+SDValue PPCTargetLowering::DAGCombineFastRecip(SDValue Op,
+ DAGCombinerInfo &DCI) const {
+ if (DCI.isAfterLegalizeVectorOps())
+ return SDValue();
+
+ EVT VT = Op.getValueType();
+
+ if ((VT == MVT::f32 && Subtarget.hasFRES()) ||
+ (VT == MVT::f64 && Subtarget.hasFRE()) ||
+ (VT == MVT::v4f32 && Subtarget.hasAltivec()) ||
+ (VT == MVT::v2f64 && Subtarget.hasVSX())) {
+
+ // Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
+ // For the reciprocal, we need to find the zero of the function:
+ // F(X) = A X - 1 [which has a zero at X = 1/A]
+ // =>
+ // X_{i+1} = X_i (2 - A X_i) = X_i + X_i (1 - A X_i) [this second form
+ // does not require additional intermediate precision]
+
+ // Convergence is quadratic, so we essentially double the number of digits
+ // correct after every iteration. The minimum architected relative
+ // accuracy is 2^-5. When hasRecipPrec(), this is 2^-14. IEEE float has
+ // 23 digits and double has 52 digits.
+ int Iterations = Subtarget.hasRecipPrec() ? 1 : 3;
+ if (VT.getScalarType() == MVT::f64)
+ ++Iterations;
+
+ SelectionDAG &DAG = DCI.DAG;
+ SDLoc dl(Op);
+
+ SDValue FPOne =
+ DAG.getConstantFP(1.0, VT.getScalarType());
+ if (VT.isVector()) {
+ assert(VT.getVectorNumElements() == 4 &&
+ "Unknown vector type");
+ FPOne = DAG.getNode(ISD::BUILD_VECTOR, dl, VT,
+ FPOne, FPOne, FPOne, FPOne);
+ }
+
+ SDValue Est = DAG.getNode(PPCISD::FRE, dl, VT, Op);
+ DCI.AddToWorklist(Est.getNode());
+
+ // Newton iterations: Est = Est + Est (1 - Arg * Est)
+ for (int i = 0; i < Iterations; ++i) {
+ SDValue NewEst = DAG.getNode(ISD::FMUL, dl, VT, Op, Est);
+ DCI.AddToWorklist(NewEst.getNode());
+
+ NewEst = DAG.getNode(ISD::FSUB, dl, VT, FPOne, NewEst);
+ DCI.AddToWorklist(NewEst.getNode());
+
+ NewEst = DAG.getNode(ISD::FMUL, dl, VT, Est, NewEst);
+ DCI.AddToWorklist(NewEst.getNode());
+
+ Est = DAG.getNode(ISD::FADD, dl, VT, Est, NewEst);
+ DCI.AddToWorklist(Est.getNode());
+ }
+
+ return Est;
+ }
+
+ return SDValue();
+}
+
+SDValue PPCTargetLowering::DAGCombineFastRecipFSQRT(SDValue Op,
+ DAGCombinerInfo &DCI) const {
+ if (DCI.isAfterLegalizeVectorOps())
+ return SDValue();
+
+ EVT VT = Op.getValueType();
+
+ if ((VT == MVT::f32 && Subtarget.hasFRSQRTES()) ||
+ (VT == MVT::f64 && Subtarget.hasFRSQRTE()) ||
+ (VT == MVT::v4f32 && Subtarget.hasAltivec()) ||
+ (VT == MVT::v2f64 && Subtarget.hasVSX())) {
+
+ // Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
+ // For the reciprocal sqrt, we need to find the zero of the function:
+ // F(X) = 1/X^2 - A [which has a zero at X = 1/sqrt(A)]
+ // =>
+ // X_{i+1} = X_i (1.5 - A X_i^2 / 2)
+ // As a result, we precompute A/2 prior to the iteration loop.
+
+ // Convergence is quadratic, so we essentially double the number of digits
+ // correct after every iteration. The minimum architected relative
+ // accuracy is 2^-5. When hasRecipPrec(), this is 2^-14. IEEE float has
+ // 23 digits and double has 52 digits.
+ int Iterations = Subtarget.hasRecipPrec() ? 1 : 3;
+ if (VT.getScalarType() == MVT::f64)
+ ++Iterations;
+
+ SelectionDAG &DAG = DCI.DAG;
+ SDLoc dl(Op);
+
+ SDValue FPThreeHalves =
+ DAG.getConstantFP(1.5, VT.getScalarType());
+ if (VT.isVector()) {
+ assert(VT.getVectorNumElements() == 4 &&
+ "Unknown vector type");
+ FPThreeHalves = DAG.getNode(ISD::BUILD_VECTOR, dl, VT,
+ FPThreeHalves, FPThreeHalves,
+ FPThreeHalves, FPThreeHalves);
+ }
+
+ SDValue Est = DAG.getNode(PPCISD::FRSQRTE, dl, VT, Op);
+ DCI.AddToWorklist(Est.getNode());
+
+ // We now need 0.5*Arg which we can write as (1.5*Arg - Arg) so that
+ // this entire sequence requires only one FP constant.
