate), Array{Union{Nothing, Array{String, 1}}, 1}})
#=    4.6 ms =# precompile(Tuple{typeof(Artifacts._artifact_str), Module, String, Base.SubString{String}, String, Base.Dict{String, Any}, Base.SHA1, Base.BinaryPlatforms.Platform, Base.Val{LazyArtifacts}})
#=   26.6 ms =# precompile(Tuple{typeof(Artifacts.__artifact_str), Module, String, Base.SubString{String}, String, Base.Dict{String, Any}, Base.SHA1, Base.BinaryPlatforms.Platform, Base.Val{LazyArtifacts}})
#=    0.0 ms =# precompile(Tuple{typeof(JLLWrappers.get_julia_libpaths)})
#=   52.2 ms =# precompile(Tuple{typeof(Base.get), Base.TTY, Symbol, Bool}) # recompile
#=    4.8 ms =# precompile(Tuple{typeof(Base.issorted), Array{String, 1}, Base.Order.ReverseOrdering{Base.Order.ForwardOrdering}})
#=   17.8 ms =# precompile(Tuple{typeof(GPUCompiler.assign_args!), Expr, Array{Any, 1}})
#=    2.0 ms =# precompile(Tuple{typeof(Base.indexed_iterate), Tuple{Array{Symbol, 1}, Array{Any, 1}}, Int64})
#=    1.7 ms =# precompile(Tuple{typeof(Base.indexed_iterate), Tuple{Array{Symbol, 1}, Array{Any, 1}}, Int64, Int64})
#=   26.7 ms =# precompile(Tuple{typeof(GPUCompiler.split_kwargs), Tuple{}, Array{Symbol, 1}, Vararg{Array{Symbol, 1}}})
#=    2.1 ms =# precompile(Tuple{typeof(Base.indexed_iterate), NTuple{4, Array{Any, 1}}, Int64})
#=    1.7 ms =# precompile(Tuple{typeof(Base.indexed_iterate), NTuple{4, Array{Any, 1}}, Int64, Int64})
#=    2.6 ms =# precompile(Tuple{typeof(Base.parse), Type{Bool}, String})
#=    1.3 ms =# precompile(Tuple{typeof(Base.:(>)), UInt32, UInt32})
#=    1.2 ms =# precompile(Tuple{typeof(Base.in), Symbol, NTuple{6, Symbol}})
#=    2.0 ms =# precompile(Tuple{typeof(Base.:(!=)), UInt32, UInt32})
#=  556.6 ms =# precompile(Tuple{typeof(GPUCompiler.compile), Symbol, GPUCompiler.CompilerJob{T, P} where P where T})
#=    4.7 ms =# precompile(Tuple{typeof(Base.getindex), Core.SimpleVector, Base.UnitRange{Int64}})
#=    1.5 ms =# precompile(Tuple{typeof(Base.Iterators.enumerate), Core.SimpleVector})
#=    2.4 ms =# precompile(Tuple{typeof(Base.iterate), Base.Iterators.Enumerate{Core.SimpleVector}, Tuple{Int64}})
#=    2.5 ms =# precompile(Tuple{typeof(Base.iterate), Base.Iterators.Enumerate{Core.SimpleVector}, Tuple{Int64, Int64}})
#=    1.5 ms =# precompile(Tuple{typeof(Base.getindex), Base.RefValue{Any}})
#=    2.1 ms =# precompile(Tuple{typeof(Base.indexed_iterate), Pair{Symbol, String}, Int64})
#=    2.0 ms =# precompile(Tuple{typeof(Base.indexed_iterate), Pair{Symbol, String}, Int64, Int64})
#=    4.0 ms =# precompile(Tuple{typeof(Base.haskey), Base.Dict{Symbol, Symbol}, Symbol})
#=    1.8 ms =# precompile(Tuple{typeof(Base.replace), String, Pair{String, String}})
#=    1.9 ms =# precompile(Tuple{Type{LLVM.MDTuple}, Ptr{LLVM.API.LLVMOpaqueMetadata}})
#=    1.5 ms =# precompile(Tuple{typeof(Base.length), Array{typeof(DataType), 1}})
#=    1.6 ms =# precompile(Tuple{Type{LLVM.PointerType}, Ptr{LLVM.API.LLVMOpaqueType}})
#=    1.6 ms =# precompile(Tuple{typeof(Base.getindex), Type{LLVM.LLVMType}})
#=    3.9 ms =# precompile(Tuple{typeof(Base.vect), LLVM.Function, Vararg{LLVM.Function}})
#=    1.9 ms =# precompile(Tuple{typeof(Base.getindex), Array{LLVM.LLVMType, 1}, Int64})
#=    2.5 ms =# precompile(Tuple{Type{LLVM.ConstantInt}, Int32})
