[llvm] [InstCombine] Handle `ICMP_EQ` when flooring by constant two (PR #73706)

[Yingwei Zheng via llvm-commits]

================
@@ -1234,14 +1234,21 @@ static Instruction *foldAddToAshr(BinaryOperator &Add) {
     return nullptr;
 
   // Rounding is done by adding -1 if the dividend (X) is negative and has any
-  // low bits set. The canonical pattern for that is an "ugt" compare with SMIN:
-  // sext (icmp ugt (X & (DivC - 1)), SMIN)
-  const APInt *MaskC;
+  // low bits set. It recognizes two canonical patterns:
+  // 1. For an 'ugt' cmp with the signed minimum value (SMIN), the
+  //    pattern is: sext (icmp ugt (X & (DivC - 1)), SMIN).
+  // 2. For an 'eq' cmp, the pattern is: sext (icmp eq (SMIN + 1), SMIN + 1).
----------------
dtcxzyw wrote:

```suggestion
  // 2. For an 'eq' cmp, the pattern is: sext (icmp eq X & (SMIN + 1), SMIN + 1).
```

https://github.com/llvm/llvm-project/pull/73706