Incorrect bounds check elimination

[yuyichao]

IIUC, @inbounds should only affect one level of inlining unless propagate_inbounds is specified, however, due to how inlining works, this is currently not true in certain cases.

julia> @inline f(a) = a[1]
f (generic function with 1 method)

julia> @inline g(a) = f(a)
g (generic function with 1 method)

julia> function k(a)
           @inbounds x = g(a)
           return x
       end
k (generic function with 1 method)

julia> @code_warntype k(Int[])
Variables:
  #self#::#k
  a::Array{Int64,1}
  x::Int64

Body:
  begin 
      $(Expr(:inbounds, true))
      x::Int64 = (Base.arrayref)(a::Array{Int64,1},1)::Int64
      $(Expr(:inbounds, :pop)) # line 3:
      return x::Int64
  end::Int64

julia> k(Int[])
0

The issue here is that the Expr(:return) is not included when adding the inbounds metadata during inlining. A simpler case where f(a) and g(a) are not oneliner and have a simpler return statement is also currently wrong on master but is fixed in #18260 (see the added test).

Maybe we can just wait for fully linear IR?

[blakejohnson]

I'll take a look at this next week.

[vtjnash]

this is basically just a special case of #9765

[yuyichao]

Is it? The issue here is that we always need to inline

return foo(...)

as

Expr(:inbounds, false)
SSAValue(n) = foo(...)
Expr(:inbounds, :pop)
return SSAValue(n)

Unless foo is guaranteed to be not affected by boundscheck status. I don't think the effect_free check is wrong here (or is used at all).