#=    3.3 ms =# precompile(Tuple{typeof(Base.getproperty), GPUCompiler.MetalCompilerTarget, Symbol})
#=    1.5 ms =# precompile(Tuple{Type{Int32}, UInt32})
#=    0.0 ms =# precompile(Tuple{typeof(LLVMDowngrader_jll.llvm_downgrade)})
#=    0.0 ms =# precompile(Tuple{LLVMDowngrader_jll.var"##llvm_downgrade#8", Bool, Bool, typeof(LLVMDowngrader_jll.llvm_downgrade)})
#=    2.7 ms =# precompile(Tuple{typeof(Base.copy), Base.EnvDict})
#=    1.6 ms =# precompile(Tuple{typeof(Base.getindex), Base.RefValue{String}})
#=    0.0 ms =# precompile(Tuple{typeof(JLLWrappers.adjust_ENV!), Base.Dict{K, V} where V where K, String, String, Bool, Bool})
#=    1.6 ms =# precompile(Tuple{Type{Base.Cmd}, Array{String, 1}})
#=    1.6 ms =# precompile(Tuple{Type{NamedTuple{(:env,), T} where T<:Tuple}, Tuple{Base.Dict{String, String}}})
#=    3.1 ms =# precompile(Tuple{typeof(LLVM.name), LLVM.Function})
#=    1.2 ms =# precompile(Tuple{typeof(Base.convert), Type{Array{UInt8, 1}}, Array{UInt8, 1}})
#=    1.2 ms =# precompile(Tuple{typeof(Base.convert), Type{Base.UUID}, Base.UUID})
#=    1.4 ms =# precompile(Tuple{typeof(Base.Ryu.writefixed), Float64, Int64})
  0.762647 seconds (3.76 M allocations: 184.910 MiB, 4.32% gc time, 81.23% compilation time: <1% of which was recompilation)
#=    1.5 ms =# precompile(Tuple{Base.var"#606#607"{Base.WeakKeyDict{Metal.MTL.MTLCommandQueue, Nothing}}, Metal.MTL.MTLCommandQueue})
ChrichriMBP:~/.julia/dev/Metal$ j --proj --trace-compile=stderr --trace-compile-timing -e"using Metal; @time Metal.@sync @metal identity(nothing);"
#=    7.9 ms =# precompile(Tuple{typeof(Base.vect), Array{String, 1}, Vararg{Array{String, 1}}})
#=    5.5 ms =# precompile(Tuple{typeof(Base.iterate), Array{Array{String, 1}, 1}})
#=    4.8 ms =# precompile(Tuple{typeof(Base.vect), Base.Dict{String, Any}, Vararg{Any}})
#=    3.3 ms =# precompile(Tuple{typeof(Base.vect), Array{String, 1}, Vararg{Any}})
#=    9.1 ms =# precompile(Tuple{typeof(Base.getindex), Type{Union{Nothing, Array{String, 1}}}, Array{String, 1}, Nothing})
#=    2.8 ms =# precompile(Tuple{typeof(Base.iterate), Array{Union{Nothing, Array{String, 1}}, 1}})
#=    4.6 ms =# precompile(Tuple{typeof(Artifacts._artifact_str), Module, String, Base.SubString{String}, String, Base.Dict{String, Any}, Base.SHA1, Base.BinaryPlatforms.Platform, Base.Val{LazyArtifacts}})
#=   30.2 ms =# precompile(Tuple{typeof(Artifacts.__artifact_str), Module, String, Base.SubString{String}, String, Base.Dict{String, Any}, Base.SHA1, Base.BinaryPlatforms.Platform, Base.Val{LazyArtifacts}})
#=    0.0 ms =# precompile(Tuple{typeof(JLLWrappers.get_julia_libpaths)})
#=   36.0 ms =# precompile(Tuple{typeof(Base.get), Base.TTY, Symbol, Bool}) # recompile
#=    5.0 ms =# precompile(Tuple{typeof(Base.issorted), Array{String, 1}, Base.Order.ReverseOrdering{Base.Order.ForwardOrdering}})
#=   18.1 ms =# precompile(Tuple{typeof(GPUCompiler.assign_args!), Expr, Array{Any, 1}})
#=    2.0 ms =# precompile(Tuple{typeof(Base.indexed_iterate), Tuple{Array{Symbol, 1}, Array{Any, 1}}, Int64})
#=    1.7 ms =# precompile(Tuple{typeof(Base.indexed_iterate), Tuple{Array{Symbol, 1}, Array{Any, 1}}, Int64, Int64})
#=   27.8 ms =# precompile(Tuple{typeof(GPUCompiler.split_kwargs), Tuple{}, Array{Symbol, 1}, Vararg{Array{Symbol, 1}}})
