@benchmark reports x5 slower with Float32 Tuple with julia 1.12 #59906
Description
@benchmark on the program below reports x5 slower (800ns vs 160ns) with Julia 1.12.1 wrt 1.11.6. Same timing if we setup tmp argument through benchmark setup argument. Tested on Ubuntu 25.04.
However, the manual @time reports the same timing in both versions.
The difference only occurs with Float32 (not with Float64, Int32 for Int64)
using BenchmarkTools
function f(tmp::Tuple{Float32, Float32})::Float32
i = 1
while i <= 1000
if i % 2 == 0
tmp = (tmp[1], tmp[2] + one(Float32))
end
i += 1
end
return tmp[1] + tmp[2]
end
function main()
tmp = (zero(Float32), zero(Float32))
res = 0
@time for i in 1:10_000
tmp = (zero(Float32), zero(Float32))
res += f(tmp)
end
println(res)
return display(@benchmark($f($tmp)))
end
main()Julia 1.12.1
Julia 1.11.6
