Fields of global constant immutables cannot be constant-folded

[TotalVerb]

Min-repro:

const x = 1
const y = (1,)
f() = x
g() = y[1]

Symptom:

julia> @code_llvm f()

define i64 @julia_f_71095() #0 {
top:
  ret i64 1
}

julia> @code_llvm g()

define i64 @julia_g_71096() #0 {
top:
  %0 = load i64, i64* inttoptr (i64 140144249280816 to i64*), align 16
  ret i64 %0
}

This makes z * im substantially slower than z * Complex(false, true).

If this cannot be easily fixed, it might be worth introducing @im for im. Note the performance difference:

julia> macro im(); :(Complex(false, true)); end
julia> f() = 2im
julia> g() = 2@im
julia> @code_llvm f()

define void @julia_f_71120(%Complex.68* noalias sret) #0 {
top:
  %1 = load i8, i8* inttoptr (i64 140144295071648 to i8*), align 32
  %2 = and i8 %1, 1
  %3 = zext i8 %2 to i64
  %4 = shl nuw nsw i64 %3, 1
  %5 = load i8, i8* inttoptr (i64 140144295071649 to i8*), align 1
  %6 = and i8 %5, 1
  %7 = zext i8 %6 to i64
  %8 = shl nuw nsw i64 %7, 1
  %.sroa.0.0..sroa_idx = getelementptr inbounds %Complex.68, %Complex.68* %0, i64 0, i32 0
  store i64 %4, i64* %.sroa.0.0..sroa_idx, align 8
  %.sroa.2.0..sroa_idx1 = getelementptr inbounds %Complex.68, %Complex.68* %0, i64 0, i32 1
  store i64 %8, i64* %.sroa.2.0..sroa_idx1, align 8
  ret void
}

julia> @code_llvm g()

define void @julia_g_71126(%Complex.68* noalias sret) #0 {
top:
  %.sroa.0.0..sroa_idx = getelementptr inbounds %Complex.68, %Complex.68* %0, i64 0, i32 0
  store i64 0, i64* %.sroa.0.0..sroa_idx, align 8
  %.sroa.2.0..sroa_idx1 = getelementptr inbounds %Complex.68, %Complex.68* %0, i64 0, i32 1
  store i64 2, i64* %.sroa.2.0..sroa_idx1, align 8
  ret void
}

[TotalVerb]

This is a regression from v0.4:

julia> f() = 2im
f (generic function with 1 method)

julia> @code_llvm f()

define void @julia_f_21209(%Complex.12* sret) {
top:
  store %Complex.12 { i64 0, i64 2 }, %Complex.12* %0, align 8
  ret void
}

[musm]

See also

julia> g(x) = 2//3 * x
g (generic function with 1 method)

julia> @code_llvm g(2f0)

; Function Attrs: uwtable
define float @julia_g_71906(float) #0 {
top:
  %1 = alloca %Rational, align 8
  call void @julia_Type_71907(%Rational* noalias nonnull sret %1, %jl_value_t* inttoptr (i64 2244157328 to %jl_value_t*), i64 2, i64 3) #1
  %2 = getelementptr inbounds %Rational, %Rational* %1, i64 0, i32 0
  %3 = load i64, i64* %2, align 8
  %4 = sitofp i64 %3 to float
  %5 = getelementptr inbounds %Rational, %Rational* %1, i64 0, i32 1
  %6 = load i64, i64* %5, align 8
  %7 = sitofp i64 %6 to float
  %8 = fdiv float %4, %7
  %9 = fmul float %8, %0
  ret float %9
}

[yuyichao]

This is a different problem....

[TotalVerb]

@yuyichao It's definitely part of this problem. Core.getfield won't be inlined as a constant.

[yuyichao]

Well #18387 (comment) is a inlining problem.

[TotalVerb]

@yuyichao I think I have a fix for that particular inlining problem that should help with (but not resolve entirely) the problem here. Will make a WIP PR shortly.

[yuyichao]

Sure, I'll be happily surprised if it helps with #18387 (comment) though since it doesn't have any constant immutable, is not a regression and is simply because the Rational constructor is too complicated to be inlined (it simplifies the ratio first).

[yuyichao]

Finally found it, the rational issue is a dup of #11522

[TotalVerb]

This was probably fixed by #21277.

[tkelman]

any way to add a regression test for it?

[TotalVerb]

A benchmark might make more sense than a regression test, since this is only a matter of performance, not correctness.

But this would work:

0.5

julia> test18387() = im.re ? 0 : 0.0
test18387 (generic function with 1 method)

julia> @inferred(test18387())
ERROR: return type Float64 does not match inferred return type Union{Float64,Int64}
 in error(::String) at ./error.jl:21

latest beta

julia> test18387() = im.re ? 0 : 0.0
test18387 (generic function with 1 method)

julia> @inferred(test18387())
0.0