[llvm] [SelectionDAG] Add computeKnownBits support for VP_LOAD_FF's second result (PR #227389)

[Craig Topper via llvm-commits]

================
@@ -4541,6 +4541,25 @@ KnownBits SelectionDAG::computeKnownBits(SDValue Op, const APInt &DemandedElts,
         Known, MF, MF.getFrameInfo().getObjectAlign(FrameIdx));
     break;
   }
+  case ISD::VP_LOAD_FF: {
+    if (Op.getResNo() != 1)
+      break;
+    // The second result of vp.load.ff is an unsigned value that is less than or
+    // equal to the EVL operand.
+    KnownBits VLKB =
+        computeKnownBits(Op.getOperand(3), DemandedElts, Depth + 1);
+    // The new VL is also bounded by the largest vector length.
+    EVT ResVT = Op->getValueType(0);
+    auto ResKB = KnownBits::makeConstant(
+        APInt(BitWidth, ResVT.getVectorMinNumElements()));
+    if (ResVT.isScalableVector()) {
+      const Function &F = getMachineFunction().getFunction();
+      ResKB = KnownBits::mul(getVScaleRange(&F, BitWidth).toKnownBits(), ResKB);
----------------
topperc wrote:

Do we need to detect overflow on this multiply? If the vscale_range attribute isn't present the upper bound is a large number. Granted in that case the lower bound 1 so toKnownBits probably can't resolve anything.

https://github.com/llvm/llvm-project/pull/227389