[llvm] [SelectionDAG] Handle `fneg`/`fabs`/`fcopysign` in `SimplifyDemandedBits` (PR #139239)

[Matt Arsenault via llvm-commits]

================
@@ -18639,17 +18639,17 @@ define bfloat @v_fabs_bf16(bfloat %a) {
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_mul_f32_e32 v0, 1.0, v0
-; GCN-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
-; GCN-NEXT:    v_and_b32_e32 v0, 0x7fffffff, v0
+; GCN-NEXT:    v_and_b32_e32 v0, 0x7fff0000, v0
+; GCN-NEXT:    v_mul_f32_e64 v0, 1.0, |v0|
----------------
arsenm wrote:

Kind of a regression? It should be neutral in code size and cycles, but probably should prefer a bit-op to an FP op (e.g. this avoids the mode dependency) 

https://github.com/llvm/llvm-project/pull/139239