JET tutorial
JET leverages the Julia compiler's inference to check user code for type instability and type errors. This tutorial will demonstrate how to use JET effectively. It presupposes the reader has working knowledge of Julia, and understands Julian concepts such as "dynamic dispatch" and "type stability" - see the Julia documentation for explanation of these concepts.
Because JET relies on the compiler's type inference, it is not able to effectively analyze type unstable code. Making your code type stable is a prerequisite for effectively using JET's type error analysis. Therefore, we will begin by showing how to use JET to fix type instabilities.
First of all, you need to install JET: JET is a standard Julia package. So you can just install it via Julia's built-in package manager and use it just like any other package:
julia> using Pkg; Pkg.add("JET")
[ some output elided ]
julia> using JETDetecting type instability with @report_opt
JET exports a function report_opt and the related macro @report_opt. It works similar to the function/macro pair (@)code_warntype from Base - except that it automatically analyses all the way down the function chain, and that it only displays any issues found.
For example, suppose we have the function:
julia> add_one_first(x) = first(x) + 1;
Any type instabilities of a given function call can be analysed thus:
julia> @report_opt add_one_first([1])No errors detectedjulia> @report_opt add_one_first(Any[1])═════ 1 possible error found ═════ ┌ add_one_first(x::Vector{Any}) @ Main ./REPL[1]:1 │ runtime dispatch detected: (%17::Any Main.:+ 1)::Any └────────────────────
You can see that add_one_first is type stable when called with a Vector{Int}, but it leads to dynamic dispatch when called with Vector{Any}.
Suppose now we have two levels of type instability, where one type instability "hides behind" another type instability, as in this example:
julia> add_one_first(x) = first(x) + 1;julia> func_var = add_one_first;julia> f(x) = func_var(x);julia> @report_opt f(Any[1])═════ 1 possible error found ═════ ┌ f(x::Vector{Any}) @ Main ./REPL[3]:2 │ runtime dispatch detected: %1::Any(x::Vector{Any})::Any └────────────────────
The dynamic dispatch we see here come from the fact that func_var is an untyped global variable. Remember that add_one_first was also type unstable when called with a Vector{Any}. So why doesn't @report_opt report that second type instability from calling add_one_first(::Vector{Any})? The reason is that because the Julia compiler does not know at compile time that func_var is equal to add_one_first, JET cannot "see through" the first type instability and see that add_one_first(Any[1]) will eventually be called.
If we fix the first instability by defining the global variable my_func_var as const:
julia> const my_func_var = add_one_first;
Then the compiler knows that add_one_first will be called, and the second type instability from this function is revealed:
julia> f(x) = my_func_var(x)f (generic function with 1 method)julia> @report_opt f(Any[1])═════ 1 possible error found ═════ ┌ f(x::Vector{Any}) @ Main ./REPL[1]:1 │┌ add_one_first(x::Vector{Any}) @ Main ./REPL[1]:1 ││ runtime dispatch detected: (%17::Any Main.:+ 1)::Any │└────────────────────
Sometimes, type instability only shows up much deeper into a call chain, several functions deep. This is not a problem for JET. In the example below, JET sees type instability ~10 function calls deep:
