Use right caller for isapplicable usage

[vchuravy]

Originally I was thinking something like:

        # 1. Check if function is inactive
        @show fargs
        @show argtypes
        inactive_arginfo = ArgInfo(nothing, pushfirst!(copy(argtypes), Core.Const(EnzymeRules.inactive)))
        inactive_atype = Tuple{typeof(EnzymeRules.inactive), atype.parameters...} 
        @show inactive_arginfo
        @show inactive_atype
        inactive_meta  = @invoke Core.Compiler.abstract_call_gf_by_type(
            interp::AbstractInterpreter,
            EnzymeRules.inactive::Any,
            inactive_arginfo::ArgInfo,
            si::StmtInfo,
            inactive_atype::Any,
            sv::AbsIntState,
            max_methods::Int,
        )

        if Core.Compiler.nmatches(inactive_meta.info) == 0

E.g. doing an abstract call against the rules function and checking how many matches we would get,
this ought to do the edge handling for us. But this change is more minimal.

[github-actions]

Benchmark Results

	main	736ddc8...	main/736ddc86c1c494...
basics/overhead	4.64 ± 0.011 ns	4.33 ± 0.01 ns	1.07
time_to_load	1.25 ± 0.026 s	1.18 ± 0.016 s	1.06

Benchmark Plots

A plot of the benchmark results have been uploaded as an artifact to the workflow run for this PR.
Go to "Actions"->"Benchmark a pull request"->[the most recent run]->"Artifacts" (at the bottom).

[vchuravy]

Doesn't help as much as I hoped. So maybe I do have to try the other variant.

[wsmoses]

x/ref #2152

[wsmoses]

technically we'll need to change this to not be exclusive (e.g. you can have both forward_rules and reverse_rules set)

[vchuravy]

The cache is split on forward vs reverse. So we only need to look at it for each.

[wsmoses]

profile.pb.gz

profile data

[wsmoses]

okay changing the inner isactive part to the native interpreter brings "indirect function call return type analysis" back to 7s from 54s

[vchuravy]

What other things have you seen here?

[wsmoses]

NoCallInfo

[vchuravy]

okay changing the inner isactive part to the native interpreter brings "indirect function call return type analysis" back to 7s from 54s

I am not sure I follow?

[wsmoses]

Before this PR the "indirect function call return type analysis" test ran in ~7 seconds. With this PR it runs in ~54 seconds. Changing

      inactive_meta  = @invoke Core.Compiler.abstract_call_gf_by_type(
            interp::AbstractInterpreter,
            EnzymeRules.inactive::Any,
            inactive_arginfo::ArgInfo,
            si::StmtInfo,
            inactive_atype::Any,
            sv::AbsIntState,

to

      ni = NativeInterpreter(interp.world)
      inactive_meta  = @invoke Core.Compiler.abstract_call_gf_by_type(
            ni::AbstractInterpreter,
            EnzymeRules.inactive::Any,
            inactive_arginfo::ArgInfo,
            si::StmtInfo,
            inactive_atype::Any,
            sv::AbsIntState,
            max_methods::Int,
        )

reduces it back to ~7s

[wsmoses]

However regardless independent of the native interpreter stuff (not pushed) this PR breaks invalidations, per CI:

has_frule     |   10     10  0.0s
autodiff(Forward, ...) custom rules: Test Failed at /home/runner/work/Enzyme.jl/Enzyme.jl/test/rules.jl:62
  Expression: (autodiff(Forward, (x->(#= /home/runner/work/Enzyme.jl/Enzyme.jl/test/rules.jl:62 =#
                f(x) ^ 2)), Duplicated(2.0, 1.0)))[1] ≈ 832.0
   Evaluated: 32.0 ≈ 832.0

Stacktrace:
 [1] macro expansion
   @ /opt/hostedtoolcache/julia/1.10.7/x64/share/julia/stdlib/v1.10/Test/src/Test.jl:672 [inlined]
 [2] macro expansion
   @ ~/work/Enzyme.jl/Enzyme.jl/test/rules.jl:62 [inlined]
 [3] macro expansion
   @ /opt/hostedtoolcache/julia/1.10.7/x64/share/julia/stdlib/v1.10/Test/src/Test.jl:1577 [inlined]
 [4] top-level scope
   @ ~/work/Enzyme.jl/Enzyme.jl/test/rules.jl:61
autodiff(Forward, ...) custom rules: Test Failed at /home/runner/work/Enzyme.jl/Enzyme.jl/test/rules.jl:70
  Expression: res[1] ≈ 80032.0
   Evaluated: 32.0 ≈ 80032.0

Stacktrace:
 [1] macro expansion
   @ /opt/hostedtoolcache/julia/1.10.7/x64/share/julia/stdlib/v1.10/Test/src/Test.jl:672 [inlined]
 [2] macro expansion
   @ ~/work/Enzyme.jl/Enzyme.jl/test/rules.jl:70 [inlined]
 [3] macro expansion
   @ /opt/hostedtoolcache/julia/1.10.7/x64/share/julia/stdlib/v1.10/Test/src/Test.jl:1577 [inlined]
 [4] top-level scope
   @ ~/work/Enzyme.jl/Enzyme.jl/test/rules.jl:61
autodiff(Forward, ...) custom rules: Test Failed at /home/runner/work/Enzyme.jl/Enzyme.jl/test/rules.jl:71
  Expression: res[2] ≈ 80096.0
   Evaluated: 96.0 ≈ 80096.0

Stacktrace:
 [1] macro expansion
   @ /opt/hostedtoolcache/julia/1.10.7/x64/share/julia/stdlib/v1.10/Test/src/Test.jl:672 [inlined]
 [2] macro expansion
   @ ~/work/Enzyme.jl/Enzyme.jl/test/rules.jl:71 [inlined]
 [3] macro expansion
   @ /opt/hostedtoolcache/julia/1.10.7/x64/share/julia/stdlib/v1.10/Test/src/Test.jl:1577 [inlined]
 [4] top-level scope
   @ ~/work/Enzyme.jl/Enzyme.jl/test/rules.jl:61
Test Summary:                       | Pass  Fail  Total  Time
autodiff(Forward, ...) custom rules |    3     3      6  2.5s
ERROR: LoadError: Some tests did not pass: 3 passed, 3 failed, 0 

I also think inference time is not fixed yet =/