Edges

Edges here are a graph-theoretic concept relating the connections between individual statements in the source code. For example, consider

julia> ex = quote
           s = 0
           k = 5
           for i = 1:3
               global s, k
               s += rand(1:5)
               k += i
           end
       end
quote
    #= REPL[2]:2 =#
    s = 0
    #= REPL[2]:3 =#
    k = 5
    #= REPL[2]:4 =#
    for i = 1:3
        #= REPL[2]:5 =#
        global s, k
        #= REPL[2]:6 =#
        s += rand(1:5)
        #= REPL[2]:7 =#
        k += i
    end
end

julia> eval(ex)

julia> s
10    # random

julia> k
11    # reproducible

We lower it,

julia> lwr = Meta.lower(Main, ex)
:($(Expr(:thunk, CodeInfo(
    @ REPL[2]:2 within `top-level scope'
1 ─       s = 0
│   @ REPL[2]:3 within `top-level scope'
│         k = 5
│   @ REPL[2]:4 within `top-level scope'
│   %3  = 1:3
│         #s1 = Base.iterate(%3)
│   %5  = #s1 === nothing
│   %6  = Base.not_int(%5)
└──       goto #4 if not %6
2 ┄ %8  = #s1
│         i = Core.getfield(%8, 1)
│   %10 = Core.getfield(%8, 2)
│   @ REPL[2]:5 within `top-level scope'
│         global k
│         global s
│   @ REPL[2]:6 within `top-level scope'
│   %13 = 1:5
│   %14 = rand(%13)
│   %15 = s + %14
│         s = %15
│   @ REPL[2]:7 within `top-level scope'
│   %17 = k + i
│         k = %17
│         #s1 = Base.iterate(%3, %10)
│   %20 = #s1 === nothing
│   %21 = Base.not_int(%20)
└──       goto #4 if not %21
3 ─       goto #2
4 ┄       return
))))

and then extract the edges:

julia> edges = CodeEdges(lwr.args[1])
CodeEdges:
  s: assigned on [1, 16], depends on [15], and used by [12, 15]
  k: assigned on [2, 18], depends on [17], and used by [11, 17]
  statement  1 depends on [15, 16] and is used by [12, 15, 16]
  statement  2 depends on [17, 18] and is used by [11, 17, 18]
  statement  3 depends on ∅ and is used by [4, 19]
  statement  4 depends on [3, 10, 19] and is used by [5, 8, 19, 20]
  statement  5 depends on [4, 19] and is used by [6]
  statement  6 depends on [5] and is used by [7]
  statement  7 depends on [6] and is used by ∅
  statement  8 depends on [4, 19] and is used by [9, 10]
  statement  9 depends on [8] and is used by [17]
  statement 10 depends on [8] and is used by [4, 19]
  statement 11 depends on [2, 18] and is used by ∅
  statement 12 depends on [1, 16] and is used by ∅
  statement 13 depends on ∅ and is used by [14]
  statement 14 depends on [13] and is used by [15]
  statement 15 depends on [1, 14, 16] and is used by [1, 16]
  statement 16 depends on [1, 15] and is used by [1, 12, 15]
  statement 17 depends on [2, 9, 18] and is used by [2, 18]
  statement 18 depends on [2, 17] and is used by [2, 11, 17]
  statement 19 depends on [3, 4, 10] and is used by [4, 5, 8, 20]
  statement 20 depends on [4, 19] and is used by [21]
  statement 21 depends on [20] and is used by [22]
  statement 22 depends on [21] and is used by ∅
  statement 23 depends on ∅ and is used by ∅
  statement 24 depends on ∅ and is used by ∅

This shows the dependencies of each line as well as the "named variables" s and k. It's worth looking specifically to see how the slot-variable #s1 gets handled, as you'll notice there is no mention of this in the "variables" section at the top. You can see that #s1 first gets assigned on line 4 (the iterate statement), which you'll notice depends on 3 (via the SSAValue printed as %3). But that line 4 also is shown as depending on 10 and 19. You can see that line 19 is the 2-argument call to iterate, and that this line depends on SSAValue %10 (the state variable). Consequently all the line-dependencies of this slot variable have been aggregated into a single list by determining the "global" influences on that slot variable.

