Erroneous test in julia/test /float16.jl?

[stensmo]

I am testing Julia on RISC-V, and hit a small snag, but the test could perhaps be removed?

julia/test/float16.jl

Line 82 in 7fa26f0

@test unsafe_trunc(Int16, NaN16) === Int16(0) #18771

Is on line 82. Julia will issue this code (Both on X86 and Risc-V)

define i16 @julia_unsafe_trunc_1513(half %"x::Float16") #0 {
top:
  %0 = fptosi half %"x::Float16" to i16
  %1 = freeze i16 %0
  ret i16 %1
}

When you check the LLVM Language Reference for fptosi https://llvm.org/docs/LangRef.html#fptosi-to-instruction you see that if the value does not fit, it's a poison value, and it's ok to replace the value with any value of that type.

On x86 the poison value is 0 and on Risc-V the posion value is -1 for this case. Why does Julia need the poison value to be 0?

[giordano]

This test was added by @timholy in #21312.

[giordano]

I feel like the test may be a victim of Hyrum's Law, and checking unsafe_trunc(Int16, NaN16) isa Int16 would be enough.

[oscardssmith]

yeah. the return value of this is very much not specified.