#=    2.0 ms =# precompile(Tuple{typeof(Base.indexed_iterate), NTuple{4, Array{Any, 1}}, Int64})
#=    1.7 ms =# precompile(Tuple{typeof(Base.indexed_iterate), NTuple{4, Array{Any, 1}}, Int64, Int64})
#=    2.7 ms =# precompile(Tuple{typeof(Base.parse), Type{Bool}, String})
#=    4.4 ms =# precompile(Tuple{typeof(Base.getindex), Core.SimpleVector, Base.UnitRange{Int64}})
#=    2.1 ms =# precompile(Tuple{typeof(Base.Iterators.enumerate), Core.SimpleVector})
#=    2.8 ms =# precompile(Tuple{typeof(Base.iterate), Base.Iterators.Enumerate{Core.SimpleVector}, Tuple{Int64}})
#=    3.1 ms =# precompile(Tuple{typeof(Base.iterate), Base.Iterators.Enumerate{Core.SimpleVector}, Tuple{Int64, Int64}})
#=    1.6 ms =# precompile(Tuple{typeof(Base.getindex), Base.RefValue{Any}})
#=    2.4 ms =# precompile(Tuple{typeof(Base.indexed_iterate), Pair{Symbol, String}, Int64})
#=    2.0 ms =# precompile(Tuple{typeof(Base.indexed_iterate), Pair{Symbol, String}, Int64, Int64})
#=    4.2 ms =# precompile(Tuple{typeof(Base.haskey), Base.Dict{Symbol, Symbol}, Symbol})
#=    2.2 ms =# precompile(Tuple{typeof(Base.replace), String, Pair{String, String}})
#=    1.8 ms =# precompile(Tuple{Type{LLVM.MDTuple}, Ptr{LLVM.API.LLVMOpaqueMetadata}})
#=    1.2 ms =# precompile(Tuple{typeof(Base.length), Array{typeof(DataType), 1}})
#=    1.5 ms =# precompile(Tuple{Type{LLVM.PointerType}, Ptr{LLVM.API.LLVMOpaqueType}})
#=    1.5 ms =# precompile(Tuple{typeof(Base.getindex), Type{LLVM.LLVMType}})
#=    4.0 ms =# precompile(Tuple{typeof(Base.vect), LLVM.Function, Vararg{LLVM.Function}})
#=    2.0 ms =# precompile(Tuple{typeof(Base.getindex), Array{LLVM.LLVMType, 1}, Int64})
#=    2.4 ms =# precompile(Tuple{Type{LLVM.ConstantInt}, Int32})
#=    3.6 ms =# precompile(Tuple{typeof(Base.getproperty), GPUCompiler.MetalCompilerTarget, Symbol})
#=    1.7 ms =# precompile(Tuple{Type{Int32}, UInt32})
#=    0.0 ms =# precompile(Tuple{typeof(LLVMDowngrader_jll.llvm_downgrade)})
#=    0.0 ms =# precompile(Tuple{LLVMDowngrader_jll.var"##llvm_downgrade#8", Bool, Bool, typeof(LLVMDowngrader_jll.llvm_downgrade)})
#=    2.3 ms =# precompile(Tuple{typeof(Base.copy), Base.EnvDict})
#=    1.4 ms =# precompile(Tuple{typeof(Base.getindex), Base.RefValue{String}})
#=    0.0 ms =# precompile(Tuple{typeof(JLLWrappers.adjust_ENV!), Base.Dict{K, V} where V where K, String, String, Bool, Bool})
#=    1.8 ms =# precompile(Tuple{Type{Base.Cmd}, Array{String, 1}})
#=    1.5 ms =# precompile(Tuple{Type{NamedTuple{(:env,), T} where T<:Tuple}, Tuple{Base.Dict{String, String}}})
#=    3.3 ms =# precompile(Tuple{typeof(LLVM.name), LLVM.Function})
#=    1.2 ms =# precompile(Tuple{typeof(Base.convert), Type{Array{UInt8, 1}}, Array{UInt8, 1}})
#=    1.3 ms =# precompile(Tuple{typeof(Base.convert), Type{Base.UUID}, Base.UUID})
#=    3.7 ms =# precompile(Tuple{typeof(Core.kwcall), NamedTuple{(:msg,), Tuple{Nothing}}, typeof(Base.time_print), Base.TTY, Float64, Int64, Int64, Int64, Int64, Float64, Float64, Bool})
#=    1.3 ms =# precompile(Tuple{typeof(Base.Ryu.writefixed), Float64, Int64})
  0.188296 seconds (15.88 k allocations: 1.108 MiB, 34.58% compilation time: <1% of which was recompilation)
#=    1.4 ms =# precompile(Tuple{Base.var"#606#607"{Base.WeakKeyDict{Metal.MTL.MTLCommandQueue, Nothing}}, Metal.MTL.MTLCommandQueue})

