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

[Matt Arsenault via llvm-commits]

================
@@ -1296,19 +1300,19 @@ let SubtargetPredicate = isGFX10Plus in {
     FLAT_Global_Atomic_Pseudo<"global_atomic_fcmpswap_x2", AVLdSt_64, f64, v2f64, AVLdSt_128>;
 } // End SubtargetPredicate = isGFX10Plus
 
-let OtherPredicates = [HasAtomicFaddNoRtnInsts] in
+let SubtargetPredicate = HasAtomicFaddNoRtnInsts in
   defm GLOBAL_ATOMIC_ADD_F32 : FLAT_Global_Atomic_Pseudo_NO_RTN <
     "global_atomic_add_f32", AVLdSt_32, f32
   >;
-let OtherPredicates = [HasAtomicBufferGlobalPkAddF16NoRtnInsts] in
+let SubtargetPredicate = HasAtomicBufferGlobalPkAddF16NoRtnInsts in
   defm GLOBAL_ATOMIC_PK_ADD_F16 : FLAT_Global_Atomic_Pseudo_NO_RTN <
     "global_atomic_pk_add_f16", AVLdSt_32, v2f16
   >;
-let OtherPredicates = [HasAtomicFaddRtnInsts] in
+let SubtargetPredicate = HasAtomicFaddRtnInsts in
   defm GLOBAL_ATOMIC_ADD_F32 : FLAT_Global_Atomic_Pseudo_RTN <
     "global_atomic_add_f32", AVLdSt_32, f32
   >;
-let OtherPredicates = [HasAtomicBufferGlobalPkAddF16Insts] in
+let SubtargetPredicate = HasAtomicBufferGlobalPkAddF16Insts in
----------------
arsenm wrote:

It's not that well formed. Really we should be more structured with what features are defined, and have a better system for combining them. The usual convention is SubtargetPredicate is the main instruction-exists option, and OtherPredicates is the side option for extra modifiers 

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