+ SDValue HalfArg = DAG.getNode(ISD::FMUL, dl, VT, FPThreeHalves, Op);
+ DCI.AddToWorklist(HalfArg.getNode());
+
+ HalfArg = DAG.getNode(ISD::FSUB, dl, VT, HalfArg, Op);
+ DCI.AddToWorklist(HalfArg.getNode());
+
+ // Newton iterations: Est = Est * (1.5 - HalfArg * Est * Est)
+ for (int i = 0; i < Iterations; ++i) {
+ SDValue NewEst = DAG.getNode(ISD::FMUL, dl, VT, Est, Est);
+ DCI.AddToWorklist(NewEst.getNode());
+
+ NewEst = DAG.getNode(ISD::FMUL, dl, VT, HalfArg, NewEst);
+ DCI.AddToWorklist(NewEst.getNode());
+
+ NewEst = DAG.getNode(ISD::FSUB, dl, VT, FPThreeHalves, NewEst);
+ DCI.AddToWorklist(NewEst.getNode());
+
+ Est = DAG.getNode(ISD::FMUL, dl, VT, Est, NewEst);
+ DCI.AddToWorklist(Est.getNode());
+ }
+
+ return Est;
+ }
+
+ return SDValue();
+}
+
+static bool isConsecutiveLSLoc(SDValue Loc, EVT VT, LSBaseSDNode *Base,
+ unsigned Bytes, int Dist,
+ SelectionDAG &DAG) {
+ if (VT.getSizeInBits() / 8 != Bytes)
+ return false;
+
+ SDValue BaseLoc = Base->getBasePtr();
+ if (Loc.getOpcode() == ISD::FrameIndex) {
+ if (BaseLoc.getOpcode() != ISD::FrameIndex)
+ return false;
+ const MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
+ int FI = cast<FrameIndexSDNode>(Loc)->getIndex();
+ int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex();
+ int FS = MFI->getObjectSize(FI);
+ int BFS = MFI->getObjectSize(BFI);
+ if (FS != BFS || FS != (int)Bytes) return false;
+ return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Bytes);
+ }
+
+ // Handle X+C
+ if (DAG.isBaseWithConstantOffset(Loc) && Loc.getOperand(0) == BaseLoc &&
+ cast<ConstantSDNode>(Loc.getOperand(1))->getSExtValue() == Dist*Bytes)
+ return true;
+
+ const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+ const GlobalValue *GV1 = nullptr;
+ const GlobalValue *GV2 = nullptr;
+ int64_t Offset1 = 0;
+ int64_t Offset2 = 0;
+ bool isGA1 = TLI.isGAPlusOffset(Loc.getNode(), GV1, Offset1);
+ bool isGA2 = TLI.isGAPlusOffset(BaseLoc.getNode(), GV2, Offset2);
+ if (isGA1 && isGA2 && GV1 == GV2)
+ return Offset1 == (Offset2 + Dist*Bytes);
+ return false;
+}
+
+// Like SelectionDAG::isConsecutiveLoad, but also works for stores, and does
+// not enforce equality of the chain operands.
+static bool isConsecutiveLS(SDNode *N, LSBaseSDNode *Base,
+ unsigned Bytes, int Dist,
+ SelectionDAG &DAG) {
+ if (LSBaseSDNode *LS = dyn_cast<LSBaseSDNode>(N)) {
+ EVT VT = LS->getMemoryVT();
+ SDValue Loc = LS->getBasePtr();
+ return isConsecutiveLSLoc(Loc, VT, Base, Bytes, Dist, DAG);
+ }
+
+ if (N->getOpcode() == ISD::INTRINSIC_W_CHAIN) {
+ EVT VT;
+ switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
+ default: return false;
+ case Intrinsic::ppc_altivec_lvx:
+ case Intrinsic::ppc_altivec_lvxl:
+ VT = MVT::v4i32;
+ break;
+ case Intrinsic::ppc_altivec_lvebx:
+ VT = MVT::i8;
+ break;
+ case Intrinsic::ppc_altivec_lvehx:
+ VT = MVT::i16;
+ break;
+ case Intrinsic::ppc_altivec_lvewx:
+ VT = MVT::i32;
+ break;
+ }
+
+ return isConsecutiveLSLoc(N->getOperand(2), VT, Base, Bytes, Dist, DAG);
+ }
+
+ if (N->getOpcode() == ISD::INTRINSIC_VOID) {
+ EVT VT;
+ switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
+ default: return false;
+ case Intrinsic::ppc_altivec_stvx:
+ case Intrinsic::ppc_altivec_stvxl:
+ VT = MVT::v4i32;
+ break;
+ case Intrinsic::ppc_altivec_stvebx:
+ VT = MVT::i8;
+ break;
+ case Intrinsic::ppc_altivec_stvehx:
+ VT = MVT::i16;
+ break;
+ case Intrinsic::ppc_altivec_stvewx:
+ VT = MVT::i32;
+ break;
+ }
+
+ return isConsecutiveLSLoc(N->getOperand(3), VT, Base, Bytes, Dist, DAG);
+ }
+
+ return false;
+}
+
+// Return true is there is a nearyby consecutive load to the one provided
+// (regardless of alignment). We search up and down the chain, looking though
+// token factors and other loads (but nothing else). As a result, a true result
+// indicates that it is safe to create a new consecutive load adjacent to the
+// load provided.