julia> @report_opt sum(Any[1])═════ 4 possible errors found ═════ ┌ sum(a::Vector{Any}) @ Base ./reducedim.jl:980 │┌ sum(a::Vector{Any}; dims::Colon, kw::@Kwargs{}) @ Base ./reducedim.jl:980 ││┌ _sum(a::Vector{Any}, ::Colon) @ Base ./reducedim.jl:984 │││┌ _sum(a::Vector{Any}, ::Colon; kw::@Kwargs{}) @ Base ./reducedim.jl:984 ││││┌ _sum(f::typeof(identity), a::Vector{Any}, ::Colon) @ Base ./reducedim.jl:985 │││││┌ _sum(f::typeof(identity), a::Vector{Any}, ::Colon; kw::@Kwargs{}) @ Base ./reducedim.jl:985 ││││││┌ mapreduce(f::typeof(identity), op::typeof(Base.add_sum), A::Vector{Any}) @ Base ./reducedim.jl:327 │││││││┌ mapreduce(f::typeof(identity), op::typeof(Base.add_sum), A::Vector{Any}; dims::Colon, init::Base._InitialValue) @ Base ./reducedim.jl:327 ││││││││┌ _mapreduce_dim(f::typeof(identity), op::typeof(Base.add_sum), ::Base._InitialValue, A::Vector{Any}, ::Colon) @ Base ./reducedim.jl:335 │││││││││┌ _mapreduce(f::typeof(identity), op::typeof(Base.add_sum), ::IndexLinear, A::Vector{Any}) @ Base ./reduce.jl:444 ││││││││││┌ mapreduce_impl(f::typeof(identity), op::typeof(Base.add_sum), A::Vector{Any}, ifirst::Int64, ilast::Int64) @ Base ./reduce.jl:277 │││││││││││┌ mapreduce_impl(f::typeof(identity), op::typeof(Base.add_sum), A::Vector{Any}, ifirst::Int64, ilast::Int64, blksize::Int64) @ Base ./reduce.jl:262 ││││││││││││ runtime dispatch detected: op::typeof(Base.add_sum)(%91::Any, %110::Any)::Any │││││││││││└──────────────────── │││││││││││┌ mapreduce_impl(f::typeof(identity), op::typeof(Base.add_sum), A::Vector{Any}, ifirst::Int64, ilast::Int64, blksize::Int64) @ Base ./reduce.jl:273 ││││││││││││ runtime dispatch detected: op::typeof(Base.add_sum)(%187::Any, %189::Any)::Any │││││││││││└──────────────────── │││││││││┌ _mapreduce(f::typeof(identity), op::typeof(Base.add_sum), ::IndexLinear, A::Vector{Any}) @ Base ./reduce.jl:437 ││││││││││ runtime dispatch detected: op::typeof(Base.add_sum)(%97::Any, %115::Any)::Any │││││││││└──────────────────── │││││││││┌ _mapreduce(f::typeof(identity), op::typeof(Base.add_sum), ::IndexLinear, A::Vector{Any}) @ Base ./reduce.jl:440 ││││││││││ runtime dispatch detected: op::typeof(Base.add_sum)(%118::Any, %139::Any)::Any │││││││││└────────────────────
As mentioned above, effective use of JET begins with liberal use of @report_opt to eliminate or reduce any dynamic dispatch, such that the Julia compiler is not blinded by dynamic dispatch.
After the program has been made as type stable as possible, it's time to use @report_call to find type errors.
Analyse methods with @report_call
The function/macro pair report_call and @report_call works just like (@)report_opt - but where the latter reports dynamic dispatch, the former finds type errors.
The @report_call macro analyses function calls like so:
julia> @report_call sum(['a'])═════ 2 possible errors found ═════ ┌ sum(a::Vector{Char}) @ Base ./reducedim.jl:980 │┌ sum(a::Vector{Char}; dims::Colon, kw::@Kwargs{}) @ Base ./reducedim.jl:980 ││┌ _sum(a::Vector{Char}, ::Colon) @ Base ./reducedim.jl:984 │││┌ _sum(a::Vector{Char}, ::Colon; kw::@Kwargs{}) @ Base ./reducedim.jl:984 ││││┌ _sum(f::typeof(identity), a::Vector{Char}, ::Colon) @ Base ./reducedim.jl:985 │││││┌ _sum(f::typeof(identity), a::Vector{Char}, ::Colon; kw::@Kwargs{}) @ Base ./reducedim.jl:985 ││││││┌ mapreduce(f::typeof(identity), op::typeof(Base.add_sum), A::Vector{Char}) @ Base ./reducedim.jl:327 │││││││┌ mapreduce(f::typeof(identity), op::typeof(Base.add_sum), A::Vector{Char}; dims::Colon, init::Base._InitialValue) @ Base ./reducedim.jl:327 ││││││││┌ _mapreduce_dim(f::typeof(identity), op::typeof(Base.add_sum), ::Base._InitialValue, A::Vector{Char}, ::Colon) @ Base ./reducedim.jl:335 │││││││││┌ _mapreduce(f::typeof(identity), op::typeof(Base.add_sum), ::IndexLinear, A::Vector{Char}) @ Base ./reduce.jl:429 ││││││││││┌ mapreduce_empty_iter(f::typeof(identity), op::typeof(Base.add_sum), itr::Vector{Char}, ItrEltype::Base.HasEltype) @ Base ./reduce.jl:377 │││││││││││┌ reduce_empty_iter(op::Base.MappingRF{typeof(identity), typeof(Base.add_sum)}, itr::Vector{Char}, ::Base.HasEltype) @ Base ./reduce.jl:381 ││││││││││││┌ reduce_empty(op::Base.MappingRF{typeof(identity), typeof(Base.add_sum)}, ::Type{Char}) @ Base ./reduce.jl:358 │││││││││││││┌ mapreduce_empty(::typeof(identity), op::typeof(Base.add_sum), T::Type{Char}) @ Base ./reduce.jl:369 ││││││││││││││┌ reduce_empty(::typeof(Base.add_sum), ::Type{Char}) @ Base ./reduce.jl:350 │││││││││││││││┌ reduce_empty(::typeof(+), ::Type{Char}) @ Base ./reduce.jl:343 ││││││││││││││││ no matching method found `zero(::Type{Char})`: zero(T::Type{Char}) │││││││││││││││└──────────────────── │││││││││┌ _mapreduce(f::typeof(identity), op::typeof(Base.add_sum), ::IndexLinear, A::Vector{Char}) @ Base ./reduce.jl:437 ││││││││││┌ add_sum(x::Char, y::Char) @ Base ./reduce.jl:24 │││││││││││ no matching method found `+(::Char, ::Char)`: (x::Char + y::Char) ││││││││││└────────────────────
In this example, JET found two possible type errors:
- If the input vector is empty, the function call will error with a
MethodErrorafter attempting to callzero(Char). - If the input vector has two or more elements, the call will error after attempting to call
+(::Char, ::Char).
Note that these type errors show up even though the input vector had exactly one element, and so neither of these errors would actually occur at runtime if sum(['a']) had been executed. This happens because JET analyses the code on a type level, only looking at the code generated with the input types. It does not analyze what will actually happen with the given input value.
Note also that the two possible errors shown are mutally exclusive - no input will lead to both errors. Nonetheless, JET is able to detect both possibilities, because it analyses all possible branches in the generated function call.
In contrast, if we analyse the same sum method on a Vector{Int} instead of Vector{Char}:
julia> @report_call sum([1])No errors detected
The two errors above do not appear. These errors do not apply to Vector{Int}, because zero(Int) is well-defined and so is +(::Int, ::Int). This illustrates that JET does not analyze methods, but rather function calls with specific types. More precisely, JET analyses so-called methodinstances, since methodinstances are compiled whereas methods are not.
Analyse whole packages with report_package
As shown above, methods cannot be analyzed, but only methodinstances, i.e. methods plus the types of their arguments. Most packages, however, define only methods, and do not contain callsites. That is, they do not have any information about the types that these methods will be called with.
However, JET is able to do limited analysis using only the method signature extracted from the method definition. For example, if I define this simple function:
first_plus_n(itr, n::Real) = first(itr) + n;, then at the very least, we can guarantee that itr isa Any and n isa Real. Hence, JET can analyze the methodinstance first_plus_n(::Any, ::Real), using only the method definition.
The JET function report_package extracts all method definitions in a package, and using the extracted signatures, runs report_call on them. For example, the package BioSymbols can be analysed like this:
julia> using JET
julia> report_package(BioSymbols)
[ output elided ]Note that report_package is less precise than @report_call, because method signatures of idiomatic Julia code are often very generic, so there is less type information in the signature itself than there is when given concrete argument types.
Usage tips
Use @report_opt before @report_call
JET works best on type-stable code. Iron out type instabilities using @report_opt before using @report_call
Filtering away false positives
It is common to find that JET finds lots of errors in your functions, which all derive from type instability and type issues in your dependencies. In fact, type issues from dependencies are often so plentiful they flood your analysis with false positives, which can make working with JET harder.
To reduce false positives, you can use the keywords ignored_modules and target_modules.Both take an iterable of modules.
