Fix underflow in highestPowOf2Divisor() (#4562)

Prevent highestPowOf2Divisor() from producing an underflow when the
argument is INT_MIN.

Co-authored-by: moerafaat <moerafaat@aucegypt.edu>

include/triton/Dialect/Triton/IR/Utility.h

@@ -31,7 +31,11 @@ template <typename Int> Int ceil(Int m, Int n) { return (m + n - 1) / n; }
 
 /// Get the highest power of 2 divisor of an integer.
 template <typename T> T highestPowOf2Divisor(T n) {
-  if (n == 0) {
+  // When n is 0 or min, return the highest power of 2. The min case is handled
+  // separately to avoid underflow when T is a signed integer. Technically
+  // in that case the correct divisor is -n, but this value is outside the
+  // range of possible values, so we take the next best alternative.
+  if (n == 0 || n == std::numeric_limits<T>::min()) {
     return (static_cast<T>(1) << (sizeof(T) * 8 - 2));
   }
   return (n & (~(n - 1)));

test/Analysis/test-alignment.mlir

@@ -863,3 +863,14 @@ tt.func public @chained_for(%8: tensor<128x64x!tt.ptr<bf16>> {tt.divisibility =
   }
   tt.return
 }
+
+// -----
+
+// CHECK-LABEL: @int_min_does_not_underflow_in_analysis
+module {
+  tt.func @int_min_does_not_underflow_in_analysis() -> i64 {
+    // CHECK: divisibility = [4611686018427387904]
+    %int_min = arith.constant -9223372036854775808 : i64
+    tt.return %int_min : i64
+  }
+}