inference: add benchmark target to check benefit of method match cache

This benchmark target corresponds to JuliaLang/julia/pull/46535.

`method_match_cache` contains many artificially generated broadcasting
operations, that will lead to constant propagations and accompanying
method match analysis on call sites with same abstract signatures, e.g.
- `Tuple{typeof(convert), Type{Int64}, Int64}`
- `Tuple{typeof(convert), Type{Float64}, Float64}`

Since we currently don't cache method match result for abstract call
signatures on the level of the runtime system, we can obtain the best
performance if we cache such abstract method match results on Julia level,
that will be revived by JuliaLang/julia/pull/46535.

src/inference/InferenceBenchmarks.jl

@@ -179,6 +179,20 @@ let # see https://github.com/JuliaLang/julia/pull/45276
         $ex
     end
 end
+let # see https://github.com/JuliaLang/julia/pull/46535
+    n = 100
+    ex = Expr(:block)
+    var = gensym()
+    push!(ex.args, :(y = sum(x)))
+    for i = 1:n
+        push!(ex.args, :(x .= $(Float64(i))))
+        push!(ex.args, :(y += sum(x)))
+    end
+    push!(ex.args, :(return y))
+    @eval global function method_match_cache(x)
+        $ex
+    end
+end
 
 const SUITE = BenchmarkGroup()
 
@@ -195,6 +209,7 @@ let g = addgroup!(SUITE, "abstract interpretation")
     g["construct_ssa!"] = @benchmarkable abs_call(CC.construct_ssa!, (Core.CodeInfo,CC.IRCode,CC.DomTree,Vector{CC.SlotInfo},Vector{Any}))
     g["domsort_ssa!"] = @benchmarkable abs_call(CC.domsort_ssa!, (CC.IRCode,CC.DomTree))
     g["quadratic"] = @benchmarkable abs_call(quadratic, (Int,))
+    g["method_match_cache"] = @benchmarkable abs_call(method_match_cache, (Float64,))
     tune_benchmarks!(g)
 end
 
@@ -210,6 +225,7 @@ let g = addgroup!(SUITE, "optimization")
     g["construct_ssa!"] = @benchmarkable f() (setup = (f = opt_call(CC.construct_ssa!, (Core.CodeInfo,CC.IRCode,CC.DomTree,Vector{CC.SlotInfo},Vector{Any}))))
     g["domsort_ssa!"] = @benchmarkable f() (setup = (f = opt_call(CC.domsort_ssa!, (CC.IRCode,CC.DomTree))))
     g["quadratic"] = @benchmarkable f() (setup = (f = opt_call(quadratic, (Int,))))
+    g["method_match_cache"] = @benchmarkable f() (setup = (f = opt_call(method_match_cache, (Float64,))))
     tune_benchmarks!(g)
 end
 
@@ -226,6 +242,7 @@ let g = addgroup!(SUITE, "allinference")
     g["construct_ssa!"] = @benchmarkable inf_call(CC.construct_ssa!, (Core.CodeInfo,CC.IRCode,CC.DomTree,Vector{CC.SlotInfo},Vector{Any}))
     g["domsort_ssa!"] = @benchmarkable inf_call(CC.domsort_ssa!, (CC.IRCode,CC.DomTree))
     g["quadratic"] = @benchmarkable inf_call(quadratic, (Int,))
+    g["method_match_cache"] = @benchmarkable inf_call(method_match_cache, (Float64,))
     tune_benchmarks!(g)
 end