+static bool findConsecutiveLoad(LoadSDNode *LD, SelectionDAG &DAG) {
+ SDValue Chain = LD->getChain();
+ EVT VT = LD->getMemoryVT();
+
+ SmallSet<SDNode *, 16> LoadRoots;
+ SmallVector<SDNode *, 8> Queue(1, Chain.getNode());
+ SmallSet<SDNode *, 16> Visited;
+
+ // First, search up the chain, branching to follow all token-factor operands.
+ // If we find a consecutive load, then we're done, otherwise, record all
+ // nodes just above the top-level loads and token factors.
+ while (!Queue.empty()) {
+ SDNode *ChainNext = Queue.pop_back_val();
+ if (!Visited.insert(ChainNext))
+ continue;
+
+ if (MemSDNode *ChainLD = dyn_cast<MemSDNode>(ChainNext)) {
+ if (isConsecutiveLS(ChainLD, LD, VT.getStoreSize(), 1, DAG))
+ return true;
+
+ if (!Visited.count(ChainLD->getChain().getNode()))
+ Queue.push_back(ChainLD->getChain().getNode());
+ } else if (ChainNext->getOpcode() == ISD::TokenFactor) {
+ for (const SDUse &O : ChainNext->ops())
+ if (!Visited.count(O.getNode()))
+ Queue.push_back(O.getNode());
+ } else
+ LoadRoots.insert(ChainNext);
+ }
+
+ // Second, search down the chain, starting from the top-level nodes recorded
+ // in the first phase. These top-level nodes are the nodes just above all
+ // loads and token factors. Starting with their uses, recursively look though
+ // all loads (just the chain uses) and token factors to find a consecutive
+ // load.
+ Visited.clear();
+ Queue.clear();
+
+ for (SmallSet<SDNode *, 16>::iterator I = LoadRoots.begin(),
+ IE = LoadRoots.end(); I != IE; ++I) {
+ Queue.push_back(*I);
+
+ while (!Queue.empty()) {
+ SDNode *LoadRoot = Queue.pop_back_val();
+ if (!Visited.insert(LoadRoot))
+ continue;
+
+ if (MemSDNode *ChainLD = dyn_cast<MemSDNode>(LoadRoot))
+ if (isConsecutiveLS(ChainLD, LD, VT.getStoreSize(), 1, DAG))
+ return true;
+
+ for (SDNode::use_iterator UI = LoadRoot->use_begin(),
+ UE = LoadRoot->use_end(); UI != UE; ++UI)
+ if (((isa<MemSDNode>(*UI) &&
+ cast<MemSDNode>(*UI)->getChain().getNode() == LoadRoot) ||
+ UI->getOpcode() == ISD::TokenFactor) && !Visited.count(*UI))
+ Queue.push_back(*UI);
+ }
+ }
+
+ return false;
+}
+
+SDValue PPCTargetLowering::DAGCombineTruncBoolExt(SDNode *N,
+ DAGCombinerInfo &DCI) const {
+ SelectionDAG &DAG = DCI.DAG;
+ SDLoc dl(N);
+
+ assert(Subtarget.useCRBits() &&
+ "Expecting to be tracking CR bits");
+ // If we're tracking CR bits, we need to be careful that we don't have:
+ // trunc(binary-ops(zext(x), zext(y)))
+ // or
+ // trunc(binary-ops(binary-ops(zext(x), zext(y)), ...)
+ // such that we're unnecessarily moving things into GPRs when it would be
+ // better to keep them in CR bits.
+
+ // Note that trunc here can be an actual i1 trunc, or can be the effective
+ // truncation that comes from a setcc or select_cc.
+ if (N->getOpcode() == ISD::TRUNCATE &&
+ N->getValueType(0) != MVT::i1)
+ return SDValue();
+
+ if (N->getOperand(0).getValueType() != MVT::i32 &&
+ N->getOperand(0).getValueType() != MVT::i64)
+ return SDValue();
+
+ if (N->getOpcode() == ISD::SETCC ||
+ N->getOpcode() == ISD::SELECT_CC) {
+ // If we're looking at a comparison, then we need to make sure that the
+ // high bits (all except for the first) don't matter the result.
