[DAGCombine] Fold insert_subvector(bitcast(x),bitcast(y)) -> bitcast(insert_subvector(x,y))

[RKSimon]

Summary

Move this combine from x86 into generic DAGCombine, which currently only manages cases where the bitcast is between types of the same scalarsize.

[nikic]

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

18266 ↗

(On Diff #190009)

Should use getVectorIdxTy?

[RKSimon]

Use getVectorIdxTy

[RKSimon]

ping?

[nikic]

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

18353 ↗

(On Diff #190902)

Can things like insert_subvector x:v7i8, (bitcast y:v2i16 to v4i8), 2 occur? Do we need to check NumElts % Scale == 0 here as well?

[nikic]

Reverse ping. Apart from the previous comment this looks good, and I saw you already added some tests for that as well :)

[RKSimon]

In D59188#1499010, @nikic wrote:

Reverse ping. Apart from the previous comment this looks good, and I saw you already added some tests for that as well :)

Sorry this completely fell off my radar (again) - will take a look

[RKSimon]

Check for (NumElts % Scale == 0) as well

Sorry for taking so long with this @nikic !

[nikic]

LGTM

llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

if (N0.isUndef() && N1.getOpcode() == ISD::INSERT_SUBVECTOR &&
     return DAG.getNode(ISD::INSERT_SUBVECTOR, SDLoc(N), VT, N0,
                        N1.getOperand(1), N2);
 
   if (!isa<ConstantSDNode>(N2))
     return SDValue();
 
   unsigned InsIdx = cast<ConstantSDNode>(N2)->getZExtValue();
 
+  // Push subvector bitcasts to the output, adjusting the index as we go.
+  // insert_subvector(bitcast(v), bitcast(s), c1)
+  // -> bitcast(insert_subvector(v, s, c2))
+  if ((N0.isUndef() || N0.getOpcode() == ISD::BITCAST) &&
+      N1.getOpcode() == ISD::BITCAST) {
+    SDValue N0Src = peekThroughBitcasts(N0);
+    SDValue N1Src = peekThroughBitcasts(N1);
+    EVT N0SrcSVT = N0Src.getValueType().getScalarType();
+    EVT N1SrcSVT = N1Src.getValueType().getScalarType();
+    if ((N0.isUndef() || N0SrcSVT == N1SrcSVT) &&
+        N0Src.getValueType().isVector() && N1Src.getValueType().isVector()) {
+      EVT NewVT;
+      SDLoc DL(N);
+      SDValue NewIdx;
+      MVT IdxVT = TLI.getVectorIdxTy(DAG.getDataLayout());
+      LLVMContext &Ctx = *DAG.getContext();
+      unsigned NumElts = VT.getVectorNumElements();
+      unsigned EltSizeInBits = VT.getScalarSizeInBits();
+      if ((EltSizeInBits % N1SrcSVT.getSizeInBits()) == 0) {
+        unsigned Scale = EltSizeInBits / N1SrcSVT.getSizeInBits();
+        NewVT = EVT::getVectorVT(Ctx, N1SrcSVT, NumElts * Scale);
+        NewIdx = DAG.getConstant(InsIdx * Scale, DL, IdxVT);
+      } else if ((N1SrcSVT.getSizeInBits() % EltSizeInBits) == 0) {
+        unsigned Scale = N1SrcSVT.getSizeInBits() / EltSizeInBits;
+        if ((NumElts % Scale) == 0 && (InsIdx % Scale) == 0) {
+          NewVT = EVT::getVectorVT(Ctx, N1SrcSVT, NumElts / Scale);
+          NewIdx = DAG.getConstant(InsIdx / Scale, DL, IdxVT);
+        }
+      }
+      if (NewIdx && hasOperation(ISD::INSERT_SUBVECTOR, NewVT)) {
+        SDValue Res = DAG.getBitcast(NewVT, N0Src);
+        Res = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, NewVT, Res, N1Src, NewIdx);
+        return DAG.getBitcast(VT, Res);
+      }
+    }
+  }
+
   // Canonicalize insert_subvector dag nodes.
   // Example:
   // (insert_subvector (insert_subvector A, Idx0), Idx1)
   // -> (insert_subvector (insert_subvector A, Idx1), Idx0)
   if (N0.getOpcode() == ISD::INSERT_SUBVECTOR && N0.hasOneUse() &&
       N1.getValueType() == N0.getOperand(1).getValueType() &&
       isa<ConstantSDNode>(N0.getOperand(2))) {
     unsigned OtherIdx = N0.getConstantOperandVal(2);

llvm/trunk/lib/Target/X86/X86ISelLowering.cpp

if (ExtIdxVal != 0) {
       // Now insert the extracted portion.
       for (int i = 0; i != SubVecNumElts; ++i)
         Mask[i + IdxVal] = i + ExtIdxVal + VecNumElts;
 
       return DAG.getVectorShuffle(OpVT, dl, Vec, SubVec.getOperand(0), Mask);
     }
   }
 
-  // Push subvector bitcasts to the output, adjusting the index as we go.
-  // insert_subvector(bitcast(v), bitcast(s), c1) ->
-  // bitcast(insert_subvector(v,s,c2))
-  // TODO: Move this to generic - which only supports same scalar sizes.
-  if ((Vec.isUndef() || Vec.getOpcode() == ISD::BITCAST) &&
-      SubVec.getOpcode() == ISD::BITCAST) {
-    SDValue VecSrc = peekThroughBitcasts(Vec);
-    SDValue SubVecSrc = peekThroughBitcasts(SubVec);
-    MVT VecSrcSVT = VecSrc.getSimpleValueType().getScalarType();
-    MVT SubVecSrcSVT = SubVecSrc.getSimpleValueType().getScalarType();
-    if (Vec.isUndef() || VecSrcSVT == SubVecSrcSVT) {
-      MVT NewOpVT;
-      SDValue NewIdx;
-      unsigned NumElts = OpVT.getVectorNumElements();
-      unsigned EltSizeInBits = OpVT.getScalarSizeInBits();
-      if ((EltSizeInBits % SubVecSrcSVT.getSizeInBits()) == 0) {
-        unsigned Scale = EltSizeInBits / SubVecSrcSVT.getSizeInBits();
-        NewOpVT = MVT::getVectorVT(SubVecSrcSVT, NumElts * Scale);
-        NewIdx = DAG.getIntPtrConstant(IdxVal * Scale, dl);
-      } else if ((SubVecSrcSVT.getSizeInBits() % EltSizeInBits) == 0) {
-        unsigned Scale = SubVecSrcSVT.getSizeInBits() / EltSizeInBits;
-        if ((IdxVal % Scale) == 0) {
-          NewOpVT = MVT::getVectorVT(SubVecSrcSVT, NumElts / Scale);
-          NewIdx = DAG.getIntPtrConstant(IdxVal / Scale, dl);
-        }
-      }
-      if (NewIdx && DAG.getTargetLoweringInfo().isOperationLegal(
-                        ISD::INSERT_SUBVECTOR, NewOpVT)) {
-        SDValue Res = DAG.getBitcast(NewOpVT, VecSrc);
-        Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, NewOpVT, Res, SubVecSrc,
-                          NewIdx);
-        return DAG.getBitcast(OpVT, Res);
-      }
-    }
-  }
-
   // Match concat_vector style patterns.
   SmallVector<SDValue, 2> SubVectorOps;
   if (collectConcatOps(N, SubVectorOps))
     if (SDValue Fold =
             combineConcatVectorOps(dl, OpVT, SubVectorOps, DAG, DCI, Subtarget))
       return Fold;
 
   // If we are inserting into both halves of the vector, the starting vector