[llvm] [InstCombine] Fix flag propagation in `foldSelectIntoOp` (PR #162003)

[Yunbo Ni via llvm-commits]

cardigan1008 wrote:

@dtcxzyw There seems to be an uncovered edge case for `fdiv` in this fix.

The current logic assumes that if a `select` has both `ninf` and `nnan`, then the operands of the folded operation must be finite. While this assumption holds for `fadd`, `fsub`, and `fmul`, it does not hold for `fdiv`.

For example:
  ```
A = 1.0, B = +Inf
  A / B = 0.0  (finite, non-NaN)
```

In the original IR, a `select ninf nnan` returning this result is valid. However, after the transformation, InstCombine produces: A / (select ninf nnan cond, B, 1.0)

If `cond` is true, the inner `select` returns `B = +Inf`, and due to the propagated `ninf`, this becomes poison. This introduces poison that was not present in the original program.

alive2: https://alive2.llvm.org/ce/z/e7be8f

https://github.com/llvm/llvm-project/pull/162003