[llvm] [InstCombine] fold icmp with add/sub instructions having the same operands (PR #143241)

[Nikita Popov via llvm-commits]

================
@@ -7728,6 +7728,45 @@ Instruction *InstCombinerImpl::visitICmpInst(ICmpInst &I) {
     }
   }
 
+  // When comparing results of sub and add instructions with identical operands,
+  // optimization is valid when the comparison type and overflow flags satisfy:
+  //   - Signed comparisons (slt/sgt/sle/sge) require 'nsw' flags
+  //   - Unsigned comparisons (ult/ugt/ule/uge) require 'nuw' flags
+  //   - Equality comparisons (eq/ne) accept either 'nsw' or 'nuw'
+  //
+  // If conditions are met, the comparison simplifies to a zero comparison on
+  // the second operand of the sub instruction with the swapped predicate.
+  // Example transformation for signed comparison:
+  //   %sub = sub nsw i8 %x, %y
+  //   %add = add nsw i8 %x, %y
+  //   %cmp = icmp sgt i8 %sub, %add  // (x - y) > (x + y)
+  // becomes:
+  //   %cmp = icmp slt i8 %y, 0      // y < 0
+  //
+  // This handles similar cases to transform.
----------------
nikic wrote:

Please make this comment much more compact, something along these lines is sufficient:
```
// icmp sgt (sub nsw x, y), (add nsw x, y)  -->  icmp slt y, 0
// and other variants with nuw/nsw flags matching the icmp predicate.
```

https://github.com/llvm/llvm-project/pull/143241