This is potentially a Base julia issue (See JuliaLang/julia#61851 (comment) cc @adienes) but I'm. reporting here since I've personally only noticed it here.

[maleadt]

This is JuliaLang/julia#61711, right? Fixed by JuliaLang/julia#61714, but not backporteable. Workaround: https://github.com/maleadt/cuTile.jl/blob/8f514cc81c68f6dfc633b0007a47d7e711e453bf/src/precompile.jl#L3-L6
I don't remember why CUDA.jl doesn't need this.

[christiangnrd]

Hmm maybe? On a freshly compiled nightly, it's much worse (6s) but with a different statement that gets precompiled. Maybe that has do do with the new data layout warning that showed up. (it shows up in both with and without repl so this is probably a separate issue)

nightly:

Details

ChrichriMBP:~/.julia/dev/Metal$ j +dev --proj --trace-compile=stderr --trace-compile-timing -e"using REPL, Metal; @time Metal.@sync @metal identity(nothing);"
#=    8.4 ms =# precompile(Tuple{typeof(Artifacts._artifact_str), Module, String, Base.SubString{String}, String, Base.Dict{String, Any}, Base.SHA1, Base.BinaryPlatforms.Platform, Base.Val{LazyArtifacts}})
#=   24.8 ms =# precompile(Tuple{typeof(Artifacts.__artifact_str), Module, String, Base.SubString{String}, String, Base.Dict{String, Any}, Base.SHA1, Base.BinaryPlatforms.Platform, Base.Val{LazyArtifacts}})
#=   19.6 ms =# precompile(Tuple{typeof(GPUCompiler.assign_args!), Expr, Array{Any, 1}})
#=    3.2 ms =# precompile(Tuple{typeof(Base.indexed_iterate), Tuple{Array{Symbol, 1}, Array{Any, 1}}, Int64})
#=    3.1 ms =# precompile(Tuple{typeof(Base.indexed_iterate), Tuple{Array{Symbol, 1}, Array{Any, 1}}, Int64, Int64})
#=   29.9 ms =# precompile(Tuple{typeof(GPUCompiler.split_kwargs), Tuple{}, Array{Symbol, 1}, Array{Symbol, 1}, Array{Symbol, 1}})
#=    1.9 ms =# precompile(Tuple{typeof(Base.indexed_iterate), NTuple{4, Array{Any, 1}}, Int64})
#=    1.7 ms =# precompile(Tuple{typeof(Base.indexed_iterate), NTuple{4, Array{Any, 1}}, Int64, Int64})
#=    2.5 ms =# precompile(Tuple{typeof(Base.parse), Type{Bool}, String})
#=    1.3 ms =# precompile(Tuple{typeof(Base.:(>)), UInt32, UInt32})
#=    4.7 ms =# precompile(Tuple{typeof(Base.:(>)), Base.VersionNumber, Base.VersionNumber})
#=    1.2 ms =# precompile(Tuple{typeof(Base.in), Symbol, NTuple{8, Symbol}})
#=    4.0 ms =# precompile(Tuple{Base.Compiler.var"#86#87"{Base.Compiler.var"#abstract_call##2#abstract_call##3", Base.Compiler.Future{Any}, Base.RefValue{Base.Compiler.CallMeta}}, Base.Compiler.NativeInterpreter, Base.Compiler.IRInterpretationState{Base.Compiler.NativeInterpreter}}) # recompile
