[Mlir-commits] [mlir] [MLIR][Utils] Fix overflow in constantTripCount for narrow types (PR #179985)

[Matthias Springer]

================
@@ -365,6 +365,13 @@ std::optional<APInt> constantTripCount(
              << (isSigned ? "isSigned" : "isUnsigned") << ")";
       return APInt(bitwidth, 0);
     }
+    // Extend both operands to a wider bitwidth to avoid overflow when computing
+    // ub - lb. For example, with i8: ub=127, lb=-128, the difference is 255,
+    // which overflows in 8-bit signed arithmetic. We need at least one extra
+    // bit.
+    unsigned extendedWidth = bitwidth + 1;
----------------
matthias-springer wrote:

Can you also extend the documentation in header file: let's add a note that the loops that we model here are not like C++ "for" loops. In C++ "for" loops, you can have different conditions like `<`, `<=`, `>` or even arbitrary C++. The loop that we model here are always `<`.

https://github.com/llvm/llvm-project/pull/179985