+ ISD::CondCode CC =
+ cast<CondCodeSDNode>(N->getOperand(
+ N->getOpcode() == ISD::SETCC ? 2 : 4))->get();
+ unsigned OpBits = N->getOperand(0).getValueSizeInBits();
+
+ if (ISD::isSignedIntSetCC(CC)) {
+ if (DAG.ComputeNumSignBits(N->getOperand(0)) != OpBits ||
+ DAG.ComputeNumSignBits(N->getOperand(1)) != OpBits)
+ return SDValue();
+ } else if (ISD::isUnsignedIntSetCC(CC)) {
+ if (!DAG.MaskedValueIsZero(N->getOperand(0),
+ APInt::getHighBitsSet(OpBits, OpBits-1)) ||
+ !DAG.MaskedValueIsZero(N->getOperand(1),
+ APInt::getHighBitsSet(OpBits, OpBits-1)))
+ return SDValue();
+ } else {
+ // This is neither a signed nor an unsigned comparison, just make sure
+ // that the high bits are equal.
+ APInt Op1Zero, Op1One;
+ APInt Op2Zero, Op2One;
+ DAG.computeKnownBits(N->getOperand(0), Op1Zero, Op1One);
+ DAG.computeKnownBits(N->getOperand(1), Op2Zero, Op2One);
+
+ // We don't really care about what is known about the first bit (if
+ // anything), so clear it in all masks prior to comparing them.
+ Op1Zero.clearBit(0); Op1One.clearBit(0);
+ Op2Zero.clearBit(0); Op2One.clearBit(0);
+
+ if (Op1Zero != Op2Zero || Op1One != Op2One)
+ return SDValue();
+ }
+ }
+
+ // We now know that the higher-order bits are irrelevant, we just need to
+ // make sure that all of the intermediate operations are bit operations, and
+ // all inputs are extensions.
+ if (N->getOperand(0).getOpcode() != ISD::AND &&
+ N->getOperand(0).getOpcode() != ISD::OR &&
+ N->getOperand(0).getOpcode() != ISD::XOR &&
+ N->getOperand(0).getOpcode() != ISD::SELECT &&
+ N->getOperand(0).getOpcode() != ISD::SELECT_CC &&
+ N->getOperand(0).getOpcode() != ISD::TRUNCATE &&
+ N->getOperand(0).getOpcode() != ISD::SIGN_EXTEND &&
+ N->getOperand(0).getOpcode() != ISD::ZERO_EXTEND &&
+ N->getOperand(0).getOpcode() != ISD::ANY_EXTEND)
+ return SDValue();
+
+ if ((N->getOpcode() == ISD::SETCC || N->getOpcode() == ISD::SELECT_CC) &&
+ N->getOperand(1).getOpcode() != ISD::AND &&
+ N->getOperand(1).getOpcode() != ISD::OR &&
+ N->getOperand(1).getOpcode() != ISD::XOR &&
+ N->getOperand(1).getOpcode() != ISD::SELECT &&
+ N->getOperand(1).getOpcode() != ISD::SELECT_CC &&
+ N->getOperand(1).getOpcode() != ISD::TRUNCATE &&
+ N->getOperand(1).getOpcode() != ISD::SIGN_EXTEND &&
+ N->getOperand(1).getOpcode() != ISD::ZERO_EXTEND &&
+ N->getOperand(1).getOpcode() != ISD::ANY_EXTEND)
+ return SDValue();
+
+ SmallVector<SDValue, 4> Inputs;
+ SmallVector<SDValue, 8> BinOps, PromOps;
+ SmallPtrSet<SDNode *, 16> Visited;
+
+ for (unsigned i = 0; i < 2; ++i) {
+ if (((N->getOperand(i).getOpcode() == ISD::SIGN_EXTEND ||
+ N->getOperand(i).getOpcode() == ISD::ZERO_EXTEND ||
+ N->getOperand(i).getOpcode() == ISD::ANY_EXTEND) &&
+ N->getOperand(i).getOperand(0).getValueType() == MVT::i1) ||
+ isa<ConstantSDNode>(N->getOperand(i)))
+ Inputs.push_back(N->getOperand(i));
+ else
+ BinOps.push_back(N->getOperand(i));
+
+ if (N->getOpcode() == ISD::TRUNCATE)
+ break;
+ }
+
+ // Visit all inputs, collect all binary operations (and, or, xor and
+ // select) that are all fed by extensions.
+ while (!BinOps.empty()) {
+ SDValue BinOp = BinOps.back();
+ BinOps.pop_back();
+
+ if (!Visited.insert(BinOp.getNode()))
+ continue;
+
+ PromOps.push_back(BinOp);
+
+ for (unsigned i = 0, ie = BinOp.getNumOperands(); i != ie; ++i) {
+ // The condition of the select is not promoted.