#=    3.6 ms =# precompile(Tuple{Base.Compiler.var"#86#87"{Base.Compiler.var"#abstract_eval_call##0#abstract_eval_call##1", Base.Compiler.Future{Base.Compiler.RTEffects}, Base.RefValue{Base.Compiler.CallMeta}}, Base.Compiler.NativeInterpreter, Base.Compiler.IRInterpretationState{Base.Compiler.NativeInterpreter}}) # recompile
#=    1.8 ms =# precompile(Tuple{typeof(Base.something), String, String})
#=    2.0 ms =# precompile(Tuple{typeof(Base.:(<=)), UInt8, UInt64})
#=    2.1 ms =# precompile(Tuple{typeof(Base.:(!=)), UInt32, UInt32})
#=    4.3 ms =# precompile(Tuple{typeof(Base.:(*)), Vararg{String, 6}})
#= 6102.4 ms =# precompile(Tuple{typeof(GPUCompiler.compile), Symbol, GPUCompiler.CompilerJob{GPUCompiler.MetalCompilerTarget, Metal.MetalCompilerParams}})
#=    6.4 ms =# precompile(Tuple{typeof(Base.getindex), Core.SimpleVector, Base.UnitRange{Int64}})
#=    1.6 ms =# precompile(Tuple{typeof(Base.Iterators.enumerate), Core.SimpleVector})
#=    3.8 ms =# precompile(Tuple{typeof(Base.iterate), Base.Iterators.Enumerate{Core.SimpleVector}, Tuple{Int64}})
#=    3.7 ms =# precompile(Tuple{typeof(Base.iterate), Base.Iterators.Enumerate{Core.SimpleVector}, Tuple{Int64, Int64}})
#=    3.9 ms =# precompile(Tuple{typeof(LLVM.triple!), LLVM.Module, String})
#=    1.4 ms =# precompile(Tuple{typeof(Base.getindex), Base.RefValue{Any}})
#=    6.2 ms =# precompile(Tuple{typeof(Base.rest), Core.SimpleVector, Int64})
#=    3.0 ms =# precompile(Tuple{typeof(Base.indexed_iterate), Pair{Symbol, String}, Int64})
#=    3.1 ms =# precompile(Tuple{typeof(Base.indexed_iterate), Pair{Symbol, String}, Int64, Int64})
#=    5.0 ms =# precompile(Tuple{typeof(Base.haskey), Base.Dict{Symbol, Symbol}, Symbol})
#=    2.9 ms =# precompile(Tuple{typeof(Base.replace), String, Pair{String, String}})
#=    1.4 ms =# precompile(Tuple{Type{LLVM.MDTuple}, Ptr{LLVM.API.LLVMOpaqueMetadata}})
#=    1.2 ms =# precompile(Tuple{typeof(Base.length), Array{typeof(DataType), 1}})
#=    1.4 ms =# precompile(Tuple{Type{LLVM.PointerType}, Ptr{LLVM.API.LLVMOpaqueType}})
warning: Linking two modules of different data layouts: '/Users/christian/.julia/scratchspaces/61eb1bfa-7361-4325-ad38-22787b887f55/compiled/v1.14/v1.22/runtime_metal-macos26.5.1-debuginfo=2-opaque.bc' is 'e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v24:32:32-v32:32:32-v48:64:64-v64:64:64-v96:128:128-v128:128:128-v192:256:256-v256:256:256-v512:512:512-v1024:1024:1024-n8:16:32-ni:10:11:12:13' whereas 'start' is 'e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v24:32:32-v32:32:32-v48:64:64-v64:64:64-v96:128:128-v128:128:128-v192:256:256-v256:256:256-v512:512:512-v1024:1024:1024-n8:16:32'