An even more useful output can be obtained from the following:

julia> LoweredCodeUtils.print_with_code(stdout, lwr.args[1], edges)
Names:
s: assigned on [1, 16], depends on [15], and used by [12, 15]
k: assigned on [2, 18], depends on [17], and used by [11, 17]
Code:
1 ─       s = 0
│             # preds: [15, 16], succs: [12, 15, 16]
│         k = 5
│             # preds: [17, 18], succs: [11, 17, 18]
│   %3  = 1:3
│             # preds: ∅, succs: [4, 19]
│         _1 = Base.iterate(%3)
│             # preds: [3, 10, 19], succs: [5, 8, 19, 20]
│   %5  = _1 === nothing
│             # preds: [4, 19], succs: [6]
│   %6  = Base.not_int(%5)
│             # preds: [5], succs: [7]
└──       goto #4 if not %6
              # preds: [6], succs: ∅
2 ┄ %8  = _1
│             # preds: [4, 19], succs: [9, 10]
│         _2 = Core.getfield(%8, 1)
│             # preds: [8], succs: [17]
│   %10 = Core.getfield(%8, 2)
│             # preds: [8], succs: [4, 19]
│         global k
│             # preds: [2, 18], succs: ∅
│         global s
│             # preds: [1, 16], succs: ∅
│   %13 = 1:5
│             # preds: ∅, succs: [14]
│   %14 = rand(%13)
│             # preds: [13], succs: [15]
│   %15 = s + %14
│             # preds: [1, 14, 16], succs: [1, 16]
│         s = %15
│             # preds: [1, 15], succs: [1, 12, 15]
│   %17 = k + _2
│             # preds: [2, 9, 18], succs: [2, 18]
│         k = %17
│             # preds: [2, 17], succs: [2, 11, 17]
│         _1 = Base.iterate(%3, %10)
│             # preds: [3, 4, 10], succs: [4, 5, 8, 20]
│   %20 = _1 === nothing
│             # preds: [4, 19], succs: [21]
│   %21 = Base.not_int(%20)
│             # preds: [20], succs: [22]
└──       goto #4 if not %21
              # preds: [21], succs: ∅
3 ─       goto #2
              # preds: ∅, succs: ∅
4 ┄       return
              # preds: ∅, succs: ∅

Here the edges are printed right after each line.

Suppose we want to evaluate just the lines needed to compute s. We can find out which lines these are with

julia> isrequired = lines_required(:s, lwr.args[1], edges)
24-element BitArray{1}:
 1
 0
 1
 1
 1
 1
 1
 1
 0
 1
 0
 0
 1
 1
 1
 1
 0
 0
 1
 1
 1
 1
 1
 0

and display them with

julia> LoweredCodeUtils.print_with_code(stdout, lwr.args[1], isrequired)
 1 t 1 ─       s = 0
 2 f │         k = 5
 3 t │   %3  = 1:3
 4 t │         _1 = Base.iterate(%3)
 5 t │   %5  = _1 === nothing
 6 t │   %6  = Base.not_int(%5)
 7 t └──       goto #4 if not %6
 8 t 2 ┄ %8  = _1
 9 f │         _2 = Core.getfield(%8, 1)
10 t │   %10 = Core.getfield(%8, 2)
11 f │         global k
12 f │         global s
13 t │   %13 = 1:5
14 t │   %14 = rand(%13)
15 t │   %15 = s + %14
16 t │         s = %15
17 f │   %17 = k + _2
18 f │         k = %17
19 t │         _1 = Base.iterate(%3, %10)
20 t │   %20 = _1 === nothing
21 t │   %21 = Base.not_int(%20)
22 t └──       goto #4 if not %21
23 t 3 ─       goto #2
24 f 4 ┄       return

We can test this with the following:

julia> using JuliaInterpreter

julia> frame = Frame(Main, lwr.args[1])
Frame for Main
   1 2  1 ─       s = 0
   2 3  │         k = 5
   3 4  │   %3  = 1:3
⋮

julia> k
11

julia> k = 0
0

julia> selective_eval_fromstart!(frame, isrequired, true)

julia> k
0

julia> s
12     # random

julia> selective_eval_fromstart!(frame, isrequired, true)

julia> k
0

julia> s
9      # random

You can see that k was not reset to its value of 11 when we ran this code selectively, but that s was updated (to a random value) each time.