[llvm] [LLVM][SVE] Implement isel for bfloat fptoi and itofp operations. (PR #129713)

[Paul Walker via llvm-commits]

================
@@ -5465,6 +5465,14 @@ multiclass sve_int_dup_fpimm_pred<string asm> {
             (!cast<Instruction>(NAME # _S) $zd, $pg, fpimm32:$imm8)>;
   def : Pat<(nxv2f64 (vselect nxv2i1:$pg, (splat_vector fpimm64:$imm8), nxv2f64:$zd)),
             (!cast<Instruction>(NAME # _D) $zd, $pg, fpimm64:$imm8)>;
+
+  // Some half precision immediates alias with bfloat (e.g. f16(1.875) == bf16(1.0)).
----------------
paulwalker-arm wrote:

Not sure I understand.  The comment is saying "some" half precision immediates alias with bfloat and the way this is achieved is by using the `fpimmbf16` complex pattern that will only let the safe ones through.

`All fpimmbf16 immediates alias with a FP16 immediate` is obvious from its use because that's how isel works.  I can remove the comment if you feel it offers no value? I only added it just in case somebody wondered why we have bfloat patterns for an instruction that doesn't really support bfloat.

https://github.com/llvm/llvm-project/pull/129713