Investigate if we need a custom hash for EnzymeCompilerParams

[vchuravy]

GPUCompiler hashes CompilerConfig which contains the CompilerParams and uses that as a key for the global dictionary that splits the global inference cache in https://github.com/JuliaGPU/GPUCompiler.jl/blob/21ca075c1e91fe0c15f1330ab487b4831013ec1f/src/jlgen.jl#L175

Enzyme.jl/src/compiler.jl

Line 2529 in 9d6963c

struct EnzymeCompilerParams <: AbstractEnzymeCompilerParams

julia> using Enzyme, GPUCompiler

julia> f1(x) = x*x
       # Returns a tuple of active returns, which in this case is simply (2.0,)
f1 (generic function with 1 method)

julia> @test first(autodiff(Reverse, f1, Active(1.0))[1]) ≈ 2.0
ERROR: LoadError: UndefVarError: `@test` not defined
in expression starting at REPL[6]:1

julia> autodiff(Reverse, f1, Active(1.0))
((2.0,),)

julia> autodiff(Reverse, f1, Const(1.0))
((nothing,),)

julia> GPUCompiler.GLOBAL_CI_CACHES
Dict{CompilerConfig, GPUCompiler.CodeCache} with 5 entries:
  CompilerConfig for EnzymeTarget         => CodeCache(IdDict{MethodInstance, Vector{CodeInstance}}())
  CompilerConfig for EnzymeTarget         => CodeCache(IdDict{MethodInstance, Vector{CodeInstance}}())
  CompilerConfig for EnzymeTarget         => CodeCache(IdDict{MethodInstance, Vector{CodeInstance}}(MethodInstance for *(::Float64, ::Float64)=>[CodeInstance(MethodInstance for *(::Float64, ::Float64), #undef, 0x0000000000000001, 0xffffffffffffffff, Float64, #undef, "\x01\0\n\0\x03\0\0\0\0\b\b\x16\x88\xc2\xe6.\x1e\x…
  CompilerConfig for NativeCompilerTarget => CodeCache(IdDict{MethodInstance, Vector{CodeInstance}}(MethodInstance for *(::Float64, ::Float64)=>[CodeInstance(MethodInstance for *(::Float64, ::Float64), #undef, 0x0000000000000001, 0xffffffffffffffff, Float64, #undef, "\x01\0\n\0\x03\0\0\0\0\b\b\x16\x88\xc2\xe6.\x1e\x…
  CompilerConfig for EnzymeTarget         => CodeCache(IdDict{MethodInstance, Vector{CodeInstance}}(MethodInstance for *(::Float64, ::Float64)=>[CodeInstance(MethodInstance for *(::Float64, ::Float64), #undef, 0x0000000000000001, 0xffffffffffffffff, Float64, #undef, "\x01\0\n\0\x03\0\0\0\0\b\b\x16\x88\xc2\xe6.\x1e\x…

julia> 

So inference unnecessarily repeats work. We need to carefully define what we need to split the cache on. cc: @maleadt

Especially for JuliaGPU/GPUCompiler.jl#534 / JuliaLang/julia#52233

I think we should probably define a AbstractCacheKey and construct that from the compiler params, requiring downstream users to make those conscious choices.