set type of statement when processing GlobalRefs

[simeonschaub]

Because we previously didn't set the type when processing GlobalRef,
the optimizer was inserting unnecessary PiNodes.

Before:

julia> x::Int = 1
1

julia> function f()
           global x = 0
           while x<10
               x += 1
           end
           x
       end
f (generic function with 1 method)

julia> @code_typed f()
CodeInfo(
1 ──       (Main.x = 0)::Any
2 ┄─ %2  = Main.x::Any
│    %3  = (isa)(%2, Int64)::Bool
└───       goto #4 if not %3
3 ── %5  = π (%2, Int64)
│    %6  = Base.slt_int(%5, 10)::Bool
└───       goto #5
4 ── %8  = (%2 < 10)::Bool
└───       goto #5
5 ┄─ %10 = φ (#3 => %6, #4 => %8)::Bool
└───       goto #10 if not %10
6 ── %12 = Main.x::Any
│    %13 = (isa)(%12, Int64)::Bool
└───       goto #8 if not %13
7 ── %15 = π (%12, Int64)
│    %16 = Base.add_int(%15, 1)::Int64
└───       goto #9
8 ── %18 = (%12 + 1)::Int64
└───       goto #9
9 ┄─ %20 = φ (#7 => %16, #8 => %18)::Int64
│          (Main.x = %20)::Any
└───       goto #2
10 ─ %23 = Main.x::Any
└───       return %23
) => Int64

This PR:

julia> @code_typed f()
CodeInfo(
1 ─      (Main.x = 0)::Any
2 ┄ %2 = Main.x::Int64
│   %3 = Base.slt_int(%2, 10)::Bool
└──      goto #4 if not %3
3 ─ %5 = Main.x::Int64
│   %6 = Base.add_int(%5, 1)::Int64
│        (Main.x = %6)::Any
└──      goto #2
4 ─ %9 = Main.x::Int64
└──      return %9
) => Int64

[aviatesk]

Inference understands + is resolved to ::Int + ::Int, so the inliner shouldn't insert that PiNode. On master, I got:

julia> x::Int = 1
1

julia> function f()
           global x = 0
           while x<10
               x += 1
           end
           x
       end
f (generic function with 1 method)

julia> @code_typed f()
CodeInfo(
1 ─      (Main.x = 0)::Any
2 ┄ %2 = Main.x::Any
│   %3 = Base.slt_int(%2, 10)::Bool
└──      goto #4 if not %3
3 ─ %5 = Main.x::Any
│   %6 = Base.add_int(%5, 1)::Int64
│        (Main.x = %6)::Any
└──      goto #2
4 ─ %9 = Main.x::Any
└──      return %9
) => Int64

so that test case should pass already.

Though we should set types for those GlobalRef statements.

[simeonschaub]

Huh, so this must have improved very recently? But you still think this change is a good idea?

[simeonschaub]

Yeah, I compared this to bfc9431 before, so apparently it must have been fixed in the last 3 days.

[aviatesk]

Interestingly, #44095 has improved the situation -- previously the inliner computes call signature by itself (so can be affected by this change) but now it just trusts the result of inference.

[aviatesk]

I couldn't come up with a test case that demonstrates what is improved by this change, but we may want to use this type information in some new optimization pass in the future.

[aviatesk]

Is the failure on macOS buildbot a known issue? I rebuilt it multiple times, but it seems to persist.

[simeonschaub]

Let's try one more time, I really can't imagine how that would be related.

[vtjnash]

Suggested change

	global x = 0
	while x < 10
	x += 1
	end
	x
	global x44200 = 0
	while x44200 < 10
	x44200 += 1
	end
	x44200

[vtjnash]

FWIW, I don't see these PiNodes on my builds, so the test does not seem to do anything for me. Could be I am looking at the wrong thing though.

[simeonschaub]

Yes, as discussed above, #44095 actually already fixed this particular test case, but I think this is worthwhile regardless.