Another unnecessary GC root

[JeffBezanson]

#15402 was and still is fixed, but I found another similar case. This is hurting the performance of #16622.

julia> g = Base.Generator(x->x, rand(5))
Base.Generator{Array{Float64,1},##3#4}(#3,[0.428471,0.0120129,0.899264,0.103639,0.700651])

julia> function h(g)
        s = 0.0
        for x in g
          s += x
        end
       s
       end 
h (generic function with 1 method)

julia> @code_llvm h(g)

Inner loop:

idxend:                                           ; preds = %if
  %25 = load double*, double** %20, align 8
  %26 = getelementptr double, double* %25, i64 %22
  %27 = load double, double* %26, align 8
  %28 = add i64 %"#temp#.08", 1
  %29 = fadd double %s.09, %27
  %30 = load %jl_value_t*, %jl_value_t** %9, align 8
  store %jl_value_t* %30, %jl_value_t** %2, align 8
  %31 = getelementptr inbounds %jl_value_t, %jl_value_t* %30, i64 1
  %32 = bitcast %jl_value_t* %31 to i64*
  %33 = load i64, i64* %32, align 8
  %34 = icmp eq i64 %"#temp#.08", %33
  br i1 %34, label %L3.loopexit, label %if

There's that most-unwelcome store again.

[yuyichao]

We've attached enough metadata for LLVM to hoist the store and opt does that in -O1 so maybe we could get llvm optimize it out by adding/tweaking LLVM passes. The better solution would still be to fix #15369 instead of pushing LLVM harder....

[JeffBezanson]

I would accept either approach, as this is quite bad. Might be related to #17321 as well.

[yuyichao]

Oh, and just to be clear by opt does that in -O1 I mean it does that on our optimized IR. opt -O3 doesn't optimize that on our unoptimized IR. Another interesting observations is that the bound check is removed by LLVM automatically in this case but it can't hoist the store at the same time (in a single pass).