Int(true) == 255 with broadcasting?

[MarcMush]

only in 1.9.0-1.9.3, I get this strange bug (not 1.8, not 1.10.0-rc1)

julia> f(n)=(n-=1)<0||[f(n)]
f (generic function with 1 method)

julia> Int.(f.(0))
255

julia> a = f.(0)
true

julia> Int.(a)
1

julia> Int.(true)
1

It seems to happen only when f is non type-stable

[BioTurboNick]

I reproduced your issue. If I provide a type annotation to f it works:

julia> f(n) = (n-=1) < 0 || [f(n)]
f (generic function with 1 method)

julia> Int.(f.(0))
255

julia> f(n)::Bool = (n-=1) < 0 || [f(n)]
f (generic function with 1 method)

julia> Int.(f.(0))
1

[BioTurboNick]

Curious...

julia> bitstring(f.(0))
"00000001"

julia> bitstring.(f.(0))
"00001111"

[giordano]

I haven't run git bisect, but

• works on 6dae654 (commit before upgrade to LLVM 14)
• doesn't work on a12c2f0 (commit of upgrade to LLVM 14)
• doesn't work on b2273d3 (commit before upgrade to LLVM 15)
• I couldn't build 190f841 (commit of upgrade to LLVM 15) because I'm getting compilation errors

Everything points to a bug in LLVM 14.

[vtjnash]

Looks like our generated IR here is wrong. It is not legal to take this bitcast, since most of the bits are undef:

julia> @code_llvm optimize=false f(0)

L4:                                               ; preds = %top
  store i1 %8, i1* %2, align 1
  %12 = bitcast i1* %2 to i8*
  %13 = bitcast [1 x i8]* %0 to i8*
  switch i8 1, label %union_move_skip [
    i8 1, label %union_move
  ]

On some versions of LLVM, it takes the liberty to replace those undef bits with -1. On other versions, it takes those liberties to instead give up and not optimize the function instead.

(can't repro this on master exactly, but can confirm the generated IR is still wrong)