+ if (BinOp.getOpcode() == ISD::SELECT && i == 0)
+ continue;
+ if (BinOp.getOpcode() == ISD::SELECT_CC && i != 2 && i != 3)
+ continue;
+
+ if (((BinOp.getOperand(i).getOpcode() == ISD::SIGN_EXTEND ||
+ BinOp.getOperand(i).getOpcode() == ISD::ZERO_EXTEND ||
+ BinOp.getOperand(i).getOpcode() == ISD::ANY_EXTEND) &&
+ BinOp.getOperand(i).getOperand(0).getValueType() == MVT::i1) ||
+ isa<ConstantSDNode>(BinOp.getOperand(i))) {
+ Inputs.push_back(BinOp.getOperand(i));
+ } else if (BinOp.getOperand(i).getOpcode() == ISD::AND ||
+ BinOp.getOperand(i).getOpcode() == ISD::OR ||
+ BinOp.getOperand(i).getOpcode() == ISD::XOR ||
+ BinOp.getOperand(i).getOpcode() == ISD::SELECT ||
+ BinOp.getOperand(i).getOpcode() == ISD::SELECT_CC ||
+ BinOp.getOperand(i).getOpcode() == ISD::TRUNCATE ||
+ BinOp.getOperand(i).getOpcode() == ISD::SIGN_EXTEND ||
+ BinOp.getOperand(i).getOpcode() == ISD::ZERO_EXTEND ||
+ BinOp.getOperand(i).getOpcode() == ISD::ANY_EXTEND) {
+ BinOps.push_back(BinOp.getOperand(i));
+ } else {
+ // We have an input that is not an extension or another binary
+ // operation; we'll abort this transformation.
+ return SDValue();
+ }
+ }
+ }
+
+ // Make sure that this is a self-contained cluster of operations (which
+ // is not quite the same thing as saying that everything has only one
+ // use).
+ for (unsigned i = 0, ie = Inputs.size(); i != ie; ++i) {
+ if (isa<ConstantSDNode>(Inputs[i]))
+ continue;
+
+ for (SDNode::use_iterator UI = Inputs[i].getNode()->use_begin(),
+ UE = Inputs[i].getNode()->use_end();
+ UI != UE; ++UI) {
+ SDNode *User = *UI;
+ if (User != N && !Visited.count(User))
+ return SDValue();
+
+ // Make sure that we're not going to promote the non-output-value
+ // operand(s) or SELECT or SELECT_CC.
+ // FIXME: Although we could sometimes handle this, and it does occur in
+ // practice that one of the condition inputs to the select is also one of
+ // the outputs, we currently can't deal with this.
+ if (User->getOpcode() == ISD::SELECT) {
+ if (User->getOperand(0) == Inputs[i])
+ return SDValue();
+ } else if (User->getOpcode() == ISD::SELECT_CC) {
+ if (User->getOperand(0) == Inputs[i] ||
+ User->getOperand(1) == Inputs[i])
+ return SDValue();
+ }
+ }
+ }
+
+ for (unsigned i = 0, ie = PromOps.size(); i != ie; ++i) {
+ for (SDNode::use_iterator UI = PromOps[i].getNode()->use_begin(),
+ UE = PromOps[i].getNode()->use_end();
+ UI != UE; ++UI) {
+ SDNode *User = *UI;
+ if (User != N && !Visited.count(User))
+ return SDValue();
+
+ // Make sure that we're not going to promote the non-output-value
+ // operand(s) or SELECT or SELECT_CC.
+ // FIXME: Although we could sometimes handle this, and it does occur in
+ // practice that one of the condition inputs to the select is also one of
+ // the outputs, we currently can't deal with this.
+ if (User->getOpcode() == ISD::SELECT) {
+ if (User->getOperand(0) == PromOps[i])
+ return SDValue();
+ } else if (User->getOpcode() == ISD::SELECT_CC) {
+ if (User->getOperand(0) == PromOps[i] ||
+ User->getOperand(1) == PromOps[i])
+ return SDValue();
+ }
+ }
+ }
+
+ // Replace all inputs with the extension operand.
+ for (unsigned i = 0, ie = Inputs.size(); i != ie; ++i) {
+ // Constants may have users outside the cluster of to-be-promoted nodes,
+ // and so we need to replace those as we do the promotions.
+ if (isa<ConstantSDNode>(Inputs[i]))
+ continue;
+ else
+ DAG.ReplaceAllUsesOfValueWith(Inputs[i], Inputs[i].getOperand(0));
+ }
+
+ // Replace all operations (these are all the same, but have a different
+ // (i1) return type). DAG.getNode will validate that the types of
+ // a binary operator match, so go through the list in reverse so that
+ // we've likely promoted both operands first. Any intermediate truncations or
+ // extensions disappear.
