Incorrect SSA value type when using `julia --code-coverage`

[KDr2]

I have a t.jl:

using CodeInfoTools

struct S x; y; end

c = CodeInfoTools.code_inferred(S, Int, Int)

@show c
@show c.code
@show c.slottypes
@show c.ssavaluetypes

When I run julia t.jl, it produces the correct output:

c = CodeInfo(
    @ /data4/zhuo/Work/julia/Libtask.jl/test/runtests.jl:3 within `S`
1 ─ %1 = Main.S::Core.Const(S)
│   %2 = %new(%1, x, y)::S
└──      return %2
)
c.code = Any[:(Main.S), :(%new(%1, _2, _3)), :(return %2)]
c.slottypes = Any[Core.Const(S), Int64, Int64]
c.ssavaluetypes = Any[Core.Const(S), S, Any]

But when I run julia --code-coverage=user t.jl, I get:

c = CodeInfo(
    @ /data4/zhuo/Work/julia/Libtask.jl/test/runtests.jl:3 within `S`
1 ─ %1 = Main.S::Nothing
│   %2 = %new(%1, x, y)::Core.Const(S)
└──      return %2
)
c.code = Any[:(Main.S), :(%new(%1, _2, _3)), :(return %2)]
c.slottypes = Any[Core.Const(S), Int64, Int64]
c.ssavaluetypes = Any[Nothing, Core.Const(S), S, Any]

There's an extra Nothing prepended to the ssavalutypes.

It's mainly caused by the --code-coverage option:

• correct: julia t.jl
• correct: julia --code-coverage=none t.jl
• incorrect: julia --code-coverage=user t.jl
• incorrect: julia --code-coverage=all t.jl

[KDr2]

The real reason is not an extra Nothing is prepended, but an expression is removed:

julia> code_typed(S, Tuple{Int, Int})
1-element Vector{Any}:
 CodeInfo(
1 ─      $(Expr(:code_coverage_effect))::Nothing # <-- this expr is removed, but its type isn't
│   %2 = Main.S::Type{S}
│   %3 = %new(%2, x, y)::S
└──      return %3
) => S

[femtomc]

@KDr2 did you end up figuring a workaround for this?

I can't tell how much of this is an issue with CodeInfoTools -- vs. the fact that Julia will insert these "marker" expressions for coverage.

[KDr2]

My workaround is not to use the --code-coverage option so far. :)

[Roger-luo]

FYI: https://github.com/JuliaCompilerPlugins/CompilerPluginTools.jl/blob/master/src/utils.jl#L82

you can just use this transform on the generated CodeInfo

[KDr2]

FYI: https://github.com/JuliaCompilerPlugins/CompilerPluginTools.jl/blob/master/src/utils.jl#L82

you can just use this transform on the generated CodeInfo

Thank you @Roger-luo, but it has no effect on this issue:

$ cat t.jl
using CodeInfoTools
using CompilerPluginTools

struct S x; y; end

c = CodeInfoTools.code_inferred(S, Int, Int)
c = CompilerPluginTools.rm_code_coverage_effect(c)

@show c
@show c.code
@show c.slottypes
@show c.ssavaluetypes


$ julia --project t.jl
c = CodeInfo(
    @ /data4/zhuo/Work/julia/Libtask.jl/t.jl:4 within `S`
1 ─ %1 = Main.S::Const(S)
│   %2 = %new(%1, x, y)::S
└──      return %2
)
c.code = Any[:(Main.S), :(%new(%1, _2, _3)), :(return %2)]
c.slottypes = Any[Const(S), Int64, Int64]
c.ssavaluetypes = Any[Const(S), S, Any] # <--- HERE is correct

$ julia --code-coverage --project t.jl
c = CodeInfo(
    @ /data4/zhuo/Work/julia/Libtask.jl/t.jl:4 within `S`
1 ─ %1 = Main.S::Nothing
│   %2 = %new(%1, x, y)::Const(S)
└──      return %2
)
c.code = Any[:(Main.S), :(%new(%1, _2, _3)), :(return %2)]
c.slottypes = Any[Const(S), Int64, Int64]
c.ssavaluetypes = Any[Nothing, Const(S), S]  # <--- HERE still a prepended Nothing

[Roger-luo]

I think you don't want to use CodeInferred since that contains the type lattice object. That pass should not be run with type lattice as the name suggested it removes the coverage effect stmts. So you should just use it on the final codeinfo. I'm not sure why there's code inferred in this package but the CodeInfo manipulation should be better written in IRCode and we should only use CodeInfo as a final product after them. Unless you have a custom abstract interpreter.

[KDr2]

But I need the type info of the CodeInfo object. And rm_code_coverage_effect doesn't indicate anything about types, i.e., it doesn't say it can/can't deal with types, so I think if a CodeInfo contains IR codes and types, it may remove coverage related things from both of them.

Anyway, I found Base.code_typed returns the same code info (with types and without errors), so I changed to that function.

Thank you.