#=    1.5 ms =# precompile(Tuple{typeof(Base.getindex), Type{LLVM.LLVMType}})
#=    6.5 ms =# precompile(Tuple{typeof(Base.vect), LLVM.Function, LLVM.Function})
#=    1.7 ms =# precompile(Tuple{typeof(Base.getindex), Array{LLVM.LLVMType, 1}, Int64})
#=    1.8 ms =# precompile(Tuple{Type{LLVM.ConstantInt}, Int32})
#=    5.2 ms =# precompile(Tuple{typeof(Base.getproperty), GPUCompiler.MetalCompilerTarget, Symbol})
#=    1.5 ms =# precompile(Tuple{Type{Int32}, UInt32})
#=    3.4 ms =# precompile(Tuple{typeof(Base.copy), Base.EnvDict})
#=    1.3 ms =# precompile(Tuple{typeof(Base.getindex), Base.RefValue{String}})
#=    2.8 ms =# precompile(Tuple{Type{Base.Cmd}, Array{String, 1}})
#=    1.4 ms =# precompile(Tuple{Type{NamedTuple{(:env,), T} where T<:Tuple}, Tuple{Base.Dict{String, String}}})
#=    4.1 ms =# precompile(Tuple{typeof(Core.kwcall), NamedTuple{(:env,), Tuple{Base.Dict{String, String}}}, Type{Base.Cmd}, Base.Cmd})
#=    3.4 ms =# precompile(Tuple{typeof(LLVM.name), LLVM.Function})
#=    1.1 ms =# precompile(Tuple{typeof(Base.convert), Type{Array{UInt8, 1}}, Array{UInt8, 1}})
#=    1.1 ms =# precompile(Tuple{typeof(Base.convert), Type{Base.UUID}, Base.UUID})
  6.308433 seconds (61.45 M allocations: 3.941 GiB, 7.13% gc time, 98.09% compilation time: <1% of which was recompilation)
#=    1.8 ms =# precompile(Tuple{Base.WeakKeyDictFinalizer{Base.WeakKeyDict{Metal.MTL.MTLCommandQueue, Nothing}}, Metal.MTL.MTLCommandQueue})

