[llvm] [ValueTracking] Add support for most `llvm.vector.reduce.*` ops. (PR #88320)

[Yingwei Zheng via llvm-commits]

================
@@ -1621,14 +1621,52 @@ static void computeKnownBitsFromOperator(const Operator *I,
         computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
         Known = KnownBits::ssub_sat(Known, Known2);
         break;
-        // for min/max reduce, any bit common to each element in the input vec
-        // is set in the output.
+        // for min/max/and/or reduce, any bit common to each element in the
+        // input vec is set in the output.
+      case Intrinsic::vector_reduce_and:
+      case Intrinsic::vector_reduce_or:
       case Intrinsic::vector_reduce_umax:
       case Intrinsic::vector_reduce_umin:
       case Intrinsic::vector_reduce_smax:
       case Intrinsic::vector_reduce_smin:
         computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
         break;
+      case Intrinsic::vector_reduce_xor:
+        computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
+        // The zeros common to all vecs are zero in the output.
+        // If the number of elements is odd, then the common ones remain. If the
+        // number of elements is even, then the common ones becomes zeros.
+        if (auto *VecTy =
+                dyn_cast<FixedVectorType>(I->getOperand(0)->getType())) {
+          // Even, so the ones become zeros.
+          if ((VecTy->getNumElements() % 2) == 0) {
+            Known.Zero |= Known.One;
+            Known.One.clearAllBits();
+          }
+
+        } else
+          Known.One.clearAllBits();
+        break;
+      case Intrinsic::vector_reduce_mul:
+      case Intrinsic::vector_reduce_add:
+        // We compute the common bits for all elements then apply the reduce op
+        // NumEle times. This is mostly useful for known high zeros.
+        if (auto *VecTy =
+                dyn_cast<FixedVectorType>(I->getOperand(0)->getType())) {
+          computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
+          KnownBits SingleKnown = Known;
+          for (unsigned i = 1, e = VecTy->getNumElements(); i < e; ++i) {
----------------
dtcxzyw wrote:

You can reduce the computational complexity from `O(n)` to `O(lgn)` with [exponentiation by squaring](https://en.wikipedia.org/wiki/Exponentiation_by_squaring).


https://github.com/llvm/llvm-project/pull/88320