[llvm] [ValueTracking] Implement `isKnownNonZero`/`computeKnownBits` for `llvm.vector.reduce.{add,mul}` (PR #88410)

[via llvm-commits]

================
@@ -2904,6 +2924,41 @@ static bool isKnownNonZeroFromOperator(const Operator *I,
       case Intrinsic::vector_reduce_smax:
       case Intrinsic::vector_reduce_smin:
         return isKnownNonZero(II->getArgOperand(0), Depth, Q);
+        // If we know the reduction doesn't overflow and all elements are
+        // non-zero, the reduction is non-zero.
+      case Intrinsic::vector_reduce_mul:
+      case Intrinsic::vector_reduce_add:
+        if (computeKnownBits(I, Depth + 1, Q).isNonZero())
+          return true;
----------------
goldsteinn wrote:

BTW, I wouldn't spend your time reviewing this unless we decide to actually try to get it in.

https://github.com/llvm/llvm-project/pull/88410