[llvm] [ValueTracking] improve `isKnownNonZero` precision for `smax` (PR #88170)

[via llvm-commits]

================
@@ -2828,23 +2828,47 @@ static bool isKnownNonZeroFromOperator(const Operator *I,
       case Intrinsic::uadd_sat:
         return isKnownNonZero(II->getArgOperand(1), DemandedElts, Depth, Q) ||
                isKnownNonZero(II->getArgOperand(0), DemandedElts, Depth, Q);
-      case Intrinsic::smin:
       case Intrinsic::smax: {
-        auto KnownOpImpliesNonZero = [&](const KnownBits &K) {
-          return II->getIntrinsicID() == Intrinsic::smin
-                     ? K.isNegative()
-                     : K.isStrictlyPositive();
-        };
-        KnownBits XKnown =
+        // if either arg is strictly positive the result is non-zero. Otherwise
+        // the result is non-zero if both ops are non-zero.
+        KnownBits Op1Known =
+            computeKnownBits(II->getArgOperand(1), DemandedElts, Depth, Q);
+        // Avoid re-computing isKnownNonZero if we already failed once.
+        bool OpsMaybeZero = false;
+        if (Op1Known.isNonNegative()) {
+          if (Op1Known.isNonZero() ||
+              isKnownNonZero(II->getArgOperand(1), DemandedElts, Depth, Q))
+            return true;
+          OpsMaybeZero = true;
----------------
goldsteinn wrote:

Sure either way works.

https://github.com/llvm/llvm-project/pull/88170