+ while (!PromOps.empty()) {
+ SDValue PromOp = PromOps.back();
+ PromOps.pop_back();
+
+ if (PromOp.getOpcode() == ISD::TRUNCATE ||
+ PromOp.getOpcode() == ISD::SIGN_EXTEND ||
+ PromOp.getOpcode() == ISD::ZERO_EXTEND ||
+ PromOp.getOpcode() == ISD::ANY_EXTEND) {
+ if (!isa<ConstantSDNode>(PromOp.getOperand(0)) &&
+ PromOp.getOperand(0).getValueType() != MVT::i1) {
+ // The operand is not yet ready (see comment below).
+ PromOps.insert(PromOps.begin(), PromOp);
+ continue;
+ }
+
+ SDValue RepValue = PromOp.getOperand(0);
+ if (isa<ConstantSDNode>(RepValue))
+ RepValue = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, RepValue);
+
+ DAG.ReplaceAllUsesOfValueWith(PromOp, RepValue);
+ continue;
+ }
+
+ unsigned C;
+ switch (PromOp.getOpcode()) {
+ default: C = 0; break;
+ case ISD::SELECT: C = 1; break;
+ case ISD::SELECT_CC: C = 2; break;
+ }
+
+ if ((!isa<ConstantSDNode>(PromOp.getOperand(C)) &&
+ PromOp.getOperand(C).getValueType() != MVT::i1) ||
+ (!isa<ConstantSDNode>(PromOp.getOperand(C+1)) &&
+ PromOp.getOperand(C+1).getValueType() != MVT::i1)) {
+ // The to-be-promoted operands of this node have not yet been
+ // promoted (this should be rare because we're going through the
+ // list backward, but if one of the operands has several users in
+ // this cluster of to-be-promoted nodes, it is possible).
+ PromOps.insert(PromOps.begin(), PromOp);
+ continue;
+ }
+
+ SmallVector<SDValue, 3> Ops(PromOp.getNode()->op_begin(),
+ PromOp.getNode()->op_end());
+
+ // If there are any constant inputs, make sure they're replaced now.
+ for (unsigned i = 0; i < 2; ++i)
+ if (isa<ConstantSDNode>(Ops[C+i]))
+ Ops[C+i] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Ops[C+i]);
+
+ DAG.ReplaceAllUsesOfValueWith(PromOp,
+ DAG.getNode(PromOp.getOpcode(), dl, MVT::i1, Ops));
+ }
+
+ // Now we're left with the initial truncation itself.
+ if (N->getOpcode() == ISD::TRUNCATE)
+ return N->getOperand(0);
+
+ // Otherwise, this is a comparison. The operands to be compared have just
+ // changed type (to i1), but everything else is the same.
+ return SDValue(N, 0);
+}
+
+SDValue PPCTargetLowering::DAGCombineExtBoolTrunc(SDNode *N,
+ DAGCombinerInfo &DCI) const {
+ SelectionDAG &DAG = DCI.DAG;
+ SDLoc dl(N);
+
+ // If we're tracking CR bits, we need to be careful that we don't have:
+ // zext(binary-ops(trunc(x), trunc(y)))
+ // or
+ // zext(binary-ops(binary-ops(trunc(x), trunc(y)), ...)
+ // such that we're unnecessarily moving things into CR bits that can more
+ // efficiently stay in GPRs. Note that if we're not certain that the high
+ // bits are set as required by the final extension, we still may need to do
+ // some masking to get the proper behavior.
+
+ // This same functionality is important on PPC64 when dealing with
+ // 32-to-64-bit extensions; these occur often when 32-bit values are used as
+ // the return values of functions. Because it is so similar, it is handled
+ // here as well.
+
+ if (N->getValueType(0) != MVT::i32 &&
+ N->getValueType(0) != MVT::i64)
+ return SDValue();
+
+ if (!((N->getOperand(0).getValueType() == MVT::i1 &&
+ Subtarget.useCRBits()) ||
+ (N->getOperand(0).getValueType() == MVT::i32 &&
+ Subtarget.isPPC64())))
+ return SDValue();
+
+ if (N->getOperand(0).getOpcode() != ISD::AND &&
+ N->getOperand(0).getOpcode() != ISD::OR &&
+ N->getOperand(0).getOpcode() != ISD::XOR &&
+ N->getOperand(0).getOpcode() != ISD::SELECT &&
+ N->getOperand(0).getOpcode() != ISD::SELECT_CC)
+ return SDValue();
+
+ SmallVector<SDValue, 4> Inputs;
+ SmallVector<SDValue, 8> BinOps(1, N->getOperand(0)), PromOps;
+ SmallPtrSet<SDNode *, 16> Visited;
+
+ // Visit all inputs, collect all binary operations (and, or, xor and
+ // select) that are all fed by truncations.
