Codegen change during 0.6 development cycle

[TotalVerb]

The output of this file

eltypestuple() = Tuple{}
eltypestuple(x, xs...) = Tuple{eltype(x), eltypes(xs...).types...}

null_safe_op(::Any, ::Type, ::Type...) = false
Base.@pure null_safe_eltype_op(f, xs...) =
    null_safe_op(f, eltypestuple(xs...).types...)

@inline function broadcast_c(f, a...)
    if null_safe_eltype_op(f, a...)
        0
    else
        1
    end
end

f() = broadcast_c(1:10, 1:10, 1:10)

code_llvm(STDOUT, f, Tuple{})

is on 0.5, quite long; on latest 0.6, also quite long.

But briefly, around when #16961 was merged, it was just

define i64 @julia_f_61750() #0 !dbg !5 {
top:
  ret i64 1
}

so something must have changed since. (The latest commit I know to have the short LLVM is 495114c, but I'll bisect to find a better bound.)

cc @JeffBezanson

Some notes:

• If @pure is removed then the generated LLVM is never short;
• @vtjnash pointed out that @pure is incorrect anyway, so we should probably still go with Refactor null_safe_op to workaround codegen changes #21290?

[Keno]

Yes, I've seen this too. Something regressed in inference's ability to constant fold type parameters.

[TotalVerb]

Although I'm not done the bisect yet, tentatively 8fee77f is the most likely cause (#20726). The symptoms match up in this particular case. While there is no actual risk of a MethodError here, there is no way for the compiler to know that, and so it must perform the call anyway. Then I think this is probably not fixable?

[TotalVerb]

Yep, the bisect confirms that

8fee77f296f5626fbcebfd87364f0ca2a3f17c87 is the first bad commit
commit 8fee77f296f5626fbcebfd87364f0ca2a3f17c87
Author: Jeff Bezanson <jeff.bezanson@gmail.com>
Date:   Tue Feb 21 15:40:09 2017 -0500

    fix #20704, `pure` annotation should not skip method errors

:040000 040000 ee47f6abb5f79d4eb4a8896eb8f5dcd22cffd8b9 ac7168b69b6c7ff5afa55c0189373d9afe36fa3c M	base
:040000 040000 ef05c2698c4935b9694e04726682b36b74f7bebb a738e9e235b557d81ebaa5d972c3ea0ab89d4e1c M	test

[JeffBezanson]

Fixed by #21310.