[AA] Make alias check more accurately.

It should be considered to be aliased only if there is at lease one AA
return non-NoAlias.
The AliasResult affect the ChainDependency adding in ScheduleDAGInstrs,
if a aliased memory location accessed in different type, TBAA may return
NoAlias of them. This would result in incorrect instr order in final cg.

Author: fengfeng09

llvm/lib/Analysis/AliasAnalysis.cpp

@@ -122,7 +122,7 @@ AliasResult AAResults::alias(const MemoryLocation &LocA,
   AAQI.Depth++;
   for (const auto &AA : AAs) {
     Result = AA->alias(LocA, LocB, AAQI, CtxI);
-    if (Result != AliasResult::MayAlias)
+    if (Result != AliasResult::NoAlias)
       break;
   }
   AAQI.Depth--;

llvm/test/CodeGen/AMDGPU/tbaa-load-store.ll

@@ -30,10 +30,10 @@ define amdgpu_kernel void @Kernel_func(i8 %a, i32 %b, i32 %c, i32 %end_bit) {
 ; GCN-NEXT:    v_min_i32_e32 v4, 1, v4
 ; GCN-NEXT:    v_bfe_u32 v3, v3, 0, v4
 ; GCN-NEXT:    v_lshlrev_b32_e32 v3, 1, v3
+; GCN-NEXT:    ds_write_b32 v1, v0
 ; GCN-NEXT:    v_sub_u32_e32 v4, v2, v3
 ; GCN-NEXT:    ds_read_u16 v3, v4
-; GCN-NEXT:    ds_write_b32 v1, v0
-; GCN-NEXT:    s_waitcnt lgkmcnt(1)
+; GCN-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN-NEXT:    v_add_u16_e32 v3, 1, v3
 ; GCN-NEXT:    v_cmp_ge_i32_e32 vcc, s1, v3
 ; GCN-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]