[ARM] When a bitcast is about to be turned into a VMOVDRR, try to combine it

with its source instead of forcing the values on GPRs.

This improves the lowering of vector code when such bitcasts happen in the
middle of vector computations.

rdar://problem/23691584 

llvm-svn: 254684

1 files changed, 55 insertions, 0 deletions

diff --git a/llvm/lib/Target/ARM/ARMISelLowering.cpp b/llvm/lib/Target/ARM/ARMISelLowering.cpp
index 33f74a3ba9f..23f7bd0f4c8 100644
--- a/llvm/lib/Target/ARM/ARMISelLowering.cpp
+++ b/llvm/lib/Target/ARM/ARMISelLowering.cpp
@@ -4139,6 +4139,56 @@ static void ExpandREAD_REGISTER(SDNode *N, SmallVectorImpl<SDValue> &Results,
   Results.push_back(Read.getOperand(0));
 }
 
+/// \p BC is a bitcast that is about to be turned into a VMOVDRR.
+/// When \p DstVT, the destination type of \p BC, is on the vector
+/// register bank and the source of bitcast, \p Op, operates on the same bank,
+/// it might be possible to combine them, such that everything stays on the
+/// vector register bank.
+/// \p return The node that would replace \p BT, if the combine
+/// is possible.
+static SDValue CombineVMOVDRRCandidateWithVecOp(const SDNode *BC,
+                                                SelectionDAG &DAG) {
+  SDValue Op = BC->getOperand(0);
+  EVT DstVT = BC->getValueType(0);
+
+  // The only vector instruction that can produce a scalar (remember,
+  // since the bitcast was about to be turned into VMOVDRR, the source
+  // type is i64) from a vector is EXTRACT_VECTOR_ELT.
+  // Moreover, we can do this combine only if there is one use.
+  // Finally, if the destination type is not a vector, there is not
+  // much point on forcing everything on the vector bank.
+  if (!DstVT.isVector() || Op.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
+      !Op.hasOneUse())
+    return SDValue();
+
+  // If the index is not constant, we will introduce an additional
+  // multiply that will stick.
+  // Give up in that case.
+  ConstantSDNode *Index = dyn_cast<ConstantSDNode>(Op.getOperand(1));
+  if (!Index)
+    return SDValue();
+  unsigned DstNumElt = DstVT.getVectorNumElements();
+
+  // Compute the new index.
+  const APInt &APIntIndex = Index->getAPIntValue();
+  APInt NewIndex(APIntIndex.getBitWidth(), DstNumElt);
+  NewIndex *= APIntIndex;
+  // Check if the new constant index fits into i32.
+  if (NewIndex.getBitWidth() > 32)
+    return SDValue();
+
+  // vMTy bitcast(i64 extractelt vNi64 src, i32 index) ->
+  // vMTy extractsubvector vNxMTy (bitcast vNi64 src), i32 index*M)
+  SDLoc dl(Op);
+  SDValue ExtractSrc = Op.getOperand(0);
+  EVT VecVT = EVT::getVectorVT(
+      *DAG.getContext(), DstVT.getScalarType(),
+      ExtractSrc.getValueType().getVectorNumElements() * DstNumElt);
+  SDValue BitCast = DAG.getNode(ISD::BITCAST, dl, VecVT, ExtractSrc);
+  return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DstVT, BitCast,
+                     DAG.getConstant(NewIndex.getZExtValue(), dl, MVT::i32));
+}
+
 /// ExpandBITCAST - If the target supports VFP, this function is called to
 /// expand a bit convert where either the source or destination type is i64 to
 /// use a VMOVDRR or VMOVRRD node.  This should not be done when the non-i64
@@ -4158,6 +4208,11 @@ static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
 
   // Turn i64->f64 into VMOVDRR.
   if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
+    // Do not force values to GPRs (this is what VMOVDRR does for the inputs)
+    // if we can combine the bitcast with its source.
+    if (SDValue Val = CombineVMOVDRRCandidateWithVecOp(N, DAG))
+      return Val;
+
     SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
                              DAG.getConstant(0, dl, MVT::i32));
     SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,