[llvm] AMDGPU: Fix mis-selecting saddr flat atomics on gfx9 (PR #156860)

[Matt Arsenault via llvm-commits]

================
@@ -57,13 +53,12 @@ entry:
 define amdgpu_kernel void @flat_atomic_cond_sub_rtn_u32(ptr %addr, i32 %in, ptr %use) {
 ; GFX12-SDAG-LABEL: flat_atomic_cond_sub_rtn_u32:
 ; GFX12-SDAG:       ; %bb.0: ; %entry
-; GFX12-SDAG-NEXT:    s_clause 0x1
 ; GFX12-SDAG-NEXT:    s_load_b96 s[0:2], s[4:5], 0x24
+; GFX12-SDAG-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX12-SDAG-NEXT:    s_load_b64 s[4:5], s[4:5], 0x34
 ; GFX12-SDAG-NEXT:    s_wait_kmcnt 0x0
-; GFX12-SDAG-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
-; GFX12-SDAG-NEXT:    v_mov_b32_e32 v2, s2
-; GFX12-SDAG-NEXT:    flat_atomic_cond_sub_u32 v2, v[0:1], v2 offset:16 th:TH_ATOMIC_RETURN
+; GFX12-SDAG-NEXT:    v_mov_b32_e32 v1, s2
+; GFX12-SDAG-NEXT:    flat_atomic_cond_sub_u32 v2, v0, v1, s[0:1] offset:16 th:TH_ATOMIC_RETURN
----------------
arsenm wrote:

Actually this shouldn't have happened, this is not a 1250 run line

https://github.com/llvm/llvm-project/pull/156860