[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

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@254684 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/ARM/ARMISelLowering.cpp b/lib/Target/ARM/ARMISelLowering.cpp
index 33f74a3ba9fdb6b8b8f46ce9feddb43fe1891a44..23f7bd0f4c8bab1e9753e19bdd5c40ebf311cd15 100644 (file)
--- a/lib/Target/ARM/ARMISelLowering.cpp
+++ b/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,

diff --git a/test/CodeGen/ARM/combine-vmovdrr.ll b/test/CodeGen/ARM/combine-vmovdrr.ll
new file mode 100644 (file)
index 0000000..358f7e3
--- /dev/null
+++ b/test/CodeGen/ARM/combine-vmovdrr.ll
@@ -0,0 +1,72 @@
+; RUN: llc %s -o - | FileCheck %s
+
+target triple = "thumbv7s-apple-ios"
+
+declare <8 x i8> @llvm.arm.neon.vtbl2(<8 x i8> %shuffle.i.i307, <8 x i8> %shuffle.i27.i308, <8 x i8> %vtbl2.i25.i)
+
+; Check that we get the motivating example:
+; The bitcasts force the values to go through the GPRs, whereas
+; they are defined on VPRs and used on VPRs.
+;
+; CHECK-LABEL: motivatingExample:
+; CHECK: vldr [[ARG2_VAL:d[0-9]+]], [r1]
+; CHECK-NEXT: vld1.32 {[[ARG1_VALlo:d[0-9]+]], [[ARG1_VALhi:d[0-9]+]]}, [r0]
+; CHECK-NEXT: vtbl.8 [[RES:d[0-9]+]], {[[ARG1_VALlo]], [[ARG1_VALhi]]}, [[ARG2_VAL]]
+; CHECK-NEXT: vstr [[RES]], [r1]
+; CHECK-NEXT: bx lr
+define void @motivatingExample(<2 x i64>* %addr, <8 x i8>* %addr2) {
+  %shuffle.i.bc.i309 = load <2 x i64>, <2 x i64>* %addr
+  %vtbl2.i25.i = load <8 x i8>, <8 x i8>* %addr2
+  %shuffle.i.extract.i310 = extractelement <2 x i64> %shuffle.i.bc.i309, i32 0
+  %shuffle.i27.extract.i311 = extractelement <2 x i64> %shuffle.i.bc.i309, i32 1
+  %tmp45 = bitcast i64 %shuffle.i.extract.i310 to <8 x i8>
+  %tmp46 = bitcast i64 %shuffle.i27.extract.i311 to <8 x i8>
+  %vtbl2.i25.i313 = tail call <8 x i8> @llvm.arm.neon.vtbl2(<8 x i8> %tmp45, <8 x i8> %tmp46, <8 x i8> %vtbl2.i25.i)
+  store <8 x i8> %vtbl2.i25.i313, <8 x i8>* %addr2
+  ret void
+}
+
+; Check that we do not perform the transformation for dynamic index.
+; CHECK-LABEL: dynamicIndex:
+; CHECK-NOT: mul
+; CHECK: pop
+define void @dynamicIndex(<2 x i64>* %addr, <8 x i8>* %addr2, i32 %index) {
+  %shuffle.i.bc.i309 = load <2 x i64>, <2 x i64>* %addr
+  %vtbl2.i25.i = load <8 x i8>, <8 x i8>* %addr2
+  %shuffle.i.extract.i310 = extractelement <2 x i64> %shuffle.i.bc.i309, i32 %index
+  %shuffle.i27.extract.i311 = extractelement <2 x i64> %shuffle.i.bc.i309, i32 1
+  %tmp45 = bitcast i64 %shuffle.i.extract.i310 to <8 x i8>
+  %tmp46 = bitcast i64 %shuffle.i27.extract.i311 to <8 x i8>
+  %vtbl2.i25.i313 = tail call <8 x i8> @llvm.arm.neon.vtbl2(<8 x i8> %tmp45, <8 x i8> %tmp46, <8 x i8> %vtbl2.i25.i)
+  store <8 x i8> %vtbl2.i25.i313, <8 x i8>* %addr2
+  ret void
+}
+
+; Check that we do not perform the transformation when there are several uses
+; of the result of the bitcast.
+; CHECK-LABEL: severalUses:
+; ARG1_VALlo is hard coded because we need to access the high part of d0,
+; i.e., s1, and we can't express that with filecheck.
+; CHECK: vld1.32 {[[ARG1_VALlo:d0]], [[ARG1_VALhi:d[0-9]+]]}, [r0]
+; CHECK-NEXT: vldr [[ARG2_VAL:d[0-9]+]], [r1]
+; s1 is actually 2 * ARG1_VALlo + 1, but we cannot express that with filecheck.
+; CHECK-NEXT: vmov [[REThi:r[0-9]+]], s1
+; We build the return value here. s0 is 2 * ARG1_VALlo.
+; CHECK-NEXT: vmov r0, s0
+; This copy is correct but actually useless. We should be able to clean it up.
+; CHECK-NEXT: vmov [[ARG1_VALloCPY:d[0-9]+]], r0, [[REThi]]
+; CHECK-NEXT: vtbl.8 [[RES:d[0-9]+]], {[[ARG1_VALloCPY]], [[ARG1_VALhi]]}, [[ARG2_VAL]]
+; CHECK-NEXT: vstr [[RES]], [r1]
+; CHECK-NEXT: mov r1, [[REThi]]
+; CHECK-NEXT: bx lr
+define i64 @severalUses(<2 x i64>* %addr, <8 x i8>* %addr2) {
+  %shuffle.i.bc.i309 = load <2 x i64>, <2 x i64>* %addr
+  %vtbl2.i25.i = load <8 x i8>, <8 x i8>* %addr2
+  %shuffle.i.extract.i310 = extractelement <2 x i64> %shuffle.i.bc.i309, i32 0
+  %shuffle.i27.extract.i311 = extractelement <2 x i64> %shuffle.i.bc.i309, i32 1
+  %tmp45 = bitcast i64 %shuffle.i.extract.i310 to <8 x i8>
+  %tmp46 = bitcast i64 %shuffle.i27.extract.i311 to <8 x i8>
+  %vtbl2.i25.i313 = tail call <8 x i8> @llvm.arm.neon.vtbl2(<8 x i8> %tmp45, <8 x i8> %tmp46, <8 x i8> %vtbl2.i25.i)
+  store <8 x i8> %vtbl2.i25.i313, <8 x i8>* %addr2
+  ret i64 %shuffle.i.extract.i310
+}