The former removes any errors that originate from any of the given modules, while the latter removes any errors except ones originating from these modules.
For example, in the REPL (which is in module Main), we can define:
julia> g(x) = first(x) + 1g (generic function with 1 method)
This throws in a Base function if we pass nothing into it:
julia> @report_call g(nothing)═════ 1 possible error found ═════ ┌ g(x::Nothing) @ Main ./REPL[1]:1 │┌ first(itr::Nothing) @ Base ./abstractarray.jl:472 ││ no matching method found `iterate(::Nothing)`: x = iterate(itr::Nothing) │└────────────────────
Since the error originates from Base, we can filter the error away by ignoring Base, or equivalently, we may retain only the ones from Main. Note that we pass (Base,) as a 1-element Tuple of modules:
julia> @report_call ignored_modules=(Base,) g(nothing)No errors detectedjulia> @report_call target_modules=(@__MODULE__,) g(nothing)No errors detected
The AnyFrameModule construct can be used to filter for (or against) any error where any of the function calls in the callchain originates from the given module. For example, in the example above, the function call begins in Main and ends in Base, so the callchain includes both modules. Ignoring AnyFrameModule(Base) or AnyFrameModule(Main) will then ignore the error:
julia> @report_call ignored_modules=(AnyFrameModule(Base),) g(nothing)No errors detectedjulia> @report_call ignored_modules=(AnyFrameModule(@__MODULE__),) g(nothing)No errors detected
Similarly, the error would be retained if target_modules would have been AnyFrameModule(Base) or AnyFrameModule(Main).
Beware that this filtering may filter away legitimate problems in your package. In the example above, if your code calls f(nothing), the error may originate from Base, but it's clearly an error in your own code to call f(nothing). So it is recommended to only filter away modules if they produce so many false positives that it makes using JET difficult.
Analyze scripts and apps by using a main function
Scripts and apps called from the command line have a single logical entry point. If you wrap the logic in a main function, JET can analyse the entire script.
For example, suppose you made this command-line script which added two numbers from the command line:
a = parse(Int, first(ARGS))
b = parse(Int, last(ARGS))
println(a + b)You could rewrite this as:
function main()
a = parse(Int, first(ARGS))
b = parse(Int, last(ARGS))
println(a + b)
end
main(), and then analyse the script with JET using @report_call main(). Using a main function has the further advantage that it makes it for other people to understand what your script does when invoked.
Alternatively, JET also provides the function report_file, which you can call like:
julia> report_file("my_script.jl"), which will be equivalent to checking @report_call main(), if the script contains a main() call at top level.
Analyze packages using a representative workload
As shown above, packages can be analysed with report_package. However, generic type signatures often used in packages lead to imprecise inference and thus imprecise analysis.
To improve analysis, you can create a file src/workload.jl in your package, which uses all (or most) functionality of the package. The function could look like:
function exercise_mypkg()
data = MyPkg.load_data()
transformed = MyPkg.transform_data(data)
# [ etc ...]
endBecause such usage necessarily requires passing concrete types to your functions, calling @report_call exercise_mypkg() leads to more precise analysis than report_package.
Furthermore, once you have written a function like exercise_mypkg, you can use a package like PrecompileTools to precompile the function, which will thus precompile all code exercised in the function, significantly reducing your package's latency.
Conversely, if you've already implemented a precompile workload, you can do the following:
using MyPkg, JET, MethodAnalysis
mis = methodinstances(MyPkg) # get all the compiled methodinstances for functions owned by the package
# Now let's filter out the ones that pass without issue
badmis = filter(mis) do mi
!isempty(JET.get_reports(report_call(mi)))
endThen you can inspect the methodinstances in badmis individually with report_call(mi).
There are two caveats to note:
methodinstances(MyPkg)only covers functions owned byMyPkg. IfMyPkgdefines an "extension" methodOtherPkg.f(...), any corresponding methodinstances would appear in the list forOtherPkg.- if you disable precompilation for your development version of the package, you'll get a greatly-reduced list of methodinstances that does not reflect ordinary usage. This style of JET analysis should only be performed on packages that are actively using their precompilation workloads.