+ while (!BinOps.empty()) {
+ SDValue BinOp = BinOps.back();
+ BinOps.pop_back();
+
+ if (!Visited.insert(BinOp.getNode()))
+ continue;
+
+ PromOps.push_back(BinOp);
+
+ for (unsigned i = 0, ie = BinOp.getNumOperands(); i != ie; ++i) {
+ // The condition of the select is not promoted.
+ if (BinOp.getOpcode() == ISD::SELECT && i == 0)
+ continue;
+ if (BinOp.getOpcode() == ISD::SELECT_CC && i != 2 && i != 3)
+ continue;
+
+ if (BinOp.getOperand(i).getOpcode() == ISD::TRUNCATE ||
+ isa<ConstantSDNode>(BinOp.getOperand(i))) {
+ Inputs.push_back(BinOp.getOperand(i));
+ } else if (BinOp.getOperand(i).getOpcode() == ISD::AND ||
+ BinOp.getOperand(i).getOpcode() == ISD::OR ||
+ BinOp.getOperand(i).getOpcode() == ISD::XOR ||
+ BinOp.getOperand(i).getOpcode() == ISD::SELECT ||
+ BinOp.getOperand(i).getOpcode() == ISD::SELECT_CC) {
+ BinOps.push_back(BinOp.getOperand(i));
+ } else {
+ // We have an input that is not a truncation or another binary
+ // operation; we'll abort this transformation.
+ return SDValue();
+ }
+ }
+ }
+
+ // Make sure that this is a self-contained cluster of operations (which
+ // is not quite the same thing as saying that everything has only one
+ // use).
+ for (unsigned i = 0, ie = Inputs.size(); i != ie; ++i) {
+ if (isa<ConstantSDNode>(Inputs[i]))
+ continue;
+
+ for (SDNode::use_iterator UI = Inputs[i].getNode()->use_begin(),
+ UE = Inputs[i].getNode()->use_end();
+ UI != UE; ++UI) {
+ SDNode *User = *UI;
+ if (User != N && !Visited.count(User))
+ return SDValue();
+
+ // Make sure that we're not going to promote the non-output-value
+ // operand(s) or SELECT or SELECT_CC.
+ // FIXME: Although we could sometimes handle this, and it does occur in
+ // practice that one of the condition inputs to the select is also one of
+ // the outputs, we currently can't deal with this.
+ if (User->getOpcode() == ISD::SELECT) {
+ if (User->getOperand(0) == Inputs[i])
+ return SDValue();
+ } else if (User->getOpcode() == ISD::SELECT_CC) {
+ if (User->getOperand(0) == Inputs[i] ||
+ User->getOperand(1) == Inputs[i])
+ return SDValue();
+ }
+ }
+ }
+
+ for (unsigned i = 0, ie = PromOps.size(); i != ie; ++i) {
+ for (SDNode::use_iterator UI = PromOps[i].getNode()->use_begin(),
+ UE = PromOps[i].getNode()->use_end();
+ UI != UE; ++UI) {
+ SDNode *User = *UI;
+ if (User != N && !Visited.count(User))
+ return SDValue();
+
+ // Make sure that we're not going to promote the non-output-value
+ // operand(s) or SELECT or SELECT_CC.
+ // FIXME: Although we could sometimes handle this, and it does occur in
+ // practice that one of the condition inputs to the select is also one of
+ // the outputs, we currently can't deal with this.
+ if (User->getOpcode() == ISD::SELECT) {
+ if (User->getOperand(0) == PromOps[i])
+ return SDValue();
+ } else if (User->getOpcode() == ISD::SELECT_CC) {
+ if (User->getOperand(0) == PromOps[i] ||
+ User->getOperand(1) == PromOps[i])
+ return SDValue();
+ }
+ }
+ }
+
+ unsigned PromBits = N->getOperand(0).getValueSizeInBits();
+ bool ReallyNeedsExt = false;
+ if (N->getOpcode() != ISD::ANY_EXTEND) {
+ // If all of the inputs are not already sign/zero extended, then
+ // we'll still need to do that at the end.
+ for (unsigned i = 0, ie = Inputs.size(); i != ie; ++i) {
+ if (isa<ConstantSDNode>(Inputs[i]))
+ continue;
+
+ unsigned OpBits =
+ Inputs[i].getOperand(0).getValueSizeInBits();
+ assert(PromBits < OpBits && "Truncation not to a smaller bit count?");
+
+ if ((N->getOpcode() == ISD::ZERO_EXTEND &&
+ !DAG.MaskedValueIsZero(Inputs[i].getOperand(0),
+ APInt::getHighBitsSet(OpBits,
+ OpBits-PromBits))) ||
+ (N->getOpcode() == ISD::SIGN_EXTEND &&
+ DAG.ComputeNumSignBits(Inputs[i].getOperand(0)) <
+ (OpBits-(PromBits-1)))) {
+ ReallyNeedsExt = true;
+ break;
+ }
+ }
+ }
+
+ // Replace all inputs, either with the truncation operand, or a
+ // truncation or extension to the final output type.
