[llvm-dev] What is TBAA's advantage over BasicAA?

[Bekket McClane via llvm-dev]

Hi all,

I’m currently studying TBAA. I understand that it provide a “struct-path” analysis, which should do well on analyzing struct-like type in high level language.
However, I got one simple question - Why don’t we just trace back the GEP chains from a given pointer, evaluate the pointer as some sort of BasePtr + Offset form, then compare if two pointers have same forms - just exactly what BasicAA do in its aliasGEP function ?
What is TBAA’s advantage over BasicAA, and its necessity of providing a decent type system, which actually looks like another form of Base + Offset?

Currently I only come up with one example: addresses of A[i].firstField and A[j].secondField, where A is SomeStruct* type. 
As the two pointers is accessing different fields of a struct, they would never alias regardless of offsets i and j against A. 
BasicAA would try to evaluate them into (A + sizeof(A) * i + Offset1) and (A + sizeof(A) * j + Offset2), and probably fail to continue since i and j are not constants.
But I don’t think this is the only reason TBAA is invented. Could someone give me examples showing TBAA performs over BasicAA in some domains?

Thank you

Bekket McClane