codegen emits bounds-checking instructions when destructuring tuple even though size is statically known

[jrevels]

There might be a better MRE for this, but here's what I've got for now:

struct Wrapper{F}
    f::F
end

@inline unwrap(w::Wrapper) = w.f
@inline unwrap(x) = x
@inline unwrapcall(w, x, y) = unwrap(w)(x, y)
@inline (w::Wrapper)(x...) = unwrapcall(w, x...)

w = Wrapper(+)

julia> @code_warntype w(1, 1)
Variables:
  w::Wrapper{Base.#+}
  x::Tuple{Int64,Int64}

Body:
  begin
      return (Base.add_int)((Core.getfield)(x::Tuple{Int64,Int64}, 1)::Int64, (Core.getfield)(x::Tuple{Int64,Int64}, 2)::Int64)::Int64
  end::Int64

julia> @code_llvm w(1, 1)

; Function Wrapper
; Location: REPL[5]
define %jl_value_t addrspace(10)* @japi1_Wrapper_61798(%jl_value_t addrspace(10)*, %jl_value_t addrspace(10)**, i32) #0 {
top:
  %3 = alloca %jl_value_t addrspace(10)**, align 8
  store volatile %jl_value_t addrspace(10)** %1, %jl_value_t addrspace(10)*** %3, align 8
; Location: REPL[5]:1
  switch i32 %2, label %pass2 [
    i32 0, label %fail
    i32 1, label %fail1
  ]

fail:                                             ; preds = %top
  call void @jl_bounds_error_tuple_int(%jl_value_t addrspace(10)** %1, i64 0, i64 1)
  unreachable

fail1:                                            ; preds = %top
  call void @jl_bounds_error_tuple_int(%jl_value_t addrspace(10)** %1, i64 1, i64 2)
  unreachable

pass2:                                            ; preds = %top
  %4 = bitcast %jl_value_t addrspace(10)** %1 to i64 addrspace(10)**
  %5 = load i64 addrspace(10)*, i64 addrspace(10)** %4, align 8
  %6 = getelementptr %jl_value_t addrspace(10)*, %jl_value_t addrspace(10)** %1, i64 1
  %7 = bitcast %jl_value_t addrspace(10)** %6 to i64 addrspace(10)**
  %8 = load i64 addrspace(10)*, i64 addrspace(10)** %7, align 8
  %9 = load i64, i64 addrspace(10)* %5, align 8
  %10 = load i64, i64 addrspace(10)* %8, align 8
  %11 = add i64 %10, %9
  %12 = call %jl_value_t addrspace(10)* @jl_box_int64(i64 signext %11)
  ret %jl_value_t addrspace(10)* %12
}

Keno pointed me to this code, so I'll start poking around there. Also kind of weird/surprising that inference doesn't inline the getfield call, but maybe my surprise is just due to my ignorance 😛

[yuyichao]

Related to that we don't specsig on vararg? (The bounds check itself can probably be removed without this)

[jrevels]

Related to that we don't specsig on vararg?

Ref #5402?

[jrevels]

Actually, the example in #5402's OP is a better MRE than the one I gave here:

julia> function f(a...)
                  z = a[1]+a[2]
                  z
              end
f (generic function with 1 method)

julia> @code_llvm f(1, 1)

define i8** @japi1_f_60877(i8**, i8***, i32) #0 !dbg !5 {
top:
  %3 = alloca i8***, align 8
  store volatile i8*** %1, i8**** %3, align 8
  switch i32 %2, label %pass2 [
    i32 0, label %fail
    i32 1, label %fail1
  ]

fail:                                             ; preds = %top
  call void @jl_bounds_error_tuple_int(i8*** %1, i64 0, i64 1)
  unreachable

fail1:                                            ; preds = %top
  call void @jl_bounds_error_tuple_int(i8*** %1, i64 1, i64 2)
  unreachable

pass2:                                            ; preds = %top
  %4 = bitcast i8*** %1 to i64**
  %5 = load i64*, i64** %4, align 8
  %6 = getelementptr i8**, i8*** %1, i64 1
  %7 = bitcast i8*** %6 to i64**
  %8 = load i64*, i64** %7, align 8
  %9 = load i64, i64* %5, align 16
  %10 = load i64, i64* %8, align 16
  %11 = add i64 %10, %9
  %12 = call i8** @jl_box_int64(i64 signext %11)
  ret i8** %12
}

Should this be closed as a dup of #5402?

[yuyichao]

Should this be closed as a dup of #5402?

This can probably be fixed without fixing #5402 but since the bounds check is really the last thing you care about in these cases I think this can be closed for now.