ChrichriMBP:~/.julia/dev/Metal$ j +dev --proj --trace-compile=stderr --trace-compile-timing -e"using Metal; @time Metal.@sync @metal identity(nothing);"
#=    8.6 ms =# precompile(Tuple{typeof(Artifacts._artifact_str), Module, String, Base.SubString{String}, String, Base.Dict{String, Any}, Base.SHA1, Base.BinaryPlatforms.Platform, Base.Val{LazyArtifacts}})
#=   25.5 ms =# precompile(Tuple{typeof(Artifacts.__artifact_str), Module, String, Base.SubString{String}, String, Base.Dict{String, Any}, Base.SHA1, Base.BinaryPlatforms.Platform, Base.Val{LazyArtifacts}})
#=   19.6 ms =# precompile(Tuple{typeof(GPUCompiler.assign_args!), Expr, Array{Any, 1}})
#=    3.3 ms =# precompile(Tuple{typeof(Base.indexed_iterate), Tuple{Array{Symbol, 1}, Array{Any, 1}}, Int64})
#=    3.0 ms =# precompile(Tuple{typeof(Base.indexed_iterate), Tuple{Array{Symbol, 1}, Array{Any, 1}}, Int64, Int64})
#=   29.7 ms =# precompile(Tuple{typeof(GPUCompiler.split_kwargs), Tuple{}, Array{Symbol, 1}, Array{Symbol, 1}, Array{Symbol, 1}})
#=    1.9 ms =# precompile(Tuple{typeof(Base.indexed_iterate), NTuple{4, Array{Any, 1}}, Int64})
#=    1.6 ms =# precompile(Tuple{typeof(Base.indexed_iterate), NTuple{4, Array{Any, 1}}, Int64, Int64})
#=    2.5 ms =# precompile(Tuple{typeof(Base.parse), Type{Bool}, String})
#=    5.8 ms =# precompile(Tuple{typeof(Base.getindex), Core.SimpleVector, Base.UnitRange{Int64}})
#=    1.4 ms =# precompile(Tuple{typeof(Base.Iterators.enumerate), Core.SimpleVector})
#=    3.5 ms =# precompile(Tuple{typeof(Base.iterate), Base.Iterators.Enumerate{Core.SimpleVector}, Tuple{Int64}})
#=    3.6 ms =# precompile(Tuple{typeof(Base.iterate), Base.Iterators.Enumerate{Core.SimpleVector}, Tuple{Int64, Int64}})
#=    3.5 ms =# precompile(Tuple{typeof(LLVM.triple!), LLVM.Module, String})
#=    1.3 ms =# precompile(Tuple{typeof(Base.getindex), Base.RefValue{Any}})
#=    5.7 ms =# precompile(Tuple{typeof(Base.rest), Core.SimpleVector, Int64})
#=    3.2 ms =# precompile(Tuple{typeof(Base.indexed_iterate), Pair{Symbol, String}, Int64})
#=    3.1 ms =# precompile(Tuple{typeof(Base.indexed_iterate), Pair{Symbol, String}, Int64, Int64})
#=    5.5 ms =# precompile(Tuple{typeof(Base.haskey), Base.Dict{Symbol, Symbol}, Symbol})
#=    2.9 ms =# precompile(Tuple{typeof(Base.replace), String, Pair{String, String}})
#=    1.5 ms =# precompile(Tuple{Type{LLVM.MDTuple}, Ptr{LLVM.API.LLVMOpaqueMetadata}})
#=    1.3 ms =# precompile(Tuple{typeof(Base.length), Array{typeof(DataType), 1}})
#=    1.9 ms =# precompile(Tuple{Type{LLVM.PointerType}, Ptr{LLVM.API.LLVMOpaqueType}})
warning: Linking two modules of different data layouts: '/Users/christian/.julia/scratchspaces/61eb1bfa-7361-4325-ad38-22787b887f55/compiled/v1.14/v1.22/runtime_metal-macos26.5.1-debuginfo=2-opaque.bc' is 'e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v24:32:32-v32:32:32-v48:64:64-v64:64:64-v96:128:128-v128:128:128-v192:256:256-v256:256:256-v512:512:512-v1024:1024:1024-n8:16:32-ni:10:11:12:13' whereas 'start' is 'e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v24:32:32-v32:32:32-v48:64:64-v64:64:64-v96:128:128-v128:128:128-v192:256:256-v256:256:256-v512:512:512-v1024:1024:1024-n8:16:32'

