passing same argument twice leads to a bounds error while compiling

[avik-pal]

julia> using Reactant

julia> foo(x, y) = x
foo (generic function with 1 method)

julia> x_ra = Reactant.to_rarray(rand(2, 2))
2×2 ConcreteRArray{Float64, 2}:
 0.806331  0.636137
 0.349127  0.0385477

julia> @compile foo(x_ra, x_ra)
ERROR: BoundsError: attempt to access Tuple{} at index [2]
Stacktrace:
 [1] getindex(t::Tuple, i::Int64)
   @ Base ./tuple.jl:31
 [2] codegen_flatten!(linear_args::Vector{Union{ReactantCore.MissingTracedValue, Reactant.TracedRArray, Reactant.TracedRNumber}}, result_stores::Dict{Tuple, Symbol})
   @ Reactant.Compiler /mnt/software/lux/Reactant.jl/src/Compiler.jl:537
 [3] compile(f::Function, args::Tuple{ConcreteRArray{Float64, 2}, ConcreteRArray{Float64, 2}}; client::Nothing, optimize::Bool, sync::Bool)
   @ Reactant.Compiler /mnt/software/lux/Reactant.jl/src/Compiler.jl:733
 [4] top-level scope
   @ /mnt/software/lux/Reactant.jl/src/Compiler.jl:457
 [5] top-level scope
   @ none:1

julia> x_ra2 = Reactant.to_rarray(rand(2, 2))
2×2 ConcreteRArray{Float64, 2}:
 0.806331  0.636137
 0.349127  0.0385477

julia> @compile foo(x_ra, x_ra2)
Reactant.Compiler.Thunk{Symbol("##foo_reactant#16369")}()

[wsmoses]

ah so I think the flatten code doesn't do the same arg deduplication that we do elsewhere (and should also fix to keep the deduplication)

[jumerckx]

julia> @compile foo(x_ra, x_ra)
2025-01-06 15:00:43.008355: I external/xla/xla/service/llvm_ir/llvm_command_line_options.cc:50] XLA (re)initializing LLVM with options fingerprint: 18078177367847095010
Reactant.Compiler.Thunk{typeof(foo), Symbol("##foo_reactant#237"), Tuple{ConcreteRArray{Float64, 2}, ConcreteRArray{Float64, 2}}, false}(foo)

julia> @code_hlo foo(x_ra, x_ra)
module {
  func.func @main(%arg0: tensor<2x2xf64>) {
    return
  }
}

This seems fixed on master?

[mofeing]

@jumerckx do you mind adding a super simple test for this?

[jumerckx]

?

Will do