+ for (unsigned i = 0, ie = Inputs.size(); i != ie; ++i) {
+ // Constant inputs need to be replaced with the to-be-promoted nodes that
+ // use them because they might have users outside of the cluster of
+ // promoted nodes.
+ if (isa<ConstantSDNode>(Inputs[i]))
+ continue;
+
+ SDValue InSrc = Inputs[i].getOperand(0);
+ if (Inputs[i].getValueType() == N->getValueType(0))
+ DAG.ReplaceAllUsesOfValueWith(Inputs[i], InSrc);
+ else if (N->getOpcode() == ISD::SIGN_EXTEND)
+ DAG.ReplaceAllUsesOfValueWith(Inputs[i],
+ DAG.getSExtOrTrunc(InSrc, dl, N->getValueType(0)));
+ else if (N->getOpcode() == ISD::ZERO_EXTEND)
+ DAG.ReplaceAllUsesOfValueWith(Inputs[i],
+ DAG.getZExtOrTrunc(InSrc, dl, N->getValueType(0)));
+ else
+ DAG.ReplaceAllUsesOfValueWith(Inputs[i],
+ DAG.getAnyExtOrTrunc(InSrc, dl, N->getValueType(0)));
+ }
+
+ // Replace all operations (these are all the same, but have a different
+ // (promoted) return type). DAG.getNode will validate that the types of
+ // a binary operator match, so go through the list in reverse so that
+ // we've likely promoted both operands first.
+ while (!PromOps.empty()) {
+ SDValue PromOp = PromOps.back();
+ PromOps.pop_back();
+
+ unsigned C;
+ switch (PromOp.getOpcode()) {
+ default: C = 0; break;
+ case ISD::SELECT: C = 1; break;
+ case ISD::SELECT_CC: C = 2; break;
+ }
+
+ if ((!isa<ConstantSDNode>(PromOp.getOperand(C)) &&
+ PromOp.getOperand(C).getValueType() != N->getValueType(0)) ||
+ (!isa<ConstantSDNode>(PromOp.getOperand(C+1)) &&
+ PromOp.getOperand(C+1).getValueType() != N->getValueType(0))) {
+ // The to-be-promoted operands of this node have not yet been
+ // promoted (this should be rare because we're going through the
+ // list backward, but if one of the operands has several users in
+ // this cluster of to-be-promoted nodes, it is possible).
+ PromOps.insert(PromOps.begin(), PromOp);
+ continue;
+ }
+
+ SmallVector<SDValue, 3> Ops(PromOp.getNode()->op_begin(),
+ PromOp.getNode()->op_end());
+
+ // If this node has constant inputs, then they'll need to be promoted here.
+ for (unsigned i = 0; i < 2; ++i) {
+ if (!isa<ConstantSDNode>(Ops[C+i]))
+ continue;
+ if (Ops[C+i].getValueType() == N->getValueType(0))
+ continue;
+
+ if (N->getOpcode() == ISD::SIGN_EXTEND)
+ Ops[C+i] = DAG.getSExtOrTrunc(Ops[C+i], dl, N->getValueType(0));
+ else if (N->getOpcode() == ISD::ZERO_EXTEND)
+ Ops[C+i] = DAG.getZExtOrTrunc(Ops[C+i], dl, N->getValueType(0));
+ else
+ Ops[C+i] = DAG.getAnyExtOrTrunc(Ops[C+i], dl, N->getValueType(0));
+ }
+
+ DAG.ReplaceAllUsesOfValueWith(PromOp,
+ DAG.getNode(PromOp.getOpcode(), dl, N->getValueType(0), Ops));
+ }
+
+ // Now we're left with the initial extension itself.
+ if (!ReallyNeedsExt)
+ return N->getOperand(0);
+
+ // To zero extend, just mask off everything except for the first bit (in the
+ // i1 case).
+ if (N->getOpcode() == ISD::ZERO_EXTEND)
+ return DAG.getNode(ISD::AND, dl, N->getValueType(0), N->getOperand(0),
+ DAG.getConstant(APInt::getLowBitsSet(
+ N->getValueSizeInBits(0), PromBits),
+ N->getValueType(0)));
+
+ assert(N->getOpcode() == ISD::SIGN_EXTEND &&
+ "Invalid extension type");
+ EVT ShiftAmountTy = getShiftAmountTy(N->getValueType(0));
+ SDValue ShiftCst =
+ DAG.getConstant(N->getValueSizeInBits(0)-PromBits, ShiftAmountTy);
+ return DAG.getNode(ISD::SRA, dl, N->getValueType(0),
+ DAG.getNode(ISD::SHL, dl, N->getValueType(0),
+ N->getOperand(0), ShiftCst), ShiftCst);
+}
+