[1.10] Function returns wrong value instead of throwing an error

[vchuravy]

julia> function f1(cond)
           x = Ref{Int}(0)
           if cond
               x[] = x
           else
               return -1
           end
       end
f1 (generic function with 1 method)

julia> @code_typed f1(true)
CodeInfo(
1 ─ %1 = %new(Base.RefValue{Int64}, 0)::Base.RefValue{Int64}
└──      goto #3 if not cond
2 ─      invoke Base.setindex!(%1::Base.RefValue{Int64}, %1::Base.RefValue{Int64})::Union{}
└──      unreachable
3 ─      return -1
) => Int64

julia> f1(true)
-1

julia> f1(false)
-1

julia> @code_llvm optimize=false f1(true)
;  @ REPL[1]:1 within `f1`
define i64 @julia_f1_305(i8 zeroext %0) #0 {
top:
  %1 = call {}*** @julia.get_pgcstack()
  %2 = bitcast {}*** %1 to {}**
  %current_task = getelementptr inbounds {}*, {}** %2, i64 -14
  %3 = bitcast {}** %current_task to i64*
  %world_age = getelementptr inbounds i64, i64* %3, i64 15
  ret i64 -1
}

julia> Base.code_ircode(f1, (Bool,))
1-element Vector{Any}:
2 1 ─ %1 = %new(Base.RefValue{Int64}, 0)::Base.RefValue{Int64}        │╻╷ Ref
3 └──      goto #3 if not _2                                          │  
4 2 ─      invoke Base.setindex!(%1::Base.RefValue{Int64}, %1::Base.RefValue{Int64})::Union{}
  └──      unreachable                                                │  
6 3 ─      return -1                                                  │  
   => Int64                                                   

This seems fixed on nightly and on 1.9, but not on 1.10

@aviatesk maybe you can take a look? I am very confused by the fact that code_ircode seems to give the right answer, but by the time we end up at LLVM someone replaced it all with a constant.

[aviatesk]

Yeah, I'll look into this today.

[oscardssmith]

is this #52856?

[vchuravy]

It looks different enough... and there is no embedded value, the value is created through new

[aviatesk]

It appears to be a bug in the LLVM optimization level. The the Julia level IR seems fine. The code gets incorrectly optimized on LLVM level:

julia> code_typed(f1, (Bool,))
1-element Vector{Any}:
 CodeInfo(
1 ─ %1 = %new(Base.RefValue{Int64}, 0)::Base.RefValue{Int64}
└──      goto #3 if not cond
2 ─      invoke Base.setindex!(%1::Base.RefValue{Int64}, %1::Base.RefValue{Int64})::Union{}
└──      unreachable
3 ─      return -1
) => Int64

julia> code_llvm(f1, (Bool,))
;  @ REPL[2]:1 within `f1`
define i64 @julia_f1_758(i8 zeroext %0) #0 {
top:
  ret i64 -1
}

[vchuravy]

@aviatesk it's before LLVM

julia> @code_llvm optimize=false f1(true)
;  @ REPL[1]:1 within `f1`
define i64 @julia_f1_305(i8 zeroext %0) #0 {
top:
  %1 = call {}*** @julia.get_pgcstack()
  %2 = bitcast {}*** %1 to {}**
  %current_task = getelementptr inbounds {}*, {}** %2, i64 -14
  %3 = bitcast {}** %current_task to i64*
  %world_age = getelementptr inbounds i64, i64* %3, i64 15
  ret i64 -1
}

[aviatesk]

Ah, I see. This seems to be another effects analysis bug:

julia> Base.infer_effects((Bool,)) do cond
           x = Ref{Int}(0)
           if cond
               x[] = x
           else
               return -1
           end
       end

[aviatesk]

Will be fixed by #53076.