For NEON vectors with 32- or 64-bit elements, select BUILD_VECTORs and
VECTOR_SHUFFLEs to REG_SEQUENCE instructions. The standard ISD::BUILD_VECTOR
node corresponds closely to REG_SEQUENCE but I couldn't use it here because
its operands do not get legalized. That is pretty awful, but I guess it
makes sense for other targets. Instead, I have added an ARM-specific version
of BUILD_VECTOR that will have its operands properly legalized.
This fixes the rest of Radar 7872877.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@105439 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/ARM/ARMISelDAGToDAG.cpp b/lib/Target/ARM/ARMISelDAGToDAG.cpp
index 8659d85..d7c2fe0 100644
--- a/lib/Target/ARM/ARMISelDAGToDAG.cpp
+++ b/lib/Target/ARM/ARMISelDAGToDAG.cpp
@@ -174,24 +174,17 @@
char ConstraintCode,
std::vector<SDValue> &OutOps);
- /// PairDRegs - Form a quad register from a pair of D registers.
- ///
+ // Form pairs of consecutive S, D, or Q registers.
+ SDNode *PairSRegs(EVT VT, SDValue V0, SDValue V1);
SDNode *PairDRegs(EVT VT, SDValue V0, SDValue V1);
-
- /// PairDRegs - Form a quad register pair from a pair of Q registers.
- ///
SDNode *PairQRegs(EVT VT, SDValue V0, SDValue V1);
- /// QuadDRegs - Form a quad register pair from a quad of D registers.
- ///
+ // Form sequences of 4 consecutive S, D, or Q registers.
+ SDNode *QuadSRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
SDNode *QuadDRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
-
- /// QuadQRegs - Form 4 consecutive Q registers.
- ///
SDNode *QuadQRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3);
- /// OctoDRegs - Form 8 consecutive D registers.
- ///
+ // Form sequences of 8 consecutive D registers.
SDNode *OctoDRegs(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3,
SDValue V4, SDValue V5, SDValue V6, SDValue V7);
};
@@ -963,6 +956,24 @@
return NULL;
}
+/// PairSRegs - Form a D register from a pair of S registers.
+///
+SDNode *ARMDAGToDAGISel::PairSRegs(EVT VT, SDValue V0, SDValue V1) {
+ DebugLoc dl = V0.getNode()->getDebugLoc();
+ SDValue SubReg0 = CurDAG->getTargetConstant(ARM::ssub_0, MVT::i32);
+ SDValue SubReg1 = CurDAG->getTargetConstant(ARM::ssub_1, MVT::i32);
+ if (llvm::ModelWithRegSequence()) {
+ const SDValue Ops[] = { V0, SubReg0, V1, SubReg1 };
+ return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
+ }
+ SDValue Undef =
+ SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, dl, VT), 0);
+ SDNode *Pair = CurDAG->getMachineNode(TargetOpcode::INSERT_SUBREG, dl,
+ VT, Undef, V0, SubReg0);
+ return CurDAG->getMachineNode(TargetOpcode::INSERT_SUBREG, dl,
+ VT, SDValue(Pair, 0), V1, SubReg1);
+}
+
/// PairDRegs - Form a quad register from a pair of D registers.
///
SDNode *ARMDAGToDAGISel::PairDRegs(EVT VT, SDValue V0, SDValue V1) {
@@ -991,6 +1002,19 @@
return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 4);
}
+/// QuadSRegs - Form 4 consecutive S registers.
+///
+SDNode *ARMDAGToDAGISel::QuadSRegs(EVT VT, SDValue V0, SDValue V1,
+ SDValue V2, SDValue V3) {
+ DebugLoc dl = V0.getNode()->getDebugLoc();
+ SDValue SubReg0 = CurDAG->getTargetConstant(ARM::ssub_0, MVT::i32);
+ SDValue SubReg1 = CurDAG->getTargetConstant(ARM::ssub_1, MVT::i32);
+ SDValue SubReg2 = CurDAG->getTargetConstant(ARM::ssub_2, MVT::i32);
+ SDValue SubReg3 = CurDAG->getTargetConstant(ARM::ssub_3, MVT::i32);
+ const SDValue Ops[] = { V0, SubReg0, V1, SubReg1, V2, SubReg2, V3, SubReg3 };
+ return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, dl, VT, Ops, 8);
+}
+
/// QuadDRegs - Form 4 consecutive D registers.
///
SDNode *ARMDAGToDAGISel::QuadDRegs(EVT VT, SDValue V0, SDValue V1,
@@ -2214,6 +2238,22 @@
SDValue Ops[] = { N->getOperand(0), N->getOperand(1), Pred, PredReg };
return CurDAG->getMachineNode(Opc, dl, VT, VT, Ops, 4);
}
+ case ARMISD::BUILD_VECTOR: {
+ EVT VecVT = N->getValueType(0);
+ EVT EltVT = VecVT.getVectorElementType();
+ unsigned NumElts = VecVT.getVectorNumElements();
+ if (EltVT.getSimpleVT() == MVT::f64) {
+ assert(NumElts == 2 && "unexpected type for BUILD_VECTOR");
+ return PairDRegs(VecVT, N->getOperand(0), N->getOperand(1));
+ }
+ assert(EltVT.getSimpleVT() == MVT::f32 &&
+ "unexpected type for BUILD_VECTOR");
+ if (NumElts == 2)
+ return PairSRegs(VecVT, N->getOperand(0), N->getOperand(1));
+ assert(NumElts == 4 && "unexpected type for BUILD_VECTOR");
+ return QuadSRegs(VecVT, N->getOperand(0), N->getOperand(1),
+ N->getOperand(2), N->getOperand(3));
+ }
case ISD::INTRINSIC_VOID:
case ISD::INTRINSIC_W_CHAIN: {