SelectionDAG shuffle nodes do not allow operands with different numbers of

elements than the result vector type.  So, when an instruction like:

%8 = shufflevector <2 x float> %4, <2 x float> %7, <4 x i32> <i32 1, i32 0, i32 3, i32 2>

is translated to a DAG, each operand is changed to a concat_vectors node that appends 2 undef elements.  That is:

shuffle [a,b], [c,d] is changed to:
shuffle [a,b,u,u], [c,d,u,u]

That's probably the right thing for x86 but for NEON, we'd much rather have:

shuffle [a,b,c,d], undef

Teach the DAG combiner how to do that transformation for ARM.  Radar 8597007.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@117482 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/ARM/ARMISelLowering.cpp b/lib/Target/ARM/ARMISelLowering.cpp
index 4a7866592d65301d98fbd4f1d9616746f5a539d8..5538a28b3a8fe9e2cadbf150cd08ec306674b1e1 100644 (file)
--- a/lib/Target/ARM/ARMISelLowering.cpp
+++ b/lib/Target/ARM/ARMISelLowering.cpp
@@ -460,6 +460,7 @@ ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
     setTargetDAGCombine(ISD::ANY_EXTEND);
     setTargetDAGCombine(ISD::SELECT_CC);
     setTargetDAGCombine(ISD::BUILD_VECTOR);
+    setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
   }
 
   computeRegisterProperties();
@@ -4531,6 +4532,59 @@ static SDValue PerformBUILD_VECTORCombine(SDNode *N, SelectionDAG &DAG) {
   return SDValue();
 }
 
+/// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
+/// ISD::VECTOR_SHUFFLE.
+static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
+  // The LLVM shufflevector instruction does not require the shuffle mask
+  // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
+  // have that requirement.  When translating to ISD::VECTOR_SHUFFLE, if the
+  // operands do not match the mask length, they are extended by concatenating
+  // them with undef vectors.  That is probably the right thing for other
+  // targets, but for NEON it is better to concatenate two double-register
+  // size vector operands into a single quad-register size vector.  Do that
+  // transformation here:
+  //   shuffle(concat(v1, undef), concat(v2, undef)) ->
+  //   shuffle(concat(v1, v2), undef)
+  SDValue Op0 = N->getOperand(0);
+  SDValue Op1 = N->getOperand(1);
+  if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
+      Op1.getOpcode() != ISD::CONCAT_VECTORS ||
+      Op0.getNumOperands() != 2 ||
+      Op1.getNumOperands() != 2)
+    return SDValue();
+  SDValue Concat0Op1 = Op0.getOperand(1);
+  SDValue Concat1Op1 = Op1.getOperand(1);
+  if (Concat0Op1.getOpcode() != ISD::UNDEF ||
+      Concat1Op1.getOpcode() != ISD::UNDEF)
+    return SDValue();
+  // Skip the transformation if any of the types are illegal.
+  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+  EVT VT = N->getValueType(0);
+  if (!TLI.isTypeLegal(VT) ||
+      !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
+      !TLI.isTypeLegal(Concat1Op1.getValueType()))
+    return SDValue();
+
+  SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
+                                  Op0.getOperand(0), Op1.getOperand(0));
+  // Translate the shuffle mask.
+  SmallVector<int, 16> NewMask;
+  unsigned NumElts = VT.getVectorNumElements();
+  unsigned HalfElts = NumElts/2;
+  ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
+  for (unsigned n = 0; n < NumElts; ++n) {
+    int MaskElt = SVN->getMaskElt(n);
+    int NewElt = -1;
+    if (MaskElt < HalfElts)
+      NewElt = MaskElt;
+    else if (MaskElt >= NumElts && MaskElt < NumElts + HalfElts)
+      NewElt = HalfElts + MaskElt - NumElts;
+    NewMask.push_back(NewElt);
+  }
+  return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
+                              DAG.getUNDEF(VT), NewMask.data());
+}
+
 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
 /// ARMISD::VDUPLANE.
 static SDValue PerformVDUPLANECombine(SDNode *N, SelectionDAG &DAG) {
@@ -4939,6 +4993,7 @@ SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
   case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
   case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
   case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI.DAG);
+  case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
   case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI.DAG);
   case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
   case ISD::SHL:

diff --git a/test/CodeGen/ARM/vrev.ll b/test/CodeGen/ARM/vrev.ll
index e1fe64b02d9d31c02113eec9143b02f12d05684d..f0f9e4e339b4294de9996632ffd6313e39beb5fd 100644 (file)
--- a/test/CodeGen/ARM/vrev.ll
+++ b/test/CodeGen/ARM/vrev.ll
@@ -129,3 +129,21 @@ define <8 x i16> @test_vrev32Q16_undef(<8 x i16>* %A) nounwind {
        %tmp2 = shufflevector <8 x i16> %tmp1, <8 x i16> undef, <8 x i32> <i32 undef, i32 0, i32 undef, i32 2, i32 5, i32 4, i32 7, i32 undef>
        ret <8 x i16> %tmp2
 }
+
+; A vcombine feeding a VREV should not obscure things.  Radar 8597007.
+
+define void @test_with_vcombine(<4 x float>* %v) nounwind {
+;CHECK: test_with_vcombine:
+;CHECK-NOT: vext
+;CHECK: vrev64.32
+  %tmp1 = load <4 x float>* %v, align 16
+  %tmp2 = bitcast <4 x float> %tmp1 to <2 x double>
+  %tmp3 = extractelement <2 x double> %tmp2, i32 0
+  %tmp4 = bitcast double %tmp3 to <2 x float>
+  %tmp5 = extractelement <2 x double> %tmp2, i32 1
+  %tmp6 = bitcast double %tmp5 to <2 x float>
+  %tmp7 = fadd <2 x float> %tmp6, %tmp6
+  %tmp8 = shufflevector <2 x float> %tmp4, <2 x float> %tmp7, <4 x i32> <i32 1, i32 0, i32 3, i32 2>
+  store <4 x float> %tmp8, <4 x float>* %v, align 16
+  ret void
+}