[Mlir-commits] [mlir] [mlir][bufferization] Relax happens-before analysis (PR #223777)

[Matthias Springer]

https://github.com/matthias-springer commented:

I found this regression with a coding agent:

```mlir
  func.func @nested_diamond_in_outer_cfg_loop(%f: f32) -> tensor<5xf32> {
    %empty = tensor.empty() : tensor<5xf32>
    %t = linalg.fill ins(%f : f32)
        outs(%empty : tensor<5xf32>) -> tensor<5xf32>

    %c0 = arith.constant 0 : index
    %c1 = arith.constant 1 : index
    %c2 = arith.constant 2 : index
    cf.br ^loop(%c0 : index)

  ^loop(%i: index):
    %take_read = arith.cmpi eq, %i, %c1 : index

    %r = scf.execute_region -> tensor<5xf32> {
      cf.cond_br %take_read, ^read, ^write

    ^read:
      %e = tensor.extract %t[%c0] : tensor<5xf32>
      "test.qux"(%e) : (f32) -> ()
      cf.br ^join(%t : tensor<5xf32>)

    ^write:
      %val = "test.bar"() : () -> f32
      %inserted = tensor.insert %val into %t[%c0] : tensor<5xf32>
      cf.br ^join(%inserted : tensor<5xf32>)

    ^join(%out: tensor<5xf32>):
      scf.yield %out : tensor<5xf32>
    }

    %next = arith.addi %i, %c1 : index
    %again = arith.cmpi ult, %next, %c2 : index
    cf.cond_br %again, ^loop(%next : index),
               ^exit(%r : tensor<5xf32>)

  ^exit(%result: tensor<5xf32>):
    func.return %result : tensor<5xf32>
  }
```

The `tensor.insert_slice` should bufferize out-of-place.


https://github.com/llvm/llvm-project/pull/223777