[llvm] [DA] Add tests where dependencies are missed due to overflow (NFC) (PR #164246)

[Sjoerd Meijer via llvm-commits]

sjoerdmeijer wrote:

> @sjoerdmeijer These are the cases I was able to find where intermediate computations cause overflow, leading to incorrect results being reported. Note that
> 
> * I didn't check all functions, especially those that based on greatest common divisor algorithms (exactSIV, exactRDIV ~and gcdMIV~).
>   
>   * It is more difficult to reason about the absence of overflow in them. I personally think we should conservatively insert overflow checks in these functions, even if we cannot find any test cases that trigger an overflow.
> * I also didn't investigate functions that I believe should be removed (e.g., banerjeeMIV).
> * These issues are not related to monotonicity checks. As far as I can tell, all of these cases are monotonic.
> * There are other correctness issues due to other reasons. I'll share them separately.

Just wanted to leave this message: thanks for the ping. I am attending the llvm dev conf, and am traveling. But I will be back in the office next week, and will be able to help with DA. Next week I am going to pick up a few things. 

https://github.com/llvm/llvm-project/pull/164246