[llvm] CodeGen: Add -denormal-fp-math-bf16 flag (PR #90425)

[Phoebe Wang via llvm-commits]

phoebewang wrote:

We don't have `bf16->float` conversion instructions. The current compiler implementation simply uses 16-bit shift, so we don't explicitly do DAZ. The dot product instructions do mention to use both DAZ and FTZ in the calculation. Together with the `float->bf16` instructions, we can say all native instructions always use DAZ and FTZ.

In the contrast, the FP16 type never does DAZ/FTZ (https://cdrdv2-public.intel.com/678970/intel-avx512-fp16.pdf), though it's not controlled by MXCSR either.

https://github.com/llvm/llvm-project/pull/90425