[SelectionDAG] Enable iX SimplifyDemandedBits to vXi1 SimplifyDemandedVectorElts simplification

[craig.topper]

Summary

Patch showing the effect of enabling bool vector oversimplification.

Non-VLX builds can simplify a kshift shuffle, but VLX builds simplify:

insert_subvector v8i zeroinitializer, v2i --> insert_subvector v8i undef, v2i

Preventing the removal of the AND to clear the upper bits of result

[RKSimon]

I haven't looked into this yet but I'm hoping this can be fixed by reusing the isTypePromotionOfi1ZeroUpBits code somehow.

[RKSimon]

Rebased - I was hoping D66489 would help more tbh...

[craig.topper]

Added a quick hack DAG combine and a FIXME to prevent the regressions.

[RKSimon]

LGTM - thanks for handling this!

llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp

if (!TLO.LegalOperations() && !VT.isVector() && !SrcVT.isVector() &&
         return TLO.CombineTo(Op,
                              TLO.DAG.getNode(ISD::SHL, dl, VT, Sign, ShAmt));
       }
     }
 
     // Bitcast from a vector using SimplifyDemanded Bits/VectorElts.
     // Demand the elt/bit if any of the original elts/bits are demanded.
     // TODO - bigendian once we have test coverage.
-    // TODO - bool vectors once SimplifyDemandedVectorElts has SETCC support.
-    if (SrcVT.isVector() && NumSrcEltBits > 1 &&
-        (BitWidth % NumSrcEltBits) == 0 &&
+    if (SrcVT.isVector() && (BitWidth % NumSrcEltBits) == 0 &&
         TLO.DAG.getDataLayout().isLittleEndian()) {
       unsigned Scale = BitWidth / NumSrcEltBits;
       unsigned NumSrcElts = SrcVT.getVectorNumElements();
       APInt DemandedSrcBits = APInt::getNullValue(NumSrcEltBits);
       APInt DemandedSrcElts = APInt::getNullValue(NumSrcElts);
       for (unsigned i = 0; i != Scale; ++i) {
         unsigned Offset = i * NumSrcEltBits;
         APInt Sub = DemandedBits.extractBits(NumSrcEltBits, Offset);

llvm/lib/Target/X86/X86ISelLowering.cpp

if ((VT == MVT::v4i1 || VT == MVT::v2i1) && SrcVT.isScalarInteger() &&
       return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, N0,
                          DAG.getIntPtrConstant(0, dl));
     }
 
     // If this is a bitcast between a MVT::v4i1/v2i1 and an illegal integer
     // type, widen both sides to avoid a trip through memory.
     if ((SrcVT == MVT::v4i1 || SrcVT == MVT::v2i1) && VT.isScalarInteger() &&
         Subtarget.hasAVX512()) {
+      // Use zeros for the widening if we already have some zeroes. This can
+      // allow SimplifyDemandedBits to remove scalar ANDs that may be down
+      // stream of this.
+      // FIXME: It might make sense to detect a concat_vectors with a mix of
+      // zeroes and undef and turn it into insert_subvector for i1 vectors as
+      // a separate combine. What we can't do is canonicalize the operands of
+      // such a concat or we'll get into a loop with SimplifyDemandedBits.
+      if (N0.getOpcode() == ISD::CONCAT_VECTORS) {
+        SDValue LastOp = N0.getOperand(N0.getNumOperands() - 1);
+        if (ISD::isBuildVectorAllZeros(LastOp.getNode())) {
+          SrcVT = LastOp.getValueType();
+          unsigned NumConcats = 8 / SrcVT.getVectorNumElements();
+          SmallVector<SDValue, 4> Ops(N0->op_begin(), N0->op_end());
+          Ops.resize(NumConcats, DAG.getConstant(0, dl, SrcVT));
+          N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i1, Ops);
+          N0 = DAG.getBitcast(MVT::i8, N0);
+          return DAG.getNode(ISD::TRUNCATE, dl, VT, N0);
+        }
+      }
+
       unsigned NumConcats = 8 / SrcVT.getVectorNumElements();
       SmallVector<SDValue, 4> Ops(NumConcats, DAG.getUNDEF(SrcVT));
       Ops[0] = N0;
       N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i1, Ops);
       N0 = DAG.getBitcast(MVT::i8, N0);
       return DAG.getNode(ISD::TRUNCATE, dl, VT, N0);
     }
   }