[llvm] [AMDGPU] Form VOPD3 pairs with pair-local literal moves (PR #223431)

[Jay Foad via llvm-commits]

================
@@ -167,12 +167,12 @@ define amdgpu_ps <2 x bfloat> @fadd_v2bf16_vv(<2 x bfloat> %a, <2 x bfloat> %b)
 ; GFX1250-NEXT:    v_dual_lshlrev_b32 v2, 16, v2 :: v_dual_lshlrev_b32 v3, 16, v3
 ; GFX1250-NEXT:    v_dual_add_f32 v0, v0, v1 :: v_dual_add_f32 v1, v2, v3
 ; GFX1250-NEXT:    v_bfe_u32 v2, v0, 16, 1
-; GFX1250-NEXT:    v_or_b32_e32 v4, 0x400000, v0
 ; GFX1250-NEXT:    v_cmp_u_f32_e32 vcc_lo, 0, v0
 ; GFX1250-NEXT:    v_bfe_u32 v3, v1, 16, 1
-; GFX1250-NEXT:    v_or_b32_e32 v5, 0x400000, v1
 ; GFX1250-NEXT:    v_add3_u32 v2, v2, v0, 0x7fff
+; GFX1250-NEXT:    v_or_b32_e32 v5, 0x400000, v1
----------------
jayfoad wrote:

Any particular reason why the order of code has changed, even though the VOPD selection has not?

https://github.com/llvm/llvm-project/pull/223431