[Mlir-commits] [mlir] Improving MinUI and MaxUI implementation in valuebounds (PR #207701)

[Matthias Springer]

================
@@ -354,20 +354,25 @@ struct MinUIOpInterface
     assert(value == minOp.getResult() && "invalid value");
 
     // ValueBoundsConstraintSet models values as signed integers (e.g. an i8
-    // 0xff is treated as -1, not 255).So, we can only derive bounds for minui
-    // if both operands are provably non-negative.
+    // 0xff is treated as -1, not 255). For an unsigned minimum it is enough
+    // that a single operand is provably non-negative: minui(x, y) is in
+    // [0, y] whenever y >= 0 (and symmetrically for x).
     bool lhsNonNegative =
         ValueBoundsConstraintSet::isProvablyNonNegative(minOp.getLhs(), cstr);
     bool rhsNonNegative =
         ValueBoundsConstraintSet::isProvablyNonNegative(minOp.getRhs(), cstr);
-    if (!lhsNonNegative || !rhsNonNegative)
+    if (!lhsNonNegative && !rhsNonNegative)
       return;
 
     cstr.bound(value) >= 0;
-    AffineExpr lhs = cstr.getExpr(minOp.getLhs());
-    AffineExpr rhs = cstr.getExpr(minOp.getRhs());
-    cstr.bound(value) <= lhs;
-    cstr.bound(value) <= rhs;
+    if (lhsNonNegative) {
+      AffineExpr lhs = cstr.getExpr(minOp.getLhs());
+      cstr.bound(value) <= lhs;
----------------
matthias-springer wrote:

This one here is much simpler than the `>= 0` bound: if the signed interpretation and the unsigned interpretation match, then `minsi` and `minui` have the same bound. I would still add a comment about this.

https://github.com/llvm/llvm-project/pull/207701