Given the following programs:

func.func @test1(%arg0: f32) -> bf16 {
    %0 = arith.truncf %arg0 fastmath<fast> : f32 to bf16
    return %0 : bf16
}

func.func @test2(%0: bf16) -> f32 {
    %1 = arith.extf %0 fastmath<fast> : bf16 to f32
    return %1 : f32
}

the convert-to-llvm pass produces the following IR:

module {
  llvm.func @test1(%arg0: f32) -> f32 {
    %0 = llvm.fptrunc %arg0 {fastmath = #arith.fastmath<fast>} : f32 to bf16
    llvm.return %arg0 : f32
  }
  llvm.func @test2(%arg0: bf16) -> f32 {
    %0 = llvm.fpext %arg0 {fastmath = #arith.fastmath<fast>} : bf16 to f32
    llvm.return %0 : f32
  }
}

https://godbolt.org/z/ahs6jaEaP

As can be seen in the output, the fastmath property is taken as is from arith and not translated to LLVMs fast math flags.
Part of the reason seems to be that fpext and fptrunc in the LLVM dialect do not support fast math flags at the moment