[flang-commits] [flang] [flang] Restructured TBAA trees in AddAliasTags pass. (PR #136725)

[Slava Zakharin via flang-commits]

vzakhari wrote:

> Makes sense to me. Does it work ok with LLVM inlining?

Almost missed the question :)

Yes, it should work same way with LLVM inlining as before, but that does not mean our current TBAA approach is valid right now.

Our TBAA is invalid for the following case:
```
subroutine callee(x,y)
  real :: x, y
  x = 1.0
  y = 2.0
end subroutine callee

subroutine test(x,y)
  real, target :: x(2), y(3)
  call callee(x(1), y(1))
  call callee(x(2), y(2))
end subroutine test
```

Consider the case where `LOC(x(1)) == LOC(y(2))`.  After LLVM inlining we will have:
```
define void @test_(ptr writeonly captures(none) %0, ptr writeonly captures(none) %1) local_unnamed_addr !dbg !4 {
  store float 1.000000e+00, ptr %0, align 4, !dbg !14, !tbaa !19
  store float 2.000000e+00, ptr %1, align 4, !dbg !25, !tbaa !26
  %3 = getelementptr i8, ptr %0, i64 4, !dbg !28
  %4 = getelementptr i8, ptr %1, i64 4, !dbg !28
  store float 1.000000e+00, ptr %3, align 4, !dbg !29, !tbaa !19
  store float 2.000000e+00, ptr %4, align 4, !dbg !31, !tbaa !26
  ret void, !dbg !32
}
```

So it says that the first and last stores do not alias, which is not true.

https://github.com/llvm/llvm-project/pull/136725