[llvm] [DA] Disable the BanerjeeMIV dependence test (PR #174733)

[Sjoerd Meijer via llvm-commits]

sjoerdmeijer wrote:

Hi @kasuga-fj , @amehsan , @Meinersbur ,
Here's a sign of life; I have been less active upstream because I am prototyping the full flow starting from Fortran code (Flang) to interchanging a loop nest in our motivating example (we had too many patches flying around, but not the full flow). The corresponding IR of our motivating example is this one: [test/Transforms/LoopInterchange/large-nested-4d.ll](https://github.com/llvm/llvm-project/blob/main/llvm/test/Transforms/LoopInterchange/large-nested-4d.ll).

After work in Flang and Delinearization, and correct subscript recognition, there is now only one dependency still in the way. I.e., DA finds an output dependency that shouldn't be there. Extracted from the pseudo code of that motivating example and regression test, it is this store:

     store GlobL(NY*i+j,NY*IL+JL)

These subscripts have multiple index variables, and so I will need a MIV test for this. Thus I am kind of pre-announcing I will need this test back one day to handle this case.

What I am investigating and fixing at the moment is to provide Banerjee with more information; there are symbols in its upperbound and it can't make the right decisions. So I am not yet working on correct issues here, but first investigating how to support this case.

Hope this extra bit of context helps. 

https://github.com/llvm/llvm-project/pull/174733