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

[Matt Arsenault via llvm-commits]

================
@@ -449,16 +451,18 @@ define amdgpu_ps void @fptrunc_f64_to_bf16_abs(double %a, ptr %out) {
 ;
 ; GFX-950-LABEL: fptrunc_f64_to_bf16_abs:
 ; GFX-950:       ; %bb.0: ; %entry
-; GFX-950-NEXT:    v_cvt_f32_f64_e64 v6, |v[0:1]|
-; GFX-950-NEXT:    v_cvt_f64_f32_e32 v[4:5], v6
-; GFX-950-NEXT:    v_and_b32_e32 v7, 1, v6
-; GFX-950-NEXT:    v_cmp_gt_f64_e64 s[2:3], |v[0:1]|, |v[4:5]|
-; GFX-950-NEXT:    v_cmp_nlg_f64_e64 s[0:1], |v[0:1]|, v[4:5]
-; GFX-950-NEXT:    v_cmp_eq_u32_e32 vcc, 1, v7
-; GFX-950-NEXT:    v_cndmask_b32_e64 v0, -1, 1, s[2:3]
-; GFX-950-NEXT:    v_add_u32_e32 v0, v6, v0
+; GFX-950-NEXT:    v_cvt_f32_f64_e64 v8, |v[0:1]|
----------------
arsenm wrote:

This looks like a regression 

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