#=    1.6 ms =# precompile(Tuple{typeof(Base.getindex), Type{LLVM.LLVMType}})
#=    6.7 ms =# precompile(Tuple{typeof(Base.vect), LLVM.Function, LLVM.Function})
#=    1.7 ms =# precompile(Tuple{typeof(Base.getindex), Array{LLVM.LLVMType, 1}, Int64})
#=    1.7 ms =# precompile(Tuple{Type{LLVM.ConstantInt}, Int32})
#=    5.1 ms =# precompile(Tuple{typeof(Base.getproperty), GPUCompiler.MetalCompilerTarget, Symbol})
#=    1.4 ms =# precompile(Tuple{Type{Int32}, UInt32})
#=    3.2 ms =# precompile(Tuple{typeof(Base.copy), Base.EnvDict})
#=    1.3 ms =# precompile(Tuple{typeof(Base.getindex), Base.RefValue{String}})
#=    2.8 ms =# precompile(Tuple{Type{Base.Cmd}, Array{String, 1}})
#=    1.3 ms =# precompile(Tuple{Type{NamedTuple{(:env,), T} where T<:Tuple}, Tuple{Base.Dict{String, String}}})
#=    4.1 ms =# precompile(Tuple{typeof(Core.kwcall), NamedTuple{(:env,), Tuple{Base.Dict{String, String}}}, Type{Base.Cmd}, Base.Cmd})
#=    3.4 ms =# precompile(Tuple{typeof(LLVM.name), LLVM.Function})
#=    1.0 ms =# precompile(Tuple{typeof(Base.convert), Type{Array{UInt8, 1}}, Array{UInt8, 1}})
#=    1.1 ms =# precompile(Tuple{typeof(Base.convert), Type{Base.UUID}, Base.UUID})
#=    6.6 ms =# precompile(Tuple{typeof(Core.kwcall), NamedTuple{(:msg,), Tuple{Nothing}}, typeof(Base.time_print), Base.TTY, Float64, Int64, Int64, Int64, Int64, Float64, Float64, Bool})
  0.174731 seconds (9.63 k allocations: 831.250 KiB, 48.07% compilation time: <1% of which was recompilation)
#=    1.6 ms =# precompile(Tuple{Base.WeakKeyDictFinalizer{Base.WeakKeyDict{Metal.MTL.MTLCommandQueue, Nothing}}, Metal.MTL.MTLCommandQueue})

[christiangnrd]

Is it expected that JuliaLang/julia#61714 is required for the workaround to work? That's the behaviour I'm seeing with the following patch applied

diff --git a/Project.toml b/Project.toml
index fbb6518d5..8961df953 100644
--- a/Project.toml
+++ b/Project.toml
@@ -23,6 +23,7 @@ ObjectiveC = "e86c9b32-1129-44ac-8ea0-90d5bb39ded9"
 PrecompileTools = "aea7be01-6a6a-4083-8856-8a6e6704d82a"
 Preferences = "21216c6a-2e73-6563-6e65-726566657250"
 Printf = "de0858da-6303-5e67-8744-51eddeeeb8d7"
+REPL = "3fa0cd96-eef1-5676-8a61-b3b8758bbffb"
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
 Random123 = "74087812-796a-5b5d-8853-05524746bad3"
 RandomNumbers = "e6cf234a-135c-5ec9-84dd-332b85af5143"
@@ -56,6 +57,7 @@ ObjectiveC = "5"
 PrecompileTools = "1"
 Preferences = "1"
 Printf = "1"
+REPL = "1.11.0"
 Random = "1"
 Random123 = "1.7.1"
 RandomNumbers = "1.6.0"
diff --git a/src/precompile.jl b/src/precompile.jl
index efd3a1a63..0bca57851 100644
--- a/src/precompile.jl
+++ b/src/precompile.jl
@@ -1,5 +1,7 @@
 using PrecompileTools: @setup_workload, @compile_workload

+import REPL
+
 Sys.isapple() && @setup_workload begin
     metallib_file = joinpath(dirname(@__DIR__), "test", "dummy.metallib")

[maleadt]

Is it expected that JuliaLang/julia#61714 is required for the workaround to work?

No. As you can see in the MWE from JuliaLang/julia#61711, depending on REPL pre-fix removed all invalidations.

[christiangnrd]

No. As you can see in the MWE from JuliaLang/julia#61711, depending on REPL pre-fix removed all invalidations.

Very strange, on 1.12 (no pr 61771) with the workaround, I get ~100ms faster precompilation of the offending invalidation, but seems like the expected outcome is no invalidation. On 1.14, I get similar (fast) ttfx both with and without a loaded REPL