[WIP][X86] `matchBinaryShuffle()`: lift same-dimensions restriction on `OR` blend

[lebedev.ri]

Summary

Does not affect any existing tests.
Not really sure if i will want to add some, or let this rot.

llvm/lib/Target/X86/X86ISelLowering.cpp

for (unsigned i = 0; i != NumMaskElts; ++i) {
       break;
     }
     if (IsBlend) {
       if (DAG.computeKnownBits(V1, DemandedZeroV1).isZero() &&
           DAG.computeKnownBits(V2, DemandedZeroV2).isZero()) {
         Shuffle = ISD::OR;
         SrcVT = DstVT = MaskVT.changeTypeToInteger();
         return true;
-      } else if (NumV1Elts == NumV2Elts && NumV1Elts == NumMaskElts) {
+      } else {
         // FIXME: handle mismatched sizes?
         // TODO: investigate if `ISD::OR` handling in
         // `TargetLowering::SimplifyDemandedVectorElts` can be improved instead.
         auto computeKnownBitsElementWise = [&DAG](SDValue V) {
           unsigned NumElts = V.getValueType().getVectorNumElements();
           KnownBits Known(NumElts);
           for (unsigned EltIdx = 0; EltIdx != NumElts; ++EltIdx) {
             APInt Mask = APInt::getOneBitSet(NumElts, EltIdx);
             KnownBits PeepholeKnown = DAG.computeKnownBits(V, Mask);
             if (PeepholeKnown.isZero())
               Known.Zero.setBit(EltIdx);
             if (PeepholeKnown.isAllOnes())
               Known.One.setBit(EltIdx);
           }
           return Known;
         };
 
         KnownBits V1Known = computeKnownBitsElementWise(V1);
         KnownBits V2Known = computeKnownBitsElementWise(V2);
 
+        for (APInt *K :
+             {&V1Known.Zero, &V1Known.One, &V2Known.Zero, &V2Known.One})
+          *K = APIntOps::ScaleBitMask(*K, NumMaskElts,
+                                      APIntOps::BitMergingApproach::Lossy);
+
         for (unsigned i = 0; i != NumMaskElts && IsBlend; ++i) {
           int M = Mask[i];
           if (M == SM_SentinelUndef)
             continue;
           if (M == SM_SentinelZero) {
             IsBlend &= V1Known.Zero[i] && V2Known.Zero[i];
             continue;
           }